diff --git a/.gitattributes b/.gitattributes index 4ba88ff626c37291ecfec397725f7150be6a5a42..0c703e813ea35a94b633c99a159d1d360399b5f3 100644 --- a/.gitattributes +++ b/.gitattributes @@ -40,3 +40,5 @@ fairseq/data/data_utils_fast.cpython-38-darwin.so filter=lfs diff=lfs merge=lfs fairseq/data/token_block_utils_fast.cpython-310-darwin.so filter=lfs diff=lfs merge=lfs -text fairseq/data/token_block_utils_fast.cpython-36m-darwin.so filter=lfs diff=lfs merge=lfs -text fairseq/data/token_block_utils_fast.cpython-38-darwin.so filter=lfs diff=lfs merge=lfs -text +fairseq/data/data_utils_fast.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text +fairseq/data/token_block_utils_fast.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text diff --git a/fairseq/__init__.py b/fairseq/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3dd29637af82124c6f90f23808908c8a884cf88e --- /dev/null +++ b/fairseq/__init__.py @@ -0,0 +1,26 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +__all__ = ['pdb'] +__version__ = '0.9.0' + +import sys + +# backwards compatibility to support `from fairseq.meters import AverageMeter` +from fairseq.logging import meters, metrics, progress_bar # noqa +sys.modules['fairseq.meters'] = meters +sys.modules['fairseq.metrics'] = metrics +sys.modules['fairseq.progress_bar'] = progress_bar + +import fairseq.criterions # noqa +import fairseq.models # noqa +import fairseq.modules # noqa +import fairseq.optim # noqa +import fairseq.optim.lr_scheduler # noqa +import fairseq.pdb # noqa +import fairseq.tasks # noqa + +import fairseq.benchmark # noqa +import fairseq.model_parallel # noqa diff --git a/fairseq/__pycache__/__init__.cpython-310.pyc b/fairseq/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b8241ad57ba17aea696541e222165f0790728443 Binary files /dev/null and b/fairseq/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/__pycache__/binarizer.cpython-310.pyc b/fairseq/__pycache__/binarizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8ea1931f58fd8c52e625c0779d761e8a7b36b7b7 Binary files /dev/null and b/fairseq/__pycache__/binarizer.cpython-310.pyc differ diff --git a/fairseq/__pycache__/checkpoint_utils.cpython-310.pyc b/fairseq/__pycache__/checkpoint_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e92332763ecbb6a3bfd414775e19138e9f75f8b3 Binary files /dev/null and b/fairseq/__pycache__/checkpoint_utils.cpython-310.pyc differ diff --git a/fairseq/__pycache__/distributed_utils.cpython-310.pyc b/fairseq/__pycache__/distributed_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2b41f7e49a7312bd8db387d16228a4f640788e30 Binary files /dev/null and b/fairseq/__pycache__/distributed_utils.cpython-310.pyc differ diff --git a/fairseq/__pycache__/file_io.cpython-310.pyc b/fairseq/__pycache__/file_io.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e2a673b2d85b90376a2a9f92df8c6c7ccaac7703 Binary files /dev/null and b/fairseq/__pycache__/file_io.cpython-310.pyc differ diff --git a/fairseq/__pycache__/file_utils.cpython-310.pyc b/fairseq/__pycache__/file_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9f88dd582b19fb3111e2fbc19e4f16b1ce9dc6c0 Binary files /dev/null and b/fairseq/__pycache__/file_utils.cpython-310.pyc differ diff --git a/fairseq/__pycache__/incremental_decoding_utils.cpython-310.pyc b/fairseq/__pycache__/incremental_decoding_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8e0294f6eb2024b858edc021f1866e6997be497d Binary files /dev/null and b/fairseq/__pycache__/incremental_decoding_utils.cpython-310.pyc differ diff --git a/fairseq/__pycache__/iterative_refinement_generator.cpython-310.pyc b/fairseq/__pycache__/iterative_refinement_generator.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..61134cc3cbc4d592b71d47b469c11ea527bed980 Binary files /dev/null and b/fairseq/__pycache__/iterative_refinement_generator.cpython-310.pyc differ diff --git a/fairseq/__pycache__/legacy_distributed_data_parallel.cpython-310.pyc b/fairseq/__pycache__/legacy_distributed_data_parallel.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ec36601deb4d0c64e423d3c6982e7ab57ec2c0fb Binary files /dev/null and b/fairseq/__pycache__/legacy_distributed_data_parallel.cpython-310.pyc differ diff --git a/fairseq/__pycache__/options.cpython-310.pyc b/fairseq/__pycache__/options.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..08c0f766d505ea734bc5a5c1e54767f0ab621b33 Binary files /dev/null and b/fairseq/__pycache__/options.cpython-310.pyc differ diff --git a/fairseq/__pycache__/pdb.cpython-310.pyc b/fairseq/__pycache__/pdb.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f3b9177cc9438cbf2c9fe93865fcd4bf07ebe966 Binary files /dev/null and b/fairseq/__pycache__/pdb.cpython-310.pyc differ diff --git a/fairseq/__pycache__/registry.cpython-310.pyc b/fairseq/__pycache__/registry.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..941bc2769852ba09c756b44050b4c31d65555772 Binary files /dev/null and b/fairseq/__pycache__/registry.cpython-310.pyc differ diff --git a/fairseq/__pycache__/search.cpython-310.pyc b/fairseq/__pycache__/search.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..03f9b5355d4a116eca6be5523a898a80de56b324 Binary files /dev/null and b/fairseq/__pycache__/search.cpython-310.pyc differ diff --git a/fairseq/__pycache__/sequence_generator.cpython-310.pyc b/fairseq/__pycache__/sequence_generator.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..56f558e5a3f3fe68973956bd0f2ae1b7d51d46ad Binary files /dev/null and b/fairseq/__pycache__/sequence_generator.cpython-310.pyc differ diff --git a/fairseq/__pycache__/tokenizer.cpython-310.pyc b/fairseq/__pycache__/tokenizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a404b3458b631f92381cb511102f523546e410d8 Binary files /dev/null and b/fairseq/__pycache__/tokenizer.cpython-310.pyc differ diff --git a/fairseq/__pycache__/utils.cpython-310.pyc b/fairseq/__pycache__/utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3555960e34025d6077ac0fa271e452c209693662 Binary files /dev/null and b/fairseq/__pycache__/utils.cpython-310.pyc differ diff --git a/fairseq/benchmark/__init__.py b/fairseq/benchmark/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..926f3ce73975745fbe7e15f307b44761228ec46e --- /dev/null +++ b/fairseq/benchmark/__init__.py @@ -0,0 +1,12 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +# import models/tasks to register them +from . import ( # noqa + dummy_lm, + dummy_masked_lm, + dummy_model, + dummy_mt, +) diff --git a/fairseq/benchmark/__pycache__/__init__.cpython-310.pyc b/fairseq/benchmark/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad26a8b829742ee2cc12f10eea6150c29e3a62e0 Binary files /dev/null and b/fairseq/benchmark/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/benchmark/__pycache__/dummy_lm.cpython-310.pyc b/fairseq/benchmark/__pycache__/dummy_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cf38e95297004a51b01faa88add542a56961dea3 Binary files /dev/null and b/fairseq/benchmark/__pycache__/dummy_lm.cpython-310.pyc differ diff --git a/fairseq/benchmark/__pycache__/dummy_masked_lm.cpython-310.pyc b/fairseq/benchmark/__pycache__/dummy_masked_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..91012312f5e6fd43b716aee6f87f35db92341a00 Binary files /dev/null and b/fairseq/benchmark/__pycache__/dummy_masked_lm.cpython-310.pyc differ diff --git a/fairseq/benchmark/__pycache__/dummy_model.cpython-310.pyc b/fairseq/benchmark/__pycache__/dummy_model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ca36f6ca77c7303cd4c725c739f67dfa3881d420 Binary files /dev/null and b/fairseq/benchmark/__pycache__/dummy_model.cpython-310.pyc differ diff --git a/fairseq/benchmark/__pycache__/dummy_mt.cpython-310.pyc b/fairseq/benchmark/__pycache__/dummy_mt.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6b126e3e7b989effaa4fe4b32f9d2033a9cc0025 Binary files /dev/null and b/fairseq/benchmark/__pycache__/dummy_mt.cpython-310.pyc differ diff --git a/fairseq/benchmark/dummy_lm.py b/fairseq/benchmark/dummy_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..92e9dc8df556ee41761065b476a7f017d1a2fe45 --- /dev/null +++ b/fairseq/benchmark/dummy_lm.py @@ -0,0 +1,118 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np +import torch + +from fairseq.data import Dictionary, FairseqDataset +from fairseq.tasks import FairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task('dummy_lm') +class DummyLMTask(FairseqTask): + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('--dict-size', default=49996, type=int) + parser.add_argument('--dataset-size', default=100000, type=int) + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments ' + 'per sample for BERT dataset') + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + dictionary.pad_to_multiple_(8) # often faster if divisible by 8 + + seq = torch.arange(args.tokens_per_sample + 1) + dictionary.pad() + 1 + + self.dummy_src = seq[:-1] + self.dummy_tgt = seq[1:] + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task. """ + dictionary = Dictionary() + for i in range(args.dict_size): + dictionary.add_symbol('word{}'.format(i)) + logger.info('dictionary: {} types'.format(len(dictionary))) + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if self.args.max_sentences is not None: + bsz = self.args.max_sentences + else: + bsz = max(1, self.args.max_tokens // self.args.tokens_per_sample) + self.datasets[split] = DummyDataset( + { + 'id': 1, + 'net_input': { + 'src_tokens': torch.stack([self.dummy_src for _ in range(bsz)]), + 'src_lengths': torch.full( + (bsz, ), self.args.tokens_per_sample, dtype=torch.long + ), + }, + 'target': torch.stack([self.dummy_tgt for _ in range(bsz)]), + 'nsentences': bsz, + 'ntokens': bsz * self.args.tokens_per_sample, + }, + num_items=self.args.dataset_size, + item_size=self.args.tokens_per_sample, + ) + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + +class DummyDataset(FairseqDataset): + + def __init__(self, batch, num_items, item_size): + super().__init__() + self.batch = batch + self.num_items = num_items + self.item_size = item_size + + def __getitem__(self, index): + return index + + def __len__(self): + return self.num_items + + def collater(self, samples): + return self.batch + + @property + def sizes(self): + return np.array([self.item_size] * self.num_items) + + def num_tokens(self, index): + return self.item_size + + def size(self, index): + return self.item_size + + def ordered_indices(self): + return np.arange(self.num_items) + + @property + def supports_prefetch(self): + return False diff --git a/fairseq/benchmark/dummy_masked_lm.py b/fairseq/benchmark/dummy_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..f2e459caa2c247e7babce1c0d0f9390106ac307c --- /dev/null +++ b/fairseq/benchmark/dummy_masked_lm.py @@ -0,0 +1,127 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np +import torch + +from fairseq.data import Dictionary, FairseqDataset +from fairseq.tasks import FairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task('dummy_masked_lm') +class DummyMaskedLMTask(FairseqTask): + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('--dict-size', default=49995, type=int) + parser.add_argument('--dataset-size', default=100000, type=int) + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments ' + 'per sample for BERT dataset') + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = dictionary.add_symbol('') + dictionary.pad_to_multiple_(8) # often faster if divisible by 8 + + mask_idx = 0 + pad_idx = 1 + seq = torch.arange(args.tokens_per_sample) + pad_idx + 1 + mask = torch.arange(2, args.tokens_per_sample, 7) # ~15% + src = seq.clone() + src[mask] = mask_idx + tgt = torch.full_like(seq, pad_idx) + tgt[mask] = seq[mask] + + self.dummy_src = src + self.dummy_tgt = tgt + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task. """ + dictionary = Dictionary() + for i in range(args.dict_size): + dictionary.add_symbol('word{}'.format(i)) + logger.info('dictionary: {} types'.format(len(dictionary))) + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if self.args.max_sentences is not None: + bsz = self.args.max_sentences + else: + bsz = max(1, self.args.max_tokens // self.args.tokens_per_sample) + self.datasets[split] = DummyDataset( + { + 'id': 1, + 'net_input': { + 'src_tokens': torch.stack([self.dummy_src for _ in range(bsz)]), + 'src_lengths': torch.full( + (bsz, ), self.args.tokens_per_sample, dtype=torch.long + ), + }, + 'target': torch.stack([self.dummy_tgt for _ in range(bsz)]), + 'nsentences': bsz, + 'ntokens': bsz * self.args.tokens_per_sample, + }, + num_items=self.args.dataset_size, + item_size=self.args.tokens_per_sample, + ) + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + +class DummyDataset(FairseqDataset): + + def __init__(self, batch, num_items, item_size): + super().__init__() + self.batch = batch + self.num_items = num_items + self.item_size = item_size + + def __getitem__(self, index): + return index + + def __len__(self): + return self.num_items + + def collater(self, samples): + return self.batch + + @property + def sizes(self): + return np.array([self.item_size] * self.num_items) + + def num_tokens(self, index): + return self.item_size + + def size(self, index): + return self.item_size + + def ordered_indices(self): + return np.arange(self.num_items) + + @property + def supports_prefetch(self): + return False diff --git a/fairseq/benchmark/dummy_model.py b/fairseq/benchmark/dummy_model.py new file mode 100644 index 0000000000000000000000000000000000000000..817cdb34bb2dcdc8a49dd13328c9d020b4e9d03c --- /dev/null +++ b/fairseq/benchmark/dummy_model.py @@ -0,0 +1,95 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +import torch.nn.functional as F + +from fairseq.data import Dictionary +from fairseq.models import ( + FairseqDecoder, + FairseqLanguageModel, + register_model, + register_model_architecture, +) + + +@register_model('dummy_model') +class DummyModel(FairseqLanguageModel): + + def __init__(self, args, encoder): + super().__init__(encoder) + self.args = args + + @staticmethod + def add_args(parser): + parser.add_argument('--num-layers', type=int, default=24) + parser.add_argument('--embed-dim', type=int, default=1024) + + @classmethod + def build_model(cls, args, task): + encoder = DummyEncoder( + num_embed=len(task.target_dictionary), + embed_dim=args.embed_dim, + num_layers=args.num_layers, + ) + return cls(args, encoder) + + def forward(self, src_tokens, masked_tokens=None, **kwargs): + return self.decoder(src_tokens, masked_tokens=masked_tokens) + + +class DummyEncoder(FairseqDecoder): + + def __init__(self, num_embed=50000, embed_dim=1024, num_layers=24): + super().__init__(Dictionary()) + self.embed = nn.Embedding( + num_embeddings=num_embed, embedding_dim=embed_dim, padding_idx=0 + ) + self.layers_a = nn.ModuleList([ + nn.Sequential( + nn.LayerNorm(embed_dim), + nn.Linear(embed_dim, 3*embed_dim), # q, k, v input projection + nn.Linear(3*embed_dim, embed_dim), # skip self-attention + nn.Linear(embed_dim, embed_dim), # output projection + nn.Dropout(), + ) + for i in range(num_layers) + ]) + self.layers_b = nn.ModuleList([ + nn.Sequential( + nn.LayerNorm(embed_dim), + nn.Linear(embed_dim, 4*embed_dim), # FFN + nn.ReLU(), + nn.Linear(4*embed_dim, embed_dim), # FFN + nn.Dropout(0.1), + ) + for i in range(num_layers) + ]) + self.out_proj = nn.Linear(embed_dim, num_embed) + + def forward(self, tokens, masked_tokens=None): + x = self.embed(tokens) + for layer_a, layer_b in zip(self.layers_a, self.layers_b): + x = x + layer_a(x) + x = x + layer_b(x) + x = self.out_proj(x) + if masked_tokens is not None: + x = x[masked_tokens] + return (x,) + + def max_positions(self): + return 1024 + + def get_normalized_probs(self, net_output, log_probs, sample=None): + logits = net_output[0].float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + + +@register_model_architecture('dummy_model', 'dummy_model') +def base_architecture(args): + pass diff --git a/fairseq/benchmark/dummy_mt.py b/fairseq/benchmark/dummy_mt.py new file mode 100644 index 0000000000000000000000000000000000000000..09f2f0c1192636d402bfd1a4d7416e08a74fad83 --- /dev/null +++ b/fairseq/benchmark/dummy_mt.py @@ -0,0 +1,120 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np +import torch + +from fairseq.data import Dictionary, FairseqDataset +from fairseq.tasks import FairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task('dummy_mt') +class DummyMTTask(FairseqTask): + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('--dict-size', default=49996, type=int) + parser.add_argument('--dataset-size', default=100000, type=int) + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments ' + 'per sample for BERT dataset') + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + dictionary.pad_to_multiple_(8) # often faster if divisible by 8 + + seq = torch.arange(args.tokens_per_sample + 1) + dictionary.pad() + 1 + + self.dummy_src = seq[:-1] + self.dummy_tgt = seq[1:] + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task. """ + dictionary = Dictionary() + for i in range(args.dict_size): + dictionary.add_symbol('word{}'.format(i)) + logger.info('dictionary: {} types'.format(len(dictionary))) + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if self.args.max_sentences is not None: + bsz = self.args.max_sentences + else: + bsz = max(1, self.args.max_tokens // self.args.tokens_per_sample) + tgt = torch.stack([self.dummy_tgt for _ in range(bsz)]) + self.datasets[split] = DummyDataset( + { + 'id': 1, + 'net_input': { + 'src_tokens': torch.stack([self.dummy_src for _ in range(bsz)]), + 'src_lengths': torch.full( + (bsz, ), self.args.tokens_per_sample, dtype=torch.long + ), + 'prev_output_tokens': tgt.clone(), + }, + 'target': tgt, + 'nsentences': bsz, + 'ntokens': bsz * self.args.tokens_per_sample, + }, + num_items=self.args.dataset_size, + item_size=self.args.tokens_per_sample, + ) + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + +class DummyDataset(FairseqDataset): + + def __init__(self, batch, num_items, item_size): + super().__init__() + self.batch = batch + self.num_items = num_items + self.item_size = item_size + + def __getitem__(self, index): + return index + + def __len__(self): + return self.num_items + + def collater(self, samples): + return self.batch + + @property + def sizes(self): + return np.array([self.item_size] * self.num_items) + + def num_tokens(self, index): + return self.item_size + + def size(self, index): + return self.item_size + + def ordered_indices(self): + return np.arange(self.num_items) + + @property + def supports_prefetch(self): + return False diff --git a/fairseq/binarizer.py b/fairseq/binarizer.py new file mode 100644 index 0000000000000000000000000000000000000000..ec3b90f211af7a8308f076389944f8a7184279ea --- /dev/null +++ b/fairseq/binarizer.py @@ -0,0 +1,104 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +from collections import Counter + +from fairseq.tokenizer import tokenize_line +import torch +from fairseq.file_io import PathManager + +def safe_readline(f): + pos = f.tell() + while True: + try: + return f.readline() + except UnicodeDecodeError: + pos -= 1 + f.seek(pos) # search where this character begins + + +class Binarizer: + @staticmethod + def binarize( + filename, + dict, + consumer, + tokenize=tokenize_line, + append_eos=True, + reverse_order=False, + offset=0, + end=-1, + already_numberized=False, + ): + nseq, ntok = 0, 0 + replaced = Counter() + + def replaced_consumer(word, idx): + if idx == dict.unk_index and word != dict.unk_word: + replaced.update([word]) + + with open(PathManager.get_local_path(filename), "r", encoding="utf-8") as f: + f.seek(offset) + # next(f) breaks f.tell(), hence readline() must be used + line = safe_readline(f) + while line: + if end > 0 and f.tell() > end: + break + if already_numberized: + id_strings = line.strip().split() + id_list = [int(id_string) for id_string in id_strings] + if reverse_order: + id_list.reverse() + if append_eos: + id_list.append(dict.eos()) + ids = torch.IntTensor(id_list) + else: + ids = dict.encode_line( + line=line, + line_tokenizer=tokenize, + add_if_not_exist=False, + consumer=replaced_consumer, + append_eos=append_eos, + reverse_order=reverse_order, + ) + nseq += 1 + ntok += len(ids) + consumer(ids) + line = f.readline() + return { + "nseq": nseq, + "nunk": sum(replaced.values()), + "ntok": ntok, + "replaced": replaced, + } + + @staticmethod + def binarize_alignments(filename, alignment_parser, consumer, offset=0, end=-1): + nseq = 0 + + with open(PathManager.get_local_path(filename), "r") as f: + f.seek(offset) + line = safe_readline(f) + while line: + if end > 0 and f.tell() > end: + break + ids = alignment_parser(line) + nseq += 1 + consumer(ids) + line = f.readline() + return {"nseq": nseq} + + @staticmethod + def find_offsets(filename, num_chunks): + with open(PathManager.get_local_path(filename), "r", encoding="utf-8") as f: + size = os.fstat(f.fileno()).st_size + chunk_size = size // num_chunks + offsets = [0 for _ in range(num_chunks + 1)] + for i in range(1, num_chunks): + f.seek(chunk_size * i) + safe_readline(f) + offsets[i] = f.tell() + return offsets diff --git a/fairseq/checkpoint_utils.py b/fairseq/checkpoint_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..af21db929ffef0c01599b87fd19431c316d6d26b --- /dev/null +++ b/fairseq/checkpoint_utils.py @@ -0,0 +1,522 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import collections +import logging +import os +import re +import traceback +from collections import OrderedDict +from typing import Union + +import torch +from fairseq.file_io import PathManager +from fairseq.models import FairseqDecoder, FairseqEncoder +from torch.serialization import default_restore_location + + +logger = logging.getLogger(__name__) + + +def save_checkpoint(args, trainer, epoch_itr, val_loss): + from fairseq import distributed_utils, meters + + # only one worker should attempt to create the required dir + if args.distributed_rank == 0: + os.makedirs(args.save_dir, exist_ok=True) + + prev_best = getattr(save_checkpoint, "best", val_loss) + if val_loss is not None: + best_function = max if args.maximize_best_checkpoint_metric else min + save_checkpoint.best = best_function(val_loss, prev_best) + + if args.no_save or not trainer.is_data_parallel_master: + return + + def is_better(a, b): + return a >= b if args.maximize_best_checkpoint_metric else a <= b + + write_timer = meters.StopwatchMeter() + write_timer.start() + + epoch = epoch_itr.epoch + end_of_epoch = epoch_itr.end_of_epoch() + updates = trainer.get_num_updates() + + suffix = getattr(args, "checkpoint_suffix", "") + checkpoint_conds = collections.OrderedDict() + checkpoint_conds["checkpoint{}{}.pt".format(epoch, suffix)] = ( + end_of_epoch + and not args.no_epoch_checkpoints + and epoch % args.save_interval == 0 + ) + checkpoint_conds["checkpoint_{}_{}{}.pt".format(epoch, updates, suffix)] = ( + not end_of_epoch + and args.save_interval_updates > 0 + and updates % args.save_interval_updates == 0 + ) + checkpoint_conds["checkpoint_best{}.pt".format(suffix)] = val_loss is not None and ( + not hasattr(save_checkpoint, "best") + or is_better(val_loss, save_checkpoint.best) + ) + if val_loss is not None and args.keep_best_checkpoints > 0: + checkpoint_conds["checkpoint.best_{}_{:.2f}.pt".format( + args.best_checkpoint_metric, val_loss)] = ( + not hasattr(save_checkpoint, "best") + or is_better(val_loss, save_checkpoint.best) + ) + checkpoint_conds["checkpoint_last{}.pt".format(suffix)] = not args.no_last_checkpoints + + extra_state = {"train_iterator": epoch_itr.state_dict(), "val_loss": val_loss} + if hasattr(save_checkpoint, "best"): + extra_state.update({"best": save_checkpoint.best}) + + checkpoints = [ + os.path.join(args.save_dir, fn) for fn, cond in checkpoint_conds.items() if cond + ] + if len(checkpoints) > 0: + trainer.save_checkpoint(checkpoints[0], extra_state) + for cp in checkpoints[1:]: + PathManager.copy(checkpoints[0], cp, overwrite=True) + + write_timer.stop() + logger.info( + "saved checkpoint {} (epoch {} @ {} updates, score {}) (writing took {} seconds)".format( + checkpoints[0], epoch, updates, val_loss, write_timer.sum + ) + ) + + if not end_of_epoch and args.keep_interval_updates > 0: + # remove old checkpoints; checkpoints are sorted in descending order + checkpoints = checkpoint_paths( + args.save_dir, pattern=r"checkpoint_\d+_(\d+)\.pt" + ) + for old_chk in checkpoints[args.keep_interval_updates :]: + if os.path.lexists(old_chk): + os.remove(old_chk) + + if args.keep_last_epochs > 0: + # remove old epoch checkpoints; checkpoints are sorted in descending order + checkpoints = checkpoint_paths(args.save_dir, pattern=r"checkpoint(\d+)\.pt") + for old_chk in checkpoints[args.keep_last_epochs :]: + if os.path.lexists(old_chk): + os.remove(old_chk) + + if args.keep_best_checkpoints > 0: + # only keep the best N checkpoints according to validation metric + checkpoints = checkpoint_paths( + args.save_dir, pattern=r"checkpoint\.best_{}_(\d+\.?\d*)\.pt".format(args.best_checkpoint_metric)) + if not args.maximize_best_checkpoint_metric: + checkpoints = checkpoints[::-1] + for old_chk in checkpoints[args.keep_best_checkpoints:]: + if os.path.lexists(old_chk): + os.remove(old_chk) + + +def load_checkpoint(args, trainer, **passthrough_args): + """ + Load a checkpoint and restore the training iterator. + + *passthrough_args* will be passed through to + ``trainer.get_train_iterator``. + """ + reset_optimizer = args.reset_optimizer + reset_lr_scheduler = args.reset_lr_scheduler + optimizer_overrides = eval(args.optimizer_overrides) + reset_meters = args.reset_meters + reset_dataloader = args.reset_dataloader + + if getattr(args, 'finetune_from_model', None) is not None \ + and (reset_optimizer or reset_lr_scheduler or reset_meters or reset_dataloader): + raise ValueError("--finetune-from-model can not be set together with either --reset-optimizer" + " or reset_lr_scheduler or reset_meters or reset_dataloader") + + suffix = getattr(args, "checkpoint_suffix", "") + if args.restore_file == "checkpoint_last.pt": # default value of restore_file is 'checkpoint_last.pt' + checkpoint_path = os.path.join(args.save_dir, "checkpoint_last{}.pt".format(suffix)) + first_launch = not PathManager.exists(checkpoint_path) + if getattr(args, 'finetune_from_model', None) is not None and first_launch: + # if there is no last checkpoint to restore, start the finetune from pretrained model + # else just use usual logic to load checkpoint, e.g. restart from last checkpoint and etc. + if PathManager.exists(args.finetune_from_model): + checkpoint_path = args.finetune_from_model + reset_optimizer = True + reset_lr_scheduler = True + reset_meters = True + reset_dataloader = True + logger.info(f'loading pretrained model from {checkpoint_path}: ' + 'optimizer, lr scheduler, meters, dataloader will be reset') + else: + raise ValueError(f'--funetune-from-model {args.finetune_from_model} does not exist') + elif getattr(args, "model_parallel_size", 1) > 1: + checkpoint_path = args.restore_file.replace(".pt", suffix + ".pt") + else: + checkpoint_path = args.restore_file + + if args.restore_file != "checkpoint_last.pt" and getattr(args, 'finetune_from_model', None): + raise ValueError( + '--finetune-from-model and --restore-file (non-default value) ' + 'can not be specified together: ' + str(args)) + + extra_state = trainer.load_checkpoint( + checkpoint_path, + reset_optimizer, + reset_lr_scheduler, + optimizer_overrides, + reset_meters=reset_meters, + ) + + if ( + extra_state is not None + and "best" in extra_state + and not reset_optimizer + and not reset_meters + ): + save_checkpoint.best = extra_state["best"] + + if extra_state is not None and not reset_dataloader: + # restore iterator from checkpoint + itr_state = extra_state["train_iterator"] + epoch_itr = trainer.get_train_iterator( + epoch=itr_state["epoch"], load_dataset=True, **passthrough_args + ) + epoch_itr.load_state_dict(itr_state) + else: + epoch_itr = trainer.get_train_iterator( + epoch=1, load_dataset=True, **passthrough_args + ) + + trainer.lr_step(epoch_itr.epoch) + + return extra_state, epoch_itr + + +def load_checkpoint_to_cpu(path, arg_overrides=None): + """Loads a checkpoint to CPU (with upgrading for backward compatibility).""" + with PathManager.open(path, "rb") as f: + state = torch.load( + f, map_location=lambda s, l: default_restore_location(s, "cpu") + ) + + args = state["args"] + if arg_overrides is not None: + for arg_name, arg_val in arg_overrides.items(): + setattr(args, arg_name, arg_val) + state = _upgrade_state_dict(state) + return state + + +def load_model_ensemble(filenames, arg_overrides=None, task=None, strict=True, suffix=''): + """Loads an ensemble of models. + + Args: + filenames (List[str]): checkpoint files to load + arg_overrides (Dict[str,Any], optional): override model args that + were used during model training + task (fairseq.tasks.FairseqTask, optional): task to use for loading + """ + ensemble, args, _task = load_model_ensemble_and_task( + filenames, arg_overrides, task, strict, suffix, + ) + return ensemble, args + + +def load_model_ensemble_and_task(filenames, arg_overrides=None, task=None, strict=True, suffix=''): + from fairseq import tasks + + ensemble = [] + for filename in filenames: + filename = filename.replace(".pt", suffix + ".pt") + if not PathManager.exists(filename): + raise IOError("Model file not found: {}".format(filename)) + state = load_checkpoint_to_cpu(filename, arg_overrides) + + args = state["args"] + if task is None: + task = tasks.setup_task(args) + + # build model for ensemble + model = task.build_model(args) + model.load_state_dict(state["model"], strict=strict, args=args) + ensemble.append(model) + return ensemble, args, task + + +def checkpoint_paths(path, pattern=r"checkpoint(\d+)\.pt"): + """Retrieves all checkpoints found in `path` directory. + + Checkpoints are identified by matching filename to the specified pattern. If + the pattern contains groups, the result will be sorted by the first group in + descending order. + """ + pt_regexp = re.compile(pattern) + files = os.listdir(path) + + entries = [] + for i, f in enumerate(files): + m = pt_regexp.fullmatch(f) + if m is not None: + idx = float(m.group(1)) if len(m.groups()) > 0 else i + entries.append((idx, m.group(0))) + return [os.path.join(path, x[1]) for x in sorted(entries, reverse=True)] + + +def torch_persistent_save(*args, **kwargs): + for i in range(3): + try: + return torch.save(*args, **kwargs) + except Exception: + if i == 2: + logger.error(traceback.format_exc()) + + +def save_state( + filename, + args, + model_state_dict, + criterion, + optimizer, + lr_scheduler, + num_updates, + optim_history=None, + extra_state=None, +): + from fairseq import utils + + if optim_history is None: + optim_history = [] + if extra_state is None: + extra_state = {} + state_dict = { + "args": args, + "model": model_state_dict or {}, + "optimizer_history": optim_history + + [ + { + "criterion_name": criterion.__class__.__name__, + "optimizer_name": optimizer.__class__.__name__, + "lr_scheduler_state": lr_scheduler.state_dict(), + "num_updates": num_updates, + } + ], + "extra_state": extra_state, + } + if utils.has_parameters(criterion): + state_dict["criterion"] = criterion.state_dict() + if not args.no_save_optimizer_state: + state_dict["last_optimizer_state"] = optimizer.state_dict() + + # convert all state to CPU + state_dict = utils.move_to_cpu(state_dict) + + with PathManager.open(filename, "wb") as f: + torch_persistent_save(state_dict, f) + + +def _upgrade_state_dict(state): + """Helper for upgrading old model checkpoints.""" + from fairseq import models, registry, tasks + + # add optimizer_history + if "optimizer_history" not in state: + state["optimizer_history"] = [ + {"criterion_name": "CrossEntropyCriterion", "best_loss": state["best_loss"]} + ] + state["last_optimizer_state"] = state["optimizer"] + del state["optimizer"] + del state["best_loss"] + # move extra_state into sub-dictionary + if "epoch" in state and "extra_state" not in state: + state["extra_state"] = { + "epoch": state["epoch"], + "batch_offset": state["batch_offset"], + "val_loss": state["val_loss"], + } + del state["epoch"] + del state["batch_offset"] + del state["val_loss"] + # reduce optimizer history's memory usage (only keep the last state) + if "optimizer" in state["optimizer_history"][-1]: + state["last_optimizer_state"] = state["optimizer_history"][-1]["optimizer"] + for optim_hist in state["optimizer_history"]: + del optim_hist["optimizer"] + # record the optimizer class name + if "optimizer_name" not in state["optimizer_history"][-1]: + state["optimizer_history"][-1]["optimizer_name"] = "FairseqNAG" + # move best_loss into lr_scheduler_state + if "lr_scheduler_state" not in state["optimizer_history"][-1]: + state["optimizer_history"][-1]["lr_scheduler_state"] = { + "best": state["optimizer_history"][-1]["best_loss"] + } + del state["optimizer_history"][-1]["best_loss"] + # keep track of number of updates + if "num_updates" not in state["optimizer_history"][-1]: + state["optimizer_history"][-1]["num_updates"] = 0 + # old model checkpoints may not have separate source/target positions + if hasattr(state["args"], "max_positions") and not hasattr( + state["args"], "max_source_positions" + ): + state["args"].max_source_positions = state["args"].max_positions + state["args"].max_target_positions = state["args"].max_positions + # use stateful training data iterator + if "train_iterator" not in state["extra_state"]: + state["extra_state"]["train_iterator"] = { + "epoch": state["extra_state"]["epoch"], + "iterations_in_epoch": state["extra_state"].get("batch_offset", 0), + } + # default to translation task + if not hasattr(state["args"], "task"): + state["args"].task = "translation" + # --raw-text and --lazy-load are deprecated + if getattr(state["args"], "raw_text", False): + state["args"].dataset_impl = "raw" + elif getattr(state["args"], "lazy_load", False): + state["args"].dataset_impl = "lazy" + # epochs start at 1 + if state["extra_state"]["train_iterator"] is not None: + state["extra_state"]["train_iterator"]["epoch"] = max( + state["extra_state"]["train_iterator"].get("epoch", 1), + 1, + ) + + # set any missing default values in the task, model or other registries + registry.set_defaults(state["args"], tasks.TASK_REGISTRY[state["args"].task]) + registry.set_defaults(state["args"], models.ARCH_MODEL_REGISTRY[state["args"].arch]) + for registry_name, REGISTRY in registry.REGISTRIES.items(): + choice = getattr(state["args"], registry_name, None) + if choice is not None: + cls = REGISTRY["registry"][choice] + registry.set_defaults(state["args"], cls) + + return state + + +def prune_state_dict(state_dict, args): + """Prune the given state_dict if desired for LayerDrop + (https://arxiv.org/abs/1909.11556). + + Training with LayerDrop allows models to be robust to pruning at inference + time. This function prunes state_dict to allow smaller models to be loaded + from a larger model and re-maps the existing state_dict for this to occur. + + It's called by functions that load models from checkpoints and does not + need to be called directly. + """ + if not args or args.arch == "ptt_transformer": + # args should not be none, but don't crash if it is. + return state_dict + + encoder_layers_to_keep = ( + args.encoder_layers_to_keep if "encoder_layers_to_keep" in vars(args) else None + ) + decoder_layers_to_keep = ( + args.decoder_layers_to_keep if "decoder_layers_to_keep" in vars(args) else None + ) + + if not encoder_layers_to_keep and not decoder_layers_to_keep: + return state_dict + + # apply pruning + logger.info( + "Pruning model to specified layer configuration - this works best if the model was trained with LayerDrop" + ) + + def create_pruning_pass(layers_to_keep, layer_name): + keep_layers = sorted( + [int(layer_string) for layer_string in layers_to_keep.split(",")] + ) + mapping_dict = {} + for i in range(len(keep_layers)): + mapping_dict[str(keep_layers[i])] = str(i) + + regex = re.compile(r"^{layer}.*\.layers\.(\d+)".format(layer=layer_name)) + return {"substitution_regex": regex, "mapping_dict": mapping_dict} + + pruning_passes = [] + if encoder_layers_to_keep: + pruning_passes.append(create_pruning_pass(encoder_layers_to_keep, "encoder")) + if decoder_layers_to_keep: + pruning_passes.append(create_pruning_pass(decoder_layers_to_keep, "decoder")) + + new_state_dict = {} + for layer_name in state_dict.keys(): + match = re.search(r"\.layers\.(\d+)\.", layer_name) + # if layer has no number in it, it is a supporting layer, such as an + # embedding + if not match: + new_state_dict[layer_name] = state_dict[layer_name] + continue + + # otherwise, layer should be pruned. + original_layer_number = match.group(1) + # figure out which mapping dict to replace from + for pruning_pass in pruning_passes: + if original_layer_number in pruning_pass["mapping_dict"] and pruning_pass[ + "substitution_regex" + ].search(layer_name): + new_layer_number = pruning_pass["mapping_dict"][original_layer_number] + substitution_match = pruning_pass["substitution_regex"].search( + layer_name + ) + new_state_key = ( + layer_name[: substitution_match.start(1)] + + new_layer_number + + layer_name[substitution_match.end(1) :] + ) + new_state_dict[new_state_key] = state_dict[layer_name] + + # Since layers are now pruned, *_layers_to_keep are no longer needed. + # This is more of "It would make it work fix" rather than a proper fix. + if "encoder_layers_to_keep" in vars(args): + args.encoder_layers_to_keep = None + if "decoder_layers_to_keep" in vars(args): + args.decoder_layers_to_keep = None + + return new_state_dict + + +def load_pretrained_component_from_model( + component: Union[FairseqEncoder, FairseqDecoder], checkpoint: str +): + """ + Load a pretrained FairseqEncoder or FairseqDecoder from checkpoint into the + provided `component` object. If state_dict fails to load, there may be a + mismatch in the architecture of the corresponding `component` found in the + `checkpoint` file. + """ + if not PathManager.exists(checkpoint): + raise IOError("Model file not found: {}".format(checkpoint)) + state = load_checkpoint_to_cpu(checkpoint) + if isinstance(component, FairseqEncoder): + component_type = "encoder" + elif isinstance(component, FairseqDecoder): + component_type = "decoder" + else: + raise ValueError( + "component to load must be either a FairseqEncoder or " + "FairseqDecoder. Loading other component types are not supported." + ) + component_state_dict = OrderedDict() + for key in state["model"].keys(): + if key.startswith(component_type): + # encoder.input_layers.0.0.weight --> input_layers.0.0.weight + component_subkey = key[len(component_type) + 1 :] + component_state_dict[component_subkey] = state["model"][key] + component.load_state_dict(component_state_dict, strict=True) + return component + + +def verify_checkpoint_directory(save_dir: str) -> None: + if not os.path.exists(save_dir): + os.makedirs(save_dir, exist_ok=True) + temp_file_path = os.path.join(save_dir, "dummy") + try: + with open(temp_file_path, "w"): + pass + except OSError as e: + logger.warning("Unable to access checkpoint save directory: {}".format(save_dir)) + raise e + else: + os.remove(temp_file_path) diff --git a/fairseq/clib/libbleu/libbleu.cpp b/fairseq/clib/libbleu/libbleu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3cf2d65b6d16e19ea299ebe43c9c25e3481d4524 --- /dev/null +++ b/fairseq/clib/libbleu/libbleu.cpp @@ -0,0 +1,141 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include +#include +#include + +typedef struct +{ + size_t reflen; + size_t predlen; + size_t match1; + size_t count1; + size_t match2; + size_t count2; + size_t match3; + size_t count3; + size_t match4; + size_t count4; +} bleu_stat; + +// left trim (remove pad) +void bleu_ltrim(size_t* len, int** sent, int pad) { + size_t start = 0; + while(start < *len) { + if (*(*sent + start) != pad) { break; } + start++; + } + *sent += start; + *len -= start; +} + +// right trim remove (eos) +void bleu_rtrim(size_t* len, int** sent, int pad, int eos) { + size_t end = *len - 1; + while (end > 0) { + if (*(*sent + end) != eos && *(*sent + end) != pad) { break; } + end--; + } + *len = end + 1; +} + +// left and right trim +void bleu_trim(size_t* len, int** sent, int pad, int eos) { + bleu_ltrim(len, sent, pad); + bleu_rtrim(len, sent, pad, eos); +} + +size_t bleu_hash(int len, int* data) { + size_t h = 14695981039346656037ul; + size_t prime = 0x100000001b3; + char* b = (char*) data; + size_t blen = sizeof(int) * len; + + while (blen-- > 0) { + h ^= *b++; + h *= prime; + } + + return h; +} + +void bleu_addngram( + size_t *ntotal, size_t *nmatch, size_t n, + size_t reflen, int* ref, size_t predlen, int* pred) { + + if (predlen < n) { return; } + + predlen = predlen - n + 1; + (*ntotal) += predlen; + + if (reflen < n) { return; } + + reflen = reflen - n + 1; + + std::map count; + while (predlen > 0) { + size_t w = bleu_hash(n, pred++); + count[w]++; + predlen--; + } + + while (reflen > 0) { + size_t w = bleu_hash(n, ref++); + if (count[w] > 0) { + (*nmatch)++; + count[w] -=1; + } + reflen--; + } +} + +extern "C" { + +#ifdef _WIN64 +__declspec(dllexport) +#endif +void bleu_zero_init(bleu_stat* stat) { + std::memset(stat, 0, sizeof(bleu_stat)); +} + +#ifdef _WIN64 +__declspec(dllexport) +#endif +void bleu_one_init(bleu_stat* stat) { + bleu_zero_init(stat); + stat->count1 = 0; + stat->count2 = 1; + stat->count3 = 1; + stat->count4 = 1; + stat->match1 = 0; + stat->match2 = 1; + stat->match3 = 1; + stat->match4 = 1; +} + +#ifdef _WIN64 +__declspec(dllexport) +#endif +void bleu_add( + bleu_stat* stat, + size_t reflen, int* ref, size_t predlen, int* pred, int pad, int eos) { + + bleu_trim(&reflen, &ref, pad, eos); + bleu_trim(&predlen, &pred, pad, eos); + stat->reflen += reflen; + stat->predlen += predlen; + + bleu_addngram(&stat->count1, &stat->match1, 1, reflen, ref, predlen, pred); + bleu_addngram(&stat->count2, &stat->match2, 2, reflen, ref, predlen, pred); + bleu_addngram(&stat->count3, &stat->match3, 3, reflen, ref, predlen, pred); + bleu_addngram(&stat->count4, &stat->match4, 4, reflen, ref, predlen, pred); +} + +} diff --git a/fairseq/clib/libbleu/module.cpp b/fairseq/clib/libbleu/module.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8ed9a84b1c028bfe9ed1d45be6857b6e79b3459f --- /dev/null +++ b/fairseq/clib/libbleu/module.cpp @@ -0,0 +1,37 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include + + +static PyMethodDef method_def[] = { + {NULL, NULL, 0, NULL} +}; + +static struct PyModuleDef module_def = { + PyModuleDef_HEAD_INIT, + "libbleu", /* name of module */ + NULL, /* module documentation, may be NULL */ + -1, /* size of per-interpreter state of the module, + or -1 if the module keeps state in global variables. */ + method_def +}; + + +#if PY_MAJOR_VERSION == 2 +PyMODINIT_FUNC init_libbleu() +#else +PyMODINIT_FUNC PyInit_libbleu() +#endif +{ + PyObject *m = PyModule_Create(&module_def); + if (!m) { + return NULL; + } + return m; +} diff --git a/fairseq/clib/libnat/edit_dist.cpp b/fairseq/clib/libnat/edit_dist.cpp new file mode 100644 index 0000000000000000000000000000000000000000..6bc6a937d6abde0cd49769c4def69ac0560096bc --- /dev/null +++ b/fairseq/clib/libnat/edit_dist.cpp @@ -0,0 +1,231 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include // @manual=//caffe2:torch_extension +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +using namespace ::std; + +vector> edit_distance2_with_dp( + vector& x, + vector& y) { + uint32_t lx = x.size(); + uint32_t ly = y.size(); + vector> d(lx + 1, vector(ly + 1)); + for (uint32_t i = 0; i < lx + 1; i++) { + d[i][0] = i; + } + for (uint32_t j = 0; j < ly + 1; j++) { + d[0][j] = j; + } + for (uint32_t i = 1; i < lx + 1; i++) { + for (uint32_t j = 1; j < ly + 1; j++) { + d[i][j] = + min(min(d[i - 1][j], d[i][j - 1]) + 1, + d[i - 1][j - 1] + 2 * (x.at(i - 1) == y.at(j - 1) ? 0 : 1)); + } + } + return d; +} + +vector> edit_distance2_backtracking( + vector>& d, + vector& x, + vector& y, + uint32_t terminal_symbol) { + vector seq; + vector> edit_seqs(x.size() + 2, vector()); + /* + edit_seqs: + 0~x.size() cell is the insertion sequences + last cell is the delete sequence + */ + + if (x.size() == 0) { + edit_seqs.at(0) = y; + return edit_seqs; + } + + uint32_t i = d.size() - 1; + uint32_t j = d.at(0).size() - 1; + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (d.at(i).at(j - 1) < d.at(i).at(j))) { + seq.push_back(1); // insert + seq.push_back(y.at(j - 1)); + j--; + } else if ((i > 0) && (d.at(i - 1).at(j) < d.at(i).at(j))) { + seq.push_back(2); // delete + seq.push_back(x.at(i - 1)); + i--; + } else { + seq.push_back(3); // keep + seq.push_back(x.at(i - 1)); + i--; + j--; + } + } + + uint32_t prev_op, op, s, word; + prev_op = 0, s = 0; + for (uint32_t k = 0; k < seq.size() / 2; k++) { + op = seq.at(seq.size() - 2 * k - 2); + word = seq.at(seq.size() - 2 * k - 1); + if (prev_op != 1) { + s++; + } + if (op == 1) // insert + { + edit_seqs.at(s - 1).push_back(word); + } else if (op == 2) // delete + { + edit_seqs.at(x.size() + 1).push_back(1); + } else { + edit_seqs.at(x.size() + 1).push_back(0); + } + + prev_op = op; + } + + for (uint32_t k = 0; k < edit_seqs.size(); k++) { + if (edit_seqs[k].size() == 0) { + edit_seqs[k].push_back(terminal_symbol); + } + } + return edit_seqs; +} + +vector> edit_distance2_backtracking_with_delete( + vector>& d, + vector& x, + vector& y, + uint32_t terminal_symbol, + uint32_t deletion_symbol) { + vector seq; + vector> edit_seqs(x.size() + 1, vector()); + /* + edit_seqs: + 0~x.size() cell is the insertion sequences + last cell is the delete sequence + */ + + if (x.size() == 0) { + edit_seqs.at(0) = y; + return edit_seqs; + } + + uint32_t i = d.size() - 1; + uint32_t j = d.at(0).size() - 1; + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (d.at(i).at(j - 1) < d.at(i).at(j))) { + seq.push_back(1); // insert + seq.push_back(y.at(j - 1)); + j--; + } else if ((i > 0) && (d.at(i - 1).at(j) < d.at(i).at(j))) { + seq.push_back(2); // delete + seq.push_back(x.at(i - 1)); + i--; + } else { + seq.push_back(3); // keep + seq.push_back(x.at(i - 1)); + i--; + j--; + } + } + + uint32_t prev_op, op, s, word; + prev_op = 0, s = 0; + for (uint32_t k = 0; k < seq.size() / 2; k++) { + op = seq.at(seq.size() - 2 * k - 2); + word = seq.at(seq.size() - 2 * k - 1); + if (prev_op != 1) { + s++; + } + if (op == 1) // insert + { + edit_seqs.at(s - 1).push_back(word); + } else if (op == 2) // delete + { + edit_seqs.at(s - 1).push_back(deletion_symbol); + } + + prev_op = op; + } + + for (uint32_t k = 0; k < edit_seqs.size(); k++) { + if (edit_seqs.at(k).size() == 0) { + edit_seqs.at(k).push_back(terminal_symbol); + } + } + return edit_seqs; +} + +vector compute_ed2( + vector>& xs, + vector>& ys) { + vector distances(xs.size()); + for (uint32_t i = 0; i < xs.size(); i++) { + vector> d = edit_distance2_with_dp(xs.at(i), ys.at(i)); + distances.at(i) = d.at(xs.at(i).size()).at(ys.at(i).size()); + } + return distances; +} + +vector>> suggested_ed2_path( + vector>& xs, + vector>& ys, + uint32_t terminal_symbol) { + vector>> seq(xs.size()); + for (uint32_t i = 0; i < xs.size(); i++) { + vector> d = edit_distance2_with_dp(xs.at(i), ys.at(i)); + seq.at(i) = + edit_distance2_backtracking(d, xs.at(i), ys.at(i), terminal_symbol); + } + return seq; +} + +vector>> suggested_ed2_path_with_delete( + vector>& xs, + vector>& ys, + uint32_t terminal_symbol, + uint32_t deletion_symbol) { + vector>> seq(xs.size()); + for (uint32_t i = 0; i < xs.size(); i++) { + vector> d = edit_distance2_with_dp(xs.at(i), ys.at(i)); + seq.at(i) = edit_distance2_backtracking_with_delete( + d, xs.at(i), ys.at(i), terminal_symbol, deletion_symbol); + } + return seq; +} + +PYBIND11_MODULE(libnat, m) { + m.def("compute_ed2", &compute_ed2, "compute_ed2"); + m.def("suggested_ed2_path", &suggested_ed2_path, "suggested_ed2_path"); + m.def( + "suggested_ed2_path_with_delete", + &suggested_ed2_path_with_delete, + "suggested_ed2_path_with_delete"); +} diff --git a/fairseq/clib/libnat_cuda/binding.cpp b/fairseq/clib/libnat_cuda/binding.cpp new file mode 100644 index 0000000000000000000000000000000000000000..aaa6244d5c6819acfae5f408280205661a3389ae --- /dev/null +++ b/fairseq/clib/libnat_cuda/binding.cpp @@ -0,0 +1,60 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +/* + This code is partially adpoted from https://github.com/1ytic/pytorch-edit-distance + */ + +#include "edit_dist.h" +#include + +#ifndef TORCH_CHECK +#define TORCH_CHECK AT_CHECK +#endif + +#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + + +torch::Tensor LevenshteinDistance( + torch::Tensor source, + torch::Tensor target, + torch::Tensor source_length, + torch::Tensor target_length) { + + CHECK_INPUT(source); + CHECK_INPUT(target); + CHECK_INPUT(source_length); + CHECK_INPUT(target_length); + return LevenshteinDistanceCuda(source, target, source_length, target_length); +} + +torch::Tensor GenerateDeletionLabel( + torch::Tensor source, + torch::Tensor operations) { + + CHECK_INPUT(source); + CHECK_INPUT(operations); + return GenerateDeletionLabelCuda(source, operations); +} + +std::pair GenerateInsertionLabel( + torch::Tensor target, + torch::Tensor operations) { + + CHECK_INPUT(target); + CHECK_INPUT(operations); + return GenerateInsertionLabelCuda(target, operations); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("levenshtein_distance", &LevenshteinDistance, "Levenshtein distance"); + m.def("generate_deletion_labels", &GenerateDeletionLabel, "Generate Deletion Label"); + m.def("generate_insertion_labels", &GenerateInsertionLabel, "Generate Insertion Label"); +} diff --git a/fairseq/clib/libnat_cuda/edit_dist.cu b/fairseq/clib/libnat_cuda/edit_dist.cu new file mode 100644 index 0000000000000000000000000000000000000000..b6486a8c22f0dc138ce80c8936082824a80a17e7 --- /dev/null +++ b/fairseq/clib/libnat_cuda/edit_dist.cu @@ -0,0 +1,332 @@ +/** +* Copyright 2017-present, Facebook, Inc. +* All rights reserved. +* +* This source code is licensed under the license found in the +* LICENSE file in the root directory of this source tree. +*/ + +#include "edit_dist.h" +#include +#include +#include +#include +#include // std::pair + +template +__global__ void generate_deletion_label_kernel( + const scalar_t* __restrict__ source, + const size_t source_size, + const size_t operation_size, + int* __restrict__ operations, + int* __restrict__ labels) { + + const int index = blockIdx.x; + const int offset = index * operation_size; + const int offset_label = index * source_size; + + for (int i = 0; i < source_size; i++) { + labels[offset_label + i] = 0; + } + + int k = 0; + for (int i = 0; i < operation_size; i++){ + if (operations[offset + i] == 0){ + break; + } else if (operations[offset + i] == 1){ + continue; + } else { + labels[offset_label + k] = 3 - operations[offset + i]; + k++; + } + } +} + +template +__global__ void generate_insertion_label_kernel( + const scalar_t* __restrict__ target, + const size_t target_size, + const size_t operation_size, + int* __restrict__ operations, + int* __restrict__ labels, + int* __restrict__ masks) { + + const int index = blockIdx.x; + const int offset = index * operation_size; + const int offset_label = index * target_size; + + int k = 0; + int u = 0; + int m = 0; + + for (int i = 0; i < target_size; i++) { + labels[offset_label + i] = 0; + masks[offset_label + i] = 0; + } + + for (int i = 0; i < operation_size-1; i++){ + if (operations[offset + i] == 0){ + break; + } else if (operations[offset + i] == 2){ + continue; + } else if (operations[offset + i] == 1){ + masks[offset_label + m] = 1; + u++; m++; + } else { + labels[offset_label + k] = u; + masks[offset_label + m] = 0; + k++; m++; + u = 0; + } + } +} + +template +__global__ void levenshtein_distance_kernel( + const scalar_t* __restrict__ source, + const scalar_t* __restrict__ target, + const int* __restrict__ source_length, + const int* __restrict__ target_length, + const size_t source_size, + const size_t target_size, + int* __restrict__ operations, + int* __restrict__ errors_curr) { + + const int index = blockIdx.x; + const int offset = index * (source_size + target_size); + const int d = index * (source_size + 1) * (target_size + 1); + const int t = target_size + 1; + + auto err_idx = [d, t](int i, int j) { return d + i * t + j; }; + auto opt_idx = [offset](int k) { return offset + k; }; + + const int hyp_len = source_length[index]; + const int ref_len = target_length[index]; + const scalar_t* hyp_begin = source + index * source_size; + const scalar_t* ref_begin = target + index * target_size; + + // dynamic programming + for (int i = 0; i <= hyp_len; i++){ + errors_curr[err_idx(i, 0)] = i; + } + for (int j = 0; j <= ref_len; j++){ + errors_curr[err_idx(0, j)] = j; + } + for (int i = 1; i <= hyp_len; i++){ + for (int j = 1; j <= ref_len; j++){ + errors_curr[err_idx(i, j)] = min( + min( + errors_curr[err_idx(i-1, j)], + errors_curr[err_idx(i, j-1)] + ) + 1, + errors_curr[err_idx(i-1, j-1)] + 2 * ( + *(hyp_begin+i-1) == *(ref_begin+j-1) ? 0 : 1 + ) + ); + } + } + + // back-tracing + int i = hyp_len; + int j = ref_len; + int o = hyp_len + ref_len; + + for (int k = 0; k < source_size + target_size; k++) { + operations[opt_idx(k)] = 0; + } + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (errors_curr[err_idx(i, j-1)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 1; j--; // insertion + } else if ((i > 0) && (errors_curr[err_idx(i-1, j)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 2; i--; // deletion + } else { + o--; operations[opt_idx(o)] = 3; i--; j--; // do nothing + } + } + + // moving to the left + for (int k = 0; k < hyp_len + ref_len; k++) { + if (k + o < hyp_len + ref_len){ + operations[opt_idx(k)] = operations[opt_idx(k+o)]; + } else{ + operations[opt_idx(k)] = 0; // padding + } + } + +} + +template +__global__ void faster_levenshtein_distance_kernel( + const scalar_t* __restrict__ source, + const scalar_t* __restrict__ target, + const int* __restrict__ source_length, + const int* __restrict__ target_length, + const size_t source_size, + const size_t target_size, + int* __restrict__ operations) { + + extern __shared__ short errors[]; + auto errors_curr = errors; + + const int index = blockIdx.x; + const int offset = index * (source_size + target_size); + const int t = target_size + 1; + + auto err_idx = [t](int i, int j) { return i * t + j; }; + auto opt_idx = [offset](int k) { return offset + k; }; + + const int hyp_len = source_length[index]; + const int ref_len = target_length[index]; + const scalar_t* hyp_begin = source + index * source_size; + const scalar_t* ref_begin = target + index * target_size; + + // dynamic programming + for (int i = 0; i <= hyp_len; i++){ + errors_curr[err_idx(i, 0)] = i; + } + for (int j = 0; j <= ref_len; j++){ + errors_curr[err_idx(0, j)] = j; + } + for (int i = 1; i <= hyp_len; i++){ + for (int j = 1; j <= ref_len; j++){ + errors_curr[err_idx(i, j)] = min( + min( + errors_curr[err_idx(i-1, j)], + errors_curr[err_idx(i, j-1)] + ) + 1, + errors_curr[err_idx(i-1, j-1)] + 2 * ( + *(hyp_begin+i-1) == *(ref_begin+j-1) ? 0 : 1 + ) + ); + } + } + + // back-tracing + int i = hyp_len; + int j = ref_len; + int o = hyp_len + ref_len; + + for (int k = 0; k < source_size + target_size; k++) { + operations[opt_idx(k)] = 0; + } + + while ((i >= 0) && (j >= 0)) { + if ((i == 0) && (j == 0)) { + break; + } + + if ((j > 0) && (errors_curr[err_idx(i, j-1)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 1; j--; // insertion + } else if ((i > 0) && (errors_curr[err_idx(i-1, j)] < errors_curr[err_idx(i, j)])) { + o--; operations[opt_idx(o)] = 2; i--; // deletion + } else { + o--; operations[opt_idx(o)] = 3; i--; j--; // do nothing + } + } + + // moving to the left + for (int k = 0; k < hyp_len + ref_len; k++) { + if (k + o < hyp_len + ref_len){ + operations[opt_idx(k)] = operations[opt_idx(k+o)]; + } else{ + operations[opt_idx(k)] = 0; // padding + } + } + +} + + +torch::Tensor GenerateDeletionLabelCuda( + torch::Tensor source, + torch::Tensor operations) { + + const auto batch_size = source.size(0); + at::TensorOptions options(source.device()); + options = options.dtype(at::ScalarType::Int); + auto labels = torch::empty({batch_size, source.size(1)}, options); + auto stream = at::cuda::getCurrentCUDAStream(source.device().index()); + + AT_DISPATCH_ALL_TYPES(source.scalar_type(), "generate_deletion_labels", ([&] { + generate_deletion_label_kernel<<>>( + source.data(), + source.size(1), + operations.size(1), + operations.data(), + labels.data()); + })); + + return labels; +} + +std::pair GenerateInsertionLabelCuda( + torch::Tensor target, + torch::Tensor operations) { + +const auto batch_size = target.size(0); +at::TensorOptions options(target.device()); +options = options.dtype(at::ScalarType::Int); +auto labels = torch::empty({batch_size, target.size(1)}, options); +auto masks = torch::empty({batch_size, target.size(1)}, options); +auto stream = at::cuda::getCurrentCUDAStream(target.device().index()); + +AT_DISPATCH_ALL_TYPES(target.scalar_type(), "generate_insertion_labels", ([&] { + generate_insertion_label_kernel<<>>( + target.data(), + target.size(1), + operations.size(1), + operations.data(), + labels.data(), + masks.data()); +})); + +return std::make_pair(labels, masks); +} + + +torch::Tensor LevenshteinDistanceCuda( + torch::Tensor source, + torch::Tensor target, + torch::Tensor source_length, + torch::Tensor target_length) { + + const auto batch_size = source.size(0); + const auto shared_size = (source.size(1) + 1) * (target.size(1) + 1) * sizeof(short); + + at::TensorOptions options(source.device()); + options = options.dtype(at::ScalarType::Int); + auto operations = torch::empty({batch_size, source.size(1) + target.size(1)}, options); + auto stream = at::cuda::getCurrentCUDAStream(source.device().index()); + + if (shared_size > 40000) { + auto distances = torch::empty({batch_size, (source.size(1) + 1) * (target.size(1) + 1)}, options); + AT_DISPATCH_ALL_TYPES(source.scalar_type(), "levenshtein_distance", ([&] { + levenshtein_distance_kernel<<>>( + source.data(), + target.data(), + source_length.data(), + target_length.data(), + source.size(1), + target.size(1), + operations.data(), + distances.data()); + })); + } else { + AT_DISPATCH_ALL_TYPES(source.scalar_type(), "faster_levenshtein_distance", ([&] { + faster_levenshtein_distance_kernel<<>>( + source.data(), + target.data(), + source_length.data(), + target_length.data(), + source.size(1), + target.size(1), + operations.data()); + })); + } + + return operations; +} diff --git a/fairseq/clib/libnat_cuda/edit_dist.h b/fairseq/clib/libnat_cuda/edit_dist.h new file mode 100644 index 0000000000000000000000000000000000000000..e3506cd34ddaa35bb724fe64a459bad8046b9a34 --- /dev/null +++ b/fairseq/clib/libnat_cuda/edit_dist.h @@ -0,0 +1,25 @@ +/** + * Copyright 2017-present, Facebook, Inc. + * All rights reserved. + * + * This source code is licensed under the license found in the + * LICENSE file in the root directory of this source tree. + */ + +#pragma once + +#include + +torch::Tensor LevenshteinDistanceCuda( + torch::Tensor source, + torch::Tensor target, + torch::Tensor source_length, + torch::Tensor target_length); + +torch::Tensor GenerateDeletionLabelCuda( + torch::Tensor source, + torch::Tensor operations); + +std::pair GenerateInsertionLabelCuda( + torch::Tensor source, + torch::Tensor operations); diff --git a/fairseq/criterions/__init__.py b/fairseq/criterions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1c28780111e6015e8c274024325d4c73b3c6c84d --- /dev/null +++ b/fairseq/criterions/__init__.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + +from fairseq import registry +from fairseq.criterions.fairseq_criterion import FairseqCriterion, LegacyFairseqCriterion + + +build_criterion, register_criterion, CRITERION_REGISTRY = registry.setup_registry( + '--criterion', + base_class=FairseqCriterion, + default='cross_entropy', +) + + +# automatically import any Python files in the criterions/ directory +for file in os.listdir(os.path.dirname(__file__)): + if file.endswith('.py') and not file.startswith('_'): + module = file[:file.find('.py')] + importlib.import_module('fairseq.criterions.' + module) diff --git a/fairseq/criterions/__pycache__/__init__.cpython-310.pyc b/fairseq/criterions/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..62ede6bb72bf14120440823b21b2551c44d87d3f Binary files /dev/null and b/fairseq/criterions/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/adaptive_loss.cpython-310.pyc b/fairseq/criterions/__pycache__/adaptive_loss.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..20b19904d968b2a4bd571b3580546398efba81d4 Binary files /dev/null and b/fairseq/criterions/__pycache__/adaptive_loss.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/composite_loss.cpython-310.pyc b/fairseq/criterions/__pycache__/composite_loss.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f40665cbdb4b02c4991ae0569caaf3543f391258 Binary files /dev/null and b/fairseq/criterions/__pycache__/composite_loss.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/cross_entropy.cpython-310.pyc b/fairseq/criterions/__pycache__/cross_entropy.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fb1ca95b5e9090a4d7925e74d69fbcce07318d6c Binary files /dev/null and b/fairseq/criterions/__pycache__/cross_entropy.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/ctc.cpython-310.pyc b/fairseq/criterions/__pycache__/ctc.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b4fb18ca67d42b0f2ff3203c1539c252b0ad2edc Binary files /dev/null and b/fairseq/criterions/__pycache__/ctc.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/fairseq_criterion.cpython-310.pyc b/fairseq/criterions/__pycache__/fairseq_criterion.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5729dbcf9d0306170b752dbb47361f7765d86004 Binary files /dev/null and b/fairseq/criterions/__pycache__/fairseq_criterion.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/label_smoothed_cross_entropy.cpython-310.pyc b/fairseq/criterions/__pycache__/label_smoothed_cross_entropy.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5dc0d09562c28e2282b353d1e654eb35f9b8bdc8 Binary files /dev/null and b/fairseq/criterions/__pycache__/label_smoothed_cross_entropy.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_alignment.cpython-310.pyc b/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_alignment.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0d594d3e8c6b25993806c059404ed0dff07f8a0d Binary files /dev/null and b/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_alignment.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/legacy_masked_lm.cpython-310.pyc b/fairseq/criterions/__pycache__/legacy_masked_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c348ab7b5fc64177f53b478a53631fd0b41cbf09 Binary files /dev/null and b/fairseq/criterions/__pycache__/legacy_masked_lm.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/masked_lm.cpython-310.pyc b/fairseq/criterions/__pycache__/masked_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..60686443f3a70397392570b43c279a612119e80e Binary files /dev/null and b/fairseq/criterions/__pycache__/masked_lm.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/nat_loss.cpython-310.pyc b/fairseq/criterions/__pycache__/nat_loss.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6490ac0dd643073634a95712bc5af09181133ac4 Binary files /dev/null and b/fairseq/criterions/__pycache__/nat_loss.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/sentence_prediction.cpython-310.pyc b/fairseq/criterions/__pycache__/sentence_prediction.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d615526f8f540d39dd347d3dd77cffce00100578 Binary files /dev/null and b/fairseq/criterions/__pycache__/sentence_prediction.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/sentence_ranking.cpython-310.pyc b/fairseq/criterions/__pycache__/sentence_ranking.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e0fd7b406e575fd7668e8874982e9e33da92e13f Binary files /dev/null and b/fairseq/criterions/__pycache__/sentence_ranking.cpython-310.pyc differ diff --git a/fairseq/criterions/__pycache__/wav2vec_criterion.cpython-310.pyc b/fairseq/criterions/__pycache__/wav2vec_criterion.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8aa015f1147a952453db5c632137bdaaa5920b36 Binary files /dev/null and b/fairseq/criterions/__pycache__/wav2vec_criterion.cpython-310.pyc differ diff --git a/fairseq/criterions/adaptive_loss.py b/fairseq/criterions/adaptive_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..1916131bb133e0b00d21730c4a386ba93483978d --- /dev/null +++ b/fairseq/criterions/adaptive_loss.py @@ -0,0 +1,101 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch.nn.functional as F + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion('adaptive_loss') +class AdaptiveLoss(FairseqCriterion): + """This is an implementation of the loss function accompanying the adaptive softmax approximation for + graphical processing units (GPU), described in the paper "Efficient softmax approximation for GPUs" + (http://arxiv.org/abs/1609.04309).""" + + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + + @classmethod + def build_criterion(cls, args, task): + if getattr(args, 'ddp_backend', None) == 'c10d': + raise Exception( + 'AdaptiveLoss is not compatible with the c10d ' + 'version of DistributedDataParallel. Please use ' + '`--ddp-backend=no_c10d` instead.' + ) + return cls(task, args.sentence_avg) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + + assert hasattr(model.decoder, 'adaptive_softmax') and model.decoder.adaptive_softmax is not None + adaptive_softmax = model.decoder.adaptive_softmax + + net_output = model(**sample['net_input']) + orig_target = model.get_targets(sample, net_output) + + nsentences = orig_target.size(0) + orig_target = orig_target.view(-1) + + bsz = orig_target.size(0) + + logits, target = adaptive_softmax(net_output[0], orig_target) + assert len(target) == len(logits) + + loss = net_output[0].new(1 if reduce else bsz).zero_() + + for i in range(len(target)): + if target[i] is not None: + assert (target[i].min() >= 0 and target[i].max() <= logits[i].size(1)) + loss += F.cross_entropy( + logits[i], + target[i], + ignore_index=self.padding_idx, + reduction='sum' if reduce else 'none', + ) + + orig = utils.strip_pad(orig_target, self.padding_idx) + ntokens = orig.numel() + sample_size = sample['target'].size(0) if self.sentence_avg else ntokens + logging_output = { + 'loss': loss.data, + 'ntokens': ntokens, + 'nsentences': nsentences, + 'sample_size': sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get('loss', 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get('ntokens', 0) for log in logging_outputs)) + sample_size = utils.item(sum(log.get('sample_size', 0) for log in logging_outputs)) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + if sample_size != ntokens: + metrics.log_scalar('nll_loss', loss_sum / ntokens / math.log(2), ntokens, round=3) + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['nll_loss'].avg)) + else: + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg)) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/composite_loss.py b/fairseq/criterions/composite_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..6671c696e97b9f9240b96d9f9da4157bcdbd74de --- /dev/null +++ b/fairseq/criterions/composite_loss.py @@ -0,0 +1,99 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from torch import nn + +from fairseq import utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion('composite_loss') +class CompositeLoss(FairseqCriterion): + """This is a composite loss that, given a list of model outputs and a list of targets, + computes an average of losses for each output-target pair""" + + def __init__(self, task, underlying_criterion): + super().__init__(task) + self.underlying_criterion = underlying_criterion + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--underlying-criterion', type=str, metavar='VAL', required=True, + help='underlying criterion to use for the composite loss') + # fmt: on + + @staticmethod + def build_underlying_criterion(args, task): + saved_criterion = args.criterion + args.criterion = args.underlying_criterion + assert saved_criterion != args.underlying_criterion + underlying_criterion = task.build_criterion(args) + args.criterion = saved_criterion + return underlying_criterion + + @classmethod + def build_criterion(cls, args, task): + underlying_criterion = CompositeLoss.build_underlying_criterion(args, task) + + class FakeModel(nn.Module): + + def __init__(self, model, net_out, target): + super().__init__() + self.model = model + self.net_out = net_out + self.target = target + + def forward(self, **unused): + return self.net_out + + def get_normalized_probs(self, net_output, log_probs, sample=None): + return self.model.get_normalized_probs(net_output, log_probs, sample=sample) + + def get_targets(self, *unused): + return self.target + + @property + def decoder(self): + return self.model.decoder + + class _CompositeLoss(FairseqCriterion): + + def __init__(self, task, underlying_criterion): + super().__init__(task) + self.underlying_criterion = underlying_criterion + + def forward(self, model, sample, reduce=True): + net_outputs = model(**sample['net_input']) + targets = sample['target'] + + bsz = targets[0].size(0) + loss = net_outputs[0][0].new(1 if reduce else bsz).float().zero_() + + sample_size = 0 + logging_output = {} + for o, t in zip(net_outputs[0], targets): + m = FakeModel(model, (o, net_outputs[1]), t) + sample['target'] = t + l, ss, logging_output = self.underlying_criterion(m, sample, reduce) + loss += l + sample_size += ss + + loss.div_(len(targets)) + sample_size /= len(targets) + + logging_output['loss'] = utils.item(loss.data) if reduce else loss.data + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + return underlying_criterion.__class__.aggregate_logging_outputs(logging_outputs) + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + underlying_criterion.__class__.reduce_metrics(logging_outputs) + + return _CompositeLoss(task, underlying_criterion) diff --git a/fairseq/criterions/cross_entropy.py b/fairseq/criterions/cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..4e750f62e3f252e1ad4e780bd5168bf1c9924de5 --- /dev/null +++ b/fairseq/criterions/cross_entropy.py @@ -0,0 +1,73 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch.nn.functional as F + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion('cross_entropy') +class CrossEntropyCriterion(FairseqCriterion): + + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample['net_input']) + loss, _ = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = sample['target'].size(0) if self.sentence_avg else sample['ntokens'] + logging_output = { + 'loss': loss.data, + 'ntokens': sample['ntokens'], + 'nsentences': sample['target'].size(0), + 'sample_size': sample_size, + } + return loss, sample_size, logging_output + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + lprobs = lprobs.view(-1, lprobs.size(-1)) + target = model.get_targets(sample, net_output).view(-1) + loss = F.nll_loss( + lprobs, + target, + ignore_index=self.padding_idx, + reduction='sum' if reduce else 'none', + ) + return loss, loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get('loss', 0) for log in logging_outputs) + ntokens = sum(log.get('ntokens', 0) for log in logging_outputs) + sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + if sample_size != ntokens: + metrics.log_scalar('nll_loss', loss_sum / ntokens / math.log(2), ntokens, round=3) + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['nll_loss'].avg)) + else: + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg)) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/ctc.py b/fairseq/criterions/ctc.py new file mode 100644 index 0000000000000000000000000000000000000000..cbf712c69d3b990ce31a6031a0b248b2423e5817 --- /dev/null +++ b/fairseq/criterions/ctc.py @@ -0,0 +1,247 @@ +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +from argparse import Namespace +import math + +import torch +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.data.data_utils import post_process +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.logging.meters import safe_round + + +@register_criterion("ctc") +class CtcCriterion(FairseqCriterion): + def __init__(self, task, wer_args, zero_infinity, sentence_avg, remove_bpe): + super().__init__(task) + self.blank_idx = task.target_dictionary.bos() + self.pad_idx = task.target_dictionary.pad() + self.eos_idx = task.target_dictionary.eos() + self.post_process = remove_bpe if remove_bpe else "letter" + + if wer_args is not None: + from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder + + wer_compute_kenlm, wer_lexicon, lm_w, ws_w = eval(wer_args) + + dec_args = Namespace() + dec_args.nbest = 1 + dec_args.criterion = "ctc" + dec_args.kenlm_model = wer_compute_kenlm + dec_args.lexicon = wer_lexicon + dec_args.beam = 50 + dec_args.beam_size_token = min(50, len(task.target_dictionary)) + dec_args.beam_threshold = min(50, len(task.target_dictionary)) + dec_args.lm_weight = lm_w + dec_args.word_score = ws_w + dec_args.unk_weight = -math.inf + dec_args.sil_weight = 0 + + self.w2l_decoder = W2lKenLMDecoder(dec_args, task.target_dictionary) + else: + self.w2l_decoder = None + + self.zero_infinity = zero_infinity + self.sentence_avg = sentence_avg + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + parser.add_argument( + "--zero-infinity", action="store_true", help="zero inf loss" + ) + try: + parser.add_argument( + "--remove-bpe", + "--post-process", + default="letter", + help="remove BPE tokens before scoring (can be set to sentencepiece, letter, and more)", + ) + except: + pass # this option might have been added from eval args + parser.add_argument( + "--wer-args", + type=str, + default=None, + help="options for wer computation on valid set using 4 gram lm. this should be a tuple of 4 elements: path to 4-gram lm, \ + path to lexicon, lm score, word score", + ) + + def forward(self, model, sample, reduce=True): + net_output = model(**sample["net_input"]) + lprobs = model.get_normalized_probs( + net_output, log_probs=True + ).contiguous() # (T, B, C) from the encoder + + if "src_lengths" in sample["net_input"]: + input_lengths = sample["net_input"]["src_lengths"] + else: + non_padding_mask = ~net_output["padding_mask"] + input_lengths = non_padding_mask.long().sum(-1) + + pad_mask = (sample["target"] != self.pad_idx) & ( + sample["target"] != self.eos_idx + ) + targets_flat = sample["target"].masked_select(pad_mask) + target_lengths = sample["target_lengths"] + + with torch.backends.cudnn.flags(enabled=False): + loss = F.ctc_loss( + lprobs, + targets_flat, + input_lengths, + target_lengths, + blank=self.blank_idx, + reduction="sum", + zero_infinity=self.zero_infinity, + ) + + ntokens = ( + sample["ntokens"] if "ntokens" in sample else target_lengths.sum().item() + ) + + sample_size = sample["target"].size(0) if self.sentence_avg else ntokens + logging_output = { + "loss": utils.item(loss.data), # * sample['ntokens'], + "ntokens": ntokens, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + } + + if not model.training: + import editdistance + + with torch.no_grad(): + lprobs_t = lprobs.transpose(0, 1).float().cpu() + + c_err = 0 + c_len = 0 + w_errs = 0 + w_len = 0 + wv_errs = 0 + for lp, t, inp_l in zip( + lprobs_t, + sample["target_label"] + if "target_label" in sample + else sample["target"], + input_lengths, + ): + lp = lp[:inp_l].unsqueeze(0) + + decoded = None + if self.w2l_decoder is not None: + decoded = self.w2l_decoder.decode(lp) + if len(decoded) < 1: + decoded = None + else: + decoded = decoded[0] + if len(decoded) < 1: + decoded = None + else: + decoded = decoded[0] + + p = (t != self.task.target_dictionary.pad()) & ( + t != self.task.target_dictionary.eos() + ) + targ = t[p] + targ_units = self.task.target_dictionary.string(targ) + targ_units_arr = targ.tolist() + + toks = lp.argmax(dim=-1).unique_consecutive() + pred_units_arr = toks[toks != self.blank_idx].tolist() + + c_err += editdistance.eval(pred_units_arr, targ_units_arr) + c_len += len(targ_units_arr) + + targ_words = post_process(targ_units, self.post_process).split() + + pred_units = self.task.target_dictionary.string(pred_units_arr) + pred_words_raw = post_process(pred_units, self.post_process).split() + + if decoded is not None and "words" in decoded: + pred_words = decoded["words"] + w_errs += editdistance.eval(pred_words, targ_words) + wv_errs += editdistance.eval(pred_words_raw, targ_words) + else: + dist = editdistance.eval(pred_words_raw, targ_words) + w_errs += dist + wv_errs += dist + + w_len += len(targ_words) + + logging_output["wv_errors"] = wv_errs + logging_output["w_errors"] = w_errs + logging_output["w_total"] = w_len + logging_output["c_errors"] = c_err + logging_output["c_total"] = c_len + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + nsentences = utils.item( + sum(log.get("nsentences", 0) for log in logging_outputs) + ) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar("ntokens", ntokens) + metrics.log_scalar("nsentences", nsentences) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + + c_errors = sum(log.get("c_errors", 0) for log in logging_outputs) + metrics.log_scalar("_c_errors", c_errors) + c_total = sum(log.get("c_total", 0) for log in logging_outputs) + metrics.log_scalar("_c_total", c_total) + w_errors = sum(log.get("w_errors", 0) for log in logging_outputs) + metrics.log_scalar("_w_errors", w_errors) + wv_errors = sum(log.get("wv_errors", 0) for log in logging_outputs) + metrics.log_scalar("_wv_errors", wv_errors) + w_total = sum(log.get("w_total", 0) for log in logging_outputs) + metrics.log_scalar("_w_total", w_total) + + if c_total > 0: + metrics.log_derived( + "uer", + lambda meters: safe_round(meters["_c_errors"].sum * 100.0 / meters["_c_total"].sum, 3) + if meters["_c_total"].sum > 0 + else float("nan"), + ) + if w_total > 0: + metrics.log_derived( + "wer", + lambda meters: safe_round(meters["_w_errors"].sum * 100.0 / meters["_w_total"].sum, 3) + if meters["_w_total"].sum > 0 + else float("nan"), + ) + metrics.log_derived( + "raw_wer", + lambda meters: safe_round(meters["_wv_errors"].sum * 100.0 / meters["_w_total"].sum, 3) + if meters["_w_total"].sum > 0 + else float("nan"), + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/fairseq_criterion.py b/fairseq/criterions/fairseq_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..9873574d474f03713b75fe22dae4302716cda467 --- /dev/null +++ b/fairseq/criterions/fairseq_criterion.py @@ -0,0 +1,119 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import inspect +from typing import Any, Dict, List + +from torch.nn.modules.loss import _Loss + +from fairseq import metrics, utils + + +class FairseqCriterion(_Loss): + + def __init__(self, task): + super().__init__() + self.task = task + if hasattr(task, 'target_dictionary'): + tgt_dict = task.target_dictionary + self.padding_idx = tgt_dict.pad() if tgt_dict is not None else -100 + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + pass + + @classmethod + def build_criterion(cls, args, task): + """Construct a criterion from command-line args.""" + # Criterions can override this, but for convenience we also try + # to automatically map argparse.Namespace keys to corresponding + # arguments in the __init__. + init_args = {} + for p in inspect.signature(cls).parameters.values(): + if ( + p.kind == p.POSITIONAL_ONLY + or p.kind == p.VAR_POSITIONAL + or p.kind == p.VAR_KEYWORD + ): + # we haven't implemented inference for these argument types, + # but PRs welcome :) + raise NotImplementedError('{} not supported'.format(p.kind)) + + assert p.kind in {p.POSITIONAL_OR_KEYWORD, p.KEYWORD_ONLY} + + if p.name == 'task': + init_args['task'] = task + elif hasattr(args, p.name): + init_args[p.name] = getattr(args, p.name) + elif p.default != p.empty: + pass # we'll use the default value + else: + raise NotImplementedError( + 'Unable to infer Criterion arguments, please implement ' + '{}.build_criterion'.format(cls.__name__) + ) + return cls(**init_args) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + raise NotImplementedError + + @staticmethod + def aggregate_logging_outputs( + logging_outputs: List[Dict[str, Any]], + ) -> Dict[str, Any]: + """Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + 'The aggregate_logging_outputs API is deprecated. ' + 'Please use the reduce_metrics API instead.' + ) + raise NotImplementedError + + @classmethod + def reduce_metrics(cls, logging_outputs: List[Dict[str, Any]]) -> None: + """Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + 'Criterions should implement the reduce_metrics API. ' + 'Falling back to deprecated aggregate_logging_outputs API.' + ) + agg_logging_outputs = cls.aggregate_logging_outputs(logging_outputs) + for k, v in agg_logging_outputs.items(): + if k in {'nsentences', 'ntokens', 'sample_size'}: + continue + metrics.log_scalar(k, v) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return False + + +class LegacyFairseqCriterion(FairseqCriterion): + + def __init__(self, args, task): + super().__init__(task=task) + self.args = args + + utils.deprecation_warning( + 'Criterions should take explicit arguments instead of an ' + 'argparse.Namespace object, please update your criterion by ' + 'extending FairseqCriterion instead of LegacyFairseqCriterion.' + ) + + @classmethod + def build_criterion(cls, args, task): + """Construct a criterion from command-line args.""" + return cls(args, task) diff --git a/fairseq/criterions/label_smoothed_cross_entropy.py b/fairseq/criterions/label_smoothed_cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..d010c3d03d82fffb6d5d6d0e346b850564534134 --- /dev/null +++ b/fairseq/criterions/label_smoothed_cross_entropy.py @@ -0,0 +1,96 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +def label_smoothed_nll_loss(lprobs, target, epsilon, ignore_index=None, reduce=True): + if target.dim() == lprobs.dim() - 1: + target = target.unsqueeze(-1) + nll_loss = -lprobs.gather(dim=-1, index=target) + smooth_loss = -lprobs.sum(dim=-1, keepdim=True) + if ignore_index is not None: + pad_mask = target.eq(ignore_index) + nll_loss.masked_fill_(pad_mask, 0.) + smooth_loss.masked_fill_(pad_mask, 0.) + else: + nll_loss = nll_loss.squeeze(-1) + smooth_loss = smooth_loss.squeeze(-1) + if reduce: + nll_loss = nll_loss.sum() + smooth_loss = smooth_loss.sum() + eps_i = epsilon / lprobs.size(-1) + loss = (1. - epsilon) * nll_loss + eps_i * smooth_loss + return loss, nll_loss + + +@register_criterion('label_smoothed_cross_entropy') +class LabelSmoothedCrossEntropyCriterion(FairseqCriterion): + + def __init__(self, task, sentence_avg, label_smoothing): + super().__init__(task) + self.sentence_avg = sentence_avg + self.eps = label_smoothing + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--label-smoothing', default=0., type=float, metavar='D', + help='epsilon for label smoothing, 0 means no label smoothing') + # fmt: on + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample['net_input']) + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = sample['target'].size(0) if self.sentence_avg else sample['ntokens'] + logging_output = { + 'loss': loss.data, + 'nll_loss': nll_loss.data, + 'ntokens': sample['ntokens'], + 'nsentences': sample['target'].size(0), + 'sample_size': sample_size, + } + return loss, sample_size, logging_output + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + lprobs = lprobs.view(-1, lprobs.size(-1)) + target = model.get_targets(sample, net_output).view(-1, 1) + loss, nll_loss = label_smoothed_nll_loss( + lprobs, target, self.eps, ignore_index=self.padding_idx, reduce=reduce, + ) + return loss, nll_loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get('loss', 0) for log in logging_outputs) + nll_loss_sum = sum(log.get('nll_loss', 0) for log in logging_outputs) + ntokens = sum(log.get('ntokens', 0) for log in logging_outputs) + sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + metrics.log_scalar('nll_loss', nll_loss_sum / ntokens / math.log(2), ntokens, round=3) + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['nll_loss'].avg)) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py b/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py new file mode 100644 index 0000000000000000000000000000000000000000..cfc7e008cd5387cbdcc835cabb1f8eaa18851064 --- /dev/null +++ b/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py @@ -0,0 +1,97 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from fairseq import metrics, utils +from fairseq.criterions import register_criterion + +from .label_smoothed_cross_entropy import LabelSmoothedCrossEntropyCriterion + + +@register_criterion('label_smoothed_cross_entropy_with_alignment') +class LabelSmoothedCrossEntropyCriterionWithAlignment(LabelSmoothedCrossEntropyCriterion): + + def __init__(self, task, sentence_avg, label_smoothing, alignment_lambda): + super().__init__(task, sentence_avg, label_smoothing) + self.alignment_lambda = alignment_lambda + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + LabelSmoothedCrossEntropyCriterion.add_args(parser) + parser.add_argument('--alignment-lambda', default=0.05, type=float, metavar='D', + help='weight for the alignment loss') + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample['net_input']) + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = sample['target'].size(0) if self.sentence_avg else sample['ntokens'] + logging_output = { + 'loss': utils.item(loss.data) if reduce else loss.data, + 'nll_loss': utils.item(nll_loss.data) if reduce else nll_loss.data, + 'ntokens': sample['ntokens'], + 'nsentences': sample['target'].size(0), + 'sample_size': sample_size, + } + + alignment_loss = None + + # Compute alignment loss only for training set and non dummy batches. + if 'alignments' in sample and sample['alignments'] is not None: + alignment_loss = self.compute_alignment_loss(sample, net_output) + + if alignment_loss is not None: + logging_output['alignment_loss'] = utils.item(alignment_loss.data) + loss += self.alignment_lambda * alignment_loss + + return loss, sample_size, logging_output + + def compute_alignment_loss(self, sample, net_output): + attn_prob = net_output[1]['attn'][0] + bsz, tgt_sz, src_sz = attn_prob.shape + attn = attn_prob.view(bsz * tgt_sz, src_sz) + + align = sample['alignments'] + align_weights = sample['align_weights'].float() + + if len(align) > 0: + # Alignment loss computation. align (shape [:, 2]) contains the src-tgt index pairs corresponding to + # the alignments. align_weights (shape [:]) contains the 1 / frequency of a tgt index for normalizing. + loss = -((attn[align[:, 1][:, None], align[:, 0][:, None]]).log() * align_weights[:, None]).sum() + else: + return None + + return loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get('loss', 0) for log in logging_outputs)) + nll_loss_sum = utils.item(sum(log.get('nll_loss', 0) for log in logging_outputs)) + alignment_loss_sum = utils.item(sum(log.get('alignment_loss', 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get('ntokens', 0) for log in logging_outputs)) + sample_size = utils.item(sum(log.get('sample_size', 0) for log in logging_outputs)) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + metrics.log_scalar('nll_loss', nll_loss_sum / ntokens / math.log(2), ntokens, round=3) + metrics.log_scalar('alignment_loss', alignment_loss_sum / sample_size / math.log(2), sample_size, round=3) + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['nll_loss'].avg)) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/legacy_masked_lm.py b/fairseq/criterions/legacy_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..10dea76e4b1a6151ff21174e6a0333f636440d45 --- /dev/null +++ b/fairseq/criterions/legacy_masked_lm.py @@ -0,0 +1,158 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F + +from fairseq import utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +def compute_cross_entropy_loss(logits, targets, ignore_index=-100): + """ + Function to compute the cross entropy loss. The default value of + ignore_index is the same as the default value for F.cross_entropy in + pytorch. + """ + assert logits.size(0) == targets.size(-1), \ + "Logits and Targets tensor shapes don't match up" + + loss = F.nll_loss( + F.log_softmax(logits, -1, dtype=torch.float32), + targets, + reduction="sum", + ignore_index=ignore_index, + ) + return loss + + +@register_criterion('legacy_masked_lm_loss') +class LegacyMaskedLmLoss(FairseqCriterion): + """ + Implementation for the loss used in masked language model (MLM) training. + This optionally also computes the next sentence prediction (NSP) loss and + adds it to the overall loss based on the specified args. There are three + cases to consider: + 1) Generic MLM training without NSP loss. In this case sentence_targets + and sentence_logits are both None. + 2) BERT training without NSP loss. In this case sentence_targets is + not None but sentence_logits is None and we should not be computing + a sentence level loss. + 3) BERT training with NSP loss. In this case both sentence_targets and + sentence_logits are not None and we should be computing a sentence + level loss. The weight of the sentence level loss is specified as + an argument. + """ + + def __init__(self, task, masked_lm_only, nsp_loss_weight): + super().__init__(task) + self.masked_lm_only = masked_lm_only + self.nsp_loss_weight = nsp_loss_weight + + @staticmethod + def add_args(parser): + """Args for MaskedLM Loss""" + # Default for masked_lm_only is False so as to not break BERT training + parser.add_argument('--masked-lm-only', default=False, + action='store_true', help='compute MLM loss only') + parser.add_argument('--nsp-loss-weight', default=1.0, type=float, + help='weight for next sentence prediction' + ' loss (default 1)') + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + lm_logits, output_metadata = model(**sample["net_input"]) + + # reshape lm_logits from (N,T,C) to (N*T,C) + lm_logits = lm_logits.view(-1, lm_logits.size(-1)) + lm_targets = sample['lm_target'].view(-1) + lm_loss = compute_cross_entropy_loss( + lm_logits, lm_targets, self.padding_idx) + + # compute the number of tokens for which loss is computed. This is used + # to normalize the loss + ntokens = utils.strip_pad(lm_targets, self.padding_idx).numel() + loss = lm_loss / ntokens + nsentences = sample['nsentences'] + # nsentences = 0 + + # Compute sentence loss if masked_lm_only is False + sentence_loss = None + if not self.masked_lm_only: + sentence_logits = output_metadata['sentence_logits'] + sentence_targets = sample['sentence_target'].view(-1) + # This needs to be recomputed due to some differences between + # TokenBlock and BlockPair dataset. This can be resolved with a + # refactor of BERTModel which we will do in the future. + # TODO: Remove this after refactor of BERTModel + nsentences = sentence_targets.size(0) + + # Check for logits being none which can happen when remove_heads + # is set to true in the BERT model. Ideally we should set + # masked_lm_only to true in this case, but that requires some + # refactor in the BERT model. + if sentence_logits is not None: + sentence_loss = compute_cross_entropy_loss( + sentence_logits, sentence_targets) + + loss += self.nsp_loss_weight * (sentence_loss / nsentences) + + # NOTE: as we are summing up per token mlm loss and per sentence nsp loss + # we don't need to use sample_size as denominator for the gradient + # here sample_size is just used for logging + sample_size = 1 + logging_output = { + 'loss': utils.item(loss.data) if reduce else loss.data, + 'lm_loss': utils.item(lm_loss.data) if reduce else lm_loss.data, + # sentence loss is not always computed + 'sentence_loss': ( + ( + utils.item(sentence_loss.data) if reduce + else sentence_loss.data + ) if sentence_loss is not None else 0.0 + ), + 'ntokens': ntokens, + 'nsentences': nsentences, + 'sample_size': sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + """Aggregate logging outputs from data parallel training.""" + lm_loss_sum = sum(log.get('lm_loss', 0) for log in logging_outputs) + sentence_loss_sum = sum( + log.get('sentence_loss', 0) for log in logging_outputs) + ntokens = sum(log.get('ntokens', 0) for log in logging_outputs) + nsentences = sum(log.get('nsentences', 0) for log in logging_outputs) + sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) + agg_loss = sum(log.get('loss', 0) for log in logging_outputs) + + agg_output = { + 'loss': agg_loss / sample_size / math.log(2) if sample_size > 0 else 0., + 'lm_loss': lm_loss_sum / ntokens / math.log(2) if ntokens > 0 else 0., + 'sentence_loss': sentence_loss_sum / nsentences / math.log(2) if nsentences > 0 else 0., + 'nll_loss': lm_loss_sum / ntokens / math.log(2) if ntokens > 0 else 0., + 'ntokens': ntokens, + 'nsentences': nsentences, + 'sample_size': sample_size, + } + return agg_output + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/masked_lm.py b/fairseq/criterions/masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..80864693ecf7935443a7f14f6e3f65b4a334d1cb --- /dev/null +++ b/fairseq/criterions/masked_lm.py @@ -0,0 +1,88 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F + +from fairseq import metrics, modules, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion('masked_lm') +class MaskedLmLoss(FairseqCriterion): + """ + Implementation for the loss used in masked language model (MLM) training. + """ + + def __init__(self, task, tpu): + super().__init__(task) + self.tpu = tpu + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + masked_tokens = sample['target'].ne(self.padding_idx) + sample_size = masked_tokens.int().sum() + + # Rare: when all tokens are masked, project all tokens. + # We use torch.where to avoid device-to-host transfers, + # except on CPU where torch.where is not well supported + # (see github.com/pytorch/pytorch/issues/26247). + if self.tpu: + masked_tokens = None # always project all tokens on TPU + elif masked_tokens.device == torch.device('cpu'): + if not masked_tokens.any(): + masked_tokens = None + else: + masked_tokens = torch.where( + masked_tokens.any(), + masked_tokens, + masked_tokens.new([True]), + ) + + logits = model(**sample['net_input'], masked_tokens=masked_tokens)[0] + targets = model.get_targets(sample, [logits]) + if masked_tokens is not None: + targets = targets[masked_tokens] + + loss = modules.cross_entropy( + logits.view(-1, logits.size(-1)), + targets.view(-1), + reduction='sum', + ignore_index=self.padding_idx, + ) + + logging_output = { + 'loss': loss if self.tpu else loss.data, + 'ntokens': sample['ntokens'], + 'nsentences': sample['nsentences'], + 'sample_size': sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get('loss', 0) for log in logging_outputs) + sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg)) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/nat_loss.py b/fairseq/criterions/nat_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..3326734d5535c72f4644ebe0e3a4b2966477dc94 --- /dev/null +++ b/fairseq/criterions/nat_loss.py @@ -0,0 +1,173 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch.nn.functional as F +import torch +from torch import Tensor + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("nat_loss") +class LabelSmoothedDualImitationCriterion(FairseqCriterion): + + def __init__(self, task, label_smoothing): + super().__init__(task) + self.label_smoothing = label_smoothing + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + parser.add_argument( + '--label-smoothing', + default=0., + type=float, + metavar='D', + help='epsilon for label smoothing, 0 means no label smoothing', + ) + + def _compute_loss( + self, outputs, targets, masks=None, label_smoothing=0.0, name="loss", factor=1.0 + ): + """ + outputs: batch x len x d_model + targets: batch x len + masks: batch x len + + policy_logprob: if there is some policy + depends on the likelihood score as rewards. + """ + + def mean_ds(x: Tensor, dim=None) -> Tensor: + return ( + x.float().mean().type_as(x) + if dim is None + else x.float().mean(dim).type_as(x) + ) + if masks is not None: + outputs, targets = outputs[masks], targets[masks] + + if masks is not None and not masks.any(): + nll_loss = torch.tensor(0) + loss = nll_loss + else: + logits = F.log_softmax(outputs, dim=-1) + if targets.dim() == 1: + losses = F.nll_loss(logits, targets.to(logits.device), reduction='none') + + else: # soft-labels + losses = F.kl_div(logits, targets.to(logits.device), reduction='none') + losses = losses.sum(-1) + + nll_loss = mean_ds(losses) + if label_smoothing > 0: + loss = nll_loss * ( + 1 - label_smoothing) - mean_ds(logits) * label_smoothing + else: + loss = nll_loss + + loss = loss * factor + return {"name": name, "loss": loss, "nll_loss": nll_loss, "factor": factor} + + def _custom_loss(self, loss, name="loss", factor=1.0): + return {"name": name, "loss": loss, "factor": factor} + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + nsentences, ntokens = sample["nsentences"], sample["ntokens"] + + # B x T + src_tokens, src_lengths = ( + sample["net_input"]["src_tokens"], + sample["net_input"]["src_lengths"], + ) + tgt_tokens, prev_output_tokens = sample["target"], sample["prev_target"] + + outputs = model(src_tokens, src_lengths, prev_output_tokens, tgt_tokens) + losses, nll_loss = [], [] + + for obj in outputs: + if outputs[obj].get("loss", None) is None: + _losses = self._compute_loss( + outputs[obj].get("out"), + outputs[obj].get("tgt"), + outputs[obj].get("mask", None), + outputs[obj].get("ls", 0.0), + name=obj + '-loss', + factor=outputs[obj].get("factor", 1.0) + ) + else: + _losses = self._custom_loss( + outputs[obj].get("loss"), + name=obj + '-loss', + factor=outputs[obj].get("factor", 1.0) + ) + + losses += [_losses] + if outputs[obj].get("nll_loss", False): + nll_loss += [_losses.get("nll_loss", 0.0)] + + loss = sum(l["loss"] for l in losses) + nll_loss = sum(l for l in nll_loss) if len(nll_loss) > 0 \ + else loss.new_tensor(0) + + # NOTE: + # we don't need to use sample_size as denominator for the gradient + # here sample_size is just used for logging + sample_size = 1 + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + + for l in losses: + logging_output[l["name"]] = ( + utils.item(l["loss"].data / l["factor"]) + if reduce + else l[["loss"]].data / l["factor"] + ) + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + sample_size = utils.item(sum(log.get("sample_size", 0) for log in logging_outputs)) + loss = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + nll_loss = utils.item(sum(log.get("nll_loss", 0) for log in logging_outputs)) + + metrics.log_scalar('loss', loss / sample_size / math.log(2), sample_size, round=3) + metrics.log_scalar('nll_loss', nll_loss / sample_size / math.log(2), sample_size, round=3) + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg)) + + for key in logging_outputs[0]: + if key[-5:] == "-loss": + val = sum(log.get(key, 0) for log in logging_outputs) + metrics.log_scalar( + key[:-5], + val / sample_size / math.log(2) if sample_size > 0 else 0.0, + sample_size, + round=3, + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/sentence_prediction.py b/fairseq/criterions/sentence_prediction.py new file mode 100644 index 0000000000000000000000000000000000000000..4ba13178560e1fc8db4435b07dc46033bd478031 --- /dev/null +++ b/fairseq/criterions/sentence_prediction.py @@ -0,0 +1,95 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion('sentence_prediction') +class SentencePredictionCriterion(FairseqCriterion): + + def __init__(self, task, classification_head_name, regression_target): + super().__init__(task) + self.classification_head_name = classification_head_name + self.regression_target = regression_target + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--classification-head-name', + default='sentence_classification_head', + help='name of the classification head to use') + # fmt: on + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, 'classification_heads') + and self.classification_head_name in model.classification_heads + ), 'model must provide sentence classification head for --criterion=sentence_prediction' + + logits, _ = model( + **sample['net_input'], + features_only=True, + classification_head_name=self.classification_head_name, + ) + targets = model.get_targets(sample, [logits]).view(-1) + sample_size = targets.numel() + + if not self.regression_target: + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + loss = F.nll_loss(lprobs, targets, reduction='sum') + else: + logits = logits.view(-1).float() + targets = targets.float() + loss = F.mse_loss(logits, targets, reduction='sum') + + logging_output = { + 'loss': loss.data, + 'ntokens': sample['ntokens'], + 'nsentences': sample_size, + 'sample_size': sample_size, + } + if not self.regression_target: + preds = logits.argmax(dim=1) + logging_output['ncorrect'] = (preds == targets).sum() + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get('loss', 0) for log in logging_outputs) + ntokens = sum(log.get('ntokens', 0) for log in logging_outputs) + nsentences = sum(log.get('nsentences', 0) for log in logging_outputs) + sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + if sample_size != ntokens: + metrics.log_scalar('nll_loss', loss_sum / ntokens / math.log(2), ntokens, round=3) + + if len(logging_outputs) > 0 and 'ncorrect' in logging_outputs[0]: + ncorrect = sum(log.get('ncorrect', 0) for log in logging_outputs) + metrics.log_scalar('accuracy', 100.0 * ncorrect / nsentences, nsentences, round=1) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/sentence_ranking.py b/fairseq/criterions/sentence_ranking.py new file mode 100644 index 0000000000000000000000000000000000000000..52a0a177d846553ca736026efc181e713fa6bd97 --- /dev/null +++ b/fairseq/criterions/sentence_ranking.py @@ -0,0 +1,116 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion('sentence_ranking') +class SentenceRankingCriterion(FairseqCriterion): + + def __init__(self, task, ranking_head_name, save_predictions, num_classes): + super().__init__(task) + self.ranking_head_name = ranking_head_name + if save_predictions is not None: + self.prediction_h = open(save_predictions, 'w') + else: + self.prediction_h = None + self.num_classes = num_classes + + def __del__(self): + if self.prediction_h is not None: + self.prediction_h.close() + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--save-predictions', metavar='FILE', + help='file to save predictions to') + parser.add_argument('--ranking-head-name', + default='sentence_classification_head', + help='name of the ranking head to use') + # fmt: on + + def forward(self, model, sample, reduce=True): + """Compute ranking loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, 'classification_heads') + and self.ranking_head_name in model.classification_heads + ), 'model must provide sentence ranking head for --criterion=sentence_ranking' + + scores = [] + for idx in range(self.num_classes): + score, _ = model( + **sample['net_input{idx}'.format(idx=idx+1)], + classification_head_name=self.ranking_head_name, + ) + scores.append(score) + + logits = torch.cat(scores, dim=1) + sample_size = logits.size(0) + + if 'target' in sample: + targets = model.get_targets(sample, [logits]).view(-1) + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + loss = F.nll_loss(lprobs, targets, reduction='sum') + else: + targets = None + loss = torch.tensor(0.0, requires_grad=True) + + if self.prediction_h is not None: + preds = logits.argmax(dim=1) + for i, (id, pred) in enumerate(zip(sample['id'].tolist(), preds.tolist())): + if targets is not None: + label = targets[i].item() + print('{}\t{}\t{}'.format(id, pred, label), file=self.prediction_h) + else: + print('{}\t{}'.format(id, pred), file=self.prediction_h) + + logging_output = { + 'loss': loss.data, + 'ntokens': sample['ntokens'], + 'nsentences': sample_size, + 'sample_size': sample_size, + } + if targets is not None: + logging_output['ncorrect'] = (logits.argmax(dim=1) == targets).sum() + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get('loss', 0) for log in logging_outputs) + ntokens = sum(log.get('ntokens', 0) for log in logging_outputs) + nsentences = sum(log.get('nsentences', 0) for log in logging_outputs) + sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + if sample_size != ntokens: + metrics.log_scalar('nll_loss', loss_sum / ntokens / math.log(2), ntokens, round=3) + + if len(logging_outputs) > 0 and 'ncorrect' in logging_outputs[0]: + ncorrect = sum(log.get('ncorrect', 0) for log in logging_outputs) + metrics.log_scalar('accuracy', 100.0 * ncorrect / nsentences, nsentences, round=1) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/criterions/wav2vec_criterion.py b/fairseq/criterions/wav2vec_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..019db622496cfb4a81170d12157db47a002912fe --- /dev/null +++ b/fairseq/criterions/wav2vec_criterion.py @@ -0,0 +1,157 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion('wav2vec') +class Wav2vecCriterion(FairseqCriterion): + + def __init__(self, task, infonce=False, loss_weights=None, log_keys=None): + super().__init__(task) + self.infonce = infonce + self.loss_weights = None if loss_weights is None else eval(loss_weights) + self.log_keys = [] if log_keys is None else eval(log_keys) + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--infonce', action='store_true', + help='if set, uses cross entropy instead of binary cross entropy (i.e. InfoNCE loss)') + parser.add_argument('--loss-weights', type=str, default=None, + help='weights for additional loss terms (not first one)') + parser.add_argument('--log-keys', type=str, default=None, + help='output keys to log') + + def forward(self, model, sample, reduce=True, log_pred=False): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample['net_input']) + logits = model.get_logits(net_output).float() + target = model.get_targets(sample, net_output) + + weights = None + if hasattr(model, 'get_target_weights') and not self.infonce: + weights = model.get_target_weights(target, net_output) + if torch.is_tensor(weights): + weights = weights.float() + + losses = [] + + if self.infonce: + loss = F.cross_entropy(logits, target, reduction="sum" if reduce else "none",) + else: + loss = F.binary_cross_entropy_with_logits(logits, target.float(), weights, reduction="sum" if reduce else "none",) + + sample_size = target.numel() if self.infonce else target.long().sum().item() + losses.append(loss) + + if self.loss_weights is not None: + assert hasattr(model, "get_extra_losses") + extra_losses = model.get_extra_losses(net_output) + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len(self.loss_weights), f'{len(extra_losses)}, {len(self.loss_weights)}' + for p, coef in zip(extra_losses, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + losses.append(p) + + logging_output = { + 'loss': loss.item() if reduce else loss, + 'ntokens': sample_size, + 'nsentences': sample['id'].numel(), + 'sample_size': sample_size, + } + + for lk in self.log_keys: + if lk in net_output: + logging_output[lk] = float((net_output[lk])) + + if len(losses) > 1: + for i, l in enumerate(losses): + logging_output[f'loss_{i}'] = l.item() + + if self.infonce: + with torch.no_grad(): + if logits.numel() == 0: + corr = 0 + count = 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == 0 + min = logits.argmin(-1) == 0 + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = max.numel() + + logging_output["correct"] = corr + logging_output["count"] = count + + if log_pred: + logging_output['logits'] = logits.cpu().numpy() + logging_output['target'] = target.cpu().numpy() + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get('loss', 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get('ntokens', 0) for log in logging_outputs)) + nsentences = utils.item(sum(log.get('nsentences', 0) for log in logging_outputs)) + sample_size = utils.item(sum(log.get('sample_size', 0) for log in logging_outputs)) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + metrics.log_scalar('ntokens', ntokens) + metrics.log_scalar('nsentences', nsentences) + + correct = sum(log.get("correct", 0) for log in logging_outputs) + metrics.log_scalar("_correct", correct) + + total = sum(log.get("count", 0) for log in logging_outputs) + metrics.log_scalar("_total", total) + + + if total > 0: + metrics.log_derived( + "accuracy", + lambda meters: round(meters["_correct"].sum / meters["_total"].sum, 5) + if meters["_total"].sum > 0 + else float("nan"), + ) + + builtin_keys = {'loss', 'ntokens', 'nsentences', 'sample_size', 'correct', 'count'} + + for k in logging_outputs[0]: + if k not in builtin_keys: + val = sum(log.get(k, 0) for log in logging_outputs) / len(logging_outputs) + if k.startswith('loss'): + metrics.log_scalar(k, val / sample_size / math.log(2), sample_size) + else: + metrics.log_scalar(k, val, round=3) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return False diff --git a/fairseq/data/__init__.py b/fairseq/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a99d9280fa1ecab0d8ba86983e0ca72dd012538e --- /dev/null +++ b/fairseq/data/__init__.py @@ -0,0 +1,114 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .dictionary import Dictionary, TruncatedDictionary + +from .fairseq_dataset import FairseqDataset, FairseqIterableDataset + +from .base_wrapper_dataset import BaseWrapperDataset + +from .add_target_dataset import AddTargetDataset +from .append_token_dataset import AppendTokenDataset +from .audio.raw_audio_dataset import FileAudioDataset +from .backtranslation_dataset import BacktranslationDataset +from .bucket_pad_length_dataset import BucketPadLengthDataset +from .colorize_dataset import ColorizeDataset +from .concat_dataset import ConcatDataset +from .concat_sentences_dataset import ConcatSentencesDataset +from .denoising_dataset import DenoisingDataset +from .id_dataset import IdDataset +from .indexed_dataset import IndexedCachedDataset, IndexedDataset, IndexedRawTextDataset, MMapIndexedDataset +from .language_pair_dataset import LanguagePairDataset +from .list_dataset import ListDataset +from .lm_context_window_dataset import LMContextWindowDataset +from .lru_cache_dataset import LRUCacheDataset +from .mask_tokens_dataset import MaskTokensDataset +from .monolingual_dataset import MonolingualDataset +from .multi_corpus_sampled_dataset import MultiCorpusSampledDataset +from .nested_dictionary_dataset import NestedDictionaryDataset +from .noising import NoisingDataset +from .numel_dataset import NumelDataset +from .num_samples_dataset import NumSamplesDataset +from .offset_tokens_dataset import OffsetTokensDataset +from .pad_dataset import LeftPadDataset, PadDataset, RightPadDataset +from .prepend_dataset import PrependDataset +from .prepend_token_dataset import PrependTokenDataset +from .raw_label_dataset import RawLabelDataset +from .replace_dataset import ReplaceDataset +from .resampling_dataset import ResamplingDataset +from .roll_dataset import RollDataset +from .round_robin_zip_datasets import RoundRobinZipDatasets +from .sort_dataset import SortDataset +from .strip_token_dataset import StripTokenDataset +from .subsample_dataset import SubsampleDataset +from .token_block_dataset import TokenBlockDataset +from .transform_eos_dataset import TransformEosDataset +from .transform_eos_lang_pair_dataset import TransformEosLangPairDataset +from .shorten_dataset import TruncateDataset, RandomCropDataset +from .multilingual.sampled_multi_dataset import SampledMultiDataset +from .multilingual.sampled_multi_epoch_dataset import SampledMultiEpochDataset +from .iterators import ( + CountingIterator, + EpochBatchIterator, + GroupedIterator, + ShardedIterator, +) + +__all__ = [ + 'AddTargetDataset', + 'AppendTokenDataset', + 'BacktranslationDataset', + 'BaseWrapperDataset', + 'BucketPadLengthDataset', + 'ColorizeDataset', + 'ConcatDataset', + 'ConcatSentencesDataset', + 'CountingIterator', + 'DenoisingDataset', + 'Dictionary', + 'EpochBatchIterator', + 'FairseqDataset', + 'FairseqIterableDataset', + 'GroupedIterator', + 'IdDataset', + 'IndexedCachedDataset', + 'IndexedDataset', + 'IndexedRawTextDataset', + 'LanguagePairDataset', + 'LeftPadDataset', + 'ListDataset', + 'LMContextWindowDataset', + 'LRUCacheDataset', + 'MaskTokensDataset', + 'MMapIndexedDataset', + 'MonolingualDataset', + 'MultiCorpusSampledDataset', + 'NestedDictionaryDataset', + 'NoisingDataset', + 'NumelDataset', + 'NumSamplesDataset', + 'OffsetTokensDataset', + 'PadDataset', + 'PrependDataset', + 'PrependTokenDataset', + 'ReplaceDataset', + 'RollDataset', + 'FileAudioDataset', + 'RawLabelDataset', + 'ResamplingDataset', + 'RightPadDataset', + 'RoundRobinZipDatasets', + 'SampledMultiDataset', + 'SampledMultiEpochDataset', + 'ShardedIterator', + 'SortDataset', + 'StripTokenDataset', + 'SubsampleDataset', + 'TokenBlockDataset', + 'TransformEosDataset', + 'TransformEosLangPairDataset', + 'TruncateDataset', + 'TruncatedDictionary', +] diff --git a/fairseq/data/__pycache__/__init__.cpython-310.pyc b/fairseq/data/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad41218b144dd8376b168030ed6977aa72f6ffb1 Binary files /dev/null and b/fairseq/data/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/add_target_dataset.cpython-310.pyc b/fairseq/data/__pycache__/add_target_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..835a27526cd7e56cd0550e4743bd6f937f9acc1a Binary files /dev/null and b/fairseq/data/__pycache__/add_target_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/append_token_dataset.cpython-310.pyc b/fairseq/data/__pycache__/append_token_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c33c9beaf9e0251846233d56197cd231630e0ff1 Binary files /dev/null and b/fairseq/data/__pycache__/append_token_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/backtranslation_dataset.cpython-310.pyc b/fairseq/data/__pycache__/backtranslation_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..297488a799d60ce75a7ca9740359a2e1dff93a68 Binary files /dev/null and b/fairseq/data/__pycache__/backtranslation_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/base_wrapper_dataset.cpython-310.pyc b/fairseq/data/__pycache__/base_wrapper_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..86578ea8c97e281c19e2d7dd259228e65633d12d Binary files /dev/null and b/fairseq/data/__pycache__/base_wrapper_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/bucket_pad_length_dataset.cpython-310.pyc b/fairseq/data/__pycache__/bucket_pad_length_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2e2ed3cc85c32a37b1e641e4bf6a347d32438af3 Binary files /dev/null and b/fairseq/data/__pycache__/bucket_pad_length_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/colorize_dataset.cpython-310.pyc b/fairseq/data/__pycache__/colorize_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..04b98140abf808a6c1f48f74fc34ef01fe2f8376 Binary files /dev/null and b/fairseq/data/__pycache__/colorize_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/concat_dataset.cpython-310.pyc b/fairseq/data/__pycache__/concat_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..346095dac46e74f06dd4b7d4f8317cf35b0154df Binary files /dev/null and b/fairseq/data/__pycache__/concat_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/concat_sentences_dataset.cpython-310.pyc b/fairseq/data/__pycache__/concat_sentences_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..02bae56218104c724543c895da6d5401b3ddac9f Binary files /dev/null and b/fairseq/data/__pycache__/concat_sentences_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/data_utils.cpython-310.pyc b/fairseq/data/__pycache__/data_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c2d9405f8ca3da8471be892c0cd925afc4d4d2cf Binary files /dev/null and b/fairseq/data/__pycache__/data_utils.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/denoising_dataset.cpython-310.pyc b/fairseq/data/__pycache__/denoising_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8a4226384d4a48dee3be21b059f1fc7ccb4895ff Binary files /dev/null and b/fairseq/data/__pycache__/denoising_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/dictionary.cpython-310.pyc b/fairseq/data/__pycache__/dictionary.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8ea1907f215670eeb17f87dd579d6c0f317fb226 Binary files /dev/null and b/fairseq/data/__pycache__/dictionary.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/fairseq_dataset.cpython-310.pyc b/fairseq/data/__pycache__/fairseq_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..127d32b6edde81c67fa2c04c305ffb129a1b5d8e Binary files /dev/null and b/fairseq/data/__pycache__/fairseq_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/id_dataset.cpython-310.pyc b/fairseq/data/__pycache__/id_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4d18d3bb4cfff3271f134e0af637b3a42e915fa4 Binary files /dev/null and b/fairseq/data/__pycache__/id_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/indexed_dataset.cpython-310.pyc b/fairseq/data/__pycache__/indexed_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..149845d065d6184dea2fad6a50eb680e63fb59f2 Binary files /dev/null and b/fairseq/data/__pycache__/indexed_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/iterators.cpython-310.pyc b/fairseq/data/__pycache__/iterators.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..06e3a9998a6a52c1d81b6636e7f0fe168db25202 Binary files /dev/null and b/fairseq/data/__pycache__/iterators.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/language_pair_dataset.cpython-310.pyc b/fairseq/data/__pycache__/language_pair_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b8083e77e134b7bb8edefe9fd3320f04db92cf6f Binary files /dev/null and b/fairseq/data/__pycache__/language_pair_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/list_dataset.cpython-310.pyc b/fairseq/data/__pycache__/list_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..271f43896a8ed9cbe067505d25d7e212f42eff8d Binary files /dev/null and b/fairseq/data/__pycache__/list_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/lm_context_window_dataset.cpython-310.pyc b/fairseq/data/__pycache__/lm_context_window_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1a3443d7ecf00d157e25790759087bf61813729c Binary files /dev/null and b/fairseq/data/__pycache__/lm_context_window_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/lru_cache_dataset.cpython-310.pyc b/fairseq/data/__pycache__/lru_cache_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3e6ec08a167582523bcb25959fde6eaae15d2081 Binary files /dev/null and b/fairseq/data/__pycache__/lru_cache_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/mask_tokens_dataset.cpython-310.pyc b/fairseq/data/__pycache__/mask_tokens_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..48280876daa2a7505f86512a6586634732b369bf Binary files /dev/null and b/fairseq/data/__pycache__/mask_tokens_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/monolingual_dataset.cpython-310.pyc b/fairseq/data/__pycache__/monolingual_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cde3ecbca18bff7283a27d71c3ad55ba7943d06c Binary files /dev/null and b/fairseq/data/__pycache__/monolingual_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/multi_corpus_sampled_dataset.cpython-310.pyc b/fairseq/data/__pycache__/multi_corpus_sampled_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..757269577c5f974b0d5e17689dc3dcf23f05d52c Binary files /dev/null and b/fairseq/data/__pycache__/multi_corpus_sampled_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/nested_dictionary_dataset.cpython-310.pyc b/fairseq/data/__pycache__/nested_dictionary_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6e68c21e44e62f9c0fe0e5dd7274d6ca40e80763 Binary files /dev/null and b/fairseq/data/__pycache__/nested_dictionary_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/noising.cpython-310.pyc b/fairseq/data/__pycache__/noising.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fe1107260d39856e578b3d2ecfd20314f988a1b2 Binary files /dev/null and b/fairseq/data/__pycache__/noising.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/num_samples_dataset.cpython-310.pyc b/fairseq/data/__pycache__/num_samples_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6ccba5e403e5a73e1b8c36f86efa228260cb8bea Binary files /dev/null and b/fairseq/data/__pycache__/num_samples_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/numel_dataset.cpython-310.pyc b/fairseq/data/__pycache__/numel_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7be8cf04ec1a091d56b0878dc3b680e5b370cc5c Binary files /dev/null and b/fairseq/data/__pycache__/numel_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/offset_tokens_dataset.cpython-310.pyc b/fairseq/data/__pycache__/offset_tokens_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bc2d683fd0b205a4e8c5605944b517c87a76d317 Binary files /dev/null and b/fairseq/data/__pycache__/offset_tokens_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/pad_dataset.cpython-310.pyc b/fairseq/data/__pycache__/pad_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8630d10e68c99d2d21040dce659d5974cf4be4b0 Binary files /dev/null and b/fairseq/data/__pycache__/pad_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/plasma_utils.cpython-310.pyc b/fairseq/data/__pycache__/plasma_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d192e6c349ac6de2d0f828668a2f6415019f26f4 Binary files /dev/null and b/fairseq/data/__pycache__/plasma_utils.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/prepend_dataset.cpython-310.pyc b/fairseq/data/__pycache__/prepend_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1f9bd78e89f268efba6fe29cd88a97662733c305 Binary files /dev/null and b/fairseq/data/__pycache__/prepend_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/prepend_token_dataset.cpython-310.pyc b/fairseq/data/__pycache__/prepend_token_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b5014bffbf45032e3771519a02152b6b4f7863b0 Binary files /dev/null and b/fairseq/data/__pycache__/prepend_token_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/raw_label_dataset.cpython-310.pyc b/fairseq/data/__pycache__/raw_label_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4899af9aca8155611df0a38c6cd57b739df2ab04 Binary files /dev/null and b/fairseq/data/__pycache__/raw_label_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/replace_dataset.cpython-310.pyc b/fairseq/data/__pycache__/replace_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4de08000e5eeb458ae1d44084ba9183cd6f9eb4c Binary files /dev/null and b/fairseq/data/__pycache__/replace_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/resampling_dataset.cpython-310.pyc b/fairseq/data/__pycache__/resampling_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..13e0318d453f6d07809fb45acbfc383199b23d27 Binary files /dev/null and b/fairseq/data/__pycache__/resampling_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/roll_dataset.cpython-310.pyc b/fairseq/data/__pycache__/roll_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0069e2fc4fdbb99f0e60bba5050eb3d1a005bee5 Binary files /dev/null and b/fairseq/data/__pycache__/roll_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/round_robin_zip_datasets.cpython-310.pyc b/fairseq/data/__pycache__/round_robin_zip_datasets.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5174abd34829c05d48745f7737f8ff3c6f35cd07 Binary files /dev/null and b/fairseq/data/__pycache__/round_robin_zip_datasets.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/shorten_dataset.cpython-310.pyc b/fairseq/data/__pycache__/shorten_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5e0f2347001583942f1a46daf2a339a831597f1e Binary files /dev/null and b/fairseq/data/__pycache__/shorten_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/sort_dataset.cpython-310.pyc b/fairseq/data/__pycache__/sort_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..25db7b3057f8f51e18b4b3566c8f18ba8efed15f Binary files /dev/null and b/fairseq/data/__pycache__/sort_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/strip_token_dataset.cpython-310.pyc b/fairseq/data/__pycache__/strip_token_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a2a991e11b6751521f5be59104663e574f61c34f Binary files /dev/null and b/fairseq/data/__pycache__/strip_token_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/subsample_dataset.cpython-310.pyc b/fairseq/data/__pycache__/subsample_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0cc376017fbe7a7d29d2c4cb243fb16da78dc757 Binary files /dev/null and b/fairseq/data/__pycache__/subsample_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/token_block_dataset.cpython-310.pyc b/fairseq/data/__pycache__/token_block_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..913ff9df16bfb459a080a4580d24497dcedc8a2a Binary files /dev/null and b/fairseq/data/__pycache__/token_block_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/transform_eos_dataset.cpython-310.pyc b/fairseq/data/__pycache__/transform_eos_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0fc937393158c870601acf92d3e97f92911ee0bf Binary files /dev/null and b/fairseq/data/__pycache__/transform_eos_dataset.cpython-310.pyc differ diff --git a/fairseq/data/__pycache__/transform_eos_lang_pair_dataset.cpython-310.pyc b/fairseq/data/__pycache__/transform_eos_lang_pair_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..491433270eef64b146e07f01bd95fc713755f9a5 Binary files /dev/null and b/fairseq/data/__pycache__/transform_eos_lang_pair_dataset.cpython-310.pyc differ diff --git a/fairseq/data/add_target_dataset.py b/fairseq/data/add_target_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..3a42dd7a2ef29f1fad139c79509923e684bfa9ad --- /dev/null +++ b/fairseq/data/add_target_dataset.py @@ -0,0 +1,56 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import BaseWrapperDataset +from . import data_utils + + +class AddTargetDataset(BaseWrapperDataset): + def __init__(self, dataset, labels, pad, eos, batch_targets, process_label=None, add_to_input=False): + super().__init__(dataset) + self.labels = labels + self.batch_targets = batch_targets + self.pad = pad + self.eos = eos + self.process_label = process_label + self.add_to_input = add_to_input + + def get_label(self, index): + return self.labels[index] if self.process_label is None else self.process_label(self.labels[index]) + + def __getitem__(self, index): + item = self.dataset[index] + item["label"] = self.get_label(index) + return item + + def size(self, index): + sz = self.dataset.size(index) + own_sz = len(self.get_label(index)) + return (sz, own_sz) + + def collater(self, samples): + collated = self.dataset.collater(samples) + if len(collated) == 0: + return collated + indices = set(collated["id"].tolist()) + target = [s["label"] for s in samples if s["id"] in indices] + + if self.batch_targets: + collated["target_lengths"] = torch.LongTensor([len(t) for t in target]) + target = data_utils.collate_tokens(target, pad_idx=self.pad, left_pad=False) + collated["ntokens"] = collated["target_lengths"].sum().item() + else: + collated["ntokens"] = sum([len(t) for t in target]) + + collated["target"] = target + + if self.add_to_input: + eos = target.new_full((target.size(0), 1), self.eos) + collated["target"] = torch.cat([target, eos], dim=-1).long() + collated["net_input"]["prev_output_tokens"] = torch.cat([eos, target], dim=-1).long() + collated["ntokens"] += target.size(0) + return collated \ No newline at end of file diff --git a/fairseq/data/append_token_dataset.py b/fairseq/data/append_token_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..7298129f62a0a61fad34045f44154ba6f2d7b864 --- /dev/null +++ b/fairseq/data/append_token_dataset.py @@ -0,0 +1,42 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class AppendTokenDataset(BaseWrapperDataset): + + def __init__(self, dataset, token=None): + super().__init__(dataset) + self.token = token + if token is not None: + self._sizes = np.array(dataset.sizes) + 1 + else: + self._sizes = dataset.sizes + + def __getitem__(self, idx): + item = self.dataset[idx] + if self.token is not None: + item = torch.cat([item, item.new([self.token])]) + return item + + @property + def sizes(self): + return self._sizes + + def num_tokens(self, index): + n = self.dataset.num_tokens(index) + if self.token is not None: + n += 1 + return n + + def size(self, index): + n = self.dataset.size(index) + if self.token is not None: + n += 1 + return n diff --git a/fairseq/data/audio/__init__.py b/fairseq/data/audio/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fairseq/data/audio/__pycache__/__init__.cpython-310.pyc b/fairseq/data/audio/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..19a86c04da258724d7f56130c5a553498627d0a8 Binary files /dev/null and b/fairseq/data/audio/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/data/audio/__pycache__/raw_audio_dataset.cpython-310.pyc b/fairseq/data/audio/__pycache__/raw_audio_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2a798bbc8bb039d5ee9ae4ba9f4c15f1c68de966 Binary files /dev/null and b/fairseq/data/audio/__pycache__/raw_audio_dataset.cpython-310.pyc differ diff --git a/fairseq/data/audio/raw_audio_dataset.py b/fairseq/data/audio/raw_audio_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..675b09564776e28f2bfaf1a81a0f915d7c34ae7c --- /dev/null +++ b/fairseq/data/audio/raw_audio_dataset.py @@ -0,0 +1,181 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import os +import logging +import numpy as np +import sys + +import torch +import torch.nn.functional as F + +from .. import FairseqDataset + +logger = logging.getLogger(__name__) + + +class RawAudioDataset(FairseqDataset): + def __init__( + self, + sample_rate, + max_sample_size=None, + min_sample_size=None, + shuffle=True, + min_length=0, + pad=False, + normalize=False, + ): + super().__init__() + + self.sample_rate = sample_rate + self.sizes = [] + self.max_sample_size = ( + max_sample_size if max_sample_size is not None else sys.maxsize + ) + self.min_sample_size = min_sample_size + self.min_length = min_length + self.pad = pad + self.shuffle = shuffle + self.normalize = normalize + + def __getitem__(self, index): + raise NotImplementedError() + + def __len__(self): + return len(self.sizes) + + def postprocess(self, feats, curr_sample_rate): + if feats.dim() == 2: + feats = feats.mean(-1) + + if curr_sample_rate != self.sample_rate: + raise Exception(f"sample rate: {curr_sample_rate}, need {self.sample_rate}") + + assert feats.dim() == 1, feats.dim() + + if self.normalize: + with torch.no_grad(): + feats = F.layer_norm(feats, feats.shape) + return feats + + def crop_to_max_size(self, wav, target_size): + size = len(wav) + diff = size - target_size + if diff <= 0: + return wav + + start = np.random.randint(0, diff + 1) + end = size - diff + start + return wav[start:end] + + def collater(self, samples): + samples = [ + s + for s in samples + if s["source"] is not None + ] + if len(samples) == 0: + return {} + + sources = [s["source"] for s in samples] + sizes = [len(s) for s in sources] + + if self.pad: + target_size = min(max(sizes), self.max_sample_size) + else: + target_size = min(min(sizes), self.max_sample_size) + + collated_sources = sources[0].new(len(sources), target_size) + padding_mask = ( + torch.BoolTensor(collated_sources.shape).fill_(False) if self.pad else None + ) + for i, (source, size) in enumerate(zip(sources, sizes)): + diff = size - target_size + if diff == 0: + collated_sources[i] = source + elif diff < 0: + assert self.pad + collated_sources[i] = torch.cat( + [source, source.new_full((-diff,), 0.0)] + ) + padding_mask[i, diff:] = True + else: + collated_sources[i] = self.crop_to_max_size(source, target_size) + + input = {"source": collated_sources} + if self.pad: + input["padding_mask"] = padding_mask + return {"id": torch.LongTensor([s["id"] for s in samples]), "net_input": input} + + def num_tokens(self, index): + return self.size(index) + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + if self.pad: + return self.sizes[index] + return min(self.sizes[index], self.max_sample_size) + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + + order.append(self.sizes) + return np.lexsort(order)[::-1] + + +class FileAudioDataset(RawAudioDataset): + def __init__( + self, + manifest_path, + sample_rate, + max_sample_size=None, + min_sample_size=None, + shuffle=True, + min_length=0, + pad=False, + normalize=False, + ): + super().__init__( + sample_rate=sample_rate, + max_sample_size=max_sample_size, + min_sample_size=min_sample_size, + shuffle=shuffle, + min_length=min_length, + pad=pad, + normalize=normalize, + ) + + self.fnames = [] + + skipped = 0 + with open(manifest_path, "r") as f: + self.root_dir = f.readline().strip() + for line in f: + items = line.strip().split("\t") + assert len(items) == 2, line + sz = int(items[1]) + if min_length is not None and sz < min_length: + skipped += 1 + continue + self.fnames.append(items[0]) + self.sizes.append(sz) + logger.info(f"loaded {len(self.fnames)}, skipped {skipped} samples") + + def __getitem__(self, index): + import soundfile as sf + + fname = os.path.join(self.root_dir, self.fnames[index]) + wav, curr_sample_rate = sf.read(fname) + feats = torch.from_numpy(wav).float() + feats = self.postprocess(feats, curr_sample_rate) + return {"id": index, "source": feats} diff --git a/fairseq/data/backtranslation_dataset.py b/fairseq/data/backtranslation_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..0007a015064cdb025b39201b3ea1647caee04ec9 --- /dev/null +++ b/fairseq/data/backtranslation_dataset.py @@ -0,0 +1,165 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from fairseq import utils + +from . import FairseqDataset + + +def backtranslate_samples(samples, collate_fn, generate_fn, cuda=True): + """Backtranslate a list of samples. + + Given an input (*samples*) of the form: + + [{'id': 1, 'source': 'hallo welt'}] + + this will return: + + [{'id': 1, 'source': 'hello world', 'target': 'hallo welt'}] + + Args: + samples (List[dict]): samples to backtranslate. Individual samples are + expected to have a 'source' key, which will become the 'target' + after backtranslation. + collate_fn (callable): function to collate samples into a mini-batch + generate_fn (callable): function to generate backtranslations + cuda (bool): use GPU for generation (default: ``True``) + + Returns: + List[dict]: an updated list of samples with a backtranslated source + """ + collated_samples = collate_fn(samples) + s = utils.move_to_cuda(collated_samples) if cuda else collated_samples + generated_sources = generate_fn(s) + + id_to_src = { + sample['id']: sample['source'] for sample in samples + } + + # Go through each tgt sentence in batch and its corresponding best + # generated hypothesis and create a backtranslation data pair + # {id: id, source: generated backtranslation, target: original tgt} + return [ + {'id': id.item(), 'target': id_to_src[id.item()], 'source': hypos[0]['tokens'].cpu()} + for id, hypos in zip(collated_samples['id'], generated_sources) + ] + + +class BacktranslationDataset(FairseqDataset): + """ + Sets up a backtranslation dataset which takes a tgt batch, generates + a src using a tgt-src backtranslation function (*backtranslation_fn*), + and returns the corresponding `{generated src, input tgt}` batch. + + Args: + tgt_dataset (~fairseq.data.FairseqDataset): the dataset to be + backtranslated. Only the source side of this dataset will be used. + After backtranslation, the source sentences in this dataset will be + returned as the targets. + src_dict (~fairseq.data.Dictionary): the dictionary of backtranslated + sentences. + tgt_dict (~fairseq.data.Dictionary, optional): the dictionary of + sentences to be backtranslated. + backtranslation_fn (callable, optional): function to call to generate + backtranslations. This is typically the `generate` method of a + :class:`~fairseq.sequence_generator.SequenceGenerator` object. + Pass in None when it is not available at initialization time, and + use set_backtranslation_fn function to set it when available. + output_collater (callable, optional): function to call on the + backtranslated samples to create the final batch + (default: ``tgt_dataset.collater``). + cuda: use GPU for generation + """ + + def __init__( + self, + tgt_dataset, + src_dict, + tgt_dict=None, + backtranslation_fn=None, + output_collater=None, + cuda=True, + **kwargs + ): + self.tgt_dataset = tgt_dataset + self.backtranslation_fn = backtranslation_fn + self.output_collater = output_collater if output_collater is not None \ + else tgt_dataset.collater + self.cuda = cuda if torch.cuda.is_available() else False + self.src_dict = src_dict + self.tgt_dict = tgt_dict + + def __getitem__(self, index): + """ + Returns a single sample from *tgt_dataset*. Note that backtranslation is + not applied in this step; use :func:`collater` instead to backtranslate + a batch of samples. + """ + return self.tgt_dataset[index] + + def __len__(self): + return len(self.tgt_dataset) + + def set_backtranslation_fn(self, backtranslation_fn): + self.backtranslation_fn = backtranslation_fn + + def collater(self, samples): + """Merge and backtranslate a list of samples to form a mini-batch. + + Using the samples from *tgt_dataset*, load a collated target sample to + feed to the backtranslation model. Then take the backtranslation with + the best score as the source and the original input as the target. + + Note: we expect *tgt_dataset* to provide a function `collater()` that + will collate samples into the format expected by *backtranslation_fn*. + After backtranslation, we will feed the new list of samples (i.e., the + `(backtranslated source, original source)` pairs) to *output_collater* + and return the result. + + Args: + samples (List[dict]): samples to backtranslate and collate + + Returns: + dict: a mini-batch with keys coming from *output_collater* + """ + if samples[0].get('is_dummy', False): + return samples + samples = backtranslate_samples( + samples=samples, + collate_fn=self.tgt_dataset.collater, + generate_fn=( + lambda net_input: self.backtranslation_fn(net_input) + ), + cuda=self.cuda, + ) + return self.output_collater(samples) + + def num_tokens(self, index): + """Just use the tgt dataset num_tokens""" + return self.tgt_dataset.num_tokens(index) + + def ordered_indices(self): + """Just use the tgt dataset ordered_indices""" + return self.tgt_dataset.ordered_indices() + + def size(self, index): + """Return an example's size as a float or tuple. This value is used + when filtering a dataset with ``--max-positions``. + + Note: we use *tgt_dataset* to approximate the length of the source + sentence, since we do not know the actual length until after + backtranslation. + """ + tgt_size = self.tgt_dataset.size(index)[0] + return (tgt_size, tgt_size) + + @property + def supports_prefetch(self): + return getattr(self.tgt_dataset, 'supports_prefetch', False) + + def prefetch(self, indices): + return self.tgt_dataset.prefetch(indices) diff --git a/fairseq/data/base_wrapper_dataset.py b/fairseq/data/base_wrapper_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..8b5326a6354641e81a2e6289e5f60e619dbadc90 --- /dev/null +++ b/fairseq/data/base_wrapper_dataset.py @@ -0,0 +1,69 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from torch.utils.data.dataloader import default_collate + +from . import FairseqDataset + + +class BaseWrapperDataset(FairseqDataset): + + def __init__(self, dataset): + super().__init__() + self.dataset = dataset + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples): + if hasattr(self.dataset, 'collater'): + return self.dataset.collater(samples) + else: + return default_collate(samples) + + @property + def sizes(self): + return self.dataset.sizes + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + return self.dataset.size(index) + + def ordered_indices(self): + return self.dataset.ordered_indices() + + @property + def supports_prefetch(self): + return getattr(self.dataset, 'supports_prefetch', False) + + def prefetch(self, indices): + self.dataset.prefetch(indices) + + def get_batch_shapes(self): + return self.dataset.get_batch_shapes() + + def batch_by_size( + self, + indices, + max_tokens=None, + max_sentences=None, + required_batch_size_multiple=1, + ): + return self.dataset.batch_by_size( + indices, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + ) + + def set_epoch(self, epoch): + super().set_epoch(epoch) + if hasattr(self.dataset, 'set_epoch'): + self.dataset.set_epoch(epoch) diff --git a/fairseq/data/bucket_pad_length_dataset.py b/fairseq/data/bucket_pad_length_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..6f53d011881b889f2fb22f0cccbca6cee1e309ac --- /dev/null +++ b/fairseq/data/bucket_pad_length_dataset.py @@ -0,0 +1,77 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch.nn.functional as F + +from fairseq.data import BaseWrapperDataset + + +class BucketPadLengthDataset(BaseWrapperDataset): + """ + Bucket and pad item lengths to the nearest bucket size. This can be used to + reduce the number of unique batch shapes, which is important on TPUs since + each new batch shape requires a recompilation. + + Args: + dataset (FairseqDatset): dataset to bucket + sizes (List[int]): all item sizes + num_buckets (int): number of buckets to create + pad_idx (int): padding symbol + left_pad (bool): if True, pad on the left; otherwise right pad + """ + + def __init__( + self, + dataset, + sizes, + num_buckets, + pad_idx, + left_pad, + ): + super().__init__(dataset) + self.pad_idx = pad_idx + self.left_pad = left_pad + + assert num_buckets > 0 + self.buckets = np.unique( + np.percentile( + sizes, + np.linspace(0, 100, num_buckets + 1), + interpolation='lower', + )[1:] + ) + + def get_bucketed_sizes(orig_sizes, buckets): + sizes = np.copy(orig_sizes) + assert np.min(sizes) >= 0 + start_val = -1 + for end_val in buckets: + mask = (sizes > start_val) & (sizes <= end_val) + sizes[mask] = end_val + start_val = end_val + return sizes + + self._bucketed_sizes = get_bucketed_sizes(sizes, self.buckets) + + def __getitem__(self, index): + item = self.dataset[index] + bucket_size = self._bucketed_sizes[index] + num_pad = bucket_size - item.size(-1) + return F.pad( + item, + (num_pad if self.left_pad else 0, 0 if self.left_pad else num_pad), + value=self.pad_idx, + ) + + @property + def sizes(self): + return self._bucketed_sizes + + def num_tokens(self, index): + return self._bucketed_sizes[index] + + def size(self, index): + return self._bucketed_sizes[index] diff --git a/fairseq/data/colorize_dataset.py b/fairseq/data/colorize_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..89e0e04142ff16aff931e0a0dc876558000db514 --- /dev/null +++ b/fairseq/data/colorize_dataset.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import BaseWrapperDataset + + +class ColorizeDataset(BaseWrapperDataset): + """ Adds 'colors' property to net input that is obtained from the provided color getter for use by models """ + def __init__(self, dataset, color_getter): + super().__init__(dataset) + self.color_getter = color_getter + + def collater(self, samples): + base_collate = super().collater(samples) + if len(base_collate) > 0: + base_collate["net_input"]["colors"] = torch.tensor( + list(self.color_getter(self.dataset, s["id"]) for s in samples), + dtype=torch.long, + ) + return base_collate diff --git a/fairseq/data/concat_dataset.py b/fairseq/data/concat_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..5ca80631f0d1e9eeffc3ec2b6782b285feae29f4 --- /dev/null +++ b/fairseq/data/concat_dataset.py @@ -0,0 +1,105 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import bisect + +import numpy as np +from torch.utils.data.dataloader import default_collate + +from . import FairseqDataset + + +class ConcatDataset(FairseqDataset): + @staticmethod + def cumsum(sequence, sample_ratios): + r, s = [], 0 + for e, ratio in zip(sequence, sample_ratios): + curr_len = int(ratio * len(e)) + r.append(curr_len + s) + s += curr_len + return r + + def __init__(self, datasets, sample_ratios=1): + super(ConcatDataset, self).__init__() + assert len(datasets) > 0, "datasets should not be an empty iterable" + self.datasets = list(datasets) + if isinstance(sample_ratios, int): + sample_ratios = [sample_ratios] * len(self.datasets) + self.sample_ratios = sample_ratios + self.cumulative_sizes = self.cumsum(self.datasets, sample_ratios) + self.real_sizes = [len(d) for d in self.datasets] + + def __len__(self): + return self.cumulative_sizes[-1] + + def __getitem__(self, idx): + dataset_idx, sample_idx = self._get_dataset_and_sample_index(idx) + return self.datasets[dataset_idx][sample_idx] + + def _get_dataset_and_sample_index(self, idx: int): + dataset_idx = bisect.bisect_right(self.cumulative_sizes, idx) + if dataset_idx == 0: + sample_idx = idx + else: + sample_idx = idx - self.cumulative_sizes[dataset_idx - 1] + sample_idx = sample_idx % self.real_sizes[dataset_idx] + return dataset_idx, sample_idx + + def collater(self, samples, **extra_args): + # For now only supports datasets with same underlying collater implementations + if hasattr(self.datasets[0], 'collater'): + return self.datasets[0].collater(samples, **extra_args) + else: + return default_collate(samples, **extra_args) + + def size(self, idx: int): + """ + Return an example's size as a float or tuple. + """ + dataset_idx, sample_idx = self._get_dataset_and_sample_index(idx) + return self.datasets[dataset_idx].size(sample_idx) + + def num_tokens(self, index: int): + return np.max(self.size(index)) + + def attr(self, attr: str, index: int): + dataset_idx = bisect.bisect_right(self.cumulative_sizes, index) + return getattr(self.datasets[dataset_idx], attr, None) + + @property + def sizes(self): + _dataset_sizes = [] + for ds, sr in zip(self.datasets, self.sample_ratios): + if isinstance(ds.sizes, np.ndarray): + _dataset_sizes.append(np.tile(ds.sizes, sr)) + else: + # Only support underlying dataset with single size array. + assert isinstance(ds.sizes, list) + _dataset_sizes.append(np.tile(ds.sizes[0], sr)) + return np.concatenate(_dataset_sizes) + + @property + def supports_prefetch(self): + return all(d.supports_prefetch for d in self.datasets) + + def ordered_indices(self): + """ + Returns indices sorted by length. So less padding is needed. + """ + return np.argsort(self.sizes) + + def prefetch(self, indices): + frm = 0 + for to, ds in zip(self.cumulative_sizes, self.datasets): + real_size = len(ds) + if getattr(ds, 'supports_prefetch', False): + ds.prefetch([(i - frm) % real_size for i in indices if frm <= i < to]) + frm = to + + def set_epoch(self, epoch): + super().set_epoch(epoch) + for ds in self.datasets: + if hasattr(ds, 'set_epoch'): + ds.set_epoch(epoch) diff --git a/fairseq/data/concat_sentences_dataset.py b/fairseq/data/concat_sentences_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..55445ee1c789bdacc4acaecd628903ced276a65c --- /dev/null +++ b/fairseq/data/concat_sentences_dataset.py @@ -0,0 +1,56 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class ConcatSentencesDataset(FairseqDataset): + + def __init__(self, *datasets): + super().__init__() + self.datasets = datasets + assert all(len(ds) == len(datasets[0]) for ds in datasets), \ + 'datasets must have the same length' + + def __getitem__(self, index): + return torch.cat([ds[index] for ds in self.datasets]) + + def __len__(self): + return len(self.datasets[0]) + + def collater(self, samples): + return self.datasets[0].collater(samples) + + @property + def sizes(self): + return sum(ds.sizes for ds in self.datasets) + + def num_tokens(self, index): + return sum(ds.num_tokens(index) for ds in self.datasets) + + def size(self, index): + return sum(ds.size(index) for ds in self.datasets) + + def ordered_indices(self): + return self.datasets[0].ordered_indices() + + @property + def supports_prefetch(self): + return any( + getattr(ds, 'supports_prefetch', False) for ds in self.datasets + ) + + def prefetch(self, indices): + for ds in self.datasets: + if getattr(ds, 'supports_prefetch', False): + ds.prefetch(indices) + + def set_epoch(self, epoch): + super().set_epoch(epoch) + for ds in self.datasets: + if hasattr(ds, 'set_epoch'): + ds.set_epoch(epoch) diff --git a/fairseq/data/data_utils.py b/fairseq/data/data_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..70d8997b19a879fe520be94ed5475349bfe9c585 --- /dev/null +++ b/fairseq/data/data_utils.py @@ -0,0 +1,396 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +try: + from collections.abc import Iterable +except ImportError: + from collections import Iterable +import contextlib +import itertools +import logging +import os + +from typing import Tuple, Optional + +import numpy as np +import torch + + +logger = logging.getLogger(__name__) + + +def infer_language_pair(path): + """Infer language pair from filename: .-.(...).idx""" + src, dst = None, None + for filename in os.listdir(path): + parts = filename.split('.') + if len(parts) >= 3 and len(parts[1].split('-')) == 2: + return parts[1].split('-') + return src, dst + + +def collate_tokens(values, pad_idx, eos_idx=None, left_pad=False, move_eos_to_beginning=False, pad_to_length=None): + """Convert a list of 1d tensors into a padded 2d tensor.""" + size = max(v.size(0) for v in values) + size = size if pad_to_length is None else max(size, pad_to_length) + res = values[0].new(len(values), size).fill_(pad_idx) + + def copy_tensor(src, dst): + assert dst.numel() == src.numel() + if move_eos_to_beginning: + if eos_idx is None: + # if no eos_idx is specified, then use the last token in src + dst[0] = src[-1] + else: + dst[0] = eos_idx + dst[1:] = src[:-1] + else: + dst.copy_(src) + + for i, v in enumerate(values): + copy_tensor(v, res[i][size - len(v):] if left_pad else res[i][:len(v)]) + return res + + +def load_indexed_dataset(path, dictionary=None, dataset_impl=None, combine=False, default='cached'): + """A helper function for loading indexed datasets. + + Args: + path (str): path to indexed dataset (e.g., 'data-bin/train') + dictionary (~fairseq.data.Dictionary): data dictionary + dataset_impl (str, optional): which dataset implementation to use. If + not provided, it will be inferred automatically. For legacy indexed + data we use the 'cached' implementation by default. + combine (bool, optional): automatically load and combine multiple + datasets. For example, if *path* is 'data-bin/train', then we will + combine 'data-bin/train', 'data-bin/train1', ... and return a + single ConcatDataset instance. + """ + from fairseq.data.concat_dataset import ConcatDataset + import fairseq.data.indexed_dataset as indexed_dataset + + datasets = [] + for k in itertools.count(): + path_k = path + (str(k) if k > 0 else '') + + dataset_impl_k = dataset_impl + if dataset_impl_k is None: + dataset_impl_k = indexed_dataset.infer_dataset_impl(path_k) + + dataset = indexed_dataset.make_dataset( + path_k, + impl=dataset_impl_k or default, + fix_lua_indexing=True, + dictionary=dictionary, + ) + if dataset is None: + break + logger.info('loaded {} examples from: {}'.format(len(dataset), path_k)) + datasets.append(dataset) + if not combine: + break + if len(datasets) == 0: + return None + elif len(datasets) == 1: + return datasets[0] + else: + return ConcatDataset(datasets) + + +@contextlib.contextmanager +def numpy_seed(seed, *addl_seeds): + """Context manager which seeds the NumPy PRNG with the specified seed and + restores the state afterward""" + if seed is None: + yield + return + if len(addl_seeds) > 0: + seed = int(hash((seed, *addl_seeds)) % 1e6) + state = np.random.get_state() + np.random.seed(seed) + try: + yield + finally: + np.random.set_state(state) + + +def collect_filtered(function, iterable, filtered): + """ + Similar to :func:`filter` but collects filtered elements in ``filtered``. + + Args: + function (callable): function that returns ``False`` for elements that + should be filtered + iterable (iterable): iterable to filter + filtered (list): list to store filtered elements + """ + for el in iterable: + if function(el): + yield el + else: + filtered.append(el) + + +def _filter_by_size_dynamic(indices, size_fn, max_positions, raise_exception=False): + def compare_leq(a, b): + return a <= b if not isinstance(a, tuple) else max(a) <= b + + def check_size(idx): + if isinstance(max_positions, float) or isinstance(max_positions, int): + return size_fn(idx) <= max_positions + elif isinstance(max_positions, dict): + idx_size = size_fn(idx) + assert isinstance(idx_size, dict) + intersect_keys = set(max_positions.keys()) & set(idx_size.keys()) + return all( + all(a is None or b is None or a <= b + for a, b in zip(idx_size[key], max_positions[key])) + for key in intersect_keys + ) + else: + # Hacky as heck, for the specific case of multilingual training with RoundRobin. + if isinstance(size_fn(idx), dict) and isinstance(max_positions, tuple): + return all( + a is None or b is None or compare_leq(a, b) + for a, b in zip(size_fn(idx).values(), max_positions) + ) + # For MultiCorpusSampledDataset, will generalize it later + if not isinstance(size_fn(idx), Iterable): + return all(size_fn(idx) <= b for b in max_positions) + return all( + a is None or b is None or a <= b + for a, b in zip(size_fn(idx), max_positions) + ) + ignored = [] + itr = collect_filtered(check_size, indices, ignored) + indices = np.fromiter(itr, dtype=np.int64, count=-1) + return indices, ignored + + +def filter_by_size(indices, dataset, max_positions, raise_exception=False): + """ + [deprecated] Filter indices based on their size. + Use `FairseqDataset::filter_indices_by_size` instead. + + Args: + indices (List[int]): ordered list of dataset indices + dataset (FairseqDataset): fairseq dataset instance + max_positions (tuple): filter elements larger than this size. + Comparisons are done component-wise. + raise_exception (bool, optional): if ``True``, raise an exception if + any elements are filtered (default: False). + """ + if isinstance(max_positions, float) or isinstance(max_positions, int): + if hasattr(dataset, 'sizes') and isinstance(dataset.sizes, np.ndarray): + ignored = indices[dataset.sizes[indices] > max_positions].tolist() + indices = indices[dataset.sizes[indices] <= max_positions] + elif hasattr(dataset, 'sizes') and isinstance(dataset.sizes, list) and len(dataset.sizes) == 1: + ignored = indices[dataset.sizes[0][indices] > max_positions].tolist() + indices = indices[dataset.sizes[0][indices] <= max_positions] + else: + indices, ignored = _filter_by_size_dynamic(indices, dataset.size, max_positions) + else: + indices, ignored = _filter_by_size_dynamic(indices, dataset.size, max_positions) + + if len(ignored) > 0 and raise_exception: + raise Exception(( + 'Size of sample #{} is invalid (={}) since max_positions={}, ' + 'skip this example with --skip-invalid-size-inputs-valid-test' + ).format(ignored[0], dataset.size(ignored[0]), max_positions)) + if len(ignored) > 0: + logger.warning(( + '{} samples have invalid sizes and will be skipped, ' + 'max_positions={}, first few sample ids={}' + ).format(len(ignored), max_positions, ignored[:10])) + return indices + + +def batch_by_size( + indices, num_tokens_fn, max_tokens=None, max_sentences=None, + required_batch_size_multiple=1, fixed_shapes=None, +): + """ + Yield mini-batches of indices bucketed by size. Batches may contain + sequences of different lengths. + + Args: + indices (List[int]): ordered list of dataset indices + num_tokens_fn (callable): function that returns the number of tokens at + a given index + max_tokens (int, optional): max number of tokens in each batch + (default: None). + max_sentences (int, optional): max number of sentences in each + batch (default: None). + required_batch_size_multiple (int, optional): require batch size to + be less than N or a multiple of N (default: 1). + fixed_shapes (List[Tuple[int, int]], optional): if given, batches will + only be created with the given shapes. *max_sentences* and + *required_batch_size_multiple* will be ignored (default: None). + """ + try: + from fairseq.data.data_utils_fast import ( + batch_by_size_fast, batch_fixed_shapes_fast, + ) + except ImportError: + raise ImportError( + 'Please build Cython components with: `pip install --editable .` ' + 'or `python setup.py build_ext --inplace`' + ) + + max_tokens = max_tokens if max_tokens is not None else -1 + max_sentences = max_sentences if max_sentences is not None else -1 + bsz_mult = required_batch_size_multiple + + if not isinstance(indices, np.ndarray): + indices = np.fromiter(indices, dtype=np.int64, count=-1) + + if fixed_shapes is None: + return batch_by_size_fast( + indices, num_tokens_fn, max_tokens, max_sentences, bsz_mult, + ) + else: + fixed_shapes = np.array(fixed_shapes, dtype=np.int64) + sort_order = np.lexsort([ + fixed_shapes[:, 1].argsort(), # length + fixed_shapes[:, 0].argsort(), # bsz + ]) + fixed_shapes_sorted = fixed_shapes[sort_order] + return batch_fixed_shapes_fast(indices, num_tokens_fn, fixed_shapes_sorted) + + +def post_process(sentence: str, symbol: str): + if symbol == "sentencepiece": + sentence = sentence.replace(" ", "").replace("\u2581", " ").strip() + elif symbol == 'wordpiece': + sentence = sentence.replace(" ", "").replace("_", " ").strip() + elif symbol == 'letter': + sentence = sentence.replace(" ", "").replace("|", " ").strip() + elif symbol == "_EOW": + sentence = sentence.replace(" ", "").replace("_EOW", " ").strip() + elif symbol is not None and symbol != 'none': + sentence = (sentence + " ").replace(symbol, "").rstrip() + return sentence + +def compute_mask_indices( + shape: Tuple[int, int], + padding_mask: Optional[torch.Tensor], + mask_prob: float, + mask_length: int, + mask_type: str = "static", + mask_other: float = 0.0, + min_masks: int = 0, + no_overlap: bool = False, + min_space: int = 0, +) -> np.ndarray: + """ + Computes random mask spans for a given shape + + Args: + shape: the the shape for which to compute masks. + should be of size 2 where first element is batch size and 2nd is timesteps + padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements + mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by + number of timesteps divided by length of mask span to mask approximately this percentage of all elements. + however due to overlaps, the actual number will be smaller (unless no_overlap is True) + mask_type: how to compute mask lengths + static = fixed size + uniform = sample from uniform distribution [mask_other, mask_length*2] + normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element + poisson = sample from possion distribution with lambda = mask length + min_masks: minimum number of masked spans + no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping + min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans + """ + + bsz, all_sz = shape + mask = np.full((bsz, all_sz), False) + + all_num_mask = int( + # add a random number for probabilistic rounding + mask_prob * all_sz / float(mask_length) + + np.random.rand() + ) + + all_num_mask = max(min_masks, all_num_mask) + + mask_idcs = [] + for i in range(bsz): + if padding_mask is not None: + sz = all_sz - padding_mask[i].long().sum().item() + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + np.random.rand() + ) + num_mask = max(min_masks, num_mask) + else: + sz = all_sz + num_mask = all_num_mask + + if mask_type == "static": + lengths = np.full(num_mask, mask_length) + elif mask_type == "uniform": + lengths = np.random.randint(mask_other, mask_length * 2 + 1, size=num_mask) + elif mask_type == "normal": + lengths = np.random.normal(mask_length, mask_other, size=num_mask) + lengths = [max(1, int(round(x))) for x in lengths] + elif mask_type == "poisson": + lengths = np.random.poisson(mask_length, size=num_mask) + lengths = [int(round(x)) for x in lengths] + else: + raise Exception("unknown mask selection " + mask_type) + + if sum(lengths) == 0: + lengths[0] = min(mask_length, sz - 1) + + if no_overlap: + mask_idc = [] + def arrange(s, e, length, keep_length): + span_start = np.random.randint(s, e-length) + mask_idc.extend(span_start + i for i in range(length)) + + new_parts = [] + if span_start - s - min_space >= keep_length: + new_parts.append((s, span_start-min_space+1)) + if e - span_start - keep_length - min_space > keep_length: + new_parts.append((span_start + length + min_space, e)) + return new_parts + + parts = [(0, sz)] + min_length = min(lengths) + for length in sorted(lengths, reverse=True): + lens = np.fromiter((e - s if e-s >= length+min_space else 0 for s, e in parts), np.int) + l_sum = np.sum(lens) + if l_sum == 0: + break + probs = lens / np.sum(lens) + c = np.random.choice(len(parts), p=probs) + s, e = parts.pop(c) + parts.extend(arrange(s, e, length, min_length)) + mask_idc = np.asarray(mask_idc) + else: + min_len = min(lengths) + if sz - min_len <= num_mask: + min_len = sz - num_mask - 1 + + mask_idc = np.random.choice(sz - min_len, num_mask, replace=False) + + mask_idc = np.asarray( + [ + mask_idc[j] + offset + for j in range(len(mask_idc)) + for offset in range(lengths[j]) + ] + ) + + mask_idcs.append(np.unique(mask_idc[mask_idc < sz])) + + min_len = min([len(m) for m in mask_idcs]) + for i, mask_idc in enumerate(mask_idcs): + if len(mask_idc) > min_len: + mask_idc = np.random.choice(mask_idc, min_len, replace=False) + mask[i, mask_idc] = True + + return mask diff --git a/fairseq/data/data_utils_fast.cpp b/fairseq/data/data_utils_fast.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b0acd994ebafcb1d54dab2a5ae2119933f9fc9b4 --- /dev/null +++ b/fairseq/data/data_utils_fast.cpp @@ -0,0 +1,31968 @@ +/* Generated by Cython 3.0.12 */ + +/* BEGIN: Cython Metadata +{ + "distutils": { + "depends": [ + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/arrayobject.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/arrayscalars.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/ndarrayobject.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/ndarraytypes.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/ufuncobject.h" + ], + "extra_compile_args": [ + "-std=c++11", + "-O3" + ], + "include_dirs": [ + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include" + ], + "language": "c++", + "name": "fairseq.data.data_utils_fast", + "sources": [ + "fairseq/data/data_utils_fast.pyx" + ] + }, + "module_name": "fairseq.data.data_utils_fast" +} +END: Cython Metadata */ + +#ifndef PY_SSIZE_T_CLEAN +#define PY_SSIZE_T_CLEAN +#endif /* PY_SSIZE_T_CLEAN */ +#if defined(CYTHON_LIMITED_API) && 0 + #ifndef Py_LIMITED_API + #if CYTHON_LIMITED_API+0 > 0x03030000 + #define Py_LIMITED_API CYTHON_LIMITED_API + #else + #define Py_LIMITED_API 0x03030000 + #endif + #endif +#endif + +#include "Python.h" +#ifndef Py_PYTHON_H + #error Python headers needed to compile C extensions, please install development version of Python. +#elif PY_VERSION_HEX < 0x02070000 || (0x03000000 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x03030000) + #error Cython requires Python 2.7+ or Python 3.3+. +#else +#if defined(CYTHON_LIMITED_API) && CYTHON_LIMITED_API +#define __PYX_EXTRA_ABI_MODULE_NAME "limited" +#else +#define __PYX_EXTRA_ABI_MODULE_NAME "" +#endif +#define CYTHON_ABI "3_0_12" __PYX_EXTRA_ABI_MODULE_NAME +#define __PYX_ABI_MODULE_NAME "_cython_" CYTHON_ABI +#define __PYX_TYPE_MODULE_PREFIX __PYX_ABI_MODULE_NAME "." +#define CYTHON_HEX_VERSION 0x03000CF0 +#define CYTHON_FUTURE_DIVISION 1 +#include +#ifndef offsetof + #define offsetof(type, member) ( (size_t) & ((type*)0) -> member ) +#endif +#if !defined(_WIN32) && !defined(WIN32) && !defined(MS_WINDOWS) + #ifndef __stdcall + #define __stdcall + #endif + #ifndef __cdecl + #define __cdecl + #endif + #ifndef __fastcall + #define __fastcall + #endif +#endif +#ifndef DL_IMPORT + #define DL_IMPORT(t) t +#endif +#ifndef DL_EXPORT + #define DL_EXPORT(t) t +#endif +#define __PYX_COMMA , +#ifndef HAVE_LONG_LONG + #define HAVE_LONG_LONG +#endif +#ifndef PY_LONG_LONG + #define PY_LONG_LONG LONG_LONG +#endif +#ifndef Py_HUGE_VAL + #define Py_HUGE_VAL HUGE_VAL +#endif +#define __PYX_LIMITED_VERSION_HEX PY_VERSION_HEX +#if defined(GRAALVM_PYTHON) + /* For very preliminary testing purposes. Most variables are set the same as PyPy. + The existence of this section does not imply that anything works or is even tested */ + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 1 + #define CYTHON_COMPILING_IN_NOGIL 0 + #undef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 0 + #undef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #if PY_VERSION_HEX < 0x03050000 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #elif !defined(CYTHON_USE_ASYNC_SLOTS) + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #undef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 0 + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #undef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #undef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 1 + #undef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 0 + #undef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 0 + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #undef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 0 + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS (PY_MAJOR_VERSION >= 3) + #endif + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #undef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #undef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 0 + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 0 + #endif + #undef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 +#elif defined(PYPY_VERSION) + #define CYTHON_COMPILING_IN_PYPY 1 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 0 + #undef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 0 + #ifndef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #endif + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #if PY_VERSION_HEX < 0x03050000 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #elif !defined(CYTHON_USE_ASYNC_SLOTS) + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #undef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 0 + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #undef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #undef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 1 + #undef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 0 + #undef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 0 + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #undef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 0 + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS (PY_MAJOR_VERSION >= 3) + #endif + #if PY_VERSION_HEX < 0x03090000 + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 0 + #elif !defined(CYTHON_PEP489_MULTI_PHASE_INIT) + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #endif + #undef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #undef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE (PY_VERSION_HEX >= 0x030400a1 && PYPY_VERSION_NUM >= 0x07030C00) + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 0 + #endif + #undef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 +#elif defined(CYTHON_LIMITED_API) + #ifdef Py_LIMITED_API + #undef __PYX_LIMITED_VERSION_HEX + #define __PYX_LIMITED_VERSION_HEX Py_LIMITED_API + #endif + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 1 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 0 + #undef CYTHON_CLINE_IN_TRACEBACK + #define CYTHON_CLINE_IN_TRACEBACK 0 + #undef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 0 + #undef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 1 + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #undef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 0 + #ifndef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #endif + #undef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #ifndef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 0 + #endif + #undef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 0 + #undef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 0 + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #undef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 0 + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS 1 + #endif + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 0 + #undef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 1 + #ifndef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 0 + #endif + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 0 + #endif + #undef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 +#elif defined(Py_GIL_DISABLED) || defined(Py_NOGIL) + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 1 + #ifndef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 1 + #endif + #ifndef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #endif + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #ifndef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #ifndef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #endif + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #ifndef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 1 + #endif + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #ifndef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 0 + #endif + #ifndef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 1 + #endif + #ifndef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 1 + #endif + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #ifndef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 1 + #endif + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS 1 + #endif + #ifndef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #endif + #ifndef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #endif + #ifndef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 1 + #endif + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 1 + #endif + #ifndef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 + #endif +#else + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 1 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 0 + #ifndef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 1 + #endif + #ifndef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #endif + #ifndef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 1 + #endif + #if PY_MAJOR_VERSION < 3 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #elif !defined(CYTHON_USE_ASYNC_SLOTS) + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #ifndef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 1 + #endif + #ifndef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 1 + #endif + #ifndef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 1 + #endif + #if PY_VERSION_HEX < 0x030300F0 || PY_VERSION_HEX >= 0x030B00A2 + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #elif !defined(CYTHON_USE_UNICODE_WRITER) + #define CYTHON_USE_UNICODE_WRITER 1 + #endif + #ifndef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 0 + #endif + #ifndef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 1 + #endif + #ifndef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 1 + #endif + #ifndef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 1 + #endif + #ifndef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL (PY_MAJOR_VERSION < 3 || PY_VERSION_HEX >= 0x03060000 && PY_VERSION_HEX < 0x030C00A6) + #endif + #ifndef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL (PY_VERSION_HEX >= 0x030700A1) + #endif + #ifndef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 1 + #endif + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS 1 + #endif + #if PY_VERSION_HEX < 0x03050000 + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 0 + #elif !defined(CYTHON_PEP489_MULTI_PHASE_INIT) + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #endif + #ifndef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #endif + #if PY_VERSION_HEX < 0x030400a1 + #undef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 0 + #elif !defined(CYTHON_USE_TP_FINALIZE) + #define CYTHON_USE_TP_FINALIZE 1 + #endif + #if PY_VERSION_HEX < 0x030600B1 + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #elif !defined(CYTHON_USE_DICT_VERSIONS) + #define CYTHON_USE_DICT_VERSIONS (PY_VERSION_HEX < 0x030C00A5) + #endif + #if PY_VERSION_HEX < 0x030700A3 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #elif !defined(CYTHON_USE_EXC_INFO_STACK) + #define CYTHON_USE_EXC_INFO_STACK 1 + #endif + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 1 + #endif + #ifndef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 1 + #endif +#endif +#if !defined(CYTHON_FAST_PYCCALL) +#define CYTHON_FAST_PYCCALL (CYTHON_FAST_PYCALL && PY_VERSION_HEX >= 0x030600B1) +#endif +#if !defined(CYTHON_VECTORCALL) +#define CYTHON_VECTORCALL (CYTHON_FAST_PYCCALL && PY_VERSION_HEX >= 0x030800B1) +#endif +#define CYTHON_BACKPORT_VECTORCALL (CYTHON_METH_FASTCALL && PY_VERSION_HEX < 0x030800B1) +#if CYTHON_USE_PYLONG_INTERNALS + #if PY_MAJOR_VERSION < 3 + #include "longintrepr.h" + #endif + #undef SHIFT + #undef BASE + #undef MASK + #ifdef SIZEOF_VOID_P + enum { __pyx_check_sizeof_voidp = 1 / (int)(SIZEOF_VOID_P == sizeof(void*)) }; + #endif +#endif +#ifndef __has_attribute + #define __has_attribute(x) 0 +#endif +#ifndef __has_cpp_attribute + #define __has_cpp_attribute(x) 0 +#endif +#ifndef CYTHON_RESTRICT + #if defined(__GNUC__) + #define CYTHON_RESTRICT __restrict__ + #elif defined(_MSC_VER) && _MSC_VER >= 1400 + #define CYTHON_RESTRICT __restrict + #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L + #define CYTHON_RESTRICT restrict + #else + #define CYTHON_RESTRICT + #endif +#endif +#ifndef CYTHON_UNUSED + #if defined(__cplusplus) + /* for clang __has_cpp_attribute(maybe_unused) is true even before C++17 + * but leads to warnings with -pedantic, since it is a C++17 feature */ + #if ((defined(_MSVC_LANG) && _MSVC_LANG >= 201703L) || __cplusplus >= 201703L) + #if __has_cpp_attribute(maybe_unused) + #define CYTHON_UNUSED [[maybe_unused]] + #endif + #endif + #endif +#endif +#ifndef CYTHON_UNUSED +# if defined(__GNUC__) +# if !(defined(__cplusplus)) || (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4)) +# define CYTHON_UNUSED __attribute__ ((__unused__)) +# else +# define CYTHON_UNUSED +# endif +# elif defined(__ICC) || (defined(__INTEL_COMPILER) && !defined(_MSC_VER)) +# define CYTHON_UNUSED __attribute__ ((__unused__)) +# else +# define CYTHON_UNUSED +# endif +#endif +#ifndef CYTHON_UNUSED_VAR +# if defined(__cplusplus) + template void CYTHON_UNUSED_VAR( const T& ) { } +# else +# define CYTHON_UNUSED_VAR(x) (void)(x) +# endif +#endif +#ifndef CYTHON_MAYBE_UNUSED_VAR + #define CYTHON_MAYBE_UNUSED_VAR(x) CYTHON_UNUSED_VAR(x) +#endif +#ifndef CYTHON_NCP_UNUSED +# if CYTHON_COMPILING_IN_CPYTHON +# define CYTHON_NCP_UNUSED +# else +# define CYTHON_NCP_UNUSED CYTHON_UNUSED +# endif +#endif +#ifndef CYTHON_USE_CPP_STD_MOVE + #if defined(__cplusplus) && (\ + __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1600)) + #define CYTHON_USE_CPP_STD_MOVE 1 + #else + #define CYTHON_USE_CPP_STD_MOVE 0 + #endif +#endif +#define __Pyx_void_to_None(void_result) ((void)(void_result), Py_INCREF(Py_None), Py_None) +#ifdef _MSC_VER + #ifndef _MSC_STDINT_H_ + #if _MSC_VER < 1300 + typedef unsigned char uint8_t; + typedef unsigned short uint16_t; + typedef unsigned int uint32_t; + #else + typedef unsigned __int8 uint8_t; + typedef unsigned __int16 uint16_t; + typedef unsigned __int32 uint32_t; + #endif + #endif + #if _MSC_VER < 1300 + #ifdef _WIN64 + typedef unsigned long long __pyx_uintptr_t; + #else + typedef unsigned int __pyx_uintptr_t; + #endif + #else + #ifdef _WIN64 + typedef unsigned __int64 __pyx_uintptr_t; + #else + typedef unsigned __int32 __pyx_uintptr_t; + #endif + #endif +#else + #include + typedef uintptr_t __pyx_uintptr_t; +#endif +#ifndef CYTHON_FALLTHROUGH + #if defined(__cplusplus) + /* for clang __has_cpp_attribute(fallthrough) is true even before C++17 + * but leads to warnings with -pedantic, since it is a C++17 feature */ + #if ((defined(_MSVC_LANG) && _MSVC_LANG >= 201703L) || __cplusplus >= 201703L) + #if __has_cpp_attribute(fallthrough) + #define CYTHON_FALLTHROUGH [[fallthrough]] + #endif + #endif + #ifndef CYTHON_FALLTHROUGH + #if __has_cpp_attribute(clang::fallthrough) + #define CYTHON_FALLTHROUGH [[clang::fallthrough]] + #elif __has_cpp_attribute(gnu::fallthrough) + #define CYTHON_FALLTHROUGH [[gnu::fallthrough]] + #endif + #endif + #endif + #ifndef CYTHON_FALLTHROUGH + #if __has_attribute(fallthrough) + #define CYTHON_FALLTHROUGH __attribute__((fallthrough)) + #else + #define CYTHON_FALLTHROUGH + #endif + #endif + #if defined(__clang__) && defined(__apple_build_version__) + #if __apple_build_version__ < 7000000 + #undef CYTHON_FALLTHROUGH + #define CYTHON_FALLTHROUGH + #endif + #endif +#endif +#ifdef __cplusplus + template + struct __PYX_IS_UNSIGNED_IMPL {static const bool value = T(0) < T(-1);}; + #define __PYX_IS_UNSIGNED(type) (__PYX_IS_UNSIGNED_IMPL::value) +#else + #define __PYX_IS_UNSIGNED(type) (((type)-1) > 0) +#endif +#if CYTHON_COMPILING_IN_PYPY == 1 + #define __PYX_NEED_TP_PRINT_SLOT (PY_VERSION_HEX >= 0x030800b4 && PY_VERSION_HEX < 0x030A0000) +#else + #define __PYX_NEED_TP_PRINT_SLOT (PY_VERSION_HEX >= 0x030800b4 && PY_VERSION_HEX < 0x03090000) +#endif +#define __PYX_REINTERPRET_FUNCION(func_pointer, other_pointer) ((func_pointer)(void(*)(void))(other_pointer)) + +#ifndef __cplusplus + #error "Cython files generated with the C++ option must be compiled with a C++ compiler." +#endif +#ifndef CYTHON_INLINE + #if defined(__clang__) + #define CYTHON_INLINE __inline__ __attribute__ ((__unused__)) + #else + #define CYTHON_INLINE inline + #endif +#endif +template +void __Pyx_call_destructor(T& x) { + x.~T(); +} +template +class __Pyx_FakeReference { + public: + __Pyx_FakeReference() : ptr(NULL) { } + __Pyx_FakeReference(const T& ref) : ptr(const_cast(&ref)) { } + T *operator->() { return ptr; } + T *operator&() { return ptr; } + operator T&() { return *ptr; } + template bool operator ==(const U& other) const { return *ptr == other; } + template bool operator !=(const U& other) const { return *ptr != other; } + template bool operator==(const __Pyx_FakeReference& other) const { return *ptr == *other.ptr; } + template bool operator!=(const __Pyx_FakeReference& other) const { return *ptr != *other.ptr; } + private: + T *ptr; +}; + +#define __PYX_BUILD_PY_SSIZE_T "n" +#define CYTHON_FORMAT_SSIZE_T "z" +#if PY_MAJOR_VERSION < 3 + #define __Pyx_BUILTIN_MODULE_NAME "__builtin__" + #define __Pyx_DefaultClassType PyClass_Type + #define __Pyx_PyCode_New(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\ + PyCode_New(a+k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos) +#else + #define __Pyx_BUILTIN_MODULE_NAME "builtins" + #define __Pyx_DefaultClassType PyType_Type +#if CYTHON_COMPILING_IN_LIMITED_API + static CYTHON_INLINE PyObject* __Pyx_PyCode_New(int a, int p, int k, int l, int s, int f, + PyObject *code, PyObject *c, PyObject* n, PyObject *v, + PyObject *fv, PyObject *cell, PyObject* fn, + PyObject *name, int fline, PyObject *lnos) { + PyObject *exception_table = NULL; + PyObject *types_module=NULL, *code_type=NULL, *result=NULL; + #if __PYX_LIMITED_VERSION_HEX < 0x030B0000 + PyObject *version_info; + PyObject *py_minor_version = NULL; + #endif + long minor_version = 0; + PyObject *type, *value, *traceback; + PyErr_Fetch(&type, &value, &traceback); + #if __PYX_LIMITED_VERSION_HEX >= 0x030B0000 + minor_version = 11; + #else + if (!(version_info = PySys_GetObject("version_info"))) goto end; + if (!(py_minor_version = PySequence_GetItem(version_info, 1))) goto end; + minor_version = PyLong_AsLong(py_minor_version); + Py_DECREF(py_minor_version); + if (minor_version == -1 && PyErr_Occurred()) goto end; + #endif + if (!(types_module = PyImport_ImportModule("types"))) goto end; + if (!(code_type = PyObject_GetAttrString(types_module, "CodeType"))) goto end; + if (minor_version <= 7) { + (void)p; + result = PyObject_CallFunction(code_type, "iiiiiOOOOOOiOO", a, k, l, s, f, code, + c, n, v, fn, name, fline, lnos, fv, cell); + } else if (minor_version <= 10) { + result = PyObject_CallFunction(code_type, "iiiiiiOOOOOOiOO", a,p, k, l, s, f, code, + c, n, v, fn, name, fline, lnos, fv, cell); + } else { + if (!(exception_table = PyBytes_FromStringAndSize(NULL, 0))) goto end; + result = PyObject_CallFunction(code_type, "iiiiiiOOOOOOOiOO", a,p, k, l, s, f, code, + c, n, v, fn, name, name, fline, lnos, exception_table, fv, cell); + } + end: + Py_XDECREF(code_type); + Py_XDECREF(exception_table); + Py_XDECREF(types_module); + if (type) { + PyErr_Restore(type, value, traceback); + } + return result; + } + #ifndef CO_OPTIMIZED + #define CO_OPTIMIZED 0x0001 + #endif + #ifndef CO_NEWLOCALS + #define CO_NEWLOCALS 0x0002 + #endif + #ifndef CO_VARARGS + #define CO_VARARGS 0x0004 + #endif + #ifndef CO_VARKEYWORDS + #define CO_VARKEYWORDS 0x0008 + #endif + #ifndef CO_ASYNC_GENERATOR + #define CO_ASYNC_GENERATOR 0x0200 + #endif + #ifndef CO_GENERATOR + #define CO_GENERATOR 0x0020 + #endif + #ifndef CO_COROUTINE + #define CO_COROUTINE 0x0080 + #endif +#elif PY_VERSION_HEX >= 0x030B0000 + static CYTHON_INLINE PyCodeObject* __Pyx_PyCode_New(int a, int p, int k, int l, int s, int f, + PyObject *code, PyObject *c, PyObject* n, PyObject *v, + PyObject *fv, PyObject *cell, PyObject* fn, + PyObject *name, int fline, PyObject *lnos) { + PyCodeObject *result; + PyObject *empty_bytes = PyBytes_FromStringAndSize("", 0); + if (!empty_bytes) return NULL; + result = + #if PY_VERSION_HEX >= 0x030C0000 + PyUnstable_Code_NewWithPosOnlyArgs + #else + PyCode_NewWithPosOnlyArgs + #endif + (a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, name, fline, lnos, empty_bytes); + Py_DECREF(empty_bytes); + return result; + } +#elif PY_VERSION_HEX >= 0x030800B2 && !CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyCode_New(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\ + PyCode_NewWithPosOnlyArgs(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos) +#else + #define __Pyx_PyCode_New(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\ + PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos) +#endif +#endif +#if PY_VERSION_HEX >= 0x030900A4 || defined(Py_IS_TYPE) + #define __Pyx_IS_TYPE(ob, type) Py_IS_TYPE(ob, type) +#else + #define __Pyx_IS_TYPE(ob, type) (((const PyObject*)ob)->ob_type == (type)) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_Is) + #define __Pyx_Py_Is(x, y) Py_Is(x, y) +#else + #define __Pyx_Py_Is(x, y) ((x) == (y)) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_IsNone) + #define __Pyx_Py_IsNone(ob) Py_IsNone(ob) +#else + #define __Pyx_Py_IsNone(ob) __Pyx_Py_Is((ob), Py_None) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_IsTrue) + #define __Pyx_Py_IsTrue(ob) Py_IsTrue(ob) +#else + #define __Pyx_Py_IsTrue(ob) __Pyx_Py_Is((ob), Py_True) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_IsFalse) + #define __Pyx_Py_IsFalse(ob) Py_IsFalse(ob) +#else + #define __Pyx_Py_IsFalse(ob) __Pyx_Py_Is((ob), Py_False) +#endif +#define __Pyx_NoneAsNull(obj) (__Pyx_Py_IsNone(obj) ? NULL : (obj)) +#if PY_VERSION_HEX >= 0x030900F0 && !CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyObject_GC_IsFinalized(o) PyObject_GC_IsFinalized(o) +#else + #define __Pyx_PyObject_GC_IsFinalized(o) _PyGC_FINALIZED(o) +#endif +#ifndef CO_COROUTINE + #define CO_COROUTINE 0x80 +#endif +#ifndef CO_ASYNC_GENERATOR + #define CO_ASYNC_GENERATOR 0x200 +#endif +#ifndef Py_TPFLAGS_CHECKTYPES + #define Py_TPFLAGS_CHECKTYPES 0 +#endif +#ifndef Py_TPFLAGS_HAVE_INDEX + #define Py_TPFLAGS_HAVE_INDEX 0 +#endif +#ifndef Py_TPFLAGS_HAVE_NEWBUFFER + #define Py_TPFLAGS_HAVE_NEWBUFFER 0 +#endif +#ifndef Py_TPFLAGS_HAVE_FINALIZE + #define Py_TPFLAGS_HAVE_FINALIZE 0 +#endif +#ifndef Py_TPFLAGS_SEQUENCE + #define Py_TPFLAGS_SEQUENCE 0 +#endif +#ifndef Py_TPFLAGS_MAPPING + #define Py_TPFLAGS_MAPPING 0 +#endif +#ifndef METH_STACKLESS + #define METH_STACKLESS 0 +#endif +#if PY_VERSION_HEX <= 0x030700A3 || !defined(METH_FASTCALL) + #ifndef METH_FASTCALL + #define METH_FASTCALL 0x80 + #endif + typedef PyObject *(*__Pyx_PyCFunctionFast) (PyObject *self, PyObject *const *args, Py_ssize_t nargs); + typedef PyObject *(*__Pyx_PyCFunctionFastWithKeywords) (PyObject *self, PyObject *const *args, + Py_ssize_t nargs, PyObject *kwnames); +#else + #if PY_VERSION_HEX >= 0x030d00A4 + # define __Pyx_PyCFunctionFast PyCFunctionFast + # define __Pyx_PyCFunctionFastWithKeywords PyCFunctionFastWithKeywords + #else + # define __Pyx_PyCFunctionFast _PyCFunctionFast + # define __Pyx_PyCFunctionFastWithKeywords _PyCFunctionFastWithKeywords + #endif +#endif +#if CYTHON_METH_FASTCALL + #define __Pyx_METH_FASTCALL METH_FASTCALL + #define __Pyx_PyCFunction_FastCall __Pyx_PyCFunctionFast + #define __Pyx_PyCFunction_FastCallWithKeywords __Pyx_PyCFunctionFastWithKeywords +#else + #define __Pyx_METH_FASTCALL METH_VARARGS + #define __Pyx_PyCFunction_FastCall PyCFunction + #define __Pyx_PyCFunction_FastCallWithKeywords PyCFunctionWithKeywords +#endif +#if CYTHON_VECTORCALL + #define __pyx_vectorcallfunc vectorcallfunc + #define __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET PY_VECTORCALL_ARGUMENTS_OFFSET + #define __Pyx_PyVectorcall_NARGS(n) PyVectorcall_NARGS((size_t)(n)) +#elif CYTHON_BACKPORT_VECTORCALL + typedef PyObject *(*__pyx_vectorcallfunc)(PyObject *callable, PyObject *const *args, + size_t nargsf, PyObject *kwnames); + #define __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET ((size_t)1 << (8 * sizeof(size_t) - 1)) + #define __Pyx_PyVectorcall_NARGS(n) ((Py_ssize_t)(((size_t)(n)) & ~__Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET)) +#else + #define __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET 0 + #define __Pyx_PyVectorcall_NARGS(n) ((Py_ssize_t)(n)) +#endif +#if PY_MAJOR_VERSION >= 0x030900B1 +#define __Pyx_PyCFunction_CheckExact(func) PyCFunction_CheckExact(func) +#else +#define __Pyx_PyCFunction_CheckExact(func) PyCFunction_Check(func) +#endif +#define __Pyx_CyOrPyCFunction_Check(func) PyCFunction_Check(func) +#if CYTHON_COMPILING_IN_CPYTHON +#define __Pyx_CyOrPyCFunction_GET_FUNCTION(func) (((PyCFunctionObject*)(func))->m_ml->ml_meth) +#elif !CYTHON_COMPILING_IN_LIMITED_API +#define __Pyx_CyOrPyCFunction_GET_FUNCTION(func) PyCFunction_GET_FUNCTION(func) +#endif +#if CYTHON_COMPILING_IN_CPYTHON +#define __Pyx_CyOrPyCFunction_GET_FLAGS(func) (((PyCFunctionObject*)(func))->m_ml->ml_flags) +static CYTHON_INLINE PyObject* __Pyx_CyOrPyCFunction_GET_SELF(PyObject *func) { + return (__Pyx_CyOrPyCFunction_GET_FLAGS(func) & METH_STATIC) ? NULL : ((PyCFunctionObject*)func)->m_self; +} +#endif +static CYTHON_INLINE int __Pyx__IsSameCFunction(PyObject *func, void *cfunc) { +#if CYTHON_COMPILING_IN_LIMITED_API + return PyCFunction_Check(func) && PyCFunction_GetFunction(func) == (PyCFunction) cfunc; +#else + return PyCFunction_Check(func) && PyCFunction_GET_FUNCTION(func) == (PyCFunction) cfunc; +#endif +} +#define __Pyx_IsSameCFunction(func, cfunc) __Pyx__IsSameCFunction(func, cfunc) +#if __PYX_LIMITED_VERSION_HEX < 0x030900B1 + #define __Pyx_PyType_FromModuleAndSpec(m, s, b) ((void)m, PyType_FromSpecWithBases(s, b)) + typedef PyObject *(*__Pyx_PyCMethod)(PyObject *, PyTypeObject *, PyObject *const *, size_t, PyObject *); +#else + #define __Pyx_PyType_FromModuleAndSpec(m, s, b) PyType_FromModuleAndSpec(m, s, b) + #define __Pyx_PyCMethod PyCMethod +#endif +#ifndef METH_METHOD + #define METH_METHOD 0x200 +#endif +#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Malloc) + #define PyObject_Malloc(s) PyMem_Malloc(s) + #define PyObject_Free(p) PyMem_Free(p) + #define PyObject_Realloc(p) PyMem_Realloc(p) +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_PyCode_HasFreeVars(co) (PyCode_GetNumFree(co) > 0) + #define __Pyx_PyFrame_SetLineNumber(frame, lineno) +#else + #define __Pyx_PyCode_HasFreeVars(co) (PyCode_GetNumFree(co) > 0) + #define __Pyx_PyFrame_SetLineNumber(frame, lineno) (frame)->f_lineno = (lineno) +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_PyThreadState_Current PyThreadState_Get() +#elif !CYTHON_FAST_THREAD_STATE + #define __Pyx_PyThreadState_Current PyThreadState_GET() +#elif PY_VERSION_HEX >= 0x030d00A1 + #define __Pyx_PyThreadState_Current PyThreadState_GetUnchecked() +#elif PY_VERSION_HEX >= 0x03060000 + #define __Pyx_PyThreadState_Current _PyThreadState_UncheckedGet() +#elif PY_VERSION_HEX >= 0x03000000 + #define __Pyx_PyThreadState_Current PyThreadState_GET() +#else + #define __Pyx_PyThreadState_Current _PyThreadState_Current +#endif +#if CYTHON_COMPILING_IN_LIMITED_API +static CYTHON_INLINE void *__Pyx_PyModule_GetState(PyObject *op) +{ + void *result; + result = PyModule_GetState(op); + if (!result) + Py_FatalError("Couldn't find the module state"); + return result; +} +#endif +#define __Pyx_PyObject_GetSlot(obj, name, func_ctype) __Pyx_PyType_GetSlot(Py_TYPE(obj), name, func_ctype) +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_PyType_GetSlot(type, name, func_ctype) ((func_ctype) PyType_GetSlot((type), Py_##name)) +#else + #define __Pyx_PyType_GetSlot(type, name, func_ctype) ((type)->name) +#endif +#if PY_VERSION_HEX < 0x030700A2 && !defined(PyThread_tss_create) && !defined(Py_tss_NEEDS_INIT) +#include "pythread.h" +#define Py_tss_NEEDS_INIT 0 +typedef int Py_tss_t; +static CYTHON_INLINE int PyThread_tss_create(Py_tss_t *key) { + *key = PyThread_create_key(); + return 0; +} +static CYTHON_INLINE Py_tss_t * PyThread_tss_alloc(void) { + Py_tss_t *key = (Py_tss_t *)PyObject_Malloc(sizeof(Py_tss_t)); + *key = Py_tss_NEEDS_INIT; + return key; +} +static CYTHON_INLINE void PyThread_tss_free(Py_tss_t *key) { + PyObject_Free(key); +} +static CYTHON_INLINE int PyThread_tss_is_created(Py_tss_t *key) { + return *key != Py_tss_NEEDS_INIT; +} +static CYTHON_INLINE void PyThread_tss_delete(Py_tss_t *key) { + PyThread_delete_key(*key); + *key = Py_tss_NEEDS_INIT; +} +static CYTHON_INLINE int PyThread_tss_set(Py_tss_t *key, void *value) { + return PyThread_set_key_value(*key, value); +} +static CYTHON_INLINE void * PyThread_tss_get(Py_tss_t *key) { + return PyThread_get_key_value(*key); +} +#endif +#if PY_MAJOR_VERSION < 3 + #if CYTHON_COMPILING_IN_PYPY + #if PYPY_VERSION_NUM < 0x07030600 + #if defined(__cplusplus) && __cplusplus >= 201402L + [[deprecated("`with nogil:` inside a nogil function will not release the GIL in PyPy2 < 7.3.6")]] + #elif defined(__GNUC__) || defined(__clang__) + __attribute__ ((__deprecated__("`with nogil:` inside a nogil function will not release the GIL in PyPy2 < 7.3.6"))) + #elif defined(_MSC_VER) + __declspec(deprecated("`with nogil:` inside a nogil function will not release the GIL in PyPy2 < 7.3.6")) + #endif + static CYTHON_INLINE int PyGILState_Check(void) { + return 0; + } + #else // PYPY_VERSION_NUM < 0x07030600 + #endif // PYPY_VERSION_NUM < 0x07030600 + #else + static CYTHON_INLINE int PyGILState_Check(void) { + PyThreadState * tstate = _PyThreadState_Current; + return tstate && (tstate == PyGILState_GetThisThreadState()); + } + #endif +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030d0000 || defined(_PyDict_NewPresized) +#define __Pyx_PyDict_NewPresized(n) ((n <= 8) ? PyDict_New() : _PyDict_NewPresized(n)) +#else +#define __Pyx_PyDict_NewPresized(n) PyDict_New() +#endif +#if PY_MAJOR_VERSION >= 3 || CYTHON_FUTURE_DIVISION + #define __Pyx_PyNumber_Divide(x,y) PyNumber_TrueDivide(x,y) + #define __Pyx_PyNumber_InPlaceDivide(x,y) PyNumber_InPlaceTrueDivide(x,y) +#else + #define __Pyx_PyNumber_Divide(x,y) PyNumber_Divide(x,y) + #define __Pyx_PyNumber_InPlaceDivide(x,y) PyNumber_InPlaceDivide(x,y) +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX > 0x030600B4 && PY_VERSION_HEX < 0x030d0000 && CYTHON_USE_UNICODE_INTERNALS +#define __Pyx_PyDict_GetItemStrWithError(dict, name) _PyDict_GetItem_KnownHash(dict, name, ((PyASCIIObject *) name)->hash) +static CYTHON_INLINE PyObject * __Pyx_PyDict_GetItemStr(PyObject *dict, PyObject *name) { + PyObject *res = __Pyx_PyDict_GetItemStrWithError(dict, name); + if (res == NULL) PyErr_Clear(); + return res; +} +#elif PY_MAJOR_VERSION >= 3 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07020000) +#define __Pyx_PyDict_GetItemStrWithError PyDict_GetItemWithError +#define __Pyx_PyDict_GetItemStr PyDict_GetItem +#else +static CYTHON_INLINE PyObject * __Pyx_PyDict_GetItemStrWithError(PyObject *dict, PyObject *name) { +#if CYTHON_COMPILING_IN_PYPY + return PyDict_GetItem(dict, name); +#else + PyDictEntry *ep; + PyDictObject *mp = (PyDictObject*) dict; + long hash = ((PyStringObject *) name)->ob_shash; + assert(hash != -1); + ep = (mp->ma_lookup)(mp, name, hash); + if (ep == NULL) { + return NULL; + } + return ep->me_value; +#endif +} +#define __Pyx_PyDict_GetItemStr PyDict_GetItem +#endif +#if CYTHON_USE_TYPE_SLOTS + #define __Pyx_PyType_GetFlags(tp) (((PyTypeObject *)tp)->tp_flags) + #define __Pyx_PyType_HasFeature(type, feature) ((__Pyx_PyType_GetFlags(type) & (feature)) != 0) + #define __Pyx_PyObject_GetIterNextFunc(obj) (Py_TYPE(obj)->tp_iternext) +#else + #define __Pyx_PyType_GetFlags(tp) (PyType_GetFlags((PyTypeObject *)tp)) + #define __Pyx_PyType_HasFeature(type, feature) PyType_HasFeature(type, feature) + #define __Pyx_PyObject_GetIterNextFunc(obj) PyIter_Next +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_SetItemOnTypeDict(tp, k, v) PyObject_GenericSetAttr((PyObject*)tp, k, v) +#else + #define __Pyx_SetItemOnTypeDict(tp, k, v) PyDict_SetItem(tp->tp_dict, k, v) +#endif +#if CYTHON_USE_TYPE_SPECS && PY_VERSION_HEX >= 0x03080000 +#define __Pyx_PyHeapTypeObject_GC_Del(obj) {\ + PyTypeObject *type = Py_TYPE((PyObject*)obj);\ + assert(__Pyx_PyType_HasFeature(type, Py_TPFLAGS_HEAPTYPE));\ + PyObject_GC_Del(obj);\ + Py_DECREF(type);\ +} +#else +#define __Pyx_PyHeapTypeObject_GC_Del(obj) PyObject_GC_Del(obj) +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define CYTHON_PEP393_ENABLED 1 + #define __Pyx_PyUnicode_READY(op) (0) + #define __Pyx_PyUnicode_GET_LENGTH(u) PyUnicode_GetLength(u) + #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_ReadChar(u, i) + #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u) ((void)u, 1114111U) + #define __Pyx_PyUnicode_KIND(u) ((void)u, (0)) + #define __Pyx_PyUnicode_DATA(u) ((void*)u) + #define __Pyx_PyUnicode_READ(k, d, i) ((void)k, PyUnicode_ReadChar((PyObject*)(d), i)) + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != PyUnicode_GetLength(u)) +#elif PY_VERSION_HEX > 0x03030000 && defined(PyUnicode_KIND) + #define CYTHON_PEP393_ENABLED 1 + #if PY_VERSION_HEX >= 0x030C0000 + #define __Pyx_PyUnicode_READY(op) (0) + #else + #define __Pyx_PyUnicode_READY(op) (likely(PyUnicode_IS_READY(op)) ?\ + 0 : _PyUnicode_Ready((PyObject *)(op))) + #endif + #define __Pyx_PyUnicode_GET_LENGTH(u) PyUnicode_GET_LENGTH(u) + #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_READ_CHAR(u, i) + #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u) PyUnicode_MAX_CHAR_VALUE(u) + #define __Pyx_PyUnicode_KIND(u) ((int)PyUnicode_KIND(u)) + #define __Pyx_PyUnicode_DATA(u) PyUnicode_DATA(u) + #define __Pyx_PyUnicode_READ(k, d, i) PyUnicode_READ(k, d, i) + #define __Pyx_PyUnicode_WRITE(k, d, i, ch) PyUnicode_WRITE(k, d, i, (Py_UCS4) ch) + #if PY_VERSION_HEX >= 0x030C0000 + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != PyUnicode_GET_LENGTH(u)) + #else + #if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x03090000 + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : ((PyCompactUnicodeObject *)(u))->wstr_length)) + #else + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : PyUnicode_GET_SIZE(u))) + #endif + #endif +#else + #define CYTHON_PEP393_ENABLED 0 + #define PyUnicode_1BYTE_KIND 1 + #define PyUnicode_2BYTE_KIND 2 + #define PyUnicode_4BYTE_KIND 4 + #define __Pyx_PyUnicode_READY(op) (0) + #define __Pyx_PyUnicode_GET_LENGTH(u) PyUnicode_GET_SIZE(u) + #define __Pyx_PyUnicode_READ_CHAR(u, i) ((Py_UCS4)(PyUnicode_AS_UNICODE(u)[i])) + #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u) ((sizeof(Py_UNICODE) == 2) ? 65535U : 1114111U) + #define __Pyx_PyUnicode_KIND(u) ((int)sizeof(Py_UNICODE)) + #define __Pyx_PyUnicode_DATA(u) ((void*)PyUnicode_AS_UNICODE(u)) + #define __Pyx_PyUnicode_READ(k, d, i) ((void)(k), (Py_UCS4)(((Py_UNICODE*)d)[i])) + #define __Pyx_PyUnicode_WRITE(k, d, i, ch) (((void)(k)), ((Py_UNICODE*)d)[i] = (Py_UNICODE) ch) + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != PyUnicode_GET_SIZE(u)) +#endif +#if CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyUnicode_Concat(a, b) PyNumber_Add(a, b) + #define __Pyx_PyUnicode_ConcatSafe(a, b) PyNumber_Add(a, b) +#else + #define __Pyx_PyUnicode_Concat(a, b) PyUnicode_Concat(a, b) + #define __Pyx_PyUnicode_ConcatSafe(a, b) ((unlikely((a) == Py_None) || unlikely((b) == Py_None)) ?\ + PyNumber_Add(a, b) : __Pyx_PyUnicode_Concat(a, b)) +#endif +#if CYTHON_COMPILING_IN_PYPY + #if !defined(PyUnicode_DecodeUnicodeEscape) + #define PyUnicode_DecodeUnicodeEscape(s, size, errors) PyUnicode_Decode(s, size, "unicode_escape", errors) + #endif + #if !defined(PyUnicode_Contains) || (PY_MAJOR_VERSION == 2 && PYPY_VERSION_NUM < 0x07030500) + #undef PyUnicode_Contains + #define PyUnicode_Contains(u, s) PySequence_Contains(u, s) + #endif + #if !defined(PyByteArray_Check) + #define PyByteArray_Check(obj) PyObject_TypeCheck(obj, &PyByteArray_Type) + #endif + #if !defined(PyObject_Format) + #define PyObject_Format(obj, fmt) PyObject_CallMethod(obj, "__format__", "O", fmt) + #endif +#endif +#define __Pyx_PyString_FormatSafe(a, b) ((unlikely((a) == Py_None || (PyString_Check(b) && !PyString_CheckExact(b)))) ? PyNumber_Remainder(a, b) : __Pyx_PyString_Format(a, b)) +#define __Pyx_PyUnicode_FormatSafe(a, b) ((unlikely((a) == Py_None || (PyUnicode_Check(b) && !PyUnicode_CheckExact(b)))) ? PyNumber_Remainder(a, b) : PyUnicode_Format(a, b)) +#if PY_MAJOR_VERSION >= 3 + #define __Pyx_PyString_Format(a, b) PyUnicode_Format(a, b) +#else + #define __Pyx_PyString_Format(a, b) PyString_Format(a, b) +#endif +#if PY_MAJOR_VERSION < 3 && !defined(PyObject_ASCII) + #define PyObject_ASCII(o) PyObject_Repr(o) +#endif +#if PY_MAJOR_VERSION >= 3 + #define PyBaseString_Type PyUnicode_Type + #define PyStringObject PyUnicodeObject + #define PyString_Type PyUnicode_Type + #define PyString_Check PyUnicode_Check + #define PyString_CheckExact PyUnicode_CheckExact +#ifndef PyObject_Unicode + #define PyObject_Unicode PyObject_Str +#endif +#endif +#if PY_MAJOR_VERSION >= 3 + #define __Pyx_PyBaseString_Check(obj) PyUnicode_Check(obj) + #define __Pyx_PyBaseString_CheckExact(obj) PyUnicode_CheckExact(obj) +#else + #define __Pyx_PyBaseString_Check(obj) (PyString_Check(obj) || PyUnicode_Check(obj)) + #define __Pyx_PyBaseString_CheckExact(obj) (PyString_CheckExact(obj) || PyUnicode_CheckExact(obj)) +#endif +#if CYTHON_COMPILING_IN_CPYTHON + #define __Pyx_PySequence_ListKeepNew(obj)\ + (likely(PyList_CheckExact(obj) && Py_REFCNT(obj) == 1) ? __Pyx_NewRef(obj) : PySequence_List(obj)) +#else + #define __Pyx_PySequence_ListKeepNew(obj) PySequence_List(obj) +#endif +#ifndef PySet_CheckExact + #define PySet_CheckExact(obj) __Pyx_IS_TYPE(obj, &PySet_Type) +#endif +#if PY_VERSION_HEX >= 0x030900A4 + #define __Pyx_SET_REFCNT(obj, refcnt) Py_SET_REFCNT(obj, refcnt) + #define __Pyx_SET_SIZE(obj, size) Py_SET_SIZE(obj, size) +#else + #define __Pyx_SET_REFCNT(obj, refcnt) Py_REFCNT(obj) = (refcnt) + #define __Pyx_SET_SIZE(obj, size) Py_SIZE(obj) = (size) +#endif +#if CYTHON_ASSUME_SAFE_MACROS + #define __Pyx_PySequence_ITEM(o, i) PySequence_ITEM(o, i) + #define __Pyx_PySequence_SIZE(seq) Py_SIZE(seq) + #define __Pyx_PyTuple_SET_ITEM(o, i, v) (PyTuple_SET_ITEM(o, i, v), (0)) + #define __Pyx_PyList_SET_ITEM(o, i, v) (PyList_SET_ITEM(o, i, v), (0)) + #define __Pyx_PyTuple_GET_SIZE(o) PyTuple_GET_SIZE(o) + #define __Pyx_PyList_GET_SIZE(o) PyList_GET_SIZE(o) + #define __Pyx_PySet_GET_SIZE(o) PySet_GET_SIZE(o) + #define __Pyx_PyBytes_GET_SIZE(o) PyBytes_GET_SIZE(o) + #define __Pyx_PyByteArray_GET_SIZE(o) PyByteArray_GET_SIZE(o) +#else + #define __Pyx_PySequence_ITEM(o, i) PySequence_GetItem(o, i) + #define __Pyx_PySequence_SIZE(seq) PySequence_Size(seq) + #define __Pyx_PyTuple_SET_ITEM(o, i, v) PyTuple_SetItem(o, i, v) + #define __Pyx_PyList_SET_ITEM(o, i, v) PyList_SetItem(o, i, v) + #define __Pyx_PyTuple_GET_SIZE(o) PyTuple_Size(o) + #define __Pyx_PyList_GET_SIZE(o) PyList_Size(o) + #define __Pyx_PySet_GET_SIZE(o) PySet_Size(o) + #define __Pyx_PyBytes_GET_SIZE(o) PyBytes_Size(o) + #define __Pyx_PyByteArray_GET_SIZE(o) PyByteArray_Size(o) +#endif +#if __PYX_LIMITED_VERSION_HEX >= 0x030d00A1 + #define __Pyx_PyImport_AddModuleRef(name) PyImport_AddModuleRef(name) +#else + static CYTHON_INLINE PyObject *__Pyx_PyImport_AddModuleRef(const char *name) { + PyObject *module = PyImport_AddModule(name); + Py_XINCREF(module); + return module; + } +#endif +#if PY_MAJOR_VERSION >= 3 + #define PyIntObject PyLongObject + #define PyInt_Type PyLong_Type + #define PyInt_Check(op) PyLong_Check(op) + #define PyInt_CheckExact(op) PyLong_CheckExact(op) + #define __Pyx_Py3Int_Check(op) PyLong_Check(op) + #define __Pyx_Py3Int_CheckExact(op) PyLong_CheckExact(op) + #define PyInt_FromString PyLong_FromString + #define PyInt_FromUnicode PyLong_FromUnicode + #define PyInt_FromLong PyLong_FromLong + #define PyInt_FromSize_t PyLong_FromSize_t + #define PyInt_FromSsize_t PyLong_FromSsize_t + #define PyInt_AsLong PyLong_AsLong + #define PyInt_AS_LONG PyLong_AS_LONG + #define PyInt_AsSsize_t PyLong_AsSsize_t + #define PyInt_AsUnsignedLongMask PyLong_AsUnsignedLongMask + #define PyInt_AsUnsignedLongLongMask PyLong_AsUnsignedLongLongMask + #define PyNumber_Int PyNumber_Long +#else + #define __Pyx_Py3Int_Check(op) (PyLong_Check(op) || PyInt_Check(op)) + #define __Pyx_Py3Int_CheckExact(op) (PyLong_CheckExact(op) || PyInt_CheckExact(op)) +#endif +#if PY_MAJOR_VERSION >= 3 + #define PyBoolObject PyLongObject +#endif +#if PY_MAJOR_VERSION >= 3 && CYTHON_COMPILING_IN_PYPY + #ifndef PyUnicode_InternFromString + #define PyUnicode_InternFromString(s) PyUnicode_FromString(s) + #endif +#endif +#if PY_VERSION_HEX < 0x030200A4 + typedef long Py_hash_t; + #define __Pyx_PyInt_FromHash_t PyInt_FromLong + #define __Pyx_PyInt_AsHash_t __Pyx_PyIndex_AsHash_t +#else + #define __Pyx_PyInt_FromHash_t PyInt_FromSsize_t + #define __Pyx_PyInt_AsHash_t __Pyx_PyIndex_AsSsize_t +#endif +#if CYTHON_USE_ASYNC_SLOTS + #if PY_VERSION_HEX >= 0x030500B1 + #define __Pyx_PyAsyncMethodsStruct PyAsyncMethods + #define __Pyx_PyType_AsAsync(obj) (Py_TYPE(obj)->tp_as_async) + #else + #define __Pyx_PyType_AsAsync(obj) ((__Pyx_PyAsyncMethodsStruct*) (Py_TYPE(obj)->tp_reserved)) + #endif +#else + #define __Pyx_PyType_AsAsync(obj) NULL +#endif +#ifndef __Pyx_PyAsyncMethodsStruct + typedef struct { + unaryfunc am_await; + unaryfunc am_aiter; + unaryfunc am_anext; + } __Pyx_PyAsyncMethodsStruct; +#endif + +#if defined(_WIN32) || defined(WIN32) || defined(MS_WINDOWS) + #if !defined(_USE_MATH_DEFINES) + #define _USE_MATH_DEFINES + #endif +#endif +#include +#ifdef NAN +#define __PYX_NAN() ((float) NAN) +#else +static CYTHON_INLINE float __PYX_NAN() { + float value; + memset(&value, 0xFF, sizeof(value)); + return value; +} +#endif +#if defined(__CYGWIN__) && defined(_LDBL_EQ_DBL) +#define __Pyx_truncl trunc +#else +#define __Pyx_truncl truncl +#endif + +#define __PYX_MARK_ERR_POS(f_index, lineno) \ + { __pyx_filename = __pyx_f[f_index]; (void)__pyx_filename; __pyx_lineno = lineno; (void)__pyx_lineno; __pyx_clineno = __LINE__; (void)__pyx_clineno; } +#define __PYX_ERR(f_index, lineno, Ln_error) \ + { __PYX_MARK_ERR_POS(f_index, lineno) goto Ln_error; } + +#ifdef CYTHON_EXTERN_C + #undef __PYX_EXTERN_C + #define __PYX_EXTERN_C CYTHON_EXTERN_C +#elif defined(__PYX_EXTERN_C) + #ifdef _MSC_VER + #pragma message ("Please do not define the '__PYX_EXTERN_C' macro externally. Use 'CYTHON_EXTERN_C' instead.") + #else + #warning Please do not define the '__PYX_EXTERN_C' macro externally. Use 'CYTHON_EXTERN_C' instead. + #endif +#else + #define __PYX_EXTERN_C extern "C++" +#endif + +#define __PYX_HAVE__fairseq__data__data_utils_fast +#define __PYX_HAVE_API__fairseq__data__data_utils_fast +/* Early includes */ +#include +#include + + /* Using NumPy API declarations from "numpy/__init__.cython-30.pxd" */ + +#include "numpy/arrayobject.h" +#include "numpy/ndarrayobject.h" +#include "numpy/ndarraytypes.h" +#include "numpy/arrayscalars.h" +#include "numpy/ufuncobject.h" +#include "pythread.h" +#include +#ifdef _OPENMP +#include +#endif /* _OPENMP */ + +#if defined(PYREX_WITHOUT_ASSERTIONS) && !defined(CYTHON_WITHOUT_ASSERTIONS) +#define CYTHON_WITHOUT_ASSERTIONS +#endif + +typedef struct {PyObject **p; const char *s; const Py_ssize_t n; const char* encoding; + const char is_unicode; const char is_str; const char intern; } __Pyx_StringTabEntry; + +#define __PYX_DEFAULT_STRING_ENCODING_IS_ASCII 0 +#define __PYX_DEFAULT_STRING_ENCODING_IS_UTF8 0 +#define __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT (PY_MAJOR_VERSION >= 3 && __PYX_DEFAULT_STRING_ENCODING_IS_UTF8) +#define __PYX_DEFAULT_STRING_ENCODING "" +#define __Pyx_PyObject_FromString __Pyx_PyBytes_FromString +#define __Pyx_PyObject_FromStringAndSize __Pyx_PyBytes_FromStringAndSize +#define __Pyx_uchar_cast(c) ((unsigned char)c) +#define __Pyx_long_cast(x) ((long)x) +#define __Pyx_fits_Py_ssize_t(v, type, is_signed) (\ + (sizeof(type) < sizeof(Py_ssize_t)) ||\ + (sizeof(type) > sizeof(Py_ssize_t) &&\ + likely(v < (type)PY_SSIZE_T_MAX ||\ + v == (type)PY_SSIZE_T_MAX) &&\ + (!is_signed || likely(v > (type)PY_SSIZE_T_MIN ||\ + v == (type)PY_SSIZE_T_MIN))) ||\ + (sizeof(type) == sizeof(Py_ssize_t) &&\ + (is_signed || likely(v < (type)PY_SSIZE_T_MAX ||\ + v == (type)PY_SSIZE_T_MAX))) ) +static CYTHON_INLINE int __Pyx_is_valid_index(Py_ssize_t i, Py_ssize_t limit) { + return (size_t) i < (size_t) limit; +} +#if defined (__cplusplus) && __cplusplus >= 201103L + #include + #define __Pyx_sst_abs(value) std::abs(value) +#elif SIZEOF_INT >= SIZEOF_SIZE_T + #define __Pyx_sst_abs(value) abs(value) +#elif SIZEOF_LONG >= SIZEOF_SIZE_T + #define __Pyx_sst_abs(value) labs(value) +#elif defined (_MSC_VER) + #define __Pyx_sst_abs(value) ((Py_ssize_t)_abs64(value)) +#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L + #define __Pyx_sst_abs(value) llabs(value) +#elif defined (__GNUC__) + #define __Pyx_sst_abs(value) __builtin_llabs(value) +#else + #define __Pyx_sst_abs(value) ((value<0) ? -value : value) +#endif +static CYTHON_INLINE Py_ssize_t __Pyx_ssize_strlen(const char *s); +static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject*); +static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject*, Py_ssize_t* length); +static CYTHON_INLINE PyObject* __Pyx_PyByteArray_FromString(const char*); +#define __Pyx_PyByteArray_FromStringAndSize(s, l) PyByteArray_FromStringAndSize((const char*)s, l) +#define __Pyx_PyBytes_FromString PyBytes_FromString +#define __Pyx_PyBytes_FromStringAndSize PyBytes_FromStringAndSize +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char*); +#if PY_MAJOR_VERSION < 3 + #define __Pyx_PyStr_FromString __Pyx_PyBytes_FromString + #define __Pyx_PyStr_FromStringAndSize __Pyx_PyBytes_FromStringAndSize +#else + #define __Pyx_PyStr_FromString __Pyx_PyUnicode_FromString + #define __Pyx_PyStr_FromStringAndSize __Pyx_PyUnicode_FromStringAndSize +#endif +#define __Pyx_PyBytes_AsWritableString(s) ((char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsWritableSString(s) ((signed char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsWritableUString(s) ((unsigned char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsString(s) ((const char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsSString(s) ((const signed char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsUString(s) ((const unsigned char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyObject_AsWritableString(s) ((char*)(__pyx_uintptr_t) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsWritableSString(s) ((signed char*)(__pyx_uintptr_t) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsWritableUString(s) ((unsigned char*)(__pyx_uintptr_t) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsSString(s) ((const signed char*) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsUString(s) ((const unsigned char*) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_FromCString(s) __Pyx_PyObject_FromString((const char*)s) +#define __Pyx_PyBytes_FromCString(s) __Pyx_PyBytes_FromString((const char*)s) +#define __Pyx_PyByteArray_FromCString(s) __Pyx_PyByteArray_FromString((const char*)s) +#define __Pyx_PyStr_FromCString(s) __Pyx_PyStr_FromString((const char*)s) +#define __Pyx_PyUnicode_FromCString(s) __Pyx_PyUnicode_FromString((const char*)s) +#define __Pyx_PyUnicode_FromOrdinal(o) PyUnicode_FromOrdinal((int)o) +#define __Pyx_PyUnicode_AsUnicode PyUnicode_AsUnicode +#define __Pyx_NewRef(obj) (Py_INCREF(obj), obj) +#define __Pyx_Owned_Py_None(b) __Pyx_NewRef(Py_None) +static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b); +static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject*); +static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject*); +static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x); +#define __Pyx_PySequence_Tuple(obj)\ + (likely(PyTuple_CheckExact(obj)) ? __Pyx_NewRef(obj) : PySequence_Tuple(obj)) +static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject*); +static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t); +static CYTHON_INLINE Py_hash_t __Pyx_PyIndex_AsHash_t(PyObject*); +#if CYTHON_ASSUME_SAFE_MACROS +#define __pyx_PyFloat_AsDouble(x) (PyFloat_CheckExact(x) ? PyFloat_AS_DOUBLE(x) : PyFloat_AsDouble(x)) +#else +#define __pyx_PyFloat_AsDouble(x) PyFloat_AsDouble(x) +#endif +#define __pyx_PyFloat_AsFloat(x) ((float) __pyx_PyFloat_AsDouble(x)) +#if PY_MAJOR_VERSION >= 3 +#define __Pyx_PyNumber_Int(x) (PyLong_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Long(x)) +#else +#define __Pyx_PyNumber_Int(x) (PyInt_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Int(x)) +#endif +#if CYTHON_USE_PYLONG_INTERNALS + #if PY_VERSION_HEX >= 0x030C00A7 + #ifndef _PyLong_SIGN_MASK + #define _PyLong_SIGN_MASK 3 + #endif + #ifndef _PyLong_NON_SIZE_BITS + #define _PyLong_NON_SIZE_BITS 3 + #endif + #define __Pyx_PyLong_Sign(x) (((PyLongObject*)x)->long_value.lv_tag & _PyLong_SIGN_MASK) + #define __Pyx_PyLong_IsNeg(x) ((__Pyx_PyLong_Sign(x) & 2) != 0) + #define __Pyx_PyLong_IsNonNeg(x) (!__Pyx_PyLong_IsNeg(x)) + #define __Pyx_PyLong_IsZero(x) (__Pyx_PyLong_Sign(x) & 1) + #define __Pyx_PyLong_IsPos(x) (__Pyx_PyLong_Sign(x) == 0) + #define __Pyx_PyLong_CompactValueUnsigned(x) (__Pyx_PyLong_Digits(x)[0]) + #define __Pyx_PyLong_DigitCount(x) ((Py_ssize_t) (((PyLongObject*)x)->long_value.lv_tag >> _PyLong_NON_SIZE_BITS)) + #define __Pyx_PyLong_SignedDigitCount(x)\ + ((1 - (Py_ssize_t) __Pyx_PyLong_Sign(x)) * __Pyx_PyLong_DigitCount(x)) + #if defined(PyUnstable_Long_IsCompact) && defined(PyUnstable_Long_CompactValue) + #define __Pyx_PyLong_IsCompact(x) PyUnstable_Long_IsCompact((PyLongObject*) x) + #define __Pyx_PyLong_CompactValue(x) PyUnstable_Long_CompactValue((PyLongObject*) x) + #else + #define __Pyx_PyLong_IsCompact(x) (((PyLongObject*)x)->long_value.lv_tag < (2 << _PyLong_NON_SIZE_BITS)) + #define __Pyx_PyLong_CompactValue(x) ((1 - (Py_ssize_t) __Pyx_PyLong_Sign(x)) * (Py_ssize_t) __Pyx_PyLong_Digits(x)[0]) + #endif + typedef Py_ssize_t __Pyx_compact_pylong; + typedef size_t __Pyx_compact_upylong; + #else + #define __Pyx_PyLong_IsNeg(x) (Py_SIZE(x) < 0) + #define __Pyx_PyLong_IsNonNeg(x) (Py_SIZE(x) >= 0) + #define __Pyx_PyLong_IsZero(x) (Py_SIZE(x) == 0) + #define __Pyx_PyLong_IsPos(x) (Py_SIZE(x) > 0) + #define __Pyx_PyLong_CompactValueUnsigned(x) ((Py_SIZE(x) == 0) ? 0 : __Pyx_PyLong_Digits(x)[0]) + #define __Pyx_PyLong_DigitCount(x) __Pyx_sst_abs(Py_SIZE(x)) + #define __Pyx_PyLong_SignedDigitCount(x) Py_SIZE(x) + #define __Pyx_PyLong_IsCompact(x) (Py_SIZE(x) == 0 || Py_SIZE(x) == 1 || Py_SIZE(x) == -1) + #define __Pyx_PyLong_CompactValue(x)\ + ((Py_SIZE(x) == 0) ? (sdigit) 0 : ((Py_SIZE(x) < 0) ? -(sdigit)__Pyx_PyLong_Digits(x)[0] : (sdigit)__Pyx_PyLong_Digits(x)[0])) + typedef sdigit __Pyx_compact_pylong; + typedef digit __Pyx_compact_upylong; + #endif + #if PY_VERSION_HEX >= 0x030C00A5 + #define __Pyx_PyLong_Digits(x) (((PyLongObject*)x)->long_value.ob_digit) + #else + #define __Pyx_PyLong_Digits(x) (((PyLongObject*)x)->ob_digit) + #endif +#endif +#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII +#include +static int __Pyx_sys_getdefaultencoding_not_ascii; +static int __Pyx_init_sys_getdefaultencoding_params(void) { + PyObject* sys; + PyObject* default_encoding = NULL; + PyObject* ascii_chars_u = NULL; + PyObject* ascii_chars_b = NULL; + const char* default_encoding_c; + sys = PyImport_ImportModule("sys"); + if (!sys) goto bad; + default_encoding = PyObject_CallMethod(sys, (char*) "getdefaultencoding", NULL); + Py_DECREF(sys); + if (!default_encoding) goto bad; + default_encoding_c = PyBytes_AsString(default_encoding); + if (!default_encoding_c) goto bad; + if (strcmp(default_encoding_c, "ascii") == 0) { + __Pyx_sys_getdefaultencoding_not_ascii = 0; + } else { + char ascii_chars[128]; + int c; + for (c = 0; c < 128; c++) { + ascii_chars[c] = (char) c; + } + __Pyx_sys_getdefaultencoding_not_ascii = 1; + ascii_chars_u = PyUnicode_DecodeASCII(ascii_chars, 128, NULL); + if (!ascii_chars_u) goto bad; + ascii_chars_b = PyUnicode_AsEncodedString(ascii_chars_u, default_encoding_c, NULL); + if (!ascii_chars_b || !PyBytes_Check(ascii_chars_b) || memcmp(ascii_chars, PyBytes_AS_STRING(ascii_chars_b), 128) != 0) { + PyErr_Format( + PyExc_ValueError, + "This module compiled with c_string_encoding=ascii, but default encoding '%.200s' is not a superset of ascii.", + default_encoding_c); + goto bad; + } + Py_DECREF(ascii_chars_u); + Py_DECREF(ascii_chars_b); + } + Py_DECREF(default_encoding); + return 0; +bad: + Py_XDECREF(default_encoding); + Py_XDECREF(ascii_chars_u); + Py_XDECREF(ascii_chars_b); + return -1; +} +#endif +#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT && PY_MAJOR_VERSION >= 3 +#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_DecodeUTF8(c_str, size, NULL) +#else +#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_Decode(c_str, size, __PYX_DEFAULT_STRING_ENCODING, NULL) +#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT +#include +static char* __PYX_DEFAULT_STRING_ENCODING; +static int __Pyx_init_sys_getdefaultencoding_params(void) { + PyObject* sys; + PyObject* default_encoding = NULL; + char* default_encoding_c; + sys = PyImport_ImportModule("sys"); + if (!sys) goto bad; + default_encoding = PyObject_CallMethod(sys, (char*) (const char*) "getdefaultencoding", NULL); + Py_DECREF(sys); + if (!default_encoding) goto bad; + default_encoding_c = PyBytes_AsString(default_encoding); + if (!default_encoding_c) goto bad; + __PYX_DEFAULT_STRING_ENCODING = (char*) malloc(strlen(default_encoding_c) + 1); + if (!__PYX_DEFAULT_STRING_ENCODING) goto bad; + strcpy(__PYX_DEFAULT_STRING_ENCODING, default_encoding_c); + Py_DECREF(default_encoding); + return 0; +bad: + Py_XDECREF(default_encoding); + return -1; +} +#endif +#endif + + +/* Test for GCC > 2.95 */ +#if defined(__GNUC__) && (__GNUC__ > 2 || (__GNUC__ == 2 && (__GNUC_MINOR__ > 95))) + #define likely(x) __builtin_expect(!!(x), 1) + #define unlikely(x) __builtin_expect(!!(x), 0) +#else /* !__GNUC__ or GCC < 2.95 */ + #define likely(x) (x) + #define unlikely(x) (x) +#endif /* __GNUC__ */ +static CYTHON_INLINE void __Pyx_pretend_to_initialize(void* ptr) { (void)ptr; } + +#if !CYTHON_USE_MODULE_STATE +static PyObject *__pyx_m = NULL; +#endif +static int __pyx_lineno; +static int __pyx_clineno = 0; +static const char * __pyx_cfilenm = __FILE__; +static const char *__pyx_filename; + +/* Header.proto */ +#if !defined(CYTHON_CCOMPLEX) + #if defined(__cplusplus) + #define CYTHON_CCOMPLEX 1 + #elif (defined(_Complex_I) && !defined(_MSC_VER)) || ((defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) && !defined(__STDC_NO_COMPLEX__) && !defined(_MSC_VER)) + #define CYTHON_CCOMPLEX 1 + #else + #define CYTHON_CCOMPLEX 0 + #endif +#endif +#if CYTHON_CCOMPLEX + #ifdef __cplusplus + #include + #else + #include + #endif +#endif +#if CYTHON_CCOMPLEX && !defined(__cplusplus) && defined(__sun__) && defined(__GNUC__) + #undef _Complex_I + #define _Complex_I 1.0fj +#endif + +/* #### Code section: filename_table ### */ + +static const char *__pyx_f[] = { + "fairseq/data/data_utils_fast.pyx", + "", + "__init__.cython-30.pxd", + "type.pxd", +}; +/* #### Code section: utility_code_proto_before_types ### */ +/* ForceInitThreads.proto */ +#ifndef __PYX_FORCE_INIT_THREADS + #define __PYX_FORCE_INIT_THREADS 0 +#endif + +/* NoFastGil.proto */ +#define __Pyx_PyGILState_Ensure PyGILState_Ensure +#define __Pyx_PyGILState_Release PyGILState_Release +#define __Pyx_FastGIL_Remember() +#define __Pyx_FastGIL_Forget() +#define __Pyx_FastGilFuncInit() + +/* BufferFormatStructs.proto */ +struct __Pyx_StructField_; +#define __PYX_BUF_FLAGS_PACKED_STRUCT (1 << 0) +typedef struct { + const char* name; + struct __Pyx_StructField_* fields; + size_t size; + size_t arraysize[8]; + int ndim; + char typegroup; + char is_unsigned; + int flags; +} __Pyx_TypeInfo; +typedef struct __Pyx_StructField_ { + __Pyx_TypeInfo* type; + const char* name; + size_t offset; +} __Pyx_StructField; +typedef struct { + __Pyx_StructField* field; + size_t parent_offset; +} __Pyx_BufFmt_StackElem; +typedef struct { + __Pyx_StructField root; + __Pyx_BufFmt_StackElem* head; + size_t fmt_offset; + size_t new_count, enc_count; + size_t struct_alignment; + int is_complex; + char enc_type; + char new_packmode; + char enc_packmode; + char is_valid_array; +} __Pyx_BufFmt_Context; + +/* Atomics.proto */ +#include +#ifndef CYTHON_ATOMICS + #define CYTHON_ATOMICS 1 +#endif +#define __PYX_CYTHON_ATOMICS_ENABLED() CYTHON_ATOMICS +#define __pyx_atomic_int_type int +#define __pyx_nonatomic_int_type int +#if CYTHON_ATOMICS && (defined(__STDC_VERSION__) &&\ + (__STDC_VERSION__ >= 201112L) &&\ + !defined(__STDC_NO_ATOMICS__)) + #include +#elif CYTHON_ATOMICS && (defined(__cplusplus) && (\ + (__cplusplus >= 201103L) ||\ + (defined(_MSC_VER) && _MSC_VER >= 1700))) + #include +#endif +#if CYTHON_ATOMICS && (defined(__STDC_VERSION__) &&\ + (__STDC_VERSION__ >= 201112L) &&\ + !defined(__STDC_NO_ATOMICS__) &&\ + ATOMIC_INT_LOCK_FREE == 2) + #undef __pyx_atomic_int_type + #define __pyx_atomic_int_type atomic_int + #define __pyx_atomic_incr_aligned(value) atomic_fetch_add_explicit(value, 1, memory_order_relaxed) + #define __pyx_atomic_decr_aligned(value) atomic_fetch_sub_explicit(value, 1, memory_order_acq_rel) + #if defined(__PYX_DEBUG_ATOMICS) && defined(_MSC_VER) + #pragma message ("Using standard C atomics") + #elif defined(__PYX_DEBUG_ATOMICS) + #warning "Using standard C atomics" + #endif +#elif CYTHON_ATOMICS && (defined(__cplusplus) && (\ + (__cplusplus >= 201103L) ||\ +\ + (defined(_MSC_VER) && _MSC_VER >= 1700)) &&\ + ATOMIC_INT_LOCK_FREE == 2) + #undef __pyx_atomic_int_type + #define __pyx_atomic_int_type std::atomic_int + #define __pyx_atomic_incr_aligned(value) std::atomic_fetch_add_explicit(value, 1, std::memory_order_relaxed) + #define __pyx_atomic_decr_aligned(value) std::atomic_fetch_sub_explicit(value, 1, std::memory_order_acq_rel) + #if defined(__PYX_DEBUG_ATOMICS) && defined(_MSC_VER) + #pragma message ("Using standard C++ atomics") + #elif defined(__PYX_DEBUG_ATOMICS) + #warning "Using standard C++ atomics" + #endif +#elif CYTHON_ATOMICS && (__GNUC__ >= 5 || (__GNUC__ == 4 &&\ + (__GNUC_MINOR__ > 1 ||\ + (__GNUC_MINOR__ == 1 && __GNUC_PATCHLEVEL__ >= 2)))) + #define __pyx_atomic_incr_aligned(value) __sync_fetch_and_add(value, 1) + #define __pyx_atomic_decr_aligned(value) __sync_fetch_and_sub(value, 1) + #ifdef __PYX_DEBUG_ATOMICS + #warning "Using GNU atomics" + #endif +#elif CYTHON_ATOMICS && defined(_MSC_VER) + #include + #undef __pyx_atomic_int_type + #define __pyx_atomic_int_type long + #undef __pyx_nonatomic_int_type + #define __pyx_nonatomic_int_type long + #pragma intrinsic (_InterlockedExchangeAdd) + #define __pyx_atomic_incr_aligned(value) _InterlockedExchangeAdd(value, 1) + #define __pyx_atomic_decr_aligned(value) _InterlockedExchangeAdd(value, -1) + #ifdef __PYX_DEBUG_ATOMICS + #pragma message ("Using MSVC atomics") + #endif +#else + #undef CYTHON_ATOMICS + #define CYTHON_ATOMICS 0 + #ifdef __PYX_DEBUG_ATOMICS + #warning "Not using atomics" + #endif +#endif +#if CYTHON_ATOMICS + #define __pyx_add_acquisition_count(memview)\ + __pyx_atomic_incr_aligned(__pyx_get_slice_count_pointer(memview)) + #define __pyx_sub_acquisition_count(memview)\ + __pyx_atomic_decr_aligned(__pyx_get_slice_count_pointer(memview)) +#else + #define __pyx_add_acquisition_count(memview)\ + __pyx_add_acquisition_count_locked(__pyx_get_slice_count_pointer(memview), memview->lock) + #define __pyx_sub_acquisition_count(memview)\ + __pyx_sub_acquisition_count_locked(__pyx_get_slice_count_pointer(memview), memview->lock) +#endif + +/* MemviewSliceStruct.proto */ +struct __pyx_memoryview_obj; +typedef struct { + struct __pyx_memoryview_obj *memview; + char *data; + Py_ssize_t shape[8]; + Py_ssize_t strides[8]; + Py_ssize_t suboffsets[8]; +} __Pyx_memviewslice; +#define __Pyx_MemoryView_Len(m) (m.shape[0]) + +/* #### Code section: numeric_typedefs ### */ + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":787 + * # in Cython to enable them only on the right systems. + * + * ctypedef npy_int8 int8_t # <<<<<<<<<<<<<< + * ctypedef npy_int16 int16_t + * ctypedef npy_int32 int32_t + */ +typedef npy_int8 __pyx_t_5numpy_int8_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":788 + * + * ctypedef npy_int8 int8_t + * ctypedef npy_int16 int16_t # <<<<<<<<<<<<<< + * ctypedef npy_int32 int32_t + * ctypedef npy_int64 int64_t + */ +typedef npy_int16 __pyx_t_5numpy_int16_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":789 + * ctypedef npy_int8 int8_t + * ctypedef npy_int16 int16_t + * ctypedef npy_int32 int32_t # <<<<<<<<<<<<<< + * ctypedef npy_int64 int64_t + * #ctypedef npy_int96 int96_t + */ +typedef npy_int32 __pyx_t_5numpy_int32_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":790 + * ctypedef npy_int16 int16_t + * ctypedef npy_int32 int32_t + * ctypedef npy_int64 int64_t # <<<<<<<<<<<<<< + * #ctypedef npy_int96 int96_t + * #ctypedef npy_int128 int128_t + */ +typedef npy_int64 __pyx_t_5numpy_int64_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":794 + * #ctypedef npy_int128 int128_t + * + * ctypedef npy_uint8 uint8_t # <<<<<<<<<<<<<< + * ctypedef npy_uint16 uint16_t + * ctypedef npy_uint32 uint32_t + */ +typedef npy_uint8 __pyx_t_5numpy_uint8_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":795 + * + * ctypedef npy_uint8 uint8_t + * ctypedef npy_uint16 uint16_t # <<<<<<<<<<<<<< + * ctypedef npy_uint32 uint32_t + * ctypedef npy_uint64 uint64_t + */ +typedef npy_uint16 __pyx_t_5numpy_uint16_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":796 + * ctypedef npy_uint8 uint8_t + * ctypedef npy_uint16 uint16_t + * ctypedef npy_uint32 uint32_t # <<<<<<<<<<<<<< + * ctypedef npy_uint64 uint64_t + * #ctypedef npy_uint96 uint96_t + */ +typedef npy_uint32 __pyx_t_5numpy_uint32_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":797 + * ctypedef npy_uint16 uint16_t + * ctypedef npy_uint32 uint32_t + * ctypedef npy_uint64 uint64_t # <<<<<<<<<<<<<< + * #ctypedef npy_uint96 uint96_t + * #ctypedef npy_uint128 uint128_t + */ +typedef npy_uint64 __pyx_t_5numpy_uint64_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":801 + * #ctypedef npy_uint128 uint128_t + * + * ctypedef npy_float32 float32_t # <<<<<<<<<<<<<< + * ctypedef npy_float64 float64_t + * #ctypedef npy_float80 float80_t + */ +typedef npy_float32 __pyx_t_5numpy_float32_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":802 + * + * ctypedef npy_float32 float32_t + * ctypedef npy_float64 float64_t # <<<<<<<<<<<<<< + * #ctypedef npy_float80 float80_t + * #ctypedef npy_float128 float128_t + */ +typedef npy_float64 __pyx_t_5numpy_float64_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":809 + * ctypedef double complex complex128_t + * + * ctypedef npy_longlong longlong_t # <<<<<<<<<<<<<< + * ctypedef npy_ulonglong ulonglong_t + * + */ +typedef npy_longlong __pyx_t_5numpy_longlong_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":810 + * + * ctypedef npy_longlong longlong_t + * ctypedef npy_ulonglong ulonglong_t # <<<<<<<<<<<<<< + * + * ctypedef npy_intp intp_t + */ +typedef npy_ulonglong __pyx_t_5numpy_ulonglong_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":812 + * ctypedef npy_ulonglong ulonglong_t + * + * ctypedef npy_intp intp_t # <<<<<<<<<<<<<< + * ctypedef npy_uintp uintp_t + * + */ +typedef npy_intp __pyx_t_5numpy_intp_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":813 + * + * ctypedef npy_intp intp_t + * ctypedef npy_uintp uintp_t # <<<<<<<<<<<<<< + * + * ctypedef npy_double float_t + */ +typedef npy_uintp __pyx_t_5numpy_uintp_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":815 + * ctypedef npy_uintp uintp_t + * + * ctypedef npy_double float_t # <<<<<<<<<<<<<< + * ctypedef npy_double double_t + * ctypedef npy_longdouble longdouble_t + */ +typedef npy_double __pyx_t_5numpy_float_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":816 + * + * ctypedef npy_double float_t + * ctypedef npy_double double_t # <<<<<<<<<<<<<< + * ctypedef npy_longdouble longdouble_t + * + */ +typedef npy_double __pyx_t_5numpy_double_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":817 + * ctypedef npy_double float_t + * ctypedef npy_double double_t + * ctypedef npy_longdouble longdouble_t # <<<<<<<<<<<<<< + * + * ctypedef float complex cfloat_t + */ +typedef npy_longdouble __pyx_t_5numpy_longdouble_t; + +/* "fairseq/data/data_utils_fast.pyx":13 + * + * DTYPE = np.int64 + * ctypedef np.int64_t DTYPE_t # <<<<<<<<<<<<<< + * + * + */ +typedef __pyx_t_5numpy_int64_t __pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t; +/* #### Code section: complex_type_declarations ### */ +/* Declarations.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + typedef ::std::complex< float > __pyx_t_float_complex; + #else + typedef float _Complex __pyx_t_float_complex; + #endif +#else + typedef struct { float real, imag; } __pyx_t_float_complex; +#endif +static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float, float); + +/* Declarations.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + typedef ::std::complex< double > __pyx_t_double_complex; + #else + typedef double _Complex __pyx_t_double_complex; + #endif +#else + typedef struct { double real, imag; } __pyx_t_double_complex; +#endif +static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double, double); + +/* Declarations.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + typedef ::std::complex< long double > __pyx_t_long_double_complex; + #else + typedef long double _Complex __pyx_t_long_double_complex; + #endif +#else + typedef struct { long double real, imag; } __pyx_t_long_double_complex; +#endif +static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double, long double); + +/* #### Code section: type_declarations ### */ + +/*--- Type declarations ---*/ +struct __pyx_array_obj; +struct __pyx_MemviewEnum_obj; +struct __pyx_memoryview_obj; +struct __pyx_memoryviewslice_obj; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1113 + * + * # Iterator API added in v1.6 + * ctypedef int (*NpyIter_IterNextFunc)(NpyIter* it) noexcept nogil # <<<<<<<<<<<<<< + * ctypedef void (*NpyIter_GetMultiIndexFunc)(NpyIter* it, npy_intp* outcoords) noexcept nogil + * + */ +typedef int (*__pyx_t_5numpy_NpyIter_IterNextFunc)(NpyIter *); + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1114 + * # Iterator API added in v1.6 + * ctypedef int (*NpyIter_IterNextFunc)(NpyIter* it) noexcept nogil + * ctypedef void (*NpyIter_GetMultiIndexFunc)(NpyIter* it, npy_intp* outcoords) noexcept nogil # <<<<<<<<<<<<<< + * + * cdef extern from "numpy/arrayobject.h": + */ +typedef void (*__pyx_t_5numpy_NpyIter_GetMultiIndexFunc)(NpyIter *, npy_intp *); + +/* "View.MemoryView":114 + * @cython.collection_type("sequence") + * @cname("__pyx_array") + * cdef class array: # <<<<<<<<<<<<<< + * + * cdef: + */ +struct __pyx_array_obj { + PyObject_HEAD + struct __pyx_vtabstruct_array *__pyx_vtab; + char *data; + Py_ssize_t len; + char *format; + int ndim; + Py_ssize_t *_shape; + Py_ssize_t *_strides; + Py_ssize_t itemsize; + PyObject *mode; + PyObject *_format; + void (*callback_free_data)(void *); + int free_data; + int dtype_is_object; +}; + + +/* "View.MemoryView":302 + * + * @cname('__pyx_MemviewEnum') + * cdef class Enum(object): # <<<<<<<<<<<<<< + * cdef object name + * def __init__(self, name): + */ +struct __pyx_MemviewEnum_obj { + PyObject_HEAD + PyObject *name; +}; + + +/* "View.MemoryView":337 + * + * @cname('__pyx_memoryview') + * cdef class memoryview: # <<<<<<<<<<<<<< + * + * cdef object obj + */ +struct __pyx_memoryview_obj { + PyObject_HEAD + struct __pyx_vtabstruct_memoryview *__pyx_vtab; + PyObject *obj; + PyObject *_size; + PyObject *_array_interface; + PyThread_type_lock lock; + __pyx_atomic_int_type acquisition_count; + Py_buffer view; + int flags; + int dtype_is_object; + __Pyx_TypeInfo *typeinfo; +}; + + +/* "View.MemoryView":952 + * @cython.collection_type("sequence") + * @cname('__pyx_memoryviewslice') + * cdef class _memoryviewslice(memoryview): # <<<<<<<<<<<<<< + * "Internal class for passing memoryview slices to Python" + * + */ +struct __pyx_memoryviewslice_obj { + struct __pyx_memoryview_obj __pyx_base; + __Pyx_memviewslice from_slice; + PyObject *from_object; + PyObject *(*to_object_func)(char *); + int (*to_dtype_func)(char *, PyObject *); +}; + + + +/* "View.MemoryView":114 + * @cython.collection_type("sequence") + * @cname("__pyx_array") + * cdef class array: # <<<<<<<<<<<<<< + * + * cdef: + */ + +struct __pyx_vtabstruct_array { + PyObject *(*get_memview)(struct __pyx_array_obj *); +}; +static struct __pyx_vtabstruct_array *__pyx_vtabptr_array; + + +/* "View.MemoryView":337 + * + * @cname('__pyx_memoryview') + * cdef class memoryview: # <<<<<<<<<<<<<< + * + * cdef object obj + */ + +struct __pyx_vtabstruct_memoryview { + char *(*get_item_pointer)(struct __pyx_memoryview_obj *, PyObject *); + PyObject *(*is_slice)(struct __pyx_memoryview_obj *, PyObject *); + PyObject *(*setitem_slice_assignment)(struct __pyx_memoryview_obj *, PyObject *, PyObject *); + PyObject *(*setitem_slice_assign_scalar)(struct __pyx_memoryview_obj *, struct __pyx_memoryview_obj *, PyObject *); + PyObject *(*setitem_indexed)(struct __pyx_memoryview_obj *, PyObject *, PyObject *); + PyObject *(*convert_item_to_object)(struct __pyx_memoryview_obj *, char *); + PyObject *(*assign_item_from_object)(struct __pyx_memoryview_obj *, char *, PyObject *); + PyObject *(*_get_base)(struct __pyx_memoryview_obj *); +}; +static struct __pyx_vtabstruct_memoryview *__pyx_vtabptr_memoryview; + + +/* "View.MemoryView":952 + * @cython.collection_type("sequence") + * @cname('__pyx_memoryviewslice') + * cdef class _memoryviewslice(memoryview): # <<<<<<<<<<<<<< + * "Internal class for passing memoryview slices to Python" + * + */ + +struct __pyx_vtabstruct__memoryviewslice { + struct __pyx_vtabstruct_memoryview __pyx_base; +}; +static struct __pyx_vtabstruct__memoryviewslice *__pyx_vtabptr__memoryviewslice; +/* #### Code section: utility_code_proto ### */ + +/* --- Runtime support code (head) --- */ +/* Refnanny.proto */ +#ifndef CYTHON_REFNANNY + #define CYTHON_REFNANNY 0 +#endif +#if CYTHON_REFNANNY + typedef struct { + void (*INCREF)(void*, PyObject*, Py_ssize_t); + void (*DECREF)(void*, PyObject*, Py_ssize_t); + void (*GOTREF)(void*, PyObject*, Py_ssize_t); + void (*GIVEREF)(void*, PyObject*, Py_ssize_t); + void* (*SetupContext)(const char*, Py_ssize_t, const char*); + void (*FinishContext)(void**); + } __Pyx_RefNannyAPIStruct; + static __Pyx_RefNannyAPIStruct *__Pyx_RefNanny = NULL; + static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname); + #define __Pyx_RefNannyDeclarations void *__pyx_refnanny = NULL; +#ifdef WITH_THREAD + #define __Pyx_RefNannySetupContext(name, acquire_gil)\ + if (acquire_gil) {\ + PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\ + __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), (__LINE__), (__FILE__));\ + PyGILState_Release(__pyx_gilstate_save);\ + } else {\ + __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), (__LINE__), (__FILE__));\ + } + #define __Pyx_RefNannyFinishContextNogil() {\ + PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\ + __Pyx_RefNannyFinishContext();\ + PyGILState_Release(__pyx_gilstate_save);\ + } +#else + #define __Pyx_RefNannySetupContext(name, acquire_gil)\ + __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), (__LINE__), (__FILE__)) + #define __Pyx_RefNannyFinishContextNogil() __Pyx_RefNannyFinishContext() +#endif + #define __Pyx_RefNannyFinishContextNogil() {\ + PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\ + __Pyx_RefNannyFinishContext();\ + PyGILState_Release(__pyx_gilstate_save);\ + } + #define __Pyx_RefNannyFinishContext()\ + __Pyx_RefNanny->FinishContext(&__pyx_refnanny) + #define __Pyx_INCREF(r) __Pyx_RefNanny->INCREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_DECREF(r) __Pyx_RefNanny->DECREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_GOTREF(r) __Pyx_RefNanny->GOTREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_GIVEREF(r) __Pyx_RefNanny->GIVEREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_XINCREF(r) do { if((r) == NULL); else {__Pyx_INCREF(r); }} while(0) + #define __Pyx_XDECREF(r) do { if((r) == NULL); else {__Pyx_DECREF(r); }} while(0) + #define __Pyx_XGOTREF(r) do { if((r) == NULL); else {__Pyx_GOTREF(r); }} while(0) + #define __Pyx_XGIVEREF(r) do { if((r) == NULL); else {__Pyx_GIVEREF(r);}} while(0) +#else + #define __Pyx_RefNannyDeclarations + #define __Pyx_RefNannySetupContext(name, acquire_gil) + #define __Pyx_RefNannyFinishContextNogil() + #define __Pyx_RefNannyFinishContext() + #define __Pyx_INCREF(r) Py_INCREF(r) + #define __Pyx_DECREF(r) Py_DECREF(r) + #define __Pyx_GOTREF(r) + #define __Pyx_GIVEREF(r) + #define __Pyx_XINCREF(r) Py_XINCREF(r) + #define __Pyx_XDECREF(r) Py_XDECREF(r) + #define __Pyx_XGOTREF(r) + #define __Pyx_XGIVEREF(r) +#endif +#define __Pyx_Py_XDECREF_SET(r, v) do {\ + PyObject *tmp = (PyObject *) r;\ + r = v; Py_XDECREF(tmp);\ + } while (0) +#define __Pyx_XDECREF_SET(r, v) do {\ + PyObject *tmp = (PyObject *) r;\ + r = v; __Pyx_XDECREF(tmp);\ + } while (0) +#define __Pyx_DECREF_SET(r, v) do {\ + PyObject *tmp = (PyObject *) r;\ + r = v; __Pyx_DECREF(tmp);\ + } while (0) +#define __Pyx_CLEAR(r) do { PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);} while(0) +#define __Pyx_XCLEAR(r) do { if((r) != NULL) {PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);}} while(0) + +/* PyErrExceptionMatches.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_PyErr_ExceptionMatches(err) __Pyx_PyErr_ExceptionMatchesInState(__pyx_tstate, err) +static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err); +#else +#define __Pyx_PyErr_ExceptionMatches(err) PyErr_ExceptionMatches(err) +#endif + +/* PyThreadStateGet.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_PyThreadState_declare PyThreadState *__pyx_tstate; +#define __Pyx_PyThreadState_assign __pyx_tstate = __Pyx_PyThreadState_Current; +#if PY_VERSION_HEX >= 0x030C00A6 +#define __Pyx_PyErr_Occurred() (__pyx_tstate->current_exception != NULL) +#define __Pyx_PyErr_CurrentExceptionType() (__pyx_tstate->current_exception ? (PyObject*) Py_TYPE(__pyx_tstate->current_exception) : (PyObject*) NULL) +#else +#define __Pyx_PyErr_Occurred() (__pyx_tstate->curexc_type != NULL) +#define __Pyx_PyErr_CurrentExceptionType() (__pyx_tstate->curexc_type) +#endif +#else +#define __Pyx_PyThreadState_declare +#define __Pyx_PyThreadState_assign +#define __Pyx_PyErr_Occurred() (PyErr_Occurred() != NULL) +#define __Pyx_PyErr_CurrentExceptionType() PyErr_Occurred() +#endif + +/* PyErrFetchRestore.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_PyErr_Clear() __Pyx_ErrRestore(NULL, NULL, NULL) +#define __Pyx_ErrRestoreWithState(type, value, tb) __Pyx_ErrRestoreInState(PyThreadState_GET(), type, value, tb) +#define __Pyx_ErrFetchWithState(type, value, tb) __Pyx_ErrFetchInState(PyThreadState_GET(), type, value, tb) +#define __Pyx_ErrRestore(type, value, tb) __Pyx_ErrRestoreInState(__pyx_tstate, type, value, tb) +#define __Pyx_ErrFetch(type, value, tb) __Pyx_ErrFetchInState(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb); +static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A6 +#define __Pyx_PyErr_SetNone(exc) (Py_INCREF(exc), __Pyx_ErrRestore((exc), NULL, NULL)) +#else +#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc) +#endif +#else +#define __Pyx_PyErr_Clear() PyErr_Clear() +#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc) +#define __Pyx_ErrRestoreWithState(type, value, tb) PyErr_Restore(type, value, tb) +#define __Pyx_ErrFetchWithState(type, value, tb) PyErr_Fetch(type, value, tb) +#define __Pyx_ErrRestoreInState(tstate, type, value, tb) PyErr_Restore(type, value, tb) +#define __Pyx_ErrFetchInState(tstate, type, value, tb) PyErr_Fetch(type, value, tb) +#define __Pyx_ErrRestore(type, value, tb) PyErr_Restore(type, value, tb) +#define __Pyx_ErrFetch(type, value, tb) PyErr_Fetch(type, value, tb) +#endif + +/* PyObjectGetAttrStr.proto */ +#if CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name); +#else +#define __Pyx_PyObject_GetAttrStr(o,n) PyObject_GetAttr(o,n) +#endif + +/* PyObjectGetAttrStrNoError.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStrNoError(PyObject* obj, PyObject* attr_name); + +/* GetBuiltinName.proto */ +static PyObject *__Pyx_GetBuiltinName(PyObject *name); + +/* TupleAndListFromArray.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyList_FromArray(PyObject *const *src, Py_ssize_t n); +static CYTHON_INLINE PyObject* __Pyx_PyTuple_FromArray(PyObject *const *src, Py_ssize_t n); +#endif + +/* IncludeStringH.proto */ +#include + +/* BytesEquals.proto */ +static CYTHON_INLINE int __Pyx_PyBytes_Equals(PyObject* s1, PyObject* s2, int equals); + +/* UnicodeEquals.proto */ +static CYTHON_INLINE int __Pyx_PyUnicode_Equals(PyObject* s1, PyObject* s2, int equals); + +/* fastcall.proto */ +#if CYTHON_AVOID_BORROWED_REFS + #define __Pyx_Arg_VARARGS(args, i) PySequence_GetItem(args, i) +#elif CYTHON_ASSUME_SAFE_MACROS + #define __Pyx_Arg_VARARGS(args, i) PyTuple_GET_ITEM(args, i) +#else + #define __Pyx_Arg_VARARGS(args, i) PyTuple_GetItem(args, i) +#endif +#if CYTHON_AVOID_BORROWED_REFS + #define __Pyx_Arg_NewRef_VARARGS(arg) __Pyx_NewRef(arg) + #define __Pyx_Arg_XDECREF_VARARGS(arg) Py_XDECREF(arg) +#else + #define __Pyx_Arg_NewRef_VARARGS(arg) arg + #define __Pyx_Arg_XDECREF_VARARGS(arg) +#endif +#define __Pyx_NumKwargs_VARARGS(kwds) PyDict_Size(kwds) +#define __Pyx_KwValues_VARARGS(args, nargs) NULL +#define __Pyx_GetKwValue_VARARGS(kw, kwvalues, s) __Pyx_PyDict_GetItemStrWithError(kw, s) +#define __Pyx_KwargsAsDict_VARARGS(kw, kwvalues) PyDict_Copy(kw) +#if CYTHON_METH_FASTCALL + #define __Pyx_Arg_FASTCALL(args, i) args[i] + #define __Pyx_NumKwargs_FASTCALL(kwds) PyTuple_GET_SIZE(kwds) + #define __Pyx_KwValues_FASTCALL(args, nargs) ((args) + (nargs)) + static CYTHON_INLINE PyObject * __Pyx_GetKwValue_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues, PyObject *s); +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 + CYTHON_UNUSED static PyObject *__Pyx_KwargsAsDict_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues); + #else + #define __Pyx_KwargsAsDict_FASTCALL(kw, kwvalues) _PyStack_AsDict(kwvalues, kw) + #endif + #define __Pyx_Arg_NewRef_FASTCALL(arg) arg /* no-op, __Pyx_Arg_FASTCALL is direct and this needs + to have the same reference counting */ + #define __Pyx_Arg_XDECREF_FASTCALL(arg) +#else + #define __Pyx_Arg_FASTCALL __Pyx_Arg_VARARGS + #define __Pyx_NumKwargs_FASTCALL __Pyx_NumKwargs_VARARGS + #define __Pyx_KwValues_FASTCALL __Pyx_KwValues_VARARGS + #define __Pyx_GetKwValue_FASTCALL __Pyx_GetKwValue_VARARGS + #define __Pyx_KwargsAsDict_FASTCALL __Pyx_KwargsAsDict_VARARGS + #define __Pyx_Arg_NewRef_FASTCALL(arg) __Pyx_Arg_NewRef_VARARGS(arg) + #define __Pyx_Arg_XDECREF_FASTCALL(arg) __Pyx_Arg_XDECREF_VARARGS(arg) +#endif +#if CYTHON_COMPILING_IN_CPYTHON && CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS +#define __Pyx_ArgsSlice_VARARGS(args, start, stop) __Pyx_PyTuple_FromArray(&__Pyx_Arg_VARARGS(args, start), stop - start) +#define __Pyx_ArgsSlice_FASTCALL(args, start, stop) __Pyx_PyTuple_FromArray(&__Pyx_Arg_FASTCALL(args, start), stop - start) +#else +#define __Pyx_ArgsSlice_VARARGS(args, start, stop) PyTuple_GetSlice(args, start, stop) +#define __Pyx_ArgsSlice_FASTCALL(args, start, stop) PyTuple_GetSlice(args, start, stop) +#endif + +/* RaiseArgTupleInvalid.proto */ +static void __Pyx_RaiseArgtupleInvalid(const char* func_name, int exact, + Py_ssize_t num_min, Py_ssize_t num_max, Py_ssize_t num_found); + +/* RaiseDoubleKeywords.proto */ +static void __Pyx_RaiseDoubleKeywordsError(const char* func_name, PyObject* kw_name); + +/* ParseKeywords.proto */ +static int __Pyx_ParseOptionalKeywords(PyObject *kwds, PyObject *const *kwvalues, + PyObject **argnames[], + PyObject *kwds2, PyObject *values[], Py_ssize_t num_pos_args, + const char* function_name); + +/* ArgTypeTest.proto */ +#define __Pyx_ArgTypeTest(obj, type, none_allowed, name, exact)\ + ((likely(__Pyx_IS_TYPE(obj, type) | (none_allowed && (obj == Py_None)))) ? 1 :\ + __Pyx__ArgTypeTest(obj, type, name, exact)) +static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact); + +/* RaiseException.proto */ +static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause); + +/* PyFunctionFastCall.proto */ +#if CYTHON_FAST_PYCALL +#if !CYTHON_VECTORCALL +#define __Pyx_PyFunction_FastCall(func, args, nargs)\ + __Pyx_PyFunction_FastCallDict((func), (args), (nargs), NULL) +static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, Py_ssize_t nargs, PyObject *kwargs); +#endif +#define __Pyx_BUILD_ASSERT_EXPR(cond)\ + (sizeof(char [1 - 2*!(cond)]) - 1) +#ifndef Py_MEMBER_SIZE +#define Py_MEMBER_SIZE(type, member) sizeof(((type *)0)->member) +#endif +#if !CYTHON_VECTORCALL +#if PY_VERSION_HEX >= 0x03080000 + #include "frameobject.h" +#if PY_VERSION_HEX >= 0x030b00a6 && !CYTHON_COMPILING_IN_LIMITED_API && !defined(PYPY_VERSION) + #ifndef Py_BUILD_CORE + #define Py_BUILD_CORE 1 + #endif + #include "internal/pycore_frame.h" +#endif + #define __Pxy_PyFrame_Initialize_Offsets() + #define __Pyx_PyFrame_GetLocalsplus(frame) ((frame)->f_localsplus) +#else + static size_t __pyx_pyframe_localsplus_offset = 0; + #include "frameobject.h" + #define __Pxy_PyFrame_Initialize_Offsets()\ + ((void)__Pyx_BUILD_ASSERT_EXPR(sizeof(PyFrameObject) == offsetof(PyFrameObject, f_localsplus) + Py_MEMBER_SIZE(PyFrameObject, f_localsplus)),\ + (void)(__pyx_pyframe_localsplus_offset = ((size_t)PyFrame_Type.tp_basicsize) - Py_MEMBER_SIZE(PyFrameObject, f_localsplus))) + #define __Pyx_PyFrame_GetLocalsplus(frame)\ + (assert(__pyx_pyframe_localsplus_offset), (PyObject **)(((char *)(frame)) + __pyx_pyframe_localsplus_offset)) +#endif +#endif +#endif + +/* PyObjectCall.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw); +#else +#define __Pyx_PyObject_Call(func, arg, kw) PyObject_Call(func, arg, kw) +#endif + +/* PyObjectCallMethO.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg); +#endif + +/* PyObjectFastCall.proto */ +#define __Pyx_PyObject_FastCall(func, args, nargs) __Pyx_PyObject_FastCallDict(func, args, (size_t)(nargs), NULL) +static CYTHON_INLINE PyObject* __Pyx_PyObject_FastCallDict(PyObject *func, PyObject **args, size_t nargs, PyObject *kwargs); + +/* RaiseUnexpectedTypeError.proto */ +static int __Pyx_RaiseUnexpectedTypeError(const char *expected, PyObject *obj); + +/* GCCDiagnostics.proto */ +#if !defined(__INTEL_COMPILER) && defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)) +#define __Pyx_HAS_GCC_DIAGNOSTIC +#endif + +/* BuildPyUnicode.proto */ +static PyObject* __Pyx_PyUnicode_BuildFromAscii(Py_ssize_t ulength, char* chars, int clength, + int prepend_sign, char padding_char); + +/* CIntToPyUnicode.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_int(int value, Py_ssize_t width, char padding_char, char format_char); + +/* CIntToPyUnicode.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_Py_ssize_t(Py_ssize_t value, Py_ssize_t width, char padding_char, char format_char); + +/* JoinPyUnicode.proto */ +static PyObject* __Pyx_PyUnicode_Join(PyObject* value_tuple, Py_ssize_t value_count, Py_ssize_t result_ulength, + Py_UCS4 max_char); + +/* StrEquals.proto */ +#if PY_MAJOR_VERSION >= 3 +#define __Pyx_PyString_Equals __Pyx_PyUnicode_Equals +#else +#define __Pyx_PyString_Equals __Pyx_PyBytes_Equals +#endif + +/* PyObjectFormatSimple.proto */ +#if CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + PyObject_Format(s, f)) +#elif PY_MAJOR_VERSION < 3 + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + likely(PyString_CheckExact(s)) ? PyUnicode_FromEncodedObject(s, NULL, "strict") :\ + PyObject_Format(s, f)) +#elif CYTHON_USE_TYPE_SLOTS + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + likely(PyLong_CheckExact(s)) ? PyLong_Type.tp_repr(s) :\ + likely(PyFloat_CheckExact(s)) ? PyFloat_Type.tp_repr(s) :\ + PyObject_Format(s, f)) +#else + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + PyObject_Format(s, f)) +#endif + +CYTHON_UNUSED static int __pyx_array_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *); /*proto*/ +/* GetAttr.proto */ +static CYTHON_INLINE PyObject *__Pyx_GetAttr(PyObject *, PyObject *); + +/* GetItemInt.proto */ +#define __Pyx_GetItemInt(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_GetItemInt_Fast(o, (Py_ssize_t)i, is_list, wraparound, boundscheck) :\ + (is_list ? (PyErr_SetString(PyExc_IndexError, "list index out of range"), (PyObject*)NULL) :\ + __Pyx_GetItemInt_Generic(o, to_py_func(i)))) +#define __Pyx_GetItemInt_List(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_GetItemInt_List_Fast(o, (Py_ssize_t)i, wraparound, boundscheck) :\ + (PyErr_SetString(PyExc_IndexError, "list index out of range"), (PyObject*)NULL)) +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_List_Fast(PyObject *o, Py_ssize_t i, + int wraparound, int boundscheck); +#define __Pyx_GetItemInt_Tuple(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_GetItemInt_Tuple_Fast(o, (Py_ssize_t)i, wraparound, boundscheck) :\ + (PyErr_SetString(PyExc_IndexError, "tuple index out of range"), (PyObject*)NULL)) +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Tuple_Fast(PyObject *o, Py_ssize_t i, + int wraparound, int boundscheck); +static PyObject *__Pyx_GetItemInt_Generic(PyObject *o, PyObject* j); +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Fast(PyObject *o, Py_ssize_t i, + int is_list, int wraparound, int boundscheck); + +/* PyObjectCallOneArg.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg); + +/* ObjectGetItem.proto */ +#if CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PyObject *__Pyx_PyObject_GetItem(PyObject *obj, PyObject *key); +#else +#define __Pyx_PyObject_GetItem(obj, key) PyObject_GetItem(obj, key) +#endif + +/* KeywordStringCheck.proto */ +static int __Pyx_CheckKeywordStrings(PyObject *kw, const char* function_name, int kw_allowed); + +/* DivInt[Py_ssize_t].proto */ +static CYTHON_INLINE Py_ssize_t __Pyx_div_Py_ssize_t(Py_ssize_t, Py_ssize_t); + +/* UnaryNegOverflows.proto */ +#define __Pyx_UNARY_NEG_WOULD_OVERFLOW(x)\ + (((x) < 0) & ((unsigned long)(x) == 0-(unsigned long)(x))) + +/* GetAttr3.proto */ +static CYTHON_INLINE PyObject *__Pyx_GetAttr3(PyObject *, PyObject *, PyObject *); + +/* PyDictVersioning.proto */ +#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS +#define __PYX_DICT_VERSION_INIT ((PY_UINT64_T) -1) +#define __PYX_GET_DICT_VERSION(dict) (((PyDictObject*)(dict))->ma_version_tag) +#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)\ + (version_var) = __PYX_GET_DICT_VERSION(dict);\ + (cache_var) = (value); +#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP) {\ + static PY_UINT64_T __pyx_dict_version = 0;\ + static PyObject *__pyx_dict_cached_value = NULL;\ + if (likely(__PYX_GET_DICT_VERSION(DICT) == __pyx_dict_version)) {\ + (VAR) = __pyx_dict_cached_value;\ + } else {\ + (VAR) = __pyx_dict_cached_value = (LOOKUP);\ + __pyx_dict_version = __PYX_GET_DICT_VERSION(DICT);\ + }\ +} +static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj); +static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj); +static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version); +#else +#define __PYX_GET_DICT_VERSION(dict) (0) +#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var) +#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP) (VAR) = (LOOKUP); +#endif + +/* GetModuleGlobalName.proto */ +#if CYTHON_USE_DICT_VERSIONS +#define __Pyx_GetModuleGlobalName(var, name) do {\ + static PY_UINT64_T __pyx_dict_version = 0;\ + static PyObject *__pyx_dict_cached_value = NULL;\ + (var) = (likely(__pyx_dict_version == __PYX_GET_DICT_VERSION(__pyx_d))) ?\ + (likely(__pyx_dict_cached_value) ? __Pyx_NewRef(__pyx_dict_cached_value) : __Pyx_GetBuiltinName(name)) :\ + __Pyx__GetModuleGlobalName(name, &__pyx_dict_version, &__pyx_dict_cached_value);\ +} while(0) +#define __Pyx_GetModuleGlobalNameUncached(var, name) do {\ + PY_UINT64_T __pyx_dict_version;\ + PyObject *__pyx_dict_cached_value;\ + (var) = __Pyx__GetModuleGlobalName(name, &__pyx_dict_version, &__pyx_dict_cached_value);\ +} while(0) +static PyObject *__Pyx__GetModuleGlobalName(PyObject *name, PY_UINT64_T *dict_version, PyObject **dict_cached_value); +#else +#define __Pyx_GetModuleGlobalName(var, name) (var) = __Pyx__GetModuleGlobalName(name) +#define __Pyx_GetModuleGlobalNameUncached(var, name) (var) = __Pyx__GetModuleGlobalName(name) +static CYTHON_INLINE PyObject *__Pyx__GetModuleGlobalName(PyObject *name); +#endif + +/* AssertionsEnabled.proto */ +#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX < 0x02070600 && !defined(Py_OptimizeFlag) + #define __Pyx_init_assertions_enabled() (0) + #define __pyx_assertions_enabled() (1) +#elif CYTHON_COMPILING_IN_LIMITED_API || (CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030C0000) + static int __pyx_assertions_enabled_flag; + #define __pyx_assertions_enabled() (__pyx_assertions_enabled_flag) + static int __Pyx_init_assertions_enabled(void) { + PyObject *builtins, *debug, *debug_str; + int flag; + builtins = PyEval_GetBuiltins(); + if (!builtins) goto bad; + debug_str = PyUnicode_FromStringAndSize("__debug__", 9); + if (!debug_str) goto bad; + debug = PyObject_GetItem(builtins, debug_str); + Py_DECREF(debug_str); + if (!debug) goto bad; + flag = PyObject_IsTrue(debug); + Py_DECREF(debug); + if (flag == -1) goto bad; + __pyx_assertions_enabled_flag = flag; + return 0; + bad: + __pyx_assertions_enabled_flag = 1; + return -1; + } +#else + #define __Pyx_init_assertions_enabled() (0) + #define __pyx_assertions_enabled() (!Py_OptimizeFlag) +#endif + +/* RaiseTooManyValuesToUnpack.proto */ +static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected); + +/* RaiseNeedMoreValuesToUnpack.proto */ +static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index); + +/* RaiseNoneIterError.proto */ +static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void); + +/* ExtTypeTest.proto */ +static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type); + +/* GetTopmostException.proto */ +#if CYTHON_USE_EXC_INFO_STACK && CYTHON_FAST_THREAD_STATE +static _PyErr_StackItem * __Pyx_PyErr_GetTopmostException(PyThreadState *tstate); +#endif + +/* SaveResetException.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_ExceptionSave(type, value, tb) __Pyx__ExceptionSave(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#define __Pyx_ExceptionReset(type, value, tb) __Pyx__ExceptionReset(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb); +#else +#define __Pyx_ExceptionSave(type, value, tb) PyErr_GetExcInfo(type, value, tb) +#define __Pyx_ExceptionReset(type, value, tb) PyErr_SetExcInfo(type, value, tb) +#endif + +/* GetException.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_GetException(type, value, tb) __Pyx__GetException(__pyx_tstate, type, value, tb) +static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#else +static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb); +#endif + +/* SwapException.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_ExceptionSwap(type, value, tb) __Pyx__ExceptionSwap(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx__ExceptionSwap(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#else +static CYTHON_INLINE void __Pyx_ExceptionSwap(PyObject **type, PyObject **value, PyObject **tb); +#endif + +/* Import.proto */ +static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level); + +/* ImportDottedModule.proto */ +static PyObject *__Pyx_ImportDottedModule(PyObject *name, PyObject *parts_tuple); +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx_ImportDottedModule_WalkParts(PyObject *module, PyObject *name, PyObject *parts_tuple); +#endif + +/* FastTypeChecks.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +#define __Pyx_TypeCheck(obj, type) __Pyx_IsSubtype(Py_TYPE(obj), (PyTypeObject *)type) +#define __Pyx_TypeCheck2(obj, type1, type2) __Pyx_IsAnySubtype2(Py_TYPE(obj), (PyTypeObject *)type1, (PyTypeObject *)type2) +static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b); +static CYTHON_INLINE int __Pyx_IsAnySubtype2(PyTypeObject *cls, PyTypeObject *a, PyTypeObject *b); +static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject *type); +static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *type1, PyObject *type2); +#else +#define __Pyx_TypeCheck(obj, type) PyObject_TypeCheck(obj, (PyTypeObject *)type) +#define __Pyx_TypeCheck2(obj, type1, type2) (PyObject_TypeCheck(obj, (PyTypeObject *)type1) || PyObject_TypeCheck(obj, (PyTypeObject *)type2)) +#define __Pyx_PyErr_GivenExceptionMatches(err, type) PyErr_GivenExceptionMatches(err, type) +#define __Pyx_PyErr_GivenExceptionMatches2(err, type1, type2) (PyErr_GivenExceptionMatches(err, type1) || PyErr_GivenExceptionMatches(err, type2)) +#endif +#define __Pyx_PyErr_ExceptionMatches2(err1, err2) __Pyx_PyErr_GivenExceptionMatches2(__Pyx_PyErr_CurrentExceptionType(), err1, err2) +#define __Pyx_PyException_Check(obj) __Pyx_TypeCheck(obj, PyExc_Exception) + +CYTHON_UNUSED static int __pyx_memoryview_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +/* ListCompAppend.proto */ +#if CYTHON_USE_PYLIST_INTERNALS && CYTHON_ASSUME_SAFE_MACROS +static CYTHON_INLINE int __Pyx_ListComp_Append(PyObject* list, PyObject* x) { + PyListObject* L = (PyListObject*) list; + Py_ssize_t len = Py_SIZE(list); + if (likely(L->allocated > len)) { + Py_INCREF(x); + #if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 + L->ob_item[len] = x; + #else + PyList_SET_ITEM(list, len, x); + #endif + __Pyx_SET_SIZE(list, len + 1); + return 0; + } + return PyList_Append(list, x); +} +#else +#define __Pyx_ListComp_Append(L,x) PyList_Append(L,x) +#endif + +/* PySequenceMultiply.proto */ +#define __Pyx_PySequence_Multiply_Left(mul, seq) __Pyx_PySequence_Multiply(seq, mul) +static CYTHON_INLINE PyObject* __Pyx_PySequence_Multiply(PyObject *seq, Py_ssize_t mul); + +/* SetItemInt.proto */ +#define __Pyx_SetItemInt(o, i, v, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_SetItemInt_Fast(o, (Py_ssize_t)i, v, is_list, wraparound, boundscheck) :\ + (is_list ? (PyErr_SetString(PyExc_IndexError, "list assignment index out of range"), -1) :\ + __Pyx_SetItemInt_Generic(o, to_py_func(i), v))) +static int __Pyx_SetItemInt_Generic(PyObject *o, PyObject *j, PyObject *v); +static CYTHON_INLINE int __Pyx_SetItemInt_Fast(PyObject *o, Py_ssize_t i, PyObject *v, + int is_list, int wraparound, int boundscheck); + +/* RaiseUnboundLocalError.proto */ +static CYTHON_INLINE void __Pyx_RaiseUnboundLocalError(const char *varname); + +/* DivInt[long].proto */ +static CYTHON_INLINE long __Pyx_div_long(long, long); + +/* PySequenceContains.proto */ +static CYTHON_INLINE int __Pyx_PySequence_ContainsTF(PyObject* item, PyObject* seq, int eq) { + int result = PySequence_Contains(seq, item); + return unlikely(result < 0) ? result : (result == (eq == Py_EQ)); +} + +/* ImportFrom.proto */ +static PyObject* __Pyx_ImportFrom(PyObject* module, PyObject* name); + +/* HasAttr.proto */ +static CYTHON_INLINE int __Pyx_HasAttr(PyObject *, PyObject *); + +/* IsLittleEndian.proto */ +static CYTHON_INLINE int __Pyx_Is_Little_Endian(void); + +/* BufferFormatCheck.proto */ +static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts); +static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx, + __Pyx_BufFmt_StackElem* stack, + __Pyx_TypeInfo* type); + +/* BufferGetAndValidate.proto */ +#define __Pyx_GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack)\ + ((obj == Py_None || obj == NULL) ?\ + (__Pyx_ZeroBuffer(buf), 0) :\ + __Pyx__GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack)) +static int __Pyx__GetBufferAndValidate(Py_buffer* buf, PyObject* obj, + __Pyx_TypeInfo* dtype, int flags, int nd, int cast, __Pyx_BufFmt_StackElem* stack); +static void __Pyx_ZeroBuffer(Py_buffer* buf); +static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info); +static Py_ssize_t __Pyx_minusones[] = { -1, -1, -1, -1, -1, -1, -1, -1 }; +static Py_ssize_t __Pyx_zeros[] = { 0, 0, 0, 0, 0, 0, 0, 0 }; + +/* BufferIndexError.proto */ +static void __Pyx_RaiseBufferIndexError(int axis); + +/* ListAppend.proto */ +#if CYTHON_USE_PYLIST_INTERNALS && CYTHON_ASSUME_SAFE_MACROS +static CYTHON_INLINE int __Pyx_PyList_Append(PyObject* list, PyObject* x) { + PyListObject* L = (PyListObject*) list; + Py_ssize_t len = Py_SIZE(list); + if (likely(L->allocated > len) & likely(len > (L->allocated >> 1))) { + Py_INCREF(x); + #if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 + L->ob_item[len] = x; + #else + PyList_SET_ITEM(list, len, x); + #endif + __Pyx_SET_SIZE(list, len + 1); + return 0; + } + return PyList_Append(list, x); +} +#else +#define __Pyx_PyList_Append(L,x) PyList_Append(L,x) +#endif + +/* SliceTupleAndList.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyList_GetSlice(PyObject* src, Py_ssize_t start, Py_ssize_t stop); +static CYTHON_INLINE PyObject* __Pyx_PyTuple_GetSlice(PyObject* src, Py_ssize_t start, Py_ssize_t stop); +#else +#define __Pyx_PyList_GetSlice(seq, start, stop) PySequence_GetSlice(seq, start, stop) +#define __Pyx_PyTuple_GetSlice(seq, start, stop) PySequence_GetSlice(seq, start, stop) +#endif + +/* PyIntCompare.proto */ +static CYTHON_INLINE int __Pyx_PyInt_BoolEqObjC(PyObject *op1, PyObject *op2, long intval, long inplace); + +/* PyObject_GenericGetAttrNoDict.proto */ +#if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static CYTHON_INLINE PyObject* __Pyx_PyObject_GenericGetAttrNoDict(PyObject* obj, PyObject* attr_name); +#else +#define __Pyx_PyObject_GenericGetAttrNoDict PyObject_GenericGetAttr +#endif + +/* PyObject_GenericGetAttr.proto */ +#if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static PyObject* __Pyx_PyObject_GenericGetAttr(PyObject* obj, PyObject* attr_name); +#else +#define __Pyx_PyObject_GenericGetAttr PyObject_GenericGetAttr +#endif + +/* IncludeStructmemberH.proto */ +#include + +/* FixUpExtensionType.proto */ +#if CYTHON_USE_TYPE_SPECS +static int __Pyx_fix_up_extension_type_from_spec(PyType_Spec *spec, PyTypeObject *type); +#endif + +/* PyObjectCallNoArg.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallNoArg(PyObject *func); + +/* PyObjectGetMethod.proto */ +static int __Pyx_PyObject_GetMethod(PyObject *obj, PyObject *name, PyObject **method); + +/* PyObjectCallMethod0.proto */ +static PyObject* __Pyx_PyObject_CallMethod0(PyObject* obj, PyObject* method_name); + +/* ValidateBasesTuple.proto */ +#if CYTHON_COMPILING_IN_CPYTHON || CYTHON_COMPILING_IN_LIMITED_API || CYTHON_USE_TYPE_SPECS +static int __Pyx_validate_bases_tuple(const char *type_name, Py_ssize_t dictoffset, PyObject *bases); +#endif + +/* PyType_Ready.proto */ +CYTHON_UNUSED static int __Pyx_PyType_Ready(PyTypeObject *t); + +/* SetVTable.proto */ +static int __Pyx_SetVtable(PyTypeObject* typeptr , void* vtable); + +/* GetVTable.proto */ +static void* __Pyx_GetVtable(PyTypeObject *type); + +/* MergeVTables.proto */ +#if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_MergeVtables(PyTypeObject *type); +#endif + +/* SetupReduce.proto */ +#if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_setup_reduce(PyObject* type_obj); +#endif + +/* TypeImport.proto */ +#ifndef __PYX_HAVE_RT_ImportType_proto_3_0_12 +#define __PYX_HAVE_RT_ImportType_proto_3_0_12 +#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L +#include +#endif +#if (defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) || __cplusplus >= 201103L +#define __PYX_GET_STRUCT_ALIGNMENT_3_0_12(s) alignof(s) +#else +#define __PYX_GET_STRUCT_ALIGNMENT_3_0_12(s) sizeof(void*) +#endif +enum __Pyx_ImportType_CheckSize_3_0_12 { + __Pyx_ImportType_CheckSize_Error_3_0_12 = 0, + __Pyx_ImportType_CheckSize_Warn_3_0_12 = 1, + __Pyx_ImportType_CheckSize_Ignore_3_0_12 = 2 +}; +static PyTypeObject *__Pyx_ImportType_3_0_12(PyObject* module, const char *module_name, const char *class_name, size_t size, size_t alignment, enum __Pyx_ImportType_CheckSize_3_0_12 check_size); +#endif + +/* FetchSharedCythonModule.proto */ +static PyObject *__Pyx_FetchSharedCythonABIModule(void); + +/* FetchCommonType.proto */ +#if !CYTHON_USE_TYPE_SPECS +static PyTypeObject* __Pyx_FetchCommonType(PyTypeObject* type); +#else +static PyTypeObject* __Pyx_FetchCommonTypeFromSpec(PyObject *module, PyType_Spec *spec, PyObject *bases); +#endif + +/* PyMethodNew.proto */ +#if CYTHON_COMPILING_IN_LIMITED_API +static PyObject *__Pyx_PyMethod_New(PyObject *func, PyObject *self, PyObject *typ) { + PyObject *typesModule=NULL, *methodType=NULL, *result=NULL; + CYTHON_UNUSED_VAR(typ); + if (!self) + return __Pyx_NewRef(func); + typesModule = PyImport_ImportModule("types"); + if (!typesModule) return NULL; + methodType = PyObject_GetAttrString(typesModule, "MethodType"); + Py_DECREF(typesModule); + if (!methodType) return NULL; + result = PyObject_CallFunctionObjArgs(methodType, func, self, NULL); + Py_DECREF(methodType); + return result; +} +#elif PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx_PyMethod_New(PyObject *func, PyObject *self, PyObject *typ) { + CYTHON_UNUSED_VAR(typ); + if (!self) + return __Pyx_NewRef(func); + return PyMethod_New(func, self); +} +#else + #define __Pyx_PyMethod_New PyMethod_New +#endif + +/* PyVectorcallFastCallDict.proto */ +#if CYTHON_METH_FASTCALL +static CYTHON_INLINE PyObject *__Pyx_PyVectorcall_FastCallDict(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw); +#endif + +/* CythonFunctionShared.proto */ +#define __Pyx_CyFunction_USED +#define __Pyx_CYFUNCTION_STATICMETHOD 0x01 +#define __Pyx_CYFUNCTION_CLASSMETHOD 0x02 +#define __Pyx_CYFUNCTION_CCLASS 0x04 +#define __Pyx_CYFUNCTION_COROUTINE 0x08 +#define __Pyx_CyFunction_GetClosure(f)\ + (((__pyx_CyFunctionObject *) (f))->func_closure) +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_CyFunction_GetClassObj(f)\ + (((__pyx_CyFunctionObject *) (f))->func_classobj) +#else + #define __Pyx_CyFunction_GetClassObj(f)\ + ((PyObject*) ((PyCMethodObject *) (f))->mm_class) +#endif +#define __Pyx_CyFunction_SetClassObj(f, classobj)\ + __Pyx__CyFunction_SetClassObj((__pyx_CyFunctionObject *) (f), (classobj)) +#define __Pyx_CyFunction_Defaults(type, f)\ + ((type *)(((__pyx_CyFunctionObject *) (f))->defaults)) +#define __Pyx_CyFunction_SetDefaultsGetter(f, g)\ + ((__pyx_CyFunctionObject *) (f))->defaults_getter = (g) +typedef struct { +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject_HEAD + PyObject *func; +#elif PY_VERSION_HEX < 0x030900B1 + PyCFunctionObject func; +#else + PyCMethodObject func; +#endif +#if CYTHON_BACKPORT_VECTORCALL + __pyx_vectorcallfunc func_vectorcall; +#endif +#if PY_VERSION_HEX < 0x030500A0 || CYTHON_COMPILING_IN_LIMITED_API + PyObject *func_weakreflist; +#endif + PyObject *func_dict; + PyObject *func_name; + PyObject *func_qualname; + PyObject *func_doc; + PyObject *func_globals; + PyObject *func_code; + PyObject *func_closure; +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + PyObject *func_classobj; +#endif + void *defaults; + int defaults_pyobjects; + size_t defaults_size; + int flags; + PyObject *defaults_tuple; + PyObject *defaults_kwdict; + PyObject *(*defaults_getter)(PyObject *); + PyObject *func_annotations; + PyObject *func_is_coroutine; +} __pyx_CyFunctionObject; +#undef __Pyx_CyOrPyCFunction_Check +#define __Pyx_CyFunction_Check(obj) __Pyx_TypeCheck(obj, __pyx_CyFunctionType) +#define __Pyx_CyOrPyCFunction_Check(obj) __Pyx_TypeCheck2(obj, __pyx_CyFunctionType, &PyCFunction_Type) +#define __Pyx_CyFunction_CheckExact(obj) __Pyx_IS_TYPE(obj, __pyx_CyFunctionType) +static CYTHON_INLINE int __Pyx__IsSameCyOrCFunction(PyObject *func, void *cfunc); +#undef __Pyx_IsSameCFunction +#define __Pyx_IsSameCFunction(func, cfunc) __Pyx__IsSameCyOrCFunction(func, cfunc) +static PyObject *__Pyx_CyFunction_Init(__pyx_CyFunctionObject* op, PyMethodDef *ml, + int flags, PyObject* qualname, + PyObject *closure, + PyObject *module, PyObject *globals, + PyObject* code); +static CYTHON_INLINE void __Pyx__CyFunction_SetClassObj(__pyx_CyFunctionObject* f, PyObject* classobj); +static CYTHON_INLINE void *__Pyx_CyFunction_InitDefaults(PyObject *m, + size_t size, + int pyobjects); +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsTuple(PyObject *m, + PyObject *tuple); +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsKwDict(PyObject *m, + PyObject *dict); +static CYTHON_INLINE void __Pyx_CyFunction_SetAnnotationsDict(PyObject *m, + PyObject *dict); +static int __pyx_CyFunction_init(PyObject *module); +#if CYTHON_METH_FASTCALL +static PyObject * __Pyx_CyFunction_Vectorcall_NOARGS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +static PyObject * __Pyx_CyFunction_Vectorcall_O(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS_METHOD(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +#if CYTHON_BACKPORT_VECTORCALL +#define __Pyx_CyFunction_func_vectorcall(f) (((__pyx_CyFunctionObject*)f)->func_vectorcall) +#else +#define __Pyx_CyFunction_func_vectorcall(f) (((PyCFunctionObject*)f)->vectorcall) +#endif +#endif + +/* CythonFunction.proto */ +static PyObject *__Pyx_CyFunction_New(PyMethodDef *ml, + int flags, PyObject* qualname, + PyObject *closure, + PyObject *module, PyObject *globals, + PyObject* code); + +/* CLineInTraceback.proto */ +#ifdef CYTHON_CLINE_IN_TRACEBACK +#define __Pyx_CLineForTraceback(tstate, c_line) (((CYTHON_CLINE_IN_TRACEBACK)) ? c_line : 0) +#else +static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line); +#endif + +/* CodeObjectCache.proto */ +#if !CYTHON_COMPILING_IN_LIMITED_API +typedef struct { + PyCodeObject* code_object; + int code_line; +} __Pyx_CodeObjectCacheEntry; +struct __Pyx_CodeObjectCache { + int count; + int max_count; + __Pyx_CodeObjectCacheEntry* entries; +}; +static struct __Pyx_CodeObjectCache __pyx_code_cache = {0,0,NULL}; +static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line); +static PyCodeObject *__pyx_find_code_object(int code_line); +static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object); +#endif + +/* AddTraceback.proto */ +static void __Pyx_AddTraceback(const char *funcname, int c_line, + int py_line, const char *filename); + +#if PY_MAJOR_VERSION < 3 + static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags); + static void __Pyx_ReleaseBuffer(Py_buffer *view); +#else + #define __Pyx_GetBuffer PyObject_GetBuffer + #define __Pyx_ReleaseBuffer PyBuffer_Release +#endif + + +/* BufferStructDeclare.proto */ +typedef struct { + Py_ssize_t shape, strides, suboffsets; +} __Pyx_Buf_DimInfo; +typedef struct { + size_t refcount; + Py_buffer pybuffer; +} __Pyx_Buffer; +typedef struct { + __Pyx_Buffer *rcbuffer; + char *data; + __Pyx_Buf_DimInfo diminfo[8]; +} __Pyx_LocalBuf_ND; + +/* MemviewSliceIsContig.proto */ +static int __pyx_memviewslice_is_contig(const __Pyx_memviewslice mvs, char order, int ndim); + +/* OverlappingSlices.proto */ +static int __pyx_slices_overlap(__Pyx_memviewslice *slice1, + __Pyx_memviewslice *slice2, + int ndim, size_t itemsize); + +/* TypeInfoCompare.proto */ +static int __pyx_typeinfo_cmp(__Pyx_TypeInfo *a, __Pyx_TypeInfo *b); + +/* MemviewSliceValidateAndInit.proto */ +static int __Pyx_ValidateAndInit_memviewslice( + int *axes_specs, + int c_or_f_flag, + int buf_flags, + int ndim, + __Pyx_TypeInfo *dtype, + __Pyx_BufFmt_StackElem stack[], + __Pyx_memviewslice *memviewslice, + PyObject *original_obj); + +/* ObjectToMemviewSlice.proto */ +static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(PyObject *, int writable_flag); + +/* MemviewDtypeToObject.proto */ +static CYTHON_INLINE PyObject *__pyx_memview_get_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(const char *itemp); +static CYTHON_INLINE int __pyx_memview_set_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(const char *itemp, PyObject *obj); + +/* ObjectToMemviewSlice.proto */ +static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(PyObject *, int writable_flag); + +/* RealImag.proto */ +#if CYTHON_CCOMPLEX + #ifdef __cplusplus + #define __Pyx_CREAL(z) ((z).real()) + #define __Pyx_CIMAG(z) ((z).imag()) + #else + #define __Pyx_CREAL(z) (__real__(z)) + #define __Pyx_CIMAG(z) (__imag__(z)) + #endif +#else + #define __Pyx_CREAL(z) ((z).real) + #define __Pyx_CIMAG(z) ((z).imag) +#endif +#if defined(__cplusplus) && CYTHON_CCOMPLEX\ + && (defined(_WIN32) || defined(__clang__) || (defined(__GNUC__) && (__GNUC__ >= 5 || __GNUC__ == 4 && __GNUC_MINOR__ >= 4 )) || __cplusplus >= 201103) + #define __Pyx_SET_CREAL(z,x) ((z).real(x)) + #define __Pyx_SET_CIMAG(z,y) ((z).imag(y)) +#else + #define __Pyx_SET_CREAL(z,x) __Pyx_CREAL(z) = (x) + #define __Pyx_SET_CIMAG(z,y) __Pyx_CIMAG(z) = (y) +#endif + +/* Arithmetic.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #define __Pyx_c_eq_float(a, b) ((a)==(b)) + #define __Pyx_c_sum_float(a, b) ((a)+(b)) + #define __Pyx_c_diff_float(a, b) ((a)-(b)) + #define __Pyx_c_prod_float(a, b) ((a)*(b)) + #define __Pyx_c_quot_float(a, b) ((a)/(b)) + #define __Pyx_c_neg_float(a) (-(a)) + #ifdef __cplusplus + #define __Pyx_c_is_zero_float(z) ((z)==(float)0) + #define __Pyx_c_conj_float(z) (::std::conj(z)) + #if 1 + #define __Pyx_c_abs_float(z) (::std::abs(z)) + #define __Pyx_c_pow_float(a, b) (::std::pow(a, b)) + #endif + #else + #define __Pyx_c_is_zero_float(z) ((z)==0) + #define __Pyx_c_conj_float(z) (conjf(z)) + #if 1 + #define __Pyx_c_abs_float(z) (cabsf(z)) + #define __Pyx_c_pow_float(a, b) (cpowf(a, b)) + #endif + #endif +#else + static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex); + static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex); + #if 1 + static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex, __pyx_t_float_complex); + #endif +#endif + +/* Arithmetic.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #define __Pyx_c_eq_double(a, b) ((a)==(b)) + #define __Pyx_c_sum_double(a, b) ((a)+(b)) + #define __Pyx_c_diff_double(a, b) ((a)-(b)) + #define __Pyx_c_prod_double(a, b) ((a)*(b)) + #define __Pyx_c_quot_double(a, b) ((a)/(b)) + #define __Pyx_c_neg_double(a) (-(a)) + #ifdef __cplusplus + #define __Pyx_c_is_zero_double(z) ((z)==(double)0) + #define __Pyx_c_conj_double(z) (::std::conj(z)) + #if 1 + #define __Pyx_c_abs_double(z) (::std::abs(z)) + #define __Pyx_c_pow_double(a, b) (::std::pow(a, b)) + #endif + #else + #define __Pyx_c_is_zero_double(z) ((z)==0) + #define __Pyx_c_conj_double(z) (conj(z)) + #if 1 + #define __Pyx_c_abs_double(z) (cabs(z)) + #define __Pyx_c_pow_double(a, b) (cpow(a, b)) + #endif + #endif +#else + static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex); + static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex); + #if 1 + static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex, __pyx_t_double_complex); + #endif +#endif + +/* Arithmetic.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #define __Pyx_c_eq_long__double(a, b) ((a)==(b)) + #define __Pyx_c_sum_long__double(a, b) ((a)+(b)) + #define __Pyx_c_diff_long__double(a, b) ((a)-(b)) + #define __Pyx_c_prod_long__double(a, b) ((a)*(b)) + #define __Pyx_c_quot_long__double(a, b) ((a)/(b)) + #define __Pyx_c_neg_long__double(a) (-(a)) + #ifdef __cplusplus + #define __Pyx_c_is_zero_long__double(z) ((z)==(long double)0) + #define __Pyx_c_conj_long__double(z) (::std::conj(z)) + #if 1 + #define __Pyx_c_abs_long__double(z) (::std::abs(z)) + #define __Pyx_c_pow_long__double(a, b) (::std::pow(a, b)) + #endif + #else + #define __Pyx_c_is_zero_long__double(z) ((z)==0) + #define __Pyx_c_conj_long__double(z) (conjl(z)) + #if 1 + #define __Pyx_c_abs_long__double(z) (cabsl(z)) + #define __Pyx_c_pow_long__double(a, b) (cpowl(a, b)) + #endif + #endif +#else + static CYTHON_INLINE int __Pyx_c_eq_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_sum_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_diff_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_prod_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_quot_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_neg_long__double(__pyx_t_long_double_complex); + static CYTHON_INLINE int __Pyx_c_is_zero_long__double(__pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_conj_long__double(__pyx_t_long_double_complex); + #if 1 + static CYTHON_INLINE long double __Pyx_c_abs_long__double(__pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_pow_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + #endif +#endif + +/* MemviewSliceCopyTemplate.proto */ +static __Pyx_memviewslice +__pyx_memoryview_copy_new_contig(const __Pyx_memviewslice *from_mvs, + const char *mode, int ndim, + size_t sizeof_dtype, int contig_flag, + int dtype_is_object); + +/* MemviewSliceInit.proto */ +#define __Pyx_BUF_MAX_NDIMS %(BUF_MAX_NDIMS)d +#define __Pyx_MEMVIEW_DIRECT 1 +#define __Pyx_MEMVIEW_PTR 2 +#define __Pyx_MEMVIEW_FULL 4 +#define __Pyx_MEMVIEW_CONTIG 8 +#define __Pyx_MEMVIEW_STRIDED 16 +#define __Pyx_MEMVIEW_FOLLOW 32 +#define __Pyx_IS_C_CONTIG 1 +#define __Pyx_IS_F_CONTIG 2 +static int __Pyx_init_memviewslice( + struct __pyx_memoryview_obj *memview, + int ndim, + __Pyx_memviewslice *memviewslice, + int memview_is_new_reference); +static CYTHON_INLINE int __pyx_add_acquisition_count_locked( + __pyx_atomic_int_type *acquisition_count, PyThread_type_lock lock); +static CYTHON_INLINE int __pyx_sub_acquisition_count_locked( + __pyx_atomic_int_type *acquisition_count, PyThread_type_lock lock); +#define __pyx_get_slice_count_pointer(memview) (&memview->acquisition_count) +#define __PYX_INC_MEMVIEW(slice, have_gil) __Pyx_INC_MEMVIEW(slice, have_gil, __LINE__) +#define __PYX_XCLEAR_MEMVIEW(slice, have_gil) __Pyx_XCLEAR_MEMVIEW(slice, have_gil, __LINE__) +static CYTHON_INLINE void __Pyx_INC_MEMVIEW(__Pyx_memviewslice *, int, int); +static CYTHON_INLINE void __Pyx_XCLEAR_MEMVIEW(__Pyx_memviewslice *, int, int); + +/* CIntFromPy.proto */ +static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *); + +/* CIntFromPy.proto */ +static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *); + +/* CIntToPy.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyInt_From_npy_int64(npy_int64 value); + +/* CIntFromPy.proto */ +static CYTHON_INLINE npy_int64 __Pyx_PyInt_As_npy_int64(PyObject *); + +/* CIntToPy.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value); + +/* None.proto */ +#include + +/* CIntToPy.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value); + +/* CIntFromPy.proto */ +static CYTHON_INLINE char __Pyx_PyInt_As_char(PyObject *); + +/* FormatTypeName.proto */ +#if CYTHON_COMPILING_IN_LIMITED_API +typedef PyObject *__Pyx_TypeName; +#define __Pyx_FMT_TYPENAME "%U" +static __Pyx_TypeName __Pyx_PyType_GetName(PyTypeObject* tp); +#define __Pyx_DECREF_TypeName(obj) Py_XDECREF(obj) +#else +typedef const char *__Pyx_TypeName; +#define __Pyx_FMT_TYPENAME "%.200s" +#define __Pyx_PyType_GetName(tp) ((tp)->tp_name) +#define __Pyx_DECREF_TypeName(obj) +#endif + +/* CheckBinaryVersion.proto */ +static unsigned long __Pyx_get_runtime_version(void); +static int __Pyx_check_binary_version(unsigned long ct_version, unsigned long rt_version, int allow_newer); + +/* InitStrings.proto */ +static int __Pyx_InitStrings(__Pyx_StringTabEntry *t); + +/* #### Code section: module_declarations ### */ +static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self); /* proto*/ +static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index); /* proto*/ +static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj); /* proto*/ +static PyObject *__pyx_memoryview_setitem_slice_assignment(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_dst, PyObject *__pyx_v_src); /* proto*/ +static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memoryview_obj *__pyx_v_self, struct __pyx_memoryview_obj *__pyx_v_dst, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryview_setitem_indexed(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp); /* proto*/ +static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryview__get_base(struct __pyx_memoryview_obj *__pyx_v_self); /* proto*/ +static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp); /* proto*/ +static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryviewslice__get_base(struct __pyx_memoryviewslice_obj *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_8itemsize_itemsize(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_9alignment_alignment(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_6fields_fields(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_5names_names(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyArray_ArrayDescr *__pyx_f_5numpy_5dtype_8subarray_subarray(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_uint64 __pyx_f_5numpy_5dtype_5flags_flags(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_7numiter_numiter(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_4size_size(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_5index_index(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_2nd_nd(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_9broadcast_10dimensions_dimensions(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE void **__pyx_f_5numpy_9broadcast_5iters_iters(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyObject *__pyx_f_5numpy_7ndarray_4base_base(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyArray_Descr *__pyx_f_5numpy_7ndarray_5descr_descr(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE int __pyx_f_5numpy_7ndarray_4ndim_ndim(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_5shape_shape(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_7strides_strides(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_7ndarray_4size_size(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE char *__pyx_f_5numpy_7ndarray_4data_data(PyArrayObject *__pyx_v_self); /* proto*/ + +/* Module declarations from "cython.view" */ + +/* Module declarations from "cython.dataclasses" */ + +/* Module declarations from "cython" */ + +/* Module declarations from "libc.string" */ + +/* Module declarations from "libc.stdio" */ + +/* Module declarations from "__builtin__" */ + +/* Module declarations from "cpython.type" */ + +/* Module declarations from "cpython" */ + +/* Module declarations from "cpython.object" */ + +/* Module declarations from "cpython.ref" */ + +/* Module declarations from "numpy" */ + +/* Module declarations from "numpy" */ + +/* Module declarations from "fairseq.data.data_utils_fast" */ +static PyObject *__pyx_collections_abc_Sequence = 0; +static PyObject *generic = 0; +static PyObject *strided = 0; +static PyObject *indirect = 0; +static PyObject *contiguous = 0; +static PyObject *indirect_contiguous = 0; +static int __pyx_memoryview_thread_locks_used; +static PyThread_type_lock __pyx_memoryview_thread_locks[8]; +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast__is_batch_full(long, long, long, long); /*proto*/ +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast_batch_by_size_fast(PyArrayObject *, PyObject *, long, long, int, int __pyx_skip_dispatch); /*proto*/ +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast__find_valid_shape(__Pyx_memviewslice, long, long); /*proto*/ +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast_batch_fixed_shapes_fast(PyArrayObject *, PyObject *, PyArrayObject *, int __pyx_skip_dispatch); /*proto*/ +static int __pyx_array_allocate_buffer(struct __pyx_array_obj *); /*proto*/ +static struct __pyx_array_obj *__pyx_array_new(PyObject *, Py_ssize_t, char *, char *, char *); /*proto*/ +static PyObject *__pyx_memoryview_new(PyObject *, int, int, __Pyx_TypeInfo *); /*proto*/ +static CYTHON_INLINE int __pyx_memoryview_check(PyObject *); /*proto*/ +static PyObject *_unellipsify(PyObject *, int); /*proto*/ +static int assert_direct_dimensions(Py_ssize_t *, int); /*proto*/ +static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_obj *, PyObject *); /*proto*/ +static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *, Py_ssize_t, Py_ssize_t, Py_ssize_t, int, int, int *, Py_ssize_t, Py_ssize_t, Py_ssize_t, int, int, int, int); /*proto*/ +static char *__pyx_pybuffer_index(Py_buffer *, char *, Py_ssize_t, Py_ssize_t); /*proto*/ +static int __pyx_memslice_transpose(__Pyx_memviewslice *); /*proto*/ +static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice, int, PyObject *(*)(char *), int (*)(char *, PyObject *), int); /*proto*/ +static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __pyx_memoryview_obj *, __Pyx_memviewslice *); /*proto*/ +static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *, __Pyx_memviewslice *); /*proto*/ +static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *); /*proto*/ +static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview_obj *, __Pyx_memviewslice *); /*proto*/ +static Py_ssize_t abs_py_ssize_t(Py_ssize_t); /*proto*/ +static char __pyx_get_best_slice_order(__Pyx_memviewslice *, int); /*proto*/ +static void _copy_strided_to_strided(char *, Py_ssize_t *, char *, Py_ssize_t *, Py_ssize_t *, Py_ssize_t *, int, size_t); /*proto*/ +static void copy_strided_to_strided(__Pyx_memviewslice *, __Pyx_memviewslice *, int, size_t); /*proto*/ +static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *, int); /*proto*/ +static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *, Py_ssize_t *, Py_ssize_t, int, char); /*proto*/ +static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *, __Pyx_memviewslice *, char, int); /*proto*/ +static int __pyx_memoryview_err_extents(int, Py_ssize_t, Py_ssize_t); /*proto*/ +static int __pyx_memoryview_err_dim(PyObject *, PyObject *, int); /*proto*/ +static int __pyx_memoryview_err(PyObject *, PyObject *); /*proto*/ +static int __pyx_memoryview_err_no_memory(void); /*proto*/ +static int __pyx_memoryview_copy_contents(__Pyx_memviewslice, __Pyx_memviewslice, int, int, int); /*proto*/ +static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *, int, int); /*proto*/ +static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *, int, int, int); /*proto*/ +static void __pyx_memoryview_refcount_objects_in_slice_with_gil(char *, Py_ssize_t *, Py_ssize_t *, int, int); /*proto*/ +static void __pyx_memoryview_refcount_objects_in_slice(char *, Py_ssize_t *, Py_ssize_t *, int, int); /*proto*/ +static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *, int, size_t, void *, int); /*proto*/ +static void __pyx_memoryview__slice_assign_scalar(char *, Py_ssize_t *, Py_ssize_t *, int, size_t, void *); /*proto*/ +static PyObject *__pyx_unpickle_Enum__set_state(struct __pyx_MemviewEnum_obj *, PyObject *); /*proto*/ +/* #### Code section: typeinfo ### */ +static __Pyx_TypeInfo __Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t = { "DTYPE_t", NULL, sizeof(__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t), { 0 }, 0, __PYX_IS_UNSIGNED(__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t) ? 'U' : 'I', __PYX_IS_UNSIGNED(__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t), 0 }; +/* #### Code section: before_global_var ### */ +#define __Pyx_MODULE_NAME "fairseq.data.data_utils_fast" +extern int __pyx_module_is_main_fairseq__data__data_utils_fast; +int __pyx_module_is_main_fairseq__data__data_utils_fast = 0; + +/* Implementation of "fairseq.data.data_utils_fast" */ +/* #### Code section: global_var ### */ +static PyObject *__pyx_builtin_range; +static PyObject *__pyx_builtin_AssertionError; +static PyObject *__pyx_builtin_max; +static PyObject *__pyx_builtin___import__; +static PyObject *__pyx_builtin_ValueError; +static PyObject *__pyx_builtin_MemoryError; +static PyObject *__pyx_builtin_enumerate; +static PyObject *__pyx_builtin_TypeError; +static PyObject *__pyx_builtin_Ellipsis; +static PyObject *__pyx_builtin_id; +static PyObject *__pyx_builtin_IndexError; +static PyObject *__pyx_builtin_ImportError; +/* #### Code section: string_decls ### */ +static const char __pyx_k_[] = ": "; +static const char __pyx_k_O[] = "O"; +static const char __pyx_k_c[] = "c"; +static const char __pyx_k__2[] = "."; +static const char __pyx_k__3[] = "*"; +static const char __pyx_k__6[] = "'"; +static const char __pyx_k__7[] = ")"; +static const char __pyx_k_gc[] = "gc"; +static const char __pyx_k_id[] = "id"; +static const char __pyx_k_np[] = "np"; +static const char __pyx_k__26[] = "?"; +static const char __pyx_k_abc[] = "abc"; +static const char __pyx_k_and[] = " and "; +static const char __pyx_k_got[] = " (got "; +static const char __pyx_k_max[] = "max"; +static const char __pyx_k_new[] = "__new__"; +static const char __pyx_k_obj[] = "obj"; +static const char __pyx_k_sys[] = "sys"; +static const char __pyx_k_base[] = "base"; +static const char __pyx_k_dict[] = "__dict__"; +static const char __pyx_k_main[] = "__main__"; +static const char __pyx_k_mode[] = "mode"; +static const char __pyx_k_name[] = "name"; +static const char __pyx_k_ndim[] = "ndim"; +static const char __pyx_k_pack[] = "pack"; +static const char __pyx_k_size[] = "size"; +static const char __pyx_k_spec[] = "__spec__"; +static const char __pyx_k_step[] = "step"; +static const char __pyx_k_stop[] = "stop"; +static const char __pyx_k_test[] = "__test__"; +static const char __pyx_k_ASCII[] = "ASCII"; +static const char __pyx_k_DTYPE[] = "DTYPE"; +static const char __pyx_k_class[] = "__class__"; +static const char __pyx_k_count[] = "count"; +static const char __pyx_k_error[] = "error"; +static const char __pyx_k_flags[] = "flags"; +static const char __pyx_k_index[] = "index"; +static const char __pyx_k_int64[] = "int64"; +static const char __pyx_k_numpy[] = "numpy"; +static const char __pyx_k_range[] = "range"; +static const char __pyx_k_shape[] = "shape"; +static const char __pyx_k_start[] = "start"; +static const char __pyx_k_enable[] = "enable"; +static const char __pyx_k_encode[] = "encode"; +static const char __pyx_k_format[] = "format"; +static const char __pyx_k_import[] = "__import__"; +static const char __pyx_k_name_2[] = "__name__"; +static const char __pyx_k_pickle[] = "pickle"; +static const char __pyx_k_reduce[] = "__reduce__"; +static const char __pyx_k_struct[] = "struct"; +static const char __pyx_k_unpack[] = "unpack"; +static const char __pyx_k_update[] = "update"; +static const char __pyx_k_disable[] = "disable"; +static const char __pyx_k_fortran[] = "fortran"; +static const char __pyx_k_indices[] = "indices"; +static const char __pyx_k_memview[] = "memview"; +static const char __pyx_k_Ellipsis[] = "Ellipsis"; +static const char __pyx_k_Sequence[] = "Sequence"; +static const char __pyx_k_bsz_mult[] = "bsz_mult"; +static const char __pyx_k_getstate[] = "__getstate__"; +static const char __pyx_k_itemsize[] = "itemsize"; +static const char __pyx_k_pyx_type[] = "__pyx_type"; +static const char __pyx_k_register[] = "register"; +static const char __pyx_k_setstate[] = "__setstate__"; +static const char __pyx_k_TypeError[] = "TypeError"; +static const char __pyx_k_enumerate[] = "enumerate"; +static const char __pyx_k_isenabled[] = "isenabled"; +static const char __pyx_k_pyx_state[] = "__pyx_state"; +static const char __pyx_k_reduce_ex[] = "__reduce_ex__"; +static const char __pyx_k_IndexError[] = "IndexError"; +static const char __pyx_k_ValueError[] = "ValueError"; +static const char __pyx_k_max_tokens[] = "max_tokens"; +static const char __pyx_k_pyx_result[] = "__pyx_result"; +static const char __pyx_k_pyx_vtable[] = "__pyx_vtable__"; +static const char __pyx_k_ImportError[] = "ImportError"; +static const char __pyx_k_MemoryError[] = "MemoryError"; +static const char __pyx_k_PickleError[] = "PickleError"; +static const char __pyx_k_collections[] = "collections"; +static const char __pyx_k_initializing[] = "_initializing"; +static const char __pyx_k_is_coroutine[] = "_is_coroutine"; +static const char __pyx_k_pyx_checksum[] = "__pyx_checksum"; +static const char __pyx_k_stringsource[] = ""; +static const char __pyx_k_version_info[] = "version_info"; +static const char __pyx_k_class_getitem[] = "__class_getitem__"; +static const char __pyx_k_max_sentences[] = "max_sentences"; +static const char __pyx_k_num_tokens_fn[] = "num_tokens_fn"; +static const char __pyx_k_reduce_cython[] = "__reduce_cython__"; +static const char __pyx_k_AssertionError[] = "AssertionError"; +static const char __pyx_k_View_MemoryView[] = "View.MemoryView"; +static const char __pyx_k_allocate_buffer[] = "allocate_buffer"; +static const char __pyx_k_collections_abc[] = "collections.abc"; +static const char __pyx_k_dtype_is_object[] = "dtype_is_object"; +static const char __pyx_k_pyx_PickleError[] = "__pyx_PickleError"; +static const char __pyx_k_setstate_cython[] = "__setstate_cython__"; +static const char __pyx_k_pyx_unpickle_Enum[] = "__pyx_unpickle_Enum"; +static const char __pyx_k_asyncio_coroutines[] = "asyncio.coroutines"; +static const char __pyx_k_batch_by_size_fast[] = "batch_by_size_fast"; +static const char __pyx_k_cline_in_traceback[] = "cline_in_traceback"; +static const char __pyx_k_strided_and_direct[] = ""; +static const char __pyx_k_fixed_shapes_sorted[] = "fixed_shapes_sorted"; +static const char __pyx_k_strided_and_indirect[] = ""; +static const char __pyx_k_Invalid_shape_in_axis[] = "Invalid shape in axis "; +static const char __pyx_k_contiguous_and_direct[] = ""; +static const char __pyx_k_Cannot_index_with_type[] = "Cannot index with type '"; +static const char __pyx_k_MemoryView_of_r_object[] = ""; +static const char __pyx_k_MemoryView_of_r_at_0x_x[] = ""; +static const char __pyx_k_batch_fixed_shapes_fast[] = "batch_fixed_shapes_fast"; +static const char __pyx_k_contiguous_and_indirect[] = ""; +static const char __pyx_k_Dimension_d_is_not_direct[] = "Dimension %d is not direct"; +static const char __pyx_k_Index_out_of_bounds_axis_d[] = "Index out of bounds (axis %d)"; +static const char __pyx_k_Step_may_not_be_zero_axis_d[] = "Step may not be zero (axis %d)"; +static const char __pyx_k_itemsize_0_for_cython_array[] = "itemsize <= 0 for cython.array"; +static const char __pyx_k_fairseq_data_data_utils_fast[] = "fairseq.data.data_utils_fast"; +static const char __pyx_k_unable_to_allocate_array_data[] = "unable to allocate array data."; +static const char __pyx_k_strided_and_direct_or_indirect[] = ""; +static const char __pyx_k_All_dimensions_preceding_dimensi[] = "All dimensions preceding dimension %d must be indexed and not sliced"; +static const char __pyx_k_Buffer_view_does_not_expose_stri[] = "Buffer view does not expose strides"; +static const char __pyx_k_Can_only_create_a_buffer_that_is[] = "Can only create a buffer that is contiguous in memory."; +static const char __pyx_k_Cannot_assign_to_read_only_memor[] = "Cannot assign to read-only memoryview"; +static const char __pyx_k_Cannot_create_writable_memory_vi[] = "Cannot create writable memory view from read-only memoryview"; +static const char __pyx_k_Cannot_transpose_memoryview_with[] = "Cannot transpose memoryview with indirect dimensions"; +static const char __pyx_k_Empty_shape_tuple_for_cython_arr[] = "Empty shape tuple for cython.array"; +static const char __pyx_k_Incompatible_checksums_0x_x_vs_0[] = "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))"; +static const char __pyx_k_Indirect_dimensions_not_supporte[] = "Indirect dimensions not supported"; +static const char __pyx_k_Invalid_mode_expected_c_or_fortr[] = "Invalid mode, expected 'c' or 'fortran', got "; +static const char __pyx_k_Out_of_bounds_on_buffer_access_a[] = "Out of bounds on buffer access (axis "; +static const char __pyx_k_Unable_to_convert_item_to_object[] = "Unable to convert item to object"; +static const char __pyx_k_fairseq_data_data_utils_fast_pyx[] = "fairseq/data/data_utils_fast.pyx"; +static const char __pyx_k_got_differing_extents_in_dimensi[] = "got differing extents in dimension "; +static const char __pyx_k_no_default___reduce___due_to_non[] = "no default __reduce__ due to non-trivial __cinit__"; +static const char __pyx_k_numpy__core_multiarray_failed_to[] = "numpy._core.multiarray failed to import"; +static const char __pyx_k_numpy__core_umath_failed_to_impo[] = "numpy._core.umath failed to import"; +static const char __pyx_k_sentence_at_index_of_size_exceed[] = "sentence at index {} of size {} exceeds max_tokens limit of {}!"; +static const char __pyx_k_unable_to_allocate_shape_and_str[] = "unable to allocate shape and strides."; +/* #### Code section: decls ### */ +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_shape, Py_ssize_t __pyx_v_itemsize, PyObject *__pyx_v_format, PyObject *__pyx_v_mode, int __pyx_v_allocate_buffer); /* proto */ +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(struct __pyx_array_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /* proto */ +static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struct __pyx_array_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_5array_7memview___get__(struct __pyx_array_obj *__pyx_v_self); /* proto */ +static Py_ssize_t __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(struct __pyx_array_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_attr); /* proto */ +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item); /* proto */ +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_12__setitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value); /* proto */ +static PyObject *__pyx_pf___pyx_array___reduce_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_array_2__setstate_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state); /* proto */ +static int __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v_name); /* proto */ +static PyObject *__pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr__(struct __pyx_MemviewEnum_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_MemviewEnum___reduce_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_MemviewEnum_2__setstate_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v___pyx_state); /* proto */ +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj, int __pyx_v_flags, int __pyx_v_dtype_is_object); /* proto */ +static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__dealloc__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4__getitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index); /* proto */ +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value); /* proto */ +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbuffer__(struct __pyx_memoryview_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_10__len__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12__repr__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14__str__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16is_c_contig(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18is_f_contig(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20copy(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22copy_fortran(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryview___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryview_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state); /* proto */ +static void __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewslice___dealloc__(struct __pyx_memoryviewslice_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryviewslice___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryviewslice_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView___pyx_unpickle_Enum(CYTHON_UNUSED PyObject *__pyx_self, PyObject *__pyx_v___pyx_type, long __pyx_v___pyx_checksum, PyObject *__pyx_v___pyx_state); /* proto */ +static PyObject *__pyx_pf_7fairseq_4data_15data_utils_fast_batch_by_size_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_indices, PyObject *__pyx_v_num_tokens_fn, long __pyx_v_max_tokens, long __pyx_v_max_sentences, int __pyx_v_bsz_mult); /* proto */ +static PyObject *__pyx_pf_7fairseq_4data_15data_utils_fast_2batch_fixed_shapes_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_indices, PyObject *__pyx_v_num_tokens_fn, PyArrayObject *__pyx_v_fixed_shapes_sorted); /* proto */ +static PyObject *__pyx_tp_new_array(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +static PyObject *__pyx_tp_new_Enum(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +static PyObject *__pyx_tp_new_memoryview(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +static PyObject *__pyx_tp_new__memoryviewslice(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +/* #### Code section: late_includes ### */ +/* #### Code section: module_state ### */ +typedef struct { + PyObject *__pyx_d; + PyObject *__pyx_b; + PyObject *__pyx_cython_runtime; + PyObject *__pyx_empty_tuple; + PyObject *__pyx_empty_bytes; + PyObject *__pyx_empty_unicode; + #ifdef __Pyx_CyFunction_USED + PyTypeObject *__pyx_CyFunctionType; + #endif + #ifdef __Pyx_FusedFunction_USED + PyTypeObject *__pyx_FusedFunctionType; + #endif + #ifdef __Pyx_Generator_USED + PyTypeObject *__pyx_GeneratorType; + #endif + #ifdef __Pyx_IterableCoroutine_USED + PyTypeObject *__pyx_IterableCoroutineType; + #endif + #ifdef __Pyx_Coroutine_USED + PyTypeObject *__pyx_CoroutineAwaitType; + #endif + #ifdef __Pyx_Coroutine_USED + PyTypeObject *__pyx_CoroutineType; + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + PyTypeObject *__pyx_ptype_7cpython_4type_type; + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + PyTypeObject *__pyx_ptype_5numpy_dtype; + PyTypeObject *__pyx_ptype_5numpy_flatiter; + PyTypeObject *__pyx_ptype_5numpy_broadcast; + PyTypeObject *__pyx_ptype_5numpy_ndarray; + PyTypeObject *__pyx_ptype_5numpy_generic; + PyTypeObject *__pyx_ptype_5numpy_number; + PyTypeObject *__pyx_ptype_5numpy_integer; + PyTypeObject *__pyx_ptype_5numpy_signedinteger; + PyTypeObject *__pyx_ptype_5numpy_unsignedinteger; + PyTypeObject *__pyx_ptype_5numpy_inexact; + PyTypeObject *__pyx_ptype_5numpy_floating; + PyTypeObject *__pyx_ptype_5numpy_complexfloating; + PyTypeObject *__pyx_ptype_5numpy_flexible; + PyTypeObject *__pyx_ptype_5numpy_character; + PyTypeObject *__pyx_ptype_5numpy_ufunc; + #if CYTHON_USE_MODULE_STATE + PyObject *__pyx_type___pyx_array; + PyObject *__pyx_type___pyx_MemviewEnum; + PyObject *__pyx_type___pyx_memoryview; + PyObject *__pyx_type___pyx_memoryviewslice; + #endif + PyTypeObject *__pyx_array_type; + PyTypeObject *__pyx_MemviewEnum_type; + PyTypeObject *__pyx_memoryview_type; + PyTypeObject *__pyx_memoryviewslice_type; + PyObject *__pyx_kp_u_; + PyObject *__pyx_n_s_ASCII; + PyObject *__pyx_kp_s_All_dimensions_preceding_dimensi; + PyObject *__pyx_n_s_AssertionError; + PyObject *__pyx_kp_s_Buffer_view_does_not_expose_stri; + PyObject *__pyx_kp_s_Can_only_create_a_buffer_that_is; + PyObject *__pyx_kp_s_Cannot_assign_to_read_only_memor; + PyObject *__pyx_kp_s_Cannot_create_writable_memory_vi; + PyObject *__pyx_kp_u_Cannot_index_with_type; + PyObject *__pyx_kp_s_Cannot_transpose_memoryview_with; + PyObject *__pyx_n_s_DTYPE; + PyObject *__pyx_kp_s_Dimension_d_is_not_direct; + PyObject *__pyx_n_s_Ellipsis; + PyObject *__pyx_kp_s_Empty_shape_tuple_for_cython_arr; + PyObject *__pyx_n_s_ImportError; + PyObject *__pyx_kp_s_Incompatible_checksums_0x_x_vs_0; + PyObject *__pyx_n_s_IndexError; + PyObject *__pyx_kp_s_Index_out_of_bounds_axis_d; + PyObject *__pyx_kp_s_Indirect_dimensions_not_supporte; + PyObject *__pyx_kp_u_Invalid_mode_expected_c_or_fortr; + PyObject *__pyx_kp_u_Invalid_shape_in_axis; + PyObject *__pyx_n_s_MemoryError; + PyObject *__pyx_kp_s_MemoryView_of_r_at_0x_x; + PyObject *__pyx_kp_s_MemoryView_of_r_object; + PyObject *__pyx_n_b_O; + PyObject *__pyx_kp_u_Out_of_bounds_on_buffer_access_a; + PyObject *__pyx_n_s_PickleError; + PyObject *__pyx_n_s_Sequence; + PyObject *__pyx_kp_s_Step_may_not_be_zero_axis_d; + PyObject *__pyx_n_s_TypeError; + PyObject *__pyx_kp_s_Unable_to_convert_item_to_object; + PyObject *__pyx_n_s_ValueError; + PyObject *__pyx_n_s_View_MemoryView; + PyObject *__pyx_kp_u__2; + PyObject *__pyx_n_s__26; + PyObject *__pyx_n_s__3; + PyObject *__pyx_kp_u__6; + PyObject *__pyx_kp_u__7; + PyObject *__pyx_n_s_abc; + PyObject *__pyx_n_s_allocate_buffer; + PyObject *__pyx_kp_u_and; + PyObject *__pyx_n_s_asyncio_coroutines; + PyObject *__pyx_n_s_base; + PyObject *__pyx_n_s_batch_by_size_fast; + PyObject *__pyx_n_s_batch_fixed_shapes_fast; + PyObject *__pyx_n_s_bsz_mult; + PyObject *__pyx_n_s_c; + PyObject *__pyx_n_u_c; + PyObject *__pyx_n_s_class; + PyObject *__pyx_n_s_class_getitem; + PyObject *__pyx_n_s_cline_in_traceback; + PyObject *__pyx_n_s_collections; + PyObject *__pyx_kp_s_collections_abc; + PyObject *__pyx_kp_s_contiguous_and_direct; + PyObject *__pyx_kp_s_contiguous_and_indirect; + PyObject *__pyx_n_s_count; + PyObject *__pyx_n_s_dict; + PyObject *__pyx_kp_u_disable; + PyObject *__pyx_n_s_dtype_is_object; + PyObject *__pyx_kp_u_enable; + PyObject *__pyx_n_s_encode; + PyObject *__pyx_n_s_enumerate; + PyObject *__pyx_n_s_error; + PyObject *__pyx_n_s_fairseq_data_data_utils_fast; + PyObject *__pyx_kp_s_fairseq_data_data_utils_fast_pyx; + PyObject *__pyx_n_s_fixed_shapes_sorted; + PyObject *__pyx_n_s_flags; + PyObject *__pyx_n_s_format; + PyObject *__pyx_n_s_fortran; + PyObject *__pyx_n_u_fortran; + PyObject *__pyx_kp_u_gc; + PyObject *__pyx_n_s_getstate; + PyObject *__pyx_kp_u_got; + PyObject *__pyx_kp_u_got_differing_extents_in_dimensi; + PyObject *__pyx_n_s_id; + PyObject *__pyx_n_s_import; + PyObject *__pyx_n_s_index; + PyObject *__pyx_n_s_indices; + PyObject *__pyx_n_s_initializing; + PyObject *__pyx_n_s_int64; + PyObject *__pyx_n_s_is_coroutine; + PyObject *__pyx_kp_u_isenabled; + PyObject *__pyx_n_s_itemsize; + PyObject *__pyx_kp_s_itemsize_0_for_cython_array; + PyObject *__pyx_n_s_main; + PyObject *__pyx_n_s_max; + PyObject *__pyx_n_s_max_sentences; + PyObject *__pyx_n_s_max_tokens; + PyObject *__pyx_n_s_memview; + PyObject *__pyx_n_s_mode; + PyObject *__pyx_n_s_name; + PyObject *__pyx_n_s_name_2; + PyObject *__pyx_n_s_ndim; + PyObject *__pyx_n_s_new; + PyObject *__pyx_kp_s_no_default___reduce___due_to_non; + PyObject *__pyx_n_s_np; + PyObject *__pyx_n_s_num_tokens_fn; + PyObject *__pyx_n_s_numpy; + PyObject *__pyx_kp_u_numpy__core_multiarray_failed_to; + PyObject *__pyx_kp_u_numpy__core_umath_failed_to_impo; + PyObject *__pyx_n_s_obj; + PyObject *__pyx_n_s_pack; + PyObject *__pyx_n_s_pickle; + PyObject *__pyx_n_s_pyx_PickleError; + PyObject *__pyx_n_s_pyx_checksum; + PyObject *__pyx_n_s_pyx_result; + PyObject *__pyx_n_s_pyx_state; + PyObject *__pyx_n_s_pyx_type; + PyObject *__pyx_n_s_pyx_unpickle_Enum; + PyObject *__pyx_n_s_pyx_vtable; + PyObject *__pyx_n_s_range; + PyObject *__pyx_n_s_reduce; + PyObject *__pyx_n_s_reduce_cython; + PyObject *__pyx_n_s_reduce_ex; + PyObject *__pyx_n_s_register; + PyObject *__pyx_kp_u_sentence_at_index_of_size_exceed; + PyObject *__pyx_n_s_setstate; + PyObject *__pyx_n_s_setstate_cython; + PyObject *__pyx_n_s_shape; + PyObject *__pyx_n_s_size; + PyObject *__pyx_n_s_spec; + PyObject *__pyx_n_s_start; + PyObject *__pyx_n_s_step; + PyObject *__pyx_n_s_stop; + PyObject *__pyx_kp_s_strided_and_direct; + PyObject *__pyx_kp_s_strided_and_direct_or_indirect; + PyObject *__pyx_kp_s_strided_and_indirect; + PyObject *__pyx_kp_s_stringsource; + PyObject *__pyx_n_s_struct; + PyObject *__pyx_n_s_sys; + PyObject *__pyx_n_s_test; + PyObject *__pyx_kp_s_unable_to_allocate_array_data; + PyObject *__pyx_kp_s_unable_to_allocate_shape_and_str; + PyObject *__pyx_n_s_unpack; + PyObject *__pyx_n_s_update; + PyObject *__pyx_n_s_version_info; + PyObject *__pyx_int_0; + PyObject *__pyx_int_1; + PyObject *__pyx_int_3; + PyObject *__pyx_int_112105877; + PyObject *__pyx_int_136983863; + PyObject *__pyx_int_184977713; + PyObject *__pyx_int_neg_1; + PyObject *__pyx_slice__5; + PyObject *__pyx_tuple__4; + PyObject *__pyx_tuple__8; + PyObject *__pyx_tuple__9; + PyObject *__pyx_tuple__10; + PyObject *__pyx_tuple__11; + PyObject *__pyx_tuple__12; + PyObject *__pyx_tuple__13; + PyObject *__pyx_tuple__14; + PyObject *__pyx_tuple__15; + PyObject *__pyx_tuple__16; + PyObject *__pyx_tuple__17; + PyObject *__pyx_tuple__18; + PyObject *__pyx_tuple__19; + PyObject *__pyx_tuple__20; + PyObject *__pyx_tuple__22; + PyObject *__pyx_tuple__24; + PyObject *__pyx_codeobj__21; + PyObject *__pyx_codeobj__23; + PyObject *__pyx_codeobj__25; +} __pyx_mstate; + +#if CYTHON_USE_MODULE_STATE +#ifdef __cplusplus +namespace { + extern struct PyModuleDef __pyx_moduledef; +} /* anonymous namespace */ +#else +static struct PyModuleDef __pyx_moduledef; +#endif + +#define __pyx_mstate(o) ((__pyx_mstate *)__Pyx_PyModule_GetState(o)) + +#define __pyx_mstate_global (__pyx_mstate(PyState_FindModule(&__pyx_moduledef))) + +#define __pyx_m (PyState_FindModule(&__pyx_moduledef)) +#else +static __pyx_mstate __pyx_mstate_global_static = +#ifdef __cplusplus + {}; +#else + {0}; +#endif +static __pyx_mstate *__pyx_mstate_global = &__pyx_mstate_global_static; +#endif +/* #### Code section: module_state_clear ### */ +#if CYTHON_USE_MODULE_STATE +static int __pyx_m_clear(PyObject *m) { + __pyx_mstate *clear_module_state = __pyx_mstate(m); + if (!clear_module_state) return 0; + Py_CLEAR(clear_module_state->__pyx_d); + Py_CLEAR(clear_module_state->__pyx_b); + Py_CLEAR(clear_module_state->__pyx_cython_runtime); + Py_CLEAR(clear_module_state->__pyx_empty_tuple); + Py_CLEAR(clear_module_state->__pyx_empty_bytes); + Py_CLEAR(clear_module_state->__pyx_empty_unicode); + #ifdef __Pyx_CyFunction_USED + Py_CLEAR(clear_module_state->__pyx_CyFunctionType); + #endif + #ifdef __Pyx_FusedFunction_USED + Py_CLEAR(clear_module_state->__pyx_FusedFunctionType); + #endif + Py_CLEAR(clear_module_state->__pyx_ptype_7cpython_4type_type); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_dtype); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_flatiter); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_broadcast); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_ndarray); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_generic); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_number); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_integer); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_signedinteger); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_unsignedinteger); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_inexact); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_floating); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_complexfloating); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_flexible); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_character); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_ufunc); + Py_CLEAR(clear_module_state->__pyx_array_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_array); + Py_CLEAR(clear_module_state->__pyx_MemviewEnum_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_MemviewEnum); + Py_CLEAR(clear_module_state->__pyx_memoryview_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_memoryview); + Py_CLEAR(clear_module_state->__pyx_memoryviewslice_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_memoryviewslice); + Py_CLEAR(clear_module_state->__pyx_kp_u_); + Py_CLEAR(clear_module_state->__pyx_n_s_ASCII); + Py_CLEAR(clear_module_state->__pyx_kp_s_All_dimensions_preceding_dimensi); + Py_CLEAR(clear_module_state->__pyx_n_s_AssertionError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Buffer_view_does_not_expose_stri); + Py_CLEAR(clear_module_state->__pyx_kp_s_Can_only_create_a_buffer_that_is); + Py_CLEAR(clear_module_state->__pyx_kp_s_Cannot_assign_to_read_only_memor); + Py_CLEAR(clear_module_state->__pyx_kp_s_Cannot_create_writable_memory_vi); + Py_CLEAR(clear_module_state->__pyx_kp_u_Cannot_index_with_type); + Py_CLEAR(clear_module_state->__pyx_kp_s_Cannot_transpose_memoryview_with); + Py_CLEAR(clear_module_state->__pyx_n_s_DTYPE); + Py_CLEAR(clear_module_state->__pyx_kp_s_Dimension_d_is_not_direct); + Py_CLEAR(clear_module_state->__pyx_n_s_Ellipsis); + Py_CLEAR(clear_module_state->__pyx_kp_s_Empty_shape_tuple_for_cython_arr); + Py_CLEAR(clear_module_state->__pyx_n_s_ImportError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0); + Py_CLEAR(clear_module_state->__pyx_n_s_IndexError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Index_out_of_bounds_axis_d); + Py_CLEAR(clear_module_state->__pyx_kp_s_Indirect_dimensions_not_supporte); + Py_CLEAR(clear_module_state->__pyx_kp_u_Invalid_mode_expected_c_or_fortr); + Py_CLEAR(clear_module_state->__pyx_kp_u_Invalid_shape_in_axis); + Py_CLEAR(clear_module_state->__pyx_n_s_MemoryError); + Py_CLEAR(clear_module_state->__pyx_kp_s_MemoryView_of_r_at_0x_x); + Py_CLEAR(clear_module_state->__pyx_kp_s_MemoryView_of_r_object); + Py_CLEAR(clear_module_state->__pyx_n_b_O); + Py_CLEAR(clear_module_state->__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + Py_CLEAR(clear_module_state->__pyx_n_s_PickleError); + Py_CLEAR(clear_module_state->__pyx_n_s_Sequence); + Py_CLEAR(clear_module_state->__pyx_kp_s_Step_may_not_be_zero_axis_d); + Py_CLEAR(clear_module_state->__pyx_n_s_TypeError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Unable_to_convert_item_to_object); + Py_CLEAR(clear_module_state->__pyx_n_s_ValueError); + Py_CLEAR(clear_module_state->__pyx_n_s_View_MemoryView); + Py_CLEAR(clear_module_state->__pyx_kp_u__2); + Py_CLEAR(clear_module_state->__pyx_n_s__26); + Py_CLEAR(clear_module_state->__pyx_n_s__3); + Py_CLEAR(clear_module_state->__pyx_kp_u__6); + Py_CLEAR(clear_module_state->__pyx_kp_u__7); + Py_CLEAR(clear_module_state->__pyx_n_s_abc); + Py_CLEAR(clear_module_state->__pyx_n_s_allocate_buffer); + Py_CLEAR(clear_module_state->__pyx_kp_u_and); + Py_CLEAR(clear_module_state->__pyx_n_s_asyncio_coroutines); + Py_CLEAR(clear_module_state->__pyx_n_s_base); + Py_CLEAR(clear_module_state->__pyx_n_s_batch_by_size_fast); + Py_CLEAR(clear_module_state->__pyx_n_s_batch_fixed_shapes_fast); + Py_CLEAR(clear_module_state->__pyx_n_s_bsz_mult); + Py_CLEAR(clear_module_state->__pyx_n_s_c); + Py_CLEAR(clear_module_state->__pyx_n_u_c); + Py_CLEAR(clear_module_state->__pyx_n_s_class); + Py_CLEAR(clear_module_state->__pyx_n_s_class_getitem); + Py_CLEAR(clear_module_state->__pyx_n_s_cline_in_traceback); + Py_CLEAR(clear_module_state->__pyx_n_s_collections); + Py_CLEAR(clear_module_state->__pyx_kp_s_collections_abc); + Py_CLEAR(clear_module_state->__pyx_kp_s_contiguous_and_direct); + Py_CLEAR(clear_module_state->__pyx_kp_s_contiguous_and_indirect); + Py_CLEAR(clear_module_state->__pyx_n_s_count); + Py_CLEAR(clear_module_state->__pyx_n_s_dict); + Py_CLEAR(clear_module_state->__pyx_kp_u_disable); + Py_CLEAR(clear_module_state->__pyx_n_s_dtype_is_object); + Py_CLEAR(clear_module_state->__pyx_kp_u_enable); + Py_CLEAR(clear_module_state->__pyx_n_s_encode); + Py_CLEAR(clear_module_state->__pyx_n_s_enumerate); + Py_CLEAR(clear_module_state->__pyx_n_s_error); + Py_CLEAR(clear_module_state->__pyx_n_s_fairseq_data_data_utils_fast); + Py_CLEAR(clear_module_state->__pyx_kp_s_fairseq_data_data_utils_fast_pyx); + Py_CLEAR(clear_module_state->__pyx_n_s_fixed_shapes_sorted); + Py_CLEAR(clear_module_state->__pyx_n_s_flags); + Py_CLEAR(clear_module_state->__pyx_n_s_format); + Py_CLEAR(clear_module_state->__pyx_n_s_fortran); + Py_CLEAR(clear_module_state->__pyx_n_u_fortran); + Py_CLEAR(clear_module_state->__pyx_kp_u_gc); + Py_CLEAR(clear_module_state->__pyx_n_s_getstate); + Py_CLEAR(clear_module_state->__pyx_kp_u_got); + Py_CLEAR(clear_module_state->__pyx_kp_u_got_differing_extents_in_dimensi); + Py_CLEAR(clear_module_state->__pyx_n_s_id); + Py_CLEAR(clear_module_state->__pyx_n_s_import); + Py_CLEAR(clear_module_state->__pyx_n_s_index); + Py_CLEAR(clear_module_state->__pyx_n_s_indices); + Py_CLEAR(clear_module_state->__pyx_n_s_initializing); + Py_CLEAR(clear_module_state->__pyx_n_s_int64); + Py_CLEAR(clear_module_state->__pyx_n_s_is_coroutine); + Py_CLEAR(clear_module_state->__pyx_kp_u_isenabled); + Py_CLEAR(clear_module_state->__pyx_n_s_itemsize); + Py_CLEAR(clear_module_state->__pyx_kp_s_itemsize_0_for_cython_array); + Py_CLEAR(clear_module_state->__pyx_n_s_main); + Py_CLEAR(clear_module_state->__pyx_n_s_max); + Py_CLEAR(clear_module_state->__pyx_n_s_max_sentences); + Py_CLEAR(clear_module_state->__pyx_n_s_max_tokens); + Py_CLEAR(clear_module_state->__pyx_n_s_memview); + Py_CLEAR(clear_module_state->__pyx_n_s_mode); + Py_CLEAR(clear_module_state->__pyx_n_s_name); + Py_CLEAR(clear_module_state->__pyx_n_s_name_2); + Py_CLEAR(clear_module_state->__pyx_n_s_ndim); + Py_CLEAR(clear_module_state->__pyx_n_s_new); + Py_CLEAR(clear_module_state->__pyx_kp_s_no_default___reduce___due_to_non); + Py_CLEAR(clear_module_state->__pyx_n_s_np); + Py_CLEAR(clear_module_state->__pyx_n_s_num_tokens_fn); + Py_CLEAR(clear_module_state->__pyx_n_s_numpy); + Py_CLEAR(clear_module_state->__pyx_kp_u_numpy__core_multiarray_failed_to); + Py_CLEAR(clear_module_state->__pyx_kp_u_numpy__core_umath_failed_to_impo); + Py_CLEAR(clear_module_state->__pyx_n_s_obj); + Py_CLEAR(clear_module_state->__pyx_n_s_pack); + Py_CLEAR(clear_module_state->__pyx_n_s_pickle); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_PickleError); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_checksum); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_result); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_state); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_type); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_unpickle_Enum); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_vtable); + Py_CLEAR(clear_module_state->__pyx_n_s_range); + Py_CLEAR(clear_module_state->__pyx_n_s_reduce); + Py_CLEAR(clear_module_state->__pyx_n_s_reduce_cython); + Py_CLEAR(clear_module_state->__pyx_n_s_reduce_ex); + Py_CLEAR(clear_module_state->__pyx_n_s_register); + Py_CLEAR(clear_module_state->__pyx_kp_u_sentence_at_index_of_size_exceed); + Py_CLEAR(clear_module_state->__pyx_n_s_setstate); + Py_CLEAR(clear_module_state->__pyx_n_s_setstate_cython); + Py_CLEAR(clear_module_state->__pyx_n_s_shape); + Py_CLEAR(clear_module_state->__pyx_n_s_size); + Py_CLEAR(clear_module_state->__pyx_n_s_spec); + Py_CLEAR(clear_module_state->__pyx_n_s_start); + Py_CLEAR(clear_module_state->__pyx_n_s_step); + Py_CLEAR(clear_module_state->__pyx_n_s_stop); + Py_CLEAR(clear_module_state->__pyx_kp_s_strided_and_direct); + Py_CLEAR(clear_module_state->__pyx_kp_s_strided_and_direct_or_indirect); + Py_CLEAR(clear_module_state->__pyx_kp_s_strided_and_indirect); + Py_CLEAR(clear_module_state->__pyx_kp_s_stringsource); + Py_CLEAR(clear_module_state->__pyx_n_s_struct); + Py_CLEAR(clear_module_state->__pyx_n_s_sys); + Py_CLEAR(clear_module_state->__pyx_n_s_test); + Py_CLEAR(clear_module_state->__pyx_kp_s_unable_to_allocate_array_data); + Py_CLEAR(clear_module_state->__pyx_kp_s_unable_to_allocate_shape_and_str); + Py_CLEAR(clear_module_state->__pyx_n_s_unpack); + Py_CLEAR(clear_module_state->__pyx_n_s_update); + Py_CLEAR(clear_module_state->__pyx_n_s_version_info); + Py_CLEAR(clear_module_state->__pyx_int_0); + Py_CLEAR(clear_module_state->__pyx_int_1); + Py_CLEAR(clear_module_state->__pyx_int_3); + Py_CLEAR(clear_module_state->__pyx_int_112105877); + Py_CLEAR(clear_module_state->__pyx_int_136983863); + Py_CLEAR(clear_module_state->__pyx_int_184977713); + Py_CLEAR(clear_module_state->__pyx_int_neg_1); + Py_CLEAR(clear_module_state->__pyx_slice__5); + Py_CLEAR(clear_module_state->__pyx_tuple__4); + Py_CLEAR(clear_module_state->__pyx_tuple__8); + Py_CLEAR(clear_module_state->__pyx_tuple__9); + Py_CLEAR(clear_module_state->__pyx_tuple__10); + Py_CLEAR(clear_module_state->__pyx_tuple__11); + Py_CLEAR(clear_module_state->__pyx_tuple__12); + Py_CLEAR(clear_module_state->__pyx_tuple__13); + Py_CLEAR(clear_module_state->__pyx_tuple__14); + Py_CLEAR(clear_module_state->__pyx_tuple__15); + Py_CLEAR(clear_module_state->__pyx_tuple__16); + Py_CLEAR(clear_module_state->__pyx_tuple__17); + Py_CLEAR(clear_module_state->__pyx_tuple__18); + Py_CLEAR(clear_module_state->__pyx_tuple__19); + Py_CLEAR(clear_module_state->__pyx_tuple__20); + Py_CLEAR(clear_module_state->__pyx_tuple__22); + Py_CLEAR(clear_module_state->__pyx_tuple__24); + Py_CLEAR(clear_module_state->__pyx_codeobj__21); + Py_CLEAR(clear_module_state->__pyx_codeobj__23); + Py_CLEAR(clear_module_state->__pyx_codeobj__25); + return 0; +} +#endif +/* #### Code section: module_state_traverse ### */ +#if CYTHON_USE_MODULE_STATE +static int __pyx_m_traverse(PyObject *m, visitproc visit, void *arg) { + __pyx_mstate *traverse_module_state = __pyx_mstate(m); + if (!traverse_module_state) return 0; + Py_VISIT(traverse_module_state->__pyx_d); + Py_VISIT(traverse_module_state->__pyx_b); + Py_VISIT(traverse_module_state->__pyx_cython_runtime); + Py_VISIT(traverse_module_state->__pyx_empty_tuple); + Py_VISIT(traverse_module_state->__pyx_empty_bytes); + Py_VISIT(traverse_module_state->__pyx_empty_unicode); + #ifdef __Pyx_CyFunction_USED + Py_VISIT(traverse_module_state->__pyx_CyFunctionType); + #endif + #ifdef __Pyx_FusedFunction_USED + Py_VISIT(traverse_module_state->__pyx_FusedFunctionType); + #endif + Py_VISIT(traverse_module_state->__pyx_ptype_7cpython_4type_type); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_dtype); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_flatiter); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_broadcast); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_ndarray); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_generic); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_number); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_integer); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_signedinteger); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_unsignedinteger); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_inexact); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_floating); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_complexfloating); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_flexible); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_character); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_ufunc); + Py_VISIT(traverse_module_state->__pyx_array_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_array); + Py_VISIT(traverse_module_state->__pyx_MemviewEnum_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_MemviewEnum); + Py_VISIT(traverse_module_state->__pyx_memoryview_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_memoryview); + Py_VISIT(traverse_module_state->__pyx_memoryviewslice_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_memoryviewslice); + Py_VISIT(traverse_module_state->__pyx_kp_u_); + Py_VISIT(traverse_module_state->__pyx_n_s_ASCII); + Py_VISIT(traverse_module_state->__pyx_kp_s_All_dimensions_preceding_dimensi); + Py_VISIT(traverse_module_state->__pyx_n_s_AssertionError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Buffer_view_does_not_expose_stri); + Py_VISIT(traverse_module_state->__pyx_kp_s_Can_only_create_a_buffer_that_is); + Py_VISIT(traverse_module_state->__pyx_kp_s_Cannot_assign_to_read_only_memor); + Py_VISIT(traverse_module_state->__pyx_kp_s_Cannot_create_writable_memory_vi); + Py_VISIT(traverse_module_state->__pyx_kp_u_Cannot_index_with_type); + Py_VISIT(traverse_module_state->__pyx_kp_s_Cannot_transpose_memoryview_with); + Py_VISIT(traverse_module_state->__pyx_n_s_DTYPE); + Py_VISIT(traverse_module_state->__pyx_kp_s_Dimension_d_is_not_direct); + Py_VISIT(traverse_module_state->__pyx_n_s_Ellipsis); + Py_VISIT(traverse_module_state->__pyx_kp_s_Empty_shape_tuple_for_cython_arr); + Py_VISIT(traverse_module_state->__pyx_n_s_ImportError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0); + Py_VISIT(traverse_module_state->__pyx_n_s_IndexError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Index_out_of_bounds_axis_d); + Py_VISIT(traverse_module_state->__pyx_kp_s_Indirect_dimensions_not_supporte); + Py_VISIT(traverse_module_state->__pyx_kp_u_Invalid_mode_expected_c_or_fortr); + Py_VISIT(traverse_module_state->__pyx_kp_u_Invalid_shape_in_axis); + Py_VISIT(traverse_module_state->__pyx_n_s_MemoryError); + Py_VISIT(traverse_module_state->__pyx_kp_s_MemoryView_of_r_at_0x_x); + Py_VISIT(traverse_module_state->__pyx_kp_s_MemoryView_of_r_object); + Py_VISIT(traverse_module_state->__pyx_n_b_O); + Py_VISIT(traverse_module_state->__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + Py_VISIT(traverse_module_state->__pyx_n_s_PickleError); + Py_VISIT(traverse_module_state->__pyx_n_s_Sequence); + Py_VISIT(traverse_module_state->__pyx_kp_s_Step_may_not_be_zero_axis_d); + Py_VISIT(traverse_module_state->__pyx_n_s_TypeError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Unable_to_convert_item_to_object); + Py_VISIT(traverse_module_state->__pyx_n_s_ValueError); + Py_VISIT(traverse_module_state->__pyx_n_s_View_MemoryView); + Py_VISIT(traverse_module_state->__pyx_kp_u__2); + Py_VISIT(traverse_module_state->__pyx_n_s__26); + Py_VISIT(traverse_module_state->__pyx_n_s__3); + Py_VISIT(traverse_module_state->__pyx_kp_u__6); + Py_VISIT(traverse_module_state->__pyx_kp_u__7); + Py_VISIT(traverse_module_state->__pyx_n_s_abc); + Py_VISIT(traverse_module_state->__pyx_n_s_allocate_buffer); + Py_VISIT(traverse_module_state->__pyx_kp_u_and); + Py_VISIT(traverse_module_state->__pyx_n_s_asyncio_coroutines); + Py_VISIT(traverse_module_state->__pyx_n_s_base); + Py_VISIT(traverse_module_state->__pyx_n_s_batch_by_size_fast); + Py_VISIT(traverse_module_state->__pyx_n_s_batch_fixed_shapes_fast); + Py_VISIT(traverse_module_state->__pyx_n_s_bsz_mult); + Py_VISIT(traverse_module_state->__pyx_n_s_c); + Py_VISIT(traverse_module_state->__pyx_n_u_c); + Py_VISIT(traverse_module_state->__pyx_n_s_class); + Py_VISIT(traverse_module_state->__pyx_n_s_class_getitem); + Py_VISIT(traverse_module_state->__pyx_n_s_cline_in_traceback); + Py_VISIT(traverse_module_state->__pyx_n_s_collections); + Py_VISIT(traverse_module_state->__pyx_kp_s_collections_abc); + Py_VISIT(traverse_module_state->__pyx_kp_s_contiguous_and_direct); + Py_VISIT(traverse_module_state->__pyx_kp_s_contiguous_and_indirect); + Py_VISIT(traverse_module_state->__pyx_n_s_count); + Py_VISIT(traverse_module_state->__pyx_n_s_dict); + Py_VISIT(traverse_module_state->__pyx_kp_u_disable); + Py_VISIT(traverse_module_state->__pyx_n_s_dtype_is_object); + Py_VISIT(traverse_module_state->__pyx_kp_u_enable); + Py_VISIT(traverse_module_state->__pyx_n_s_encode); + Py_VISIT(traverse_module_state->__pyx_n_s_enumerate); + Py_VISIT(traverse_module_state->__pyx_n_s_error); + Py_VISIT(traverse_module_state->__pyx_n_s_fairseq_data_data_utils_fast); + Py_VISIT(traverse_module_state->__pyx_kp_s_fairseq_data_data_utils_fast_pyx); + Py_VISIT(traverse_module_state->__pyx_n_s_fixed_shapes_sorted); + Py_VISIT(traverse_module_state->__pyx_n_s_flags); + Py_VISIT(traverse_module_state->__pyx_n_s_format); + Py_VISIT(traverse_module_state->__pyx_n_s_fortran); + Py_VISIT(traverse_module_state->__pyx_n_u_fortran); + Py_VISIT(traverse_module_state->__pyx_kp_u_gc); + Py_VISIT(traverse_module_state->__pyx_n_s_getstate); + Py_VISIT(traverse_module_state->__pyx_kp_u_got); + Py_VISIT(traverse_module_state->__pyx_kp_u_got_differing_extents_in_dimensi); + Py_VISIT(traverse_module_state->__pyx_n_s_id); + Py_VISIT(traverse_module_state->__pyx_n_s_import); + Py_VISIT(traverse_module_state->__pyx_n_s_index); + Py_VISIT(traverse_module_state->__pyx_n_s_indices); + Py_VISIT(traverse_module_state->__pyx_n_s_initializing); + Py_VISIT(traverse_module_state->__pyx_n_s_int64); + Py_VISIT(traverse_module_state->__pyx_n_s_is_coroutine); + Py_VISIT(traverse_module_state->__pyx_kp_u_isenabled); + Py_VISIT(traverse_module_state->__pyx_n_s_itemsize); + Py_VISIT(traverse_module_state->__pyx_kp_s_itemsize_0_for_cython_array); + Py_VISIT(traverse_module_state->__pyx_n_s_main); + Py_VISIT(traverse_module_state->__pyx_n_s_max); + Py_VISIT(traverse_module_state->__pyx_n_s_max_sentences); + Py_VISIT(traverse_module_state->__pyx_n_s_max_tokens); + Py_VISIT(traverse_module_state->__pyx_n_s_memview); + Py_VISIT(traverse_module_state->__pyx_n_s_mode); + Py_VISIT(traverse_module_state->__pyx_n_s_name); + Py_VISIT(traverse_module_state->__pyx_n_s_name_2); + Py_VISIT(traverse_module_state->__pyx_n_s_ndim); + Py_VISIT(traverse_module_state->__pyx_n_s_new); + Py_VISIT(traverse_module_state->__pyx_kp_s_no_default___reduce___due_to_non); + Py_VISIT(traverse_module_state->__pyx_n_s_np); + Py_VISIT(traverse_module_state->__pyx_n_s_num_tokens_fn); + Py_VISIT(traverse_module_state->__pyx_n_s_numpy); + Py_VISIT(traverse_module_state->__pyx_kp_u_numpy__core_multiarray_failed_to); + Py_VISIT(traverse_module_state->__pyx_kp_u_numpy__core_umath_failed_to_impo); + Py_VISIT(traverse_module_state->__pyx_n_s_obj); + Py_VISIT(traverse_module_state->__pyx_n_s_pack); + Py_VISIT(traverse_module_state->__pyx_n_s_pickle); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_PickleError); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_checksum); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_result); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_state); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_type); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_unpickle_Enum); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_vtable); + Py_VISIT(traverse_module_state->__pyx_n_s_range); + Py_VISIT(traverse_module_state->__pyx_n_s_reduce); + Py_VISIT(traverse_module_state->__pyx_n_s_reduce_cython); + Py_VISIT(traverse_module_state->__pyx_n_s_reduce_ex); + Py_VISIT(traverse_module_state->__pyx_n_s_register); + Py_VISIT(traverse_module_state->__pyx_kp_u_sentence_at_index_of_size_exceed); + Py_VISIT(traverse_module_state->__pyx_n_s_setstate); + Py_VISIT(traverse_module_state->__pyx_n_s_setstate_cython); + Py_VISIT(traverse_module_state->__pyx_n_s_shape); + Py_VISIT(traverse_module_state->__pyx_n_s_size); + Py_VISIT(traverse_module_state->__pyx_n_s_spec); + Py_VISIT(traverse_module_state->__pyx_n_s_start); + Py_VISIT(traverse_module_state->__pyx_n_s_step); + Py_VISIT(traverse_module_state->__pyx_n_s_stop); + Py_VISIT(traverse_module_state->__pyx_kp_s_strided_and_direct); + Py_VISIT(traverse_module_state->__pyx_kp_s_strided_and_direct_or_indirect); + Py_VISIT(traverse_module_state->__pyx_kp_s_strided_and_indirect); + Py_VISIT(traverse_module_state->__pyx_kp_s_stringsource); + Py_VISIT(traverse_module_state->__pyx_n_s_struct); + Py_VISIT(traverse_module_state->__pyx_n_s_sys); + Py_VISIT(traverse_module_state->__pyx_n_s_test); + Py_VISIT(traverse_module_state->__pyx_kp_s_unable_to_allocate_array_data); + Py_VISIT(traverse_module_state->__pyx_kp_s_unable_to_allocate_shape_and_str); + Py_VISIT(traverse_module_state->__pyx_n_s_unpack); + Py_VISIT(traverse_module_state->__pyx_n_s_update); + Py_VISIT(traverse_module_state->__pyx_n_s_version_info); + Py_VISIT(traverse_module_state->__pyx_int_0); + Py_VISIT(traverse_module_state->__pyx_int_1); + Py_VISIT(traverse_module_state->__pyx_int_3); + Py_VISIT(traverse_module_state->__pyx_int_112105877); + Py_VISIT(traverse_module_state->__pyx_int_136983863); + Py_VISIT(traverse_module_state->__pyx_int_184977713); + Py_VISIT(traverse_module_state->__pyx_int_neg_1); + Py_VISIT(traverse_module_state->__pyx_slice__5); + Py_VISIT(traverse_module_state->__pyx_tuple__4); + Py_VISIT(traverse_module_state->__pyx_tuple__8); + Py_VISIT(traverse_module_state->__pyx_tuple__9); + Py_VISIT(traverse_module_state->__pyx_tuple__10); + Py_VISIT(traverse_module_state->__pyx_tuple__11); + Py_VISIT(traverse_module_state->__pyx_tuple__12); + Py_VISIT(traverse_module_state->__pyx_tuple__13); + Py_VISIT(traverse_module_state->__pyx_tuple__14); + Py_VISIT(traverse_module_state->__pyx_tuple__15); + Py_VISIT(traverse_module_state->__pyx_tuple__16); + Py_VISIT(traverse_module_state->__pyx_tuple__17); + Py_VISIT(traverse_module_state->__pyx_tuple__18); + Py_VISIT(traverse_module_state->__pyx_tuple__19); + Py_VISIT(traverse_module_state->__pyx_tuple__20); + Py_VISIT(traverse_module_state->__pyx_tuple__22); + Py_VISIT(traverse_module_state->__pyx_tuple__24); + Py_VISIT(traverse_module_state->__pyx_codeobj__21); + Py_VISIT(traverse_module_state->__pyx_codeobj__23); + Py_VISIT(traverse_module_state->__pyx_codeobj__25); + return 0; +} +#endif +/* #### Code section: module_state_defines ### */ +#define __pyx_d __pyx_mstate_global->__pyx_d +#define __pyx_b __pyx_mstate_global->__pyx_b +#define __pyx_cython_runtime __pyx_mstate_global->__pyx_cython_runtime +#define __pyx_empty_tuple __pyx_mstate_global->__pyx_empty_tuple +#define __pyx_empty_bytes __pyx_mstate_global->__pyx_empty_bytes +#define __pyx_empty_unicode __pyx_mstate_global->__pyx_empty_unicode +#ifdef __Pyx_CyFunction_USED +#define __pyx_CyFunctionType __pyx_mstate_global->__pyx_CyFunctionType +#endif +#ifdef __Pyx_FusedFunction_USED +#define __pyx_FusedFunctionType __pyx_mstate_global->__pyx_FusedFunctionType +#endif +#ifdef __Pyx_Generator_USED +#define __pyx_GeneratorType __pyx_mstate_global->__pyx_GeneratorType +#endif +#ifdef __Pyx_IterableCoroutine_USED +#define __pyx_IterableCoroutineType __pyx_mstate_global->__pyx_IterableCoroutineType +#endif +#ifdef __Pyx_Coroutine_USED +#define __pyx_CoroutineAwaitType __pyx_mstate_global->__pyx_CoroutineAwaitType +#endif +#ifdef __Pyx_Coroutine_USED +#define __pyx_CoroutineType __pyx_mstate_global->__pyx_CoroutineType +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#define __pyx_ptype_7cpython_4type_type __pyx_mstate_global->__pyx_ptype_7cpython_4type_type +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#define __pyx_ptype_5numpy_dtype __pyx_mstate_global->__pyx_ptype_5numpy_dtype +#define __pyx_ptype_5numpy_flatiter __pyx_mstate_global->__pyx_ptype_5numpy_flatiter +#define __pyx_ptype_5numpy_broadcast __pyx_mstate_global->__pyx_ptype_5numpy_broadcast +#define __pyx_ptype_5numpy_ndarray __pyx_mstate_global->__pyx_ptype_5numpy_ndarray +#define __pyx_ptype_5numpy_generic __pyx_mstate_global->__pyx_ptype_5numpy_generic +#define __pyx_ptype_5numpy_number __pyx_mstate_global->__pyx_ptype_5numpy_number +#define __pyx_ptype_5numpy_integer __pyx_mstate_global->__pyx_ptype_5numpy_integer +#define __pyx_ptype_5numpy_signedinteger __pyx_mstate_global->__pyx_ptype_5numpy_signedinteger +#define __pyx_ptype_5numpy_unsignedinteger __pyx_mstate_global->__pyx_ptype_5numpy_unsignedinteger +#define __pyx_ptype_5numpy_inexact __pyx_mstate_global->__pyx_ptype_5numpy_inexact +#define __pyx_ptype_5numpy_floating __pyx_mstate_global->__pyx_ptype_5numpy_floating +#define __pyx_ptype_5numpy_complexfloating __pyx_mstate_global->__pyx_ptype_5numpy_complexfloating +#define __pyx_ptype_5numpy_flexible __pyx_mstate_global->__pyx_ptype_5numpy_flexible +#define __pyx_ptype_5numpy_character __pyx_mstate_global->__pyx_ptype_5numpy_character +#define __pyx_ptype_5numpy_ufunc __pyx_mstate_global->__pyx_ptype_5numpy_ufunc +#if CYTHON_USE_MODULE_STATE +#define __pyx_type___pyx_array __pyx_mstate_global->__pyx_type___pyx_array +#define __pyx_type___pyx_MemviewEnum __pyx_mstate_global->__pyx_type___pyx_MemviewEnum +#define __pyx_type___pyx_memoryview __pyx_mstate_global->__pyx_type___pyx_memoryview +#define __pyx_type___pyx_memoryviewslice __pyx_mstate_global->__pyx_type___pyx_memoryviewslice +#endif +#define __pyx_array_type __pyx_mstate_global->__pyx_array_type +#define __pyx_MemviewEnum_type __pyx_mstate_global->__pyx_MemviewEnum_type +#define __pyx_memoryview_type __pyx_mstate_global->__pyx_memoryview_type +#define __pyx_memoryviewslice_type __pyx_mstate_global->__pyx_memoryviewslice_type +#define __pyx_kp_u_ __pyx_mstate_global->__pyx_kp_u_ +#define __pyx_n_s_ASCII __pyx_mstate_global->__pyx_n_s_ASCII +#define __pyx_kp_s_All_dimensions_preceding_dimensi __pyx_mstate_global->__pyx_kp_s_All_dimensions_preceding_dimensi +#define __pyx_n_s_AssertionError __pyx_mstate_global->__pyx_n_s_AssertionError +#define __pyx_kp_s_Buffer_view_does_not_expose_stri __pyx_mstate_global->__pyx_kp_s_Buffer_view_does_not_expose_stri +#define __pyx_kp_s_Can_only_create_a_buffer_that_is __pyx_mstate_global->__pyx_kp_s_Can_only_create_a_buffer_that_is +#define __pyx_kp_s_Cannot_assign_to_read_only_memor __pyx_mstate_global->__pyx_kp_s_Cannot_assign_to_read_only_memor +#define __pyx_kp_s_Cannot_create_writable_memory_vi __pyx_mstate_global->__pyx_kp_s_Cannot_create_writable_memory_vi +#define __pyx_kp_u_Cannot_index_with_type __pyx_mstate_global->__pyx_kp_u_Cannot_index_with_type +#define __pyx_kp_s_Cannot_transpose_memoryview_with __pyx_mstate_global->__pyx_kp_s_Cannot_transpose_memoryview_with +#define __pyx_n_s_DTYPE __pyx_mstate_global->__pyx_n_s_DTYPE +#define __pyx_kp_s_Dimension_d_is_not_direct __pyx_mstate_global->__pyx_kp_s_Dimension_d_is_not_direct +#define __pyx_n_s_Ellipsis __pyx_mstate_global->__pyx_n_s_Ellipsis +#define __pyx_kp_s_Empty_shape_tuple_for_cython_arr __pyx_mstate_global->__pyx_kp_s_Empty_shape_tuple_for_cython_arr +#define __pyx_n_s_ImportError __pyx_mstate_global->__pyx_n_s_ImportError +#define __pyx_kp_s_Incompatible_checksums_0x_x_vs_0 __pyx_mstate_global->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0 +#define __pyx_n_s_IndexError __pyx_mstate_global->__pyx_n_s_IndexError +#define __pyx_kp_s_Index_out_of_bounds_axis_d __pyx_mstate_global->__pyx_kp_s_Index_out_of_bounds_axis_d +#define __pyx_kp_s_Indirect_dimensions_not_supporte __pyx_mstate_global->__pyx_kp_s_Indirect_dimensions_not_supporte +#define __pyx_kp_u_Invalid_mode_expected_c_or_fortr __pyx_mstate_global->__pyx_kp_u_Invalid_mode_expected_c_or_fortr +#define __pyx_kp_u_Invalid_shape_in_axis __pyx_mstate_global->__pyx_kp_u_Invalid_shape_in_axis +#define __pyx_n_s_MemoryError __pyx_mstate_global->__pyx_n_s_MemoryError +#define __pyx_kp_s_MemoryView_of_r_at_0x_x __pyx_mstate_global->__pyx_kp_s_MemoryView_of_r_at_0x_x +#define __pyx_kp_s_MemoryView_of_r_object __pyx_mstate_global->__pyx_kp_s_MemoryView_of_r_object +#define __pyx_n_b_O __pyx_mstate_global->__pyx_n_b_O +#define __pyx_kp_u_Out_of_bounds_on_buffer_access_a __pyx_mstate_global->__pyx_kp_u_Out_of_bounds_on_buffer_access_a +#define __pyx_n_s_PickleError __pyx_mstate_global->__pyx_n_s_PickleError +#define __pyx_n_s_Sequence __pyx_mstate_global->__pyx_n_s_Sequence +#define __pyx_kp_s_Step_may_not_be_zero_axis_d __pyx_mstate_global->__pyx_kp_s_Step_may_not_be_zero_axis_d +#define __pyx_n_s_TypeError __pyx_mstate_global->__pyx_n_s_TypeError +#define __pyx_kp_s_Unable_to_convert_item_to_object __pyx_mstate_global->__pyx_kp_s_Unable_to_convert_item_to_object +#define __pyx_n_s_ValueError __pyx_mstate_global->__pyx_n_s_ValueError +#define __pyx_n_s_View_MemoryView __pyx_mstate_global->__pyx_n_s_View_MemoryView +#define __pyx_kp_u__2 __pyx_mstate_global->__pyx_kp_u__2 +#define __pyx_n_s__26 __pyx_mstate_global->__pyx_n_s__26 +#define __pyx_n_s__3 __pyx_mstate_global->__pyx_n_s__3 +#define __pyx_kp_u__6 __pyx_mstate_global->__pyx_kp_u__6 +#define __pyx_kp_u__7 __pyx_mstate_global->__pyx_kp_u__7 +#define __pyx_n_s_abc __pyx_mstate_global->__pyx_n_s_abc +#define __pyx_n_s_allocate_buffer __pyx_mstate_global->__pyx_n_s_allocate_buffer +#define __pyx_kp_u_and __pyx_mstate_global->__pyx_kp_u_and +#define __pyx_n_s_asyncio_coroutines __pyx_mstate_global->__pyx_n_s_asyncio_coroutines +#define __pyx_n_s_base __pyx_mstate_global->__pyx_n_s_base +#define __pyx_n_s_batch_by_size_fast __pyx_mstate_global->__pyx_n_s_batch_by_size_fast +#define __pyx_n_s_batch_fixed_shapes_fast __pyx_mstate_global->__pyx_n_s_batch_fixed_shapes_fast +#define __pyx_n_s_bsz_mult __pyx_mstate_global->__pyx_n_s_bsz_mult +#define __pyx_n_s_c __pyx_mstate_global->__pyx_n_s_c +#define __pyx_n_u_c __pyx_mstate_global->__pyx_n_u_c +#define __pyx_n_s_class __pyx_mstate_global->__pyx_n_s_class +#define __pyx_n_s_class_getitem __pyx_mstate_global->__pyx_n_s_class_getitem +#define __pyx_n_s_cline_in_traceback __pyx_mstate_global->__pyx_n_s_cline_in_traceback +#define __pyx_n_s_collections __pyx_mstate_global->__pyx_n_s_collections +#define __pyx_kp_s_collections_abc __pyx_mstate_global->__pyx_kp_s_collections_abc +#define __pyx_kp_s_contiguous_and_direct __pyx_mstate_global->__pyx_kp_s_contiguous_and_direct +#define __pyx_kp_s_contiguous_and_indirect __pyx_mstate_global->__pyx_kp_s_contiguous_and_indirect +#define __pyx_n_s_count __pyx_mstate_global->__pyx_n_s_count +#define __pyx_n_s_dict __pyx_mstate_global->__pyx_n_s_dict +#define __pyx_kp_u_disable __pyx_mstate_global->__pyx_kp_u_disable +#define __pyx_n_s_dtype_is_object __pyx_mstate_global->__pyx_n_s_dtype_is_object +#define __pyx_kp_u_enable __pyx_mstate_global->__pyx_kp_u_enable +#define __pyx_n_s_encode __pyx_mstate_global->__pyx_n_s_encode +#define __pyx_n_s_enumerate __pyx_mstate_global->__pyx_n_s_enumerate +#define __pyx_n_s_error __pyx_mstate_global->__pyx_n_s_error +#define __pyx_n_s_fairseq_data_data_utils_fast __pyx_mstate_global->__pyx_n_s_fairseq_data_data_utils_fast +#define __pyx_kp_s_fairseq_data_data_utils_fast_pyx __pyx_mstate_global->__pyx_kp_s_fairseq_data_data_utils_fast_pyx +#define __pyx_n_s_fixed_shapes_sorted __pyx_mstate_global->__pyx_n_s_fixed_shapes_sorted +#define __pyx_n_s_flags __pyx_mstate_global->__pyx_n_s_flags +#define __pyx_n_s_format __pyx_mstate_global->__pyx_n_s_format +#define __pyx_n_s_fortran __pyx_mstate_global->__pyx_n_s_fortran +#define __pyx_n_u_fortran __pyx_mstate_global->__pyx_n_u_fortran +#define __pyx_kp_u_gc __pyx_mstate_global->__pyx_kp_u_gc +#define __pyx_n_s_getstate __pyx_mstate_global->__pyx_n_s_getstate +#define __pyx_kp_u_got __pyx_mstate_global->__pyx_kp_u_got +#define __pyx_kp_u_got_differing_extents_in_dimensi __pyx_mstate_global->__pyx_kp_u_got_differing_extents_in_dimensi +#define __pyx_n_s_id __pyx_mstate_global->__pyx_n_s_id +#define __pyx_n_s_import __pyx_mstate_global->__pyx_n_s_import +#define __pyx_n_s_index __pyx_mstate_global->__pyx_n_s_index +#define __pyx_n_s_indices __pyx_mstate_global->__pyx_n_s_indices +#define __pyx_n_s_initializing __pyx_mstate_global->__pyx_n_s_initializing +#define __pyx_n_s_int64 __pyx_mstate_global->__pyx_n_s_int64 +#define __pyx_n_s_is_coroutine __pyx_mstate_global->__pyx_n_s_is_coroutine +#define __pyx_kp_u_isenabled __pyx_mstate_global->__pyx_kp_u_isenabled +#define __pyx_n_s_itemsize __pyx_mstate_global->__pyx_n_s_itemsize +#define __pyx_kp_s_itemsize_0_for_cython_array __pyx_mstate_global->__pyx_kp_s_itemsize_0_for_cython_array +#define __pyx_n_s_main __pyx_mstate_global->__pyx_n_s_main +#define __pyx_n_s_max __pyx_mstate_global->__pyx_n_s_max +#define __pyx_n_s_max_sentences __pyx_mstate_global->__pyx_n_s_max_sentences +#define __pyx_n_s_max_tokens __pyx_mstate_global->__pyx_n_s_max_tokens +#define __pyx_n_s_memview __pyx_mstate_global->__pyx_n_s_memview +#define __pyx_n_s_mode __pyx_mstate_global->__pyx_n_s_mode +#define __pyx_n_s_name __pyx_mstate_global->__pyx_n_s_name +#define __pyx_n_s_name_2 __pyx_mstate_global->__pyx_n_s_name_2 +#define __pyx_n_s_ndim __pyx_mstate_global->__pyx_n_s_ndim +#define __pyx_n_s_new __pyx_mstate_global->__pyx_n_s_new +#define __pyx_kp_s_no_default___reduce___due_to_non __pyx_mstate_global->__pyx_kp_s_no_default___reduce___due_to_non +#define __pyx_n_s_np __pyx_mstate_global->__pyx_n_s_np +#define __pyx_n_s_num_tokens_fn __pyx_mstate_global->__pyx_n_s_num_tokens_fn +#define __pyx_n_s_numpy __pyx_mstate_global->__pyx_n_s_numpy +#define __pyx_kp_u_numpy__core_multiarray_failed_to __pyx_mstate_global->__pyx_kp_u_numpy__core_multiarray_failed_to +#define __pyx_kp_u_numpy__core_umath_failed_to_impo __pyx_mstate_global->__pyx_kp_u_numpy__core_umath_failed_to_impo +#define __pyx_n_s_obj __pyx_mstate_global->__pyx_n_s_obj +#define __pyx_n_s_pack __pyx_mstate_global->__pyx_n_s_pack +#define __pyx_n_s_pickle __pyx_mstate_global->__pyx_n_s_pickle +#define __pyx_n_s_pyx_PickleError __pyx_mstate_global->__pyx_n_s_pyx_PickleError +#define __pyx_n_s_pyx_checksum __pyx_mstate_global->__pyx_n_s_pyx_checksum +#define __pyx_n_s_pyx_result __pyx_mstate_global->__pyx_n_s_pyx_result +#define __pyx_n_s_pyx_state __pyx_mstate_global->__pyx_n_s_pyx_state +#define __pyx_n_s_pyx_type __pyx_mstate_global->__pyx_n_s_pyx_type +#define __pyx_n_s_pyx_unpickle_Enum __pyx_mstate_global->__pyx_n_s_pyx_unpickle_Enum +#define __pyx_n_s_pyx_vtable __pyx_mstate_global->__pyx_n_s_pyx_vtable +#define __pyx_n_s_range __pyx_mstate_global->__pyx_n_s_range +#define __pyx_n_s_reduce __pyx_mstate_global->__pyx_n_s_reduce +#define __pyx_n_s_reduce_cython __pyx_mstate_global->__pyx_n_s_reduce_cython +#define __pyx_n_s_reduce_ex __pyx_mstate_global->__pyx_n_s_reduce_ex +#define __pyx_n_s_register __pyx_mstate_global->__pyx_n_s_register +#define __pyx_kp_u_sentence_at_index_of_size_exceed __pyx_mstate_global->__pyx_kp_u_sentence_at_index_of_size_exceed +#define __pyx_n_s_setstate __pyx_mstate_global->__pyx_n_s_setstate +#define __pyx_n_s_setstate_cython __pyx_mstate_global->__pyx_n_s_setstate_cython +#define __pyx_n_s_shape __pyx_mstate_global->__pyx_n_s_shape +#define __pyx_n_s_size __pyx_mstate_global->__pyx_n_s_size +#define __pyx_n_s_spec __pyx_mstate_global->__pyx_n_s_spec +#define __pyx_n_s_start __pyx_mstate_global->__pyx_n_s_start +#define __pyx_n_s_step __pyx_mstate_global->__pyx_n_s_step +#define __pyx_n_s_stop __pyx_mstate_global->__pyx_n_s_stop +#define __pyx_kp_s_strided_and_direct __pyx_mstate_global->__pyx_kp_s_strided_and_direct +#define __pyx_kp_s_strided_and_direct_or_indirect __pyx_mstate_global->__pyx_kp_s_strided_and_direct_or_indirect +#define __pyx_kp_s_strided_and_indirect __pyx_mstate_global->__pyx_kp_s_strided_and_indirect +#define __pyx_kp_s_stringsource __pyx_mstate_global->__pyx_kp_s_stringsource +#define __pyx_n_s_struct __pyx_mstate_global->__pyx_n_s_struct +#define __pyx_n_s_sys __pyx_mstate_global->__pyx_n_s_sys +#define __pyx_n_s_test __pyx_mstate_global->__pyx_n_s_test +#define __pyx_kp_s_unable_to_allocate_array_data __pyx_mstate_global->__pyx_kp_s_unable_to_allocate_array_data +#define __pyx_kp_s_unable_to_allocate_shape_and_str __pyx_mstate_global->__pyx_kp_s_unable_to_allocate_shape_and_str +#define __pyx_n_s_unpack __pyx_mstate_global->__pyx_n_s_unpack +#define __pyx_n_s_update __pyx_mstate_global->__pyx_n_s_update +#define __pyx_n_s_version_info __pyx_mstate_global->__pyx_n_s_version_info +#define __pyx_int_0 __pyx_mstate_global->__pyx_int_0 +#define __pyx_int_1 __pyx_mstate_global->__pyx_int_1 +#define __pyx_int_3 __pyx_mstate_global->__pyx_int_3 +#define __pyx_int_112105877 __pyx_mstate_global->__pyx_int_112105877 +#define __pyx_int_136983863 __pyx_mstate_global->__pyx_int_136983863 +#define __pyx_int_184977713 __pyx_mstate_global->__pyx_int_184977713 +#define __pyx_int_neg_1 __pyx_mstate_global->__pyx_int_neg_1 +#define __pyx_slice__5 __pyx_mstate_global->__pyx_slice__5 +#define __pyx_tuple__4 __pyx_mstate_global->__pyx_tuple__4 +#define __pyx_tuple__8 __pyx_mstate_global->__pyx_tuple__8 +#define __pyx_tuple__9 __pyx_mstate_global->__pyx_tuple__9 +#define __pyx_tuple__10 __pyx_mstate_global->__pyx_tuple__10 +#define __pyx_tuple__11 __pyx_mstate_global->__pyx_tuple__11 +#define __pyx_tuple__12 __pyx_mstate_global->__pyx_tuple__12 +#define __pyx_tuple__13 __pyx_mstate_global->__pyx_tuple__13 +#define __pyx_tuple__14 __pyx_mstate_global->__pyx_tuple__14 +#define __pyx_tuple__15 __pyx_mstate_global->__pyx_tuple__15 +#define __pyx_tuple__16 __pyx_mstate_global->__pyx_tuple__16 +#define __pyx_tuple__17 __pyx_mstate_global->__pyx_tuple__17 +#define __pyx_tuple__18 __pyx_mstate_global->__pyx_tuple__18 +#define __pyx_tuple__19 __pyx_mstate_global->__pyx_tuple__19 +#define __pyx_tuple__20 __pyx_mstate_global->__pyx_tuple__20 +#define __pyx_tuple__22 __pyx_mstate_global->__pyx_tuple__22 +#define __pyx_tuple__24 __pyx_mstate_global->__pyx_tuple__24 +#define __pyx_codeobj__21 __pyx_mstate_global->__pyx_codeobj__21 +#define __pyx_codeobj__23 __pyx_mstate_global->__pyx_codeobj__23 +#define __pyx_codeobj__25 __pyx_mstate_global->__pyx_codeobj__25 +/* #### Code section: module_code ### */ + +/* "View.MemoryView":131 + * cdef bint dtype_is_object + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, # <<<<<<<<<<<<<< + * mode="c", bint allocate_buffer=True): + * + */ + +/* Python wrapper */ +static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/ +static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds) { + PyObject *__pyx_v_shape = 0; + Py_ssize_t __pyx_v_itemsize; + PyObject *__pyx_v_format = 0; + PyObject *__pyx_v_mode = 0; + int __pyx_v_allocate_buffer; + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[5] = {0,0,0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__cinit__ (wrapper)", 0); + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return -1; + #endif + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_shape,&__pyx_n_s_itemsize,&__pyx_n_s_format,&__pyx_n_s_mode,&__pyx_n_s_allocate_buffer,0}; + values[3] = __Pyx_Arg_NewRef_VARARGS(((PyObject *)__pyx_n_s_c)); + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 5: values[4] = __Pyx_Arg_VARARGS(__pyx_args, 4); + CYTHON_FALLTHROUGH; + case 4: values[3] = __Pyx_Arg_VARARGS(__pyx_args, 3); + CYTHON_FALLTHROUGH; + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_VARARGS(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_shape)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_itemsize)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 1); __PYX_ERR(1, 131, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_format)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 2); __PYX_ERR(1, 131, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 3: + if (kw_args > 0) { + PyObject* value = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_mode); + if (value) { values[3] = __Pyx_Arg_NewRef_VARARGS(value); kw_args--; } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 4: + if (kw_args > 0) { + PyObject* value = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_allocate_buffer); + if (value) { values[4] = __Pyx_Arg_NewRef_VARARGS(value); kw_args--; } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__cinit__") < 0)) __PYX_ERR(1, 131, __pyx_L3_error) + } + } else { + switch (__pyx_nargs) { + case 5: values[4] = __Pyx_Arg_VARARGS(__pyx_args, 4); + CYTHON_FALLTHROUGH; + case 4: values[3] = __Pyx_Arg_VARARGS(__pyx_args, 3); + CYTHON_FALLTHROUGH; + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + break; + default: goto __pyx_L5_argtuple_error; + } + } + __pyx_v_shape = ((PyObject*)values[0]); + __pyx_v_itemsize = __Pyx_PyIndex_AsSsize_t(values[1]); if (unlikely((__pyx_v_itemsize == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + __pyx_v_format = values[2]; + __pyx_v_mode = values[3]; + if (values[4]) { + __pyx_v_allocate_buffer = __Pyx_PyObject_IsTrue(values[4]); if (unlikely((__pyx_v_allocate_buffer == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 132, __pyx_L3_error) + } else { + + /* "View.MemoryView":132 + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, + * mode="c", bint allocate_buffer=True): # <<<<<<<<<<<<<< + * + * cdef int idx + */ + __pyx_v_allocate_buffer = ((int)1); + } + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, __pyx_nargs); __PYX_ERR(1, 131, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.array.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return -1; + __pyx_L4_argument_unpacking_done:; + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_shape), (&PyTuple_Type), 1, "shape", 1))) __PYX_ERR(1, 131, __pyx_L1_error) + if (unlikely(((PyObject *)__pyx_v_format) == Py_None)) { + PyErr_Format(PyExc_TypeError, "Argument '%.200s' must not be None", "format"); __PYX_ERR(1, 131, __pyx_L1_error) + } + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(((struct __pyx_array_obj *)__pyx_v_self), __pyx_v_shape, __pyx_v_itemsize, __pyx_v_format, __pyx_v_mode, __pyx_v_allocate_buffer); + + /* "View.MemoryView":131 + * cdef bint dtype_is_object + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, # <<<<<<<<<<<<<< + * mode="c", bint allocate_buffer=True): + * + */ + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __pyx_r = -1; + __pyx_L0:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_shape, Py_ssize_t __pyx_v_itemsize, PyObject *__pyx_v_format, PyObject *__pyx_v_mode, int __pyx_v_allocate_buffer) { + int __pyx_v_idx; + Py_ssize_t __pyx_v_dim; + char __pyx_v_order; + int __pyx_r; + __Pyx_RefNannyDeclarations + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + unsigned int __pyx_t_7; + char *__pyx_t_8; + int __pyx_t_9; + Py_ssize_t __pyx_t_10; + Py_UCS4 __pyx_t_11; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__cinit__", 0); + __Pyx_INCREF(__pyx_v_format); + + /* "View.MemoryView":137 + * cdef Py_ssize_t dim + * + * self.ndim = len(shape) # <<<<<<<<<<<<<< + * self.itemsize = itemsize + * + */ + if (unlikely(__pyx_v_shape == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(1, 137, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_PyTuple_GET_SIZE(__pyx_v_shape); if (unlikely(__pyx_t_1 == ((Py_ssize_t)-1))) __PYX_ERR(1, 137, __pyx_L1_error) + __pyx_v_self->ndim = ((int)__pyx_t_1); + + /* "View.MemoryView":138 + * + * self.ndim = len(shape) + * self.itemsize = itemsize # <<<<<<<<<<<<<< + * + * if not self.ndim: + */ + __pyx_v_self->itemsize = __pyx_v_itemsize; + + /* "View.MemoryView":140 + * self.itemsize = itemsize + * + * if not self.ndim: # <<<<<<<<<<<<<< + * raise ValueError, "Empty shape tuple for cython.array" + * + */ + __pyx_t_2 = (!(__pyx_v_self->ndim != 0)); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":141 + * + * if not self.ndim: + * raise ValueError, "Empty shape tuple for cython.array" # <<<<<<<<<<<<<< + * + * if itemsize <= 0: + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Empty_shape_tuple_for_cython_arr, 0, 0); + __PYX_ERR(1, 141, __pyx_L1_error) + + /* "View.MemoryView":140 + * self.itemsize = itemsize + * + * if not self.ndim: # <<<<<<<<<<<<<< + * raise ValueError, "Empty shape tuple for cython.array" + * + */ + } + + /* "View.MemoryView":143 + * raise ValueError, "Empty shape tuple for cython.array" + * + * if itemsize <= 0: # <<<<<<<<<<<<<< + * raise ValueError, "itemsize <= 0 for cython.array" + * + */ + __pyx_t_2 = (__pyx_v_itemsize <= 0); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":144 + * + * if itemsize <= 0: + * raise ValueError, "itemsize <= 0 for cython.array" # <<<<<<<<<<<<<< + * + * if not isinstance(format, bytes): + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_itemsize_0_for_cython_array, 0, 0); + __PYX_ERR(1, 144, __pyx_L1_error) + + /* "View.MemoryView":143 + * raise ValueError, "Empty shape tuple for cython.array" + * + * if itemsize <= 0: # <<<<<<<<<<<<<< + * raise ValueError, "itemsize <= 0 for cython.array" + * + */ + } + + /* "View.MemoryView":146 + * raise ValueError, "itemsize <= 0 for cython.array" + * + * if not isinstance(format, bytes): # <<<<<<<<<<<<<< + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string + */ + __pyx_t_2 = PyBytes_Check(__pyx_v_format); + __pyx_t_3 = (!__pyx_t_2); + if (__pyx_t_3) { + + /* "View.MemoryView":147 + * + * if not isinstance(format, bytes): + * format = format.encode('ASCII') # <<<<<<<<<<<<<< + * self._format = format # keep a reference to the byte string + * self.format = self._format + */ + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_format, __pyx_n_s_encode); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 147, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_6 = NULL; + __pyx_t_7 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_5))) { + __pyx_t_6 = PyMethod_GET_SELF(__pyx_t_5); + if (likely(__pyx_t_6)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_5); + __Pyx_INCREF(__pyx_t_6); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_5, function); + __pyx_t_7 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_6, __pyx_n_s_ASCII}; + __pyx_t_4 = __Pyx_PyObject_FastCall(__pyx_t_5, __pyx_callargs+1-__pyx_t_7, 1+__pyx_t_7); + __Pyx_XDECREF(__pyx_t_6); __pyx_t_6 = 0; + if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 147, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + __Pyx_DECREF_SET(__pyx_v_format, __pyx_t_4); + __pyx_t_4 = 0; + + /* "View.MemoryView":146 + * raise ValueError, "itemsize <= 0 for cython.array" + * + * if not isinstance(format, bytes): # <<<<<<<<<<<<<< + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string + */ + } + + /* "View.MemoryView":148 + * if not isinstance(format, bytes): + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string # <<<<<<<<<<<<<< + * self.format = self._format + * + */ + if (!(likely(PyBytes_CheckExact(__pyx_v_format))||((__pyx_v_format) == Py_None) || __Pyx_RaiseUnexpectedTypeError("bytes", __pyx_v_format))) __PYX_ERR(1, 148, __pyx_L1_error) + __pyx_t_4 = __pyx_v_format; + __Pyx_INCREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + __Pyx_GOTREF(__pyx_v_self->_format); + __Pyx_DECREF(__pyx_v_self->_format); + __pyx_v_self->_format = ((PyObject*)__pyx_t_4); + __pyx_t_4 = 0; + + /* "View.MemoryView":149 + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string + * self.format = self._format # <<<<<<<<<<<<<< + * + * + */ + if (unlikely(__pyx_v_self->_format == Py_None)) { + PyErr_SetString(PyExc_TypeError, "expected bytes, NoneType found"); + __PYX_ERR(1, 149, __pyx_L1_error) + } + __pyx_t_8 = __Pyx_PyBytes_AsWritableString(__pyx_v_self->_format); if (unlikely((!__pyx_t_8) && PyErr_Occurred())) __PYX_ERR(1, 149, __pyx_L1_error) + __pyx_v_self->format = __pyx_t_8; + + /* "View.MemoryView":152 + * + * + * self._shape = PyObject_Malloc(sizeof(Py_ssize_t)*self.ndim*2) # <<<<<<<<<<<<<< + * self._strides = self._shape + self.ndim + * + */ + __pyx_v_self->_shape = ((Py_ssize_t *)PyObject_Malloc((((sizeof(Py_ssize_t)) * __pyx_v_self->ndim) * 2))); + + /* "View.MemoryView":153 + * + * self._shape = PyObject_Malloc(sizeof(Py_ssize_t)*self.ndim*2) + * self._strides = self._shape + self.ndim # <<<<<<<<<<<<<< + * + * if not self._shape: + */ + __pyx_v_self->_strides = (__pyx_v_self->_shape + __pyx_v_self->ndim); + + /* "View.MemoryView":155 + * self._strides = self._shape + self.ndim + * + * if not self._shape: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate shape and strides." + * + */ + __pyx_t_3 = (!(__pyx_v_self->_shape != 0)); + if (unlikely(__pyx_t_3)) { + + /* "View.MemoryView":156 + * + * if not self._shape: + * raise MemoryError, "unable to allocate shape and strides." # <<<<<<<<<<<<<< + * + * + */ + __Pyx_Raise(__pyx_builtin_MemoryError, __pyx_kp_s_unable_to_allocate_shape_and_str, 0, 0); + __PYX_ERR(1, 156, __pyx_L1_error) + + /* "View.MemoryView":155 + * self._strides = self._shape + self.ndim + * + * if not self._shape: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate shape and strides." + * + */ + } + + /* "View.MemoryView":159 + * + * + * for idx, dim in enumerate(shape): # <<<<<<<<<<<<<< + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + */ + __pyx_t_9 = 0; + __pyx_t_4 = __pyx_v_shape; __Pyx_INCREF(__pyx_t_4); + __pyx_t_1 = 0; + for (;;) { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_4); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 159, __pyx_L1_error) + #endif + if (__pyx_t_1 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_4, __pyx_t_1); __Pyx_INCREF(__pyx_t_5); __pyx_t_1++; if (unlikely((0 < 0))) __PYX_ERR(1, 159, __pyx_L1_error) + #else + __pyx_t_5 = __Pyx_PySequence_ITEM(__pyx_t_4, __pyx_t_1); __pyx_t_1++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 159, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + #endif + __pyx_t_10 = __Pyx_PyIndex_AsSsize_t(__pyx_t_5); if (unlikely((__pyx_t_10 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 159, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __pyx_v_dim = __pyx_t_10; + __pyx_v_idx = __pyx_t_9; + __pyx_t_9 = (__pyx_t_9 + 1); + + /* "View.MemoryView":160 + * + * for idx, dim in enumerate(shape): + * if dim <= 0: # <<<<<<<<<<<<<< + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + * self._shape[idx] = dim + */ + __pyx_t_3 = (__pyx_v_dim <= 0); + if (unlikely(__pyx_t_3)) { + + /* "View.MemoryView":161 + * for idx, dim in enumerate(shape): + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." # <<<<<<<<<<<<<< + * self._shape[idx] = dim + * + */ + __pyx_t_5 = PyTuple_New(5); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_10 = 0; + __pyx_t_11 = 127; + __Pyx_INCREF(__pyx_kp_u_Invalid_shape_in_axis); + __pyx_t_10 += 22; + __Pyx_GIVEREF(__pyx_kp_u_Invalid_shape_in_axis); + PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_kp_u_Invalid_shape_in_axis); + __pyx_t_6 = __Pyx_PyUnicode_From_int(__pyx_v_idx, 0, ' ', 'd'); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __pyx_t_10 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_6); + __Pyx_GIVEREF(__pyx_t_6); + PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_t_6); + __pyx_t_6 = 0; + __Pyx_INCREF(__pyx_kp_u_); + __pyx_t_10 += 2; + __Pyx_GIVEREF(__pyx_kp_u_); + PyTuple_SET_ITEM(__pyx_t_5, 2, __pyx_kp_u_); + __pyx_t_6 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_dim, 0, ' ', 'd'); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __pyx_t_10 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_6); + __Pyx_GIVEREF(__pyx_t_6); + PyTuple_SET_ITEM(__pyx_t_5, 3, __pyx_t_6); + __pyx_t_6 = 0; + __Pyx_INCREF(__pyx_kp_u__2); + __pyx_t_10 += 1; + __Pyx_GIVEREF(__pyx_kp_u__2); + PyTuple_SET_ITEM(__pyx_t_5, 4, __pyx_kp_u__2); + __pyx_t_6 = __Pyx_PyUnicode_Join(__pyx_t_5, 5, __pyx_t_10, __pyx_t_11); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_t_6, 0, 0); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __PYX_ERR(1, 161, __pyx_L1_error) + + /* "View.MemoryView":160 + * + * for idx, dim in enumerate(shape): + * if dim <= 0: # <<<<<<<<<<<<<< + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + * self._shape[idx] = dim + */ + } + + /* "View.MemoryView":162 + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + * self._shape[idx] = dim # <<<<<<<<<<<<<< + * + * cdef char order + */ + (__pyx_v_self->_shape[__pyx_v_idx]) = __pyx_v_dim; + + /* "View.MemoryView":159 + * + * + * for idx, dim in enumerate(shape): # <<<<<<<<<<<<<< + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + */ + } + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "View.MemoryView":165 + * + * cdef char order + * if mode == 'c': # <<<<<<<<<<<<<< + * order = b'C' + * self.mode = u'c' + */ + __pyx_t_3 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_c, Py_EQ)); if (unlikely((__pyx_t_3 < 0))) __PYX_ERR(1, 165, __pyx_L1_error) + if (__pyx_t_3) { + + /* "View.MemoryView":166 + * cdef char order + * if mode == 'c': + * order = b'C' # <<<<<<<<<<<<<< + * self.mode = u'c' + * elif mode == 'fortran': + */ + __pyx_v_order = 'C'; + + /* "View.MemoryView":167 + * if mode == 'c': + * order = b'C' + * self.mode = u'c' # <<<<<<<<<<<<<< + * elif mode == 'fortran': + * order = b'F' + */ + __Pyx_INCREF(__pyx_n_u_c); + __Pyx_GIVEREF(__pyx_n_u_c); + __Pyx_GOTREF(__pyx_v_self->mode); + __Pyx_DECREF(__pyx_v_self->mode); + __pyx_v_self->mode = __pyx_n_u_c; + + /* "View.MemoryView":165 + * + * cdef char order + * if mode == 'c': # <<<<<<<<<<<<<< + * order = b'C' + * self.mode = u'c' + */ + goto __pyx_L11; + } + + /* "View.MemoryView":168 + * order = b'C' + * self.mode = u'c' + * elif mode == 'fortran': # <<<<<<<<<<<<<< + * order = b'F' + * self.mode = u'fortran' + */ + __pyx_t_3 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_fortran, Py_EQ)); if (unlikely((__pyx_t_3 < 0))) __PYX_ERR(1, 168, __pyx_L1_error) + if (likely(__pyx_t_3)) { + + /* "View.MemoryView":169 + * self.mode = u'c' + * elif mode == 'fortran': + * order = b'F' # <<<<<<<<<<<<<< + * self.mode = u'fortran' + * else: + */ + __pyx_v_order = 'F'; + + /* "View.MemoryView":170 + * elif mode == 'fortran': + * order = b'F' + * self.mode = u'fortran' # <<<<<<<<<<<<<< + * else: + * raise ValueError, f"Invalid mode, expected 'c' or 'fortran', got {mode}" + */ + __Pyx_INCREF(__pyx_n_u_fortran); + __Pyx_GIVEREF(__pyx_n_u_fortran); + __Pyx_GOTREF(__pyx_v_self->mode); + __Pyx_DECREF(__pyx_v_self->mode); + __pyx_v_self->mode = __pyx_n_u_fortran; + + /* "View.MemoryView":168 + * order = b'C' + * self.mode = u'c' + * elif mode == 'fortran': # <<<<<<<<<<<<<< + * order = b'F' + * self.mode = u'fortran' + */ + goto __pyx_L11; + } + + /* "View.MemoryView":172 + * self.mode = u'fortran' + * else: + * raise ValueError, f"Invalid mode, expected 'c' or 'fortran', got {mode}" # <<<<<<<<<<<<<< + * + * self.len = fill_contig_strides_array(self._shape, self._strides, itemsize, self.ndim, order) + */ + /*else*/ { + __pyx_t_4 = __Pyx_PyObject_FormatSimple(__pyx_v_mode, __pyx_empty_unicode); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 172, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_6 = __Pyx_PyUnicode_Concat(__pyx_kp_u_Invalid_mode_expected_c_or_fortr, __pyx_t_4); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 172, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_t_6, 0, 0); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __PYX_ERR(1, 172, __pyx_L1_error) + } + __pyx_L11:; + + /* "View.MemoryView":174 + * raise ValueError, f"Invalid mode, expected 'c' or 'fortran', got {mode}" + * + * self.len = fill_contig_strides_array(self._shape, self._strides, itemsize, self.ndim, order) # <<<<<<<<<<<<<< + * + * self.free_data = allocate_buffer + */ + __pyx_v_self->len = __pyx_fill_contig_strides_array(__pyx_v_self->_shape, __pyx_v_self->_strides, __pyx_v_itemsize, __pyx_v_self->ndim, __pyx_v_order); + + /* "View.MemoryView":176 + * self.len = fill_contig_strides_array(self._shape, self._strides, itemsize, self.ndim, order) + * + * self.free_data = allocate_buffer # <<<<<<<<<<<<<< + * self.dtype_is_object = format == b'O' + * + */ + __pyx_v_self->free_data = __pyx_v_allocate_buffer; + + /* "View.MemoryView":177 + * + * self.free_data = allocate_buffer + * self.dtype_is_object = format == b'O' # <<<<<<<<<<<<<< + * + * if allocate_buffer: + */ + __pyx_t_6 = PyObject_RichCompare(__pyx_v_format, __pyx_n_b_O, Py_EQ); __Pyx_XGOTREF(__pyx_t_6); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 177, __pyx_L1_error) + __pyx_t_3 = __Pyx_PyObject_IsTrue(__pyx_t_6); if (unlikely((__pyx_t_3 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 177, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __pyx_v_self->dtype_is_object = __pyx_t_3; + + /* "View.MemoryView":179 + * self.dtype_is_object = format == b'O' + * + * if allocate_buffer: # <<<<<<<<<<<<<< + * _allocate_buffer(self) + * + */ + if (__pyx_v_allocate_buffer) { + + /* "View.MemoryView":180 + * + * if allocate_buffer: + * _allocate_buffer(self) # <<<<<<<<<<<<<< + * + * @cname('getbuffer') + */ + __pyx_t_9 = __pyx_array_allocate_buffer(__pyx_v_self); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 180, __pyx_L1_error) + + /* "View.MemoryView":179 + * self.dtype_is_object = format == b'O' + * + * if allocate_buffer: # <<<<<<<<<<<<<< + * _allocate_buffer(self) + * + */ + } + + /* "View.MemoryView":131 + * cdef bint dtype_is_object + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, # <<<<<<<<<<<<<< + * mode="c", bint allocate_buffer=True): + * + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_AddTraceback("View.MemoryView.array.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_format); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":182 + * _allocate_buffer(self) + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + */ + +/* Python wrapper */ +CYTHON_UNUSED static int __pyx_array_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +CYTHON_UNUSED static int __pyx_array_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getbuffer__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(((struct __pyx_array_obj *)__pyx_v_self), ((Py_buffer *)__pyx_v_info), ((int)__pyx_v_flags)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(struct __pyx_array_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + int __pyx_v_bufmode; + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + char *__pyx_t_2; + Py_ssize_t __pyx_t_3; + int __pyx_t_4; + Py_ssize_t *__pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + if (unlikely(__pyx_v_info == NULL)) { + PyErr_SetString(PyExc_BufferError, "PyObject_GetBuffer: view==NULL argument is obsolete"); + return -1; + } + __Pyx_RefNannySetupContext("__getbuffer__", 0); + __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(__pyx_v_info->obj); + + /* "View.MemoryView":184 + * @cname('getbuffer') + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 # <<<<<<<<<<<<<< + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": + */ + __pyx_v_bufmode = -1; + + /* "View.MemoryView":185 + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): # <<<<<<<<<<<<<< + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + */ + __pyx_t_1 = ((__pyx_v_flags & ((PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS) | PyBUF_ANY_CONTIGUOUS)) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":186 + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": # <<<<<<<<<<<<<< + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": + */ + __pyx_t_1 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_c, Py_EQ)); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 186, __pyx_L1_error) + if (__pyx_t_1) { + + /* "View.MemoryView":187 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS # <<<<<<<<<<<<<< + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + */ + __pyx_v_bufmode = (PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS); + + /* "View.MemoryView":186 + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": # <<<<<<<<<<<<<< + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": + */ + goto __pyx_L4; + } + + /* "View.MemoryView":188 + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": # <<<<<<<<<<<<<< + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): + */ + __pyx_t_1 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_fortran, Py_EQ)); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 188, __pyx_L1_error) + if (__pyx_t_1) { + + /* "View.MemoryView":189 + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS # <<<<<<<<<<<<<< + * if not (flags & bufmode): + * raise ValueError, "Can only create a buffer that is contiguous in memory." + */ + __pyx_v_bufmode = (PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS); + + /* "View.MemoryView":188 + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": # <<<<<<<<<<<<<< + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): + */ + } + __pyx_L4:; + + /* "View.MemoryView":190 + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): # <<<<<<<<<<<<<< + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data + */ + __pyx_t_1 = (!((__pyx_v_flags & __pyx_v_bufmode) != 0)); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":191 + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): + * raise ValueError, "Can only create a buffer that is contiguous in memory." # <<<<<<<<<<<<<< + * info.buf = self.data + * info.len = self.len + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Can_only_create_a_buffer_that_is, 0, 0); + __PYX_ERR(1, 191, __pyx_L1_error) + + /* "View.MemoryView":190 + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): # <<<<<<<<<<<<<< + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data + */ + } + + /* "View.MemoryView":185 + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): # <<<<<<<<<<<<<< + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + */ + } + + /* "View.MemoryView":192 + * if not (flags & bufmode): + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data # <<<<<<<<<<<<<< + * info.len = self.len + * + */ + __pyx_t_2 = __pyx_v_self->data; + __pyx_v_info->buf = __pyx_t_2; + + /* "View.MemoryView":193 + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data + * info.len = self.len # <<<<<<<<<<<<<< + * + * if flags & PyBUF_STRIDES: + */ + __pyx_t_3 = __pyx_v_self->len; + __pyx_v_info->len = __pyx_t_3; + + /* "View.MemoryView":195 + * info.len = self.len + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.ndim = self.ndim + * info.shape = self._shape + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_STRIDES) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":196 + * + * if flags & PyBUF_STRIDES: + * info.ndim = self.ndim # <<<<<<<<<<<<<< + * info.shape = self._shape + * info.strides = self._strides + */ + __pyx_t_4 = __pyx_v_self->ndim; + __pyx_v_info->ndim = __pyx_t_4; + + /* "View.MemoryView":197 + * if flags & PyBUF_STRIDES: + * info.ndim = self.ndim + * info.shape = self._shape # <<<<<<<<<<<<<< + * info.strides = self._strides + * else: + */ + __pyx_t_5 = __pyx_v_self->_shape; + __pyx_v_info->shape = __pyx_t_5; + + /* "View.MemoryView":198 + * info.ndim = self.ndim + * info.shape = self._shape + * info.strides = self._strides # <<<<<<<<<<<<<< + * else: + * info.ndim = 1 + */ + __pyx_t_5 = __pyx_v_self->_strides; + __pyx_v_info->strides = __pyx_t_5; + + /* "View.MemoryView":195 + * info.len = self.len + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.ndim = self.ndim + * info.shape = self._shape + */ + goto __pyx_L6; + } + + /* "View.MemoryView":200 + * info.strides = self._strides + * else: + * info.ndim = 1 # <<<<<<<<<<<<<< + * info.shape = &self.len if flags & PyBUF_ND else NULL + * info.strides = NULL + */ + /*else*/ { + __pyx_v_info->ndim = 1; + + /* "View.MemoryView":201 + * else: + * info.ndim = 1 + * info.shape = &self.len if flags & PyBUF_ND else NULL # <<<<<<<<<<<<<< + * info.strides = NULL + * + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_ND) != 0); + if (__pyx_t_1) { + __pyx_t_5 = (&__pyx_v_self->len); + } else { + __pyx_t_5 = NULL; + } + __pyx_v_info->shape = __pyx_t_5; + + /* "View.MemoryView":202 + * info.ndim = 1 + * info.shape = &self.len if flags & PyBUF_ND else NULL + * info.strides = NULL # <<<<<<<<<<<<<< + * + * info.suboffsets = NULL + */ + __pyx_v_info->strides = NULL; + } + __pyx_L6:; + + /* "View.MemoryView":204 + * info.strides = NULL + * + * info.suboffsets = NULL # <<<<<<<<<<<<<< + * info.itemsize = self.itemsize + * info.readonly = 0 + */ + __pyx_v_info->suboffsets = NULL; + + /* "View.MemoryView":205 + * + * info.suboffsets = NULL + * info.itemsize = self.itemsize # <<<<<<<<<<<<<< + * info.readonly = 0 + * info.format = self.format if flags & PyBUF_FORMAT else NULL + */ + __pyx_t_3 = __pyx_v_self->itemsize; + __pyx_v_info->itemsize = __pyx_t_3; + + /* "View.MemoryView":206 + * info.suboffsets = NULL + * info.itemsize = self.itemsize + * info.readonly = 0 # <<<<<<<<<<<<<< + * info.format = self.format if flags & PyBUF_FORMAT else NULL + * info.obj = self + */ + __pyx_v_info->readonly = 0; + + /* "View.MemoryView":207 + * info.itemsize = self.itemsize + * info.readonly = 0 + * info.format = self.format if flags & PyBUF_FORMAT else NULL # <<<<<<<<<<<<<< + * info.obj = self + * + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0); + if (__pyx_t_1) { + __pyx_t_2 = __pyx_v_self->format; + } else { + __pyx_t_2 = NULL; + } + __pyx_v_info->format = __pyx_t_2; + + /* "View.MemoryView":208 + * info.readonly = 0 + * info.format = self.format if flags & PyBUF_FORMAT else NULL + * info.obj = self # <<<<<<<<<<<<<< + * + * def __dealloc__(array self): + */ + __Pyx_INCREF((PyObject *)__pyx_v_self); + __Pyx_GIVEREF((PyObject *)__pyx_v_self); + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); + __pyx_v_info->obj = ((PyObject *)__pyx_v_self); + + /* "View.MemoryView":182 + * _allocate_buffer(self) + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.array.__getbuffer__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + if (__pyx_v_info->obj != NULL) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + goto __pyx_L2; + __pyx_L0:; + if (__pyx_v_info->obj == Py_None) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + __pyx_L2:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":210 + * info.obj = self + * + * def __dealloc__(array self): # <<<<<<<<<<<<<< + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + */ + +/* Python wrapper */ +static void __pyx_array___dealloc__(PyObject *__pyx_v_self); /*proto*/ +static void __pyx_array___dealloc__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__dealloc__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); +} + +static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struct __pyx_array_obj *__pyx_v_self) { + int __pyx_t_1; + int __pyx_t_2; + + /* "View.MemoryView":211 + * + * def __dealloc__(array self): + * if self.callback_free_data != NULL: # <<<<<<<<<<<<<< + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + */ + __pyx_t_1 = (__pyx_v_self->callback_free_data != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":212 + * def __dealloc__(array self): + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) # <<<<<<<<<<<<<< + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: + */ + __pyx_v_self->callback_free_data(__pyx_v_self->data); + + /* "View.MemoryView":211 + * + * def __dealloc__(array self): + * if self.callback_free_data != NULL: # <<<<<<<<<<<<<< + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":213 + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: # <<<<<<<<<<<<<< + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + */ + if (__pyx_v_self->free_data) { + } else { + __pyx_t_1 = __pyx_v_self->free_data; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_self->data != NULL); + __pyx_t_1 = __pyx_t_2; + __pyx_L4_bool_binop_done:; + if (__pyx_t_1) { + + /* "View.MemoryView":214 + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) + */ + if (__pyx_v_self->dtype_is_object) { + + /* "View.MemoryView":215 + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) # <<<<<<<<<<<<<< + * free(self.data) + * PyObject_Free(self._shape) + */ + __pyx_memoryview_refcount_objects_in_slice(__pyx_v_self->data, __pyx_v_self->_shape, __pyx_v_self->_strides, __pyx_v_self->ndim, 0); + + /* "View.MemoryView":214 + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) + */ + } + + /* "View.MemoryView":216 + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) # <<<<<<<<<<<<<< + * PyObject_Free(self._shape) + * + */ + free(__pyx_v_self->data); + + /* "View.MemoryView":213 + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: # <<<<<<<<<<<<<< + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + */ + } + __pyx_L3:; + + /* "View.MemoryView":217 + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) + * PyObject_Free(self._shape) # <<<<<<<<<<<<<< + * + * @property + */ + PyObject_Free(__pyx_v_self->_shape); + + /* "View.MemoryView":210 + * info.obj = self + * + * def __dealloc__(array self): # <<<<<<<<<<<<<< + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + */ + + /* function exit code */ +} + +/* "View.MemoryView":219 + * PyObject_Free(self._shape) + * + * @property # <<<<<<<<<<<<<< + * def memview(self): + * return self.get_memview() + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_5array_7memview_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_5array_7memview_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_5array_7memview___get__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_5array_7memview___get__(struct __pyx_array_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":221 + * @property + * def memview(self): + * return self.get_memview() # <<<<<<<<<<<<<< + * + * @cname('get_memview') + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = ((struct __pyx_vtabstruct_array *)__pyx_v_self->__pyx_vtab)->get_memview(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 221, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":219 + * PyObject_Free(self._shape) + * + * @property # <<<<<<<<<<<<<< + * def memview(self): + * return self.get_memview() + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.array.memview.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":224 + * + * @cname('get_memview') + * cdef get_memview(self): # <<<<<<<<<<<<<< + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE + * return memoryview(self, flags, self.dtype_is_object) + */ + +static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self) { + int __pyx_v_flags; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("get_memview", 1); + + /* "View.MemoryView":225 + * @cname('get_memview') + * cdef get_memview(self): + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE # <<<<<<<<<<<<<< + * return memoryview(self, flags, self.dtype_is_object) + * + */ + __pyx_v_flags = ((PyBUF_ANY_CONTIGUOUS | PyBUF_FORMAT) | PyBUF_WRITABLE); + + /* "View.MemoryView":226 + * cdef get_memview(self): + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE + * return memoryview(self, flags, self.dtype_is_object) # <<<<<<<<<<<<<< + * + * def __len__(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF((PyObject *)__pyx_v_self); + __Pyx_GIVEREF((PyObject *)__pyx_v_self); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, ((PyObject *)__pyx_v_self))) __PYX_ERR(1, 226, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_1)) __PYX_ERR(1, 226, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2)) __PYX_ERR(1, 226, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":224 + * + * @cname('get_memview') + * cdef get_memview(self): # <<<<<<<<<<<<<< + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE + * return memoryview(self, flags, self.dtype_is_object) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.array.get_memview", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":228 + * return memoryview(self, flags, self.dtype_is_object) + * + * def __len__(self): # <<<<<<<<<<<<<< + * return self._shape[0] + * + */ + +/* Python wrapper */ +static Py_ssize_t __pyx_array___len__(PyObject *__pyx_v_self); /*proto*/ +static Py_ssize_t __pyx_array___len__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + Py_ssize_t __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__len__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static Py_ssize_t __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(struct __pyx_array_obj *__pyx_v_self) { + Py_ssize_t __pyx_r; + + /* "View.MemoryView":229 + * + * def __len__(self): + * return self._shape[0] # <<<<<<<<<<<<<< + * + * def __getattr__(self, attr): + */ + __pyx_r = (__pyx_v_self->_shape[0]); + goto __pyx_L0; + + /* "View.MemoryView":228 + * return memoryview(self, flags, self.dtype_is_object) + * + * def __len__(self): # <<<<<<<<<<<<<< + * return self._shape[0] + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":231 + * return self._shape[0] + * + * def __getattr__(self, attr): # <<<<<<<<<<<<<< + * return getattr(self.memview, attr) + * + */ + +/* Python wrapper */ +static PyObject *__pyx_array___getattr__(PyObject *__pyx_v_self, PyObject *__pyx_v_attr); /*proto*/ +static PyObject *__pyx_array___getattr__(PyObject *__pyx_v_self, PyObject *__pyx_v_attr) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getattr__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(((struct __pyx_array_obj *)__pyx_v_self), ((PyObject *)__pyx_v_attr)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_attr) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__getattr__", 1); + + /* "View.MemoryView":232 + * + * def __getattr__(self, attr): + * return getattr(self.memview, attr) # <<<<<<<<<<<<<< + * + * def __getitem__(self, item): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 232, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_GetAttr(__pyx_t_1, __pyx_v_attr); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 232, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":231 + * return self._shape[0] + * + * def __getattr__(self, attr): # <<<<<<<<<<<<<< + * return getattr(self.memview, attr) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.array.__getattr__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":234 + * return getattr(self.memview, attr) + * + * def __getitem__(self, item): # <<<<<<<<<<<<<< + * return self.memview[item] + * + */ + +/* Python wrapper */ +static PyObject *__pyx_array___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item); /*proto*/ +static PyObject *__pyx_array___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__(((struct __pyx_array_obj *)__pyx_v_self), ((PyObject *)__pyx_v_item)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__getitem__", 1); + + /* "View.MemoryView":235 + * + * def __getitem__(self, item): + * return self.memview[item] # <<<<<<<<<<<<<< + * + * def __setitem__(self, item, value): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 235, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetItem(__pyx_t_1, __pyx_v_item); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 235, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":234 + * return getattr(self.memview, attr) + * + * def __getitem__(self, item): # <<<<<<<<<<<<<< + * return self.memview[item] + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.array.__getitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":237 + * return self.memview[item] + * + * def __setitem__(self, item, value): # <<<<<<<<<<<<<< + * self.memview[item] = value + * + */ + +/* Python wrapper */ +static int __pyx_array___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value); /*proto*/ +static int __pyx_array___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_12__setitem__(((struct __pyx_array_obj *)__pyx_v_self), ((PyObject *)__pyx_v_item), ((PyObject *)__pyx_v_value)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_12__setitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setitem__", 1); + + /* "View.MemoryView":238 + * + * def __setitem__(self, item, value): + * self.memview[item] = value # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 238, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + if (unlikely((PyObject_SetItem(__pyx_t_1, __pyx_v_item, __pyx_v_value) < 0))) __PYX_ERR(1, 238, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "View.MemoryView":237 + * return self.memview[item] + * + * def __setitem__(self, item, value): # <<<<<<<<<<<<<< + * self.memview[item] = value + * + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.array.__setitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_array_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_array_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_array___reduce_cython__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_array___reduce_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":2 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 2, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.array.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_array_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_array_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + CYTHON_UNUSED PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 3, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 3, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 3, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.array.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_array_2__setstate_cython__(((struct __pyx_array_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_array_2__setstate_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":4 + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 4, __pyx_L1_error) + + /* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.array.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":248 + * + * @cname("__pyx_array_allocate_buffer") + * cdef int _allocate_buffer(array self) except -1: # <<<<<<<<<<<<<< + * + * + */ + +static int __pyx_array_allocate_buffer(struct __pyx_array_obj *__pyx_v_self) { + Py_ssize_t __pyx_v_i; + PyObject **__pyx_v_p; + int __pyx_r; + int __pyx_t_1; + Py_ssize_t __pyx_t_2; + Py_ssize_t __pyx_t_3; + Py_ssize_t __pyx_t_4; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + + /* "View.MemoryView":254 + * cdef PyObject **p + * + * self.free_data = True # <<<<<<<<<<<<<< + * self.data = malloc(self.len) + * if not self.data: + */ + __pyx_v_self->free_data = 1; + + /* "View.MemoryView":255 + * + * self.free_data = True + * self.data = malloc(self.len) # <<<<<<<<<<<<<< + * if not self.data: + * raise MemoryError, "unable to allocate array data." + */ + __pyx_v_self->data = ((char *)malloc(__pyx_v_self->len)); + + /* "View.MemoryView":256 + * self.free_data = True + * self.data = malloc(self.len) + * if not self.data: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate array data." + * + */ + __pyx_t_1 = (!(__pyx_v_self->data != 0)); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":257 + * self.data = malloc(self.len) + * if not self.data: + * raise MemoryError, "unable to allocate array data." # <<<<<<<<<<<<<< + * + * if self.dtype_is_object: + */ + __Pyx_Raise(__pyx_builtin_MemoryError, __pyx_kp_s_unable_to_allocate_array_data, 0, 0); + __PYX_ERR(1, 257, __pyx_L1_error) + + /* "View.MemoryView":256 + * self.free_data = True + * self.data = malloc(self.len) + * if not self.data: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate array data." + * + */ + } + + /* "View.MemoryView":259 + * raise MemoryError, "unable to allocate array data." + * + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * p = self.data + * for i in range(self.len // self.itemsize): + */ + if (__pyx_v_self->dtype_is_object) { + + /* "View.MemoryView":260 + * + * if self.dtype_is_object: + * p = self.data # <<<<<<<<<<<<<< + * for i in range(self.len // self.itemsize): + * p[i] = Py_None + */ + __pyx_v_p = ((PyObject **)__pyx_v_self->data); + + /* "View.MemoryView":261 + * if self.dtype_is_object: + * p = self.data + * for i in range(self.len // self.itemsize): # <<<<<<<<<<<<<< + * p[i] = Py_None + * Py_INCREF(Py_None) + */ + if (unlikely(__pyx_v_self->itemsize == 0)) { + PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero"); + __PYX_ERR(1, 261, __pyx_L1_error) + } + else if (sizeof(Py_ssize_t) == sizeof(long) && (!(((Py_ssize_t)-1) > 0)) && unlikely(__pyx_v_self->itemsize == (Py_ssize_t)-1) && unlikely(__Pyx_UNARY_NEG_WOULD_OVERFLOW(__pyx_v_self->len))) { + PyErr_SetString(PyExc_OverflowError, "value too large to perform division"); + __PYX_ERR(1, 261, __pyx_L1_error) + } + __pyx_t_2 = __Pyx_div_Py_ssize_t(__pyx_v_self->len, __pyx_v_self->itemsize); + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":262 + * p = self.data + * for i in range(self.len // self.itemsize): + * p[i] = Py_None # <<<<<<<<<<<<<< + * Py_INCREF(Py_None) + * return 0 + */ + (__pyx_v_p[__pyx_v_i]) = Py_None; + + /* "View.MemoryView":263 + * for i in range(self.len // self.itemsize): + * p[i] = Py_None + * Py_INCREF(Py_None) # <<<<<<<<<<<<<< + * return 0 + * + */ + Py_INCREF(Py_None); + } + + /* "View.MemoryView":259 + * raise MemoryError, "unable to allocate array data." + * + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * p = self.data + * for i in range(self.len // self.itemsize): + */ + } + + /* "View.MemoryView":264 + * p[i] = Py_None + * Py_INCREF(Py_None) + * return 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":248 + * + * @cname("__pyx_array_allocate_buffer") + * cdef int _allocate_buffer(array self) except -1: # <<<<<<<<<<<<<< + * + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._allocate_buffer", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":268 + * + * @cname("__pyx_array_new") + * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format, char *c_mode, char *buf): # <<<<<<<<<<<<<< + * cdef array result + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + */ + +static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize_t __pyx_v_itemsize, char *__pyx_v_format, char *__pyx_v_c_mode, char *__pyx_v_buf) { + struct __pyx_array_obj *__pyx_v_result = 0; + PyObject *__pyx_v_mode = 0; + struct __pyx_array_obj *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("array_cwrapper", 1); + + /* "View.MemoryView":270 + * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format, char *c_mode, char *buf): + * cdef array result + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. # <<<<<<<<<<<<<< + * + * if buf is NULL: + */ + __pyx_t_2 = ((__pyx_v_c_mode[0]) == 'f'); + if (__pyx_t_2) { + __Pyx_INCREF(__pyx_n_s_fortran); + __pyx_t_1 = __pyx_n_s_fortran; + } else { + __Pyx_INCREF(__pyx_n_s_c); + __pyx_t_1 = __pyx_n_s_c; + } + __pyx_v_mode = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":272 + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + * + * if buf is NULL: # <<<<<<<<<<<<<< + * result = array.__new__(array, shape, itemsize, format, mode) + * else: + */ + __pyx_t_2 = (__pyx_v_buf == NULL); + if (__pyx_t_2) { + + /* "View.MemoryView":273 + * + * if buf is NULL: + * result = array.__new__(array, shape, itemsize, format, mode) # <<<<<<<<<<<<<< + * else: + * result = array.__new__(array, shape, itemsize, format, mode, allocate_buffer=False) + */ + __pyx_t_1 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = PyTuple_New(4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_INCREF(__pyx_v_shape); + __Pyx_GIVEREF(__pyx_v_shape); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_v_shape)) __PYX_ERR(1, 273, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 1, __pyx_t_1)) __PYX_ERR(1, 273, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 2, __pyx_t_3)) __PYX_ERR(1, 273, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_mode); + __Pyx_GIVEREF(__pyx_v_mode); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 3, __pyx_v_mode)) __PYX_ERR(1, 273, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_3 = 0; + __pyx_t_3 = ((PyObject *)__pyx_tp_new_array(((PyTypeObject *)__pyx_array_type), __pyx_t_4, NULL)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF((PyObject *)__pyx_t_3); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_v_result = ((struct __pyx_array_obj *)__pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":272 + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + * + * if buf is NULL: # <<<<<<<<<<<<<< + * result = array.__new__(array, shape, itemsize, format, mode) + * else: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":275 + * result = array.__new__(array, shape, itemsize, format, mode) + * else: + * result = array.__new__(array, shape, itemsize, format, mode, allocate_buffer=False) # <<<<<<<<<<<<<< + * result.data = buf + * + */ + /*else*/ { + __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_1 = PyTuple_New(4); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_v_shape); + __Pyx_GIVEREF(__pyx_v_shape); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_shape)) __PYX_ERR(1, 275, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_t_3)) __PYX_ERR(1, 275, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_4); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 2, __pyx_t_4)) __PYX_ERR(1, 275, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_mode); + __Pyx_GIVEREF(__pyx_v_mode); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 3, __pyx_v_mode)) __PYX_ERR(1, 275, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_4 = 0; + __pyx_t_4 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + if (PyDict_SetItem(__pyx_t_4, __pyx_n_s_allocate_buffer, Py_False) < 0) __PYX_ERR(1, 275, __pyx_L1_error) + __pyx_t_3 = ((PyObject *)__pyx_tp_new_array(((PyTypeObject *)__pyx_array_type), __pyx_t_1, __pyx_t_4)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF((PyObject *)__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_v_result = ((struct __pyx_array_obj *)__pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":276 + * else: + * result = array.__new__(array, shape, itemsize, format, mode, allocate_buffer=False) + * result.data = buf # <<<<<<<<<<<<<< + * + * return result + */ + __pyx_v_result->data = __pyx_v_buf; + } + __pyx_L3:; + + /* "View.MemoryView":278 + * result.data = buf + * + * return result # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = __pyx_v_result; + goto __pyx_L0; + + /* "View.MemoryView":268 + * + * @cname("__pyx_array_new") + * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format, char *c_mode, char *buf): # <<<<<<<<<<<<<< + * cdef array result + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.array_cwrapper", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XDECREF(__pyx_v_mode); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":304 + * cdef class Enum(object): + * cdef object name + * def __init__(self, name): # <<<<<<<<<<<<<< + * self.name = name + * def __repr__(self): + */ + +/* Python wrapper */ +static int __pyx_MemviewEnum___init__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/ +static int __pyx_MemviewEnum___init__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds) { + PyObject *__pyx_v_name = 0; + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__init__ (wrapper)", 0); + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return -1; + #endif + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_name,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_VARARGS(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_name)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 304, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__init__") < 0)) __PYX_ERR(1, 304, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + } + __pyx_v_name = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__init__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 304, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.Enum.__init__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return -1; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self), __pyx_v_name); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v_name) { + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__init__", 1); + + /* "View.MemoryView":305 + * cdef object name + * def __init__(self, name): + * self.name = name # <<<<<<<<<<<<<< + * def __repr__(self): + * return self.name + */ + __Pyx_INCREF(__pyx_v_name); + __Pyx_GIVEREF(__pyx_v_name); + __Pyx_GOTREF(__pyx_v_self->name); + __Pyx_DECREF(__pyx_v_self->name); + __pyx_v_self->name = __pyx_v_name; + + /* "View.MemoryView":304 + * cdef class Enum(object): + * cdef object name + * def __init__(self, name): # <<<<<<<<<<<<<< + * self.name = name + * def __repr__(self): + */ + + /* function exit code */ + __pyx_r = 0; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":306 + * def __init__(self, name): + * self.name = name + * def __repr__(self): # <<<<<<<<<<<<<< + * return self.name + * + */ + +/* Python wrapper */ +static PyObject *__pyx_MemviewEnum___repr__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_MemviewEnum___repr__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__repr__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr__(struct __pyx_MemviewEnum_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__repr__", 1); + + /* "View.MemoryView":307 + * self.name = name + * def __repr__(self): + * return self.name # <<<<<<<<<<<<<< + * + * cdef generic = Enum("") + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->name); + __pyx_r = __pyx_v_self->name; + goto __pyx_L0; + + /* "View.MemoryView":306 + * def __init__(self, name): + * self.name = name + * def __repr__(self): # <<<<<<<<<<<<<< + * return self.name + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_MemviewEnum_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_MemviewEnum_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_MemviewEnum___reduce_cython__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_MemviewEnum___reduce_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self) { + PyObject *__pyx_v_state = 0; + PyObject *__pyx_v__dict = 0; + int __pyx_v_use_setstate; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":5 + * cdef object _dict + * cdef bint use_setstate + * state = (self.name,) # <<<<<<<<<<<<<< + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: + */ + __pyx_t_1 = PyTuple_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_v_self->name); + __Pyx_GIVEREF(__pyx_v_self->name); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_self->name)) __PYX_ERR(1, 5, __pyx_L1_error); + __pyx_v_state = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "(tree fragment)":6 + * cdef bint use_setstate + * state = (self.name,) + * _dict = getattr(self, '__dict__', None) # <<<<<<<<<<<<<< + * if _dict is not None: + * state += (_dict,) + */ + __pyx_t_1 = __Pyx_GetAttr3(((PyObject *)__pyx_v_self), __pyx_n_s_dict, Py_None); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v__dict = __pyx_t_1; + __pyx_t_1 = 0; + + /* "(tree fragment)":7 + * state = (self.name,) + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: # <<<<<<<<<<<<<< + * state += (_dict,) + * use_setstate = True + */ + __pyx_t_2 = (__pyx_v__dict != Py_None); + if (__pyx_t_2) { + + /* "(tree fragment)":8 + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: + * state += (_dict,) # <<<<<<<<<<<<<< + * use_setstate = True + * else: + */ + __pyx_t_1 = PyTuple_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 8, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_v__dict); + __Pyx_GIVEREF(__pyx_v__dict); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v__dict)) __PYX_ERR(1, 8, __pyx_L1_error); + __pyx_t_3 = PyNumber_InPlaceAdd(__pyx_v_state, __pyx_t_1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 8, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF_SET(__pyx_v_state, ((PyObject*)__pyx_t_3)); + __pyx_t_3 = 0; + + /* "(tree fragment)":9 + * if _dict is not None: + * state += (_dict,) + * use_setstate = True # <<<<<<<<<<<<<< + * else: + * use_setstate = self.name is not None + */ + __pyx_v_use_setstate = 1; + + /* "(tree fragment)":7 + * state = (self.name,) + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: # <<<<<<<<<<<<<< + * state += (_dict,) + * use_setstate = True + */ + goto __pyx_L3; + } + + /* "(tree fragment)":11 + * use_setstate = True + * else: + * use_setstate = self.name is not None # <<<<<<<<<<<<<< + * if use_setstate: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + */ + /*else*/ { + __pyx_t_2 = (__pyx_v_self->name != Py_None); + __pyx_v_use_setstate = __pyx_t_2; + } + __pyx_L3:; + + /* "(tree fragment)":12 + * else: + * use_setstate = self.name is not None + * if use_setstate: # <<<<<<<<<<<<<< + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + * else: + */ + if (__pyx_v_use_setstate) { + + /* "(tree fragment)":13 + * use_setstate = self.name is not None + * if use_setstate: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state # <<<<<<<<<<<<<< + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_GetModuleGlobalName(__pyx_t_3, __pyx_n_s_pyx_unpickle_Enum); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 = PyTuple_New(3); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + __Pyx_GIVEREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, ((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))))) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_136983863); + __Pyx_GIVEREF(__pyx_int_136983863); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_136983863)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(Py_None); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 2, Py_None)) __PYX_ERR(1, 13, __pyx_L1_error); + __pyx_t_4 = PyTuple_New(3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_t_3)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 1, __pyx_t_1)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_state); + __Pyx_GIVEREF(__pyx_v_state); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 2, __pyx_v_state)) __PYX_ERR(1, 13, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_1 = 0; + __pyx_r = __pyx_t_4; + __pyx_t_4 = 0; + goto __pyx_L0; + + /* "(tree fragment)":12 + * else: + * use_setstate = self.name is not None + * if use_setstate: # <<<<<<<<<<<<<< + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + * else: + */ + } + + /* "(tree fragment)":15 + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * __pyx_unpickle_Enum__set_state(self, __pyx_state) + */ + /*else*/ { + __Pyx_XDECREF(__pyx_r); + __Pyx_GetModuleGlobalName(__pyx_t_4, __pyx_n_s_pyx_unpickle_Enum); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_1 = PyTuple_New(3); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + __Pyx_GIVEREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, ((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))))) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_136983863); + __Pyx_GIVEREF(__pyx_int_136983863); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_136983863)) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_state); + __Pyx_GIVEREF(__pyx_v_state); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 2, __pyx_v_state)) __PYX_ERR(1, 15, __pyx_L1_error); + __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_4); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_4)) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_1)) __PYX_ERR(1, 15, __pyx_L1_error); + __pyx_t_4 = 0; + __pyx_t_1 = 0; + __pyx_r = __pyx_t_3; + __pyx_t_3 = 0; + goto __pyx_L0; + } + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.Enum.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_state); + __Pyx_XDECREF(__pyx_v__dict); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":16 + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state(self, __pyx_state) + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_MemviewEnum_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_MemviewEnum_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 16, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 16, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 16, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.Enum.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_MemviewEnum_2__setstate_cython__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_MemviewEnum_2__setstate_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":17 + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + * def __setstate_cython__(self, __pyx_state): + * __pyx_unpickle_Enum__set_state(self, __pyx_state) # <<<<<<<<<<<<<< + */ + if (!(likely(PyTuple_CheckExact(__pyx_v___pyx_state))||((__pyx_v___pyx_state) == Py_None) || __Pyx_RaiseUnexpectedTypeError("tuple", __pyx_v___pyx_state))) __PYX_ERR(1, 17, __pyx_L1_error) + __pyx_t_1 = __pyx_unpickle_Enum__set_state(__pyx_v_self, ((PyObject*)__pyx_v___pyx_state)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 17, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":16 + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state(self, __pyx_state) + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.Enum.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":349 + * cdef __Pyx_TypeInfo *typeinfo + * + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): # <<<<<<<<<<<<<< + * self.obj = obj + * self.flags = flags + */ + +/* Python wrapper */ +static int __pyx_memoryview___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/ +static int __pyx_memoryview___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds) { + PyObject *__pyx_v_obj = 0; + int __pyx_v_flags; + int __pyx_v_dtype_is_object; + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[3] = {0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__cinit__ (wrapper)", 0); + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return -1; + #endif + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_obj,&__pyx_n_s_flags,&__pyx_n_s_dtype_is_object,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_VARARGS(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_obj)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_flags)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, 1); __PYX_ERR(1, 349, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (kw_args > 0) { + PyObject* value = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_dtype_is_object); + if (value) { values[2] = __Pyx_Arg_NewRef_VARARGS(value); kw_args--; } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__cinit__") < 0)) __PYX_ERR(1, 349, __pyx_L3_error) + } + } else { + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + break; + default: goto __pyx_L5_argtuple_error; + } + } + __pyx_v_obj = values[0]; + __pyx_v_flags = __Pyx_PyInt_As_int(values[1]); if (unlikely((__pyx_v_flags == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + if (values[2]) { + __pyx_v_dtype_is_object = __Pyx_PyObject_IsTrue(values[2]); if (unlikely((__pyx_v_dtype_is_object == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + } else { + __pyx_v_dtype_is_object = ((int)0); + } + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, __pyx_nargs); __PYX_ERR(1, 349, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.memoryview.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return -1; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit__(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_obj, __pyx_v_flags, __pyx_v_dtype_is_object); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj, int __pyx_v_flags, int __pyx_v_dtype_is_object) { + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + Py_intptr_t __pyx_t_4; + size_t __pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__cinit__", 1); + + /* "View.MemoryView":350 + * + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): + * self.obj = obj # <<<<<<<<<<<<<< + * self.flags = flags + * if type(self) is memoryview or obj is not None: + */ + __Pyx_INCREF(__pyx_v_obj); + __Pyx_GIVEREF(__pyx_v_obj); + __Pyx_GOTREF(__pyx_v_self->obj); + __Pyx_DECREF(__pyx_v_self->obj); + __pyx_v_self->obj = __pyx_v_obj; + + /* "View.MemoryView":351 + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): + * self.obj = obj + * self.flags = flags # <<<<<<<<<<<<<< + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) + */ + __pyx_v_self->flags = __pyx_v_flags; + + /* "View.MemoryView":352 + * self.obj = obj + * self.flags = flags + * if type(self) is memoryview or obj is not None: # <<<<<<<<<<<<<< + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: + */ + __pyx_t_2 = (((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))) == ((PyObject *)__pyx_memoryview_type)); + if (!__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_obj != Py_None); + __pyx_t_1 = __pyx_t_2; + __pyx_L4_bool_binop_done:; + if (__pyx_t_1) { + + /* "View.MemoryView":353 + * self.flags = flags + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) # <<<<<<<<<<<<<< + * if self.view.obj == NULL: + * (<__pyx_buffer *> &self.view).obj = Py_None + */ + __pyx_t_3 = __Pyx_GetBuffer(__pyx_v_obj, (&__pyx_v_self->view), __pyx_v_flags); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 353, __pyx_L1_error) + + /* "View.MemoryView":354 + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: # <<<<<<<<<<<<<< + * (<__pyx_buffer *> &self.view).obj = Py_None + * Py_INCREF(Py_None) + */ + __pyx_t_1 = (((PyObject *)__pyx_v_self->view.obj) == NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":355 + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: + * (<__pyx_buffer *> &self.view).obj = Py_None # <<<<<<<<<<<<<< + * Py_INCREF(Py_None) + * + */ + ((Py_buffer *)(&__pyx_v_self->view))->obj = Py_None; + + /* "View.MemoryView":356 + * if self.view.obj == NULL: + * (<__pyx_buffer *> &self.view).obj = Py_None + * Py_INCREF(Py_None) # <<<<<<<<<<<<<< + * + * if not __PYX_CYTHON_ATOMICS_ENABLED(): + */ + Py_INCREF(Py_None); + + /* "View.MemoryView":354 + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: # <<<<<<<<<<<<<< + * (<__pyx_buffer *> &self.view).obj = Py_None + * Py_INCREF(Py_None) + */ + } + + /* "View.MemoryView":352 + * self.obj = obj + * self.flags = flags + * if type(self) is memoryview or obj is not None: # <<<<<<<<<<<<<< + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: + */ + } + + /* "View.MemoryView":358 + * Py_INCREF(Py_None) + * + * if not __PYX_CYTHON_ATOMICS_ENABLED(): # <<<<<<<<<<<<<< + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: + */ + __pyx_t_1 = (!__PYX_CYTHON_ATOMICS_ENABLED()); + if (__pyx_t_1) { + + /* "View.MemoryView":360 + * if not __PYX_CYTHON_ATOMICS_ENABLED(): + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: # <<<<<<<<<<<<<< + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + */ + __pyx_t_1 = (__pyx_memoryview_thread_locks_used < 8); + if (__pyx_t_1) { + + /* "View.MemoryView":361 + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: + */ + __pyx_v_self->lock = (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]); + + /* "View.MemoryView":362 + * if __pyx_memoryview_thread_locks_used < 8: + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 # <<<<<<<<<<<<<< + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() + */ + __pyx_memoryview_thread_locks_used = (__pyx_memoryview_thread_locks_used + 1); + + /* "View.MemoryView":360 + * if not __PYX_CYTHON_ATOMICS_ENABLED(): + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: # <<<<<<<<<<<<<< + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + */ + } + + /* "View.MemoryView":363 + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: # <<<<<<<<<<<<<< + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: + */ + __pyx_t_1 = (__pyx_v_self->lock == NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":364 + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() # <<<<<<<<<<<<<< + * if self.lock is NULL: + * raise MemoryError + */ + __pyx_v_self->lock = PyThread_allocate_lock(); + + /* "View.MemoryView":365 + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + __pyx_t_1 = (__pyx_v_self->lock == NULL); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":366 + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: + * raise MemoryError # <<<<<<<<<<<<<< + * + * if flags & PyBUF_FORMAT: + */ + PyErr_NoMemory(); __PYX_ERR(1, 366, __pyx_L1_error) + + /* "View.MemoryView":365 + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + } + + /* "View.MemoryView":363 + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: # <<<<<<<<<<<<<< + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: + */ + } + + /* "View.MemoryView":358 + * Py_INCREF(Py_None) + * + * if not __PYX_CYTHON_ATOMICS_ENABLED(): # <<<<<<<<<<<<<< + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: + */ + } + + /* "View.MemoryView":368 + * raise MemoryError + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":369 + * + * if flags & PyBUF_FORMAT: + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') # <<<<<<<<<<<<<< + * else: + * self.dtype_is_object = dtype_is_object + */ + __pyx_t_2 = ((__pyx_v_self->view.format[0]) == 'O'); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L12_bool_binop_done; + } + __pyx_t_2 = ((__pyx_v_self->view.format[1]) == '\x00'); + __pyx_t_1 = __pyx_t_2; + __pyx_L12_bool_binop_done:; + __pyx_v_self->dtype_is_object = __pyx_t_1; + + /* "View.MemoryView":368 + * raise MemoryError + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') + * else: + */ + goto __pyx_L11; + } + + /* "View.MemoryView":371 + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') + * else: + * self.dtype_is_object = dtype_is_object # <<<<<<<<<<<<<< + * + * assert (&self.acquisition_count) % sizeof(__pyx_atomic_int_type) == 0 + */ + /*else*/ { + __pyx_v_self->dtype_is_object = __pyx_v_dtype_is_object; + } + __pyx_L11:; + + /* "View.MemoryView":373 + * self.dtype_is_object = dtype_is_object + * + * assert (&self.acquisition_count) % sizeof(__pyx_atomic_int_type) == 0 # <<<<<<<<<<<<<< + * self.typeinfo = NULL + * + */ + #ifndef CYTHON_WITHOUT_ASSERTIONS + if (unlikely(__pyx_assertions_enabled())) { + __pyx_t_4 = ((Py_intptr_t)((void *)(&__pyx_v_self->acquisition_count))); + __pyx_t_5 = (sizeof(__pyx_atomic_int_type)); + if (unlikely(__pyx_t_5 == 0)) { + PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero"); + __PYX_ERR(1, 373, __pyx_L1_error) + } + __pyx_t_1 = ((__pyx_t_4 % __pyx_t_5) == 0); + if (unlikely(!__pyx_t_1)) { + __Pyx_Raise(__pyx_builtin_AssertionError, 0, 0, 0); + __PYX_ERR(1, 373, __pyx_L1_error) + } + } + #else + if ((1)); else __PYX_ERR(1, 373, __pyx_L1_error) + #endif + + /* "View.MemoryView":374 + * + * assert (&self.acquisition_count) % sizeof(__pyx_atomic_int_type) == 0 + * self.typeinfo = NULL # <<<<<<<<<<<<<< + * + * def __dealloc__(memoryview self): + */ + __pyx_v_self->typeinfo = NULL; + + /* "View.MemoryView":349 + * cdef __Pyx_TypeInfo *typeinfo + * + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): # <<<<<<<<<<<<<< + * self.obj = obj + * self.flags = flags + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":376 + * self.typeinfo = NULL + * + * def __dealloc__(memoryview self): # <<<<<<<<<<<<<< + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + */ + +/* Python wrapper */ +static void __pyx_memoryview___dealloc__(PyObject *__pyx_v_self); /*proto*/ +static void __pyx_memoryview___dealloc__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__dealloc__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__dealloc__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); +} + +static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__dealloc__(struct __pyx_memoryview_obj *__pyx_v_self) { + int __pyx_v_i; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + PyThread_type_lock __pyx_t_5; + PyThread_type_lock __pyx_t_6; + + /* "View.MemoryView":377 + * + * def __dealloc__(memoryview self): + * if self.obj is not None: # <<<<<<<<<<<<<< + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + */ + __pyx_t_1 = (__pyx_v_self->obj != Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":378 + * def __dealloc__(memoryview self): + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) # <<<<<<<<<<<<<< + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + * + */ + __Pyx_ReleaseBuffer((&__pyx_v_self->view)); + + /* "View.MemoryView":377 + * + * def __dealloc__(memoryview self): + * if self.obj is not None: # <<<<<<<<<<<<<< + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":379 + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: # <<<<<<<<<<<<<< + * + * (<__pyx_buffer *> &self.view).obj = NULL + */ + __pyx_t_1 = (((Py_buffer *)(&__pyx_v_self->view))->obj == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":381 + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + * + * (<__pyx_buffer *> &self.view).obj = NULL # <<<<<<<<<<<<<< + * Py_DECREF(Py_None) + * + */ + ((Py_buffer *)(&__pyx_v_self->view))->obj = NULL; + + /* "View.MemoryView":382 + * + * (<__pyx_buffer *> &self.view).obj = NULL + * Py_DECREF(Py_None) # <<<<<<<<<<<<<< + * + * cdef int i + */ + Py_DECREF(Py_None); + + /* "View.MemoryView":379 + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: # <<<<<<<<<<<<<< + * + * (<__pyx_buffer *> &self.view).obj = NULL + */ + } + __pyx_L3:; + + /* "View.MemoryView":386 + * cdef int i + * global __pyx_memoryview_thread_locks_used + * if self.lock != NULL: # <<<<<<<<<<<<<< + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: + */ + __pyx_t_1 = (__pyx_v_self->lock != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":387 + * global __pyx_memoryview_thread_locks_used + * if self.lock != NULL: + * for i in range(__pyx_memoryview_thread_locks_used): # <<<<<<<<<<<<<< + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 + */ + __pyx_t_2 = __pyx_memoryview_thread_locks_used; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":388 + * if self.lock != NULL: + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: + */ + __pyx_t_1 = ((__pyx_memoryview_thread_locks[__pyx_v_i]) == __pyx_v_self->lock); + if (__pyx_t_1) { + + /* "View.MemoryView":389 + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 # <<<<<<<<<<<<<< + * if i != __pyx_memoryview_thread_locks_used: + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + */ + __pyx_memoryview_thread_locks_used = (__pyx_memoryview_thread_locks_used - 1); + + /* "View.MemoryView":390 + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + */ + __pyx_t_1 = (__pyx_v_i != __pyx_memoryview_thread_locks_used); + if (__pyx_t_1) { + + /* "View.MemoryView":392 + * if i != __pyx_memoryview_thread_locks_used: + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) # <<<<<<<<<<<<<< + * break + * else: + */ + __pyx_t_5 = (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]); + __pyx_t_6 = (__pyx_memoryview_thread_locks[__pyx_v_i]); + + /* "View.MemoryView":391 + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + * break + */ + (__pyx_memoryview_thread_locks[__pyx_v_i]) = __pyx_t_5; + (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]) = __pyx_t_6; + + /* "View.MemoryView":390 + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + */ + } + + /* "View.MemoryView":393 + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + * break # <<<<<<<<<<<<<< + * else: + * PyThread_free_lock(self.lock) + */ + goto __pyx_L6_break; + + /* "View.MemoryView":388 + * if self.lock != NULL: + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: + */ + } + } + /*else*/ { + + /* "View.MemoryView":395 + * break + * else: + * PyThread_free_lock(self.lock) # <<<<<<<<<<<<<< + * + * cdef char *get_item_pointer(memoryview self, object index) except NULL: + */ + PyThread_free_lock(__pyx_v_self->lock); + } + __pyx_L6_break:; + + /* "View.MemoryView":386 + * cdef int i + * global __pyx_memoryview_thread_locks_used + * if self.lock != NULL: # <<<<<<<<<<<<<< + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: + */ + } + + /* "View.MemoryView":376 + * self.typeinfo = NULL + * + * def __dealloc__(memoryview self): # <<<<<<<<<<<<<< + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + */ + + /* function exit code */ +} + +/* "View.MemoryView":397 + * PyThread_free_lock(self.lock) + * + * cdef char *get_item_pointer(memoryview self, object index) except NULL: # <<<<<<<<<<<<<< + * cdef Py_ssize_t dim + * cdef char *itemp = self.view.buf + */ + +static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index) { + Py_ssize_t __pyx_v_dim; + char *__pyx_v_itemp; + PyObject *__pyx_v_idx = NULL; + char *__pyx_r; + __Pyx_RefNannyDeclarations + Py_ssize_t __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + Py_ssize_t __pyx_t_3; + PyObject *(*__pyx_t_4)(PyObject *); + PyObject *__pyx_t_5 = NULL; + Py_ssize_t __pyx_t_6; + char *__pyx_t_7; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("get_item_pointer", 1); + + /* "View.MemoryView":399 + * cdef char *get_item_pointer(memoryview self, object index) except NULL: + * cdef Py_ssize_t dim + * cdef char *itemp = self.view.buf # <<<<<<<<<<<<<< + * + * for dim, idx in enumerate(index): + */ + __pyx_v_itemp = ((char *)__pyx_v_self->view.buf); + + /* "View.MemoryView":401 + * cdef char *itemp = self.view.buf + * + * for dim, idx in enumerate(index): # <<<<<<<<<<<<<< + * itemp = pybuffer_index(&self.view, itemp, idx, dim) + * + */ + __pyx_t_1 = 0; + if (likely(PyList_CheckExact(__pyx_v_index)) || PyTuple_CheckExact(__pyx_v_index)) { + __pyx_t_2 = __pyx_v_index; __Pyx_INCREF(__pyx_t_2); + __pyx_t_3 = 0; + __pyx_t_4 = NULL; + } else { + __pyx_t_3 = -1; __pyx_t_2 = PyObject_GetIter(__pyx_v_index); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 401, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = __Pyx_PyObject_GetIterNextFunc(__pyx_t_2); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 401, __pyx_L1_error) + } + for (;;) { + if (likely(!__pyx_t_4)) { + if (likely(PyList_CheckExact(__pyx_t_2))) { + { + Py_ssize_t __pyx_temp = __Pyx_PyList_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #endif + if (__pyx_t_3 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_5 = PyList_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely((0 < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #else + __pyx_t_5 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 401, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + #endif + } else { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #endif + if (__pyx_t_3 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely((0 < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #else + __pyx_t_5 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 401, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + #endif + } + } else { + __pyx_t_5 = __pyx_t_4(__pyx_t_2); + if (unlikely(!__pyx_t_5)) { + PyObject* exc_type = PyErr_Occurred(); + if (exc_type) { + if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear(); + else __PYX_ERR(1, 401, __pyx_L1_error) + } + break; + } + __Pyx_GOTREF(__pyx_t_5); + } + __Pyx_XDECREF_SET(__pyx_v_idx, __pyx_t_5); + __pyx_t_5 = 0; + __pyx_v_dim = __pyx_t_1; + __pyx_t_1 = (__pyx_t_1 + 1); + + /* "View.MemoryView":402 + * + * for dim, idx in enumerate(index): + * itemp = pybuffer_index(&self.view, itemp, idx, dim) # <<<<<<<<<<<<<< + * + * return itemp + */ + __pyx_t_6 = __Pyx_PyIndex_AsSsize_t(__pyx_v_idx); if (unlikely((__pyx_t_6 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 402, __pyx_L1_error) + __pyx_t_7 = __pyx_pybuffer_index((&__pyx_v_self->view), __pyx_v_itemp, __pyx_t_6, __pyx_v_dim); if (unlikely(__pyx_t_7 == ((char *)NULL))) __PYX_ERR(1, 402, __pyx_L1_error) + __pyx_v_itemp = __pyx_t_7; + + /* "View.MemoryView":401 + * cdef char *itemp = self.view.buf + * + * for dim, idx in enumerate(index): # <<<<<<<<<<<<<< + * itemp = pybuffer_index(&self.view, itemp, idx, dim) + * + */ + } + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":404 + * itemp = pybuffer_index(&self.view, itemp, idx, dim) + * + * return itemp # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = __pyx_v_itemp; + goto __pyx_L0; + + /* "View.MemoryView":397 + * PyThread_free_lock(self.lock) + * + * cdef char *get_item_pointer(memoryview self, object index) except NULL: # <<<<<<<<<<<<<< + * cdef Py_ssize_t dim + * cdef char *itemp = self.view.buf + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.memoryview.get_item_pointer", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_idx); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":407 + * + * + * def __getitem__(memoryview self, object index): # <<<<<<<<<<<<<< + * if index is Ellipsis: + * return self + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index); /*proto*/ +static PyObject *__pyx_memoryview___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4__getitem__(((struct __pyx_memoryview_obj *)__pyx_v_self), ((PyObject *)__pyx_v_index)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4__getitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index) { + PyObject *__pyx_v_have_slices = NULL; + PyObject *__pyx_v_indices = NULL; + char *__pyx_v_itemp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + char *__pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__getitem__", 1); + + /* "View.MemoryView":408 + * + * def __getitem__(memoryview self, object index): + * if index is Ellipsis: # <<<<<<<<<<<<<< + * return self + * + */ + __pyx_t_1 = (__pyx_v_index == __pyx_builtin_Ellipsis); + if (__pyx_t_1) { + + /* "View.MemoryView":409 + * def __getitem__(memoryview self, object index): + * if index is Ellipsis: + * return self # <<<<<<<<<<<<<< + * + * have_slices, indices = _unellipsify(index, self.view.ndim) + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_self); + __pyx_r = ((PyObject *)__pyx_v_self); + goto __pyx_L0; + + /* "View.MemoryView":408 + * + * def __getitem__(memoryview self, object index): + * if index is Ellipsis: # <<<<<<<<<<<<<< + * return self + * + */ + } + + /* "View.MemoryView":411 + * return self + * + * have_slices, indices = _unellipsify(index, self.view.ndim) # <<<<<<<<<<<<<< + * + * cdef char *itemp + */ + __pyx_t_2 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 411, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + if (likely(__pyx_t_2 != Py_None)) { + PyObject* sequence = __pyx_t_2; + Py_ssize_t size = __Pyx_PySequence_SIZE(sequence); + if (unlikely(size != 2)) { + if (size > 2) __Pyx_RaiseTooManyValuesError(2); + else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size); + __PYX_ERR(1, 411, __pyx_L1_error) + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_3 = PyTuple_GET_ITEM(sequence, 0); + __pyx_t_4 = PyTuple_GET_ITEM(sequence, 1); + __Pyx_INCREF(__pyx_t_3); + __Pyx_INCREF(__pyx_t_4); + #else + __pyx_t_3 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 411, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 411, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + #endif + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + } else { + __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 411, __pyx_L1_error) + } + __pyx_v_have_slices = __pyx_t_3; + __pyx_t_3 = 0; + __pyx_v_indices = __pyx_t_4; + __pyx_t_4 = 0; + + /* "View.MemoryView":414 + * + * cdef char *itemp + * if have_slices: # <<<<<<<<<<<<<< + * return memview_slice(self, indices) + * else: + */ + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 414, __pyx_L1_error) + if (__pyx_t_1) { + + /* "View.MemoryView":415 + * cdef char *itemp + * if have_slices: + * return memview_slice(self, indices) # <<<<<<<<<<<<<< + * else: + * itemp = self.get_item_pointer(indices) + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = ((PyObject *)__pyx_memview_slice(__pyx_v_self, __pyx_v_indices)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 415, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":414 + * + * cdef char *itemp + * if have_slices: # <<<<<<<<<<<<<< + * return memview_slice(self, indices) + * else: + */ + } + + /* "View.MemoryView":417 + * return memview_slice(self, indices) + * else: + * itemp = self.get_item_pointer(indices) # <<<<<<<<<<<<<< + * return self.convert_item_to_object(itemp) + * + */ + /*else*/ { + __pyx_t_5 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_indices); if (unlikely(__pyx_t_5 == ((char *)NULL))) __PYX_ERR(1, 417, __pyx_L1_error) + __pyx_v_itemp = __pyx_t_5; + + /* "View.MemoryView":418 + * else: + * itemp = self.get_item_pointer(indices) + * return self.convert_item_to_object(itemp) # <<<<<<<<<<<<<< + * + * def __setitem__(memoryview self, object index, object value): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->convert_item_to_object(__pyx_v_self, __pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 418, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + } + + /* "View.MemoryView":407 + * + * + * def __getitem__(memoryview self, object index): # <<<<<<<<<<<<<< + * if index is Ellipsis: + * return self + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.memoryview.__getitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_have_slices); + __Pyx_XDECREF(__pyx_v_indices); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":420 + * return self.convert_item_to_object(itemp) + * + * def __setitem__(memoryview self, object index, object value): # <<<<<<<<<<<<<< + * if self.view.readonly: + * raise TypeError, "Cannot assign to read-only memoryview" + */ + +/* Python wrapper */ +static int __pyx_memoryview___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value); /*proto*/ +static int __pyx_memoryview___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setitem__(((struct __pyx_memoryview_obj *)__pyx_v_self), ((PyObject *)__pyx_v_index), ((PyObject *)__pyx_v_value)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value) { + PyObject *__pyx_v_have_slices = NULL; + PyObject *__pyx_v_obj = NULL; + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setitem__", 0); + __Pyx_INCREF(__pyx_v_index); + + /* "View.MemoryView":421 + * + * def __setitem__(memoryview self, object index, object value): + * if self.view.readonly: # <<<<<<<<<<<<<< + * raise TypeError, "Cannot assign to read-only memoryview" + * + */ + if (unlikely(__pyx_v_self->view.readonly)) { + + /* "View.MemoryView":422 + * def __setitem__(memoryview self, object index, object value): + * if self.view.readonly: + * raise TypeError, "Cannot assign to read-only memoryview" # <<<<<<<<<<<<<< + * + * have_slices, index = _unellipsify(index, self.view.ndim) + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_Cannot_assign_to_read_only_memor, 0, 0); + __PYX_ERR(1, 422, __pyx_L1_error) + + /* "View.MemoryView":421 + * + * def __setitem__(memoryview self, object index, object value): + * if self.view.readonly: # <<<<<<<<<<<<<< + * raise TypeError, "Cannot assign to read-only memoryview" + * + */ + } + + /* "View.MemoryView":424 + * raise TypeError, "Cannot assign to read-only memoryview" + * + * have_slices, index = _unellipsify(index, self.view.ndim) # <<<<<<<<<<<<<< + * + * if have_slices: + */ + __pyx_t_1 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 424, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + if (likely(__pyx_t_1 != Py_None)) { + PyObject* sequence = __pyx_t_1; + Py_ssize_t size = __Pyx_PySequence_SIZE(sequence); + if (unlikely(size != 2)) { + if (size > 2) __Pyx_RaiseTooManyValuesError(2); + else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size); + __PYX_ERR(1, 424, __pyx_L1_error) + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_2 = PyTuple_GET_ITEM(sequence, 0); + __pyx_t_3 = PyTuple_GET_ITEM(sequence, 1); + __Pyx_INCREF(__pyx_t_2); + __Pyx_INCREF(__pyx_t_3); + #else + __pyx_t_2 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 424, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 424, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + #endif + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + } else { + __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 424, __pyx_L1_error) + } + __pyx_v_have_slices = __pyx_t_2; + __pyx_t_2 = 0; + __Pyx_DECREF_SET(__pyx_v_index, __pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":426 + * have_slices, index = _unellipsify(index, self.view.ndim) + * + * if have_slices: # <<<<<<<<<<<<<< + * obj = self.is_slice(value) + * if obj is not None: + */ + __pyx_t_4 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely((__pyx_t_4 < 0))) __PYX_ERR(1, 426, __pyx_L1_error) + if (__pyx_t_4) { + + /* "View.MemoryView":427 + * + * if have_slices: + * obj = self.is_slice(value) # <<<<<<<<<<<<<< + * if obj is not None: + * self.setitem_slice_assignment(self[index], obj) + */ + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->is_slice(__pyx_v_self, __pyx_v_value); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 427, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_obj = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":428 + * if have_slices: + * obj = self.is_slice(value) + * if obj is not None: # <<<<<<<<<<<<<< + * self.setitem_slice_assignment(self[index], obj) + * else: + */ + __pyx_t_4 = (__pyx_v_obj != Py_None); + if (__pyx_t_4) { + + /* "View.MemoryView":429 + * obj = self.is_slice(value) + * if obj is not None: + * self.setitem_slice_assignment(self[index], obj) # <<<<<<<<<<<<<< + * else: + * self.setitem_slice_assign_scalar(self[index], value) + */ + __pyx_t_1 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 429, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assignment(__pyx_v_self, __pyx_t_1, __pyx_v_obj); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 429, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + + /* "View.MemoryView":428 + * if have_slices: + * obj = self.is_slice(value) + * if obj is not None: # <<<<<<<<<<<<<< + * self.setitem_slice_assignment(self[index], obj) + * else: + */ + goto __pyx_L5; + } + + /* "View.MemoryView":431 + * self.setitem_slice_assignment(self[index], obj) + * else: + * self.setitem_slice_assign_scalar(self[index], value) # <<<<<<<<<<<<<< + * else: + * self.setitem_indexed(index, value) + */ + /*else*/ { + __pyx_t_3 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 431, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_memoryview_type))))) __PYX_ERR(1, 431, __pyx_L1_error) + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assign_scalar(__pyx_v_self, ((struct __pyx_memoryview_obj *)__pyx_t_3), __pyx_v_value); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 431, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + } + __pyx_L5:; + + /* "View.MemoryView":426 + * have_slices, index = _unellipsify(index, self.view.ndim) + * + * if have_slices: # <<<<<<<<<<<<<< + * obj = self.is_slice(value) + * if obj is not None: + */ + goto __pyx_L4; + } + + /* "View.MemoryView":433 + * self.setitem_slice_assign_scalar(self[index], value) + * else: + * self.setitem_indexed(index, value) # <<<<<<<<<<<<<< + * + * cdef is_slice(self, obj): + */ + /*else*/ { + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_indexed(__pyx_v_self, __pyx_v_index, __pyx_v_value); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 433, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + } + __pyx_L4:; + + /* "View.MemoryView":420 + * return self.convert_item_to_object(itemp) + * + * def __setitem__(memoryview self, object index, object value): # <<<<<<<<<<<<<< + * if self.view.readonly: + * raise TypeError, "Cannot assign to read-only memoryview" + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.__setitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_have_slices); + __Pyx_XDECREF(__pyx_v_obj); + __Pyx_XDECREF(__pyx_v_index); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":435 + * self.setitem_indexed(index, value) + * + * cdef is_slice(self, obj): # <<<<<<<<<<<<<< + * if not isinstance(obj, memoryview): + * try: + */ + +static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_t_9; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("is_slice", 0); + __Pyx_INCREF(__pyx_v_obj); + + /* "View.MemoryView":436 + * + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): # <<<<<<<<<<<<<< + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + */ + __pyx_t_1 = __Pyx_TypeCheck(__pyx_v_obj, __pyx_memoryview_type); + __pyx_t_2 = (!__pyx_t_1); + if (__pyx_t_2) { + + /* "View.MemoryView":437 + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): + * try: # <<<<<<<<<<<<<< + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_3, &__pyx_t_4, &__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_4); + __Pyx_XGOTREF(__pyx_t_5); + /*try:*/ { + + /* "View.MemoryView":438 + * if not isinstance(obj, memoryview): + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, # <<<<<<<<<<<<<< + * self.dtype_is_object) + * except TypeError: + */ + __pyx_t_6 = __Pyx_PyInt_From_int(((__pyx_v_self->flags & (~PyBUF_WRITABLE)) | PyBUF_ANY_CONTIGUOUS)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 438, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_6); + + /* "View.MemoryView":439 + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) # <<<<<<<<<<<<<< + * except TypeError: + * return None + */ + __pyx_t_7 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 439, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_7); + + /* "View.MemoryView":438 + * if not isinstance(obj, memoryview): + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, # <<<<<<<<<<<<<< + * self.dtype_is_object) + * except TypeError: + */ + __pyx_t_8 = PyTuple_New(3); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 438, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_INCREF(__pyx_v_obj); + __Pyx_GIVEREF(__pyx_v_obj); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_8, 0, __pyx_v_obj)) __PYX_ERR(1, 438, __pyx_L4_error); + __Pyx_GIVEREF(__pyx_t_6); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_8, 1, __pyx_t_6)) __PYX_ERR(1, 438, __pyx_L4_error); + __Pyx_GIVEREF(__pyx_t_7); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_8, 2, __pyx_t_7)) __PYX_ERR(1, 438, __pyx_L4_error); + __pyx_t_6 = 0; + __pyx_t_7 = 0; + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_8, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 438, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __Pyx_DECREF_SET(__pyx_v_obj, __pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":437 + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): + * try: # <<<<<<<<<<<<<< + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + */ + } + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + goto __pyx_L9_try_end; + __pyx_L4_error:; + __Pyx_XDECREF(__pyx_t_6); __pyx_t_6 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_XDECREF(__pyx_t_8); __pyx_t_8 = 0; + + /* "View.MemoryView":440 + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + * except TypeError: # <<<<<<<<<<<<<< + * return None + * + */ + __pyx_t_9 = __Pyx_PyErr_ExceptionMatches(__pyx_builtin_TypeError); + if (__pyx_t_9) { + __Pyx_AddTraceback("View.MemoryView.memoryview.is_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_7, &__pyx_t_8, &__pyx_t_6) < 0) __PYX_ERR(1, 440, __pyx_L6_except_error) + __Pyx_XGOTREF(__pyx_t_7); + __Pyx_XGOTREF(__pyx_t_8); + __Pyx_XGOTREF(__pyx_t_6); + + /* "View.MemoryView":441 + * self.dtype_is_object) + * except TypeError: + * return None # <<<<<<<<<<<<<< + * + * return obj + */ + __Pyx_XDECREF(__pyx_r); + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L7_except_return; + } + goto __pyx_L6_except_error; + + /* "View.MemoryView":437 + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): + * try: # <<<<<<<<<<<<<< + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + */ + __pyx_L6_except_error:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_XGIVEREF(__pyx_t_5); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_4, __pyx_t_5); + goto __pyx_L1_error; + __pyx_L7_except_return:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_XGIVEREF(__pyx_t_5); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_4, __pyx_t_5); + goto __pyx_L0; + __pyx_L9_try_end:; + } + + /* "View.MemoryView":436 + * + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): # <<<<<<<<<<<<<< + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + */ + } + + /* "View.MemoryView":443 + * return None + * + * return obj # <<<<<<<<<<<<<< + * + * cdef setitem_slice_assignment(self, dst, src): + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_obj); + __pyx_r = __pyx_v_obj; + goto __pyx_L0; + + /* "View.MemoryView":435 + * self.setitem_indexed(index, value) + * + * cdef is_slice(self, obj): # <<<<<<<<<<<<<< + * if not isinstance(obj, memoryview): + * try: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("View.MemoryView.memoryview.is_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_obj); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":445 + * return obj + * + * cdef setitem_slice_assignment(self, dst, src): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice dst_slice + * cdef __Pyx_memviewslice src_slice + */ + +static PyObject *__pyx_memoryview_setitem_slice_assignment(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_dst, PyObject *__pyx_v_src) { + __Pyx_memviewslice __pyx_v_dst_slice; + __Pyx_memviewslice __pyx_v_src_slice; + __Pyx_memviewslice __pyx_v_msrc; + __Pyx_memviewslice __pyx_v_mdst; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_t_3; + int __pyx_t_4; + int __pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("setitem_slice_assignment", 1); + + /* "View.MemoryView":448 + * cdef __Pyx_memviewslice dst_slice + * cdef __Pyx_memviewslice src_slice + * cdef __Pyx_memviewslice msrc = get_slice_from_memview(src, &src_slice)[0] # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice mdst = get_slice_from_memview(dst, &dst_slice)[0] + * + */ + if (!(likely(((__pyx_v_src) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_src, __pyx_memoryview_type))))) __PYX_ERR(1, 448, __pyx_L1_error) + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_src), (&__pyx_v_src_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 448, __pyx_L1_error) + __pyx_v_msrc = (__pyx_t_1[0]); + + /* "View.MemoryView":449 + * cdef __Pyx_memviewslice src_slice + * cdef __Pyx_memviewslice msrc = get_slice_from_memview(src, &src_slice)[0] + * cdef __Pyx_memviewslice mdst = get_slice_from_memview(dst, &dst_slice)[0] # <<<<<<<<<<<<<< + * + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) + */ + if (!(likely(((__pyx_v_dst) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_dst, __pyx_memoryview_type))))) __PYX_ERR(1, 449, __pyx_L1_error) + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_dst), (&__pyx_v_dst_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 449, __pyx_L1_error) + __pyx_v_mdst = (__pyx_t_1[0]); + + /* "View.MemoryView":451 + * cdef __Pyx_memviewslice mdst = get_slice_from_memview(dst, &dst_slice)[0] + * + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) # <<<<<<<<<<<<<< + * + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): + */ + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_v_src, __pyx_n_s_ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = __Pyx_PyInt_As_int(__pyx_t_2); if (unlikely((__pyx_t_3 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_v_dst, __pyx_n_s_ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_t_2); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_5 = __pyx_memoryview_copy_contents(__pyx_v_msrc, __pyx_v_mdst, __pyx_t_3, __pyx_t_4, __pyx_v_self->dtype_is_object); if (unlikely(__pyx_t_5 == ((int)-1))) __PYX_ERR(1, 451, __pyx_L1_error) + + /* "View.MemoryView":445 + * return obj + * + * cdef setitem_slice_assignment(self, dst, src): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice dst_slice + * cdef __Pyx_memviewslice src_slice + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.setitem_slice_assignment", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":453 + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) + * + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): # <<<<<<<<<<<<<< + * cdef int array[128] + * cdef void *tmp = NULL + */ + +static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memoryview_obj *__pyx_v_self, struct __pyx_memoryview_obj *__pyx_v_dst, PyObject *__pyx_v_value) { + int __pyx_v_array[0x80]; + void *__pyx_v_tmp; + void *__pyx_v_item; + __Pyx_memviewslice *__pyx_v_dst_slice; + __Pyx_memviewslice __pyx_v_tmp_slice; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + int __pyx_t_5; + char const *__pyx_t_6; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + PyObject *__pyx_t_9 = NULL; + PyObject *__pyx_t_10 = NULL; + PyObject *__pyx_t_11 = NULL; + PyObject *__pyx_t_12 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("setitem_slice_assign_scalar", 1); + + /* "View.MemoryView":455 + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): + * cdef int array[128] + * cdef void *tmp = NULL # <<<<<<<<<<<<<< + * cdef void *item + * + */ + __pyx_v_tmp = NULL; + + /* "View.MemoryView":460 + * cdef __Pyx_memviewslice *dst_slice + * cdef __Pyx_memviewslice tmp_slice + * dst_slice = get_slice_from_memview(dst, &tmp_slice) # <<<<<<<<<<<<<< + * + * if self.view.itemsize > sizeof(array): + */ + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_dst, (&__pyx_v_tmp_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 460, __pyx_L1_error) + __pyx_v_dst_slice = __pyx_t_1; + + /* "View.MemoryView":462 + * dst_slice = get_slice_from_memview(dst, &tmp_slice) + * + * if self.view.itemsize > sizeof(array): # <<<<<<<<<<<<<< + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: + */ + __pyx_t_2 = (((size_t)__pyx_v_self->view.itemsize) > (sizeof(__pyx_v_array))); + if (__pyx_t_2) { + + /* "View.MemoryView":463 + * + * if self.view.itemsize > sizeof(array): + * tmp = PyMem_Malloc(self.view.itemsize) # <<<<<<<<<<<<<< + * if tmp == NULL: + * raise MemoryError + */ + __pyx_v_tmp = PyMem_Malloc(__pyx_v_self->view.itemsize); + + /* "View.MemoryView":464 + * if self.view.itemsize > sizeof(array): + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * item = tmp + */ + __pyx_t_2 = (__pyx_v_tmp == NULL); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":465 + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: + * raise MemoryError # <<<<<<<<<<<<<< + * item = tmp + * else: + */ + PyErr_NoMemory(); __PYX_ERR(1, 465, __pyx_L1_error) + + /* "View.MemoryView":464 + * if self.view.itemsize > sizeof(array): + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * item = tmp + */ + } + + /* "View.MemoryView":466 + * if tmp == NULL: + * raise MemoryError + * item = tmp # <<<<<<<<<<<<<< + * else: + * item = array + */ + __pyx_v_item = __pyx_v_tmp; + + /* "View.MemoryView":462 + * dst_slice = get_slice_from_memview(dst, &tmp_slice) + * + * if self.view.itemsize > sizeof(array): # <<<<<<<<<<<<<< + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":468 + * item = tmp + * else: + * item = array # <<<<<<<<<<<<<< + * + * try: + */ + /*else*/ { + __pyx_v_item = ((void *)__pyx_v_array); + } + __pyx_L3:; + + /* "View.MemoryView":470 + * item = array + * + * try: # <<<<<<<<<<<<<< + * if self.dtype_is_object: + * ( item)[0] = value + */ + /*try:*/ { + + /* "View.MemoryView":471 + * + * try: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * ( item)[0] = value + * else: + */ + if (__pyx_v_self->dtype_is_object) { + + /* "View.MemoryView":472 + * try: + * if self.dtype_is_object: + * ( item)[0] = value # <<<<<<<<<<<<<< + * else: + * self.assign_item_from_object( item, value) + */ + (((PyObject **)__pyx_v_item)[0]) = ((PyObject *)__pyx_v_value); + + /* "View.MemoryView":471 + * + * try: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * ( item)[0] = value + * else: + */ + goto __pyx_L8; + } + + /* "View.MemoryView":474 + * ( item)[0] = value + * else: + * self.assign_item_from_object( item, value) # <<<<<<<<<<<<<< + * + * + */ + /*else*/ { + __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, ((char *)__pyx_v_item), __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 474, __pyx_L6_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + __pyx_L8:; + + /* "View.MemoryView":478 + * + * + * if self.view.suboffsets != NULL: # <<<<<<<<<<<<<< + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, + */ + __pyx_t_2 = (__pyx_v_self->view.suboffsets != NULL); + if (__pyx_t_2) { + + /* "View.MemoryView":479 + * + * if self.view.suboffsets != NULL: + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) # <<<<<<<<<<<<<< + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, + * item, self.dtype_is_object) + */ + __pyx_t_4 = assert_direct_dimensions(__pyx_v_self->view.suboffsets, __pyx_v_self->view.ndim); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 479, __pyx_L6_error) + + /* "View.MemoryView":478 + * + * + * if self.view.suboffsets != NULL: # <<<<<<<<<<<<<< + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, + */ + } + + /* "View.MemoryView":480 + * if self.view.suboffsets != NULL: + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, # <<<<<<<<<<<<<< + * item, self.dtype_is_object) + * finally: + */ + __pyx_memoryview_slice_assign_scalar(__pyx_v_dst_slice, __pyx_v_dst->view.ndim, __pyx_v_self->view.itemsize, __pyx_v_item, __pyx_v_self->dtype_is_object); + } + + /* "View.MemoryView":483 + * item, self.dtype_is_object) + * finally: + * PyMem_Free(tmp) # <<<<<<<<<<<<<< + * + * cdef setitem_indexed(self, index, value): + */ + /*finally:*/ { + /*normal exit:*/{ + PyMem_Free(__pyx_v_tmp); + goto __pyx_L7; + } + __pyx_L6_error:; + /*exception exit:*/{ + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __pyx_t_7 = 0; __pyx_t_8 = 0; __pyx_t_9 = 0; __pyx_t_10 = 0; __pyx_t_11 = 0; __pyx_t_12 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + if (PY_MAJOR_VERSION >= 3) __Pyx_ExceptionSwap(&__pyx_t_10, &__pyx_t_11, &__pyx_t_12); + if ((PY_MAJOR_VERSION < 3) || unlikely(__Pyx_GetException(&__pyx_t_7, &__pyx_t_8, &__pyx_t_9) < 0)) __Pyx_ErrFetch(&__pyx_t_7, &__pyx_t_8, &__pyx_t_9); + __Pyx_XGOTREF(__pyx_t_7); + __Pyx_XGOTREF(__pyx_t_8); + __Pyx_XGOTREF(__pyx_t_9); + __Pyx_XGOTREF(__pyx_t_10); + __Pyx_XGOTREF(__pyx_t_11); + __Pyx_XGOTREF(__pyx_t_12); + __pyx_t_4 = __pyx_lineno; __pyx_t_5 = __pyx_clineno; __pyx_t_6 = __pyx_filename; + { + PyMem_Free(__pyx_v_tmp); + } + if (PY_MAJOR_VERSION >= 3) { + __Pyx_XGIVEREF(__pyx_t_10); + __Pyx_XGIVEREF(__pyx_t_11); + __Pyx_XGIVEREF(__pyx_t_12); + __Pyx_ExceptionReset(__pyx_t_10, __pyx_t_11, __pyx_t_12); + } + __Pyx_XGIVEREF(__pyx_t_7); + __Pyx_XGIVEREF(__pyx_t_8); + __Pyx_XGIVEREF(__pyx_t_9); + __Pyx_ErrRestore(__pyx_t_7, __pyx_t_8, __pyx_t_9); + __pyx_t_7 = 0; __pyx_t_8 = 0; __pyx_t_9 = 0; __pyx_t_10 = 0; __pyx_t_11 = 0; __pyx_t_12 = 0; + __pyx_lineno = __pyx_t_4; __pyx_clineno = __pyx_t_5; __pyx_filename = __pyx_t_6; + goto __pyx_L1_error; + } + __pyx_L7:; + } + + /* "View.MemoryView":453 + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) + * + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): # <<<<<<<<<<<<<< + * cdef int array[128] + * cdef void *tmp = NULL + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.setitem_slice_assign_scalar", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":485 + * PyMem_Free(tmp) + * + * cdef setitem_indexed(self, index, value): # <<<<<<<<<<<<<< + * cdef char *itemp = self.get_item_pointer(index) + * self.assign_item_from_object(itemp, value) + */ + +static PyObject *__pyx_memoryview_setitem_indexed(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value) { + char *__pyx_v_itemp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + char *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("setitem_indexed", 1); + + /* "View.MemoryView":486 + * + * cdef setitem_indexed(self, index, value): + * cdef char *itemp = self.get_item_pointer(index) # <<<<<<<<<<<<<< + * self.assign_item_from_object(itemp, value) + * + */ + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_index); if (unlikely(__pyx_t_1 == ((char *)NULL))) __PYX_ERR(1, 486, __pyx_L1_error) + __pyx_v_itemp = __pyx_t_1; + + /* "View.MemoryView":487 + * cdef setitem_indexed(self, index, value): + * cdef char *itemp = self.get_item_pointer(index) + * self.assign_item_from_object(itemp, value) # <<<<<<<<<<<<<< + * + * cdef convert_item_to_object(self, char *itemp): + */ + __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 487, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":485 + * PyMem_Free(tmp) + * + * cdef setitem_indexed(self, index, value): # <<<<<<<<<<<<<< + * cdef char *itemp = self.get_item_pointer(index) + * self.assign_item_from_object(itemp, value) + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.setitem_indexed", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":489 + * self.assign_item_from_object(itemp, value) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + +static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp) { + PyObject *__pyx_v_struct = NULL; + PyObject *__pyx_v_bytesitem = 0; + PyObject *__pyx_v_result = NULL; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + unsigned int __pyx_t_8; + Py_ssize_t __pyx_t_9; + int __pyx_t_10; + int __pyx_t_11; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("convert_item_to_object", 1); + + /* "View.MemoryView":492 + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + * import struct # <<<<<<<<<<<<<< + * cdef bytes bytesitem + * + */ + __pyx_t_1 = __Pyx_ImportDottedModule(__pyx_n_s_struct, NULL); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 492, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_struct = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":495 + * cdef bytes bytesitem + * + * bytesitem = itemp[:self.view.itemsize] # <<<<<<<<<<<<<< + * try: + * result = struct.unpack(self.view.format, bytesitem) + */ + __pyx_t_1 = __Pyx_PyBytes_FromStringAndSize(__pyx_v_itemp + 0, __pyx_v_self->view.itemsize - 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 495, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_bytesitem = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":496 + * + * bytesitem = itemp[:self.view.itemsize] + * try: # <<<<<<<<<<<<<< + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_2, &__pyx_t_3, &__pyx_t_4); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_4); + /*try:*/ { + + /* "View.MemoryView":497 + * bytesitem = itemp[:self.view.itemsize] + * try: + * result = struct.unpack(self.view.format, bytesitem) # <<<<<<<<<<<<<< + * except struct.error: + * raise ValueError, "Unable to convert item to object" + */ + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_unpack); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 497, __pyx_L3_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_6 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 497, __pyx_L3_error) + __Pyx_GOTREF(__pyx_t_6); + __pyx_t_7 = NULL; + __pyx_t_8 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_5))) { + __pyx_t_7 = PyMethod_GET_SELF(__pyx_t_5); + if (likely(__pyx_t_7)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_5); + __Pyx_INCREF(__pyx_t_7); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_5, function); + __pyx_t_8 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[3] = {__pyx_t_7, __pyx_t_6, __pyx_v_bytesitem}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_5, __pyx_callargs+1-__pyx_t_8, 2+__pyx_t_8); + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 497, __pyx_L3_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + __pyx_v_result = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":496 + * + * bytesitem = itemp[:self.view.itemsize] + * try: # <<<<<<<<<<<<<< + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + */ + } + + /* "View.MemoryView":501 + * raise ValueError, "Unable to convert item to object" + * else: + * if len(self.view.format) == 1: # <<<<<<<<<<<<<< + * return result[0] + * return result + */ + /*else:*/ { + __pyx_t_9 = __Pyx_ssize_strlen(__pyx_v_self->view.format); if (unlikely(__pyx_t_9 == ((Py_ssize_t)-1))) __PYX_ERR(1, 501, __pyx_L5_except_error) + __pyx_t_10 = (__pyx_t_9 == 1); + if (__pyx_t_10) { + + /* "View.MemoryView":502 + * else: + * if len(self.view.format) == 1: + * return result[0] # <<<<<<<<<<<<<< + * return result + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_GetItemInt(__pyx_v_result, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 502, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L6_except_return; + + /* "View.MemoryView":501 + * raise ValueError, "Unable to convert item to object" + * else: + * if len(self.view.format) == 1: # <<<<<<<<<<<<<< + * return result[0] + * return result + */ + } + + /* "View.MemoryView":503 + * if len(self.view.format) == 1: + * return result[0] + * return result # <<<<<<<<<<<<<< + * + * cdef assign_item_from_object(self, char *itemp, object value): + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_result); + __pyx_r = __pyx_v_result; + goto __pyx_L6_except_return; + } + __pyx_L3_error:; + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_6); __pyx_t_6 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":498 + * try: + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: # <<<<<<<<<<<<<< + * raise ValueError, "Unable to convert item to object" + * else: + */ + __Pyx_ErrFetch(&__pyx_t_1, &__pyx_t_5, &__pyx_t_6); + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_error); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 498, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_11 = __Pyx_PyErr_GivenExceptionMatches(__pyx_t_1, __pyx_t_7); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_ErrRestore(__pyx_t_1, __pyx_t_5, __pyx_t_6); + __pyx_t_1 = 0; __pyx_t_5 = 0; __pyx_t_6 = 0; + if (__pyx_t_11) { + __Pyx_AddTraceback("View.MemoryView.memoryview.convert_item_to_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_6, &__pyx_t_5, &__pyx_t_1) < 0) __PYX_ERR(1, 498, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_1); + + /* "View.MemoryView":499 + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + * raise ValueError, "Unable to convert item to object" # <<<<<<<<<<<<<< + * else: + * if len(self.view.format) == 1: + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Unable_to_convert_item_to_object, 0, 0); + __PYX_ERR(1, 499, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "View.MemoryView":496 + * + * bytesitem = itemp[:self.view.itemsize] + * try: # <<<<<<<<<<<<<< + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_ExceptionReset(__pyx_t_2, __pyx_t_3, __pyx_t_4); + goto __pyx_L1_error; + __pyx_L6_except_return:; + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_ExceptionReset(__pyx_t_2, __pyx_t_3, __pyx_t_4); + goto __pyx_L0; + } + + /* "View.MemoryView":489 + * self.assign_item_from_object(itemp, value) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_AddTraceback("View.MemoryView.memoryview.convert_item_to_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_struct); + __Pyx_XDECREF(__pyx_v_bytesitem); + __Pyx_XDECREF(__pyx_v_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":505 + * return result + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + +static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value) { + PyObject *__pyx_v_struct = NULL; + char __pyx_v_c; + PyObject *__pyx_v_bytesvalue = 0; + Py_ssize_t __pyx_v_i; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + unsigned int __pyx_t_6; + Py_ssize_t __pyx_t_7; + PyObject *__pyx_t_8 = NULL; + char *__pyx_t_9; + char *__pyx_t_10; + char *__pyx_t_11; + char *__pyx_t_12; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("assign_item_from_object", 1); + + /* "View.MemoryView":508 + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + * import struct # <<<<<<<<<<<<<< + * cdef char c + * cdef bytes bytesvalue + */ + __pyx_t_1 = __Pyx_ImportDottedModule(__pyx_n_s_struct, NULL); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 508, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_struct = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":513 + * cdef Py_ssize_t i + * + * if isinstance(value, tuple): # <<<<<<<<<<<<<< + * bytesvalue = struct.pack(self.view.format, *value) + * else: + */ + __pyx_t_2 = PyTuple_Check(__pyx_v_value); + if (__pyx_t_2) { + + /* "View.MemoryView":514 + * + * if isinstance(value, tuple): + * bytesvalue = struct.pack(self.view.format, *value) # <<<<<<<<<<<<<< + * else: + * bytesvalue = struct.pack(self.view.format, value) + */ + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = PyTuple_New(1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_3 = __Pyx_PySequence_Tuple(__pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_5 = PyNumber_Add(__pyx_t_4, __pyx_t_3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_t_3 = __Pyx_PyObject_Call(__pyx_t_1, __pyx_t_5, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + if (!(likely(PyBytes_CheckExact(__pyx_t_3))||((__pyx_t_3) == Py_None) || __Pyx_RaiseUnexpectedTypeError("bytes", __pyx_t_3))) __PYX_ERR(1, 514, __pyx_L1_error) + __pyx_v_bytesvalue = ((PyObject*)__pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":513 + * cdef Py_ssize_t i + * + * if isinstance(value, tuple): # <<<<<<<<<<<<<< + * bytesvalue = struct.pack(self.view.format, *value) + * else: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":516 + * bytesvalue = struct.pack(self.view.format, *value) + * else: + * bytesvalue = struct.pack(self.view.format, value) # <<<<<<<<<<<<<< + * + * for i, c in enumerate(bytesvalue): + */ + /*else*/ { + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 516, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_1 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 516, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_4 = NULL; + __pyx_t_6 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_5))) { + __pyx_t_4 = PyMethod_GET_SELF(__pyx_t_5); + if (likely(__pyx_t_4)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_5); + __Pyx_INCREF(__pyx_t_4); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_5, function); + __pyx_t_6 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[3] = {__pyx_t_4, __pyx_t_1, __pyx_v_value}; + __pyx_t_3 = __Pyx_PyObject_FastCall(__pyx_t_5, __pyx_callargs+1-__pyx_t_6, 2+__pyx_t_6); + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 516, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + if (!(likely(PyBytes_CheckExact(__pyx_t_3))||((__pyx_t_3) == Py_None) || __Pyx_RaiseUnexpectedTypeError("bytes", __pyx_t_3))) __PYX_ERR(1, 516, __pyx_L1_error) + __pyx_v_bytesvalue = ((PyObject*)__pyx_t_3); + __pyx_t_3 = 0; + } + __pyx_L3:; + + /* "View.MemoryView":518 + * bytesvalue = struct.pack(self.view.format, value) + * + * for i, c in enumerate(bytesvalue): # <<<<<<<<<<<<<< + * itemp[i] = c + * + */ + __pyx_t_7 = 0; + if (unlikely(__pyx_v_bytesvalue == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' is not iterable"); + __PYX_ERR(1, 518, __pyx_L1_error) + } + __Pyx_INCREF(__pyx_v_bytesvalue); + __pyx_t_8 = __pyx_v_bytesvalue; + __pyx_t_10 = PyBytes_AS_STRING(__pyx_t_8); + __pyx_t_11 = (__pyx_t_10 + PyBytes_GET_SIZE(__pyx_t_8)); + for (__pyx_t_12 = __pyx_t_10; __pyx_t_12 < __pyx_t_11; __pyx_t_12++) { + __pyx_t_9 = __pyx_t_12; + __pyx_v_c = (__pyx_t_9[0]); + + /* "View.MemoryView":519 + * + * for i, c in enumerate(bytesvalue): + * itemp[i] = c # <<<<<<<<<<<<<< + * + * @cname('getbuffer') + */ + __pyx_v_i = __pyx_t_7; + + /* "View.MemoryView":518 + * bytesvalue = struct.pack(self.view.format, value) + * + * for i, c in enumerate(bytesvalue): # <<<<<<<<<<<<<< + * itemp[i] = c + * + */ + __pyx_t_7 = (__pyx_t_7 + 1); + + /* "View.MemoryView":519 + * + * for i, c in enumerate(bytesvalue): + * itemp[i] = c # <<<<<<<<<<<<<< + * + * @cname('getbuffer') + */ + (__pyx_v_itemp[__pyx_v_i]) = __pyx_v_c; + } + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + + /* "View.MemoryView":505 + * return result + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("View.MemoryView.memoryview.assign_item_from_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_struct); + __Pyx_XDECREF(__pyx_v_bytesvalue); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":521 + * itemp[i] = c + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: + */ + +/* Python wrapper */ +CYTHON_UNUSED static int __pyx_memoryview_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +CYTHON_UNUSED static int __pyx_memoryview_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getbuffer__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbuffer__(((struct __pyx_memoryview_obj *)__pyx_v_self), ((Py_buffer *)__pyx_v_info), ((int)__pyx_v_flags)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbuffer__(struct __pyx_memoryview_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + Py_ssize_t *__pyx_t_3; + char *__pyx_t_4; + void *__pyx_t_5; + int __pyx_t_6; + Py_ssize_t __pyx_t_7; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + if (unlikely(__pyx_v_info == NULL)) { + PyErr_SetString(PyExc_BufferError, "PyObject_GetBuffer: view==NULL argument is obsolete"); + return -1; + } + __Pyx_RefNannySetupContext("__getbuffer__", 0); + __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(__pyx_v_info->obj); + + /* "View.MemoryView":523 + * @cname('getbuffer') + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: # <<<<<<<<<<<<<< + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + */ + __pyx_t_2 = ((__pyx_v_flags & PyBUF_WRITABLE) != 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_1 = __pyx_v_self->view.readonly; + __pyx_L4_bool_binop_done:; + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":524 + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: + * raise ValueError, "Cannot create writable memory view from read-only memoryview" # <<<<<<<<<<<<<< + * + * if flags & PyBUF_ND: + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Cannot_create_writable_memory_vi, 0, 0); + __PYX_ERR(1, 524, __pyx_L1_error) + + /* "View.MemoryView":523 + * @cname('getbuffer') + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: # <<<<<<<<<<<<<< + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + */ + } + + /* "View.MemoryView":526 + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + * if flags & PyBUF_ND: # <<<<<<<<<<<<<< + * info.shape = self.view.shape + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_ND) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":527 + * + * if flags & PyBUF_ND: + * info.shape = self.view.shape # <<<<<<<<<<<<<< + * else: + * info.shape = NULL + */ + __pyx_t_3 = __pyx_v_self->view.shape; + __pyx_v_info->shape = __pyx_t_3; + + /* "View.MemoryView":526 + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + * if flags & PyBUF_ND: # <<<<<<<<<<<<<< + * info.shape = self.view.shape + * else: + */ + goto __pyx_L6; + } + + /* "View.MemoryView":529 + * info.shape = self.view.shape + * else: + * info.shape = NULL # <<<<<<<<<<<<<< + * + * if flags & PyBUF_STRIDES: + */ + /*else*/ { + __pyx_v_info->shape = NULL; + } + __pyx_L6:; + + /* "View.MemoryView":531 + * info.shape = NULL + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.strides = self.view.strides + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_STRIDES) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":532 + * + * if flags & PyBUF_STRIDES: + * info.strides = self.view.strides # <<<<<<<<<<<<<< + * else: + * info.strides = NULL + */ + __pyx_t_3 = __pyx_v_self->view.strides; + __pyx_v_info->strides = __pyx_t_3; + + /* "View.MemoryView":531 + * info.shape = NULL + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.strides = self.view.strides + * else: + */ + goto __pyx_L7; + } + + /* "View.MemoryView":534 + * info.strides = self.view.strides + * else: + * info.strides = NULL # <<<<<<<<<<<<<< + * + * if flags & PyBUF_INDIRECT: + */ + /*else*/ { + __pyx_v_info->strides = NULL; + } + __pyx_L7:; + + /* "View.MemoryView":536 + * info.strides = NULL + * + * if flags & PyBUF_INDIRECT: # <<<<<<<<<<<<<< + * info.suboffsets = self.view.suboffsets + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_INDIRECT) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":537 + * + * if flags & PyBUF_INDIRECT: + * info.suboffsets = self.view.suboffsets # <<<<<<<<<<<<<< + * else: + * info.suboffsets = NULL + */ + __pyx_t_3 = __pyx_v_self->view.suboffsets; + __pyx_v_info->suboffsets = __pyx_t_3; + + /* "View.MemoryView":536 + * info.strides = NULL + * + * if flags & PyBUF_INDIRECT: # <<<<<<<<<<<<<< + * info.suboffsets = self.view.suboffsets + * else: + */ + goto __pyx_L8; + } + + /* "View.MemoryView":539 + * info.suboffsets = self.view.suboffsets + * else: + * info.suboffsets = NULL # <<<<<<<<<<<<<< + * + * if flags & PyBUF_FORMAT: + */ + /*else*/ { + __pyx_v_info->suboffsets = NULL; + } + __pyx_L8:; + + /* "View.MemoryView":541 + * info.suboffsets = NULL + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * info.format = self.view.format + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":542 + * + * if flags & PyBUF_FORMAT: + * info.format = self.view.format # <<<<<<<<<<<<<< + * else: + * info.format = NULL + */ + __pyx_t_4 = __pyx_v_self->view.format; + __pyx_v_info->format = __pyx_t_4; + + /* "View.MemoryView":541 + * info.suboffsets = NULL + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * info.format = self.view.format + * else: + */ + goto __pyx_L9; + } + + /* "View.MemoryView":544 + * info.format = self.view.format + * else: + * info.format = NULL # <<<<<<<<<<<<<< + * + * info.buf = self.view.buf + */ + /*else*/ { + __pyx_v_info->format = NULL; + } + __pyx_L9:; + + /* "View.MemoryView":546 + * info.format = NULL + * + * info.buf = self.view.buf # <<<<<<<<<<<<<< + * info.ndim = self.view.ndim + * info.itemsize = self.view.itemsize + */ + __pyx_t_5 = __pyx_v_self->view.buf; + __pyx_v_info->buf = __pyx_t_5; + + /* "View.MemoryView":547 + * + * info.buf = self.view.buf + * info.ndim = self.view.ndim # <<<<<<<<<<<<<< + * info.itemsize = self.view.itemsize + * info.len = self.view.len + */ + __pyx_t_6 = __pyx_v_self->view.ndim; + __pyx_v_info->ndim = __pyx_t_6; + + /* "View.MemoryView":548 + * info.buf = self.view.buf + * info.ndim = self.view.ndim + * info.itemsize = self.view.itemsize # <<<<<<<<<<<<<< + * info.len = self.view.len + * info.readonly = self.view.readonly + */ + __pyx_t_7 = __pyx_v_self->view.itemsize; + __pyx_v_info->itemsize = __pyx_t_7; + + /* "View.MemoryView":549 + * info.ndim = self.view.ndim + * info.itemsize = self.view.itemsize + * info.len = self.view.len # <<<<<<<<<<<<<< + * info.readonly = self.view.readonly + * info.obj = self + */ + __pyx_t_7 = __pyx_v_self->view.len; + __pyx_v_info->len = __pyx_t_7; + + /* "View.MemoryView":550 + * info.itemsize = self.view.itemsize + * info.len = self.view.len + * info.readonly = self.view.readonly # <<<<<<<<<<<<<< + * info.obj = self + * + */ + __pyx_t_1 = __pyx_v_self->view.readonly; + __pyx_v_info->readonly = __pyx_t_1; + + /* "View.MemoryView":551 + * info.len = self.view.len + * info.readonly = self.view.readonly + * info.obj = self # <<<<<<<<<<<<<< + * + * + */ + __Pyx_INCREF((PyObject *)__pyx_v_self); + __Pyx_GIVEREF((PyObject *)__pyx_v_self); + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); + __pyx_v_info->obj = ((PyObject *)__pyx_v_self); + + /* "View.MemoryView":521 + * itemp[i] = c + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__getbuffer__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + if (__pyx_v_info->obj != NULL) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + goto __pyx_L2; + __pyx_L0:; + if (__pyx_v_info->obj == Py_None) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + __pyx_L2:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":554 + * + * + * @property # <<<<<<<<<<<<<< + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_1T_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_1T_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + struct __pyx_memoryviewslice_obj *__pyx_v_result = 0; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":556 + * @property + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) # <<<<<<<<<<<<<< + * transpose_memslice(&result.from_slice) + * return result + */ + __pyx_t_1 = __pyx_memoryview_copy_object(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 556, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + if (!(likely(((__pyx_t_1) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_1, __pyx_memoryviewslice_type))))) __PYX_ERR(1, 556, __pyx_L1_error) + __pyx_v_result = ((struct __pyx_memoryviewslice_obj *)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":557 + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) + * transpose_memslice(&result.from_slice) # <<<<<<<<<<<<<< + * return result + * + */ + __pyx_t_2 = __pyx_memslice_transpose((&__pyx_v_result->from_slice)); if (unlikely(__pyx_t_2 == ((int)-1))) __PYX_ERR(1, 557, __pyx_L1_error) + + /* "View.MemoryView":558 + * cdef _memoryviewslice result = memoryview_copy(self) + * transpose_memslice(&result.from_slice) + * return result # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = ((PyObject *)__pyx_v_result); + goto __pyx_L0; + + /* "View.MemoryView":554 + * + * + * @property # <<<<<<<<<<<<<< + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.T.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":560 + * return result + * + * @property # <<<<<<<<<<<<<< + * def base(self): + * return self._get_base() + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4base_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4base_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":562 + * @property + * def base(self): + * return self._get_base() # <<<<<<<<<<<<<< + * + * cdef _get_base(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->_get_base(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 562, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":560 + * return result + * + * @property # <<<<<<<<<<<<<< + * def base(self): + * return self._get_base() + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.base.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":564 + * return self._get_base() + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.obj + * + */ + +static PyObject *__pyx_memoryview__get_base(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("_get_base", 1); + + /* "View.MemoryView":565 + * + * cdef _get_base(self): + * return self.obj # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->obj); + __pyx_r = __pyx_v_self->obj; + goto __pyx_L0; + + /* "View.MemoryView":564 + * return self._get_base() + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.obj + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":567 + * return self.obj + * + * @property # <<<<<<<<<<<<<< + * def shape(self): + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_5shape_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_5shape_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_7genexpr__pyx_v_length; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + PyObject *__pyx_t_5 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":569 + * @property + * def shape(self): + * return tuple([length for length in self.view.shape[:self.view.ndim]]) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + { /* enter inner scope */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = (__pyx_v_self->view.shape + __pyx_v_self->view.ndim); + for (__pyx_t_4 = __pyx_v_self->view.shape; __pyx_t_4 < __pyx_t_3; __pyx_t_4++) { + __pyx_t_2 = __pyx_t_4; + __pyx_7genexpr__pyx_v_length = (__pyx_t_2[0]); + __pyx_t_5 = PyInt_FromSsize_t(__pyx_7genexpr__pyx_v_length); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + if (unlikely(__Pyx_ListComp_Append(__pyx_t_1, (PyObject*)__pyx_t_5))) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + } /* exit inner scope */ + __pyx_t_5 = PyList_AsTuple(((PyObject*)__pyx_t_1)); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_r = __pyx_t_5; + __pyx_t_5 = 0; + goto __pyx_L0; + + /* "View.MemoryView":567 + * return self.obj + * + * @property # <<<<<<<<<<<<<< + * def shape(self): + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.memoryview.shape.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":571 + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def strides(self): + * if self.view.strides == NULL: + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_7strides_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_7strides_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_8genexpr1__pyx_v_stride; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + Py_ssize_t *__pyx_t_5; + PyObject *__pyx_t_6 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":573 + * @property + * def strides(self): + * if self.view.strides == NULL: # <<<<<<<<<<<<<< + * + * raise ValueError, "Buffer view does not expose strides" + */ + __pyx_t_1 = (__pyx_v_self->view.strides == NULL); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":575 + * if self.view.strides == NULL: + * + * raise ValueError, "Buffer view does not expose strides" # <<<<<<<<<<<<<< + * + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Buffer_view_does_not_expose_stri, 0, 0); + __PYX_ERR(1, 575, __pyx_L1_error) + + /* "View.MemoryView":573 + * @property + * def strides(self): + * if self.view.strides == NULL: # <<<<<<<<<<<<<< + * + * raise ValueError, "Buffer view does not expose strides" + */ + } + + /* "View.MemoryView":577 + * raise ValueError, "Buffer view does not expose strides" + * + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + { /* enter inner scope */ + __pyx_t_2 = PyList_New(0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = (__pyx_v_self->view.strides + __pyx_v_self->view.ndim); + for (__pyx_t_5 = __pyx_v_self->view.strides; __pyx_t_5 < __pyx_t_4; __pyx_t_5++) { + __pyx_t_3 = __pyx_t_5; + __pyx_8genexpr1__pyx_v_stride = (__pyx_t_3[0]); + __pyx_t_6 = PyInt_FromSsize_t(__pyx_8genexpr1__pyx_v_stride); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + if (unlikely(__Pyx_ListComp_Append(__pyx_t_2, (PyObject*)__pyx_t_6))) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + } + } /* exit inner scope */ + __pyx_t_6 = PyList_AsTuple(((PyObject*)__pyx_t_2)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_6; + __pyx_t_6 = 0; + goto __pyx_L0; + + /* "View.MemoryView":571 + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def strides(self): + * if self.view.strides == NULL: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_AddTraceback("View.MemoryView.memoryview.strides.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":579 + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def suboffsets(self): + * if self.view.suboffsets == NULL: + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_10suboffsets_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_10suboffsets_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_8genexpr2__pyx_v_suboffset; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + Py_ssize_t *__pyx_t_5; + PyObject *__pyx_t_6 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":581 + * @property + * def suboffsets(self): + * if self.view.suboffsets == NULL: # <<<<<<<<<<<<<< + * return (-1,) * self.view.ndim + * + */ + __pyx_t_1 = (__pyx_v_self->view.suboffsets == NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":582 + * def suboffsets(self): + * if self.view.suboffsets == NULL: + * return (-1,) * self.view.ndim # <<<<<<<<<<<<<< + * + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __Pyx_PySequence_Multiply(__pyx_tuple__4, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 582, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":581 + * @property + * def suboffsets(self): + * if self.view.suboffsets == NULL: # <<<<<<<<<<<<<< + * return (-1,) * self.view.ndim + * + */ + } + + /* "View.MemoryView":584 + * return (-1,) * self.view.ndim + * + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + { /* enter inner scope */ + __pyx_t_2 = PyList_New(0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = (__pyx_v_self->view.suboffsets + __pyx_v_self->view.ndim); + for (__pyx_t_5 = __pyx_v_self->view.suboffsets; __pyx_t_5 < __pyx_t_4; __pyx_t_5++) { + __pyx_t_3 = __pyx_t_5; + __pyx_8genexpr2__pyx_v_suboffset = (__pyx_t_3[0]); + __pyx_t_6 = PyInt_FromSsize_t(__pyx_8genexpr2__pyx_v_suboffset); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + if (unlikely(__Pyx_ListComp_Append(__pyx_t_2, (PyObject*)__pyx_t_6))) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + } + } /* exit inner scope */ + __pyx_t_6 = PyList_AsTuple(((PyObject*)__pyx_t_2)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_6; + __pyx_t_6 = 0; + goto __pyx_L0; + + /* "View.MemoryView":579 + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def suboffsets(self): + * if self.view.suboffsets == NULL: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_AddTraceback("View.MemoryView.memoryview.suboffsets.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":586 + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def ndim(self): + * return self.view.ndim + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4ndim_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4ndim_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":588 + * @property + * def ndim(self): + * return self.view.ndim # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_self->view.ndim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 588, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":586 + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def ndim(self): + * return self.view.ndim + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.ndim.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":590 + * return self.view.ndim + * + * @property # <<<<<<<<<<<<<< + * def itemsize(self): + * return self.view.itemsize + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_8itemsize_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_8itemsize_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":592 + * @property + * def itemsize(self): + * return self.view.itemsize # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 592, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":590 + * return self.view.ndim + * + * @property # <<<<<<<<<<<<<< + * def itemsize(self): + * return self.view.itemsize + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.itemsize.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":594 + * return self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def nbytes(self): + * return self.size * self.view.itemsize + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_6nbytes_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_6nbytes_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":596 + * @property + * def nbytes(self): + * return self.size * self.view.itemsize # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_size); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 596, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 596, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyNumber_Multiply(__pyx_t_1, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 596, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_3; + __pyx_t_3 = 0; + goto __pyx_L0; + + /* "View.MemoryView":594 + * return self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def nbytes(self): + * return self.size * self.view.itemsize + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.nbytes.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":598 + * return self.size * self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def size(self): + * if self._size is None: + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4size_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4size_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_v_result = NULL; + PyObject *__pyx_v_length = NULL; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + PyObject *__pyx_t_5 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":600 + * @property + * def size(self): + * if self._size is None: # <<<<<<<<<<<<<< + * result = 1 + * + */ + __pyx_t_1 = (__pyx_v_self->_size == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":601 + * def size(self): + * if self._size is None: + * result = 1 # <<<<<<<<<<<<<< + * + * for length in self.view.shape[:self.view.ndim]: + */ + __Pyx_INCREF(__pyx_int_1); + __pyx_v_result = __pyx_int_1; + + /* "View.MemoryView":603 + * result = 1 + * + * for length in self.view.shape[:self.view.ndim]: # <<<<<<<<<<<<<< + * result *= length + * + */ + __pyx_t_3 = (__pyx_v_self->view.shape + __pyx_v_self->view.ndim); + for (__pyx_t_4 = __pyx_v_self->view.shape; __pyx_t_4 < __pyx_t_3; __pyx_t_4++) { + __pyx_t_2 = __pyx_t_4; + __pyx_t_5 = PyInt_FromSsize_t((__pyx_t_2[0])); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 603, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_XDECREF_SET(__pyx_v_length, __pyx_t_5); + __pyx_t_5 = 0; + + /* "View.MemoryView":604 + * + * for length in self.view.shape[:self.view.ndim]: + * result *= length # <<<<<<<<<<<<<< + * + * self._size = result + */ + __pyx_t_5 = PyNumber_InPlaceMultiply(__pyx_v_result, __pyx_v_length); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 604, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF_SET(__pyx_v_result, __pyx_t_5); + __pyx_t_5 = 0; + } + + /* "View.MemoryView":606 + * result *= length + * + * self._size = result # <<<<<<<<<<<<<< + * + * return self._size + */ + __Pyx_INCREF(__pyx_v_result); + __Pyx_GIVEREF(__pyx_v_result); + __Pyx_GOTREF(__pyx_v_self->_size); + __Pyx_DECREF(__pyx_v_self->_size); + __pyx_v_self->_size = __pyx_v_result; + + /* "View.MemoryView":600 + * @property + * def size(self): + * if self._size is None: # <<<<<<<<<<<<<< + * result = 1 + * + */ + } + + /* "View.MemoryView":608 + * self._size = result + * + * return self._size # <<<<<<<<<<<<<< + * + * def __len__(self): + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->_size); + __pyx_r = __pyx_v_self->_size; + goto __pyx_L0; + + /* "View.MemoryView":598 + * return self.size * self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def size(self): + * if self._size is None: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.memoryview.size.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_result); + __Pyx_XDECREF(__pyx_v_length); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":610 + * return self._size + * + * def __len__(self): # <<<<<<<<<<<<<< + * if self.view.ndim >= 1: + * return self.view.shape[0] + */ + +/* Python wrapper */ +static Py_ssize_t __pyx_memoryview___len__(PyObject *__pyx_v_self); /*proto*/ +static Py_ssize_t __pyx_memoryview___len__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + Py_ssize_t __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__len__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_10__len__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_10__len__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_r; + int __pyx_t_1; + + /* "View.MemoryView":611 + * + * def __len__(self): + * if self.view.ndim >= 1: # <<<<<<<<<<<<<< + * return self.view.shape[0] + * + */ + __pyx_t_1 = (__pyx_v_self->view.ndim >= 1); + if (__pyx_t_1) { + + /* "View.MemoryView":612 + * def __len__(self): + * if self.view.ndim >= 1: + * return self.view.shape[0] # <<<<<<<<<<<<<< + * + * return 0 + */ + __pyx_r = (__pyx_v_self->view.shape[0]); + goto __pyx_L0; + + /* "View.MemoryView":611 + * + * def __len__(self): + * if self.view.ndim >= 1: # <<<<<<<<<<<<<< + * return self.view.shape[0] + * + */ + } + + /* "View.MemoryView":614 + * return self.view.shape[0] + * + * return 0 # <<<<<<<<<<<<<< + * + * def __repr__(self): + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":610 + * return self._size + * + * def __len__(self): # <<<<<<<<<<<<<< + * if self.view.ndim >= 1: + * return self.view.shape[0] + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":616 + * return 0 + * + * def __repr__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__, + * id(self)) + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview___repr__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_memoryview___repr__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__repr__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12__repr__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12__repr__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__repr__", 1); + + /* "View.MemoryView":617 + * + * def __repr__(self): + * return "" % (self.base.__class__.__name__, # <<<<<<<<<<<<<< + * id(self)) + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":618 + * def __repr__(self): + * return "" % (self.base.__class__.__name__, + * id(self)) # <<<<<<<<<<<<<< + * + * def __str__(self): + */ + __pyx_t_2 = __Pyx_PyObject_CallOneArg(__pyx_builtin_id, ((PyObject *)__pyx_v_self)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 618, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + + /* "View.MemoryView":617 + * + * def __repr__(self): + * return "" % (self.base.__class__.__name__, # <<<<<<<<<<<<<< + * id(self)) + * + */ + __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1)) __PYX_ERR(1, 617, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_2)) __PYX_ERR(1, 617, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_at_0x_x, __pyx_t_3); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":616 + * return 0 + * + * def __repr__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__, + * id(self)) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.__repr__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":620 + * id(self)) + * + * def __str__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__,) + * + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview___str__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_memoryview___str__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__str__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14__str__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14__str__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__str__", 1); + + /* "View.MemoryView":621 + * + * def __str__(self): + * return "" % (self.base.__class__.__name__,) # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_2 = PyTuple_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 0, __pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_object, __pyx_t_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":620 + * id(self)) + * + * def __str__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__,) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.__str__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":624 + * + * + * def is_c_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_is_c_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_is_c_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("is_c_contig (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("is_c_contig", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "is_c_contig", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16is_c_contig(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16is_c_contig(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice *__pyx_v_mslice; + __Pyx_memviewslice __pyx_v_tmp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("is_c_contig", 1); + + /* "View.MemoryView":627 + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) # <<<<<<<<<<<<<< + * return slice_is_contig(mslice[0], 'C', self.view.ndim) + * + */ + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 627, __pyx_L1_error) + __pyx_v_mslice = __pyx_t_1; + + /* "View.MemoryView":628 + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) + * return slice_is_contig(mslice[0], 'C', self.view.ndim) # <<<<<<<<<<<<<< + * + * def is_f_contig(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'C', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 628, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":624 + * + * + * def is_c_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.is_c_contig", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":630 + * return slice_is_contig(mslice[0], 'C', self.view.ndim) + * + * def is_f_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_is_f_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_is_f_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("is_f_contig (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("is_f_contig", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "is_f_contig", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18is_f_contig(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18is_f_contig(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice *__pyx_v_mslice; + __Pyx_memviewslice __pyx_v_tmp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("is_f_contig", 1); + + /* "View.MemoryView":633 + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) # <<<<<<<<<<<<<< + * return slice_is_contig(mslice[0], 'F', self.view.ndim) + * + */ + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 633, __pyx_L1_error) + __pyx_v_mslice = __pyx_t_1; + + /* "View.MemoryView":634 + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) + * return slice_is_contig(mslice[0], 'F', self.view.ndim) # <<<<<<<<<<<<<< + * + * def copy(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'F', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 634, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":630 + * return slice_is_contig(mslice[0], 'C', self.view.ndim) + * + * def is_f_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.is_f_contig", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":636 + * return slice_is_contig(mslice[0], 'F', self.view.ndim) + * + * def copy(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice mslice + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_copy(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_copy(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("copy (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("copy", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "copy", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20copy(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20copy(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice __pyx_v_mslice; + int __pyx_v_flags; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("copy", 1); + + /* "View.MemoryView":638 + * def copy(self): + * cdef __Pyx_memviewslice mslice + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS # <<<<<<<<<<<<<< + * + * slice_copy(self, &mslice) + */ + __pyx_v_flags = (__pyx_v_self->flags & (~PyBUF_F_CONTIGUOUS)); + + /* "View.MemoryView":640 + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS + * + * slice_copy(self, &mslice) # <<<<<<<<<<<<<< + * mslice = slice_copy_contig(&mslice, "c", self.view.ndim, + * self.view.itemsize, + */ + __pyx_memoryview_slice_copy(__pyx_v_self, (&__pyx_v_mslice)); + + /* "View.MemoryView":641 + * + * slice_copy(self, &mslice) + * mslice = slice_copy_contig(&mslice, "c", self.view.ndim, # <<<<<<<<<<<<<< + * self.view.itemsize, + * flags|PyBUF_C_CONTIGUOUS, + */ + __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_mslice), ((char *)"c"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_C_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 641, __pyx_L1_error) + __pyx_v_mslice = __pyx_t_1; + + /* "View.MemoryView":646 + * self.dtype_is_object) + * + * return memoryview_copy_from_slice(self, &mslice) # <<<<<<<<<<<<<< + * + * def copy_fortran(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_mslice)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 646, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":636 + * return slice_is_contig(mslice[0], 'F', self.view.ndim) + * + * def copy(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice mslice + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.copy", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":648 + * return memoryview_copy_from_slice(self, &mslice) + * + * def copy_fortran(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice src, dst + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_copy_fortran(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_copy_fortran(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("copy_fortran (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("copy_fortran", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "copy_fortran", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22copy_fortran(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22copy_fortran(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice __pyx_v_src; + __Pyx_memviewslice __pyx_v_dst; + int __pyx_v_flags; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("copy_fortran", 1); + + /* "View.MemoryView":650 + * def copy_fortran(self): + * cdef __Pyx_memviewslice src, dst + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS # <<<<<<<<<<<<<< + * + * slice_copy(self, &src) + */ + __pyx_v_flags = (__pyx_v_self->flags & (~PyBUF_C_CONTIGUOUS)); + + /* "View.MemoryView":652 + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS + * + * slice_copy(self, &src) # <<<<<<<<<<<<<< + * dst = slice_copy_contig(&src, "fortran", self.view.ndim, + * self.view.itemsize, + */ + __pyx_memoryview_slice_copy(__pyx_v_self, (&__pyx_v_src)); + + /* "View.MemoryView":653 + * + * slice_copy(self, &src) + * dst = slice_copy_contig(&src, "fortran", self.view.ndim, # <<<<<<<<<<<<<< + * self.view.itemsize, + * flags|PyBUF_F_CONTIGUOUS, + */ + __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_src), ((char *)"fortran"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_F_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 653, __pyx_L1_error) + __pyx_v_dst = __pyx_t_1; + + /* "View.MemoryView":658 + * self.dtype_is_object) + * + * return memoryview_copy_from_slice(self, &dst) # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_dst)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 658, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":648 + * return memoryview_copy_from_slice(self, &mslice) + * + * def copy_fortran(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice src, dst + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.copy_fortran", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryview_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryview_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_memoryview___reduce_cython__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryview___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":2 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 2, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryview_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryview_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + CYTHON_UNUSED PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 3, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 3, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 3, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.memoryview.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_memoryview_2__setstate_cython__(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryview_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":4 + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 4, __pyx_L1_error) + + /* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":662 + * + * @cname('__pyx_memoryview_new') + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): # <<<<<<<<<<<<<< + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo + */ + +static PyObject *__pyx_memoryview_new(PyObject *__pyx_v_o, int __pyx_v_flags, int __pyx_v_dtype_is_object, __Pyx_TypeInfo *__pyx_v_typeinfo) { + struct __pyx_memoryview_obj *__pyx_v_result = 0; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_cwrapper", 1); + + /* "View.MemoryView":663 + * @cname('__pyx_memoryview_new') + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): + * cdef memoryview result = memoryview(o, flags, dtype_is_object) # <<<<<<<<<<<<<< + * result.typeinfo = typeinfo + * return result + */ + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF(__pyx_v_o); + __Pyx_GIVEREF(__pyx_v_o); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_o)) __PYX_ERR(1, 663, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_1)) __PYX_ERR(1, 663, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2)) __PYX_ERR(1, 663, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_v_result = ((struct __pyx_memoryview_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":664 + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo # <<<<<<<<<<<<<< + * return result + * + */ + __pyx_v_result->typeinfo = __pyx_v_typeinfo; + + /* "View.MemoryView":665 + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo + * return result # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_check') + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = ((PyObject *)__pyx_v_result); + goto __pyx_L0; + + /* "View.MemoryView":662 + * + * @cname('__pyx_memoryview_new') + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): # <<<<<<<<<<<<<< + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview_cwrapper", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":668 + * + * @cname('__pyx_memoryview_check') + * cdef inline bint memoryview_check(object o) noexcept: # <<<<<<<<<<<<<< + * return isinstance(o, memoryview) + * + */ + +static CYTHON_INLINE int __pyx_memoryview_check(PyObject *__pyx_v_o) { + int __pyx_r; + int __pyx_t_1; + + /* "View.MemoryView":669 + * @cname('__pyx_memoryview_check') + * cdef inline bint memoryview_check(object o) noexcept: + * return isinstance(o, memoryview) # <<<<<<<<<<<<<< + * + * cdef tuple _unellipsify(object index, int ndim): + */ + __pyx_t_1 = __Pyx_TypeCheck(__pyx_v_o, __pyx_memoryview_type); + __pyx_r = __pyx_t_1; + goto __pyx_L0; + + /* "View.MemoryView":668 + * + * @cname('__pyx_memoryview_check') + * cdef inline bint memoryview_check(object o) noexcept: # <<<<<<<<<<<<<< + * return isinstance(o, memoryview) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":671 + * return isinstance(o, memoryview) + * + * cdef tuple _unellipsify(object index, int ndim): # <<<<<<<<<<<<<< + * """ + * Replace all ellipses with full slices and fill incomplete indices with + */ + +static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) { + Py_ssize_t __pyx_v_idx; + PyObject *__pyx_v_tup = NULL; + PyObject *__pyx_v_result = NULL; + int __pyx_v_have_slices; + int __pyx_v_seen_ellipsis; + PyObject *__pyx_v_item = NULL; + Py_ssize_t __pyx_v_nslices; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + Py_ssize_t __pyx_t_4; + Py_ssize_t __pyx_t_5; + Py_UCS4 __pyx_t_6; + PyObject *__pyx_t_7 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_unellipsify", 1); + + /* "View.MemoryView":677 + * """ + * cdef Py_ssize_t idx + * tup = index if isinstance(index, tuple) else (index,) # <<<<<<<<<<<<<< + * + * result = [slice(None)] * ndim + */ + __pyx_t_2 = PyTuple_Check(__pyx_v_index); + if (__pyx_t_2) { + __Pyx_INCREF(((PyObject*)__pyx_v_index)); + __pyx_t_1 = __pyx_v_index; + } else { + __pyx_t_3 = PyTuple_New(1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 677, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF(__pyx_v_index); + __Pyx_GIVEREF(__pyx_v_index); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_index)) __PYX_ERR(1, 677, __pyx_L1_error); + __pyx_t_1 = __pyx_t_3; + __pyx_t_3 = 0; + } + __pyx_v_tup = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":679 + * tup = index if isinstance(index, tuple) else (index,) + * + * result = [slice(None)] * ndim # <<<<<<<<<<<<<< + * have_slices = False + * seen_ellipsis = False + */ + __pyx_t_1 = PyList_New(1 * ((__pyx_v_ndim<0) ? 0:__pyx_v_ndim)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 679, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + { Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < __pyx_v_ndim; __pyx_temp++) { + __Pyx_INCREF(__pyx_slice__5); + __Pyx_GIVEREF(__pyx_slice__5); + if (__Pyx_PyList_SET_ITEM(__pyx_t_1, __pyx_temp, __pyx_slice__5)) __PYX_ERR(1, 679, __pyx_L1_error); + } + } + __pyx_v_result = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":680 + * + * result = [slice(None)] * ndim + * have_slices = False # <<<<<<<<<<<<<< + * seen_ellipsis = False + * idx = 0 + */ + __pyx_v_have_slices = 0; + + /* "View.MemoryView":681 + * result = [slice(None)] * ndim + * have_slices = False + * seen_ellipsis = False # <<<<<<<<<<<<<< + * idx = 0 + * for item in tup: + */ + __pyx_v_seen_ellipsis = 0; + + /* "View.MemoryView":682 + * have_slices = False + * seen_ellipsis = False + * idx = 0 # <<<<<<<<<<<<<< + * for item in tup: + * if item is Ellipsis: + */ + __pyx_v_idx = 0; + + /* "View.MemoryView":683 + * seen_ellipsis = False + * idx = 0 + * for item in tup: # <<<<<<<<<<<<<< + * if item is Ellipsis: + * if not seen_ellipsis: + */ + if (unlikely(__pyx_v_tup == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable"); + __PYX_ERR(1, 683, __pyx_L1_error) + } + __pyx_t_1 = __pyx_v_tup; __Pyx_INCREF(__pyx_t_1); + __pyx_t_4 = 0; + for (;;) { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_1); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 683, __pyx_L1_error) + #endif + if (__pyx_t_4 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_3 = PyTuple_GET_ITEM(__pyx_t_1, __pyx_t_4); __Pyx_INCREF(__pyx_t_3); __pyx_t_4++; if (unlikely((0 < 0))) __PYX_ERR(1, 683, __pyx_L1_error) + #else + __pyx_t_3 = __Pyx_PySequence_ITEM(__pyx_t_1, __pyx_t_4); __pyx_t_4++; if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 683, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + #endif + __Pyx_XDECREF_SET(__pyx_v_item, __pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":684 + * idx = 0 + * for item in tup: + * if item is Ellipsis: # <<<<<<<<<<<<<< + * if not seen_ellipsis: + * idx += ndim - len(tup) + */ + __pyx_t_2 = (__pyx_v_item == __pyx_builtin_Ellipsis); + if (__pyx_t_2) { + + /* "View.MemoryView":685 + * for item in tup: + * if item is Ellipsis: + * if not seen_ellipsis: # <<<<<<<<<<<<<< + * idx += ndim - len(tup) + * seen_ellipsis = True + */ + __pyx_t_2 = (!__pyx_v_seen_ellipsis); + if (__pyx_t_2) { + + /* "View.MemoryView":686 + * if item is Ellipsis: + * if not seen_ellipsis: + * idx += ndim - len(tup) # <<<<<<<<<<<<<< + * seen_ellipsis = True + * have_slices = True + */ + if (unlikely(__pyx_v_tup == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(1, 686, __pyx_L1_error) + } + __pyx_t_5 = __Pyx_PyTuple_GET_SIZE(__pyx_v_tup); if (unlikely(__pyx_t_5 == ((Py_ssize_t)-1))) __PYX_ERR(1, 686, __pyx_L1_error) + __pyx_v_idx = (__pyx_v_idx + (__pyx_v_ndim - __pyx_t_5)); + + /* "View.MemoryView":687 + * if not seen_ellipsis: + * idx += ndim - len(tup) + * seen_ellipsis = True # <<<<<<<<<<<<<< + * have_slices = True + * else: + */ + __pyx_v_seen_ellipsis = 1; + + /* "View.MemoryView":685 + * for item in tup: + * if item is Ellipsis: + * if not seen_ellipsis: # <<<<<<<<<<<<<< + * idx += ndim - len(tup) + * seen_ellipsis = True + */ + } + + /* "View.MemoryView":688 + * idx += ndim - len(tup) + * seen_ellipsis = True + * have_slices = True # <<<<<<<<<<<<<< + * else: + * if isinstance(item, slice): + */ + __pyx_v_have_slices = 1; + + /* "View.MemoryView":684 + * idx = 0 + * for item in tup: + * if item is Ellipsis: # <<<<<<<<<<<<<< + * if not seen_ellipsis: + * idx += ndim - len(tup) + */ + goto __pyx_L5; + } + + /* "View.MemoryView":690 + * have_slices = True + * else: + * if isinstance(item, slice): # <<<<<<<<<<<<<< + * have_slices = True + * elif not PyIndex_Check(item): + */ + /*else*/ { + __pyx_t_2 = PySlice_Check(__pyx_v_item); + if (__pyx_t_2) { + + /* "View.MemoryView":691 + * else: + * if isinstance(item, slice): + * have_slices = True # <<<<<<<<<<<<<< + * elif not PyIndex_Check(item): + * raise TypeError, f"Cannot index with type '{type(item)}'" + */ + __pyx_v_have_slices = 1; + + /* "View.MemoryView":690 + * have_slices = True + * else: + * if isinstance(item, slice): # <<<<<<<<<<<<<< + * have_slices = True + * elif not PyIndex_Check(item): + */ + goto __pyx_L7; + } + + /* "View.MemoryView":692 + * if isinstance(item, slice): + * have_slices = True + * elif not PyIndex_Check(item): # <<<<<<<<<<<<<< + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item + */ + __pyx_t_2 = (!(PyIndex_Check(__pyx_v_item) != 0)); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":693 + * have_slices = True + * elif not PyIndex_Check(item): + * raise TypeError, f"Cannot index with type '{type(item)}'" # <<<<<<<<<<<<<< + * result[idx] = item + * idx += 1 + */ + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 693, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_5 = 0; + __pyx_t_6 = 127; + __Pyx_INCREF(__pyx_kp_u_Cannot_index_with_type); + __pyx_t_5 += 24; + __Pyx_GIVEREF(__pyx_kp_u_Cannot_index_with_type); + PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_kp_u_Cannot_index_with_type); + __pyx_t_7 = __Pyx_PyObject_FormatSimple(((PyObject *)Py_TYPE(__pyx_v_item)), __pyx_empty_unicode); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 693, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_6 = (__Pyx_PyUnicode_MAX_CHAR_VALUE(__pyx_t_7) > __pyx_t_6) ? __Pyx_PyUnicode_MAX_CHAR_VALUE(__pyx_t_7) : __pyx_t_6; + __pyx_t_5 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_7); + __pyx_t_7 = 0; + __Pyx_INCREF(__pyx_kp_u__6); + __pyx_t_5 += 1; + __Pyx_GIVEREF(__pyx_kp_u__6); + PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_kp_u__6); + __pyx_t_7 = __Pyx_PyUnicode_Join(__pyx_t_3, 3, __pyx_t_5, __pyx_t_6); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 693, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_t_7, 0, 0); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __PYX_ERR(1, 693, __pyx_L1_error) + + /* "View.MemoryView":692 + * if isinstance(item, slice): + * have_slices = True + * elif not PyIndex_Check(item): # <<<<<<<<<<<<<< + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item + */ + } + __pyx_L7:; + + /* "View.MemoryView":694 + * elif not PyIndex_Check(item): + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item # <<<<<<<<<<<<<< + * idx += 1 + * + */ + if (unlikely((__Pyx_SetItemInt(__pyx_v_result, __pyx_v_idx, __pyx_v_item, Py_ssize_t, 1, PyInt_FromSsize_t, 1, 1, 1) < 0))) __PYX_ERR(1, 694, __pyx_L1_error) + } + __pyx_L5:; + + /* "View.MemoryView":695 + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item + * idx += 1 # <<<<<<<<<<<<<< + * + * nslices = ndim - idx + */ + __pyx_v_idx = (__pyx_v_idx + 1); + + /* "View.MemoryView":683 + * seen_ellipsis = False + * idx = 0 + * for item in tup: # <<<<<<<<<<<<<< + * if item is Ellipsis: + * if not seen_ellipsis: + */ + } + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "View.MemoryView":697 + * idx += 1 + * + * nslices = ndim - idx # <<<<<<<<<<<<<< + * return have_slices or nslices, tuple(result) + * + */ + __pyx_v_nslices = (__pyx_v_ndim - __pyx_v_idx); + + /* "View.MemoryView":698 + * + * nslices = ndim - idx + * return have_slices or nslices, tuple(result) # <<<<<<<<<<<<<< + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + */ + __Pyx_XDECREF(__pyx_r); + if (!__pyx_v_have_slices) { + } else { + __pyx_t_7 = __Pyx_PyBool_FromLong(__pyx_v_have_slices); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_1 = __pyx_t_7; + __pyx_t_7 = 0; + goto __pyx_L9_bool_binop_done; + } + __pyx_t_7 = PyInt_FromSsize_t(__pyx_v_nslices); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_1 = __pyx_t_7; + __pyx_t_7 = 0; + __pyx_L9_bool_binop_done:; + __pyx_t_7 = PyList_AsTuple(__pyx_v_result); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1)) __PYX_ERR(1, 698, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_7); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_7 = 0; + __pyx_r = ((PyObject*)__pyx_t_3); + __pyx_t_3 = 0; + goto __pyx_L0; + + /* "View.MemoryView":671 + * return isinstance(o, memoryview) + * + * cdef tuple _unellipsify(object index, int ndim): # <<<<<<<<<<<<<< + * """ + * Replace all ellipses with full slices and fill incomplete indices with + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_AddTraceback("View.MemoryView._unellipsify", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_tup); + __Pyx_XDECREF(__pyx_v_result); + __Pyx_XDECREF(__pyx_v_item); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":700 + * return have_slices or nslices, tuple(result) + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: # <<<<<<<<<<<<<< + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: + */ + +static int assert_direct_dimensions(Py_ssize_t *__pyx_v_suboffsets, int __pyx_v_ndim) { + Py_ssize_t __pyx_v_suboffset; + int __pyx_r; + Py_ssize_t *__pyx_t_1; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + int __pyx_t_4; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + + /* "View.MemoryView":701 + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + * for suboffset in suboffsets[:ndim]: # <<<<<<<<<<<<<< + * if suboffset >= 0: + * raise ValueError, "Indirect dimensions not supported" + */ + __pyx_t_2 = (__pyx_v_suboffsets + __pyx_v_ndim); + for (__pyx_t_3 = __pyx_v_suboffsets; __pyx_t_3 < __pyx_t_2; __pyx_t_3++) { + __pyx_t_1 = __pyx_t_3; + __pyx_v_suboffset = (__pyx_t_1[0]); + + /* "View.MemoryView":702 + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * raise ValueError, "Indirect dimensions not supported" + * return 0 # return type just used as an error flag + */ + __pyx_t_4 = (__pyx_v_suboffset >= 0); + if (unlikely(__pyx_t_4)) { + + /* "View.MemoryView":703 + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: + * raise ValueError, "Indirect dimensions not supported" # <<<<<<<<<<<<<< + * return 0 # return type just used as an error flag + * + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Indirect_dimensions_not_supporte, 0, 0); + __PYX_ERR(1, 703, __pyx_L1_error) + + /* "View.MemoryView":702 + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * raise ValueError, "Indirect dimensions not supported" + * return 0 # return type just used as an error flag + */ + } + } + + /* "View.MemoryView":704 + * if suboffset >= 0: + * raise ValueError, "Indirect dimensions not supported" + * return 0 # return type just used as an error flag # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":700 + * return have_slices or nslices, tuple(result) + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: # <<<<<<<<<<<<<< + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.assert_direct_dimensions", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":711 + * + * @cname('__pyx_memview_slice') + * cdef memoryview memview_slice(memoryview memview, object indices): # <<<<<<<<<<<<<< + * cdef int new_ndim = 0, suboffset_dim = -1, dim + * cdef bint negative_step + */ + +static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_obj *__pyx_v_memview, PyObject *__pyx_v_indices) { + int __pyx_v_new_ndim; + int __pyx_v_suboffset_dim; + int __pyx_v_dim; + __Pyx_memviewslice __pyx_v_src; + __Pyx_memviewslice __pyx_v_dst; + __Pyx_memviewslice *__pyx_v_p_src; + struct __pyx_memoryviewslice_obj *__pyx_v_memviewsliceobj = 0; + __Pyx_memviewslice *__pyx_v_p_dst; + int *__pyx_v_p_suboffset_dim; + Py_ssize_t __pyx_v_start; + Py_ssize_t __pyx_v_stop; + Py_ssize_t __pyx_v_step; + Py_ssize_t __pyx_v_cindex; + int __pyx_v_have_start; + int __pyx_v_have_stop; + int __pyx_v_have_step; + PyObject *__pyx_v_index = NULL; + struct __pyx_memoryview_obj *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + struct __pyx_memoryview_obj *__pyx_t_3; + char *__pyx_t_4; + int __pyx_t_5; + Py_ssize_t __pyx_t_6; + PyObject *(*__pyx_t_7)(PyObject *); + PyObject *__pyx_t_8 = NULL; + Py_ssize_t __pyx_t_9; + int __pyx_t_10; + Py_ssize_t __pyx_t_11; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memview_slice", 1); + + /* "View.MemoryView":712 + * @cname('__pyx_memview_slice') + * cdef memoryview memview_slice(memoryview memview, object indices): + * cdef int new_ndim = 0, suboffset_dim = -1, dim # <<<<<<<<<<<<<< + * cdef bint negative_step + * cdef __Pyx_memviewslice src, dst + */ + __pyx_v_new_ndim = 0; + __pyx_v_suboffset_dim = -1; + + /* "View.MemoryView":719 + * + * + * memset(&dst, 0, sizeof(dst)) # <<<<<<<<<<<<<< + * + * cdef _memoryviewslice memviewsliceobj + */ + (void)(memset((&__pyx_v_dst), 0, (sizeof(__pyx_v_dst)))); + + /* "View.MemoryView":723 + * cdef _memoryviewslice memviewsliceobj + * + * assert memview.view.ndim > 0 # <<<<<<<<<<<<<< + * + * if isinstance(memview, _memoryviewslice): + */ + #ifndef CYTHON_WITHOUT_ASSERTIONS + if (unlikely(__pyx_assertions_enabled())) { + __pyx_t_1 = (__pyx_v_memview->view.ndim > 0); + if (unlikely(!__pyx_t_1)) { + __Pyx_Raise(__pyx_builtin_AssertionError, 0, 0, 0); + __PYX_ERR(1, 723, __pyx_L1_error) + } + } + #else + if ((1)); else __PYX_ERR(1, 723, __pyx_L1_error) + #endif + + /* "View.MemoryView":725 + * assert memview.view.ndim > 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * memviewsliceobj = memview + * p_src = &memviewsliceobj.from_slice + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":726 + * + * if isinstance(memview, _memoryviewslice): + * memviewsliceobj = memview # <<<<<<<<<<<<<< + * p_src = &memviewsliceobj.from_slice + * else: + */ + if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 726, __pyx_L1_error) + __pyx_t_2 = ((PyObject *)__pyx_v_memview); + __Pyx_INCREF(__pyx_t_2); + __pyx_v_memviewsliceobj = ((struct __pyx_memoryviewslice_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":727 + * if isinstance(memview, _memoryviewslice): + * memviewsliceobj = memview + * p_src = &memviewsliceobj.from_slice # <<<<<<<<<<<<<< + * else: + * slice_copy(memview, &src) + */ + __pyx_v_p_src = (&__pyx_v_memviewsliceobj->from_slice); + + /* "View.MemoryView":725 + * assert memview.view.ndim > 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * memviewsliceobj = memview + * p_src = &memviewsliceobj.from_slice + */ + goto __pyx_L3; + } + + /* "View.MemoryView":729 + * p_src = &memviewsliceobj.from_slice + * else: + * slice_copy(memview, &src) # <<<<<<<<<<<<<< + * p_src = &src + * + */ + /*else*/ { + __pyx_memoryview_slice_copy(__pyx_v_memview, (&__pyx_v_src)); + + /* "View.MemoryView":730 + * else: + * slice_copy(memview, &src) + * p_src = &src # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_p_src = (&__pyx_v_src); + } + __pyx_L3:; + + /* "View.MemoryView":736 + * + * + * dst.memview = p_src.memview # <<<<<<<<<<<<<< + * dst.data = p_src.data + * + */ + __pyx_t_3 = __pyx_v_p_src->memview; + __pyx_v_dst.memview = __pyx_t_3; + + /* "View.MemoryView":737 + * + * dst.memview = p_src.memview + * dst.data = p_src.data # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_4 = __pyx_v_p_src->data; + __pyx_v_dst.data = __pyx_t_4; + + /* "View.MemoryView":742 + * + * + * cdef __Pyx_memviewslice *p_dst = &dst # <<<<<<<<<<<<<< + * cdef int *p_suboffset_dim = &suboffset_dim + * cdef Py_ssize_t start, stop, step, cindex + */ + __pyx_v_p_dst = (&__pyx_v_dst); + + /* "View.MemoryView":743 + * + * cdef __Pyx_memviewslice *p_dst = &dst + * cdef int *p_suboffset_dim = &suboffset_dim # <<<<<<<<<<<<<< + * cdef Py_ssize_t start, stop, step, cindex + * cdef bint have_start, have_stop, have_step + */ + __pyx_v_p_suboffset_dim = (&__pyx_v_suboffset_dim); + + /* "View.MemoryView":747 + * cdef bint have_start, have_stop, have_step + * + * for dim, index in enumerate(indices): # <<<<<<<<<<<<<< + * if PyIndex_Check(index): + * cindex = index + */ + __pyx_t_5 = 0; + if (likely(PyList_CheckExact(__pyx_v_indices)) || PyTuple_CheckExact(__pyx_v_indices)) { + __pyx_t_2 = __pyx_v_indices; __Pyx_INCREF(__pyx_t_2); + __pyx_t_6 = 0; + __pyx_t_7 = NULL; + } else { + __pyx_t_6 = -1; __pyx_t_2 = PyObject_GetIter(__pyx_v_indices); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 747, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_7 = __Pyx_PyObject_GetIterNextFunc(__pyx_t_2); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 747, __pyx_L1_error) + } + for (;;) { + if (likely(!__pyx_t_7)) { + if (likely(PyList_CheckExact(__pyx_t_2))) { + { + Py_ssize_t __pyx_temp = __Pyx_PyList_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #endif + if (__pyx_t_6 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_8 = PyList_GET_ITEM(__pyx_t_2, __pyx_t_6); __Pyx_INCREF(__pyx_t_8); __pyx_t_6++; if (unlikely((0 < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #else + __pyx_t_8 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_6); __pyx_t_6++; if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 747, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + #endif + } else { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #endif + if (__pyx_t_6 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_8 = PyTuple_GET_ITEM(__pyx_t_2, __pyx_t_6); __Pyx_INCREF(__pyx_t_8); __pyx_t_6++; if (unlikely((0 < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #else + __pyx_t_8 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_6); __pyx_t_6++; if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 747, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + #endif + } + } else { + __pyx_t_8 = __pyx_t_7(__pyx_t_2); + if (unlikely(!__pyx_t_8)) { + PyObject* exc_type = PyErr_Occurred(); + if (exc_type) { + if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear(); + else __PYX_ERR(1, 747, __pyx_L1_error) + } + break; + } + __Pyx_GOTREF(__pyx_t_8); + } + __Pyx_XDECREF_SET(__pyx_v_index, __pyx_t_8); + __pyx_t_8 = 0; + __pyx_v_dim = __pyx_t_5; + __pyx_t_5 = (__pyx_t_5 + 1); + + /* "View.MemoryView":748 + * + * for dim, index in enumerate(indices): + * if PyIndex_Check(index): # <<<<<<<<<<<<<< + * cindex = index + * slice_memviewslice( + */ + __pyx_t_1 = (PyIndex_Check(__pyx_v_index) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":749 + * for dim, index in enumerate(indices): + * if PyIndex_Check(index): + * cindex = index # <<<<<<<<<<<<<< + * slice_memviewslice( + * p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim], + */ + __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_v_index); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 749, __pyx_L1_error) + __pyx_v_cindex = __pyx_t_9; + + /* "View.MemoryView":750 + * if PyIndex_Check(index): + * cindex = index + * slice_memviewslice( # <<<<<<<<<<<<<< + * p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim], + * dim, new_ndim, p_suboffset_dim, + */ + __pyx_t_10 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_v_cindex, 0, 0, 0, 0, 0, 0); if (unlikely(__pyx_t_10 == ((int)-1))) __PYX_ERR(1, 750, __pyx_L1_error) + + /* "View.MemoryView":748 + * + * for dim, index in enumerate(indices): + * if PyIndex_Check(index): # <<<<<<<<<<<<<< + * cindex = index + * slice_memviewslice( + */ + goto __pyx_L6; + } + + /* "View.MemoryView":756 + * 0, 0, 0, # have_{start,stop,step} + * False) + * elif index is None: # <<<<<<<<<<<<<< + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 + */ + __pyx_t_1 = (__pyx_v_index == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":757 + * False) + * elif index is None: + * p_dst.shape[new_ndim] = 1 # <<<<<<<<<<<<<< + * p_dst.strides[new_ndim] = 0 + * p_dst.suboffsets[new_ndim] = -1 + */ + (__pyx_v_p_dst->shape[__pyx_v_new_ndim]) = 1; + + /* "View.MemoryView":758 + * elif index is None: + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 # <<<<<<<<<<<<<< + * p_dst.suboffsets[new_ndim] = -1 + * new_ndim += 1 + */ + (__pyx_v_p_dst->strides[__pyx_v_new_ndim]) = 0; + + /* "View.MemoryView":759 + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 + * p_dst.suboffsets[new_ndim] = -1 # <<<<<<<<<<<<<< + * new_ndim += 1 + * else: + */ + (__pyx_v_p_dst->suboffsets[__pyx_v_new_ndim]) = -1L; + + /* "View.MemoryView":760 + * p_dst.strides[new_ndim] = 0 + * p_dst.suboffsets[new_ndim] = -1 + * new_ndim += 1 # <<<<<<<<<<<<<< + * else: + * start = index.start or 0 + */ + __pyx_v_new_ndim = (__pyx_v_new_ndim + 1); + + /* "View.MemoryView":756 + * 0, 0, 0, # have_{start,stop,step} + * False) + * elif index is None: # <<<<<<<<<<<<<< + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 + */ + goto __pyx_L6; + } + + /* "View.MemoryView":762 + * new_ndim += 1 + * else: + * start = index.start or 0 # <<<<<<<<<<<<<< + * stop = index.stop or 0 + * step = index.step or 0 + */ + /*else*/ { + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 762, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_8); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 762, __pyx_L1_error) + if (!__pyx_t_1) { + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + } else { + __pyx_t_11 = __Pyx_PyIndex_AsSsize_t(__pyx_t_8); if (unlikely((__pyx_t_11 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 762, __pyx_L1_error) + __pyx_t_9 = __pyx_t_11; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L7_bool_binop_done; + } + __pyx_t_9 = 0; + __pyx_L7_bool_binop_done:; + __pyx_v_start = __pyx_t_9; + + /* "View.MemoryView":763 + * else: + * start = index.start or 0 + * stop = index.stop or 0 # <<<<<<<<<<<<<< + * step = index.step or 0 + * + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 763, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_8); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 763, __pyx_L1_error) + if (!__pyx_t_1) { + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + } else { + __pyx_t_11 = __Pyx_PyIndex_AsSsize_t(__pyx_t_8); if (unlikely((__pyx_t_11 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 763, __pyx_L1_error) + __pyx_t_9 = __pyx_t_11; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L9_bool_binop_done; + } + __pyx_t_9 = 0; + __pyx_L9_bool_binop_done:; + __pyx_v_stop = __pyx_t_9; + + /* "View.MemoryView":764 + * start = index.start or 0 + * stop = index.stop or 0 + * step = index.step or 0 # <<<<<<<<<<<<<< + * + * have_start = index.start is not None + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 764, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_8); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 764, __pyx_L1_error) + if (!__pyx_t_1) { + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + } else { + __pyx_t_11 = __Pyx_PyIndex_AsSsize_t(__pyx_t_8); if (unlikely((__pyx_t_11 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 764, __pyx_L1_error) + __pyx_t_9 = __pyx_t_11; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L11_bool_binop_done; + } + __pyx_t_9 = 0; + __pyx_L11_bool_binop_done:; + __pyx_v_step = __pyx_t_9; + + /* "View.MemoryView":766 + * step = index.step or 0 + * + * have_start = index.start is not None # <<<<<<<<<<<<<< + * have_stop = index.stop is not None + * have_step = index.step is not None + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 766, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = (__pyx_t_8 != Py_None); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __pyx_v_have_start = __pyx_t_1; + + /* "View.MemoryView":767 + * + * have_start = index.start is not None + * have_stop = index.stop is not None # <<<<<<<<<<<<<< + * have_step = index.step is not None + * + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 767, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = (__pyx_t_8 != Py_None); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __pyx_v_have_stop = __pyx_t_1; + + /* "View.MemoryView":768 + * have_start = index.start is not None + * have_stop = index.stop is not None + * have_step = index.step is not None # <<<<<<<<<<<<<< + * + * slice_memviewslice( + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 768, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = (__pyx_t_8 != Py_None); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __pyx_v_have_step = __pyx_t_1; + + /* "View.MemoryView":770 + * have_step = index.step is not None + * + * slice_memviewslice( # <<<<<<<<<<<<<< + * p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim], + * dim, new_ndim, p_suboffset_dim, + */ + __pyx_t_10 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_v_start, __pyx_v_stop, __pyx_v_step, __pyx_v_have_start, __pyx_v_have_stop, __pyx_v_have_step, 1); if (unlikely(__pyx_t_10 == ((int)-1))) __PYX_ERR(1, 770, __pyx_L1_error) + + /* "View.MemoryView":776 + * have_start, have_stop, have_step, + * True) + * new_ndim += 1 # <<<<<<<<<<<<<< + * + * if isinstance(memview, _memoryviewslice): + */ + __pyx_v_new_ndim = (__pyx_v_new_ndim + 1); + } + __pyx_L6:; + + /* "View.MemoryView":747 + * cdef bint have_start, have_stop, have_step + * + * for dim, index in enumerate(indices): # <<<<<<<<<<<<<< + * if PyIndex_Check(index): + * cindex = index + */ + } + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":778 + * new_ndim += 1 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":779 + * + * if isinstance(memview, _memoryviewslice): + * return memoryview_fromslice(dst, new_ndim, # <<<<<<<<<<<<<< + * memviewsliceobj.to_object_func, + * memviewsliceobj.to_dtype_func, + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + + /* "View.MemoryView":780 + * if isinstance(memview, _memoryviewslice): + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, # <<<<<<<<<<<<<< + * memviewsliceobj.to_dtype_func, + * memview.dtype_is_object) + */ + if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 780, __pyx_L1_error) } + + /* "View.MemoryView":781 + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, + * memviewsliceobj.to_dtype_func, # <<<<<<<<<<<<<< + * memview.dtype_is_object) + * else: + */ + if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 781, __pyx_L1_error) } + + /* "View.MemoryView":779 + * + * if isinstance(memview, _memoryviewslice): + * return memoryview_fromslice(dst, new_ndim, # <<<<<<<<<<<<<< + * memviewsliceobj.to_object_func, + * memviewsliceobj.to_dtype_func, + */ + __pyx_t_2 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, __pyx_v_memviewsliceobj->to_object_func, __pyx_v_memviewsliceobj->to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 779, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + if (!(likely(((__pyx_t_2) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_2, __pyx_memoryview_type))))) __PYX_ERR(1, 779, __pyx_L1_error) + __pyx_r = ((struct __pyx_memoryview_obj *)__pyx_t_2); + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":778 + * new_ndim += 1 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, + */ + } + + /* "View.MemoryView":784 + * memview.dtype_is_object) + * else: + * return memoryview_fromslice(dst, new_ndim, NULL, NULL, # <<<<<<<<<<<<<< + * memview.dtype_is_object) + * + */ + /*else*/ { + __Pyx_XDECREF((PyObject *)__pyx_r); + + /* "View.MemoryView":785 + * else: + * return memoryview_fromslice(dst, new_ndim, NULL, NULL, + * memview.dtype_is_object) # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, NULL, NULL, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 784, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + + /* "View.MemoryView":784 + * memview.dtype_is_object) + * else: + * return memoryview_fromslice(dst, new_ndim, NULL, NULL, # <<<<<<<<<<<<<< + * memview.dtype_is_object) + * + */ + if (!(likely(((__pyx_t_2) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_2, __pyx_memoryview_type))))) __PYX_ERR(1, 784, __pyx_L1_error) + __pyx_r = ((struct __pyx_memoryview_obj *)__pyx_t_2); + __pyx_t_2 = 0; + goto __pyx_L0; + } + + /* "View.MemoryView":711 + * + * @cname('__pyx_memview_slice') + * cdef memoryview memview_slice(memoryview memview, object indices): # <<<<<<<<<<<<<< + * cdef int new_ndim = 0, suboffset_dim = -1, dim + * cdef bint negative_step + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("View.MemoryView.memview_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_memviewsliceobj); + __Pyx_XDECREF(__pyx_v_index); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":793 + * + * @cname('__pyx_memoryview_slice_memviewslice') + * cdef int slice_memviewslice( # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * Py_ssize_t shape, Py_ssize_t stride, Py_ssize_t suboffset, + */ + +static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst, Py_ssize_t __pyx_v_shape, Py_ssize_t __pyx_v_stride, Py_ssize_t __pyx_v_suboffset, int __pyx_v_dim, int __pyx_v_new_ndim, int *__pyx_v_suboffset_dim, Py_ssize_t __pyx_v_start, Py_ssize_t __pyx_v_stop, Py_ssize_t __pyx_v_step, int __pyx_v_have_start, int __pyx_v_have_stop, int __pyx_v_have_step, int __pyx_v_is_slice) { + Py_ssize_t __pyx_v_new_shape; + int __pyx_v_negative_step; + int __pyx_r; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":813 + * cdef bint negative_step + * + * if not is_slice: # <<<<<<<<<<<<<< + * + * if start < 0: + */ + __pyx_t_1 = (!__pyx_v_is_slice); + if (__pyx_t_1) { + + /* "View.MemoryView":815 + * if not is_slice: + * + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if not 0 <= start < shape: + */ + __pyx_t_1 = (__pyx_v_start < 0); + if (__pyx_t_1) { + + /* "View.MemoryView":816 + * + * if start < 0: + * start += shape # <<<<<<<<<<<<<< + * if not 0 <= start < shape: + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) + */ + __pyx_v_start = (__pyx_v_start + __pyx_v_shape); + + /* "View.MemoryView":815 + * if not is_slice: + * + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if not 0 <= start < shape: + */ + } + + /* "View.MemoryView":817 + * if start < 0: + * start += shape + * if not 0 <= start < shape: # <<<<<<<<<<<<<< + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) + * else: + */ + __pyx_t_1 = (0 <= __pyx_v_start); + if (__pyx_t_1) { + __pyx_t_1 = (__pyx_v_start < __pyx_v_shape); + } + __pyx_t_2 = (!__pyx_t_1); + if (__pyx_t_2) { + + /* "View.MemoryView":818 + * start += shape + * if not 0 <= start < shape: + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) # <<<<<<<<<<<<<< + * else: + * + */ + __pyx_t_3 = __pyx_memoryview_err_dim(PyExc_IndexError, __pyx_kp_s_Index_out_of_bounds_axis_d, __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 818, __pyx_L1_error) + + /* "View.MemoryView":817 + * if start < 0: + * start += shape + * if not 0 <= start < shape: # <<<<<<<<<<<<<< + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) + * else: + */ + } + + /* "View.MemoryView":813 + * cdef bint negative_step + * + * if not is_slice: # <<<<<<<<<<<<<< + * + * if start < 0: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":821 + * else: + * + * if have_step: # <<<<<<<<<<<<<< + * negative_step = step < 0 + * if step == 0: + */ + /*else*/ { + __pyx_t_2 = (__pyx_v_have_step != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":822 + * + * if have_step: + * negative_step = step < 0 # <<<<<<<<<<<<<< + * if step == 0: + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + */ + __pyx_v_negative_step = (__pyx_v_step < 0); + + /* "View.MemoryView":823 + * if have_step: + * negative_step = step < 0 + * if step == 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + * else: + */ + __pyx_t_2 = (__pyx_v_step == 0); + if (__pyx_t_2) { + + /* "View.MemoryView":824 + * negative_step = step < 0 + * if step == 0: + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) # <<<<<<<<<<<<<< + * else: + * negative_step = False + */ + __pyx_t_3 = __pyx_memoryview_err_dim(PyExc_ValueError, __pyx_kp_s_Step_may_not_be_zero_axis_d, __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 824, __pyx_L1_error) + + /* "View.MemoryView":823 + * if have_step: + * negative_step = step < 0 + * if step == 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + * else: + */ + } + + /* "View.MemoryView":821 + * else: + * + * if have_step: # <<<<<<<<<<<<<< + * negative_step = step < 0 + * if step == 0: + */ + goto __pyx_L6; + } + + /* "View.MemoryView":826 + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + * else: + * negative_step = False # <<<<<<<<<<<<<< + * step = 1 + * + */ + /*else*/ { + __pyx_v_negative_step = 0; + + /* "View.MemoryView":827 + * else: + * negative_step = False + * step = 1 # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_step = 1; + } + __pyx_L6:; + + /* "View.MemoryView":830 + * + * + * if have_start: # <<<<<<<<<<<<<< + * if start < 0: + * start += shape + */ + __pyx_t_2 = (__pyx_v_have_start != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":831 + * + * if have_start: + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if start < 0: + */ + __pyx_t_2 = (__pyx_v_start < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":832 + * if have_start: + * if start < 0: + * start += shape # <<<<<<<<<<<<<< + * if start < 0: + * start = 0 + */ + __pyx_v_start = (__pyx_v_start + __pyx_v_shape); + + /* "View.MemoryView":833 + * if start < 0: + * start += shape + * if start < 0: # <<<<<<<<<<<<<< + * start = 0 + * elif start >= shape: + */ + __pyx_t_2 = (__pyx_v_start < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":834 + * start += shape + * if start < 0: + * start = 0 # <<<<<<<<<<<<<< + * elif start >= shape: + * if negative_step: + */ + __pyx_v_start = 0; + + /* "View.MemoryView":833 + * if start < 0: + * start += shape + * if start < 0: # <<<<<<<<<<<<<< + * start = 0 + * elif start >= shape: + */ + } + + /* "View.MemoryView":831 + * + * if have_start: + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if start < 0: + */ + goto __pyx_L9; + } + + /* "View.MemoryView":835 + * if start < 0: + * start = 0 + * elif start >= shape: # <<<<<<<<<<<<<< + * if negative_step: + * start = shape - 1 + */ + __pyx_t_2 = (__pyx_v_start >= __pyx_v_shape); + if (__pyx_t_2) { + + /* "View.MemoryView":836 + * start = 0 + * elif start >= shape: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + if (__pyx_v_negative_step) { + + /* "View.MemoryView":837 + * elif start >= shape: + * if negative_step: + * start = shape - 1 # <<<<<<<<<<<<<< + * else: + * start = shape + */ + __pyx_v_start = (__pyx_v_shape - 1); + + /* "View.MemoryView":836 + * start = 0 + * elif start >= shape: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + goto __pyx_L11; + } + + /* "View.MemoryView":839 + * start = shape - 1 + * else: + * start = shape # <<<<<<<<<<<<<< + * else: + * if negative_step: + */ + /*else*/ { + __pyx_v_start = __pyx_v_shape; + } + __pyx_L11:; + + /* "View.MemoryView":835 + * if start < 0: + * start = 0 + * elif start >= shape: # <<<<<<<<<<<<<< + * if negative_step: + * start = shape - 1 + */ + } + __pyx_L9:; + + /* "View.MemoryView":830 + * + * + * if have_start: # <<<<<<<<<<<<<< + * if start < 0: + * start += shape + */ + goto __pyx_L8; + } + + /* "View.MemoryView":841 + * start = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + /*else*/ { + if (__pyx_v_negative_step) { + + /* "View.MemoryView":842 + * else: + * if negative_step: + * start = shape - 1 # <<<<<<<<<<<<<< + * else: + * start = 0 + */ + __pyx_v_start = (__pyx_v_shape - 1); + + /* "View.MemoryView":841 + * start = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + goto __pyx_L12; + } + + /* "View.MemoryView":844 + * start = shape - 1 + * else: + * start = 0 # <<<<<<<<<<<<<< + * + * if have_stop: + */ + /*else*/ { + __pyx_v_start = 0; + } + __pyx_L12:; + } + __pyx_L8:; + + /* "View.MemoryView":846 + * start = 0 + * + * if have_stop: # <<<<<<<<<<<<<< + * if stop < 0: + * stop += shape + */ + __pyx_t_2 = (__pyx_v_have_stop != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":847 + * + * if have_stop: + * if stop < 0: # <<<<<<<<<<<<<< + * stop += shape + * if stop < 0: + */ + __pyx_t_2 = (__pyx_v_stop < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":848 + * if have_stop: + * if stop < 0: + * stop += shape # <<<<<<<<<<<<<< + * if stop < 0: + * stop = 0 + */ + __pyx_v_stop = (__pyx_v_stop + __pyx_v_shape); + + /* "View.MemoryView":849 + * if stop < 0: + * stop += shape + * if stop < 0: # <<<<<<<<<<<<<< + * stop = 0 + * elif stop > shape: + */ + __pyx_t_2 = (__pyx_v_stop < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":850 + * stop += shape + * if stop < 0: + * stop = 0 # <<<<<<<<<<<<<< + * elif stop > shape: + * stop = shape + */ + __pyx_v_stop = 0; + + /* "View.MemoryView":849 + * if stop < 0: + * stop += shape + * if stop < 0: # <<<<<<<<<<<<<< + * stop = 0 + * elif stop > shape: + */ + } + + /* "View.MemoryView":847 + * + * if have_stop: + * if stop < 0: # <<<<<<<<<<<<<< + * stop += shape + * if stop < 0: + */ + goto __pyx_L14; + } + + /* "View.MemoryView":851 + * if stop < 0: + * stop = 0 + * elif stop > shape: # <<<<<<<<<<<<<< + * stop = shape + * else: + */ + __pyx_t_2 = (__pyx_v_stop > __pyx_v_shape); + if (__pyx_t_2) { + + /* "View.MemoryView":852 + * stop = 0 + * elif stop > shape: + * stop = shape # <<<<<<<<<<<<<< + * else: + * if negative_step: + */ + __pyx_v_stop = __pyx_v_shape; + + /* "View.MemoryView":851 + * if stop < 0: + * stop = 0 + * elif stop > shape: # <<<<<<<<<<<<<< + * stop = shape + * else: + */ + } + __pyx_L14:; + + /* "View.MemoryView":846 + * start = 0 + * + * if have_stop: # <<<<<<<<<<<<<< + * if stop < 0: + * stop += shape + */ + goto __pyx_L13; + } + + /* "View.MemoryView":854 + * stop = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * stop = -1 + * else: + */ + /*else*/ { + if (__pyx_v_negative_step) { + + /* "View.MemoryView":855 + * else: + * if negative_step: + * stop = -1 # <<<<<<<<<<<<<< + * else: + * stop = shape + */ + __pyx_v_stop = -1L; + + /* "View.MemoryView":854 + * stop = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * stop = -1 + * else: + */ + goto __pyx_L16; + } + + /* "View.MemoryView":857 + * stop = -1 + * else: + * stop = shape # <<<<<<<<<<<<<< + * + * + */ + /*else*/ { + __pyx_v_stop = __pyx_v_shape; + } + __pyx_L16:; + } + __pyx_L13:; + + /* "View.MemoryView":861 + * + * with cython.cdivision(True): + * new_shape = (stop - start) // step # <<<<<<<<<<<<<< + * + * if (stop - start) - step * new_shape: + */ + __pyx_v_new_shape = ((__pyx_v_stop - __pyx_v_start) / __pyx_v_step); + + /* "View.MemoryView":863 + * new_shape = (stop - start) // step + * + * if (stop - start) - step * new_shape: # <<<<<<<<<<<<<< + * new_shape += 1 + * + */ + __pyx_t_2 = (((__pyx_v_stop - __pyx_v_start) - (__pyx_v_step * __pyx_v_new_shape)) != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":864 + * + * if (stop - start) - step * new_shape: + * new_shape += 1 # <<<<<<<<<<<<<< + * + * if new_shape < 0: + */ + __pyx_v_new_shape = (__pyx_v_new_shape + 1); + + /* "View.MemoryView":863 + * new_shape = (stop - start) // step + * + * if (stop - start) - step * new_shape: # <<<<<<<<<<<<<< + * new_shape += 1 + * + */ + } + + /* "View.MemoryView":866 + * new_shape += 1 + * + * if new_shape < 0: # <<<<<<<<<<<<<< + * new_shape = 0 + * + */ + __pyx_t_2 = (__pyx_v_new_shape < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":867 + * + * if new_shape < 0: + * new_shape = 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_new_shape = 0; + + /* "View.MemoryView":866 + * new_shape += 1 + * + * if new_shape < 0: # <<<<<<<<<<<<<< + * new_shape = 0 + * + */ + } + + /* "View.MemoryView":870 + * + * + * dst.strides[new_ndim] = stride * step # <<<<<<<<<<<<<< + * dst.shape[new_ndim] = new_shape + * dst.suboffsets[new_ndim] = suboffset + */ + (__pyx_v_dst->strides[__pyx_v_new_ndim]) = (__pyx_v_stride * __pyx_v_step); + + /* "View.MemoryView":871 + * + * dst.strides[new_ndim] = stride * step + * dst.shape[new_ndim] = new_shape # <<<<<<<<<<<<<< + * dst.suboffsets[new_ndim] = suboffset + * + */ + (__pyx_v_dst->shape[__pyx_v_new_ndim]) = __pyx_v_new_shape; + + /* "View.MemoryView":872 + * dst.strides[new_ndim] = stride * step + * dst.shape[new_ndim] = new_shape + * dst.suboffsets[new_ndim] = suboffset # <<<<<<<<<<<<<< + * + * + */ + (__pyx_v_dst->suboffsets[__pyx_v_new_ndim]) = __pyx_v_suboffset; + } + __pyx_L3:; + + /* "View.MemoryView":875 + * + * + * if suboffset_dim[0] < 0: # <<<<<<<<<<<<<< + * dst.data += start * stride + * else: + */ + __pyx_t_2 = ((__pyx_v_suboffset_dim[0]) < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":876 + * + * if suboffset_dim[0] < 0: + * dst.data += start * stride # <<<<<<<<<<<<<< + * else: + * dst.suboffsets[suboffset_dim[0]] += start * stride + */ + __pyx_v_dst->data = (__pyx_v_dst->data + (__pyx_v_start * __pyx_v_stride)); + + /* "View.MemoryView":875 + * + * + * if suboffset_dim[0] < 0: # <<<<<<<<<<<<<< + * dst.data += start * stride + * else: + */ + goto __pyx_L19; + } + + /* "View.MemoryView":878 + * dst.data += start * stride + * else: + * dst.suboffsets[suboffset_dim[0]] += start * stride # <<<<<<<<<<<<<< + * + * if suboffset >= 0: + */ + /*else*/ { + __pyx_t_3 = (__pyx_v_suboffset_dim[0]); + (__pyx_v_dst->suboffsets[__pyx_t_3]) = ((__pyx_v_dst->suboffsets[__pyx_t_3]) + (__pyx_v_start * __pyx_v_stride)); + } + __pyx_L19:; + + /* "View.MemoryView":880 + * dst.suboffsets[suboffset_dim[0]] += start * stride + * + * if suboffset >= 0: # <<<<<<<<<<<<<< + * if not is_slice: + * if new_ndim == 0: + */ + __pyx_t_2 = (__pyx_v_suboffset >= 0); + if (__pyx_t_2) { + + /* "View.MemoryView":881 + * + * if suboffset >= 0: + * if not is_slice: # <<<<<<<<<<<<<< + * if new_ndim == 0: + * dst.data = ( dst.data)[0] + suboffset + */ + __pyx_t_2 = (!__pyx_v_is_slice); + if (__pyx_t_2) { + + /* "View.MemoryView":882 + * if suboffset >= 0: + * if not is_slice: + * if new_ndim == 0: # <<<<<<<<<<<<<< + * dst.data = ( dst.data)[0] + suboffset + * else: + */ + __pyx_t_2 = (__pyx_v_new_ndim == 0); + if (__pyx_t_2) { + + /* "View.MemoryView":883 + * if not is_slice: + * if new_ndim == 0: + * dst.data = ( dst.data)[0] + suboffset # <<<<<<<<<<<<<< + * else: + * _err_dim(PyExc_IndexError, "All dimensions preceding dimension %d " + */ + __pyx_v_dst->data = ((((char **)__pyx_v_dst->data)[0]) + __pyx_v_suboffset); + + /* "View.MemoryView":882 + * if suboffset >= 0: + * if not is_slice: + * if new_ndim == 0: # <<<<<<<<<<<<<< + * dst.data = ( dst.data)[0] + suboffset + * else: + */ + goto __pyx_L22; + } + + /* "View.MemoryView":885 + * dst.data = ( dst.data)[0] + suboffset + * else: + * _err_dim(PyExc_IndexError, "All dimensions preceding dimension %d " # <<<<<<<<<<<<<< + * "must be indexed and not sliced", dim) + * else: + */ + /*else*/ { + + /* "View.MemoryView":886 + * else: + * _err_dim(PyExc_IndexError, "All dimensions preceding dimension %d " + * "must be indexed and not sliced", dim) # <<<<<<<<<<<<<< + * else: + * suboffset_dim[0] = new_ndim + */ + __pyx_t_3 = __pyx_memoryview_err_dim(PyExc_IndexError, __pyx_kp_s_All_dimensions_preceding_dimensi, __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 885, __pyx_L1_error) + } + __pyx_L22:; + + /* "View.MemoryView":881 + * + * if suboffset >= 0: + * if not is_slice: # <<<<<<<<<<<<<< + * if new_ndim == 0: + * dst.data = ( dst.data)[0] + suboffset + */ + goto __pyx_L21; + } + + /* "View.MemoryView":888 + * "must be indexed and not sliced", dim) + * else: + * suboffset_dim[0] = new_ndim # <<<<<<<<<<<<<< + * + * return 0 + */ + /*else*/ { + (__pyx_v_suboffset_dim[0]) = __pyx_v_new_ndim; + } + __pyx_L21:; + + /* "View.MemoryView":880 + * dst.suboffsets[suboffset_dim[0]] += start * stride + * + * if suboffset >= 0: # <<<<<<<<<<<<<< + * if not is_slice: + * if new_ndim == 0: + */ + } + + /* "View.MemoryView":890 + * suboffset_dim[0] = new_ndim + * + * return 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":793 + * + * @cname('__pyx_memoryview_slice_memviewslice') + * cdef int slice_memviewslice( # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * Py_ssize_t shape, Py_ssize_t stride, Py_ssize_t suboffset, + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.slice_memviewslice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":896 + * + * @cname('__pyx_pybuffer_index') + * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index, # <<<<<<<<<<<<<< + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 + */ + +static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, Py_ssize_t __pyx_v_index, Py_ssize_t __pyx_v_dim) { + Py_ssize_t __pyx_v_shape; + Py_ssize_t __pyx_v_stride; + Py_ssize_t __pyx_v_suboffset; + Py_ssize_t __pyx_v_itemsize; + char *__pyx_v_resultp; + char *__pyx_r; + __Pyx_RefNannyDeclarations + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + Py_UCS4 __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("pybuffer_index", 1); + + /* "View.MemoryView":898 + * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index, + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 # <<<<<<<<<<<<<< + * cdef Py_ssize_t itemsize = view.itemsize + * cdef char *resultp + */ + __pyx_v_suboffset = -1L; + + /* "View.MemoryView":899 + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 + * cdef Py_ssize_t itemsize = view.itemsize # <<<<<<<<<<<<<< + * cdef char *resultp + * + */ + __pyx_t_1 = __pyx_v_view->itemsize; + __pyx_v_itemsize = __pyx_t_1; + + /* "View.MemoryView":902 + * cdef char *resultp + * + * if view.ndim == 0: # <<<<<<<<<<<<<< + * shape = view.len // itemsize + * stride = itemsize + */ + __pyx_t_2 = (__pyx_v_view->ndim == 0); + if (__pyx_t_2) { + + /* "View.MemoryView":903 + * + * if view.ndim == 0: + * shape = view.len // itemsize # <<<<<<<<<<<<<< + * stride = itemsize + * else: + */ + if (unlikely(__pyx_v_itemsize == 0)) { + PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero"); + __PYX_ERR(1, 903, __pyx_L1_error) + } + else if (sizeof(Py_ssize_t) == sizeof(long) && (!(((Py_ssize_t)-1) > 0)) && unlikely(__pyx_v_itemsize == (Py_ssize_t)-1) && unlikely(__Pyx_UNARY_NEG_WOULD_OVERFLOW(__pyx_v_view->len))) { + PyErr_SetString(PyExc_OverflowError, "value too large to perform division"); + __PYX_ERR(1, 903, __pyx_L1_error) + } + __pyx_v_shape = __Pyx_div_Py_ssize_t(__pyx_v_view->len, __pyx_v_itemsize); + + /* "View.MemoryView":904 + * if view.ndim == 0: + * shape = view.len // itemsize + * stride = itemsize # <<<<<<<<<<<<<< + * else: + * shape = view.shape[dim] + */ + __pyx_v_stride = __pyx_v_itemsize; + + /* "View.MemoryView":902 + * cdef char *resultp + * + * if view.ndim == 0: # <<<<<<<<<<<<<< + * shape = view.len // itemsize + * stride = itemsize + */ + goto __pyx_L3; + } + + /* "View.MemoryView":906 + * stride = itemsize + * else: + * shape = view.shape[dim] # <<<<<<<<<<<<<< + * stride = view.strides[dim] + * if view.suboffsets != NULL: + */ + /*else*/ { + __pyx_v_shape = (__pyx_v_view->shape[__pyx_v_dim]); + + /* "View.MemoryView":907 + * else: + * shape = view.shape[dim] + * stride = view.strides[dim] # <<<<<<<<<<<<<< + * if view.suboffsets != NULL: + * suboffset = view.suboffsets[dim] + */ + __pyx_v_stride = (__pyx_v_view->strides[__pyx_v_dim]); + + /* "View.MemoryView":908 + * shape = view.shape[dim] + * stride = view.strides[dim] + * if view.suboffsets != NULL: # <<<<<<<<<<<<<< + * suboffset = view.suboffsets[dim] + * + */ + __pyx_t_2 = (__pyx_v_view->suboffsets != NULL); + if (__pyx_t_2) { + + /* "View.MemoryView":909 + * stride = view.strides[dim] + * if view.suboffsets != NULL: + * suboffset = view.suboffsets[dim] # <<<<<<<<<<<<<< + * + * if index < 0: + */ + __pyx_v_suboffset = (__pyx_v_view->suboffsets[__pyx_v_dim]); + + /* "View.MemoryView":908 + * shape = view.shape[dim] + * stride = view.strides[dim] + * if view.suboffsets != NULL: # <<<<<<<<<<<<<< + * suboffset = view.suboffsets[dim] + * + */ + } + } + __pyx_L3:; + + /* "View.MemoryView":911 + * suboffset = view.suboffsets[dim] + * + * if index < 0: # <<<<<<<<<<<<<< + * index += view.shape[dim] + * if index < 0: + */ + __pyx_t_2 = (__pyx_v_index < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":912 + * + * if index < 0: + * index += view.shape[dim] # <<<<<<<<<<<<<< + * if index < 0: + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + */ + __pyx_v_index = (__pyx_v_index + (__pyx_v_view->shape[__pyx_v_dim])); + + /* "View.MemoryView":913 + * if index < 0: + * index += view.shape[dim] + * if index < 0: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + __pyx_t_2 = (__pyx_v_index < 0); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":914 + * index += view.shape[dim] + * if index < 0: + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" # <<<<<<<<<<<<<< + * + * if index >= shape: + */ + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 914, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 = 0; + __pyx_t_4 = 127; + __Pyx_INCREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_1 += 37; + __Pyx_GIVEREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_5 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_dim, 0, ' ', 'd'); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 914, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_1 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_5); + __Pyx_GIVEREF(__pyx_t_5); + PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_5); + __pyx_t_5 = 0; + __Pyx_INCREF(__pyx_kp_u__7); + __pyx_t_1 += 1; + __Pyx_GIVEREF(__pyx_kp_u__7); + PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_kp_u__7); + __pyx_t_5 = __Pyx_PyUnicode_Join(__pyx_t_3, 3, __pyx_t_1, __pyx_t_4); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 914, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_Raise(__pyx_builtin_IndexError, __pyx_t_5, 0, 0); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __PYX_ERR(1, 914, __pyx_L1_error) + + /* "View.MemoryView":913 + * if index < 0: + * index += view.shape[dim] + * if index < 0: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + } + + /* "View.MemoryView":911 + * suboffset = view.suboffsets[dim] + * + * if index < 0: # <<<<<<<<<<<<<< + * index += view.shape[dim] + * if index < 0: + */ + } + + /* "View.MemoryView":916 + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + * if index >= shape: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + __pyx_t_2 = (__pyx_v_index >= __pyx_v_shape); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":917 + * + * if index >= shape: + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" # <<<<<<<<<<<<<< + * + * resultp = bufp + index * stride + */ + __pyx_t_5 = PyTuple_New(3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 917, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_1 = 0; + __pyx_t_4 = 127; + __Pyx_INCREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_1 += 37; + __Pyx_GIVEREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_3 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_dim, 0, ' ', 'd'); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 917, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_3); + PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_t_3); + __pyx_t_3 = 0; + __Pyx_INCREF(__pyx_kp_u__7); + __pyx_t_1 += 1; + __Pyx_GIVEREF(__pyx_kp_u__7); + PyTuple_SET_ITEM(__pyx_t_5, 2, __pyx_kp_u__7); + __pyx_t_3 = __Pyx_PyUnicode_Join(__pyx_t_5, 3, __pyx_t_1, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 917, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_Raise(__pyx_builtin_IndexError, __pyx_t_3, 0, 0); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __PYX_ERR(1, 917, __pyx_L1_error) + + /* "View.MemoryView":916 + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + * if index >= shape: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + } + + /* "View.MemoryView":919 + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + * resultp = bufp + index * stride # <<<<<<<<<<<<<< + * if suboffset >= 0: + * resultp = ( resultp)[0] + suboffset + */ + __pyx_v_resultp = (__pyx_v_bufp + (__pyx_v_index * __pyx_v_stride)); + + /* "View.MemoryView":920 + * + * resultp = bufp + index * stride + * if suboffset >= 0: # <<<<<<<<<<<<<< + * resultp = ( resultp)[0] + suboffset + * + */ + __pyx_t_2 = (__pyx_v_suboffset >= 0); + if (__pyx_t_2) { + + /* "View.MemoryView":921 + * resultp = bufp + index * stride + * if suboffset >= 0: + * resultp = ( resultp)[0] + suboffset # <<<<<<<<<<<<<< + * + * return resultp + */ + __pyx_v_resultp = ((((char **)__pyx_v_resultp)[0]) + __pyx_v_suboffset); + + /* "View.MemoryView":920 + * + * resultp = bufp + index * stride + * if suboffset >= 0: # <<<<<<<<<<<<<< + * resultp = ( resultp)[0] + suboffset + * + */ + } + + /* "View.MemoryView":923 + * resultp = ( resultp)[0] + suboffset + * + * return resultp # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = __pyx_v_resultp; + goto __pyx_L0; + + /* "View.MemoryView":896 + * + * @cname('__pyx_pybuffer_index') + * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index, # <<<<<<<<<<<<<< + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.pybuffer_index", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":929 + * + * @cname('__pyx_memslice_transpose') + * cdef int transpose_memslice(__Pyx_memviewslice *memslice) except -1 nogil: # <<<<<<<<<<<<<< + * cdef int ndim = memslice.memview.view.ndim + * + */ + +static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) { + int __pyx_v_ndim; + Py_ssize_t *__pyx_v_shape; + Py_ssize_t *__pyx_v_strides; + int __pyx_v_i; + int __pyx_v_j; + int __pyx_r; + int __pyx_t_1; + Py_ssize_t *__pyx_t_2; + long __pyx_t_3; + long __pyx_t_4; + Py_ssize_t __pyx_t_5; + Py_ssize_t __pyx_t_6; + int __pyx_t_7; + int __pyx_t_8; + int __pyx_t_9; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":930 + * @cname('__pyx_memslice_transpose') + * cdef int transpose_memslice(__Pyx_memviewslice *memslice) except -1 nogil: + * cdef int ndim = memslice.memview.view.ndim # <<<<<<<<<<<<<< + * + * cdef Py_ssize_t *shape = memslice.shape + */ + __pyx_t_1 = __pyx_v_memslice->memview->view.ndim; + __pyx_v_ndim = __pyx_t_1; + + /* "View.MemoryView":932 + * cdef int ndim = memslice.memview.view.ndim + * + * cdef Py_ssize_t *shape = memslice.shape # <<<<<<<<<<<<<< + * cdef Py_ssize_t *strides = memslice.strides + * + */ + __pyx_t_2 = __pyx_v_memslice->shape; + __pyx_v_shape = __pyx_t_2; + + /* "View.MemoryView":933 + * + * cdef Py_ssize_t *shape = memslice.shape + * cdef Py_ssize_t *strides = memslice.strides # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = __pyx_v_memslice->strides; + __pyx_v_strides = __pyx_t_2; + + /* "View.MemoryView":937 + * + * cdef int i, j + * for i in range(ndim // 2): # <<<<<<<<<<<<<< + * j = ndim - 1 - i + * strides[i], strides[j] = strides[j], strides[i] + */ + __pyx_t_3 = __Pyx_div_long(__pyx_v_ndim, 2); + __pyx_t_4 = __pyx_t_3; + for (__pyx_t_1 = 0; __pyx_t_1 < __pyx_t_4; __pyx_t_1+=1) { + __pyx_v_i = __pyx_t_1; + + /* "View.MemoryView":938 + * cdef int i, j + * for i in range(ndim // 2): + * j = ndim - 1 - i # <<<<<<<<<<<<<< + * strides[i], strides[j] = strides[j], strides[i] + * shape[i], shape[j] = shape[j], shape[i] + */ + __pyx_v_j = ((__pyx_v_ndim - 1) - __pyx_v_i); + + /* "View.MemoryView":939 + * for i in range(ndim // 2): + * j = ndim - 1 - i + * strides[i], strides[j] = strides[j], strides[i] # <<<<<<<<<<<<<< + * shape[i], shape[j] = shape[j], shape[i] + * + */ + __pyx_t_5 = (__pyx_v_strides[__pyx_v_j]); + __pyx_t_6 = (__pyx_v_strides[__pyx_v_i]); + (__pyx_v_strides[__pyx_v_i]) = __pyx_t_5; + (__pyx_v_strides[__pyx_v_j]) = __pyx_t_6; + + /* "View.MemoryView":940 + * j = ndim - 1 - i + * strides[i], strides[j] = strides[j], strides[i] + * shape[i], shape[j] = shape[j], shape[i] # <<<<<<<<<<<<<< + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: + */ + __pyx_t_6 = (__pyx_v_shape[__pyx_v_j]); + __pyx_t_5 = (__pyx_v_shape[__pyx_v_i]); + (__pyx_v_shape[__pyx_v_i]) = __pyx_t_6; + (__pyx_v_shape[__pyx_v_j]) = __pyx_t_5; + + /* "View.MemoryView":942 + * shape[i], shape[j] = shape[j], shape[i] + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: # <<<<<<<<<<<<<< + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") + * + */ + __pyx_t_8 = ((__pyx_v_memslice->suboffsets[__pyx_v_i]) >= 0); + if (!__pyx_t_8) { + } else { + __pyx_t_7 = __pyx_t_8; + goto __pyx_L6_bool_binop_done; + } + __pyx_t_8 = ((__pyx_v_memslice->suboffsets[__pyx_v_j]) >= 0); + __pyx_t_7 = __pyx_t_8; + __pyx_L6_bool_binop_done:; + if (__pyx_t_7) { + + /* "View.MemoryView":943 + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") # <<<<<<<<<<<<<< + * + * return 0 + */ + __pyx_t_9 = __pyx_memoryview_err(PyExc_ValueError, __pyx_kp_s_Cannot_transpose_memoryview_with); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 943, __pyx_L1_error) + + /* "View.MemoryView":942 + * shape[i], shape[j] = shape[j], shape[i] + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: # <<<<<<<<<<<<<< + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") + * + */ + } + } + + /* "View.MemoryView":945 + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") + * + * return 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":929 + * + * @cname('__pyx_memslice_transpose') + * cdef int transpose_memslice(__Pyx_memviewslice *memslice) except -1 nogil: # <<<<<<<<<<<<<< + * cdef int ndim = memslice.memview.view.ndim + * + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.transpose_memslice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":963 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * def __dealloc__(self): # <<<<<<<<<<<<<< + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + */ + +/* Python wrapper */ +static void __pyx_memoryviewslice___dealloc__(PyObject *__pyx_v_self); /*proto*/ +static void __pyx_memoryviewslice___dealloc__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__dealloc__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewslice___dealloc__(((struct __pyx_memoryviewslice_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); +} + +static void __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewslice___dealloc__(struct __pyx_memoryviewslice_obj *__pyx_v_self) { + + /* "View.MemoryView":964 + * + * def __dealloc__(self): + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) # <<<<<<<<<<<<<< + * + * cdef convert_item_to_object(self, char *itemp): + */ + __PYX_XCLEAR_MEMVIEW((&__pyx_v_self->from_slice), 1); + + /* "View.MemoryView":963 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * def __dealloc__(self): # <<<<<<<<<<<<<< + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + */ + + /* function exit code */ +} + +/* "View.MemoryView":966 + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * if self.to_object_func != NULL: + * return self.to_object_func(itemp) + */ + +static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("convert_item_to_object", 1); + + /* "View.MemoryView":967 + * + * cdef convert_item_to_object(self, char *itemp): + * if self.to_object_func != NULL: # <<<<<<<<<<<<<< + * return self.to_object_func(itemp) + * else: + */ + __pyx_t_1 = (__pyx_v_self->to_object_func != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":968 + * cdef convert_item_to_object(self, char *itemp): + * if self.to_object_func != NULL: + * return self.to_object_func(itemp) # <<<<<<<<<<<<<< + * else: + * return memoryview.convert_item_to_object(self, itemp) + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_v_self->to_object_func(__pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 968, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":967 + * + * cdef convert_item_to_object(self, char *itemp): + * if self.to_object_func != NULL: # <<<<<<<<<<<<<< + * return self.to_object_func(itemp) + * else: + */ + } + + /* "View.MemoryView":970 + * return self.to_object_func(itemp) + * else: + * return memoryview.convert_item_to_object(self, itemp) # <<<<<<<<<<<<<< + * + * cdef assign_item_from_object(self, char *itemp, object value): + */ + /*else*/ { + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_memoryview_convert_item_to_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 970, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + } + + /* "View.MemoryView":966 + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * if self.to_object_func != NULL: + * return self.to_object_func(itemp) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.convert_item_to_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":972 + * return memoryview.convert_item_to_object(self, itemp) + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * if self.to_dtype_func != NULL: + * self.to_dtype_func(itemp, value) + */ + +static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("assign_item_from_object", 1); + + /* "View.MemoryView":973 + * + * cdef assign_item_from_object(self, char *itemp, object value): + * if self.to_dtype_func != NULL: # <<<<<<<<<<<<<< + * self.to_dtype_func(itemp, value) + * else: + */ + __pyx_t_1 = (__pyx_v_self->to_dtype_func != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":974 + * cdef assign_item_from_object(self, char *itemp, object value): + * if self.to_dtype_func != NULL: + * self.to_dtype_func(itemp, value) # <<<<<<<<<<<<<< + * else: + * memoryview.assign_item_from_object(self, itemp, value) + */ + __pyx_t_2 = __pyx_v_self->to_dtype_func(__pyx_v_itemp, __pyx_v_value); if (unlikely(__pyx_t_2 == ((int)0))) __PYX_ERR(1, 974, __pyx_L1_error) + + /* "View.MemoryView":973 + * + * cdef assign_item_from_object(self, char *itemp, object value): + * if self.to_dtype_func != NULL: # <<<<<<<<<<<<<< + * self.to_dtype_func(itemp, value) + * else: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":976 + * self.to_dtype_func(itemp, value) + * else: + * memoryview.assign_item_from_object(self, itemp, value) # <<<<<<<<<<<<<< + * + * cdef _get_base(self): + */ + /*else*/ { + __pyx_t_3 = __pyx_memoryview_assign_item_from_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 976, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + __pyx_L3:; + + /* "View.MemoryView":972 + * return memoryview.convert_item_to_object(self, itemp) + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * if self.to_dtype_func != NULL: + * self.to_dtype_func(itemp, value) + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.assign_item_from_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":978 + * memoryview.assign_item_from_object(self, itemp, value) + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.from_object + * + */ + +static PyObject *__pyx_memoryviewslice__get_base(struct __pyx_memoryviewslice_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("_get_base", 1); + + /* "View.MemoryView":979 + * + * cdef _get_base(self): + * return self.from_object # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->from_object); + __pyx_r = __pyx_v_self->from_object; + goto __pyx_L0; + + /* "View.MemoryView":978 + * memoryview.assign_item_from_object(self, itemp, value) + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.from_object + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryviewslice_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryviewslice_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_memoryviewslice___reduce_cython__(((struct __pyx_memoryviewslice_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryviewslice___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":2 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 2, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryviewslice_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryviewslice_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + CYTHON_UNUSED PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 3, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 3, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 3, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_memoryviewslice_2__setstate_cython__(((struct __pyx_memoryviewslice_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryviewslice_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":4 + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 4, __pyx_L1_error) + + /* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":999 + * + * @cname('__pyx_memoryview_fromslice') + * cdef memoryview_fromslice(__Pyx_memviewslice memviewslice, # <<<<<<<<<<<<<< + * int ndim, + * object (*to_object_func)(char *), + */ + +static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewslice, int __pyx_v_ndim, PyObject *(*__pyx_v_to_object_func)(char *), int (*__pyx_v_to_dtype_func)(char *, PyObject *), int __pyx_v_dtype_is_object) { + struct __pyx_memoryviewslice_obj *__pyx_v_result = 0; + Py_ssize_t __pyx_v_suboffset; + PyObject *__pyx_v_length = NULL; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + __Pyx_TypeInfo *__pyx_t_4; + Py_buffer __pyx_t_5; + Py_ssize_t *__pyx_t_6; + Py_ssize_t *__pyx_t_7; + Py_ssize_t *__pyx_t_8; + Py_ssize_t __pyx_t_9; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_fromslice", 1); + + /* "View.MemoryView":1007 + * cdef _memoryviewslice result + * + * if memviewslice.memview == Py_None: # <<<<<<<<<<<<<< + * return None + * + */ + __pyx_t_1 = (((PyObject *)__pyx_v_memviewslice.memview) == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":1008 + * + * if memviewslice.memview == Py_None: + * return None # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + + /* "View.MemoryView":1007 + * cdef _memoryviewslice result + * + * if memviewslice.memview == Py_None: # <<<<<<<<<<<<<< + * return None + * + */ + } + + /* "View.MemoryView":1013 + * + * + * result = _memoryviewslice.__new__(_memoryviewslice, None, 0, dtype_is_object) # <<<<<<<<<<<<<< + * + * result.from_slice = memviewslice + */ + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1013, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(Py_None); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, Py_None)) __PYX_ERR(1, 1013, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_0); + __Pyx_GIVEREF(__pyx_int_0); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_0)) __PYX_ERR(1, 1013, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error); + __pyx_t_2 = 0; + __pyx_t_2 = ((PyObject *)__pyx_tp_new__memoryviewslice(((PyTypeObject *)__pyx_memoryviewslice_type), __pyx_t_3, NULL)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error) + __Pyx_GOTREF((PyObject *)__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_v_result = ((struct __pyx_memoryviewslice_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":1015 + * result = _memoryviewslice.__new__(_memoryviewslice, None, 0, dtype_is_object) + * + * result.from_slice = memviewslice # <<<<<<<<<<<<<< + * __PYX_INC_MEMVIEW(&memviewslice, 1) + * + */ + __pyx_v_result->from_slice = __pyx_v_memviewslice; + + /* "View.MemoryView":1016 + * + * result.from_slice = memviewslice + * __PYX_INC_MEMVIEW(&memviewslice, 1) # <<<<<<<<<<<<<< + * + * result.from_object = ( memviewslice.memview)._get_base() + */ + __PYX_INC_MEMVIEW((&__pyx_v_memviewslice), 1); + + /* "View.MemoryView":1018 + * __PYX_INC_MEMVIEW(&memviewslice, 1) + * + * result.from_object = ( memviewslice.memview)._get_base() # <<<<<<<<<<<<<< + * result.typeinfo = memviewslice.memview.typeinfo + * + */ + __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)((struct __pyx_memoryview_obj *)__pyx_v_memviewslice.memview)->__pyx_vtab)->_get_base(((struct __pyx_memoryview_obj *)__pyx_v_memviewslice.memview)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1018, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_2); + __Pyx_GOTREF(__pyx_v_result->from_object); + __Pyx_DECREF(__pyx_v_result->from_object); + __pyx_v_result->from_object = __pyx_t_2; + __pyx_t_2 = 0; + + /* "View.MemoryView":1019 + * + * result.from_object = ( memviewslice.memview)._get_base() + * result.typeinfo = memviewslice.memview.typeinfo # <<<<<<<<<<<<<< + * + * result.view = memviewslice.memview.view + */ + __pyx_t_4 = __pyx_v_memviewslice.memview->typeinfo; + __pyx_v_result->__pyx_base.typeinfo = __pyx_t_4; + + /* "View.MemoryView":1021 + * result.typeinfo = memviewslice.memview.typeinfo + * + * result.view = memviewslice.memview.view # <<<<<<<<<<<<<< + * result.view.buf = memviewslice.data + * result.view.ndim = ndim + */ + __pyx_t_5 = __pyx_v_memviewslice.memview->view; + __pyx_v_result->__pyx_base.view = __pyx_t_5; + + /* "View.MemoryView":1022 + * + * result.view = memviewslice.memview.view + * result.view.buf = memviewslice.data # <<<<<<<<<<<<<< + * result.view.ndim = ndim + * (<__pyx_buffer *> &result.view).obj = Py_None + */ + __pyx_v_result->__pyx_base.view.buf = ((void *)__pyx_v_memviewslice.data); + + /* "View.MemoryView":1023 + * result.view = memviewslice.memview.view + * result.view.buf = memviewslice.data + * result.view.ndim = ndim # <<<<<<<<<<<<<< + * (<__pyx_buffer *> &result.view).obj = Py_None + * Py_INCREF(Py_None) + */ + __pyx_v_result->__pyx_base.view.ndim = __pyx_v_ndim; + + /* "View.MemoryView":1024 + * result.view.buf = memviewslice.data + * result.view.ndim = ndim + * (<__pyx_buffer *> &result.view).obj = Py_None # <<<<<<<<<<<<<< + * Py_INCREF(Py_None) + * + */ + ((Py_buffer *)(&__pyx_v_result->__pyx_base.view))->obj = Py_None; + + /* "View.MemoryView":1025 + * result.view.ndim = ndim + * (<__pyx_buffer *> &result.view).obj = Py_None + * Py_INCREF(Py_None) # <<<<<<<<<<<<<< + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: + */ + Py_INCREF(Py_None); + + /* "View.MemoryView":1027 + * Py_INCREF(Py_None) + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: # <<<<<<<<<<<<<< + * result.flags = PyBUF_RECORDS + * else: + */ + __pyx_t_1 = ((((struct __pyx_memoryview_obj *)__pyx_v_memviewslice.memview)->flags & PyBUF_WRITABLE) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":1028 + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: + * result.flags = PyBUF_RECORDS # <<<<<<<<<<<<<< + * else: + * result.flags = PyBUF_RECORDS_RO + */ + __pyx_v_result->__pyx_base.flags = PyBUF_RECORDS; + + /* "View.MemoryView":1027 + * Py_INCREF(Py_None) + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: # <<<<<<<<<<<<<< + * result.flags = PyBUF_RECORDS + * else: + */ + goto __pyx_L4; + } + + /* "View.MemoryView":1030 + * result.flags = PyBUF_RECORDS + * else: + * result.flags = PyBUF_RECORDS_RO # <<<<<<<<<<<<<< + * + * result.view.shape = result.from_slice.shape + */ + /*else*/ { + __pyx_v_result->__pyx_base.flags = PyBUF_RECORDS_RO; + } + __pyx_L4:; + + /* "View.MemoryView":1032 + * result.flags = PyBUF_RECORDS_RO + * + * result.view.shape = result.from_slice.shape # <<<<<<<<<<<<<< + * result.view.strides = result.from_slice.strides + * + */ + __pyx_v_result->__pyx_base.view.shape = ((Py_ssize_t *)__pyx_v_result->from_slice.shape); + + /* "View.MemoryView":1033 + * + * result.view.shape = result.from_slice.shape + * result.view.strides = result.from_slice.strides # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_result->__pyx_base.view.strides = ((Py_ssize_t *)__pyx_v_result->from_slice.strides); + + /* "View.MemoryView":1036 + * + * + * result.view.suboffsets = NULL # <<<<<<<<<<<<<< + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: + */ + __pyx_v_result->__pyx_base.view.suboffsets = NULL; + + /* "View.MemoryView":1037 + * + * result.view.suboffsets = NULL + * for suboffset in result.from_slice.suboffsets[:ndim]: # <<<<<<<<<<<<<< + * if suboffset >= 0: + * result.view.suboffsets = result.from_slice.suboffsets + */ + __pyx_t_7 = (__pyx_v_result->from_slice.suboffsets + __pyx_v_ndim); + for (__pyx_t_8 = __pyx_v_result->from_slice.suboffsets; __pyx_t_8 < __pyx_t_7; __pyx_t_8++) { + __pyx_t_6 = __pyx_t_8; + __pyx_v_suboffset = (__pyx_t_6[0]); + + /* "View.MemoryView":1038 + * result.view.suboffsets = NULL + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * result.view.suboffsets = result.from_slice.suboffsets + * break + */ + __pyx_t_1 = (__pyx_v_suboffset >= 0); + if (__pyx_t_1) { + + /* "View.MemoryView":1039 + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: + * result.view.suboffsets = result.from_slice.suboffsets # <<<<<<<<<<<<<< + * break + * + */ + __pyx_v_result->__pyx_base.view.suboffsets = ((Py_ssize_t *)__pyx_v_result->from_slice.suboffsets); + + /* "View.MemoryView":1040 + * if suboffset >= 0: + * result.view.suboffsets = result.from_slice.suboffsets + * break # <<<<<<<<<<<<<< + * + * result.view.len = result.view.itemsize + */ + goto __pyx_L6_break; + + /* "View.MemoryView":1038 + * result.view.suboffsets = NULL + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * result.view.suboffsets = result.from_slice.suboffsets + * break + */ + } + } + __pyx_L6_break:; + + /* "View.MemoryView":1042 + * break + * + * result.view.len = result.view.itemsize # <<<<<<<<<<<<<< + * for length in result.view.shape[:ndim]: + * result.view.len *= length + */ + __pyx_t_9 = __pyx_v_result->__pyx_base.view.itemsize; + __pyx_v_result->__pyx_base.view.len = __pyx_t_9; + + /* "View.MemoryView":1043 + * + * result.view.len = result.view.itemsize + * for length in result.view.shape[:ndim]: # <<<<<<<<<<<<<< + * result.view.len *= length + * + */ + __pyx_t_7 = (__pyx_v_result->__pyx_base.view.shape + __pyx_v_ndim); + for (__pyx_t_8 = __pyx_v_result->__pyx_base.view.shape; __pyx_t_8 < __pyx_t_7; __pyx_t_8++) { + __pyx_t_6 = __pyx_t_8; + __pyx_t_2 = PyInt_FromSsize_t((__pyx_t_6[0])); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1043, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_XDECREF_SET(__pyx_v_length, __pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":1044 + * result.view.len = result.view.itemsize + * for length in result.view.shape[:ndim]: + * result.view.len *= length # <<<<<<<<<<<<<< + * + * result.to_object_func = to_object_func + */ + __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_result->__pyx_base.view.len); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1044, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyNumber_InPlaceMultiply(__pyx_t_2, __pyx_v_length); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1044, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_t_3); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 1044, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_v_result->__pyx_base.view.len = __pyx_t_9; + } + + /* "View.MemoryView":1046 + * result.view.len *= length + * + * result.to_object_func = to_object_func # <<<<<<<<<<<<<< + * result.to_dtype_func = to_dtype_func + * + */ + __pyx_v_result->to_object_func = __pyx_v_to_object_func; + + /* "View.MemoryView":1047 + * + * result.to_object_func = to_object_func + * result.to_dtype_func = to_dtype_func # <<<<<<<<<<<<<< + * + * return result + */ + __pyx_v_result->to_dtype_func = __pyx_v_to_dtype_func; + + /* "View.MemoryView":1049 + * result.to_dtype_func = to_dtype_func + * + * return result # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_get_slice_from_memoryview') + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = ((PyObject *)__pyx_v_result); + goto __pyx_L0; + + /* "View.MemoryView":999 + * + * @cname('__pyx_memoryview_fromslice') + * cdef memoryview_fromslice(__Pyx_memviewslice memviewslice, # <<<<<<<<<<<<<< + * int ndim, + * object (*to_object_func)(char *), + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview_fromslice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XDECREF(__pyx_v_length); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1052 + * + * @cname('__pyx_memoryview_get_slice_from_memoryview') + * cdef __Pyx_memviewslice *get_slice_from_memview(memoryview memview, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + */ + +static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __pyx_memoryview_obj *__pyx_v_memview, __Pyx_memviewslice *__pyx_v_mslice) { + struct __pyx_memoryviewslice_obj *__pyx_v_obj = 0; + __Pyx_memviewslice *__pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("get_slice_from_memview", 1); + + /* "View.MemoryView":1055 + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * obj = memview + * return &obj.from_slice + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":1056 + * cdef _memoryviewslice obj + * if isinstance(memview, _memoryviewslice): + * obj = memview # <<<<<<<<<<<<<< + * return &obj.from_slice + * else: + */ + if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 1056, __pyx_L1_error) + __pyx_t_2 = ((PyObject *)__pyx_v_memview); + __Pyx_INCREF(__pyx_t_2); + __pyx_v_obj = ((struct __pyx_memoryviewslice_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":1057 + * if isinstance(memview, _memoryviewslice): + * obj = memview + * return &obj.from_slice # <<<<<<<<<<<<<< + * else: + * slice_copy(memview, mslice) + */ + __pyx_r = (&__pyx_v_obj->from_slice); + goto __pyx_L0; + + /* "View.MemoryView":1055 + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * obj = memview + * return &obj.from_slice + */ + } + + /* "View.MemoryView":1059 + * return &obj.from_slice + * else: + * slice_copy(memview, mslice) # <<<<<<<<<<<<<< + * return mslice + * + */ + /*else*/ { + __pyx_memoryview_slice_copy(__pyx_v_memview, __pyx_v_mslice); + + /* "View.MemoryView":1060 + * else: + * slice_copy(memview, mslice) + * return mslice # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_slice_copy') + */ + __pyx_r = __pyx_v_mslice; + goto __pyx_L0; + } + + /* "View.MemoryView":1052 + * + * @cname('__pyx_memoryview_get_slice_from_memoryview') + * cdef __Pyx_memviewslice *get_slice_from_memview(memoryview memview, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.get_slice_from_memview", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_obj); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1063 + * + * @cname('__pyx_memoryview_slice_copy') + * cdef void slice_copy(memoryview memview, __Pyx_memviewslice *dst) noexcept: # <<<<<<<<<<<<<< + * cdef int dim + * cdef (Py_ssize_t*) shape, strides, suboffsets + */ + +static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_memview, __Pyx_memviewslice *__pyx_v_dst) { + int __pyx_v_dim; + Py_ssize_t *__pyx_v_shape; + Py_ssize_t *__pyx_v_strides; + Py_ssize_t *__pyx_v_suboffsets; + Py_ssize_t *__pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + Py_ssize_t __pyx_t_5; + int __pyx_t_6; + + /* "View.MemoryView":1067 + * cdef (Py_ssize_t*) shape, strides, suboffsets + * + * shape = memview.view.shape # <<<<<<<<<<<<<< + * strides = memview.view.strides + * suboffsets = memview.view.suboffsets + */ + __pyx_t_1 = __pyx_v_memview->view.shape; + __pyx_v_shape = __pyx_t_1; + + /* "View.MemoryView":1068 + * + * shape = memview.view.shape + * strides = memview.view.strides # <<<<<<<<<<<<<< + * suboffsets = memview.view.suboffsets + * + */ + __pyx_t_1 = __pyx_v_memview->view.strides; + __pyx_v_strides = __pyx_t_1; + + /* "View.MemoryView":1069 + * shape = memview.view.shape + * strides = memview.view.strides + * suboffsets = memview.view.suboffsets # <<<<<<<<<<<<<< + * + * dst.memview = <__pyx_memoryview *> memview + */ + __pyx_t_1 = __pyx_v_memview->view.suboffsets; + __pyx_v_suboffsets = __pyx_t_1; + + /* "View.MemoryView":1071 + * suboffsets = memview.view.suboffsets + * + * dst.memview = <__pyx_memoryview *> memview # <<<<<<<<<<<<<< + * dst.data = memview.view.buf + * + */ + __pyx_v_dst->memview = ((struct __pyx_memoryview_obj *)__pyx_v_memview); + + /* "View.MemoryView":1072 + * + * dst.memview = <__pyx_memoryview *> memview + * dst.data = memview.view.buf # <<<<<<<<<<<<<< + * + * for dim in range(memview.view.ndim): + */ + __pyx_v_dst->data = ((char *)__pyx_v_memview->view.buf); + + /* "View.MemoryView":1074 + * dst.data = memview.view.buf + * + * for dim in range(memview.view.ndim): # <<<<<<<<<<<<<< + * dst.shape[dim] = shape[dim] + * dst.strides[dim] = strides[dim] + */ + __pyx_t_2 = __pyx_v_memview->view.ndim; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_dim = __pyx_t_4; + + /* "View.MemoryView":1075 + * + * for dim in range(memview.view.ndim): + * dst.shape[dim] = shape[dim] # <<<<<<<<<<<<<< + * dst.strides[dim] = strides[dim] + * dst.suboffsets[dim] = suboffsets[dim] if suboffsets else -1 + */ + (__pyx_v_dst->shape[__pyx_v_dim]) = (__pyx_v_shape[__pyx_v_dim]); + + /* "View.MemoryView":1076 + * for dim in range(memview.view.ndim): + * dst.shape[dim] = shape[dim] + * dst.strides[dim] = strides[dim] # <<<<<<<<<<<<<< + * dst.suboffsets[dim] = suboffsets[dim] if suboffsets else -1 + * + */ + (__pyx_v_dst->strides[__pyx_v_dim]) = (__pyx_v_strides[__pyx_v_dim]); + + /* "View.MemoryView":1077 + * dst.shape[dim] = shape[dim] + * dst.strides[dim] = strides[dim] + * dst.suboffsets[dim] = suboffsets[dim] if suboffsets else -1 # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_copy_object') + */ + __pyx_t_6 = (__pyx_v_suboffsets != 0); + if (__pyx_t_6) { + __pyx_t_5 = (__pyx_v_suboffsets[__pyx_v_dim]); + } else { + __pyx_t_5 = -1L; + } + (__pyx_v_dst->suboffsets[__pyx_v_dim]) = __pyx_t_5; + } + + /* "View.MemoryView":1063 + * + * @cname('__pyx_memoryview_slice_copy') + * cdef void slice_copy(memoryview memview, __Pyx_memviewslice *dst) noexcept: # <<<<<<<<<<<<<< + * cdef int dim + * cdef (Py_ssize_t*) shape, strides, suboffsets + */ + + /* function exit code */ +} + +/* "View.MemoryView":1080 + * + * @cname('__pyx_memoryview_copy_object') + * cdef memoryview_copy(memoryview memview): # <<<<<<<<<<<<<< + * "Create a new memoryview object" + * cdef __Pyx_memviewslice memviewslice + */ + +static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *__pyx_v_memview) { + __Pyx_memviewslice __pyx_v_memviewslice; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_copy", 1); + + /* "View.MemoryView":1083 + * "Create a new memoryview object" + * cdef __Pyx_memviewslice memviewslice + * slice_copy(memview, &memviewslice) # <<<<<<<<<<<<<< + * return memoryview_copy_from_slice(memview, &memviewslice) + * + */ + __pyx_memoryview_slice_copy(__pyx_v_memview, (&__pyx_v_memviewslice)); + + /* "View.MemoryView":1084 + * cdef __Pyx_memviewslice memviewslice + * slice_copy(memview, &memviewslice) + * return memoryview_copy_from_slice(memview, &memviewslice) # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_copy_object_from_slice') + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __pyx_memoryview_copy_object_from_slice(__pyx_v_memview, (&__pyx_v_memviewslice)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1084, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":1080 + * + * @cname('__pyx_memoryview_copy_object') + * cdef memoryview_copy(memoryview memview): # <<<<<<<<<<<<<< + * "Create a new memoryview object" + * cdef __Pyx_memviewslice memviewslice + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview_copy", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1087 + * + * @cname('__pyx_memoryview_copy_object_from_slice') + * cdef memoryview_copy_from_slice(memoryview memview, __Pyx_memviewslice *memviewslice): # <<<<<<<<<<<<<< + * """ + * Create a new memoryview object from a given memoryview object and slice. + */ + +static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview_obj *__pyx_v_memview, __Pyx_memviewslice *__pyx_v_memviewslice) { + PyObject *(*__pyx_v_to_object_func)(char *); + int (*__pyx_v_to_dtype_func)(char *, PyObject *); + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *(*__pyx_t_2)(char *); + int (*__pyx_t_3)(char *, PyObject *); + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_copy_from_slice", 1); + + /* "View.MemoryView":1094 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * to_object_func = (<_memoryviewslice> memview).to_object_func + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":1095 + * + * if isinstance(memview, _memoryviewslice): + * to_object_func = (<_memoryviewslice> memview).to_object_func # <<<<<<<<<<<<<< + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + * else: + */ + __pyx_t_2 = ((struct __pyx_memoryviewslice_obj *)__pyx_v_memview)->to_object_func; + __pyx_v_to_object_func = __pyx_t_2; + + /* "View.MemoryView":1096 + * if isinstance(memview, _memoryviewslice): + * to_object_func = (<_memoryviewslice> memview).to_object_func + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func # <<<<<<<<<<<<<< + * else: + * to_object_func = NULL + */ + __pyx_t_3 = ((struct __pyx_memoryviewslice_obj *)__pyx_v_memview)->to_dtype_func; + __pyx_v_to_dtype_func = __pyx_t_3; + + /* "View.MemoryView":1094 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * to_object_func = (<_memoryviewslice> memview).to_object_func + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1098 + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + * else: + * to_object_func = NULL # <<<<<<<<<<<<<< + * to_dtype_func = NULL + * + */ + /*else*/ { + __pyx_v_to_object_func = NULL; + + /* "View.MemoryView":1099 + * else: + * to_object_func = NULL + * to_dtype_func = NULL # <<<<<<<<<<<<<< + * + * return memoryview_fromslice(memviewslice[0], memview.view.ndim, + */ + __pyx_v_to_dtype_func = NULL; + } + __pyx_L3:; + + /* "View.MemoryView":1101 + * to_dtype_func = NULL + * + * return memoryview_fromslice(memviewslice[0], memview.view.ndim, # <<<<<<<<<<<<<< + * to_object_func, to_dtype_func, + * memview.dtype_is_object) + */ + __Pyx_XDECREF(__pyx_r); + + /* "View.MemoryView":1103 + * return memoryview_fromslice(memviewslice[0], memview.view.ndim, + * to_object_func, to_dtype_func, + * memview.dtype_is_object) # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_4 = __pyx_memoryview_fromslice((__pyx_v_memviewslice[0]), __pyx_v_memview->view.ndim, __pyx_v_to_object_func, __pyx_v_to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1101, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_r = __pyx_t_4; + __pyx_t_4 = 0; + goto __pyx_L0; + + /* "View.MemoryView":1087 + * + * @cname('__pyx_memoryview_copy_object_from_slice') + * cdef memoryview_copy_from_slice(memoryview memview, __Pyx_memviewslice *memviewslice): # <<<<<<<<<<<<<< + * """ + * Create a new memoryview object from a given memoryview object and slice. + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.memoryview_copy_from_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1109 + * + * + * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) noexcept nogil: # <<<<<<<<<<<<<< + * return -arg if arg < 0 else arg + * + */ + +static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) { + Py_ssize_t __pyx_r; + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + + /* "View.MemoryView":1110 + * + * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) noexcept nogil: + * return -arg if arg < 0 else arg # <<<<<<<<<<<<<< + * + * @cname('__pyx_get_best_slice_order') + */ + __pyx_t_2 = (__pyx_v_arg < 0); + if (__pyx_t_2) { + __pyx_t_1 = (-__pyx_v_arg); + } else { + __pyx_t_1 = __pyx_v_arg; + } + __pyx_r = __pyx_t_1; + goto __pyx_L0; + + /* "View.MemoryView":1109 + * + * + * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) noexcept nogil: # <<<<<<<<<<<<<< + * return -arg if arg < 0 else arg + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1113 + * + * @cname('__pyx_get_best_slice_order') + * cdef char get_best_order(__Pyx_memviewslice *mslice, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * Figure out the best memory access order for a given slice. + */ + +static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int __pyx_v_ndim) { + int __pyx_v_i; + Py_ssize_t __pyx_v_c_stride; + Py_ssize_t __pyx_v_f_stride; + char __pyx_r; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + + /* "View.MemoryView":1118 + * """ + * cdef int i + * cdef Py_ssize_t c_stride = 0 # <<<<<<<<<<<<<< + * cdef Py_ssize_t f_stride = 0 + * + */ + __pyx_v_c_stride = 0; + + /* "View.MemoryView":1119 + * cdef int i + * cdef Py_ssize_t c_stride = 0 + * cdef Py_ssize_t f_stride = 0 # <<<<<<<<<<<<<< + * + * for i in range(ndim - 1, -1, -1): + */ + __pyx_v_f_stride = 0; + + /* "View.MemoryView":1121 + * cdef Py_ssize_t f_stride = 0 + * + * for i in range(ndim - 1, -1, -1): # <<<<<<<<<<<<<< + * if mslice.shape[i] > 1: + * c_stride = mslice.strides[i] + */ + for (__pyx_t_1 = (__pyx_v_ndim - 1); __pyx_t_1 > -1; __pyx_t_1-=1) { + __pyx_v_i = __pyx_t_1; + + /* "View.MemoryView":1122 + * + * for i in range(ndim - 1, -1, -1): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * c_stride = mslice.strides[i] + * break + */ + __pyx_t_2 = ((__pyx_v_mslice->shape[__pyx_v_i]) > 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1123 + * for i in range(ndim - 1, -1, -1): + * if mslice.shape[i] > 1: + * c_stride = mslice.strides[i] # <<<<<<<<<<<<<< + * break + * + */ + __pyx_v_c_stride = (__pyx_v_mslice->strides[__pyx_v_i]); + + /* "View.MemoryView":1124 + * if mslice.shape[i] > 1: + * c_stride = mslice.strides[i] + * break # <<<<<<<<<<<<<< + * + * for i in range(ndim): + */ + goto __pyx_L4_break; + + /* "View.MemoryView":1122 + * + * for i in range(ndim - 1, -1, -1): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * c_stride = mslice.strides[i] + * break + */ + } + } + __pyx_L4_break:; + + /* "View.MemoryView":1126 + * break + * + * for i in range(ndim): # <<<<<<<<<<<<<< + * if mslice.shape[i] > 1: + * f_stride = mslice.strides[i] + */ + __pyx_t_1 = __pyx_v_ndim; + __pyx_t_3 = __pyx_t_1; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1127 + * + * for i in range(ndim): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * f_stride = mslice.strides[i] + * break + */ + __pyx_t_2 = ((__pyx_v_mslice->shape[__pyx_v_i]) > 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1128 + * for i in range(ndim): + * if mslice.shape[i] > 1: + * f_stride = mslice.strides[i] # <<<<<<<<<<<<<< + * break + * + */ + __pyx_v_f_stride = (__pyx_v_mslice->strides[__pyx_v_i]); + + /* "View.MemoryView":1129 + * if mslice.shape[i] > 1: + * f_stride = mslice.strides[i] + * break # <<<<<<<<<<<<<< + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): + */ + goto __pyx_L7_break; + + /* "View.MemoryView":1127 + * + * for i in range(ndim): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * f_stride = mslice.strides[i] + * break + */ + } + } + __pyx_L7_break:; + + /* "View.MemoryView":1131 + * break + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): # <<<<<<<<<<<<<< + * return 'C' + * else: + */ + __pyx_t_2 = (abs_py_ssize_t(__pyx_v_c_stride) <= abs_py_ssize_t(__pyx_v_f_stride)); + if (__pyx_t_2) { + + /* "View.MemoryView":1132 + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): + * return 'C' # <<<<<<<<<<<<<< + * else: + * return 'F' + */ + __pyx_r = 'C'; + goto __pyx_L0; + + /* "View.MemoryView":1131 + * break + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): # <<<<<<<<<<<<<< + * return 'C' + * else: + */ + } + + /* "View.MemoryView":1134 + * return 'C' + * else: + * return 'F' # <<<<<<<<<<<<<< + * + * @cython.cdivision(True) + */ + /*else*/ { + __pyx_r = 'F'; + goto __pyx_L0; + } + + /* "View.MemoryView":1113 + * + * @cname('__pyx_get_best_slice_order') + * cdef char get_best_order(__Pyx_memviewslice *mslice, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * Figure out the best memory access order for a given slice. + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1137 + * + * @cython.cdivision(True) + * cdef void _copy_strided_to_strided(char *src_data, Py_ssize_t *src_strides, # <<<<<<<<<<<<<< + * char *dst_data, Py_ssize_t *dst_strides, + * Py_ssize_t *src_shape, Py_ssize_t *dst_shape, + */ + +static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v_src_strides, char *__pyx_v_dst_data, Py_ssize_t *__pyx_v_dst_strides, Py_ssize_t *__pyx_v_src_shape, Py_ssize_t *__pyx_v_dst_shape, int __pyx_v_ndim, size_t __pyx_v_itemsize) { + CYTHON_UNUSED Py_ssize_t __pyx_v_i; + CYTHON_UNUSED Py_ssize_t __pyx_v_src_extent; + Py_ssize_t __pyx_v_dst_extent; + Py_ssize_t __pyx_v_src_stride; + Py_ssize_t __pyx_v_dst_stride; + int __pyx_t_1; + int __pyx_t_2; + Py_ssize_t __pyx_t_3; + Py_ssize_t __pyx_t_4; + Py_ssize_t __pyx_t_5; + + /* "View.MemoryView":1144 + * + * cdef Py_ssize_t i + * cdef Py_ssize_t src_extent = src_shape[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t dst_extent = dst_shape[0] + * cdef Py_ssize_t src_stride = src_strides[0] + */ + __pyx_v_src_extent = (__pyx_v_src_shape[0]); + + /* "View.MemoryView":1145 + * cdef Py_ssize_t i + * cdef Py_ssize_t src_extent = src_shape[0] + * cdef Py_ssize_t dst_extent = dst_shape[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t src_stride = src_strides[0] + * cdef Py_ssize_t dst_stride = dst_strides[0] + */ + __pyx_v_dst_extent = (__pyx_v_dst_shape[0]); + + /* "View.MemoryView":1146 + * cdef Py_ssize_t src_extent = src_shape[0] + * cdef Py_ssize_t dst_extent = dst_shape[0] + * cdef Py_ssize_t src_stride = src_strides[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t dst_stride = dst_strides[0] + * + */ + __pyx_v_src_stride = (__pyx_v_src_strides[0]); + + /* "View.MemoryView":1147 + * cdef Py_ssize_t dst_extent = dst_shape[0] + * cdef Py_ssize_t src_stride = src_strides[0] + * cdef Py_ssize_t dst_stride = dst_strides[0] # <<<<<<<<<<<<<< + * + * if ndim == 1: + */ + __pyx_v_dst_stride = (__pyx_v_dst_strides[0]); + + /* "View.MemoryView":1149 + * cdef Py_ssize_t dst_stride = dst_strides[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): + */ + __pyx_t_1 = (__pyx_v_ndim == 1); + if (__pyx_t_1) { + + /* "View.MemoryView":1150 + * + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and # <<<<<<<<<<<<<< + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) + */ + __pyx_t_2 = (__pyx_v_src_stride > 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L5_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_dst_stride > 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L5_bool_binop_done; + } + + /* "View.MemoryView":1151 + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): # <<<<<<<<<<<<<< + * memcpy(dst_data, src_data, itemsize * dst_extent) + * else: + */ + __pyx_t_2 = (((size_t)__pyx_v_src_stride) == __pyx_v_itemsize); + if (__pyx_t_2) { + __pyx_t_2 = (__pyx_v_itemsize == ((size_t)__pyx_v_dst_stride)); + } + __pyx_t_1 = __pyx_t_2; + __pyx_L5_bool_binop_done:; + + /* "View.MemoryView":1150 + * + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and # <<<<<<<<<<<<<< + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) + */ + if (__pyx_t_1) { + + /* "View.MemoryView":1152 + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) # <<<<<<<<<<<<<< + * else: + * for i in range(dst_extent): + */ + (void)(memcpy(__pyx_v_dst_data, __pyx_v_src_data, (__pyx_v_itemsize * __pyx_v_dst_extent))); + + /* "View.MemoryView":1150 + * + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and # <<<<<<<<<<<<<< + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) + */ + goto __pyx_L4; + } + + /* "View.MemoryView":1154 + * memcpy(dst_data, src_data, itemsize * dst_extent) + * else: + * for i in range(dst_extent): # <<<<<<<<<<<<<< + * memcpy(dst_data, src_data, itemsize) + * src_data += src_stride + */ + /*else*/ { + __pyx_t_3 = __pyx_v_dst_extent; + __pyx_t_4 = __pyx_t_3; + for (__pyx_t_5 = 0; __pyx_t_5 < __pyx_t_4; __pyx_t_5+=1) { + __pyx_v_i = __pyx_t_5; + + /* "View.MemoryView":1155 + * else: + * for i in range(dst_extent): + * memcpy(dst_data, src_data, itemsize) # <<<<<<<<<<<<<< + * src_data += src_stride + * dst_data += dst_stride + */ + (void)(memcpy(__pyx_v_dst_data, __pyx_v_src_data, __pyx_v_itemsize)); + + /* "View.MemoryView":1156 + * for i in range(dst_extent): + * memcpy(dst_data, src_data, itemsize) + * src_data += src_stride # <<<<<<<<<<<<<< + * dst_data += dst_stride + * else: + */ + __pyx_v_src_data = (__pyx_v_src_data + __pyx_v_src_stride); + + /* "View.MemoryView":1157 + * memcpy(dst_data, src_data, itemsize) + * src_data += src_stride + * dst_data += dst_stride # <<<<<<<<<<<<<< + * else: + * for i in range(dst_extent): + */ + __pyx_v_dst_data = (__pyx_v_dst_data + __pyx_v_dst_stride); + } + } + __pyx_L4:; + + /* "View.MemoryView":1149 + * cdef Py_ssize_t dst_stride = dst_strides[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1159 + * dst_data += dst_stride + * else: + * for i in range(dst_extent): # <<<<<<<<<<<<<< + * _copy_strided_to_strided(src_data, src_strides + 1, + * dst_data, dst_strides + 1, + */ + /*else*/ { + __pyx_t_3 = __pyx_v_dst_extent; + __pyx_t_4 = __pyx_t_3; + for (__pyx_t_5 = 0; __pyx_t_5 < __pyx_t_4; __pyx_t_5+=1) { + __pyx_v_i = __pyx_t_5; + + /* "View.MemoryView":1160 + * else: + * for i in range(dst_extent): + * _copy_strided_to_strided(src_data, src_strides + 1, # <<<<<<<<<<<<<< + * dst_data, dst_strides + 1, + * src_shape + 1, dst_shape + 1, + */ + _copy_strided_to_strided(__pyx_v_src_data, (__pyx_v_src_strides + 1), __pyx_v_dst_data, (__pyx_v_dst_strides + 1), (__pyx_v_src_shape + 1), (__pyx_v_dst_shape + 1), (__pyx_v_ndim - 1), __pyx_v_itemsize); + + /* "View.MemoryView":1164 + * src_shape + 1, dst_shape + 1, + * ndim - 1, itemsize) + * src_data += src_stride # <<<<<<<<<<<<<< + * dst_data += dst_stride + * + */ + __pyx_v_src_data = (__pyx_v_src_data + __pyx_v_src_stride); + + /* "View.MemoryView":1165 + * ndim - 1, itemsize) + * src_data += src_stride + * dst_data += dst_stride # <<<<<<<<<<<<<< + * + * cdef void copy_strided_to_strided(__Pyx_memviewslice *src, + */ + __pyx_v_dst_data = (__pyx_v_dst_data + __pyx_v_dst_stride); + } + } + __pyx_L3:; + + /* "View.MemoryView":1137 + * + * @cython.cdivision(True) + * cdef void _copy_strided_to_strided(char *src_data, Py_ssize_t *src_strides, # <<<<<<<<<<<<<< + * char *dst_data, Py_ssize_t *dst_strides, + * Py_ssize_t *src_shape, Py_ssize_t *dst_shape, + */ + + /* function exit code */ +} + +/* "View.MemoryView":1167 + * dst_data += dst_stride + * + * cdef void copy_strided_to_strided(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * int ndim, size_t itemsize) noexcept nogil: + */ + +static void copy_strided_to_strided(__Pyx_memviewslice *__pyx_v_src, __Pyx_memviewslice *__pyx_v_dst, int __pyx_v_ndim, size_t __pyx_v_itemsize) { + + /* "View.MemoryView":1170 + * __Pyx_memviewslice *dst, + * int ndim, size_t itemsize) noexcept nogil: + * _copy_strided_to_strided(src.data, src.strides, dst.data, dst.strides, # <<<<<<<<<<<<<< + * src.shape, dst.shape, ndim, itemsize) + * + */ + _copy_strided_to_strided(__pyx_v_src->data, __pyx_v_src->strides, __pyx_v_dst->data, __pyx_v_dst->strides, __pyx_v_src->shape, __pyx_v_dst->shape, __pyx_v_ndim, __pyx_v_itemsize); + + /* "View.MemoryView":1167 + * dst_data += dst_stride + * + * cdef void copy_strided_to_strided(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * int ndim, size_t itemsize) noexcept nogil: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1174 + * + * @cname('__pyx_memoryview_slice_get_size') + * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * "Return the size of the memory occupied by the slice in number of bytes" + * cdef Py_ssize_t shape, size = src.memview.view.itemsize + */ + +static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_src, int __pyx_v_ndim) { + Py_ssize_t __pyx_v_shape; + Py_ssize_t __pyx_v_size; + Py_ssize_t __pyx_r; + Py_ssize_t __pyx_t_1; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + + /* "View.MemoryView":1176 + * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) noexcept nogil: + * "Return the size of the memory occupied by the slice in number of bytes" + * cdef Py_ssize_t shape, size = src.memview.view.itemsize # <<<<<<<<<<<<<< + * + * for shape in src.shape[:ndim]: + */ + __pyx_t_1 = __pyx_v_src->memview->view.itemsize; + __pyx_v_size = __pyx_t_1; + + /* "View.MemoryView":1178 + * cdef Py_ssize_t shape, size = src.memview.view.itemsize + * + * for shape in src.shape[:ndim]: # <<<<<<<<<<<<<< + * size *= shape + * + */ + __pyx_t_3 = (__pyx_v_src->shape + __pyx_v_ndim); + for (__pyx_t_4 = __pyx_v_src->shape; __pyx_t_4 < __pyx_t_3; __pyx_t_4++) { + __pyx_t_2 = __pyx_t_4; + __pyx_v_shape = (__pyx_t_2[0]); + + /* "View.MemoryView":1179 + * + * for shape in src.shape[:ndim]: + * size *= shape # <<<<<<<<<<<<<< + * + * return size + */ + __pyx_v_size = (__pyx_v_size * __pyx_v_shape); + } + + /* "View.MemoryView":1181 + * size *= shape + * + * return size # <<<<<<<<<<<<<< + * + * @cname('__pyx_fill_contig_strides_array') + */ + __pyx_r = __pyx_v_size; + goto __pyx_L0; + + /* "View.MemoryView":1174 + * + * @cname('__pyx_memoryview_slice_get_size') + * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * "Return the size of the memory occupied by the slice in number of bytes" + * cdef Py_ssize_t shape, size = src.memview.view.itemsize + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1184 + * + * @cname('__pyx_fill_contig_strides_array') + * cdef Py_ssize_t fill_contig_strides_array( # <<<<<<<<<<<<<< + * Py_ssize_t *shape, Py_ssize_t *strides, Py_ssize_t stride, + * int ndim, char order) noexcept nogil: + */ + +static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, Py_ssize_t __pyx_v_stride, int __pyx_v_ndim, char __pyx_v_order) { + int __pyx_v_idx; + Py_ssize_t __pyx_r; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + + /* "View.MemoryView":1193 + * cdef int idx + * + * if order == 'F': # <<<<<<<<<<<<<< + * for idx in range(ndim): + * strides[idx] = stride + */ + __pyx_t_1 = (__pyx_v_order == 'F'); + if (__pyx_t_1) { + + /* "View.MemoryView":1194 + * + * if order == 'F': + * for idx in range(ndim): # <<<<<<<<<<<<<< + * strides[idx] = stride + * stride *= shape[idx] + */ + __pyx_t_2 = __pyx_v_ndim; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_idx = __pyx_t_4; + + /* "View.MemoryView":1195 + * if order == 'F': + * for idx in range(ndim): + * strides[idx] = stride # <<<<<<<<<<<<<< + * stride *= shape[idx] + * else: + */ + (__pyx_v_strides[__pyx_v_idx]) = __pyx_v_stride; + + /* "View.MemoryView":1196 + * for idx in range(ndim): + * strides[idx] = stride + * stride *= shape[idx] # <<<<<<<<<<<<<< + * else: + * for idx in range(ndim - 1, -1, -1): + */ + __pyx_v_stride = (__pyx_v_stride * (__pyx_v_shape[__pyx_v_idx])); + } + + /* "View.MemoryView":1193 + * cdef int idx + * + * if order == 'F': # <<<<<<<<<<<<<< + * for idx in range(ndim): + * strides[idx] = stride + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1198 + * stride *= shape[idx] + * else: + * for idx in range(ndim - 1, -1, -1): # <<<<<<<<<<<<<< + * strides[idx] = stride + * stride *= shape[idx] + */ + /*else*/ { + for (__pyx_t_2 = (__pyx_v_ndim - 1); __pyx_t_2 > -1; __pyx_t_2-=1) { + __pyx_v_idx = __pyx_t_2; + + /* "View.MemoryView":1199 + * else: + * for idx in range(ndim - 1, -1, -1): + * strides[idx] = stride # <<<<<<<<<<<<<< + * stride *= shape[idx] + * + */ + (__pyx_v_strides[__pyx_v_idx]) = __pyx_v_stride; + + /* "View.MemoryView":1200 + * for idx in range(ndim - 1, -1, -1): + * strides[idx] = stride + * stride *= shape[idx] # <<<<<<<<<<<<<< + * + * return stride + */ + __pyx_v_stride = (__pyx_v_stride * (__pyx_v_shape[__pyx_v_idx])); + } + } + __pyx_L3:; + + /* "View.MemoryView":1202 + * stride *= shape[idx] + * + * return stride # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_copy_data_to_temp') + */ + __pyx_r = __pyx_v_stride; + goto __pyx_L0; + + /* "View.MemoryView":1184 + * + * @cname('__pyx_fill_contig_strides_array') + * cdef Py_ssize_t fill_contig_strides_array( # <<<<<<<<<<<<<< + * Py_ssize_t *shape, Py_ssize_t *strides, Py_ssize_t stride, + * int ndim, char order) noexcept nogil: + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1205 + * + * @cname('__pyx_memoryview_copy_data_to_temp') + * cdef void *copy_data_to_temp(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *tmpslice, + * char order, + */ + +static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src, __Pyx_memviewslice *__pyx_v_tmpslice, char __pyx_v_order, int __pyx_v_ndim) { + int __pyx_v_i; + void *__pyx_v_result; + size_t __pyx_v_itemsize; + size_t __pyx_v_size; + void *__pyx_r; + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + struct __pyx_memoryview_obj *__pyx_t_4; + int __pyx_t_5; + int __pyx_t_6; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":1216 + * cdef void *result + * + * cdef size_t itemsize = src.memview.view.itemsize # <<<<<<<<<<<<<< + * cdef size_t size = slice_get_size(src, ndim) + * + */ + __pyx_t_1 = __pyx_v_src->memview->view.itemsize; + __pyx_v_itemsize = __pyx_t_1; + + /* "View.MemoryView":1217 + * + * cdef size_t itemsize = src.memview.view.itemsize + * cdef size_t size = slice_get_size(src, ndim) # <<<<<<<<<<<<<< + * + * result = malloc(size) + */ + __pyx_v_size = __pyx_memoryview_slice_get_size(__pyx_v_src, __pyx_v_ndim); + + /* "View.MemoryView":1219 + * cdef size_t size = slice_get_size(src, ndim) + * + * result = malloc(size) # <<<<<<<<<<<<<< + * if not result: + * _err_no_memory() + */ + __pyx_v_result = malloc(__pyx_v_size); + + /* "View.MemoryView":1220 + * + * result = malloc(size) + * if not result: # <<<<<<<<<<<<<< + * _err_no_memory() + * + */ + __pyx_t_2 = (!(__pyx_v_result != 0)); + if (__pyx_t_2) { + + /* "View.MemoryView":1221 + * result = malloc(size) + * if not result: + * _err_no_memory() # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_3 = __pyx_memoryview_err_no_memory(); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 1221, __pyx_L1_error) + + /* "View.MemoryView":1220 + * + * result = malloc(size) + * if not result: # <<<<<<<<<<<<<< + * _err_no_memory() + * + */ + } + + /* "View.MemoryView":1224 + * + * + * tmpslice.data = result # <<<<<<<<<<<<<< + * tmpslice.memview = src.memview + * for i in range(ndim): + */ + __pyx_v_tmpslice->data = ((char *)__pyx_v_result); + + /* "View.MemoryView":1225 + * + * tmpslice.data = result + * tmpslice.memview = src.memview # <<<<<<<<<<<<<< + * for i in range(ndim): + * tmpslice.shape[i] = src.shape[i] + */ + __pyx_t_4 = __pyx_v_src->memview; + __pyx_v_tmpslice->memview = __pyx_t_4; + + /* "View.MemoryView":1226 + * tmpslice.data = result + * tmpslice.memview = src.memview + * for i in range(ndim): # <<<<<<<<<<<<<< + * tmpslice.shape[i] = src.shape[i] + * tmpslice.suboffsets[i] = -1 + */ + __pyx_t_3 = __pyx_v_ndim; + __pyx_t_5 = __pyx_t_3; + for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) { + __pyx_v_i = __pyx_t_6; + + /* "View.MemoryView":1227 + * tmpslice.memview = src.memview + * for i in range(ndim): + * tmpslice.shape[i] = src.shape[i] # <<<<<<<<<<<<<< + * tmpslice.suboffsets[i] = -1 + * + */ + (__pyx_v_tmpslice->shape[__pyx_v_i]) = (__pyx_v_src->shape[__pyx_v_i]); + + /* "View.MemoryView":1228 + * for i in range(ndim): + * tmpslice.shape[i] = src.shape[i] + * tmpslice.suboffsets[i] = -1 # <<<<<<<<<<<<<< + * + * fill_contig_strides_array(&tmpslice.shape[0], &tmpslice.strides[0], itemsize, ndim, order) + */ + (__pyx_v_tmpslice->suboffsets[__pyx_v_i]) = -1L; + } + + /* "View.MemoryView":1230 + * tmpslice.suboffsets[i] = -1 + * + * fill_contig_strides_array(&tmpslice.shape[0], &tmpslice.strides[0], itemsize, ndim, order) # <<<<<<<<<<<<<< + * + * + */ + (void)(__pyx_fill_contig_strides_array((&(__pyx_v_tmpslice->shape[0])), (&(__pyx_v_tmpslice->strides[0])), __pyx_v_itemsize, __pyx_v_ndim, __pyx_v_order)); + + /* "View.MemoryView":1233 + * + * + * for i in range(ndim): # <<<<<<<<<<<<<< + * if tmpslice.shape[i] == 1: + * tmpslice.strides[i] = 0 + */ + __pyx_t_3 = __pyx_v_ndim; + __pyx_t_5 = __pyx_t_3; + for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) { + __pyx_v_i = __pyx_t_6; + + /* "View.MemoryView":1234 + * + * for i in range(ndim): + * if tmpslice.shape[i] == 1: # <<<<<<<<<<<<<< + * tmpslice.strides[i] = 0 + * + */ + __pyx_t_2 = ((__pyx_v_tmpslice->shape[__pyx_v_i]) == 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1235 + * for i in range(ndim): + * if tmpslice.shape[i] == 1: + * tmpslice.strides[i] = 0 # <<<<<<<<<<<<<< + * + * if slice_is_contig(src[0], order, ndim): + */ + (__pyx_v_tmpslice->strides[__pyx_v_i]) = 0; + + /* "View.MemoryView":1234 + * + * for i in range(ndim): + * if tmpslice.shape[i] == 1: # <<<<<<<<<<<<<< + * tmpslice.strides[i] = 0 + * + */ + } + } + + /* "View.MemoryView":1237 + * tmpslice.strides[i] = 0 + * + * if slice_is_contig(src[0], order, ndim): # <<<<<<<<<<<<<< + * memcpy(result, src.data, size) + * else: + */ + __pyx_t_2 = __pyx_memviewslice_is_contig((__pyx_v_src[0]), __pyx_v_order, __pyx_v_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1238 + * + * if slice_is_contig(src[0], order, ndim): + * memcpy(result, src.data, size) # <<<<<<<<<<<<<< + * else: + * copy_strided_to_strided(src, tmpslice, ndim, itemsize) + */ + (void)(memcpy(__pyx_v_result, __pyx_v_src->data, __pyx_v_size)); + + /* "View.MemoryView":1237 + * tmpslice.strides[i] = 0 + * + * if slice_is_contig(src[0], order, ndim): # <<<<<<<<<<<<<< + * memcpy(result, src.data, size) + * else: + */ + goto __pyx_L9; + } + + /* "View.MemoryView":1240 + * memcpy(result, src.data, size) + * else: + * copy_strided_to_strided(src, tmpslice, ndim, itemsize) # <<<<<<<<<<<<<< + * + * return result + */ + /*else*/ { + copy_strided_to_strided(__pyx_v_src, __pyx_v_tmpslice, __pyx_v_ndim, __pyx_v_itemsize); + } + __pyx_L9:; + + /* "View.MemoryView":1242 + * copy_strided_to_strided(src, tmpslice, ndim, itemsize) + * + * return result # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = __pyx_v_result; + goto __pyx_L0; + + /* "View.MemoryView":1205 + * + * @cname('__pyx_memoryview_copy_data_to_temp') + * cdef void *copy_data_to_temp(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *tmpslice, + * char order, + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.copy_data_to_temp", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1247 + * + * @cname('__pyx_memoryview_err_extents') + * cdef int _err_extents(int i, Py_ssize_t extent1, # <<<<<<<<<<<<<< + * Py_ssize_t extent2) except -1 with gil: + * raise ValueError, f"got differing extents in dimension {i} (got {extent1} and {extent2})" + */ + +static int __pyx_memoryview_err_extents(int __pyx_v_i, Py_ssize_t __pyx_v_extent1, Py_ssize_t __pyx_v_extent2) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + Py_ssize_t __pyx_t_2; + Py_UCS4 __pyx_t_3; + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_RefNannySetupContext("_err_extents", 0); + + /* "View.MemoryView":1249 + * cdef int _err_extents(int i, Py_ssize_t extent1, + * Py_ssize_t extent2) except -1 with gil: + * raise ValueError, f"got differing extents in dimension {i} (got {extent1} and {extent2})" # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_err_dim') + */ + __pyx_t_1 = PyTuple_New(7); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = 0; + __pyx_t_3 = 127; + __Pyx_INCREF(__pyx_kp_u_got_differing_extents_in_dimensi); + __pyx_t_2 += 35; + __Pyx_GIVEREF(__pyx_kp_u_got_differing_extents_in_dimensi); + PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_kp_u_got_differing_extents_in_dimensi); + __pyx_t_4 = __Pyx_PyUnicode_From_int(__pyx_v_i, 0, ' ', 'd'); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_2 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_t_4); + __pyx_t_4 = 0; + __Pyx_INCREF(__pyx_kp_u_got); + __pyx_t_2 += 6; + __Pyx_GIVEREF(__pyx_kp_u_got); + PyTuple_SET_ITEM(__pyx_t_1, 2, __pyx_kp_u_got); + __pyx_t_4 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_extent1, 0, ' ', 'd'); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_2 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + PyTuple_SET_ITEM(__pyx_t_1, 3, __pyx_t_4); + __pyx_t_4 = 0; + __Pyx_INCREF(__pyx_kp_u_and); + __pyx_t_2 += 5; + __Pyx_GIVEREF(__pyx_kp_u_and); + PyTuple_SET_ITEM(__pyx_t_1, 4, __pyx_kp_u_and); + __pyx_t_4 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_extent2, 0, ' ', 'd'); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_2 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + PyTuple_SET_ITEM(__pyx_t_1, 5, __pyx_t_4); + __pyx_t_4 = 0; + __Pyx_INCREF(__pyx_kp_u__7); + __pyx_t_2 += 1; + __Pyx_GIVEREF(__pyx_kp_u__7); + PyTuple_SET_ITEM(__pyx_t_1, 6, __pyx_kp_u__7); + __pyx_t_4 = __Pyx_PyUnicode_Join(__pyx_t_1, 7, __pyx_t_2, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_t_4, 0, 0); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __PYX_ERR(1, 1249, __pyx_L1_error) + + /* "View.MemoryView":1247 + * + * @cname('__pyx_memoryview_err_extents') + * cdef int _err_extents(int i, Py_ssize_t extent1, # <<<<<<<<<<<<<< + * Py_ssize_t extent2) except -1 with gil: + * raise ValueError, f"got differing extents in dimension {i} (got {extent1} and {extent2})" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView._err_extents", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __Pyx_RefNannyFinishContext(); + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1252 + * + * @cname('__pyx_memoryview_err_dim') + * cdef int _err_dim(PyObject *error, str msg, int dim) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg % dim + * + */ + +static int __pyx_memoryview_err_dim(PyObject *__pyx_v_error, PyObject *__pyx_v_msg, int __pyx_v_dim) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_RefNannySetupContext("_err_dim", 0); + __Pyx_INCREF(__pyx_v_msg); + + /* "View.MemoryView":1253 + * @cname('__pyx_memoryview_err_dim') + * cdef int _err_dim(PyObject *error, str msg, int dim) except -1 with gil: + * raise error, msg % dim # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_err') + */ + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_dim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1253, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyString_FormatSafe(__pyx_v_msg, __pyx_t_1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1253, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_Raise(((PyObject *)__pyx_v_error), __pyx_t_2, 0, 0); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __PYX_ERR(1, 1253, __pyx_L1_error) + + /* "View.MemoryView":1252 + * + * @cname('__pyx_memoryview_err_dim') + * cdef int _err_dim(PyObject *error, str msg, int dim) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg % dim + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView._err_dim", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __Pyx_XDECREF(__pyx_v_msg); + __Pyx_RefNannyFinishContext(); + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1256 + * + * @cname('__pyx_memoryview_err') + * cdef int _err(PyObject *error, str msg) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg + * + */ + +static int __pyx_memoryview_err(PyObject *__pyx_v_error, PyObject *__pyx_v_msg) { + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_RefNannySetupContext("_err", 0); + __Pyx_INCREF(__pyx_v_msg); + + /* "View.MemoryView":1257 + * @cname('__pyx_memoryview_err') + * cdef int _err(PyObject *error, str msg) except -1 with gil: + * raise error, msg # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_err_no_memory') + */ + __Pyx_Raise(((PyObject *)__pyx_v_error), __pyx_v_msg, 0, 0); + __PYX_ERR(1, 1257, __pyx_L1_error) + + /* "View.MemoryView":1256 + * + * @cname('__pyx_memoryview_err') + * cdef int _err(PyObject *error, str msg) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._err", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __Pyx_XDECREF(__pyx_v_msg); + __Pyx_RefNannyFinishContext(); + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1260 + * + * @cname('__pyx_memoryview_err_no_memory') + * cdef int _err_no_memory() except -1 with gil: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + +static int __pyx_memoryview_err_no_memory(void) { + int __pyx_r; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + + /* "View.MemoryView":1261 + * @cname('__pyx_memoryview_err_no_memory') + * cdef int _err_no_memory() except -1 with gil: + * raise MemoryError # <<<<<<<<<<<<<< + * + * + */ + PyErr_NoMemory(); __PYX_ERR(1, 1261, __pyx_L1_error) + + /* "View.MemoryView":1260 + * + * @cname('__pyx_memoryview_err_no_memory') + * cdef int _err_no_memory() except -1 with gil: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._err_no_memory", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1265 + * + * @cname('__pyx_memoryview_copy_contents') + * cdef int memoryview_copy_contents(__Pyx_memviewslice src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice dst, + * int src_ndim, int dst_ndim, + */ + +static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_memviewslice __pyx_v_dst, int __pyx_v_src_ndim, int __pyx_v_dst_ndim, int __pyx_v_dtype_is_object) { + void *__pyx_v_tmpdata; + size_t __pyx_v_itemsize; + int __pyx_v_i; + char __pyx_v_order; + int __pyx_v_broadcasting; + int __pyx_v_direct_copy; + __Pyx_memviewslice __pyx_v_tmp; + int __pyx_v_ndim; + int __pyx_r; + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + int __pyx_t_5; + int __pyx_t_6; + void *__pyx_t_7; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":1273 + * Check for overlapping memory and verify the shapes. + * """ + * cdef void *tmpdata = NULL # <<<<<<<<<<<<<< + * cdef size_t itemsize = src.memview.view.itemsize + * cdef int i + */ + __pyx_v_tmpdata = NULL; + + /* "View.MemoryView":1274 + * """ + * cdef void *tmpdata = NULL + * cdef size_t itemsize = src.memview.view.itemsize # <<<<<<<<<<<<<< + * cdef int i + * cdef char order = get_best_order(&src, src_ndim) + */ + __pyx_t_1 = __pyx_v_src.memview->view.itemsize; + __pyx_v_itemsize = __pyx_t_1; + + /* "View.MemoryView":1276 + * cdef size_t itemsize = src.memview.view.itemsize + * cdef int i + * cdef char order = get_best_order(&src, src_ndim) # <<<<<<<<<<<<<< + * cdef bint broadcasting = False + * cdef bint direct_copy = False + */ + __pyx_v_order = __pyx_get_best_slice_order((&__pyx_v_src), __pyx_v_src_ndim); + + /* "View.MemoryView":1277 + * cdef int i + * cdef char order = get_best_order(&src, src_ndim) + * cdef bint broadcasting = False # <<<<<<<<<<<<<< + * cdef bint direct_copy = False + * cdef __Pyx_memviewslice tmp + */ + __pyx_v_broadcasting = 0; + + /* "View.MemoryView":1278 + * cdef char order = get_best_order(&src, src_ndim) + * cdef bint broadcasting = False + * cdef bint direct_copy = False # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice tmp + * + */ + __pyx_v_direct_copy = 0; + + /* "View.MemoryView":1281 + * cdef __Pyx_memviewslice tmp + * + * if src_ndim < dst_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: + */ + __pyx_t_2 = (__pyx_v_src_ndim < __pyx_v_dst_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1282 + * + * if src_ndim < dst_ndim: + * broadcast_leading(&src, src_ndim, dst_ndim) # <<<<<<<<<<<<<< + * elif dst_ndim < src_ndim: + * broadcast_leading(&dst, dst_ndim, src_ndim) + */ + __pyx_memoryview_broadcast_leading((&__pyx_v_src), __pyx_v_src_ndim, __pyx_v_dst_ndim); + + /* "View.MemoryView":1281 + * cdef __Pyx_memviewslice tmp + * + * if src_ndim < dst_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1283 + * if src_ndim < dst_ndim: + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&dst, dst_ndim, src_ndim) + * + */ + __pyx_t_2 = (__pyx_v_dst_ndim < __pyx_v_src_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1284 + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: + * broadcast_leading(&dst, dst_ndim, src_ndim) # <<<<<<<<<<<<<< + * + * cdef int ndim = max(src_ndim, dst_ndim) + */ + __pyx_memoryview_broadcast_leading((&__pyx_v_dst), __pyx_v_dst_ndim, __pyx_v_src_ndim); + + /* "View.MemoryView":1283 + * if src_ndim < dst_ndim: + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&dst, dst_ndim, src_ndim) + * + */ + } + __pyx_L3:; + + /* "View.MemoryView":1286 + * broadcast_leading(&dst, dst_ndim, src_ndim) + * + * cdef int ndim = max(src_ndim, dst_ndim) # <<<<<<<<<<<<<< + * + * for i in range(ndim): + */ + __pyx_t_3 = __pyx_v_dst_ndim; + __pyx_t_4 = __pyx_v_src_ndim; + __pyx_t_2 = (__pyx_t_3 > __pyx_t_4); + if (__pyx_t_2) { + __pyx_t_5 = __pyx_t_3; + } else { + __pyx_t_5 = __pyx_t_4; + } + __pyx_v_ndim = __pyx_t_5; + + /* "View.MemoryView":1288 + * cdef int ndim = max(src_ndim, dst_ndim) + * + * for i in range(ndim): # <<<<<<<<<<<<<< + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: + */ + __pyx_t_5 = __pyx_v_ndim; + __pyx_t_3 = __pyx_t_5; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1289 + * + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: # <<<<<<<<<<<<<< + * if src.shape[i] == 1: + * broadcasting = True + */ + __pyx_t_2 = ((__pyx_v_src.shape[__pyx_v_i]) != (__pyx_v_dst.shape[__pyx_v_i])); + if (__pyx_t_2) { + + /* "View.MemoryView":1290 + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: # <<<<<<<<<<<<<< + * broadcasting = True + * src.strides[i] = 0 + */ + __pyx_t_2 = ((__pyx_v_src.shape[__pyx_v_i]) == 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1291 + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: + * broadcasting = True # <<<<<<<<<<<<<< + * src.strides[i] = 0 + * else: + */ + __pyx_v_broadcasting = 1; + + /* "View.MemoryView":1292 + * if src.shape[i] == 1: + * broadcasting = True + * src.strides[i] = 0 # <<<<<<<<<<<<<< + * else: + * _err_extents(i, dst.shape[i], src.shape[i]) + */ + (__pyx_v_src.strides[__pyx_v_i]) = 0; + + /* "View.MemoryView":1290 + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: # <<<<<<<<<<<<<< + * broadcasting = True + * src.strides[i] = 0 + */ + goto __pyx_L7; + } + + /* "View.MemoryView":1294 + * src.strides[i] = 0 + * else: + * _err_extents(i, dst.shape[i], src.shape[i]) # <<<<<<<<<<<<<< + * + * if src.suboffsets[i] >= 0: + */ + /*else*/ { + __pyx_t_6 = __pyx_memoryview_err_extents(__pyx_v_i, (__pyx_v_dst.shape[__pyx_v_i]), (__pyx_v_src.shape[__pyx_v_i])); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1294, __pyx_L1_error) + } + __pyx_L7:; + + /* "View.MemoryView":1289 + * + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: # <<<<<<<<<<<<<< + * if src.shape[i] == 1: + * broadcasting = True + */ + } + + /* "View.MemoryView":1296 + * _err_extents(i, dst.shape[i], src.shape[i]) + * + * if src.suboffsets[i] >= 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + */ + __pyx_t_2 = ((__pyx_v_src.suboffsets[__pyx_v_i]) >= 0); + if (__pyx_t_2) { + + /* "View.MemoryView":1297 + * + * if src.suboffsets[i] >= 0: + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) # <<<<<<<<<<<<<< + * + * if slices_overlap(&src, &dst, ndim, itemsize): + */ + __pyx_t_6 = __pyx_memoryview_err_dim(PyExc_ValueError, __pyx_kp_s_Dimension_d_is_not_direct, __pyx_v_i); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1297, __pyx_L1_error) + + /* "View.MemoryView":1296 + * _err_extents(i, dst.shape[i], src.shape[i]) + * + * if src.suboffsets[i] >= 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + */ + } + } + + /* "View.MemoryView":1299 + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + * if slices_overlap(&src, &dst, ndim, itemsize): # <<<<<<<<<<<<<< + * + * if not slice_is_contig(src, order, ndim): + */ + __pyx_t_2 = __pyx_slices_overlap((&__pyx_v_src), (&__pyx_v_dst), __pyx_v_ndim, __pyx_v_itemsize); + if (__pyx_t_2) { + + /* "View.MemoryView":1301 + * if slices_overlap(&src, &dst, ndim, itemsize): + * + * if not slice_is_contig(src, order, ndim): # <<<<<<<<<<<<<< + * order = get_best_order(&dst, ndim) + * + */ + __pyx_t_2 = (!__pyx_memviewslice_is_contig(__pyx_v_src, __pyx_v_order, __pyx_v_ndim)); + if (__pyx_t_2) { + + /* "View.MemoryView":1302 + * + * if not slice_is_contig(src, order, ndim): + * order = get_best_order(&dst, ndim) # <<<<<<<<<<<<<< + * + * tmpdata = copy_data_to_temp(&src, &tmp, order, ndim) + */ + __pyx_v_order = __pyx_get_best_slice_order((&__pyx_v_dst), __pyx_v_ndim); + + /* "View.MemoryView":1301 + * if slices_overlap(&src, &dst, ndim, itemsize): + * + * if not slice_is_contig(src, order, ndim): # <<<<<<<<<<<<<< + * order = get_best_order(&dst, ndim) + * + */ + } + + /* "View.MemoryView":1304 + * order = get_best_order(&dst, ndim) + * + * tmpdata = copy_data_to_temp(&src, &tmp, order, ndim) # <<<<<<<<<<<<<< + * src = tmp + * + */ + __pyx_t_7 = __pyx_memoryview_copy_data_to_temp((&__pyx_v_src), (&__pyx_v_tmp), __pyx_v_order, __pyx_v_ndim); if (unlikely(__pyx_t_7 == ((void *)NULL))) __PYX_ERR(1, 1304, __pyx_L1_error) + __pyx_v_tmpdata = __pyx_t_7; + + /* "View.MemoryView":1305 + * + * tmpdata = copy_data_to_temp(&src, &tmp, order, ndim) + * src = tmp # <<<<<<<<<<<<<< + * + * if not broadcasting: + */ + __pyx_v_src = __pyx_v_tmp; + + /* "View.MemoryView":1299 + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + * if slices_overlap(&src, &dst, ndim, itemsize): # <<<<<<<<<<<<<< + * + * if not slice_is_contig(src, order, ndim): + */ + } + + /* "View.MemoryView":1307 + * src = tmp + * + * if not broadcasting: # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = (!__pyx_v_broadcasting); + if (__pyx_t_2) { + + /* "View.MemoryView":1310 + * + * + * if slice_is_contig(src, 'C', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): + */ + __pyx_t_2 = __pyx_memviewslice_is_contig(__pyx_v_src, 'C', __pyx_v_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1311 + * + * if slice_is_contig(src, 'C', ndim): + * direct_copy = slice_is_contig(dst, 'C', ndim) # <<<<<<<<<<<<<< + * elif slice_is_contig(src, 'F', ndim): + * direct_copy = slice_is_contig(dst, 'F', ndim) + */ + __pyx_v_direct_copy = __pyx_memviewslice_is_contig(__pyx_v_dst, 'C', __pyx_v_ndim); + + /* "View.MemoryView":1310 + * + * + * if slice_is_contig(src, 'C', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): + */ + goto __pyx_L12; + } + + /* "View.MemoryView":1312 + * if slice_is_contig(src, 'C', ndim): + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + */ + __pyx_t_2 = __pyx_memviewslice_is_contig(__pyx_v_src, 'F', __pyx_v_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1313 + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): + * direct_copy = slice_is_contig(dst, 'F', ndim) # <<<<<<<<<<<<<< + * + * if direct_copy: + */ + __pyx_v_direct_copy = __pyx_memviewslice_is_contig(__pyx_v_dst, 'F', __pyx_v_ndim); + + /* "View.MemoryView":1312 + * if slice_is_contig(src, 'C', ndim): + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + */ + } + __pyx_L12:; + + /* "View.MemoryView":1315 + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + * if direct_copy: # <<<<<<<<<<<<<< + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + */ + if (__pyx_v_direct_copy) { + + /* "View.MemoryView":1317 + * if direct_copy: + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) # <<<<<<<<<<<<<< + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 0); + + /* "View.MemoryView":1318 + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) # <<<<<<<<<<<<<< + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * free(tmpdata) + */ + (void)(memcpy(__pyx_v_dst.data, __pyx_v_src.data, __pyx_memoryview_slice_get_size((&__pyx_v_src), __pyx_v_ndim))); + + /* "View.MemoryView":1319 + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) # <<<<<<<<<<<<<< + * free(tmpdata) + * return 0 + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 1); + + /* "View.MemoryView":1320 + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * free(tmpdata) # <<<<<<<<<<<<<< + * return 0 + * + */ + free(__pyx_v_tmpdata); + + /* "View.MemoryView":1321 + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * free(tmpdata) + * return 0 # <<<<<<<<<<<<<< + * + * if order == 'F' == get_best_order(&dst, ndim): + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":1315 + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + * if direct_copy: # <<<<<<<<<<<<<< + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + */ + } + + /* "View.MemoryView":1307 + * src = tmp + * + * if not broadcasting: # <<<<<<<<<<<<<< + * + * + */ + } + + /* "View.MemoryView":1323 + * return 0 + * + * if order == 'F' == get_best_order(&dst, ndim): # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = (__pyx_v_order == 'F'); + if (__pyx_t_2) { + __pyx_t_2 = ('F' == __pyx_get_best_slice_order((&__pyx_v_dst), __pyx_v_ndim)); + } + if (__pyx_t_2) { + + /* "View.MemoryView":1326 + * + * + * transpose_memslice(&src) # <<<<<<<<<<<<<< + * transpose_memslice(&dst) + * + */ + __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_src)); if (unlikely(__pyx_t_5 == ((int)-1))) __PYX_ERR(1, 1326, __pyx_L1_error) + + /* "View.MemoryView":1327 + * + * transpose_memslice(&src) + * transpose_memslice(&dst) # <<<<<<<<<<<<<< + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + */ + __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_dst)); if (unlikely(__pyx_t_5 == ((int)-1))) __PYX_ERR(1, 1327, __pyx_L1_error) + + /* "View.MemoryView":1323 + * return 0 + * + * if order == 'F' == get_best_order(&dst, ndim): # <<<<<<<<<<<<<< + * + * + */ + } + + /* "View.MemoryView":1329 + * transpose_memslice(&dst) + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) # <<<<<<<<<<<<<< + * copy_strided_to_strided(&src, &dst, ndim, itemsize) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 0); + + /* "View.MemoryView":1330 + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * copy_strided_to_strided(&src, &dst, ndim, itemsize) # <<<<<<<<<<<<<< + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * + */ + copy_strided_to_strided((&__pyx_v_src), (&__pyx_v_dst), __pyx_v_ndim, __pyx_v_itemsize); + + /* "View.MemoryView":1331 + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * copy_strided_to_strided(&src, &dst, ndim, itemsize) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) # <<<<<<<<<<<<<< + * + * free(tmpdata) + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 1); + + /* "View.MemoryView":1333 + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * + * free(tmpdata) # <<<<<<<<<<<<<< + * return 0 + * + */ + free(__pyx_v_tmpdata); + + /* "View.MemoryView":1334 + * + * free(tmpdata) + * return 0 # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_broadcast_leading') + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":1265 + * + * @cname('__pyx_memoryview_copy_contents') + * cdef int memoryview_copy_contents(__Pyx_memviewslice src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice dst, + * int src_ndim, int dst_ndim, + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.memoryview_copy_contents", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1337 + * + * @cname('__pyx_memoryview_broadcast_leading') + * cdef void broadcast_leading(__Pyx_memviewslice *mslice, # <<<<<<<<<<<<<< + * int ndim, + * int ndim_other) noexcept nogil: + */ + +static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslice, int __pyx_v_ndim, int __pyx_v_ndim_other) { + int __pyx_v_i; + int __pyx_v_offset; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + + /* "View.MemoryView":1341 + * int ndim_other) noexcept nogil: + * cdef int i + * cdef int offset = ndim_other - ndim # <<<<<<<<<<<<<< + * + * for i in range(ndim - 1, -1, -1): + */ + __pyx_v_offset = (__pyx_v_ndim_other - __pyx_v_ndim); + + /* "View.MemoryView":1343 + * cdef int offset = ndim_other - ndim + * + * for i in range(ndim - 1, -1, -1): # <<<<<<<<<<<<<< + * mslice.shape[i + offset] = mslice.shape[i] + * mslice.strides[i + offset] = mslice.strides[i] + */ + for (__pyx_t_1 = (__pyx_v_ndim - 1); __pyx_t_1 > -1; __pyx_t_1-=1) { + __pyx_v_i = __pyx_t_1; + + /* "View.MemoryView":1344 + * + * for i in range(ndim - 1, -1, -1): + * mslice.shape[i + offset] = mslice.shape[i] # <<<<<<<<<<<<<< + * mslice.strides[i + offset] = mslice.strides[i] + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] + */ + (__pyx_v_mslice->shape[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->shape[__pyx_v_i]); + + /* "View.MemoryView":1345 + * for i in range(ndim - 1, -1, -1): + * mslice.shape[i + offset] = mslice.shape[i] + * mslice.strides[i + offset] = mslice.strides[i] # <<<<<<<<<<<<<< + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] + * + */ + (__pyx_v_mslice->strides[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->strides[__pyx_v_i]); + + /* "View.MemoryView":1346 + * mslice.shape[i + offset] = mslice.shape[i] + * mslice.strides[i + offset] = mslice.strides[i] + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] # <<<<<<<<<<<<<< + * + * for i in range(offset): + */ + (__pyx_v_mslice->suboffsets[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->suboffsets[__pyx_v_i]); + } + + /* "View.MemoryView":1348 + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] + * + * for i in range(offset): # <<<<<<<<<<<<<< + * mslice.shape[i] = 1 + * mslice.strides[i] = mslice.strides[0] + */ + __pyx_t_1 = __pyx_v_offset; + __pyx_t_2 = __pyx_t_1; + for (__pyx_t_3 = 0; __pyx_t_3 < __pyx_t_2; __pyx_t_3+=1) { + __pyx_v_i = __pyx_t_3; + + /* "View.MemoryView":1349 + * + * for i in range(offset): + * mslice.shape[i] = 1 # <<<<<<<<<<<<<< + * mslice.strides[i] = mslice.strides[0] + * mslice.suboffsets[i] = -1 + */ + (__pyx_v_mslice->shape[__pyx_v_i]) = 1; + + /* "View.MemoryView":1350 + * for i in range(offset): + * mslice.shape[i] = 1 + * mslice.strides[i] = mslice.strides[0] # <<<<<<<<<<<<<< + * mslice.suboffsets[i] = -1 + * + */ + (__pyx_v_mslice->strides[__pyx_v_i]) = (__pyx_v_mslice->strides[0]); + + /* "View.MemoryView":1351 + * mslice.shape[i] = 1 + * mslice.strides[i] = mslice.strides[0] + * mslice.suboffsets[i] = -1 # <<<<<<<<<<<<<< + * + * + */ + (__pyx_v_mslice->suboffsets[__pyx_v_i]) = -1L; + } + + /* "View.MemoryView":1337 + * + * @cname('__pyx_memoryview_broadcast_leading') + * cdef void broadcast_leading(__Pyx_memviewslice *mslice, # <<<<<<<<<<<<<< + * int ndim, + * int ndim_other) noexcept nogil: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1359 + * + * @cname('__pyx_memoryview_refcount_copying') + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: # <<<<<<<<<<<<<< + * + * if dtype_is_object: + */ + +static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *__pyx_v_dst, int __pyx_v_dtype_is_object, int __pyx_v_ndim, int __pyx_v_inc) { + + /* "View.MemoryView":1361 + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: + * + * if dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice_with_gil(dst.data, dst.shape, dst.strides, ndim, inc) + * + */ + if (__pyx_v_dtype_is_object) { + + /* "View.MemoryView":1362 + * + * if dtype_is_object: + * refcount_objects_in_slice_with_gil(dst.data, dst.shape, dst.strides, ndim, inc) # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil') + */ + __pyx_memoryview_refcount_objects_in_slice_with_gil(__pyx_v_dst->data, __pyx_v_dst->shape, __pyx_v_dst->strides, __pyx_v_ndim, __pyx_v_inc); + + /* "View.MemoryView":1361 + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: + * + * if dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice_with_gil(dst.data, dst.shape, dst.strides, ndim, inc) + * + */ + } + + /* "View.MemoryView":1359 + * + * @cname('__pyx_memoryview_refcount_copying') + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: # <<<<<<<<<<<<<< + * + * if dtype_is_object: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1365 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil') + * cdef void refcount_objects_in_slice_with_gil(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * bint inc) noexcept with gil: + */ + +static void __pyx_memoryview_refcount_objects_in_slice_with_gil(char *__pyx_v_data, Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, int __pyx_v_ndim, int __pyx_v_inc) { + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + + /* "View.MemoryView":1368 + * Py_ssize_t *strides, int ndim, + * bint inc) noexcept with gil: + * refcount_objects_in_slice(data, shape, strides, ndim, inc) # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_refcount_objects_in_slice') + */ + __pyx_memoryview_refcount_objects_in_slice(__pyx_v_data, __pyx_v_shape, __pyx_v_strides, __pyx_v_ndim, __pyx_v_inc); + + /* "View.MemoryView":1365 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil') + * cdef void refcount_objects_in_slice_with_gil(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * bint inc) noexcept with gil: + */ + + /* function exit code */ + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif +} + +/* "View.MemoryView":1371 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice') + * cdef void refcount_objects_in_slice(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, bint inc) noexcept: + * cdef Py_ssize_t i + */ + +static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, int __pyx_v_ndim, int __pyx_v_inc) { + CYTHON_UNUSED Py_ssize_t __pyx_v_i; + Py_ssize_t __pyx_v_stride; + Py_ssize_t __pyx_t_1; + Py_ssize_t __pyx_t_2; + Py_ssize_t __pyx_t_3; + int __pyx_t_4; + + /* "View.MemoryView":1374 + * Py_ssize_t *strides, int ndim, bint inc) noexcept: + * cdef Py_ssize_t i + * cdef Py_ssize_t stride = strides[0] # <<<<<<<<<<<<<< + * + * for i in range(shape[0]): + */ + __pyx_v_stride = (__pyx_v_strides[0]); + + /* "View.MemoryView":1376 + * cdef Py_ssize_t stride = strides[0] + * + * for i in range(shape[0]): # <<<<<<<<<<<<<< + * if ndim == 1: + * if inc: + */ + __pyx_t_1 = (__pyx_v_shape[0]); + __pyx_t_2 = __pyx_t_1; + for (__pyx_t_3 = 0; __pyx_t_3 < __pyx_t_2; __pyx_t_3+=1) { + __pyx_v_i = __pyx_t_3; + + /* "View.MemoryView":1377 + * + * for i in range(shape[0]): + * if ndim == 1: # <<<<<<<<<<<<<< + * if inc: + * Py_INCREF(( data)[0]) + */ + __pyx_t_4 = (__pyx_v_ndim == 1); + if (__pyx_t_4) { + + /* "View.MemoryView":1378 + * for i in range(shape[0]): + * if ndim == 1: + * if inc: # <<<<<<<<<<<<<< + * Py_INCREF(( data)[0]) + * else: + */ + if (__pyx_v_inc) { + + /* "View.MemoryView":1379 + * if ndim == 1: + * if inc: + * Py_INCREF(( data)[0]) # <<<<<<<<<<<<<< + * else: + * Py_DECREF(( data)[0]) + */ + Py_INCREF((((PyObject **)__pyx_v_data)[0])); + + /* "View.MemoryView":1378 + * for i in range(shape[0]): + * if ndim == 1: + * if inc: # <<<<<<<<<<<<<< + * Py_INCREF(( data)[0]) + * else: + */ + goto __pyx_L6; + } + + /* "View.MemoryView":1381 + * Py_INCREF(( data)[0]) + * else: + * Py_DECREF(( data)[0]) # <<<<<<<<<<<<<< + * else: + * refcount_objects_in_slice(data, shape + 1, strides + 1, ndim - 1, inc) + */ + /*else*/ { + Py_DECREF((((PyObject **)__pyx_v_data)[0])); + } + __pyx_L6:; + + /* "View.MemoryView":1377 + * + * for i in range(shape[0]): + * if ndim == 1: # <<<<<<<<<<<<<< + * if inc: + * Py_INCREF(( data)[0]) + */ + goto __pyx_L5; + } + + /* "View.MemoryView":1383 + * Py_DECREF(( data)[0]) + * else: + * refcount_objects_in_slice(data, shape + 1, strides + 1, ndim - 1, inc) # <<<<<<<<<<<<<< + * + * data += stride + */ + /*else*/ { + __pyx_memoryview_refcount_objects_in_slice(__pyx_v_data, (__pyx_v_shape + 1), (__pyx_v_strides + 1), (__pyx_v_ndim - 1), __pyx_v_inc); + } + __pyx_L5:; + + /* "View.MemoryView":1385 + * refcount_objects_in_slice(data, shape + 1, strides + 1, ndim - 1, inc) + * + * data += stride # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_data = (__pyx_v_data + __pyx_v_stride); + } + + /* "View.MemoryView":1371 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice') + * cdef void refcount_objects_in_slice(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, bint inc) noexcept: + * cdef Py_ssize_t i + */ + + /* function exit code */ +} + +/* "View.MemoryView":1391 + * + * @cname('__pyx_memoryview_slice_assign_scalar') + * cdef void slice_assign_scalar(__Pyx_memviewslice *dst, int ndim, # <<<<<<<<<<<<<< + * size_t itemsize, void *item, + * bint dtype_is_object) noexcept nogil: + */ + +static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *__pyx_v_dst, int __pyx_v_ndim, size_t __pyx_v_itemsize, void *__pyx_v_item, int __pyx_v_dtype_is_object) { + + /* "View.MemoryView":1394 + * size_t itemsize, void *item, + * bint dtype_is_object) noexcept nogil: + * refcount_copying(dst, dtype_is_object, ndim, inc=False) # <<<<<<<<<<<<<< + * _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim, itemsize, item) + * refcount_copying(dst, dtype_is_object, ndim, inc=True) + */ + __pyx_memoryview_refcount_copying(__pyx_v_dst, __pyx_v_dtype_is_object, __pyx_v_ndim, 0); + + /* "View.MemoryView":1395 + * bint dtype_is_object) noexcept nogil: + * refcount_copying(dst, dtype_is_object, ndim, inc=False) + * _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim, itemsize, item) # <<<<<<<<<<<<<< + * refcount_copying(dst, dtype_is_object, ndim, inc=True) + * + */ + __pyx_memoryview__slice_assign_scalar(__pyx_v_dst->data, __pyx_v_dst->shape, __pyx_v_dst->strides, __pyx_v_ndim, __pyx_v_itemsize, __pyx_v_item); + + /* "View.MemoryView":1396 + * refcount_copying(dst, dtype_is_object, ndim, inc=False) + * _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim, itemsize, item) + * refcount_copying(dst, dtype_is_object, ndim, inc=True) # <<<<<<<<<<<<<< + * + * + */ + __pyx_memoryview_refcount_copying(__pyx_v_dst, __pyx_v_dtype_is_object, __pyx_v_ndim, 1); + + /* "View.MemoryView":1391 + * + * @cname('__pyx_memoryview_slice_assign_scalar') + * cdef void slice_assign_scalar(__Pyx_memviewslice *dst, int ndim, # <<<<<<<<<<<<<< + * size_t itemsize, void *item, + * bint dtype_is_object) noexcept nogil: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1400 + * + * @cname('__pyx_memoryview__slice_assign_scalar') + * cdef void _slice_assign_scalar(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * size_t itemsize, void *item) noexcept nogil: + */ + +static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, int __pyx_v_ndim, size_t __pyx_v_itemsize, void *__pyx_v_item) { + CYTHON_UNUSED Py_ssize_t __pyx_v_i; + Py_ssize_t __pyx_v_stride; + Py_ssize_t __pyx_v_extent; + int __pyx_t_1; + Py_ssize_t __pyx_t_2; + Py_ssize_t __pyx_t_3; + Py_ssize_t __pyx_t_4; + + /* "View.MemoryView":1404 + * size_t itemsize, void *item) noexcept nogil: + * cdef Py_ssize_t i + * cdef Py_ssize_t stride = strides[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t extent = shape[0] + * + */ + __pyx_v_stride = (__pyx_v_strides[0]); + + /* "View.MemoryView":1405 + * cdef Py_ssize_t i + * cdef Py_ssize_t stride = strides[0] + * cdef Py_ssize_t extent = shape[0] # <<<<<<<<<<<<<< + * + * if ndim == 1: + */ + __pyx_v_extent = (__pyx_v_shape[0]); + + /* "View.MemoryView":1407 + * cdef Py_ssize_t extent = shape[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * for i in range(extent): + * memcpy(data, item, itemsize) + */ + __pyx_t_1 = (__pyx_v_ndim == 1); + if (__pyx_t_1) { + + /* "View.MemoryView":1408 + * + * if ndim == 1: + * for i in range(extent): # <<<<<<<<<<<<<< + * memcpy(data, item, itemsize) + * data += stride + */ + __pyx_t_2 = __pyx_v_extent; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1409 + * if ndim == 1: + * for i in range(extent): + * memcpy(data, item, itemsize) # <<<<<<<<<<<<<< + * data += stride + * else: + */ + (void)(memcpy(__pyx_v_data, __pyx_v_item, __pyx_v_itemsize)); + + /* "View.MemoryView":1410 + * for i in range(extent): + * memcpy(data, item, itemsize) + * data += stride # <<<<<<<<<<<<<< + * else: + * for i in range(extent): + */ + __pyx_v_data = (__pyx_v_data + __pyx_v_stride); + } + + /* "View.MemoryView":1407 + * cdef Py_ssize_t extent = shape[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * for i in range(extent): + * memcpy(data, item, itemsize) + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1412 + * data += stride + * else: + * for i in range(extent): # <<<<<<<<<<<<<< + * _slice_assign_scalar(data, shape + 1, strides + 1, ndim - 1, itemsize, item) + * data += stride + */ + /*else*/ { + __pyx_t_2 = __pyx_v_extent; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1413 + * else: + * for i in range(extent): + * _slice_assign_scalar(data, shape + 1, strides + 1, ndim - 1, itemsize, item) # <<<<<<<<<<<<<< + * data += stride + * + */ + __pyx_memoryview__slice_assign_scalar(__pyx_v_data, (__pyx_v_shape + 1), (__pyx_v_strides + 1), (__pyx_v_ndim - 1), __pyx_v_itemsize, __pyx_v_item); + + /* "View.MemoryView":1414 + * for i in range(extent): + * _slice_assign_scalar(data, shape + 1, strides + 1, ndim - 1, itemsize, item) + * data += stride # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_data = (__pyx_v_data + __pyx_v_stride); + } + } + __pyx_L3:; + + /* "View.MemoryView":1400 + * + * @cname('__pyx_memoryview__slice_assign_scalar') + * cdef void _slice_assign_scalar(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * size_t itemsize, void *item) noexcept nogil: + */ + + /* function exit code */ +} + +/* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_1__pyx_unpickle_Enum(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_15View_dot_MemoryView_1__pyx_unpickle_Enum = {"__pyx_unpickle_Enum", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_15View_dot_MemoryView_1__pyx_unpickle_Enum, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_15View_dot_MemoryView_1__pyx_unpickle_Enum(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyObject *__pyx_v___pyx_type = 0; + long __pyx_v___pyx_checksum; + PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[3] = {0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__pyx_unpickle_Enum (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_type,&__pyx_n_s_pyx_checksum,&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_type)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_checksum)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_Enum", 1, 3, 3, 1); __PYX_ERR(1, 1, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_Enum", 1, 3, 3, 2); __PYX_ERR(1, 1, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__pyx_unpickle_Enum") < 0)) __PYX_ERR(1, 1, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 3)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + } + __pyx_v___pyx_type = values[0]; + __pyx_v___pyx_checksum = __Pyx_PyInt_As_long(values[1]); if (unlikely((__pyx_v___pyx_checksum == (long)-1) && PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + __pyx_v___pyx_state = values[2]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_Enum", 1, 3, 3, __pyx_nargs); __PYX_ERR(1, 1, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.__pyx_unpickle_Enum", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf_15View_dot_MemoryView___pyx_unpickle_Enum(__pyx_self, __pyx_v___pyx_type, __pyx_v___pyx_checksum, __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView___pyx_unpickle_Enum(CYTHON_UNUSED PyObject *__pyx_self, PyObject *__pyx_v___pyx_type, long __pyx_v___pyx_checksum, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_v___pyx_PickleError = 0; + PyObject *__pyx_v___pyx_result = 0; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + unsigned int __pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__pyx_unpickle_Enum", 1); + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + */ + __pyx_t_1 = __Pyx_PyInt_From_long(__pyx_v___pyx_checksum); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = (__Pyx_PySequence_ContainsTF(__pyx_t_1, __pyx_tuple__8, Py_NE)); if (unlikely((__pyx_t_2 < 0))) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + if (__pyx_t_2) { + + /* "(tree fragment)":5 + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): + * from pickle import PickleError as __pyx_PickleError # <<<<<<<<<<<<<< + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) + */ + __pyx_t_1 = PyList_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_n_s_PickleError); + __Pyx_GIVEREF(__pyx_n_s_PickleError); + if (__Pyx_PyList_SET_ITEM(__pyx_t_1, 0, __pyx_n_s_PickleError)) __PYX_ERR(1, 5, __pyx_L1_error); + __pyx_t_3 = __Pyx_Import(__pyx_n_s_pickle, __pyx_t_1, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_ImportFrom(__pyx_t_3, __pyx_n_s_PickleError); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_t_1); + __pyx_v___pyx_PickleError = __pyx_t_1; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + + /* "(tree fragment)":6 + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum # <<<<<<<<<<<<<< + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: + */ + __pyx_t_3 = __Pyx_PyInt_From_long(__pyx_v___pyx_checksum); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 = __Pyx_PyString_Format(__pyx_kp_s_Incompatible_checksums_0x_x_vs_0, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_Raise(__pyx_v___pyx_PickleError, __pyx_t_1, 0, 0); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __PYX_ERR(1, 6, __pyx_L1_error) + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + */ + } + + /* "(tree fragment)":7 + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) # <<<<<<<<<<<<<< + * if __pyx_state is not None: + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + */ + __pyx_t_3 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_MemviewEnum_type), __pyx_n_s_new); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = NULL; + __pyx_t_5 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_3))) { + __pyx_t_4 = PyMethod_GET_SELF(__pyx_t_3); + if (likely(__pyx_t_4)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_3); + __Pyx_INCREF(__pyx_t_4); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_3, function); + __pyx_t_5 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_4, __pyx_v___pyx_type}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_3, __pyx_callargs+1-__pyx_t_5, 1+__pyx_t_5); + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + __pyx_v___pyx_result = __pyx_t_1; + __pyx_t_1 = 0; + + /* "(tree fragment)":8 + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + */ + __pyx_t_2 = (__pyx_v___pyx_state != Py_None); + if (__pyx_t_2) { + + /* "(tree fragment)":9 + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) # <<<<<<<<<<<<<< + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + */ + if (!(likely(PyTuple_CheckExact(__pyx_v___pyx_state))||((__pyx_v___pyx_state) == Py_None) || __Pyx_RaiseUnexpectedTypeError("tuple", __pyx_v___pyx_state))) __PYX_ERR(1, 9, __pyx_L1_error) + __pyx_t_1 = __pyx_unpickle_Enum__set_state(((struct __pyx_MemviewEnum_obj *)__pyx_v___pyx_result), ((PyObject*)__pyx_v___pyx_state)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":8 + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + */ + } + + /* "(tree fragment)":10 + * if __pyx_state is not None: + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result # <<<<<<<<<<<<<< + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v___pyx_result); + __pyx_r = __pyx_v___pyx_result; + goto __pyx_L0; + + /* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.__pyx_unpickle_Enum", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v___pyx_PickleError); + __Pyx_XDECREF(__pyx_v___pyx_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":11 + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): # <<<<<<<<<<<<<< + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + */ + +static PyObject *__pyx_unpickle_Enum__set_state(struct __pyx_MemviewEnum_obj *__pyx_v___pyx_result, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + Py_ssize_t __pyx_t_3; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + unsigned int __pyx_t_8; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__pyx_unpickle_Enum__set_state", 1); + + /* "(tree fragment)":12 + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] # <<<<<<<<<<<<<< + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + * __pyx_result.__dict__.update(__pyx_state[1]) + */ + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 12, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_GIVEREF(__pyx_t_1); + __Pyx_GOTREF(__pyx_v___pyx_result->name); + __Pyx_DECREF(__pyx_v___pyx_result->name); + __pyx_v___pyx_result->name = __pyx_t_1; + __pyx_t_1 = 0; + + /* "(tree fragment)":13 + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): # <<<<<<<<<<<<<< + * __pyx_result.__dict__.update(__pyx_state[1]) + */ + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(1, 13, __pyx_L1_error) + } + __pyx_t_3 = __Pyx_PyTuple_GET_SIZE(__pyx_v___pyx_state); if (unlikely(__pyx_t_3 == ((Py_ssize_t)-1))) __PYX_ERR(1, 13, __pyx_L1_error) + __pyx_t_4 = (__pyx_t_3 > 1); + if (__pyx_t_4) { + } else { + __pyx_t_2 = __pyx_t_4; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_4 = __Pyx_HasAttr(((PyObject *)__pyx_v___pyx_result), __pyx_n_s_dict); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 13, __pyx_L1_error) + __pyx_t_2 = __pyx_t_4; + __pyx_L4_bool_binop_done:; + if (__pyx_t_2) { + + /* "(tree fragment)":14 + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + * __pyx_result.__dict__.update(__pyx_state[1]) # <<<<<<<<<<<<<< + */ + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v___pyx_result), __pyx_n_s_dict); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_6 = __Pyx_PyObject_GetAttrStr(__pyx_t_5, __pyx_n_s_update); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 14, __pyx_L1_error) + } + __pyx_t_5 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 1, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_7 = NULL; + __pyx_t_8 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_6))) { + __pyx_t_7 = PyMethod_GET_SELF(__pyx_t_6); + if (likely(__pyx_t_7)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_6); + __Pyx_INCREF(__pyx_t_7); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_6, function); + __pyx_t_8 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_7, __pyx_t_5}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_6, __pyx_callargs+1-__pyx_t_8, 1+__pyx_t_8); + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + } + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":13 + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): # <<<<<<<<<<<<<< + * __pyx_result.__dict__.update(__pyx_state[1]) + */ + } + + /* "(tree fragment)":11 + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): # <<<<<<<<<<<<<< + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_AddTraceback("View.MemoryView.__pyx_unpickle_Enum__set_state", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":287 + * + * @property + * cdef inline npy_intp itemsize(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ELSIZE(self) + * + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_8itemsize_itemsize(PyArray_Descr *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":288 + * @property + * cdef inline npy_intp itemsize(self) noexcept nogil: + * return PyDataType_ELSIZE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyDataType_ELSIZE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":287 + * + * @property + * cdef inline npy_intp itemsize(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ELSIZE(self) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":291 + * + * @property + * cdef inline npy_intp alignment(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ALIGNMENT(self) + * + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_9alignment_alignment(PyArray_Descr *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":292 + * @property + * cdef inline npy_intp alignment(self) noexcept nogil: + * return PyDataType_ALIGNMENT(self) # <<<<<<<<<<<<<< + * + * # Use fields/names with care as they may be NULL. You must check + */ + __pyx_r = PyDataType_ALIGNMENT(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":291 + * + * @property + * cdef inline npy_intp alignment(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ALIGNMENT(self) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":297 + * # for this using PyDataType_HASFIELDS. + * @property + * cdef inline object fields(self): # <<<<<<<<<<<<<< + * return PyDataType_FIELDS(self) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_6fields_fields(PyArray_Descr *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1; + __Pyx_RefNannySetupContext("fields", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":298 + * @property + * cdef inline object fields(self): + * return PyDataType_FIELDS(self) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyDataType_FIELDS(__pyx_v_self); + __Pyx_INCREF(((PyObject *)__pyx_t_1)); + __pyx_r = ((PyObject *)__pyx_t_1); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":297 + * # for this using PyDataType_HASFIELDS. + * @property + * cdef inline object fields(self): # <<<<<<<<<<<<<< + * return PyDataType_FIELDS(self) + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":301 + * + * @property + * cdef inline tuple names(self): # <<<<<<<<<<<<<< + * return PyDataType_NAMES(self) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_5names_names(PyArray_Descr *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1; + __Pyx_RefNannySetupContext("names", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":302 + * @property + * cdef inline tuple names(self): + * return PyDataType_NAMES(self) # <<<<<<<<<<<<<< + * + * # Use PyDataType_HASSUBARRAY to test whether this field is + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyDataType_NAMES(__pyx_v_self); + __Pyx_INCREF(((PyObject*)__pyx_t_1)); + __pyx_r = ((PyObject*)__pyx_t_1); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":301 + * + * @property + * cdef inline tuple names(self): # <<<<<<<<<<<<<< + * return PyDataType_NAMES(self) + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":308 + * # this field via the inline helper method PyDataType_SHAPE. + * @property + * cdef inline PyArray_ArrayDescr* subarray(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_SUBARRAY(self) + * + */ + +static CYTHON_INLINE PyArray_ArrayDescr *__pyx_f_5numpy_5dtype_8subarray_subarray(PyArray_Descr *__pyx_v_self) { + PyArray_ArrayDescr *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":309 + * @property + * cdef inline PyArray_ArrayDescr* subarray(self) noexcept nogil: + * return PyDataType_SUBARRAY(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyDataType_SUBARRAY(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":308 + * # this field via the inline helper method PyDataType_SHAPE. + * @property + * cdef inline PyArray_ArrayDescr* subarray(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_SUBARRAY(self) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":312 + * + * @property + * cdef inline npy_uint64 flags(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The data types flags.""" + * return PyDataType_FLAGS(self) + */ + +static CYTHON_INLINE npy_uint64 __pyx_f_5numpy_5dtype_5flags_flags(PyArray_Descr *__pyx_v_self) { + npy_uint64 __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":314 + * cdef inline npy_uint64 flags(self) noexcept nogil: + * """The data types flags.""" + * return PyDataType_FLAGS(self) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyDataType_FLAGS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":312 + * + * @property + * cdef inline npy_uint64 flags(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The data types flags.""" + * return PyDataType_FLAGS(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":324 + * + * @property + * cdef inline int numiter(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of arrays that need to be broadcast to the same shape.""" + * return PyArray_MultiIter_NUMITER(self) + */ + +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_7numiter_numiter(PyArrayMultiIterObject *__pyx_v_self) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":326 + * cdef inline int numiter(self) noexcept nogil: + * """The number of arrays that need to be broadcast to the same shape.""" + * return PyArray_MultiIter_NUMITER(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_NUMITER(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":324 + * + * @property + * cdef inline int numiter(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of arrays that need to be broadcast to the same shape.""" + * return PyArray_MultiIter_NUMITER(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":329 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The total broadcasted size.""" + * return PyArray_MultiIter_SIZE(self) + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_4size_size(PyArrayMultiIterObject *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":331 + * cdef inline npy_intp size(self) noexcept nogil: + * """The total broadcasted size.""" + * return PyArray_MultiIter_SIZE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_SIZE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":329 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The total broadcasted size.""" + * return PyArray_MultiIter_SIZE(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":334 + * + * @property + * cdef inline npy_intp index(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The current (1-d) index into the broadcasted result.""" + * return PyArray_MultiIter_INDEX(self) + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_5index_index(PyArrayMultiIterObject *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":336 + * cdef inline npy_intp index(self) noexcept nogil: + * """The current (1-d) index into the broadcasted result.""" + * return PyArray_MultiIter_INDEX(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_INDEX(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":334 + * + * @property + * cdef inline npy_intp index(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The current (1-d) index into the broadcasted result.""" + * return PyArray_MultiIter_INDEX(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":339 + * + * @property + * cdef inline int nd(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of dimensions in the broadcasted result.""" + * return PyArray_MultiIter_NDIM(self) + */ + +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_2nd_nd(PyArrayMultiIterObject *__pyx_v_self) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":341 + * cdef inline int nd(self) noexcept nogil: + * """The number of dimensions in the broadcasted result.""" + * return PyArray_MultiIter_NDIM(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_NDIM(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":339 + * + * @property + * cdef inline int nd(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of dimensions in the broadcasted result.""" + * return PyArray_MultiIter_NDIM(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":344 + * + * @property + * cdef inline npy_intp* dimensions(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The shape of the broadcasted result.""" + * return PyArray_MultiIter_DIMS(self) + */ + +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_9broadcast_10dimensions_dimensions(PyArrayMultiIterObject *__pyx_v_self) { + npy_intp *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":346 + * cdef inline npy_intp* dimensions(self) noexcept nogil: + * """The shape of the broadcasted result.""" + * return PyArray_MultiIter_DIMS(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_DIMS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":344 + * + * @property + * cdef inline npy_intp* dimensions(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The shape of the broadcasted result.""" + * return PyArray_MultiIter_DIMS(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":349 + * + * @property + * cdef inline void** iters(self) noexcept nogil: # <<<<<<<<<<<<<< + * """An array of iterator objects that holds the iterators for the arrays to be broadcast together. + * On return, the iterators are adjusted for broadcasting.""" + */ + +static CYTHON_INLINE void **__pyx_f_5numpy_9broadcast_5iters_iters(PyArrayMultiIterObject *__pyx_v_self) { + void **__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":352 + * """An array of iterator objects that holds the iterators for the arrays to be broadcast together. + * On return, the iterators are adjusted for broadcasting.""" + * return PyArray_MultiIter_ITERS(self) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyArray_MultiIter_ITERS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":349 + * + * @property + * cdef inline void** iters(self) noexcept nogil: # <<<<<<<<<<<<<< + * """An array of iterator objects that holds the iterators for the arrays to be broadcast together. + * On return, the iterators are adjusted for broadcasting.""" + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":367 + * + * @property + * cdef inline PyObject* base(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a borrowed reference to the object owning the data/memory. + * """ + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_7ndarray_4base_base(PyArrayObject *__pyx_v_self) { + PyObject *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":370 + * """Returns a borrowed reference to the object owning the data/memory. + * """ + * return PyArray_BASE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_BASE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":367 + * + * @property + * cdef inline PyObject* base(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a borrowed reference to the object owning the data/memory. + * """ + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":373 + * + * @property + * cdef inline dtype descr(self): # <<<<<<<<<<<<<< + * """Returns an owned reference to the dtype of the array. + * """ + */ + +static CYTHON_INLINE PyArray_Descr *__pyx_f_5numpy_7ndarray_5descr_descr(PyArrayObject *__pyx_v_self) { + PyArray_Descr *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyArray_Descr *__pyx_t_1; + __Pyx_RefNannySetupContext("descr", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":376 + * """Returns an owned reference to the dtype of the array. + * """ + * return PyArray_DESCR(self) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __pyx_t_1 = PyArray_DESCR(__pyx_v_self); + __Pyx_INCREF((PyObject *)((PyArray_Descr *)__pyx_t_1)); + __pyx_r = ((PyArray_Descr *)__pyx_t_1); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":373 + * + * @property + * cdef inline dtype descr(self): # <<<<<<<<<<<<<< + * """Returns an owned reference to the dtype of the array. + * """ + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":379 + * + * @property + * cdef inline int ndim(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the number of dimensions in the array. + * """ + */ + +static CYTHON_INLINE int __pyx_f_5numpy_7ndarray_4ndim_ndim(PyArrayObject *__pyx_v_self) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":382 + * """Returns the number of dimensions in the array. + * """ + * return PyArray_NDIM(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_NDIM(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":379 + * + * @property + * cdef inline int ndim(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the number of dimensions in the array. + * """ + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":385 + * + * @property + * cdef inline npy_intp *shape(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the dimensions/shape of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_5shape_shape(PyArrayObject *__pyx_v_self) { + npy_intp *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":390 + * Can return NULL for 0-dimensional arrays. + * """ + * return PyArray_DIMS(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_DIMS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":385 + * + * @property + * cdef inline npy_intp *shape(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the dimensions/shape of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":393 + * + * @property + * cdef inline npy_intp *strides(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the strides of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_7strides_strides(PyArrayObject *__pyx_v_self) { + npy_intp *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":397 + * The number of elements matches the number of dimensions of the array (ndim). + * """ + * return PyArray_STRIDES(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_STRIDES(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":393 + * + * @property + * cdef inline npy_intp *strides(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the strides of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":400 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the total size (in number of elements) of the array. + * """ + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_7ndarray_4size_size(PyArrayObject *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":403 + * """Returns the total size (in number of elements) of the array. + * """ + * return PyArray_SIZE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_SIZE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":400 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the total size (in number of elements) of the array. + * """ + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":406 + * + * @property + * cdef inline char* data(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The pointer to the data buffer as a char*. + * This is provided for legacy reasons to avoid direct struct field access. + */ + +static CYTHON_INLINE char *__pyx_f_5numpy_7ndarray_4data_data(PyArrayObject *__pyx_v_self) { + char *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":412 + * of `PyArray_DATA()` instead, which returns a 'void*'. + * """ + * return PyArray_BYTES(self) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyArray_BYTES(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":406 + * + * @property + * cdef inline char* data(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The pointer to the data buffer as a char*. + * This is provided for legacy reasons to avoid direct struct field access. + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":824 + * ctypedef long double complex clongdouble_t + * + * cdef inline object PyArray_MultiIterNew1(a): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(1, a) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew1(PyObject *__pyx_v_a) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew1", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":825 + * + * cdef inline object PyArray_MultiIterNew1(a): + * return PyArray_MultiIterNew(1, a) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew2(a, b): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(1, ((void *)__pyx_v_a)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 825, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":824 + * ctypedef long double complex clongdouble_t + * + * cdef inline object PyArray_MultiIterNew1(a): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(1, a) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew1", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":827 + * return PyArray_MultiIterNew(1, a) + * + * cdef inline object PyArray_MultiIterNew2(a, b): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(2, a, b) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew2(PyObject *__pyx_v_a, PyObject *__pyx_v_b) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew2", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":828 + * + * cdef inline object PyArray_MultiIterNew2(a, b): + * return PyArray_MultiIterNew(2, a, b) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(2, ((void *)__pyx_v_a), ((void *)__pyx_v_b)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 828, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":827 + * return PyArray_MultiIterNew(1, a) + * + * cdef inline object PyArray_MultiIterNew2(a, b): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(2, a, b) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew2", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":830 + * return PyArray_MultiIterNew(2, a, b) + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(3, a, b, c) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew3(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew3", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":831 + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): + * return PyArray_MultiIterNew(3, a, b, c) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(3, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 831, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":830 + * return PyArray_MultiIterNew(2, a, b) + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(3, a, b, c) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew3", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":833 + * return PyArray_MultiIterNew(3, a, b, c) + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(4, a, b, c, d) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew4(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew4", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":834 + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): + * return PyArray_MultiIterNew(4, a, b, c, d) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(4, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 834, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":833 + * return PyArray_MultiIterNew(3, a, b, c) + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(4, a, b, c, d) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew4", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":836 + * return PyArray_MultiIterNew(4, a, b, c, d) + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew5(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d, PyObject *__pyx_v_e) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew5", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":837 + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): + * return PyArray_MultiIterNew(5, a, b, c, d, e) # <<<<<<<<<<<<<< + * + * cdef inline tuple PyDataType_SHAPE(dtype d): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(5, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d), ((void *)__pyx_v_e)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 837, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":836 + * return PyArray_MultiIterNew(4, a, b, c, d) + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew5", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":839 + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + * cdef inline tuple PyDataType_SHAPE(dtype d): # <<<<<<<<<<<<<< + * if PyDataType_HASSUBARRAY(d): + * return d.subarray.shape + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyDataType_SHAPE(PyArray_Descr *__pyx_v_d) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2; + __Pyx_RefNannySetupContext("PyDataType_SHAPE", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":840 + * + * cdef inline tuple PyDataType_SHAPE(dtype d): + * if PyDataType_HASSUBARRAY(d): # <<<<<<<<<<<<<< + * return d.subarray.shape + * else: + */ + __pyx_t_1 = PyDataType_HASSUBARRAY(__pyx_v_d); + if (__pyx_t_1) { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":841 + * cdef inline tuple PyDataType_SHAPE(dtype d): + * if PyDataType_HASSUBARRAY(d): + * return d.subarray.shape # <<<<<<<<<<<<<< + * else: + * return () + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_f_5numpy_5dtype_8subarray_subarray(__pyx_v_d)->shape; + __Pyx_INCREF(((PyObject*)__pyx_t_2)); + __pyx_r = ((PyObject*)__pyx_t_2); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":840 + * + * cdef inline tuple PyDataType_SHAPE(dtype d): + * if PyDataType_HASSUBARRAY(d): # <<<<<<<<<<<<<< + * return d.subarray.shape + * else: + */ + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":843 + * return d.subarray.shape + * else: + * return () # <<<<<<<<<<<<<< + * + * + */ + /*else*/ { + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_empty_tuple); + __pyx_r = __pyx_empty_tuple; + goto __pyx_L0; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":839 + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + * cdef inline tuple PyDataType_SHAPE(dtype d): # <<<<<<<<<<<<<< + * if PyDataType_HASSUBARRAY(d): + * return d.subarray.shape + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1027 + * int _import_umath() except -1 + * + * cdef inline void set_array_base(ndarray arr, object base) except *: # <<<<<<<<<<<<<< + * Py_INCREF(base) # important to do this before stealing the reference below! + * PyArray_SetBaseObject(arr, base) + */ + +static CYTHON_INLINE void __pyx_f_5numpy_set_array_base(PyArrayObject *__pyx_v_arr, PyObject *__pyx_v_base) { + int __pyx_t_1; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1028 + * + * cdef inline void set_array_base(ndarray arr, object base) except *: + * Py_INCREF(base) # important to do this before stealing the reference below! # <<<<<<<<<<<<<< + * PyArray_SetBaseObject(arr, base) + * + */ + Py_INCREF(__pyx_v_base); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1029 + * cdef inline void set_array_base(ndarray arr, object base) except *: + * Py_INCREF(base) # important to do this before stealing the reference below! + * PyArray_SetBaseObject(arr, base) # <<<<<<<<<<<<<< + * + * cdef inline object get_array_base(ndarray arr): + */ + __pyx_t_1 = PyArray_SetBaseObject(__pyx_v_arr, __pyx_v_base); if (unlikely(__pyx_t_1 == ((int)-1))) __PYX_ERR(2, 1029, __pyx_L1_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1027 + * int _import_umath() except -1 + * + * cdef inline void set_array_base(ndarray arr, object base) except *: # <<<<<<<<<<<<<< + * Py_INCREF(base) # important to do this before stealing the reference below! + * PyArray_SetBaseObject(arr, base) + */ + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("numpy.set_array_base", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_L0:; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1031 + * PyArray_SetBaseObject(arr, base) + * + * cdef inline object get_array_base(ndarray arr): # <<<<<<<<<<<<<< + * base = PyArray_BASE(arr) + * if base is NULL: + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_get_array_base(PyArrayObject *__pyx_v_arr) { + PyObject *__pyx_v_base; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + __Pyx_RefNannySetupContext("get_array_base", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1032 + * + * cdef inline object get_array_base(ndarray arr): + * base = PyArray_BASE(arr) # <<<<<<<<<<<<<< + * if base is NULL: + * return None + */ + __pyx_v_base = PyArray_BASE(__pyx_v_arr); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1033 + * cdef inline object get_array_base(ndarray arr): + * base = PyArray_BASE(arr) + * if base is NULL: # <<<<<<<<<<<<<< + * return None + * return base + */ + __pyx_t_1 = (__pyx_v_base == NULL); + if (__pyx_t_1) { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1034 + * base = PyArray_BASE(arr) + * if base is NULL: + * return None # <<<<<<<<<<<<<< + * return base + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1033 + * cdef inline object get_array_base(ndarray arr): + * base = PyArray_BASE(arr) + * if base is NULL: # <<<<<<<<<<<<<< + * return None + * return base + */ + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1035 + * if base is NULL: + * return None + * return base # <<<<<<<<<<<<<< + * + * # Versions of the import_* functions which are more suitable for + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(((PyObject *)__pyx_v_base)); + __pyx_r = ((PyObject *)__pyx_v_base); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1031 + * PyArray_SetBaseObject(arr, base) + * + * cdef inline object get_array_base(ndarray arr): # <<<<<<<<<<<<<< + * base = PyArray_BASE(arr) + * if base is NULL: + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1039 + * # Versions of the import_* functions which are more suitable for + * # Cython code. + * cdef inline int import_array() except -1: # <<<<<<<<<<<<<< + * try: + * __pyx_import_array() + */ + +static CYTHON_INLINE int __pyx_f_5numpy_import_array(void) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("import_array", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1040 + * # Cython code. + * cdef inline int import_array() except -1: + * try: # <<<<<<<<<<<<<< + * __pyx_import_array() + * except Exception: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1041 + * cdef inline int import_array() except -1: + * try: + * __pyx_import_array() # <<<<<<<<<<<<<< + * except Exception: + * raise ImportError("numpy._core.multiarray failed to import") + */ + __pyx_t_4 = _import_array(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(2, 1041, __pyx_L3_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1040 + * # Cython code. + * cdef inline int import_array() except -1: + * try: # <<<<<<<<<<<<<< + * __pyx_import_array() + * except Exception: + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L8_try_end; + __pyx_L3_error:; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1042 + * try: + * __pyx_import_array() + * except Exception: # <<<<<<<<<<<<<< + * raise ImportError("numpy._core.multiarray failed to import") + * + */ + __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0]))); + if (__pyx_t_4) { + __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(2, 1042, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_7); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1043 + * __pyx_import_array() + * except Exception: + * raise ImportError("numpy._core.multiarray failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_umath() except -1: + */ + __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__9, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(2, 1043, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_Raise(__pyx_t_8, 0, 0, 0); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __PYX_ERR(2, 1043, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1040 + * # Cython code. + * cdef inline int import_array() except -1: + * try: # <<<<<<<<<<<<<< + * __pyx_import_array() + * except Exception: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L8_try_end:; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1039 + * # Versions of the import_* functions which are more suitable for + * # Cython code. + * cdef inline int import_array() except -1: # <<<<<<<<<<<<<< + * try: + * __pyx_import_array() + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1045 + * raise ImportError("numpy._core.multiarray failed to import") + * + * cdef inline int import_umath() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + +static CYTHON_INLINE int __pyx_f_5numpy_import_umath(void) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("import_umath", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1046 + * + * cdef inline int import_umath() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1047 + * cdef inline int import_umath() except -1: + * try: + * _import_umath() # <<<<<<<<<<<<<< + * except Exception: + * raise ImportError("numpy._core.umath failed to import") + */ + __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(2, 1047, __pyx_L3_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1046 + * + * cdef inline int import_umath() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L8_try_end; + __pyx_L3_error:; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1048 + * try: + * _import_umath() + * except Exception: # <<<<<<<<<<<<<< + * raise ImportError("numpy._core.umath failed to import") + * + */ + __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0]))); + if (__pyx_t_4) { + __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(2, 1048, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_7); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1049 + * _import_umath() + * except Exception: + * raise ImportError("numpy._core.umath failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_ufunc() except -1: + */ + __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__10, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(2, 1049, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_Raise(__pyx_t_8, 0, 0, 0); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __PYX_ERR(2, 1049, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1046 + * + * cdef inline int import_umath() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L8_try_end:; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1045 + * raise ImportError("numpy._core.multiarray failed to import") + * + * cdef inline int import_umath() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1051 + * raise ImportError("numpy._core.umath failed to import") + * + * cdef inline int import_ufunc() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + +static CYTHON_INLINE int __pyx_f_5numpy_import_ufunc(void) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("import_ufunc", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1052 + * + * cdef inline int import_ufunc() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1053 + * cdef inline int import_ufunc() except -1: + * try: + * _import_umath() # <<<<<<<<<<<<<< + * except Exception: + * raise ImportError("numpy._core.umath failed to import") + */ + __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(2, 1053, __pyx_L3_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1052 + * + * cdef inline int import_ufunc() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L8_try_end; + __pyx_L3_error:; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1054 + * try: + * _import_umath() + * except Exception: # <<<<<<<<<<<<<< + * raise ImportError("numpy._core.umath failed to import") + * + */ + __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0]))); + if (__pyx_t_4) { + __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(2, 1054, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_7); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1055 + * _import_umath() + * except Exception: + * raise ImportError("numpy._core.umath failed to import") # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__10, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(2, 1055, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_Raise(__pyx_t_8, 0, 0, 0); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __PYX_ERR(2, 1055, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1052 + * + * cdef inline int import_ufunc() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L8_try_end:; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1051 + * raise ImportError("numpy._core.umath failed to import") + * + * cdef inline int import_ufunc() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1058 + * + * + * cdef inline bint is_timedelta64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.timedelta64)` + */ + +static CYTHON_INLINE int __pyx_f_5numpy_is_timedelta64_object(PyObject *__pyx_v_obj) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1070 + * bool + * """ + * return PyObject_TypeCheck(obj, &PyTimedeltaArrType_Type) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyObject_TypeCheck(__pyx_v_obj, (&PyTimedeltaArrType_Type)); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1058 + * + * + * cdef inline bint is_timedelta64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.timedelta64)` + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1073 + * + * + * cdef inline bint is_datetime64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.datetime64)` + */ + +static CYTHON_INLINE int __pyx_f_5numpy_is_datetime64_object(PyObject *__pyx_v_obj) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1085 + * bool + * """ + * return PyObject_TypeCheck(obj, &PyDatetimeArrType_Type) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyObject_TypeCheck(__pyx_v_obj, (&PyDatetimeArrType_Type)); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1073 + * + * + * cdef inline bint is_datetime64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.datetime64)` + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1088 + * + * + * cdef inline npy_datetime get_datetime64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy datetime64 object + */ + +static CYTHON_INLINE npy_datetime __pyx_f_5numpy_get_datetime64_value(PyObject *__pyx_v_obj) { + npy_datetime __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1095 + * also needed. That can be found using `get_datetime64_unit`. + * """ + * return (obj).obval # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = ((PyDatetimeScalarObject *)__pyx_v_obj)->obval; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1088 + * + * + * cdef inline npy_datetime get_datetime64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy datetime64 object + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1098 + * + * + * cdef inline npy_timedelta get_timedelta64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy timedelta64 object + */ + +static CYTHON_INLINE npy_timedelta __pyx_f_5numpy_get_timedelta64_value(PyObject *__pyx_v_obj) { + npy_timedelta __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1102 + * returns the int64 value underlying scalar numpy timedelta64 object + * """ + * return (obj).obval # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = ((PyTimedeltaScalarObject *)__pyx_v_obj)->obval; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1098 + * + * + * cdef inline npy_timedelta get_timedelta64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy timedelta64 object + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1105 + * + * + * cdef inline NPY_DATETIMEUNIT get_datetime64_unit(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the unit part of the dtype for a numpy datetime64 object. + */ + +static CYTHON_INLINE NPY_DATETIMEUNIT __pyx_f_5numpy_get_datetime64_unit(PyObject *__pyx_v_obj) { + NPY_DATETIMEUNIT __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1109 + * returns the unit part of the dtype for a numpy datetime64 object. + * """ + * return (obj).obmeta.base # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = ((NPY_DATETIMEUNIT)((PyDatetimeScalarObject *)__pyx_v_obj)->obmeta.base); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1105 + * + * + * cdef inline NPY_DATETIMEUNIT get_datetime64_unit(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the unit part of the dtype for a numpy datetime64 object. + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "fairseq/data/data_utils_fast.pyx":16 + * + * + * cdef _is_batch_full(long num_sentences, long num_tokens, long max_tokens, long max_sentences): # <<<<<<<<<<<<<< + * if num_sentences == 0: + * return 0 + */ + +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast__is_batch_full(long __pyx_v_num_sentences, long __pyx_v_num_tokens, long __pyx_v_max_tokens, long __pyx_v_max_sentences) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + __Pyx_RefNannySetupContext("_is_batch_full", 1); + + /* "fairseq/data/data_utils_fast.pyx":17 + * + * cdef _is_batch_full(long num_sentences, long num_tokens, long max_tokens, long max_sentences): + * if num_sentences == 0: # <<<<<<<<<<<<<< + * return 0 + * if max_sentences > 0 and num_sentences == max_sentences: + */ + __pyx_t_1 = (__pyx_v_num_sentences == 0); + if (__pyx_t_1) { + + /* "fairseq/data/data_utils_fast.pyx":18 + * cdef _is_batch_full(long num_sentences, long num_tokens, long max_tokens, long max_sentences): + * if num_sentences == 0: + * return 0 # <<<<<<<<<<<<<< + * if max_sentences > 0 and num_sentences == max_sentences: + * return 1 + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_int_0); + __pyx_r = __pyx_int_0; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":17 + * + * cdef _is_batch_full(long num_sentences, long num_tokens, long max_tokens, long max_sentences): + * if num_sentences == 0: # <<<<<<<<<<<<<< + * return 0 + * if max_sentences > 0 and num_sentences == max_sentences: + */ + } + + /* "fairseq/data/data_utils_fast.pyx":19 + * if num_sentences == 0: + * return 0 + * if max_sentences > 0 and num_sentences == max_sentences: # <<<<<<<<<<<<<< + * return 1 + * if max_tokens > 0 and num_tokens > max_tokens: + */ + __pyx_t_2 = (__pyx_v_max_sentences > 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L5_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_num_sentences == __pyx_v_max_sentences); + __pyx_t_1 = __pyx_t_2; + __pyx_L5_bool_binop_done:; + if (__pyx_t_1) { + + /* "fairseq/data/data_utils_fast.pyx":20 + * return 0 + * if max_sentences > 0 and num_sentences == max_sentences: + * return 1 # <<<<<<<<<<<<<< + * if max_tokens > 0 and num_tokens > max_tokens: + * return 1 + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_int_1); + __pyx_r = __pyx_int_1; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":19 + * if num_sentences == 0: + * return 0 + * if max_sentences > 0 and num_sentences == max_sentences: # <<<<<<<<<<<<<< + * return 1 + * if max_tokens > 0 and num_tokens > max_tokens: + */ + } + + /* "fairseq/data/data_utils_fast.pyx":21 + * if max_sentences > 0 and num_sentences == max_sentences: + * return 1 + * if max_tokens > 0 and num_tokens > max_tokens: # <<<<<<<<<<<<<< + * return 1 + * return 0 + */ + __pyx_t_2 = (__pyx_v_max_tokens > 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L8_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_num_tokens > __pyx_v_max_tokens); + __pyx_t_1 = __pyx_t_2; + __pyx_L8_bool_binop_done:; + if (__pyx_t_1) { + + /* "fairseq/data/data_utils_fast.pyx":22 + * return 1 + * if max_tokens > 0 and num_tokens > max_tokens: + * return 1 # <<<<<<<<<<<<<< + * return 0 + * + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_int_1); + __pyx_r = __pyx_int_1; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":21 + * if max_sentences > 0 and num_sentences == max_sentences: + * return 1 + * if max_tokens > 0 and num_tokens > max_tokens: # <<<<<<<<<<<<<< + * return 1 + * return 0 + */ + } + + /* "fairseq/data/data_utils_fast.pyx":23 + * if max_tokens > 0 and num_tokens > max_tokens: + * return 1 + * return 0 # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_int_0); + __pyx_r = __pyx_int_0; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":16 + * + * + * cdef _is_batch_full(long num_sentences, long num_tokens, long max_tokens, long max_sentences): # <<<<<<<<<<<<<< + * if num_sentences == 0: + * return 0 + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/data_utils_fast.pyx":27 + * + * @cython.cdivision(True) + * cpdef list batch_by_size_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + +static PyObject *__pyx_pw_7fairseq_4data_15data_utils_fast_1batch_by_size_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast_batch_by_size_fast(PyArrayObject *__pyx_v_indices, PyObject *__pyx_v_num_tokens_fn, long __pyx_v_max_tokens, long __pyx_v_max_sentences, int __pyx_v_bsz_mult, CYTHON_UNUSED int __pyx_skip_dispatch) { + long __pyx_v_sample_len; + PyObject *__pyx_v_sample_lens = 0; + PyObject *__pyx_v_batch = 0; + PyObject *__pyx_v_batches = 0; + long __pyx_v_mod_len; + long __pyx_v_i; + long __pyx_v_idx; + long __pyx_v_num_tokens; + __Pyx_memviewslice __pyx_v_indices_view = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_LocalBuf_ND __pyx_pybuffernd_indices; + __Pyx_Buffer __pyx_pybuffer_indices; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + __Pyx_memviewslice __pyx_t_2 = { 0, 0, { 0 }, { 0 }, { 0 } }; + Py_ssize_t __pyx_t_3; + Py_ssize_t __pyx_t_4; + long __pyx_t_5; + Py_ssize_t __pyx_t_6; + int __pyx_t_7; + PyObject *__pyx_t_8 = NULL; + PyObject *__pyx_t_9 = NULL; + PyObject *__pyx_t_10 = NULL; + unsigned int __pyx_t_11; + long __pyx_t_12; + int __pyx_t_13; + long __pyx_t_14; + long __pyx_t_15; + int __pyx_t_16; + int __pyx_t_17; + PyObject *__pyx_t_18 = NULL; + PyObject *__pyx_t_19 = NULL; + Py_ssize_t __pyx_t_20; + Py_ssize_t __pyx_t_21; + Py_ssize_t __pyx_t_22; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("batch_by_size_fast", 1); + __pyx_pybuffer_indices.pybuffer.buf = NULL; + __pyx_pybuffer_indices.refcount = 0; + __pyx_pybuffernd_indices.data = NULL; + __pyx_pybuffernd_indices.rcbuffer = &__pyx_pybuffer_indices; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 27, __pyx_L1_error) + } + __pyx_pybuffernd_indices.diminfo[0].strides = __pyx_pybuffernd_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_indices.diminfo[0].shape = __pyx_pybuffernd_indices.rcbuffer->pybuffer.shape[0]; + + /* "fairseq/data/data_utils_fast.pyx":34 + * int bsz_mult, + * ): + * cdef long sample_len = 0 # <<<<<<<<<<<<<< + * cdef list sample_lens = [] + * cdef list batch = [] + */ + __pyx_v_sample_len = 0; + + /* "fairseq/data/data_utils_fast.pyx":35 + * ): + * cdef long sample_len = 0 + * cdef list sample_lens = [] # <<<<<<<<<<<<<< + * cdef list batch = [] + * cdef list batches = [] + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 35, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_sample_lens = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":36 + * cdef long sample_len = 0 + * cdef list sample_lens = [] + * cdef list batch = [] # <<<<<<<<<<<<<< + * cdef list batches = [] + * cdef long mod_len + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 36, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_batch = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":37 + * cdef list sample_lens = [] + * cdef list batch = [] + * cdef list batches = [] # <<<<<<<<<<<<<< + * cdef long mod_len + * cdef long i + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 37, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_batches = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":42 + * cdef long idx + * cdef long num_tokens + * cdef DTYPE_t[:] indices_view = indices # <<<<<<<<<<<<<< + * + * for i in range(len(indices_view)): + */ + __pyx_t_2 = __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(((PyObject *)__pyx_v_indices), PyBUF_WRITABLE); if (unlikely(!__pyx_t_2.memview)) __PYX_ERR(0, 42, __pyx_L1_error) + __pyx_v_indices_view = __pyx_t_2; + __pyx_t_2.memview = NULL; + __pyx_t_2.data = NULL; + + /* "fairseq/data/data_utils_fast.pyx":44 + * cdef DTYPE_t[:] indices_view = indices + * + * for i in range(len(indices_view)): # <<<<<<<<<<<<<< + * idx = indices_view[i] + * num_tokens = num_tokens_fn(idx) + */ + __pyx_t_3 = __Pyx_MemoryView_Len(__pyx_v_indices_view); + __pyx_t_4 = __pyx_t_3; + for (__pyx_t_5 = 0; __pyx_t_5 < __pyx_t_4; __pyx_t_5+=1) { + __pyx_v_i = __pyx_t_5; + + /* "fairseq/data/data_utils_fast.pyx":45 + * + * for i in range(len(indices_view)): + * idx = indices_view[i] # <<<<<<<<<<<<<< + * num_tokens = num_tokens_fn(idx) + * sample_lens.append(num_tokens) + */ + __pyx_t_6 = __pyx_v_i; + __pyx_t_7 = -1; + if (__pyx_t_6 < 0) { + __pyx_t_6 += __pyx_v_indices_view.shape[0]; + if (unlikely(__pyx_t_6 < 0)) __pyx_t_7 = 0; + } else if (unlikely(__pyx_t_6 >= __pyx_v_indices_view.shape[0])) __pyx_t_7 = 0; + if (unlikely(__pyx_t_7 != -1)) { + __Pyx_RaiseBufferIndexError(__pyx_t_7); + __PYX_ERR(0, 45, __pyx_L1_error) + } + __pyx_v_idx = (*((__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_indices_view.data + __pyx_t_6 * __pyx_v_indices_view.strides[0]) ))); + + /* "fairseq/data/data_utils_fast.pyx":46 + * for i in range(len(indices_view)): + * idx = indices_view[i] + * num_tokens = num_tokens_fn(idx) # <<<<<<<<<<<<<< + * sample_lens.append(num_tokens) + * sample_len = max(sample_len, num_tokens) + */ + __pyx_t_8 = __Pyx_PyInt_From_long(__pyx_v_idx); if (unlikely(!__pyx_t_8)) __PYX_ERR(0, 46, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_INCREF(__pyx_v_num_tokens_fn); + __pyx_t_9 = __pyx_v_num_tokens_fn; __pyx_t_10 = NULL; + __pyx_t_11 = 0; + #if CYTHON_UNPACK_METHODS + if (unlikely(PyMethod_Check(__pyx_t_9))) { + __pyx_t_10 = PyMethod_GET_SELF(__pyx_t_9); + if (likely(__pyx_t_10)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_9); + __Pyx_INCREF(__pyx_t_10); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_9, function); + __pyx_t_11 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_10, __pyx_t_8}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_9, __pyx_callargs+1-__pyx_t_11, 1+__pyx_t_11); + __Pyx_XDECREF(__pyx_t_10); __pyx_t_10 = 0; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 46, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + } + __pyx_t_12 = __Pyx_PyInt_As_long(__pyx_t_1); if (unlikely((__pyx_t_12 == (long)-1) && PyErr_Occurred())) __PYX_ERR(0, 46, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_v_num_tokens = __pyx_t_12; + + /* "fairseq/data/data_utils_fast.pyx":47 + * idx = indices_view[i] + * num_tokens = num_tokens_fn(idx) + * sample_lens.append(num_tokens) # <<<<<<<<<<<<<< + * sample_len = max(sample_len, num_tokens) + * + */ + __pyx_t_1 = __Pyx_PyInt_From_long(__pyx_v_num_tokens); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 47, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_sample_lens, __pyx_t_1); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 47, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":48 + * num_tokens = num_tokens_fn(idx) + * sample_lens.append(num_tokens) + * sample_len = max(sample_len, num_tokens) # <<<<<<<<<<<<<< + * + * assert max_tokens <= 0 or sample_len <= max_tokens, ( + */ + __pyx_t_12 = __pyx_v_num_tokens; + __pyx_t_14 = __pyx_v_sample_len; + __pyx_t_16 = (__pyx_t_12 > __pyx_t_14); + if (__pyx_t_16) { + __pyx_t_15 = __pyx_t_12; + } else { + __pyx_t_15 = __pyx_t_14; + } + __pyx_v_sample_len = __pyx_t_15; + + /* "fairseq/data/data_utils_fast.pyx":50 + * sample_len = max(sample_len, num_tokens) + * + * assert max_tokens <= 0 or sample_len <= max_tokens, ( # <<<<<<<<<<<<<< + * "sentence at index {} of size {} exceeds max_tokens " + * "limit of {}!".format(idx, sample_len, max_tokens) + */ + #ifndef CYTHON_WITHOUT_ASSERTIONS + if (unlikely(__pyx_assertions_enabled())) { + __pyx_t_17 = (__pyx_v_max_tokens <= 0); + if (!__pyx_t_17) { + } else { + __pyx_t_16 = __pyx_t_17; + goto __pyx_L5_bool_binop_done; + } + __pyx_t_17 = (__pyx_v_sample_len <= __pyx_v_max_tokens); + __pyx_t_16 = __pyx_t_17; + __pyx_L5_bool_binop_done:; + if (unlikely(!__pyx_t_16)) { + + /* "fairseq/data/data_utils_fast.pyx":52 + * assert max_tokens <= 0 or sample_len <= max_tokens, ( + * "sentence at index {} of size {} exceeds max_tokens " + * "limit of {}!".format(idx, sample_len, max_tokens) # <<<<<<<<<<<<<< + * ) + * num_tokens = (len(batch) + 1) * sample_len + */ + __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_kp_u_sentence_at_index_of_size_exceed, __pyx_n_s_format); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 52, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __pyx_t_8 = __Pyx_PyInt_From_long(__pyx_v_idx); if (unlikely(!__pyx_t_8)) __PYX_ERR(0, 52, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_10 = __Pyx_PyInt_From_long(__pyx_v_sample_len); if (unlikely(!__pyx_t_10)) __PYX_ERR(0, 52, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_10); + __pyx_t_18 = __Pyx_PyInt_From_long(__pyx_v_max_tokens); if (unlikely(!__pyx_t_18)) __PYX_ERR(0, 52, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_18); + __pyx_t_19 = NULL; + __pyx_t_11 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_9))) { + __pyx_t_19 = PyMethod_GET_SELF(__pyx_t_9); + if (likely(__pyx_t_19)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_9); + __Pyx_INCREF(__pyx_t_19); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_9, function); + __pyx_t_11 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[4] = {__pyx_t_19, __pyx_t_8, __pyx_t_10, __pyx_t_18}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_9, __pyx_callargs+1-__pyx_t_11, 3+__pyx_t_11); + __Pyx_XDECREF(__pyx_t_19); __pyx_t_19 = 0; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0; + __Pyx_DECREF(__pyx_t_18); __pyx_t_18 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 52, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + } + __pyx_t_9 = PyTuple_Pack(1, __pyx_t_1); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 52, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_Raise(__pyx_builtin_AssertionError, __pyx_t_9, 0, 0); + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + __PYX_ERR(0, 50, __pyx_L1_error) + } + } + #else + if ((1)); else __PYX_ERR(0, 50, __pyx_L1_error) + #endif + + /* "fairseq/data/data_utils_fast.pyx":54 + * "limit of {}!".format(idx, sample_len, max_tokens) + * ) + * num_tokens = (len(batch) + 1) * sample_len # <<<<<<<<<<<<<< + * + * if _is_batch_full(len(batch), num_tokens, max_tokens, max_sentences): + */ + __pyx_t_20 = __Pyx_PyList_GET_SIZE(__pyx_v_batch); if (unlikely(__pyx_t_20 == ((Py_ssize_t)-1))) __PYX_ERR(0, 54, __pyx_L1_error) + __pyx_v_num_tokens = ((__pyx_t_20 + 1) * __pyx_v_sample_len); + + /* "fairseq/data/data_utils_fast.pyx":56 + * num_tokens = (len(batch) + 1) * sample_len + * + * if _is_batch_full(len(batch), num_tokens, max_tokens, max_sentences): # <<<<<<<<<<<<<< + * mod_len = max( + * bsz_mult * (len(batch) // bsz_mult), + */ + __pyx_t_20 = __Pyx_PyList_GET_SIZE(__pyx_v_batch); if (unlikely(__pyx_t_20 == ((Py_ssize_t)-1))) __PYX_ERR(0, 56, __pyx_L1_error) + __pyx_t_9 = __pyx_f_7fairseq_4data_15data_utils_fast__is_batch_full(__pyx_t_20, __pyx_v_num_tokens, __pyx_v_max_tokens, __pyx_v_max_sentences); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 56, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __pyx_t_16 = __Pyx_PyObject_IsTrue(__pyx_t_9); if (unlikely((__pyx_t_16 < 0))) __PYX_ERR(0, 56, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + if (__pyx_t_16) { + + /* "fairseq/data/data_utils_fast.pyx":59 + * mod_len = max( + * bsz_mult * (len(batch) // bsz_mult), + * len(batch) % bsz_mult, # <<<<<<<<<<<<<< + * ) + * batches.append(batch[:mod_len]) + */ + __pyx_t_20 = __Pyx_PyList_GET_SIZE(__pyx_v_batch); if (unlikely(__pyx_t_20 == ((Py_ssize_t)-1))) __PYX_ERR(0, 59, __pyx_L1_error) + __pyx_t_21 = (__pyx_t_20 % __pyx_v_bsz_mult); + + /* "fairseq/data/data_utils_fast.pyx":58 + * if _is_batch_full(len(batch), num_tokens, max_tokens, max_sentences): + * mod_len = max( + * bsz_mult * (len(batch) // bsz_mult), # <<<<<<<<<<<<<< + * len(batch) % bsz_mult, + * ) + */ + __pyx_t_20 = __Pyx_PyList_GET_SIZE(__pyx_v_batch); if (unlikely(__pyx_t_20 == ((Py_ssize_t)-1))) __PYX_ERR(0, 58, __pyx_L1_error) + __pyx_t_22 = (__pyx_v_bsz_mult * (__pyx_t_20 / __pyx_v_bsz_mult)); + + /* "fairseq/data/data_utils_fast.pyx":59 + * mod_len = max( + * bsz_mult * (len(batch) // bsz_mult), + * len(batch) % bsz_mult, # <<<<<<<<<<<<<< + * ) + * batches.append(batch[:mod_len]) + */ + __pyx_t_16 = (__pyx_t_21 > __pyx_t_22); + if (__pyx_t_16) { + __pyx_t_20 = __pyx_t_21; + } else { + __pyx_t_20 = __pyx_t_22; + } + __pyx_v_mod_len = __pyx_t_20; + + /* "fairseq/data/data_utils_fast.pyx":61 + * len(batch) % bsz_mult, + * ) + * batches.append(batch[:mod_len]) # <<<<<<<<<<<<<< + * batch = batch[mod_len:] + * sample_lens = sample_lens[mod_len:] + */ + __pyx_t_9 = __Pyx_PyList_GetSlice(__pyx_v_batch, 0, __pyx_v_mod_len); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 61, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_batches, __pyx_t_9); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 61, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + + /* "fairseq/data/data_utils_fast.pyx":62 + * ) + * batches.append(batch[:mod_len]) + * batch = batch[mod_len:] # <<<<<<<<<<<<<< + * sample_lens = sample_lens[mod_len:] + * sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 + */ + __pyx_t_9 = __Pyx_PyList_GetSlice(__pyx_v_batch, __pyx_v_mod_len, PY_SSIZE_T_MAX); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 62, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __Pyx_DECREF_SET(__pyx_v_batch, ((PyObject*)__pyx_t_9)); + __pyx_t_9 = 0; + + /* "fairseq/data/data_utils_fast.pyx":63 + * batches.append(batch[:mod_len]) + * batch = batch[mod_len:] + * sample_lens = sample_lens[mod_len:] # <<<<<<<<<<<<<< + * sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 + * batch.append(idx) + */ + __pyx_t_9 = __Pyx_PyList_GetSlice(__pyx_v_sample_lens, __pyx_v_mod_len, PY_SSIZE_T_MAX); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 63, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __Pyx_DECREF_SET(__pyx_v_sample_lens, ((PyObject*)__pyx_t_9)); + __pyx_t_9 = 0; + + /* "fairseq/data/data_utils_fast.pyx":64 + * batch = batch[mod_len:] + * sample_lens = sample_lens[mod_len:] + * sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 # <<<<<<<<<<<<<< + * batch.append(idx) + * if len(batch) > 0: + */ + __pyx_t_20 = __Pyx_PyList_GET_SIZE(__pyx_v_sample_lens); if (unlikely(__pyx_t_20 == ((Py_ssize_t)-1))) __PYX_ERR(0, 64, __pyx_L1_error) + __pyx_t_16 = (__pyx_t_20 > 0); + if (__pyx_t_16) { + __pyx_t_9 = __Pyx_PyObject_CallOneArg(__pyx_builtin_max, __pyx_v_sample_lens); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 64, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __pyx_t_12 = __Pyx_PyInt_As_long(__pyx_t_9); if (unlikely((__pyx_t_12 == (long)-1) && PyErr_Occurred())) __PYX_ERR(0, 64, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + __pyx_t_15 = __pyx_t_12; + } else { + __pyx_t_15 = 0; + } + __pyx_v_sample_len = __pyx_t_15; + + /* "fairseq/data/data_utils_fast.pyx":56 + * num_tokens = (len(batch) + 1) * sample_len + * + * if _is_batch_full(len(batch), num_tokens, max_tokens, max_sentences): # <<<<<<<<<<<<<< + * mod_len = max( + * bsz_mult * (len(batch) // bsz_mult), + */ + } + + /* "fairseq/data/data_utils_fast.pyx":65 + * sample_lens = sample_lens[mod_len:] + * sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 + * batch.append(idx) # <<<<<<<<<<<<<< + * if len(batch) > 0: + * batches.append(batch) + */ + __pyx_t_9 = __Pyx_PyInt_From_long(__pyx_v_idx); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 65, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_batch, __pyx_t_9); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 65, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + } + + /* "fairseq/data/data_utils_fast.pyx":66 + * sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 + * batch.append(idx) + * if len(batch) > 0: # <<<<<<<<<<<<<< + * batches.append(batch) + * return batches + */ + __pyx_t_3 = __Pyx_PyList_GET_SIZE(__pyx_v_batch); if (unlikely(__pyx_t_3 == ((Py_ssize_t)-1))) __PYX_ERR(0, 66, __pyx_L1_error) + __pyx_t_16 = (__pyx_t_3 > 0); + if (__pyx_t_16) { + + /* "fairseq/data/data_utils_fast.pyx":67 + * batch.append(idx) + * if len(batch) > 0: + * batches.append(batch) # <<<<<<<<<<<<<< + * return batches + * + */ + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_batches, __pyx_v_batch); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 67, __pyx_L1_error) + + /* "fairseq/data/data_utils_fast.pyx":66 + * sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 + * batch.append(idx) + * if len(batch) > 0: # <<<<<<<<<<<<<< + * batches.append(batch) + * return batches + */ + } + + /* "fairseq/data/data_utils_fast.pyx":68 + * if len(batch) > 0: + * batches.append(batch) + * return batches # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_batches); + __pyx_r = __pyx_v_batches; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":27 + * + * @cython.cdivision(True) + * cpdef list batch_by_size_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __PYX_XCLEAR_MEMVIEW(&__pyx_t_2, 1); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_XDECREF(__pyx_t_9); + __Pyx_XDECREF(__pyx_t_10); + __Pyx_XDECREF(__pyx_t_18); + __Pyx_XDECREF(__pyx_t_19); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.data_utils_fast.batch_by_size_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XDECREF(__pyx_v_sample_lens); + __Pyx_XDECREF(__pyx_v_batch); + __Pyx_XDECREF(__pyx_v_batches); + __PYX_XCLEAR_MEMVIEW(&__pyx_v_indices_view, 1); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* Python wrapper */ +static PyObject *__pyx_pw_7fairseq_4data_15data_utils_fast_1batch_by_size_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_7fairseq_4data_15data_utils_fast_1batch_by_size_fast = {"batch_by_size_fast", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_15data_utils_fast_1batch_by_size_fast, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_7fairseq_4data_15data_utils_fast_1batch_by_size_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyArrayObject *__pyx_v_indices = 0; + PyObject *__pyx_v_num_tokens_fn = 0; + long __pyx_v_max_tokens; + long __pyx_v_max_sentences; + int __pyx_v_bsz_mult; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[5] = {0,0,0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("batch_by_size_fast (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_indices,&__pyx_n_s_num_tokens_fn,&__pyx_n_s_max_tokens,&__pyx_n_s_max_sentences,&__pyx_n_s_bsz_mult,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 5: values[4] = __Pyx_Arg_FASTCALL(__pyx_args, 4); + CYTHON_FALLTHROUGH; + case 4: values[3] = __Pyx_Arg_FASTCALL(__pyx_args, 3); + CYTHON_FALLTHROUGH; + case 3: values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_indices)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 27, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_num_tokens_fn)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 27, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("batch_by_size_fast", 1, 5, 5, 1); __PYX_ERR(0, 27, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_max_tokens)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 27, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("batch_by_size_fast", 1, 5, 5, 2); __PYX_ERR(0, 27, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 3: + if (likely((values[3] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_max_sentences)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[3]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 27, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("batch_by_size_fast", 1, 5, 5, 3); __PYX_ERR(0, 27, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 4: + if (likely((values[4] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_bsz_mult)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[4]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 27, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("batch_by_size_fast", 1, 5, 5, 4); __PYX_ERR(0, 27, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "batch_by_size_fast") < 0)) __PYX_ERR(0, 27, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 5)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + values[3] = __Pyx_Arg_FASTCALL(__pyx_args, 3); + values[4] = __Pyx_Arg_FASTCALL(__pyx_args, 4); + } + __pyx_v_indices = ((PyArrayObject *)values[0]); + __pyx_v_num_tokens_fn = values[1]; + __pyx_v_max_tokens = __Pyx_PyInt_As_long(values[2]); if (unlikely((__pyx_v_max_tokens == (long)-1) && PyErr_Occurred())) __PYX_ERR(0, 30, __pyx_L3_error) + __pyx_v_max_sentences = __Pyx_PyInt_As_long(values[3]); if (unlikely((__pyx_v_max_sentences == (long)-1) && PyErr_Occurred())) __PYX_ERR(0, 31, __pyx_L3_error) + __pyx_v_bsz_mult = __Pyx_PyInt_As_int(values[4]); if (unlikely((__pyx_v_bsz_mult == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 32, __pyx_L3_error) + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("batch_by_size_fast", 1, 5, 5, __pyx_nargs); __PYX_ERR(0, 27, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("fairseq.data.data_utils_fast.batch_by_size_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_indices), __pyx_ptype_5numpy_ndarray, 1, "indices", 0))) __PYX_ERR(0, 28, __pyx_L1_error) + __pyx_r = __pyx_pf_7fairseq_4data_15data_utils_fast_batch_by_size_fast(__pyx_self, __pyx_v_indices, __pyx_v_num_tokens_fn, __pyx_v_max_tokens, __pyx_v_max_sentences, __pyx_v_bsz_mult); + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __pyx_r = NULL; + __pyx_L0:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_7fairseq_4data_15data_utils_fast_batch_by_size_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_indices, PyObject *__pyx_v_num_tokens_fn, long __pyx_v_max_tokens, long __pyx_v_max_sentences, int __pyx_v_bsz_mult) { + __Pyx_LocalBuf_ND __pyx_pybuffernd_indices; + __Pyx_Buffer __pyx_pybuffer_indices; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("batch_by_size_fast", 1); + __pyx_pybuffer_indices.pybuffer.buf = NULL; + __pyx_pybuffer_indices.refcount = 0; + __pyx_pybuffernd_indices.data = NULL; + __pyx_pybuffernd_indices.rcbuffer = &__pyx_pybuffer_indices; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 27, __pyx_L1_error) + } + __pyx_pybuffernd_indices.diminfo[0].strides = __pyx_pybuffernd_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_indices.diminfo[0].shape = __pyx_pybuffernd_indices.rcbuffer->pybuffer.shape[0]; + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __pyx_f_7fairseq_4data_15data_utils_fast_batch_by_size_fast(__pyx_v_indices, __pyx_v_num_tokens_fn, __pyx_v_max_tokens, __pyx_v_max_sentences, __pyx_v_bsz_mult, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 27, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.data_utils_fast.batch_by_size_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/data_utils_fast.pyx":71 + * + * + * cdef _find_valid_shape( # <<<<<<<<<<<<<< + * DTYPE_t[:, :] shapes_view, + * long num_sentences, + */ + +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast__find_valid_shape(__Pyx_memviewslice __pyx_v_shapes_view, long __pyx_v_num_sentences, long __pyx_v_num_tokens) { + Py_ssize_t __pyx_v_i; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + Py_ssize_t __pyx_t_1; + Py_ssize_t __pyx_t_2; + Py_ssize_t __pyx_t_3; + int __pyx_t_4; + Py_ssize_t __pyx_t_5; + Py_ssize_t __pyx_t_6; + int __pyx_t_7; + int __pyx_t_8; + PyObject *__pyx_t_9 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_find_valid_shape", 1); + + /* "fairseq/data/data_utils_fast.pyx":77 + * ): + * """Return index of first valid shape of -1 if none is found.""" + * for i in range(shapes_view.shape[0]): # <<<<<<<<<<<<<< + * if num_sentences <= shapes_view[i][0] and num_tokens <= shapes_view[i][1]: + * return i + */ + __pyx_t_1 = (__pyx_v_shapes_view.shape[0]); + __pyx_t_2 = __pyx_t_1; + for (__pyx_t_3 = 0; __pyx_t_3 < __pyx_t_2; __pyx_t_3+=1) { + __pyx_v_i = __pyx_t_3; + + /* "fairseq/data/data_utils_fast.pyx":78 + * """Return index of first valid shape of -1 if none is found.""" + * for i in range(shapes_view.shape[0]): + * if num_sentences <= shapes_view[i][0] and num_tokens <= shapes_view[i][1]: # <<<<<<<<<<<<<< + * return i + * return -1 + */ + __pyx_t_5 = __pyx_v_i; + __pyx_t_6 = 0; + __pyx_t_7 = -1; + if (__pyx_t_5 < 0) { + __pyx_t_5 += __pyx_v_shapes_view.shape[0]; + if (unlikely(__pyx_t_5 < 0)) __pyx_t_7 = 0; + } else if (unlikely(__pyx_t_5 >= __pyx_v_shapes_view.shape[0])) __pyx_t_7 = 0; + if (__pyx_t_6 < 0) { + __pyx_t_6 += __pyx_v_shapes_view.shape[1]; + if (unlikely(__pyx_t_6 < 0)) __pyx_t_7 = 1; + } else if (unlikely(__pyx_t_6 >= __pyx_v_shapes_view.shape[1])) __pyx_t_7 = 1; + if (unlikely(__pyx_t_7 != -1)) { + __Pyx_RaiseBufferIndexError(__pyx_t_7); + __PYX_ERR(0, 78, __pyx_L1_error) + } + __pyx_t_8 = (__pyx_v_num_sentences <= (*((__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_shapes_view.data + __pyx_t_5 * __pyx_v_shapes_view.strides[0]) ) + __pyx_t_6 * __pyx_v_shapes_view.strides[1]) )))); + if (__pyx_t_8) { + } else { + __pyx_t_4 = __pyx_t_8; + goto __pyx_L6_bool_binop_done; + } + __pyx_t_6 = __pyx_v_i; + __pyx_t_5 = 1; + __pyx_t_7 = -1; + if (__pyx_t_6 < 0) { + __pyx_t_6 += __pyx_v_shapes_view.shape[0]; + if (unlikely(__pyx_t_6 < 0)) __pyx_t_7 = 0; + } else if (unlikely(__pyx_t_6 >= __pyx_v_shapes_view.shape[0])) __pyx_t_7 = 0; + if (__pyx_t_5 < 0) { + __pyx_t_5 += __pyx_v_shapes_view.shape[1]; + if (unlikely(__pyx_t_5 < 0)) __pyx_t_7 = 1; + } else if (unlikely(__pyx_t_5 >= __pyx_v_shapes_view.shape[1])) __pyx_t_7 = 1; + if (unlikely(__pyx_t_7 != -1)) { + __Pyx_RaiseBufferIndexError(__pyx_t_7); + __PYX_ERR(0, 78, __pyx_L1_error) + } + __pyx_t_8 = (__pyx_v_num_tokens <= (*((__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_shapes_view.data + __pyx_t_6 * __pyx_v_shapes_view.strides[0]) ) + __pyx_t_5 * __pyx_v_shapes_view.strides[1]) )))); + __pyx_t_4 = __pyx_t_8; + __pyx_L6_bool_binop_done:; + if (__pyx_t_4) { + + /* "fairseq/data/data_utils_fast.pyx":79 + * for i in range(shapes_view.shape[0]): + * if num_sentences <= shapes_view[i][0] and num_tokens <= shapes_view[i][1]: + * return i # <<<<<<<<<<<<<< + * return -1 + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_9 = PyInt_FromSsize_t(__pyx_v_i); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 79, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __pyx_r = __pyx_t_9; + __pyx_t_9 = 0; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":78 + * """Return index of first valid shape of -1 if none is found.""" + * for i in range(shapes_view.shape[0]): + * if num_sentences <= shapes_view[i][0] and num_tokens <= shapes_view[i][1]: # <<<<<<<<<<<<<< + * return i + * return -1 + */ + } + } + + /* "fairseq/data/data_utils_fast.pyx":80 + * if num_sentences <= shapes_view[i][0] and num_tokens <= shapes_view[i][1]: + * return i + * return -1 # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_int_neg_1); + __pyx_r = __pyx_int_neg_1; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":71 + * + * + * cdef _find_valid_shape( # <<<<<<<<<<<<<< + * DTYPE_t[:, :] shapes_view, + * long num_sentences, + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_9); + __Pyx_AddTraceback("fairseq.data.data_utils_fast._find_valid_shape", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/data_utils_fast.pyx":84 + * + * @cython.cdivision(True) + * cpdef list batch_fixed_shapes_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + +static PyObject *__pyx_pw_7fairseq_4data_15data_utils_fast_3batch_fixed_shapes_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_f_7fairseq_4data_15data_utils_fast_batch_fixed_shapes_fast(PyArrayObject *__pyx_v_indices, PyObject *__pyx_v_num_tokens_fn, PyArrayObject *__pyx_v_fixed_shapes_sorted, CYTHON_UNUSED int __pyx_skip_dispatch) { + long __pyx_v_sample_len; + PyObject *__pyx_v_sample_lens = 0; + PyObject *__pyx_v_batch = 0; + PyObject *__pyx_v_batches = 0; + long __pyx_v_i; + long __pyx_v_idx; + long __pyx_v_num_tokens; + __Pyx_memviewslice __pyx_v_indices_view = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_memviewslice __pyx_v_shapes_view = { 0, 0, { 0 }, { 0 }, { 0 } }; + PyObject *__pyx_v_shape_idx = NULL; + __Pyx_LocalBuf_ND __pyx_pybuffernd_fixed_shapes_sorted; + __Pyx_Buffer __pyx_pybuffer_fixed_shapes_sorted; + __Pyx_LocalBuf_ND __pyx_pybuffernd_indices; + __Pyx_Buffer __pyx_pybuffer_indices; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + __Pyx_memviewslice __pyx_t_2 = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_memviewslice __pyx_t_3 = { 0, 0, { 0 }, { 0 }, { 0 } }; + Py_ssize_t __pyx_t_4; + Py_ssize_t __pyx_t_5; + long __pyx_t_6; + Py_ssize_t __pyx_t_7; + int __pyx_t_8; + PyObject *__pyx_t_9 = NULL; + PyObject *__pyx_t_10 = NULL; + PyObject *__pyx_t_11 = NULL; + unsigned int __pyx_t_12; + long __pyx_t_13; + int __pyx_t_14; + long __pyx_t_15; + long __pyx_t_16; + int __pyx_t_17; + Py_ssize_t __pyx_t_18; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("batch_fixed_shapes_fast", 1); + __pyx_pybuffer_indices.pybuffer.buf = NULL; + __pyx_pybuffer_indices.refcount = 0; + __pyx_pybuffernd_indices.data = NULL; + __pyx_pybuffernd_indices.rcbuffer = &__pyx_pybuffer_indices; + __pyx_pybuffer_fixed_shapes_sorted.pybuffer.buf = NULL; + __pyx_pybuffer_fixed_shapes_sorted.refcount = 0; + __pyx_pybuffernd_fixed_shapes_sorted.data = NULL; + __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer = &__pyx_pybuffer_fixed_shapes_sorted; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 84, __pyx_L1_error) + } + __pyx_pybuffernd_indices.diminfo[0].strides = __pyx_pybuffernd_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_indices.diminfo[0].shape = __pyx_pybuffernd_indices.rcbuffer->pybuffer.shape[0]; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer, (PyObject*)__pyx_v_fixed_shapes_sorted, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 84, __pyx_L1_error) + } + __pyx_pybuffernd_fixed_shapes_sorted.diminfo[0].strides = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_fixed_shapes_sorted.diminfo[0].shape = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_fixed_shapes_sorted.diminfo[1].strides = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_fixed_shapes_sorted.diminfo[1].shape = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.shape[1]; + + /* "fairseq/data/data_utils_fast.pyx":89 + * np.ndarray[DTYPE_t, ndim=2] fixed_shapes_sorted, + * ): + * cdef long sample_len = 0 # <<<<<<<<<<<<<< + * cdef list sample_lens = [] + * cdef list batch = [] + */ + __pyx_v_sample_len = 0; + + /* "fairseq/data/data_utils_fast.pyx":90 + * ): + * cdef long sample_len = 0 + * cdef list sample_lens = [] # <<<<<<<<<<<<<< + * cdef list batch = [] + * cdef list batches = [] + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 90, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_sample_lens = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":91 + * cdef long sample_len = 0 + * cdef list sample_lens = [] + * cdef list batch = [] # <<<<<<<<<<<<<< + * cdef list batches = [] + * cdef long mod_len + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 91, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_batch = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":92 + * cdef list sample_lens = [] + * cdef list batch = [] + * cdef list batches = [] # <<<<<<<<<<<<<< + * cdef long mod_len + * cdef long i + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 92, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_batches = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":97 + * cdef long idx + * cdef long num_tokens + * cdef DTYPE_t[:] indices_view = indices # <<<<<<<<<<<<<< + * cdef DTYPE_t[:, :] shapes_view = fixed_shapes_sorted + * + */ + __pyx_t_2 = __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(((PyObject *)__pyx_v_indices), PyBUF_WRITABLE); if (unlikely(!__pyx_t_2.memview)) __PYX_ERR(0, 97, __pyx_L1_error) + __pyx_v_indices_view = __pyx_t_2; + __pyx_t_2.memview = NULL; + __pyx_t_2.data = NULL; + + /* "fairseq/data/data_utils_fast.pyx":98 + * cdef long num_tokens + * cdef DTYPE_t[:] indices_view = indices + * cdef DTYPE_t[:, :] shapes_view = fixed_shapes_sorted # <<<<<<<<<<<<<< + * + * for i in range(len(indices_view)): + */ + __pyx_t_3 = __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(((PyObject *)__pyx_v_fixed_shapes_sorted), PyBUF_WRITABLE); if (unlikely(!__pyx_t_3.memview)) __PYX_ERR(0, 98, __pyx_L1_error) + __pyx_v_shapes_view = __pyx_t_3; + __pyx_t_3.memview = NULL; + __pyx_t_3.data = NULL; + + /* "fairseq/data/data_utils_fast.pyx":100 + * cdef DTYPE_t[:, :] shapes_view = fixed_shapes_sorted + * + * for i in range(len(indices_view)): # <<<<<<<<<<<<<< + * idx = indices_view[i] + * num_tokens = num_tokens_fn(idx) + */ + __pyx_t_4 = __Pyx_MemoryView_Len(__pyx_v_indices_view); + __pyx_t_5 = __pyx_t_4; + for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) { + __pyx_v_i = __pyx_t_6; + + /* "fairseq/data/data_utils_fast.pyx":101 + * + * for i in range(len(indices_view)): + * idx = indices_view[i] # <<<<<<<<<<<<<< + * num_tokens = num_tokens_fn(idx) + * sample_lens.append(num_tokens) + */ + __pyx_t_7 = __pyx_v_i; + __pyx_t_8 = -1; + if (__pyx_t_7 < 0) { + __pyx_t_7 += __pyx_v_indices_view.shape[0]; + if (unlikely(__pyx_t_7 < 0)) __pyx_t_8 = 0; + } else if (unlikely(__pyx_t_7 >= __pyx_v_indices_view.shape[0])) __pyx_t_8 = 0; + if (unlikely(__pyx_t_8 != -1)) { + __Pyx_RaiseBufferIndexError(__pyx_t_8); + __PYX_ERR(0, 101, __pyx_L1_error) + } + __pyx_v_idx = (*((__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_indices_view.data + __pyx_t_7 * __pyx_v_indices_view.strides[0]) ))); + + /* "fairseq/data/data_utils_fast.pyx":102 + * for i in range(len(indices_view)): + * idx = indices_view[i] + * num_tokens = num_tokens_fn(idx) # <<<<<<<<<<<<<< + * sample_lens.append(num_tokens) + * sample_len = max(sample_len, num_tokens) + */ + __pyx_t_9 = __Pyx_PyInt_From_long(__pyx_v_idx); if (unlikely(!__pyx_t_9)) __PYX_ERR(0, 102, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_9); + __Pyx_INCREF(__pyx_v_num_tokens_fn); + __pyx_t_10 = __pyx_v_num_tokens_fn; __pyx_t_11 = NULL; + __pyx_t_12 = 0; + #if CYTHON_UNPACK_METHODS + if (unlikely(PyMethod_Check(__pyx_t_10))) { + __pyx_t_11 = PyMethod_GET_SELF(__pyx_t_10); + if (likely(__pyx_t_11)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_10); + __Pyx_INCREF(__pyx_t_11); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_10, function); + __pyx_t_12 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_11, __pyx_t_9}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_10, __pyx_callargs+1-__pyx_t_12, 1+__pyx_t_12); + __Pyx_XDECREF(__pyx_t_11); __pyx_t_11 = 0; + __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 102, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0; + } + __pyx_t_13 = __Pyx_PyInt_As_long(__pyx_t_1); if (unlikely((__pyx_t_13 == (long)-1) && PyErr_Occurred())) __PYX_ERR(0, 102, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_v_num_tokens = __pyx_t_13; + + /* "fairseq/data/data_utils_fast.pyx":103 + * idx = indices_view[i] + * num_tokens = num_tokens_fn(idx) + * sample_lens.append(num_tokens) # <<<<<<<<<<<<<< + * sample_len = max(sample_len, num_tokens) + * + */ + __pyx_t_1 = __Pyx_PyInt_From_long(__pyx_v_num_tokens); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 103, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_14 = __Pyx_PyList_Append(__pyx_v_sample_lens, __pyx_t_1); if (unlikely(__pyx_t_14 == ((int)-1))) __PYX_ERR(0, 103, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":104 + * num_tokens = num_tokens_fn(idx) + * sample_lens.append(num_tokens) + * sample_len = max(sample_len, num_tokens) # <<<<<<<<<<<<<< + * + * shape_idx = _find_valid_shape(shapes_view, len(batch) + 1, sample_len) + */ + __pyx_t_13 = __pyx_v_num_tokens; + __pyx_t_15 = __pyx_v_sample_len; + __pyx_t_17 = (__pyx_t_13 > __pyx_t_15); + if (__pyx_t_17) { + __pyx_t_16 = __pyx_t_13; + } else { + __pyx_t_16 = __pyx_t_15; + } + __pyx_v_sample_len = __pyx_t_16; + + /* "fairseq/data/data_utils_fast.pyx":106 + * sample_len = max(sample_len, num_tokens) + * + * shape_idx = _find_valid_shape(shapes_view, len(batch) + 1, sample_len) # <<<<<<<<<<<<<< + * if shape_idx == -1: + * batches.append(batch) + */ + __pyx_t_18 = __Pyx_PyList_GET_SIZE(__pyx_v_batch); if (unlikely(__pyx_t_18 == ((Py_ssize_t)-1))) __PYX_ERR(0, 106, __pyx_L1_error) + __pyx_t_1 = __pyx_f_7fairseq_4data_15data_utils_fast__find_valid_shape(__pyx_v_shapes_view, (__pyx_t_18 + 1), __pyx_v_sample_len); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 106, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_XDECREF_SET(__pyx_v_shape_idx, __pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":107 + * + * shape_idx = _find_valid_shape(shapes_view, len(batch) + 1, sample_len) + * if shape_idx == -1: # <<<<<<<<<<<<<< + * batches.append(batch) + * batch = [] + */ + __pyx_t_17 = (__Pyx_PyInt_BoolEqObjC(__pyx_v_shape_idx, __pyx_int_neg_1, -1L, 0)); if (unlikely((__pyx_t_17 < 0))) __PYX_ERR(0, 107, __pyx_L1_error) + if (__pyx_t_17) { + + /* "fairseq/data/data_utils_fast.pyx":108 + * shape_idx = _find_valid_shape(shapes_view, len(batch) + 1, sample_len) + * if shape_idx == -1: + * batches.append(batch) # <<<<<<<<<<<<<< + * batch = [] + * sample_lens = [] + */ + __pyx_t_14 = __Pyx_PyList_Append(__pyx_v_batches, __pyx_v_batch); if (unlikely(__pyx_t_14 == ((int)-1))) __PYX_ERR(0, 108, __pyx_L1_error) + + /* "fairseq/data/data_utils_fast.pyx":109 + * if shape_idx == -1: + * batches.append(batch) + * batch = [] # <<<<<<<<<<<<<< + * sample_lens = [] + * sample_len = 0 + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 109, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF_SET(__pyx_v_batch, ((PyObject*)__pyx_t_1)); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":110 + * batches.append(batch) + * batch = [] + * sample_lens = [] # <<<<<<<<<<<<<< + * sample_len = 0 + * shapes_view = fixed_shapes_sorted + */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 110, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF_SET(__pyx_v_sample_lens, ((PyObject*)__pyx_t_1)); + __pyx_t_1 = 0; + + /* "fairseq/data/data_utils_fast.pyx":111 + * batch = [] + * sample_lens = [] + * sample_len = 0 # <<<<<<<<<<<<<< + * shapes_view = fixed_shapes_sorted + * elif shape_idx > 0: + */ + __pyx_v_sample_len = 0; + + /* "fairseq/data/data_utils_fast.pyx":112 + * sample_lens = [] + * sample_len = 0 + * shapes_view = fixed_shapes_sorted # <<<<<<<<<<<<<< + * elif shape_idx > 0: + * # small optimization for the next call to _find_valid_shape + */ + __pyx_t_3 = __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(((PyObject *)__pyx_v_fixed_shapes_sorted), PyBUF_WRITABLE); if (unlikely(!__pyx_t_3.memview)) __PYX_ERR(0, 112, __pyx_L1_error) + __PYX_XCLEAR_MEMVIEW(&__pyx_v_shapes_view, 1); + __pyx_v_shapes_view = __pyx_t_3; + __pyx_t_3.memview = NULL; + __pyx_t_3.data = NULL; + + /* "fairseq/data/data_utils_fast.pyx":107 + * + * shape_idx = _find_valid_shape(shapes_view, len(batch) + 1, sample_len) + * if shape_idx == -1: # <<<<<<<<<<<<<< + * batches.append(batch) + * batch = [] + */ + goto __pyx_L5; + } + + /* "fairseq/data/data_utils_fast.pyx":113 + * sample_len = 0 + * shapes_view = fixed_shapes_sorted + * elif shape_idx > 0: # <<<<<<<<<<<<<< + * # small optimization for the next call to _find_valid_shape + * shapes_view = shapes_view[shape_idx:] + */ + __pyx_t_1 = PyObject_RichCompare(__pyx_v_shape_idx, __pyx_int_0, Py_GT); __Pyx_XGOTREF(__pyx_t_1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 113, __pyx_L1_error) + __pyx_t_17 = __Pyx_PyObject_IsTrue(__pyx_t_1); if (unlikely((__pyx_t_17 < 0))) __PYX_ERR(0, 113, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + if (__pyx_t_17) { + + /* "fairseq/data/data_utils_fast.pyx":115 + * elif shape_idx > 0: + * # small optimization for the next call to _find_valid_shape + * shapes_view = shapes_view[shape_idx:] # <<<<<<<<<<<<<< + * + * batch.append(idx) + */ + __pyx_t_18 = __Pyx_PyIndex_AsSsize_t(__pyx_v_shape_idx); if (unlikely((__pyx_t_18 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(0, 115, __pyx_L1_error) + __pyx_t_3.data = __pyx_v_shapes_view.data; + __pyx_t_3.memview = __pyx_v_shapes_view.memview; + __PYX_INC_MEMVIEW(&__pyx_t_3, 1); + __pyx_t_8 = -1; + if (unlikely(__pyx_memoryview_slice_memviewslice( + &__pyx_t_3, + __pyx_v_shapes_view.shape[0], __pyx_v_shapes_view.strides[0], __pyx_v_shapes_view.suboffsets[0], + 0, + 0, + &__pyx_t_8, + __pyx_t_18, + 0, + 0, + 1, + 0, + 0, + 1) < 0)) +{ + __PYX_ERR(0, 115, __pyx_L1_error) +} + +__pyx_t_3.shape[1] = __pyx_v_shapes_view.shape[1]; +__pyx_t_3.strides[1] = __pyx_v_shapes_view.strides[1]; + __pyx_t_3.suboffsets[1] = -1; + +__PYX_XCLEAR_MEMVIEW(&__pyx_v_shapes_view, 1); + __pyx_v_shapes_view = __pyx_t_3; + __pyx_t_3.memview = NULL; + __pyx_t_3.data = NULL; + + /* "fairseq/data/data_utils_fast.pyx":113 + * sample_len = 0 + * shapes_view = fixed_shapes_sorted + * elif shape_idx > 0: # <<<<<<<<<<<<<< + * # small optimization for the next call to _find_valid_shape + * shapes_view = shapes_view[shape_idx:] + */ + } + __pyx_L5:; + + /* "fairseq/data/data_utils_fast.pyx":117 + * shapes_view = shapes_view[shape_idx:] + * + * batch.append(idx) # <<<<<<<<<<<<<< + * + * if len(batch) > 0: + */ + __pyx_t_1 = __Pyx_PyInt_From_long(__pyx_v_idx); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_14 = __Pyx_PyList_Append(__pyx_v_batch, __pyx_t_1); if (unlikely(__pyx_t_14 == ((int)-1))) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + } + + /* "fairseq/data/data_utils_fast.pyx":119 + * batch.append(idx) + * + * if len(batch) > 0: # <<<<<<<<<<<<<< + * batches.append(batch) + * + */ + __pyx_t_4 = __Pyx_PyList_GET_SIZE(__pyx_v_batch); if (unlikely(__pyx_t_4 == ((Py_ssize_t)-1))) __PYX_ERR(0, 119, __pyx_L1_error) + __pyx_t_17 = (__pyx_t_4 > 0); + if (__pyx_t_17) { + + /* "fairseq/data/data_utils_fast.pyx":120 + * + * if len(batch) > 0: + * batches.append(batch) # <<<<<<<<<<<<<< + * + * return batches + */ + __pyx_t_14 = __Pyx_PyList_Append(__pyx_v_batches, __pyx_v_batch); if (unlikely(__pyx_t_14 == ((int)-1))) __PYX_ERR(0, 120, __pyx_L1_error) + + /* "fairseq/data/data_utils_fast.pyx":119 + * batch.append(idx) + * + * if len(batch) > 0: # <<<<<<<<<<<<<< + * batches.append(batch) + * + */ + } + + /* "fairseq/data/data_utils_fast.pyx":122 + * batches.append(batch) + * + * return batches # <<<<<<<<<<<<<< + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_batches); + __pyx_r = __pyx_v_batches; + goto __pyx_L0; + + /* "fairseq/data/data_utils_fast.pyx":84 + * + * @cython.cdivision(True) + * cpdef list batch_fixed_shapes_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __PYX_XCLEAR_MEMVIEW(&__pyx_t_2, 1); + __PYX_XCLEAR_MEMVIEW(&__pyx_t_3, 1); + __Pyx_XDECREF(__pyx_t_9); + __Pyx_XDECREF(__pyx_t_10); + __Pyx_XDECREF(__pyx_t_11); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.data_utils_fast.batch_fixed_shapes_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XDECREF(__pyx_v_sample_lens); + __Pyx_XDECREF(__pyx_v_batch); + __Pyx_XDECREF(__pyx_v_batches); + __PYX_XCLEAR_MEMVIEW(&__pyx_v_indices_view, 1); + __PYX_XCLEAR_MEMVIEW(&__pyx_v_shapes_view, 1); + __Pyx_XDECREF(__pyx_v_shape_idx); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* Python wrapper */ +static PyObject *__pyx_pw_7fairseq_4data_15data_utils_fast_3batch_fixed_shapes_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_7fairseq_4data_15data_utils_fast_3batch_fixed_shapes_fast = {"batch_fixed_shapes_fast", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_15data_utils_fast_3batch_fixed_shapes_fast, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_7fairseq_4data_15data_utils_fast_3batch_fixed_shapes_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyArrayObject *__pyx_v_indices = 0; + PyObject *__pyx_v_num_tokens_fn = 0; + PyArrayObject *__pyx_v_fixed_shapes_sorted = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[3] = {0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("batch_fixed_shapes_fast (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_indices,&__pyx_n_s_num_tokens_fn,&__pyx_n_s_fixed_shapes_sorted,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_indices)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 84, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_num_tokens_fn)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 84, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("batch_fixed_shapes_fast", 1, 3, 3, 1); __PYX_ERR(0, 84, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_fixed_shapes_sorted)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 84, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("batch_fixed_shapes_fast", 1, 3, 3, 2); __PYX_ERR(0, 84, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "batch_fixed_shapes_fast") < 0)) __PYX_ERR(0, 84, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 3)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + } + __pyx_v_indices = ((PyArrayObject *)values[0]); + __pyx_v_num_tokens_fn = values[1]; + __pyx_v_fixed_shapes_sorted = ((PyArrayObject *)values[2]); + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("batch_fixed_shapes_fast", 1, 3, 3, __pyx_nargs); __PYX_ERR(0, 84, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("fairseq.data.data_utils_fast.batch_fixed_shapes_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_indices), __pyx_ptype_5numpy_ndarray, 1, "indices", 0))) __PYX_ERR(0, 85, __pyx_L1_error) + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_fixed_shapes_sorted), __pyx_ptype_5numpy_ndarray, 1, "fixed_shapes_sorted", 0))) __PYX_ERR(0, 87, __pyx_L1_error) + __pyx_r = __pyx_pf_7fairseq_4data_15data_utils_fast_2batch_fixed_shapes_fast(__pyx_self, __pyx_v_indices, __pyx_v_num_tokens_fn, __pyx_v_fixed_shapes_sorted); + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __pyx_r = NULL; + __pyx_L0:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_7fairseq_4data_15data_utils_fast_2batch_fixed_shapes_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_indices, PyObject *__pyx_v_num_tokens_fn, PyArrayObject *__pyx_v_fixed_shapes_sorted) { + __Pyx_LocalBuf_ND __pyx_pybuffernd_fixed_shapes_sorted; + __Pyx_Buffer __pyx_pybuffer_fixed_shapes_sorted; + __Pyx_LocalBuf_ND __pyx_pybuffernd_indices; + __Pyx_Buffer __pyx_pybuffer_indices; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("batch_fixed_shapes_fast", 1); + __pyx_pybuffer_indices.pybuffer.buf = NULL; + __pyx_pybuffer_indices.refcount = 0; + __pyx_pybuffernd_indices.data = NULL; + __pyx_pybuffernd_indices.rcbuffer = &__pyx_pybuffer_indices; + __pyx_pybuffer_fixed_shapes_sorted.pybuffer.buf = NULL; + __pyx_pybuffer_fixed_shapes_sorted.refcount = 0; + __pyx_pybuffernd_fixed_shapes_sorted.data = NULL; + __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer = &__pyx_pybuffer_fixed_shapes_sorted; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 84, __pyx_L1_error) + } + __pyx_pybuffernd_indices.diminfo[0].strides = __pyx_pybuffernd_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_indices.diminfo[0].shape = __pyx_pybuffernd_indices.rcbuffer->pybuffer.shape[0]; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer, (PyObject*)__pyx_v_fixed_shapes_sorted, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 84, __pyx_L1_error) + } + __pyx_pybuffernd_fixed_shapes_sorted.diminfo[0].strides = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_fixed_shapes_sorted.diminfo[0].shape = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_fixed_shapes_sorted.diminfo[1].strides = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_fixed_shapes_sorted.diminfo[1].shape = __pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer.shape[1]; + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __pyx_f_7fairseq_4data_15data_utils_fast_batch_fixed_shapes_fast(__pyx_v_indices, __pyx_v_num_tokens_fn, __pyx_v_fixed_shapes_sorted, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 84, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.data_utils_fast.batch_fixed_shapes_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_fixed_shapes_sorted.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_indices.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} +static struct __pyx_vtabstruct_array __pyx_vtable_array; + +static PyObject *__pyx_tp_new_array(PyTypeObject *t, PyObject *a, PyObject *k) { + struct __pyx_array_obj *p; + PyObject *o; + #if CYTHON_COMPILING_IN_LIMITED_API + allocfunc alloc_func = (allocfunc)PyType_GetSlot(t, Py_tp_alloc); + o = alloc_func(t, 0); + #else + if (likely(!__Pyx_PyType_HasFeature(t, Py_TPFLAGS_IS_ABSTRACT))) { + o = (*t->tp_alloc)(t, 0); + } else { + o = (PyObject *) PyBaseObject_Type.tp_new(t, __pyx_empty_tuple, 0); + } + if (unlikely(!o)) return 0; + #endif + p = ((struct __pyx_array_obj *)o); + p->__pyx_vtab = __pyx_vtabptr_array; + p->mode = ((PyObject*)Py_None); Py_INCREF(Py_None); + p->_format = ((PyObject*)Py_None); Py_INCREF(Py_None); + if (unlikely(__pyx_array___cinit__(o, a, k) < 0)) goto bad; + return o; + bad: + Py_DECREF(o); o = 0; + return NULL; +} + +static void __pyx_tp_dealloc_array(PyObject *o) { + struct __pyx_array_obj *p = (struct __pyx_array_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && (!PyType_IS_GC(Py_TYPE(o)) || !__Pyx_PyObject_GC_IsFinalized(o))) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc_array) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + { + PyObject *etype, *eval, *etb; + PyErr_Fetch(&etype, &eval, &etb); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1); + __pyx_array___dealloc__(o); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1); + PyErr_Restore(etype, eval, etb); + } + Py_CLEAR(p->mode); + Py_CLEAR(p->_format); + #if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + (*Py_TYPE(o)->tp_free)(o); + #else + { + freefunc tp_free = (freefunc)PyType_GetSlot(Py_TYPE(o), Py_tp_free); + if (tp_free) tp_free(o); + } + #endif +} +static PyObject *__pyx_sq_item_array(PyObject *o, Py_ssize_t i) { + PyObject *r; + PyObject *x = PyInt_FromSsize_t(i); if(!x) return 0; + r = Py_TYPE(o)->tp_as_mapping->mp_subscript(o, x); + Py_DECREF(x); + return r; +} + +static int __pyx_mp_ass_subscript_array(PyObject *o, PyObject *i, PyObject *v) { + if (v) { + return __pyx_array___setitem__(o, i, v); + } + else { + __Pyx_TypeName o_type_name; + o_type_name = __Pyx_PyType_GetName(Py_TYPE(o)); + PyErr_Format(PyExc_NotImplementedError, + "Subscript deletion not supported by " __Pyx_FMT_TYPENAME, o_type_name); + __Pyx_DECREF_TypeName(o_type_name); + return -1; + } +} + +static PyObject *__pyx_tp_getattro_array(PyObject *o, PyObject *n) { + PyObject *v = __Pyx_PyObject_GenericGetAttr(o, n); + if (!v && PyErr_ExceptionMatches(PyExc_AttributeError)) { + PyErr_Clear(); + v = __pyx_array___getattr__(o, n); + } + return v; +} + +static PyObject *__pyx_getprop___pyx_array_memview(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_5array_7memview_1__get__(o); +} + +static PyMethodDef __pyx_methods_array[] = { + {"__getattr__", (PyCFunction)__pyx_array___getattr__, METH_O|METH_COEXIST, 0}, + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_array_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_array_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; + +static struct PyGetSetDef __pyx_getsets_array[] = { + {(char *)"memview", __pyx_getprop___pyx_array_memview, 0, (char *)0, 0}, + {0, 0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +#if !CYTHON_COMPILING_IN_LIMITED_API + +static PyBufferProcs __pyx_tp_as_buffer_array = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_array_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; +#endif +static PyType_Slot __pyx_type___pyx_array_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc_array}, + {Py_sq_length, (void *)__pyx_array___len__}, + {Py_sq_item, (void *)__pyx_sq_item_array}, + {Py_mp_length, (void *)__pyx_array___len__}, + {Py_mp_subscript, (void *)__pyx_array___getitem__}, + {Py_mp_ass_subscript, (void *)__pyx_mp_ass_subscript_array}, + {Py_tp_getattro, (void *)__pyx_tp_getattro_array}, + #if defined(Py_bf_getbuffer) + {Py_bf_getbuffer, (void *)__pyx_array_getbuffer}, + #endif + {Py_tp_methods, (void *)__pyx_methods_array}, + {Py_tp_getset, (void *)__pyx_getsets_array}, + {Py_tp_new, (void *)__pyx_tp_new_array}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_array_spec = { + "fairseq.data.data_utils_fast.array", + sizeof(struct __pyx_array_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_SEQUENCE, + __pyx_type___pyx_array_slots, +}; +#else + +static PySequenceMethods __pyx_tp_as_sequence_array = { + __pyx_array___len__, /*sq_length*/ + 0, /*sq_concat*/ + 0, /*sq_repeat*/ + __pyx_sq_item_array, /*sq_item*/ + 0, /*sq_slice*/ + 0, /*sq_ass_item*/ + 0, /*sq_ass_slice*/ + 0, /*sq_contains*/ + 0, /*sq_inplace_concat*/ + 0, /*sq_inplace_repeat*/ +}; + +static PyMappingMethods __pyx_tp_as_mapping_array = { + __pyx_array___len__, /*mp_length*/ + __pyx_array___getitem__, /*mp_subscript*/ + __pyx_mp_ass_subscript_array, /*mp_ass_subscript*/ +}; + +static PyBufferProcs __pyx_tp_as_buffer_array = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_array_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; + +static PyTypeObject __pyx_type___pyx_array = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.data_utils_fast.""array", /*tp_name*/ + sizeof(struct __pyx_array_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc_array, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + 0, /*tp_repr*/ + 0, /*tp_as_number*/ + &__pyx_tp_as_sequence_array, /*tp_as_sequence*/ + &__pyx_tp_as_mapping_array, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + 0, /*tp_str*/ + __pyx_tp_getattro_array, /*tp_getattro*/ + 0, /*tp_setattro*/ + &__pyx_tp_as_buffer_array, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_SEQUENCE, /*tp_flags*/ + 0, /*tp_doc*/ + 0, /*tp_traverse*/ + 0, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods_array, /*tp_methods*/ + 0, /*tp_members*/ + __pyx_getsets_array, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + 0, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new_array, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif + +static PyObject *__pyx_tp_new_Enum(PyTypeObject *t, CYTHON_UNUSED PyObject *a, CYTHON_UNUSED PyObject *k) { + struct __pyx_MemviewEnum_obj *p; + PyObject *o; + #if CYTHON_COMPILING_IN_LIMITED_API + allocfunc alloc_func = (allocfunc)PyType_GetSlot(t, Py_tp_alloc); + o = alloc_func(t, 0); + #else + if (likely(!__Pyx_PyType_HasFeature(t, Py_TPFLAGS_IS_ABSTRACT))) { + o = (*t->tp_alloc)(t, 0); + } else { + o = (PyObject *) PyBaseObject_Type.tp_new(t, __pyx_empty_tuple, 0); + } + if (unlikely(!o)) return 0; + #endif + p = ((struct __pyx_MemviewEnum_obj *)o); + p->name = Py_None; Py_INCREF(Py_None); + return o; +} + +static void __pyx_tp_dealloc_Enum(PyObject *o) { + struct __pyx_MemviewEnum_obj *p = (struct __pyx_MemviewEnum_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && !__Pyx_PyObject_GC_IsFinalized(o)) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc_Enum) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + PyObject_GC_UnTrack(o); + Py_CLEAR(p->name); + #if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + (*Py_TYPE(o)->tp_free)(o); + #else + { + freefunc tp_free = (freefunc)PyType_GetSlot(Py_TYPE(o), Py_tp_free); + if (tp_free) tp_free(o); + } + #endif +} + +static int __pyx_tp_traverse_Enum(PyObject *o, visitproc v, void *a) { + int e; + struct __pyx_MemviewEnum_obj *p = (struct __pyx_MemviewEnum_obj *)o; + if (p->name) { + e = (*v)(p->name, a); if (e) return e; + } + return 0; +} + +static int __pyx_tp_clear_Enum(PyObject *o) { + PyObject* tmp; + struct __pyx_MemviewEnum_obj *p = (struct __pyx_MemviewEnum_obj *)o; + tmp = ((PyObject*)p->name); + p->name = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + return 0; +} + +static PyObject *__pyx_specialmethod___pyx_MemviewEnum___repr__(PyObject *self, CYTHON_UNUSED PyObject *arg) { + return __pyx_MemviewEnum___repr__(self); +} + +static PyMethodDef __pyx_methods_Enum[] = { + {"__repr__", (PyCFunction)__pyx_specialmethod___pyx_MemviewEnum___repr__, METH_NOARGS|METH_COEXIST, 0}, + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_MemviewEnum_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_MemviewEnum_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +static PyType_Slot __pyx_type___pyx_MemviewEnum_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc_Enum}, + {Py_tp_repr, (void *)__pyx_MemviewEnum___repr__}, + {Py_tp_traverse, (void *)__pyx_tp_traverse_Enum}, + {Py_tp_clear, (void *)__pyx_tp_clear_Enum}, + {Py_tp_methods, (void *)__pyx_methods_Enum}, + {Py_tp_init, (void *)__pyx_MemviewEnum___init__}, + {Py_tp_new, (void *)__pyx_tp_new_Enum}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_MemviewEnum_spec = { + "fairseq.data.data_utils_fast.Enum", + sizeof(struct __pyx_MemviewEnum_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, + __pyx_type___pyx_MemviewEnum_slots, +}; +#else + +static PyTypeObject __pyx_type___pyx_MemviewEnum = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.data_utils_fast.""Enum", /*tp_name*/ + sizeof(struct __pyx_MemviewEnum_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc_Enum, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + __pyx_MemviewEnum___repr__, /*tp_repr*/ + 0, /*tp_as_number*/ + 0, /*tp_as_sequence*/ + 0, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + 0, /*tp_str*/ + 0, /*tp_getattro*/ + 0, /*tp_setattro*/ + 0, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, /*tp_flags*/ + 0, /*tp_doc*/ + __pyx_tp_traverse_Enum, /*tp_traverse*/ + __pyx_tp_clear_Enum, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods_Enum, /*tp_methods*/ + 0, /*tp_members*/ + 0, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + __pyx_MemviewEnum___init__, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new_Enum, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif +static struct __pyx_vtabstruct_memoryview __pyx_vtable_memoryview; + +static PyObject *__pyx_tp_new_memoryview(PyTypeObject *t, PyObject *a, PyObject *k) { + struct __pyx_memoryview_obj *p; + PyObject *o; + #if CYTHON_COMPILING_IN_LIMITED_API + allocfunc alloc_func = (allocfunc)PyType_GetSlot(t, Py_tp_alloc); + o = alloc_func(t, 0); + #else + if (likely(!__Pyx_PyType_HasFeature(t, Py_TPFLAGS_IS_ABSTRACT))) { + o = (*t->tp_alloc)(t, 0); + } else { + o = (PyObject *) PyBaseObject_Type.tp_new(t, __pyx_empty_tuple, 0); + } + if (unlikely(!o)) return 0; + #endif + p = ((struct __pyx_memoryview_obj *)o); + p->__pyx_vtab = __pyx_vtabptr_memoryview; + p->obj = Py_None; Py_INCREF(Py_None); + p->_size = Py_None; Py_INCREF(Py_None); + p->_array_interface = Py_None; Py_INCREF(Py_None); + p->view.obj = NULL; + if (unlikely(__pyx_memoryview___cinit__(o, a, k) < 0)) goto bad; + return o; + bad: + Py_DECREF(o); o = 0; + return NULL; +} + +static void __pyx_tp_dealloc_memoryview(PyObject *o) { + struct __pyx_memoryview_obj *p = (struct __pyx_memoryview_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && !__Pyx_PyObject_GC_IsFinalized(o)) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc_memoryview) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + PyObject_GC_UnTrack(o); + { + PyObject *etype, *eval, *etb; + PyErr_Fetch(&etype, &eval, &etb); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1); + __pyx_memoryview___dealloc__(o); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1); + PyErr_Restore(etype, eval, etb); + } + Py_CLEAR(p->obj); + Py_CLEAR(p->_size); + Py_CLEAR(p->_array_interface); + #if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + (*Py_TYPE(o)->tp_free)(o); + #else + { + freefunc tp_free = (freefunc)PyType_GetSlot(Py_TYPE(o), Py_tp_free); + if (tp_free) tp_free(o); + } + #endif +} + +static int __pyx_tp_traverse_memoryview(PyObject *o, visitproc v, void *a) { + int e; + struct __pyx_memoryview_obj *p = (struct __pyx_memoryview_obj *)o; + if (p->obj) { + e = (*v)(p->obj, a); if (e) return e; + } + if (p->_size) { + e = (*v)(p->_size, a); if (e) return e; + } + if (p->_array_interface) { + e = (*v)(p->_array_interface, a); if (e) return e; + } + if (p->view.obj) { + e = (*v)(p->view.obj, a); if (e) return e; + } + return 0; +} + +static int __pyx_tp_clear_memoryview(PyObject *o) { + PyObject* tmp; + struct __pyx_memoryview_obj *p = (struct __pyx_memoryview_obj *)o; + tmp = ((PyObject*)p->obj); + p->obj = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + tmp = ((PyObject*)p->_size); + p->_size = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + tmp = ((PyObject*)p->_array_interface); + p->_array_interface = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + Py_CLEAR(p->view.obj); + return 0; +} +static PyObject *__pyx_sq_item_memoryview(PyObject *o, Py_ssize_t i) { + PyObject *r; + PyObject *x = PyInt_FromSsize_t(i); if(!x) return 0; + r = Py_TYPE(o)->tp_as_mapping->mp_subscript(o, x); + Py_DECREF(x); + return r; +} + +static int __pyx_mp_ass_subscript_memoryview(PyObject *o, PyObject *i, PyObject *v) { + if (v) { + return __pyx_memoryview___setitem__(o, i, v); + } + else { + __Pyx_TypeName o_type_name; + o_type_name = __Pyx_PyType_GetName(Py_TYPE(o)); + PyErr_Format(PyExc_NotImplementedError, + "Subscript deletion not supported by " __Pyx_FMT_TYPENAME, o_type_name); + __Pyx_DECREF_TypeName(o_type_name); + return -1; + } +} + +static PyObject *__pyx_getprop___pyx_memoryview_T(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_1T_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_base(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_4base_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_shape(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_5shape_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_strides(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_7strides_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_suboffsets(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_10suboffsets_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_ndim(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_4ndim_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_itemsize(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_8itemsize_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_nbytes(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_6nbytes_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_size(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_4size_1__get__(o); +} + +static PyObject *__pyx_specialmethod___pyx_memoryview___repr__(PyObject *self, CYTHON_UNUSED PyObject *arg) { + return __pyx_memoryview___repr__(self); +} + +static PyMethodDef __pyx_methods_memoryview[] = { + {"__repr__", (PyCFunction)__pyx_specialmethod___pyx_memoryview___repr__, METH_NOARGS|METH_COEXIST, 0}, + {"is_c_contig", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_is_c_contig, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"is_f_contig", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_is_f_contig, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"copy", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_copy, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"copy_fortran", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_copy_fortran, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryview_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryview_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; + +static struct PyGetSetDef __pyx_getsets_memoryview[] = { + {(char *)"T", __pyx_getprop___pyx_memoryview_T, 0, (char *)0, 0}, + {(char *)"base", __pyx_getprop___pyx_memoryview_base, 0, (char *)0, 0}, + {(char *)"shape", __pyx_getprop___pyx_memoryview_shape, 0, (char *)0, 0}, + {(char *)"strides", __pyx_getprop___pyx_memoryview_strides, 0, (char *)0, 0}, + {(char *)"suboffsets", __pyx_getprop___pyx_memoryview_suboffsets, 0, (char *)0, 0}, + {(char *)"ndim", __pyx_getprop___pyx_memoryview_ndim, 0, (char *)0, 0}, + {(char *)"itemsize", __pyx_getprop___pyx_memoryview_itemsize, 0, (char *)0, 0}, + {(char *)"nbytes", __pyx_getprop___pyx_memoryview_nbytes, 0, (char *)0, 0}, + {(char *)"size", __pyx_getprop___pyx_memoryview_size, 0, (char *)0, 0}, + {0, 0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +#if !CYTHON_COMPILING_IN_LIMITED_API + +static PyBufferProcs __pyx_tp_as_buffer_memoryview = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_memoryview_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; +#endif +static PyType_Slot __pyx_type___pyx_memoryview_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc_memoryview}, + {Py_tp_repr, (void *)__pyx_memoryview___repr__}, + {Py_sq_length, (void *)__pyx_memoryview___len__}, + {Py_sq_item, (void *)__pyx_sq_item_memoryview}, + {Py_mp_length, (void *)__pyx_memoryview___len__}, + {Py_mp_subscript, (void *)__pyx_memoryview___getitem__}, + {Py_mp_ass_subscript, (void *)__pyx_mp_ass_subscript_memoryview}, + {Py_tp_str, (void *)__pyx_memoryview___str__}, + #if defined(Py_bf_getbuffer) + {Py_bf_getbuffer, (void *)__pyx_memoryview_getbuffer}, + #endif + {Py_tp_traverse, (void *)__pyx_tp_traverse_memoryview}, + {Py_tp_clear, (void *)__pyx_tp_clear_memoryview}, + {Py_tp_methods, (void *)__pyx_methods_memoryview}, + {Py_tp_getset, (void *)__pyx_getsets_memoryview}, + {Py_tp_new, (void *)__pyx_tp_new_memoryview}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_memoryview_spec = { + "fairseq.data.data_utils_fast.memoryview", + sizeof(struct __pyx_memoryview_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, + __pyx_type___pyx_memoryview_slots, +}; +#else + +static PySequenceMethods __pyx_tp_as_sequence_memoryview = { + __pyx_memoryview___len__, /*sq_length*/ + 0, /*sq_concat*/ + 0, /*sq_repeat*/ + __pyx_sq_item_memoryview, /*sq_item*/ + 0, /*sq_slice*/ + 0, /*sq_ass_item*/ + 0, /*sq_ass_slice*/ + 0, /*sq_contains*/ + 0, /*sq_inplace_concat*/ + 0, /*sq_inplace_repeat*/ +}; + +static PyMappingMethods __pyx_tp_as_mapping_memoryview = { + __pyx_memoryview___len__, /*mp_length*/ + __pyx_memoryview___getitem__, /*mp_subscript*/ + __pyx_mp_ass_subscript_memoryview, /*mp_ass_subscript*/ +}; + +static PyBufferProcs __pyx_tp_as_buffer_memoryview = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_memoryview_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; + +static PyTypeObject __pyx_type___pyx_memoryview = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.data_utils_fast.""memoryview", /*tp_name*/ + sizeof(struct __pyx_memoryview_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc_memoryview, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + __pyx_memoryview___repr__, /*tp_repr*/ + 0, /*tp_as_number*/ + &__pyx_tp_as_sequence_memoryview, /*tp_as_sequence*/ + &__pyx_tp_as_mapping_memoryview, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + __pyx_memoryview___str__, /*tp_str*/ + 0, /*tp_getattro*/ + 0, /*tp_setattro*/ + &__pyx_tp_as_buffer_memoryview, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, /*tp_flags*/ + 0, /*tp_doc*/ + __pyx_tp_traverse_memoryview, /*tp_traverse*/ + __pyx_tp_clear_memoryview, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods_memoryview, /*tp_methods*/ + 0, /*tp_members*/ + __pyx_getsets_memoryview, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + 0, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new_memoryview, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif +static struct __pyx_vtabstruct__memoryviewslice __pyx_vtable__memoryviewslice; + +static PyObject *__pyx_tp_new__memoryviewslice(PyTypeObject *t, PyObject *a, PyObject *k) { + struct __pyx_memoryviewslice_obj *p; + PyObject *o = __pyx_tp_new_memoryview(t, a, k); + if (unlikely(!o)) return 0; + p = ((struct __pyx_memoryviewslice_obj *)o); + p->__pyx_base.__pyx_vtab = (struct __pyx_vtabstruct_memoryview*)__pyx_vtabptr__memoryviewslice; + new((void*)&(p->from_slice)) __Pyx_memviewslice(); + p->from_object = Py_None; Py_INCREF(Py_None); + p->from_slice.memview = NULL; + return o; +} + +static void __pyx_tp_dealloc__memoryviewslice(PyObject *o) { + struct __pyx_memoryviewslice_obj *p = (struct __pyx_memoryviewslice_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && !__Pyx_PyObject_GC_IsFinalized(o)) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc__memoryviewslice) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + PyObject_GC_UnTrack(o); + { + PyObject *etype, *eval, *etb; + PyErr_Fetch(&etype, &eval, &etb); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1); + __pyx_memoryviewslice___dealloc__(o); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1); + PyErr_Restore(etype, eval, etb); + } + __Pyx_call_destructor(p->from_slice); + Py_CLEAR(p->from_object); + PyObject_GC_Track(o); + __pyx_tp_dealloc_memoryview(o); +} + +static int __pyx_tp_traverse__memoryviewslice(PyObject *o, visitproc v, void *a) { + int e; + struct __pyx_memoryviewslice_obj *p = (struct __pyx_memoryviewslice_obj *)o; + e = __pyx_tp_traverse_memoryview(o, v, a); if (e) return e; + if (p->from_object) { + e = (*v)(p->from_object, a); if (e) return e; + } + return 0; +} + +static int __pyx_tp_clear__memoryviewslice(PyObject *o) { + PyObject* tmp; + struct __pyx_memoryviewslice_obj *p = (struct __pyx_memoryviewslice_obj *)o; + __pyx_tp_clear_memoryview(o); + tmp = ((PyObject*)p->from_object); + p->from_object = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + __PYX_XCLEAR_MEMVIEW(&p->from_slice, 1); + return 0; +} + +static PyMethodDef __pyx_methods__memoryviewslice[] = { + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryviewslice_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryviewslice_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +static PyType_Slot __pyx_type___pyx_memoryviewslice_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc__memoryviewslice}, + {Py_tp_doc, (void *)PyDoc_STR("Internal class for passing memoryview slices to Python")}, + {Py_tp_traverse, (void *)__pyx_tp_traverse__memoryviewslice}, + {Py_tp_clear, (void *)__pyx_tp_clear__memoryviewslice}, + {Py_tp_methods, (void *)__pyx_methods__memoryviewslice}, + {Py_tp_new, (void *)__pyx_tp_new__memoryviewslice}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_memoryviewslice_spec = { + "fairseq.data.data_utils_fast._memoryviewslice", + sizeof(struct __pyx_memoryviewslice_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC|Py_TPFLAGS_SEQUENCE, + __pyx_type___pyx_memoryviewslice_slots, +}; +#else + +static PyTypeObject __pyx_type___pyx_memoryviewslice = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.data_utils_fast.""_memoryviewslice", /*tp_name*/ + sizeof(struct __pyx_memoryviewslice_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc__memoryviewslice, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + #if CYTHON_COMPILING_IN_PYPY || 0 + __pyx_memoryview___repr__, /*tp_repr*/ + #else + 0, /*tp_repr*/ + #endif + 0, /*tp_as_number*/ + 0, /*tp_as_sequence*/ + 0, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + #if CYTHON_COMPILING_IN_PYPY || 0 + __pyx_memoryview___str__, /*tp_str*/ + #else + 0, /*tp_str*/ + #endif + 0, /*tp_getattro*/ + 0, /*tp_setattro*/ + 0, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC|Py_TPFLAGS_SEQUENCE, /*tp_flags*/ + PyDoc_STR("Internal class for passing memoryview slices to Python"), /*tp_doc*/ + __pyx_tp_traverse__memoryviewslice, /*tp_traverse*/ + __pyx_tp_clear__memoryviewslice, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods__memoryviewslice, /*tp_methods*/ + 0, /*tp_members*/ + 0, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + 0, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new__memoryviewslice, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif + +static PyMethodDef __pyx_methods[] = { + {0, 0, 0, 0} +}; +#ifndef CYTHON_SMALL_CODE +#if defined(__clang__) + #define CYTHON_SMALL_CODE +#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3)) + #define CYTHON_SMALL_CODE __attribute__((cold)) +#else + #define CYTHON_SMALL_CODE +#endif +#endif +/* #### Code section: pystring_table ### */ + +static int __Pyx_CreateStringTabAndInitStrings(void) { + __Pyx_StringTabEntry __pyx_string_tab[] = { + {&__pyx_kp_u_, __pyx_k_, sizeof(__pyx_k_), 0, 1, 0, 0}, + {&__pyx_n_s_ASCII, __pyx_k_ASCII, sizeof(__pyx_k_ASCII), 0, 0, 1, 1}, + {&__pyx_kp_s_All_dimensions_preceding_dimensi, __pyx_k_All_dimensions_preceding_dimensi, sizeof(__pyx_k_All_dimensions_preceding_dimensi), 0, 0, 1, 0}, + {&__pyx_n_s_AssertionError, __pyx_k_AssertionError, sizeof(__pyx_k_AssertionError), 0, 0, 1, 1}, + {&__pyx_kp_s_Buffer_view_does_not_expose_stri, __pyx_k_Buffer_view_does_not_expose_stri, sizeof(__pyx_k_Buffer_view_does_not_expose_stri), 0, 0, 1, 0}, + {&__pyx_kp_s_Can_only_create_a_buffer_that_is, __pyx_k_Can_only_create_a_buffer_that_is, sizeof(__pyx_k_Can_only_create_a_buffer_that_is), 0, 0, 1, 0}, + {&__pyx_kp_s_Cannot_assign_to_read_only_memor, __pyx_k_Cannot_assign_to_read_only_memor, sizeof(__pyx_k_Cannot_assign_to_read_only_memor), 0, 0, 1, 0}, + {&__pyx_kp_s_Cannot_create_writable_memory_vi, __pyx_k_Cannot_create_writable_memory_vi, sizeof(__pyx_k_Cannot_create_writable_memory_vi), 0, 0, 1, 0}, + {&__pyx_kp_u_Cannot_index_with_type, __pyx_k_Cannot_index_with_type, sizeof(__pyx_k_Cannot_index_with_type), 0, 1, 0, 0}, + {&__pyx_kp_s_Cannot_transpose_memoryview_with, __pyx_k_Cannot_transpose_memoryview_with, sizeof(__pyx_k_Cannot_transpose_memoryview_with), 0, 0, 1, 0}, + {&__pyx_n_s_DTYPE, __pyx_k_DTYPE, sizeof(__pyx_k_DTYPE), 0, 0, 1, 1}, + {&__pyx_kp_s_Dimension_d_is_not_direct, __pyx_k_Dimension_d_is_not_direct, sizeof(__pyx_k_Dimension_d_is_not_direct), 0, 0, 1, 0}, + {&__pyx_n_s_Ellipsis, __pyx_k_Ellipsis, sizeof(__pyx_k_Ellipsis), 0, 0, 1, 1}, + {&__pyx_kp_s_Empty_shape_tuple_for_cython_arr, __pyx_k_Empty_shape_tuple_for_cython_arr, sizeof(__pyx_k_Empty_shape_tuple_for_cython_arr), 0, 0, 1, 0}, + {&__pyx_n_s_ImportError, __pyx_k_ImportError, sizeof(__pyx_k_ImportError), 0, 0, 1, 1}, + {&__pyx_kp_s_Incompatible_checksums_0x_x_vs_0, __pyx_k_Incompatible_checksums_0x_x_vs_0, sizeof(__pyx_k_Incompatible_checksums_0x_x_vs_0), 0, 0, 1, 0}, + {&__pyx_n_s_IndexError, __pyx_k_IndexError, sizeof(__pyx_k_IndexError), 0, 0, 1, 1}, + {&__pyx_kp_s_Index_out_of_bounds_axis_d, __pyx_k_Index_out_of_bounds_axis_d, sizeof(__pyx_k_Index_out_of_bounds_axis_d), 0, 0, 1, 0}, + {&__pyx_kp_s_Indirect_dimensions_not_supporte, __pyx_k_Indirect_dimensions_not_supporte, sizeof(__pyx_k_Indirect_dimensions_not_supporte), 0, 0, 1, 0}, + {&__pyx_kp_u_Invalid_mode_expected_c_or_fortr, __pyx_k_Invalid_mode_expected_c_or_fortr, sizeof(__pyx_k_Invalid_mode_expected_c_or_fortr), 0, 1, 0, 0}, + {&__pyx_kp_u_Invalid_shape_in_axis, __pyx_k_Invalid_shape_in_axis, sizeof(__pyx_k_Invalid_shape_in_axis), 0, 1, 0, 0}, + {&__pyx_n_s_MemoryError, __pyx_k_MemoryError, sizeof(__pyx_k_MemoryError), 0, 0, 1, 1}, + {&__pyx_kp_s_MemoryView_of_r_at_0x_x, __pyx_k_MemoryView_of_r_at_0x_x, sizeof(__pyx_k_MemoryView_of_r_at_0x_x), 0, 0, 1, 0}, + {&__pyx_kp_s_MemoryView_of_r_object, __pyx_k_MemoryView_of_r_object, sizeof(__pyx_k_MemoryView_of_r_object), 0, 0, 1, 0}, + {&__pyx_n_b_O, __pyx_k_O, sizeof(__pyx_k_O), 0, 0, 0, 1}, + {&__pyx_kp_u_Out_of_bounds_on_buffer_access_a, __pyx_k_Out_of_bounds_on_buffer_access_a, sizeof(__pyx_k_Out_of_bounds_on_buffer_access_a), 0, 1, 0, 0}, + {&__pyx_n_s_PickleError, __pyx_k_PickleError, sizeof(__pyx_k_PickleError), 0, 0, 1, 1}, + {&__pyx_n_s_Sequence, __pyx_k_Sequence, sizeof(__pyx_k_Sequence), 0, 0, 1, 1}, + {&__pyx_kp_s_Step_may_not_be_zero_axis_d, __pyx_k_Step_may_not_be_zero_axis_d, sizeof(__pyx_k_Step_may_not_be_zero_axis_d), 0, 0, 1, 0}, + {&__pyx_n_s_TypeError, __pyx_k_TypeError, sizeof(__pyx_k_TypeError), 0, 0, 1, 1}, + {&__pyx_kp_s_Unable_to_convert_item_to_object, __pyx_k_Unable_to_convert_item_to_object, sizeof(__pyx_k_Unable_to_convert_item_to_object), 0, 0, 1, 0}, + {&__pyx_n_s_ValueError, __pyx_k_ValueError, sizeof(__pyx_k_ValueError), 0, 0, 1, 1}, + {&__pyx_n_s_View_MemoryView, __pyx_k_View_MemoryView, sizeof(__pyx_k_View_MemoryView), 0, 0, 1, 1}, + {&__pyx_kp_u__2, __pyx_k__2, sizeof(__pyx_k__2), 0, 1, 0, 0}, + {&__pyx_n_s__26, __pyx_k__26, sizeof(__pyx_k__26), 0, 0, 1, 1}, + {&__pyx_n_s__3, __pyx_k__3, sizeof(__pyx_k__3), 0, 0, 1, 1}, + {&__pyx_kp_u__6, __pyx_k__6, sizeof(__pyx_k__6), 0, 1, 0, 0}, + {&__pyx_kp_u__7, __pyx_k__7, sizeof(__pyx_k__7), 0, 1, 0, 0}, + {&__pyx_n_s_abc, __pyx_k_abc, sizeof(__pyx_k_abc), 0, 0, 1, 1}, + {&__pyx_n_s_allocate_buffer, __pyx_k_allocate_buffer, sizeof(__pyx_k_allocate_buffer), 0, 0, 1, 1}, + {&__pyx_kp_u_and, __pyx_k_and, sizeof(__pyx_k_and), 0, 1, 0, 0}, + {&__pyx_n_s_asyncio_coroutines, __pyx_k_asyncio_coroutines, sizeof(__pyx_k_asyncio_coroutines), 0, 0, 1, 1}, + {&__pyx_n_s_base, __pyx_k_base, sizeof(__pyx_k_base), 0, 0, 1, 1}, + {&__pyx_n_s_batch_by_size_fast, __pyx_k_batch_by_size_fast, sizeof(__pyx_k_batch_by_size_fast), 0, 0, 1, 1}, + {&__pyx_n_s_batch_fixed_shapes_fast, __pyx_k_batch_fixed_shapes_fast, sizeof(__pyx_k_batch_fixed_shapes_fast), 0, 0, 1, 1}, + {&__pyx_n_s_bsz_mult, __pyx_k_bsz_mult, sizeof(__pyx_k_bsz_mult), 0, 0, 1, 1}, + {&__pyx_n_s_c, __pyx_k_c, sizeof(__pyx_k_c), 0, 0, 1, 1}, + {&__pyx_n_u_c, __pyx_k_c, sizeof(__pyx_k_c), 0, 1, 0, 1}, + {&__pyx_n_s_class, __pyx_k_class, sizeof(__pyx_k_class), 0, 0, 1, 1}, + {&__pyx_n_s_class_getitem, __pyx_k_class_getitem, sizeof(__pyx_k_class_getitem), 0, 0, 1, 1}, + {&__pyx_n_s_cline_in_traceback, __pyx_k_cline_in_traceback, sizeof(__pyx_k_cline_in_traceback), 0, 0, 1, 1}, + {&__pyx_n_s_collections, __pyx_k_collections, sizeof(__pyx_k_collections), 0, 0, 1, 1}, + {&__pyx_kp_s_collections_abc, __pyx_k_collections_abc, sizeof(__pyx_k_collections_abc), 0, 0, 1, 0}, + {&__pyx_kp_s_contiguous_and_direct, __pyx_k_contiguous_and_direct, sizeof(__pyx_k_contiguous_and_direct), 0, 0, 1, 0}, + {&__pyx_kp_s_contiguous_and_indirect, __pyx_k_contiguous_and_indirect, sizeof(__pyx_k_contiguous_and_indirect), 0, 0, 1, 0}, + {&__pyx_n_s_count, __pyx_k_count, sizeof(__pyx_k_count), 0, 0, 1, 1}, + {&__pyx_n_s_dict, __pyx_k_dict, sizeof(__pyx_k_dict), 0, 0, 1, 1}, + {&__pyx_kp_u_disable, __pyx_k_disable, sizeof(__pyx_k_disable), 0, 1, 0, 0}, + {&__pyx_n_s_dtype_is_object, __pyx_k_dtype_is_object, sizeof(__pyx_k_dtype_is_object), 0, 0, 1, 1}, + {&__pyx_kp_u_enable, __pyx_k_enable, sizeof(__pyx_k_enable), 0, 1, 0, 0}, + {&__pyx_n_s_encode, __pyx_k_encode, sizeof(__pyx_k_encode), 0, 0, 1, 1}, + {&__pyx_n_s_enumerate, __pyx_k_enumerate, sizeof(__pyx_k_enumerate), 0, 0, 1, 1}, + {&__pyx_n_s_error, __pyx_k_error, sizeof(__pyx_k_error), 0, 0, 1, 1}, + {&__pyx_n_s_fairseq_data_data_utils_fast, __pyx_k_fairseq_data_data_utils_fast, sizeof(__pyx_k_fairseq_data_data_utils_fast), 0, 0, 1, 1}, + {&__pyx_kp_s_fairseq_data_data_utils_fast_pyx, __pyx_k_fairseq_data_data_utils_fast_pyx, sizeof(__pyx_k_fairseq_data_data_utils_fast_pyx), 0, 0, 1, 0}, + {&__pyx_n_s_fixed_shapes_sorted, __pyx_k_fixed_shapes_sorted, sizeof(__pyx_k_fixed_shapes_sorted), 0, 0, 1, 1}, + {&__pyx_n_s_flags, __pyx_k_flags, sizeof(__pyx_k_flags), 0, 0, 1, 1}, + {&__pyx_n_s_format, __pyx_k_format, sizeof(__pyx_k_format), 0, 0, 1, 1}, + {&__pyx_n_s_fortran, __pyx_k_fortran, sizeof(__pyx_k_fortran), 0, 0, 1, 1}, + {&__pyx_n_u_fortran, __pyx_k_fortran, sizeof(__pyx_k_fortran), 0, 1, 0, 1}, + {&__pyx_kp_u_gc, __pyx_k_gc, sizeof(__pyx_k_gc), 0, 1, 0, 0}, + {&__pyx_n_s_getstate, __pyx_k_getstate, sizeof(__pyx_k_getstate), 0, 0, 1, 1}, + {&__pyx_kp_u_got, __pyx_k_got, sizeof(__pyx_k_got), 0, 1, 0, 0}, + {&__pyx_kp_u_got_differing_extents_in_dimensi, __pyx_k_got_differing_extents_in_dimensi, sizeof(__pyx_k_got_differing_extents_in_dimensi), 0, 1, 0, 0}, + {&__pyx_n_s_id, __pyx_k_id, sizeof(__pyx_k_id), 0, 0, 1, 1}, + {&__pyx_n_s_import, __pyx_k_import, sizeof(__pyx_k_import), 0, 0, 1, 1}, + {&__pyx_n_s_index, __pyx_k_index, sizeof(__pyx_k_index), 0, 0, 1, 1}, + {&__pyx_n_s_indices, __pyx_k_indices, sizeof(__pyx_k_indices), 0, 0, 1, 1}, + {&__pyx_n_s_initializing, __pyx_k_initializing, sizeof(__pyx_k_initializing), 0, 0, 1, 1}, + {&__pyx_n_s_int64, __pyx_k_int64, sizeof(__pyx_k_int64), 0, 0, 1, 1}, + {&__pyx_n_s_is_coroutine, __pyx_k_is_coroutine, sizeof(__pyx_k_is_coroutine), 0, 0, 1, 1}, + {&__pyx_kp_u_isenabled, __pyx_k_isenabled, sizeof(__pyx_k_isenabled), 0, 1, 0, 0}, + {&__pyx_n_s_itemsize, __pyx_k_itemsize, sizeof(__pyx_k_itemsize), 0, 0, 1, 1}, + {&__pyx_kp_s_itemsize_0_for_cython_array, __pyx_k_itemsize_0_for_cython_array, sizeof(__pyx_k_itemsize_0_for_cython_array), 0, 0, 1, 0}, + {&__pyx_n_s_main, __pyx_k_main, sizeof(__pyx_k_main), 0, 0, 1, 1}, + {&__pyx_n_s_max, __pyx_k_max, sizeof(__pyx_k_max), 0, 0, 1, 1}, + {&__pyx_n_s_max_sentences, __pyx_k_max_sentences, sizeof(__pyx_k_max_sentences), 0, 0, 1, 1}, + {&__pyx_n_s_max_tokens, __pyx_k_max_tokens, sizeof(__pyx_k_max_tokens), 0, 0, 1, 1}, + {&__pyx_n_s_memview, __pyx_k_memview, sizeof(__pyx_k_memview), 0, 0, 1, 1}, + {&__pyx_n_s_mode, __pyx_k_mode, sizeof(__pyx_k_mode), 0, 0, 1, 1}, + {&__pyx_n_s_name, __pyx_k_name, sizeof(__pyx_k_name), 0, 0, 1, 1}, + {&__pyx_n_s_name_2, __pyx_k_name_2, sizeof(__pyx_k_name_2), 0, 0, 1, 1}, + {&__pyx_n_s_ndim, __pyx_k_ndim, sizeof(__pyx_k_ndim), 0, 0, 1, 1}, + {&__pyx_n_s_new, __pyx_k_new, sizeof(__pyx_k_new), 0, 0, 1, 1}, + {&__pyx_kp_s_no_default___reduce___due_to_non, __pyx_k_no_default___reduce___due_to_non, sizeof(__pyx_k_no_default___reduce___due_to_non), 0, 0, 1, 0}, + {&__pyx_n_s_np, __pyx_k_np, sizeof(__pyx_k_np), 0, 0, 1, 1}, + {&__pyx_n_s_num_tokens_fn, __pyx_k_num_tokens_fn, sizeof(__pyx_k_num_tokens_fn), 0, 0, 1, 1}, + {&__pyx_n_s_numpy, __pyx_k_numpy, sizeof(__pyx_k_numpy), 0, 0, 1, 1}, + {&__pyx_kp_u_numpy__core_multiarray_failed_to, __pyx_k_numpy__core_multiarray_failed_to, sizeof(__pyx_k_numpy__core_multiarray_failed_to), 0, 1, 0, 0}, + {&__pyx_kp_u_numpy__core_umath_failed_to_impo, __pyx_k_numpy__core_umath_failed_to_impo, sizeof(__pyx_k_numpy__core_umath_failed_to_impo), 0, 1, 0, 0}, + {&__pyx_n_s_obj, __pyx_k_obj, sizeof(__pyx_k_obj), 0, 0, 1, 1}, + {&__pyx_n_s_pack, __pyx_k_pack, sizeof(__pyx_k_pack), 0, 0, 1, 1}, + {&__pyx_n_s_pickle, __pyx_k_pickle, sizeof(__pyx_k_pickle), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_PickleError, __pyx_k_pyx_PickleError, sizeof(__pyx_k_pyx_PickleError), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_checksum, __pyx_k_pyx_checksum, sizeof(__pyx_k_pyx_checksum), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_result, __pyx_k_pyx_result, sizeof(__pyx_k_pyx_result), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_state, __pyx_k_pyx_state, sizeof(__pyx_k_pyx_state), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_type, __pyx_k_pyx_type, sizeof(__pyx_k_pyx_type), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_unpickle_Enum, __pyx_k_pyx_unpickle_Enum, sizeof(__pyx_k_pyx_unpickle_Enum), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_vtable, __pyx_k_pyx_vtable, sizeof(__pyx_k_pyx_vtable), 0, 0, 1, 1}, + {&__pyx_n_s_range, __pyx_k_range, sizeof(__pyx_k_range), 0, 0, 1, 1}, + {&__pyx_n_s_reduce, __pyx_k_reduce, sizeof(__pyx_k_reduce), 0, 0, 1, 1}, + {&__pyx_n_s_reduce_cython, __pyx_k_reduce_cython, sizeof(__pyx_k_reduce_cython), 0, 0, 1, 1}, + {&__pyx_n_s_reduce_ex, __pyx_k_reduce_ex, sizeof(__pyx_k_reduce_ex), 0, 0, 1, 1}, + {&__pyx_n_s_register, __pyx_k_register, sizeof(__pyx_k_register), 0, 0, 1, 1}, + {&__pyx_kp_u_sentence_at_index_of_size_exceed, __pyx_k_sentence_at_index_of_size_exceed, sizeof(__pyx_k_sentence_at_index_of_size_exceed), 0, 1, 0, 0}, + {&__pyx_n_s_setstate, __pyx_k_setstate, sizeof(__pyx_k_setstate), 0, 0, 1, 1}, + {&__pyx_n_s_setstate_cython, __pyx_k_setstate_cython, sizeof(__pyx_k_setstate_cython), 0, 0, 1, 1}, + {&__pyx_n_s_shape, __pyx_k_shape, sizeof(__pyx_k_shape), 0, 0, 1, 1}, + {&__pyx_n_s_size, __pyx_k_size, sizeof(__pyx_k_size), 0, 0, 1, 1}, + {&__pyx_n_s_spec, __pyx_k_spec, sizeof(__pyx_k_spec), 0, 0, 1, 1}, + {&__pyx_n_s_start, __pyx_k_start, sizeof(__pyx_k_start), 0, 0, 1, 1}, + {&__pyx_n_s_step, __pyx_k_step, sizeof(__pyx_k_step), 0, 0, 1, 1}, + {&__pyx_n_s_stop, __pyx_k_stop, sizeof(__pyx_k_stop), 0, 0, 1, 1}, + {&__pyx_kp_s_strided_and_direct, __pyx_k_strided_and_direct, sizeof(__pyx_k_strided_and_direct), 0, 0, 1, 0}, + {&__pyx_kp_s_strided_and_direct_or_indirect, __pyx_k_strided_and_direct_or_indirect, sizeof(__pyx_k_strided_and_direct_or_indirect), 0, 0, 1, 0}, + {&__pyx_kp_s_strided_and_indirect, __pyx_k_strided_and_indirect, sizeof(__pyx_k_strided_and_indirect), 0, 0, 1, 0}, + {&__pyx_kp_s_stringsource, __pyx_k_stringsource, sizeof(__pyx_k_stringsource), 0, 0, 1, 0}, + {&__pyx_n_s_struct, __pyx_k_struct, sizeof(__pyx_k_struct), 0, 0, 1, 1}, + {&__pyx_n_s_sys, __pyx_k_sys, sizeof(__pyx_k_sys), 0, 0, 1, 1}, + {&__pyx_n_s_test, __pyx_k_test, sizeof(__pyx_k_test), 0, 0, 1, 1}, + {&__pyx_kp_s_unable_to_allocate_array_data, __pyx_k_unable_to_allocate_array_data, sizeof(__pyx_k_unable_to_allocate_array_data), 0, 0, 1, 0}, + {&__pyx_kp_s_unable_to_allocate_shape_and_str, __pyx_k_unable_to_allocate_shape_and_str, sizeof(__pyx_k_unable_to_allocate_shape_and_str), 0, 0, 1, 0}, + {&__pyx_n_s_unpack, __pyx_k_unpack, sizeof(__pyx_k_unpack), 0, 0, 1, 1}, + {&__pyx_n_s_update, __pyx_k_update, sizeof(__pyx_k_update), 0, 0, 1, 1}, + {&__pyx_n_s_version_info, __pyx_k_version_info, sizeof(__pyx_k_version_info), 0, 0, 1, 1}, + {0, 0, 0, 0, 0, 0, 0} + }; + return __Pyx_InitStrings(__pyx_string_tab); +} +/* #### Code section: cached_builtins ### */ +static CYTHON_SMALL_CODE int __Pyx_InitCachedBuiltins(void) { + __pyx_builtin_range = __Pyx_GetBuiltinName(__pyx_n_s_range); if (!__pyx_builtin_range) __PYX_ERR(0, 44, __pyx_L1_error) + __pyx_builtin_AssertionError = __Pyx_GetBuiltinName(__pyx_n_s_AssertionError); if (!__pyx_builtin_AssertionError) __PYX_ERR(0, 50, __pyx_L1_error) + __pyx_builtin_max = __Pyx_GetBuiltinName(__pyx_n_s_max); if (!__pyx_builtin_max) __PYX_ERR(0, 64, __pyx_L1_error) + __pyx_builtin___import__ = __Pyx_GetBuiltinName(__pyx_n_s_import); if (!__pyx_builtin___import__) __PYX_ERR(1, 100, __pyx_L1_error) + __pyx_builtin_ValueError = __Pyx_GetBuiltinName(__pyx_n_s_ValueError); if (!__pyx_builtin_ValueError) __PYX_ERR(1, 141, __pyx_L1_error) + __pyx_builtin_MemoryError = __Pyx_GetBuiltinName(__pyx_n_s_MemoryError); if (!__pyx_builtin_MemoryError) __PYX_ERR(1, 156, __pyx_L1_error) + __pyx_builtin_enumerate = __Pyx_GetBuiltinName(__pyx_n_s_enumerate); if (!__pyx_builtin_enumerate) __PYX_ERR(1, 159, __pyx_L1_error) + __pyx_builtin_TypeError = __Pyx_GetBuiltinName(__pyx_n_s_TypeError); if (!__pyx_builtin_TypeError) __PYX_ERR(1, 2, __pyx_L1_error) + __pyx_builtin_Ellipsis = __Pyx_GetBuiltinName(__pyx_n_s_Ellipsis); if (!__pyx_builtin_Ellipsis) __PYX_ERR(1, 408, __pyx_L1_error) + __pyx_builtin_id = __Pyx_GetBuiltinName(__pyx_n_s_id); if (!__pyx_builtin_id) __PYX_ERR(1, 618, __pyx_L1_error) + __pyx_builtin_IndexError = __Pyx_GetBuiltinName(__pyx_n_s_IndexError); if (!__pyx_builtin_IndexError) __PYX_ERR(1, 914, __pyx_L1_error) + __pyx_builtin_ImportError = __Pyx_GetBuiltinName(__pyx_n_s_ImportError); if (!__pyx_builtin_ImportError) __PYX_ERR(2, 1043, __pyx_L1_error) + return 0; + __pyx_L1_error:; + return -1; +} +/* #### Code section: cached_constants ### */ + +static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_InitCachedConstants", 0); + + /* "View.MemoryView":582 + * def suboffsets(self): + * if self.view.suboffsets == NULL: + * return (-1,) * self.view.ndim # <<<<<<<<<<<<<< + * + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + */ + __pyx_tuple__4 = PyTuple_New(1); if (unlikely(!__pyx_tuple__4)) __PYX_ERR(1, 582, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__4); + __Pyx_INCREF(__pyx_int_neg_1); + __Pyx_GIVEREF(__pyx_int_neg_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_tuple__4, 0, __pyx_int_neg_1)) __PYX_ERR(1, 582, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_tuple__4); + + /* "View.MemoryView":679 + * tup = index if isinstance(index, tuple) else (index,) + * + * result = [slice(None)] * ndim # <<<<<<<<<<<<<< + * have_slices = False + * seen_ellipsis = False + */ + __pyx_slice__5 = PySlice_New(Py_None, Py_None, Py_None); if (unlikely(!__pyx_slice__5)) __PYX_ERR(1, 679, __pyx_L1_error) + __Pyx_GOTREF(__pyx_slice__5); + __Pyx_GIVEREF(__pyx_slice__5); + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + */ + __pyx_tuple__8 = PyTuple_Pack(3, __pyx_int_136983863, __pyx_int_112105877, __pyx_int_184977713); if (unlikely(!__pyx_tuple__8)) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__8); + __Pyx_GIVEREF(__pyx_tuple__8); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1043 + * __pyx_import_array() + * except Exception: + * raise ImportError("numpy._core.multiarray failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_umath() except -1: + */ + __pyx_tuple__9 = PyTuple_Pack(1, __pyx_kp_u_numpy__core_multiarray_failed_to); if (unlikely(!__pyx_tuple__9)) __PYX_ERR(2, 1043, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__9); + __Pyx_GIVEREF(__pyx_tuple__9); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1049 + * _import_umath() + * except Exception: + * raise ImportError("numpy._core.umath failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_ufunc() except -1: + */ + __pyx_tuple__10 = PyTuple_Pack(1, __pyx_kp_u_numpy__core_umath_failed_to_impo); if (unlikely(!__pyx_tuple__10)) __PYX_ERR(2, 1049, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__10); + __Pyx_GIVEREF(__pyx_tuple__10); + + /* "View.MemoryView":100 + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: + * if __import__("sys").version_info >= (3, 3): # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + */ + __pyx_tuple__11 = PyTuple_Pack(1, __pyx_n_s_sys); if (unlikely(!__pyx_tuple__11)) __PYX_ERR(1, 100, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__11); + __Pyx_GIVEREF(__pyx_tuple__11); + __pyx_tuple__12 = PyTuple_Pack(2, __pyx_int_3, __pyx_int_3); if (unlikely(!__pyx_tuple__12)) __PYX_ERR(1, 100, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__12); + __Pyx_GIVEREF(__pyx_tuple__12); + + /* "View.MemoryView":101 + * try: + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence # <<<<<<<<<<<<<< + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence + */ + __pyx_tuple__13 = PyTuple_Pack(1, __pyx_kp_s_collections_abc); if (unlikely(!__pyx_tuple__13)) __PYX_ERR(1, 101, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__13); + __Pyx_GIVEREF(__pyx_tuple__13); + + /* "View.MemoryView":103 + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence # <<<<<<<<<<<<<< + * except: + * + */ + __pyx_tuple__14 = PyTuple_Pack(1, __pyx_n_s_collections); if (unlikely(!__pyx_tuple__14)) __PYX_ERR(1, 103, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__14); + __Pyx_GIVEREF(__pyx_tuple__14); + + /* "View.MemoryView":309 + * return self.name + * + * cdef generic = Enum("") # <<<<<<<<<<<<<< + * cdef strided = Enum("") # default + * cdef indirect = Enum("") + */ + __pyx_tuple__15 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct_or_indirect); if (unlikely(!__pyx_tuple__15)) __PYX_ERR(1, 309, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__15); + __Pyx_GIVEREF(__pyx_tuple__15); + + /* "View.MemoryView":310 + * + * cdef generic = Enum("") + * cdef strided = Enum("") # default # <<<<<<<<<<<<<< + * cdef indirect = Enum("") + * + */ + __pyx_tuple__16 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct); if (unlikely(!__pyx_tuple__16)) __PYX_ERR(1, 310, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__16); + __Pyx_GIVEREF(__pyx_tuple__16); + + /* "View.MemoryView":311 + * cdef generic = Enum("") + * cdef strided = Enum("") # default + * cdef indirect = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_tuple__17 = PyTuple_Pack(1, __pyx_kp_s_strided_and_indirect); if (unlikely(!__pyx_tuple__17)) __PYX_ERR(1, 311, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__17); + __Pyx_GIVEREF(__pyx_tuple__17); + + /* "View.MemoryView":314 + * + * + * cdef contiguous = Enum("") # <<<<<<<<<<<<<< + * cdef indirect_contiguous = Enum("") + * + */ + __pyx_tuple__18 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_direct); if (unlikely(!__pyx_tuple__18)) __PYX_ERR(1, 314, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__18); + __Pyx_GIVEREF(__pyx_tuple__18); + + /* "View.MemoryView":315 + * + * cdef contiguous = Enum("") + * cdef indirect_contiguous = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_tuple__19 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_indirect); if (unlikely(!__pyx_tuple__19)) __PYX_ERR(1, 315, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__19); + __Pyx_GIVEREF(__pyx_tuple__19); + + /* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + __pyx_tuple__20 = PyTuple_Pack(5, __pyx_n_s_pyx_type, __pyx_n_s_pyx_checksum, __pyx_n_s_pyx_state, __pyx_n_s_pyx_PickleError, __pyx_n_s_pyx_result); if (unlikely(!__pyx_tuple__20)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__20); + __Pyx_GIVEREF(__pyx_tuple__20); + __pyx_codeobj__21 = (PyObject*)__Pyx_PyCode_New(3, 0, 0, 5, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__20, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_stringsource, __pyx_n_s_pyx_unpickle_Enum, 1, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__21)) __PYX_ERR(1, 1, __pyx_L1_error) + + /* "fairseq/data/data_utils_fast.pyx":27 + * + * @cython.cdivision(True) + * cpdef list batch_by_size_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + __pyx_tuple__22 = PyTuple_Pack(5, __pyx_n_s_indices, __pyx_n_s_num_tokens_fn, __pyx_n_s_max_tokens, __pyx_n_s_max_sentences, __pyx_n_s_bsz_mult); if (unlikely(!__pyx_tuple__22)) __PYX_ERR(0, 27, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__22); + __Pyx_GIVEREF(__pyx_tuple__22); + __pyx_codeobj__23 = (PyObject*)__Pyx_PyCode_New(5, 0, 0, 5, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__22, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_fairseq_data_data_utils_fast_pyx, __pyx_n_s_batch_by_size_fast, 27, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__23)) __PYX_ERR(0, 27, __pyx_L1_error) + + /* "fairseq/data/data_utils_fast.pyx":84 + * + * @cython.cdivision(True) + * cpdef list batch_fixed_shapes_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + __pyx_tuple__24 = PyTuple_Pack(3, __pyx_n_s_indices, __pyx_n_s_num_tokens_fn, __pyx_n_s_fixed_shapes_sorted); if (unlikely(!__pyx_tuple__24)) __PYX_ERR(0, 84, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__24); + __Pyx_GIVEREF(__pyx_tuple__24); + __pyx_codeobj__25 = (PyObject*)__Pyx_PyCode_New(3, 0, 0, 3, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__24, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_fairseq_data_data_utils_fast_pyx, __pyx_n_s_batch_fixed_shapes_fast, 84, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__25)) __PYX_ERR(0, 84, __pyx_L1_error) + __Pyx_RefNannyFinishContext(); + return 0; + __pyx_L1_error:; + __Pyx_RefNannyFinishContext(); + return -1; +} +/* #### Code section: init_constants ### */ + +static CYTHON_SMALL_CODE int __Pyx_InitConstants(void) { + if (__Pyx_CreateStringTabAndInitStrings() < 0) __PYX_ERR(0, 1, __pyx_L1_error); + __pyx_int_0 = PyInt_FromLong(0); if (unlikely(!__pyx_int_0)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_1 = PyInt_FromLong(1); if (unlikely(!__pyx_int_1)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_3 = PyInt_FromLong(3); if (unlikely(!__pyx_int_3)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_112105877 = PyInt_FromLong(112105877L); if (unlikely(!__pyx_int_112105877)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_136983863 = PyInt_FromLong(136983863L); if (unlikely(!__pyx_int_136983863)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_184977713 = PyInt_FromLong(184977713L); if (unlikely(!__pyx_int_184977713)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_neg_1 = PyInt_FromLong(-1); if (unlikely(!__pyx_int_neg_1)) __PYX_ERR(0, 1, __pyx_L1_error) + return 0; + __pyx_L1_error:; + return -1; +} +/* #### Code section: init_globals ### */ + +static CYTHON_SMALL_CODE int __Pyx_InitGlobals(void) { + /* AssertionsEnabled.init */ + if (likely(__Pyx_init_assertions_enabled() == 0)); else + +if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 1, __pyx_L1_error) + + /* NumpyImportArray.init */ + /* + * Cython has automatically inserted a call to _import_array since + * you didn't include one when you cimported numpy. To disable this + * add the line + * numpy._import_array + */ +#ifdef NPY_FEATURE_VERSION +#ifndef NO_IMPORT_ARRAY +if (unlikely(_import_array() == -1)) { + PyErr_SetString(PyExc_ImportError, "numpy.core.multiarray failed to import " + "(auto-generated because you didn't call 'numpy.import_array()' after cimporting numpy; " + "use 'numpy._import_array' to disable if you are certain you don't need it)."); +} +#endif +#endif + +if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 1, __pyx_L1_error) + + return 0; + __pyx_L1_error:; + return -1; +} +/* #### Code section: init_module ### */ + +static CYTHON_SMALL_CODE int __Pyx_modinit_global_init_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_variable_export_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_function_export_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_type_init_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_type_import_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_variable_import_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_function_import_code(void); /*proto*/ + +static int __Pyx_modinit_global_init_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_global_init_code", 0); + /*--- Global init code ---*/ + __pyx_collections_abc_Sequence = Py_None; Py_INCREF(Py_None); + generic = Py_None; Py_INCREF(Py_None); + strided = Py_None; Py_INCREF(Py_None); + indirect = Py_None; Py_INCREF(Py_None); + contiguous = Py_None; Py_INCREF(Py_None); + indirect_contiguous = Py_None; Py_INCREF(Py_None); + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_variable_export_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_variable_export_code", 0); + /*--- Variable export code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_function_export_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_function_export_code", 0); + /*--- Function export code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_type_init_code(void) { + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__Pyx_modinit_type_init_code", 0); + /*--- Type init code ---*/ + __pyx_vtabptr_array = &__pyx_vtable_array; + __pyx_vtable_array.get_memview = (PyObject *(*)(struct __pyx_array_obj *))__pyx_array_get_memview; + #if CYTHON_USE_TYPE_SPECS + __pyx_array_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_array_spec, NULL); if (unlikely(!__pyx_array_type)) __PYX_ERR(1, 114, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + __pyx_array_type->tp_as_buffer = &__pyx_tp_as_buffer_array; + if (!__pyx_array_type->tp_as_buffer->bf_releasebuffer && __pyx_array_type->tp_base->tp_as_buffer && __pyx_array_type->tp_base->tp_as_buffer->bf_releasebuffer) { + __pyx_array_type->tp_as_buffer->bf_releasebuffer = __pyx_array_type->tp_base->tp_as_buffer->bf_releasebuffer; + } + #elif defined(Py_bf_getbuffer) && defined(Py_bf_releasebuffer) + /* PY_VERSION_HEX >= 0x03090000 || Py_LIMITED_API >= 0x030B0000 */ + #elif defined(_MSC_VER) + #pragma message ("The buffer protocol is not supported in the Limited C-API < 3.11.") + #else + #warning "The buffer protocol is not supported in the Limited C-API < 3.11." + #endif + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_array_spec, __pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #else + __pyx_array_type = &__pyx_type___pyx_array; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_array_type->tp_print = 0; + #endif + if (__Pyx_SetVtable(__pyx_array_type, __pyx_vtabptr_array) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_MergeVtables(__pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #endif + #if CYTHON_USE_TYPE_SPECS + __pyx_MemviewEnum_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_MemviewEnum_spec, NULL); if (unlikely(!__pyx_MemviewEnum_type)) __PYX_ERR(1, 302, __pyx_L1_error) + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_MemviewEnum_spec, __pyx_MemviewEnum_type) < 0) __PYX_ERR(1, 302, __pyx_L1_error) + #else + __pyx_MemviewEnum_type = &__pyx_type___pyx_MemviewEnum; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_MemviewEnum_type) < 0) __PYX_ERR(1, 302, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_MemviewEnum_type->tp_print = 0; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_MemviewEnum_type->tp_dictoffset && __pyx_MemviewEnum_type->tp_getattro == PyObject_GenericGetAttr)) { + __pyx_MemviewEnum_type->tp_getattro = __Pyx_PyObject_GenericGetAttr; + } + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_MemviewEnum_type) < 0) __PYX_ERR(1, 302, __pyx_L1_error) + #endif + __pyx_vtabptr_memoryview = &__pyx_vtable_memoryview; + __pyx_vtable_memoryview.get_item_pointer = (char *(*)(struct __pyx_memoryview_obj *, PyObject *))__pyx_memoryview_get_item_pointer; + __pyx_vtable_memoryview.is_slice = (PyObject *(*)(struct __pyx_memoryview_obj *, PyObject *))__pyx_memoryview_is_slice; + __pyx_vtable_memoryview.setitem_slice_assignment = (PyObject *(*)(struct __pyx_memoryview_obj *, PyObject *, PyObject *))__pyx_memoryview_setitem_slice_assignment; + __pyx_vtable_memoryview.setitem_slice_assign_scalar = (PyObject *(*)(struct __pyx_memoryview_obj *, struct __pyx_memoryview_obj *, PyObject *))__pyx_memoryview_setitem_slice_assign_scalar; + __pyx_vtable_memoryview.setitem_indexed = (PyObject *(*)(struct __pyx_memoryview_obj *, PyObject *, PyObject *))__pyx_memoryview_setitem_indexed; + __pyx_vtable_memoryview.convert_item_to_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *))__pyx_memoryview_convert_item_to_object; + __pyx_vtable_memoryview.assign_item_from_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *, PyObject *))__pyx_memoryview_assign_item_from_object; + __pyx_vtable_memoryview._get_base = (PyObject *(*)(struct __pyx_memoryview_obj *))__pyx_memoryview__get_base; + #if CYTHON_USE_TYPE_SPECS + __pyx_memoryview_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_memoryview_spec, NULL); if (unlikely(!__pyx_memoryview_type)) __PYX_ERR(1, 337, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + __pyx_memoryview_type->tp_as_buffer = &__pyx_tp_as_buffer_memoryview; + if (!__pyx_memoryview_type->tp_as_buffer->bf_releasebuffer && __pyx_memoryview_type->tp_base->tp_as_buffer && __pyx_memoryview_type->tp_base->tp_as_buffer->bf_releasebuffer) { + __pyx_memoryview_type->tp_as_buffer->bf_releasebuffer = __pyx_memoryview_type->tp_base->tp_as_buffer->bf_releasebuffer; + } + #elif defined(Py_bf_getbuffer) && defined(Py_bf_releasebuffer) + /* PY_VERSION_HEX >= 0x03090000 || Py_LIMITED_API >= 0x030B0000 */ + #elif defined(_MSC_VER) + #pragma message ("The buffer protocol is not supported in the Limited C-API < 3.11.") + #else + #warning "The buffer protocol is not supported in the Limited C-API < 3.11." + #endif + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_memoryview_spec, __pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #else + __pyx_memoryview_type = &__pyx_type___pyx_memoryview; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_memoryview_type->tp_print = 0; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_memoryview_type->tp_dictoffset && __pyx_memoryview_type->tp_getattro == PyObject_GenericGetAttr)) { + __pyx_memoryview_type->tp_getattro = __Pyx_PyObject_GenericGetAttr; + } + #endif + if (__Pyx_SetVtable(__pyx_memoryview_type, __pyx_vtabptr_memoryview) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_MergeVtables(__pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #endif + __pyx_vtabptr__memoryviewslice = &__pyx_vtable__memoryviewslice; + __pyx_vtable__memoryviewslice.__pyx_base = *__pyx_vtabptr_memoryview; + __pyx_vtable__memoryviewslice.__pyx_base.convert_item_to_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *))__pyx_memoryviewslice_convert_item_to_object; + __pyx_vtable__memoryviewslice.__pyx_base.assign_item_from_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *, PyObject *))__pyx_memoryviewslice_assign_item_from_object; + __pyx_vtable__memoryviewslice.__pyx_base._get_base = (PyObject *(*)(struct __pyx_memoryview_obj *))__pyx_memoryviewslice__get_base; + #if CYTHON_USE_TYPE_SPECS + __pyx_t_1 = PyTuple_Pack(1, (PyObject *)__pyx_memoryview_type); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 952, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_memoryviewslice_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_memoryviewslice_spec, __pyx_t_1); + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + if (unlikely(!__pyx_memoryviewslice_type)) __PYX_ERR(1, 952, __pyx_L1_error) + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_memoryviewslice_spec, __pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #else + __pyx_memoryviewslice_type = &__pyx_type___pyx_memoryviewslice; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + __pyx_memoryviewslice_type->tp_base = __pyx_memoryview_type; + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_memoryviewslice_type->tp_print = 0; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_memoryviewslice_type->tp_dictoffset && __pyx_memoryviewslice_type->tp_getattro == PyObject_GenericGetAttr)) { + __pyx_memoryviewslice_type->tp_getattro = __Pyx_PyObject_GenericGetAttr; + } + #endif + if (__Pyx_SetVtable(__pyx_memoryviewslice_type, __pyx_vtabptr__memoryviewslice) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_MergeVtables(__pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #endif + __Pyx_RefNannyFinishContext(); + return 0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_RefNannyFinishContext(); + return -1; +} + +static int __Pyx_modinit_type_import_code(void) { + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__Pyx_modinit_type_import_code", 0); + /*--- Type import code ---*/ + __pyx_t_1 = PyImport_ImportModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_t_1)) __PYX_ERR(3, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_ptype_7cpython_4type_type = __Pyx_ImportType_3_0_12(__pyx_t_1, __Pyx_BUILTIN_MODULE_NAME, "type", + #if defined(PYPY_VERSION_NUM) && PYPY_VERSION_NUM < 0x050B0000 + sizeof(PyTypeObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyTypeObject), + #elif CYTHON_COMPILING_IN_LIMITED_API + sizeof(PyTypeObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyTypeObject), + #else + sizeof(PyHeapTypeObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyHeapTypeObject), + #endif + __Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_7cpython_4type_type) __PYX_ERR(3, 9, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = PyImport_ImportModule("numpy"); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 272, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_ptype_5numpy_dtype = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "dtype", sizeof(PyArray_Descr), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArray_Descr),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_dtype) __PYX_ERR(2, 272, __pyx_L1_error) + __pyx_ptype_5numpy_flatiter = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "flatiter", sizeof(PyArrayIterObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArrayIterObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_flatiter) __PYX_ERR(2, 317, __pyx_L1_error) + __pyx_ptype_5numpy_broadcast = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "broadcast", sizeof(PyArrayMultiIterObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArrayMultiIterObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_broadcast) __PYX_ERR(2, 321, __pyx_L1_error) + __pyx_ptype_5numpy_ndarray = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "ndarray", sizeof(PyArrayObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArrayObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_ndarray) __PYX_ERR(2, 360, __pyx_L1_error) + __pyx_ptype_5numpy_generic = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "generic", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_generic) __PYX_ERR(2, 865, __pyx_L1_error) + __pyx_ptype_5numpy_number = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "number", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_number) __PYX_ERR(2, 867, __pyx_L1_error) + __pyx_ptype_5numpy_integer = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "integer", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_integer) __PYX_ERR(2, 869, __pyx_L1_error) + __pyx_ptype_5numpy_signedinteger = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "signedinteger", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_signedinteger) __PYX_ERR(2, 871, __pyx_L1_error) + __pyx_ptype_5numpy_unsignedinteger = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "unsignedinteger", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_unsignedinteger) __PYX_ERR(2, 873, __pyx_L1_error) + __pyx_ptype_5numpy_inexact = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "inexact", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_inexact) __PYX_ERR(2, 875, __pyx_L1_error) + __pyx_ptype_5numpy_floating = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "floating", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_floating) __PYX_ERR(2, 877, __pyx_L1_error) + __pyx_ptype_5numpy_complexfloating = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "complexfloating", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_complexfloating) __PYX_ERR(2, 879, __pyx_L1_error) + __pyx_ptype_5numpy_flexible = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "flexible", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_flexible) __PYX_ERR(2, 881, __pyx_L1_error) + __pyx_ptype_5numpy_character = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "character", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_character) __PYX_ERR(2, 883, __pyx_L1_error) + __pyx_ptype_5numpy_ufunc = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "ufunc", sizeof(PyUFuncObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyUFuncObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_ufunc) __PYX_ERR(2, 947, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_RefNannyFinishContext(); + return 0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_RefNannyFinishContext(); + return -1; +} + +static int __Pyx_modinit_variable_import_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_variable_import_code", 0); + /*--- Variable import code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_function_import_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_function_import_code", 0); + /*--- Function import code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + + +#if PY_MAJOR_VERSION >= 3 +#if CYTHON_PEP489_MULTI_PHASE_INIT +static PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def); /*proto*/ +static int __pyx_pymod_exec_data_utils_fast(PyObject* module); /*proto*/ +static PyModuleDef_Slot __pyx_moduledef_slots[] = { + {Py_mod_create, (void*)__pyx_pymod_create}, + {Py_mod_exec, (void*)__pyx_pymod_exec_data_utils_fast}, + {0, NULL} +}; +#endif + +#ifdef __cplusplus +namespace { + struct PyModuleDef __pyx_moduledef = + #else + static struct PyModuleDef __pyx_moduledef = + #endif + { + PyModuleDef_HEAD_INIT, + "data_utils_fast", + 0, /* m_doc */ + #if CYTHON_PEP489_MULTI_PHASE_INIT + 0, /* m_size */ + #elif CYTHON_USE_MODULE_STATE + sizeof(__pyx_mstate), /* m_size */ + #else + -1, /* m_size */ + #endif + __pyx_methods /* m_methods */, + #if CYTHON_PEP489_MULTI_PHASE_INIT + __pyx_moduledef_slots, /* m_slots */ + #else + NULL, /* m_reload */ + #endif + #if CYTHON_USE_MODULE_STATE + __pyx_m_traverse, /* m_traverse */ + __pyx_m_clear, /* m_clear */ + NULL /* m_free */ + #else + NULL, /* m_traverse */ + NULL, /* m_clear */ + NULL /* m_free */ + #endif + }; + #ifdef __cplusplus +} /* anonymous namespace */ +#endif +#endif + +#ifndef CYTHON_NO_PYINIT_EXPORT +#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC +#elif PY_MAJOR_VERSION < 3 +#ifdef __cplusplus +#define __Pyx_PyMODINIT_FUNC extern "C" void +#else +#define __Pyx_PyMODINIT_FUNC void +#endif +#else +#ifdef __cplusplus +#define __Pyx_PyMODINIT_FUNC extern "C" PyObject * +#else +#define __Pyx_PyMODINIT_FUNC PyObject * +#endif +#endif + + +#if PY_MAJOR_VERSION < 3 +__Pyx_PyMODINIT_FUNC initdata_utils_fast(void) CYTHON_SMALL_CODE; /*proto*/ +__Pyx_PyMODINIT_FUNC initdata_utils_fast(void) +#else +__Pyx_PyMODINIT_FUNC PyInit_data_utils_fast(void) CYTHON_SMALL_CODE; /*proto*/ +__Pyx_PyMODINIT_FUNC PyInit_data_utils_fast(void) +#if CYTHON_PEP489_MULTI_PHASE_INIT +{ + return PyModuleDef_Init(&__pyx_moduledef); +} +static CYTHON_SMALL_CODE int __Pyx_check_single_interpreter(void) { + #if PY_VERSION_HEX >= 0x030700A1 + static PY_INT64_T main_interpreter_id = -1; + PY_INT64_T current_id = PyInterpreterState_GetID(PyThreadState_Get()->interp); + if (main_interpreter_id == -1) { + main_interpreter_id = current_id; + return (unlikely(current_id == -1)) ? -1 : 0; + } else if (unlikely(main_interpreter_id != current_id)) + #else + static PyInterpreterState *main_interpreter = NULL; + PyInterpreterState *current_interpreter = PyThreadState_Get()->interp; + if (!main_interpreter) { + main_interpreter = current_interpreter; + } else if (unlikely(main_interpreter != current_interpreter)) + #endif + { + PyErr_SetString( + PyExc_ImportError, + "Interpreter change detected - this module can only be loaded into one interpreter per process."); + return -1; + } + return 0; +} +#if CYTHON_COMPILING_IN_LIMITED_API +static CYTHON_SMALL_CODE int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *module, const char* from_name, const char* to_name, int allow_none) +#else +static CYTHON_SMALL_CODE int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *moddict, const char* from_name, const char* to_name, int allow_none) +#endif +{ + PyObject *value = PyObject_GetAttrString(spec, from_name); + int result = 0; + if (likely(value)) { + if (allow_none || value != Py_None) { +#if CYTHON_COMPILING_IN_LIMITED_API + result = PyModule_AddObject(module, to_name, value); +#else + result = PyDict_SetItemString(moddict, to_name, value); +#endif + } + Py_DECREF(value); + } else if (PyErr_ExceptionMatches(PyExc_AttributeError)) { + PyErr_Clear(); + } else { + result = -1; + } + return result; +} +static CYTHON_SMALL_CODE PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def) { + PyObject *module = NULL, *moddict, *modname; + CYTHON_UNUSED_VAR(def); + if (__Pyx_check_single_interpreter()) + return NULL; + if (__pyx_m) + return __Pyx_NewRef(__pyx_m); + modname = PyObject_GetAttrString(spec, "name"); + if (unlikely(!modname)) goto bad; + module = PyModule_NewObject(modname); + Py_DECREF(modname); + if (unlikely(!module)) goto bad; +#if CYTHON_COMPILING_IN_LIMITED_API + moddict = module; +#else + moddict = PyModule_GetDict(module); + if (unlikely(!moddict)) goto bad; +#endif + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "loader", "__loader__", 1) < 0)) goto bad; + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "origin", "__file__", 1) < 0)) goto bad; + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "parent", "__package__", 1) < 0)) goto bad; + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "submodule_search_locations", "__path__", 0) < 0)) goto bad; + return module; +bad: + Py_XDECREF(module); + return NULL; +} + + +static CYTHON_SMALL_CODE int __pyx_pymod_exec_data_utils_fast(PyObject *__pyx_pyinit_module) +#endif +#endif +{ + int stringtab_initialized = 0; + #if CYTHON_USE_MODULE_STATE + int pystate_addmodule_run = 0; + #endif + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + int __pyx_t_6; + PyObject *__pyx_t_7 = NULL; + static PyThread_type_lock __pyx_t_8[8]; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannyDeclarations + #if CYTHON_PEP489_MULTI_PHASE_INIT + if (__pyx_m) { + if (__pyx_m == __pyx_pyinit_module) return 0; + PyErr_SetString(PyExc_RuntimeError, "Module 'data_utils_fast' has already been imported. Re-initialisation is not supported."); + return -1; + } + #elif PY_MAJOR_VERSION >= 3 + if (__pyx_m) return __Pyx_NewRef(__pyx_m); + #endif + /*--- Module creation code ---*/ + #if CYTHON_PEP489_MULTI_PHASE_INIT + __pyx_m = __pyx_pyinit_module; + Py_INCREF(__pyx_m); + #else + #if PY_MAJOR_VERSION < 3 + __pyx_m = Py_InitModule4("data_utils_fast", __pyx_methods, 0, 0, PYTHON_API_VERSION); Py_XINCREF(__pyx_m); + if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error) + #elif CYTHON_USE_MODULE_STATE + __pyx_t_1 = PyModule_Create(&__pyx_moduledef); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error) + { + int add_module_result = PyState_AddModule(__pyx_t_1, &__pyx_moduledef); + __pyx_t_1 = 0; /* transfer ownership from __pyx_t_1 to "data_utils_fast" pseudovariable */ + if (unlikely((add_module_result < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + pystate_addmodule_run = 1; + } + #else + __pyx_m = PyModule_Create(&__pyx_moduledef); + if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #endif + CYTHON_UNUSED_VAR(__pyx_t_1); + __pyx_d = PyModule_GetDict(__pyx_m); if (unlikely(!__pyx_d)) __PYX_ERR(0, 1, __pyx_L1_error) + Py_INCREF(__pyx_d); + __pyx_b = __Pyx_PyImport_AddModuleRef(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_b)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_cython_runtime = __Pyx_PyImport_AddModuleRef((const char *) "cython_runtime"); if (unlikely(!__pyx_cython_runtime)) __PYX_ERR(0, 1, __pyx_L1_error) + if (PyObject_SetAttrString(__pyx_m, "__builtins__", __pyx_b) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #if CYTHON_REFNANNY +__Pyx_RefNanny = __Pyx_RefNannyImportAPI("refnanny"); +if (!__Pyx_RefNanny) { + PyErr_Clear(); + __Pyx_RefNanny = __Pyx_RefNannyImportAPI("Cython.Runtime.refnanny"); + if (!__Pyx_RefNanny) + Py_FatalError("failed to import 'refnanny' module"); +} +#endif + __Pyx_RefNannySetupContext("__Pyx_PyMODINIT_FUNC PyInit_data_utils_fast(void)", 0); + if (__Pyx_check_binary_version(__PYX_LIMITED_VERSION_HEX, __Pyx_get_runtime_version(), CYTHON_COMPILING_IN_LIMITED_API) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #ifdef __Pxy_PyFrame_Initialize_Offsets + __Pxy_PyFrame_Initialize_Offsets(); + #endif + __pyx_empty_tuple = PyTuple_New(0); if (unlikely(!__pyx_empty_tuple)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_empty_bytes = PyBytes_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_bytes)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_empty_unicode = PyUnicode_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_unicode)) __PYX_ERR(0, 1, __pyx_L1_error) + #ifdef __Pyx_CyFunction_USED + if (__pyx_CyFunction_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_FusedFunction_USED + if (__pyx_FusedFunction_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_Coroutine_USED + if (__pyx_Coroutine_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_Generator_USED + if (__pyx_Generator_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_AsyncGen_USED + if (__pyx_AsyncGen_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_StopAsyncIteration_USED + if (__pyx_StopAsyncIteration_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + /*--- Library function declarations ---*/ + /*--- Threads initialization code ---*/ + #if defined(WITH_THREAD) && PY_VERSION_HEX < 0x030700F0 && defined(__PYX_FORCE_INIT_THREADS) && __PYX_FORCE_INIT_THREADS + PyEval_InitThreads(); + #endif + /*--- Initialize various global constants etc. ---*/ + if (__Pyx_InitConstants() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + stringtab_initialized = 1; + if (__Pyx_InitGlobals() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #if PY_MAJOR_VERSION < 3 && (__PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT) + if (__Pyx_init_sys_getdefaultencoding_params() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + if (__pyx_module_is_main_fairseq__data__data_utils_fast) { + if (PyObject_SetAttr(__pyx_m, __pyx_n_s_name_2, __pyx_n_s_main) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + } + #if PY_MAJOR_VERSION >= 3 + { + PyObject *modules = PyImport_GetModuleDict(); if (unlikely(!modules)) __PYX_ERR(0, 1, __pyx_L1_error) + if (!PyDict_GetItemString(modules, "fairseq.data.data_utils_fast")) { + if (unlikely((PyDict_SetItemString(modules, "fairseq.data.data_utils_fast", __pyx_m) < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + } + } + #endif + /*--- Builtin init code ---*/ + if (__Pyx_InitCachedBuiltins() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + /*--- Constants init code ---*/ + if (__Pyx_InitCachedConstants() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + /*--- Global type/function init code ---*/ + (void)__Pyx_modinit_global_init_code(); + (void)__Pyx_modinit_variable_export_code(); + (void)__Pyx_modinit_function_export_code(); + if (unlikely((__Pyx_modinit_type_init_code() < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + if (unlikely((__Pyx_modinit_type_import_code() < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + (void)__Pyx_modinit_variable_import_code(); + (void)__Pyx_modinit_function_import_code(); + /*--- Execution code ---*/ + #if defined(__Pyx_Generator_USED) || defined(__Pyx_Coroutine_USED) + if (__Pyx_patch_abc() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + + /* "View.MemoryView":99 + * + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: # <<<<<<<<<<<<<< + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "View.MemoryView":100 + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: + * if __import__("sys").version_info >= (3, 3): # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + */ + __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin___import__, __pyx_tuple__11, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_version_info); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_t_4 = PyObject_RichCompare(__pyx_t_5, __pyx_tuple__12, Py_GE); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely((__pyx_t_6 < 0))) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + if (__pyx_t_6) { + + /* "View.MemoryView":101 + * try: + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence # <<<<<<<<<<<<<< + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence + */ + __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin___import__, __pyx_tuple__13, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 101, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_abc); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 101, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_t_4 = __Pyx_PyObject_GetAttrStr(__pyx_t_5, __pyx_n_s_Sequence); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 101, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XGOTREF(__pyx_collections_abc_Sequence); + __Pyx_DECREF_SET(__pyx_collections_abc_Sequence, __pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + __pyx_t_4 = 0; + + /* "View.MemoryView":100 + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: + * if __import__("sys").version_info >= (3, 3): # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + */ + goto __pyx_L8; + } + + /* "View.MemoryView":103 + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence # <<<<<<<<<<<<<< + * except: + * + */ + /*else*/ { + __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin___import__, __pyx_tuple__14, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 103, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_Sequence); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 103, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XGOTREF(__pyx_collections_abc_Sequence); + __Pyx_DECREF_SET(__pyx_collections_abc_Sequence, __pyx_t_5); + __Pyx_GIVEREF(__pyx_t_5); + __pyx_t_5 = 0; + } + __pyx_L8:; + + /* "View.MemoryView":99 + * + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: # <<<<<<<<<<<<<< + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L7_try_end; + __pyx_L2_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + + /* "View.MemoryView":104 + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence + * except: # <<<<<<<<<<<<<< + * + * __pyx_collections_abc_Sequence = None + */ + /*except:*/ { + __Pyx_AddTraceback("View.MemoryView", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_4, &__pyx_t_7) < 0) __PYX_ERR(1, 104, __pyx_L4_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_4); + __Pyx_XGOTREF(__pyx_t_7); + + /* "View.MemoryView":106 + * except: + * + * __pyx_collections_abc_Sequence = None # <<<<<<<<<<<<<< + * + * + */ + __Pyx_INCREF(Py_None); + __Pyx_XGOTREF(__pyx_collections_abc_Sequence); + __Pyx_DECREF_SET(__pyx_collections_abc_Sequence, Py_None); + __Pyx_GIVEREF(Py_None); + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + goto __pyx_L3_exception_handled; + } + + /* "View.MemoryView":99 + * + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: # <<<<<<<<<<<<<< + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + */ + __pyx_L4_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L3_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + __pyx_L7_try_end:; + } + + /* "View.MemoryView":241 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_3, &__pyx_t_2, &__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_1); + /*try:*/ { + + /* "View.MemoryView":242 + * + * try: + * count = __pyx_collections_abc_Sequence.count # <<<<<<<<<<<<<< + * index = __pyx_collections_abc_Sequence.index + * except: + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_count); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 242, __pyx_L11_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_array_type, __pyx_n_s_count, __pyx_t_7) < 0) __PYX_ERR(1, 242, __pyx_L11_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_array_type); + + /* "View.MemoryView":243 + * try: + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index # <<<<<<<<<<<<<< + * except: + * pass + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_index); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 243, __pyx_L11_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_array_type, __pyx_n_s_index, __pyx_t_7) < 0) __PYX_ERR(1, 243, __pyx_L11_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_array_type); + + /* "View.MemoryView":241 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + } + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + goto __pyx_L16_try_end; + __pyx_L11_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":244 + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + * except: # <<<<<<<<<<<<<< + * pass + * + */ + /*except:*/ { + __Pyx_ErrRestore(0,0,0); + goto __pyx_L12_exception_handled; + } + __pyx_L12_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_2, __pyx_t_1); + __pyx_L16_try_end:; + } + + /* "View.MemoryView":309 + * return self.name + * + * cdef generic = Enum("") # <<<<<<<<<<<<<< + * cdef strided = Enum("") # default + * cdef indirect = Enum("") + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__15, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 309, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(generic); + __Pyx_DECREF_SET(generic, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":310 + * + * cdef generic = Enum("") + * cdef strided = Enum("") # default # <<<<<<<<<<<<<< + * cdef indirect = Enum("") + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__16, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 310, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(strided); + __Pyx_DECREF_SET(strided, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":311 + * cdef generic = Enum("") + * cdef strided = Enum("") # default + * cdef indirect = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__17, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 311, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(indirect); + __Pyx_DECREF_SET(indirect, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":314 + * + * + * cdef contiguous = Enum("") # <<<<<<<<<<<<<< + * cdef indirect_contiguous = Enum("") + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__18, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 314, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(contiguous); + __Pyx_DECREF_SET(contiguous, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":315 + * + * cdef contiguous = Enum("") + * cdef indirect_contiguous = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__19, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 315, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(indirect_contiguous); + __Pyx_DECREF_SET(indirect_contiguous, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":323 + * + * + * cdef int __pyx_memoryview_thread_locks_used = 0 # <<<<<<<<<<<<<< + * cdef PyThread_type_lock[8] __pyx_memoryview_thread_locks = [ + * PyThread_allocate_lock(), + */ + __pyx_memoryview_thread_locks_used = 0; + + /* "View.MemoryView":324 + * + * cdef int __pyx_memoryview_thread_locks_used = 0 + * cdef PyThread_type_lock[8] __pyx_memoryview_thread_locks = [ # <<<<<<<<<<<<<< + * PyThread_allocate_lock(), + * PyThread_allocate_lock(), + */ + __pyx_t_8[0] = PyThread_allocate_lock(); + __pyx_t_8[1] = PyThread_allocate_lock(); + __pyx_t_8[2] = PyThread_allocate_lock(); + __pyx_t_8[3] = PyThread_allocate_lock(); + __pyx_t_8[4] = PyThread_allocate_lock(); + __pyx_t_8[5] = PyThread_allocate_lock(); + __pyx_t_8[6] = PyThread_allocate_lock(); + __pyx_t_8[7] = PyThread_allocate_lock(); + memcpy(&(__pyx_memoryview_thread_locks[0]), __pyx_t_8, sizeof(__pyx_memoryview_thread_locks[0]) * (8)); + + /* "View.MemoryView":982 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "View.MemoryView":983 + * + * try: + * count = __pyx_collections_abc_Sequence.count # <<<<<<<<<<<<<< + * index = __pyx_collections_abc_Sequence.index + * except: + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_count); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 983, __pyx_L17_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_memoryviewslice_type, __pyx_n_s_count, __pyx_t_7) < 0) __PYX_ERR(1, 983, __pyx_L17_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_memoryviewslice_type); + + /* "View.MemoryView":984 + * try: + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index # <<<<<<<<<<<<<< + * except: + * pass + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_index); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 984, __pyx_L17_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_memoryviewslice_type, __pyx_n_s_index, __pyx_t_7) < 0) __PYX_ERR(1, 984, __pyx_L17_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_memoryviewslice_type); + + /* "View.MemoryView":982 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L22_try_end; + __pyx_L17_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":985 + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + * except: # <<<<<<<<<<<<<< + * pass + * + */ + /*except:*/ { + __Pyx_ErrRestore(0,0,0); + goto __pyx_L18_exception_handled; + } + __pyx_L18_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + __pyx_L22_try_end:; + } + + /* "View.MemoryView":988 + * pass + * + * try: # <<<<<<<<<<<<<< + * if __pyx_collections_abc_Sequence: + * + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_3, &__pyx_t_2, &__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_1); + /*try:*/ { + + /* "View.MemoryView":989 + * + * try: + * if __pyx_collections_abc_Sequence: # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_collections_abc_Sequence); if (unlikely((__pyx_t_6 < 0))) __PYX_ERR(1, 989, __pyx_L23_error) + if (__pyx_t_6) { + + /* "View.MemoryView":993 + * + * + * __pyx_collections_abc_Sequence.register(_memoryviewslice) # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence.register(array) + * except: + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_register); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 993, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_4 = __Pyx_PyObject_CallOneArg(__pyx_t_7, ((PyObject *)__pyx_memoryviewslice_type)); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 993, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "View.MemoryView":994 + * + * __pyx_collections_abc_Sequence.register(_memoryviewslice) + * __pyx_collections_abc_Sequence.register(array) # <<<<<<<<<<<<<< + * except: + * pass # ignore failure, it's a minor issue + */ + __pyx_t_4 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_register); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 994, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_7 = __Pyx_PyObject_CallOneArg(__pyx_t_4, ((PyObject *)__pyx_array_type)); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 994, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":989 + * + * try: + * if __pyx_collections_abc_Sequence: # <<<<<<<<<<<<<< + * + * + */ + } + + /* "View.MemoryView":988 + * pass + * + * try: # <<<<<<<<<<<<<< + * if __pyx_collections_abc_Sequence: + * + */ + } + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + goto __pyx_L28_try_end; + __pyx_L23_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":995 + * __pyx_collections_abc_Sequence.register(_memoryviewslice) + * __pyx_collections_abc_Sequence.register(array) + * except: # <<<<<<<<<<<<<< + * pass # ignore failure, it's a minor issue + * + */ + /*except:*/ { + __Pyx_ErrRestore(0,0,0); + goto __pyx_L24_exception_handled; + } + __pyx_L24_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_2, __pyx_t_1); + __pyx_L28_try_end:; + } + + /* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + __pyx_t_7 = PyCFunction_NewEx(&__pyx_mdef_15View_dot_MemoryView_1__pyx_unpickle_Enum, NULL, __pyx_n_s_View_MemoryView); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_pyx_unpickle_Enum, __pyx_t_7) < 0) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/data_utils_fast.pyx":7 + * # LICENSE file in the root directory of this source tree. + * + * import numpy as np # <<<<<<<<<<<<<< + * + * cimport cython + */ + __pyx_t_7 = __Pyx_ImportDottedModule(__pyx_n_s_numpy, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_np, __pyx_t_7) < 0) __PYX_ERR(0, 7, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/data_utils_fast.pyx":12 + * cimport numpy as np + * + * DTYPE = np.int64 # <<<<<<<<<<<<<< + * ctypedef np.int64_t DTYPE_t + * + */ + __Pyx_GetModuleGlobalName(__pyx_t_7, __pyx_n_s_np); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_4 = __Pyx_PyObject_GetAttrStr(__pyx_t_7, __pyx_n_s_int64); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + if (PyDict_SetItem(__pyx_d, __pyx_n_s_DTYPE, __pyx_t_4) < 0) __PYX_ERR(0, 12, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "fairseq/data/data_utils_fast.pyx":27 + * + * @cython.cdivision(True) + * cpdef list batch_by_size_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + __pyx_t_4 = __Pyx_CyFunction_New(&__pyx_mdef_7fairseq_4data_15data_utils_fast_1batch_by_size_fast, 0, __pyx_n_s_batch_by_size_fast, NULL, __pyx_n_s_fairseq_data_data_utils_fast, __pyx_d, ((PyObject *)__pyx_codeobj__23)); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 27, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_batch_by_size_fast, __pyx_t_4) < 0) __PYX_ERR(0, 27, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "fairseq/data/data_utils_fast.pyx":84 + * + * @cython.cdivision(True) + * cpdef list batch_fixed_shapes_fast( # <<<<<<<<<<<<<< + * np.ndarray[DTYPE_t, ndim=1] indices, + * num_tokens_fn, + */ + __pyx_t_4 = __Pyx_CyFunction_New(&__pyx_mdef_7fairseq_4data_15data_utils_fast_3batch_fixed_shapes_fast, 0, __pyx_n_s_batch_fixed_shapes_fast, NULL, __pyx_n_s_fairseq_data_data_utils_fast, __pyx_d, ((PyObject *)__pyx_codeobj__25)); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 84, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_batch_fixed_shapes_fast, __pyx_t_4) < 0) __PYX_ERR(0, 84, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "fairseq/data/data_utils_fast.pyx":1 + * # cython: language_level=3 # <<<<<<<<<<<<<< + * # Copyright (c) Facebook, Inc. and its affiliates. + * # + */ + __pyx_t_4 = __Pyx_PyDict_NewPresized(0); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_test, __pyx_t_4) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /*--- Wrapped vars code ---*/ + + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_7); + if (__pyx_m) { + if (__pyx_d && stringtab_initialized) { + __Pyx_AddTraceback("init fairseq.data.data_utils_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + } + #if !CYTHON_USE_MODULE_STATE + Py_CLEAR(__pyx_m); + #else + Py_DECREF(__pyx_m); + if (pystate_addmodule_run) { + PyObject *tp, *value, *tb; + PyErr_Fetch(&tp, &value, &tb); + PyState_RemoveModule(&__pyx_moduledef); + PyErr_Restore(tp, value, tb); + } + #endif + } else if (!PyErr_Occurred()) { + PyErr_SetString(PyExc_ImportError, "init fairseq.data.data_utils_fast"); + } + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + #if CYTHON_PEP489_MULTI_PHASE_INIT + return (__pyx_m != NULL) ? 0 : -1; + #elif PY_MAJOR_VERSION >= 3 + return __pyx_m; + #else + return; + #endif +} +/* #### Code section: cleanup_globals ### */ +/* #### Code section: cleanup_module ### */ +/* #### Code section: main_method ### */ +/* #### Code section: utility_code_pragmas ### */ +#ifdef _MSC_VER +#pragma warning( push ) +/* Warning 4127: conditional expression is constant + * Cython uses constant conditional expressions to allow in inline functions to be optimized at + * compile-time, so this warning is not useful + */ +#pragma warning( disable : 4127 ) +#endif + + + +/* #### Code section: utility_code_def ### */ + +/* --- Runtime support code --- */ +/* Refnanny */ +#if CYTHON_REFNANNY +static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname) { + PyObject *m = NULL, *p = NULL; + void *r = NULL; + m = PyImport_ImportModule(modname); + if (!m) goto end; + p = PyObject_GetAttrString(m, "RefNannyAPI"); + if (!p) goto end; + r = PyLong_AsVoidPtr(p); +end: + Py_XDECREF(p); + Py_XDECREF(m); + return (__Pyx_RefNannyAPIStruct *)r; +} +#endif + +/* PyErrExceptionMatches */ +#if CYTHON_FAST_THREAD_STATE +static int __Pyx_PyErr_ExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) { + Py_ssize_t i, n; + n = PyTuple_GET_SIZE(tuple); +#if PY_MAJOR_VERSION >= 3 + for (i=0; i= 0x030C00A6 + PyObject *current_exception = tstate->current_exception; + if (unlikely(!current_exception)) return 0; + exc_type = (PyObject*) Py_TYPE(current_exception); + if (exc_type == err) return 1; +#else + exc_type = tstate->curexc_type; + if (exc_type == err) return 1; + if (unlikely(!exc_type)) return 0; +#endif + #if CYTHON_AVOID_BORROWED_REFS + Py_INCREF(exc_type); + #endif + if (unlikely(PyTuple_Check(err))) { + result = __Pyx_PyErr_ExceptionMatchesTuple(exc_type, err); + } else { + result = __Pyx_PyErr_GivenExceptionMatches(exc_type, err); + } + #if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(exc_type); + #endif + return result; +} +#endif + +/* PyErrFetchRestore */ +#if CYTHON_FAST_THREAD_STATE +static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) { +#if PY_VERSION_HEX >= 0x030C00A6 + PyObject *tmp_value; + assert(type == NULL || (value != NULL && type == (PyObject*) Py_TYPE(value))); + if (value) { + #if CYTHON_COMPILING_IN_CPYTHON + if (unlikely(((PyBaseExceptionObject*) value)->traceback != tb)) + #endif + PyException_SetTraceback(value, tb); + } + tmp_value = tstate->current_exception; + tstate->current_exception = value; + Py_XDECREF(tmp_value); + Py_XDECREF(type); + Py_XDECREF(tb); +#else + PyObject *tmp_type, *tmp_value, *tmp_tb; + tmp_type = tstate->curexc_type; + tmp_value = tstate->curexc_value; + tmp_tb = tstate->curexc_traceback; + tstate->curexc_type = type; + tstate->curexc_value = value; + tstate->curexc_traceback = tb; + Py_XDECREF(tmp_type); + Py_XDECREF(tmp_value); + Py_XDECREF(tmp_tb); +#endif +} +static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) { +#if PY_VERSION_HEX >= 0x030C00A6 + PyObject* exc_value; + exc_value = tstate->current_exception; + tstate->current_exception = 0; + *value = exc_value; + *type = NULL; + *tb = NULL; + if (exc_value) { + *type = (PyObject*) Py_TYPE(exc_value); + Py_INCREF(*type); + #if CYTHON_COMPILING_IN_CPYTHON + *tb = ((PyBaseExceptionObject*) exc_value)->traceback; + Py_XINCREF(*tb); + #else + *tb = PyException_GetTraceback(exc_value); + #endif + } +#else + *type = tstate->curexc_type; + *value = tstate->curexc_value; + *tb = tstate->curexc_traceback; + tstate->curexc_type = 0; + tstate->curexc_value = 0; + tstate->curexc_traceback = 0; +#endif +} +#endif + +/* PyObjectGetAttrStr */ +#if CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name) { + PyTypeObject* tp = Py_TYPE(obj); + if (likely(tp->tp_getattro)) + return tp->tp_getattro(obj, attr_name); +#if PY_MAJOR_VERSION < 3 + if (likely(tp->tp_getattr)) + return tp->tp_getattr(obj, PyString_AS_STRING(attr_name)); +#endif + return PyObject_GetAttr(obj, attr_name); +} +#endif + +/* PyObjectGetAttrStrNoError */ +#if __PYX_LIMITED_VERSION_HEX < 0x030d00A1 +static void __Pyx_PyObject_GetAttrStr_ClearAttributeError(void) { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + if (likely(__Pyx_PyErr_ExceptionMatches(PyExc_AttributeError))) + __Pyx_PyErr_Clear(); +} +#endif +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStrNoError(PyObject* obj, PyObject* attr_name) { + PyObject *result; +#if __PYX_LIMITED_VERSION_HEX >= 0x030d00A1 + (void) PyObject_GetOptionalAttr(obj, attr_name, &result); + return result; +#else +#if CYTHON_COMPILING_IN_CPYTHON && CYTHON_USE_TYPE_SLOTS && PY_VERSION_HEX >= 0x030700B1 + PyTypeObject* tp = Py_TYPE(obj); + if (likely(tp->tp_getattro == PyObject_GenericGetAttr)) { + return _PyObject_GenericGetAttrWithDict(obj, attr_name, NULL, 1); + } +#endif + result = __Pyx_PyObject_GetAttrStr(obj, attr_name); + if (unlikely(!result)) { + __Pyx_PyObject_GetAttrStr_ClearAttributeError(); + } + return result; +#endif +} + +/* GetBuiltinName */ +static PyObject *__Pyx_GetBuiltinName(PyObject *name) { + PyObject* result = __Pyx_PyObject_GetAttrStrNoError(__pyx_b, name); + if (unlikely(!result) && !PyErr_Occurred()) { + PyErr_Format(PyExc_NameError, +#if PY_MAJOR_VERSION >= 3 + "name '%U' is not defined", name); +#else + "name '%.200s' is not defined", PyString_AS_STRING(name)); +#endif + } + return result; +} + +/* TupleAndListFromArray */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE void __Pyx_copy_object_array(PyObject *const *CYTHON_RESTRICT src, PyObject** CYTHON_RESTRICT dest, Py_ssize_t length) { + PyObject *v; + Py_ssize_t i; + for (i = 0; i < length; i++) { + v = dest[i] = src[i]; + Py_INCREF(v); + } +} +static CYTHON_INLINE PyObject * +__Pyx_PyTuple_FromArray(PyObject *const *src, Py_ssize_t n) +{ + PyObject *res; + if (n <= 0) { + Py_INCREF(__pyx_empty_tuple); + return __pyx_empty_tuple; + } + res = PyTuple_New(n); + if (unlikely(res == NULL)) return NULL; + __Pyx_copy_object_array(src, ((PyTupleObject*)res)->ob_item, n); + return res; +} +static CYTHON_INLINE PyObject * +__Pyx_PyList_FromArray(PyObject *const *src, Py_ssize_t n) +{ + PyObject *res; + if (n <= 0) { + return PyList_New(0); + } + res = PyList_New(n); + if (unlikely(res == NULL)) return NULL; + __Pyx_copy_object_array(src, ((PyListObject*)res)->ob_item, n); + return res; +} +#endif + +/* BytesEquals */ +static CYTHON_INLINE int __Pyx_PyBytes_Equals(PyObject* s1, PyObject* s2, int equals) { +#if CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API + return PyObject_RichCompareBool(s1, s2, equals); +#else + if (s1 == s2) { + return (equals == Py_EQ); + } else if (PyBytes_CheckExact(s1) & PyBytes_CheckExact(s2)) { + const char *ps1, *ps2; + Py_ssize_t length = PyBytes_GET_SIZE(s1); + if (length != PyBytes_GET_SIZE(s2)) + return (equals == Py_NE); + ps1 = PyBytes_AS_STRING(s1); + ps2 = PyBytes_AS_STRING(s2); + if (ps1[0] != ps2[0]) { + return (equals == Py_NE); + } else if (length == 1) { + return (equals == Py_EQ); + } else { + int result; +#if CYTHON_USE_UNICODE_INTERNALS && (PY_VERSION_HEX < 0x030B0000) + Py_hash_t hash1, hash2; + hash1 = ((PyBytesObject*)s1)->ob_shash; + hash2 = ((PyBytesObject*)s2)->ob_shash; + if (hash1 != hash2 && hash1 != -1 && hash2 != -1) { + return (equals == Py_NE); + } +#endif + result = memcmp(ps1, ps2, (size_t)length); + return (equals == Py_EQ) ? (result == 0) : (result != 0); + } + } else if ((s1 == Py_None) & PyBytes_CheckExact(s2)) { + return (equals == Py_NE); + } else if ((s2 == Py_None) & PyBytes_CheckExact(s1)) { + return (equals == Py_NE); + } else { + int result; + PyObject* py_result = PyObject_RichCompare(s1, s2, equals); + if (!py_result) + return -1; + result = __Pyx_PyObject_IsTrue(py_result); + Py_DECREF(py_result); + return result; + } +#endif +} + +/* UnicodeEquals */ +static CYTHON_INLINE int __Pyx_PyUnicode_Equals(PyObject* s1, PyObject* s2, int equals) { +#if CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API + return PyObject_RichCompareBool(s1, s2, equals); +#else +#if PY_MAJOR_VERSION < 3 + PyObject* owned_ref = NULL; +#endif + int s1_is_unicode, s2_is_unicode; + if (s1 == s2) { + goto return_eq; + } + s1_is_unicode = PyUnicode_CheckExact(s1); + s2_is_unicode = PyUnicode_CheckExact(s2); +#if PY_MAJOR_VERSION < 3 + if ((s1_is_unicode & (!s2_is_unicode)) && PyString_CheckExact(s2)) { + owned_ref = PyUnicode_FromObject(s2); + if (unlikely(!owned_ref)) + return -1; + s2 = owned_ref; + s2_is_unicode = 1; + } else if ((s2_is_unicode & (!s1_is_unicode)) && PyString_CheckExact(s1)) { + owned_ref = PyUnicode_FromObject(s1); + if (unlikely(!owned_ref)) + return -1; + s1 = owned_ref; + s1_is_unicode = 1; + } else if (((!s2_is_unicode) & (!s1_is_unicode))) { + return __Pyx_PyBytes_Equals(s1, s2, equals); + } +#endif + if (s1_is_unicode & s2_is_unicode) { + Py_ssize_t length; + int kind; + void *data1, *data2; + if (unlikely(__Pyx_PyUnicode_READY(s1) < 0) || unlikely(__Pyx_PyUnicode_READY(s2) < 0)) + return -1; + length = __Pyx_PyUnicode_GET_LENGTH(s1); + if (length != __Pyx_PyUnicode_GET_LENGTH(s2)) { + goto return_ne; + } +#if CYTHON_USE_UNICODE_INTERNALS + { + Py_hash_t hash1, hash2; + #if CYTHON_PEP393_ENABLED + hash1 = ((PyASCIIObject*)s1)->hash; + hash2 = ((PyASCIIObject*)s2)->hash; + #else + hash1 = ((PyUnicodeObject*)s1)->hash; + hash2 = ((PyUnicodeObject*)s2)->hash; + #endif + if (hash1 != hash2 && hash1 != -1 && hash2 != -1) { + goto return_ne; + } + } +#endif + kind = __Pyx_PyUnicode_KIND(s1); + if (kind != __Pyx_PyUnicode_KIND(s2)) { + goto return_ne; + } + data1 = __Pyx_PyUnicode_DATA(s1); + data2 = __Pyx_PyUnicode_DATA(s2); + if (__Pyx_PyUnicode_READ(kind, data1, 0) != __Pyx_PyUnicode_READ(kind, data2, 0)) { + goto return_ne; + } else if (length == 1) { + goto return_eq; + } else { + int result = memcmp(data1, data2, (size_t)(length * kind)); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + return (equals == Py_EQ) ? (result == 0) : (result != 0); + } + } else if ((s1 == Py_None) & s2_is_unicode) { + goto return_ne; + } else if ((s2 == Py_None) & s1_is_unicode) { + goto return_ne; + } else { + int result; + PyObject* py_result = PyObject_RichCompare(s1, s2, equals); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + if (!py_result) + return -1; + result = __Pyx_PyObject_IsTrue(py_result); + Py_DECREF(py_result); + return result; + } +return_eq: + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + return (equals == Py_EQ); +return_ne: + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + return (equals == Py_NE); +#endif +} + +/* fastcall */ +#if CYTHON_METH_FASTCALL +static CYTHON_INLINE PyObject * __Pyx_GetKwValue_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues, PyObject *s) +{ + Py_ssize_t i, n = PyTuple_GET_SIZE(kwnames); + for (i = 0; i < n; i++) + { + if (s == PyTuple_GET_ITEM(kwnames, i)) return kwvalues[i]; + } + for (i = 0; i < n; i++) + { + int eq = __Pyx_PyUnicode_Equals(s, PyTuple_GET_ITEM(kwnames, i), Py_EQ); + if (unlikely(eq != 0)) { + if (unlikely(eq < 0)) return NULL; + return kwvalues[i]; + } + } + return NULL; +} +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 +CYTHON_UNUSED static PyObject *__Pyx_KwargsAsDict_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues) { + Py_ssize_t i, nkwargs = PyTuple_GET_SIZE(kwnames); + PyObject *dict; + dict = PyDict_New(); + if (unlikely(!dict)) + return NULL; + for (i=0; i= 3 + "%s() got multiple values for keyword argument '%U'", func_name, kw_name); + #else + "%s() got multiple values for keyword argument '%s'", func_name, + PyString_AsString(kw_name)); + #endif +} + +/* ParseKeywords */ +static int __Pyx_ParseOptionalKeywords( + PyObject *kwds, + PyObject *const *kwvalues, + PyObject **argnames[], + PyObject *kwds2, + PyObject *values[], + Py_ssize_t num_pos_args, + const char* function_name) +{ + PyObject *key = 0, *value = 0; + Py_ssize_t pos = 0; + PyObject*** name; + PyObject*** first_kw_arg = argnames + num_pos_args; + int kwds_is_tuple = CYTHON_METH_FASTCALL && likely(PyTuple_Check(kwds)); + while (1) { + Py_XDECREF(key); key = NULL; + Py_XDECREF(value); value = NULL; + if (kwds_is_tuple) { + Py_ssize_t size; +#if CYTHON_ASSUME_SAFE_MACROS + size = PyTuple_GET_SIZE(kwds); +#else + size = PyTuple_Size(kwds); + if (size < 0) goto bad; +#endif + if (pos >= size) break; +#if CYTHON_AVOID_BORROWED_REFS + key = __Pyx_PySequence_ITEM(kwds, pos); + if (!key) goto bad; +#elif CYTHON_ASSUME_SAFE_MACROS + key = PyTuple_GET_ITEM(kwds, pos); +#else + key = PyTuple_GetItem(kwds, pos); + if (!key) goto bad; +#endif + value = kwvalues[pos]; + pos++; + } + else + { + if (!PyDict_Next(kwds, &pos, &key, &value)) break; +#if CYTHON_AVOID_BORROWED_REFS + Py_INCREF(key); +#endif + } + name = first_kw_arg; + while (*name && (**name != key)) name++; + if (*name) { + values[name-argnames] = value; +#if CYTHON_AVOID_BORROWED_REFS + Py_INCREF(value); + Py_DECREF(key); +#endif + key = NULL; + value = NULL; + continue; + } +#if !CYTHON_AVOID_BORROWED_REFS + Py_INCREF(key); +#endif + Py_INCREF(value); + name = first_kw_arg; + #if PY_MAJOR_VERSION < 3 + if (likely(PyString_Check(key))) { + while (*name) { + if ((CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**name) == PyString_GET_SIZE(key)) + && _PyString_Eq(**name, key)) { + values[name-argnames] = value; +#if CYTHON_AVOID_BORROWED_REFS + value = NULL; +#endif + break; + } + name++; + } + if (*name) continue; + else { + PyObject*** argname = argnames; + while (argname != first_kw_arg) { + if ((**argname == key) || ( + (CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**argname) == PyString_GET_SIZE(key)) + && _PyString_Eq(**argname, key))) { + goto arg_passed_twice; + } + argname++; + } + } + } else + #endif + if (likely(PyUnicode_Check(key))) { + while (*name) { + int cmp = ( + #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3 + (__Pyx_PyUnicode_GET_LENGTH(**name) != __Pyx_PyUnicode_GET_LENGTH(key)) ? 1 : + #endif + PyUnicode_Compare(**name, key) + ); + if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad; + if (cmp == 0) { + values[name-argnames] = value; +#if CYTHON_AVOID_BORROWED_REFS + value = NULL; +#endif + break; + } + name++; + } + if (*name) continue; + else { + PyObject*** argname = argnames; + while (argname != first_kw_arg) { + int cmp = (**argname == key) ? 0 : + #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3 + (__Pyx_PyUnicode_GET_LENGTH(**argname) != __Pyx_PyUnicode_GET_LENGTH(key)) ? 1 : + #endif + PyUnicode_Compare(**argname, key); + if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad; + if (cmp == 0) goto arg_passed_twice; + argname++; + } + } + } else + goto invalid_keyword_type; + if (kwds2) { + if (unlikely(PyDict_SetItem(kwds2, key, value))) goto bad; + } else { + goto invalid_keyword; + } + } + Py_XDECREF(key); + Py_XDECREF(value); + return 0; +arg_passed_twice: + __Pyx_RaiseDoubleKeywordsError(function_name, key); + goto bad; +invalid_keyword_type: + PyErr_Format(PyExc_TypeError, + "%.200s() keywords must be strings", function_name); + goto bad; +invalid_keyword: + #if PY_MAJOR_VERSION < 3 + PyErr_Format(PyExc_TypeError, + "%.200s() got an unexpected keyword argument '%.200s'", + function_name, PyString_AsString(key)); + #else + PyErr_Format(PyExc_TypeError, + "%s() got an unexpected keyword argument '%U'", + function_name, key); + #endif +bad: + Py_XDECREF(key); + Py_XDECREF(value); + return -1; +} + +/* ArgTypeTest */ +static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact) +{ + __Pyx_TypeName type_name; + __Pyx_TypeName obj_type_name; + if (unlikely(!type)) { + PyErr_SetString(PyExc_SystemError, "Missing type object"); + return 0; + } + else if (exact) { + #if PY_MAJOR_VERSION == 2 + if ((type == &PyBaseString_Type) && likely(__Pyx_PyBaseString_CheckExact(obj))) return 1; + #endif + } + else { + if (likely(__Pyx_TypeCheck(obj, type))) return 1; + } + type_name = __Pyx_PyType_GetName(type); + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, + "Argument '%.200s' has incorrect type (expected " __Pyx_FMT_TYPENAME + ", got " __Pyx_FMT_TYPENAME ")", name, type_name, obj_type_name); + __Pyx_DECREF_TypeName(type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return 0; +} + +/* RaiseException */ +#if PY_MAJOR_VERSION < 3 +static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause) { + __Pyx_PyThreadState_declare + CYTHON_UNUSED_VAR(cause); + Py_XINCREF(type); + if (!value || value == Py_None) + value = NULL; + else + Py_INCREF(value); + if (!tb || tb == Py_None) + tb = NULL; + else { + Py_INCREF(tb); + if (!PyTraceBack_Check(tb)) { + PyErr_SetString(PyExc_TypeError, + "raise: arg 3 must be a traceback or None"); + goto raise_error; + } + } + if (PyType_Check(type)) { +#if CYTHON_COMPILING_IN_PYPY + if (!value) { + Py_INCREF(Py_None); + value = Py_None; + } +#endif + PyErr_NormalizeException(&type, &value, &tb); + } else { + if (value) { + PyErr_SetString(PyExc_TypeError, + "instance exception may not have a separate value"); + goto raise_error; + } + value = type; + type = (PyObject*) Py_TYPE(type); + Py_INCREF(type); + if (!PyType_IsSubtype((PyTypeObject *)type, (PyTypeObject *)PyExc_BaseException)) { + PyErr_SetString(PyExc_TypeError, + "raise: exception class must be a subclass of BaseException"); + goto raise_error; + } + } + __Pyx_PyThreadState_assign + __Pyx_ErrRestore(type, value, tb); + return; +raise_error: + Py_XDECREF(value); + Py_XDECREF(type); + Py_XDECREF(tb); + return; +} +#else +static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause) { + PyObject* owned_instance = NULL; + if (tb == Py_None) { + tb = 0; + } else if (tb && !PyTraceBack_Check(tb)) { + PyErr_SetString(PyExc_TypeError, + "raise: arg 3 must be a traceback or None"); + goto bad; + } + if (value == Py_None) + value = 0; + if (PyExceptionInstance_Check(type)) { + if (value) { + PyErr_SetString(PyExc_TypeError, + "instance exception may not have a separate value"); + goto bad; + } + value = type; + type = (PyObject*) Py_TYPE(value); + } else if (PyExceptionClass_Check(type)) { + PyObject *instance_class = NULL; + if (value && PyExceptionInstance_Check(value)) { + instance_class = (PyObject*) Py_TYPE(value); + if (instance_class != type) { + int is_subclass = PyObject_IsSubclass(instance_class, type); + if (!is_subclass) { + instance_class = NULL; + } else if (unlikely(is_subclass == -1)) { + goto bad; + } else { + type = instance_class; + } + } + } + if (!instance_class) { + PyObject *args; + if (!value) + args = PyTuple_New(0); + else if (PyTuple_Check(value)) { + Py_INCREF(value); + args = value; + } else + args = PyTuple_Pack(1, value); + if (!args) + goto bad; + owned_instance = PyObject_Call(type, args, NULL); + Py_DECREF(args); + if (!owned_instance) + goto bad; + value = owned_instance; + if (!PyExceptionInstance_Check(value)) { + PyErr_Format(PyExc_TypeError, + "calling %R should have returned an instance of " + "BaseException, not %R", + type, Py_TYPE(value)); + goto bad; + } + } + } else { + PyErr_SetString(PyExc_TypeError, + "raise: exception class must be a subclass of BaseException"); + goto bad; + } + if (cause) { + PyObject *fixed_cause; + if (cause == Py_None) { + fixed_cause = NULL; + } else if (PyExceptionClass_Check(cause)) { + fixed_cause = PyObject_CallObject(cause, NULL); + if (fixed_cause == NULL) + goto bad; + } else if (PyExceptionInstance_Check(cause)) { + fixed_cause = cause; + Py_INCREF(fixed_cause); + } else { + PyErr_SetString(PyExc_TypeError, + "exception causes must derive from " + "BaseException"); + goto bad; + } + PyException_SetCause(value, fixed_cause); + } + PyErr_SetObject(type, value); + if (tb) { + #if PY_VERSION_HEX >= 0x030C00A6 + PyException_SetTraceback(value, tb); + #elif CYTHON_FAST_THREAD_STATE + PyThreadState *tstate = __Pyx_PyThreadState_Current; + PyObject* tmp_tb = tstate->curexc_traceback; + if (tb != tmp_tb) { + Py_INCREF(tb); + tstate->curexc_traceback = tb; + Py_XDECREF(tmp_tb); + } +#else + PyObject *tmp_type, *tmp_value, *tmp_tb; + PyErr_Fetch(&tmp_type, &tmp_value, &tmp_tb); + Py_INCREF(tb); + PyErr_Restore(tmp_type, tmp_value, tb); + Py_XDECREF(tmp_tb); +#endif + } +bad: + Py_XDECREF(owned_instance); + return; +} +#endif + +/* PyFunctionFastCall */ +#if CYTHON_FAST_PYCALL && !CYTHON_VECTORCALL +static PyObject* __Pyx_PyFunction_FastCallNoKw(PyCodeObject *co, PyObject **args, Py_ssize_t na, + PyObject *globals) { + PyFrameObject *f; + PyThreadState *tstate = __Pyx_PyThreadState_Current; + PyObject **fastlocals; + Py_ssize_t i; + PyObject *result; + assert(globals != NULL); + /* XXX Perhaps we should create a specialized + PyFrame_New() that doesn't take locals, but does + take builtins without sanity checking them. + */ + assert(tstate != NULL); + f = PyFrame_New(tstate, co, globals, NULL); + if (f == NULL) { + return NULL; + } + fastlocals = __Pyx_PyFrame_GetLocalsplus(f); + for (i = 0; i < na; i++) { + Py_INCREF(*args); + fastlocals[i] = *args++; + } + result = PyEval_EvalFrameEx(f,0); + ++tstate->recursion_depth; + Py_DECREF(f); + --tstate->recursion_depth; + return result; +} +static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, Py_ssize_t nargs, PyObject *kwargs) { + PyCodeObject *co = (PyCodeObject *)PyFunction_GET_CODE(func); + PyObject *globals = PyFunction_GET_GLOBALS(func); + PyObject *argdefs = PyFunction_GET_DEFAULTS(func); + PyObject *closure; +#if PY_MAJOR_VERSION >= 3 + PyObject *kwdefs; +#endif + PyObject *kwtuple, **k; + PyObject **d; + Py_ssize_t nd; + Py_ssize_t nk; + PyObject *result; + assert(kwargs == NULL || PyDict_Check(kwargs)); + nk = kwargs ? PyDict_Size(kwargs) : 0; + #if PY_MAJOR_VERSION < 3 + if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object"))) { + return NULL; + } + #else + if (unlikely(Py_EnterRecursiveCall(" while calling a Python object"))) { + return NULL; + } + #endif + if ( +#if PY_MAJOR_VERSION >= 3 + co->co_kwonlyargcount == 0 && +#endif + likely(kwargs == NULL || nk == 0) && + co->co_flags == (CO_OPTIMIZED | CO_NEWLOCALS | CO_NOFREE)) { + if (argdefs == NULL && co->co_argcount == nargs) { + result = __Pyx_PyFunction_FastCallNoKw(co, args, nargs, globals); + goto done; + } + else if (nargs == 0 && argdefs != NULL + && co->co_argcount == Py_SIZE(argdefs)) { + /* function called with no arguments, but all parameters have + a default value: use default values as arguments .*/ + args = &PyTuple_GET_ITEM(argdefs, 0); + result =__Pyx_PyFunction_FastCallNoKw(co, args, Py_SIZE(argdefs), globals); + goto done; + } + } + if (kwargs != NULL) { + Py_ssize_t pos, i; + kwtuple = PyTuple_New(2 * nk); + if (kwtuple == NULL) { + result = NULL; + goto done; + } + k = &PyTuple_GET_ITEM(kwtuple, 0); + pos = i = 0; + while (PyDict_Next(kwargs, &pos, &k[i], &k[i+1])) { + Py_INCREF(k[i]); + Py_INCREF(k[i+1]); + i += 2; + } + nk = i / 2; + } + else { + kwtuple = NULL; + k = NULL; + } + closure = PyFunction_GET_CLOSURE(func); +#if PY_MAJOR_VERSION >= 3 + kwdefs = PyFunction_GET_KW_DEFAULTS(func); +#endif + if (argdefs != NULL) { + d = &PyTuple_GET_ITEM(argdefs, 0); + nd = Py_SIZE(argdefs); + } + else { + d = NULL; + nd = 0; + } +#if PY_MAJOR_VERSION >= 3 + result = PyEval_EvalCodeEx((PyObject*)co, globals, (PyObject *)NULL, + args, (int)nargs, + k, (int)nk, + d, (int)nd, kwdefs, closure); +#else + result = PyEval_EvalCodeEx(co, globals, (PyObject *)NULL, + args, (int)nargs, + k, (int)nk, + d, (int)nd, closure); +#endif + Py_XDECREF(kwtuple); +done: + Py_LeaveRecursiveCall(); + return result; +} +#endif + +/* PyObjectCall */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw) { + PyObject *result; + ternaryfunc call = Py_TYPE(func)->tp_call; + if (unlikely(!call)) + return PyObject_Call(func, arg, kw); + #if PY_MAJOR_VERSION < 3 + if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object"))) + return NULL; + #else + if (unlikely(Py_EnterRecursiveCall(" while calling a Python object"))) + return NULL; + #endif + result = (*call)(func, arg, kw); + Py_LeaveRecursiveCall(); + if (unlikely(!result) && unlikely(!PyErr_Occurred())) { + PyErr_SetString( + PyExc_SystemError, + "NULL result without error in PyObject_Call"); + } + return result; +} +#endif + +/* PyObjectCallMethO */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg) { + PyObject *self, *result; + PyCFunction cfunc; + cfunc = __Pyx_CyOrPyCFunction_GET_FUNCTION(func); + self = __Pyx_CyOrPyCFunction_GET_SELF(func); + #if PY_MAJOR_VERSION < 3 + if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object"))) + return NULL; + #else + if (unlikely(Py_EnterRecursiveCall(" while calling a Python object"))) + return NULL; + #endif + result = cfunc(self, arg); + Py_LeaveRecursiveCall(); + if (unlikely(!result) && unlikely(!PyErr_Occurred())) { + PyErr_SetString( + PyExc_SystemError, + "NULL result without error in PyObject_Call"); + } + return result; +} +#endif + +/* PyObjectFastCall */ +#if PY_VERSION_HEX < 0x03090000 || CYTHON_COMPILING_IN_LIMITED_API +static PyObject* __Pyx_PyObject_FastCall_fallback(PyObject *func, PyObject **args, size_t nargs, PyObject *kwargs) { + PyObject *argstuple; + PyObject *result = 0; + size_t i; + argstuple = PyTuple_New((Py_ssize_t)nargs); + if (unlikely(!argstuple)) return NULL; + for (i = 0; i < nargs; i++) { + Py_INCREF(args[i]); + if (__Pyx_PyTuple_SET_ITEM(argstuple, (Py_ssize_t)i, args[i]) < 0) goto bad; + } + result = __Pyx_PyObject_Call(func, argstuple, kwargs); + bad: + Py_DECREF(argstuple); + return result; +} +#endif +static CYTHON_INLINE PyObject* __Pyx_PyObject_FastCallDict(PyObject *func, PyObject **args, size_t _nargs, PyObject *kwargs) { + Py_ssize_t nargs = __Pyx_PyVectorcall_NARGS(_nargs); +#if CYTHON_COMPILING_IN_CPYTHON + if (nargs == 0 && kwargs == NULL) { + if (__Pyx_CyOrPyCFunction_Check(func) && likely( __Pyx_CyOrPyCFunction_GET_FLAGS(func) & METH_NOARGS)) + return __Pyx_PyObject_CallMethO(func, NULL); + } + else if (nargs == 1 && kwargs == NULL) { + if (__Pyx_CyOrPyCFunction_Check(func) && likely( __Pyx_CyOrPyCFunction_GET_FLAGS(func) & METH_O)) + return __Pyx_PyObject_CallMethO(func, args[0]); + } +#endif + #if PY_VERSION_HEX < 0x030800B1 + #if CYTHON_FAST_PYCCALL + if (PyCFunction_Check(func)) { + if (kwargs) { + return _PyCFunction_FastCallDict(func, args, nargs, kwargs); + } else { + return _PyCFunction_FastCallKeywords(func, args, nargs, NULL); + } + } + #if PY_VERSION_HEX >= 0x030700A1 + if (!kwargs && __Pyx_IS_TYPE(func, &PyMethodDescr_Type)) { + return _PyMethodDescr_FastCallKeywords(func, args, nargs, NULL); + } + #endif + #endif + #if CYTHON_FAST_PYCALL + if (PyFunction_Check(func)) { + return __Pyx_PyFunction_FastCallDict(func, args, nargs, kwargs); + } + #endif + #endif + if (kwargs == NULL) { + #if CYTHON_VECTORCALL + #if PY_VERSION_HEX < 0x03090000 + vectorcallfunc f = _PyVectorcall_Function(func); + #else + vectorcallfunc f = PyVectorcall_Function(func); + #endif + if (f) { + return f(func, args, (size_t)nargs, NULL); + } + #elif defined(__Pyx_CyFunction_USED) && CYTHON_BACKPORT_VECTORCALL + if (__Pyx_CyFunction_CheckExact(func)) { + __pyx_vectorcallfunc f = __Pyx_CyFunction_func_vectorcall(func); + if (f) return f(func, args, (size_t)nargs, NULL); + } + #endif + } + if (nargs == 0) { + return __Pyx_PyObject_Call(func, __pyx_empty_tuple, kwargs); + } + #if PY_VERSION_HEX >= 0x03090000 && !CYTHON_COMPILING_IN_LIMITED_API + return PyObject_VectorcallDict(func, args, (size_t)nargs, kwargs); + #else + return __Pyx_PyObject_FastCall_fallback(func, args, (size_t)nargs, kwargs); + #endif +} + +/* RaiseUnexpectedTypeError */ +static int +__Pyx_RaiseUnexpectedTypeError(const char *expected, PyObject *obj) +{ + __Pyx_TypeName obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, "Expected %s, got " __Pyx_FMT_TYPENAME, + expected, obj_type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return 0; +} + +/* CIntToDigits */ +static const char DIGIT_PAIRS_10[2*10*10+1] = { + "00010203040506070809" + "10111213141516171819" + "20212223242526272829" + "30313233343536373839" + "40414243444546474849" + "50515253545556575859" + "60616263646566676869" + "70717273747576777879" + "80818283848586878889" + "90919293949596979899" +}; +static const char DIGIT_PAIRS_8[2*8*8+1] = { + "0001020304050607" + "1011121314151617" + "2021222324252627" + "3031323334353637" + "4041424344454647" + "5051525354555657" + "6061626364656667" + "7071727374757677" +}; +static const char DIGITS_HEX[2*16+1] = { + "0123456789abcdef" + "0123456789ABCDEF" +}; + +/* BuildPyUnicode */ +static PyObject* __Pyx_PyUnicode_BuildFromAscii(Py_ssize_t ulength, char* chars, int clength, + int prepend_sign, char padding_char) { + PyObject *uval; + Py_ssize_t uoffset = ulength - clength; +#if CYTHON_USE_UNICODE_INTERNALS + Py_ssize_t i; +#if CYTHON_PEP393_ENABLED + void *udata; + uval = PyUnicode_New(ulength, 127); + if (unlikely(!uval)) return NULL; + udata = PyUnicode_DATA(uval); +#else + Py_UNICODE *udata; + uval = PyUnicode_FromUnicode(NULL, ulength); + if (unlikely(!uval)) return NULL; + udata = PyUnicode_AS_UNICODE(uval); +#endif + if (uoffset > 0) { + i = 0; + if (prepend_sign) { + __Pyx_PyUnicode_WRITE(PyUnicode_1BYTE_KIND, udata, 0, '-'); + i++; + } + for (; i < uoffset; i++) { + __Pyx_PyUnicode_WRITE(PyUnicode_1BYTE_KIND, udata, i, padding_char); + } + } + for (i=0; i < clength; i++) { + __Pyx_PyUnicode_WRITE(PyUnicode_1BYTE_KIND, udata, uoffset+i, chars[i]); + } +#else + { + PyObject *sign = NULL, *padding = NULL; + uval = NULL; + if (uoffset > 0) { + prepend_sign = !!prepend_sign; + if (uoffset > prepend_sign) { + padding = PyUnicode_FromOrdinal(padding_char); + if (likely(padding) && uoffset > prepend_sign + 1) { + PyObject *tmp; + PyObject *repeat = PyInt_FromSsize_t(uoffset - prepend_sign); + if (unlikely(!repeat)) goto done_or_error; + tmp = PyNumber_Multiply(padding, repeat); + Py_DECREF(repeat); + Py_DECREF(padding); + padding = tmp; + } + if (unlikely(!padding)) goto done_or_error; + } + if (prepend_sign) { + sign = PyUnicode_FromOrdinal('-'); + if (unlikely(!sign)) goto done_or_error; + } + } + uval = PyUnicode_DecodeASCII(chars, clength, NULL); + if (likely(uval) && padding) { + PyObject *tmp = PyNumber_Add(padding, uval); + Py_DECREF(uval); + uval = tmp; + } + if (likely(uval) && sign) { + PyObject *tmp = PyNumber_Add(sign, uval); + Py_DECREF(uval); + uval = tmp; + } +done_or_error: + Py_XDECREF(padding); + Py_XDECREF(sign); + } +#endif + return uval; +} + +/* CIntToPyUnicode */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_int(int value, Py_ssize_t width, char padding_char, char format_char) { + char digits[sizeof(int)*3+2]; + char *dpos, *end = digits + sizeof(int)*3+2; + const char *hex_digits = DIGITS_HEX; + Py_ssize_t length, ulength; + int prepend_sign, last_one_off; + int remaining; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const int neg_one = (int) -1, const_zero = (int) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (format_char == 'X') { + hex_digits += 16; + format_char = 'x'; + } + remaining = value; + last_one_off = 0; + dpos = end; + do { + int digit_pos; + switch (format_char) { + case 'o': + digit_pos = abs((int)(remaining % (8*8))); + remaining = (int) (remaining / (8*8)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_8 + digit_pos * 2, 2); + last_one_off = (digit_pos < 8); + break; + case 'd': + digit_pos = abs((int)(remaining % (10*10))); + remaining = (int) (remaining / (10*10)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_10 + digit_pos * 2, 2); + last_one_off = (digit_pos < 10); + break; + case 'x': + *(--dpos) = hex_digits[abs((int)(remaining % 16))]; + remaining = (int) (remaining / 16); + break; + default: + assert(0); + break; + } + } while (unlikely(remaining != 0)); + assert(!last_one_off || *dpos == '0'); + dpos += last_one_off; + length = end - dpos; + ulength = length; + prepend_sign = 0; + if (!is_unsigned && value <= neg_one) { + if (padding_char == ' ' || width <= length + 1) { + *(--dpos) = '-'; + ++length; + } else { + prepend_sign = 1; + } + ++ulength; + } + if (width > ulength) { + ulength = width; + } + if (ulength == 1) { + return PyUnicode_FromOrdinal(*dpos); + } + return __Pyx_PyUnicode_BuildFromAscii(ulength, dpos, (int) length, prepend_sign, padding_char); +} + +/* CIntToPyUnicode */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_Py_ssize_t(Py_ssize_t value, Py_ssize_t width, char padding_char, char format_char) { + char digits[sizeof(Py_ssize_t)*3+2]; + char *dpos, *end = digits + sizeof(Py_ssize_t)*3+2; + const char *hex_digits = DIGITS_HEX; + Py_ssize_t length, ulength; + int prepend_sign, last_one_off; + Py_ssize_t remaining; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const Py_ssize_t neg_one = (Py_ssize_t) -1, const_zero = (Py_ssize_t) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (format_char == 'X') { + hex_digits += 16; + format_char = 'x'; + } + remaining = value; + last_one_off = 0; + dpos = end; + do { + int digit_pos; + switch (format_char) { + case 'o': + digit_pos = abs((int)(remaining % (8*8))); + remaining = (Py_ssize_t) (remaining / (8*8)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_8 + digit_pos * 2, 2); + last_one_off = (digit_pos < 8); + break; + case 'd': + digit_pos = abs((int)(remaining % (10*10))); + remaining = (Py_ssize_t) (remaining / (10*10)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_10 + digit_pos * 2, 2); + last_one_off = (digit_pos < 10); + break; + case 'x': + *(--dpos) = hex_digits[abs((int)(remaining % 16))]; + remaining = (Py_ssize_t) (remaining / 16); + break; + default: + assert(0); + break; + } + } while (unlikely(remaining != 0)); + assert(!last_one_off || *dpos == '0'); + dpos += last_one_off; + length = end - dpos; + ulength = length; + prepend_sign = 0; + if (!is_unsigned && value <= neg_one) { + if (padding_char == ' ' || width <= length + 1) { + *(--dpos) = '-'; + ++length; + } else { + prepend_sign = 1; + } + ++ulength; + } + if (width > ulength) { + ulength = width; + } + if (ulength == 1) { + return PyUnicode_FromOrdinal(*dpos); + } + return __Pyx_PyUnicode_BuildFromAscii(ulength, dpos, (int) length, prepend_sign, padding_char); +} + +/* JoinPyUnicode */ +static PyObject* __Pyx_PyUnicode_Join(PyObject* value_tuple, Py_ssize_t value_count, Py_ssize_t result_ulength, + Py_UCS4 max_char) { +#if CYTHON_USE_UNICODE_INTERNALS && CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + PyObject *result_uval; + int result_ukind, kind_shift; + Py_ssize_t i, char_pos; + void *result_udata; + CYTHON_MAYBE_UNUSED_VAR(max_char); +#if CYTHON_PEP393_ENABLED + result_uval = PyUnicode_New(result_ulength, max_char); + if (unlikely(!result_uval)) return NULL; + result_ukind = (max_char <= 255) ? PyUnicode_1BYTE_KIND : (max_char <= 65535) ? PyUnicode_2BYTE_KIND : PyUnicode_4BYTE_KIND; + kind_shift = (result_ukind == PyUnicode_4BYTE_KIND) ? 2 : result_ukind - 1; + result_udata = PyUnicode_DATA(result_uval); +#else + result_uval = PyUnicode_FromUnicode(NULL, result_ulength); + if (unlikely(!result_uval)) return NULL; + result_ukind = sizeof(Py_UNICODE); + kind_shift = (result_ukind == 4) ? 2 : result_ukind - 1; + result_udata = PyUnicode_AS_UNICODE(result_uval); +#endif + assert(kind_shift == 2 || kind_shift == 1 || kind_shift == 0); + char_pos = 0; + for (i=0; i < value_count; i++) { + int ukind; + Py_ssize_t ulength; + void *udata; + PyObject *uval = PyTuple_GET_ITEM(value_tuple, i); + if (unlikely(__Pyx_PyUnicode_READY(uval))) + goto bad; + ulength = __Pyx_PyUnicode_GET_LENGTH(uval); + if (unlikely(!ulength)) + continue; + if (unlikely((PY_SSIZE_T_MAX >> kind_shift) - ulength < char_pos)) + goto overflow; + ukind = __Pyx_PyUnicode_KIND(uval); + udata = __Pyx_PyUnicode_DATA(uval); + if (!CYTHON_PEP393_ENABLED || ukind == result_ukind) { + memcpy((char *)result_udata + (char_pos << kind_shift), udata, (size_t) (ulength << kind_shift)); + } else { + #if PY_VERSION_HEX >= 0x030d0000 + if (unlikely(PyUnicode_CopyCharacters(result_uval, char_pos, uval, 0, ulength) < 0)) goto bad; + #elif CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030300F0 || defined(_PyUnicode_FastCopyCharacters) + _PyUnicode_FastCopyCharacters(result_uval, char_pos, uval, 0, ulength); + #else + Py_ssize_t j; + for (j=0; j < ulength; j++) { + Py_UCS4 uchar = __Pyx_PyUnicode_READ(ukind, udata, j); + __Pyx_PyUnicode_WRITE(result_ukind, result_udata, char_pos+j, uchar); + } + #endif + } + char_pos += ulength; + } + return result_uval; +overflow: + PyErr_SetString(PyExc_OverflowError, "join() result is too long for a Python string"); +bad: + Py_DECREF(result_uval); + return NULL; +#else + CYTHON_UNUSED_VAR(max_char); + CYTHON_UNUSED_VAR(result_ulength); + CYTHON_UNUSED_VAR(value_count); + return PyUnicode_Join(__pyx_empty_unicode, value_tuple); +#endif +} + +/* GetAttr */ +static CYTHON_INLINE PyObject *__Pyx_GetAttr(PyObject *o, PyObject *n) { +#if CYTHON_USE_TYPE_SLOTS +#if PY_MAJOR_VERSION >= 3 + if (likely(PyUnicode_Check(n))) +#else + if (likely(PyString_Check(n))) +#endif + return __Pyx_PyObject_GetAttrStr(o, n); +#endif + return PyObject_GetAttr(o, n); +} + +/* GetItemInt */ +static PyObject *__Pyx_GetItemInt_Generic(PyObject *o, PyObject* j) { + PyObject *r; + if (unlikely(!j)) return NULL; + r = PyObject_GetItem(o, j); + Py_DECREF(j); + return r; +} +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_List_Fast(PyObject *o, Py_ssize_t i, + CYTHON_NCP_UNUSED int wraparound, + CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + Py_ssize_t wrapped_i = i; + if (wraparound & unlikely(i < 0)) { + wrapped_i += PyList_GET_SIZE(o); + } + if ((!boundscheck) || likely(__Pyx_is_valid_index(wrapped_i, PyList_GET_SIZE(o)))) { + PyObject *r = PyList_GET_ITEM(o, wrapped_i); + Py_INCREF(r); + return r; + } + return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i)); +#else + return PySequence_GetItem(o, i); +#endif +} +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Tuple_Fast(PyObject *o, Py_ssize_t i, + CYTHON_NCP_UNUSED int wraparound, + CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + Py_ssize_t wrapped_i = i; + if (wraparound & unlikely(i < 0)) { + wrapped_i += PyTuple_GET_SIZE(o); + } + if ((!boundscheck) || likely(__Pyx_is_valid_index(wrapped_i, PyTuple_GET_SIZE(o)))) { + PyObject *r = PyTuple_GET_ITEM(o, wrapped_i); + Py_INCREF(r); + return r; + } + return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i)); +#else + return PySequence_GetItem(o, i); +#endif +} +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Fast(PyObject *o, Py_ssize_t i, int is_list, + CYTHON_NCP_UNUSED int wraparound, + CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS && CYTHON_USE_TYPE_SLOTS + if (is_list || PyList_CheckExact(o)) { + Py_ssize_t n = ((!wraparound) | likely(i >= 0)) ? i : i + PyList_GET_SIZE(o); + if ((!boundscheck) || (likely(__Pyx_is_valid_index(n, PyList_GET_SIZE(o))))) { + PyObject *r = PyList_GET_ITEM(o, n); + Py_INCREF(r); + return r; + } + } + else if (PyTuple_CheckExact(o)) { + Py_ssize_t n = ((!wraparound) | likely(i >= 0)) ? i : i + PyTuple_GET_SIZE(o); + if ((!boundscheck) || likely(__Pyx_is_valid_index(n, PyTuple_GET_SIZE(o)))) { + PyObject *r = PyTuple_GET_ITEM(o, n); + Py_INCREF(r); + return r; + } + } else { + PyMappingMethods *mm = Py_TYPE(o)->tp_as_mapping; + PySequenceMethods *sm = Py_TYPE(o)->tp_as_sequence; + if (mm && mm->mp_subscript) { + PyObject *r, *key = PyInt_FromSsize_t(i); + if (unlikely(!key)) return NULL; + r = mm->mp_subscript(o, key); + Py_DECREF(key); + return r; + } + if (likely(sm && sm->sq_item)) { + if (wraparound && unlikely(i < 0) && likely(sm->sq_length)) { + Py_ssize_t l = sm->sq_length(o); + if (likely(l >= 0)) { + i += l; + } else { + if (!PyErr_ExceptionMatches(PyExc_OverflowError)) + return NULL; + PyErr_Clear(); + } + } + return sm->sq_item(o, i); + } + } +#else + if (is_list || !PyMapping_Check(o)) { + return PySequence_GetItem(o, i); + } +#endif + return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i)); +} + +/* PyObjectCallOneArg */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg) { + PyObject *args[2] = {NULL, arg}; + return __Pyx_PyObject_FastCall(func, args+1, 1 | __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET); +} + +/* ObjectGetItem */ +#if CYTHON_USE_TYPE_SLOTS +static PyObject *__Pyx_PyObject_GetIndex(PyObject *obj, PyObject *index) { + PyObject *runerr = NULL; + Py_ssize_t key_value; + key_value = __Pyx_PyIndex_AsSsize_t(index); + if (likely(key_value != -1 || !(runerr = PyErr_Occurred()))) { + return __Pyx_GetItemInt_Fast(obj, key_value, 0, 1, 1); + } + if (PyErr_GivenExceptionMatches(runerr, PyExc_OverflowError)) { + __Pyx_TypeName index_type_name = __Pyx_PyType_GetName(Py_TYPE(index)); + PyErr_Clear(); + PyErr_Format(PyExc_IndexError, + "cannot fit '" __Pyx_FMT_TYPENAME "' into an index-sized integer", index_type_name); + __Pyx_DECREF_TypeName(index_type_name); + } + return NULL; +} +static PyObject *__Pyx_PyObject_GetItem_Slow(PyObject *obj, PyObject *key) { + __Pyx_TypeName obj_type_name; + if (likely(PyType_Check(obj))) { + PyObject *meth = __Pyx_PyObject_GetAttrStrNoError(obj, __pyx_n_s_class_getitem); + if (!meth) { + PyErr_Clear(); + } else { + PyObject *result = __Pyx_PyObject_CallOneArg(meth, key); + Py_DECREF(meth); + return result; + } + } + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, + "'" __Pyx_FMT_TYPENAME "' object is not subscriptable", obj_type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return NULL; +} +static PyObject *__Pyx_PyObject_GetItem(PyObject *obj, PyObject *key) { + PyTypeObject *tp = Py_TYPE(obj); + PyMappingMethods *mm = tp->tp_as_mapping; + PySequenceMethods *sm = tp->tp_as_sequence; + if (likely(mm && mm->mp_subscript)) { + return mm->mp_subscript(obj, key); + } + if (likely(sm && sm->sq_item)) { + return __Pyx_PyObject_GetIndex(obj, key); + } + return __Pyx_PyObject_GetItem_Slow(obj, key); +} +#endif + +/* KeywordStringCheck */ +static int __Pyx_CheckKeywordStrings( + PyObject *kw, + const char* function_name, + int kw_allowed) +{ + PyObject* key = 0; + Py_ssize_t pos = 0; +#if CYTHON_COMPILING_IN_PYPY + if (!kw_allowed && PyDict_Next(kw, &pos, &key, 0)) + goto invalid_keyword; + return 1; +#else + if (CYTHON_METH_FASTCALL && likely(PyTuple_Check(kw))) { + Py_ssize_t kwsize; +#if CYTHON_ASSUME_SAFE_MACROS + kwsize = PyTuple_GET_SIZE(kw); +#else + kwsize = PyTuple_Size(kw); + if (kwsize < 0) return 0; +#endif + if (unlikely(kwsize == 0)) + return 1; + if (!kw_allowed) { +#if CYTHON_ASSUME_SAFE_MACROS + key = PyTuple_GET_ITEM(kw, 0); +#else + key = PyTuple_GetItem(kw, pos); + if (!key) return 0; +#endif + goto invalid_keyword; + } +#if PY_VERSION_HEX < 0x03090000 + for (pos = 0; pos < kwsize; pos++) { +#if CYTHON_ASSUME_SAFE_MACROS + key = PyTuple_GET_ITEM(kw, pos); +#else + key = PyTuple_GetItem(kw, pos); + if (!key) return 0; +#endif + if (unlikely(!PyUnicode_Check(key))) + goto invalid_keyword_type; + } +#endif + return 1; + } + while (PyDict_Next(kw, &pos, &key, 0)) { + #if PY_MAJOR_VERSION < 3 + if (unlikely(!PyString_Check(key))) + #endif + if (unlikely(!PyUnicode_Check(key))) + goto invalid_keyword_type; + } + if (!kw_allowed && unlikely(key)) + goto invalid_keyword; + return 1; +invalid_keyword_type: + PyErr_Format(PyExc_TypeError, + "%.200s() keywords must be strings", function_name); + return 0; +#endif +invalid_keyword: + #if PY_MAJOR_VERSION < 3 + PyErr_Format(PyExc_TypeError, + "%.200s() got an unexpected keyword argument '%.200s'", + function_name, PyString_AsString(key)); + #else + PyErr_Format(PyExc_TypeError, + "%s() got an unexpected keyword argument '%U'", + function_name, key); + #endif + return 0; +} + +/* DivInt[Py_ssize_t] */ +static CYTHON_INLINE Py_ssize_t __Pyx_div_Py_ssize_t(Py_ssize_t a, Py_ssize_t b) { + Py_ssize_t q = a / b; + Py_ssize_t r = a - q*b; + q -= ((r != 0) & ((r ^ b) < 0)); + return q; +} + +/* GetAttr3 */ +#if __PYX_LIMITED_VERSION_HEX < 0x030d00A1 +static PyObject *__Pyx_GetAttr3Default(PyObject *d) { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + if (unlikely(!__Pyx_PyErr_ExceptionMatches(PyExc_AttributeError))) + return NULL; + __Pyx_PyErr_Clear(); + Py_INCREF(d); + return d; +} +#endif +static CYTHON_INLINE PyObject *__Pyx_GetAttr3(PyObject *o, PyObject *n, PyObject *d) { + PyObject *r; +#if __PYX_LIMITED_VERSION_HEX >= 0x030d00A1 + int res = PyObject_GetOptionalAttr(o, n, &r); + return (res != 0) ? r : __Pyx_NewRef(d); +#else + #if CYTHON_USE_TYPE_SLOTS + if (likely(PyString_Check(n))) { + r = __Pyx_PyObject_GetAttrStrNoError(o, n); + if (unlikely(!r) && likely(!PyErr_Occurred())) { + r = __Pyx_NewRef(d); + } + return r; + } + #endif + r = PyObject_GetAttr(o, n); + return (likely(r)) ? r : __Pyx_GetAttr3Default(d); +#endif +} + +/* PyDictVersioning */ +#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj) { + PyObject *dict = Py_TYPE(obj)->tp_dict; + return likely(dict) ? __PYX_GET_DICT_VERSION(dict) : 0; +} +static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj) { + PyObject **dictptr = NULL; + Py_ssize_t offset = Py_TYPE(obj)->tp_dictoffset; + if (offset) { +#if CYTHON_COMPILING_IN_CPYTHON + dictptr = (likely(offset > 0)) ? (PyObject **) ((char *)obj + offset) : _PyObject_GetDictPtr(obj); +#else + dictptr = _PyObject_GetDictPtr(obj); +#endif + } + return (dictptr && *dictptr) ? __PYX_GET_DICT_VERSION(*dictptr) : 0; +} +static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version) { + PyObject *dict = Py_TYPE(obj)->tp_dict; + if (unlikely(!dict) || unlikely(tp_dict_version != __PYX_GET_DICT_VERSION(dict))) + return 0; + return obj_dict_version == __Pyx_get_object_dict_version(obj); +} +#endif + +/* GetModuleGlobalName */ +#if CYTHON_USE_DICT_VERSIONS +static PyObject *__Pyx__GetModuleGlobalName(PyObject *name, PY_UINT64_T *dict_version, PyObject **dict_cached_value) +#else +static CYTHON_INLINE PyObject *__Pyx__GetModuleGlobalName(PyObject *name) +#endif +{ + PyObject *result; +#if !CYTHON_AVOID_BORROWED_REFS +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030500A1 && PY_VERSION_HEX < 0x030d0000 + result = _PyDict_GetItem_KnownHash(__pyx_d, name, ((PyASCIIObject *) name)->hash); + __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version) + if (likely(result)) { + return __Pyx_NewRef(result); + } else if (unlikely(PyErr_Occurred())) { + return NULL; + } +#elif CYTHON_COMPILING_IN_LIMITED_API + if (unlikely(!__pyx_m)) { + return NULL; + } + result = PyObject_GetAttr(__pyx_m, name); + if (likely(result)) { + return result; + } +#else + result = PyDict_GetItem(__pyx_d, name); + __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version) + if (likely(result)) { + return __Pyx_NewRef(result); + } +#endif +#else + result = PyObject_GetItem(__pyx_d, name); + __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version) + if (likely(result)) { + return __Pyx_NewRef(result); + } + PyErr_Clear(); +#endif + return __Pyx_GetBuiltinName(name); +} + +/* RaiseTooManyValuesToUnpack */ +static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected) { + PyErr_Format(PyExc_ValueError, + "too many values to unpack (expected %" CYTHON_FORMAT_SSIZE_T "d)", expected); +} + +/* RaiseNeedMoreValuesToUnpack */ +static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index) { + PyErr_Format(PyExc_ValueError, + "need more than %" CYTHON_FORMAT_SSIZE_T "d value%.1s to unpack", + index, (index == 1) ? "" : "s"); +} + +/* RaiseNoneIterError */ +static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable"); +} + +/* ExtTypeTest */ +static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type) { + __Pyx_TypeName obj_type_name; + __Pyx_TypeName type_name; + if (unlikely(!type)) { + PyErr_SetString(PyExc_SystemError, "Missing type object"); + return 0; + } + if (likely(__Pyx_TypeCheck(obj, type))) + return 1; + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + type_name = __Pyx_PyType_GetName(type); + PyErr_Format(PyExc_TypeError, + "Cannot convert " __Pyx_FMT_TYPENAME " to " __Pyx_FMT_TYPENAME, + obj_type_name, type_name); + __Pyx_DECREF_TypeName(obj_type_name); + __Pyx_DECREF_TypeName(type_name); + return 0; +} + +/* GetTopmostException */ +#if CYTHON_USE_EXC_INFO_STACK && CYTHON_FAST_THREAD_STATE +static _PyErr_StackItem * +__Pyx_PyErr_GetTopmostException(PyThreadState *tstate) +{ + _PyErr_StackItem *exc_info = tstate->exc_info; + while ((exc_info->exc_value == NULL || exc_info->exc_value == Py_None) && + exc_info->previous_item != NULL) + { + exc_info = exc_info->previous_item; + } + return exc_info; +} +#endif + +/* SaveResetException */ +#if CYTHON_FAST_THREAD_STATE +static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) { + #if CYTHON_USE_EXC_INFO_STACK && PY_VERSION_HEX >= 0x030B00a4 + _PyErr_StackItem *exc_info = __Pyx_PyErr_GetTopmostException(tstate); + PyObject *exc_value = exc_info->exc_value; + if (exc_value == NULL || exc_value == Py_None) { + *value = NULL; + *type = NULL; + *tb = NULL; + } else { + *value = exc_value; + Py_INCREF(*value); + *type = (PyObject*) Py_TYPE(exc_value); + Py_INCREF(*type); + *tb = PyException_GetTraceback(exc_value); + } + #elif CYTHON_USE_EXC_INFO_STACK + _PyErr_StackItem *exc_info = __Pyx_PyErr_GetTopmostException(tstate); + *type = exc_info->exc_type; + *value = exc_info->exc_value; + *tb = exc_info->exc_traceback; + Py_XINCREF(*type); + Py_XINCREF(*value); + Py_XINCREF(*tb); + #else + *type = tstate->exc_type; + *value = tstate->exc_value; + *tb = tstate->exc_traceback; + Py_XINCREF(*type); + Py_XINCREF(*value); + Py_XINCREF(*tb); + #endif +} +static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) { + #if CYTHON_USE_EXC_INFO_STACK && PY_VERSION_HEX >= 0x030B00a4 + _PyErr_StackItem *exc_info = tstate->exc_info; + PyObject *tmp_value = exc_info->exc_value; + exc_info->exc_value = value; + Py_XDECREF(tmp_value); + Py_XDECREF(type); + Py_XDECREF(tb); + #else + PyObject *tmp_type, *tmp_value, *tmp_tb; + #if CYTHON_USE_EXC_INFO_STACK + _PyErr_StackItem *exc_info = tstate->exc_info; + tmp_type = exc_info->exc_type; + tmp_value = exc_info->exc_value; + tmp_tb = exc_info->exc_traceback; + exc_info->exc_type = type; + exc_info->exc_value = value; + exc_info->exc_traceback = tb; + #else + tmp_type = tstate->exc_type; + tmp_value = tstate->exc_value; + tmp_tb = tstate->exc_traceback; + tstate->exc_type = type; + tstate->exc_value = value; + tstate->exc_traceback = tb; + #endif + Py_XDECREF(tmp_type); + Py_XDECREF(tmp_value); + Py_XDECREF(tmp_tb); + #endif +} +#endif + +/* GetException */ +#if CYTHON_FAST_THREAD_STATE +static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) +#else +static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb) +#endif +{ + PyObject *local_type = NULL, *local_value, *local_tb = NULL; +#if CYTHON_FAST_THREAD_STATE + PyObject *tmp_type, *tmp_value, *tmp_tb; + #if PY_VERSION_HEX >= 0x030C00A6 + local_value = tstate->current_exception; + tstate->current_exception = 0; + if (likely(local_value)) { + local_type = (PyObject*) Py_TYPE(local_value); + Py_INCREF(local_type); + local_tb = PyException_GetTraceback(local_value); + } + #else + local_type = tstate->curexc_type; + local_value = tstate->curexc_value; + local_tb = tstate->curexc_traceback; + tstate->curexc_type = 0; + tstate->curexc_value = 0; + tstate->curexc_traceback = 0; + #endif +#else + PyErr_Fetch(&local_type, &local_value, &local_tb); +#endif + PyErr_NormalizeException(&local_type, &local_value, &local_tb); +#if CYTHON_FAST_THREAD_STATE && PY_VERSION_HEX >= 0x030C00A6 + if (unlikely(tstate->current_exception)) +#elif CYTHON_FAST_THREAD_STATE + if (unlikely(tstate->curexc_type)) +#else + if (unlikely(PyErr_Occurred())) +#endif + goto bad; + #if PY_MAJOR_VERSION >= 3 + if (local_tb) { + if (unlikely(PyException_SetTraceback(local_value, local_tb) < 0)) + goto bad; + } + #endif + Py_XINCREF(local_tb); + Py_XINCREF(local_type); + Py_XINCREF(local_value); + *type = local_type; + *value = local_value; + *tb = local_tb; +#if CYTHON_FAST_THREAD_STATE + #if CYTHON_USE_EXC_INFO_STACK + { + _PyErr_StackItem *exc_info = tstate->exc_info; + #if PY_VERSION_HEX >= 0x030B00a4 + tmp_value = exc_info->exc_value; + exc_info->exc_value = local_value; + tmp_type = NULL; + tmp_tb = NULL; + Py_XDECREF(local_type); + Py_XDECREF(local_tb); + #else + tmp_type = exc_info->exc_type; + tmp_value = exc_info->exc_value; + tmp_tb = exc_info->exc_traceback; + exc_info->exc_type = local_type; + exc_info->exc_value = local_value; + exc_info->exc_traceback = local_tb; + #endif + } + #else + tmp_type = tstate->exc_type; + tmp_value = tstate->exc_value; + tmp_tb = tstate->exc_traceback; + tstate->exc_type = local_type; + tstate->exc_value = local_value; + tstate->exc_traceback = local_tb; + #endif + Py_XDECREF(tmp_type); + Py_XDECREF(tmp_value); + Py_XDECREF(tmp_tb); +#else + PyErr_SetExcInfo(local_type, local_value, local_tb); +#endif + return 0; +bad: + *type = 0; + *value = 0; + *tb = 0; + Py_XDECREF(local_type); + Py_XDECREF(local_value); + Py_XDECREF(local_tb); + return -1; +} + +/* SwapException */ +#if CYTHON_FAST_THREAD_STATE +static CYTHON_INLINE void __Pyx__ExceptionSwap(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) { + PyObject *tmp_type, *tmp_value, *tmp_tb; + #if CYTHON_USE_EXC_INFO_STACK && PY_VERSION_HEX >= 0x030B00a4 + _PyErr_StackItem *exc_info = tstate->exc_info; + tmp_value = exc_info->exc_value; + exc_info->exc_value = *value; + if (tmp_value == NULL || tmp_value == Py_None) { + Py_XDECREF(tmp_value); + tmp_value = NULL; + tmp_type = NULL; + tmp_tb = NULL; + } else { + tmp_type = (PyObject*) Py_TYPE(tmp_value); + Py_INCREF(tmp_type); + #if CYTHON_COMPILING_IN_CPYTHON + tmp_tb = ((PyBaseExceptionObject*) tmp_value)->traceback; + Py_XINCREF(tmp_tb); + #else + tmp_tb = PyException_GetTraceback(tmp_value); + #endif + } + #elif CYTHON_USE_EXC_INFO_STACK + _PyErr_StackItem *exc_info = tstate->exc_info; + tmp_type = exc_info->exc_type; + tmp_value = exc_info->exc_value; + tmp_tb = exc_info->exc_traceback; + exc_info->exc_type = *type; + exc_info->exc_value = *value; + exc_info->exc_traceback = *tb; + #else + tmp_type = tstate->exc_type; + tmp_value = tstate->exc_value; + tmp_tb = tstate->exc_traceback; + tstate->exc_type = *type; + tstate->exc_value = *value; + tstate->exc_traceback = *tb; + #endif + *type = tmp_type; + *value = tmp_value; + *tb = tmp_tb; +} +#else +static CYTHON_INLINE void __Pyx_ExceptionSwap(PyObject **type, PyObject **value, PyObject **tb) { + PyObject *tmp_type, *tmp_value, *tmp_tb; + PyErr_GetExcInfo(&tmp_type, &tmp_value, &tmp_tb); + PyErr_SetExcInfo(*type, *value, *tb); + *type = tmp_type; + *value = tmp_value; + *tb = tmp_tb; +} +#endif + +/* Import */ +static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level) { + PyObject *module = 0; + PyObject *empty_dict = 0; + PyObject *empty_list = 0; + #if PY_MAJOR_VERSION < 3 + PyObject *py_import; + py_import = __Pyx_PyObject_GetAttrStr(__pyx_b, __pyx_n_s_import); + if (unlikely(!py_import)) + goto bad; + if (!from_list) { + empty_list = PyList_New(0); + if (unlikely(!empty_list)) + goto bad; + from_list = empty_list; + } + #endif + empty_dict = PyDict_New(); + if (unlikely(!empty_dict)) + goto bad; + { + #if PY_MAJOR_VERSION >= 3 + if (level == -1) { + if (strchr(__Pyx_MODULE_NAME, '.') != NULL) { + module = PyImport_ImportModuleLevelObject( + name, __pyx_d, empty_dict, from_list, 1); + if (unlikely(!module)) { + if (unlikely(!PyErr_ExceptionMatches(PyExc_ImportError))) + goto bad; + PyErr_Clear(); + } + } + level = 0; + } + #endif + if (!module) { + #if PY_MAJOR_VERSION < 3 + PyObject *py_level = PyInt_FromLong(level); + if (unlikely(!py_level)) + goto bad; + module = PyObject_CallFunctionObjArgs(py_import, + name, __pyx_d, empty_dict, from_list, py_level, (PyObject *)NULL); + Py_DECREF(py_level); + #else + module = PyImport_ImportModuleLevelObject( + name, __pyx_d, empty_dict, from_list, level); + #endif + } + } +bad: + Py_XDECREF(empty_dict); + Py_XDECREF(empty_list); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(py_import); + #endif + return module; +} + +/* ImportDottedModule */ +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx__ImportDottedModule_Error(PyObject *name, PyObject *parts_tuple, Py_ssize_t count) { + PyObject *partial_name = NULL, *slice = NULL, *sep = NULL; + if (unlikely(PyErr_Occurred())) { + PyErr_Clear(); + } + if (likely(PyTuple_GET_SIZE(parts_tuple) == count)) { + partial_name = name; + } else { + slice = PySequence_GetSlice(parts_tuple, 0, count); + if (unlikely(!slice)) + goto bad; + sep = PyUnicode_FromStringAndSize(".", 1); + if (unlikely(!sep)) + goto bad; + partial_name = PyUnicode_Join(sep, slice); + } + PyErr_Format( +#if PY_MAJOR_VERSION < 3 + PyExc_ImportError, + "No module named '%s'", PyString_AS_STRING(partial_name)); +#else +#if PY_VERSION_HEX >= 0x030600B1 + PyExc_ModuleNotFoundError, +#else + PyExc_ImportError, +#endif + "No module named '%U'", partial_name); +#endif +bad: + Py_XDECREF(sep); + Py_XDECREF(slice); + Py_XDECREF(partial_name); + return NULL; +} +#endif +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx__ImportDottedModule_Lookup(PyObject *name) { + PyObject *imported_module; +#if PY_VERSION_HEX < 0x030700A1 || (CYTHON_COMPILING_IN_PYPY && PYPY_VERSION_NUM < 0x07030400) + PyObject *modules = PyImport_GetModuleDict(); + if (unlikely(!modules)) + return NULL; + imported_module = __Pyx_PyDict_GetItemStr(modules, name); + Py_XINCREF(imported_module); +#else + imported_module = PyImport_GetModule(name); +#endif + return imported_module; +} +#endif +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx_ImportDottedModule_WalkParts(PyObject *module, PyObject *name, PyObject *parts_tuple) { + Py_ssize_t i, nparts; + nparts = PyTuple_GET_SIZE(parts_tuple); + for (i=1; i < nparts && module; i++) { + PyObject *part, *submodule; +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + part = PyTuple_GET_ITEM(parts_tuple, i); +#else + part = PySequence_ITEM(parts_tuple, i); +#endif + submodule = __Pyx_PyObject_GetAttrStrNoError(module, part); +#if !(CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS) + Py_DECREF(part); +#endif + Py_DECREF(module); + module = submodule; + } + if (unlikely(!module)) { + return __Pyx__ImportDottedModule_Error(name, parts_tuple, i); + } + return module; +} +#endif +static PyObject *__Pyx__ImportDottedModule(PyObject *name, PyObject *parts_tuple) { +#if PY_MAJOR_VERSION < 3 + PyObject *module, *from_list, *star = __pyx_n_s__3; + CYTHON_UNUSED_VAR(parts_tuple); + from_list = PyList_New(1); + if (unlikely(!from_list)) + return NULL; + Py_INCREF(star); + PyList_SET_ITEM(from_list, 0, star); + module = __Pyx_Import(name, from_list, 0); + Py_DECREF(from_list); + return module; +#else + PyObject *imported_module; + PyObject *module = __Pyx_Import(name, NULL, 0); + if (!parts_tuple || unlikely(!module)) + return module; + imported_module = __Pyx__ImportDottedModule_Lookup(name); + if (likely(imported_module)) { + Py_DECREF(module); + return imported_module; + } + PyErr_Clear(); + return __Pyx_ImportDottedModule_WalkParts(module, name, parts_tuple); +#endif +} +static PyObject *__Pyx_ImportDottedModule(PyObject *name, PyObject *parts_tuple) { +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030400B1 + PyObject *module = __Pyx__ImportDottedModule_Lookup(name); + if (likely(module)) { + PyObject *spec = __Pyx_PyObject_GetAttrStrNoError(module, __pyx_n_s_spec); + if (likely(spec)) { + PyObject *unsafe = __Pyx_PyObject_GetAttrStrNoError(spec, __pyx_n_s_initializing); + if (likely(!unsafe || !__Pyx_PyObject_IsTrue(unsafe))) { + Py_DECREF(spec); + spec = NULL; + } + Py_XDECREF(unsafe); + } + if (likely(!spec)) { + PyErr_Clear(); + return module; + } + Py_DECREF(spec); + Py_DECREF(module); + } else if (PyErr_Occurred()) { + PyErr_Clear(); + } +#endif + return __Pyx__ImportDottedModule(name, parts_tuple); +} + +/* FastTypeChecks */ +#if CYTHON_COMPILING_IN_CPYTHON +static int __Pyx_InBases(PyTypeObject *a, PyTypeObject *b) { + while (a) { + a = __Pyx_PyType_GetSlot(a, tp_base, PyTypeObject*); + if (a == b) + return 1; + } + return b == &PyBaseObject_Type; +} +static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b) { + PyObject *mro; + if (a == b) return 1; + mro = a->tp_mro; + if (likely(mro)) { + Py_ssize_t i, n; + n = PyTuple_GET_SIZE(mro); + for (i = 0; i < n; i++) { + if (PyTuple_GET_ITEM(mro, i) == (PyObject *)b) + return 1; + } + return 0; + } + return __Pyx_InBases(a, b); +} +static CYTHON_INLINE int __Pyx_IsAnySubtype2(PyTypeObject *cls, PyTypeObject *a, PyTypeObject *b) { + PyObject *mro; + if (cls == a || cls == b) return 1; + mro = cls->tp_mro; + if (likely(mro)) { + Py_ssize_t i, n; + n = PyTuple_GET_SIZE(mro); + for (i = 0; i < n; i++) { + PyObject *base = PyTuple_GET_ITEM(mro, i); + if (base == (PyObject *)a || base == (PyObject *)b) + return 1; + } + return 0; + } + return __Pyx_InBases(cls, a) || __Pyx_InBases(cls, b); +} +#if PY_MAJOR_VERSION == 2 +static int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject* exc_type2) { + PyObject *exception, *value, *tb; + int res; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&exception, &value, &tb); + res = exc_type1 ? PyObject_IsSubclass(err, exc_type1) : 0; + if (unlikely(res == -1)) { + PyErr_WriteUnraisable(err); + res = 0; + } + if (!res) { + res = PyObject_IsSubclass(err, exc_type2); + if (unlikely(res == -1)) { + PyErr_WriteUnraisable(err); + res = 0; + } + } + __Pyx_ErrRestore(exception, value, tb); + return res; +} +#else +static CYTHON_INLINE int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject *exc_type2) { + if (exc_type1) { + return __Pyx_IsAnySubtype2((PyTypeObject*)err, (PyTypeObject*)exc_type1, (PyTypeObject*)exc_type2); + } else { + return __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type2); + } +} +#endif +static int __Pyx_PyErr_GivenExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) { + Py_ssize_t i, n; + assert(PyExceptionClass_Check(exc_type)); + n = PyTuple_GET_SIZE(tuple); +#if PY_MAJOR_VERSION >= 3 + for (i=0; itp_as_sequence && type->tp_as_sequence->sq_repeat)) { + return type->tp_as_sequence->sq_repeat(seq, mul); + } else +#endif + { + return __Pyx_PySequence_Multiply_Generic(seq, mul); + } +} + +/* SetItemInt */ +static int __Pyx_SetItemInt_Generic(PyObject *o, PyObject *j, PyObject *v) { + int r; + if (unlikely(!j)) return -1; + r = PyObject_SetItem(o, j, v); + Py_DECREF(j); + return r; +} +static CYTHON_INLINE int __Pyx_SetItemInt_Fast(PyObject *o, Py_ssize_t i, PyObject *v, int is_list, + CYTHON_NCP_UNUSED int wraparound, CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS && CYTHON_USE_TYPE_SLOTS + if (is_list || PyList_CheckExact(o)) { + Py_ssize_t n = (!wraparound) ? i : ((likely(i >= 0)) ? i : i + PyList_GET_SIZE(o)); + if ((!boundscheck) || likely(__Pyx_is_valid_index(n, PyList_GET_SIZE(o)))) { + PyObject* old = PyList_GET_ITEM(o, n); + Py_INCREF(v); + PyList_SET_ITEM(o, n, v); + Py_DECREF(old); + return 1; + } + } else { + PyMappingMethods *mm = Py_TYPE(o)->tp_as_mapping; + PySequenceMethods *sm = Py_TYPE(o)->tp_as_sequence; + if (mm && mm->mp_ass_subscript) { + int r; + PyObject *key = PyInt_FromSsize_t(i); + if (unlikely(!key)) return -1; + r = mm->mp_ass_subscript(o, key, v); + Py_DECREF(key); + return r; + } + if (likely(sm && sm->sq_ass_item)) { + if (wraparound && unlikely(i < 0) && likely(sm->sq_length)) { + Py_ssize_t l = sm->sq_length(o); + if (likely(l >= 0)) { + i += l; + } else { + if (!PyErr_ExceptionMatches(PyExc_OverflowError)) + return -1; + PyErr_Clear(); + } + } + return sm->sq_ass_item(o, i, v); + } + } +#else + if (is_list || !PyMapping_Check(o)) + { + return PySequence_SetItem(o, i, v); + } +#endif + return __Pyx_SetItemInt_Generic(o, PyInt_FromSsize_t(i), v); +} + +/* RaiseUnboundLocalError */ +static CYTHON_INLINE void __Pyx_RaiseUnboundLocalError(const char *varname) { + PyErr_Format(PyExc_UnboundLocalError, "local variable '%s' referenced before assignment", varname); +} + +/* DivInt[long] */ +static CYTHON_INLINE long __Pyx_div_long(long a, long b) { + long q = a / b; + long r = a - q*b; + q -= ((r != 0) & ((r ^ b) < 0)); + return q; +} + +/* ImportFrom */ +static PyObject* __Pyx_ImportFrom(PyObject* module, PyObject* name) { + PyObject* value = __Pyx_PyObject_GetAttrStr(module, name); + if (unlikely(!value) && PyErr_ExceptionMatches(PyExc_AttributeError)) { + const char* module_name_str = 0; + PyObject* module_name = 0; + PyObject* module_dot = 0; + PyObject* full_name = 0; + PyErr_Clear(); + module_name_str = PyModule_GetName(module); + if (unlikely(!module_name_str)) { goto modbad; } + module_name = PyUnicode_FromString(module_name_str); + if (unlikely(!module_name)) { goto modbad; } + module_dot = PyUnicode_Concat(module_name, __pyx_kp_u__2); + if (unlikely(!module_dot)) { goto modbad; } + full_name = PyUnicode_Concat(module_dot, name); + if (unlikely(!full_name)) { goto modbad; } + #if PY_VERSION_HEX < 0x030700A1 || (CYTHON_COMPILING_IN_PYPY && PYPY_VERSION_NUM < 0x07030400) + { + PyObject *modules = PyImport_GetModuleDict(); + if (unlikely(!modules)) + goto modbad; + value = PyObject_GetItem(modules, full_name); + } + #else + value = PyImport_GetModule(full_name); + #endif + modbad: + Py_XDECREF(full_name); + Py_XDECREF(module_dot); + Py_XDECREF(module_name); + } + if (unlikely(!value)) { + PyErr_Format(PyExc_ImportError, + #if PY_MAJOR_VERSION < 3 + "cannot import name %.230s", PyString_AS_STRING(name)); + #else + "cannot import name %S", name); + #endif + } + return value; +} + +/* HasAttr */ +static CYTHON_INLINE int __Pyx_HasAttr(PyObject *o, PyObject *n) { + PyObject *r; + if (unlikely(!__Pyx_PyBaseString_Check(n))) { + PyErr_SetString(PyExc_TypeError, + "hasattr(): attribute name must be string"); + return -1; + } + r = __Pyx_GetAttr(o, n); + if (!r) { + PyErr_Clear(); + return 0; + } else { + Py_DECREF(r); + return 1; + } +} + +/* IsLittleEndian */ +static CYTHON_INLINE int __Pyx_Is_Little_Endian(void) +{ + union { + uint32_t u32; + uint8_t u8[4]; + } S; + S.u32 = 0x01020304; + return S.u8[0] == 4; +} + +/* BufferFormatCheck */ +static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx, + __Pyx_BufFmt_StackElem* stack, + __Pyx_TypeInfo* type) { + stack[0].field = &ctx->root; + stack[0].parent_offset = 0; + ctx->root.type = type; + ctx->root.name = "buffer dtype"; + ctx->root.offset = 0; + ctx->head = stack; + ctx->head->field = &ctx->root; + ctx->fmt_offset = 0; + ctx->head->parent_offset = 0; + ctx->new_packmode = '@'; + ctx->enc_packmode = '@'; + ctx->new_count = 1; + ctx->enc_count = 0; + ctx->enc_type = 0; + ctx->is_complex = 0; + ctx->is_valid_array = 0; + ctx->struct_alignment = 0; + while (type->typegroup == 'S') { + ++ctx->head; + ctx->head->field = type->fields; + ctx->head->parent_offset = 0; + type = type->fields->type; + } +} +static int __Pyx_BufFmt_ParseNumber(const char** ts) { + int count; + const char* t = *ts; + if (*t < '0' || *t > '9') { + return -1; + } else { + count = *t++ - '0'; + while (*t >= '0' && *t <= '9') { + count *= 10; + count += *t++ - '0'; + } + } + *ts = t; + return count; +} +static int __Pyx_BufFmt_ExpectNumber(const char **ts) { + int number = __Pyx_BufFmt_ParseNumber(ts); + if (number == -1) + PyErr_Format(PyExc_ValueError,\ + "Does not understand character buffer dtype format string ('%c')", **ts); + return number; +} +static void __Pyx_BufFmt_RaiseUnexpectedChar(char ch) { + PyErr_Format(PyExc_ValueError, + "Unexpected format string character: '%c'", ch); +} +static const char* __Pyx_BufFmt_DescribeTypeChar(char ch, int is_complex) { + switch (ch) { + case '?': return "'bool'"; + case 'c': return "'char'"; + case 'b': return "'signed char'"; + case 'B': return "'unsigned char'"; + case 'h': return "'short'"; + case 'H': return "'unsigned short'"; + case 'i': return "'int'"; + case 'I': return "'unsigned int'"; + case 'l': return "'long'"; + case 'L': return "'unsigned long'"; + case 'q': return "'long long'"; + case 'Q': return "'unsigned long long'"; + case 'f': return (is_complex ? "'complex float'" : "'float'"); + case 'd': return (is_complex ? "'complex double'" : "'double'"); + case 'g': return (is_complex ? "'complex long double'" : "'long double'"); + case 'T': return "a struct"; + case 'O': return "Python object"; + case 'P': return "a pointer"; + case 's': case 'p': return "a string"; + case 0: return "end"; + default: return "unparsable format string"; + } +} +static size_t __Pyx_BufFmt_TypeCharToStandardSize(char ch, int is_complex) { + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return 2; + case 'i': case 'I': case 'l': case 'L': return 4; + case 'q': case 'Q': return 8; + case 'f': return (is_complex ? 8 : 4); + case 'd': return (is_complex ? 16 : 8); + case 'g': { + PyErr_SetString(PyExc_ValueError, "Python does not define a standard format string size for long double ('g').."); + return 0; + } + case 'O': case 'P': return sizeof(void*); + default: + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } +} +static size_t __Pyx_BufFmt_TypeCharToNativeSize(char ch, int is_complex) { + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return sizeof(short); + case 'i': case 'I': return sizeof(int); + case 'l': case 'L': return sizeof(long); + #ifdef HAVE_LONG_LONG + case 'q': case 'Q': return sizeof(PY_LONG_LONG); + #endif + case 'f': return sizeof(float) * (is_complex ? 2 : 1); + case 'd': return sizeof(double) * (is_complex ? 2 : 1); + case 'g': return sizeof(long double) * (is_complex ? 2 : 1); + case 'O': case 'P': return sizeof(void*); + default: { + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } + } +} +typedef struct { char c; short x; } __Pyx_st_short; +typedef struct { char c; int x; } __Pyx_st_int; +typedef struct { char c; long x; } __Pyx_st_long; +typedef struct { char c; float x; } __Pyx_st_float; +typedef struct { char c; double x; } __Pyx_st_double; +typedef struct { char c; long double x; } __Pyx_st_longdouble; +typedef struct { char c; void *x; } __Pyx_st_void_p; +#ifdef HAVE_LONG_LONG +typedef struct { char c; PY_LONG_LONG x; } __Pyx_st_longlong; +#endif +static size_t __Pyx_BufFmt_TypeCharToAlignment(char ch, int is_complex) { + CYTHON_UNUSED_VAR(is_complex); + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return sizeof(__Pyx_st_short) - sizeof(short); + case 'i': case 'I': return sizeof(__Pyx_st_int) - sizeof(int); + case 'l': case 'L': return sizeof(__Pyx_st_long) - sizeof(long); +#ifdef HAVE_LONG_LONG + case 'q': case 'Q': return sizeof(__Pyx_st_longlong) - sizeof(PY_LONG_LONG); +#endif + case 'f': return sizeof(__Pyx_st_float) - sizeof(float); + case 'd': return sizeof(__Pyx_st_double) - sizeof(double); + case 'g': return sizeof(__Pyx_st_longdouble) - sizeof(long double); + case 'P': case 'O': return sizeof(__Pyx_st_void_p) - sizeof(void*); + default: + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } +} +/* These are for computing the padding at the end of the struct to align + on the first member of the struct. This will probably the same as above, + but we don't have any guarantees. + */ +typedef struct { short x; char c; } __Pyx_pad_short; +typedef struct { int x; char c; } __Pyx_pad_int; +typedef struct { long x; char c; } __Pyx_pad_long; +typedef struct { float x; char c; } __Pyx_pad_float; +typedef struct { double x; char c; } __Pyx_pad_double; +typedef struct { long double x; char c; } __Pyx_pad_longdouble; +typedef struct { void *x; char c; } __Pyx_pad_void_p; +#ifdef HAVE_LONG_LONG +typedef struct { PY_LONG_LONG x; char c; } __Pyx_pad_longlong; +#endif +static size_t __Pyx_BufFmt_TypeCharToPadding(char ch, int is_complex) { + CYTHON_UNUSED_VAR(is_complex); + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return sizeof(__Pyx_pad_short) - sizeof(short); + case 'i': case 'I': return sizeof(__Pyx_pad_int) - sizeof(int); + case 'l': case 'L': return sizeof(__Pyx_pad_long) - sizeof(long); +#ifdef HAVE_LONG_LONG + case 'q': case 'Q': return sizeof(__Pyx_pad_longlong) - sizeof(PY_LONG_LONG); +#endif + case 'f': return sizeof(__Pyx_pad_float) - sizeof(float); + case 'd': return sizeof(__Pyx_pad_double) - sizeof(double); + case 'g': return sizeof(__Pyx_pad_longdouble) - sizeof(long double); + case 'P': case 'O': return sizeof(__Pyx_pad_void_p) - sizeof(void*); + default: + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } +} +static char __Pyx_BufFmt_TypeCharToGroup(char ch, int is_complex) { + switch (ch) { + case 'c': + return 'H'; + case 'b': case 'h': case 'i': + case 'l': case 'q': case 's': case 'p': + return 'I'; + case '?': case 'B': case 'H': case 'I': case 'L': case 'Q': + return 'U'; + case 'f': case 'd': case 'g': + return (is_complex ? 'C' : 'R'); + case 'O': + return 'O'; + case 'P': + return 'P'; + default: { + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } + } +} +static void __Pyx_BufFmt_RaiseExpected(__Pyx_BufFmt_Context* ctx) { + if (ctx->head == NULL || ctx->head->field == &ctx->root) { + const char* expected; + const char* quote; + if (ctx->head == NULL) { + expected = "end"; + quote = ""; + } else { + expected = ctx->head->field->type->name; + quote = "'"; + } + PyErr_Format(PyExc_ValueError, + "Buffer dtype mismatch, expected %s%s%s but got %s", + quote, expected, quote, + __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex)); + } else { + __Pyx_StructField* field = ctx->head->field; + __Pyx_StructField* parent = (ctx->head - 1)->field; + PyErr_Format(PyExc_ValueError, + "Buffer dtype mismatch, expected '%s' but got %s in '%s.%s'", + field->type->name, __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex), + parent->type->name, field->name); + } +} +static int __Pyx_BufFmt_ProcessTypeChunk(__Pyx_BufFmt_Context* ctx) { + char group; + size_t size, offset, arraysize = 1; + if (ctx->enc_type == 0) return 0; + if (ctx->head->field->type->arraysize[0]) { + int i, ndim = 0; + if (ctx->enc_type == 's' || ctx->enc_type == 'p') { + ctx->is_valid_array = ctx->head->field->type->ndim == 1; + ndim = 1; + if (ctx->enc_count != ctx->head->field->type->arraysize[0]) { + PyErr_Format(PyExc_ValueError, + "Expected a dimension of size %zu, got %zu", + ctx->head->field->type->arraysize[0], ctx->enc_count); + return -1; + } + } + if (!ctx->is_valid_array) { + PyErr_Format(PyExc_ValueError, "Expected %d dimensions, got %d", + ctx->head->field->type->ndim, ndim); + return -1; + } + for (i = 0; i < ctx->head->field->type->ndim; i++) { + arraysize *= ctx->head->field->type->arraysize[i]; + } + ctx->is_valid_array = 0; + ctx->enc_count = 1; + } + group = __Pyx_BufFmt_TypeCharToGroup(ctx->enc_type, ctx->is_complex); + do { + __Pyx_StructField* field = ctx->head->field; + __Pyx_TypeInfo* type = field->type; + if (ctx->enc_packmode == '@' || ctx->enc_packmode == '^') { + size = __Pyx_BufFmt_TypeCharToNativeSize(ctx->enc_type, ctx->is_complex); + } else { + size = __Pyx_BufFmt_TypeCharToStandardSize(ctx->enc_type, ctx->is_complex); + } + if (ctx->enc_packmode == '@') { + size_t align_at = __Pyx_BufFmt_TypeCharToAlignment(ctx->enc_type, ctx->is_complex); + size_t align_mod_offset; + if (align_at == 0) return -1; + align_mod_offset = ctx->fmt_offset % align_at; + if (align_mod_offset > 0) ctx->fmt_offset += align_at - align_mod_offset; + if (ctx->struct_alignment == 0) + ctx->struct_alignment = __Pyx_BufFmt_TypeCharToPadding(ctx->enc_type, + ctx->is_complex); + } + if (type->size != size || type->typegroup != group) { + if (type->typegroup == 'C' && type->fields != NULL) { + size_t parent_offset = ctx->head->parent_offset + field->offset; + ++ctx->head; + ctx->head->field = type->fields; + ctx->head->parent_offset = parent_offset; + continue; + } + if ((type->typegroup == 'H' || group == 'H') && type->size == size) { + } else { + __Pyx_BufFmt_RaiseExpected(ctx); + return -1; + } + } + offset = ctx->head->parent_offset + field->offset; + if (ctx->fmt_offset != offset) { + PyErr_Format(PyExc_ValueError, + "Buffer dtype mismatch; next field is at offset %" CYTHON_FORMAT_SSIZE_T "d but %" CYTHON_FORMAT_SSIZE_T "d expected", + (Py_ssize_t)ctx->fmt_offset, (Py_ssize_t)offset); + return -1; + } + ctx->fmt_offset += size; + if (arraysize) + ctx->fmt_offset += (arraysize - 1) * size; + --ctx->enc_count; + while (1) { + if (field == &ctx->root) { + ctx->head = NULL; + if (ctx->enc_count != 0) { + __Pyx_BufFmt_RaiseExpected(ctx); + return -1; + } + break; + } + ctx->head->field = ++field; + if (field->type == NULL) { + --ctx->head; + field = ctx->head->field; + continue; + } else if (field->type->typegroup == 'S') { + size_t parent_offset = ctx->head->parent_offset + field->offset; + if (field->type->fields->type == NULL) continue; + field = field->type->fields; + ++ctx->head; + ctx->head->field = field; + ctx->head->parent_offset = parent_offset; + break; + } else { + break; + } + } + } while (ctx->enc_count); + ctx->enc_type = 0; + ctx->is_complex = 0; + return 0; +} +static int +__pyx_buffmt_parse_array(__Pyx_BufFmt_Context* ctx, const char** tsp) +{ + const char *ts = *tsp; + int i = 0, number, ndim; + ++ts; + if (ctx->new_count != 1) { + PyErr_SetString(PyExc_ValueError, + "Cannot handle repeated arrays in format string"); + return -1; + } + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return -1; + ndim = ctx->head->field->type->ndim; + while (*ts && *ts != ')') { + switch (*ts) { + case ' ': case '\f': case '\r': case '\n': case '\t': case '\v': continue; + default: break; + } + number = __Pyx_BufFmt_ExpectNumber(&ts); + if (number == -1) return -1; + if (i < ndim && (size_t) number != ctx->head->field->type->arraysize[i]) { + PyErr_Format(PyExc_ValueError, + "Expected a dimension of size %zu, got %d", + ctx->head->field->type->arraysize[i], number); + return -1; + } + if (*ts != ',' && *ts != ')') { + PyErr_Format(PyExc_ValueError, + "Expected a comma in format string, got '%c'", *ts); + return -1; + } + if (*ts == ',') ts++; + i++; + } + if (i != ndim) { + PyErr_Format(PyExc_ValueError, "Expected %d dimension(s), got %d", + ctx->head->field->type->ndim, i); + return -1; + } + if (!*ts) { + PyErr_SetString(PyExc_ValueError, + "Unexpected end of format string, expected ')'"); + return -1; + } + ctx->is_valid_array = 1; + ctx->new_count = 1; + *tsp = ++ts; + return 0; +} +static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts) { + int got_Z = 0; + while (1) { + switch(*ts) { + case 0: + if (ctx->enc_type != 0 && ctx->head == NULL) { + __Pyx_BufFmt_RaiseExpected(ctx); + return NULL; + } + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + if (ctx->head != NULL) { + __Pyx_BufFmt_RaiseExpected(ctx); + return NULL; + } + return ts; + case ' ': + case '\r': + case '\n': + ++ts; + break; + case '<': + if (!__Pyx_Is_Little_Endian()) { + PyErr_SetString(PyExc_ValueError, "Little-endian buffer not supported on big-endian compiler"); + return NULL; + } + ctx->new_packmode = '='; + ++ts; + break; + case '>': + case '!': + if (__Pyx_Is_Little_Endian()) { + PyErr_SetString(PyExc_ValueError, "Big-endian buffer not supported on little-endian compiler"); + return NULL; + } + ctx->new_packmode = '='; + ++ts; + break; + case '=': + case '@': + case '^': + ctx->new_packmode = *ts++; + break; + case 'T': + { + const char* ts_after_sub; + size_t i, struct_count = ctx->new_count; + size_t struct_alignment = ctx->struct_alignment; + ctx->new_count = 1; + ++ts; + if (*ts != '{') { + PyErr_SetString(PyExc_ValueError, "Buffer acquisition: Expected '{' after 'T'"); + return NULL; + } + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->enc_type = 0; + ctx->enc_count = 0; + ctx->struct_alignment = 0; + ++ts; + ts_after_sub = ts; + for (i = 0; i != struct_count; ++i) { + ts_after_sub = __Pyx_BufFmt_CheckString(ctx, ts); + if (!ts_after_sub) return NULL; + } + ts = ts_after_sub; + if (struct_alignment) ctx->struct_alignment = struct_alignment; + } + break; + case '}': + { + size_t alignment = ctx->struct_alignment; + ++ts; + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->enc_type = 0; + if (alignment && ctx->fmt_offset % alignment) { + ctx->fmt_offset += alignment - (ctx->fmt_offset % alignment); + } + } + return ts; + case 'x': + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->fmt_offset += ctx->new_count; + ctx->new_count = 1; + ctx->enc_count = 0; + ctx->enc_type = 0; + ctx->enc_packmode = ctx->new_packmode; + ++ts; + break; + case 'Z': + got_Z = 1; + ++ts; + if (*ts != 'f' && *ts != 'd' && *ts != 'g') { + __Pyx_BufFmt_RaiseUnexpectedChar('Z'); + return NULL; + } + CYTHON_FALLTHROUGH; + case '?': case 'c': case 'b': case 'B': case 'h': case 'H': case 'i': case 'I': + case 'l': case 'L': case 'q': case 'Q': + case 'f': case 'd': case 'g': + case 'O': case 'p': + if ((ctx->enc_type == *ts) && (got_Z == ctx->is_complex) && + (ctx->enc_packmode == ctx->new_packmode) && (!ctx->is_valid_array)) { + ctx->enc_count += ctx->new_count; + ctx->new_count = 1; + got_Z = 0; + ++ts; + break; + } + CYTHON_FALLTHROUGH; + case 's': + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->enc_count = ctx->new_count; + ctx->enc_packmode = ctx->new_packmode; + ctx->enc_type = *ts; + ctx->is_complex = got_Z; + ++ts; + ctx->new_count = 1; + got_Z = 0; + break; + case ':': + ++ts; + while(*ts != ':') ++ts; + ++ts; + break; + case '(': + if (__pyx_buffmt_parse_array(ctx, &ts) < 0) return NULL; + break; + default: + { + int number = __Pyx_BufFmt_ExpectNumber(&ts); + if (number == -1) return NULL; + ctx->new_count = (size_t)number; + } + } + } +} + +/* BufferGetAndValidate */ + static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info) { + if (unlikely(info->buf == NULL)) return; + if (info->suboffsets == __Pyx_minusones) info->suboffsets = NULL; + __Pyx_ReleaseBuffer(info); +} +static void __Pyx_ZeroBuffer(Py_buffer* buf) { + buf->buf = NULL; + buf->obj = NULL; + buf->strides = __Pyx_zeros; + buf->shape = __Pyx_zeros; + buf->suboffsets = __Pyx_minusones; +} +static int __Pyx__GetBufferAndValidate( + Py_buffer* buf, PyObject* obj, __Pyx_TypeInfo* dtype, int flags, + int nd, int cast, __Pyx_BufFmt_StackElem* stack) +{ + buf->buf = NULL; + if (unlikely(__Pyx_GetBuffer(obj, buf, flags) == -1)) { + __Pyx_ZeroBuffer(buf); + return -1; + } + if (unlikely(buf->ndim != nd)) { + PyErr_Format(PyExc_ValueError, + "Buffer has wrong number of dimensions (expected %d, got %d)", + nd, buf->ndim); + goto fail; + } + if (!cast) { + __Pyx_BufFmt_Context ctx; + __Pyx_BufFmt_Init(&ctx, stack, dtype); + if (!__Pyx_BufFmt_CheckString(&ctx, buf->format)) goto fail; + } + if (unlikely((size_t)buf->itemsize != dtype->size)) { + PyErr_Format(PyExc_ValueError, + "Item size of buffer (%" CYTHON_FORMAT_SSIZE_T "d byte%s) does not match size of '%s' (%" CYTHON_FORMAT_SSIZE_T "d byte%s)", + buf->itemsize, (buf->itemsize > 1) ? "s" : "", + dtype->name, (Py_ssize_t)dtype->size, (dtype->size > 1) ? "s" : ""); + goto fail; + } + if (buf->suboffsets == NULL) buf->suboffsets = __Pyx_minusones; + return 0; +fail:; + __Pyx_SafeReleaseBuffer(buf); + return -1; +} + +/* BufferIndexError */ + static void __Pyx_RaiseBufferIndexError(int axis) { + PyErr_Format(PyExc_IndexError, + "Out of bounds on buffer access (axis %d)", axis); +} + +/* SliceTupleAndList */ + #if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE void __Pyx_crop_slice(Py_ssize_t* _start, Py_ssize_t* _stop, Py_ssize_t* _length) { + Py_ssize_t start = *_start, stop = *_stop, length = *_length; + if (start < 0) { + start += length; + if (start < 0) + start = 0; + } + if (stop < 0) + stop += length; + else if (stop > length) + stop = length; + *_length = stop - start; + *_start = start; + *_stop = stop; +} +static CYTHON_INLINE PyObject* __Pyx_PyList_GetSlice( + PyObject* src, Py_ssize_t start, Py_ssize_t stop) { + Py_ssize_t length = PyList_GET_SIZE(src); + __Pyx_crop_slice(&start, &stop, &length); + if (length <= 0) { + return PyList_New(0); + } + return __Pyx_PyList_FromArray(((PyListObject*)src)->ob_item + start, length); +} +static CYTHON_INLINE PyObject* __Pyx_PyTuple_GetSlice( + PyObject* src, Py_ssize_t start, Py_ssize_t stop) { + Py_ssize_t length = PyTuple_GET_SIZE(src); + __Pyx_crop_slice(&start, &stop, &length); + return __Pyx_PyTuple_FromArray(((PyTupleObject*)src)->ob_item + start, length); +} +#endif + +/* PyIntCompare */ + static CYTHON_INLINE int __Pyx_PyInt_BoolEqObjC(PyObject *op1, PyObject *op2, long intval, long inplace) { + CYTHON_MAYBE_UNUSED_VAR(intval); + CYTHON_UNUSED_VAR(inplace); + if (op1 == op2) { + return 1; + } + #if PY_MAJOR_VERSION < 3 + if (likely(PyInt_CheckExact(op1))) { + const long b = intval; + long a = PyInt_AS_LONG(op1); + return (a == b); + } + #endif + #if CYTHON_USE_PYLONG_INTERNALS + if (likely(PyLong_CheckExact(op1))) { + int unequal; + unsigned long uintval; + Py_ssize_t size = __Pyx_PyLong_DigitCount(op1); + const digit* digits = __Pyx_PyLong_Digits(op1); + if (intval == 0) { + return (__Pyx_PyLong_IsZero(op1) == 1); + } else if (intval < 0) { + if (__Pyx_PyLong_IsNonNeg(op1)) + return 0; + intval = -intval; + } else { + if (__Pyx_PyLong_IsNeg(op1)) + return 0; + } + uintval = (unsigned long) intval; +#if PyLong_SHIFT * 4 < SIZEOF_LONG*8 + if (uintval >> (PyLong_SHIFT * 4)) { + unequal = (size != 5) || (digits[0] != (uintval & (unsigned long) PyLong_MASK)) + | (digits[1] != ((uintval >> (1 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)) | (digits[2] != ((uintval >> (2 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)) | (digits[3] != ((uintval >> (3 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)) | (digits[4] != ((uintval >> (4 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)); + } else +#endif +#if PyLong_SHIFT * 3 < SIZEOF_LONG*8 + if (uintval >> (PyLong_SHIFT * 3)) { + unequal = (size != 4) || (digits[0] != (uintval & (unsigned long) PyLong_MASK)) + | (digits[1] != ((uintval >> (1 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)) | (digits[2] != ((uintval >> (2 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)) | (digits[3] != ((uintval >> (3 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)); + } else +#endif +#if PyLong_SHIFT * 2 < SIZEOF_LONG*8 + if (uintval >> (PyLong_SHIFT * 2)) { + unequal = (size != 3) || (digits[0] != (uintval & (unsigned long) PyLong_MASK)) + | (digits[1] != ((uintval >> (1 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)) | (digits[2] != ((uintval >> (2 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)); + } else +#endif +#if PyLong_SHIFT * 1 < SIZEOF_LONG*8 + if (uintval >> (PyLong_SHIFT * 1)) { + unequal = (size != 2) || (digits[0] != (uintval & (unsigned long) PyLong_MASK)) + | (digits[1] != ((uintval >> (1 * PyLong_SHIFT)) & (unsigned long) PyLong_MASK)); + } else +#endif + unequal = (size != 1) || (((unsigned long) digits[0]) != (uintval & (unsigned long) PyLong_MASK)); + return (unequal == 0); + } + #endif + if (PyFloat_CheckExact(op1)) { + const long b = intval; +#if CYTHON_COMPILING_IN_LIMITED_API + double a = __pyx_PyFloat_AsDouble(op1); +#else + double a = PyFloat_AS_DOUBLE(op1); +#endif + return ((double)a == (double)b); + } + return __Pyx_PyObject_IsTrueAndDecref( + PyObject_RichCompare(op1, op2, Py_EQ)); +} + +/* PyObject_GenericGetAttrNoDict */ + #if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static PyObject *__Pyx_RaiseGenericGetAttributeError(PyTypeObject *tp, PyObject *attr_name) { + __Pyx_TypeName type_name = __Pyx_PyType_GetName(tp); + PyErr_Format(PyExc_AttributeError, +#if PY_MAJOR_VERSION >= 3 + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%U'", + type_name, attr_name); +#else + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%.400s'", + type_name, PyString_AS_STRING(attr_name)); +#endif + __Pyx_DECREF_TypeName(type_name); + return NULL; +} +static CYTHON_INLINE PyObject* __Pyx_PyObject_GenericGetAttrNoDict(PyObject* obj, PyObject* attr_name) { + PyObject *descr; + PyTypeObject *tp = Py_TYPE(obj); + if (unlikely(!PyString_Check(attr_name))) { + return PyObject_GenericGetAttr(obj, attr_name); + } + assert(!tp->tp_dictoffset); + descr = _PyType_Lookup(tp, attr_name); + if (unlikely(!descr)) { + return __Pyx_RaiseGenericGetAttributeError(tp, attr_name); + } + Py_INCREF(descr); + #if PY_MAJOR_VERSION < 3 + if (likely(PyType_HasFeature(Py_TYPE(descr), Py_TPFLAGS_HAVE_CLASS))) + #endif + { + descrgetfunc f = Py_TYPE(descr)->tp_descr_get; + if (unlikely(f)) { + PyObject *res = f(descr, obj, (PyObject *)tp); + Py_DECREF(descr); + return res; + } + } + return descr; +} +#endif + +/* PyObject_GenericGetAttr */ + #if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static PyObject* __Pyx_PyObject_GenericGetAttr(PyObject* obj, PyObject* attr_name) { + if (unlikely(Py_TYPE(obj)->tp_dictoffset)) { + return PyObject_GenericGetAttr(obj, attr_name); + } + return __Pyx_PyObject_GenericGetAttrNoDict(obj, attr_name); +} +#endif + +/* FixUpExtensionType */ + #if CYTHON_USE_TYPE_SPECS +static int __Pyx_fix_up_extension_type_from_spec(PyType_Spec *spec, PyTypeObject *type) { +#if PY_VERSION_HEX > 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + CYTHON_UNUSED_VAR(spec); + CYTHON_UNUSED_VAR(type); +#else + const PyType_Slot *slot = spec->slots; + while (slot && slot->slot && slot->slot != Py_tp_members) + slot++; + if (slot && slot->slot == Py_tp_members) { + int changed = 0; +#if !(PY_VERSION_HEX <= 0x030900b1 && CYTHON_COMPILING_IN_CPYTHON) + const +#endif + PyMemberDef *memb = (PyMemberDef*) slot->pfunc; + while (memb && memb->name) { + if (memb->name[0] == '_' && memb->name[1] == '_') { +#if PY_VERSION_HEX < 0x030900b1 + if (strcmp(memb->name, "__weaklistoffset__") == 0) { + assert(memb->type == T_PYSSIZET); + assert(memb->flags == READONLY); + type->tp_weaklistoffset = memb->offset; + changed = 1; + } + else if (strcmp(memb->name, "__dictoffset__") == 0) { + assert(memb->type == T_PYSSIZET); + assert(memb->flags == READONLY); + type->tp_dictoffset = memb->offset; + changed = 1; + } +#if CYTHON_METH_FASTCALL + else if (strcmp(memb->name, "__vectorcalloffset__") == 0) { + assert(memb->type == T_PYSSIZET); + assert(memb->flags == READONLY); +#if PY_VERSION_HEX >= 0x030800b4 + type->tp_vectorcall_offset = memb->offset; +#else + type->tp_print = (printfunc) memb->offset; +#endif + changed = 1; + } +#endif +#else + if ((0)); +#endif +#if PY_VERSION_HEX <= 0x030900b1 && CYTHON_COMPILING_IN_CPYTHON + else if (strcmp(memb->name, "__module__") == 0) { + PyObject *descr; + assert(memb->type == T_OBJECT); + assert(memb->flags == 0 || memb->flags == READONLY); + descr = PyDescr_NewMember(type, memb); + if (unlikely(!descr)) + return -1; + if (unlikely(PyDict_SetItem(type->tp_dict, PyDescr_NAME(descr), descr) < 0)) { + Py_DECREF(descr); + return -1; + } + Py_DECREF(descr); + changed = 1; + } +#endif + } + memb++; + } + if (changed) + PyType_Modified(type); + } +#endif + return 0; +} +#endif + +/* PyObjectCallNoArg */ + static CYTHON_INLINE PyObject* __Pyx_PyObject_CallNoArg(PyObject *func) { + PyObject *arg[2] = {NULL, NULL}; + return __Pyx_PyObject_FastCall(func, arg + 1, 0 | __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET); +} + +/* PyObjectGetMethod */ + static int __Pyx_PyObject_GetMethod(PyObject *obj, PyObject *name, PyObject **method) { + PyObject *attr; +#if CYTHON_UNPACK_METHODS && CYTHON_COMPILING_IN_CPYTHON && CYTHON_USE_PYTYPE_LOOKUP + __Pyx_TypeName type_name; + PyTypeObject *tp = Py_TYPE(obj); + PyObject *descr; + descrgetfunc f = NULL; + PyObject **dictptr, *dict; + int meth_found = 0; + assert (*method == NULL); + if (unlikely(tp->tp_getattro != PyObject_GenericGetAttr)) { + attr = __Pyx_PyObject_GetAttrStr(obj, name); + goto try_unpack; + } + if (unlikely(tp->tp_dict == NULL) && unlikely(PyType_Ready(tp) < 0)) { + return 0; + } + descr = _PyType_Lookup(tp, name); + if (likely(descr != NULL)) { + Py_INCREF(descr); +#if defined(Py_TPFLAGS_METHOD_DESCRIPTOR) && Py_TPFLAGS_METHOD_DESCRIPTOR + if (__Pyx_PyType_HasFeature(Py_TYPE(descr), Py_TPFLAGS_METHOD_DESCRIPTOR)) +#elif PY_MAJOR_VERSION >= 3 + #ifdef __Pyx_CyFunction_USED + if (likely(PyFunction_Check(descr) || __Pyx_IS_TYPE(descr, &PyMethodDescr_Type) || __Pyx_CyFunction_Check(descr))) + #else + if (likely(PyFunction_Check(descr) || __Pyx_IS_TYPE(descr, &PyMethodDescr_Type))) + #endif +#else + #ifdef __Pyx_CyFunction_USED + if (likely(PyFunction_Check(descr) || __Pyx_CyFunction_Check(descr))) + #else + if (likely(PyFunction_Check(descr))) + #endif +#endif + { + meth_found = 1; + } else { + f = Py_TYPE(descr)->tp_descr_get; + if (f != NULL && PyDescr_IsData(descr)) { + attr = f(descr, obj, (PyObject *)Py_TYPE(obj)); + Py_DECREF(descr); + goto try_unpack; + } + } + } + dictptr = _PyObject_GetDictPtr(obj); + if (dictptr != NULL && (dict = *dictptr) != NULL) { + Py_INCREF(dict); + attr = __Pyx_PyDict_GetItemStr(dict, name); + if (attr != NULL) { + Py_INCREF(attr); + Py_DECREF(dict); + Py_XDECREF(descr); + goto try_unpack; + } + Py_DECREF(dict); + } + if (meth_found) { + *method = descr; + return 1; + } + if (f != NULL) { + attr = f(descr, obj, (PyObject *)Py_TYPE(obj)); + Py_DECREF(descr); + goto try_unpack; + } + if (likely(descr != NULL)) { + *method = descr; + return 0; + } + type_name = __Pyx_PyType_GetName(tp); + PyErr_Format(PyExc_AttributeError, +#if PY_MAJOR_VERSION >= 3 + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%U'", + type_name, name); +#else + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%.400s'", + type_name, PyString_AS_STRING(name)); +#endif + __Pyx_DECREF_TypeName(type_name); + return 0; +#else + attr = __Pyx_PyObject_GetAttrStr(obj, name); + goto try_unpack; +#endif +try_unpack: +#if CYTHON_UNPACK_METHODS + if (likely(attr) && PyMethod_Check(attr) && likely(PyMethod_GET_SELF(attr) == obj)) { + PyObject *function = PyMethod_GET_FUNCTION(attr); + Py_INCREF(function); + Py_DECREF(attr); + *method = function; + return 1; + } +#endif + *method = attr; + return 0; +} + +/* PyObjectCallMethod0 */ + static PyObject* __Pyx_PyObject_CallMethod0(PyObject* obj, PyObject* method_name) { + PyObject *method = NULL, *result = NULL; + int is_method = __Pyx_PyObject_GetMethod(obj, method_name, &method); + if (likely(is_method)) { + result = __Pyx_PyObject_CallOneArg(method, obj); + Py_DECREF(method); + return result; + } + if (unlikely(!method)) goto bad; + result = __Pyx_PyObject_CallNoArg(method); + Py_DECREF(method); +bad: + return result; +} + +/* ValidateBasesTuple */ + #if CYTHON_COMPILING_IN_CPYTHON || CYTHON_COMPILING_IN_LIMITED_API || CYTHON_USE_TYPE_SPECS +static int __Pyx_validate_bases_tuple(const char *type_name, Py_ssize_t dictoffset, PyObject *bases) { + Py_ssize_t i, n; +#if CYTHON_ASSUME_SAFE_MACROS + n = PyTuple_GET_SIZE(bases); +#else + n = PyTuple_Size(bases); + if (n < 0) return -1; +#endif + for (i = 1; i < n; i++) + { +#if CYTHON_AVOID_BORROWED_REFS + PyObject *b0 = PySequence_GetItem(bases, i); + if (!b0) return -1; +#elif CYTHON_ASSUME_SAFE_MACROS + PyObject *b0 = PyTuple_GET_ITEM(bases, i); +#else + PyObject *b0 = PyTuple_GetItem(bases, i); + if (!b0) return -1; +#endif + PyTypeObject *b; +#if PY_MAJOR_VERSION < 3 + if (PyClass_Check(b0)) + { + PyErr_Format(PyExc_TypeError, "base class '%.200s' is an old-style class", + PyString_AS_STRING(((PyClassObject*)b0)->cl_name)); +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + return -1; + } +#endif + b = (PyTypeObject*) b0; + if (!__Pyx_PyType_HasFeature(b, Py_TPFLAGS_HEAPTYPE)) + { + __Pyx_TypeName b_name = __Pyx_PyType_GetName(b); + PyErr_Format(PyExc_TypeError, + "base class '" __Pyx_FMT_TYPENAME "' is not a heap type", b_name); + __Pyx_DECREF_TypeName(b_name); +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + return -1; + } + if (dictoffset == 0) + { + Py_ssize_t b_dictoffset = 0; +#if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + b_dictoffset = b->tp_dictoffset; +#else + PyObject *py_b_dictoffset = PyObject_GetAttrString((PyObject*)b, "__dictoffset__"); + if (!py_b_dictoffset) goto dictoffset_return; + b_dictoffset = PyLong_AsSsize_t(py_b_dictoffset); + Py_DECREF(py_b_dictoffset); + if (b_dictoffset == -1 && PyErr_Occurred()) goto dictoffset_return; +#endif + if (b_dictoffset) { + { + __Pyx_TypeName b_name = __Pyx_PyType_GetName(b); + PyErr_Format(PyExc_TypeError, + "extension type '%.200s' has no __dict__ slot, " + "but base type '" __Pyx_FMT_TYPENAME "' has: " + "either add 'cdef dict __dict__' to the extension type " + "or add '__slots__ = [...]' to the base type", + type_name, b_name); + __Pyx_DECREF_TypeName(b_name); + } +#if !(CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY) + dictoffset_return: +#endif +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + return -1; + } + } +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + } + return 0; +} +#endif + +/* PyType_Ready */ + static int __Pyx_PyType_Ready(PyTypeObject *t) { +#if CYTHON_USE_TYPE_SPECS || !(CYTHON_COMPILING_IN_CPYTHON || CYTHON_COMPILING_IN_LIMITED_API) || defined(PYSTON_MAJOR_VERSION) + (void)__Pyx_PyObject_CallMethod0; +#if CYTHON_USE_TYPE_SPECS + (void)__Pyx_validate_bases_tuple; +#endif + return PyType_Ready(t); +#else + int r; + PyObject *bases = __Pyx_PyType_GetSlot(t, tp_bases, PyObject*); + if (bases && unlikely(__Pyx_validate_bases_tuple(t->tp_name, t->tp_dictoffset, bases) == -1)) + return -1; +#if PY_VERSION_HEX >= 0x03050000 && !defined(PYSTON_MAJOR_VERSION) + { + int gc_was_enabled; + #if PY_VERSION_HEX >= 0x030A00b1 + gc_was_enabled = PyGC_Disable(); + (void)__Pyx_PyObject_CallMethod0; + #else + PyObject *ret, *py_status; + PyObject *gc = NULL; + #if PY_VERSION_HEX >= 0x030700a1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM+0 >= 0x07030400) + gc = PyImport_GetModule(__pyx_kp_u_gc); + #endif + if (unlikely(!gc)) gc = PyImport_Import(__pyx_kp_u_gc); + if (unlikely(!gc)) return -1; + py_status = __Pyx_PyObject_CallMethod0(gc, __pyx_kp_u_isenabled); + if (unlikely(!py_status)) { + Py_DECREF(gc); + return -1; + } + gc_was_enabled = __Pyx_PyObject_IsTrue(py_status); + Py_DECREF(py_status); + if (gc_was_enabled > 0) { + ret = __Pyx_PyObject_CallMethod0(gc, __pyx_kp_u_disable); + if (unlikely(!ret)) { + Py_DECREF(gc); + return -1; + } + Py_DECREF(ret); + } else if (unlikely(gc_was_enabled == -1)) { + Py_DECREF(gc); + return -1; + } + #endif + t->tp_flags |= Py_TPFLAGS_HEAPTYPE; +#if PY_VERSION_HEX >= 0x030A0000 + t->tp_flags |= Py_TPFLAGS_IMMUTABLETYPE; +#endif +#else + (void)__Pyx_PyObject_CallMethod0; +#endif + r = PyType_Ready(t); +#if PY_VERSION_HEX >= 0x03050000 && !defined(PYSTON_MAJOR_VERSION) + t->tp_flags &= ~Py_TPFLAGS_HEAPTYPE; + #if PY_VERSION_HEX >= 0x030A00b1 + if (gc_was_enabled) + PyGC_Enable(); + #else + if (gc_was_enabled) { + PyObject *tp, *v, *tb; + PyErr_Fetch(&tp, &v, &tb); + ret = __Pyx_PyObject_CallMethod0(gc, __pyx_kp_u_enable); + if (likely(ret || r == -1)) { + Py_XDECREF(ret); + PyErr_Restore(tp, v, tb); + } else { + Py_XDECREF(tp); + Py_XDECREF(v); + Py_XDECREF(tb); + r = -1; + } + } + Py_DECREF(gc); + #endif + } +#endif + return r; +#endif +} + +/* SetVTable */ + static int __Pyx_SetVtable(PyTypeObject *type, void *vtable) { + PyObject *ob = PyCapsule_New(vtable, 0, 0); + if (unlikely(!ob)) + goto bad; +#if CYTHON_COMPILING_IN_LIMITED_API + if (unlikely(PyObject_SetAttr((PyObject *) type, __pyx_n_s_pyx_vtable, ob) < 0)) +#else + if (unlikely(PyDict_SetItem(type->tp_dict, __pyx_n_s_pyx_vtable, ob) < 0)) +#endif + goto bad; + Py_DECREF(ob); + return 0; +bad: + Py_XDECREF(ob); + return -1; +} + +/* GetVTable */ + static void* __Pyx_GetVtable(PyTypeObject *type) { + void* ptr; +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *ob = PyObject_GetAttr((PyObject *)type, __pyx_n_s_pyx_vtable); +#else + PyObject *ob = PyObject_GetItem(type->tp_dict, __pyx_n_s_pyx_vtable); +#endif + if (!ob) + goto bad; + ptr = PyCapsule_GetPointer(ob, 0); + if (!ptr && !PyErr_Occurred()) + PyErr_SetString(PyExc_RuntimeError, "invalid vtable found for imported type"); + Py_DECREF(ob); + return ptr; +bad: + Py_XDECREF(ob); + return NULL; +} + +/* MergeVTables */ + #if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_MergeVtables(PyTypeObject *type) { + int i; + void** base_vtables; + __Pyx_TypeName tp_base_name; + __Pyx_TypeName base_name; + void* unknown = (void*)-1; + PyObject* bases = type->tp_bases; + int base_depth = 0; + { + PyTypeObject* base = type->tp_base; + while (base) { + base_depth += 1; + base = base->tp_base; + } + } + base_vtables = (void**) malloc(sizeof(void*) * (size_t)(base_depth + 1)); + base_vtables[0] = unknown; + for (i = 1; i < PyTuple_GET_SIZE(bases); i++) { + void* base_vtable = __Pyx_GetVtable(((PyTypeObject*)PyTuple_GET_ITEM(bases, i))); + if (base_vtable != NULL) { + int j; + PyTypeObject* base = type->tp_base; + for (j = 0; j < base_depth; j++) { + if (base_vtables[j] == unknown) { + base_vtables[j] = __Pyx_GetVtable(base); + base_vtables[j + 1] = unknown; + } + if (base_vtables[j] == base_vtable) { + break; + } else if (base_vtables[j] == NULL) { + goto bad; + } + base = base->tp_base; + } + } + } + PyErr_Clear(); + free(base_vtables); + return 0; +bad: + tp_base_name = __Pyx_PyType_GetName(type->tp_base); + base_name = __Pyx_PyType_GetName((PyTypeObject*)PyTuple_GET_ITEM(bases, i)); + PyErr_Format(PyExc_TypeError, + "multiple bases have vtable conflict: '" __Pyx_FMT_TYPENAME "' and '" __Pyx_FMT_TYPENAME "'", tp_base_name, base_name); + __Pyx_DECREF_TypeName(tp_base_name); + __Pyx_DECREF_TypeName(base_name); + free(base_vtables); + return -1; +} +#endif + +/* SetupReduce */ + #if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_setup_reduce_is_named(PyObject* meth, PyObject* name) { + int ret; + PyObject *name_attr; + name_attr = __Pyx_PyObject_GetAttrStrNoError(meth, __pyx_n_s_name_2); + if (likely(name_attr)) { + ret = PyObject_RichCompareBool(name_attr, name, Py_EQ); + } else { + ret = -1; + } + if (unlikely(ret < 0)) { + PyErr_Clear(); + ret = 0; + } + Py_XDECREF(name_attr); + return ret; +} +static int __Pyx_setup_reduce(PyObject* type_obj) { + int ret = 0; + PyObject *object_reduce = NULL; + PyObject *object_getstate = NULL; + PyObject *object_reduce_ex = NULL; + PyObject *reduce = NULL; + PyObject *reduce_ex = NULL; + PyObject *reduce_cython = NULL; + PyObject *setstate = NULL; + PyObject *setstate_cython = NULL; + PyObject *getstate = NULL; +#if CYTHON_USE_PYTYPE_LOOKUP + getstate = _PyType_Lookup((PyTypeObject*)type_obj, __pyx_n_s_getstate); +#else + getstate = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_getstate); + if (!getstate && PyErr_Occurred()) { + goto __PYX_BAD; + } +#endif + if (getstate) { +#if CYTHON_USE_PYTYPE_LOOKUP + object_getstate = _PyType_Lookup(&PyBaseObject_Type, __pyx_n_s_getstate); +#else + object_getstate = __Pyx_PyObject_GetAttrStrNoError((PyObject*)&PyBaseObject_Type, __pyx_n_s_getstate); + if (!object_getstate && PyErr_Occurred()) { + goto __PYX_BAD; + } +#endif + if (object_getstate != getstate) { + goto __PYX_GOOD; + } + } +#if CYTHON_USE_PYTYPE_LOOKUP + object_reduce_ex = _PyType_Lookup(&PyBaseObject_Type, __pyx_n_s_reduce_ex); if (!object_reduce_ex) goto __PYX_BAD; +#else + object_reduce_ex = __Pyx_PyObject_GetAttrStr((PyObject*)&PyBaseObject_Type, __pyx_n_s_reduce_ex); if (!object_reduce_ex) goto __PYX_BAD; +#endif + reduce_ex = __Pyx_PyObject_GetAttrStr(type_obj, __pyx_n_s_reduce_ex); if (unlikely(!reduce_ex)) goto __PYX_BAD; + if (reduce_ex == object_reduce_ex) { +#if CYTHON_USE_PYTYPE_LOOKUP + object_reduce = _PyType_Lookup(&PyBaseObject_Type, __pyx_n_s_reduce); if (!object_reduce) goto __PYX_BAD; +#else + object_reduce = __Pyx_PyObject_GetAttrStr((PyObject*)&PyBaseObject_Type, __pyx_n_s_reduce); if (!object_reduce) goto __PYX_BAD; +#endif + reduce = __Pyx_PyObject_GetAttrStr(type_obj, __pyx_n_s_reduce); if (unlikely(!reduce)) goto __PYX_BAD; + if (reduce == object_reduce || __Pyx_setup_reduce_is_named(reduce, __pyx_n_s_reduce_cython)) { + reduce_cython = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_reduce_cython); + if (likely(reduce_cython)) { + ret = PyDict_SetItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_reduce, reduce_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + ret = PyDict_DelItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_reduce_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + } else if (reduce == object_reduce || PyErr_Occurred()) { + goto __PYX_BAD; + } + setstate = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_setstate); + if (!setstate) PyErr_Clear(); + if (!setstate || __Pyx_setup_reduce_is_named(setstate, __pyx_n_s_setstate_cython)) { + setstate_cython = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_setstate_cython); + if (likely(setstate_cython)) { + ret = PyDict_SetItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_setstate, setstate_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + ret = PyDict_DelItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_setstate_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + } else if (!setstate || PyErr_Occurred()) { + goto __PYX_BAD; + } + } + PyType_Modified((PyTypeObject*)type_obj); + } + } + goto __PYX_GOOD; +__PYX_BAD: + if (!PyErr_Occurred()) { + __Pyx_TypeName type_obj_name = + __Pyx_PyType_GetName((PyTypeObject*)type_obj); + PyErr_Format(PyExc_RuntimeError, + "Unable to initialize pickling for " __Pyx_FMT_TYPENAME, type_obj_name); + __Pyx_DECREF_TypeName(type_obj_name); + } + ret = -1; +__PYX_GOOD: +#if !CYTHON_USE_PYTYPE_LOOKUP + Py_XDECREF(object_reduce); + Py_XDECREF(object_reduce_ex); + Py_XDECREF(object_getstate); + Py_XDECREF(getstate); +#endif + Py_XDECREF(reduce); + Py_XDECREF(reduce_ex); + Py_XDECREF(reduce_cython); + Py_XDECREF(setstate); + Py_XDECREF(setstate_cython); + return ret; +} +#endif + +/* TypeImport */ + #ifndef __PYX_HAVE_RT_ImportType_3_0_12 +#define __PYX_HAVE_RT_ImportType_3_0_12 +static PyTypeObject *__Pyx_ImportType_3_0_12(PyObject *module, const char *module_name, const char *class_name, + size_t size, size_t alignment, enum __Pyx_ImportType_CheckSize_3_0_12 check_size) +{ + PyObject *result = 0; + char warning[200]; + Py_ssize_t basicsize; + Py_ssize_t itemsize; +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *py_basicsize; + PyObject *py_itemsize; +#endif + result = PyObject_GetAttrString(module, class_name); + if (!result) + goto bad; + if (!PyType_Check(result)) { + PyErr_Format(PyExc_TypeError, + "%.200s.%.200s is not a type object", + module_name, class_name); + goto bad; + } +#if !CYTHON_COMPILING_IN_LIMITED_API + basicsize = ((PyTypeObject *)result)->tp_basicsize; + itemsize = ((PyTypeObject *)result)->tp_itemsize; +#else + py_basicsize = PyObject_GetAttrString(result, "__basicsize__"); + if (!py_basicsize) + goto bad; + basicsize = PyLong_AsSsize_t(py_basicsize); + Py_DECREF(py_basicsize); + py_basicsize = 0; + if (basicsize == (Py_ssize_t)-1 && PyErr_Occurred()) + goto bad; + py_itemsize = PyObject_GetAttrString(result, "__itemsize__"); + if (!py_itemsize) + goto bad; + itemsize = PyLong_AsSsize_t(py_itemsize); + Py_DECREF(py_itemsize); + py_itemsize = 0; + if (itemsize == (Py_ssize_t)-1 && PyErr_Occurred()) + goto bad; +#endif + if (itemsize) { + if (size % alignment) { + alignment = size % alignment; + } + if (itemsize < (Py_ssize_t)alignment) + itemsize = (Py_ssize_t)alignment; + } + if ((size_t)(basicsize + itemsize) < size) { + PyErr_Format(PyExc_ValueError, + "%.200s.%.200s size changed, may indicate binary incompatibility. " + "Expected %zd from C header, got %zd from PyObject", + module_name, class_name, size, basicsize+itemsize); + goto bad; + } + if (check_size == __Pyx_ImportType_CheckSize_Error_3_0_12 && + ((size_t)basicsize > size || (size_t)(basicsize + itemsize) < size)) { + PyErr_Format(PyExc_ValueError, + "%.200s.%.200s size changed, may indicate binary incompatibility. " + "Expected %zd from C header, got %zd-%zd from PyObject", + module_name, class_name, size, basicsize, basicsize+itemsize); + goto bad; + } + else if (check_size == __Pyx_ImportType_CheckSize_Warn_3_0_12 && (size_t)basicsize > size) { + PyOS_snprintf(warning, sizeof(warning), + "%s.%s size changed, may indicate binary incompatibility. " + "Expected %zd from C header, got %zd from PyObject", + module_name, class_name, size, basicsize); + if (PyErr_WarnEx(NULL, warning, 0) < 0) goto bad; + } + return (PyTypeObject *)result; +bad: + Py_XDECREF(result); + return NULL; +} +#endif + +/* FetchSharedCythonModule */ + static PyObject *__Pyx_FetchSharedCythonABIModule(void) { + return __Pyx_PyImport_AddModuleRef((char*) __PYX_ABI_MODULE_NAME); +} + +/* FetchCommonType */ + static int __Pyx_VerifyCachedType(PyObject *cached_type, + const char *name, + Py_ssize_t basicsize, + Py_ssize_t expected_basicsize) { + if (!PyType_Check(cached_type)) { + PyErr_Format(PyExc_TypeError, + "Shared Cython type %.200s is not a type object", name); + return -1; + } + if (basicsize != expected_basicsize) { + PyErr_Format(PyExc_TypeError, + "Shared Cython type %.200s has the wrong size, try recompiling", + name); + return -1; + } + return 0; +} +#if !CYTHON_USE_TYPE_SPECS +static PyTypeObject* __Pyx_FetchCommonType(PyTypeObject* type) { + PyObject* abi_module; + const char* object_name; + PyTypeObject *cached_type = NULL; + abi_module = __Pyx_FetchSharedCythonABIModule(); + if (!abi_module) return NULL; + object_name = strrchr(type->tp_name, '.'); + object_name = object_name ? object_name+1 : type->tp_name; + cached_type = (PyTypeObject*) PyObject_GetAttrString(abi_module, object_name); + if (cached_type) { + if (__Pyx_VerifyCachedType( + (PyObject *)cached_type, + object_name, + cached_type->tp_basicsize, + type->tp_basicsize) < 0) { + goto bad; + } + goto done; + } + if (!PyErr_ExceptionMatches(PyExc_AttributeError)) goto bad; + PyErr_Clear(); + if (PyType_Ready(type) < 0) goto bad; + if (PyObject_SetAttrString(abi_module, object_name, (PyObject *)type) < 0) + goto bad; + Py_INCREF(type); + cached_type = type; +done: + Py_DECREF(abi_module); + return cached_type; +bad: + Py_XDECREF(cached_type); + cached_type = NULL; + goto done; +} +#else +static PyTypeObject *__Pyx_FetchCommonTypeFromSpec(PyObject *module, PyType_Spec *spec, PyObject *bases) { + PyObject *abi_module, *cached_type = NULL; + const char* object_name = strrchr(spec->name, '.'); + object_name = object_name ? object_name+1 : spec->name; + abi_module = __Pyx_FetchSharedCythonABIModule(); + if (!abi_module) return NULL; + cached_type = PyObject_GetAttrString(abi_module, object_name); + if (cached_type) { + Py_ssize_t basicsize; +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *py_basicsize; + py_basicsize = PyObject_GetAttrString(cached_type, "__basicsize__"); + if (unlikely(!py_basicsize)) goto bad; + basicsize = PyLong_AsSsize_t(py_basicsize); + Py_DECREF(py_basicsize); + py_basicsize = 0; + if (unlikely(basicsize == (Py_ssize_t)-1) && PyErr_Occurred()) goto bad; +#else + basicsize = likely(PyType_Check(cached_type)) ? ((PyTypeObject*) cached_type)->tp_basicsize : -1; +#endif + if (__Pyx_VerifyCachedType( + cached_type, + object_name, + basicsize, + spec->basicsize) < 0) { + goto bad; + } + goto done; + } + if (!PyErr_ExceptionMatches(PyExc_AttributeError)) goto bad; + PyErr_Clear(); + CYTHON_UNUSED_VAR(module); + cached_type = __Pyx_PyType_FromModuleAndSpec(abi_module, spec, bases); + if (unlikely(!cached_type)) goto bad; + if (unlikely(__Pyx_fix_up_extension_type_from_spec(spec, (PyTypeObject *) cached_type) < 0)) goto bad; + if (PyObject_SetAttrString(abi_module, object_name, cached_type) < 0) goto bad; +done: + Py_DECREF(abi_module); + assert(cached_type == NULL || PyType_Check(cached_type)); + return (PyTypeObject *) cached_type; +bad: + Py_XDECREF(cached_type); + cached_type = NULL; + goto done; +} +#endif + +/* PyVectorcallFastCallDict */ + #if CYTHON_METH_FASTCALL +static PyObject *__Pyx_PyVectorcall_FastCallDict_kw(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw) +{ + PyObject *res = NULL; + PyObject *kwnames; + PyObject **newargs; + PyObject **kwvalues; + Py_ssize_t i, pos; + size_t j; + PyObject *key, *value; + unsigned long keys_are_strings; + Py_ssize_t nkw = PyDict_GET_SIZE(kw); + newargs = (PyObject **)PyMem_Malloc((nargs + (size_t)nkw) * sizeof(args[0])); + if (unlikely(newargs == NULL)) { + PyErr_NoMemory(); + return NULL; + } + for (j = 0; j < nargs; j++) newargs[j] = args[j]; + kwnames = PyTuple_New(nkw); + if (unlikely(kwnames == NULL)) { + PyMem_Free(newargs); + return NULL; + } + kwvalues = newargs + nargs; + pos = i = 0; + keys_are_strings = Py_TPFLAGS_UNICODE_SUBCLASS; + while (PyDict_Next(kw, &pos, &key, &value)) { + keys_are_strings &= Py_TYPE(key)->tp_flags; + Py_INCREF(key); + Py_INCREF(value); + PyTuple_SET_ITEM(kwnames, i, key); + kwvalues[i] = value; + i++; + } + if (unlikely(!keys_are_strings)) { + PyErr_SetString(PyExc_TypeError, "keywords must be strings"); + goto cleanup; + } + res = vc(func, newargs, nargs, kwnames); +cleanup: + Py_DECREF(kwnames); + for (i = 0; i < nkw; i++) + Py_DECREF(kwvalues[i]); + PyMem_Free(newargs); + return res; +} +static CYTHON_INLINE PyObject *__Pyx_PyVectorcall_FastCallDict(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw) +{ + if (likely(kw == NULL) || PyDict_GET_SIZE(kw) == 0) { + return vc(func, args, nargs, NULL); + } + return __Pyx_PyVectorcall_FastCallDict_kw(func, vc, args, nargs, kw); +} +#endif + +/* CythonFunctionShared */ + #if CYTHON_COMPILING_IN_LIMITED_API +static CYTHON_INLINE int __Pyx__IsSameCyOrCFunction(PyObject *func, void *cfunc) { + if (__Pyx_CyFunction_Check(func)) { + return PyCFunction_GetFunction(((__pyx_CyFunctionObject*)func)->func) == (PyCFunction) cfunc; + } else if (PyCFunction_Check(func)) { + return PyCFunction_GetFunction(func) == (PyCFunction) cfunc; + } + return 0; +} +#else +static CYTHON_INLINE int __Pyx__IsSameCyOrCFunction(PyObject *func, void *cfunc) { + return __Pyx_CyOrPyCFunction_Check(func) && __Pyx_CyOrPyCFunction_GET_FUNCTION(func) == (PyCFunction) cfunc; +} +#endif +static CYTHON_INLINE void __Pyx__CyFunction_SetClassObj(__pyx_CyFunctionObject* f, PyObject* classobj) { +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + __Pyx_Py_XDECREF_SET( + __Pyx_CyFunction_GetClassObj(f), + ((classobj) ? __Pyx_NewRef(classobj) : NULL)); +#else + __Pyx_Py_XDECREF_SET( + ((PyCMethodObject *) (f))->mm_class, + (PyTypeObject*)((classobj) ? __Pyx_NewRef(classobj) : NULL)); +#endif +} +static PyObject * +__Pyx_CyFunction_get_doc(__pyx_CyFunctionObject *op, void *closure) +{ + CYTHON_UNUSED_VAR(closure); + if (unlikely(op->func_doc == NULL)) { +#if CYTHON_COMPILING_IN_LIMITED_API + op->func_doc = PyObject_GetAttrString(op->func, "__doc__"); + if (unlikely(!op->func_doc)) return NULL; +#else + if (((PyCFunctionObject*)op)->m_ml->ml_doc) { +#if PY_MAJOR_VERSION >= 3 + op->func_doc = PyUnicode_FromString(((PyCFunctionObject*)op)->m_ml->ml_doc); +#else + op->func_doc = PyString_FromString(((PyCFunctionObject*)op)->m_ml->ml_doc); +#endif + if (unlikely(op->func_doc == NULL)) + return NULL; + } else { + Py_INCREF(Py_None); + return Py_None; + } +#endif + } + Py_INCREF(op->func_doc); + return op->func_doc; +} +static int +__Pyx_CyFunction_set_doc(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (value == NULL) { + value = Py_None; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_doc, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_name(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (unlikely(op->func_name == NULL)) { +#if CYTHON_COMPILING_IN_LIMITED_API + op->func_name = PyObject_GetAttrString(op->func, "__name__"); +#elif PY_MAJOR_VERSION >= 3 + op->func_name = PyUnicode_InternFromString(((PyCFunctionObject*)op)->m_ml->ml_name); +#else + op->func_name = PyString_InternFromString(((PyCFunctionObject*)op)->m_ml->ml_name); +#endif + if (unlikely(op->func_name == NULL)) + return NULL; + } + Py_INCREF(op->func_name); + return op->func_name; +} +static int +__Pyx_CyFunction_set_name(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); +#if PY_MAJOR_VERSION >= 3 + if (unlikely(value == NULL || !PyUnicode_Check(value))) +#else + if (unlikely(value == NULL || !PyString_Check(value))) +#endif + { + PyErr_SetString(PyExc_TypeError, + "__name__ must be set to a string object"); + return -1; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_name, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_qualname(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + Py_INCREF(op->func_qualname); + return op->func_qualname; +} +static int +__Pyx_CyFunction_set_qualname(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); +#if PY_MAJOR_VERSION >= 3 + if (unlikely(value == NULL || !PyUnicode_Check(value))) +#else + if (unlikely(value == NULL || !PyString_Check(value))) +#endif + { + PyErr_SetString(PyExc_TypeError, + "__qualname__ must be set to a string object"); + return -1; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_qualname, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_dict(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (unlikely(op->func_dict == NULL)) { + op->func_dict = PyDict_New(); + if (unlikely(op->func_dict == NULL)) + return NULL; + } + Py_INCREF(op->func_dict); + return op->func_dict; +} +static int +__Pyx_CyFunction_set_dict(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (unlikely(value == NULL)) { + PyErr_SetString(PyExc_TypeError, + "function's dictionary may not be deleted"); + return -1; + } + if (unlikely(!PyDict_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "setting function's dictionary to a non-dict"); + return -1; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_dict, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_globals(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + Py_INCREF(op->func_globals); + return op->func_globals; +} +static PyObject * +__Pyx_CyFunction_get_closure(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(op); + CYTHON_UNUSED_VAR(context); + Py_INCREF(Py_None); + return Py_None; +} +static PyObject * +__Pyx_CyFunction_get_code(__pyx_CyFunctionObject *op, void *context) +{ + PyObject* result = (op->func_code) ? op->func_code : Py_None; + CYTHON_UNUSED_VAR(context); + Py_INCREF(result); + return result; +} +static int +__Pyx_CyFunction_init_defaults(__pyx_CyFunctionObject *op) { + int result = 0; + PyObject *res = op->defaults_getter((PyObject *) op); + if (unlikely(!res)) + return -1; + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + op->defaults_tuple = PyTuple_GET_ITEM(res, 0); + Py_INCREF(op->defaults_tuple); + op->defaults_kwdict = PyTuple_GET_ITEM(res, 1); + Py_INCREF(op->defaults_kwdict); + #else + op->defaults_tuple = __Pyx_PySequence_ITEM(res, 0); + if (unlikely(!op->defaults_tuple)) result = -1; + else { + op->defaults_kwdict = __Pyx_PySequence_ITEM(res, 1); + if (unlikely(!op->defaults_kwdict)) result = -1; + } + #endif + Py_DECREF(res); + return result; +} +static int +__Pyx_CyFunction_set_defaults(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + if (!value) { + value = Py_None; + } else if (unlikely(value != Py_None && !PyTuple_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "__defaults__ must be set to a tuple object"); + return -1; + } + PyErr_WarnEx(PyExc_RuntimeWarning, "changes to cyfunction.__defaults__ will not " + "currently affect the values used in function calls", 1); + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->defaults_tuple, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_defaults(__pyx_CyFunctionObject *op, void *context) { + PyObject* result = op->defaults_tuple; + CYTHON_UNUSED_VAR(context); + if (unlikely(!result)) { + if (op->defaults_getter) { + if (unlikely(__Pyx_CyFunction_init_defaults(op) < 0)) return NULL; + result = op->defaults_tuple; + } else { + result = Py_None; + } + } + Py_INCREF(result); + return result; +} +static int +__Pyx_CyFunction_set_kwdefaults(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + if (!value) { + value = Py_None; + } else if (unlikely(value != Py_None && !PyDict_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "__kwdefaults__ must be set to a dict object"); + return -1; + } + PyErr_WarnEx(PyExc_RuntimeWarning, "changes to cyfunction.__kwdefaults__ will not " + "currently affect the values used in function calls", 1); + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->defaults_kwdict, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_kwdefaults(__pyx_CyFunctionObject *op, void *context) { + PyObject* result = op->defaults_kwdict; + CYTHON_UNUSED_VAR(context); + if (unlikely(!result)) { + if (op->defaults_getter) { + if (unlikely(__Pyx_CyFunction_init_defaults(op) < 0)) return NULL; + result = op->defaults_kwdict; + } else { + result = Py_None; + } + } + Py_INCREF(result); + return result; +} +static int +__Pyx_CyFunction_set_annotations(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + if (!value || value == Py_None) { + value = NULL; + } else if (unlikely(!PyDict_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "__annotations__ must be set to a dict object"); + return -1; + } + Py_XINCREF(value); + __Pyx_Py_XDECREF_SET(op->func_annotations, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_annotations(__pyx_CyFunctionObject *op, void *context) { + PyObject* result = op->func_annotations; + CYTHON_UNUSED_VAR(context); + if (unlikely(!result)) { + result = PyDict_New(); + if (unlikely(!result)) return NULL; + op->func_annotations = result; + } + Py_INCREF(result); + return result; +} +static PyObject * +__Pyx_CyFunction_get_is_coroutine(__pyx_CyFunctionObject *op, void *context) { + int is_coroutine; + CYTHON_UNUSED_VAR(context); + if (op->func_is_coroutine) { + return __Pyx_NewRef(op->func_is_coroutine); + } + is_coroutine = op->flags & __Pyx_CYFUNCTION_COROUTINE; +#if PY_VERSION_HEX >= 0x03050000 + if (is_coroutine) { + PyObject *module, *fromlist, *marker = __pyx_n_s_is_coroutine; + fromlist = PyList_New(1); + if (unlikely(!fromlist)) return NULL; + Py_INCREF(marker); +#if CYTHON_ASSUME_SAFE_MACROS + PyList_SET_ITEM(fromlist, 0, marker); +#else + if (unlikely(PyList_SetItem(fromlist, 0, marker) < 0)) { + Py_DECREF(marker); + Py_DECREF(fromlist); + return NULL; + } +#endif + module = PyImport_ImportModuleLevelObject(__pyx_n_s_asyncio_coroutines, NULL, NULL, fromlist, 0); + Py_DECREF(fromlist); + if (unlikely(!module)) goto ignore; + op->func_is_coroutine = __Pyx_PyObject_GetAttrStr(module, marker); + Py_DECREF(module); + if (likely(op->func_is_coroutine)) { + return __Pyx_NewRef(op->func_is_coroutine); + } +ignore: + PyErr_Clear(); + } +#endif + op->func_is_coroutine = __Pyx_PyBool_FromLong(is_coroutine); + return __Pyx_NewRef(op->func_is_coroutine); +} +#if CYTHON_COMPILING_IN_LIMITED_API +static PyObject * +__Pyx_CyFunction_get_module(__pyx_CyFunctionObject *op, void *context) { + CYTHON_UNUSED_VAR(context); + return PyObject_GetAttrString(op->func, "__module__"); +} +static int +__Pyx_CyFunction_set_module(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + return PyObject_SetAttrString(op->func, "__module__", value); +} +#endif +static PyGetSetDef __pyx_CyFunction_getsets[] = { + {(char *) "func_doc", (getter)__Pyx_CyFunction_get_doc, (setter)__Pyx_CyFunction_set_doc, 0, 0}, + {(char *) "__doc__", (getter)__Pyx_CyFunction_get_doc, (setter)__Pyx_CyFunction_set_doc, 0, 0}, + {(char *) "func_name", (getter)__Pyx_CyFunction_get_name, (setter)__Pyx_CyFunction_set_name, 0, 0}, + {(char *) "__name__", (getter)__Pyx_CyFunction_get_name, (setter)__Pyx_CyFunction_set_name, 0, 0}, + {(char *) "__qualname__", (getter)__Pyx_CyFunction_get_qualname, (setter)__Pyx_CyFunction_set_qualname, 0, 0}, + {(char *) "func_dict", (getter)__Pyx_CyFunction_get_dict, (setter)__Pyx_CyFunction_set_dict, 0, 0}, + {(char *) "__dict__", (getter)__Pyx_CyFunction_get_dict, (setter)__Pyx_CyFunction_set_dict, 0, 0}, + {(char *) "func_globals", (getter)__Pyx_CyFunction_get_globals, 0, 0, 0}, + {(char *) "__globals__", (getter)__Pyx_CyFunction_get_globals, 0, 0, 0}, + {(char *) "func_closure", (getter)__Pyx_CyFunction_get_closure, 0, 0, 0}, + {(char *) "__closure__", (getter)__Pyx_CyFunction_get_closure, 0, 0, 0}, + {(char *) "func_code", (getter)__Pyx_CyFunction_get_code, 0, 0, 0}, + {(char *) "__code__", (getter)__Pyx_CyFunction_get_code, 0, 0, 0}, + {(char *) "func_defaults", (getter)__Pyx_CyFunction_get_defaults, (setter)__Pyx_CyFunction_set_defaults, 0, 0}, + {(char *) "__defaults__", (getter)__Pyx_CyFunction_get_defaults, (setter)__Pyx_CyFunction_set_defaults, 0, 0}, + {(char *) "__kwdefaults__", (getter)__Pyx_CyFunction_get_kwdefaults, (setter)__Pyx_CyFunction_set_kwdefaults, 0, 0}, + {(char *) "__annotations__", (getter)__Pyx_CyFunction_get_annotations, (setter)__Pyx_CyFunction_set_annotations, 0, 0}, + {(char *) "_is_coroutine", (getter)__Pyx_CyFunction_get_is_coroutine, 0, 0, 0}, +#if CYTHON_COMPILING_IN_LIMITED_API + {"__module__", (getter)__Pyx_CyFunction_get_module, (setter)__Pyx_CyFunction_set_module, 0, 0}, +#endif + {0, 0, 0, 0, 0} +}; +static PyMemberDef __pyx_CyFunction_members[] = { +#if !CYTHON_COMPILING_IN_LIMITED_API + {(char *) "__module__", T_OBJECT, offsetof(PyCFunctionObject, m_module), 0, 0}, +#endif +#if CYTHON_USE_TYPE_SPECS + {(char *) "__dictoffset__", T_PYSSIZET, offsetof(__pyx_CyFunctionObject, func_dict), READONLY, 0}, +#if CYTHON_METH_FASTCALL +#if CYTHON_BACKPORT_VECTORCALL + {(char *) "__vectorcalloffset__", T_PYSSIZET, offsetof(__pyx_CyFunctionObject, func_vectorcall), READONLY, 0}, +#else +#if !CYTHON_COMPILING_IN_LIMITED_API + {(char *) "__vectorcalloffset__", T_PYSSIZET, offsetof(PyCFunctionObject, vectorcall), READONLY, 0}, +#endif +#endif +#endif +#if PY_VERSION_HEX < 0x030500A0 || CYTHON_COMPILING_IN_LIMITED_API + {(char *) "__weaklistoffset__", T_PYSSIZET, offsetof(__pyx_CyFunctionObject, func_weakreflist), READONLY, 0}, +#else + {(char *) "__weaklistoffset__", T_PYSSIZET, offsetof(PyCFunctionObject, m_weakreflist), READONLY, 0}, +#endif +#endif + {0, 0, 0, 0, 0} +}; +static PyObject * +__Pyx_CyFunction_reduce(__pyx_CyFunctionObject *m, PyObject *args) +{ + CYTHON_UNUSED_VAR(args); +#if PY_MAJOR_VERSION >= 3 + Py_INCREF(m->func_qualname); + return m->func_qualname; +#else + return PyString_FromString(((PyCFunctionObject*)m)->m_ml->ml_name); +#endif +} +static PyMethodDef __pyx_CyFunction_methods[] = { + {"__reduce__", (PyCFunction)__Pyx_CyFunction_reduce, METH_VARARGS, 0}, + {0, 0, 0, 0} +}; +#if PY_VERSION_HEX < 0x030500A0 || CYTHON_COMPILING_IN_LIMITED_API +#define __Pyx_CyFunction_weakreflist(cyfunc) ((cyfunc)->func_weakreflist) +#else +#define __Pyx_CyFunction_weakreflist(cyfunc) (((PyCFunctionObject*)cyfunc)->m_weakreflist) +#endif +static PyObject *__Pyx_CyFunction_Init(__pyx_CyFunctionObject *op, PyMethodDef *ml, int flags, PyObject* qualname, + PyObject *closure, PyObject *module, PyObject* globals, PyObject* code) { +#if !CYTHON_COMPILING_IN_LIMITED_API + PyCFunctionObject *cf = (PyCFunctionObject*) op; +#endif + if (unlikely(op == NULL)) + return NULL; +#if CYTHON_COMPILING_IN_LIMITED_API + op->func = PyCFunction_NewEx(ml, (PyObject*)op, module); + if (unlikely(!op->func)) return NULL; +#endif + op->flags = flags; + __Pyx_CyFunction_weakreflist(op) = NULL; +#if !CYTHON_COMPILING_IN_LIMITED_API + cf->m_ml = ml; + cf->m_self = (PyObject *) op; +#endif + Py_XINCREF(closure); + op->func_closure = closure; +#if !CYTHON_COMPILING_IN_LIMITED_API + Py_XINCREF(module); + cf->m_module = module; +#endif + op->func_dict = NULL; + op->func_name = NULL; + Py_INCREF(qualname); + op->func_qualname = qualname; + op->func_doc = NULL; +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + op->func_classobj = NULL; +#else + ((PyCMethodObject*)op)->mm_class = NULL; +#endif + op->func_globals = globals; + Py_INCREF(op->func_globals); + Py_XINCREF(code); + op->func_code = code; + op->defaults_pyobjects = 0; + op->defaults_size = 0; + op->defaults = NULL; + op->defaults_tuple = NULL; + op->defaults_kwdict = NULL; + op->defaults_getter = NULL; + op->func_annotations = NULL; + op->func_is_coroutine = NULL; +#if CYTHON_METH_FASTCALL + switch (ml->ml_flags & (METH_VARARGS | METH_FASTCALL | METH_NOARGS | METH_O | METH_KEYWORDS | METH_METHOD)) { + case METH_NOARGS: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_NOARGS; + break; + case METH_O: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_O; + break; + case METH_METHOD | METH_FASTCALL | METH_KEYWORDS: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS_METHOD; + break; + case METH_FASTCALL | METH_KEYWORDS: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS; + break; + case METH_VARARGS | METH_KEYWORDS: + __Pyx_CyFunction_func_vectorcall(op) = NULL; + break; + default: + PyErr_SetString(PyExc_SystemError, "Bad call flags for CyFunction"); + Py_DECREF(op); + return NULL; + } +#endif + return (PyObject *) op; +} +static int +__Pyx_CyFunction_clear(__pyx_CyFunctionObject *m) +{ + Py_CLEAR(m->func_closure); +#if CYTHON_COMPILING_IN_LIMITED_API + Py_CLEAR(m->func); +#else + Py_CLEAR(((PyCFunctionObject*)m)->m_module); +#endif + Py_CLEAR(m->func_dict); + Py_CLEAR(m->func_name); + Py_CLEAR(m->func_qualname); + Py_CLEAR(m->func_doc); + Py_CLEAR(m->func_globals); + Py_CLEAR(m->func_code); +#if !CYTHON_COMPILING_IN_LIMITED_API +#if PY_VERSION_HEX < 0x030900B1 + Py_CLEAR(__Pyx_CyFunction_GetClassObj(m)); +#else + { + PyObject *cls = (PyObject*) ((PyCMethodObject *) (m))->mm_class; + ((PyCMethodObject *) (m))->mm_class = NULL; + Py_XDECREF(cls); + } +#endif +#endif + Py_CLEAR(m->defaults_tuple); + Py_CLEAR(m->defaults_kwdict); + Py_CLEAR(m->func_annotations); + Py_CLEAR(m->func_is_coroutine); + if (m->defaults) { + PyObject **pydefaults = __Pyx_CyFunction_Defaults(PyObject *, m); + int i; + for (i = 0; i < m->defaults_pyobjects; i++) + Py_XDECREF(pydefaults[i]); + PyObject_Free(m->defaults); + m->defaults = NULL; + } + return 0; +} +static void __Pyx__CyFunction_dealloc(__pyx_CyFunctionObject *m) +{ + if (__Pyx_CyFunction_weakreflist(m) != NULL) + PyObject_ClearWeakRefs((PyObject *) m); + __Pyx_CyFunction_clear(m); + __Pyx_PyHeapTypeObject_GC_Del(m); +} +static void __Pyx_CyFunction_dealloc(__pyx_CyFunctionObject *m) +{ + PyObject_GC_UnTrack(m); + __Pyx__CyFunction_dealloc(m); +} +static int __Pyx_CyFunction_traverse(__pyx_CyFunctionObject *m, visitproc visit, void *arg) +{ + Py_VISIT(m->func_closure); +#if CYTHON_COMPILING_IN_LIMITED_API + Py_VISIT(m->func); +#else + Py_VISIT(((PyCFunctionObject*)m)->m_module); +#endif + Py_VISIT(m->func_dict); + Py_VISIT(m->func_name); + Py_VISIT(m->func_qualname); + Py_VISIT(m->func_doc); + Py_VISIT(m->func_globals); + Py_VISIT(m->func_code); +#if !CYTHON_COMPILING_IN_LIMITED_API + Py_VISIT(__Pyx_CyFunction_GetClassObj(m)); +#endif + Py_VISIT(m->defaults_tuple); + Py_VISIT(m->defaults_kwdict); + Py_VISIT(m->func_is_coroutine); + if (m->defaults) { + PyObject **pydefaults = __Pyx_CyFunction_Defaults(PyObject *, m); + int i; + for (i = 0; i < m->defaults_pyobjects; i++) + Py_VISIT(pydefaults[i]); + } + return 0; +} +static PyObject* +__Pyx_CyFunction_repr(__pyx_CyFunctionObject *op) +{ +#if PY_MAJOR_VERSION >= 3 + return PyUnicode_FromFormat("", + op->func_qualname, (void *)op); +#else + return PyString_FromFormat("", + PyString_AsString(op->func_qualname), (void *)op); +#endif +} +static PyObject * __Pyx_CyFunction_CallMethod(PyObject *func, PyObject *self, PyObject *arg, PyObject *kw) { +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *f = ((__pyx_CyFunctionObject*)func)->func; + PyObject *py_name = NULL; + PyCFunction meth; + int flags; + meth = PyCFunction_GetFunction(f); + if (unlikely(!meth)) return NULL; + flags = PyCFunction_GetFlags(f); + if (unlikely(flags < 0)) return NULL; +#else + PyCFunctionObject* f = (PyCFunctionObject*)func; + PyCFunction meth = f->m_ml->ml_meth; + int flags = f->m_ml->ml_flags; +#endif + Py_ssize_t size; + switch (flags & (METH_VARARGS | METH_KEYWORDS | METH_NOARGS | METH_O)) { + case METH_VARARGS: + if (likely(kw == NULL || PyDict_Size(kw) == 0)) + return (*meth)(self, arg); + break; + case METH_VARARGS | METH_KEYWORDS: + return (*(PyCFunctionWithKeywords)(void*)meth)(self, arg, kw); + case METH_NOARGS: + if (likely(kw == NULL || PyDict_Size(kw) == 0)) { +#if CYTHON_ASSUME_SAFE_MACROS + size = PyTuple_GET_SIZE(arg); +#else + size = PyTuple_Size(arg); + if (unlikely(size < 0)) return NULL; +#endif + if (likely(size == 0)) + return (*meth)(self, NULL); +#if CYTHON_COMPILING_IN_LIMITED_API + py_name = __Pyx_CyFunction_get_name((__pyx_CyFunctionObject*)func, NULL); + if (!py_name) return NULL; + PyErr_Format(PyExc_TypeError, + "%.200S() takes no arguments (%" CYTHON_FORMAT_SSIZE_T "d given)", + py_name, size); + Py_DECREF(py_name); +#else + PyErr_Format(PyExc_TypeError, + "%.200s() takes no arguments (%" CYTHON_FORMAT_SSIZE_T "d given)", + f->m_ml->ml_name, size); +#endif + return NULL; + } + break; + case METH_O: + if (likely(kw == NULL || PyDict_Size(kw) == 0)) { +#if CYTHON_ASSUME_SAFE_MACROS + size = PyTuple_GET_SIZE(arg); +#else + size = PyTuple_Size(arg); + if (unlikely(size < 0)) return NULL; +#endif + if (likely(size == 1)) { + PyObject *result, *arg0; + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + arg0 = PyTuple_GET_ITEM(arg, 0); + #else + arg0 = __Pyx_PySequence_ITEM(arg, 0); if (unlikely(!arg0)) return NULL; + #endif + result = (*meth)(self, arg0); + #if !(CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS) + Py_DECREF(arg0); + #endif + return result; + } +#if CYTHON_COMPILING_IN_LIMITED_API + py_name = __Pyx_CyFunction_get_name((__pyx_CyFunctionObject*)func, NULL); + if (!py_name) return NULL; + PyErr_Format(PyExc_TypeError, + "%.200S() takes exactly one argument (%" CYTHON_FORMAT_SSIZE_T "d given)", + py_name, size); + Py_DECREF(py_name); +#else + PyErr_Format(PyExc_TypeError, + "%.200s() takes exactly one argument (%" CYTHON_FORMAT_SSIZE_T "d given)", + f->m_ml->ml_name, size); +#endif + return NULL; + } + break; + default: + PyErr_SetString(PyExc_SystemError, "Bad call flags for CyFunction"); + return NULL; + } +#if CYTHON_COMPILING_IN_LIMITED_API + py_name = __Pyx_CyFunction_get_name((__pyx_CyFunctionObject*)func, NULL); + if (!py_name) return NULL; + PyErr_Format(PyExc_TypeError, "%.200S() takes no keyword arguments", + py_name); + Py_DECREF(py_name); +#else + PyErr_Format(PyExc_TypeError, "%.200s() takes no keyword arguments", + f->m_ml->ml_name); +#endif + return NULL; +} +static CYTHON_INLINE PyObject *__Pyx_CyFunction_Call(PyObject *func, PyObject *arg, PyObject *kw) { + PyObject *self, *result; +#if CYTHON_COMPILING_IN_LIMITED_API + self = PyCFunction_GetSelf(((__pyx_CyFunctionObject*)func)->func); + if (unlikely(!self) && PyErr_Occurred()) return NULL; +#else + self = ((PyCFunctionObject*)func)->m_self; +#endif + result = __Pyx_CyFunction_CallMethod(func, self, arg, kw); + return result; +} +static PyObject *__Pyx_CyFunction_CallAsMethod(PyObject *func, PyObject *args, PyObject *kw) { + PyObject *result; + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *) func; +#if CYTHON_METH_FASTCALL + __pyx_vectorcallfunc vc = __Pyx_CyFunction_func_vectorcall(cyfunc); + if (vc) { +#if CYTHON_ASSUME_SAFE_MACROS + return __Pyx_PyVectorcall_FastCallDict(func, vc, &PyTuple_GET_ITEM(args, 0), (size_t)PyTuple_GET_SIZE(args), kw); +#else + (void) &__Pyx_PyVectorcall_FastCallDict; + return PyVectorcall_Call(func, args, kw); +#endif + } +#endif + if ((cyfunc->flags & __Pyx_CYFUNCTION_CCLASS) && !(cyfunc->flags & __Pyx_CYFUNCTION_STATICMETHOD)) { + Py_ssize_t argc; + PyObject *new_args; + PyObject *self; +#if CYTHON_ASSUME_SAFE_MACROS + argc = PyTuple_GET_SIZE(args); +#else + argc = PyTuple_Size(args); + if (unlikely(!argc) < 0) return NULL; +#endif + new_args = PyTuple_GetSlice(args, 1, argc); + if (unlikely(!new_args)) + return NULL; + self = PyTuple_GetItem(args, 0); + if (unlikely(!self)) { + Py_DECREF(new_args); +#if PY_MAJOR_VERSION > 2 + PyErr_Format(PyExc_TypeError, + "unbound method %.200S() needs an argument", + cyfunc->func_qualname); +#else + PyErr_SetString(PyExc_TypeError, + "unbound method needs an argument"); +#endif + return NULL; + } + result = __Pyx_CyFunction_CallMethod(func, self, new_args, kw); + Py_DECREF(new_args); + } else { + result = __Pyx_CyFunction_Call(func, args, kw); + } + return result; +} +#if CYTHON_METH_FASTCALL +static CYTHON_INLINE int __Pyx_CyFunction_Vectorcall_CheckArgs(__pyx_CyFunctionObject *cyfunc, Py_ssize_t nargs, PyObject *kwnames) +{ + int ret = 0; + if ((cyfunc->flags & __Pyx_CYFUNCTION_CCLASS) && !(cyfunc->flags & __Pyx_CYFUNCTION_STATICMETHOD)) { + if (unlikely(nargs < 1)) { + PyErr_Format(PyExc_TypeError, "%.200s() needs an argument", + ((PyCFunctionObject*)cyfunc)->m_ml->ml_name); + return -1; + } + ret = 1; + } + if (unlikely(kwnames) && unlikely(PyTuple_GET_SIZE(kwnames))) { + PyErr_Format(PyExc_TypeError, + "%.200s() takes no keyword arguments", ((PyCFunctionObject*)cyfunc)->m_ml->ml_name); + return -1; + } + return ret; +} +static PyObject * __Pyx_CyFunction_Vectorcall_NOARGS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, kwnames)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + if (unlikely(nargs != 0)) { + PyErr_Format(PyExc_TypeError, + "%.200s() takes no arguments (%" CYTHON_FORMAT_SSIZE_T "d given)", + def->ml_name, nargs); + return NULL; + } + return def->ml_meth(self, NULL); +} +static PyObject * __Pyx_CyFunction_Vectorcall_O(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, kwnames)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + if (unlikely(nargs != 1)) { + PyErr_Format(PyExc_TypeError, + "%.200s() takes exactly one argument (%" CYTHON_FORMAT_SSIZE_T "d given)", + def->ml_name, nargs); + return NULL; + } + return def->ml_meth(self, args[0]); +} +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, NULL)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + return ((__Pyx_PyCFunctionFastWithKeywords)(void(*)(void))def->ml_meth)(self, args, nargs, kwnames); +} +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS_METHOD(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; + PyTypeObject *cls = (PyTypeObject *) __Pyx_CyFunction_GetClassObj(cyfunc); +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, NULL)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + return ((__Pyx_PyCMethod)(void(*)(void))def->ml_meth)(self, cls, args, (size_t)nargs, kwnames); +} +#endif +#if CYTHON_USE_TYPE_SPECS +static PyType_Slot __pyx_CyFunctionType_slots[] = { + {Py_tp_dealloc, (void *)__Pyx_CyFunction_dealloc}, + {Py_tp_repr, (void *)__Pyx_CyFunction_repr}, + {Py_tp_call, (void *)__Pyx_CyFunction_CallAsMethod}, + {Py_tp_traverse, (void *)__Pyx_CyFunction_traverse}, + {Py_tp_clear, (void *)__Pyx_CyFunction_clear}, + {Py_tp_methods, (void *)__pyx_CyFunction_methods}, + {Py_tp_members, (void *)__pyx_CyFunction_members}, + {Py_tp_getset, (void *)__pyx_CyFunction_getsets}, + {Py_tp_descr_get, (void *)__Pyx_PyMethod_New}, + {0, 0}, +}; +static PyType_Spec __pyx_CyFunctionType_spec = { + __PYX_TYPE_MODULE_PREFIX "cython_function_or_method", + sizeof(__pyx_CyFunctionObject), + 0, +#ifdef Py_TPFLAGS_METHOD_DESCRIPTOR + Py_TPFLAGS_METHOD_DESCRIPTOR | +#endif +#if (defined(_Py_TPFLAGS_HAVE_VECTORCALL) && CYTHON_METH_FASTCALL) + _Py_TPFLAGS_HAVE_VECTORCALL | +#endif + Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HAVE_GC | Py_TPFLAGS_BASETYPE, + __pyx_CyFunctionType_slots +}; +#else +static PyTypeObject __pyx_CyFunctionType_type = { + PyVarObject_HEAD_INIT(0, 0) + __PYX_TYPE_MODULE_PREFIX "cython_function_or_method", + sizeof(__pyx_CyFunctionObject), + 0, + (destructor) __Pyx_CyFunction_dealloc, +#if !CYTHON_METH_FASTCALL + 0, +#elif CYTHON_BACKPORT_VECTORCALL + (printfunc)offsetof(__pyx_CyFunctionObject, func_vectorcall), +#else + offsetof(PyCFunctionObject, vectorcall), +#endif + 0, + 0, +#if PY_MAJOR_VERSION < 3 + 0, +#else + 0, +#endif + (reprfunc) __Pyx_CyFunction_repr, + 0, + 0, + 0, + 0, + __Pyx_CyFunction_CallAsMethod, + 0, + 0, + 0, + 0, +#ifdef Py_TPFLAGS_METHOD_DESCRIPTOR + Py_TPFLAGS_METHOD_DESCRIPTOR | +#endif +#if defined(_Py_TPFLAGS_HAVE_VECTORCALL) && CYTHON_METH_FASTCALL + _Py_TPFLAGS_HAVE_VECTORCALL | +#endif + Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HAVE_GC | Py_TPFLAGS_BASETYPE, + 0, + (traverseproc) __Pyx_CyFunction_traverse, + (inquiry) __Pyx_CyFunction_clear, + 0, +#if PY_VERSION_HEX < 0x030500A0 + offsetof(__pyx_CyFunctionObject, func_weakreflist), +#else + offsetof(PyCFunctionObject, m_weakreflist), +#endif + 0, + 0, + __pyx_CyFunction_methods, + __pyx_CyFunction_members, + __pyx_CyFunction_getsets, + 0, + 0, + __Pyx_PyMethod_New, + 0, + offsetof(__pyx_CyFunctionObject, func_dict), + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, +#if PY_VERSION_HEX >= 0x030400a1 + 0, +#endif +#if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, +#endif +#if __PYX_NEED_TP_PRINT_SLOT + 0, +#endif +#if PY_VERSION_HEX >= 0x030C0000 + 0, +#endif +#if PY_VERSION_HEX >= 0x030d00A4 + 0, +#endif +#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, +#endif +}; +#endif +static int __pyx_CyFunction_init(PyObject *module) { +#if CYTHON_USE_TYPE_SPECS + __pyx_CyFunctionType = __Pyx_FetchCommonTypeFromSpec(module, &__pyx_CyFunctionType_spec, NULL); +#else + CYTHON_UNUSED_VAR(module); + __pyx_CyFunctionType = __Pyx_FetchCommonType(&__pyx_CyFunctionType_type); +#endif + if (unlikely(__pyx_CyFunctionType == NULL)) { + return -1; + } + return 0; +} +static CYTHON_INLINE void *__Pyx_CyFunction_InitDefaults(PyObject *func, size_t size, int pyobjects) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->defaults = PyObject_Malloc(size); + if (unlikely(!m->defaults)) + return PyErr_NoMemory(); + memset(m->defaults, 0, size); + m->defaults_pyobjects = pyobjects; + m->defaults_size = size; + return m->defaults; +} +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsTuple(PyObject *func, PyObject *tuple) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->defaults_tuple = tuple; + Py_INCREF(tuple); +} +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsKwDict(PyObject *func, PyObject *dict) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->defaults_kwdict = dict; + Py_INCREF(dict); +} +static CYTHON_INLINE void __Pyx_CyFunction_SetAnnotationsDict(PyObject *func, PyObject *dict) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->func_annotations = dict; + Py_INCREF(dict); +} + +/* CythonFunction */ + static PyObject *__Pyx_CyFunction_New(PyMethodDef *ml, int flags, PyObject* qualname, + PyObject *closure, PyObject *module, PyObject* globals, PyObject* code) { + PyObject *op = __Pyx_CyFunction_Init( + PyObject_GC_New(__pyx_CyFunctionObject, __pyx_CyFunctionType), + ml, flags, qualname, closure, module, globals, code + ); + if (likely(op)) { + PyObject_GC_Track(op); + } + return op; +} + +/* CLineInTraceback */ + #ifndef CYTHON_CLINE_IN_TRACEBACK +static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line) { + PyObject *use_cline; + PyObject *ptype, *pvalue, *ptraceback; +#if CYTHON_COMPILING_IN_CPYTHON + PyObject **cython_runtime_dict; +#endif + CYTHON_MAYBE_UNUSED_VAR(tstate); + if (unlikely(!__pyx_cython_runtime)) { + return c_line; + } + __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback); +#if CYTHON_COMPILING_IN_CPYTHON + cython_runtime_dict = _PyObject_GetDictPtr(__pyx_cython_runtime); + if (likely(cython_runtime_dict)) { + __PYX_PY_DICT_LOOKUP_IF_MODIFIED( + use_cline, *cython_runtime_dict, + __Pyx_PyDict_GetItemStr(*cython_runtime_dict, __pyx_n_s_cline_in_traceback)) + } else +#endif + { + PyObject *use_cline_obj = __Pyx_PyObject_GetAttrStrNoError(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback); + if (use_cline_obj) { + use_cline = PyObject_Not(use_cline_obj) ? Py_False : Py_True; + Py_DECREF(use_cline_obj); + } else { + PyErr_Clear(); + use_cline = NULL; + } + } + if (!use_cline) { + c_line = 0; + (void) PyObject_SetAttr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback, Py_False); + } + else if (use_cline == Py_False || (use_cline != Py_True && PyObject_Not(use_cline) != 0)) { + c_line = 0; + } + __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback); + return c_line; +} +#endif + +/* CodeObjectCache */ + #if !CYTHON_COMPILING_IN_LIMITED_API +static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line) { + int start = 0, mid = 0, end = count - 1; + if (end >= 0 && code_line > entries[end].code_line) { + return count; + } + while (start < end) { + mid = start + (end - start) / 2; + if (code_line < entries[mid].code_line) { + end = mid; + } else if (code_line > entries[mid].code_line) { + start = mid + 1; + } else { + return mid; + } + } + if (code_line <= entries[mid].code_line) { + return mid; + } else { + return mid + 1; + } +} +static PyCodeObject *__pyx_find_code_object(int code_line) { + PyCodeObject* code_object; + int pos; + if (unlikely(!code_line) || unlikely(!__pyx_code_cache.entries)) { + return NULL; + } + pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line); + if (unlikely(pos >= __pyx_code_cache.count) || unlikely(__pyx_code_cache.entries[pos].code_line != code_line)) { + return NULL; + } + code_object = __pyx_code_cache.entries[pos].code_object; + Py_INCREF(code_object); + return code_object; +} +static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object) { + int pos, i; + __Pyx_CodeObjectCacheEntry* entries = __pyx_code_cache.entries; + if (unlikely(!code_line)) { + return; + } + if (unlikely(!entries)) { + entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Malloc(64*sizeof(__Pyx_CodeObjectCacheEntry)); + if (likely(entries)) { + __pyx_code_cache.entries = entries; + __pyx_code_cache.max_count = 64; + __pyx_code_cache.count = 1; + entries[0].code_line = code_line; + entries[0].code_object = code_object; + Py_INCREF(code_object); + } + return; + } + pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line); + if ((pos < __pyx_code_cache.count) && unlikely(__pyx_code_cache.entries[pos].code_line == code_line)) { + PyCodeObject* tmp = entries[pos].code_object; + entries[pos].code_object = code_object; + Py_DECREF(tmp); + return; + } + if (__pyx_code_cache.count == __pyx_code_cache.max_count) { + int new_max = __pyx_code_cache.max_count + 64; + entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Realloc( + __pyx_code_cache.entries, ((size_t)new_max) * sizeof(__Pyx_CodeObjectCacheEntry)); + if (unlikely(!entries)) { + return; + } + __pyx_code_cache.entries = entries; + __pyx_code_cache.max_count = new_max; + } + for (i=__pyx_code_cache.count; i>pos; i--) { + entries[i] = entries[i-1]; + } + entries[pos].code_line = code_line; + entries[pos].code_object = code_object; + __pyx_code_cache.count++; + Py_INCREF(code_object); +} +#endif + +/* AddTraceback */ + #include "compile.h" +#include "frameobject.h" +#include "traceback.h" +#if PY_VERSION_HEX >= 0x030b00a6 && !CYTHON_COMPILING_IN_LIMITED_API && !defined(PYPY_VERSION) + #ifndef Py_BUILD_CORE + #define Py_BUILD_CORE 1 + #endif + #include "internal/pycore_frame.h" +#endif +#if CYTHON_COMPILING_IN_LIMITED_API +static PyObject *__Pyx_PyCode_Replace_For_AddTraceback(PyObject *code, PyObject *scratch_dict, + PyObject *firstlineno, PyObject *name) { + PyObject *replace = NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "co_firstlineno", firstlineno))) return NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "co_name", name))) return NULL; + replace = PyObject_GetAttrString(code, "replace"); + if (likely(replace)) { + PyObject *result; + result = PyObject_Call(replace, __pyx_empty_tuple, scratch_dict); + Py_DECREF(replace); + return result; + } + PyErr_Clear(); + #if __PYX_LIMITED_VERSION_HEX < 0x030780000 + { + PyObject *compiled = NULL, *result = NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "code", code))) return NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "type", (PyObject*)(&PyType_Type)))) return NULL; + compiled = Py_CompileString( + "out = type(code)(\n" + " code.co_argcount, code.co_kwonlyargcount, code.co_nlocals, code.co_stacksize,\n" + " code.co_flags, code.co_code, code.co_consts, code.co_names,\n" + " code.co_varnames, code.co_filename, co_name, co_firstlineno,\n" + " code.co_lnotab)\n", "", Py_file_input); + if (!compiled) return NULL; + result = PyEval_EvalCode(compiled, scratch_dict, scratch_dict); + Py_DECREF(compiled); + if (!result) PyErr_Print(); + Py_DECREF(result); + result = PyDict_GetItemString(scratch_dict, "out"); + if (result) Py_INCREF(result); + return result; + } + #else + return NULL; + #endif +} +static void __Pyx_AddTraceback(const char *funcname, int c_line, + int py_line, const char *filename) { + PyObject *code_object = NULL, *py_py_line = NULL, *py_funcname = NULL, *dict = NULL; + PyObject *replace = NULL, *getframe = NULL, *frame = NULL; + PyObject *exc_type, *exc_value, *exc_traceback; + int success = 0; + if (c_line) { + (void) __pyx_cfilenm; + (void) __Pyx_CLineForTraceback(__Pyx_PyThreadState_Current, c_line); + } + PyErr_Fetch(&exc_type, &exc_value, &exc_traceback); + code_object = Py_CompileString("_getframe()", filename, Py_eval_input); + if (unlikely(!code_object)) goto bad; + py_py_line = PyLong_FromLong(py_line); + if (unlikely(!py_py_line)) goto bad; + py_funcname = PyUnicode_FromString(funcname); + if (unlikely(!py_funcname)) goto bad; + dict = PyDict_New(); + if (unlikely(!dict)) goto bad; + { + PyObject *old_code_object = code_object; + code_object = __Pyx_PyCode_Replace_For_AddTraceback(code_object, dict, py_py_line, py_funcname); + Py_DECREF(old_code_object); + } + if (unlikely(!code_object)) goto bad; + getframe = PySys_GetObject("_getframe"); + if (unlikely(!getframe)) goto bad; + if (unlikely(PyDict_SetItemString(dict, "_getframe", getframe))) goto bad; + frame = PyEval_EvalCode(code_object, dict, dict); + if (unlikely(!frame) || frame == Py_None) goto bad; + success = 1; + bad: + PyErr_Restore(exc_type, exc_value, exc_traceback); + Py_XDECREF(code_object); + Py_XDECREF(py_py_line); + Py_XDECREF(py_funcname); + Py_XDECREF(dict); + Py_XDECREF(replace); + if (success) { + PyTraceBack_Here( + (struct _frame*)frame); + } + Py_XDECREF(frame); +} +#else +static PyCodeObject* __Pyx_CreateCodeObjectForTraceback( + const char *funcname, int c_line, + int py_line, const char *filename) { + PyCodeObject *py_code = NULL; + PyObject *py_funcname = NULL; + #if PY_MAJOR_VERSION < 3 + PyObject *py_srcfile = NULL; + py_srcfile = PyString_FromString(filename); + if (!py_srcfile) goto bad; + #endif + if (c_line) { + #if PY_MAJOR_VERSION < 3 + py_funcname = PyString_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line); + if (!py_funcname) goto bad; + #else + py_funcname = PyUnicode_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line); + if (!py_funcname) goto bad; + funcname = PyUnicode_AsUTF8(py_funcname); + if (!funcname) goto bad; + #endif + } + else { + #if PY_MAJOR_VERSION < 3 + py_funcname = PyString_FromString(funcname); + if (!py_funcname) goto bad; + #endif + } + #if PY_MAJOR_VERSION < 3 + py_code = __Pyx_PyCode_New( + 0, + 0, + 0, + 0, + 0, + 0, + __pyx_empty_bytes, /*PyObject *code,*/ + __pyx_empty_tuple, /*PyObject *consts,*/ + __pyx_empty_tuple, /*PyObject *names,*/ + __pyx_empty_tuple, /*PyObject *varnames,*/ + __pyx_empty_tuple, /*PyObject *freevars,*/ + __pyx_empty_tuple, /*PyObject *cellvars,*/ + py_srcfile, /*PyObject *filename,*/ + py_funcname, /*PyObject *name,*/ + py_line, + __pyx_empty_bytes /*PyObject *lnotab*/ + ); + Py_DECREF(py_srcfile); + #else + py_code = PyCode_NewEmpty(filename, funcname, py_line); + #endif + Py_XDECREF(py_funcname); + return py_code; +bad: + Py_XDECREF(py_funcname); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(py_srcfile); + #endif + return NULL; +} +static void __Pyx_AddTraceback(const char *funcname, int c_line, + int py_line, const char *filename) { + PyCodeObject *py_code = 0; + PyFrameObject *py_frame = 0; + PyThreadState *tstate = __Pyx_PyThreadState_Current; + PyObject *ptype, *pvalue, *ptraceback; + if (c_line) { + c_line = __Pyx_CLineForTraceback(tstate, c_line); + } + py_code = __pyx_find_code_object(c_line ? -c_line : py_line); + if (!py_code) { + __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback); + py_code = __Pyx_CreateCodeObjectForTraceback( + funcname, c_line, py_line, filename); + if (!py_code) { + /* If the code object creation fails, then we should clear the + fetched exception references and propagate the new exception */ + Py_XDECREF(ptype); + Py_XDECREF(pvalue); + Py_XDECREF(ptraceback); + goto bad; + } + __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback); + __pyx_insert_code_object(c_line ? -c_line : py_line, py_code); + } + py_frame = PyFrame_New( + tstate, /*PyThreadState *tstate,*/ + py_code, /*PyCodeObject *code,*/ + __pyx_d, /*PyObject *globals,*/ + 0 /*PyObject *locals*/ + ); + if (!py_frame) goto bad; + __Pyx_PyFrame_SetLineNumber(py_frame, py_line); + PyTraceBack_Here(py_frame); +bad: + Py_XDECREF(py_code); + Py_XDECREF(py_frame); +} +#endif + +#if PY_MAJOR_VERSION < 3 +static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags) { + __Pyx_TypeName obj_type_name; + if (PyObject_CheckBuffer(obj)) return PyObject_GetBuffer(obj, view, flags); + if (__Pyx_TypeCheck(obj, __pyx_array_type)) return __pyx_array_getbuffer(obj, view, flags); + if (__Pyx_TypeCheck(obj, __pyx_memoryview_type)) return __pyx_memoryview_getbuffer(obj, view, flags); + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, + "'" __Pyx_FMT_TYPENAME "' does not have the buffer interface", + obj_type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return -1; +} +static void __Pyx_ReleaseBuffer(Py_buffer *view) { + PyObject *obj = view->obj; + if (!obj) return; + if (PyObject_CheckBuffer(obj)) { + PyBuffer_Release(view); + return; + } + if ((0)) {} + view->obj = NULL; + Py_DECREF(obj); +} +#endif + + + /* MemviewSliceIsContig */ + static int +__pyx_memviewslice_is_contig(const __Pyx_memviewslice mvs, char order, int ndim) +{ + int i, index, step, start; + Py_ssize_t itemsize = mvs.memview->view.itemsize; + if (order == 'F') { + step = 1; + start = 0; + } else { + step = -1; + start = ndim - 1; + } + for (i = 0; i < ndim; i++) { + index = start + step * i; + if (mvs.suboffsets[index] >= 0 || mvs.strides[index] != itemsize) + return 0; + itemsize *= mvs.shape[index]; + } + return 1; +} + +/* OverlappingSlices */ + static void +__pyx_get_array_memory_extents(__Pyx_memviewslice *slice, + void **out_start, void **out_end, + int ndim, size_t itemsize) +{ + char *start, *end; + int i; + start = end = slice->data; + for (i = 0; i < ndim; i++) { + Py_ssize_t stride = slice->strides[i]; + Py_ssize_t extent = slice->shape[i]; + if (extent == 0) { + *out_start = *out_end = start; + return; + } else { + if (stride > 0) + end += stride * (extent - 1); + else + start += stride * (extent - 1); + } + } + *out_start = start; + *out_end = end + itemsize; +} +static int +__pyx_slices_overlap(__Pyx_memviewslice *slice1, + __Pyx_memviewslice *slice2, + int ndim, size_t itemsize) +{ + void *start1, *end1, *start2, *end2; + __pyx_get_array_memory_extents(slice1, &start1, &end1, ndim, itemsize); + __pyx_get_array_memory_extents(slice2, &start2, &end2, ndim, itemsize); + return (start1 < end2) && (start2 < end1); +} + +/* CIntFromPyVerify */ + #define __PYX_VERIFY_RETURN_INT(target_type, func_type, func_value)\ + __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 0) +#define __PYX_VERIFY_RETURN_INT_EXC(target_type, func_type, func_value)\ + __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 1) +#define __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, exc)\ + {\ + func_type value = func_value;\ + if (sizeof(target_type) < sizeof(func_type)) {\ + if (unlikely(value != (func_type) (target_type) value)) {\ + func_type zero = 0;\ + if (exc && unlikely(value == (func_type)-1 && PyErr_Occurred()))\ + return (target_type) -1;\ + if (is_unsigned && unlikely(value < zero))\ + goto raise_neg_overflow;\ + else\ + goto raise_overflow;\ + }\ + }\ + return (target_type) value;\ + } + +/* TypeInfoCompare */ + static int +__pyx_typeinfo_cmp(__Pyx_TypeInfo *a, __Pyx_TypeInfo *b) +{ + int i; + if (!a || !b) + return 0; + if (a == b) + return 1; + if (a->size != b->size || a->typegroup != b->typegroup || + a->is_unsigned != b->is_unsigned || a->ndim != b->ndim) { + if (a->typegroup == 'H' || b->typegroup == 'H') { + return a->size == b->size; + } else { + return 0; + } + } + if (a->ndim) { + for (i = 0; i < a->ndim; i++) + if (a->arraysize[i] != b->arraysize[i]) + return 0; + } + if (a->typegroup == 'S') { + if (a->flags != b->flags) + return 0; + if (a->fields || b->fields) { + if (!(a->fields && b->fields)) + return 0; + for (i = 0; a->fields[i].type && b->fields[i].type; i++) { + __Pyx_StructField *field_a = a->fields + i; + __Pyx_StructField *field_b = b->fields + i; + if (field_a->offset != field_b->offset || + !__pyx_typeinfo_cmp(field_a->type, field_b->type)) + return 0; + } + return !a->fields[i].type && !b->fields[i].type; + } + } + return 1; +} + +/* MemviewSliceValidateAndInit */ + static int +__pyx_check_strides(Py_buffer *buf, int dim, int ndim, int spec) +{ + if (buf->shape[dim] <= 1) + return 1; + if (buf->strides) { + if (spec & __Pyx_MEMVIEW_CONTIG) { + if (spec & (__Pyx_MEMVIEW_PTR|__Pyx_MEMVIEW_FULL)) { + if (unlikely(buf->strides[dim] != sizeof(void *))) { + PyErr_Format(PyExc_ValueError, + "Buffer is not indirectly contiguous " + "in dimension %d.", dim); + goto fail; + } + } else if (unlikely(buf->strides[dim] != buf->itemsize)) { + PyErr_SetString(PyExc_ValueError, + "Buffer and memoryview are not contiguous " + "in the same dimension."); + goto fail; + } + } + if (spec & __Pyx_MEMVIEW_FOLLOW) { + Py_ssize_t stride = buf->strides[dim]; + if (stride < 0) + stride = -stride; + if (unlikely(stride < buf->itemsize)) { + PyErr_SetString(PyExc_ValueError, + "Buffer and memoryview are not contiguous " + "in the same dimension."); + goto fail; + } + } + } else { + if (unlikely(spec & __Pyx_MEMVIEW_CONTIG && dim != ndim - 1)) { + PyErr_Format(PyExc_ValueError, + "C-contiguous buffer is not contiguous in " + "dimension %d", dim); + goto fail; + } else if (unlikely(spec & (__Pyx_MEMVIEW_PTR))) { + PyErr_Format(PyExc_ValueError, + "C-contiguous buffer is not indirect in " + "dimension %d", dim); + goto fail; + } else if (unlikely(buf->suboffsets)) { + PyErr_SetString(PyExc_ValueError, + "Buffer exposes suboffsets but no strides"); + goto fail; + } + } + return 1; +fail: + return 0; +} +static int +__pyx_check_suboffsets(Py_buffer *buf, int dim, int ndim, int spec) +{ + CYTHON_UNUSED_VAR(ndim); + if (spec & __Pyx_MEMVIEW_DIRECT) { + if (unlikely(buf->suboffsets && buf->suboffsets[dim] >= 0)) { + PyErr_Format(PyExc_ValueError, + "Buffer not compatible with direct access " + "in dimension %d.", dim); + goto fail; + } + } + if (spec & __Pyx_MEMVIEW_PTR) { + if (unlikely(!buf->suboffsets || (buf->suboffsets[dim] < 0))) { + PyErr_Format(PyExc_ValueError, + "Buffer is not indirectly accessible " + "in dimension %d.", dim); + goto fail; + } + } + return 1; +fail: + return 0; +} +static int +__pyx_verify_contig(Py_buffer *buf, int ndim, int c_or_f_flag) +{ + int i; + if (c_or_f_flag & __Pyx_IS_F_CONTIG) { + Py_ssize_t stride = 1; + for (i = 0; i < ndim; i++) { + if (unlikely(stride * buf->itemsize != buf->strides[i] && buf->shape[i] > 1)) { + PyErr_SetString(PyExc_ValueError, + "Buffer not fortran contiguous."); + goto fail; + } + stride = stride * buf->shape[i]; + } + } else if (c_or_f_flag & __Pyx_IS_C_CONTIG) { + Py_ssize_t stride = 1; + for (i = ndim - 1; i >- 1; i--) { + if (unlikely(stride * buf->itemsize != buf->strides[i] && buf->shape[i] > 1)) { + PyErr_SetString(PyExc_ValueError, + "Buffer not C contiguous."); + goto fail; + } + stride = stride * buf->shape[i]; + } + } + return 1; +fail: + return 0; +} +static int __Pyx_ValidateAndInit_memviewslice( + int *axes_specs, + int c_or_f_flag, + int buf_flags, + int ndim, + __Pyx_TypeInfo *dtype, + __Pyx_BufFmt_StackElem stack[], + __Pyx_memviewslice *memviewslice, + PyObject *original_obj) +{ + struct __pyx_memoryview_obj *memview, *new_memview; + __Pyx_RefNannyDeclarations + Py_buffer *buf; + int i, spec = 0, retval = -1; + __Pyx_BufFmt_Context ctx; + int from_memoryview = __pyx_memoryview_check(original_obj); + __Pyx_RefNannySetupContext("ValidateAndInit_memviewslice", 0); + if (from_memoryview && __pyx_typeinfo_cmp(dtype, ((struct __pyx_memoryview_obj *) + original_obj)->typeinfo)) { + memview = (struct __pyx_memoryview_obj *) original_obj; + new_memview = NULL; + } else { + memview = (struct __pyx_memoryview_obj *) __pyx_memoryview_new( + original_obj, buf_flags, 0, dtype); + new_memview = memview; + if (unlikely(!memview)) + goto fail; + } + buf = &memview->view; + if (unlikely(buf->ndim != ndim)) { + PyErr_Format(PyExc_ValueError, + "Buffer has wrong number of dimensions (expected %d, got %d)", + ndim, buf->ndim); + goto fail; + } + if (new_memview) { + __Pyx_BufFmt_Init(&ctx, stack, dtype); + if (unlikely(!__Pyx_BufFmt_CheckString(&ctx, buf->format))) goto fail; + } + if (unlikely((unsigned) buf->itemsize != dtype->size)) { + PyErr_Format(PyExc_ValueError, + "Item size of buffer (%" CYTHON_FORMAT_SSIZE_T "u byte%s) " + "does not match size of '%s' (%" CYTHON_FORMAT_SSIZE_T "u byte%s)", + buf->itemsize, + (buf->itemsize > 1) ? "s" : "", + dtype->name, + dtype->size, + (dtype->size > 1) ? "s" : ""); + goto fail; + } + if (buf->len > 0) { + for (i = 0; i < ndim; i++) { + spec = axes_specs[i]; + if (unlikely(!__pyx_check_strides(buf, i, ndim, spec))) + goto fail; + if (unlikely(!__pyx_check_suboffsets(buf, i, ndim, spec))) + goto fail; + } + if (unlikely(buf->strides && !__pyx_verify_contig(buf, ndim, c_or_f_flag))) + goto fail; + } + if (unlikely(__Pyx_init_memviewslice(memview, ndim, memviewslice, + new_memview != NULL) == -1)) { + goto fail; + } + retval = 0; + goto no_fail; +fail: + Py_XDECREF(new_memview); + retval = -1; +no_fail: + __Pyx_RefNannyFinishContext(); + return retval; +} + +/* ObjectToMemviewSlice */ + static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(PyObject *obj, int writable_flag) { + __Pyx_memviewslice result = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_BufFmt_StackElem stack[1]; + int axes_specs[] = { (__Pyx_MEMVIEW_DIRECT | __Pyx_MEMVIEW_STRIDED) }; + int retcode; + if (obj == Py_None) { + result.memview = (struct __pyx_memoryview_obj *) Py_None; + return result; + } + retcode = __Pyx_ValidateAndInit_memviewslice(axes_specs, 0, + PyBUF_RECORDS_RO | writable_flag, 1, + &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, stack, + &result, obj); + if (unlikely(retcode == -1)) + goto __pyx_fail; + return result; +__pyx_fail: + result.memview = NULL; + result.data = NULL; + return result; +} + +/* MemviewDtypeToObject */ + static CYTHON_INLINE PyObject *__pyx_memview_get_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(const char *itemp) { + return (PyObject *) __Pyx_PyInt_From_npy_int64(*(__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t *) itemp); +} +static CYTHON_INLINE int __pyx_memview_set_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(const char *itemp, PyObject *obj) { + __pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t value = __Pyx_PyInt_As_npy_int64(obj); + if (unlikely((value == ((npy_int64)-1)) && PyErr_Occurred())) + return 0; + *(__pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t *) itemp = value; + return 1; +} + +/* ObjectToMemviewSlice */ + static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t(PyObject *obj, int writable_flag) { + __Pyx_memviewslice result = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_BufFmt_StackElem stack[1]; + int axes_specs[] = { (__Pyx_MEMVIEW_DIRECT | __Pyx_MEMVIEW_STRIDED), (__Pyx_MEMVIEW_DIRECT | __Pyx_MEMVIEW_STRIDED) }; + int retcode; + if (obj == Py_None) { + result.memview = (struct __pyx_memoryview_obj *) Py_None; + return result; + } + retcode = __Pyx_ValidateAndInit_memviewslice(axes_specs, 0, + PyBUF_RECORDS_RO | writable_flag, 2, + &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_15data_utils_fast_DTYPE_t, stack, + &result, obj); + if (unlikely(retcode == -1)) + goto __pyx_fail; + return result; +__pyx_fail: + result.memview = NULL; + result.data = NULL; + return result; +} + +/* Declarations */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) { + return ::std::complex< float >(x, y); + } + #else + static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) { + return x + y*(__pyx_t_float_complex)_Complex_I; + } + #endif +#else + static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) { + __pyx_t_float_complex z; + z.real = x; + z.imag = y; + return z; + } +#endif + +/* Arithmetic */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) +#else + static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + return (a.real == b.real) && (a.imag == b.imag); + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + z.real = a.real + b.real; + z.imag = a.imag + b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + z.real = a.real - b.real; + z.imag = a.imag - b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + z.real = a.real * b.real - a.imag * b.imag; + z.imag = a.real * b.imag + a.imag * b.real; + return z; + } + #if 1 + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + if (b.imag == 0) { + return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real); + } else if (fabsf(b.real) >= fabsf(b.imag)) { + if (b.real == 0 && b.imag == 0) { + return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.imag); + } else { + float r = b.imag / b.real; + float s = (float)(1.0) / (b.real + b.imag * r); + return __pyx_t_float_complex_from_parts( + (a.real + a.imag * r) * s, (a.imag - a.real * r) * s); + } + } else { + float r = b.real / b.imag; + float s = (float)(1.0) / (b.imag + b.real * r); + return __pyx_t_float_complex_from_parts( + (a.real * r + a.imag) * s, (a.imag * r - a.real) * s); + } + } + #else + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + if (b.imag == 0) { + return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real); + } else { + float denom = b.real * b.real + b.imag * b.imag; + return __pyx_t_float_complex_from_parts( + (a.real * b.real + a.imag * b.imag) / denom, + (a.imag * b.real - a.real * b.imag) / denom); + } + } + #endif + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex a) { + __pyx_t_float_complex z; + z.real = -a.real; + z.imag = -a.imag; + return z; + } + static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex a) { + return (a.real == 0) && (a.imag == 0); + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex a) { + __pyx_t_float_complex z; + z.real = a.real; + z.imag = -a.imag; + return z; + } + #if 1 + static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex z) { + #if !defined(HAVE_HYPOT) || defined(_MSC_VER) + return sqrtf(z.real*z.real + z.imag*z.imag); + #else + return hypotf(z.real, z.imag); + #endif + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + float r, lnr, theta, z_r, z_theta; + if (b.imag == 0 && b.real == (int)b.real) { + if (b.real < 0) { + float denom = a.real * a.real + a.imag * a.imag; + a.real = a.real / denom; + a.imag = -a.imag / denom; + b.real = -b.real; + } + switch ((int)b.real) { + case 0: + z.real = 1; + z.imag = 0; + return z; + case 1: + return a; + case 2: + return __Pyx_c_prod_float(a, a); + case 3: + z = __Pyx_c_prod_float(a, a); + return __Pyx_c_prod_float(z, a); + case 4: + z = __Pyx_c_prod_float(a, a); + return __Pyx_c_prod_float(z, z); + } + } + if (a.imag == 0) { + if (a.real == 0) { + return a; + } else if ((b.imag == 0) && (a.real >= 0)) { + z.real = powf(a.real, b.real); + z.imag = 0; + return z; + } else if (a.real > 0) { + r = a.real; + theta = 0; + } else { + r = -a.real; + theta = atan2f(0.0, -1.0); + } + } else { + r = __Pyx_c_abs_float(a); + theta = atan2f(a.imag, a.real); + } + lnr = logf(r); + z_r = expf(lnr * b.real - theta * b.imag); + z_theta = theta * b.real + lnr * b.imag; + z.real = z_r * cosf(z_theta); + z.imag = z_r * sinf(z_theta); + return z; + } + #endif +#endif + +/* Declarations */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) { + return ::std::complex< double >(x, y); + } + #else + static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) { + return x + y*(__pyx_t_double_complex)_Complex_I; + } + #endif +#else + static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) { + __pyx_t_double_complex z; + z.real = x; + z.imag = y; + return z; + } +#endif + +/* Arithmetic */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) +#else + static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + return (a.real == b.real) && (a.imag == b.imag); + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + z.real = a.real + b.real; + z.imag = a.imag + b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + z.real = a.real - b.real; + z.imag = a.imag - b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + z.real = a.real * b.real - a.imag * b.imag; + z.imag = a.real * b.imag + a.imag * b.real; + return z; + } + #if 1 + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + if (b.imag == 0) { + return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else if (fabs(b.real) >= fabs(b.imag)) { + if (b.real == 0 && b.imag == 0) { + return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.imag); + } else { + double r = b.imag / b.real; + double s = (double)(1.0) / (b.real + b.imag * r); + return __pyx_t_double_complex_from_parts( + (a.real + a.imag * r) * s, (a.imag - a.real * r) * s); + } + } else { + double r = b.real / b.imag; + double s = (double)(1.0) / (b.imag + b.real * r); + return __pyx_t_double_complex_from_parts( + (a.real * r + a.imag) * s, (a.imag * r - a.real) * s); + } + } + #else + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + if (b.imag == 0) { + return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else { + double denom = b.real * b.real + b.imag * b.imag; + return __pyx_t_double_complex_from_parts( + (a.real * b.real + a.imag * b.imag) / denom, + (a.imag * b.real - a.real * b.imag) / denom); + } + } + #endif + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex a) { + __pyx_t_double_complex z; + z.real = -a.real; + z.imag = -a.imag; + return z; + } + static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex a) { + return (a.real == 0) && (a.imag == 0); + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex a) { + __pyx_t_double_complex z; + z.real = a.real; + z.imag = -a.imag; + return z; + } + #if 1 + static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex z) { + #if !defined(HAVE_HYPOT) || defined(_MSC_VER) + return sqrt(z.real*z.real + z.imag*z.imag); + #else + return hypot(z.real, z.imag); + #endif + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + double r, lnr, theta, z_r, z_theta; + if (b.imag == 0 && b.real == (int)b.real) { + if (b.real < 0) { + double denom = a.real * a.real + a.imag * a.imag; + a.real = a.real / denom; + a.imag = -a.imag / denom; + b.real = -b.real; + } + switch ((int)b.real) { + case 0: + z.real = 1; + z.imag = 0; + return z; + case 1: + return a; + case 2: + return __Pyx_c_prod_double(a, a); + case 3: + z = __Pyx_c_prod_double(a, a); + return __Pyx_c_prod_double(z, a); + case 4: + z = __Pyx_c_prod_double(a, a); + return __Pyx_c_prod_double(z, z); + } + } + if (a.imag == 0) { + if (a.real == 0) { + return a; + } else if ((b.imag == 0) && (a.real >= 0)) { + z.real = pow(a.real, b.real); + z.imag = 0; + return z; + } else if (a.real > 0) { + r = a.real; + theta = 0; + } else { + r = -a.real; + theta = atan2(0.0, -1.0); + } + } else { + r = __Pyx_c_abs_double(a); + theta = atan2(a.imag, a.real); + } + lnr = log(r); + z_r = exp(lnr * b.real - theta * b.imag); + z_theta = theta * b.real + lnr * b.imag; + z.real = z_r * cos(z_theta); + z.imag = z_r * sin(z_theta); + return z; + } + #endif +#endif + +/* Declarations */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double x, long double y) { + return ::std::complex< long double >(x, y); + } + #else + static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double x, long double y) { + return x + y*(__pyx_t_long_double_complex)_Complex_I; + } + #endif +#else + static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double x, long double y) { + __pyx_t_long_double_complex z; + z.real = x; + z.imag = y; + return z; + } +#endif + +/* Arithmetic */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) +#else + static CYTHON_INLINE int __Pyx_c_eq_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + return (a.real == b.real) && (a.imag == b.imag); + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_sum_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + z.real = a.real + b.real; + z.imag = a.imag + b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_diff_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + z.real = a.real - b.real; + z.imag = a.imag - b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_prod_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + z.real = a.real * b.real - a.imag * b.imag; + z.imag = a.real * b.imag + a.imag * b.real; + return z; + } + #if 1 + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_quot_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + if (b.imag == 0) { + return __pyx_t_long_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else if (fabsl(b.real) >= fabsl(b.imag)) { + if (b.real == 0 && b.imag == 0) { + return __pyx_t_long_double_complex_from_parts(a.real / b.real, a.imag / b.imag); + } else { + long double r = b.imag / b.real; + long double s = (long double)(1.0) / (b.real + b.imag * r); + return __pyx_t_long_double_complex_from_parts( + (a.real + a.imag * r) * s, (a.imag - a.real * r) * s); + } + } else { + long double r = b.real / b.imag; + long double s = (long double)(1.0) / (b.imag + b.real * r); + return __pyx_t_long_double_complex_from_parts( + (a.real * r + a.imag) * s, (a.imag * r - a.real) * s); + } + } + #else + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_quot_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + if (b.imag == 0) { + return __pyx_t_long_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else { + long double denom = b.real * b.real + b.imag * b.imag; + return __pyx_t_long_double_complex_from_parts( + (a.real * b.real + a.imag * b.imag) / denom, + (a.imag * b.real - a.real * b.imag) / denom); + } + } + #endif + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_neg_long__double(__pyx_t_long_double_complex a) { + __pyx_t_long_double_complex z; + z.real = -a.real; + z.imag = -a.imag; + return z; + } + static CYTHON_INLINE int __Pyx_c_is_zero_long__double(__pyx_t_long_double_complex a) { + return (a.real == 0) && (a.imag == 0); + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_conj_long__double(__pyx_t_long_double_complex a) { + __pyx_t_long_double_complex z; + z.real = a.real; + z.imag = -a.imag; + return z; + } + #if 1 + static CYTHON_INLINE long double __Pyx_c_abs_long__double(__pyx_t_long_double_complex z) { + #if !defined(HAVE_HYPOT) || defined(_MSC_VER) + return sqrtl(z.real*z.real + z.imag*z.imag); + #else + return hypotl(z.real, z.imag); + #endif + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_pow_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + long double r, lnr, theta, z_r, z_theta; + if (b.imag == 0 && b.real == (int)b.real) { + if (b.real < 0) { + long double denom = a.real * a.real + a.imag * a.imag; + a.real = a.real / denom; + a.imag = -a.imag / denom; + b.real = -b.real; + } + switch ((int)b.real) { + case 0: + z.real = 1; + z.imag = 0; + return z; + case 1: + return a; + case 2: + return __Pyx_c_prod_long__double(a, a); + case 3: + z = __Pyx_c_prod_long__double(a, a); + return __Pyx_c_prod_long__double(z, a); + case 4: + z = __Pyx_c_prod_long__double(a, a); + return __Pyx_c_prod_long__double(z, z); + } + } + if (a.imag == 0) { + if (a.real == 0) { + return a; + } else if ((b.imag == 0) && (a.real >= 0)) { + z.real = powl(a.real, b.real); + z.imag = 0; + return z; + } else if (a.real > 0) { + r = a.real; + theta = 0; + } else { + r = -a.real; + theta = atan2l(0.0, -1.0); + } + } else { + r = __Pyx_c_abs_long__double(a); + theta = atan2l(a.imag, a.real); + } + lnr = logl(r); + z_r = expl(lnr * b.real - theta * b.imag); + z_theta = theta * b.real + lnr * b.imag; + z.real = z_r * cosl(z_theta); + z.imag = z_r * sinl(z_theta); + return z; + } + #endif +#endif + +/* MemviewSliceCopyTemplate */ + static __Pyx_memviewslice +__pyx_memoryview_copy_new_contig(const __Pyx_memviewslice *from_mvs, + const char *mode, int ndim, + size_t sizeof_dtype, int contig_flag, + int dtype_is_object) +{ + __Pyx_RefNannyDeclarations + int i; + __Pyx_memviewslice new_mvs = { 0, 0, { 0 }, { 0 }, { 0 } }; + struct __pyx_memoryview_obj *from_memview = from_mvs->memview; + Py_buffer *buf = &from_memview->view; + PyObject *shape_tuple = NULL; + PyObject *temp_int = NULL; + struct __pyx_array_obj *array_obj = NULL; + struct __pyx_memoryview_obj *memview_obj = NULL; + __Pyx_RefNannySetupContext("__pyx_memoryview_copy_new_contig", 0); + for (i = 0; i < ndim; i++) { + if (unlikely(from_mvs->suboffsets[i] >= 0)) { + PyErr_Format(PyExc_ValueError, "Cannot copy memoryview slice with " + "indirect dimensions (axis %d)", i); + goto fail; + } + } + shape_tuple = PyTuple_New(ndim); + if (unlikely(!shape_tuple)) { + goto fail; + } + __Pyx_GOTREF(shape_tuple); + for(i = 0; i < ndim; i++) { + temp_int = PyInt_FromSsize_t(from_mvs->shape[i]); + if(unlikely(!temp_int)) { + goto fail; + } else { + PyTuple_SET_ITEM(shape_tuple, i, temp_int); + temp_int = NULL; + } + } + array_obj = __pyx_array_new(shape_tuple, sizeof_dtype, buf->format, (char *) mode, NULL); + if (unlikely(!array_obj)) { + goto fail; + } + __Pyx_GOTREF(array_obj); + memview_obj = (struct __pyx_memoryview_obj *) __pyx_memoryview_new( + (PyObject *) array_obj, contig_flag, + dtype_is_object, + from_mvs->memview->typeinfo); + if (unlikely(!memview_obj)) + goto fail; + if (unlikely(__Pyx_init_memviewslice(memview_obj, ndim, &new_mvs, 1) < 0)) + goto fail; + if (unlikely(__pyx_memoryview_copy_contents(*from_mvs, new_mvs, ndim, ndim, + dtype_is_object) < 0)) + goto fail; + goto no_fail; +fail: + __Pyx_XDECREF(new_mvs.memview); + new_mvs.memview = NULL; + new_mvs.data = NULL; +no_fail: + __Pyx_XDECREF(shape_tuple); + __Pyx_XDECREF(temp_int); + __Pyx_XDECREF(array_obj); + __Pyx_RefNannyFinishContext(); + return new_mvs; +} + +/* MemviewSliceInit */ + static int +__Pyx_init_memviewslice(struct __pyx_memoryview_obj *memview, + int ndim, + __Pyx_memviewslice *memviewslice, + int memview_is_new_reference) +{ + __Pyx_RefNannyDeclarations + int i, retval=-1; + Py_buffer *buf = &memview->view; + __Pyx_RefNannySetupContext("init_memviewslice", 0); + if (unlikely(memviewslice->memview || memviewslice->data)) { + PyErr_SetString(PyExc_ValueError, + "memviewslice is already initialized!"); + goto fail; + } + if (buf->strides) { + for (i = 0; i < ndim; i++) { + memviewslice->strides[i] = buf->strides[i]; + } + } else { + Py_ssize_t stride = buf->itemsize; + for (i = ndim - 1; i >= 0; i--) { + memviewslice->strides[i] = stride; + stride *= buf->shape[i]; + } + } + for (i = 0; i < ndim; i++) { + memviewslice->shape[i] = buf->shape[i]; + if (buf->suboffsets) { + memviewslice->suboffsets[i] = buf->suboffsets[i]; + } else { + memviewslice->suboffsets[i] = -1; + } + } + memviewslice->memview = memview; + memviewslice->data = (char *)buf->buf; + if (__pyx_add_acquisition_count(memview) == 0 && !memview_is_new_reference) { + Py_INCREF(memview); + } + retval = 0; + goto no_fail; +fail: + memviewslice->memview = 0; + memviewslice->data = 0; + retval = -1; +no_fail: + __Pyx_RefNannyFinishContext(); + return retval; +} +#ifndef Py_NO_RETURN +#define Py_NO_RETURN +#endif +static void __pyx_fatalerror(const char *fmt, ...) Py_NO_RETURN { + va_list vargs; + char msg[200]; +#if PY_VERSION_HEX >= 0x030A0000 || defined(HAVE_STDARG_PROTOTYPES) + va_start(vargs, fmt); +#else + va_start(vargs); +#endif + vsnprintf(msg, 200, fmt, vargs); + va_end(vargs); + Py_FatalError(msg); +} +static CYTHON_INLINE int +__pyx_add_acquisition_count_locked(__pyx_atomic_int_type *acquisition_count, + PyThread_type_lock lock) +{ + int result; + PyThread_acquire_lock(lock, 1); + result = (*acquisition_count)++; + PyThread_release_lock(lock); + return result; +} +static CYTHON_INLINE int +__pyx_sub_acquisition_count_locked(__pyx_atomic_int_type *acquisition_count, + PyThread_type_lock lock) +{ + int result; + PyThread_acquire_lock(lock, 1); + result = (*acquisition_count)--; + PyThread_release_lock(lock); + return result; +} +static CYTHON_INLINE void +__Pyx_INC_MEMVIEW(__Pyx_memviewslice *memslice, int have_gil, int lineno) +{ + __pyx_nonatomic_int_type old_acquisition_count; + struct __pyx_memoryview_obj *memview = memslice->memview; + if (unlikely(!memview || (PyObject *) memview == Py_None)) { + return; + } + old_acquisition_count = __pyx_add_acquisition_count(memview); + if (unlikely(old_acquisition_count <= 0)) { + if (likely(old_acquisition_count == 0)) { + if (have_gil) { + Py_INCREF((PyObject *) memview); + } else { + PyGILState_STATE _gilstate = PyGILState_Ensure(); + Py_INCREF((PyObject *) memview); + PyGILState_Release(_gilstate); + } + } else { + __pyx_fatalerror("Acquisition count is %d (line %d)", + old_acquisition_count+1, lineno); + } + } +} +static CYTHON_INLINE void __Pyx_XCLEAR_MEMVIEW(__Pyx_memviewslice *memslice, + int have_gil, int lineno) { + __pyx_nonatomic_int_type old_acquisition_count; + struct __pyx_memoryview_obj *memview = memslice->memview; + if (unlikely(!memview || (PyObject *) memview == Py_None)) { + memslice->memview = NULL; + return; + } + old_acquisition_count = __pyx_sub_acquisition_count(memview); + memslice->data = NULL; + if (likely(old_acquisition_count > 1)) { + memslice->memview = NULL; + } else if (likely(old_acquisition_count == 1)) { + if (have_gil) { + Py_CLEAR(memslice->memview); + } else { + PyGILState_STATE _gilstate = PyGILState_Ensure(); + Py_CLEAR(memslice->memview); + PyGILState_Release(_gilstate); + } + } else { + __pyx_fatalerror("Acquisition count is %d (line %d)", + old_acquisition_count-1, lineno); + } +} + +/* CIntFromPy */ + static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const long neg_one = (long) -1, const_zero = (long) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(long) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(long, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (long) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + long val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (long) -1; + val = __Pyx_PyInt_As_long(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(long, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(long) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) >= 2 * PyLong_SHIFT)) { + return (long) (((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(long) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) >= 3 * PyLong_SHIFT)) { + return (long) (((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(long) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) >= 4 * PyLong_SHIFT)) { + return (long) (((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (long) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(long) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(long, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(long) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(long, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(long, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(long) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 2 * PyLong_SHIFT)) { + return (long) (((long)-1)*(((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(long) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 2 * PyLong_SHIFT)) { + return (long) ((((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(long) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 3 * PyLong_SHIFT)) { + return (long) (((long)-1)*(((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(long) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 3 * PyLong_SHIFT)) { + return (long) ((((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(long) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 4 * PyLong_SHIFT)) { + return (long) (((long)-1)*(((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(long) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 4 * PyLong_SHIFT)) { + return (long) ((((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(long) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(long, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(long) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(long, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + long val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (long) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (long) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (long) -1; + } else { + stepval = v; + } + v = NULL; + val = (long) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(long) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((long) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(long) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((long) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((long) 1) << (sizeof(long) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (long) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to long"); + return (long) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to long"); + return (long) -1; +} + +/* CIntFromPy */ + static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const int neg_one = (int) -1, const_zero = (int) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(int) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(int, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (int) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + int val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (int) -1; + val = __Pyx_PyInt_As_int(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(int, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(int) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) >= 2 * PyLong_SHIFT)) { + return (int) (((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(int) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) >= 3 * PyLong_SHIFT)) { + return (int) (((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(int) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) >= 4 * PyLong_SHIFT)) { + return (int) (((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (int) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(int) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(int, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(int) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(int, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(int, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(int) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 2 * PyLong_SHIFT)) { + return (int) (((int)-1)*(((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(int) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 2 * PyLong_SHIFT)) { + return (int) ((((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(int) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 3 * PyLong_SHIFT)) { + return (int) (((int)-1)*(((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(int) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 3 * PyLong_SHIFT)) { + return (int) ((((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(int) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 4 * PyLong_SHIFT)) { + return (int) (((int)-1)*(((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(int) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 4 * PyLong_SHIFT)) { + return (int) ((((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(int) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(int, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(int) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(int, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + int val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (int) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (int) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (int) -1; + } else { + stepval = v; + } + v = NULL; + val = (int) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(int) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((int) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(int) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((int) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((int) 1) << (sizeof(int) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (int) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to int"); + return (int) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to int"); + return (int) -1; +} + +/* CIntToPy */ + static CYTHON_INLINE PyObject* __Pyx_PyInt_From_npy_int64(npy_int64 value) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const npy_int64 neg_one = (npy_int64) -1, const_zero = (npy_int64) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (is_unsigned) { + if (sizeof(npy_int64) < sizeof(long)) { + return PyInt_FromLong((long) value); + } else if (sizeof(npy_int64) <= sizeof(unsigned long)) { + return PyLong_FromUnsignedLong((unsigned long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(npy_int64) <= sizeof(unsigned PY_LONG_LONG)) { + return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value); +#endif + } + } else { + if (sizeof(npy_int64) <= sizeof(long)) { + return PyInt_FromLong((long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(npy_int64) <= sizeof(PY_LONG_LONG)) { + return PyLong_FromLongLong((PY_LONG_LONG) value); +#endif + } + } + { + unsigned char *bytes = (unsigned char *)&value; +#if !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX >= 0x030d00A4 + if (is_unsigned) { + return PyLong_FromUnsignedNativeBytes(bytes, sizeof(value), -1); + } else { + return PyLong_FromNativeBytes(bytes, sizeof(value), -1); + } +#elif !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX < 0x030d0000 + int one = 1; int little = (int)*(unsigned char *)&one; + return _PyLong_FromByteArray(bytes, sizeof(npy_int64), + little, !is_unsigned); +#else + int one = 1; int little = (int)*(unsigned char *)&one; + PyObject *from_bytes, *result = NULL; + PyObject *py_bytes = NULL, *arg_tuple = NULL, *kwds = NULL, *order_str = NULL; + from_bytes = PyObject_GetAttrString((PyObject*)&PyLong_Type, "from_bytes"); + if (!from_bytes) return NULL; + py_bytes = PyBytes_FromStringAndSize((char*)bytes, sizeof(npy_int64)); + if (!py_bytes) goto limited_bad; + order_str = PyUnicode_FromString(little ? "little" : "big"); + if (!order_str) goto limited_bad; + arg_tuple = PyTuple_Pack(2, py_bytes, order_str); + if (!arg_tuple) goto limited_bad; + if (!is_unsigned) { + kwds = PyDict_New(); + if (!kwds) goto limited_bad; + if (PyDict_SetItemString(kwds, "signed", __Pyx_NewRef(Py_True))) goto limited_bad; + } + result = PyObject_Call(from_bytes, arg_tuple, kwds); + limited_bad: + Py_XDECREF(kwds); + Py_XDECREF(arg_tuple); + Py_XDECREF(order_str); + Py_XDECREF(py_bytes); + Py_XDECREF(from_bytes); + return result; +#endif + } +} + +/* CIntFromPy */ + static CYTHON_INLINE npy_int64 __Pyx_PyInt_As_npy_int64(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const npy_int64 neg_one = (npy_int64) -1, const_zero = (npy_int64) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(npy_int64) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (npy_int64) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + npy_int64 val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (npy_int64) -1; + val = __Pyx_PyInt_As_npy_int64(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(npy_int64, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(npy_int64) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) >= 2 * PyLong_SHIFT)) { + return (npy_int64) (((((npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(npy_int64) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) >= 3 * PyLong_SHIFT)) { + return (npy_int64) (((((((npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(npy_int64) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) >= 4 * PyLong_SHIFT)) { + return (npy_int64) (((((((((npy_int64)digits[3]) << PyLong_SHIFT) | (npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (npy_int64) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(npy_int64) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(npy_int64) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(npy_int64, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(npy_int64) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 2 * PyLong_SHIFT)) { + return (npy_int64) (((npy_int64)-1)*(((((npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(npy_int64) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 2 * PyLong_SHIFT)) { + return (npy_int64) ((((((npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(npy_int64) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 3 * PyLong_SHIFT)) { + return (npy_int64) (((npy_int64)-1)*(((((((npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(npy_int64) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 3 * PyLong_SHIFT)) { + return (npy_int64) ((((((((npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(npy_int64) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 4 * PyLong_SHIFT)) { + return (npy_int64) (((npy_int64)-1)*(((((((((npy_int64)digits[3]) << PyLong_SHIFT) | (npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(npy_int64) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 4 * PyLong_SHIFT)) { + return (npy_int64) ((((((((((npy_int64)digits[3]) << PyLong_SHIFT) | (npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(npy_int64) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(npy_int64) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + npy_int64 val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (npy_int64) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (npy_int64) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (npy_int64) -1; + } else { + stepval = v; + } + v = NULL; + val = (npy_int64) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(npy_int64) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((npy_int64) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(npy_int64) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((npy_int64) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((npy_int64) 1) << (sizeof(npy_int64) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (npy_int64) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to npy_int64"); + return (npy_int64) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to npy_int64"); + return (npy_int64) -1; +} + +/* CIntToPy */ + static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const long neg_one = (long) -1, const_zero = (long) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (is_unsigned) { + if (sizeof(long) < sizeof(long)) { + return PyInt_FromLong((long) value); + } else if (sizeof(long) <= sizeof(unsigned long)) { + return PyLong_FromUnsignedLong((unsigned long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) { + return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value); +#endif + } + } else { + if (sizeof(long) <= sizeof(long)) { + return PyInt_FromLong((long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) { + return PyLong_FromLongLong((PY_LONG_LONG) value); +#endif + } + } + { + unsigned char *bytes = (unsigned char *)&value; +#if !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX >= 0x030d00A4 + if (is_unsigned) { + return PyLong_FromUnsignedNativeBytes(bytes, sizeof(value), -1); + } else { + return PyLong_FromNativeBytes(bytes, sizeof(value), -1); + } +#elif !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX < 0x030d0000 + int one = 1; int little = (int)*(unsigned char *)&one; + return _PyLong_FromByteArray(bytes, sizeof(long), + little, !is_unsigned); +#else + int one = 1; int little = (int)*(unsigned char *)&one; + PyObject *from_bytes, *result = NULL; + PyObject *py_bytes = NULL, *arg_tuple = NULL, *kwds = NULL, *order_str = NULL; + from_bytes = PyObject_GetAttrString((PyObject*)&PyLong_Type, "from_bytes"); + if (!from_bytes) return NULL; + py_bytes = PyBytes_FromStringAndSize((char*)bytes, sizeof(long)); + if (!py_bytes) goto limited_bad; + order_str = PyUnicode_FromString(little ? "little" : "big"); + if (!order_str) goto limited_bad; + arg_tuple = PyTuple_Pack(2, py_bytes, order_str); + if (!arg_tuple) goto limited_bad; + if (!is_unsigned) { + kwds = PyDict_New(); + if (!kwds) goto limited_bad; + if (PyDict_SetItemString(kwds, "signed", __Pyx_NewRef(Py_True))) goto limited_bad; + } + result = PyObject_Call(from_bytes, arg_tuple, kwds); + limited_bad: + Py_XDECREF(kwds); + Py_XDECREF(arg_tuple); + Py_XDECREF(order_str); + Py_XDECREF(py_bytes); + Py_XDECREF(from_bytes); + return result; +#endif + } +} + +/* CIntToPy */ + static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const int neg_one = (int) -1, const_zero = (int) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (is_unsigned) { + if (sizeof(int) < sizeof(long)) { + return PyInt_FromLong((long) value); + } else if (sizeof(int) <= sizeof(unsigned long)) { + return PyLong_FromUnsignedLong((unsigned long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) { + return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value); +#endif + } + } else { + if (sizeof(int) <= sizeof(long)) { + return PyInt_FromLong((long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) { + return PyLong_FromLongLong((PY_LONG_LONG) value); +#endif + } + } + { + unsigned char *bytes = (unsigned char *)&value; +#if !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX >= 0x030d00A4 + if (is_unsigned) { + return PyLong_FromUnsignedNativeBytes(bytes, sizeof(value), -1); + } else { + return PyLong_FromNativeBytes(bytes, sizeof(value), -1); + } +#elif !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX < 0x030d0000 + int one = 1; int little = (int)*(unsigned char *)&one; + return _PyLong_FromByteArray(bytes, sizeof(int), + little, !is_unsigned); +#else + int one = 1; int little = (int)*(unsigned char *)&one; + PyObject *from_bytes, *result = NULL; + PyObject *py_bytes = NULL, *arg_tuple = NULL, *kwds = NULL, *order_str = NULL; + from_bytes = PyObject_GetAttrString((PyObject*)&PyLong_Type, "from_bytes"); + if (!from_bytes) return NULL; + py_bytes = PyBytes_FromStringAndSize((char*)bytes, sizeof(int)); + if (!py_bytes) goto limited_bad; + order_str = PyUnicode_FromString(little ? "little" : "big"); + if (!order_str) goto limited_bad; + arg_tuple = PyTuple_Pack(2, py_bytes, order_str); + if (!arg_tuple) goto limited_bad; + if (!is_unsigned) { + kwds = PyDict_New(); + if (!kwds) goto limited_bad; + if (PyDict_SetItemString(kwds, "signed", __Pyx_NewRef(Py_True))) goto limited_bad; + } + result = PyObject_Call(from_bytes, arg_tuple, kwds); + limited_bad: + Py_XDECREF(kwds); + Py_XDECREF(arg_tuple); + Py_XDECREF(order_str); + Py_XDECREF(py_bytes); + Py_XDECREF(from_bytes); + return result; +#endif + } +} + +/* CIntFromPy */ + static CYTHON_INLINE char __Pyx_PyInt_As_char(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const char neg_one = (char) -1, const_zero = (char) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(char) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(char, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (char) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + char val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (char) -1; + val = __Pyx_PyInt_As_char(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(char, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(char) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) >= 2 * PyLong_SHIFT)) { + return (char) (((((char)digits[1]) << PyLong_SHIFT) | (char)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(char) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) >= 3 * PyLong_SHIFT)) { + return (char) (((((((char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(char) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) >= 4 * PyLong_SHIFT)) { + return (char) (((((((((char)digits[3]) << PyLong_SHIFT) | (char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (char) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(char) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(char, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(char) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(char, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(char, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(char) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 2 * PyLong_SHIFT)) { + return (char) (((char)-1)*(((((char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(char) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 2 * PyLong_SHIFT)) { + return (char) ((((((char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(char) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 3 * PyLong_SHIFT)) { + return (char) (((char)-1)*(((((((char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(char) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 3 * PyLong_SHIFT)) { + return (char) ((((((((char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(char) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 4 * PyLong_SHIFT)) { + return (char) (((char)-1)*(((((((((char)digits[3]) << PyLong_SHIFT) | (char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(char) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 4 * PyLong_SHIFT)) { + return (char) ((((((((((char)digits[3]) << PyLong_SHIFT) | (char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(char) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(char, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(char) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(char, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + char val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (char) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (char) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (char) -1; + } else { + stepval = v; + } + v = NULL; + val = (char) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(char) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((char) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(char) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((char) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((char) 1) << (sizeof(char) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (char) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to char"); + return (char) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to char"); + return (char) -1; +} + +/* FormatTypeName */ + #if CYTHON_COMPILING_IN_LIMITED_API +static __Pyx_TypeName +__Pyx_PyType_GetName(PyTypeObject* tp) +{ + PyObject *name = __Pyx_PyObject_GetAttrStr((PyObject *)tp, + __pyx_n_s_name_2); + if (unlikely(name == NULL) || unlikely(!PyUnicode_Check(name))) { + PyErr_Clear(); + Py_XDECREF(name); + name = __Pyx_NewRef(__pyx_n_s__26); + } + return name; +} +#endif + +/* CheckBinaryVersion */ + static unsigned long __Pyx_get_runtime_version(void) { +#if __PYX_LIMITED_VERSION_HEX >= 0x030B00A4 + return Py_Version & ~0xFFUL; +#else + const char* rt_version = Py_GetVersion(); + unsigned long version = 0; + unsigned long factor = 0x01000000UL; + unsigned int digit = 0; + int i = 0; + while (factor) { + while ('0' <= rt_version[i] && rt_version[i] <= '9') { + digit = digit * 10 + (unsigned int) (rt_version[i] - '0'); + ++i; + } + version += factor * digit; + if (rt_version[i] != '.') + break; + digit = 0; + factor >>= 8; + ++i; + } + return version; +#endif +} +static int __Pyx_check_binary_version(unsigned long ct_version, unsigned long rt_version, int allow_newer) { + const unsigned long MAJOR_MINOR = 0xFFFF0000UL; + if ((rt_version & MAJOR_MINOR) == (ct_version & MAJOR_MINOR)) + return 0; + if (likely(allow_newer && (rt_version & MAJOR_MINOR) > (ct_version & MAJOR_MINOR))) + return 1; + { + char message[200]; + PyOS_snprintf(message, sizeof(message), + "compile time Python version %d.%d " + "of module '%.100s' " + "%s " + "runtime version %d.%d", + (int) (ct_version >> 24), (int) ((ct_version >> 16) & 0xFF), + __Pyx_MODULE_NAME, + (allow_newer) ? "was newer than" : "does not match", + (int) (rt_version >> 24), (int) ((rt_version >> 16) & 0xFF) + ); + return PyErr_WarnEx(NULL, message, 1); + } +} + +/* InitStrings */ + #if PY_MAJOR_VERSION >= 3 +static int __Pyx_InitString(__Pyx_StringTabEntry t, PyObject **str) { + if (t.is_unicode | t.is_str) { + if (t.intern) { + *str = PyUnicode_InternFromString(t.s); + } else if (t.encoding) { + *str = PyUnicode_Decode(t.s, t.n - 1, t.encoding, NULL); + } else { + *str = PyUnicode_FromStringAndSize(t.s, t.n - 1); + } + } else { + *str = PyBytes_FromStringAndSize(t.s, t.n - 1); + } + if (!*str) + return -1; + if (PyObject_Hash(*str) == -1) + return -1; + return 0; +} +#endif +static int __Pyx_InitStrings(__Pyx_StringTabEntry *t) { + while (t->p) { + #if PY_MAJOR_VERSION >= 3 + __Pyx_InitString(*t, t->p); + #else + if (t->is_unicode) { + *t->p = PyUnicode_DecodeUTF8(t->s, t->n - 1, NULL); + } else if (t->intern) { + *t->p = PyString_InternFromString(t->s); + } else { + *t->p = PyString_FromStringAndSize(t->s, t->n - 1); + } + if (!*t->p) + return -1; + if (PyObject_Hash(*t->p) == -1) + return -1; + #endif + ++t; + } + return 0; +} + +#include +static CYTHON_INLINE Py_ssize_t __Pyx_ssize_strlen(const char *s) { + size_t len = strlen(s); + if (unlikely(len > (size_t) PY_SSIZE_T_MAX)) { + PyErr_SetString(PyExc_OverflowError, "byte string is too long"); + return -1; + } + return (Py_ssize_t) len; +} +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char* c_str) { + Py_ssize_t len = __Pyx_ssize_strlen(c_str); + if (unlikely(len < 0)) return NULL; + return __Pyx_PyUnicode_FromStringAndSize(c_str, len); +} +static CYTHON_INLINE PyObject* __Pyx_PyByteArray_FromString(const char* c_str) { + Py_ssize_t len = __Pyx_ssize_strlen(c_str); + if (unlikely(len < 0)) return NULL; + return PyByteArray_FromStringAndSize(c_str, len); +} +static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject* o) { + Py_ssize_t ignore; + return __Pyx_PyObject_AsStringAndSize(o, &ignore); +} +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT +#if !CYTHON_PEP393_ENABLED +static const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) { + char* defenc_c; + PyObject* defenc = _PyUnicode_AsDefaultEncodedString(o, NULL); + if (!defenc) return NULL; + defenc_c = PyBytes_AS_STRING(defenc); +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII + { + char* end = defenc_c + PyBytes_GET_SIZE(defenc); + char* c; + for (c = defenc_c; c < end; c++) { + if ((unsigned char) (*c) >= 128) { + PyUnicode_AsASCIIString(o); + return NULL; + } + } + } +#endif + *length = PyBytes_GET_SIZE(defenc); + return defenc_c; +} +#else +static CYTHON_INLINE const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) { + if (unlikely(__Pyx_PyUnicode_READY(o) == -1)) return NULL; +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII + if (likely(PyUnicode_IS_ASCII(o))) { + *length = PyUnicode_GET_LENGTH(o); + return PyUnicode_AsUTF8(o); + } else { + PyUnicode_AsASCIIString(o); + return NULL; + } +#else + return PyUnicode_AsUTF8AndSize(o, length); +#endif +} +#endif +#endif +static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject* o, Py_ssize_t *length) { +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT + if ( +#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII + __Pyx_sys_getdefaultencoding_not_ascii && +#endif + PyUnicode_Check(o)) { + return __Pyx_PyUnicode_AsStringAndSize(o, length); + } else +#endif +#if (!CYTHON_COMPILING_IN_PYPY && !CYTHON_COMPILING_IN_LIMITED_API) || (defined(PyByteArray_AS_STRING) && defined(PyByteArray_GET_SIZE)) + if (PyByteArray_Check(o)) { + *length = PyByteArray_GET_SIZE(o); + return PyByteArray_AS_STRING(o); + } else +#endif + { + char* result; + int r = PyBytes_AsStringAndSize(o, &result, length); + if (unlikely(r < 0)) { + return NULL; + } else { + return result; + } + } +} +static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject* x) { + int is_true = x == Py_True; + if (is_true | (x == Py_False) | (x == Py_None)) return is_true; + else return PyObject_IsTrue(x); +} +static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject* x) { + int retval; + if (unlikely(!x)) return -1; + retval = __Pyx_PyObject_IsTrue(x); + Py_DECREF(x); + return retval; +} +static PyObject* __Pyx_PyNumber_IntOrLongWrongResultType(PyObject* result, const char* type_name) { + __Pyx_TypeName result_type_name = __Pyx_PyType_GetName(Py_TYPE(result)); +#if PY_MAJOR_VERSION >= 3 + if (PyLong_Check(result)) { + if (PyErr_WarnFormat(PyExc_DeprecationWarning, 1, + "__int__ returned non-int (type " __Pyx_FMT_TYPENAME "). " + "The ability to return an instance of a strict subclass of int is deprecated, " + "and may be removed in a future version of Python.", + result_type_name)) { + __Pyx_DECREF_TypeName(result_type_name); + Py_DECREF(result); + return NULL; + } + __Pyx_DECREF_TypeName(result_type_name); + return result; + } +#endif + PyErr_Format(PyExc_TypeError, + "__%.4s__ returned non-%.4s (type " __Pyx_FMT_TYPENAME ")", + type_name, type_name, result_type_name); + __Pyx_DECREF_TypeName(result_type_name); + Py_DECREF(result); + return NULL; +} +static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x) { +#if CYTHON_USE_TYPE_SLOTS + PyNumberMethods *m; +#endif + const char *name = NULL; + PyObject *res = NULL; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x) || PyLong_Check(x))) +#else + if (likely(PyLong_Check(x))) +#endif + return __Pyx_NewRef(x); +#if CYTHON_USE_TYPE_SLOTS + m = Py_TYPE(x)->tp_as_number; + #if PY_MAJOR_VERSION < 3 + if (m && m->nb_int) { + name = "int"; + res = m->nb_int(x); + } + else if (m && m->nb_long) { + name = "long"; + res = m->nb_long(x); + } + #else + if (likely(m && m->nb_int)) { + name = "int"; + res = m->nb_int(x); + } + #endif +#else + if (!PyBytes_CheckExact(x) && !PyUnicode_CheckExact(x)) { + res = PyNumber_Int(x); + } +#endif + if (likely(res)) { +#if PY_MAJOR_VERSION < 3 + if (unlikely(!PyInt_Check(res) && !PyLong_Check(res))) { +#else + if (unlikely(!PyLong_CheckExact(res))) { +#endif + return __Pyx_PyNumber_IntOrLongWrongResultType(res, name); + } + } + else if (!PyErr_Occurred()) { + PyErr_SetString(PyExc_TypeError, + "an integer is required"); + } + return res; +} +static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject* b) { + Py_ssize_t ival; + PyObject *x; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_CheckExact(b))) { + if (sizeof(Py_ssize_t) >= sizeof(long)) + return PyInt_AS_LONG(b); + else + return PyInt_AsSsize_t(b); + } +#endif + if (likely(PyLong_CheckExact(b))) { + #if CYTHON_USE_PYLONG_INTERNALS + if (likely(__Pyx_PyLong_IsCompact(b))) { + return __Pyx_PyLong_CompactValue(b); + } else { + const digit* digits = __Pyx_PyLong_Digits(b); + const Py_ssize_t size = __Pyx_PyLong_SignedDigitCount(b); + switch (size) { + case 2: + if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) { + return (Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case -2: + if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) { + return -(Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case 3: + if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) { + return (Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case -3: + if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) { + return -(Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case 4: + if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) { + return (Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case -4: + if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) { + return -(Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + } + } + #endif + return PyLong_AsSsize_t(b); + } + x = PyNumber_Index(b); + if (!x) return -1; + ival = PyInt_AsSsize_t(x); + Py_DECREF(x); + return ival; +} +static CYTHON_INLINE Py_hash_t __Pyx_PyIndex_AsHash_t(PyObject* o) { + if (sizeof(Py_hash_t) == sizeof(Py_ssize_t)) { + return (Py_hash_t) __Pyx_PyIndex_AsSsize_t(o); +#if PY_MAJOR_VERSION < 3 + } else if (likely(PyInt_CheckExact(o))) { + return PyInt_AS_LONG(o); +#endif + } else { + Py_ssize_t ival; + PyObject *x; + x = PyNumber_Index(o); + if (!x) return -1; + ival = PyInt_AsLong(x); + Py_DECREF(x); + return ival; + } +} +static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b) { + return b ? __Pyx_NewRef(Py_True) : __Pyx_NewRef(Py_False); +} +static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t ival) { + return PyInt_FromSize_t(ival); +} + + +/* #### Code section: utility_code_pragmas_end ### */ +#ifdef _MSC_VER +#pragma warning( pop ) +#endif + + + +/* #### Code section: end ### */ +#endif /* Py_PYTHON_H */ diff --git a/fairseq/data/data_utils_fast.cpython-310-x86_64-linux-gnu.so b/fairseq/data/data_utils_fast.cpython-310-x86_64-linux-gnu.so new file mode 100644 index 0000000000000000000000000000000000000000..3f9f034218ccae9ed274144e459e40d10b0d43bf --- /dev/null +++ b/fairseq/data/data_utils_fast.cpython-310-x86_64-linux-gnu.so @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c75ab2d7aba9471ec9d32a4c76dc68e51a9fd8ee6d3553637c68f70fceb2451 +size 258392 diff --git a/fairseq/data/data_utils_fast.pyx b/fairseq/data/data_utils_fast.pyx new file mode 100644 index 0000000000000000000000000000000000000000..c1f97bf5b661e1b597a447075f282ea56ccca796 --- /dev/null +++ b/fairseq/data/data_utils_fast.pyx @@ -0,0 +1,122 @@ +# cython: language_level=3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np + +cimport cython +cimport numpy as np + +DTYPE = np.int64 +ctypedef np.int64_t DTYPE_t + + +cdef _is_batch_full(long num_sentences, long num_tokens, long max_tokens, long max_sentences): + if num_sentences == 0: + return 0 + if max_sentences > 0 and num_sentences == max_sentences: + return 1 + if max_tokens > 0 and num_tokens > max_tokens: + return 1 + return 0 + + +@cython.cdivision(True) +cpdef list batch_by_size_fast( + np.ndarray[DTYPE_t, ndim=1] indices, + num_tokens_fn, + long max_tokens, + long max_sentences, + int bsz_mult, +): + cdef long sample_len = 0 + cdef list sample_lens = [] + cdef list batch = [] + cdef list batches = [] + cdef long mod_len + cdef long i + cdef long idx + cdef long num_tokens + cdef DTYPE_t[:] indices_view = indices + + for i in range(len(indices_view)): + idx = indices_view[i] + num_tokens = num_tokens_fn(idx) + sample_lens.append(num_tokens) + sample_len = max(sample_len, num_tokens) + + assert max_tokens <= 0 or sample_len <= max_tokens, ( + "sentence at index {} of size {} exceeds max_tokens " + "limit of {}!".format(idx, sample_len, max_tokens) + ) + num_tokens = (len(batch) + 1) * sample_len + + if _is_batch_full(len(batch), num_tokens, max_tokens, max_sentences): + mod_len = max( + bsz_mult * (len(batch) // bsz_mult), + len(batch) % bsz_mult, + ) + batches.append(batch[:mod_len]) + batch = batch[mod_len:] + sample_lens = sample_lens[mod_len:] + sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 + batch.append(idx) + if len(batch) > 0: + batches.append(batch) + return batches + + +cdef _find_valid_shape( + DTYPE_t[:, :] shapes_view, + long num_sentences, + long num_tokens, +): + """Return index of first valid shape of -1 if none is found.""" + for i in range(shapes_view.shape[0]): + if num_sentences <= shapes_view[i][0] and num_tokens <= shapes_view[i][1]: + return i + return -1 + + +@cython.cdivision(True) +cpdef list batch_fixed_shapes_fast( + np.ndarray[DTYPE_t, ndim=1] indices, + num_tokens_fn, + np.ndarray[DTYPE_t, ndim=2] fixed_shapes_sorted, +): + cdef long sample_len = 0 + cdef list sample_lens = [] + cdef list batch = [] + cdef list batches = [] + cdef long mod_len + cdef long i + cdef long idx + cdef long num_tokens + cdef DTYPE_t[:] indices_view = indices + cdef DTYPE_t[:, :] shapes_view = fixed_shapes_sorted + + for i in range(len(indices_view)): + idx = indices_view[i] + num_tokens = num_tokens_fn(idx) + sample_lens.append(num_tokens) + sample_len = max(sample_len, num_tokens) + + shape_idx = _find_valid_shape(shapes_view, len(batch) + 1, sample_len) + if shape_idx == -1: + batches.append(batch) + batch = [] + sample_lens = [] + sample_len = 0 + shapes_view = fixed_shapes_sorted + elif shape_idx > 0: + # small optimization for the next call to _find_valid_shape + shapes_view = shapes_view[shape_idx:] + + batch.append(idx) + + if len(batch) > 0: + batches.append(batch) + + return batches diff --git a/fairseq/data/denoising_dataset.py b/fairseq/data/denoising_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..8dc240c1ebebeba2cec15dcba9d8b2f27d8fa050 --- /dev/null +++ b/fairseq/data/denoising_dataset.py @@ -0,0 +1,407 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import math + +from . import data_utils, FairseqDataset + + +def collate( + samples, + pad_idx, + eos_idx, + vocab, + left_pad_source=False, + left_pad_target=False, + input_feeding=True, + pad_to_length=None, +): + assert input_feeding + if len(samples) == 0: + return {} + + def merge(key, left_pad, move_eos_to_beginning=False, pad_to_length=None): + return data_utils.collate_tokens( + [s[key] for s in samples], + pad_idx, + eos_idx=None, # use eos_idx of each sample instead of vocab.eos() + left_pad=left_pad, + move_eos_to_beginning=move_eos_to_beginning, + pad_to_length=pad_to_length, + ) + + id = torch.LongTensor([s['id'] for s in samples]) + src_tokens = merge( + 'source', left_pad=left_pad_source, + pad_to_length=pad_to_length['source'] if pad_to_length is not None else None, + ) + # sort by descending source length + src_lengths = torch.LongTensor([s['source'].numel() for s in samples]) + src_lengths, sort_order = src_lengths.sort(descending=True) + id = id.index_select(0, sort_order) + src_tokens = src_tokens.index_select(0, sort_order) + + prev_output_tokens = None + target = None + if samples[0].get('target', None) is not None: + target = merge( + 'target', left_pad=left_pad_target, + pad_to_length=pad_to_length['target'] if pad_to_length is not None else None, + ) + target = target.index_select(0, sort_order) + ntokens = sum(len(s['target']) for s in samples) + + if input_feeding: + # we create a shifted version of targets for feeding the + # previous output token(s) into the next decoder step + prev_output_tokens = merge( + 'target', + left_pad=left_pad_target, + move_eos_to_beginning=True, + pad_to_length=pad_to_length['target'] if pad_to_length is not None else None, + ) + prev_output_tokens = prev_output_tokens.index_select(0, sort_order) + else: + ntokens = sum(len(s['source']) for s in samples) + + batch = { + 'id': id, + 'ntokens': ntokens, + 'net_input': { + 'src_tokens': src_tokens, + 'src_lengths': src_lengths, + }, + 'target': target, + 'nsentences': samples[0]['source'].size(0), + 'sort_order': sort_order, + } + if prev_output_tokens is not None: + batch['net_input']['prev_output_tokens'] = prev_output_tokens + + return batch + + +class DenoisingDataset(FairseqDataset): + """ + A wrapper around TokenBlockDataset for BART dataset. + + Args: + dataset (TokenBlockDataset): dataset to wrap + sizes (List[int]): sentence lengths + vocab (~fairseq.data.Dictionary): vocabulary + mask_idx (int): dictionary index used for masked token + mask_whole_words: only mask whole words. This should be a byte mask + over vocab indices, indicating whether it is the beginning of a + word. We will extend any mask to encompass the whole word. + shuffle (bool, optional): shuffle the elements before batching. + Default: ``True`` + seed: Seed for random number generator for reproducibility. + args: argparse arguments. + """ + + def __init__( + self, + dataset, + sizes, + vocab, + mask_idx, + mask_whole_words, + shuffle, + seed, + args, + eos=None, + item_transform_func=None, + ): + self.dataset = dataset + + self.sizes = sizes + + self.vocab = vocab + self.shuffle = shuffle + self.seed = seed + self.mask_idx = mask_idx + self.mask_whole_word = mask_whole_words + self.mask_ratio = args.mask + self.random_ratio = args.mask_random + self.insert_ratio = args.insert + self.rotate_ratio = args.rotate + self.permute_sentence_ratio = args.permute_sentences + self.eos = (eos if eos is not None else vocab.eos()) + self.item_transform_func = item_transform_func + + if args.bpe != 'gpt2': + self.full_stop_index = self.vocab.eos() + else: + assert args.bpe == 'gpt2' + self.full_stop_index = self.vocab.index('13') + + self.replace_length = args.replace_length + if self.replace_length not in [-1, 0, 1]: + raise ValueError(f'invalid arg: replace_length={self.replace_length}') + if args.mask_length not in ['subword', 'word', 'span-poisson']: + raise ValueError(f'invalid arg: mask-length={args.mask_length}') + if args.mask_length == 'subword' and args.replace_length not in [0, 1]: + raise ValueError(f'if using subwords, use replace-length=1 or 0') + + self.mask_span_distribution = None + if args.mask_length == 'span-poisson': + _lambda = args.poisson_lambda + + lambda_to_the_k = 1 + e_to_the_minus_lambda = math.exp(-_lambda) + k_factorial = 1 + ps = [] + for k in range(0, 128): + ps.append(e_to_the_minus_lambda * lambda_to_the_k / k_factorial) + lambda_to_the_k *= _lambda + k_factorial *= (k + 1) + if ps[-1] < 0.0000001: + break + ps = torch.FloatTensor(ps) + self.mask_span_distribution = torch.distributions.Categorical(ps) + + self.epoch = 0 + + def set_epoch(self, epoch, **unused): + self.epoch = epoch + + def __getitem__(self, index): + with data_utils.numpy_seed(self.seed, self.epoch, index): + tokens = self.dataset[index] + assert tokens[-1] == self.eos + source, target = tokens, tokens.clone() + + if self.permute_sentence_ratio > 0.0: + source = self.permute_sentences(source, self.permute_sentence_ratio) + + if self.mask_ratio > 0: + source = self.add_whole_word_mask(source, self.mask_ratio) + + if self.insert_ratio > 0: + source = self.add_insertion_noise(source, self.insert_ratio) + + if self.rotate_ratio > 0.0 and np.random.random() < self.rotate_ratio: + source = self.add_rolling_noise(source) + # there can additional changes to make: + if self.item_transform_func is not None: + source, target = self.item_transform_func(source, target) + + assert (source >= 0).all() + assert (source[1:-1] >= 1).all() + assert (source <= len(self.vocab)).all() + assert source[0] == self.vocab.bos() + assert source[-1] == self.eos + return { + 'id': index, + 'source': source, + 'target': target, + } + + def __len__(self): + return len(self.dataset) + + def permute_sentences(self, source, p=1.0): + full_stops = (source == self.full_stop_index) + # Pretend it ends with a full stop so last span is a sentence + full_stops[-2] = 1 + + # Tokens that are full stops, where the previous token is not + sentence_ends = (full_stops[1:] * ~full_stops[:-1]).nonzero() + 2 + result = source.clone() + + num_sentences = sentence_ends.size(0) + num_to_permute = math.ceil((num_sentences * 2 * p) / 2.0) + substitutions = torch.randperm(num_sentences)[:num_to_permute] + ordering = torch.arange(0, num_sentences) + ordering[substitutions] = substitutions[torch.randperm(num_to_permute)] + + # Ignore at start + index = 1 + for i in ordering: + sentence = source[(sentence_ends[i - 1] if i > 0 else 1):sentence_ends[i]] + result[index:index + sentence.size(0)] = sentence + index += sentence.size(0) + return result + + def word_starts(self, source): + if self.mask_whole_word is not None: + is_word_start = self.mask_whole_word.gather(0, source) + else: + is_word_start = torch.ones(source.size()) + is_word_start[0] = 0 + is_word_start[-1] = 0 + return is_word_start + + def add_whole_word_mask(self, source, p): + is_word_start = self.word_starts(source) + num_to_mask = int(math.ceil(is_word_start.float().sum() * p)) + num_inserts = 0 + if num_to_mask == 0: + return source + + if self.mask_span_distribution is not None: + lengths = self.mask_span_distribution.sample(sample_shape=(num_to_mask,)) + + # Make sure we have enough to mask + cum_length = torch.cumsum(lengths, 0) + while cum_length[-1] < num_to_mask: + lengths = torch.cat([lengths, self.mask_span_distribution.sample(sample_shape=(num_to_mask,))], dim=0) + cum_length = torch.cumsum(lengths, 0) + + # Trim to masking budget + i = 0 + while cum_length[i] < num_to_mask: + i += 1 + lengths[i] = num_to_mask - (0 if i == 0 else cum_length[i - 1]) + num_to_mask = i + 1 + lengths = lengths[:num_to_mask] + + # Handle 0-length mask (inserts) separately + lengths = lengths[lengths > 0] + num_inserts = num_to_mask - lengths.size(0) + num_to_mask -= num_inserts + if num_to_mask == 0: + return self.add_insertion_noise(source, num_inserts / source.size(0)) + + assert (lengths > 0).all() + else: + lengths = torch.ones((num_to_mask,)).long() + assert is_word_start[-1] == 0 + word_starts = is_word_start.nonzero() + indices = word_starts[torch.randperm(word_starts.size(0))[:num_to_mask]].squeeze(1) + mask_random = torch.FloatTensor(num_to_mask).uniform_() < self.random_ratio + + source_length = source.size(0) + assert source_length - 1 not in indices + to_keep = torch.ones(source_length, dtype=torch.bool) + is_word_start[-1] = 255 # acts as a long length, so spans don't go over the end of doc + if self.replace_length == 0: + to_keep[indices] = 0 + else: + # keep index, but replace it with [MASK] + source[indices] = self.mask_idx + source[indices[mask_random]] = torch.randint(1, len(self.vocab), size=(mask_random.sum(),)) + + if self.mask_span_distribution is not None: + assert len(lengths.size()) == 1 + assert lengths.size() == indices.size() + lengths -= 1 + while indices.size(0) > 0: + assert lengths.size() == indices.size() + lengths -= is_word_start[indices + 1].long() + uncompleted = lengths >= 0 + indices = indices[uncompleted] + 1 + mask_random = mask_random[uncompleted] + lengths = lengths[uncompleted] + if self.replace_length != -1: + # delete token + to_keep[indices] = 0 + else: + # keep index, but replace it with [MASK] + source[indices] = self.mask_idx + source[indices[mask_random]] = torch.randint(1, len(self.vocab), size=(mask_random.sum(),)) + else: + # A bit faster when all lengths are 1 + while indices.size(0) > 0: + uncompleted = is_word_start[indices + 1] == 0 + indices = indices[uncompleted] + 1 + mask_random = mask_random[uncompleted] + if self.replace_length != -1: + # delete token + to_keep[indices] = 0 + else: + # keep index, but replace it with [MASK] + source[indices] = self.mask_idx + source[indices[mask_random]] = torch.randint(1, len(self.vocab), size=(mask_random.sum(),)) + + assert source_length - 1 not in indices + + source = source[to_keep] + + if num_inserts > 0: + source = self.add_insertion_noise(source, num_inserts / source.size(0)) + + return source + + def add_permuted_noise(self, tokens, p): + num_words = len(tokens) + num_to_permute = math.ceil(((num_words * 2) * p) / 2.0) + substitutions = torch.randperm(num_words - 2)[:num_to_permute] + 1 + tokens[substitutions] = tokens[substitutions[torch.randperm(num_to_permute)]] + return tokens + + def add_rolling_noise(self, tokens): + offset = np.random.randint(1, max(1, tokens.size(-1) - 1) + 1) + tokens = torch.cat( + (tokens[0:1], tokens[offset:-1], tokens[1:offset], tokens[-1:]), + dim=0, + ) + return tokens + + def add_insertion_noise(self, tokens, p): + if p == 0.0: + return tokens + + num_tokens = len(tokens) + n = int(math.ceil(num_tokens * p)) + + noise_indices = torch.randperm(num_tokens + n - 2)[:n] + 1 + noise_mask = torch.zeros(size=(num_tokens + n,), dtype=torch.bool) + noise_mask[noise_indices] = 1 + result = torch.LongTensor(n + len(tokens)).fill_(-1) + + num_random = int(math.ceil(n * self.random_ratio)) + result[noise_indices[num_random:]] = self.mask_idx + result[noise_indices[:num_random]] = torch.randint(low=1, high=len(self.vocab), size=(num_random,)) + + result[~noise_mask] = tokens + + assert (result >= 0).all() + return result + + def collater(self, samples, pad_to_length=None): + """Merge a list of samples to form a mini-batch. + Args: + samples (List[dict]): samples to collate + Returns: + dict: a mini-batch of data + """ + return collate( + samples, self.vocab.pad(), self.eos, self.vocab, + pad_to_length=pad_to_length) + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return self.sizes[index] + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return self.sizes[index] + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + indices = np.random.permutation(len(self)) + else: + indices = np.arange(len(self)) + return indices[np.argsort(self.sizes[indices], kind='mergesort')] + + def prefetch(self, indices): + self.src.prefetch(indices) + self.tgt.prefetch(indices) + + @property + def supports_prefetch(self): + return ( + hasattr(self.src, 'supports_prefetch') + and self.src.supports_prefetch + and hasattr(self.tgt, 'supports_prefetch') + and self.tgt.supports_prefetch + ) diff --git a/fairseq/data/dictionary.py b/fairseq/data/dictionary.py new file mode 100644 index 0000000000000000000000000000000000000000..01a6a8148688173d991017b39338068711e90b23 --- /dev/null +++ b/fairseq/data/dictionary.py @@ -0,0 +1,388 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +from collections import Counter +from multiprocessing import Pool + +import torch +from fairseq import utils +from fairseq.binarizer import safe_readline +from fairseq.data import data_utils +from fairseq.file_io import PathManager +from fairseq.tokenizer import tokenize_line + + +class Dictionary(object): + """A mapping from symbols to consecutive integers""" + + def __init__( + self, + *, # begin keyword-only arguments + pad="", + eos="", + unk="", + bos="", + extra_special_symbols=None, + ): + self.unk_word, self.pad_word, self.eos_word = unk, pad, eos + self.symbols = [] + self.count = [] + self.indices = {} + self.bos_index = self.add_symbol(bos) + self.pad_index = self.add_symbol(pad) + self.eos_index = self.add_symbol(eos) + self.unk_index = self.add_symbol(unk) + if extra_special_symbols: + for s in extra_special_symbols: + self.add_symbol(s) + self.nspecial = len(self.symbols) + + def __eq__(self, other): + return self.indices == other.indices + + def __getitem__(self, idx): + if idx < len(self.symbols): + return self.symbols[idx] + return self.unk_word + + def __len__(self): + """Returns the number of symbols in the dictionary""" + return len(self.symbols) + + def __contains__(self, sym): + return sym in self.indices + + def index(self, sym): + """Returns the index of the specified symbol""" + assert isinstance(sym, str) + if sym in self.indices: + return self.indices[sym] + return self.unk_index + + def string( + self, + tensor, + bpe_symbol=None, + escape_unk=False, + extra_symbols_to_ignore=None, + unk_string=None, + ): + """Helper for converting a tensor of token indices to a string. + + Can optionally remove BPE symbols or escape words. + """ + if torch.is_tensor(tensor) and tensor.dim() == 2: + return "\n".join( + self.string(t, bpe_symbol, escape_unk, extra_symbols_to_ignore) + for t in tensor + ) + + extra_symbols_to_ignore = set(extra_symbols_to_ignore or []) + extra_symbols_to_ignore.add(self.eos()) + + def token_string(i): + if i == self.unk(): + if unk_string is not None: + return unk_string + else: + return self.unk_string(escape_unk) + else: + return self[i] + + if hasattr(self, "bos_index"): + extra_symbols_to_ignore.add(self.bos()) + + sent = " ".join( + token_string(i) + for i in tensor + if utils.item(i) not in extra_symbols_to_ignore + ) + + return data_utils.post_process(sent, bpe_symbol) + + def unk_string(self, escape=False): + """Return unknown string, optionally escaped as: <>""" + if escape: + return "<{}>".format(self.unk_word) + else: + return self.unk_word + + def add_symbol(self, word, n=1, overwrite=False): + """Adds a word to the dictionary""" + if word in self.indices and not overwrite: + idx = self.indices[word] + self.count[idx] = self.count[idx] + n + return idx + else: + idx = len(self.symbols) + self.indices[word] = idx + self.symbols.append(word) + self.count.append(n) + return idx + + def update(self, new_dict): + """Updates counts from new dictionary.""" + for word in new_dict.symbols: + idx2 = new_dict.indices[word] + if word in self.indices: + idx = self.indices[word] + self.count[idx] = self.count[idx] + new_dict.count[idx2] + else: + idx = len(self.symbols) + self.indices[word] = idx + self.symbols.append(word) + self.count.append(new_dict.count[idx2]) + + def finalize(self, threshold=-1, nwords=-1, padding_factor=8): + """Sort symbols by frequency in descending order, ignoring special ones. + + Args: + - threshold defines the minimum word count + - nwords defines the total number of words in the final dictionary, + including special symbols + - padding_factor can be used to pad the dictionary size to be a + multiple of 8, which is important on some hardware (e.g., Nvidia + Tensor Cores). + """ + if nwords <= 0: + nwords = len(self) + + new_indices = dict(zip(self.symbols[: self.nspecial], range(self.nspecial))) + new_symbols = self.symbols[: self.nspecial] + new_count = self.count[: self.nspecial] + + c = Counter( + dict( + sorted(zip(self.symbols[self.nspecial :], self.count[self.nspecial :])) + ) + ) + for symbol, count in c.most_common(nwords - self.nspecial): + if count >= threshold: + new_indices[symbol] = len(new_symbols) + new_symbols.append(symbol) + new_count.append(count) + else: + break + + assert len(new_symbols) == len(new_indices) + + self.count = list(new_count) + self.symbols = list(new_symbols) + self.indices = new_indices + + self.pad_to_multiple_(padding_factor) + + def pad_to_multiple_(self, padding_factor): + """Pad Dictionary size to be a multiple of *padding_factor*.""" + if padding_factor > 1: + i = 0 + while len(self) % padding_factor != 0: + symbol = "madeupword{:04d}".format(i) + self.add_symbol(symbol, n=0) + i += 1 + + def bos(self): + """Helper to get index of beginning-of-sentence symbol""" + return self.bos_index + + def pad(self): + """Helper to get index of pad symbol""" + return self.pad_index + + def eos(self): + """Helper to get index of end-of-sentence symbol""" + return self.eos_index + + def unk(self): + """Helper to get index of unk symbol""" + return self.unk_index + + @classmethod + def load(cls, f): + """Loads the dictionary from a text file with the format: + + ``` + + + ... + ``` + """ + d = cls() + d.add_from_file(f) + return d + + def add_from_file(self, f): + """ + Loads a pre-existing dictionary from a text file and adds its symbols + to this instance. + """ + if isinstance(f, str): + try: + with PathManager.open(f, "r", encoding="utf-8") as fd: + self.add_from_file(fd) + except FileNotFoundError as fnfe: + raise fnfe + except UnicodeError: + raise Exception( + "Incorrect encoding detected in {}, please " + "rebuild the dataset".format(f) + ) + return + + lines = f.readlines() + indices_start_line = self._load_meta(lines) + + for line in lines[indices_start_line:]: + try: + line, field = line.rstrip().rsplit(" ", 1) + if field == "#fairseq:overwrite": + overwrite = True + line, field = line.rsplit(" ", 1) + else: + overwrite = False + count = int(field) + word = line + if word in self and not overwrite: + raise RuntimeError( + "Duplicate word found when loading Dictionary: '{}'. " + "Duplicate words can overwrite earlier ones by adding the " + "#fairseq:overwrite flag at the end of the corresponding row " + "in the dictionary file. If using the Camembert model, please " + "download an updated copy of the model file." + .format(word) + ) + self.add_symbol(word, n=count, overwrite=overwrite) + except ValueError: + raise ValueError( + "Incorrect dictionary format, expected ' [flags]'" + ) + + def _save(self, f, kv_iterator): + if isinstance(f, str): + PathManager.mkdirs(os.path.dirname(f)) + with PathManager.open(f, "w", encoding="utf-8") as fd: + return self.save(fd) + for k, v in kv_iterator: + print("{} {}".format(k, v), file=f) + + def _get_meta(self): + return [], [] + + def _load_meta(self, lines): + return 0 + + def save(self, f): + """Stores dictionary into a text file""" + ex_keys, ex_vals = self._get_meta() + self._save( + f, + zip( + ex_keys + self.symbols[self.nspecial :], + ex_vals + self.count[self.nspecial :], + ), + ) + + def dummy_sentence(self, length): + t = torch.Tensor(length).uniform_(self.nspecial + 1, len(self)).long() + t[-1] = self.eos() + return t + + def encode_line( + self, + line, + line_tokenizer=tokenize_line, + add_if_not_exist=True, + consumer=None, + append_eos=True, + reverse_order=False, + ): + words = line_tokenizer(line) + if reverse_order: + words = list(reversed(words)) + nwords = len(words) + ids = torch.IntTensor(nwords + 1 if append_eos else nwords) + + for i, word in enumerate(words): + if add_if_not_exist: + idx = self.add_symbol(word) + else: + idx = self.index(word) + if consumer is not None: + consumer(word, idx) + ids[i] = idx + if append_eos: + ids[nwords] = self.eos_index + return ids + + @staticmethod + def _add_file_to_dictionary_single_worker( + filename, tokenize, eos_word, worker_id=0, num_workers=1 + ): + counter = Counter() + with open(PathManager.get_local_path(filename), "r", encoding="utf-8") as f: + size = os.fstat(f.fileno()).st_size + chunk_size = size // num_workers + offset = worker_id * chunk_size + end = offset + chunk_size + f.seek(offset) + if offset > 0: + safe_readline(f) # drop first incomplete line + line = f.readline() + while line: + for word in tokenize(line): + counter.update([word]) + counter.update([eos_word]) + if f.tell() > end: + break + line = f.readline() + return counter + + @staticmethod + def add_file_to_dictionary(filename, dict, tokenize, num_workers): + def merge_result(counter): + for w, c in sorted(counter.items()): + dict.add_symbol(w, c) + + if num_workers > 1: + pool = Pool(processes=num_workers) + results = [] + for worker_id in range(num_workers): + results.append( + pool.apply_async( + Dictionary._add_file_to_dictionary_single_worker, + (filename, tokenize, dict.eos_word, worker_id, num_workers), + ) + ) + pool.close() + pool.join() + for r in results: + merge_result(r.get()) + else: + merge_result( + Dictionary._add_file_to_dictionary_single_worker( + filename, tokenize, dict.eos_word + ) + ) + + +class TruncatedDictionary(object): + def __init__(self, wrapped_dict, length): + self.__class__ = type( + wrapped_dict.__class__.__name__, + (self.__class__, wrapped_dict.__class__), + {}, + ) + self.__dict__ = wrapped_dict.__dict__ + self.wrapped_dict = wrapped_dict + self.length = min(len(self.wrapped_dict), length) + + def __len__(self): + return self.length + + def __getitem__(self, i): + if i < self.length: + return self.wrapped_dict[i] + return self.wrapped_dict.unk() diff --git a/fairseq/data/encoders/__init__.py b/fairseq/data/encoders/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c0909b669757e4e5731be8ac28cf9c221a1583ea --- /dev/null +++ b/fairseq/data/encoders/__init__.py @@ -0,0 +1,29 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import importlib +import os + +from fairseq import registry + + +build_tokenizer, register_tokenizer, TOKENIZER_REGISTRY = registry.setup_registry( + '--tokenizer', + default=None, +) + + +build_bpe, register_bpe, BPE_REGISTRY = registry.setup_registry( + '--bpe', + default=None, +) + + +# automatically import any Python files in the encoders/ directory +for file in os.listdir(os.path.dirname(__file__)): + if file.endswith('.py') and not file.startswith('_'): + module = file[:file.find('.py')] + importlib.import_module('fairseq.data.encoders.' + module) diff --git a/fairseq/data/encoders/__pycache__/__init__.cpython-310.pyc b/fairseq/data/encoders/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c8307bad166f28075f5e2b15b8d69c245c9df306 Binary files /dev/null and b/fairseq/data/encoders/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/byte_bpe.cpython-310.pyc b/fairseq/data/encoders/__pycache__/byte_bpe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1ecf22e48c8e24bc874d96b5cdc766cf2b7397ef Binary files /dev/null and b/fairseq/data/encoders/__pycache__/byte_bpe.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/byte_utils.cpython-310.pyc b/fairseq/data/encoders/__pycache__/byte_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..66590e961ff2266dc568ba039186ee13b7b0c375 Binary files /dev/null and b/fairseq/data/encoders/__pycache__/byte_utils.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/bytes.cpython-310.pyc b/fairseq/data/encoders/__pycache__/bytes.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ca307a60e60e93ad624f4f1db8517d4d4c9ffc11 Binary files /dev/null and b/fairseq/data/encoders/__pycache__/bytes.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/characters.cpython-310.pyc b/fairseq/data/encoders/__pycache__/characters.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6a732cb9648b5625fbf05071cb927163c66da9db Binary files /dev/null and b/fairseq/data/encoders/__pycache__/characters.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/fastbpe.cpython-310.pyc b/fairseq/data/encoders/__pycache__/fastbpe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..10a3272ef6da700340d12792ed52a426ff22e943 Binary files /dev/null and b/fairseq/data/encoders/__pycache__/fastbpe.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/gpt2_bpe.cpython-310.pyc b/fairseq/data/encoders/__pycache__/gpt2_bpe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5db4a1da60595bb4efb19e0b0b8acaff5a992ec5 Binary files /dev/null and b/fairseq/data/encoders/__pycache__/gpt2_bpe.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/gpt2_bpe_utils.cpython-310.pyc b/fairseq/data/encoders/__pycache__/gpt2_bpe_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..81b680034357b9d785a811627b1cb96b22ca4bfb Binary files /dev/null and b/fairseq/data/encoders/__pycache__/gpt2_bpe_utils.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/hf_bert_bpe.cpython-310.pyc b/fairseq/data/encoders/__pycache__/hf_bert_bpe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..88c5567a8611dc856426f576154c24b67f88b9d0 Binary files /dev/null and b/fairseq/data/encoders/__pycache__/hf_bert_bpe.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/hf_byte_bpe.cpython-310.pyc b/fairseq/data/encoders/__pycache__/hf_byte_bpe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ddc7736205b6914ebb6589133a914cdb1ee0d810 Binary files /dev/null and b/fairseq/data/encoders/__pycache__/hf_byte_bpe.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/moses_tokenizer.cpython-310.pyc b/fairseq/data/encoders/__pycache__/moses_tokenizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c7b725560f01c3f168086307af2c876c4a592d9d Binary files /dev/null and b/fairseq/data/encoders/__pycache__/moses_tokenizer.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/nltk_tokenizer.cpython-310.pyc b/fairseq/data/encoders/__pycache__/nltk_tokenizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..98a6e22dd0a6c31887af105374af7190a0a66cbf Binary files /dev/null and b/fairseq/data/encoders/__pycache__/nltk_tokenizer.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/sentencepiece_bpe.cpython-310.pyc b/fairseq/data/encoders/__pycache__/sentencepiece_bpe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5c67c3c92ba28a3f527e7c5543168341d79f481a Binary files /dev/null and b/fairseq/data/encoders/__pycache__/sentencepiece_bpe.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/space_tokenizer.cpython-310.pyc b/fairseq/data/encoders/__pycache__/space_tokenizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f210725b0ccf738576e3a13b36e1e1e56872ec0e Binary files /dev/null and b/fairseq/data/encoders/__pycache__/space_tokenizer.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/subword_nmt_bpe.cpython-310.pyc b/fairseq/data/encoders/__pycache__/subword_nmt_bpe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a5bd7d27f9f6d10186aea7729112e78cfb5571cc Binary files /dev/null and b/fairseq/data/encoders/__pycache__/subword_nmt_bpe.cpython-310.pyc differ diff --git a/fairseq/data/encoders/__pycache__/utils.cpython-310.pyc b/fairseq/data/encoders/__pycache__/utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..91d23e5dc4c756aa47b63049f9cfec3cd400a8bd Binary files /dev/null and b/fairseq/data/encoders/__pycache__/utils.cpython-310.pyc differ diff --git a/fairseq/data/encoders/byte_bpe.py b/fairseq/data/encoders/byte_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..1d78ff91504d5af66a958e3fc7422b89df95d3dc --- /dev/null +++ b/fairseq/data/encoders/byte_bpe.py @@ -0,0 +1,38 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.data.encoders.byte_utils import (byte_encode, smart_byte_decode, + SPACE, SPACE_ESCAPE) + + +@register_bpe('byte_bpe') +class ByteBPE(object): + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--sentencepiece-model-path', type=str, + help='path to sentencepiece model') + # fmt: on + + def __init__(self, args): + vocab = file_utils.cached_path(args.sentencepiece_model_path) + try: + import sentencepiece as spm + self.sp = spm.SentencePieceProcessor() + self.sp.Load(vocab) + except ImportError: + raise ImportError('Please install sentencepiece with: pip install sentencepiece') + + def encode(self, x: str) -> str: + byte_encoded = byte_encode(x) + return SPACE.join(self.sp.EncodeAsPieces(byte_encoded)) + + @staticmethod + def decode(x: str) -> str: + unescaped = x.replace(SPACE, '').replace(SPACE_ESCAPE, SPACE) + return smart_byte_decode(unescaped) diff --git a/fairseq/data/encoders/byte_utils.py b/fairseq/data/encoders/byte_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7c4bb7471359f54414b83166617475cf58375fca --- /dev/null +++ b/fairseq/data/encoders/byte_utils.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + +WHITESPACE_NORMALIZER = re.compile(r'\s+') +SPACE = chr(32) +SPACE_ESCAPE = chr(9601) +# excluding non-breaking space (160) here +PRINTABLE_LATIN = set( + list(range(32, 126 + 1)) + list(range(161, 172 + 1)) + + list(range(174, 255 + 1)) +) +BYTE_TO_BCHAR = { + b: chr(b) if b in PRINTABLE_LATIN else chr(256 + b) for b in range(256) +} +BCHAR_TO_BYTE = {bc: b for b, bc in BYTE_TO_BCHAR.items()} + + +def byte_encode(x: str) -> str: + normalized = WHITESPACE_NORMALIZER.sub(SPACE, x) + return ''.join([BYTE_TO_BCHAR[b] for b in normalized.encode('utf-8')]) + + +def byte_decode(x: str) -> str: + try: + return bytes([BCHAR_TO_BYTE[bc] for bc in x]).decode('utf-8') + except ValueError: + return '' + + +def smart_byte_decode(x: str) -> str: + output = byte_decode(x) + if output == '': + # DP the best recovery (max valid chars) if it's broken + n_bytes = len(x) + f = [0 for _ in range(n_bytes + 1)] + pt = [0 for _ in range(n_bytes + 1)] + for i in range(1, n_bytes + 1): + f[i], pt[i] = f[i - 1], i - 1 + for j in range(1, min(4, i) + 1): + if f[i - j] + 1 > f[i] and len(byte_decode(x[i - j: i])) > 0: + f[i], pt[i] = f[i - j] + 1, i - j + cur_pt = n_bytes + while cur_pt > 0: + if f[cur_pt] == f[pt[cur_pt]] + 1: + output = byte_decode(x[pt[cur_pt]: cur_pt]) + output + cur_pt = pt[cur_pt] + return output diff --git a/fairseq/data/encoders/bytes.py b/fairseq/data/encoders/bytes.py new file mode 100644 index 0000000000000000000000000000000000000000..8bace19c53542bf872c3a02b322c0f39842d423f --- /dev/null +++ b/fairseq/data/encoders/bytes.py @@ -0,0 +1,30 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.data.encoders import register_bpe +from fairseq.data.encoders.byte_utils import (byte_encode, smart_byte_decode, + SPACE, SPACE_ESCAPE) + + +@register_bpe('bytes') +class Bytes(object): + def __init__(self, args): + pass + + @staticmethod + def add_args(parser): + pass + + @staticmethod + def encode(x: str) -> str: + encoded = byte_encode(x) + escaped = encoded.replace(SPACE, SPACE_ESCAPE) + return SPACE.join(list(escaped)) + + @staticmethod + def decode(x: str) -> str: + unescaped = x.replace(SPACE, '').replace(SPACE_ESCAPE, SPACE) + return smart_byte_decode(unescaped) diff --git a/fairseq/data/encoders/characters.py b/fairseq/data/encoders/characters.py new file mode 100644 index 0000000000000000000000000000000000000000..db6a58a6502ace4709838ad574757074e893e1a3 --- /dev/null +++ b/fairseq/data/encoders/characters.py @@ -0,0 +1,29 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.data.encoders import register_bpe + +SPACE = chr(32) +SPACE_ESCAPE = chr(9601) + + +@register_bpe('characters') +class Characters(object): + def __init__(self, args): + pass + + @staticmethod + def add_args(parser): + pass + + @staticmethod + def encode(x: str) -> str: + escaped = x.replace(SPACE, SPACE_ESCAPE) + return SPACE.join(list(escaped)) + + @staticmethod + def decode(x: str) -> str: + return x.replace(SPACE, '').replace(SPACE_ESCAPE, SPACE) diff --git a/fairseq/data/encoders/fastbpe.py b/fairseq/data/encoders/fastbpe.py new file mode 100644 index 0000000000000000000000000000000000000000..ea0badd544d506bd773b39693c02179ed54ca0b1 --- /dev/null +++ b/fairseq/data/encoders/fastbpe.py @@ -0,0 +1,35 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe + + +@register_bpe('fastbpe') +class fastBPE(object): + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--bpe-codes', type=str, + help='path to fastBPE BPE') + # fmt: on + + def __init__(self, args): + if args.bpe_codes is None: + raise ValueError('--bpe-codes is required for --bpe=fastbpe') + codes = file_utils.cached_path(args.bpe_codes) + try: + import fastBPE + self.bpe = fastBPE.fastBPE(codes) + self.bpe_symbol = "@@ " + except ImportError: + raise ImportError('Please install fastBPE with: pip install fastBPE') + + def encode(self, x: str) -> str: + return self.bpe.apply([x])[0] + + def decode(self, x: str) -> str: + return (x + ' ').replace(self.bpe_symbol, '').rstrip() diff --git a/fairseq/data/encoders/gpt2_bpe.py b/fairseq/data/encoders/gpt2_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..54e0593d00e23bc235e557df29909325f2a17056 --- /dev/null +++ b/fairseq/data/encoders/gpt2_bpe.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe + +from .gpt2_bpe_utils import get_encoder + + +DEFAULT_ENCODER_JSON = 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json' +DEFAULT_VOCAB_BPE = 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe' + + +@register_bpe('gpt2') +class GPT2BPE(object): + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--gpt2-encoder-json', type=str, + default=DEFAULT_ENCODER_JSON, + help='path to encoder.json') + parser.add_argument('--gpt2-vocab-bpe', type=str, + default=DEFAULT_VOCAB_BPE, + help='path to vocab.bpe') + # fmt: on + + def __init__(self, args): + encoder_json = file_utils.cached_path( + getattr(args, 'gpt2_encoder_json', DEFAULT_ENCODER_JSON) + ) + vocab_bpe = file_utils.cached_path( + getattr(args, 'gpt2_vocab_bpe', DEFAULT_VOCAB_BPE) + ) + self.bpe = get_encoder(encoder_json, vocab_bpe) + + def encode(self, x: str) -> str: + return ' '.join(map(str, self.bpe.encode(x))) + + def decode(self, x: str) -> str: + return self.bpe.decode([ + int(tok) if tok not in {'', ''} else tok + for tok in x.split() + ]) + + def is_beginning_of_word(self, x: str) -> bool: + return self.decode(x).startswith(' ') diff --git a/fairseq/data/encoders/gpt2_bpe_utils.py b/fairseq/data/encoders/gpt2_bpe_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..1917f823141306e2e7f5a4a19c59d35a370a8711 --- /dev/null +++ b/fairseq/data/encoders/gpt2_bpe_utils.py @@ -0,0 +1,127 @@ +""" +Byte pair encoding utilities from GPT-2. + +Original source: https://github.com/openai/gpt-2/blob/master/src/encoder.py +Original license: MIT +""" + +from functools import lru_cache +import json + + +@lru_cache() +def bytes_to_unicode(): + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a signficant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8+n) + n += 1 + cs = [chr(n) for n in cs] + return dict(zip(bs, cs)) + +def get_pairs(word): + """Return set of symbol pairs in a word. + Word is represented as tuple of symbols (symbols being variable-length strings). + """ + pairs = set() + prev_char = word[0] + for char in word[1:]: + pairs.add((prev_char, char)) + prev_char = char + return pairs + +class Encoder: + + def __init__(self, encoder, bpe_merges, errors='replace'): + self.encoder = encoder + self.decoder = {v:k for k,v in self.encoder.items()} + self.errors = errors # how to handle errors in decoding + self.byte_encoder = bytes_to_unicode() + self.byte_decoder = {v:k for k, v in self.byte_encoder.items()} + self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges)))) + self.cache = {} + + try: + import regex as re + self.re = re + except ImportError: + raise ImportError('Please install regex with: pip install regex') + + # Should haved added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions + self.pat = self.re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""") + + def bpe(self, token): + if token in self.cache: + return self.cache[token] + word = tuple(token) + pairs = get_pairs(word) + + if not pairs: + return token + + while True: + bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf'))) + if bigram not in self.bpe_ranks: + break + first, second = bigram + new_word = [] + i = 0 + while i < len(word): + try: + j = word.index(first, i) + new_word.extend(word[i:j]) + i = j + except: + new_word.extend(word[i:]) + break + + if word[i] == first and i < len(word)-1 and word[i+1] == second: + new_word.append(first+second) + i += 2 + else: + new_word.append(word[i]) + i += 1 + new_word = tuple(new_word) + word = new_word + if len(word) == 1: + break + else: + pairs = get_pairs(word) + word = ' '.join(word) + self.cache[token] = word + return word + + def encode(self, text): + bpe_tokens = [] + for token in self.re.findall(self.pat, text): + token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8')) + bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' ')) + return bpe_tokens + + def decode(self, tokens): + text = ''.join([self.decoder.get(token, token) for token in tokens]) + text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors=self.errors) + return text + +def get_encoder(encoder_json_path, vocab_bpe_path): + with open(encoder_json_path, 'r') as f: + encoder = json.load(f) + with open(vocab_bpe_path, 'r', encoding="utf-8") as f: + bpe_data = f.read() + bpe_merges = [tuple(merge_str.split()) for merge_str in bpe_data.split('\n')[1:-1]] + return Encoder( + encoder=encoder, + bpe_merges=bpe_merges, + ) diff --git a/fairseq/data/encoders/hf_bert_bpe.py b/fairseq/data/encoders/hf_bert_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..16adc45aeea4e2c61ed5d99a3561113c7db6411b --- /dev/null +++ b/fairseq/data/encoders/hf_bert_bpe.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data.encoders import register_bpe + + +@register_bpe('bert') +class BertBPE(object): + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--bpe-cased', action='store_true', + help='set for cased BPE', + default=False) + parser.add_argument('--bpe-vocab-file', type=str, + help='bpe vocab file.') + # fmt: on + + def __init__(self, args): + try: + from transformers import BertTokenizer + except ImportError: + raise ImportError( + 'Please install transformers with: pip install transformers' + ) + + if 'bpe_vocab_file' in args: + self.bert_tokenizer = BertTokenizer( + args.bpe_vocab_file, + do_lower_case=not args.bpe_cased + ) + else: + vocab_file_name = 'bert-base-cased' if args.bpe_cased else 'bert-base-uncased' + self.bert_tokenizer = BertTokenizer.from_pretrained(vocab_file_name) + + def encode(self, x: str) -> str: + return ' '.join(self.bert_tokenizer.tokenize(x)) + + def decode(self, x: str) -> str: + return self.bert_tokenizer.clean_up_tokenization( + self.bert_tokenizer.convert_tokens_to_string(x.split(' ')) + ) + + def is_beginning_of_word(self, x: str) -> bool: + return not x.startswith('##') diff --git a/fairseq/data/encoders/hf_byte_bpe.py b/fairseq/data/encoders/hf_byte_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..2767df044e473ebde9e69536a923e8fa5d9ff500 --- /dev/null +++ b/fairseq/data/encoders/hf_byte_bpe.py @@ -0,0 +1,46 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data.encoders import register_bpe + + +@register_bpe('hf_byte_bpe') +class HuggingFaceByteLevelBPE(object): + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--bpe-merges', help='path to merges.txt') + parser.add_argument('--bpe-vocab', help='path to vocab.json') + parser.add_argument('--bpe-add-prefix-space', action='store_true', + help='add prefix space before encoding') + # fmt: on + + def __init__(self, args): + try: + from tokenizers import ByteLevelBPETokenizer + except ImportError: + raise ImportError( + 'Please install huggingface/tokenizers with: ' + 'pip install tokenizers' + ) + + self.bpe = ByteLevelBPETokenizer( + args.bpe_vocab, + args.bpe_merges, + add_prefix_space=getattr(args, 'bpe_add_prefix_space', False), + ) + + def encode(self, x: str) -> str: + return ' '.join(map(str, self.bpe.encode(x).ids)) + + def decode(self, x: str) -> str: + return self.bpe.decode([ + int(tok) if tok not in {'', ''} else tok + for tok in x.split() + ]) + + def is_beginning_of_word(self, x: str) -> bool: + return self.decode(x).startswith(' ') diff --git a/fairseq/data/encoders/moses_tokenizer.py b/fairseq/data/encoders/moses_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..b1e7478b9deecc12aaf2bfac066572fca109d0cd --- /dev/null +++ b/fairseq/data/encoders/moses_tokenizer.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data.encoders import register_tokenizer + + +@register_tokenizer('moses') +class MosesTokenizer(object): + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--moses-source-lang', metavar='SRC', + help='source language') + parser.add_argument('--moses-target-lang', metavar='TARGET', + help='target language') + parser.add_argument('--moses-no-dash-splits', action='store_true', default=False, + help='don\'t apply dash split rules') + parser.add_argument('--moses-no-escape', action='store_true', default=False, + help='don\'t perform HTML escaping on apostrophy, quotes, etc.') + # fmt: on + + def __init__(self, args): + self.args = args + + if getattr(args, 'moses_source_lang', None) is None: + args.moses_source_lang = getattr(args, 'source_lang', 'en') + if getattr(args, 'moses_target_lang', None) is None: + args.moses_target_lang = getattr(args, 'target_lang', 'en') + + try: + from sacremoses import MosesTokenizer, MosesDetokenizer + self.tok = MosesTokenizer(args.moses_source_lang) + self.detok = MosesDetokenizer(args.moses_target_lang) + except ImportError: + raise ImportError('Please install Moses tokenizer with: pip install sacremoses') + + def encode(self, x: str) -> str: + return self.tok.tokenize( + x, + aggressive_dash_splits=(not self.args.moses_no_dash_splits), + return_str=True, + escape=(not self.args.moses_no_escape), + ) + + def decode(self, x: str) -> str: + return self.detok.detokenize(x.split()) diff --git a/fairseq/data/encoders/nltk_tokenizer.py b/fairseq/data/encoders/nltk_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..3db8ee56525196a68383dee5c04a22835e45625b --- /dev/null +++ b/fairseq/data/encoders/nltk_tokenizer.py @@ -0,0 +1,23 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data.encoders import register_tokenizer + + +@register_tokenizer('nltk') +class NLTKTokenizer(object): + + def __init__(self, source_lang=None, target_lang=None): + try: + from nltk.tokenize import word_tokenize + self.word_tokenize = word_tokenize + except ImportError: + raise ImportError('Please install nltk with: pip install nltk') + + def encode(self, x: str) -> str: + return ' '.join(self.word_tokenize(x)) + + def decode(self, x: str) -> str: + return x diff --git a/fairseq/data/encoders/sentencepiece_bpe.py b/fairseq/data/encoders/sentencepiece_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..e5ff5db389a4718b028eef75b80616e1f0702b0b --- /dev/null +++ b/fairseq/data/encoders/sentencepiece_bpe.py @@ -0,0 +1,43 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe + + +@register_bpe('sentencepiece') +class SentencepieceBPE(object): + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--sentencepiece-model', type=str, + help='path to sentencepiece model') + # fmt: on + + def __init__(self, args): + sentencepiece_model = file_utils.cached_path(args.sentencepiece_model) + try: + import sentencepiece as spm + self.sp = spm.SentencePieceProcessor() + self.sp.Load(sentencepiece_model) + except ImportError: + raise ImportError('Please install sentencepiece with: pip install sentencepiece') + + def encode(self, x: str) -> str: + return ' '.join(self.sp.EncodeAsPieces(x)) + + def decode(self, x: str) -> str: + return x.replace(' ', '').replace('\u2581', ' ').strip() + + def is_beginning_of_word(self, x: str) -> bool: + if x in ['', '', '', '']: + # special elements are always considered beginnings + # HACK: this logic is already present in fairseq/tasks/masked_lm.py + # but these special tokens are also contained in the sentencepiece + # vocabulary which causes duplicate special tokens. This hack makes + # sure that they are all taken into account. + return True + return x.startswith('\u2581') diff --git a/fairseq/data/encoders/space_tokenizer.py b/fairseq/data/encoders/space_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..670001a8e8514e910c91281a88876261514f9374 --- /dev/null +++ b/fairseq/data/encoders/space_tokenizer.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + +from fairseq.data.encoders import register_tokenizer + + +@register_tokenizer('space') +class SpaceTokenizer(object): + + def __init__(self, source_lang=None, target_lang=None): + self.space_tok = re.compile(r"\s+") + + def encode(self, x: str) -> str: + return self.space_tok.sub(' ', x) + + def decode(self, x: str) -> str: + return x diff --git a/fairseq/data/encoders/subword_nmt_bpe.py b/fairseq/data/encoders/subword_nmt_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..78f19b43eaec7745ef86f693359946c7b60bd14c --- /dev/null +++ b/fairseq/data/encoders/subword_nmt_bpe.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe + + +@register_bpe('subword_nmt') +class SubwordNMTBPE(object): + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--bpe-codes', type=str, + help='path to subword NMT BPE') + parser.add_argument('--bpe-separator', default='@@', + help='BPE separator') + # fmt: on + + def __init__(self, args): + if args.bpe_codes is None: + raise ValueError('--bpe-codes is required for --bpe=subword_nmt') + codes = file_utils.cached_path(args.bpe_codes) + try: + from subword_nmt import apply_bpe + bpe_parser = apply_bpe.create_parser() + bpe_args = bpe_parser.parse_args([ + '--codes', codes, + '--separator', args.bpe_separator, + ]) + self.bpe = apply_bpe.BPE( + bpe_args.codes, + bpe_args.merges, + bpe_args.separator, + None, + bpe_args.glossaries, + ) + self.bpe_symbol = bpe_args.separator + ' ' + except ImportError: + raise ImportError('Please install subword_nmt with: pip install subword-nmt') + + def encode(self, x: str) -> str: + return self.bpe.process_line(x) + + def decode(self, x: str) -> str: + return (x + ' ').replace(self.bpe_symbol, '').rstrip() diff --git a/fairseq/data/encoders/utils.py b/fairseq/data/encoders/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a0e491c143c17b8364cb1d4a730c7415fb116907 --- /dev/null +++ b/fairseq/data/encoders/utils.py @@ -0,0 +1,28 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.data import encoders + + +def get_whole_word_mask(args, dictionary): + bpe = encoders.build_bpe(args) + if bpe is not None: + def is_beginning_of_word(i): + if i < dictionary.nspecial: + # special elements are always considered beginnings + return True + tok = dictionary[i] + if tok.startswith('madeupword'): + return True + try: + return bpe.is_beginning_of_word(tok) + except ValueError: + return True + mask_whole_words = torch.ByteTensor(list( + map(is_beginning_of_word, range(len(dictionary))) + )) + return mask_whole_words + return None diff --git a/fairseq/data/fairseq_dataset.py b/fairseq/data/fairseq_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..900bfaff101d9b22bcb60fab78546517ed910a8d --- /dev/null +++ b/fairseq/data/fairseq_dataset.py @@ -0,0 +1,162 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch.utils.data + +from fairseq.data import data_utils + + +class EpochListening: + """Mixin for receiving updates whenever the epoch increments.""" + + def set_epoch(self, epoch): + """Will receive the updated epoch number at the beginning of the epoch.""" + pass + + +class FairseqDataset(torch.utils.data.Dataset, EpochListening): + """A dataset that provides helpers for batching.""" + + def __getitem__(self, index): + raise NotImplementedError + + def __len__(self): + raise NotImplementedError + + def collater(self, samples): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch suitable for forwarding with a Model + """ + raise NotImplementedError + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + raise NotImplementedError + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + raise NotImplementedError + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + return np.arange(len(self)) + + @property + def supports_prefetch(self): + """Whether this dataset supports prefetching.""" + return False + + def attr(self, attr: str, index: int): + return getattr(self, attr, None) + + def prefetch(self, indices): + """Prefetch the data required for this epoch.""" + raise NotImplementedError + + def get_batch_shapes(self): + """ + Return a list of valid batch shapes, for example:: + + [(8, 512), (16, 256), (32, 128)] + + The first dimension of each tuple is the batch size and can be ``None`` + to automatically infer the max batch size based on ``--max-tokens``. + The second dimension of each tuple is the max supported length as given + by :func:`fairseq.data.FairseqDataset.num_tokens`. + + This will be used by :func:`fairseq.data.FairseqDataset.batch_by_size` + to restrict batch shapes. This is useful on TPUs to avoid too many + dynamic shapes (and recompilations). + """ + return None + + def batch_by_size( + self, + indices, + max_tokens=None, + max_sentences=None, + required_batch_size_multiple=1, + ): + """ + Given an ordered set of indices, return batches according to + *max_tokens*, *max_sentences* and *required_batch_size_multiple*. + """ + from fairseq.data import data_utils + + fixed_shapes = self.get_batch_shapes() + if fixed_shapes is not None: + + def adjust_bsz(bsz, num_tokens): + if bsz is None: + assert max_tokens is not None, 'Must specify --max-tokens' + bsz = max_tokens // num_tokens + if max_sentences is not None: + bsz = min(bsz, max_sentences) + elif ( + bsz >= required_batch_size_multiple + and bsz % required_batch_size_multiple != 0 + ): + bsz -= (bsz % required_batch_size_multiple) + return bsz + + fixed_shapes = np.array([ + [adjust_bsz(bsz, num_tokens), num_tokens] + for (bsz, num_tokens) in fixed_shapes + ]) + + return data_utils.batch_by_size( + indices, + num_tokens_fn=self.num_tokens, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + fixed_shapes=fixed_shapes, + ) + + def filter_indices_by_size(self, indices, max_sizes): + """ Filter a list of sample indices. Remove those that are longer + than specified in max_sizes. + + WARNING: don't update, override method in child classes + + Args: + indices (np.array): original array of sample indices + max_sizes (int or list[int] or tuple[int]): max sample size, + can be defined separately for src and tgt (then list or tuple) + + Returns: + np.array: filtered sample array + list: list of removed indices + """ + if isinstance(max_sizes, float) or isinstance(max_sizes, int): + if hasattr(self, 'sizes') and isinstance(self.sizes, np.ndarray): + ignored = indices[self.sizes[indices] > max_sizes].tolist() + indices = indices[self.sizes[indices] <= max_sizes] + elif hasattr(self, 'sizes') and isinstance(self.sizes, list) and len(self.sizes) == 1: + ignored = indices[self.sizes[0][indices] > max_sizes].tolist() + indices = indices[self.sizes[0][indices] <= max_sizes] + else: + indices, ignored = data_utils._filter_by_size_dynamic(indices, self.size, max_sizes) + else: + indices, ignored = data_utils._filter_by_size_dynamic(indices, self.size, max_sizes) + return indices, ignored + + +class FairseqIterableDataset(torch.utils.data.IterableDataset, EpochListening): + """For datasets that need to be read sequentially, usually because the data + is being streamed or otherwise can't be manipulated on a single machine. + """ + + def __iter__(self): + raise NotImplementedError diff --git a/fairseq/data/id_dataset.py b/fairseq/data/id_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..6a73ba1ff74e53d8789581c24880c0b6485cffa7 --- /dev/null +++ b/fairseq/data/id_dataset.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class IdDataset(FairseqDataset): + + def __getitem__(self, index): + return index + + def __len__(self): + return 0 + + def collater(self, samples): + return torch.tensor(samples) diff --git a/fairseq/data/indexed_dataset.py b/fairseq/data/indexed_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..12497989bbe3a52bd70dc222538cb4fb5eb70b43 --- /dev/null +++ b/fairseq/data/indexed_dataset.py @@ -0,0 +1,523 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from functools import lru_cache +import os +import shutil +import struct + +import numpy as np +import torch + +from . import FairseqDataset + + +def __best_fitting_dtype(vocab_size=None): + if vocab_size is not None and vocab_size < 65500: + return np.uint16 + else: + return np.int32 + + +def get_available_dataset_impl(): + return ['raw', 'lazy', 'cached', 'mmap'] + + +def infer_dataset_impl(path): + if IndexedRawTextDataset.exists(path): + return 'raw' + elif IndexedDataset.exists(path): + with open(index_file_path(path), 'rb') as f: + magic = f.read(8) + if magic == IndexedDataset._HDR_MAGIC: + return 'cached' + elif magic == MMapIndexedDataset.Index._HDR_MAGIC[:8]: + return 'mmap' + else: + return None + else: + return None + + +def make_builder(out_file, impl, vocab_size=None): + if impl == 'mmap': + return MMapIndexedDatasetBuilder(out_file, dtype=__best_fitting_dtype(vocab_size)) + else: + return IndexedDatasetBuilder(out_file) + + +def make_dataset(path, impl, fix_lua_indexing=False, dictionary=None): + if impl == 'raw' and IndexedRawTextDataset.exists(path): + assert dictionary is not None + return IndexedRawTextDataset(path, dictionary) + elif impl == 'lazy' and IndexedDataset.exists(path): + return IndexedDataset(path, fix_lua_indexing=fix_lua_indexing) + elif impl == 'cached' and IndexedDataset.exists(path): + return IndexedCachedDataset(path, fix_lua_indexing=fix_lua_indexing) + elif impl == 'mmap' and MMapIndexedDataset.exists(path): + return MMapIndexedDataset(path) + return None + + +def dataset_exists(path, impl): + if impl == 'raw': + return IndexedRawTextDataset.exists(path) + elif impl == 'mmap': + return MMapIndexedDataset.exists(path) + else: + return IndexedDataset.exists(path) + + +def read_longs(f, n): + a = np.empty(n, dtype=np.int64) + f.readinto(a) + return a + + +def write_longs(f, a): + f.write(np.array(a, dtype=np.int64)) + + +dtypes = { + 1: np.uint8, + 2: np.int8, + 3: np.int16, + 4: np.int32, + 5: np.int64, + 6: np.float, + 7: np.double, + 8: np.uint16 +} + + +def code(dtype): + for k in dtypes.keys(): + if dtypes[k] == dtype: + return k + raise ValueError(dtype) + + +def index_file_path(prefix_path): + return prefix_path + '.idx' + + +def data_file_path(prefix_path): + return prefix_path + '.bin' + + +class IndexedDataset(FairseqDataset): + """Loader for TorchNet IndexedDataset""" + _HDR_MAGIC = b'TNTIDX\x00\x00' + + def __init__(self, path, fix_lua_indexing=False): + super().__init__() + self.path = path + self.fix_lua_indexing = fix_lua_indexing + self.data_file = None + self.read_index(path) + + def read_index(self, path): + with open(index_file_path(path), 'rb') as f: + magic = f.read(8) + assert magic == self._HDR_MAGIC, ( + 'Index file doesn\'t match expected format. ' + 'Make sure that --dataset-impl is configured properly.' + ) + version = f.read(8) + assert struct.unpack('= self._len: + raise IndexError('index out of range') + + def __del__(self): + if self.data_file: + self.data_file.close() + + @lru_cache(maxsize=8) + def __getitem__(self, i): + if not self.data_file: + self.read_data(self.path) + self.check_index(i) + tensor_size = self.sizes[self.dim_offsets[i]:self.dim_offsets[i + 1]] + a = np.empty(tensor_size, dtype=self.dtype) + self.data_file.seek(self.data_offsets[i] * self.element_size) + self.data_file.readinto(a) + item = torch.from_numpy(a).long() + if self.fix_lua_indexing: + item -= 1 # subtract 1 for 0-based indexing + return item + + def __len__(self): + return self._len + + def num_tokens(self, index): + return self.sizes[index] + + def size(self, index): + return self.sizes[index] + + @staticmethod + def exists(path): + return ( + os.path.exists(index_file_path(path)) and os.path.exists(data_file_path(path)) + ) + + @property + def supports_prefetch(self): + return False # avoid prefetching to save memory + + +class IndexedCachedDataset(IndexedDataset): + + def __init__(self, path, fix_lua_indexing=False): + super().__init__(path, fix_lua_indexing=fix_lua_indexing) + self.cache = None + self.cache_index = {} + + @property + def supports_prefetch(self): + return True + + def prefetch(self, indices): + if all(i in self.cache_index for i in indices): + return + if not self.data_file: + self.read_data(self.path) + indices = sorted(set(indices)) + total_size = 0 + for i in indices: + total_size += self.data_offsets[i + 1] - self.data_offsets[i] + self.cache = np.empty(total_size, dtype=self.dtype) + ptx = 0 + self.cache_index.clear() + for i in indices: + self.cache_index[i] = ptx + size = self.data_offsets[i + 1] - self.data_offsets[i] + a = self.cache[ptx: ptx + size] + self.data_file.seek(self.data_offsets[i] * self.element_size) + self.data_file.readinto(a) + ptx += size + if self.data_file: + # close and delete data file after prefetch so we can pickle + self.data_file.close() + self.data_file = None + + @lru_cache(maxsize=8) + def __getitem__(self, i): + self.check_index(i) + tensor_size = self.sizes[self.dim_offsets[i]:self.dim_offsets[i + 1]] + a = np.empty(tensor_size, dtype=self.dtype) + ptx = self.cache_index[i] + np.copyto(a, self.cache[ptx: ptx + a.size]) + item = torch.from_numpy(a).long() + if self.fix_lua_indexing: + item -= 1 # subtract 1 for 0-based indexing + return item + + +class IndexedRawTextDataset(FairseqDataset): + """Takes a text file as input and binarizes it in memory at instantiation. + Original lines are also kept in memory""" + + def __init__(self, path, dictionary, append_eos=True, reverse_order=False): + self.tokens_list = [] + self.lines = [] + self.sizes = [] + self.append_eos = append_eos + self.reverse_order = reverse_order + self.read_data(path, dictionary) + self.size = len(self.tokens_list) + + def read_data(self, path, dictionary): + with open(path, 'r', encoding='utf-8') as f: + for line in f: + self.lines.append(line.strip('\n')) + tokens = dictionary.encode_line( + line, add_if_not_exist=False, + append_eos=self.append_eos, reverse_order=self.reverse_order, + ).long() + self.tokens_list.append(tokens) + self.sizes.append(len(tokens)) + self.sizes = np.array(self.sizes) + + def check_index(self, i): + if i < 0 or i >= self.size: + raise IndexError('index out of range') + + @lru_cache(maxsize=8) + def __getitem__(self, i): + self.check_index(i) + return self.tokens_list[i] + + def get_original_text(self, i): + self.check_index(i) + return self.lines[i] + + def __del__(self): + pass + + def __len__(self): + return self.size + + def num_tokens(self, index): + return self.sizes[index] + + def size(self, index): + return self.sizes[index] + + @staticmethod + def exists(path): + return os.path.exists(path) + + +class IndexedDatasetBuilder(object): + element_sizes = { + np.uint8: 1, + np.int8: 1, + np.int16: 2, + np.int32: 4, + np.int64: 8, + np.float: 4, + np.double: 8 + } + + def __init__(self, out_file, dtype=np.int32): + self.out_file = open(out_file, 'wb') + self.dtype = dtype + self.data_offsets = [0] + self.dim_offsets = [0] + self.sizes = [] + self.element_size = self.element_sizes[self.dtype] + + def add_item(self, tensor): + # +1 for Lua compatibility + bytes = self.out_file.write(np.array(tensor.numpy() + 1, dtype=self.dtype)) + self.data_offsets.append(self.data_offsets[-1] + bytes / self.element_size) + for s in tensor.size(): + self.sizes.append(s) + self.dim_offsets.append(self.dim_offsets[-1] + len(tensor.size())) + + def merge_file_(self, another_file): + index = IndexedDataset(another_file) + assert index.dtype == self.dtype + + begin = self.data_offsets[-1] + for offset in index.data_offsets[1:]: + self.data_offsets.append(begin + offset) + self.sizes.extend(index.sizes) + begin = self.dim_offsets[-1] + for dim_offset in index.dim_offsets[1:]: + self.dim_offsets.append(begin + dim_offset) + + with open(data_file_path(another_file), 'rb') as f: + while True: + data = f.read(1024) + if data: + self.out_file.write(data) + else: + break + + def finalize(self, index_file): + self.out_file.close() + index = open(index_file, 'wb') + index.write(b'TNTIDX\x00\x00') + index.write(struct.pack('= self.total: + return + self.n += 1 + yield x + + def __next__(self): + return next(self.itr) + + def has_next(self): + """Whether the iterator has been exhausted.""" + return self.n < len(self) + + def skip(self, num_to_skip): + """Fast-forward the iterator by skipping *num_to_skip* elements.""" + next(itertools.islice(self.itr, num_to_skip, num_to_skip), None) + return self + + def take(self, n): + """ + Truncates the iterator to n elements at most. + """ + self.total = min(self.total, n) + + # Propagate this change to the underlying iterator + if hasattr(self.iterable, "take"): + self.iterable.take(n) + + +class EpochBatchIterating(object): + def __len__(self) -> int: + raise NotImplementedError + + @property + def next_epoch_idx(self): + raise NotImplementedError + + def next_epoch_itr(self, shuffle=True, fix_batches_to_gpus=False): + """Return a new iterator over the dataset. + + Args: + shuffle (bool, optional): shuffle batches before returning the + iterator (default: True). + fix_batches_to_gpus: ensure that batches are always + allocated to the same shards across epochs. Requires + that :attr:`dataset` supports prefetching (default: False). + """ + raise NotImplementedError + + def end_of_epoch(self) -> bool: + """Returns whether the most recent epoch iterator has been exhausted""" + raise NotImplementedError + + @property + def iterations_in_epoch(self) -> int: + """The number of consumed batches in the current epoch.""" + raise NotImplementedError + + def state_dict(self): + """Returns a dictionary containing a whole state of the iterator.""" + raise NotImplementedError + + def load_state_dict(self, state_dict): + """Copies the state of the iterator from the given *state_dict*.""" + raise NotImplementedError + + +class StreamingEpochBatchIterator(EpochBatchIterating): + def __init__( + self, dataset, epoch=1, num_shards=1, shard_id=0, + ): + assert isinstance(dataset, torch.utils.data.IterableDataset) + self.dataset = dataset + self.epoch = max(epoch, 1) # we use 1-based indexing for epochs + self._current_epoch_iterator = None + self.num_shards = num_shards + self.shard_id = shard_id + + @property + def next_epoch_idx(self): + """Return the epoch index after *next_epoch_itr* is called.""" + if self._current_epoch_iterator is not None and self.end_of_epoch(): + return self.epoch + 1 + else: + return self.epoch + + def next_epoch_itr(self, shuffle=True, fix_batches_to_gpus=False): + self.epoch = self.next_epoch_idx + self.dataset.set_epoch(self.epoch) + self._current_epoch_iterator = CountingIterator( + iterable=ShardedIterator( + iterable=self.dataset, + num_shards=self.num_shards, + shard_id=self.shard_id, + ), + ) + return self._current_epoch_iterator + + def end_of_epoch(self) -> bool: + return not self._current_epoch_iterator.has_next() + + @property + def iterations_in_epoch(self) -> int: + if self._current_epoch_iterator is not None: + return self._current_epoch_iterator.n + return 0 + + def state_dict(self): + return { + 'epoch': self.epoch, + } + + def load_state_dict(self, state_dict): + self.epoch = state_dict['epoch'] + + +class EpochBatchIterator(EpochBatchIterating): + """A multi-epoch iterator over a :class:`torch.utils.data.Dataset`. + + Compared to :class:`torch.utils.data.DataLoader`, this iterator: + + - can be reused across multiple epochs with the :func:`next_epoch_itr` + method (optionally shuffled between epochs) + - can be serialized/deserialized with the :func:`state_dict` and + :func:`load_state_dict` methods + - supports sharding with the *num_shards* and *shard_id* arguments + + Args: + dataset (~torch.utils.data.Dataset): dataset from which to load the data + collate_fn (callable): merges a list of samples to form a mini-batch + batch_sampler (~torch.utils.data.Sampler or a callable): an iterator over batches of + indices, or a callable to create such an iterator (~torch.utils.data.Sampler). + A callable batch_sampler will be called for each epoch to enable per epoch dynamic + batch iterators defined by this callable batch_sampler. + seed (int, optional): seed for random number generator for + reproducibility (default: 1). + num_shards (int, optional): shard the data iterator into N + shards (default: 1). + shard_id (int, optional): which shard of the data iterator to + return (default: 0). + num_workers (int, optional): how many subprocesses to use for data + loading. 0 means the data will be loaded in the main process + (default: 0). + epoch (int, optional): the epoch to start the iterator from + (default: 1). + buffer_size (int, optional): the number of batches to keep ready in the + queue. Helps speeding up dataloading. When buffer_size is zero, the + default torch.utils.data.DataLoader preloading is used. + timeout (int, optional): if positive, the timeout value for collecting a batch + from workers. Should always be non-negative. (default: ``0``) + """ + + def __init__( + self, dataset, collate_fn, batch_sampler, seed=1, num_shards=1, shard_id=0, + num_workers=0, epoch=1, buffer_size=0, timeout=0, + ): + assert isinstance(dataset, torch.utils.data.Dataset) + self.dataset = dataset + self.collate_fn = collate_fn + self.batch_sampler = batch_sampler + self._frozen_batches = tuple(batch_sampler) if not callable(batch_sampler) else None + self.seed = seed + self.num_shards = num_shards + self.shard_id = shard_id + self.num_workers = num_workers + # This upper limit here is to prevent people from abusing this feature + # in a shared computing environment. + self.buffer_size = min(buffer_size, 20) + self.timeout = timeout + + self.epoch = max(epoch, 1) # we use 1-based indexing for epochs + self.shuffle = True + self._cur_epoch_itr = None + self._next_epoch_itr = None + self._supports_prefetch = getattr(dataset, 'supports_prefetch', False) + + @property + def frozen_batches(self): + if self._frozen_batches is None: + self._frozen_batches = tuple(self.batch_sampler(self.dataset, self.epoch)) + return self._frozen_batches + + def __len__(self): + return int(math.ceil(len(self.frozen_batches) / float(self.num_shards))) + + @property + def n(self): + return self.iterations_in_epoch + + @property + def next_epoch_idx(self): + """Return the epoch index after *next_epoch_itr* is called.""" + if self._next_epoch_itr is not None: + return self.epoch + elif self._cur_epoch_itr is not None and self.end_of_epoch(): + return self.epoch + 1 + else: + return self.epoch + + def next_epoch_itr(self, shuffle=True, fix_batches_to_gpus=False): + """Return a new iterator over the dataset. + + Args: + shuffle (bool, optional): shuffle batches before returning the + iterator (default: True). + fix_batches_to_gpus: ensure that batches are always + allocated to the same shards across epochs. Requires + that :attr:`dataset` supports prefetching (default: False). + """ + self.epoch = self.next_epoch_idx + self.dataset.set_epoch(self.epoch) + if self._next_epoch_itr is not None: + self._cur_epoch_itr = self._next_epoch_itr + self._next_epoch_itr = None + else: + if callable(self.batch_sampler): + # reset _frozen_batches to refresh the next epoch + self._frozen_batches = None + self._cur_epoch_itr = self._get_iterator_for_epoch( + self.epoch, shuffle, fix_batches_to_gpus=fix_batches_to_gpus, + ) + self.shuffle = shuffle + return self._cur_epoch_itr + + def end_of_epoch(self) -> bool: + """Returns whether the most recent epoch iterator has been exhausted""" + return not self._cur_epoch_itr.has_next() + + @property + def iterations_in_epoch(self): + """The number of consumed batches in the current epoch.""" + if self._cur_epoch_itr is not None: + return self._cur_epoch_itr.n + elif self._next_epoch_itr is not None: + return self._next_epoch_itr.n + return 0 + + def state_dict(self): + """Returns a dictionary containing a whole state of the iterator.""" + return { + 'epoch': self.epoch, + 'iterations_in_epoch': self.iterations_in_epoch, + 'shuffle': self.shuffle, + } + + def load_state_dict(self, state_dict): + """Copies the state of the iterator from the given *state_dict*.""" + self.epoch = state_dict['epoch'] + itr_pos = state_dict.get('iterations_in_epoch', 0) + if itr_pos > 0: + # fast-forward epoch iterator + self._next_epoch_itr = self._get_iterator_for_epoch( + self.epoch, + shuffle=state_dict.get('shuffle', True), + offset=itr_pos, + ) + if self._next_epoch_itr is None: + # we finished the epoch, increment epoch counter + self.epoch += 1 + else: + self._next_epoch_itr = None + + def _get_iterator_for_epoch(self, epoch, shuffle, fix_batches_to_gpus=False, offset=0): + + def shuffle_batches(batches, seed): + with data_utils.numpy_seed(seed): + np.random.shuffle(batches) + return batches + + if self._supports_prefetch: + batches = self.frozen_batches + + if shuffle and not fix_batches_to_gpus: + batches = shuffle_batches(list(batches), self.seed + epoch) + + batches = list(ShardedIterator( + batches, self.num_shards, self.shard_id, fill_value=[] + )) + self.dataset.prefetch([i for s in batches for i in s]) + + if shuffle and fix_batches_to_gpus: + batches = shuffle_batches(batches, self.seed + epoch + self.shard_id) + else: + if shuffle: + batches = shuffle_batches(list(self.frozen_batches), self.seed + epoch) + else: + batches = self.frozen_batches + batches = list(ShardedIterator( + batches, self.num_shards, self.shard_id, fill_value=[] + )) + + if offset > 0 and offset >= len(batches): + return None + + if self.num_workers > 0: + os.environ['PYTHONWARNINGS'] = 'ignore:semaphore_tracker:UserWarning' + + # Create data loader + itr = torch.utils.data.DataLoader( + self.dataset, + collate_fn=self.collate_fn, + batch_sampler=batches[offset:], + num_workers=self.num_workers, + timeout=self.timeout, + ) + + # Wrap with a BufferedIterator if needed + if self.buffer_size > 0: + itr = BufferedIterator(self.buffer_size, itr) + + # Wrap with CoutingIterator + itr = CountingIterator(itr, start=offset) + return itr + + +class GroupedIterator(CountingIterator): + """Wrapper around an iterable that returns groups (chunks) of items. + + Args: + iterable (iterable): iterable to wrap + chunk_size (int): size of each chunk + + Attributes: + n (int): number of elements consumed from this iterator + """ + + def __init__(self, iterable, chunk_size): + itr = _chunk_iterator(iterable, chunk_size) + super().__init__( + itr, + start=int(math.ceil(getattr(iterable, 'n', 0) / float(chunk_size))), + total=int(math.ceil(len(iterable) / float(chunk_size))), + ) + self.chunk_size = chunk_size + + +def _chunk_iterator(itr, chunk_size): + chunk = [] + for x in itr: + chunk.append(x) + if len(chunk) == chunk_size: + yield chunk + chunk = [] + if len(chunk) > 0: + yield chunk + + +class ShardedIterator(CountingIterator): + """A sharded wrapper around an iterable, padded to length. + + Args: + iterable (iterable): iterable to wrap + num_shards (int): number of shards to split the iterable into + shard_id (int): which shard to iterator over + fill_value (Any, optional): padding value when the iterable doesn't + evenly divide *num_shards* (default: None). + + Attributes: + n (int): number of elements consumed from this iterator + """ + + def __init__(self, iterable, num_shards, shard_id, fill_value=None): + if shard_id < 0 or shard_id >= num_shards: + raise ValueError('shard_id must be between 0 and num_shards') + sharded_len = int(math.ceil(len(iterable) / float(num_shards))) + itr = map( + operator.itemgetter(1), + itertools.zip_longest( + range(sharded_len), + itertools.islice(iterable, shard_id, len(iterable), num_shards), + fillvalue=fill_value, + ), + ) + super().__init__( + itr, + start=int(math.ceil(getattr(iterable, 'n', 0) / float(num_shards))), + total=sharded_len, + ) + + +class BackgroundConsumer(Thread): + def __init__(self, queue, source, max_len): + Thread.__init__(self) + + self._queue = queue + self._source = source + self._max_len = max_len + self.count = 0 + + def run(self): + try: + self._source_iter = iter(self._source) + for _ in range(len(self._source)): + item = next(self._source_iter) + self._queue.put(item) + + # Stop if we reached the maximum length + self.count += 1 + if self._max_len is not None and self.count >= self._max_len: + break + + # Signal the consumer we are done. + self._queue.put(_sentinel) + except Exception as e: + self._queue.put(e) + + del self._source_iter + + +class BufferedIterator(object): + def __init__(self, size, iterable): + self._queue = queue.Queue(size) + self._iterable = iterable + self.max_len = None + self._consumer = None + + self.start_time = time.time() + self.warning_time = None + + def _create_consumer(self): + self._consumer = BackgroundConsumer( + self._queue, + self._iterable, + self.max_len + ) + self._consumer.daemon = True + self._consumer.start() + + def __iter__(self): + return self + + def __len__(self): + return len(self._iterable) + + def take(self, n): + self.max_len = n + + def __next__(self): + # Create consumer if not created yet + if self._consumer is None: + self._create_consumer() + + # Notify the user if there is a data loading bottleneck + if self._queue.qsize() < min(2, max(1, self._queue.maxsize // 2)): + if time.time() - self.start_time > 5 * 60: + if self.warning_time is None or time.time() - self.warning_time > 15 * 60: + logger.info( + "Data loading buffer is empty or nearly empty. This may " + "indicate a data loading bottleneck, and increasing the " + "number of workers (--num-workers) may help." + ) + self.warning_time = time.time() + + # Get next example + item = self._queue.get(True) + if isinstance(item, Exception): + raise item + if item is _sentinel: + raise StopIteration() + return item diff --git a/fairseq/data/language_pair_dataset.py b/fairseq/data/language_pair_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..7576e07d346352b673039d651ecf0b13104b7adc --- /dev/null +++ b/fairseq/data/language_pair_dataset.py @@ -0,0 +1,420 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np +import torch + +from fairseq.data import data_utils, FairseqDataset + + +logger = logging.getLogger(__name__) + + +def collate( + samples, + pad_idx, + eos_idx, + left_pad_source=True, + left_pad_target=False, + input_feeding=True, + pad_to_length=None, +): + if len(samples) == 0: + return {} + + def merge(key, left_pad, move_eos_to_beginning=False, pad_to_length=None): + return data_utils.collate_tokens( + [s[key] for s in samples], + pad_idx, eos_idx, left_pad, move_eos_to_beginning, + pad_to_length=pad_to_length, + ) + + def check_alignment(alignment, src_len, tgt_len): + if alignment is None or len(alignment) == 0: + return False + if alignment[:, 0].max().item() >= src_len - 1 or alignment[:, 1].max().item() >= tgt_len - 1: + logger.warning("alignment size mismatch found, skipping alignment!") + return False + return True + + def compute_alignment_weights(alignments): + """ + Given a tensor of shape [:, 2] containing the source-target indices + corresponding to the alignments, a weight vector containing the + inverse frequency of each target index is computed. + For e.g. if alignments = [[5, 7], [2, 3], [1, 3], [4, 2]], then + a tensor containing [1., 0.5, 0.5, 1] should be returned (since target + index 3 is repeated twice) + """ + align_tgt = alignments[:, 1] + _, align_tgt_i, align_tgt_c = torch.unique(align_tgt, return_inverse=True, return_counts=True) + align_weights = align_tgt_c[align_tgt_i[np.arange(len(align_tgt))]] + return 1. / align_weights.float() + + id = torch.LongTensor([s['id'] for s in samples]) + src_tokens = merge( + 'source', left_pad=left_pad_source, + pad_to_length=pad_to_length['source'] if pad_to_length is not None else None + ) + # sort by descending source length + src_lengths = torch.LongTensor([ + s['source'].ne(pad_idx).long().sum() for s in samples + ]) + src_lengths, sort_order = src_lengths.sort(descending=True) + id = id.index_select(0, sort_order) + src_tokens = src_tokens.index_select(0, sort_order) + + prev_output_tokens = None + target = None + if samples[0].get('target', None) is not None: + target = merge( + 'target', left_pad=left_pad_target, + pad_to_length=pad_to_length['target'] if pad_to_length is not None else None, + ) + target = target.index_select(0, sort_order) + tgt_lengths = torch.LongTensor([ + s['target'].ne(pad_idx).long().sum() for s in samples + ]).index_select(0, sort_order) + ntokens = tgt_lengths.sum().item() + + if samples[0].get('prev_output_tokens', None) is not None: + prev_output_tokens = merge('prev_output_tokens', left_pad=left_pad_target) + elif input_feeding: + # we create a shifted version of targets for feeding the + # previous output token(s) into the next decoder step + prev_output_tokens = merge( + 'target', + left_pad=left_pad_target, + move_eos_to_beginning=True, + pad_to_length=pad_to_length['target'] if pad_to_length is not None else None, + ) + else: + ntokens = src_lengths.sum().item() + + batch = { + 'id': id, + 'nsentences': len(samples), + 'ntokens': ntokens, + 'net_input': { + 'src_tokens': src_tokens, + 'src_lengths': src_lengths, + }, + 'target': target, + } + if prev_output_tokens is not None: + batch['net_input']['prev_output_tokens'] = prev_output_tokens.index_select(0, sort_order) + + if samples[0].get('alignment', None) is not None: + bsz, tgt_sz = batch['target'].shape + src_sz = batch['net_input']['src_tokens'].shape[1] + + offsets = torch.zeros((len(sort_order), 2), dtype=torch.long) + offsets[:, 1] += (torch.arange(len(sort_order), dtype=torch.long) * tgt_sz) + if left_pad_source: + offsets[:, 0] += (src_sz - src_lengths) + if left_pad_target: + offsets[:, 1] += (tgt_sz - tgt_lengths) + + alignments = [ + alignment + offset + for align_idx, offset, src_len, tgt_len in zip(sort_order, offsets, src_lengths, tgt_lengths) + for alignment in [samples[align_idx]['alignment'].view(-1, 2)] + if check_alignment(alignment, src_len, tgt_len) + ] + + if len(alignments) > 0: + alignments = torch.cat(alignments, dim=0) + align_weights = compute_alignment_weights(alignments) + + batch['alignments'] = alignments + batch['align_weights'] = align_weights + + return batch + + +class LanguagePairDataset(FairseqDataset): + """ + A pair of torch.utils.data.Datasets. + + Args: + src (torch.utils.data.Dataset): source dataset to wrap + src_sizes (List[int]): source sentence lengths + src_dict (~fairseq.data.Dictionary): source vocabulary + tgt (torch.utils.data.Dataset, optional): target dataset to wrap + tgt_sizes (List[int], optional): target sentence lengths + tgt_dict (~fairseq.data.Dictionary, optional): target vocabulary + left_pad_source (bool, optional): pad source tensors on the left side + (default: True). + left_pad_target (bool, optional): pad target tensors on the left side + (default: False). + shuffle (bool, optional): shuffle dataset elements before batching + (default: True). + input_feeding (bool, optional): create a shifted version of the targets + to be passed into the model for teacher forcing (default: True). + remove_eos_from_source (bool, optional): if set, removes eos from end + of source if it's present (default: False). + append_eos_to_target (bool, optional): if set, appends eos to end of + target if it's absent (default: False). + align_dataset (torch.utils.data.Dataset, optional): dataset + containing alignments. + append_bos (bool, optional): if set, appends bos to the beginning of + source/target sentence. + num_buckets (int, optional): if set to a value greater than 0, then + batches will be bucketed into the given number of batch shapes. + src_lang_id (int, optional): source language ID, if set, the collated batch + will contain a field 'src_lang_id' in 'net_input' which indicates the + source language of the samples. + tgt_lang_id (int, optional): target language ID, if set, the collated batch + will contain a field 'tgt_lang_id' which indicates the target language + of the samples. + """ + + def __init__( + self, src, src_sizes, src_dict, + tgt=None, tgt_sizes=None, tgt_dict=None, + left_pad_source=True, left_pad_target=False, + shuffle=True, input_feeding=True, + remove_eos_from_source=False, append_eos_to_target=False, + align_dataset=None, + append_bos=False, eos=None, + num_buckets=0, + src_lang_id=None, + tgt_lang_id=None, + ): + if tgt_dict is not None: + assert src_dict.pad() == tgt_dict.pad() + assert src_dict.eos() == tgt_dict.eos() + assert src_dict.unk() == tgt_dict.unk() + if tgt is not None: + assert len(src) == len(tgt), "Source and target must contain the same number of examples" + self.src = src + self.tgt = tgt + self.src_sizes = np.array(src_sizes) + self.tgt_sizes = np.array(tgt_sizes) if tgt_sizes is not None else None + self.src_dict = src_dict + self.tgt_dict = tgt_dict + self.left_pad_source = left_pad_source + self.left_pad_target = left_pad_target + self.shuffle = shuffle + self.input_feeding = input_feeding + self.remove_eos_from_source = remove_eos_from_source + self.append_eos_to_target = append_eos_to_target + self.align_dataset = align_dataset + if self.align_dataset is not None: + assert self.tgt_sizes is not None, "Both source and target needed when alignments are provided" + self.append_bos = append_bos + self.eos = (eos if eos is not None else src_dict.eos()) + self.src_lang_id = src_lang_id + self.tgt_lang_id = tgt_lang_id + if num_buckets > 0: + from fairseq.data import BucketPadLengthDataset + self.src = BucketPadLengthDataset( + self.src, + sizes=self.src_sizes, + num_buckets=num_buckets, + pad_idx=self.src_dict.pad(), + left_pad=self.left_pad_source, + ) + self.src_sizes = self.src.sizes + logger.info('bucketing source lengths: {}'.format(list(self.src.buckets))) + if self.tgt is not None: + self.tgt = BucketPadLengthDataset( + self.tgt, + sizes=self.tgt_sizes, + num_buckets=num_buckets, + pad_idx=self.tgt_dict.pad(), + left_pad=self.left_pad_target, + ) + self.tgt_sizes = self.tgt.sizes + logger.info('bucketing target lengths: {}'.format(list(self.tgt.buckets))) + + # determine bucket sizes using self.num_tokens, which will return + # the padded lengths (thanks to BucketPadLengthDataset) + num_tokens = np.vectorize(self.num_tokens, otypes=[np.long]) + self.bucketed_num_tokens = num_tokens(np.arange(len(self.src))) + self.buckets = [ + (None, num_tokens) + for num_tokens in np.unique(self.bucketed_num_tokens) + ] + else: + self.buckets = None + + def get_batch_shapes(self): + return self.buckets + + def __getitem__(self, index): + tgt_item = self.tgt[index] if self.tgt is not None else None + src_item = self.src[index] + # Append EOS to end of tgt sentence if it does not have an EOS and remove + # EOS from end of src sentence if it exists. This is useful when we use + # use existing datasets for opposite directions i.e., when we want to + # use tgt_dataset as src_dataset and vice versa + if self.append_eos_to_target: + eos = self.tgt_dict.eos() if self.tgt_dict else self.src_dict.eos() + if self.tgt and self.tgt[index][-1] != eos: + tgt_item = torch.cat([self.tgt[index], torch.LongTensor([eos])]) + + if self.append_bos: + bos = self.tgt_dict.bos() if self.tgt_dict else self.src_dict.bos() + if self.tgt and self.tgt[index][0] != bos: + tgt_item = torch.cat([torch.LongTensor([bos]), self.tgt[index]]) + + bos = self.src_dict.bos() + if self.src[index][0] != bos: + src_item = torch.cat([torch.LongTensor([bos]), self.src[index]]) + + if self.remove_eos_from_source: + eos = self.src_dict.eos() + if self.src[index][-1] == eos: + src_item = self.src[index][:-1] + + example = { + 'id': index, + 'source': src_item, + 'target': tgt_item, + } + if self.align_dataset is not None: + example['alignment'] = self.align_dataset[index] + return example + + def __len__(self): + return len(self.src) + + def collater(self, samples, pad_to_length=None): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + pad_to_length (dict, optional): a dictionary of + {'source': source_pad_to_length, 'target': target_pad_to_length} + to indicate the max length to pad to in source and target respectively. + + Returns: + dict: a mini-batch with the following keys: + + - `id` (LongTensor): example IDs in the original input order + - `ntokens` (int): total number of tokens in the batch + - `net_input` (dict): the input to the Model, containing keys: + + - `src_tokens` (LongTensor): a padded 2D Tensor of tokens in + the source sentence of shape `(bsz, src_len)`. Padding will + appear on the left if *left_pad_source* is ``True``. + - `src_lengths` (LongTensor): 1D Tensor of the unpadded + lengths of each source sentence of shape `(bsz)` + - `prev_output_tokens` (LongTensor): a padded 2D Tensor of + tokens in the target sentence, shifted right by one + position for teacher forcing, of shape `(bsz, tgt_len)`. + This key will not be present if *input_feeding* is + ``False``. Padding will appear on the left if + *left_pad_target* is ``True``. + - `src_lang_id` (LongTensor): a long Tensor which contains source + language IDs of each sample in the batch + + - `target` (LongTensor): a padded 2D Tensor of tokens in the + target sentence of shape `(bsz, tgt_len)`. Padding will appear + on the left if *left_pad_target* is ``True``. + - `tgt_lang_id` (LongTensor): a long Tensor which contains target language + IDs of each sample in the batch + """ + res = collate( + samples, + pad_idx=self.src_dict.pad(), + eos_idx=self.eos, + left_pad_source=self.left_pad_source, + left_pad_target=self.left_pad_target, + input_feeding=self.input_feeding, + pad_to_length=pad_to_length, + ) + if self.src_lang_id is not None or self.tgt_lang_id is not None: + src_tokens = res['net_input']['src_tokens'] + bsz = src_tokens.size(0) + if self.src_lang_id is not None: + res['net_input']['src_lang_id'] = torch.LongTensor( + [[self.src_lang_id]] + ).expand(bsz, 1).to(src_tokens) + if self.tgt_lang_id is not None: + res['tgt_lang_id'] = torch.LongTensor( + [[self.tgt_lang_id]] + ).expand(bsz, 1).to(src_tokens) + return res + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return max(self.src_sizes[index], self.tgt_sizes[index] if self.tgt_sizes is not None else 0) + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return (self.src_sizes[index], self.tgt_sizes[index] if self.tgt_sizes is not None else 0) + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + indices = np.random.permutation(len(self)) + else: + indices = np.arange(len(self)) + if self.buckets is None: + # sort by target length, then source length + if self.tgt_sizes is not None: + indices = indices[ + np.argsort(self.tgt_sizes[indices], kind='mergesort') + ] + return indices[np.argsort(self.src_sizes[indices], kind='mergesort')] + else: + # sort by bucketed_num_tokens, which is: + # max(padded_src_len, padded_tgt_len) + return indices[ + np.argsort(self.bucketed_num_tokens[indices], kind='mergesort') + ] + + @property + def supports_prefetch(self): + return ( + getattr(self.src, 'supports_prefetch', False) + and (getattr(self.tgt, 'supports_prefetch', False) or self.tgt is None) + ) + + def prefetch(self, indices): + self.src.prefetch(indices) + if self.tgt is not None: + self.tgt.prefetch(indices) + if self.align_dataset is not None: + self.align_dataset.prefetch(indices) + + def filter_indices_by_size(self, indices, max_sizes): + """ Filter a list of sample indices. Remove those that are longer + than specified in max_sizes. + + Args: + indices (np.array): original array of sample indices + max_sizes (int or list[int] or tuple[int]): max sample size, + can be defined separately for src and tgt (then list or tuple) + + Returns: + np.array: filtered sample array + list: list of removed indices + """ + if max_sizes is None: + return indices, [] + if type(max_sizes) in (int, float): + max_src_size, max_tgt_size = max_sizes, max_sizes + else: + max_src_size, max_tgt_size = max_sizes + if self.tgt_sizes is None: + ignored = indices[self.src_sizes[indices] > max_src_size] + else: + ignored = indices[(self.src_sizes[indices] > max_src_size) | + (self.tgt_sizes[indices] > max_tgt_size)] + if len(ignored) > 0: + if self.tgt_sizes is None: + indices = indices[self.src_sizes[indices] <= max_src_size] + else: + indices = indices[(self.src_sizes[indices] <= max_src_size) & + (self.tgt_sizes[indices] <= max_tgt_size)] + return indices, ignored.tolist() diff --git a/fairseq/data/legacy/__init__.py b/fairseq/data/legacy/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1acaafeb09176dfda2a8bb30bab7e0ea764faa23 --- /dev/null +++ b/fairseq/data/legacy/__init__.py @@ -0,0 +1,15 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .masked_lm_dictionary import BertDictionary, MaskedLMDictionary +from .block_pair_dataset import BlockPairDataset +from .masked_lm_dataset import MaskedLMDataset + +__all__ = [ + 'BertDictionary', + 'BlockPairDataset', + 'MaskedLMDataset', + 'MaskedLMDictionary', +] diff --git a/fairseq/data/legacy/__pycache__/__init__.cpython-310.pyc b/fairseq/data/legacy/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3e082e96f212ee24f1a2b14ef14d019cc67655fc Binary files /dev/null and b/fairseq/data/legacy/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/data/legacy/__pycache__/block_pair_dataset.cpython-310.pyc b/fairseq/data/legacy/__pycache__/block_pair_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..de1e6d9c4ed07e64b0f03d3e32a5695d22429d8a Binary files /dev/null and b/fairseq/data/legacy/__pycache__/block_pair_dataset.cpython-310.pyc differ diff --git a/fairseq/data/legacy/__pycache__/masked_lm_dataset.cpython-310.pyc b/fairseq/data/legacy/__pycache__/masked_lm_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7f1f03ce979359de67ab2f27bf8aa51576132639 Binary files /dev/null and b/fairseq/data/legacy/__pycache__/masked_lm_dataset.cpython-310.pyc differ diff --git a/fairseq/data/legacy/__pycache__/masked_lm_dictionary.cpython-310.pyc b/fairseq/data/legacy/__pycache__/masked_lm_dictionary.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f233d68f91b112a112a9195c39af770e7d5a1563 Binary files /dev/null and b/fairseq/data/legacy/__pycache__/masked_lm_dictionary.cpython-310.pyc differ diff --git a/fairseq/data/legacy/block_pair_dataset.py b/fairseq/data/legacy/block_pair_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..b9fc8141471984f644edbc97915647897f1de617 --- /dev/null +++ b/fairseq/data/legacy/block_pair_dataset.py @@ -0,0 +1,312 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import numpy as np +import torch + +from fairseq.data import FairseqDataset + + +class BlockPairDataset(FairseqDataset): + """Break a Dataset of tokens into sentence pair blocks for next sentence + prediction as well as masked language model. + + High-level logics are: + 1. break input tensor to tensor blocks + 2. pair the blocks with 50% next sentence and 50% random sentence + 3. return paired blocks as well as related segment labels + + Args: + dataset (~torch.utils.data.Dataset): dataset to break into blocks + sizes: array of sentence lengths + dictionary: dictionary for the task + block_size: maximum block size + break_mode: mode for breaking copurs into block pairs. currently we support + 2 modes + doc: respect document boundaries and each part of the pair should belong to on document + none: don't respect any boundary and cut tokens evenly + short_seq_prob: probability for generating shorter block pairs + doc_break_size: Size for empty line separating documents. Typically 1 if + the sentences have eos, 0 otherwise. + """ + + def __init__( + self, + dataset, + dictionary, + sizes, + block_size, + break_mode="doc", + short_seq_prob=0.1, + doc_break_size=1, + ): + super().__init__() + self.dataset = dataset + self.pad = dictionary.pad() + self.eos = dictionary.eos() + self.cls = dictionary.cls() + self.mask = dictionary.mask() + self.sep = dictionary.sep() + self.break_mode = break_mode + self.dictionary = dictionary + self.short_seq_prob = short_seq_prob + self.block_indices = [] + + assert len(dataset) == len(sizes) + + if break_mode == "doc": + cur_doc = [] + for sent_id, sz in enumerate(sizes): + assert doc_break_size == 0 or sz != 0, ( + "when doc_break_size is non-zero, we expect documents to be" + "separated by a blank line with a single eos." + ) + # empty line as document separator + if sz == doc_break_size: + if len(cur_doc) == 0: + continue + self.block_indices.append(cur_doc) + cur_doc = [] + else: + cur_doc.append(sent_id) + max_num_tokens = block_size - 3 # Account for [CLS], [SEP], [SEP] + self.sent_pairs = [] + self.sizes = [] + for doc_id, doc in enumerate(self.block_indices): + self._generate_sentence_pair(doc, doc_id, max_num_tokens, sizes) + elif break_mode is None or break_mode == "none": + # each block should have half of the block size since we are constructing block pair + sent_length = (block_size - 3) // 2 + total_len = sum(dataset.sizes) + length = math.ceil(total_len / sent_length) + + def block_at(i): + start = i * sent_length + end = min(start + sent_length, total_len) + return (start, end) + + sent_indices = np.array([block_at(i) for i in range(length)]) + sent_sizes = np.array([e - s for s, e in sent_indices]) + dataset_index = self._sent_to_dataset_index(sent_sizes) + + # pair sentences + self._pair_sentences(dataset_index) + else: + raise ValueError("Invalid break_mode: " + break_mode) + + def _pair_sentences(self, dataset_index): + """ + Give a list of evenly cut blocks/sentences, pair these sentences with 50% + consecutive sentences and 50% random sentences. + This is used for none break mode + """ + # pair sentences + for sent_id, sent in enumerate(dataset_index): + next_sent_label = ( + 1 if np.random.rand() > 0.5 and sent_id != len(dataset_index) - 1 else 0 + ) + if next_sent_label: + next_sent = dataset_index[sent_id + 1] + else: + next_sent = dataset_index[ + self._skip_sampling(len(dataset_index), [sent_id, sent_id + 1]) + ] + self.sent_pairs.append((sent, next_sent, next_sent_label)) + + # The current blocks don't include the special tokens but the + # sizes already account for this + self.sizes.append(3 + sent[3] + next_sent[3]) + + def _sent_to_dataset_index(self, sent_sizes): + """ + Build index mapping block indices to the underlying dataset indices + """ + dataset_index = [] + ds_idx, ds_remaining = -1, 0 + for to_consume in sent_sizes: + sent_size = to_consume + if ds_remaining == 0: + ds_idx += 1 + ds_remaining = sent_sizes[ds_idx] + start_ds_idx = ds_idx + start_offset = sent_sizes[ds_idx] - ds_remaining + while to_consume > ds_remaining: + to_consume -= ds_remaining + ds_idx += 1 + ds_remaining = sent_sizes[ds_idx] + ds_remaining -= to_consume + dataset_index.append( + ( + start_ds_idx, # starting index in dataset + start_offset, # starting offset within starting index + ds_idx, # ending index in dataset + sent_size, # sentence length + ) + ) + assert ds_remaining == 0 + assert ds_idx == len(self.dataset) - 1 + return dataset_index + + def _generate_sentence_pair(self, doc, doc_id, max_num_tokens, sizes): + """ + Go through a single document and genrate sentence paris from it + """ + current_chunk = [] + current_length = 0 + curr = 0 + # To provide more randomness, we decrease target seq length for parts of + # samples (10% by default). Note that max_num_tokens is the hard threshold + # for batching and will never be changed. + target_seq_length = max_num_tokens + if np.random.random() < self.short_seq_prob: + target_seq_length = np.random.randint(2, max_num_tokens) + # loop through all sentences in document + while curr < len(doc): + sent_id = doc[curr] + current_chunk.append(sent_id) + current_length = sum(sizes[current_chunk]) + # split chunk and generate pair when exceed target_seq_length or + # finish the loop + if curr == len(doc) - 1 or current_length >= target_seq_length: + # split the chunk into 2 parts + a_end = 1 + if len(current_chunk) > 2: + a_end = np.random.randint(1, len(current_chunk) - 1) + sent_a = current_chunk[:a_end] + len_a = sum(sizes[sent_a]) + # generate next sentence label, note that if there is only 1 sentence + # in current chunk, label is always 0 + next_sent_label = ( + 1 if np.random.rand() > 0.5 and len(current_chunk) != 1 else 0 + ) + if not next_sent_label: + # if next sentence label is 0, sample sent_b from a random doc + target_b_length = target_seq_length - len_a + rand_doc_id = self._skip_sampling(len(self.block_indices), [doc_id]) + random_doc = self.block_indices[rand_doc_id] + random_start = np.random.randint(0, len(random_doc)) + sent_b = [] + len_b = 0 + for j in range(random_start, len(random_doc)): + sent_b.append(random_doc[j]) + len_b = sum(sizes[sent_b]) + if len_b >= target_b_length: + break + # return the second part of the chunk since it's not used + num_unused_segments = len(current_chunk) - a_end + curr -= num_unused_segments + else: + # if next sentence label is 1, use the second part of chunk as sent_B + sent_b = current_chunk[a_end:] + len_b = sum(sizes[sent_b]) + # currently sent_a and sent_B may be longer than max_num_tokens, + # truncate them and return block idx and offsets for them + sent_a, sent_b = self._truncate_sentences( + sent_a, sent_b, max_num_tokens + ) + self.sent_pairs.append((sent_a, sent_b, next_sent_label)) + self.sizes.append(3 + sent_a[3] + sent_b[3]) + current_chunk = [] + curr += 1 + + def _skip_sampling(self, total, skip_ids): + """ + Generate a random integer which is not in skip_ids. Sample range is [0, total) + TODO: ids in skip_ids should be consecutive, we can extend it to more generic version later + """ + rand_id = np.random.randint(total - len(skip_ids)) + return rand_id if rand_id < min(skip_ids) else rand_id + len(skip_ids) + + def _truncate_sentences(self, sent_a, sent_b, max_num_tokens): + """ + Trancate a pair of sentence to limit total length under max_num_tokens + Logics: + 1. Truncate longer sentence + 2. Tokens to be truncated could be at the beginning or the end of the sentnce + Returns: + Truncated sentences represented by dataset idx + """ + len_a, len_b = sum(self.dataset.sizes[sent_a]), sum(self.dataset.sizes[sent_b]) + front_cut_a = front_cut_b = end_cut_a = end_cut_b = 0 + + while True: + total_length = ( + len_a + len_b - front_cut_a - front_cut_b - end_cut_a - end_cut_b + ) + if total_length <= max_num_tokens: + break + + if len_a - front_cut_a - end_cut_a > len_b - front_cut_b - end_cut_b: + if np.random.rand() < 0.5: + front_cut_a += 1 + else: + end_cut_a += 1 + else: + if np.random.rand() < 0.5: + front_cut_b += 1 + else: + end_cut_b += 1 + + # calculate ds indices as well as offsets and return + truncated_sent_a = self._cut_sentence(sent_a, front_cut_a, end_cut_a) + truncated_sent_b = self._cut_sentence(sent_b, front_cut_b, end_cut_b) + return truncated_sent_a, truncated_sent_b + + def _cut_sentence(self, sent, front_cut, end_cut): + """ + Cut a sentence based on the numbers of tokens to be cut from beginning and end + Represent the sentence as dataset idx and return + """ + start_ds_idx, end_ds_idx, offset = sent[0], sent[-1], 0 + target_len = sum(self.dataset.sizes[sent]) - front_cut - end_cut + while front_cut > 0: + if self.dataset.sizes[start_ds_idx] > front_cut: + offset += front_cut + break + else: + front_cut -= self.dataset.sizes[start_ds_idx] + start_ds_idx += 1 + while end_cut > 0: + if self.dataset.sizes[end_ds_idx] > end_cut: + break + else: + end_cut -= self.dataset.sizes[end_ds_idx] + end_ds_idx -= 1 + return start_ds_idx, offset, end_ds_idx, target_len + + def _fetch_block(self, start_ds_idx, offset, end_ds_idx, length): + """ + Fetch a block of tokens based on its dataset idx + """ + buffer = torch.cat( + [self.dataset[idx] for idx in range(start_ds_idx, end_ds_idx + 1)] + ) + s, e = offset, offset + length + return buffer[s:e] + + def __getitem__(self, index): + block1, block2, next_sent_label = self.sent_pairs[index] + block1 = self._fetch_block(*block1) + block2 = self._fetch_block(*block2) + return block1, block2, next_sent_label + + def __len__(self): + return len(self.sizes) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + prefetch_idx = set() + for index in indices: + for block1, block2, _ in [self.sent_pairs[index]]: + for ds_idx in range(block1[0], block1[2] + 1): + prefetch_idx.add(ds_idx) + for ds_idx in range(block2[0], block2[2] + 1): + prefetch_idx.add(ds_idx) + self.dataset.prefetch(prefetch_idx) diff --git a/fairseq/data/legacy/masked_lm_dataset.py b/fairseq/data/legacy/masked_lm_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..953aa85dd42bcaacd2b0ff2be9309395c212eb13 --- /dev/null +++ b/fairseq/data/legacy/masked_lm_dataset.py @@ -0,0 +1,322 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import numpy as np +import torch + +from typing import Dict, List, Tuple + +from fairseq.data import FairseqDataset, data_utils + +from fairseq.data import Dictionary +from fairseq.data.legacy.block_pair_dataset import BlockPairDataset +from fairseq.data.token_block_dataset import TokenBlockDataset +from fairseq.data.concat_dataset import ConcatDataset + + +class MaskedLMDataset(FairseqDataset): + """ + A wrapper Dataset for masked language modelling. The dataset + wraps around TokenBlockDataset or BlockedPairDataset and creates a batch + where the input blocks are masked according to the specified masking + probability. Additionally the batch can also contain sentence level targets + if this is specified. + + Args: + dataset: Dataset which generates blocks of data. Only BlockPairDataset + and TokenBlockDataset are supported. + sizes: Sentence lengths + vocab: Dictionary with the vocabulary and special tokens. + pad_idx: Id of padding token in dictionary + mask_idx: Id of mask token in dictionary + classif_token_idx: Id of classification token in dictionary. This is the + token associated with the sentence embedding (Eg: CLS for BERT) + sep_token_idx: Id of separator token in dictionary + (Eg: SEP in BERT) + seed: Seed for random number generator for reproducibility. + shuffle: Shuffle the elements before batching. + has_pairs: Specifies whether the underlying dataset + generates a pair of blocks along with a sentence_target or not. + Setting it to True assumes that the underlying dataset generates a + label for the pair of sentences which is surfaced as + sentence_target. The default value assumes a single block with no + sentence target. + segment_id: An optional segment id for filling in the segment labels + when we are in the single block setting (Eg: XLM). Default is 0. + masking_ratio: specifies what percentage of the blocks should be masked. + masking_prob: specifies the probability of a given token being + replaced with the "MASK" token. + random_token_prob: specifies the probability of a given token being + replaced by a random token from the vocabulary. + """ + + def __init__( + self, + dataset: FairseqDataset, + sizes: np.ndarray, + vocab: Dictionary, + pad_idx: int, + mask_idx: int, + classif_token_idx: int, + sep_token_idx: int, + seed: int = 1, + shuffle: bool = True, + has_pairs: bool = True, + segment_id: int = 0, + masking_ratio: float = 0.15, + masking_prob: float = 0.8, + random_token_prob: float = 0.1 + ): + # Make sure the input datasets are the ones supported + assert ( + isinstance(dataset, TokenBlockDataset) or + isinstance(dataset, BlockPairDataset) or + isinstance(dataset, ConcatDataset) + ), "MaskedLMDataset only wraps TokenBlockDataset or BlockPairDataset or " \ + "ConcatDataset" + + self.dataset = dataset + self.sizes = np.array(sizes) + self.vocab = vocab + self.pad_idx = pad_idx + self.mask_idx = mask_idx + self.classif_token_idx = classif_token_idx + self.sep_token_idx = sep_token_idx + self.shuffle = shuffle + self.seed = seed + self.has_pairs = has_pairs + self.segment_id = segment_id + self.masking_ratio = masking_ratio + self.masking_prob = masking_prob + self.random_token_prob = random_token_prob + + # If we have only one block then sizes needs to be updated to include + # the classification token + if not has_pairs: + self.sizes = self.sizes + 1 + + def __getitem__( + self, + index: int + ): + # if has_pairs, then expect 2 blocks and a sentence target + if self.has_pairs: + (block_one, block_two, sentence_target) = self.dataset[index] + else: + block_one = self.dataset[index] + + return { + "id": index, + "block_one": block_one, + "block_two": block_two if self.has_pairs else None, + "sentence_target": sentence_target if self.has_pairs else None, + } + + def __len__(self): + return len(self.dataset) + + def _mask_block( + self, + sentence: np.ndarray, + mask_idx: int, + pad_idx: int, + dictionary_token_range: Tuple, + ): + """ + Mask tokens for Masked Language Model training + Samples mask_ratio tokens that will be predicted by LM. + + Note:This function may not be efficient enough since we had multiple + conversions between np and torch, we can replace them with torch + operators later. + + Args: + sentence: 1d tensor to be masked + mask_idx: index to use for masking the sentence + pad_idx: index to use for masking the target for tokens we aren't + predicting + dictionary_token_range: range of indices in dictionary which can + be used for random word replacement + (e.g. without special characters) + Return: + masked_sent: masked sentence + target: target with words which we are not predicting replaced + by pad_idx + """ + masked_sent = np.copy(sentence) + sent_length = len(sentence) + mask_num = math.ceil(sent_length * self.masking_ratio) + mask = np.random.choice(sent_length, mask_num, replace=False) + target = np.copy(sentence) + + for i in range(sent_length): + if i in mask: + rand = np.random.random() + + # replace with mask if probability is less than masking_prob + # (Eg: 0.8) + if rand < self.masking_prob: + masked_sent[i] = mask_idx + + # replace with random token if probability is less than + # masking_prob + random_token_prob (Eg: 0.9) + elif rand < (self.masking_prob + self.random_token_prob): + # sample random token from dictionary + masked_sent[i] = ( + np.random.randint( + dictionary_token_range[0], dictionary_token_range[1] + ) + ) + else: + target[i] = pad_idx + + return masked_sent, target + + def _collate( + self, + samples: List[Dict], + pad_idx: int, + eos_idx: int + ): + """ + Does the heavy lifting for creating a batch from the input list of + examples. The logic is as follows: + 1. Mask the input blocks. In case has_pair is True then we have 2 + blocks to mask. + 2. Prepend the first masked block tensor with the special token + used as sentence embedding. Eg: CLS in BERT. This happens + irrespective of the value of has_pair. + 3. If has_pair is True, then append the first masked block with the + special separator token (eg: SEP for BERT) and compute segment + label accordingly. In this case, also append the second masked + block with this special separator token and compute its segment + label. + 4. For the targets tensor, prepend and append with padding index + accordingly. + 5. Concatenate all tensors. + """ + if len(samples) == 0: + return {} + # To ensure determinism, we reset the state of the PRNG after every + # batch based on the seed and the first id of the batch. This ensures + # that across epochs we get the same mask for the same example. This + # is needed for reproducibility and is how BERT does masking + # TODO: Can we add deteminism without this constraint? + with data_utils.numpy_seed(self.seed + samples[0]["id"]): + for s in samples: + + # token range is needed for replacing with random token during + # masking + token_range = (self.vocab.nspecial, len(self.vocab)) + + # mask according to specified probabilities. + masked_blk_one, masked_tgt_one = self._mask_block( + s["block_one"], self.mask_idx, self.pad_idx, token_range, + ) + + tokens = np.concatenate([ + [self.classif_token_idx], masked_blk_one + ]) + targets = np.concatenate([[self.pad_idx], masked_tgt_one]) + segments = np.ones(len(tokens)) * self.segment_id + + # if has_pairs is True then we need to add the SEP token to both + # the blocks after masking and re-compute segments based on the new + # lengths. + if self.has_pairs: + tokens_one = np.concatenate([tokens, [self.sep_token_idx]]) + targets_one = np.concatenate([targets, [self.pad_idx]]) + + masked_blk_two, masked_tgt_two = self._mask_block( + s["block_two"], self.mask_idx, self.pad_idx, token_range) + tokens_two = np.concatenate( + [masked_blk_two, [self.sep_token_idx]]) + targets_two = np.concatenate([masked_tgt_two, [self.pad_idx]]) + + # block + 1 sep + 1 special (CLS) + segments_one = np.zeros(len(tokens_one)) + # block + 1 sep + segments_two = np.ones(len(tokens_two)) + + tokens = np.concatenate([tokens_one, tokens_two]) + targets = np.concatenate([targets_one, targets_two]) + segments = np.concatenate([segments_one, segments_two]) + + s["source"] = torch.LongTensor(tokens) + s["segment_labels"] = torch.LongTensor(segments) + s["lm_target"] = torch.LongTensor(targets) + + def merge(key): + return data_utils.collate_tokens( + [s[key] for s in samples], pad_idx, eos_idx, left_pad=False + ) + return { + "id": torch.LongTensor([s["id"] for s in samples]), + "ntokens": sum(len(s["source"]) for s in samples), + "net_input": { + "src_tokens": merge("source"), + "segment_labels": merge("segment_labels"), + }, + "lm_target": merge("lm_target"), + "sentence_target": torch.LongTensor( + [s["sentence_target"] for s in samples] + ) if self.has_pairs else None, + "nsentences": len(samples), + } + + def collater( + self, + samples: List[Dict] + ): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch of data + """ + return self._collate(samples, self.vocab.pad(), self.vocab.eos()) + + def num_tokens( + self, + index: int + ): + """ + Return the number of tokens in a sample. This value is used to + enforce max-tokens during batching. + """ + return self.sizes[index] + + def size( + self, + index: int + ): + """ + Return an example's size as a float or tuple. This value is used when + filtering a dataset with max-positions. + """ + return self.sizes[index] + + def ordered_indices(self): + """ + Return an ordered list of indices. Batches will be constructed based + on this order. + """ + if self.shuffle: + return np.random.permutation(len(self)) + else: + order = [np.arange(len(self))] + order.append(self.sizes) + return np.lexsort(order) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + self.dataset.prefetch(indices) diff --git a/fairseq/data/legacy/masked_lm_dictionary.py b/fairseq/data/legacy/masked_lm_dictionary.py new file mode 100644 index 0000000000000000000000000000000000000000..bff4bcb5ec085ced94aa198dff9c9a3bfe61b1ca --- /dev/null +++ b/fairseq/data/legacy/masked_lm_dictionary.py @@ -0,0 +1,58 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data import Dictionary + + +class MaskedLMDictionary(Dictionary): + """ + Dictionary for Masked Language Modelling tasks. This extends Dictionary by + adding the mask symbol. + """ + def __init__( + self, + pad='', + eos='', + unk='', + mask='', + ): + super().__init__(pad=pad, eos=eos, unk=unk) + self.mask_word = mask + self.mask_index = self.add_symbol(mask) + self.nspecial = len(self.symbols) + + def mask(self): + """Helper to get index of mask symbol""" + return self.mask_index + + +class BertDictionary(MaskedLMDictionary): + """ + Dictionary for BERT task. This extends MaskedLMDictionary by adding support + for cls and sep symbols. + """ + def __init__( + self, + pad='', + eos='', + unk='', + mask='', + cls='', + sep='' + ): + super().__init__(pad=pad, eos=eos, unk=unk, mask=mask) + self.cls_word = cls + self.sep_word = sep + self.cls_index = self.add_symbol(cls) + self.sep_index = self.add_symbol(sep) + self.nspecial = len(self.symbols) + + def cls(self): + """Helper to get index of cls symbol""" + return self.cls_index + + def sep(self): + """Helper to get index of sep symbol""" + return self.sep_index diff --git a/fairseq/data/list_dataset.py b/fairseq/data/list_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..b96bba343747280e1867d8368449458de1c91b1d --- /dev/null +++ b/fairseq/data/list_dataset.py @@ -0,0 +1,33 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class ListDataset(BaseWrapperDataset): + + def __init__(self, dataset, sizes=None): + super().__init__(dataset) + self._sizes = sizes + + def __iter__(self): + for x in self.dataset: + yield x + + def collater(self, samples): + return samples + + @property + def sizes(self): + return self._sizes + + def num_tokens(self, index): + return self.sizes[index] + + def size(self, index): + return self.sizes[index] + + def set_epoch(self, epoch): + pass diff --git a/fairseq/data/lm_context_window_dataset.py b/fairseq/data/lm_context_window_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..17ba08bc7f5a7026e234967b20cde4e9d53a0069 --- /dev/null +++ b/fairseq/data/lm_context_window_dataset.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from fairseq.data.monolingual_dataset import MonolingualDataset + +from . import FairseqDataset + + +class LMContextWindowDataset(FairseqDataset): + """Wraps a MonolingualDataset and provides more context for evaluation.""" + + def __init__(self, dataset, tokens_per_sample, context_window, pad_idx): + assert isinstance(dataset, MonolingualDataset) + assert context_window > 0 + self.dataset = dataset + self.tokens_per_sample = tokens_per_sample + self.context_window = context_window + self.pad_idx = pad_idx + self.prev_tokens = np.empty([0]) + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples): + sample = self.dataset.collater(samples) + + pad = self.pad_idx + max_sample_len = self.tokens_per_sample + self.context_window + + bsz, tsz = sample['net_input']['src_tokens'].shape + start_idxs = [0] * bsz + toks = sample['net_input']['src_tokens'] + lengths = sample['net_input']['src_lengths'] + tgt = sample['target'] + new_toks = np.empty([bsz, tsz + self.context_window], dtype=np.int64) + new_tgt = np.full([bsz, tsz + self.context_window], pad, dtype=np.int64) + sample_lens = toks.ne(pad).long().sum(dim=1).cpu() + for i in range(bsz): + sample_len = sample_lens[i] + extra = len(self.prev_tokens) + sample_len - max_sample_len + if extra > 0: + self.prev_tokens = self.prev_tokens[extra:] + pads = np.full(self.context_window - len(self.prev_tokens), pad) + new_toks[i] = np.concatenate([self.prev_tokens, toks[i].numpy(), pads]) + new_tgt[i, len(self.prev_tokens):len(self.prev_tokens) + len(tgt[i])] = tgt[i] + start_idxs[i] = len(self.prev_tokens) + lengths[i] += len(self.prev_tokens) + self.prev_tokens = new_toks[i][new_toks[i] != pad][-self.context_window:] + sample['net_input']['src_tokens'] = torch.from_numpy(new_toks) + sample['target'] = torch.from_numpy(new_tgt) + sample['start_indices'] = start_idxs + + return sample + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + return self.dataset.size(index) + + def ordered_indices(self): + # NOTE we don't shuffle the data to retain access to the previous dataset elements + return np.arange(len(self.dataset)) + + @property + def supports_prefetch(self): + return getattr(self.dataset, 'supports_prefetch', False) + + def prefetch(self, indices): + return self.dataset.prefetch(indices) diff --git a/fairseq/data/lru_cache_dataset.py b/fairseq/data/lru_cache_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..833a2c75cb7c2280dd63a80c020f16b385f9317d --- /dev/null +++ b/fairseq/data/lru_cache_dataset.py @@ -0,0 +1,22 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from functools import lru_cache + +from . import BaseWrapperDataset + + +class LRUCacheDataset(BaseWrapperDataset): + + def __init__(self, dataset, token=None): + super().__init__(dataset) + + @lru_cache(maxsize=8) + def __getitem__(self, index): + return self.dataset[index] + + @lru_cache(maxsize=8) + def collater(self, samples): + return self.dataset.collater(samples) diff --git a/fairseq/data/mask_tokens_dataset.py b/fairseq/data/mask_tokens_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..28bc3bc9cf0bf32c22560f3c5d4c26b41fae8684 --- /dev/null +++ b/fairseq/data/mask_tokens_dataset.py @@ -0,0 +1,173 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from functools import lru_cache + +import numpy as np +import torch + +from fairseq.data import data_utils, Dictionary + +from . import BaseWrapperDataset, LRUCacheDataset + + +class MaskTokensDataset(BaseWrapperDataset): + """ + A wrapper Dataset for masked language modeling. + + Input items are masked according to the specified masking probability. + + Args: + dataset: Dataset to wrap. + sizes: Sentence lengths + vocab: Dictionary with the vocabulary and special tokens. + pad_idx: Id of pad token in vocab + mask_idx: Id of mask token in vocab + return_masked_tokens: controls whether to return the non-masked tokens + (the default) or to return a tensor with the original masked token + IDs (and *pad_idx* elsewhere). The latter is useful as targets for + masked LM training. + seed: Seed for random number generator for reproducibility. + mask_prob: probability of replacing a token with *mask_idx*. + leave_unmasked_prob: probability that a masked token is unmasked. + random_token_prob: probability of replacing a masked token with a + random token from the vocabulary. + freq_weighted_replacement: sample random replacement words based on + word frequencies in the vocab. + mask_whole_words: only mask whole words. This should be a byte mask + over vocab indices, indicating whether it is the beginning of a + word. We will extend any mask to encompass the whole word. + bpe: BPE to use for whole-word masking. + """ + + @classmethod + def apply_mask(cls, dataset: torch.utils.data.Dataset, *args, **kwargs): + """Return the source and target datasets for masked LM training.""" + dataset = LRUCacheDataset(dataset) + return ( + LRUCacheDataset(cls(dataset, *args, **kwargs, return_masked_tokens=False)), + LRUCacheDataset(cls(dataset, *args, **kwargs, return_masked_tokens=True)), + ) + + def __init__( + self, + dataset: torch.utils.data.Dataset, + vocab: Dictionary, + pad_idx: int, + mask_idx: int, + return_masked_tokens: bool = False, + seed: int = 1, + mask_prob: float = 0.15, + leave_unmasked_prob: float = 0.1, + random_token_prob: float = 0.1, + freq_weighted_replacement: bool = False, + mask_whole_words: torch.Tensor = None, + ): + assert 0.0 < mask_prob < 1.0 + assert 0.0 <= random_token_prob <= 1.0 + assert 0.0 <= leave_unmasked_prob <= 1.0 + assert random_token_prob + leave_unmasked_prob <= 1.0 + + self.dataset = dataset + self.vocab = vocab + self.pad_idx = pad_idx + self.mask_idx = mask_idx + self.return_masked_tokens = return_masked_tokens + self.seed = seed + self.mask_prob = mask_prob + self.leave_unmasked_prob = leave_unmasked_prob + self.random_token_prob = random_token_prob + self.mask_whole_words = mask_whole_words + + if random_token_prob > 0.0: + if freq_weighted_replacement: + weights = np.array(self.vocab.count) + else: + weights = np.ones(len(self.vocab)) + weights[:self.vocab.nspecial] = 0 + self.weights = weights / weights.sum() + + self.epoch = 0 + + def set_epoch(self, epoch, **unused): + super().set_epoch(epoch) + self.epoch = epoch + + @lru_cache(maxsize=8) + def __getitem__(self, index: int): + with data_utils.numpy_seed(self.seed, self.epoch, index): + item = self.dataset[index] + sz = len(item) + + assert self.mask_idx not in item, \ + 'Dataset contains mask_idx (={}), this is not expected!'.format( + self.mask_idx, + ) + + if self.mask_whole_words is not None: + word_begins_mask = self.mask_whole_words.gather(0, item) + word_begins_idx = word_begins_mask.nonzero().view(-1) + sz = len(word_begins_idx) + words = np.split(word_begins_mask, word_begins_idx)[1:] + assert len(words) == sz + word_lens = list(map(len, words)) + + # decide elements to mask + mask = np.full(sz, False) + num_mask = int( + # add a random number for probabilistic rounding + self.mask_prob * sz + np.random.rand() + ) + mask[np.random.choice(sz, num_mask, replace=False)] = True + + if self.return_masked_tokens: + # exit early if we're just returning the masked tokens + # (i.e., the targets for masked LM training) + if self.mask_whole_words is not None: + mask = np.repeat(mask, word_lens) + new_item = np.full(len(mask), self.pad_idx) + new_item[mask] = item[torch.from_numpy(mask.astype(np.uint8)) == 1] + return torch.from_numpy(new_item) + + # decide unmasking and random replacement + rand_or_unmask_prob = self.random_token_prob + self.leave_unmasked_prob + if rand_or_unmask_prob > 0.0: + rand_or_unmask = mask & (np.random.rand(sz) < rand_or_unmask_prob) + if self.random_token_prob == 0.0: + unmask = rand_or_unmask + rand_mask = None + elif self.leave_unmasked_prob == 0.0: + unmask = None + rand_mask = rand_or_unmask + else: + unmask_prob = self.leave_unmasked_prob / rand_or_unmask_prob + decision = np.random.rand(sz) < unmask_prob + unmask = rand_or_unmask & decision + rand_mask = rand_or_unmask & (~decision) + else: + unmask = rand_mask = None + + if unmask is not None: + mask = mask ^ unmask + + if self.mask_whole_words is not None: + mask = np.repeat(mask, word_lens) + + new_item = np.copy(item) + new_item[mask] = self.mask_idx + if rand_mask is not None: + num_rand = rand_mask.sum() + if num_rand > 0: + if self.mask_whole_words is not None: + rand_mask = np.repeat(rand_mask, word_lens) + num_rand = rand_mask.sum() + + new_item[rand_mask] = np.random.choice( + len(self.vocab), + num_rand, + p=self.weights, + ) + + return torch.from_numpy(new_item) diff --git a/fairseq/data/monolingual_dataset.py b/fairseq/data/monolingual_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..76c377237470fe01331bddc1a5d9f059641e817e --- /dev/null +++ b/fairseq/data/monolingual_dataset.py @@ -0,0 +1,200 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import data_utils, FairseqDataset + + +def collate(samples, pad_idx, eos_idx): + if len(samples) == 0: + return {} + + def merge(key, is_list=False): + if is_list: + res = [] + for i in range(len(samples[0][key])): + res.append(data_utils.collate_tokens( + [s[key][i] for s in samples], pad_idx, eos_idx, left_pad=False, + )) + return res + else: + return data_utils.collate_tokens( + [s[key] for s in samples], pad_idx, eos_idx, left_pad=False, + ) + + src_tokens = merge('source') + if samples[0]['target'] is not None: + is_target_list = isinstance(samples[0]['target'], list) + target = merge('target', is_target_list) + else: + target = src_tokens + + return { + 'id': torch.LongTensor([s['id'] for s in samples]), + 'nsentences': len(samples), + 'ntokens': sum(len(s['source']) for s in samples), + 'net_input': { + 'src_tokens': src_tokens, + 'src_lengths': torch.LongTensor([ + s['source'].numel() for s in samples + ]), + }, + 'target': target, + } + + +class MonolingualDataset(FairseqDataset): + """ + A wrapper around torch.utils.data.Dataset for monolingual data. + + Args: + dataset (torch.utils.data.Dataset): dataset to wrap + sizes (List[int]): sentence lengths + vocab (~fairseq.data.Dictionary): vocabulary + shuffle (bool, optional): shuffle the elements before batching + (default: True). + """ + + def __init__(self, dataset, sizes, src_vocab, tgt_vocab, add_eos_for_other_targets, shuffle, + targets=None, add_bos_token=False): + self.dataset = dataset + self.sizes = np.array(sizes) + self.vocab = src_vocab + self.tgt_vocab = tgt_vocab + self.add_eos_for_other_targets = add_eos_for_other_targets + self.shuffle = shuffle + self.add_bos_token = add_bos_token + + assert targets is None or all(t in {'self', 'future', 'past'} for t in targets), \ + "targets must be none or one of 'self', 'future', 'past'" + if targets is not None and len(targets) == 0: + targets = None + self.targets = targets + + def __getitem__(self, index): + if self.targets is not None: + # *future_target* is the original sentence + # *source* is shifted right by 1 (maybe left-padded with eos) + # *past_target* is shifted right by 2 (left-padded as needed) + # + # Left-to-right language models should condition on *source* and + # predict *future_target*. + # Right-to-left language models should condition on *source* and + # predict *past_target*. + source, future_target, past_target = self.dataset[index] + source, target = self._make_source_target(source, future_target, past_target) + else: + source = self.dataset[index] + target = None + source, target = self._maybe_add_bos(source, target) + return {'id': index, 'source': source, 'target': target} + + def __len__(self): + return len(self.dataset) + + def _make_source_target(self, source, future_target, past_target): + if self.targets is not None: + target = [] + + if self.add_eos_for_other_targets and (('self' in self.targets) or ('past' in self.targets)) \ + and source[-1] != self.vocab.eos(): + # append eos at the end of source + source = torch.cat([source, source.new([self.vocab.eos()])]) + + if 'future' in self.targets: + future_target = torch.cat([future_target, future_target.new([self.vocab.pad()])]) + if 'past' in self.targets: + # first token is before the start of sentence which is only used in "none" break mode when + # add_eos_for_other_targets is False + past_target = torch.cat([past_target.new([self.vocab.pad()]), past_target[1:], source[-2, None]]) + + for t in self.targets: + if t == 'self': + target.append(source) + elif t == 'future': + target.append(future_target) + elif t == 'past': + target.append(past_target) + else: + raise Exception('invalid target ' + t) + + if len(target) == 1: + target = target[0] + else: + target = future_target + + return source, self._filter_vocab(target) + + def _maybe_add_bos(self, source, target): + if self.add_bos_token: + source = torch.cat([source.new([self.vocab.bos()]), source]) + if target is not None: + target = torch.cat([target.new([self.tgt_vocab.bos()]), target]) + return source, target + + def _filter_vocab(self, target): + if len(self.tgt_vocab) != len(self.vocab): + def _filter(target): + mask = target.ge(len(self.tgt_vocab)) + if mask.any(): + target[mask] = self.tgt_vocab.unk() + return target + + if isinstance(target, list): + return [_filter(t) for t in target] + return _filter(target) + return target + + def collater(self, samples): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch with the following keys: + + - `id` (LongTensor): example IDs in the original input order + - `ntokens` (int): total number of tokens in the batch + - `net_input` (dict): the input to the Model, containing keys: + + - `src_tokens` (LongTensor): a padded 2D Tensor of tokens in + the source sentence of shape `(bsz, src_len)`. Padding will + appear on the right. + + - `target` (LongTensor): a padded 2D Tensor of tokens in the + target sentence of shape `(bsz, tgt_len)`. Padding will appear + on the right. + """ + return collate(samples, self.vocab.pad(), self.vocab.eos()) + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return self.sizes[index] + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return self.sizes[index] + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + order.append(self.sizes) + return np.lexsort(order) + + @property + def supports_prefetch(self): + return getattr(self.dataset, 'supports_prefetch', False) + + def prefetch(self, indices): + self.dataset.prefetch(indices) diff --git a/fairseq/data/multi_corpus_dataset.py b/fairseq/data/multi_corpus_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..02d269a17c176187fea30680c33b22a0850326d4 --- /dev/null +++ b/fairseq/data/multi_corpus_dataset.py @@ -0,0 +1,149 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from collections import OrderedDict +from typing import Dict, List + +import numpy as np +from fairseq.data import data_utils + +from . import FairseqDataset + + +logger = logging.getLogger(__name__) + + +class MultiCorpusDataset(FairseqDataset): + """ + Stores multiple instances of FairseqDataset together. Requires each instance + to be the same dataset, as the collate method needs to work on batches with + samples from each dataset. + + Allows specifying a distribution over the datasets to use. Note that unlike + MultiCorpusSampledDataset, this distribution allows sampling for each item, + rather than on a batch level. + + Each time ordered_indices() is called, a new sample is generated with + the specified distribution. + + Args: + datasets: a OrderedDict of FairseqDataset instances. + distribution: a List containing the probability of getting an utterance from + corresponding dataset + """ + + def __init__( + self, datasets: Dict[str, FairseqDataset], distribution: List[float], seed: int + ): + super().__init__() + assert isinstance(datasets, OrderedDict) + assert len(datasets) == len(distribution) + self.datasets = datasets + self.distribution = distribution + self.seed = seed + + # Avoid repeated conversions to list later + self.dataset_list = list(datasets.values()) + self.total_num_instances = 0 + + first_dataset = list(self.datasets.values())[0] + + self.dataset_offsets = [] + for dataset in datasets.values(): + assert isinstance(dataset, FairseqDataset) + assert type(dataset) is type(first_dataset) + self.dataset_offsets.append(self.total_num_instances) + self.total_num_instances += len(dataset) + + def ordered_indices(self): + with data_utils.numpy_seed(self.seed, self.epoch): + # Used to store the order of indices of each dataset to use + indices = [ + np.random.permutation(len(dataset)) + for dataset in self.datasets.values() + ] + # Keep track of which samples we've used for each dataset + counters = [0 for _ in self.datasets] + + return np.array( + [ + self._sample(indices, counters) + for _ in range(self.total_num_instances) + ], + dtype=np.int64, + ) + + def _sample(self, indices, counters): + # First pick dataset + dataset_idx = np.random.choice(len(self.distribution), p=self.distribution) + + # Then get dataset internal index + idx = indices[dataset_idx][counters[dataset_idx]] + + # Convert to multi-datasets index + idx += self.dataset_offsets[dataset_idx] + + counters[dataset_idx] += 1 + + # Reset if we reach end + if counters[dataset_idx] == len(self.dataset_list[dataset_idx]): + counters[dataset_idx] = 0 + indices[dataset_idx] = np.random.permutation( + len(self.dataset_list[dataset_idx]) + ) + + return idx + + def _map_index(self, index: int): + """ + If dataset A has length N and dataset B has length M + then index 1 maps to index 1 of dataset A, and index N + 1 + maps to index 1 of B. + """ + counter = 0 + for key, dataset in self.datasets.items(): + if index < counter + len(dataset): + return index - counter, key + counter += len(dataset) + raise ValueError( + "Invalid index: {}, max: {}".format(index, self.total_num_instances) + ) + + def __len__(self): + """ + Length of this dataset is the sum of individual datasets + """ + return self.total_num_instances + + def __getitem__(self, index): + index, key = self._map_index(index) + return self.datasets[key][index] + + def collater(self, samples): + """ + Since we enforce all datsets to be the same, collating is just + picking the first one and doing collate. + """ + if len(samples) == 0: + return None + + return list(self.datasets.values())[0].collater(samples) + + def num_tokens(self, index: int): + index, key = self._map_index(index) + return self.datasets[key].num_tokens(index) + + def size(self, index: int): + index, key = self._map_index(index) + return self.datasets[key].size(index) + + def set_epoch(self, epoch, **unused): + super().set_epoch(epoch) + self.epoch = epoch + + @property + def supports_prefetch(self): + return False diff --git a/fairseq/data/multi_corpus_sampled_dataset.py b/fairseq/data/multi_corpus_sampled_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..ad8e951cc905a73fea28b4fac449e307cadfa52f --- /dev/null +++ b/fairseq/data/multi_corpus_sampled_dataset.py @@ -0,0 +1,145 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict +from typing import Callable, Dict, List + +import numpy as np + +from . import FairseqDataset + + +def uniform_sampler(x): + # Sample from uniform distribution + return np.random.choice(x, 1).item() + + +class MultiCorpusSampledDataset(FairseqDataset): + """ + Stores multiple instances of FairseqDataset together and in every iteration + creates a batch by first sampling a dataset according to a specified + probability distribution and then getting instances from that dataset. + + Args: + datasets: an OrderedDict of FairseqDataset instances. + sampling_func: A function for sampling over list of dataset keys. + The default strategy is to sample uniformly. + """ + + def __init__( + self, + datasets: Dict[str, FairseqDataset], + sampling_func: Callable[[List], int] = None, + ): + super().__init__() + assert isinstance(datasets, OrderedDict) + self.datasets = datasets + if sampling_func is None: + sampling_func = uniform_sampler + self.sampling_func = sampling_func + + self.total_num_instances = 0 + for _, dataset in datasets.items(): + assert isinstance(dataset, FairseqDataset) + self.total_num_instances += len(dataset) + + self._ordered_indices = None + + def __len__(self): + """ + Length of this dataset is the sum of individual datasets + """ + return self.total_num_instances + + def ordered_indices(self): + """ + Ordered indices for batching. Here we call the underlying + dataset's ordered_indices() so that we get the same random ordering + as we would have from using the underlying dataset directly. + """ + if self._ordered_indices is None: + self._ordered_indices = OrderedDict( + [ + (key, dataset.ordered_indices()) + for key, dataset in self.datasets.items() + ] + ) + return np.arange(len(self)) + + def _map_index_to_dataset(self, key: int, index: int): + """ + Different underlying datasets have different lengths. In order to ensure + we are not accessing an index outside the range of the current dataset + size, we wrap around. This function should be called after we have + created an ordering for this and all underlying datasets. + """ + assert ( + self._ordered_indices is not None + ), "Must call MultiCorpusSampledDataset.ordered_indices() first" + mapped_index = index % len(self.datasets[key]) + return self._ordered_indices[key][mapped_index] + + def __getitem__(self, index: int): + """ + Get the item associated with index from each underlying dataset. + Since index is in the range of [0, TotalNumInstances], we need to + map the index to the dataset before retrieving the item. + """ + return OrderedDict( + [ + (key, dataset[self._map_index_to_dataset(key, index)]) + for key, dataset in self.datasets.items() + ] + ) + + def collater(self, samples: List[Dict]): + """ + Generate a mini-batch for this dataset. + To convert this into a regular mini-batch we use the following + logic: + 1. Select a dataset using the specified probability distribution. + 2. Call the collater function of the selected dataset. + """ + if len(samples) == 0: + return None + + selected_key = self.sampling_func(list(self.datasets.keys())) + selected_samples = [sample[selected_key] for sample in samples] + return self.datasets[selected_key].collater(selected_samples) + + def num_tokens(self, index: int): + """ + Return an example's length (number of tokens), used for batching. Here + we return the max across all examples at index across all underlying + datasets. + """ + return max( + dataset.num_tokens(self._map_index_to_dataset(key, index)) + for key, dataset in self.datasets.items() + ) + + def size(self, index: int): + """ + Return an example's size as a float or tuple. Here we return the max + across all underlying datasets. This value is used when filtering a + dataset with max-positions. + """ + return max( + dataset.size(self._map_index_to_dataset(key, index)) + for key, dataset in self.datasets.items() + ) + + @property + def supports_prefetch(self): + return all( + getattr(dataset, "supports_prefetch", False) + for dataset in self.datasets.values() + ) + + def prefetch(self, indices): + for key, dataset in self.datasets.items(): + dataset.prefetch( + [self._map_index_to_dataset(key, index) for index in indices] + ) diff --git a/fairseq/data/multilingual/__init__.py b/fairseq/data/multilingual/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6264236915a7269a4d920ee8213004374dd86a9a --- /dev/null +++ b/fairseq/data/multilingual/__init__.py @@ -0,0 +1,4 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. diff --git a/fairseq/data/multilingual/__pycache__/__init__.cpython-310.pyc b/fairseq/data/multilingual/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..54cb89885e6765ef5d75a07cc0955d6781d67cee Binary files /dev/null and b/fairseq/data/multilingual/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/data/multilingual/__pycache__/multilingual_data_manager.cpython-310.pyc b/fairseq/data/multilingual/__pycache__/multilingual_data_manager.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..890640754681ebbcf3deb6e47d6292469995e302 Binary files /dev/null and b/fairseq/data/multilingual/__pycache__/multilingual_data_manager.cpython-310.pyc differ diff --git a/fairseq/data/multilingual/__pycache__/sampled_multi_dataset.cpython-310.pyc b/fairseq/data/multilingual/__pycache__/sampled_multi_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e9246b8078356a443c46a02773e990a6ba6b6305 Binary files /dev/null and b/fairseq/data/multilingual/__pycache__/sampled_multi_dataset.cpython-310.pyc differ diff --git a/fairseq/data/multilingual/__pycache__/sampled_multi_epoch_dataset.cpython-310.pyc b/fairseq/data/multilingual/__pycache__/sampled_multi_epoch_dataset.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3686a68c049f73511ba8b5ea0b253bddbe733fcf Binary files /dev/null and b/fairseq/data/multilingual/__pycache__/sampled_multi_epoch_dataset.cpython-310.pyc differ diff --git a/fairseq/data/multilingual/__pycache__/sampling_method.cpython-310.pyc b/fairseq/data/multilingual/__pycache__/sampling_method.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..821bed25db6dc5330c3623a2fd0b8d53fca12869 Binary files /dev/null and b/fairseq/data/multilingual/__pycache__/sampling_method.cpython-310.pyc differ diff --git a/fairseq/data/multilingual/multilingual_data_manager.py b/fairseq/data/multilingual/multilingual_data_manager.py new file mode 100644 index 0000000000000000000000000000000000000000..b31028c1912e3f88131609b4094f08f20ab3625a --- /dev/null +++ b/fairseq/data/multilingual/multilingual_data_manager.py @@ -0,0 +1,836 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools +import logging +import os + +import numpy as np +from collections import OrderedDict + +import json +from fairseq import options, utils +from fairseq.options import eval_str_dict, csv_str_list + +from fairseq.data import ( + Dictionary, + AppendTokenDataset, + ConcatDataset, + data_utils, + indexed_dataset, + LanguagePairDataset, + PrependTokenDataset, + StripTokenDataset, + TruncateDataset, + SampledMultiDataset, + TransformEosLangPairDataset, + SampledMultiEpochDataset, +) +from fairseq.data.multilingual.sampled_multi_dataset import CollateFormat +from fairseq.file_io import PathManager + +logger = logging.getLogger(__name__) + + +def _lang_token(lang: str, style='__{}__'): + return style.format(lang) + + +def _lang_token_index(dic: Dictionary, lang: str, style='__{}__'): + """Return language token index.""" + idx = dic.index(_lang_token(lang, style)) + assert idx != dic.unk_index, \ + 'cannot find language token for lang {}'.format(lang) + return idx + + +def _lang_id(dic: Dictionary, lang: str): + """Return language ID index.""" + idx = dic.index(lang) + assert idx != dic.unk_index, \ + 'cannot find language ID for lang {}'.format(lang) + return idx + + +def load_sampling_weights(from_file): + with open(from_file) as f: + weights = json.load(f) + return weights + + +class MultilingualDatasetManager(object): + def __init__(self, args, lang_pairs, langs, dicts, sampling_method): + super().__init__() + self.args = args + self.seed = args.seed + self.lang_pairs = lang_pairs + self.langs = langs + self.dicts = dicts + self.lang_dict = self.create_lang_dictionary(self.langs) + self.sampling_method = sampling_method + self.sampling_scheduler = None + self._has_sharded_data = False + self._num_shards_dict = {} + + @classmethod + def setup_data_manager(cls, args, lang_pairs, langs, dicts, sampling_method): + return MultilingualDatasetManager(args, lang_pairs, langs, dicts, sampling_method) + + @staticmethod + def add_args(parser): + parser.add_argument('data', help='colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner') + parser.add_argument('--langs', default=None, type=csv_str_list, + help='a list of languages comma sperated languages which can appear in lang-pairs; ' + 'note that the ordering determines language token IDs', + ) + parser.add_argument('--lang-dict', default=None, type=str, + help='an external file which contains a list of ' + 'languages which can appear in lang-pairs; ' + 'note that the ordering determines language token IDs; ' + '--langs and --lang-dict are two exclusive options') + parser.add_argument('--lang-tok-style', default='multilingual', + type=str, choices=['multilingual', 'mbart'], + help='language token styles') + + parser.add_argument('--load-alignments', action='store_true', + help='load the binarized alignments') + parser.add_argument('--left-pad-source', default='True', type=str, metavar='BOOL', + help='pad the source on the left') + parser.add_argument('--left-pad-target', default='False', type=str, metavar='BOOL', + help='pad the target on the left') + parser.add_argument('--max-source-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the source sequence') + parser.add_argument('--max-target-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the target sequence') + parser.add_argument('--upsample-primary', default=1, type=int, + help='amount to upsample primary dataset') + parser.add_argument('--truncate-source', action='store_true', default=False, + help='truncate source to max-source-positions') + parser.add_argument('--encoder-langtok', default=None, type=str, choices=['src', 'tgt'], + metavar='SRCTGT', + help='prepend to the beginning of source sentence the source or target ' + 'language token. (src/tgt)') + parser.add_argument('--decoder-langtok', action='store_true', + help='prepend to the beginning of target sentence the target language token') + parser.add_argument('--lang-tok-replacing-bos-eos', action='store_true', default=False) + parser.add_argument('--enable-lang-ids', default=False, action='store_true', + help='whether to include language IDs in samples') + parser.add_argument('--enable-reservsed-directions-shared-datasets', default=False, action='store_true', + help='whether to allow datasets be used in reversed directions') + + parser.add_argument('--extra-data', help='a dictionary of data name to this path, \ + e.g. {"mined", path_to_mined_data, "denoised": path_to_denoised_data}', + type=lambda uf: eval_str_dict(uf, type=str), + default=None) + parser.add_argument('--extra-lang-pairs', help='a dictionary of data name to the language pairs they serve, \ + e.g. {"mined": comma-separated-lang-pairs, "denoised": comma-separated-lang-pairs}', + type=lambda uf: eval_str_dict(uf, type=str), + default=None) + parser.add_argument('--langtoks-specs', + help='a list of comma separated data types that a set of language tokens to be specialized for, \ + e.g. "main,dae,mined". There will be a set of language tokens added to the vocab to \ + distinguish languages in different training data types. If not specified, default language \ + tokens per languages will be added', + default='main', + type=csv_str_list, + ) + parser.add_argument('--langtoks', help='a dictionary of how to add language tokens, \ + e.g. {"mined": (None, "tgt"), "mono_dae": ("src.dae", "tgt"), "main": \ + ("src", "tgt")}, or {"mined": ("src.mined", "tgt")}', + default=None, + type=lambda uf: eval_str_dict(uf, type=str), + ) + parser.add_argument('--sampling-weights-from-file', + help='a file contain a python dictionary of how to sample data sets, \ + e.g. { "main:en_XX-es_XX": 0.2, "mined:en_XX-pt_XX": 0.5, \ + "mono_dae:es_XX-es_XX: 0.3, "main:en_xx-fr_XX": 0.8 }', + default=None, type=str, + ) + parser.add_argument('--sampling-weights', help='a dictionary of how to sample data sets, \ + e.g. { "main:en_XX-es_XX": 0.2, "mined:en_XX-pt_XX": 0.5, \ + "mono_dae:es_XX-es_XX: 0.3, "main:en_xx-fr_XX": 0.8 }', + default=None, + type=lambda uf: eval_str_dict(uf, type=str), + ) + parser.add_argument('--virtual-epoch-size', default=1000000, type=int, + help='virtual epoch size to speed up data loading') + parser.add_argument('--virtual-data-size', default=None, type=int, + help='virtual data size of the whole joint dataset to speed' + 'up data loading and have specific dynamic sampling strategy interval') + + @classmethod + def load_langs(cls, args, **kwargs): + if args.lang_dict and args.langs: + raise ValueError('--langs and --lang-dict can not both be specified') + if args.lang_dict is None and args.langs is None: + logger.warning( + 'External language dictionary is not provided; ' + 'use lang-pairs to infer the set of supported languages. ' + 'The language ordering is not stable which might cause ' + 'misalignment in pretraining and finetuning.') + # infer from lang_pairs as it is + langs = list({x for lang_pair in args.lang_pairs for x in lang_pair.split('-')}) + langs = sorted(langs) + logger.info(f'inferred language list: {langs}') + elif args.lang_dict: + with PathManager.open(args.lang_dict, "r", encoding="utf-8") as f: + langs = [lang.strip() for lang in f.readlines() if lang.strip()] + logger.info(f'loaded language list from {args.lang_dict} as they are ordered in file') + elif args.langs: + langs = args.langs + logger.info(f'parsed the language list as they are ordered in the option: {langs}') + return langs + + def has_sharded_data(self, split): + return self._has_sharded_data and split == getattr(self.args, "train_subset", None) + + def _shared_collater(self): + return ( + not (self.args.extra_data and 'mono_dae' in self.args.extra_data) + and (not self.args.lang_tok_replacing_bos_eos) + ) + + @classmethod + def prepare(cls, load_dictionary, args, **kargs): + args.left_pad_source = options.eval_bool(args.left_pad_source) + args.left_pad_target = options.eval_bool(args.left_pad_target) + + if not hasattr(args, 'shuffle_instance'): + args.shuffle_instance = False + if args.langtoks is None: + args.langtoks = {} + if 'main' not in args.langtoks: + src_langtok_spec = args.encoder_langtok if args.encoder_langtok else None + tgt_langtok_spec = 'tgt' if args.decoder_langtok else None + args.langtoks['main'] = (src_langtok_spec, tgt_langtok_spec) + + def check_langs(langs, pairs): + messages = [] + for src, tgt in pairs: + if src not in langs or tgt not in langs: + messages.append(f'language pair {src}-{tgt} contains languages ' + 'that are not in the language dictionary') + if len(messages) > 0: + raise ValueError(' '.join(messages) + f"; langs: {langs}") + + if args.lang_pairs is None: + raise ValueError('--lang-pairs is required. List all the language pairs in the training objective.') + if isinstance(args.lang_pairs, str): + args.lang_pairs = args.lang_pairs.split(',') + if args.source_lang is not None or args.target_lang is not None: + training = False + else: + training = True + sorted_langs = cls.load_langs(args, **kargs) + check_langs( + sorted_langs, + ([p.split('-') for p in args.lang_pairs] if training + else [(args.source_lang, args.target_lang)]) + ) + + # load dictionaries + if training: + extra_lang_pairs = ( + list({p for _, v in args.extra_lang_pairs.items() for p in v.split(',')}) + if args.extra_lang_pairs else [] + ) + langs_to_load_dicts = sorted({x for p in args.lang_pairs + extra_lang_pairs for x in p.split('-')}) + else: + langs_to_load_dicts = sorted([args.source_lang, args.target_lang]) + + dicts = OrderedDict() + supported_langtok_specs = args.langtoks_specs + for lang in langs_to_load_dicts: + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dicts[lang] = load_dictionary(os.path.join(paths[0], 'dict.{}.txt'.format(lang))) + if len(dicts) > 0: + assert dicts[lang].pad() == dicts[langs_to_load_dicts[0]].pad() + assert dicts[lang].eos() == dicts[langs_to_load_dicts[0]].eos() + assert dicts[lang].unk() == dicts[langs_to_load_dicts[0]].unk() + + # keep the langs consistent for all experiments with the same lang dict + # for finetuning regardless of whether lang_tok is required or not just add the tokens to the dicts + for spec in supported_langtok_specs: + for lang_to_add in sorted_langs: + dicts[lang].add_symbol( + MultilingualDatasetManager.get_lang_tok(lang_to_add, args, spec) + ) + if args.lang_tok_style == 'mbart' or (args.extra_data and 'mono_dae' in args.extra_data): + dicts[lang].add_symbol('') + logger.info('[{}] dictionary: {} types'.format(lang, len(dicts[lang]))) + return sorted_langs, dicts, training + + TOKEN_STYLES = { + 'mbart': '[{}]', + 'multilingual': '__{}__' + } + + @classmethod + def create_lang_dictionary(cls, langs): + unk = '' + # hack to remove symbols other than unk as they are not needed by lang dict + lang_dict = Dictionary( + pad=unk, + eos=unk, + unk=unk, + bos=unk, + ) + for lang in langs: + lang_dict.add_symbol(lang) + return lang_dict + + @classmethod + def get_lang_tok_style(cls, args): + return cls.TOKEN_STYLES[args.lang_tok_style] + + @classmethod + def get_lang_tok(cls, lang, args, spec=''): + if spec is None: + return None + if spec.endswith('dae'): + lang = f'{lang}_dae' + elif spec.endswith('mined'): + lang = f'{lang}_mined' + return _lang_token(lang, cls.get_lang_tok_style(args)) + + @classmethod + def get_langtok_index(cls, lang_tok, dic): + idx = dic.index(lang_tok) + assert idx != dic.unk_index, \ + 'cannot find language token {} in the dictionary'.format(lang_tok) + return idx + + def get_encoder_langtok(self, src_lang, tgt_lang, spec=None): + if spec is None: + return None + if spec and spec.startswith('src'): + if src_lang is None: + return None + langtok = self.get_lang_tok(src_lang, self.args, spec) + else: + if tgt_lang is None: + return None + langtok = self.get_lang_tok(tgt_lang, self.args, spec) + return self.get_langtok_index(langtok, self.dicts[src_lang if src_lang else tgt_lang]) + + def get_decoder_langtok(self, tgt_lang, spec=None): + if spec is None: + return None + langtok = self.get_lang_tok(tgt_lang, self.args, spec) + return self.get_langtok_index(langtok, self.dicts[tgt_lang]) + + @classmethod + def load_data(cls, path, vdict, impl): + dataset = data_utils.load_indexed_dataset(path, vdict, impl) + return dataset + + @classmethod + def split_exists(cls, split, src, tgt, lang, data_path, dataset_impl): + filename = os.path.join(data_path, '{}.{}-{}.{}'.format(split, src, tgt, lang)) + return indexed_dataset.dataset_exists(filename, impl=dataset_impl) + + @classmethod + def mono_split_exists(cls, split, lang, data_path, dataset_impl): + filename = os.path.join(data_path, '{}.{}'.format(split, lang)) + return indexed_dataset.dataset_exists(filename, impl=dataset_impl) + + @classmethod + def bitext_split_exists(cls, split, src, tgt, data_path, dataset_impl): + src_exists = cls.split_exists(split, src, tgt, lang=src, data_path=data_path, dataset_impl=dataset_impl) \ + or cls.split_exists(split, tgt, src, lang=src, data_path=data_path, dataset_impl=dataset_impl) + + tgt_exists = cls.split_exists(split, src, tgt, lang=tgt, data_path=data_path, dataset_impl=dataset_impl) \ + or cls.split_exists(split, tgt, src, lang=tgt, data_path=data_path, dataset_impl=dataset_impl) + return src_exists and tgt_exists + + @classmethod + def get_split_num_shards(cls, split, src, tgt, data_paths, dataset_impl): + return sum( + 1 for path in data_paths + if cls.bitext_split_exists(split, src, tgt, path, dataset_impl) + ) + + @classmethod + def get_mono_split_num_shards(cls, split, lang, data_paths, dataset_impl): + return sum( + 1 for path in data_paths + if cls.mono_split_exists(split, lang, path, dataset_impl) + ) + + def load_lang_dataset( + self, + data_path, split, + src, src_dict, + tgt, tgt_dict, + combine, dataset_impl, upsample_primary, + max_source_positions, + prepend_bos=False, load_alignments=False, + truncate_source=False, + ): + + src_datasets = [] + tgt_datasets = [] + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else '') + + # infer langcode + if self.split_exists(split_k, src, tgt, src, data_path, dataset_impl): + prefix = os.path.join(data_path, '{}.{}-{}.'.format(split_k, src, tgt)) + elif self.split_exists(split_k, tgt, src, src, data_path, dataset_impl): + prefix = os.path.join(data_path, '{}.{}-{}.'.format(split_k, tgt, src)) + else: + if k > 0: + break + else: + logger.error(f"Dataset not found: {data_path}, {split_k}, {src}, {tgt}") + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, data_path)) + + src_dataset = self.load_data(prefix + src, src_dict, dataset_impl) + if truncate_source: + src_dataset = AppendTokenDataset( + TruncateDataset( + StripTokenDataset(src_dataset, src_dict.eos()), + max_source_positions - 1, + ), + src_dict.eos(), + ) + src_datasets.append(src_dataset) + tgt_datasets.append( + self.load_data(prefix + tgt, tgt_dict, dataset_impl) + ) + + logger.info('{} {} {}-{} {} examples'.format( + data_path, split_k, src, tgt, len(src_datasets[-1]) + )) + + if not combine: + break + + assert len(src_datasets) == len(tgt_datasets) + + if len(src_datasets) == 1: + src_dataset, tgt_dataset = src_datasets[0], tgt_datasets[0] + else: + sample_ratios = [1] * len(src_datasets) + sample_ratios[0] = upsample_primary + src_dataset = ConcatDataset(src_datasets, sample_ratios) + tgt_dataset = ConcatDataset(tgt_datasets, sample_ratios) + + if prepend_bos: + assert hasattr(src_dict, "bos_index") and hasattr(tgt_dict, "bos_index") + src_dataset = PrependTokenDataset(src_dataset, src_dict.bos()) + tgt_dataset = PrependTokenDataset(tgt_dataset, tgt_dict.bos()) + + align_dataset = None + if load_alignments: + align_path = os.path.join(data_path, '{}.align.{}-{}'.format(split, src, tgt)) + if indexed_dataset.dataset_exists(align_path, impl=dataset_impl): + align_dataset = data_utils.load_indexed_dataset(align_path, None, dataset_impl) + + return src_dataset, tgt_dataset, align_dataset + + def load_langpair_dataset( + self, + data_path, split, + src, src_dict, + tgt, tgt_dict, + combine, dataset_impl, upsample_primary, + left_pad_source, left_pad_target, max_source_positions, + max_target_positions, prepend_bos=False, load_alignments=False, + truncate_source=False, + src_dataset_transform_func=lambda dataset: dataset, + tgt_dataset_transform_func=lambda dataset: dataset, + src_lang_id=None, + tgt_lang_id=None, + langpairs_sharing_datasets=None, + ): + norm_direction = "-".join(sorted([src, tgt])) + if langpairs_sharing_datasets is not None: + src_dataset = langpairs_sharing_datasets.get((data_path, split, norm_direction, src), 'NotInCache') + tgt_dataset = langpairs_sharing_datasets.get((data_path, split, norm_direction, tgt), 'NotInCache') + align_dataset = langpairs_sharing_datasets.get((data_path, split, norm_direction, src, tgt), 'NotInCache') + + # a hack: any one is not in cache, we need to reload them + if ( + langpairs_sharing_datasets is None + or src_dataset == 'NotInCache' + or tgt_dataset == 'NotInCache' + or align_dataset == 'NotInCache' + or split != getattr(self.args, "train_subset", None) + ): + # source and target datasets can be reused in reversed directions to save memory + # reversed directions of valid and test data will not share source and target datasets + src_dataset, tgt_dataset, align_dataset = self.load_lang_dataset( + data_path, split, + src, src_dict, + tgt, tgt_dict, + combine, dataset_impl, upsample_primary, + max_source_positions=max_source_positions, + prepend_bos=prepend_bos, load_alignments=load_alignments, + truncate_source=truncate_source, + ) + src_dataset = src_dataset_transform_func(src_dataset) + tgt_dataset = tgt_dataset_transform_func(tgt_dataset) + if langpairs_sharing_datasets is not None: + langpairs_sharing_datasets[(data_path, split, norm_direction, src)] = src_dataset + langpairs_sharing_datasets[(data_path, split, norm_direction, tgt)] = tgt_dataset + langpairs_sharing_datasets[(data_path, split, norm_direction, src, tgt)] = align_dataset + if align_dataset is None: + # no align data so flag the reverse direction as well in sharing + langpairs_sharing_datasets[(data_path, split, norm_direction, tgt, src)] = align_dataset + else: + logger.info(f"Reusing source and target datasets of [{split}] {tgt}-{src} for reversed direction: " + f"[{split}] {src}-{tgt}: src length={len(src_dataset)}; tgt length={len(tgt_dataset)}") + + return LanguagePairDataset( + src_dataset, src_dataset.sizes, src_dict, + tgt_dataset, tgt_dataset.sizes, tgt_dict, + left_pad_source=left_pad_source, + left_pad_target=left_pad_target, + align_dataset=align_dataset, + src_lang_id=src_lang_id, + tgt_lang_id=tgt_lang_id, + ) + + def src_dataset_tranform_func(self, src_lang, tgt_lang, dataset, spec=None): + if self.args.lang_tok_replacing_bos_eos: + # it is handled by self.alter_dataset_langtok + # TODO: Unifiy with alter_dataset_langtok + return dataset + if spec is None: + return dataset + tok = self.get_encoder_langtok(src_lang, tgt_lang, spec) + if tok: + return PrependTokenDataset(dataset, tok) + return dataset + + def tgt_dataset_tranform_func(self, source_lang, target_lang, dataset, spec=None): + if self.args.lang_tok_replacing_bos_eos: + # TODO: Unifiy with alter_dataset_langtok + # It is handled by self.alter_dataset_langtok. + # The complication in self.alter_dataset_langtok + # makes a unified framework difficult. + return dataset + # if not self.args.decoder_langtok: + if not spec: + return dataset + tok = self.get_decoder_langtok(target_lang, spec) + if tok: + return PrependTokenDataset(dataset, tok) + return dataset + + def alter_dataset_langtok(self, lang_pair_dataset, + src_eos=None, src_lang=None, + tgt_eos=None, tgt_lang=None, + src_langtok_spec=None, tgt_langtok_spec=None, + ): + if src_langtok_spec is None and tgt_langtok_spec is None: + return lang_pair_dataset + + new_src_eos = None + if src_langtok_spec is not None and src_eos is not None \ + and (src_lang is not None or tgt_lang is not None): + new_src_eos = self.get_encoder_langtok(src_lang, tgt_lang, src_langtok_spec) + else: + src_eos = None + + new_tgt_bos = None + if tgt_langtok_spec and tgt_eos is not None and tgt_lang is not None: + new_tgt_bos = self.get_decoder_langtok(tgt_lang, tgt_langtok_spec) + else: + tgt_eos = None + + return TransformEosLangPairDataset( + lang_pair_dataset, + src_eos=src_eos, + new_src_eos=new_src_eos, + tgt_bos=tgt_eos, + new_tgt_bos=new_tgt_bos, + ) + + def load_a_dataset( + self, + split, + data_path, + src, src_dict, + tgt, tgt_dict, + combine, + prepend_bos=False, + langpairs_sharing_datasets=None, + data_category=None, + **extra_kwargs, + ): + dataset_impl = self.args.dataset_impl + upsample_primary = self.args.upsample_primary + left_pad_source = self.args.left_pad_source + left_pad_target = self.args.left_pad_target + max_source_positions = self.args.max_source_positions + max_target_positions = self.args.max_target_positions + load_alignments = self.args.load_alignments + truncate_source = self.args.truncate_source + src_dataset_transform_func = self.src_dataset_tranform_func + tgt_dataset_transform_func = self.tgt_dataset_tranform_func + enable_lang_ids = self.args.enable_lang_ids + lang_dictionary = self.lang_dict + src_langtok_spec, tgt_langtok_spec = extra_kwargs['langtok_spec'] + + src_langtok = self.get_encoder_langtok(src, tgt, src_langtok_spec) + tgt_langtok = self.get_decoder_langtok(tgt, tgt_langtok_spec) + logger.info(f'{data_category}:{src}-{tgt} src_langtok: {src_langtok}; tgt_langtok: {tgt_langtok}') + + langpair_ds = self.load_langpair_dataset( + data_path, split, + src, src_dict, + tgt, tgt_dict, + combine, dataset_impl, upsample_primary, + left_pad_source, left_pad_target, max_source_positions, + max_target_positions, prepend_bos, load_alignments, + truncate_source, + src_dataset_transform_func=lambda dataset: src_dataset_transform_func(src, tgt, dataset, src_langtok_spec), + tgt_dataset_transform_func=lambda dataset: tgt_dataset_transform_func(src, tgt, dataset, tgt_langtok_spec), + src_lang_id=_lang_id(lang_dictionary, src) if enable_lang_ids and lang_dictionary is not None else None, + tgt_lang_id=_lang_id(lang_dictionary, tgt) if enable_lang_ids and lang_dictionary is not None else None, + langpairs_sharing_datasets=langpairs_sharing_datasets, + ) + if langpair_ds.tgt_sizes is None: + # hack to use src_sizes as the sizes for the whole pair dataset for ConcatDataset + langpair_ds.sizes = langpair_ds.src_sizes + else: + # use the max of two sides to define the size to help max positions filtering + langpair_ds.sizes = np.vstack([langpair_ds.src_sizes, langpair_ds.tgt_sizes]).max(axis=0) + assert langpair_ds.sizes.shape == langpair_ds.src_sizes.shape + # TODO: handle modified lang toks for mined data and dae data + if self.args.lang_tok_replacing_bos_eos: + ds = self.alter_dataset_langtok( + langpair_ds, + src_eos=self.dicts[src if src else tgt].eos(), + src_lang=src, + tgt_eos=self.dicts[tgt].eos(), + tgt_lang=tgt, + src_langtok_spec=src_langtok_spec, + tgt_langtok_spec=tgt_langtok_spec, + ) + else: + ds = langpair_ds + return ds + + def load_split_langpair_datasets( + self, + split, + data_param_list, + ): + datasets = [] + langpairs_sharing_datasets = {} if self.args.enable_reservsed_directions_shared_datasets else None + for param in data_param_list: + ds = self.load_a_dataset(split=split, langpairs_sharing_datasets=langpairs_sharing_datasets, **param) + datasets.append(ds) + return datasets + + def get_data_paths_and_lang_pairs(self, split): + datapaths = { + 'main': self.args.data, + } + lang_pairs = { + 'main': self.lang_pairs + } + if split == getattr(self.args, "train_subset", None): + # only training data can have extra data and extra language pairs + if self.args.extra_data: + extra_datapaths = self.args.extra_data + datapaths.update(extra_datapaths) + if self.args.extra_lang_pairs: + extra_lang_pairs = {k: v.split(',') for k, v in self.args.extra_lang_pairs.items()} + lang_pairs.update(extra_lang_pairs) + return datapaths, lang_pairs + + @classmethod + def get_dataset_key(cls, data_category, src, tgt): + return f'{data_category}:{src}-{tgt}' + + def get_split_num_data_shards(self, split): + if split in self._num_shards_dict: + return self._num_shards_dict[split] + num_shards_dict = {} + data_paths, lang_pairs = self.get_data_paths_and_lang_pairs(split) + + for data_category, paths in data_paths.items(): + if data_category not in lang_pairs: + continue + paths = utils.split_paths(paths) + lang_dirs = [lang_pair.split('-') for lang_pair in lang_pairs[data_category]] + lang_dirs = [x if len(x) > 1 else (x[0], x[0]) for x in lang_dirs] + for src, tgt in lang_dirs: + # monolingual data ruqires tgt only + assert src is not None or 'mono_' in data_category, (f'error: src={src}, ' + 'tgt={tgt} for data_category={data_category}') + key = self.get_dataset_key(data_category, src, tgt) + if 'mono_' in data_category: + num_shards_dict[key] = self.get_mono_split_num_shards( + split, tgt, paths, self.args.dataset_impl) + else: + num_shards_dict[key] = self.get_split_num_shards( + split, src, tgt, paths, self.args.dataset_impl) + self._num_shards_dict[split] = num_shards_dict + logger.info(f"[{split}] num of shards: {num_shards_dict}") + return num_shards_dict + + def get_split_data_path(self, paths, epoch, shard_epoch, num_shards): + shard = epoch if shard_epoch is None else shard_epoch + shard = (shard - 1) % num_shards + path = paths[shard] + return path + + def get_split_data_param_list(self, split, epoch, shard_epoch=None): + # TODO: to extend with extra datasets and keys and loop over different shard data paths + param_list = [] + data_paths, lang_pairs = self.get_data_paths_and_lang_pairs(split) + logger.info(f'langtoks settings: {self.args.langtoks}') + split_num_shards_dict = self.get_split_num_data_shards(split) + for data_category, paths in data_paths.items(): + if data_category not in lang_pairs: + continue + paths = utils.split_paths(paths) + assert len(paths) > 0 + if len(paths) > 1: + self._has_sharded_data = True + + if data_category in self.args.langtoks: + lang_tok_spec = self.args.langtoks[data_category] + else: + # default to None + lang_tok_spec = (None, None) + + # infer langcode + lang_dirs = [lang_pair.split('-') for lang_pair in lang_pairs[data_category]] + lang_dirs = [x if len(x) > 1 else (x[0], x[0]) for x in lang_dirs] + for src, tgt in lang_dirs: + assert src is not None or data_category == 'mono_dae', (f'error: src={src}, ' + 'tgt={tgt} for data_category={data_category}') + # logger.info(f"preparing param for {data_category}: {src} - {tgt}") + key = self.get_dataset_key(data_category, src, tgt) + data_path = self.get_split_data_path( + paths, epoch, shard_epoch, split_num_shards_dict[key]) + param_list.append( + { + 'key': key, + 'data_path': data_path, + 'split': split, + 'src': src, + 'src_dict': self.dicts[src] if src and data_category != 'mono_dae' else None, + 'tgt': tgt, + 'tgt_dict': self.dicts[tgt], + 'data_category': data_category, + 'langtok_spec': lang_tok_spec, + } + ) + return param_list + + def get_train_dataset_sizes(self, data_param_list, datasets): + num_shards = [ + self.get_split_num_data_shards(param['split'])[param['key']] for param in data_param_list] + data_sizes = [(key, len(d) * num_shard) for (key, d), num_shard in zip(datasets, num_shards)] + logger.info(f'data sizes multiplied by num_shards used in sampling ratios: {data_sizes}') + return [s for _, s in data_sizes] + + def get_train_sampling_ratios(self, data_param_list, datasets, epoch=1): + data_sizes = self.get_train_dataset_sizes(data_param_list, datasets) + sampling_func = self.sampling_method.sampling_method_selector() + sample_ratios = sampling_func(data_sizes) if sampling_func is not None else None + return sample_ratios + + def get_sampling_ratios(self, data_param_list, datasets, epoch): + if self.args.sampling_weights_from_file: + weights = load_sampling_weights(self.args.sampling_weights_from_file) + sample_ratios = [weights[k] for k, _ in datasets] + logger.info('| ignoring --sampling-weights when loadding sampling weights ' + f'from file {self.args.sampling_weights_from_file}') + elif self.args.sampling_weights: + sample_ratios = [self.args.sampling_weights[k] for k, _ in datasets] + else: + sample_ratios = self.get_train_sampling_ratios(data_param_list, datasets, epoch) + + if sample_ratios is not None: + logger.info('| Upsample ratios: {}'.format( + list(zip(map(lambda x: x['key'], data_param_list), sample_ratios)) + )) + assert len(sample_ratios) == len(datasets) + return sample_ratios + + def load_split_datasets( + self, + split, + training, + epoch=1, combine=False, shard_epoch=None, **kwargs, + ): + data_param_list = self.get_split_data_param_list( + split, epoch, shard_epoch=shard_epoch, + ) + langpairs_sharing_datasets = {} if self.args.enable_reservsed_directions_shared_datasets else None + datasets = [ + ( + param['key'], + self.load_a_dataset( + combine=combine, + langpairs_sharing_datasets=langpairs_sharing_datasets, + **param + ), + ) + for param in data_param_list + ] + return datasets, data_param_list + + def load_into_sampled_multi_epoch_dataset( + self, split, datasets, data_param_list, + epoch, shard_epoch=None + ): + sample_ratios = self.get_sampling_ratios(data_param_list, datasets, epoch) + return SampledMultiEpochDataset( + OrderedDict(datasets), + epoch=epoch, + shard_epoch=shard_epoch, + # valid and test datasets will be degerate to concating datasets: + sampling_ratios=sample_ratios, + eval_key=None, + batch_by_size=True, + collate_format=CollateFormat.single, + virtual_size=self.args.virtual_data_size, + split=split, + virtual_epoch_size=self.args.virtual_epoch_size, + # if not using lang_tok altering, simplified to use the same collater + shared_collater=self._shared_collater(), + ) + + def load_into_concat_dataset(self, split, datasets, data_param_list): + if self.args.lang_tok_replacing_bos_eos: + # TODO: to investigate why TransformEosLangPairDataset doesn't work with ConcatDataset + return SampledMultiDataset( + OrderedDict(datasets), + sampling_ratios=None, + eval_key=None, + batch_by_size=True, + collate_format=CollateFormat.single, + virtual_size=None, + split=split, + ) + return ConcatDataset([d for _, d in datasets]) + + def load_sampled_multi_epoch_dataset( + self, + split, + training, + epoch=0, combine=False, shard_epoch=None, **kwargs + ): + datasets, data_param_list = self.load_split_datasets( + split, training, + epoch, combine, shard_epoch=shard_epoch, **kwargs + ) + if training and split == getattr(self.args, "train_subset", None): + return self.load_into_sampled_multi_epoch_dataset( + split, datasets, data_param_list, epoch, shard_epoch=shard_epoch) + else: + return self.load_into_concat_dataset(split, datasets, data_param_list) diff --git a/fairseq/data/multilingual/sampled_multi_dataset.py b/fairseq/data/multilingual/sampled_multi_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..95eab280f030bfbd8ced892f4e0c59c4ffed8cfe --- /dev/null +++ b/fairseq/data/multilingual/sampled_multi_dataset.py @@ -0,0 +1,396 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import List +from enum import Enum +from collections import OrderedDict +from collections import defaultdict +from bisect import bisect_right +import hashlib +import logging +import datetime +import time + +import numpy as np +import torch + +from fairseq import distributed_utils +from fairseq.data import plasma_utils, FairseqDataset + + +def get_time_gap(s, e): + return (datetime.datetime.fromtimestamp(e) - datetime.datetime.fromtimestamp(s)).__str__() + + +logger = logging.getLogger(__name__) + + +def default_virtual_size_func(datasets, ratios, max_scale_up=1.5): + sizes = [len(d) for d in datasets] + if ratios is None: + return sum(sizes) + largest_idx = np.argmax(sizes) + largest_r = ratios[largest_idx] + largest_s = sizes[largest_idx] + # set virtual sizes relative to the largest dataset + virtual_sizes = [(r / largest_r) * largest_s for r in ratios] + vsize = sum(virtual_sizes) + max_size = sum(sizes) * max_scale_up + return int(vsize if vsize < max_size else max_size) + + +class CollateFormat(Enum): + single = 1 + ordered_dict = 2 + + +class SampledMultiDataset(FairseqDataset): + """Samples from multiple sub-datasets according to given sampling ratios. + Args: + datasets ( + List[~torch.utils.data.Dataset] + or OrderedDict[str, ~torch.utils.data.Dataset] + ): datasets + sampling_ratios (List[float]): list of probability of each dataset to be sampled + (default: None, which corresponds to concating all dataset together). + batch_by_size (bool): whether or not to batch by sequence length + (default: True). + seed (int): RNG seed to use (default: 2). + epoch (int): starting epoch number (default: 1). + eval_key (str, optional): a key used at evaluation time that causes + this instance to pass-through batches from *datasets[eval_key]*. + collate_format (CollateFormat): collater output format, either CollateFormat.ordered_dict or + CollateFormat.single (default: CollateFormat.single) where CollateFormat.single configures + the collater to output batches of data mixed from all sub-datasets, + and CollateFormat.ordered_dict configures the collater to output a dictionary of batches indexed by keys + of sub-datasets. + Note that not all sub-datasets will present in a single batch in both formats. + virtual_size (int, or callable): the expected virtual size of the dataset (default: default_virtual_size_func). + split (str): the split of the data, e.g. 'train', 'valid' or 'test'. + shared_collater (bool): whether or not to all sub-datasets have the same collater. + """ + + def __init__( + self, + datasets, + sampling_ratios=None, + batch_by_size=False, + seed=2, + epoch=1, + eval_key=None, + collate_format=CollateFormat.single, + virtual_size=default_virtual_size_func, + split='', + shared_collater=False, + ): + super().__init__() + self.batch_by_size = batch_by_size + self.shared_collater = shared_collater + + if isinstance(datasets, OrderedDict): + self.keys = list(datasets.keys()) + datasets = list(datasets.values()) + elif isinstance(datasets, List): + self.keys = list(range(len(datasets))) + else: + raise AssertionError() + self.datasets = datasets + self.split = split + + self.eval_key = eval_key + if self.eval_key is not None: + self.collate_format = CollateFormat.single + else: + self.collate_format = collate_format + + self.seed = seed + self._cur_epoch = None + self._cur_indices = None + self._sizes = None + self._ordered_indices = None + self.virtual_size_per_dataset = None + # caching properties + self._reset_cached_properties() + self.setup_sampling(sampling_ratios, virtual_size) + self.cumulated_sizes = None + self.virtual_size_per_dataset = None + self._size_cache = {} + self.set_epoch(epoch) + + def _clean_if_not_none(self, var_list): + for v in var_list: + if v is not None: + del v + + def _reset_cached_properties(self): + self._clean_if_not_none([ + self._sizes, self._ordered_indices, self._cur_indices + ]) + self._sizes = None + self._ordered_indices = None + self._cur_indices = None + + def setup_sampling(self, sample_ratios, virtual_size): + sizes = [len(d) for d in self.datasets] + if sample_ratios is None: + # default back to concating datasets + self.sample_ratios = None + self.virtual_size = sum(sizes) + else: + if not isinstance(sample_ratios, np.ndarray): + sample_ratios = np.array(sample_ratios) + self.sample_ratios = plasma_utils.PlasmaArray(sample_ratios) + virtual_size = default_virtual_size_func if virtual_size is None else virtual_size + self.virtual_size = ( + virtual_size(self.datasets, self.sample_ratios.array) if callable(virtual_size) + else virtual_size) + + def adjust_sampling(self, epoch, sampling_ratios, virtual_size): + if sampling_ratios is not None: + sampling_ratios = self._sync_sample_ratios(sampling_ratios) + self.setup_sampling(sampling_ratios, virtual_size) + + def _sync_sample_ratios(self, ratios): + # in case the ratios are not precisely the same across processes + # also to ensure every procresses update the ratios in the same pace + ratios = torch.DoubleTensor(ratios) + if torch.distributed.is_initialized(): + if torch.cuda.is_available(): + distributed_utils.all_reduce(ratios.cuda()) + else: + distributed_utils.all_reduce(ratios) + ret = ratios.cpu() + ret = ret.numpy() + return ret + + def random_choice_in_dataset(self, rng, dataset, choice_size): + if hasattr(dataset, 'random_choice_in_dataset'): + return dataset.random_choice_in_dataset(rng, choice_size) + dataset_size = len(dataset) + return rng.choice(dataset_size, choice_size, replace=(choice_size > dataset_size)) + + def get_virtual_indices(self, rng, datasets, sample_ratios, virtual_size): + def get_counts(sample_ratios): + counts = np.array([virtual_size * r for r in sample_ratios], dtype=np.int64) + diff = virtual_size - counts.sum() + assert diff >= 0 + # due to round-offs, the size might not match the desired sizes + if diff > 0: + dataset_indices = rng.choice(len(sample_ratios), size=diff, p=sample_ratios) + for i in dataset_indices: + counts[i] += 1 + return counts + + def get_in_dataset_indices(datasets, sizes, sample_ratios): + counts = get_counts(sample_ratios) + # uniformally sample desired counts for each dataset + # if the desired counts are large, sample with replacement: + indices = [ + self.random_choice_in_dataset(rng, d, c) + for c, d in zip(counts, datasets)] + return indices + + sizes = [len(d) for d in datasets] + if sample_ratios is None: + # default back to concating datasets + in_dataset_indices = [list(range(s)) for s in sizes] + virtual_sizes_per_dataset = sizes + else: + sample_ratios = sample_ratios.array + ratios = sample_ratios / sample_ratios.sum() + in_dataset_indices = get_in_dataset_indices(datasets, sizes, ratios) + virtual_sizes_per_dataset = [len(d) for d in in_dataset_indices] + virtual_sizes_per_dataset = np.array(virtual_sizes_per_dataset, np.int64) + cumulative_sizes = np.cumsum(virtual_sizes_per_dataset) + assert sum(virtual_sizes_per_dataset) == virtual_size + assert cumulative_sizes[-1] == virtual_size + if virtual_size < sum(sizes): + logger.warning( + f'virtual data size ({virtual_size}) is less than real data size ({sum(sizes)}).' + ' If virtual size << real data size, there could be data coverage issue.' + ) + in_dataset_indices = np.hstack(in_dataset_indices) + return in_dataset_indices, cumulative_sizes, virtual_sizes_per_dataset + + def _get_dataset_and_index(self, index): + i = bisect_right(self.cumulated_sizes.array, index) + return i, self._cur_indices.array[index] + + def __getitem__(self, index): + ds_idx, ds_sample_idx = self._get_dataset_and_index(index) + ret = (ds_idx, self.datasets[ds_idx][ds_sample_idx]) + return ret + + def num_tokens(self, index): + ds_idx, ds_sample_idx = self._get_dataset_and_index(index) + return self.datasets[ds_idx].num_tokens(ds_sample_idx) + + def size(self, index): + if self._sizes is not None: + return self._sizes[index] + ds_idx, ds_sample_idx = self._get_dataset_and_index(index) + return self.datasets[ds_idx].size(ds_sample_idx) + + def __len__(self): + return self.virtual_size + + def collater(self, samples, **extra_args): + """Merge a list of samples to form a mini-batch.""" + if len(samples) == 0: + return None + if self.collate_format == 'ordered_dict': + collect_samples = [[] for _ in range(len(self.datasets))] + for (i, sample) in samples: + collect_samples[i].append(sample) + return OrderedDict([ + (self.keys[i], dataset.collater(collect_samples[i])) + for i, (key, dataset) in enumerate(zip(self.keys, self.datasets)) + if len(collect_samples[i]) > 0 + ]) + elif self.shared_collater: + return self.datasets[0].collater( + [s for _, s in samples] + ) + else: + samples_dict = defaultdict(list) + pad_to_length = defaultdict(int) if 'pad_to_length' not in extra_args else extra_args['pad_to_length'] + for ds_idx, s in samples: + pad_to_length['source'] = max(pad_to_length['source'], s['source'].size(0)) + if s['target'] is not None: + pad_to_length['target'] = max(pad_to_length['target'], s['target'].size(0)) + samples_dict[ds_idx].append(s) + batches = [ + self.datasets[i].collater(samples_dict[i], pad_to_length=pad_to_length) + for i in range(len(self.datasets)) + if len(samples_dict[i]) > 0 + ] + + def straight_data(tensors): + batch = torch.cat(tensors, dim=0) + return batch + + src_lengths = straight_data([b['net_input']['src_lengths'] for b in batches]) + src_lengths, sort_order = src_lengths.sort(descending=True) + + def straight_order(tensors): + batch = straight_data(tensors) + return batch.index_select(0, sort_order) + + batch = { + 'id': straight_order([b['id'] for b in batches]), + 'nsentences': sum(b['nsentences'] for b in batches), + 'ntokens': sum(b['ntokens'] for b in batches), + 'net_input': { + 'src_tokens': straight_order([b['net_input']['src_tokens'] for b in batches]), + 'src_lengths': src_lengths, + }, + 'target': straight_order([b['target'] for b in batches]) if batches[0]['target'] is not None else None, + } + if 'prev_output_tokens' in batches[0]['net_input']: + batch['net_input']['prev_output_tokens'] = straight_order( + [b['net_input']['prev_output_tokens'] for b in batches]) + if 'src_lang_id' in batches[0]['net_input']: + batch['net_input']['src_lang_id'] = straight_order([b['net_input']['src_lang_id'] for b in batches]) + if 'tgt_lang_id' in batches[0]: + batch['tgt_lang_id'] = straight_order([b['tgt_lang_id'] for b in batches]) + return batch + + @property + def sizes(self): + if self._sizes is not None: + return self._sizes + start_time = time.time() + size_cache = self._size_cache + ret = [] + for i in range(len(self)): + ds_idx, ds_sample_idx = self._get_dataset_and_index(i) + if (ds_idx, ds_sample_idx) in size_cache: + ret.append(size_cache[(ds_idx, ds_sample_idx)]) + else: + s = self.datasets[ds_idx].size(ds_sample_idx) + size_cache[(ds_idx, ds_sample_idx)] = s + ret.append(s) + logger.debug(f'sizes() calling time: {get_time_gap(start_time, time.time())}') + self._sizes = np.array(ret, np.int64) + return self._sizes + + def ordered_indices(self): + if self._ordered_indices is not None: + return self._ordered_indices + + if self.batch_by_size: + # No need to do shuffle as the data items are already randomized + indices = np.arange(len(self)) + sizes = self.sizes + tgt_sizes = sizes[:, 1] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else None + src_sizes = sizes[:, 0] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else sizes + + # sort by target length, then source length + if tgt_sizes is not None: + indices = indices[ + np.argsort(tgt_sizes[indices], kind='mergesort') + ] + sort_indices = indices[np.argsort(src_sizes[indices], kind='mergesort')] + else: + sort_indices = np.arange(len(self)) + self._ordered_indices = sort_indices + return sort_indices + + def prefetch(self, indices): + prefetch_indices = [[] for _ in range(len(self.datasets))] + for i in indices: + ds_idx, ds_sample_idx = self._get_dataset_and_index(i) + prefetch_indices[ds_idx].append(ds_sample_idx) + for i in range(len(prefetch_indices)): + self.datasets[i].prefetch(prefetch_indices[i]) + + def set_epoch(self, epoch): + super().set_epoch(epoch) + if epoch == self._cur_epoch: + # re-enter so return + return + for d in self.datasets: + if hasattr(d, 'set_epoch'): + d.set_epoch(epoch) + self._cur_epoch = epoch + self._establish_virtual_datasets() + + def _establish_virtual_datasets(self): + if self.sample_ratios is None and self._cur_indices is not None: + # not a samping dataset, no need to resample if indices are already established + return + self._reset_cached_properties() + + start_time = time.time() + # Generate a weighted sample of indices as a function of the + # random seed and the current epoch. + rng = np.random.RandomState( + [ + int(hashlib.sha1(str(self.__class__.__name__).encode('utf-8')).hexdigest(), 16) % (2 ** 32), + self.seed % (2 ** 32), # global seed + self._cur_epoch, # epoch index, + ] + ) + indices, cumulated_sizes, virtual_size_per_dataset = self.get_virtual_indices( + rng, self.datasets, self.sample_ratios, self.virtual_size) + + self._clean_if_not_none([ + self.cumulated_sizes, self.virtual_size_per_dataset + ]) + self._cur_indices = plasma_utils.PlasmaArray(indices) + self.cumulated_sizes = plasma_utils.PlasmaArray(cumulated_sizes) + self.virtual_size_per_dataset = plasma_utils.PlasmaArray(virtual_size_per_dataset) + + raw_sizes = [len(d) for d in self.datasets] + sampled_sizes = self.virtual_size_per_dataset.array + logger.info(f'[{self.split}] Raw sizes: {str(dict(zip(self.keys, raw_sizes)))}; ' + f'raw total size: {sum(raw_sizes)}') + logger.info(f'[{self.split}] Resampled sizes: {str(dict(zip(self.keys, sampled_sizes)))}; ' + f'resampled total size: {sum(sampled_sizes)}') + if self.sample_ratios is not None: + logger.info(f'[{self.split}] Upsampling ratios: {str(dict(zip(self.keys, self.sample_ratios.array)))}') + else: + logger.info(f'[{self.split}] A concat dataset') + logger.debug(f'[{self.split}] virtual dataset established time: {get_time_gap(start_time, time.time())}') diff --git a/fairseq/data/multilingual/sampled_multi_epoch_dataset.py b/fairseq/data/multilingual/sampled_multi_epoch_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..fdd47e50910d8bc9a37f6936264591e8ae8413ae --- /dev/null +++ b/fairseq/data/multilingual/sampled_multi_epoch_dataset.py @@ -0,0 +1,258 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import hashlib +import math +import logging +import time + +import numpy as np +import torch +from fairseq import distributed_utils +from fairseq.data import plasma_utils, SampledMultiDataset +from .sampled_multi_dataset import default_virtual_size_func, get_time_gap, CollateFormat + + +logger = logging.getLogger(__name__) + + +class SampledMultiEpochDataset(SampledMultiDataset): + """Samples from multiple sub-datasets according to sampling ratios + using virtual epoch sizes to speed up dataloading. + Args: + datasets ( + List[~torch.utils.data.Dataset] + or OrderedDict[str, ~torch.utils.data.Dataset] + ): datasets + sampling_ratios (List[float]): list of probability of each dataset to be sampled + (default: None, which corresponds to concating all dataset together). + batch_by_size (bool): whether or not to batch by sequence length + (default: True). + seed (int): RNG seed to use (default: 2). + epoch (int): starting epoch number (default: 1). + eval_key (str, optional): a key used at evaluation time that causes + this instance to pass-through batches from *datasets[eval_key]*. + collate_format (CollateFormat): collater output format, either CollateFormat.ordered_dict or + CollateFormat.single (default: CollateFormat.single) where CollateFormat.single configures + the collater to output batches of data mixed from all sub-datasets, + and CollateFormat.ordered_dict configures the collater to output a dictionary of batches indexed by keys + of sub-datasets. + Note that not all sub-datasets will present in a single batch in both formats. + virtual_size (int, or callable): the expected virtual size of the dataset (default: default_virtual_size_func). + split (str): the split of the data, e.g. 'train', 'valid' or 'test'. + virtual_epoch_size (int): virtual epoch size, the dataset will go through the data by + this virtual epoch size one by one to speed up data loading, e.g. indicing and filtering + can be performed whenever a virtual epoch is loaded without waiting for the whole dataset to be loaded. + shared_collater (bool): whether or not to all sub-datasets have the same collater. + shard_epoch (int): the real epoch number for shard selection. + """ + def __init__( + self, + datasets, + sampling_ratios=None, + batch_by_size=False, + seed=2, + epoch=1, + eval_key=None, + collate_format=CollateFormat.single, + virtual_size=default_virtual_size_func, + split='', + virtual_epoch_size=None, + shared_collater=False, + shard_epoch=1, + ): + self.virtual_epoch_size = virtual_epoch_size + self._current_epoch_start_index = None + self._epoch_sizes = None + self._epoch_ordered_indices = None + self._random_globa_indices = None + self.shard_epoch = shard_epoch if shard_epoch is not None else 1 + self.load_next_shard = None + super().__init__( + datasets=datasets, + sampling_ratios=sampling_ratios, + batch_by_size=batch_by_size, + seed=seed, + epoch=epoch, + eval_key=eval_key, + collate_format=collate_format, + virtual_size=virtual_size, + split=split, + shared_collater=shared_collater, + ) + + def _setup(self, epoch): + self.virtual_epoch_size = self.virtual_epoch_size if self.virtual_epoch_size is not None else self.virtual_size + if self.virtual_epoch_size > self.virtual_size: + logger.warning(f'virtual epoch size {self.virtual_epoch_size} ' + f'is greater than virtual dataset size {self.virtual_size}') + self.virtual_epoch_size = self.virtual_size + self.num_virtual_epochs = math.ceil(self.virtual_size / self.virtual_epoch_size) + self._current_epoch_start_index = self._get_epoch_start_index(epoch) + logger.info(f'virtual epoch size {self.virtual_epoch_size}; virtual dataset size {self.virtual_size}') + + def _map_epoch_index_to_global(self, index): + index = self._current_epoch_start_index + index + # add randomness + return self._random_globa_indices.array[index] + + def __getitem__(self, index): + i = self._map_epoch_index_to_global(index) + return super().__getitem__(i) + + def num_tokens(self, index): + i = self._map_epoch_index_to_global(index) + return super().num_tokens(i) + + def size(self, index): + if self._epoch_sizes is not None: + return self._epoch_sizes.array[index] + index = self._map_epoch_index_to_global(index) + ds_idx, ds_sample_idx = self._get_dataset_and_index(index) + return self.datasets[ds_idx].size(ds_sample_idx) + + def __len__(self): + return ( + self.virtual_epoch_size + if self._current_epoch_start_index + self.virtual_epoch_size < self.virtual_size + else self.virtual_size - self._current_epoch_start_index + ) + + @property + def sizes(self): + if self._epoch_sizes is not None: + return self._epoch_sizes.array + start_time = time.time() + + size_cache = self._size_cache + ret = [] + for i in range(len(self)): + index = self._map_epoch_index_to_global(i) + ds_idx, ds_sample_idx = self._get_dataset_and_index(index) + + if (ds_idx, ds_sample_idx) in size_cache: + ret.append(size_cache[(ds_idx, ds_sample_idx)]) + else: + s = self.datasets[ds_idx].size(ds_sample_idx) + s = (s, s) if not isinstance(s, tuple) else s + size_cache[(ds_idx, ds_sample_idx)] = s + ret.append(s) + self._epoch_sizes = plasma_utils.PlasmaArray(np.array(ret, np.int64)) + logger.info(f'sizes() calling time: {get_time_gap(start_time, time.time())}') + return self._epoch_sizes.array + + def ordered_indices(self): + if self._epoch_ordered_indices is not None: + return self._epoch_ordered_indices.array + + if self.batch_by_size: + # No need to do shuffle as the data items are already randomized + indices = np.arange(len(self)) + sizes = self.sizes + tgt_sizes = sizes[:, 1] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else None + src_sizes = sizes[:, 0] if len(sizes.shape) > 0 and sizes.shape[1] > 1 else sizes + + # sort by target length, then source length + if tgt_sizes is not None: + indices = indices[ + np.argsort(tgt_sizes[indices], kind='mergesort') + ] + sort_indices = indices[np.argsort(src_sizes[indices], kind='mergesort')] + else: + sort_indices = np.arange(len(self)) + self._epoch_ordered_indices = plasma_utils.PlasmaArray(sort_indices) + return self._epoch_ordered_indices.array + + def prefetch(self, indices): + prefetch_indices = [[] for _ in range(len(self.datasets))] + for i in indices: + index = self._map_epoch_index_to_global(i) + ds_idx, ds_sample_idx = self._get_dataset_and_index(index) + prefetch_indices[ds_idx].append(ds_sample_idx) + for i in range(len(prefetch_indices)): + self.datasets[i].prefetch(prefetch_indices[i]) + + def set_epoch(self, epoch): + if self._current_epoch_start_index is None: + self._setup(epoch) + self._next_virtual_epoch(epoch) + if epoch == self._cur_epoch: + # re-enter so return + return + self._next_virtual_epoch(epoch) + + def _get_epoch_start_index(self, epoch): + assert epoch >= 1 # fairseq is using 1-based epoch everywhere + return ((epoch - 1) % self.num_virtual_epochs) * self.virtual_epoch_size + + def _next_global_indices(self, epoch): + rng = np.random.RandomState( + [ + int(hashlib.sha1(str(self.__class__.__name__).encode('utf-8')).hexdigest(), 16) % (2 ** 32), + self.seed % (2 ** 32), # global seed + epoch, # epoch index, + ] + ) + del self._random_globa_indices + self._random_globa_indices = plasma_utils.PlasmaArray( + rng.choice(self.virtual_size, self.virtual_size, replace=False)) + if self.load_next_shard is None: + self.load_next_shard = False + else: + # increase shard epoch for next loading + self.shard_epoch += 1 + self.load_next_shard = True + # a hack to avoid possible out of sync of shard epoch number + # TODO: to confirm whether this is needed; without it, CUDA event error is occassionally observed + synced_shard_epoch = self._sync_shard_epoch(self.shard_epoch) + logger.info('to load next epoch/shard in next load_dataset: ' + f'epoch={epoch}/shard_epoch={self.shard_epoch}[synced={synced_shard_epoch}]') + + def _sync_shard_epoch(self, shard_epoch): + # in case the ratios are not precisely the same across processes + # also to ensure every procresses update the ratios in the same pace + shard_epoch = torch.DoubleTensor([shard_epoch]) + if torch.distributed.is_initialized(): + if torch.cuda.is_available(): + distributed_utils.all_reduce(shard_epoch.cuda()) + else: + distributed_utils.all_reduce(shard_epoch) + ret = shard_epoch.cpu() + ret = ret.numpy() + return ret + + def _sync_epoch(self, epoch): + # in case the ratios are not precisely the same across processes + # also to ensure every procresses update the ratios in the same pace + epoch = torch.DoubleTensor([epoch]) + if torch.distributed.is_initialized(): + if torch.cuda.is_available(): + distributed_utils.all_reduce(epoch.cuda()) + else: + distributed_utils.all_reduce(epoch) + ret = epoch.cpu() + ret = ret.numpy() + return ret + + def _next_virtual_epoch(self, epoch): + index = self._get_epoch_start_index(epoch) + if index == 0 or self._random_globa_indices is None: + # need to start from the beginning, + # so call super().set_epoch(epoch) to establish the global virtual indices + logger.info('establishing a new set of global virtual indices for ' + f'epoch={epoch}/shard_epoch={self.shard_epoch}') + super().set_epoch(epoch) + self._next_global_indices(epoch) + else: + self._cur_epoch = epoch + # reset cache sizes and ordered_indices for the epoch after moving to a new epoch + + self._clean_if_not_none([ + self._epoch_sizes, self._epoch_ordered_indices, self._size_cache + ]) + self._epoch_sizes = None + self._epoch_ordered_indices = None + self._current_epoch_start_index = index + self._size_cache = {} diff --git a/fairseq/data/multilingual/sampling_method.py b/fairseq/data/multilingual/sampling_method.py new file mode 100644 index 0000000000000000000000000000000000000000..6a9d39f7a6e48104201d9a8019abd9124aa775b8 --- /dev/null +++ b/fairseq/data/multilingual/sampling_method.py @@ -0,0 +1,66 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import List +import logging + + +logger = logging.getLogger(__name__) + + +def uniform(dataset_sizes: List[int]): + return [1.0] * len(dataset_sizes) + + +def temperature_sampling(dataset_sizes, temp): + total_size = sum(dataset_sizes) + return [(size / total_size) ** (1.0/temp) for size in dataset_sizes] + + +def make_temperature_sampling(temp=1.0): + def sampling_func(dataset_sizes): + return temperature_sampling(dataset_sizes, temp) + return sampling_func + + +def make_ratio_sampling(ratios): + def sampling_func(dataset_sizes): + return ratios + return sampling_func + + +class SamplingMethod: + @staticmethod + def add_arguments(parser): + parser.add_argument( + '--sampling-method', + choices=['uniform', 'temperature', 'concat', 'RoundRobin', ], + type=str, + default='concat', + help='The method to sample data per language pairs') + parser.add_argument('--sampling-temperature', default=1.5, type=float, + help='only work with --sampling-method temperature') + + @staticmethod + def build_sampler(args, task): + return SamplingMethod(args, task) + + def __init__(self, args, task): + self.args = args + self.task = task + + def is_adaptive(self): + return False + + def sampling_method_selector(self): + args = self.args + logger.info(f'selected sampler: {args.sampling_method}') + if args.sampling_method == 'uniform': + return uniform + elif args.sampling_method == 'temperature' or self.is_adaptive(): + return make_temperature_sampling(float(args.sampling_temperature)) + else: + # default to concating all data set together + return None diff --git a/fairseq/data/nested_dictionary_dataset.py b/fairseq/data/nested_dictionary_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..2795f895ddeeb70b53368c3624d466436810c7f9 --- /dev/null +++ b/fairseq/data/nested_dictionary_dataset.py @@ -0,0 +1,116 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict + +import torch +from torch.utils.data.dataloader import default_collate + +from . import FairseqDataset + + +def _flatten(dico, prefix=None): + """Flatten a nested dictionary.""" + new_dico = OrderedDict() + if isinstance(dico, dict): + prefix = prefix + '.' if prefix is not None else '' + for k, v in dico.items(): + if v is None: + continue + new_dico.update(_flatten(v, prefix + k)) + elif isinstance(dico, list): + for i, v in enumerate(dico): + new_dico.update(_flatten(v, prefix + '.[' + str(i) + ']')) + else: + new_dico = OrderedDict({prefix: dico}) + return new_dico + + +def _unflatten(dico): + """Unflatten a flattened dictionary into a nested dictionary.""" + new_dico = OrderedDict() + for full_k, v in dico.items(): + full_k = full_k.split('.') + node = new_dico + for k in full_k[:-1]: + if k.startswith('[') and k.endswith(']'): + k = int(k[1:-1]) + if k not in node: + node[k] = OrderedDict() + node = node[k] + node[full_k[-1]] = v + return new_dico + + +class NestedDictionaryDataset(FairseqDataset): + + def __init__(self, defn, sizes=None): + super().__init__() + self.defn = _flatten(defn) + self.sizes = [sizes] if not isinstance(sizes, (list, tuple)) else sizes + + first = None + for v in self.defn.values(): + if not isinstance(v, (FairseqDataset, torch.utils.data.Dataset, )): + raise ValueError('Expected Dataset but found: {}'.format(v.__class__)) + first = first or v + if len(v) > 0: + assert len(v) == len(first), 'dataset lengths must match' + + self._len = len(first) + + def __getitem__(self, index): + return OrderedDict((k, ds[index]) for k, ds in self.defn.items()) + + def __len__(self): + return self._len + + def collater(self, samples): + """Merge a list of samples to form a mini-batch. + + Args: + samples (List[dict]): samples to collate + + Returns: + dict: a mini-batch suitable for forwarding with a Model + """ + if len(samples) == 0: + return {} + sample = OrderedDict() + for k, ds in self.defn.items(): + try: + sample[k] = ds.collater([s[k] for s in samples]) + except NotImplementedError: + sample[k] = default_collate([s[k] for s in samples]) + return _unflatten(sample) + + def num_tokens(self, index): + """Return the number of tokens in a sample. This value is used to + enforce ``--max-tokens`` during batching.""" + return max(s[index] for s in self.sizes) + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + if len(self.sizes) == 1: + return self.sizes[0][index] + else: + return (s[index] for s in self.sizes) + + @property + def supports_prefetch(self): + """Whether this dataset supports prefetching.""" + return any(ds.supports_prefetch for ds in self.defn.values()) + + def prefetch(self, indices): + """Prefetch the data required for this epoch.""" + for ds in self.defn.values(): + if getattr(ds, 'supports_prefetch', False): + ds.prefetch(indices) + + def set_epoch(self, epoch): + super().set_epoch(epoch) + for ds in self.defn.values(): + ds.set_epoch(epoch) diff --git a/fairseq/data/noising.py b/fairseq/data/noising.py new file mode 100644 index 0000000000000000000000000000000000000000..5801ae6eac1a0c96750e24bf8d43803f125e03c1 --- /dev/null +++ b/fairseq/data/noising.py @@ -0,0 +1,315 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import numpy as np + +from fairseq.data import data_utils + + +class WordNoising(object): + """Generate a noisy version of a sentence, without changing words themselves.""" + def __init__(self, dictionary, bpe_cont_marker="@@", bpe_end_marker=None): + self.dictionary = dictionary + self.bpe_end = None + if bpe_cont_marker: + self.bpe_end = np.array([ + not self.dictionary[i].endswith(bpe_cont_marker) + for i in range(len(self.dictionary)) + ]) + elif bpe_end_marker: + self.bpe_end = np.array([ + self.dictionary[i].endswith(bpe_end_marker) + for i in range(len(self.dictionary)) + ]) + + self.get_word_idx = ( + self._get_bpe_word_idx + if self.bpe_end is not None + else self._get_token_idx + ) + + def noising(self, x, lengths, noising_prob=0.0): + raise NotImplementedError() + + def _get_bpe_word_idx(self, x): + """ + Given a list of BPE tokens, for every index in the tokens list, + return the index of the word grouping that it belongs to. + For example, for input x corresponding to ["how", "are", "y@@", "ou"], + return [[0], [1], [2], [2]]. + """ + # x: (T x B) + bpe_end = self.bpe_end[x] + + if (x.size(0) == 1 and x.size(1) == 1): + # Special case when we only have one word in x. If x = [[N]], + # bpe_end is a scalar (bool) instead of a 2-dim array of bools, + # which makes the sum operation below fail. + return np.array([[0]]) + + # do a reduce front sum to generate word ids + word_idx = bpe_end[::-1].cumsum(0)[::-1] + word_idx = word_idx.max(0)[None, :] - word_idx + return word_idx + + def _get_token_idx(self, x): + """ + This is to extend noising functions to be able to apply to non-bpe + tokens, e.g. word or characters. + """ + x = torch.t(x) + word_idx = np.array([range(len(x_i)) for x_i in x]) + return np.transpose(word_idx) + + +class WordDropout(WordNoising): + """Randomly drop input words. If not passing blank_idx (default is None), + then dropped words will be removed. Otherwise, it will be replaced by the + blank_idx.""" + + def __init__(self, dictionary, default_dropout_prob=0.1, bpe_cont_marker="@@", bpe_end_marker=None): + super().__init__(dictionary, bpe_cont_marker, bpe_end_marker) + self.default_dropout_prob = default_dropout_prob + + def noising(self, x, lengths, dropout_prob=None, blank_idx=None): + if dropout_prob is None: + dropout_prob = self.default_dropout_prob + # x: (T x B), lengths: B + if dropout_prob == 0: + return x, lengths + + assert 0 < dropout_prob < 1 + + # be sure to drop entire words + word_idx = self.get_word_idx(x) + sentences = [] + modified_lengths = [] + for i in range(lengths.size(0)): + # Since dropout probabilities need to apply over non-pad tokens, + # it is not trivial to generate the keep mask without consider + # input lengths; otherwise, this could be done outside the loop + + # We want to drop whole words based on word_idx grouping + num_words = max(word_idx[:, i]) + 1 + + # ith example: [x0, x1, ..., eos, pad, ..., pad] + # We should only generate keep probs for non-EOS tokens. Thus if the + # input sentence ends in EOS, the last word idx is not included in + # the dropout mask generation and we append True to always keep EOS. + # Otherwise, just generate the dropout mask for all word idx + # positions. + has_eos = x[lengths[i] - 1, i] == self.dictionary.eos() + if has_eos: # has eos? + keep = np.random.rand(num_words - 1) >= dropout_prob + keep = np.append(keep, [True]) # keep EOS symbol + else: + keep = np.random.rand(num_words) >= dropout_prob + + words = x[:lengths[i], i].tolist() + + # TODO: speed up the following loop + # drop words from the input according to keep + new_s = [ + w if keep[word_idx[j, i]] else blank_idx + for j, w in enumerate(words) + ] + new_s = [w for w in new_s if w is not None] + # we need to have at least one word in the sentence (more than the + # start / end sentence symbols) + if len(new_s) <= 1: + # insert at beginning in case the only token left is EOS + # EOS should be at end of list. + new_s.insert(0, words[np.random.randint(0, len(words))]) + assert len(new_s) >= 1 and ( + not has_eos # Either don't have EOS at end or last token is EOS + or (len(new_s) >= 2 and new_s[-1] == self.dictionary.eos()) + ), "New sentence is invalid." + sentences.append(new_s) + modified_lengths.append(len(new_s)) + # re-construct input + modified_lengths = torch.LongTensor(modified_lengths) + modified_x = torch.LongTensor( + modified_lengths.max(), + modified_lengths.size(0) + ).fill_(self.dictionary.pad()) + for i in range(modified_lengths.size(0)): + modified_x[:modified_lengths[i], i].copy_(torch.LongTensor(sentences[i])) + + return modified_x, modified_lengths + + +class WordShuffle(WordNoising): + """Shuffle words by no more than k positions.""" + + def __init__(self, dictionary, default_max_shuffle_distance=3, bpe_cont_marker="@@", bpe_end_marker=None): + super().__init__(dictionary, bpe_cont_marker, bpe_end_marker) + self.default_max_shuffle_distance = 3 + + def noising(self, x, lengths, max_shuffle_distance=None): + if max_shuffle_distance is None: + max_shuffle_distance = self.default_max_shuffle_distance + # x: (T x B), lengths: B + if max_shuffle_distance == 0: + return x, lengths + + # max_shuffle_distance < 1 will return the same sequence + assert max_shuffle_distance > 1 + + # define noise word scores + noise = np.random.uniform( + 0, + max_shuffle_distance, + size=(x.size(0), x.size(1)), + ) + noise[0] = -1 # do not move start sentence symbol + # be sure to shuffle entire words + word_idx = self.get_word_idx(x) + x2 = x.clone() + for i in range(lengths.size(0)): + length_no_eos = lengths[i] + if x[lengths[i] - 1, i] == self.dictionary.eos(): + length_no_eos = lengths[i] - 1 + # generate a random permutation + scores = word_idx[:length_no_eos, i] + noise[word_idx[:length_no_eos, i], i] + # ensure no reordering inside a word + scores += 1e-6 * np.arange(length_no_eos.item()) + permutation = scores.argsort() + # shuffle words + x2[:length_no_eos, i].copy_( + x2[:length_no_eos, i][torch.from_numpy(permutation)] + ) + return x2, lengths + + +class UnsupervisedMTNoising(WordNoising): + """ + Implements the default configuration for noising in UnsupervisedMT + (github.com/facebookresearch/UnsupervisedMT) + """ + def __init__( + self, + dictionary, + max_word_shuffle_distance, + word_dropout_prob, + word_blanking_prob, + bpe_cont_marker="@@", + bpe_end_marker=None, + ): + super().__init__(dictionary) + self.max_word_shuffle_distance = max_word_shuffle_distance + self.word_dropout_prob = word_dropout_prob + self.word_blanking_prob = word_blanking_prob + + self.word_dropout = WordDropout( + dictionary=dictionary, + bpe_cont_marker=bpe_cont_marker, + bpe_end_marker=bpe_end_marker, + ) + self.word_shuffle = WordShuffle( + dictionary=dictionary, + bpe_cont_marker=bpe_cont_marker, + bpe_end_marker=bpe_end_marker, + ) + + def noising(self, x, lengths): + # 1. Word Shuffle + noisy_src_tokens, noisy_src_lengths = self.word_shuffle.noising( + x=x, + lengths=lengths, + max_shuffle_distance=self.max_word_shuffle_distance, + ) + # 2. Word Dropout + noisy_src_tokens, noisy_src_lengths = self.word_dropout.noising( + x=noisy_src_tokens, + lengths=noisy_src_lengths, + dropout_prob=self.word_dropout_prob, + ) + # 3. Word Blanking + noisy_src_tokens, noisy_src_lengths = self.word_dropout.noising( + x=noisy_src_tokens, + lengths=noisy_src_lengths, + dropout_prob=self.word_blanking_prob, + blank_idx=self.dictionary.unk(), + ) + + return noisy_src_tokens + + +class NoisingDataset(torch.utils.data.Dataset): + def __init__( + self, + src_dataset, + src_dict, + seed, + noiser=None, + noising_class=UnsupervisedMTNoising, + **kwargs + ): + """ + Wrap a :class:`~torch.utils.data.Dataset` and apply noise to the + samples based on the supplied noising configuration. + + Args: + src_dataset (~torch.utils.data.Dataset): dataset to wrap. + to build self.src_dataset -- + a LanguagePairDataset with src dataset as the source dataset and + None as the target dataset. Should NOT have padding so that + src_lengths are accurately calculated by language_pair_dataset + collate function. + We use language_pair_dataset here to encapsulate the tgt_dataset + so we can re-use the LanguagePairDataset collater to format the + batches in the structure that SequenceGenerator expects. + src_dict (~fairseq.data.Dictionary): source dictionary + seed (int): seed to use when generating random noise + noiser (WordNoising): a pre-initialized :class:`WordNoising` + instance. If this is None, a new instance will be created using + *noising_class* and *kwargs*. + noising_class (class, optional): class to use to initialize a + default :class:`WordNoising` instance. + kwargs (dict, optional): arguments to initialize the default + :class:`WordNoising` instance given by *noiser*. + """ + self.src_dataset = src_dataset + self.src_dict = src_dict + self.seed = seed + self.noiser = noiser if noiser is not None else noising_class( + dictionary=src_dict, **kwargs, + ) + + def __getitem__(self, index): + """ + Returns a single noisy sample. Multiple samples are fed to the collater + create a noising dataset batch. + """ + src_tokens = self.src_dataset[index] + src_lengths = torch.LongTensor([len(src_tokens)]) + src_tokens = src_tokens.unsqueeze(0) + + # Transpose src tokens to fit expected shape of x in noising function + # (batch size, sequence length) -> (sequence length, batch size) + src_tokens_t = torch.t(src_tokens) + + with data_utils.numpy_seed(self.seed + index): + noisy_src_tokens = self.noiser.noising(src_tokens_t, src_lengths) + + # Transpose back to expected src_tokens format + # (sequence length, 1) -> (1, sequence length) + noisy_src_tokens = torch.t(noisy_src_tokens) + return noisy_src_tokens[0] + + def __len__(self): + """ + The length of the noising dataset is the length of src. + """ + return len(self.src_dataset) + + @property + def supports_prefetch(self): + return self.src_dataset.supports_prefetch + + def prefetch(self, indices): + if self.src_dataset.supports_prefetch: + self.src_dataset.prefetch(indices) diff --git a/fairseq/data/num_samples_dataset.py b/fairseq/data/num_samples_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..9d7ea440192d5ba91efcc5fa6dfae781c60c74f0 --- /dev/null +++ b/fairseq/data/num_samples_dataset.py @@ -0,0 +1,18 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import FairseqDataset + + +class NumSamplesDataset(FairseqDataset): + + def __getitem__(self, index): + return 1 + + def __len__(self): + return 0 + + def collater(self, samples): + return sum(samples) diff --git a/fairseq/data/numel_dataset.py b/fairseq/data/numel_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..50087e5857cb9c988ff4eadba1bf59cf5527a47f --- /dev/null +++ b/fairseq/data/numel_dataset.py @@ -0,0 +1,32 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class NumelDataset(BaseWrapperDataset): + + def __init__(self, dataset, reduce=False): + super().__init__(dataset) + self.reduce = reduce + + def __getitem__(self, index): + item = self.dataset[index] + if torch.is_tensor(item): + return torch.numel(item) + else: + return np.size(item) + + def __len__(self): + return len(self.dataset) + + def collater(self, samples): + if self.reduce: + return sum(samples) + else: + return torch.tensor(samples) diff --git a/fairseq/data/offset_tokens_dataset.py b/fairseq/data/offset_tokens_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..a6fd559a304d2f9ef06704dee7dbec19a3843b10 --- /dev/null +++ b/fairseq/data/offset_tokens_dataset.py @@ -0,0 +1,16 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class OffsetTokensDataset(BaseWrapperDataset): + + def __init__(self, dataset, offset): + super().__init__(dataset) + self.offset = offset + + def __getitem__(self, idx): + return self.dataset[idx] + self.offset diff --git a/fairseq/data/pad_dataset.py b/fairseq/data/pad_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..4c13b549aab2fbe09feae8b9054934840d89512b --- /dev/null +++ b/fairseq/data/pad_dataset.py @@ -0,0 +1,31 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data import data_utils + +from . import BaseWrapperDataset + + +class PadDataset(BaseWrapperDataset): + + def __init__(self, dataset, pad_idx, left_pad): + super().__init__(dataset) + self.pad_idx = pad_idx + self.left_pad = left_pad + + def collater(self, samples): + return data_utils.collate_tokens(samples, self.pad_idx, left_pad=self.left_pad) + + +class LeftPadDataset(PadDataset): + + def __init__(self, dataset, pad_idx): + super().__init__(dataset, pad_idx, left_pad=True) + + +class RightPadDataset(PadDataset): + + def __init__(self, dataset, pad_idx): + super().__init__(dataset, pad_idx, left_pad=False) diff --git a/fairseq/data/plasma_utils.py b/fairseq/data/plasma_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..33f250eea943ee492a400f331b2bcbe8d69f9018 --- /dev/null +++ b/fairseq/data/plasma_utils.py @@ -0,0 +1,86 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import subprocess +import tempfile + + +class PlasmaArray(object): + """ + Wrapper around numpy arrays that automatically moves the data to shared + memory upon serialization. This is particularly helpful when passing numpy + arrays through multiprocessing, so that data is not unnecessarily + duplicated or pickled. + """ + + def __init__(self, array): + super().__init__() + self.array = array + self.disable = array.nbytes < 134217728 # disable for arrays <128MB + self.object_id = None + self.path = None + + # variables with underscores shouldn't be pickled + self._client = None + self._server = None + self._server_tmp = None + self._plasma = None + + @property + def plasma(self): + if self._plasma is None and not self.disable: + try: + import pyarrow.plasma as plasma + self._plasma = plasma + except ImportError: + self._plasma = None + return self._plasma + + def start_server(self): + if self.plasma is None or self._server is not None: + return + assert self.object_id is None + assert self.path is None + self._server_tmp = tempfile.NamedTemporaryFile() + self.path = self._server_tmp.name + self._server = subprocess.Popen([ + 'plasma_store', + '-m', str(int(1.05 * self.array.nbytes)), + '-s', self.path, + ]) + + @property + def client(self): + if self._client is None: + assert self.path is not None + self._client = self.plasma.connect(self.path) + return self._client + + def __getstate__(self): + if self.plasma is None: + return self.__dict__ + if self.object_id is None: + self.start_server() + self.object_id = self.client.put(self.array) + state = self.__dict__.copy() + del state['array'] + state['_client'] = None + state['_server'] = None + state['_server_tmp'] = None + state['_plasma'] = None + return state + + def __setstate__(self, state): + self.__dict__.update(state) + if self.plasma is None: + return + self.array = self.client.get(self.object_id) + + def __del__(self): + if self._server is not None: + self._server.kill() + self._server = None + self._server_tmp.close() + self._server_tmp = None diff --git a/fairseq/data/prepend_dataset.py b/fairseq/data/prepend_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..ad74784d2d7920e4a6225282d95543ce16ea50d9 --- /dev/null +++ b/fairseq/data/prepend_dataset.py @@ -0,0 +1,28 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class PrependDataset(BaseWrapperDataset): + def __init__(self, dataset, prepend_getter, ensure_first_token_is=None): + super().__init__(dataset) + self.prepend_getter = prepend_getter + self.ensure_first_token = ensure_first_token_is + + def __getitem__(self, idx): + item = self.dataset[idx] + is_tuple = isinstance(item, tuple) + src = item[0] if is_tuple else item + + assert self.ensure_first_token is None or src[0] == self.ensure_first_token + prepend_idx = self.prepend_getter(self.dataset, idx) + assert isinstance(prepend_idx, int) + src[0] = prepend_idx + item = tuple((src,) + item[1:]) if is_tuple else src + return item diff --git a/fairseq/data/prepend_token_dataset.py b/fairseq/data/prepend_token_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..9dac71badf9f182ed5a1afdca99815a63717e214 --- /dev/null +++ b/fairseq/data/prepend_token_dataset.py @@ -0,0 +1,42 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from . import BaseWrapperDataset + + +class PrependTokenDataset(BaseWrapperDataset): + + def __init__(self, dataset, token=None): + super().__init__(dataset) + self.token = token + if token is not None: + self._sizes = np.array(dataset.sizes) + 1 + else: + self._sizes = dataset.sizes + + def __getitem__(self, idx): + item = self.dataset[idx] + if self.token is not None: + item = torch.cat([item.new([self.token]), item]) + return item + + @property + def sizes(self): + return self._sizes + + def num_tokens(self, index): + n = self.dataset.num_tokens(index) + if self.token is not None: + n += 1 + return n + + def size(self, index): + n = self.dataset.size(index) + if self.token is not None: + n += 1 + return n diff --git a/fairseq/data/raw_label_dataset.py b/fairseq/data/raw_label_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..e67170f1a5e1dc5f861dd651ec126a05a09cb085 --- /dev/null +++ b/fairseq/data/raw_label_dataset.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class RawLabelDataset(FairseqDataset): + + def __init__(self, labels): + super().__init__() + self.labels = labels + + def __getitem__(self, index): + return self.labels[index] + + def __len__(self): + return len(self.labels) + + def collater(self, samples): + return torch.tensor(samples) diff --git a/fairseq/data/replace_dataset.py b/fairseq/data/replace_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..3bc52f0fb52b62ce689494973ed49acda462be87 --- /dev/null +++ b/fairseq/data/replace_dataset.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class ReplaceDataset(BaseWrapperDataset): + """Replaces tokens found in the dataset by a specified replacement token + + Args: + dataset (~torch.utils.data.Dataset): dataset to replace tokens in + replace_map(Dictionary[int,int]): map of token to replace -> replacement token + offsets (List[int]): do not replace tokens before (from left if pos, right if neg) this offset. should be + as many as the number of objects returned by the underlying dataset __getitem__ method. + """ + + def __init__(self, dataset, replace_map, offsets): + super().__init__(dataset) + assert len(replace_map) > 0 + self.replace_map = replace_map + self.offsets = offsets + + def __getitem__(self, index): + item = self.dataset[index] + is_tuple = isinstance(item, tuple) + srcs = item if is_tuple else [item] + + for offset, src in zip(self.offsets, srcs): + for k, v in self.replace_map.items(): + src_off = src[offset:] if offset >= 0 else src[:offset] + src_off.masked_fill_(src_off == k, v) + + item = srcs if is_tuple else srcs[0] + return item diff --git a/fairseq/data/resampling_dataset.py b/fairseq/data/resampling_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..a2c9b31d7953840c89ce5a898ba66cc12c5f4f1b --- /dev/null +++ b/fairseq/data/resampling_dataset.py @@ -0,0 +1,136 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np + +from fairseq.data import BaseWrapperDataset, plasma_utils + + +logger = logging.getLogger(__name__) + + +class ResamplingDataset(BaseWrapperDataset): + """Randomly samples from a given dataset at each epoch. + + Sampling is done with or without replacement, depending on the "replace" + parameter. + + Optionally, the epoch size can be rescaled. This is potentially desirable + to increase per-epoch coverage of the base dataset (since sampling with + replacement means that many items in the dataset will be left out). In the + case of sampling without replacement, size_ratio should be strictly less + than 1. + + Args: + dataset (~torch.utils.data.Dataset): dataset on which to sample. + weights (List[float]): list of probability weights + (default: None, which corresponds to uniform sampling). + replace (bool): sampling mode; True for "with replacement", or False + for "without replacement" (default: True) + size_ratio (float): the ratio to subsample to; must be positive + (default: 1.0). + batch_by_size (bool): whether or not to batch by sequence length + (default: True). + seed (int): RNG seed to use (default: 0). + epoch (int): starting epoch number (default: 1). + """ + + def __init__( + self, + dataset, + weights=None, + replace=True, + size_ratio=1.0, + batch_by_size=True, + seed=0, + epoch=1, + ): + super().__init__(dataset) + + if weights is None: + self.weights = None + + else: + assert len(weights) == len(dataset) + weights_arr = np.array(weights, dtype=np.float64) + weights_arr /= weights_arr.sum() + self.weights = plasma_utils.PlasmaArray(weights_arr) + + self.replace = replace + + assert size_ratio > 0.0 + if not self.replace: + assert size_ratio < 1.0 + self.size_ratio = float(size_ratio) + self.actual_size = np.ceil(len(dataset) * self.size_ratio).astype(int) + + self.batch_by_size = batch_by_size + self.seed = seed + + self._cur_epoch = None + self._cur_indices = None + + self.set_epoch(epoch) + + def __getitem__(self, index): + return self.dataset[self._cur_indices.array[index]] + + def __len__(self): + return self.actual_size + + @property + def sizes(self): + if isinstance(self.dataset.sizes, list): + return [s[self._cur_indices.array] for s in self.dataset.sizes] + return self.dataset.sizes[self._cur_indices.array] + + def num_tokens(self, index): + return self.dataset.num_tokens(self._cur_indices.array[index]) + + def size(self, index): + return self.dataset.size(self._cur_indices.array[index]) + + def ordered_indices(self): + if self.batch_by_size: + order = [ + np.arange(len(self)), + self.sizes, + ] # No need to handle `self.shuffle == True` + return np.lexsort(order) + else: + return np.arange(len(self)) + + def prefetch(self, indices): + self.dataset.prefetch(self._cur_indices.array[indices]) + + def set_epoch(self, epoch): + logger.debug('ResamplingDataset.set_epoch: {}'.format(epoch)) + super().set_epoch(epoch) + + if epoch == self._cur_epoch: + return + + self._cur_epoch = epoch + + # Generate a weighted sample of indices as a function of the + # random seed and the current epoch. + + rng = np.random.RandomState( + [ + 42, # magic number + self.seed % (2 ** 32), # global seed + self._cur_epoch, # epoch index + ] + ) + self._cur_indices = plasma_utils.PlasmaArray( + rng.choice( + len(self.dataset), + self.actual_size, + replace=self.replace, + p=(None if self.weights is None else self.weights.array), + ) + ) diff --git a/fairseq/data/roll_dataset.py b/fairseq/data/roll_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..d07800d0f6625446b24d031a112204c962ef751c --- /dev/null +++ b/fairseq/data/roll_dataset.py @@ -0,0 +1,19 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import BaseWrapperDataset + + +class RollDataset(BaseWrapperDataset): + + def __init__(self, dataset, shifts): + super().__init__(dataset) + self.shifts = shifts + + def __getitem__(self, index): + item = self.dataset[index] + return torch.roll(item, self.shifts) diff --git a/fairseq/data/round_robin_zip_datasets.py b/fairseq/data/round_robin_zip_datasets.py new file mode 100644 index 0000000000000000000000000000000000000000..5bfc966ce8fb33080152c6c910969b5350610b71 --- /dev/null +++ b/fairseq/data/round_robin_zip_datasets.py @@ -0,0 +1,110 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict + +import numpy as np + +from . import FairseqDataset + + +class RoundRobinZipDatasets(FairseqDataset): + """Zip multiple :class:`~fairseq.data.FairseqDataset` instances together. + + Shorter datasets are repeated in a round-robin fashion to match the length + of the longest one. + + Args: + datasets (Dict[~fairseq.data.FairseqDataset]): a dictionary of + :class:`~fairseq.data.FairseqDataset` instances. + eval_key (str, optional): a key used at evaluation time that causes + this instance to pass-through batches from *datasets[eval_key]*. + """ + + def __init__(self, datasets, eval_key=None): + super().__init__() + assert isinstance(datasets, OrderedDict) + self.datasets = datasets + self.eval_key = eval_key + + self.longest_dataset = None + self.longest_dataset_key = None + for key, dataset in datasets.items(): + assert isinstance(dataset, FairseqDataset) + if self.longest_dataset is None or len(dataset) > len(self.longest_dataset): + self.longest_dataset = dataset + self.longest_dataset_key = key + + self._ordered_indices = None + + def _map_index(self, key, index): + assert self._ordered_indices is not None, \ + 'Must call RoundRobinZipDatasets.ordered_indices() first' + return self._ordered_indices[key][index % len(self.datasets[key])] + + def __getitem__(self, index): + if self.eval_key is None: + return OrderedDict([ + (key, dataset[self._map_index(key, index)]) + for key, dataset in self.datasets.items() + ]) + else: + # at evaluation time it's useful to pass-through batches from a single key + return self.datasets[self.eval_key][self._map_index(self.eval_key, index)] + + def __len__(self): + return len(self.longest_dataset) + + def collater(self, samples): + """Merge a list of samples to form a mini-batch.""" + if len(samples) == 0: + return None + if self.eval_key is None: + return OrderedDict([ + (key, dataset.collater([sample[key] for sample in samples])) + for key, dataset in self.datasets.items() + ]) + else: + # at evaluation time it's useful to pass-through batches from a single key + return self.datasets[self.eval_key].collater(samples) + + def num_tokens(self, index): + """Return an example's length (number of tokens), used for batching.""" + # TODO make it configurable whether to use max() or sum() here + return max( + dataset.num_tokens(self._map_index(key, index)) + for key, dataset in self.datasets.items() + ) + + def size(self, index): + """Return an example's size as a float or tuple. This value is used when + filtering a dataset with ``--max-positions``.""" + return { + key: dataset.size(self._map_index(key, index)) + for key, dataset in self.datasets.items() + } + + def ordered_indices(self): + """Ordered indices for batching.""" + if self._ordered_indices is None: + # Call the underlying dataset's ordered_indices() here, so that we + # get the same random ordering as we would have from using the + # underlying dataset directly. + self._ordered_indices = OrderedDict([ + (key, dataset.ordered_indices()) + for key, dataset in self.datasets.items() + ]) + return np.arange(len(self)) + + @property + def supports_prefetch(self): + return all( + getattr(dataset, 'supports_prefetch', False) + for dataset in self.datasets.values() + ) + + def prefetch(self, indices): + for key, dataset in self.datasets.items(): + dataset.prefetch([self._map_index(key, index) for index in indices]) diff --git a/fairseq/data/shorten_dataset.py b/fairseq/data/shorten_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..9c84219dc7bccb2562d52a2039616236e05f782a --- /dev/null +++ b/fairseq/data/shorten_dataset.py @@ -0,0 +1,74 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +from fairseq.data import data_utils + +from . import BaseWrapperDataset + + +class TruncateDataset(BaseWrapperDataset): + """Truncate a sequence by returning the first truncation_length tokens + """ + + def __init__(self, dataset, truncation_length): + super().__init__(dataset) + assert truncation_length is not None + self.truncation_length = truncation_length + self.dataset = dataset + + def __getitem__(self, index): + item = self.dataset[index] + item_len = item.size(0) + if item_len > self.truncation_length: + item = item[:self.truncation_length] + return item + + @property + def sizes(self): + return np.minimum(self.dataset.sizes, self.truncation_length) + + def __len__(self): + return len(self.dataset) + + +class RandomCropDataset(TruncateDataset): + """Truncate a sequence by returning a random crop of truncation_length tokens + """ + + def __init__(self, dataset, truncation_length, seed=1): + super().__init__(dataset, truncation_length) + self.seed = seed + self.epoch = 0 + + def set_epoch(self, epoch, **unused): + super().set_epoch(epoch) + self.epoch = epoch + + def __getitem__(self, index): + with data_utils.numpy_seed(self.seed, self.epoch, index): + item = self.dataset[index] + item_len = item.size(0) + excess = item_len - self.truncation_length + if excess > 0: + start_idx = np.random.randint(0, excess) + item = item[start_idx:start_idx+self.truncation_length] + return item + +def maybe_shorten_dataset( + dataset, + split, + shorten_data_split_list, + shorten_method, + tokens_per_sample, + seed, +): + truncate_split = split in shorten_data_split_list.split(',') \ + or len(shorten_data_split_list) == 0 + if shorten_method == 'truncate' and truncate_split: + dataset = TruncateDataset(dataset, tokens_per_sample) + elif shorten_method == 'random_crop' and truncate_split: + dataset = RandomCropDataset(dataset, tokens_per_sample, seed) + return dataset diff --git a/fairseq/data/sort_dataset.py b/fairseq/data/sort_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..9b510b93a0f22d2c9ba98ce36d5a79921f456a77 --- /dev/null +++ b/fairseq/data/sort_dataset.py @@ -0,0 +1,22 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np + +from . import BaseWrapperDataset + + +class SortDataset(BaseWrapperDataset): + + def __init__(self, dataset, sort_order): + super().__init__(dataset) + if not isinstance(sort_order, (list, tuple)): + sort_order = [sort_order] + self.sort_order = sort_order + + assert all(len(so) == len(dataset) for so in sort_order) + + def ordered_indices(self): + return np.lexsort(self.sort_order) diff --git a/fairseq/data/strip_token_dataset.py b/fairseq/data/strip_token_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..e388db0e5fc6575a8d6fa60b7d5f546de7d4c1fd --- /dev/null +++ b/fairseq/data/strip_token_dataset.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import BaseWrapperDataset + + +class StripTokenDataset(BaseWrapperDataset): + + def __init__(self, dataset, id_to_strip): + super().__init__(dataset) + self.id_to_strip = id_to_strip + + def __getitem__(self, index): + item = self.dataset[index] + while len(item) > 0 and item[-1] == self.id_to_strip: + item = item[:-1] + while len(item) > 0 and item[0] == self.id_to_strip: + item = item[1:] + return item diff --git a/fairseq/data/subsample_dataset.py b/fairseq/data/subsample_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..e395674a55572065662095aada5eb1675edfafbb --- /dev/null +++ b/fairseq/data/subsample_dataset.py @@ -0,0 +1,71 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import numpy as np + +from . import BaseWrapperDataset + + +logger = logging.getLogger(__name__) + + +class SubsampleDataset(BaseWrapperDataset): + """Subsamples a given dataset by a specified ratio. Subsampling is done on the number of examples + + Args: + dataset (~torch.utils.data.Dataset): dataset to subsample + size_ratio(float): the ratio to subsample to. must be between 0 and 1 (exclusive) + """ + + def __init__(self, dataset, size_ratio): + super().__init__(dataset) + assert size_ratio < 1 + self.actual_size = np.ceil(len(dataset) * size_ratio).astype(int) + self.indices = np.random.choice( + list(range(len(self.dataset))), self.actual_size, replace=False + ) + logger.info( + "subsampled dataset from {} to {} (ratio={})".format( + len(self.dataset), self.actual_size, size_ratio + ) + ) + + def __getitem__(self, index): + return self.dataset[self.indices[index]] + + def __len__(self): + return self.actual_size + + def collater(self, samples): + return self.dataset.collater(samples) + + @property + def sizes(self): + return self.dataset.sizes[self.indices] + + @property + def name(self): + return self.dataset.name + + def num_tokens(self, index): + return self.dataset.num_tokens(self.indices[index]) + + def size(self, index): + return self.dataset.size(self.indices[index]) + + def ordered_indices(self): + """Return an ordered list of indices. Batches will be constructed based + on this order.""" + if self.shuffle: + order = [np.random.permutation(len(self))] + else: + order = [np.arange(len(self))] + order.append(self.sizes) + return np.lexsort(order) + + def prefetch(self, indices): + self.dataset.prefetch(self.indices[indices]) diff --git a/fairseq/data/token_block_dataset.py b/fairseq/data/token_block_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..4e2f5cc482870cb85c0bc1b47d0ced5ff1b57c59 --- /dev/null +++ b/fairseq/data/token_block_dataset.py @@ -0,0 +1,166 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch + +from fairseq.data import FairseqDataset, plasma_utils + + +class TokenBlockDataset(FairseqDataset): + """Break a Dataset of tokens into blocks. + + Args: + dataset (~torch.utils.data.Dataset): dataset to break into blocks + sizes (List[int]): sentence lengths (required for 'complete' and 'eos') + block_size (int): maximum block size (ignored in 'eos' break mode) + break_mode (str, optional): Mode used for breaking tokens. Values can + be one of: + - 'none': break tokens into equally sized blocks (up to block_size) + - 'complete': break tokens into blocks (up to block_size) such that + blocks contains complete sentences, although block_size may be + exceeded if some sentences exceed block_size + - 'complete_doc': similar to 'complete' mode, but do not + cross document boundaries + - 'eos': each block contains one sentence (block_size is ignored) + include_targets (bool, optional): return next tokens as targets + (default: False). + document_sep_len (int, optional): document separator size (required for + 'complete_doc' break mode). Typically 1 if the sentences have eos + and 0 otherwise. + """ + def __init__( + self, + dataset, + sizes, + block_size, + pad, + eos, + break_mode=None, + include_targets=False, + document_sep_len=1, + ): + try: + from fairseq.data.token_block_utils_fast import ( + _get_slice_indices_fast, + _get_block_to_dataset_index_fast, + ) + except ImportError: + raise ImportError( + 'Please build Cython components with: `pip install --editable .` ' + 'or `python setup.py build_ext --inplace`' + ) + + super().__init__() + self.dataset = dataset + self.pad = pad + self.eos = eos + self.include_targets = include_targets + + assert len(dataset) == len(sizes) + assert len(dataset) > 0 + + if isinstance(sizes, list): + sizes = np.array(sizes, dtype=np.int64) + else: + if torch.is_tensor(sizes): + sizes = sizes.numpy() + sizes = sizes.astype(np.int64) + + break_mode = break_mode if break_mode is not None else 'none' + + # For "eos" break-mode, block_size is not required parameters. + if break_mode == "eos" and block_size is None: + block_size = 0 + + slice_indices = _get_slice_indices_fast(sizes, break_mode, block_size, document_sep_len) + self._sizes = slice_indices[:, 1] - slice_indices[:, 0] + + # build index mapping block indices to the underlying dataset indices + if break_mode == "eos": + # much faster version for eos break mode + block_to_dataset_index = np.stack( + [ + np.arange(len(sizes)), # starting index in dataset + np.zeros( + len(sizes), dtype=np.long + ), # starting offset within starting index + np.arange(len(sizes)), # ending index in dataset + ], + 1, + ) + else: + block_to_dataset_index = _get_block_to_dataset_index_fast( + sizes, + slice_indices, + ) + self._slice_indices = plasma_utils.PlasmaArray(slice_indices) + self._sizes = plasma_utils.PlasmaArray(self._sizes) + self._block_to_dataset_index = plasma_utils.PlasmaArray(block_to_dataset_index) + + @property + def slice_indices(self): + return self._slice_indices.array + + @property + def sizes(self): + return self._sizes.array + + @property + def block_to_dataset_index(self): + return self._block_to_dataset_index.array + + def attr(self, attr: str, index: int): + start_ds_idx, _, _ = self.block_to_dataset_index[index] + return self.dataset.attr(attr, start_ds_idx) + + def __getitem__(self, index): + start_ds_idx, start_offset, end_ds_idx = self.block_to_dataset_index[index] + + buffer = torch.cat( + [self.dataset[idx] for idx in range(start_ds_idx, end_ds_idx + 1)] + ) + + slice_s, slice_e = self.slice_indices[index] + length = slice_e - slice_s + s, e = start_offset, start_offset + length + item = buffer[s:e] + + if self.include_targets: + # *target* is the original sentence (=item) + # *source* is shifted right by 1 (maybe left-padded with eos) + # *past_target* is shifted right by 2 (left-padded as needed) + if s == 0: + source = torch.cat([item.new([self.eos]), buffer[0 : e - 1]]) + past_target = torch.cat( + [item.new([self.pad, self.eos]), buffer[0 : e - 2]] + ) + else: + source = buffer[s - 1 : e - 1] + if s == 1: + past_target = torch.cat([item.new([self.eos]), buffer[0 : e - 2]]) + else: + past_target = buffer[s - 2 : e - 2] + + return source, item, past_target + + return item + + def __len__(self): + return len(self.slice_indices) + + @property + def supports_prefetch(self): + return getattr(self.dataset, "supports_prefetch", False) + + def prefetch(self, indices): + self.dataset.prefetch( + { + ds_idx + for index in indices + for start_ds_idx, _, end_ds_idx in [self.block_to_dataset_index[index]] + for ds_idx in range(start_ds_idx, end_ds_idx + 1) + } + ) diff --git a/fairseq/data/token_block_utils_fast.cpp b/fairseq/data/token_block_utils_fast.cpp new file mode 100644 index 0000000000000000000000000000000000000000..a99e960fd2d0615707a9de9b2d26bad8b9e3db60 --- /dev/null +++ b/fairseq/data/token_block_utils_fast.cpp @@ -0,0 +1,34438 @@ +/* Generated by Cython 3.0.12 */ + +/* BEGIN: Cython Metadata +{ + "distutils": { + "depends": [ + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/arrayobject.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/arrayscalars.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/ndarrayobject.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/ndarraytypes.h", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include/numpy/ufuncobject.h" + ], + "extra_compile_args": [ + "-std=c++11", + "-O3" + ], + "include_dirs": [ + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include", + "/tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/_core/include" + ], + "language": "c++", + "name": "fairseq.data.token_block_utils_fast", + "sources": [ + "fairseq/data/token_block_utils_fast.pyx" + ] + }, + "module_name": "fairseq.data.token_block_utils_fast" +} +END: Cython Metadata */ + +#ifndef PY_SSIZE_T_CLEAN +#define PY_SSIZE_T_CLEAN +#endif /* PY_SSIZE_T_CLEAN */ +#if defined(CYTHON_LIMITED_API) && 0 + #ifndef Py_LIMITED_API + #if CYTHON_LIMITED_API+0 > 0x03030000 + #define Py_LIMITED_API CYTHON_LIMITED_API + #else + #define Py_LIMITED_API 0x03030000 + #endif + #endif +#endif + +#include "Python.h" +#ifndef Py_PYTHON_H + #error Python headers needed to compile C extensions, please install development version of Python. +#elif PY_VERSION_HEX < 0x02070000 || (0x03000000 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x03030000) + #error Cython requires Python 2.7+ or Python 3.3+. +#else +#if defined(CYTHON_LIMITED_API) && CYTHON_LIMITED_API +#define __PYX_EXTRA_ABI_MODULE_NAME "limited" +#else +#define __PYX_EXTRA_ABI_MODULE_NAME "" +#endif +#define CYTHON_ABI "3_0_12" __PYX_EXTRA_ABI_MODULE_NAME +#define __PYX_ABI_MODULE_NAME "_cython_" CYTHON_ABI +#define __PYX_TYPE_MODULE_PREFIX __PYX_ABI_MODULE_NAME "." +#define CYTHON_HEX_VERSION 0x03000CF0 +#define CYTHON_FUTURE_DIVISION 1 +#include +#ifndef offsetof + #define offsetof(type, member) ( (size_t) & ((type*)0) -> member ) +#endif +#if !defined(_WIN32) && !defined(WIN32) && !defined(MS_WINDOWS) + #ifndef __stdcall + #define __stdcall + #endif + #ifndef __cdecl + #define __cdecl + #endif + #ifndef __fastcall + #define __fastcall + #endif +#endif +#ifndef DL_IMPORT + #define DL_IMPORT(t) t +#endif +#ifndef DL_EXPORT + #define DL_EXPORT(t) t +#endif +#define __PYX_COMMA , +#ifndef HAVE_LONG_LONG + #define HAVE_LONG_LONG +#endif +#ifndef PY_LONG_LONG + #define PY_LONG_LONG LONG_LONG +#endif +#ifndef Py_HUGE_VAL + #define Py_HUGE_VAL HUGE_VAL +#endif +#define __PYX_LIMITED_VERSION_HEX PY_VERSION_HEX +#if defined(GRAALVM_PYTHON) + /* For very preliminary testing purposes. Most variables are set the same as PyPy. + The existence of this section does not imply that anything works or is even tested */ + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 1 + #define CYTHON_COMPILING_IN_NOGIL 0 + #undef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 0 + #undef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #if PY_VERSION_HEX < 0x03050000 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #elif !defined(CYTHON_USE_ASYNC_SLOTS) + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #undef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 0 + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #undef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #undef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 1 + #undef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 0 + #undef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 0 + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #undef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 0 + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS (PY_MAJOR_VERSION >= 3) + #endif + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #undef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #undef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 0 + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 0 + #endif + #undef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 +#elif defined(PYPY_VERSION) + #define CYTHON_COMPILING_IN_PYPY 1 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 0 + #undef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 0 + #ifndef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #endif + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #if PY_VERSION_HEX < 0x03050000 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #elif !defined(CYTHON_USE_ASYNC_SLOTS) + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #undef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 0 + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #undef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #undef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 1 + #undef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 0 + #undef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 0 + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #undef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 0 + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS (PY_MAJOR_VERSION >= 3) + #endif + #if PY_VERSION_HEX < 0x03090000 + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 0 + #elif !defined(CYTHON_PEP489_MULTI_PHASE_INIT) + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #endif + #undef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #undef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE (PY_VERSION_HEX >= 0x030400a1 && PYPY_VERSION_NUM >= 0x07030C00) + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 0 + #endif + #undef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 +#elif defined(CYTHON_LIMITED_API) + #ifdef Py_LIMITED_API + #undef __PYX_LIMITED_VERSION_HEX + #define __PYX_LIMITED_VERSION_HEX Py_LIMITED_API + #endif + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 1 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 0 + #undef CYTHON_CLINE_IN_TRACEBACK + #define CYTHON_CLINE_IN_TRACEBACK 0 + #undef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 0 + #undef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 1 + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #undef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 0 + #ifndef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #endif + #undef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #ifndef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 0 + #endif + #undef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 0 + #undef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 0 + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #undef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 0 + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS 1 + #endif + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 0 + #undef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 1 + #ifndef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 0 + #endif + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 0 + #endif + #undef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 +#elif defined(Py_GIL_DISABLED) || defined(Py_NOGIL) + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 0 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 1 + #ifndef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 1 + #endif + #ifndef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #endif + #undef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 0 + #ifndef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #ifndef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 0 + #endif + #undef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 0 + #ifndef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 1 + #endif + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #ifndef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 0 + #endif + #ifndef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 1 + #endif + #ifndef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 1 + #endif + #undef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 0 + #undef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL 0 + #ifndef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL 1 + #endif + #undef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 0 + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS 1 + #endif + #ifndef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #endif + #ifndef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #endif + #ifndef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 1 + #endif + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 1 + #endif + #ifndef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 0 + #endif +#else + #define CYTHON_COMPILING_IN_PYPY 0 + #define CYTHON_COMPILING_IN_CPYTHON 1 + #define CYTHON_COMPILING_IN_LIMITED_API 0 + #define CYTHON_COMPILING_IN_GRAAL 0 + #define CYTHON_COMPILING_IN_NOGIL 0 + #ifndef CYTHON_USE_TYPE_SLOTS + #define CYTHON_USE_TYPE_SLOTS 1 + #endif + #ifndef CYTHON_USE_TYPE_SPECS + #define CYTHON_USE_TYPE_SPECS 0 + #endif + #ifndef CYTHON_USE_PYTYPE_LOOKUP + #define CYTHON_USE_PYTYPE_LOOKUP 1 + #endif + #if PY_MAJOR_VERSION < 3 + #undef CYTHON_USE_ASYNC_SLOTS + #define CYTHON_USE_ASYNC_SLOTS 0 + #elif !defined(CYTHON_USE_ASYNC_SLOTS) + #define CYTHON_USE_ASYNC_SLOTS 1 + #endif + #ifndef CYTHON_USE_PYLONG_INTERNALS + #define CYTHON_USE_PYLONG_INTERNALS 1 + #endif + #ifndef CYTHON_USE_PYLIST_INTERNALS + #define CYTHON_USE_PYLIST_INTERNALS 1 + #endif + #ifndef CYTHON_USE_UNICODE_INTERNALS + #define CYTHON_USE_UNICODE_INTERNALS 1 + #endif + #if PY_VERSION_HEX < 0x030300F0 || PY_VERSION_HEX >= 0x030B00A2 + #undef CYTHON_USE_UNICODE_WRITER + #define CYTHON_USE_UNICODE_WRITER 0 + #elif !defined(CYTHON_USE_UNICODE_WRITER) + #define CYTHON_USE_UNICODE_WRITER 1 + #endif + #ifndef CYTHON_AVOID_BORROWED_REFS + #define CYTHON_AVOID_BORROWED_REFS 0 + #endif + #ifndef CYTHON_ASSUME_SAFE_MACROS + #define CYTHON_ASSUME_SAFE_MACROS 1 + #endif + #ifndef CYTHON_UNPACK_METHODS + #define CYTHON_UNPACK_METHODS 1 + #endif + #ifndef CYTHON_FAST_THREAD_STATE + #define CYTHON_FAST_THREAD_STATE 1 + #endif + #ifndef CYTHON_FAST_GIL + #define CYTHON_FAST_GIL (PY_MAJOR_VERSION < 3 || PY_VERSION_HEX >= 0x03060000 && PY_VERSION_HEX < 0x030C00A6) + #endif + #ifndef CYTHON_METH_FASTCALL + #define CYTHON_METH_FASTCALL (PY_VERSION_HEX >= 0x030700A1) + #endif + #ifndef CYTHON_FAST_PYCALL + #define CYTHON_FAST_PYCALL 1 + #endif + #ifndef CYTHON_PEP487_INIT_SUBCLASS + #define CYTHON_PEP487_INIT_SUBCLASS 1 + #endif + #if PY_VERSION_HEX < 0x03050000 + #undef CYTHON_PEP489_MULTI_PHASE_INIT + #define CYTHON_PEP489_MULTI_PHASE_INIT 0 + #elif !defined(CYTHON_PEP489_MULTI_PHASE_INIT) + #define CYTHON_PEP489_MULTI_PHASE_INIT 1 + #endif + #ifndef CYTHON_USE_MODULE_STATE + #define CYTHON_USE_MODULE_STATE 0 + #endif + #if PY_VERSION_HEX < 0x030400a1 + #undef CYTHON_USE_TP_FINALIZE + #define CYTHON_USE_TP_FINALIZE 0 + #elif !defined(CYTHON_USE_TP_FINALIZE) + #define CYTHON_USE_TP_FINALIZE 1 + #endif + #if PY_VERSION_HEX < 0x030600B1 + #undef CYTHON_USE_DICT_VERSIONS + #define CYTHON_USE_DICT_VERSIONS 0 + #elif !defined(CYTHON_USE_DICT_VERSIONS) + #define CYTHON_USE_DICT_VERSIONS (PY_VERSION_HEX < 0x030C00A5) + #endif + #if PY_VERSION_HEX < 0x030700A3 + #undef CYTHON_USE_EXC_INFO_STACK + #define CYTHON_USE_EXC_INFO_STACK 0 + #elif !defined(CYTHON_USE_EXC_INFO_STACK) + #define CYTHON_USE_EXC_INFO_STACK 1 + #endif + #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC + #define CYTHON_UPDATE_DESCRIPTOR_DOC 1 + #endif + #ifndef CYTHON_USE_FREELISTS + #define CYTHON_USE_FREELISTS 1 + #endif +#endif +#if !defined(CYTHON_FAST_PYCCALL) +#define CYTHON_FAST_PYCCALL (CYTHON_FAST_PYCALL && PY_VERSION_HEX >= 0x030600B1) +#endif +#if !defined(CYTHON_VECTORCALL) +#define CYTHON_VECTORCALL (CYTHON_FAST_PYCCALL && PY_VERSION_HEX >= 0x030800B1) +#endif +#define CYTHON_BACKPORT_VECTORCALL (CYTHON_METH_FASTCALL && PY_VERSION_HEX < 0x030800B1) +#if CYTHON_USE_PYLONG_INTERNALS + #if PY_MAJOR_VERSION < 3 + #include "longintrepr.h" + #endif + #undef SHIFT + #undef BASE + #undef MASK + #ifdef SIZEOF_VOID_P + enum { __pyx_check_sizeof_voidp = 1 / (int)(SIZEOF_VOID_P == sizeof(void*)) }; + #endif +#endif +#ifndef __has_attribute + #define __has_attribute(x) 0 +#endif +#ifndef __has_cpp_attribute + #define __has_cpp_attribute(x) 0 +#endif +#ifndef CYTHON_RESTRICT + #if defined(__GNUC__) + #define CYTHON_RESTRICT __restrict__ + #elif defined(_MSC_VER) && _MSC_VER >= 1400 + #define CYTHON_RESTRICT __restrict + #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L + #define CYTHON_RESTRICT restrict + #else + #define CYTHON_RESTRICT + #endif +#endif +#ifndef CYTHON_UNUSED + #if defined(__cplusplus) + /* for clang __has_cpp_attribute(maybe_unused) is true even before C++17 + * but leads to warnings with -pedantic, since it is a C++17 feature */ + #if ((defined(_MSVC_LANG) && _MSVC_LANG >= 201703L) || __cplusplus >= 201703L) + #if __has_cpp_attribute(maybe_unused) + #define CYTHON_UNUSED [[maybe_unused]] + #endif + #endif + #endif +#endif +#ifndef CYTHON_UNUSED +# if defined(__GNUC__) +# if !(defined(__cplusplus)) || (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4)) +# define CYTHON_UNUSED __attribute__ ((__unused__)) +# else +# define CYTHON_UNUSED +# endif +# elif defined(__ICC) || (defined(__INTEL_COMPILER) && !defined(_MSC_VER)) +# define CYTHON_UNUSED __attribute__ ((__unused__)) +# else +# define CYTHON_UNUSED +# endif +#endif +#ifndef CYTHON_UNUSED_VAR +# if defined(__cplusplus) + template void CYTHON_UNUSED_VAR( const T& ) { } +# else +# define CYTHON_UNUSED_VAR(x) (void)(x) +# endif +#endif +#ifndef CYTHON_MAYBE_UNUSED_VAR + #define CYTHON_MAYBE_UNUSED_VAR(x) CYTHON_UNUSED_VAR(x) +#endif +#ifndef CYTHON_NCP_UNUSED +# if CYTHON_COMPILING_IN_CPYTHON +# define CYTHON_NCP_UNUSED +# else +# define CYTHON_NCP_UNUSED CYTHON_UNUSED +# endif +#endif +#ifndef CYTHON_USE_CPP_STD_MOVE + #if defined(__cplusplus) && (\ + __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1600)) + #define CYTHON_USE_CPP_STD_MOVE 1 + #else + #define CYTHON_USE_CPP_STD_MOVE 0 + #endif +#endif +#define __Pyx_void_to_None(void_result) ((void)(void_result), Py_INCREF(Py_None), Py_None) +#ifdef _MSC_VER + #ifndef _MSC_STDINT_H_ + #if _MSC_VER < 1300 + typedef unsigned char uint8_t; + typedef unsigned short uint16_t; + typedef unsigned int uint32_t; + #else + typedef unsigned __int8 uint8_t; + typedef unsigned __int16 uint16_t; + typedef unsigned __int32 uint32_t; + #endif + #endif + #if _MSC_VER < 1300 + #ifdef _WIN64 + typedef unsigned long long __pyx_uintptr_t; + #else + typedef unsigned int __pyx_uintptr_t; + #endif + #else + #ifdef _WIN64 + typedef unsigned __int64 __pyx_uintptr_t; + #else + typedef unsigned __int32 __pyx_uintptr_t; + #endif + #endif +#else + #include + typedef uintptr_t __pyx_uintptr_t; +#endif +#ifndef CYTHON_FALLTHROUGH + #if defined(__cplusplus) + /* for clang __has_cpp_attribute(fallthrough) is true even before C++17 + * but leads to warnings with -pedantic, since it is a C++17 feature */ + #if ((defined(_MSVC_LANG) && _MSVC_LANG >= 201703L) || __cplusplus >= 201703L) + #if __has_cpp_attribute(fallthrough) + #define CYTHON_FALLTHROUGH [[fallthrough]] + #endif + #endif + #ifndef CYTHON_FALLTHROUGH + #if __has_cpp_attribute(clang::fallthrough) + #define CYTHON_FALLTHROUGH [[clang::fallthrough]] + #elif __has_cpp_attribute(gnu::fallthrough) + #define CYTHON_FALLTHROUGH [[gnu::fallthrough]] + #endif + #endif + #endif + #ifndef CYTHON_FALLTHROUGH + #if __has_attribute(fallthrough) + #define CYTHON_FALLTHROUGH __attribute__((fallthrough)) + #else + #define CYTHON_FALLTHROUGH + #endif + #endif + #if defined(__clang__) && defined(__apple_build_version__) + #if __apple_build_version__ < 7000000 + #undef CYTHON_FALLTHROUGH + #define CYTHON_FALLTHROUGH + #endif + #endif +#endif +#ifdef __cplusplus + template + struct __PYX_IS_UNSIGNED_IMPL {static const bool value = T(0) < T(-1);}; + #define __PYX_IS_UNSIGNED(type) (__PYX_IS_UNSIGNED_IMPL::value) +#else + #define __PYX_IS_UNSIGNED(type) (((type)-1) > 0) +#endif +#if CYTHON_COMPILING_IN_PYPY == 1 + #define __PYX_NEED_TP_PRINT_SLOT (PY_VERSION_HEX >= 0x030800b4 && PY_VERSION_HEX < 0x030A0000) +#else + #define __PYX_NEED_TP_PRINT_SLOT (PY_VERSION_HEX >= 0x030800b4 && PY_VERSION_HEX < 0x03090000) +#endif +#define __PYX_REINTERPRET_FUNCION(func_pointer, other_pointer) ((func_pointer)(void(*)(void))(other_pointer)) + +#ifndef __cplusplus + #error "Cython files generated with the C++ option must be compiled with a C++ compiler." +#endif +#ifndef CYTHON_INLINE + #if defined(__clang__) + #define CYTHON_INLINE __inline__ __attribute__ ((__unused__)) + #else + #define CYTHON_INLINE inline + #endif +#endif +template +void __Pyx_call_destructor(T& x) { + x.~T(); +} +template +class __Pyx_FakeReference { + public: + __Pyx_FakeReference() : ptr(NULL) { } + __Pyx_FakeReference(const T& ref) : ptr(const_cast(&ref)) { } + T *operator->() { return ptr; } + T *operator&() { return ptr; } + operator T&() { return *ptr; } + template bool operator ==(const U& other) const { return *ptr == other; } + template bool operator !=(const U& other) const { return *ptr != other; } + template bool operator==(const __Pyx_FakeReference& other) const { return *ptr == *other.ptr; } + template bool operator!=(const __Pyx_FakeReference& other) const { return *ptr != *other.ptr; } + private: + T *ptr; +}; + +#define __PYX_BUILD_PY_SSIZE_T "n" +#define CYTHON_FORMAT_SSIZE_T "z" +#if PY_MAJOR_VERSION < 3 + #define __Pyx_BUILTIN_MODULE_NAME "__builtin__" + #define __Pyx_DefaultClassType PyClass_Type + #define __Pyx_PyCode_New(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\ + PyCode_New(a+k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos) +#else + #define __Pyx_BUILTIN_MODULE_NAME "builtins" + #define __Pyx_DefaultClassType PyType_Type +#if CYTHON_COMPILING_IN_LIMITED_API + static CYTHON_INLINE PyObject* __Pyx_PyCode_New(int a, int p, int k, int l, int s, int f, + PyObject *code, PyObject *c, PyObject* n, PyObject *v, + PyObject *fv, PyObject *cell, PyObject* fn, + PyObject *name, int fline, PyObject *lnos) { + PyObject *exception_table = NULL; + PyObject *types_module=NULL, *code_type=NULL, *result=NULL; + #if __PYX_LIMITED_VERSION_HEX < 0x030B0000 + PyObject *version_info; + PyObject *py_minor_version = NULL; + #endif + long minor_version = 0; + PyObject *type, *value, *traceback; + PyErr_Fetch(&type, &value, &traceback); + #if __PYX_LIMITED_VERSION_HEX >= 0x030B0000 + minor_version = 11; + #else + if (!(version_info = PySys_GetObject("version_info"))) goto end; + if (!(py_minor_version = PySequence_GetItem(version_info, 1))) goto end; + minor_version = PyLong_AsLong(py_minor_version); + Py_DECREF(py_minor_version); + if (minor_version == -1 && PyErr_Occurred()) goto end; + #endif + if (!(types_module = PyImport_ImportModule("types"))) goto end; + if (!(code_type = PyObject_GetAttrString(types_module, "CodeType"))) goto end; + if (minor_version <= 7) { + (void)p; + result = PyObject_CallFunction(code_type, "iiiiiOOOOOOiOO", a, k, l, s, f, code, + c, n, v, fn, name, fline, lnos, fv, cell); + } else if (minor_version <= 10) { + result = PyObject_CallFunction(code_type, "iiiiiiOOOOOOiOO", a,p, k, l, s, f, code, + c, n, v, fn, name, fline, lnos, fv, cell); + } else { + if (!(exception_table = PyBytes_FromStringAndSize(NULL, 0))) goto end; + result = PyObject_CallFunction(code_type, "iiiiiiOOOOOOOiOO", a,p, k, l, s, f, code, + c, n, v, fn, name, name, fline, lnos, exception_table, fv, cell); + } + end: + Py_XDECREF(code_type); + Py_XDECREF(exception_table); + Py_XDECREF(types_module); + if (type) { + PyErr_Restore(type, value, traceback); + } + return result; + } + #ifndef CO_OPTIMIZED + #define CO_OPTIMIZED 0x0001 + #endif + #ifndef CO_NEWLOCALS + #define CO_NEWLOCALS 0x0002 + #endif + #ifndef CO_VARARGS + #define CO_VARARGS 0x0004 + #endif + #ifndef CO_VARKEYWORDS + #define CO_VARKEYWORDS 0x0008 + #endif + #ifndef CO_ASYNC_GENERATOR + #define CO_ASYNC_GENERATOR 0x0200 + #endif + #ifndef CO_GENERATOR + #define CO_GENERATOR 0x0020 + #endif + #ifndef CO_COROUTINE + #define CO_COROUTINE 0x0080 + #endif +#elif PY_VERSION_HEX >= 0x030B0000 + static CYTHON_INLINE PyCodeObject* __Pyx_PyCode_New(int a, int p, int k, int l, int s, int f, + PyObject *code, PyObject *c, PyObject* n, PyObject *v, + PyObject *fv, PyObject *cell, PyObject* fn, + PyObject *name, int fline, PyObject *lnos) { + PyCodeObject *result; + PyObject *empty_bytes = PyBytes_FromStringAndSize("", 0); + if (!empty_bytes) return NULL; + result = + #if PY_VERSION_HEX >= 0x030C0000 + PyUnstable_Code_NewWithPosOnlyArgs + #else + PyCode_NewWithPosOnlyArgs + #endif + (a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, name, fline, lnos, empty_bytes); + Py_DECREF(empty_bytes); + return result; + } +#elif PY_VERSION_HEX >= 0x030800B2 && !CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyCode_New(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\ + PyCode_NewWithPosOnlyArgs(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos) +#else + #define __Pyx_PyCode_New(a, p, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos)\ + PyCode_New(a, k, l, s, f, code, c, n, v, fv, cell, fn, name, fline, lnos) +#endif +#endif +#if PY_VERSION_HEX >= 0x030900A4 || defined(Py_IS_TYPE) + #define __Pyx_IS_TYPE(ob, type) Py_IS_TYPE(ob, type) +#else + #define __Pyx_IS_TYPE(ob, type) (((const PyObject*)ob)->ob_type == (type)) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_Is) + #define __Pyx_Py_Is(x, y) Py_Is(x, y) +#else + #define __Pyx_Py_Is(x, y) ((x) == (y)) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_IsNone) + #define __Pyx_Py_IsNone(ob) Py_IsNone(ob) +#else + #define __Pyx_Py_IsNone(ob) __Pyx_Py_Is((ob), Py_None) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_IsTrue) + #define __Pyx_Py_IsTrue(ob) Py_IsTrue(ob) +#else + #define __Pyx_Py_IsTrue(ob) __Pyx_Py_Is((ob), Py_True) +#endif +#if PY_VERSION_HEX >= 0x030A00B1 || defined(Py_IsFalse) + #define __Pyx_Py_IsFalse(ob) Py_IsFalse(ob) +#else + #define __Pyx_Py_IsFalse(ob) __Pyx_Py_Is((ob), Py_False) +#endif +#define __Pyx_NoneAsNull(obj) (__Pyx_Py_IsNone(obj) ? NULL : (obj)) +#if PY_VERSION_HEX >= 0x030900F0 && !CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyObject_GC_IsFinalized(o) PyObject_GC_IsFinalized(o) +#else + #define __Pyx_PyObject_GC_IsFinalized(o) _PyGC_FINALIZED(o) +#endif +#ifndef CO_COROUTINE + #define CO_COROUTINE 0x80 +#endif +#ifndef CO_ASYNC_GENERATOR + #define CO_ASYNC_GENERATOR 0x200 +#endif +#ifndef Py_TPFLAGS_CHECKTYPES + #define Py_TPFLAGS_CHECKTYPES 0 +#endif +#ifndef Py_TPFLAGS_HAVE_INDEX + #define Py_TPFLAGS_HAVE_INDEX 0 +#endif +#ifndef Py_TPFLAGS_HAVE_NEWBUFFER + #define Py_TPFLAGS_HAVE_NEWBUFFER 0 +#endif +#ifndef Py_TPFLAGS_HAVE_FINALIZE + #define Py_TPFLAGS_HAVE_FINALIZE 0 +#endif +#ifndef Py_TPFLAGS_SEQUENCE + #define Py_TPFLAGS_SEQUENCE 0 +#endif +#ifndef Py_TPFLAGS_MAPPING + #define Py_TPFLAGS_MAPPING 0 +#endif +#ifndef METH_STACKLESS + #define METH_STACKLESS 0 +#endif +#if PY_VERSION_HEX <= 0x030700A3 || !defined(METH_FASTCALL) + #ifndef METH_FASTCALL + #define METH_FASTCALL 0x80 + #endif + typedef PyObject *(*__Pyx_PyCFunctionFast) (PyObject *self, PyObject *const *args, Py_ssize_t nargs); + typedef PyObject *(*__Pyx_PyCFunctionFastWithKeywords) (PyObject *self, PyObject *const *args, + Py_ssize_t nargs, PyObject *kwnames); +#else + #if PY_VERSION_HEX >= 0x030d00A4 + # define __Pyx_PyCFunctionFast PyCFunctionFast + # define __Pyx_PyCFunctionFastWithKeywords PyCFunctionFastWithKeywords + #else + # define __Pyx_PyCFunctionFast _PyCFunctionFast + # define __Pyx_PyCFunctionFastWithKeywords _PyCFunctionFastWithKeywords + #endif +#endif +#if CYTHON_METH_FASTCALL + #define __Pyx_METH_FASTCALL METH_FASTCALL + #define __Pyx_PyCFunction_FastCall __Pyx_PyCFunctionFast + #define __Pyx_PyCFunction_FastCallWithKeywords __Pyx_PyCFunctionFastWithKeywords +#else + #define __Pyx_METH_FASTCALL METH_VARARGS + #define __Pyx_PyCFunction_FastCall PyCFunction + #define __Pyx_PyCFunction_FastCallWithKeywords PyCFunctionWithKeywords +#endif +#if CYTHON_VECTORCALL + #define __pyx_vectorcallfunc vectorcallfunc + #define __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET PY_VECTORCALL_ARGUMENTS_OFFSET + #define __Pyx_PyVectorcall_NARGS(n) PyVectorcall_NARGS((size_t)(n)) +#elif CYTHON_BACKPORT_VECTORCALL + typedef PyObject *(*__pyx_vectorcallfunc)(PyObject *callable, PyObject *const *args, + size_t nargsf, PyObject *kwnames); + #define __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET ((size_t)1 << (8 * sizeof(size_t) - 1)) + #define __Pyx_PyVectorcall_NARGS(n) ((Py_ssize_t)(((size_t)(n)) & ~__Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET)) +#else + #define __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET 0 + #define __Pyx_PyVectorcall_NARGS(n) ((Py_ssize_t)(n)) +#endif +#if PY_MAJOR_VERSION >= 0x030900B1 +#define __Pyx_PyCFunction_CheckExact(func) PyCFunction_CheckExact(func) +#else +#define __Pyx_PyCFunction_CheckExact(func) PyCFunction_Check(func) +#endif +#define __Pyx_CyOrPyCFunction_Check(func) PyCFunction_Check(func) +#if CYTHON_COMPILING_IN_CPYTHON +#define __Pyx_CyOrPyCFunction_GET_FUNCTION(func) (((PyCFunctionObject*)(func))->m_ml->ml_meth) +#elif !CYTHON_COMPILING_IN_LIMITED_API +#define __Pyx_CyOrPyCFunction_GET_FUNCTION(func) PyCFunction_GET_FUNCTION(func) +#endif +#if CYTHON_COMPILING_IN_CPYTHON +#define __Pyx_CyOrPyCFunction_GET_FLAGS(func) (((PyCFunctionObject*)(func))->m_ml->ml_flags) +static CYTHON_INLINE PyObject* __Pyx_CyOrPyCFunction_GET_SELF(PyObject *func) { + return (__Pyx_CyOrPyCFunction_GET_FLAGS(func) & METH_STATIC) ? NULL : ((PyCFunctionObject*)func)->m_self; +} +#endif +static CYTHON_INLINE int __Pyx__IsSameCFunction(PyObject *func, void *cfunc) { +#if CYTHON_COMPILING_IN_LIMITED_API + return PyCFunction_Check(func) && PyCFunction_GetFunction(func) == (PyCFunction) cfunc; +#else + return PyCFunction_Check(func) && PyCFunction_GET_FUNCTION(func) == (PyCFunction) cfunc; +#endif +} +#define __Pyx_IsSameCFunction(func, cfunc) __Pyx__IsSameCFunction(func, cfunc) +#if __PYX_LIMITED_VERSION_HEX < 0x030900B1 + #define __Pyx_PyType_FromModuleAndSpec(m, s, b) ((void)m, PyType_FromSpecWithBases(s, b)) + typedef PyObject *(*__Pyx_PyCMethod)(PyObject *, PyTypeObject *, PyObject *const *, size_t, PyObject *); +#else + #define __Pyx_PyType_FromModuleAndSpec(m, s, b) PyType_FromModuleAndSpec(m, s, b) + #define __Pyx_PyCMethod PyCMethod +#endif +#ifndef METH_METHOD + #define METH_METHOD 0x200 +#endif +#if CYTHON_COMPILING_IN_PYPY && !defined(PyObject_Malloc) + #define PyObject_Malloc(s) PyMem_Malloc(s) + #define PyObject_Free(p) PyMem_Free(p) + #define PyObject_Realloc(p) PyMem_Realloc(p) +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_PyCode_HasFreeVars(co) (PyCode_GetNumFree(co) > 0) + #define __Pyx_PyFrame_SetLineNumber(frame, lineno) +#else + #define __Pyx_PyCode_HasFreeVars(co) (PyCode_GetNumFree(co) > 0) + #define __Pyx_PyFrame_SetLineNumber(frame, lineno) (frame)->f_lineno = (lineno) +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_PyThreadState_Current PyThreadState_Get() +#elif !CYTHON_FAST_THREAD_STATE + #define __Pyx_PyThreadState_Current PyThreadState_GET() +#elif PY_VERSION_HEX >= 0x030d00A1 + #define __Pyx_PyThreadState_Current PyThreadState_GetUnchecked() +#elif PY_VERSION_HEX >= 0x03060000 + #define __Pyx_PyThreadState_Current _PyThreadState_UncheckedGet() +#elif PY_VERSION_HEX >= 0x03000000 + #define __Pyx_PyThreadState_Current PyThreadState_GET() +#else + #define __Pyx_PyThreadState_Current _PyThreadState_Current +#endif +#if CYTHON_COMPILING_IN_LIMITED_API +static CYTHON_INLINE void *__Pyx_PyModule_GetState(PyObject *op) +{ + void *result; + result = PyModule_GetState(op); + if (!result) + Py_FatalError("Couldn't find the module state"); + return result; +} +#endif +#define __Pyx_PyObject_GetSlot(obj, name, func_ctype) __Pyx_PyType_GetSlot(Py_TYPE(obj), name, func_ctype) +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_PyType_GetSlot(type, name, func_ctype) ((func_ctype) PyType_GetSlot((type), Py_##name)) +#else + #define __Pyx_PyType_GetSlot(type, name, func_ctype) ((type)->name) +#endif +#if PY_VERSION_HEX < 0x030700A2 && !defined(PyThread_tss_create) && !defined(Py_tss_NEEDS_INIT) +#include "pythread.h" +#define Py_tss_NEEDS_INIT 0 +typedef int Py_tss_t; +static CYTHON_INLINE int PyThread_tss_create(Py_tss_t *key) { + *key = PyThread_create_key(); + return 0; +} +static CYTHON_INLINE Py_tss_t * PyThread_tss_alloc(void) { + Py_tss_t *key = (Py_tss_t *)PyObject_Malloc(sizeof(Py_tss_t)); + *key = Py_tss_NEEDS_INIT; + return key; +} +static CYTHON_INLINE void PyThread_tss_free(Py_tss_t *key) { + PyObject_Free(key); +} +static CYTHON_INLINE int PyThread_tss_is_created(Py_tss_t *key) { + return *key != Py_tss_NEEDS_INIT; +} +static CYTHON_INLINE void PyThread_tss_delete(Py_tss_t *key) { + PyThread_delete_key(*key); + *key = Py_tss_NEEDS_INIT; +} +static CYTHON_INLINE int PyThread_tss_set(Py_tss_t *key, void *value) { + return PyThread_set_key_value(*key, value); +} +static CYTHON_INLINE void * PyThread_tss_get(Py_tss_t *key) { + return PyThread_get_key_value(*key); +} +#endif +#if PY_MAJOR_VERSION < 3 + #if CYTHON_COMPILING_IN_PYPY + #if PYPY_VERSION_NUM < 0x07030600 + #if defined(__cplusplus) && __cplusplus >= 201402L + [[deprecated("`with nogil:` inside a nogil function will not release the GIL in PyPy2 < 7.3.6")]] + #elif defined(__GNUC__) || defined(__clang__) + __attribute__ ((__deprecated__("`with nogil:` inside a nogil function will not release the GIL in PyPy2 < 7.3.6"))) + #elif defined(_MSC_VER) + __declspec(deprecated("`with nogil:` inside a nogil function will not release the GIL in PyPy2 < 7.3.6")) + #endif + static CYTHON_INLINE int PyGILState_Check(void) { + return 0; + } + #else // PYPY_VERSION_NUM < 0x07030600 + #endif // PYPY_VERSION_NUM < 0x07030600 + #else + static CYTHON_INLINE int PyGILState_Check(void) { + PyThreadState * tstate = _PyThreadState_Current; + return tstate && (tstate == PyGILState_GetThisThreadState()); + } + #endif +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030d0000 || defined(_PyDict_NewPresized) +#define __Pyx_PyDict_NewPresized(n) ((n <= 8) ? PyDict_New() : _PyDict_NewPresized(n)) +#else +#define __Pyx_PyDict_NewPresized(n) PyDict_New() +#endif +#if PY_MAJOR_VERSION >= 3 || CYTHON_FUTURE_DIVISION + #define __Pyx_PyNumber_Divide(x,y) PyNumber_TrueDivide(x,y) + #define __Pyx_PyNumber_InPlaceDivide(x,y) PyNumber_InPlaceTrueDivide(x,y) +#else + #define __Pyx_PyNumber_Divide(x,y) PyNumber_Divide(x,y) + #define __Pyx_PyNumber_InPlaceDivide(x,y) PyNumber_InPlaceDivide(x,y) +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX > 0x030600B4 && PY_VERSION_HEX < 0x030d0000 && CYTHON_USE_UNICODE_INTERNALS +#define __Pyx_PyDict_GetItemStrWithError(dict, name) _PyDict_GetItem_KnownHash(dict, name, ((PyASCIIObject *) name)->hash) +static CYTHON_INLINE PyObject * __Pyx_PyDict_GetItemStr(PyObject *dict, PyObject *name) { + PyObject *res = __Pyx_PyDict_GetItemStrWithError(dict, name); + if (res == NULL) PyErr_Clear(); + return res; +} +#elif PY_MAJOR_VERSION >= 3 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07020000) +#define __Pyx_PyDict_GetItemStrWithError PyDict_GetItemWithError +#define __Pyx_PyDict_GetItemStr PyDict_GetItem +#else +static CYTHON_INLINE PyObject * __Pyx_PyDict_GetItemStrWithError(PyObject *dict, PyObject *name) { +#if CYTHON_COMPILING_IN_PYPY + return PyDict_GetItem(dict, name); +#else + PyDictEntry *ep; + PyDictObject *mp = (PyDictObject*) dict; + long hash = ((PyStringObject *) name)->ob_shash; + assert(hash != -1); + ep = (mp->ma_lookup)(mp, name, hash); + if (ep == NULL) { + return NULL; + } + return ep->me_value; +#endif +} +#define __Pyx_PyDict_GetItemStr PyDict_GetItem +#endif +#if CYTHON_USE_TYPE_SLOTS + #define __Pyx_PyType_GetFlags(tp) (((PyTypeObject *)tp)->tp_flags) + #define __Pyx_PyType_HasFeature(type, feature) ((__Pyx_PyType_GetFlags(type) & (feature)) != 0) + #define __Pyx_PyObject_GetIterNextFunc(obj) (Py_TYPE(obj)->tp_iternext) +#else + #define __Pyx_PyType_GetFlags(tp) (PyType_GetFlags((PyTypeObject *)tp)) + #define __Pyx_PyType_HasFeature(type, feature) PyType_HasFeature(type, feature) + #define __Pyx_PyObject_GetIterNextFunc(obj) PyIter_Next +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_SetItemOnTypeDict(tp, k, v) PyObject_GenericSetAttr((PyObject*)tp, k, v) +#else + #define __Pyx_SetItemOnTypeDict(tp, k, v) PyDict_SetItem(tp->tp_dict, k, v) +#endif +#if CYTHON_USE_TYPE_SPECS && PY_VERSION_HEX >= 0x03080000 +#define __Pyx_PyHeapTypeObject_GC_Del(obj) {\ + PyTypeObject *type = Py_TYPE((PyObject*)obj);\ + assert(__Pyx_PyType_HasFeature(type, Py_TPFLAGS_HEAPTYPE));\ + PyObject_GC_Del(obj);\ + Py_DECREF(type);\ +} +#else +#define __Pyx_PyHeapTypeObject_GC_Del(obj) PyObject_GC_Del(obj) +#endif +#if CYTHON_COMPILING_IN_LIMITED_API + #define CYTHON_PEP393_ENABLED 1 + #define __Pyx_PyUnicode_READY(op) (0) + #define __Pyx_PyUnicode_GET_LENGTH(u) PyUnicode_GetLength(u) + #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_ReadChar(u, i) + #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u) ((void)u, 1114111U) + #define __Pyx_PyUnicode_KIND(u) ((void)u, (0)) + #define __Pyx_PyUnicode_DATA(u) ((void*)u) + #define __Pyx_PyUnicode_READ(k, d, i) ((void)k, PyUnicode_ReadChar((PyObject*)(d), i)) + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != PyUnicode_GetLength(u)) +#elif PY_VERSION_HEX > 0x03030000 && defined(PyUnicode_KIND) + #define CYTHON_PEP393_ENABLED 1 + #if PY_VERSION_HEX >= 0x030C0000 + #define __Pyx_PyUnicode_READY(op) (0) + #else + #define __Pyx_PyUnicode_READY(op) (likely(PyUnicode_IS_READY(op)) ?\ + 0 : _PyUnicode_Ready((PyObject *)(op))) + #endif + #define __Pyx_PyUnicode_GET_LENGTH(u) PyUnicode_GET_LENGTH(u) + #define __Pyx_PyUnicode_READ_CHAR(u, i) PyUnicode_READ_CHAR(u, i) + #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u) PyUnicode_MAX_CHAR_VALUE(u) + #define __Pyx_PyUnicode_KIND(u) ((int)PyUnicode_KIND(u)) + #define __Pyx_PyUnicode_DATA(u) PyUnicode_DATA(u) + #define __Pyx_PyUnicode_READ(k, d, i) PyUnicode_READ(k, d, i) + #define __Pyx_PyUnicode_WRITE(k, d, i, ch) PyUnicode_WRITE(k, d, i, (Py_UCS4) ch) + #if PY_VERSION_HEX >= 0x030C0000 + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != PyUnicode_GET_LENGTH(u)) + #else + #if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x03090000 + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : ((PyCompactUnicodeObject *)(u))->wstr_length)) + #else + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != (likely(PyUnicode_IS_READY(u)) ? PyUnicode_GET_LENGTH(u) : PyUnicode_GET_SIZE(u))) + #endif + #endif +#else + #define CYTHON_PEP393_ENABLED 0 + #define PyUnicode_1BYTE_KIND 1 + #define PyUnicode_2BYTE_KIND 2 + #define PyUnicode_4BYTE_KIND 4 + #define __Pyx_PyUnicode_READY(op) (0) + #define __Pyx_PyUnicode_GET_LENGTH(u) PyUnicode_GET_SIZE(u) + #define __Pyx_PyUnicode_READ_CHAR(u, i) ((Py_UCS4)(PyUnicode_AS_UNICODE(u)[i])) + #define __Pyx_PyUnicode_MAX_CHAR_VALUE(u) ((sizeof(Py_UNICODE) == 2) ? 65535U : 1114111U) + #define __Pyx_PyUnicode_KIND(u) ((int)sizeof(Py_UNICODE)) + #define __Pyx_PyUnicode_DATA(u) ((void*)PyUnicode_AS_UNICODE(u)) + #define __Pyx_PyUnicode_READ(k, d, i) ((void)(k), (Py_UCS4)(((Py_UNICODE*)d)[i])) + #define __Pyx_PyUnicode_WRITE(k, d, i, ch) (((void)(k)), ((Py_UNICODE*)d)[i] = (Py_UNICODE) ch) + #define __Pyx_PyUnicode_IS_TRUE(u) (0 != PyUnicode_GET_SIZE(u)) +#endif +#if CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyUnicode_Concat(a, b) PyNumber_Add(a, b) + #define __Pyx_PyUnicode_ConcatSafe(a, b) PyNumber_Add(a, b) +#else + #define __Pyx_PyUnicode_Concat(a, b) PyUnicode_Concat(a, b) + #define __Pyx_PyUnicode_ConcatSafe(a, b) ((unlikely((a) == Py_None) || unlikely((b) == Py_None)) ?\ + PyNumber_Add(a, b) : __Pyx_PyUnicode_Concat(a, b)) +#endif +#if CYTHON_COMPILING_IN_PYPY + #if !defined(PyUnicode_DecodeUnicodeEscape) + #define PyUnicode_DecodeUnicodeEscape(s, size, errors) PyUnicode_Decode(s, size, "unicode_escape", errors) + #endif + #if !defined(PyUnicode_Contains) || (PY_MAJOR_VERSION == 2 && PYPY_VERSION_NUM < 0x07030500) + #undef PyUnicode_Contains + #define PyUnicode_Contains(u, s) PySequence_Contains(u, s) + #endif + #if !defined(PyByteArray_Check) + #define PyByteArray_Check(obj) PyObject_TypeCheck(obj, &PyByteArray_Type) + #endif + #if !defined(PyObject_Format) + #define PyObject_Format(obj, fmt) PyObject_CallMethod(obj, "__format__", "O", fmt) + #endif +#endif +#define __Pyx_PyString_FormatSafe(a, b) ((unlikely((a) == Py_None || (PyString_Check(b) && !PyString_CheckExact(b)))) ? PyNumber_Remainder(a, b) : __Pyx_PyString_Format(a, b)) +#define __Pyx_PyUnicode_FormatSafe(a, b) ((unlikely((a) == Py_None || (PyUnicode_Check(b) && !PyUnicode_CheckExact(b)))) ? PyNumber_Remainder(a, b) : PyUnicode_Format(a, b)) +#if PY_MAJOR_VERSION >= 3 + #define __Pyx_PyString_Format(a, b) PyUnicode_Format(a, b) +#else + #define __Pyx_PyString_Format(a, b) PyString_Format(a, b) +#endif +#if PY_MAJOR_VERSION < 3 && !defined(PyObject_ASCII) + #define PyObject_ASCII(o) PyObject_Repr(o) +#endif +#if PY_MAJOR_VERSION >= 3 + #define PyBaseString_Type PyUnicode_Type + #define PyStringObject PyUnicodeObject + #define PyString_Type PyUnicode_Type + #define PyString_Check PyUnicode_Check + #define PyString_CheckExact PyUnicode_CheckExact +#ifndef PyObject_Unicode + #define PyObject_Unicode PyObject_Str +#endif +#endif +#if PY_MAJOR_VERSION >= 3 + #define __Pyx_PyBaseString_Check(obj) PyUnicode_Check(obj) + #define __Pyx_PyBaseString_CheckExact(obj) PyUnicode_CheckExact(obj) +#else + #define __Pyx_PyBaseString_Check(obj) (PyString_Check(obj) || PyUnicode_Check(obj)) + #define __Pyx_PyBaseString_CheckExact(obj) (PyString_CheckExact(obj) || PyUnicode_CheckExact(obj)) +#endif +#if CYTHON_COMPILING_IN_CPYTHON + #define __Pyx_PySequence_ListKeepNew(obj)\ + (likely(PyList_CheckExact(obj) && Py_REFCNT(obj) == 1) ? __Pyx_NewRef(obj) : PySequence_List(obj)) +#else + #define __Pyx_PySequence_ListKeepNew(obj) PySequence_List(obj) +#endif +#ifndef PySet_CheckExact + #define PySet_CheckExact(obj) __Pyx_IS_TYPE(obj, &PySet_Type) +#endif +#if PY_VERSION_HEX >= 0x030900A4 + #define __Pyx_SET_REFCNT(obj, refcnt) Py_SET_REFCNT(obj, refcnt) + #define __Pyx_SET_SIZE(obj, size) Py_SET_SIZE(obj, size) +#else + #define __Pyx_SET_REFCNT(obj, refcnt) Py_REFCNT(obj) = (refcnt) + #define __Pyx_SET_SIZE(obj, size) Py_SIZE(obj) = (size) +#endif +#if CYTHON_ASSUME_SAFE_MACROS + #define __Pyx_PySequence_ITEM(o, i) PySequence_ITEM(o, i) + #define __Pyx_PySequence_SIZE(seq) Py_SIZE(seq) + #define __Pyx_PyTuple_SET_ITEM(o, i, v) (PyTuple_SET_ITEM(o, i, v), (0)) + #define __Pyx_PyList_SET_ITEM(o, i, v) (PyList_SET_ITEM(o, i, v), (0)) + #define __Pyx_PyTuple_GET_SIZE(o) PyTuple_GET_SIZE(o) + #define __Pyx_PyList_GET_SIZE(o) PyList_GET_SIZE(o) + #define __Pyx_PySet_GET_SIZE(o) PySet_GET_SIZE(o) + #define __Pyx_PyBytes_GET_SIZE(o) PyBytes_GET_SIZE(o) + #define __Pyx_PyByteArray_GET_SIZE(o) PyByteArray_GET_SIZE(o) +#else + #define __Pyx_PySequence_ITEM(o, i) PySequence_GetItem(o, i) + #define __Pyx_PySequence_SIZE(seq) PySequence_Size(seq) + #define __Pyx_PyTuple_SET_ITEM(o, i, v) PyTuple_SetItem(o, i, v) + #define __Pyx_PyList_SET_ITEM(o, i, v) PyList_SetItem(o, i, v) + #define __Pyx_PyTuple_GET_SIZE(o) PyTuple_Size(o) + #define __Pyx_PyList_GET_SIZE(o) PyList_Size(o) + #define __Pyx_PySet_GET_SIZE(o) PySet_Size(o) + #define __Pyx_PyBytes_GET_SIZE(o) PyBytes_Size(o) + #define __Pyx_PyByteArray_GET_SIZE(o) PyByteArray_Size(o) +#endif +#if __PYX_LIMITED_VERSION_HEX >= 0x030d00A1 + #define __Pyx_PyImport_AddModuleRef(name) PyImport_AddModuleRef(name) +#else + static CYTHON_INLINE PyObject *__Pyx_PyImport_AddModuleRef(const char *name) { + PyObject *module = PyImport_AddModule(name); + Py_XINCREF(module); + return module; + } +#endif +#if PY_MAJOR_VERSION >= 3 + #define PyIntObject PyLongObject + #define PyInt_Type PyLong_Type + #define PyInt_Check(op) PyLong_Check(op) + #define PyInt_CheckExact(op) PyLong_CheckExact(op) + #define __Pyx_Py3Int_Check(op) PyLong_Check(op) + #define __Pyx_Py3Int_CheckExact(op) PyLong_CheckExact(op) + #define PyInt_FromString PyLong_FromString + #define PyInt_FromUnicode PyLong_FromUnicode + #define PyInt_FromLong PyLong_FromLong + #define PyInt_FromSize_t PyLong_FromSize_t + #define PyInt_FromSsize_t PyLong_FromSsize_t + #define PyInt_AsLong PyLong_AsLong + #define PyInt_AS_LONG PyLong_AS_LONG + #define PyInt_AsSsize_t PyLong_AsSsize_t + #define PyInt_AsUnsignedLongMask PyLong_AsUnsignedLongMask + #define PyInt_AsUnsignedLongLongMask PyLong_AsUnsignedLongLongMask + #define PyNumber_Int PyNumber_Long +#else + #define __Pyx_Py3Int_Check(op) (PyLong_Check(op) || PyInt_Check(op)) + #define __Pyx_Py3Int_CheckExact(op) (PyLong_CheckExact(op) || PyInt_CheckExact(op)) +#endif +#if PY_MAJOR_VERSION >= 3 + #define PyBoolObject PyLongObject +#endif +#if PY_MAJOR_VERSION >= 3 && CYTHON_COMPILING_IN_PYPY + #ifndef PyUnicode_InternFromString + #define PyUnicode_InternFromString(s) PyUnicode_FromString(s) + #endif +#endif +#if PY_VERSION_HEX < 0x030200A4 + typedef long Py_hash_t; + #define __Pyx_PyInt_FromHash_t PyInt_FromLong + #define __Pyx_PyInt_AsHash_t __Pyx_PyIndex_AsHash_t +#else + #define __Pyx_PyInt_FromHash_t PyInt_FromSsize_t + #define __Pyx_PyInt_AsHash_t __Pyx_PyIndex_AsSsize_t +#endif +#if CYTHON_USE_ASYNC_SLOTS + #if PY_VERSION_HEX >= 0x030500B1 + #define __Pyx_PyAsyncMethodsStruct PyAsyncMethods + #define __Pyx_PyType_AsAsync(obj) (Py_TYPE(obj)->tp_as_async) + #else + #define __Pyx_PyType_AsAsync(obj) ((__Pyx_PyAsyncMethodsStruct*) (Py_TYPE(obj)->tp_reserved)) + #endif +#else + #define __Pyx_PyType_AsAsync(obj) NULL +#endif +#ifndef __Pyx_PyAsyncMethodsStruct + typedef struct { + unaryfunc am_await; + unaryfunc am_aiter; + unaryfunc am_anext; + } __Pyx_PyAsyncMethodsStruct; +#endif + +#if defined(_WIN32) || defined(WIN32) || defined(MS_WINDOWS) + #if !defined(_USE_MATH_DEFINES) + #define _USE_MATH_DEFINES + #endif +#endif +#include +#ifdef NAN +#define __PYX_NAN() ((float) NAN) +#else +static CYTHON_INLINE float __PYX_NAN() { + float value; + memset(&value, 0xFF, sizeof(value)); + return value; +} +#endif +#if defined(__CYGWIN__) && defined(_LDBL_EQ_DBL) +#define __Pyx_truncl trunc +#else +#define __Pyx_truncl truncl +#endif + +#define __PYX_MARK_ERR_POS(f_index, lineno) \ + { __pyx_filename = __pyx_f[f_index]; (void)__pyx_filename; __pyx_lineno = lineno; (void)__pyx_lineno; __pyx_clineno = __LINE__; (void)__pyx_clineno; } +#define __PYX_ERR(f_index, lineno, Ln_error) \ + { __PYX_MARK_ERR_POS(f_index, lineno) goto Ln_error; } + +#ifdef CYTHON_EXTERN_C + #undef __PYX_EXTERN_C + #define __PYX_EXTERN_C CYTHON_EXTERN_C +#elif defined(__PYX_EXTERN_C) + #ifdef _MSC_VER + #pragma message ("Please do not define the '__PYX_EXTERN_C' macro externally. Use 'CYTHON_EXTERN_C' instead.") + #else + #warning Please do not define the '__PYX_EXTERN_C' macro externally. Use 'CYTHON_EXTERN_C' instead. + #endif +#else + #define __PYX_EXTERN_C extern "C++" +#endif + +#define __PYX_HAVE__fairseq__data__token_block_utils_fast +#define __PYX_HAVE_API__fairseq__data__token_block_utils_fast +/* Early includes */ +#include +#include +#include + + /* Using NumPy API declarations from "numpy/__init__.cython-30.pxd" */ + +#include "numpy/arrayobject.h" +#include "numpy/ndarrayobject.h" +#include "numpy/ndarraytypes.h" +#include "numpy/arrayscalars.h" +#include "numpy/ufuncobject.h" +#include "pythread.h" +#include +#ifdef _OPENMP +#include +#endif /* _OPENMP */ + +#if defined(PYREX_WITHOUT_ASSERTIONS) && !defined(CYTHON_WITHOUT_ASSERTIONS) +#define CYTHON_WITHOUT_ASSERTIONS +#endif + +typedef struct {PyObject **p; const char *s; const Py_ssize_t n; const char* encoding; + const char is_unicode; const char is_str; const char intern; } __Pyx_StringTabEntry; + +#define __PYX_DEFAULT_STRING_ENCODING_IS_ASCII 0 +#define __PYX_DEFAULT_STRING_ENCODING_IS_UTF8 0 +#define __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT (PY_MAJOR_VERSION >= 3 && __PYX_DEFAULT_STRING_ENCODING_IS_UTF8) +#define __PYX_DEFAULT_STRING_ENCODING "" +#define __Pyx_PyObject_FromString __Pyx_PyBytes_FromString +#define __Pyx_PyObject_FromStringAndSize __Pyx_PyBytes_FromStringAndSize +#define __Pyx_uchar_cast(c) ((unsigned char)c) +#define __Pyx_long_cast(x) ((long)x) +#define __Pyx_fits_Py_ssize_t(v, type, is_signed) (\ + (sizeof(type) < sizeof(Py_ssize_t)) ||\ + (sizeof(type) > sizeof(Py_ssize_t) &&\ + likely(v < (type)PY_SSIZE_T_MAX ||\ + v == (type)PY_SSIZE_T_MAX) &&\ + (!is_signed || likely(v > (type)PY_SSIZE_T_MIN ||\ + v == (type)PY_SSIZE_T_MIN))) ||\ + (sizeof(type) == sizeof(Py_ssize_t) &&\ + (is_signed || likely(v < (type)PY_SSIZE_T_MAX ||\ + v == (type)PY_SSIZE_T_MAX))) ) +static CYTHON_INLINE int __Pyx_is_valid_index(Py_ssize_t i, Py_ssize_t limit) { + return (size_t) i < (size_t) limit; +} +#if defined (__cplusplus) && __cplusplus >= 201103L + #include + #define __Pyx_sst_abs(value) std::abs(value) +#elif SIZEOF_INT >= SIZEOF_SIZE_T + #define __Pyx_sst_abs(value) abs(value) +#elif SIZEOF_LONG >= SIZEOF_SIZE_T + #define __Pyx_sst_abs(value) labs(value) +#elif defined (_MSC_VER) + #define __Pyx_sst_abs(value) ((Py_ssize_t)_abs64(value)) +#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L + #define __Pyx_sst_abs(value) llabs(value) +#elif defined (__GNUC__) + #define __Pyx_sst_abs(value) __builtin_llabs(value) +#else + #define __Pyx_sst_abs(value) ((value<0) ? -value : value) +#endif +static CYTHON_INLINE Py_ssize_t __Pyx_ssize_strlen(const char *s); +static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject*); +static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject*, Py_ssize_t* length); +static CYTHON_INLINE PyObject* __Pyx_PyByteArray_FromString(const char*); +#define __Pyx_PyByteArray_FromStringAndSize(s, l) PyByteArray_FromStringAndSize((const char*)s, l) +#define __Pyx_PyBytes_FromString PyBytes_FromString +#define __Pyx_PyBytes_FromStringAndSize PyBytes_FromStringAndSize +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char*); +#if PY_MAJOR_VERSION < 3 + #define __Pyx_PyStr_FromString __Pyx_PyBytes_FromString + #define __Pyx_PyStr_FromStringAndSize __Pyx_PyBytes_FromStringAndSize +#else + #define __Pyx_PyStr_FromString __Pyx_PyUnicode_FromString + #define __Pyx_PyStr_FromStringAndSize __Pyx_PyUnicode_FromStringAndSize +#endif +#define __Pyx_PyBytes_AsWritableString(s) ((char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsWritableSString(s) ((signed char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsWritableUString(s) ((unsigned char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsString(s) ((const char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsSString(s) ((const signed char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyBytes_AsUString(s) ((const unsigned char*) PyBytes_AS_STRING(s)) +#define __Pyx_PyObject_AsWritableString(s) ((char*)(__pyx_uintptr_t) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsWritableSString(s) ((signed char*)(__pyx_uintptr_t) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsWritableUString(s) ((unsigned char*)(__pyx_uintptr_t) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsSString(s) ((const signed char*) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_AsUString(s) ((const unsigned char*) __Pyx_PyObject_AsString(s)) +#define __Pyx_PyObject_FromCString(s) __Pyx_PyObject_FromString((const char*)s) +#define __Pyx_PyBytes_FromCString(s) __Pyx_PyBytes_FromString((const char*)s) +#define __Pyx_PyByteArray_FromCString(s) __Pyx_PyByteArray_FromString((const char*)s) +#define __Pyx_PyStr_FromCString(s) __Pyx_PyStr_FromString((const char*)s) +#define __Pyx_PyUnicode_FromCString(s) __Pyx_PyUnicode_FromString((const char*)s) +#define __Pyx_PyUnicode_FromOrdinal(o) PyUnicode_FromOrdinal((int)o) +#define __Pyx_PyUnicode_AsUnicode PyUnicode_AsUnicode +#define __Pyx_NewRef(obj) (Py_INCREF(obj), obj) +#define __Pyx_Owned_Py_None(b) __Pyx_NewRef(Py_None) +static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b); +static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject*); +static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject*); +static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x); +#define __Pyx_PySequence_Tuple(obj)\ + (likely(PyTuple_CheckExact(obj)) ? __Pyx_NewRef(obj) : PySequence_Tuple(obj)) +static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject*); +static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t); +static CYTHON_INLINE Py_hash_t __Pyx_PyIndex_AsHash_t(PyObject*); +#if CYTHON_ASSUME_SAFE_MACROS +#define __pyx_PyFloat_AsDouble(x) (PyFloat_CheckExact(x) ? PyFloat_AS_DOUBLE(x) : PyFloat_AsDouble(x)) +#else +#define __pyx_PyFloat_AsDouble(x) PyFloat_AsDouble(x) +#endif +#define __pyx_PyFloat_AsFloat(x) ((float) __pyx_PyFloat_AsDouble(x)) +#if PY_MAJOR_VERSION >= 3 +#define __Pyx_PyNumber_Int(x) (PyLong_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Long(x)) +#else +#define __Pyx_PyNumber_Int(x) (PyInt_CheckExact(x) ? __Pyx_NewRef(x) : PyNumber_Int(x)) +#endif +#if CYTHON_USE_PYLONG_INTERNALS + #if PY_VERSION_HEX >= 0x030C00A7 + #ifndef _PyLong_SIGN_MASK + #define _PyLong_SIGN_MASK 3 + #endif + #ifndef _PyLong_NON_SIZE_BITS + #define _PyLong_NON_SIZE_BITS 3 + #endif + #define __Pyx_PyLong_Sign(x) (((PyLongObject*)x)->long_value.lv_tag & _PyLong_SIGN_MASK) + #define __Pyx_PyLong_IsNeg(x) ((__Pyx_PyLong_Sign(x) & 2) != 0) + #define __Pyx_PyLong_IsNonNeg(x) (!__Pyx_PyLong_IsNeg(x)) + #define __Pyx_PyLong_IsZero(x) (__Pyx_PyLong_Sign(x) & 1) + #define __Pyx_PyLong_IsPos(x) (__Pyx_PyLong_Sign(x) == 0) + #define __Pyx_PyLong_CompactValueUnsigned(x) (__Pyx_PyLong_Digits(x)[0]) + #define __Pyx_PyLong_DigitCount(x) ((Py_ssize_t) (((PyLongObject*)x)->long_value.lv_tag >> _PyLong_NON_SIZE_BITS)) + #define __Pyx_PyLong_SignedDigitCount(x)\ + ((1 - (Py_ssize_t) __Pyx_PyLong_Sign(x)) * __Pyx_PyLong_DigitCount(x)) + #if defined(PyUnstable_Long_IsCompact) && defined(PyUnstable_Long_CompactValue) + #define __Pyx_PyLong_IsCompact(x) PyUnstable_Long_IsCompact((PyLongObject*) x) + #define __Pyx_PyLong_CompactValue(x) PyUnstable_Long_CompactValue((PyLongObject*) x) + #else + #define __Pyx_PyLong_IsCompact(x) (((PyLongObject*)x)->long_value.lv_tag < (2 << _PyLong_NON_SIZE_BITS)) + #define __Pyx_PyLong_CompactValue(x) ((1 - (Py_ssize_t) __Pyx_PyLong_Sign(x)) * (Py_ssize_t) __Pyx_PyLong_Digits(x)[0]) + #endif + typedef Py_ssize_t __Pyx_compact_pylong; + typedef size_t __Pyx_compact_upylong; + #else + #define __Pyx_PyLong_IsNeg(x) (Py_SIZE(x) < 0) + #define __Pyx_PyLong_IsNonNeg(x) (Py_SIZE(x) >= 0) + #define __Pyx_PyLong_IsZero(x) (Py_SIZE(x) == 0) + #define __Pyx_PyLong_IsPos(x) (Py_SIZE(x) > 0) + #define __Pyx_PyLong_CompactValueUnsigned(x) ((Py_SIZE(x) == 0) ? 0 : __Pyx_PyLong_Digits(x)[0]) + #define __Pyx_PyLong_DigitCount(x) __Pyx_sst_abs(Py_SIZE(x)) + #define __Pyx_PyLong_SignedDigitCount(x) Py_SIZE(x) + #define __Pyx_PyLong_IsCompact(x) (Py_SIZE(x) == 0 || Py_SIZE(x) == 1 || Py_SIZE(x) == -1) + #define __Pyx_PyLong_CompactValue(x)\ + ((Py_SIZE(x) == 0) ? (sdigit) 0 : ((Py_SIZE(x) < 0) ? -(sdigit)__Pyx_PyLong_Digits(x)[0] : (sdigit)__Pyx_PyLong_Digits(x)[0])) + typedef sdigit __Pyx_compact_pylong; + typedef digit __Pyx_compact_upylong; + #endif + #if PY_VERSION_HEX >= 0x030C00A5 + #define __Pyx_PyLong_Digits(x) (((PyLongObject*)x)->long_value.ob_digit) + #else + #define __Pyx_PyLong_Digits(x) (((PyLongObject*)x)->ob_digit) + #endif +#endif +#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII +#include +static int __Pyx_sys_getdefaultencoding_not_ascii; +static int __Pyx_init_sys_getdefaultencoding_params(void) { + PyObject* sys; + PyObject* default_encoding = NULL; + PyObject* ascii_chars_u = NULL; + PyObject* ascii_chars_b = NULL; + const char* default_encoding_c; + sys = PyImport_ImportModule("sys"); + if (!sys) goto bad; + default_encoding = PyObject_CallMethod(sys, (char*) "getdefaultencoding", NULL); + Py_DECREF(sys); + if (!default_encoding) goto bad; + default_encoding_c = PyBytes_AsString(default_encoding); + if (!default_encoding_c) goto bad; + if (strcmp(default_encoding_c, "ascii") == 0) { + __Pyx_sys_getdefaultencoding_not_ascii = 0; + } else { + char ascii_chars[128]; + int c; + for (c = 0; c < 128; c++) { + ascii_chars[c] = (char) c; + } + __Pyx_sys_getdefaultencoding_not_ascii = 1; + ascii_chars_u = PyUnicode_DecodeASCII(ascii_chars, 128, NULL); + if (!ascii_chars_u) goto bad; + ascii_chars_b = PyUnicode_AsEncodedString(ascii_chars_u, default_encoding_c, NULL); + if (!ascii_chars_b || !PyBytes_Check(ascii_chars_b) || memcmp(ascii_chars, PyBytes_AS_STRING(ascii_chars_b), 128) != 0) { + PyErr_Format( + PyExc_ValueError, + "This module compiled with c_string_encoding=ascii, but default encoding '%.200s' is not a superset of ascii.", + default_encoding_c); + goto bad; + } + Py_DECREF(ascii_chars_u); + Py_DECREF(ascii_chars_b); + } + Py_DECREF(default_encoding); + return 0; +bad: + Py_XDECREF(default_encoding); + Py_XDECREF(ascii_chars_u); + Py_XDECREF(ascii_chars_b); + return -1; +} +#endif +#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT && PY_MAJOR_VERSION >= 3 +#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_DecodeUTF8(c_str, size, NULL) +#else +#define __Pyx_PyUnicode_FromStringAndSize(c_str, size) PyUnicode_Decode(c_str, size, __PYX_DEFAULT_STRING_ENCODING, NULL) +#if __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT +#include +static char* __PYX_DEFAULT_STRING_ENCODING; +static int __Pyx_init_sys_getdefaultencoding_params(void) { + PyObject* sys; + PyObject* default_encoding = NULL; + char* default_encoding_c; + sys = PyImport_ImportModule("sys"); + if (!sys) goto bad; + default_encoding = PyObject_CallMethod(sys, (char*) (const char*) "getdefaultencoding", NULL); + Py_DECREF(sys); + if (!default_encoding) goto bad; + default_encoding_c = PyBytes_AsString(default_encoding); + if (!default_encoding_c) goto bad; + __PYX_DEFAULT_STRING_ENCODING = (char*) malloc(strlen(default_encoding_c) + 1); + if (!__PYX_DEFAULT_STRING_ENCODING) goto bad; + strcpy(__PYX_DEFAULT_STRING_ENCODING, default_encoding_c); + Py_DECREF(default_encoding); + return 0; +bad: + Py_XDECREF(default_encoding); + return -1; +} +#endif +#endif + + +/* Test for GCC > 2.95 */ +#if defined(__GNUC__) && (__GNUC__ > 2 || (__GNUC__ == 2 && (__GNUC_MINOR__ > 95))) + #define likely(x) __builtin_expect(!!(x), 1) + #define unlikely(x) __builtin_expect(!!(x), 0) +#else /* !__GNUC__ or GCC < 2.95 */ + #define likely(x) (x) + #define unlikely(x) (x) +#endif /* __GNUC__ */ +static CYTHON_INLINE void __Pyx_pretend_to_initialize(void* ptr) { (void)ptr; } + +#if !CYTHON_USE_MODULE_STATE +static PyObject *__pyx_m = NULL; +#endif +static int __pyx_lineno; +static int __pyx_clineno = 0; +static const char * __pyx_cfilenm = __FILE__; +static const char *__pyx_filename; + +/* Header.proto */ +#if !defined(CYTHON_CCOMPLEX) + #if defined(__cplusplus) + #define CYTHON_CCOMPLEX 1 + #elif (defined(_Complex_I) && !defined(_MSC_VER)) || ((defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) && !defined(__STDC_NO_COMPLEX__) && !defined(_MSC_VER)) + #define CYTHON_CCOMPLEX 1 + #else + #define CYTHON_CCOMPLEX 0 + #endif +#endif +#if CYTHON_CCOMPLEX + #ifdef __cplusplus + #include + #else + #include + #endif +#endif +#if CYTHON_CCOMPLEX && !defined(__cplusplus) && defined(__sun__) && defined(__GNUC__) + #undef _Complex_I + #define _Complex_I 1.0fj +#endif + +/* #### Code section: filename_table ### */ + +static const char *__pyx_f[] = { + "fairseq/data/token_block_utils_fast.pyx", + "", + "__init__.cython-30.pxd", + "type.pxd", +}; +/* #### Code section: utility_code_proto_before_types ### */ +/* ForceInitThreads.proto */ +#ifndef __PYX_FORCE_INIT_THREADS + #define __PYX_FORCE_INIT_THREADS 0 +#endif + +/* NoFastGil.proto */ +#define __Pyx_PyGILState_Ensure PyGILState_Ensure +#define __Pyx_PyGILState_Release PyGILState_Release +#define __Pyx_FastGIL_Remember() +#define __Pyx_FastGIL_Forget() +#define __Pyx_FastGilFuncInit() + +/* BufferFormatStructs.proto */ +struct __Pyx_StructField_; +#define __PYX_BUF_FLAGS_PACKED_STRUCT (1 << 0) +typedef struct { + const char* name; + struct __Pyx_StructField_* fields; + size_t size; + size_t arraysize[8]; + int ndim; + char typegroup; + char is_unsigned; + int flags; +} __Pyx_TypeInfo; +typedef struct __Pyx_StructField_ { + __Pyx_TypeInfo* type; + const char* name; + size_t offset; +} __Pyx_StructField; +typedef struct { + __Pyx_StructField* field; + size_t parent_offset; +} __Pyx_BufFmt_StackElem; +typedef struct { + __Pyx_StructField root; + __Pyx_BufFmt_StackElem* head; + size_t fmt_offset; + size_t new_count, enc_count; + size_t struct_alignment; + int is_complex; + char enc_type; + char new_packmode; + char enc_packmode; + char is_valid_array; +} __Pyx_BufFmt_Context; + +/* Atomics.proto */ +#include +#ifndef CYTHON_ATOMICS + #define CYTHON_ATOMICS 1 +#endif +#define __PYX_CYTHON_ATOMICS_ENABLED() CYTHON_ATOMICS +#define __pyx_atomic_int_type int +#define __pyx_nonatomic_int_type int +#if CYTHON_ATOMICS && (defined(__STDC_VERSION__) &&\ + (__STDC_VERSION__ >= 201112L) &&\ + !defined(__STDC_NO_ATOMICS__)) + #include +#elif CYTHON_ATOMICS && (defined(__cplusplus) && (\ + (__cplusplus >= 201103L) ||\ + (defined(_MSC_VER) && _MSC_VER >= 1700))) + #include +#endif +#if CYTHON_ATOMICS && (defined(__STDC_VERSION__) &&\ + (__STDC_VERSION__ >= 201112L) &&\ + !defined(__STDC_NO_ATOMICS__) &&\ + ATOMIC_INT_LOCK_FREE == 2) + #undef __pyx_atomic_int_type + #define __pyx_atomic_int_type atomic_int + #define __pyx_atomic_incr_aligned(value) atomic_fetch_add_explicit(value, 1, memory_order_relaxed) + #define __pyx_atomic_decr_aligned(value) atomic_fetch_sub_explicit(value, 1, memory_order_acq_rel) + #if defined(__PYX_DEBUG_ATOMICS) && defined(_MSC_VER) + #pragma message ("Using standard C atomics") + #elif defined(__PYX_DEBUG_ATOMICS) + #warning "Using standard C atomics" + #endif +#elif CYTHON_ATOMICS && (defined(__cplusplus) && (\ + (__cplusplus >= 201103L) ||\ +\ + (defined(_MSC_VER) && _MSC_VER >= 1700)) &&\ + ATOMIC_INT_LOCK_FREE == 2) + #undef __pyx_atomic_int_type + #define __pyx_atomic_int_type std::atomic_int + #define __pyx_atomic_incr_aligned(value) std::atomic_fetch_add_explicit(value, 1, std::memory_order_relaxed) + #define __pyx_atomic_decr_aligned(value) std::atomic_fetch_sub_explicit(value, 1, std::memory_order_acq_rel) + #if defined(__PYX_DEBUG_ATOMICS) && defined(_MSC_VER) + #pragma message ("Using standard C++ atomics") + #elif defined(__PYX_DEBUG_ATOMICS) + #warning "Using standard C++ atomics" + #endif +#elif CYTHON_ATOMICS && (__GNUC__ >= 5 || (__GNUC__ == 4 &&\ + (__GNUC_MINOR__ > 1 ||\ + (__GNUC_MINOR__ == 1 && __GNUC_PATCHLEVEL__ >= 2)))) + #define __pyx_atomic_incr_aligned(value) __sync_fetch_and_add(value, 1) + #define __pyx_atomic_decr_aligned(value) __sync_fetch_and_sub(value, 1) + #ifdef __PYX_DEBUG_ATOMICS + #warning "Using GNU atomics" + #endif +#elif CYTHON_ATOMICS && defined(_MSC_VER) + #include + #undef __pyx_atomic_int_type + #define __pyx_atomic_int_type long + #undef __pyx_nonatomic_int_type + #define __pyx_nonatomic_int_type long + #pragma intrinsic (_InterlockedExchangeAdd) + #define __pyx_atomic_incr_aligned(value) _InterlockedExchangeAdd(value, 1) + #define __pyx_atomic_decr_aligned(value) _InterlockedExchangeAdd(value, -1) + #ifdef __PYX_DEBUG_ATOMICS + #pragma message ("Using MSVC atomics") + #endif +#else + #undef CYTHON_ATOMICS + #define CYTHON_ATOMICS 0 + #ifdef __PYX_DEBUG_ATOMICS + #warning "Not using atomics" + #endif +#endif +#if CYTHON_ATOMICS + #define __pyx_add_acquisition_count(memview)\ + __pyx_atomic_incr_aligned(__pyx_get_slice_count_pointer(memview)) + #define __pyx_sub_acquisition_count(memview)\ + __pyx_atomic_decr_aligned(__pyx_get_slice_count_pointer(memview)) +#else + #define __pyx_add_acquisition_count(memview)\ + __pyx_add_acquisition_count_locked(__pyx_get_slice_count_pointer(memview), memview->lock) + #define __pyx_sub_acquisition_count(memview)\ + __pyx_sub_acquisition_count_locked(__pyx_get_slice_count_pointer(memview), memview->lock) +#endif + +/* MemviewSliceStruct.proto */ +struct __pyx_memoryview_obj; +typedef struct { + struct __pyx_memoryview_obj *memview; + char *data; + Py_ssize_t shape[8]; + Py_ssize_t strides[8]; + Py_ssize_t suboffsets[8]; +} __Pyx_memviewslice; +#define __Pyx_MemoryView_Len(m) (m.shape[0]) + +/* #### Code section: numeric_typedefs ### */ + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":787 + * # in Cython to enable them only on the right systems. + * + * ctypedef npy_int8 int8_t # <<<<<<<<<<<<<< + * ctypedef npy_int16 int16_t + * ctypedef npy_int32 int32_t + */ +typedef npy_int8 __pyx_t_5numpy_int8_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":788 + * + * ctypedef npy_int8 int8_t + * ctypedef npy_int16 int16_t # <<<<<<<<<<<<<< + * ctypedef npy_int32 int32_t + * ctypedef npy_int64 int64_t + */ +typedef npy_int16 __pyx_t_5numpy_int16_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":789 + * ctypedef npy_int8 int8_t + * ctypedef npy_int16 int16_t + * ctypedef npy_int32 int32_t # <<<<<<<<<<<<<< + * ctypedef npy_int64 int64_t + * #ctypedef npy_int96 int96_t + */ +typedef npy_int32 __pyx_t_5numpy_int32_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":790 + * ctypedef npy_int16 int16_t + * ctypedef npy_int32 int32_t + * ctypedef npy_int64 int64_t # <<<<<<<<<<<<<< + * #ctypedef npy_int96 int96_t + * #ctypedef npy_int128 int128_t + */ +typedef npy_int64 __pyx_t_5numpy_int64_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":794 + * #ctypedef npy_int128 int128_t + * + * ctypedef npy_uint8 uint8_t # <<<<<<<<<<<<<< + * ctypedef npy_uint16 uint16_t + * ctypedef npy_uint32 uint32_t + */ +typedef npy_uint8 __pyx_t_5numpy_uint8_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":795 + * + * ctypedef npy_uint8 uint8_t + * ctypedef npy_uint16 uint16_t # <<<<<<<<<<<<<< + * ctypedef npy_uint32 uint32_t + * ctypedef npy_uint64 uint64_t + */ +typedef npy_uint16 __pyx_t_5numpy_uint16_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":796 + * ctypedef npy_uint8 uint8_t + * ctypedef npy_uint16 uint16_t + * ctypedef npy_uint32 uint32_t # <<<<<<<<<<<<<< + * ctypedef npy_uint64 uint64_t + * #ctypedef npy_uint96 uint96_t + */ +typedef npy_uint32 __pyx_t_5numpy_uint32_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":797 + * ctypedef npy_uint16 uint16_t + * ctypedef npy_uint32 uint32_t + * ctypedef npy_uint64 uint64_t # <<<<<<<<<<<<<< + * #ctypedef npy_uint96 uint96_t + * #ctypedef npy_uint128 uint128_t + */ +typedef npy_uint64 __pyx_t_5numpy_uint64_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":801 + * #ctypedef npy_uint128 uint128_t + * + * ctypedef npy_float32 float32_t # <<<<<<<<<<<<<< + * ctypedef npy_float64 float64_t + * #ctypedef npy_float80 float80_t + */ +typedef npy_float32 __pyx_t_5numpy_float32_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":802 + * + * ctypedef npy_float32 float32_t + * ctypedef npy_float64 float64_t # <<<<<<<<<<<<<< + * #ctypedef npy_float80 float80_t + * #ctypedef npy_float128 float128_t + */ +typedef npy_float64 __pyx_t_5numpy_float64_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":809 + * ctypedef double complex complex128_t + * + * ctypedef npy_longlong longlong_t # <<<<<<<<<<<<<< + * ctypedef npy_ulonglong ulonglong_t + * + */ +typedef npy_longlong __pyx_t_5numpy_longlong_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":810 + * + * ctypedef npy_longlong longlong_t + * ctypedef npy_ulonglong ulonglong_t # <<<<<<<<<<<<<< + * + * ctypedef npy_intp intp_t + */ +typedef npy_ulonglong __pyx_t_5numpy_ulonglong_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":812 + * ctypedef npy_ulonglong ulonglong_t + * + * ctypedef npy_intp intp_t # <<<<<<<<<<<<<< + * ctypedef npy_uintp uintp_t + * + */ +typedef npy_intp __pyx_t_5numpy_intp_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":813 + * + * ctypedef npy_intp intp_t + * ctypedef npy_uintp uintp_t # <<<<<<<<<<<<<< + * + * ctypedef npy_double float_t + */ +typedef npy_uintp __pyx_t_5numpy_uintp_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":815 + * ctypedef npy_uintp uintp_t + * + * ctypedef npy_double float_t # <<<<<<<<<<<<<< + * ctypedef npy_double double_t + * ctypedef npy_longdouble longdouble_t + */ +typedef npy_double __pyx_t_5numpy_float_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":816 + * + * ctypedef npy_double float_t + * ctypedef npy_double double_t # <<<<<<<<<<<<<< + * ctypedef npy_longdouble longdouble_t + * + */ +typedef npy_double __pyx_t_5numpy_double_t; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":817 + * ctypedef npy_double float_t + * ctypedef npy_double double_t + * ctypedef npy_longdouble longdouble_t # <<<<<<<<<<<<<< + * + * ctypedef float complex cfloat_t + */ +typedef npy_longdouble __pyx_t_5numpy_longdouble_t; + +/* "fairseq/data/token_block_utils_fast.pyx":16 + * + * DTYPE = np.int64 + * ctypedef np.int64_t DTYPE_t # <<<<<<<<<<<<<< + * + * + */ +typedef __pyx_t_5numpy_int64_t __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t; +/* #### Code section: complex_type_declarations ### */ +/* Declarations.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + typedef ::std::complex< float > __pyx_t_float_complex; + #else + typedef float _Complex __pyx_t_float_complex; + #endif +#else + typedef struct { float real, imag; } __pyx_t_float_complex; +#endif +static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float, float); + +/* Declarations.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + typedef ::std::complex< double > __pyx_t_double_complex; + #else + typedef double _Complex __pyx_t_double_complex; + #endif +#else + typedef struct { double real, imag; } __pyx_t_double_complex; +#endif +static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double, double); + +/* Declarations.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + typedef ::std::complex< long double > __pyx_t_long_double_complex; + #else + typedef long double _Complex __pyx_t_long_double_complex; + #endif +#else + typedef struct { long double real, imag; } __pyx_t_long_double_complex; +#endif +static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double, long double); + +/* #### Code section: type_declarations ### */ + +/*--- Type declarations ---*/ +struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; +struct __pyx_array_obj; +struct __pyx_MemviewEnum_obj; +struct __pyx_memoryview_obj; +struct __pyx_memoryviewslice_obj; + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1113 + * + * # Iterator API added in v1.6 + * ctypedef int (*NpyIter_IterNextFunc)(NpyIter* it) noexcept nogil # <<<<<<<<<<<<<< + * ctypedef void (*NpyIter_GetMultiIndexFunc)(NpyIter* it, npy_intp* outcoords) noexcept nogil + * + */ +typedef int (*__pyx_t_5numpy_NpyIter_IterNextFunc)(NpyIter *); + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1114 + * # Iterator API added in v1.6 + * ctypedef int (*NpyIter_IterNextFunc)(NpyIter* it) noexcept nogil + * ctypedef void (*NpyIter_GetMultiIndexFunc)(NpyIter* it, npy_intp* outcoords) noexcept nogil # <<<<<<<<<<<<<< + * + * cdef extern from "numpy/arrayobject.h": + */ +typedef void (*__pyx_t_5numpy_NpyIter_GetMultiIndexFunc)(NpyIter *, npy_intp *); + +/* "fairseq/data/token_block_utils_fast.pyx":139 + * + * + * cdef class DatasetSearcher(object): # <<<<<<<<<<<<<< + * """Helper for mapping "flat" indices to indices and offsets in an + * underlying dataset.""" + */ +struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher { + PyObject_HEAD + struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_vtab; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t current_i; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t current_offset; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t current_index; + __Pyx_memviewslice sizes; +}; + + +/* "View.MemoryView":114 + * @cython.collection_type("sequence") + * @cname("__pyx_array") + * cdef class array: # <<<<<<<<<<<<<< + * + * cdef: + */ +struct __pyx_array_obj { + PyObject_HEAD + struct __pyx_vtabstruct_array *__pyx_vtab; + char *data; + Py_ssize_t len; + char *format; + int ndim; + Py_ssize_t *_shape; + Py_ssize_t *_strides; + Py_ssize_t itemsize; + PyObject *mode; + PyObject *_format; + void (*callback_free_data)(void *); + int free_data; + int dtype_is_object; +}; + + +/* "View.MemoryView":302 + * + * @cname('__pyx_MemviewEnum') + * cdef class Enum(object): # <<<<<<<<<<<<<< + * cdef object name + * def __init__(self, name): + */ +struct __pyx_MemviewEnum_obj { + PyObject_HEAD + PyObject *name; +}; + + +/* "View.MemoryView":337 + * + * @cname('__pyx_memoryview') + * cdef class memoryview: # <<<<<<<<<<<<<< + * + * cdef object obj + */ +struct __pyx_memoryview_obj { + PyObject_HEAD + struct __pyx_vtabstruct_memoryview *__pyx_vtab; + PyObject *obj; + PyObject *_size; + PyObject *_array_interface; + PyThread_type_lock lock; + __pyx_atomic_int_type acquisition_count; + Py_buffer view; + int flags; + int dtype_is_object; + __Pyx_TypeInfo *typeinfo; +}; + + +/* "View.MemoryView":952 + * @cython.collection_type("sequence") + * @cname('__pyx_memoryviewslice') + * cdef class _memoryviewslice(memoryview): # <<<<<<<<<<<<<< + * "Internal class for passing memoryview slices to Python" + * + */ +struct __pyx_memoryviewslice_obj { + struct __pyx_memoryview_obj __pyx_base; + __Pyx_memviewslice from_slice; + PyObject *from_object; + PyObject *(*to_object_func)(char *); + int (*to_dtype_func)(char *, PyObject *); +}; + + + +/* "fairseq/data/token_block_utils_fast.pyx":139 + * + * + * cdef class DatasetSearcher(object): # <<<<<<<<<<<<<< + * """Helper for mapping "flat" indices to indices and offsets in an + * underlying dataset.""" + */ + +struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher { + PyObject *(*reset)(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *); + int (*step)(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t); + PyObject *(*seek)(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t); +}; +static struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_vtabptr_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; + + +/* "View.MemoryView":114 + * @cython.collection_type("sequence") + * @cname("__pyx_array") + * cdef class array: # <<<<<<<<<<<<<< + * + * cdef: + */ + +struct __pyx_vtabstruct_array { + PyObject *(*get_memview)(struct __pyx_array_obj *); +}; +static struct __pyx_vtabstruct_array *__pyx_vtabptr_array; + + +/* "View.MemoryView":337 + * + * @cname('__pyx_memoryview') + * cdef class memoryview: # <<<<<<<<<<<<<< + * + * cdef object obj + */ + +struct __pyx_vtabstruct_memoryview { + char *(*get_item_pointer)(struct __pyx_memoryview_obj *, PyObject *); + PyObject *(*is_slice)(struct __pyx_memoryview_obj *, PyObject *); + PyObject *(*setitem_slice_assignment)(struct __pyx_memoryview_obj *, PyObject *, PyObject *); + PyObject *(*setitem_slice_assign_scalar)(struct __pyx_memoryview_obj *, struct __pyx_memoryview_obj *, PyObject *); + PyObject *(*setitem_indexed)(struct __pyx_memoryview_obj *, PyObject *, PyObject *); + PyObject *(*convert_item_to_object)(struct __pyx_memoryview_obj *, char *); + PyObject *(*assign_item_from_object)(struct __pyx_memoryview_obj *, char *, PyObject *); + PyObject *(*_get_base)(struct __pyx_memoryview_obj *); +}; +static struct __pyx_vtabstruct_memoryview *__pyx_vtabptr_memoryview; + + +/* "View.MemoryView":952 + * @cython.collection_type("sequence") + * @cname('__pyx_memoryviewslice') + * cdef class _memoryviewslice(memoryview): # <<<<<<<<<<<<<< + * "Internal class for passing memoryview slices to Python" + * + */ + +struct __pyx_vtabstruct__memoryviewslice { + struct __pyx_vtabstruct_memoryview __pyx_base; +}; +static struct __pyx_vtabstruct__memoryviewslice *__pyx_vtabptr__memoryviewslice; +/* #### Code section: utility_code_proto ### */ + +/* --- Runtime support code (head) --- */ +/* Refnanny.proto */ +#ifndef CYTHON_REFNANNY + #define CYTHON_REFNANNY 0 +#endif +#if CYTHON_REFNANNY + typedef struct { + void (*INCREF)(void*, PyObject*, Py_ssize_t); + void (*DECREF)(void*, PyObject*, Py_ssize_t); + void (*GOTREF)(void*, PyObject*, Py_ssize_t); + void (*GIVEREF)(void*, PyObject*, Py_ssize_t); + void* (*SetupContext)(const char*, Py_ssize_t, const char*); + void (*FinishContext)(void**); + } __Pyx_RefNannyAPIStruct; + static __Pyx_RefNannyAPIStruct *__Pyx_RefNanny = NULL; + static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname); + #define __Pyx_RefNannyDeclarations void *__pyx_refnanny = NULL; +#ifdef WITH_THREAD + #define __Pyx_RefNannySetupContext(name, acquire_gil)\ + if (acquire_gil) {\ + PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\ + __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), (__LINE__), (__FILE__));\ + PyGILState_Release(__pyx_gilstate_save);\ + } else {\ + __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), (__LINE__), (__FILE__));\ + } + #define __Pyx_RefNannyFinishContextNogil() {\ + PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\ + __Pyx_RefNannyFinishContext();\ + PyGILState_Release(__pyx_gilstate_save);\ + } +#else + #define __Pyx_RefNannySetupContext(name, acquire_gil)\ + __pyx_refnanny = __Pyx_RefNanny->SetupContext((name), (__LINE__), (__FILE__)) + #define __Pyx_RefNannyFinishContextNogil() __Pyx_RefNannyFinishContext() +#endif + #define __Pyx_RefNannyFinishContextNogil() {\ + PyGILState_STATE __pyx_gilstate_save = PyGILState_Ensure();\ + __Pyx_RefNannyFinishContext();\ + PyGILState_Release(__pyx_gilstate_save);\ + } + #define __Pyx_RefNannyFinishContext()\ + __Pyx_RefNanny->FinishContext(&__pyx_refnanny) + #define __Pyx_INCREF(r) __Pyx_RefNanny->INCREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_DECREF(r) __Pyx_RefNanny->DECREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_GOTREF(r) __Pyx_RefNanny->GOTREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_GIVEREF(r) __Pyx_RefNanny->GIVEREF(__pyx_refnanny, (PyObject *)(r), (__LINE__)) + #define __Pyx_XINCREF(r) do { if((r) == NULL); else {__Pyx_INCREF(r); }} while(0) + #define __Pyx_XDECREF(r) do { if((r) == NULL); else {__Pyx_DECREF(r); }} while(0) + #define __Pyx_XGOTREF(r) do { if((r) == NULL); else {__Pyx_GOTREF(r); }} while(0) + #define __Pyx_XGIVEREF(r) do { if((r) == NULL); else {__Pyx_GIVEREF(r);}} while(0) +#else + #define __Pyx_RefNannyDeclarations + #define __Pyx_RefNannySetupContext(name, acquire_gil) + #define __Pyx_RefNannyFinishContextNogil() + #define __Pyx_RefNannyFinishContext() + #define __Pyx_INCREF(r) Py_INCREF(r) + #define __Pyx_DECREF(r) Py_DECREF(r) + #define __Pyx_GOTREF(r) + #define __Pyx_GIVEREF(r) + #define __Pyx_XINCREF(r) Py_XINCREF(r) + #define __Pyx_XDECREF(r) Py_XDECREF(r) + #define __Pyx_XGOTREF(r) + #define __Pyx_XGIVEREF(r) +#endif +#define __Pyx_Py_XDECREF_SET(r, v) do {\ + PyObject *tmp = (PyObject *) r;\ + r = v; Py_XDECREF(tmp);\ + } while (0) +#define __Pyx_XDECREF_SET(r, v) do {\ + PyObject *tmp = (PyObject *) r;\ + r = v; __Pyx_XDECREF(tmp);\ + } while (0) +#define __Pyx_DECREF_SET(r, v) do {\ + PyObject *tmp = (PyObject *) r;\ + r = v; __Pyx_DECREF(tmp);\ + } while (0) +#define __Pyx_CLEAR(r) do { PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);} while(0) +#define __Pyx_XCLEAR(r) do { if((r) != NULL) {PyObject* tmp = ((PyObject*)(r)); r = NULL; __Pyx_DECREF(tmp);}} while(0) + +/* PyErrExceptionMatches.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_PyErr_ExceptionMatches(err) __Pyx_PyErr_ExceptionMatchesInState(__pyx_tstate, err) +static CYTHON_INLINE int __Pyx_PyErr_ExceptionMatchesInState(PyThreadState* tstate, PyObject* err); +#else +#define __Pyx_PyErr_ExceptionMatches(err) PyErr_ExceptionMatches(err) +#endif + +/* PyThreadStateGet.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_PyThreadState_declare PyThreadState *__pyx_tstate; +#define __Pyx_PyThreadState_assign __pyx_tstate = __Pyx_PyThreadState_Current; +#if PY_VERSION_HEX >= 0x030C00A6 +#define __Pyx_PyErr_Occurred() (__pyx_tstate->current_exception != NULL) +#define __Pyx_PyErr_CurrentExceptionType() (__pyx_tstate->current_exception ? (PyObject*) Py_TYPE(__pyx_tstate->current_exception) : (PyObject*) NULL) +#else +#define __Pyx_PyErr_Occurred() (__pyx_tstate->curexc_type != NULL) +#define __Pyx_PyErr_CurrentExceptionType() (__pyx_tstate->curexc_type) +#endif +#else +#define __Pyx_PyThreadState_declare +#define __Pyx_PyThreadState_assign +#define __Pyx_PyErr_Occurred() (PyErr_Occurred() != NULL) +#define __Pyx_PyErr_CurrentExceptionType() PyErr_Occurred() +#endif + +/* PyErrFetchRestore.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_PyErr_Clear() __Pyx_ErrRestore(NULL, NULL, NULL) +#define __Pyx_ErrRestoreWithState(type, value, tb) __Pyx_ErrRestoreInState(PyThreadState_GET(), type, value, tb) +#define __Pyx_ErrFetchWithState(type, value, tb) __Pyx_ErrFetchInState(PyThreadState_GET(), type, value, tb) +#define __Pyx_ErrRestore(type, value, tb) __Pyx_ErrRestoreInState(__pyx_tstate, type, value, tb) +#define __Pyx_ErrFetch(type, value, tb) __Pyx_ErrFetchInState(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb); +static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A6 +#define __Pyx_PyErr_SetNone(exc) (Py_INCREF(exc), __Pyx_ErrRestore((exc), NULL, NULL)) +#else +#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc) +#endif +#else +#define __Pyx_PyErr_Clear() PyErr_Clear() +#define __Pyx_PyErr_SetNone(exc) PyErr_SetNone(exc) +#define __Pyx_ErrRestoreWithState(type, value, tb) PyErr_Restore(type, value, tb) +#define __Pyx_ErrFetchWithState(type, value, tb) PyErr_Fetch(type, value, tb) +#define __Pyx_ErrRestoreInState(tstate, type, value, tb) PyErr_Restore(type, value, tb) +#define __Pyx_ErrFetchInState(tstate, type, value, tb) PyErr_Fetch(type, value, tb) +#define __Pyx_ErrRestore(type, value, tb) PyErr_Restore(type, value, tb) +#define __Pyx_ErrFetch(type, value, tb) PyErr_Fetch(type, value, tb) +#endif + +/* PyObjectGetAttrStr.proto */ +#if CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name); +#else +#define __Pyx_PyObject_GetAttrStr(o,n) PyObject_GetAttr(o,n) +#endif + +/* PyObjectGetAttrStrNoError.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStrNoError(PyObject* obj, PyObject* attr_name); + +/* GetBuiltinName.proto */ +static PyObject *__Pyx_GetBuiltinName(PyObject *name); + +/* TupleAndListFromArray.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyList_FromArray(PyObject *const *src, Py_ssize_t n); +static CYTHON_INLINE PyObject* __Pyx_PyTuple_FromArray(PyObject *const *src, Py_ssize_t n); +#endif + +/* IncludeStringH.proto */ +#include + +/* BytesEquals.proto */ +static CYTHON_INLINE int __Pyx_PyBytes_Equals(PyObject* s1, PyObject* s2, int equals); + +/* UnicodeEquals.proto */ +static CYTHON_INLINE int __Pyx_PyUnicode_Equals(PyObject* s1, PyObject* s2, int equals); + +/* fastcall.proto */ +#if CYTHON_AVOID_BORROWED_REFS + #define __Pyx_Arg_VARARGS(args, i) PySequence_GetItem(args, i) +#elif CYTHON_ASSUME_SAFE_MACROS + #define __Pyx_Arg_VARARGS(args, i) PyTuple_GET_ITEM(args, i) +#else + #define __Pyx_Arg_VARARGS(args, i) PyTuple_GetItem(args, i) +#endif +#if CYTHON_AVOID_BORROWED_REFS + #define __Pyx_Arg_NewRef_VARARGS(arg) __Pyx_NewRef(arg) + #define __Pyx_Arg_XDECREF_VARARGS(arg) Py_XDECREF(arg) +#else + #define __Pyx_Arg_NewRef_VARARGS(arg) arg + #define __Pyx_Arg_XDECREF_VARARGS(arg) +#endif +#define __Pyx_NumKwargs_VARARGS(kwds) PyDict_Size(kwds) +#define __Pyx_KwValues_VARARGS(args, nargs) NULL +#define __Pyx_GetKwValue_VARARGS(kw, kwvalues, s) __Pyx_PyDict_GetItemStrWithError(kw, s) +#define __Pyx_KwargsAsDict_VARARGS(kw, kwvalues) PyDict_Copy(kw) +#if CYTHON_METH_FASTCALL + #define __Pyx_Arg_FASTCALL(args, i) args[i] + #define __Pyx_NumKwargs_FASTCALL(kwds) PyTuple_GET_SIZE(kwds) + #define __Pyx_KwValues_FASTCALL(args, nargs) ((args) + (nargs)) + static CYTHON_INLINE PyObject * __Pyx_GetKwValue_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues, PyObject *s); +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 + CYTHON_UNUSED static PyObject *__Pyx_KwargsAsDict_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues); + #else + #define __Pyx_KwargsAsDict_FASTCALL(kw, kwvalues) _PyStack_AsDict(kwvalues, kw) + #endif + #define __Pyx_Arg_NewRef_FASTCALL(arg) arg /* no-op, __Pyx_Arg_FASTCALL is direct and this needs + to have the same reference counting */ + #define __Pyx_Arg_XDECREF_FASTCALL(arg) +#else + #define __Pyx_Arg_FASTCALL __Pyx_Arg_VARARGS + #define __Pyx_NumKwargs_FASTCALL __Pyx_NumKwargs_VARARGS + #define __Pyx_KwValues_FASTCALL __Pyx_KwValues_VARARGS + #define __Pyx_GetKwValue_FASTCALL __Pyx_GetKwValue_VARARGS + #define __Pyx_KwargsAsDict_FASTCALL __Pyx_KwargsAsDict_VARARGS + #define __Pyx_Arg_NewRef_FASTCALL(arg) __Pyx_Arg_NewRef_VARARGS(arg) + #define __Pyx_Arg_XDECREF_FASTCALL(arg) __Pyx_Arg_XDECREF_VARARGS(arg) +#endif +#if CYTHON_COMPILING_IN_CPYTHON && CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS +#define __Pyx_ArgsSlice_VARARGS(args, start, stop) __Pyx_PyTuple_FromArray(&__Pyx_Arg_VARARGS(args, start), stop - start) +#define __Pyx_ArgsSlice_FASTCALL(args, start, stop) __Pyx_PyTuple_FromArray(&__Pyx_Arg_FASTCALL(args, start), stop - start) +#else +#define __Pyx_ArgsSlice_VARARGS(args, start, stop) PyTuple_GetSlice(args, start, stop) +#define __Pyx_ArgsSlice_FASTCALL(args, start, stop) PyTuple_GetSlice(args, start, stop) +#endif + +/* RaiseArgTupleInvalid.proto */ +static void __Pyx_RaiseArgtupleInvalid(const char* func_name, int exact, + Py_ssize_t num_min, Py_ssize_t num_max, Py_ssize_t num_found); + +/* RaiseDoubleKeywords.proto */ +static void __Pyx_RaiseDoubleKeywordsError(const char* func_name, PyObject* kw_name); + +/* ParseKeywords.proto */ +static int __Pyx_ParseOptionalKeywords(PyObject *kwds, PyObject *const *kwvalues, + PyObject **argnames[], + PyObject *kwds2, PyObject *values[], Py_ssize_t num_pos_args, + const char* function_name); + +/* ArgTypeTest.proto */ +#define __Pyx_ArgTypeTest(obj, type, none_allowed, name, exact)\ + ((likely(__Pyx_IS_TYPE(obj, type) | (none_allowed && (obj == Py_None)))) ? 1 :\ + __Pyx__ArgTypeTest(obj, type, name, exact)) +static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact); + +/* RaiseException.proto */ +static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause); + +/* PyFunctionFastCall.proto */ +#if CYTHON_FAST_PYCALL +#if !CYTHON_VECTORCALL +#define __Pyx_PyFunction_FastCall(func, args, nargs)\ + __Pyx_PyFunction_FastCallDict((func), (args), (nargs), NULL) +static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, Py_ssize_t nargs, PyObject *kwargs); +#endif +#define __Pyx_BUILD_ASSERT_EXPR(cond)\ + (sizeof(char [1 - 2*!(cond)]) - 1) +#ifndef Py_MEMBER_SIZE +#define Py_MEMBER_SIZE(type, member) sizeof(((type *)0)->member) +#endif +#if !CYTHON_VECTORCALL +#if PY_VERSION_HEX >= 0x03080000 + #include "frameobject.h" +#if PY_VERSION_HEX >= 0x030b00a6 && !CYTHON_COMPILING_IN_LIMITED_API && !defined(PYPY_VERSION) + #ifndef Py_BUILD_CORE + #define Py_BUILD_CORE 1 + #endif + #include "internal/pycore_frame.h" +#endif + #define __Pxy_PyFrame_Initialize_Offsets() + #define __Pyx_PyFrame_GetLocalsplus(frame) ((frame)->f_localsplus) +#else + static size_t __pyx_pyframe_localsplus_offset = 0; + #include "frameobject.h" + #define __Pxy_PyFrame_Initialize_Offsets()\ + ((void)__Pyx_BUILD_ASSERT_EXPR(sizeof(PyFrameObject) == offsetof(PyFrameObject, f_localsplus) + Py_MEMBER_SIZE(PyFrameObject, f_localsplus)),\ + (void)(__pyx_pyframe_localsplus_offset = ((size_t)PyFrame_Type.tp_basicsize) - Py_MEMBER_SIZE(PyFrameObject, f_localsplus))) + #define __Pyx_PyFrame_GetLocalsplus(frame)\ + (assert(__pyx_pyframe_localsplus_offset), (PyObject **)(((char *)(frame)) + __pyx_pyframe_localsplus_offset)) +#endif +#endif +#endif + +/* PyObjectCall.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw); +#else +#define __Pyx_PyObject_Call(func, arg, kw) PyObject_Call(func, arg, kw) +#endif + +/* PyObjectCallMethO.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg); +#endif + +/* PyObjectFastCall.proto */ +#define __Pyx_PyObject_FastCall(func, args, nargs) __Pyx_PyObject_FastCallDict(func, args, (size_t)(nargs), NULL) +static CYTHON_INLINE PyObject* __Pyx_PyObject_FastCallDict(PyObject *func, PyObject **args, size_t nargs, PyObject *kwargs); + +/* RaiseUnexpectedTypeError.proto */ +static int __Pyx_RaiseUnexpectedTypeError(const char *expected, PyObject *obj); + +/* GCCDiagnostics.proto */ +#if !defined(__INTEL_COMPILER) && defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)) +#define __Pyx_HAS_GCC_DIAGNOSTIC +#endif + +/* BuildPyUnicode.proto */ +static PyObject* __Pyx_PyUnicode_BuildFromAscii(Py_ssize_t ulength, char* chars, int clength, + int prepend_sign, char padding_char); + +/* CIntToPyUnicode.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_int(int value, Py_ssize_t width, char padding_char, char format_char); + +/* CIntToPyUnicode.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_Py_ssize_t(Py_ssize_t value, Py_ssize_t width, char padding_char, char format_char); + +/* JoinPyUnicode.proto */ +static PyObject* __Pyx_PyUnicode_Join(PyObject* value_tuple, Py_ssize_t value_count, Py_ssize_t result_ulength, + Py_UCS4 max_char); + +/* StrEquals.proto */ +#if PY_MAJOR_VERSION >= 3 +#define __Pyx_PyString_Equals __Pyx_PyUnicode_Equals +#else +#define __Pyx_PyString_Equals __Pyx_PyBytes_Equals +#endif + +/* PyObjectFormatSimple.proto */ +#if CYTHON_COMPILING_IN_PYPY + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + PyObject_Format(s, f)) +#elif PY_MAJOR_VERSION < 3 + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + likely(PyString_CheckExact(s)) ? PyUnicode_FromEncodedObject(s, NULL, "strict") :\ + PyObject_Format(s, f)) +#elif CYTHON_USE_TYPE_SLOTS + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + likely(PyLong_CheckExact(s)) ? PyLong_Type.tp_repr(s) :\ + likely(PyFloat_CheckExact(s)) ? PyFloat_Type.tp_repr(s) :\ + PyObject_Format(s, f)) +#else + #define __Pyx_PyObject_FormatSimple(s, f) (\ + likely(PyUnicode_CheckExact(s)) ? (Py_INCREF(s), s) :\ + PyObject_Format(s, f)) +#endif + +CYTHON_UNUSED static int __pyx_array_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *); /*proto*/ +/* GetAttr.proto */ +static CYTHON_INLINE PyObject *__Pyx_GetAttr(PyObject *, PyObject *); + +/* GetItemInt.proto */ +#define __Pyx_GetItemInt(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_GetItemInt_Fast(o, (Py_ssize_t)i, is_list, wraparound, boundscheck) :\ + (is_list ? (PyErr_SetString(PyExc_IndexError, "list index out of range"), (PyObject*)NULL) :\ + __Pyx_GetItemInt_Generic(o, to_py_func(i)))) +#define __Pyx_GetItemInt_List(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_GetItemInt_List_Fast(o, (Py_ssize_t)i, wraparound, boundscheck) :\ + (PyErr_SetString(PyExc_IndexError, "list index out of range"), (PyObject*)NULL)) +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_List_Fast(PyObject *o, Py_ssize_t i, + int wraparound, int boundscheck); +#define __Pyx_GetItemInt_Tuple(o, i, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_GetItemInt_Tuple_Fast(o, (Py_ssize_t)i, wraparound, boundscheck) :\ + (PyErr_SetString(PyExc_IndexError, "tuple index out of range"), (PyObject*)NULL)) +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Tuple_Fast(PyObject *o, Py_ssize_t i, + int wraparound, int boundscheck); +static PyObject *__Pyx_GetItemInt_Generic(PyObject *o, PyObject* j); +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Fast(PyObject *o, Py_ssize_t i, + int is_list, int wraparound, int boundscheck); + +/* PyObjectCallOneArg.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg); + +/* ObjectGetItem.proto */ +#if CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PyObject *__Pyx_PyObject_GetItem(PyObject *obj, PyObject *key); +#else +#define __Pyx_PyObject_GetItem(obj, key) PyObject_GetItem(obj, key) +#endif + +/* KeywordStringCheck.proto */ +static int __Pyx_CheckKeywordStrings(PyObject *kw, const char* function_name, int kw_allowed); + +/* DivInt[Py_ssize_t].proto */ +static CYTHON_INLINE Py_ssize_t __Pyx_div_Py_ssize_t(Py_ssize_t, Py_ssize_t); + +/* UnaryNegOverflows.proto */ +#define __Pyx_UNARY_NEG_WOULD_OVERFLOW(x)\ + (((x) < 0) & ((unsigned long)(x) == 0-(unsigned long)(x))) + +/* GetAttr3.proto */ +static CYTHON_INLINE PyObject *__Pyx_GetAttr3(PyObject *, PyObject *, PyObject *); + +/* PyDictVersioning.proto */ +#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS +#define __PYX_DICT_VERSION_INIT ((PY_UINT64_T) -1) +#define __PYX_GET_DICT_VERSION(dict) (((PyDictObject*)(dict))->ma_version_tag) +#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var)\ + (version_var) = __PYX_GET_DICT_VERSION(dict);\ + (cache_var) = (value); +#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP) {\ + static PY_UINT64_T __pyx_dict_version = 0;\ + static PyObject *__pyx_dict_cached_value = NULL;\ + if (likely(__PYX_GET_DICT_VERSION(DICT) == __pyx_dict_version)) {\ + (VAR) = __pyx_dict_cached_value;\ + } else {\ + (VAR) = __pyx_dict_cached_value = (LOOKUP);\ + __pyx_dict_version = __PYX_GET_DICT_VERSION(DICT);\ + }\ +} +static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj); +static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj); +static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version); +#else +#define __PYX_GET_DICT_VERSION(dict) (0) +#define __PYX_UPDATE_DICT_CACHE(dict, value, cache_var, version_var) +#define __PYX_PY_DICT_LOOKUP_IF_MODIFIED(VAR, DICT, LOOKUP) (VAR) = (LOOKUP); +#endif + +/* GetModuleGlobalName.proto */ +#if CYTHON_USE_DICT_VERSIONS +#define __Pyx_GetModuleGlobalName(var, name) do {\ + static PY_UINT64_T __pyx_dict_version = 0;\ + static PyObject *__pyx_dict_cached_value = NULL;\ + (var) = (likely(__pyx_dict_version == __PYX_GET_DICT_VERSION(__pyx_d))) ?\ + (likely(__pyx_dict_cached_value) ? __Pyx_NewRef(__pyx_dict_cached_value) : __Pyx_GetBuiltinName(name)) :\ + __Pyx__GetModuleGlobalName(name, &__pyx_dict_version, &__pyx_dict_cached_value);\ +} while(0) +#define __Pyx_GetModuleGlobalNameUncached(var, name) do {\ + PY_UINT64_T __pyx_dict_version;\ + PyObject *__pyx_dict_cached_value;\ + (var) = __Pyx__GetModuleGlobalName(name, &__pyx_dict_version, &__pyx_dict_cached_value);\ +} while(0) +static PyObject *__Pyx__GetModuleGlobalName(PyObject *name, PY_UINT64_T *dict_version, PyObject **dict_cached_value); +#else +#define __Pyx_GetModuleGlobalName(var, name) (var) = __Pyx__GetModuleGlobalName(name) +#define __Pyx_GetModuleGlobalNameUncached(var, name) (var) = __Pyx__GetModuleGlobalName(name) +static CYTHON_INLINE PyObject *__Pyx__GetModuleGlobalName(PyObject *name); +#endif + +/* AssertionsEnabled.proto */ +#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX < 0x02070600 && !defined(Py_OptimizeFlag) + #define __Pyx_init_assertions_enabled() (0) + #define __pyx_assertions_enabled() (1) +#elif CYTHON_COMPILING_IN_LIMITED_API || (CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030C0000) + static int __pyx_assertions_enabled_flag; + #define __pyx_assertions_enabled() (__pyx_assertions_enabled_flag) + static int __Pyx_init_assertions_enabled(void) { + PyObject *builtins, *debug, *debug_str; + int flag; + builtins = PyEval_GetBuiltins(); + if (!builtins) goto bad; + debug_str = PyUnicode_FromStringAndSize("__debug__", 9); + if (!debug_str) goto bad; + debug = PyObject_GetItem(builtins, debug_str); + Py_DECREF(debug_str); + if (!debug) goto bad; + flag = PyObject_IsTrue(debug); + Py_DECREF(debug); + if (flag == -1) goto bad; + __pyx_assertions_enabled_flag = flag; + return 0; + bad: + __pyx_assertions_enabled_flag = 1; + return -1; + } +#else + #define __Pyx_init_assertions_enabled() (0) + #define __pyx_assertions_enabled() (!Py_OptimizeFlag) +#endif + +/* RaiseTooManyValuesToUnpack.proto */ +static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected); + +/* RaiseNeedMoreValuesToUnpack.proto */ +static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index); + +/* RaiseNoneIterError.proto */ +static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void); + +/* ExtTypeTest.proto */ +static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type); + +/* GetTopmostException.proto */ +#if CYTHON_USE_EXC_INFO_STACK && CYTHON_FAST_THREAD_STATE +static _PyErr_StackItem * __Pyx_PyErr_GetTopmostException(PyThreadState *tstate); +#endif + +/* SaveResetException.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_ExceptionSave(type, value, tb) __Pyx__ExceptionSave(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#define __Pyx_ExceptionReset(type, value, tb) __Pyx__ExceptionReset(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb); +#else +#define __Pyx_ExceptionSave(type, value, tb) PyErr_GetExcInfo(type, value, tb) +#define __Pyx_ExceptionReset(type, value, tb) PyErr_SetExcInfo(type, value, tb) +#endif + +/* GetException.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_GetException(type, value, tb) __Pyx__GetException(__pyx_tstate, type, value, tb) +static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#else +static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb); +#endif + +/* SwapException.proto */ +#if CYTHON_FAST_THREAD_STATE +#define __Pyx_ExceptionSwap(type, value, tb) __Pyx__ExceptionSwap(__pyx_tstate, type, value, tb) +static CYTHON_INLINE void __Pyx__ExceptionSwap(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb); +#else +static CYTHON_INLINE void __Pyx_ExceptionSwap(PyObject **type, PyObject **value, PyObject **tb); +#endif + +/* Import.proto */ +static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level); + +/* ImportDottedModule.proto */ +static PyObject *__Pyx_ImportDottedModule(PyObject *name, PyObject *parts_tuple); +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx_ImportDottedModule_WalkParts(PyObject *module, PyObject *name, PyObject *parts_tuple); +#endif + +/* FastTypeChecks.proto */ +#if CYTHON_COMPILING_IN_CPYTHON +#define __Pyx_TypeCheck(obj, type) __Pyx_IsSubtype(Py_TYPE(obj), (PyTypeObject *)type) +#define __Pyx_TypeCheck2(obj, type1, type2) __Pyx_IsAnySubtype2(Py_TYPE(obj), (PyTypeObject *)type1, (PyTypeObject *)type2) +static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b); +static CYTHON_INLINE int __Pyx_IsAnySubtype2(PyTypeObject *cls, PyTypeObject *a, PyTypeObject *b); +static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches(PyObject *err, PyObject *type); +static CYTHON_INLINE int __Pyx_PyErr_GivenExceptionMatches2(PyObject *err, PyObject *type1, PyObject *type2); +#else +#define __Pyx_TypeCheck(obj, type) PyObject_TypeCheck(obj, (PyTypeObject *)type) +#define __Pyx_TypeCheck2(obj, type1, type2) (PyObject_TypeCheck(obj, (PyTypeObject *)type1) || PyObject_TypeCheck(obj, (PyTypeObject *)type2)) +#define __Pyx_PyErr_GivenExceptionMatches(err, type) PyErr_GivenExceptionMatches(err, type) +#define __Pyx_PyErr_GivenExceptionMatches2(err, type1, type2) (PyErr_GivenExceptionMatches(err, type1) || PyErr_GivenExceptionMatches(err, type2)) +#endif +#define __Pyx_PyErr_ExceptionMatches2(err1, err2) __Pyx_PyErr_GivenExceptionMatches2(__Pyx_PyErr_CurrentExceptionType(), err1, err2) +#define __Pyx_PyException_Check(obj) __Pyx_TypeCheck(obj, PyExc_Exception) + +CYTHON_UNUSED static int __pyx_memoryview_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +/* ListCompAppend.proto */ +#if CYTHON_USE_PYLIST_INTERNALS && CYTHON_ASSUME_SAFE_MACROS +static CYTHON_INLINE int __Pyx_ListComp_Append(PyObject* list, PyObject* x) { + PyListObject* L = (PyListObject*) list; + Py_ssize_t len = Py_SIZE(list); + if (likely(L->allocated > len)) { + Py_INCREF(x); + #if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 + L->ob_item[len] = x; + #else + PyList_SET_ITEM(list, len, x); + #endif + __Pyx_SET_SIZE(list, len + 1); + return 0; + } + return PyList_Append(list, x); +} +#else +#define __Pyx_ListComp_Append(L,x) PyList_Append(L,x) +#endif + +/* PySequenceMultiply.proto */ +#define __Pyx_PySequence_Multiply_Left(mul, seq) __Pyx_PySequence_Multiply(seq, mul) +static CYTHON_INLINE PyObject* __Pyx_PySequence_Multiply(PyObject *seq, Py_ssize_t mul); + +/* SetItemInt.proto */ +#define __Pyx_SetItemInt(o, i, v, type, is_signed, to_py_func, is_list, wraparound, boundscheck)\ + (__Pyx_fits_Py_ssize_t(i, type, is_signed) ?\ + __Pyx_SetItemInt_Fast(o, (Py_ssize_t)i, v, is_list, wraparound, boundscheck) :\ + (is_list ? (PyErr_SetString(PyExc_IndexError, "list assignment index out of range"), -1) :\ + __Pyx_SetItemInt_Generic(o, to_py_func(i), v))) +static int __Pyx_SetItemInt_Generic(PyObject *o, PyObject *j, PyObject *v); +static CYTHON_INLINE int __Pyx_SetItemInt_Fast(PyObject *o, Py_ssize_t i, PyObject *v, + int is_list, int wraparound, int boundscheck); + +/* RaiseUnboundLocalError.proto */ +static CYTHON_INLINE void __Pyx_RaiseUnboundLocalError(const char *varname); + +/* DivInt[long].proto */ +static CYTHON_INLINE long __Pyx_div_long(long, long); + +/* PySequenceContains.proto */ +static CYTHON_INLINE int __Pyx_PySequence_ContainsTF(PyObject* item, PyObject* seq, int eq) { + int result = PySequence_Contains(seq, item); + return unlikely(result < 0) ? result : (result == (eq == Py_EQ)); +} + +/* ImportFrom.proto */ +static PyObject* __Pyx_ImportFrom(PyObject* module, PyObject* name); + +/* HasAttr.proto */ +static CYTHON_INLINE int __Pyx_HasAttr(PyObject *, PyObject *); + +/* IsLittleEndian.proto */ +static CYTHON_INLINE int __Pyx_Is_Little_Endian(void); + +/* BufferFormatCheck.proto */ +static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts); +static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx, + __Pyx_BufFmt_StackElem* stack, + __Pyx_TypeInfo* type); + +/* BufferGetAndValidate.proto */ +#define __Pyx_GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack)\ + ((obj == Py_None || obj == NULL) ?\ + (__Pyx_ZeroBuffer(buf), 0) :\ + __Pyx__GetBufferAndValidate(buf, obj, dtype, flags, nd, cast, stack)) +static int __Pyx__GetBufferAndValidate(Py_buffer* buf, PyObject* obj, + __Pyx_TypeInfo* dtype, int flags, int nd, int cast, __Pyx_BufFmt_StackElem* stack); +static void __Pyx_ZeroBuffer(Py_buffer* buf); +static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info); +static Py_ssize_t __Pyx_minusones[] = { -1, -1, -1, -1, -1, -1, -1, -1 }; +static Py_ssize_t __Pyx_zeros[] = { 0, 0, 0, 0, 0, 0, 0, 0 }; + +/* BufferFallbackError.proto */ +static void __Pyx_RaiseBufferFallbackError(void); + +/* ListAppend.proto */ +#if CYTHON_USE_PYLIST_INTERNALS && CYTHON_ASSUME_SAFE_MACROS +static CYTHON_INLINE int __Pyx_PyList_Append(PyObject* list, PyObject* x) { + PyListObject* L = (PyListObject*) list; + Py_ssize_t len = Py_SIZE(list); + if (likely(L->allocated > len) & likely(len > (L->allocated >> 1))) { + Py_INCREF(x); + #if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 + L->ob_item[len] = x; + #else + PyList_SET_ITEM(list, len, x); + #endif + __Pyx_SET_SIZE(list, len + 1); + return 0; + } + return PyList_Append(list, x); +} +#else +#define __Pyx_PyList_Append(L,x) PyList_Append(L,x) +#endif + +/* PyIntBinop.proto */ +#if !CYTHON_COMPILING_IN_PYPY +static PyObject* __Pyx_PyInt_SubtractObjC(PyObject *op1, PyObject *op2, long intval, int inplace, int zerodivision_check); +#else +#define __Pyx_PyInt_SubtractObjC(op1, op2, intval, inplace, zerodivision_check)\ + (inplace ? PyNumber_InPlaceSubtract(op1, op2) : PyNumber_Subtract(op1, op2)) +#endif + +/* SliceObject.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetSlice( + PyObject* obj, Py_ssize_t cstart, Py_ssize_t cstop, + PyObject** py_start, PyObject** py_stop, PyObject** py_slice, + int has_cstart, int has_cstop, int wraparound); + +/* PyObject_GenericGetAttrNoDict.proto */ +#if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static CYTHON_INLINE PyObject* __Pyx_PyObject_GenericGetAttrNoDict(PyObject* obj, PyObject* attr_name); +#else +#define __Pyx_PyObject_GenericGetAttrNoDict PyObject_GenericGetAttr +#endif + +/* PyObject_GenericGetAttr.proto */ +#if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static PyObject* __Pyx_PyObject_GenericGetAttr(PyObject* obj, PyObject* attr_name); +#else +#define __Pyx_PyObject_GenericGetAttr PyObject_GenericGetAttr +#endif + +/* IncludeStructmemberH.proto */ +#include + +/* FixUpExtensionType.proto */ +#if CYTHON_USE_TYPE_SPECS +static int __Pyx_fix_up_extension_type_from_spec(PyType_Spec *spec, PyTypeObject *type); +#endif + +/* PyObjectCallNoArg.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallNoArg(PyObject *func); + +/* PyObjectGetMethod.proto */ +static int __Pyx_PyObject_GetMethod(PyObject *obj, PyObject *name, PyObject **method); + +/* PyObjectCallMethod0.proto */ +static PyObject* __Pyx_PyObject_CallMethod0(PyObject* obj, PyObject* method_name); + +/* ValidateBasesTuple.proto */ +#if CYTHON_COMPILING_IN_CPYTHON || CYTHON_COMPILING_IN_LIMITED_API || CYTHON_USE_TYPE_SPECS +static int __Pyx_validate_bases_tuple(const char *type_name, Py_ssize_t dictoffset, PyObject *bases); +#endif + +/* PyType_Ready.proto */ +CYTHON_UNUSED static int __Pyx_PyType_Ready(PyTypeObject *t); + +/* SetVTable.proto */ +static int __Pyx_SetVtable(PyTypeObject* typeptr , void* vtable); + +/* GetVTable.proto */ +static void* __Pyx_GetVtable(PyTypeObject *type); + +/* MergeVTables.proto */ +#if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_MergeVtables(PyTypeObject *type); +#endif + +/* SetupReduce.proto */ +#if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_setup_reduce(PyObject* type_obj); +#endif + +/* TypeImport.proto */ +#ifndef __PYX_HAVE_RT_ImportType_proto_3_0_12 +#define __PYX_HAVE_RT_ImportType_proto_3_0_12 +#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L +#include +#endif +#if (defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L) || __cplusplus >= 201103L +#define __PYX_GET_STRUCT_ALIGNMENT_3_0_12(s) alignof(s) +#else +#define __PYX_GET_STRUCT_ALIGNMENT_3_0_12(s) sizeof(void*) +#endif +enum __Pyx_ImportType_CheckSize_3_0_12 { + __Pyx_ImportType_CheckSize_Error_3_0_12 = 0, + __Pyx_ImportType_CheckSize_Warn_3_0_12 = 1, + __Pyx_ImportType_CheckSize_Ignore_3_0_12 = 2 +}; +static PyTypeObject *__Pyx_ImportType_3_0_12(PyObject* module, const char *module_name, const char *class_name, size_t size, size_t alignment, enum __Pyx_ImportType_CheckSize_3_0_12 check_size); +#endif + +/* FetchSharedCythonModule.proto */ +static PyObject *__Pyx_FetchSharedCythonABIModule(void); + +/* FetchCommonType.proto */ +#if !CYTHON_USE_TYPE_SPECS +static PyTypeObject* __Pyx_FetchCommonType(PyTypeObject* type); +#else +static PyTypeObject* __Pyx_FetchCommonTypeFromSpec(PyObject *module, PyType_Spec *spec, PyObject *bases); +#endif + +/* PyMethodNew.proto */ +#if CYTHON_COMPILING_IN_LIMITED_API +static PyObject *__Pyx_PyMethod_New(PyObject *func, PyObject *self, PyObject *typ) { + PyObject *typesModule=NULL, *methodType=NULL, *result=NULL; + CYTHON_UNUSED_VAR(typ); + if (!self) + return __Pyx_NewRef(func); + typesModule = PyImport_ImportModule("types"); + if (!typesModule) return NULL; + methodType = PyObject_GetAttrString(typesModule, "MethodType"); + Py_DECREF(typesModule); + if (!methodType) return NULL; + result = PyObject_CallFunctionObjArgs(methodType, func, self, NULL); + Py_DECREF(methodType); + return result; +} +#elif PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx_PyMethod_New(PyObject *func, PyObject *self, PyObject *typ) { + CYTHON_UNUSED_VAR(typ); + if (!self) + return __Pyx_NewRef(func); + return PyMethod_New(func, self); +} +#else + #define __Pyx_PyMethod_New PyMethod_New +#endif + +/* PyVectorcallFastCallDict.proto */ +#if CYTHON_METH_FASTCALL +static CYTHON_INLINE PyObject *__Pyx_PyVectorcall_FastCallDict(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw); +#endif + +/* CythonFunctionShared.proto */ +#define __Pyx_CyFunction_USED +#define __Pyx_CYFUNCTION_STATICMETHOD 0x01 +#define __Pyx_CYFUNCTION_CLASSMETHOD 0x02 +#define __Pyx_CYFUNCTION_CCLASS 0x04 +#define __Pyx_CYFUNCTION_COROUTINE 0x08 +#define __Pyx_CyFunction_GetClosure(f)\ + (((__pyx_CyFunctionObject *) (f))->func_closure) +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + #define __Pyx_CyFunction_GetClassObj(f)\ + (((__pyx_CyFunctionObject *) (f))->func_classobj) +#else + #define __Pyx_CyFunction_GetClassObj(f)\ + ((PyObject*) ((PyCMethodObject *) (f))->mm_class) +#endif +#define __Pyx_CyFunction_SetClassObj(f, classobj)\ + __Pyx__CyFunction_SetClassObj((__pyx_CyFunctionObject *) (f), (classobj)) +#define __Pyx_CyFunction_Defaults(type, f)\ + ((type *)(((__pyx_CyFunctionObject *) (f))->defaults)) +#define __Pyx_CyFunction_SetDefaultsGetter(f, g)\ + ((__pyx_CyFunctionObject *) (f))->defaults_getter = (g) +typedef struct { +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject_HEAD + PyObject *func; +#elif PY_VERSION_HEX < 0x030900B1 + PyCFunctionObject func; +#else + PyCMethodObject func; +#endif +#if CYTHON_BACKPORT_VECTORCALL + __pyx_vectorcallfunc func_vectorcall; +#endif +#if PY_VERSION_HEX < 0x030500A0 || CYTHON_COMPILING_IN_LIMITED_API + PyObject *func_weakreflist; +#endif + PyObject *func_dict; + PyObject *func_name; + PyObject *func_qualname; + PyObject *func_doc; + PyObject *func_globals; + PyObject *func_code; + PyObject *func_closure; +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + PyObject *func_classobj; +#endif + void *defaults; + int defaults_pyobjects; + size_t defaults_size; + int flags; + PyObject *defaults_tuple; + PyObject *defaults_kwdict; + PyObject *(*defaults_getter)(PyObject *); + PyObject *func_annotations; + PyObject *func_is_coroutine; +} __pyx_CyFunctionObject; +#undef __Pyx_CyOrPyCFunction_Check +#define __Pyx_CyFunction_Check(obj) __Pyx_TypeCheck(obj, __pyx_CyFunctionType) +#define __Pyx_CyOrPyCFunction_Check(obj) __Pyx_TypeCheck2(obj, __pyx_CyFunctionType, &PyCFunction_Type) +#define __Pyx_CyFunction_CheckExact(obj) __Pyx_IS_TYPE(obj, __pyx_CyFunctionType) +static CYTHON_INLINE int __Pyx__IsSameCyOrCFunction(PyObject *func, void *cfunc); +#undef __Pyx_IsSameCFunction +#define __Pyx_IsSameCFunction(func, cfunc) __Pyx__IsSameCyOrCFunction(func, cfunc) +static PyObject *__Pyx_CyFunction_Init(__pyx_CyFunctionObject* op, PyMethodDef *ml, + int flags, PyObject* qualname, + PyObject *closure, + PyObject *module, PyObject *globals, + PyObject* code); +static CYTHON_INLINE void __Pyx__CyFunction_SetClassObj(__pyx_CyFunctionObject* f, PyObject* classobj); +static CYTHON_INLINE void *__Pyx_CyFunction_InitDefaults(PyObject *m, + size_t size, + int pyobjects); +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsTuple(PyObject *m, + PyObject *tuple); +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsKwDict(PyObject *m, + PyObject *dict); +static CYTHON_INLINE void __Pyx_CyFunction_SetAnnotationsDict(PyObject *m, + PyObject *dict); +static int __pyx_CyFunction_init(PyObject *module); +#if CYTHON_METH_FASTCALL +static PyObject * __Pyx_CyFunction_Vectorcall_NOARGS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +static PyObject * __Pyx_CyFunction_Vectorcall_O(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS_METHOD(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames); +#if CYTHON_BACKPORT_VECTORCALL +#define __Pyx_CyFunction_func_vectorcall(f) (((__pyx_CyFunctionObject*)f)->func_vectorcall) +#else +#define __Pyx_CyFunction_func_vectorcall(f) (((PyCFunctionObject*)f)->vectorcall) +#endif +#endif + +/* CythonFunction.proto */ +static PyObject *__Pyx_CyFunction_New(PyMethodDef *ml, + int flags, PyObject* qualname, + PyObject *closure, + PyObject *module, PyObject *globals, + PyObject* code); + +/* CLineInTraceback.proto */ +#ifdef CYTHON_CLINE_IN_TRACEBACK +#define __Pyx_CLineForTraceback(tstate, c_line) (((CYTHON_CLINE_IN_TRACEBACK)) ? c_line : 0) +#else +static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line); +#endif + +/* CodeObjectCache.proto */ +#if !CYTHON_COMPILING_IN_LIMITED_API +typedef struct { + PyCodeObject* code_object; + int code_line; +} __Pyx_CodeObjectCacheEntry; +struct __Pyx_CodeObjectCache { + int count; + int max_count; + __Pyx_CodeObjectCacheEntry* entries; +}; +static struct __Pyx_CodeObjectCache __pyx_code_cache = {0,0,NULL}; +static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line); +static PyCodeObject *__pyx_find_code_object(int code_line); +static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object); +#endif + +/* AddTraceback.proto */ +static void __Pyx_AddTraceback(const char *funcname, int c_line, + int py_line, const char *filename); + +#if PY_MAJOR_VERSION < 3 + static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags); + static void __Pyx_ReleaseBuffer(Py_buffer *view); +#else + #define __Pyx_GetBuffer PyObject_GetBuffer + #define __Pyx_ReleaseBuffer PyBuffer_Release +#endif + + +/* BufferStructDeclare.proto */ +typedef struct { + Py_ssize_t shape, strides, suboffsets; +} __Pyx_Buf_DimInfo; +typedef struct { + size_t refcount; + Py_buffer pybuffer; +} __Pyx_Buffer; +typedef struct { + __Pyx_Buffer *rcbuffer; + char *data; + __Pyx_Buf_DimInfo diminfo[8]; +} __Pyx_LocalBuf_ND; + +/* MemviewSliceIsContig.proto */ +static int __pyx_memviewslice_is_contig(const __Pyx_memviewslice mvs, char order, int ndim); + +/* OverlappingSlices.proto */ +static int __pyx_slices_overlap(__Pyx_memviewslice *slice1, + __Pyx_memviewslice *slice2, + int ndim, size_t itemsize); + +/* MemviewDtypeToObject.proto */ +static CYTHON_INLINE PyObject *__pyx_memview_get_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(const char *itemp); +static CYTHON_INLINE int __pyx_memview_set_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(const char *itemp, PyObject *obj); + +/* TypeInfoCompare.proto */ +static int __pyx_typeinfo_cmp(__Pyx_TypeInfo *a, __Pyx_TypeInfo *b); + +/* MemviewSliceValidateAndInit.proto */ +static int __Pyx_ValidateAndInit_memviewslice( + int *axes_specs, + int c_or_f_flag, + int buf_flags, + int ndim, + __Pyx_TypeInfo *dtype, + __Pyx_BufFmt_StackElem stack[], + __Pyx_memviewslice *memviewslice, + PyObject *original_obj); + +/* ObjectToMemviewSlice.proto */ +static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(PyObject *, int writable_flag); + +/* ObjectToMemviewSlice.proto */ +static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(PyObject *, int writable_flag); + +/* RealImag.proto */ +#if CYTHON_CCOMPLEX + #ifdef __cplusplus + #define __Pyx_CREAL(z) ((z).real()) + #define __Pyx_CIMAG(z) ((z).imag()) + #else + #define __Pyx_CREAL(z) (__real__(z)) + #define __Pyx_CIMAG(z) (__imag__(z)) + #endif +#else + #define __Pyx_CREAL(z) ((z).real) + #define __Pyx_CIMAG(z) ((z).imag) +#endif +#if defined(__cplusplus) && CYTHON_CCOMPLEX\ + && (defined(_WIN32) || defined(__clang__) || (defined(__GNUC__) && (__GNUC__ >= 5 || __GNUC__ == 4 && __GNUC_MINOR__ >= 4 )) || __cplusplus >= 201103) + #define __Pyx_SET_CREAL(z,x) ((z).real(x)) + #define __Pyx_SET_CIMAG(z,y) ((z).imag(y)) +#else + #define __Pyx_SET_CREAL(z,x) __Pyx_CREAL(z) = (x) + #define __Pyx_SET_CIMAG(z,y) __Pyx_CIMAG(z) = (y) +#endif + +/* Arithmetic.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #define __Pyx_c_eq_float(a, b) ((a)==(b)) + #define __Pyx_c_sum_float(a, b) ((a)+(b)) + #define __Pyx_c_diff_float(a, b) ((a)-(b)) + #define __Pyx_c_prod_float(a, b) ((a)*(b)) + #define __Pyx_c_quot_float(a, b) ((a)/(b)) + #define __Pyx_c_neg_float(a) (-(a)) + #ifdef __cplusplus + #define __Pyx_c_is_zero_float(z) ((z)==(float)0) + #define __Pyx_c_conj_float(z) (::std::conj(z)) + #if 1 + #define __Pyx_c_abs_float(z) (::std::abs(z)) + #define __Pyx_c_pow_float(a, b) (::std::pow(a, b)) + #endif + #else + #define __Pyx_c_is_zero_float(z) ((z)==0) + #define __Pyx_c_conj_float(z) (conjf(z)) + #if 1 + #define __Pyx_c_abs_float(z) (cabsf(z)) + #define __Pyx_c_pow_float(a, b) (cpowf(a, b)) + #endif + #endif +#else + static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex, __pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex); + static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex); + #if 1 + static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex); + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex, __pyx_t_float_complex); + #endif +#endif + +/* Arithmetic.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #define __Pyx_c_eq_double(a, b) ((a)==(b)) + #define __Pyx_c_sum_double(a, b) ((a)+(b)) + #define __Pyx_c_diff_double(a, b) ((a)-(b)) + #define __Pyx_c_prod_double(a, b) ((a)*(b)) + #define __Pyx_c_quot_double(a, b) ((a)/(b)) + #define __Pyx_c_neg_double(a) (-(a)) + #ifdef __cplusplus + #define __Pyx_c_is_zero_double(z) ((z)==(double)0) + #define __Pyx_c_conj_double(z) (::std::conj(z)) + #if 1 + #define __Pyx_c_abs_double(z) (::std::abs(z)) + #define __Pyx_c_pow_double(a, b) (::std::pow(a, b)) + #endif + #else + #define __Pyx_c_is_zero_double(z) ((z)==0) + #define __Pyx_c_conj_double(z) (conj(z)) + #if 1 + #define __Pyx_c_abs_double(z) (cabs(z)) + #define __Pyx_c_pow_double(a, b) (cpow(a, b)) + #endif + #endif +#else + static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex, __pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex); + static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex); + #if 1 + static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex); + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex, __pyx_t_double_complex); + #endif +#endif + +/* Arithmetic.proto */ +#if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #define __Pyx_c_eq_long__double(a, b) ((a)==(b)) + #define __Pyx_c_sum_long__double(a, b) ((a)+(b)) + #define __Pyx_c_diff_long__double(a, b) ((a)-(b)) + #define __Pyx_c_prod_long__double(a, b) ((a)*(b)) + #define __Pyx_c_quot_long__double(a, b) ((a)/(b)) + #define __Pyx_c_neg_long__double(a) (-(a)) + #ifdef __cplusplus + #define __Pyx_c_is_zero_long__double(z) ((z)==(long double)0) + #define __Pyx_c_conj_long__double(z) (::std::conj(z)) + #if 1 + #define __Pyx_c_abs_long__double(z) (::std::abs(z)) + #define __Pyx_c_pow_long__double(a, b) (::std::pow(a, b)) + #endif + #else + #define __Pyx_c_is_zero_long__double(z) ((z)==0) + #define __Pyx_c_conj_long__double(z) (conjl(z)) + #if 1 + #define __Pyx_c_abs_long__double(z) (cabsl(z)) + #define __Pyx_c_pow_long__double(a, b) (cpowl(a, b)) + #endif + #endif +#else + static CYTHON_INLINE int __Pyx_c_eq_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_sum_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_diff_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_prod_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_quot_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_neg_long__double(__pyx_t_long_double_complex); + static CYTHON_INLINE int __Pyx_c_is_zero_long__double(__pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_conj_long__double(__pyx_t_long_double_complex); + #if 1 + static CYTHON_INLINE long double __Pyx_c_abs_long__double(__pyx_t_long_double_complex); + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_pow_long__double(__pyx_t_long_double_complex, __pyx_t_long_double_complex); + #endif +#endif + +/* MemviewSliceCopyTemplate.proto */ +static __Pyx_memviewslice +__pyx_memoryview_copy_new_contig(const __Pyx_memviewslice *from_mvs, + const char *mode, int ndim, + size_t sizeof_dtype, int contig_flag, + int dtype_is_object); + +/* MemviewSliceInit.proto */ +#define __Pyx_BUF_MAX_NDIMS %(BUF_MAX_NDIMS)d +#define __Pyx_MEMVIEW_DIRECT 1 +#define __Pyx_MEMVIEW_PTR 2 +#define __Pyx_MEMVIEW_FULL 4 +#define __Pyx_MEMVIEW_CONTIG 8 +#define __Pyx_MEMVIEW_STRIDED 16 +#define __Pyx_MEMVIEW_FOLLOW 32 +#define __Pyx_IS_C_CONTIG 1 +#define __Pyx_IS_F_CONTIG 2 +static int __Pyx_init_memviewslice( + struct __pyx_memoryview_obj *memview, + int ndim, + __Pyx_memviewslice *memviewslice, + int memview_is_new_reference); +static CYTHON_INLINE int __pyx_add_acquisition_count_locked( + __pyx_atomic_int_type *acquisition_count, PyThread_type_lock lock); +static CYTHON_INLINE int __pyx_sub_acquisition_count_locked( + __pyx_atomic_int_type *acquisition_count, PyThread_type_lock lock); +#define __pyx_get_slice_count_pointer(memview) (&memview->acquisition_count) +#define __PYX_INC_MEMVIEW(slice, have_gil) __Pyx_INC_MEMVIEW(slice, have_gil, __LINE__) +#define __PYX_XCLEAR_MEMVIEW(slice, have_gil) __Pyx_XCLEAR_MEMVIEW(slice, have_gil, __LINE__) +static CYTHON_INLINE void __Pyx_INC_MEMVIEW(__Pyx_memviewslice *, int, int); +static CYTHON_INLINE void __Pyx_XCLEAR_MEMVIEW(__Pyx_memviewslice *, int, int); + +/* CIntToPy.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyInt_From_npy_int64(npy_int64 value); + +/* CIntFromPy.proto */ +static CYTHON_INLINE npy_int64 __Pyx_PyInt_As_npy_int64(PyObject *); + +/* CIntFromPy.proto */ +static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *); + +/* CIntFromPy.proto */ +static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *); + +/* CIntToPy.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value); + +/* None.proto */ +#include + +/* CIntToPy.proto */ +static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value); + +/* CIntFromPy.proto */ +static CYTHON_INLINE char __Pyx_PyInt_As_char(PyObject *); + +/* FormatTypeName.proto */ +#if CYTHON_COMPILING_IN_LIMITED_API +typedef PyObject *__Pyx_TypeName; +#define __Pyx_FMT_TYPENAME "%U" +static __Pyx_TypeName __Pyx_PyType_GetName(PyTypeObject* tp); +#define __Pyx_DECREF_TypeName(obj) Py_XDECREF(obj) +#else +typedef const char *__Pyx_TypeName; +#define __Pyx_FMT_TYPENAME "%.200s" +#define __Pyx_PyType_GetName(tp) ((tp)->tp_name) +#define __Pyx_DECREF_TypeName(obj) +#endif + +/* CheckBinaryVersion.proto */ +static unsigned long __Pyx_get_runtime_version(void); +static int __Pyx_check_binary_version(unsigned long ct_version, unsigned long rt_version, int allow_newer); + +/* InitStrings.proto */ +static int __Pyx_InitStrings(__Pyx_StringTabEntry *t); + +/* #### Code section: module_declarations ### */ +static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self); /* proto*/ +static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index); /* proto*/ +static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj); /* proto*/ +static PyObject *__pyx_memoryview_setitem_slice_assignment(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_dst, PyObject *__pyx_v_src); /* proto*/ +static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memoryview_obj *__pyx_v_self, struct __pyx_memoryview_obj *__pyx_v_dst, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryview_setitem_indexed(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp); /* proto*/ +static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryview__get_base(struct __pyx_memoryview_obj *__pyx_v_self); /* proto*/ +static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp); /* proto*/ +static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value); /* proto*/ +static PyObject *__pyx_memoryviewslice__get_base(struct __pyx_memoryviewslice_obj *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_8itemsize_itemsize(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_9alignment_alignment(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_6fields_fields(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_5names_names(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyArray_ArrayDescr *__pyx_f_5numpy_5dtype_8subarray_subarray(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_uint64 __pyx_f_5numpy_5dtype_5flags_flags(PyArray_Descr *__pyx_v_self); /* proto*/ +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_7numiter_numiter(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_4size_size(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_5index_index(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_2nd_nd(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_9broadcast_10dimensions_dimensions(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE void **__pyx_f_5numpy_9broadcast_5iters_iters(PyArrayMultiIterObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyObject *__pyx_f_5numpy_7ndarray_4base_base(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE PyArray_Descr *__pyx_f_5numpy_7ndarray_5descr_descr(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE int __pyx_f_5numpy_7ndarray_4ndim_ndim(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_5shape_shape(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_7strides_strides(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE npy_intp __pyx_f_5numpy_7ndarray_4size_size(PyArrayObject *__pyx_v_self); /* proto*/ +static CYTHON_INLINE char *__pyx_f_5numpy_7ndarray_4data_data(PyArrayObject *__pyx_v_self); /* proto*/ +static PyObject *__pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_reset(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self); /* proto*/ +static int __pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_step(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_i); /* proto*/ +static PyObject *__pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_seek(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_i); /* proto*/ + +/* Module declarations from "libc.math" */ + +/* Module declarations from "cython.view" */ + +/* Module declarations from "cython.dataclasses" */ + +/* Module declarations from "cython" */ + +/* Module declarations from "libc.string" */ + +/* Module declarations from "libc.stdio" */ + +/* Module declarations from "__builtin__" */ + +/* Module declarations from "cpython.type" */ + +/* Module declarations from "cpython" */ + +/* Module declarations from "cpython.object" */ + +/* Module declarations from "cpython.ref" */ + +/* Module declarations from "numpy" */ + +/* Module declarations from "numpy" */ + +/* Module declarations from "fairseq.data.token_block_utils_fast" */ +static PyObject *__pyx_collections_abc_Sequence = 0; +static PyObject *generic = 0; +static PyObject *strided = 0; +static PyObject *indirect = 0; +static PyObject *contiguous = 0; +static PyObject *indirect_contiguous = 0; +static int __pyx_memoryview_thread_locks_used; +static PyThread_type_lock __pyx_memoryview_thread_locks[8]; +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__get_slice_indices_none_mode(PyArrayObject *, int); /*proto*/ +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__fast_convert_to_np_array(PyObject *); /*proto*/ +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__get_slice_indices_fast(PyArrayObject *, PyObject *, int, int, int __pyx_skip_dispatch); /*proto*/ +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__get_block_to_dataset_index_fast(PyArrayObject *, PyArrayObject *, int __pyx_skip_dispatch); /*proto*/ +static PyObject *__pyx_f_7fairseq_4data_22token_block_utils_fast___pyx_unpickle_DatasetSearcher__set_state(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *, PyObject *); /*proto*/ +static int __pyx_array_allocate_buffer(struct __pyx_array_obj *); /*proto*/ +static struct __pyx_array_obj *__pyx_array_new(PyObject *, Py_ssize_t, char *, char *, char *); /*proto*/ +static PyObject *__pyx_memoryview_new(PyObject *, int, int, __Pyx_TypeInfo *); /*proto*/ +static CYTHON_INLINE int __pyx_memoryview_check(PyObject *); /*proto*/ +static PyObject *_unellipsify(PyObject *, int); /*proto*/ +static int assert_direct_dimensions(Py_ssize_t *, int); /*proto*/ +static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_obj *, PyObject *); /*proto*/ +static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *, Py_ssize_t, Py_ssize_t, Py_ssize_t, int, int, int *, Py_ssize_t, Py_ssize_t, Py_ssize_t, int, int, int, int); /*proto*/ +static char *__pyx_pybuffer_index(Py_buffer *, char *, Py_ssize_t, Py_ssize_t); /*proto*/ +static int __pyx_memslice_transpose(__Pyx_memviewslice *); /*proto*/ +static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice, int, PyObject *(*)(char *), int (*)(char *, PyObject *), int); /*proto*/ +static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __pyx_memoryview_obj *, __Pyx_memviewslice *); /*proto*/ +static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *, __Pyx_memviewslice *); /*proto*/ +static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *); /*proto*/ +static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview_obj *, __Pyx_memviewslice *); /*proto*/ +static Py_ssize_t abs_py_ssize_t(Py_ssize_t); /*proto*/ +static char __pyx_get_best_slice_order(__Pyx_memviewslice *, int); /*proto*/ +static void _copy_strided_to_strided(char *, Py_ssize_t *, char *, Py_ssize_t *, Py_ssize_t *, Py_ssize_t *, int, size_t); /*proto*/ +static void copy_strided_to_strided(__Pyx_memviewslice *, __Pyx_memviewslice *, int, size_t); /*proto*/ +static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *, int); /*proto*/ +static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *, Py_ssize_t *, Py_ssize_t, int, char); /*proto*/ +static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *, __Pyx_memviewslice *, char, int); /*proto*/ +static int __pyx_memoryview_err_extents(int, Py_ssize_t, Py_ssize_t); /*proto*/ +static int __pyx_memoryview_err_dim(PyObject *, PyObject *, int); /*proto*/ +static int __pyx_memoryview_err(PyObject *, PyObject *); /*proto*/ +static int __pyx_memoryview_err_no_memory(void); /*proto*/ +static int __pyx_memoryview_copy_contents(__Pyx_memviewslice, __Pyx_memviewslice, int, int, int); /*proto*/ +static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *, int, int); /*proto*/ +static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *, int, int, int); /*proto*/ +static void __pyx_memoryview_refcount_objects_in_slice_with_gil(char *, Py_ssize_t *, Py_ssize_t *, int, int); /*proto*/ +static void __pyx_memoryview_refcount_objects_in_slice(char *, Py_ssize_t *, Py_ssize_t *, int, int); /*proto*/ +static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *, int, size_t, void *, int); /*proto*/ +static void __pyx_memoryview__slice_assign_scalar(char *, Py_ssize_t *, Py_ssize_t *, int, size_t, void *); /*proto*/ +static PyObject *__pyx_unpickle_Enum__set_state(struct __pyx_MemviewEnum_obj *, PyObject *); /*proto*/ +/* #### Code section: typeinfo ### */ +static __Pyx_TypeInfo __Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t = { "DTYPE_t", NULL, sizeof(__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t), { 0 }, 0, __PYX_IS_UNSIGNED(__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t) ? 'U' : 'I', __PYX_IS_UNSIGNED(__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t), 0 }; +/* #### Code section: before_global_var ### */ +#define __Pyx_MODULE_NAME "fairseq.data.token_block_utils_fast" +extern int __pyx_module_is_main_fairseq__data__token_block_utils_fast; +int __pyx_module_is_main_fairseq__data__token_block_utils_fast = 0; + +/* Implementation of "fairseq.data.token_block_utils_fast" */ +/* #### Code section: global_var ### */ +static PyObject *__pyx_builtin_range; +static PyObject *__pyx_builtin_ValueError; +static PyObject *__pyx_builtin_AssertionError; +static PyObject *__pyx_builtin___import__; +static PyObject *__pyx_builtin_MemoryError; +static PyObject *__pyx_builtin_enumerate; +static PyObject *__pyx_builtin_TypeError; +static PyObject *__pyx_builtin_Ellipsis; +static PyObject *__pyx_builtin_id; +static PyObject *__pyx_builtin_IndexError; +static PyObject *__pyx_builtin_ImportError; +/* #### Code section: string_decls ### */ +static const char __pyx_k_[] = ": "; +static const char __pyx_k_O[] = "O"; +static const char __pyx_k_c[] = "c"; +static const char __pyx_k__2[] = "."; +static const char __pyx_k__3[] = "*"; +static const char __pyx_k__6[] = "'"; +static const char __pyx_k__7[] = ")"; +static const char __pyx_k_gc[] = "gc"; +static const char __pyx_k_id[] = "id"; +static const char __pyx_k_np[] = "np"; +static const char __pyx_k__35[] = "?"; +static const char __pyx_k_abc[] = "abc"; +static const char __pyx_k_and[] = " and "; +static const char __pyx_k_eos[] = "eos"; +static const char __pyx_k_got[] = " (got "; +static const char __pyx_k_new[] = "__new__"; +static const char __pyx_k_obj[] = "obj"; +static const char __pyx_k_sum[] = "sum"; +static const char __pyx_k_sys[] = "sys"; +static const char __pyx_k_axis[] = "axis"; +static const char __pyx_k_base[] = "base"; +static const char __pyx_k_dict[] = "__dict__"; +static const char __pyx_k_main[] = "__main__"; +static const char __pyx_k_mode[] = "mode"; +static const char __pyx_k_name[] = "name"; +static const char __pyx_k_ndim[] = "ndim"; +static const char __pyx_k_none[] = "none"; +static const char __pyx_k_pack[] = "pack"; +static const char __pyx_k_self[] = "self"; +static const char __pyx_k_size[] = "size"; +static const char __pyx_k_spec[] = "__spec__"; +static const char __pyx_k_step[] = "step"; +static const char __pyx_k_stop[] = "stop"; +static const char __pyx_k_test[] = "__test__"; +static const char __pyx_k_ASCII[] = "ASCII"; +static const char __pyx_k_DTYPE[] = "DTYPE"; +static const char __pyx_k_chain[] = "chain"; +static const char __pyx_k_class[] = "__class__"; +static const char __pyx_k_count[] = "count"; +static const char __pyx_k_dtype[] = "dtype"; +static const char __pyx_k_error[] = "error"; +static const char __pyx_k_flags[] = "flags"; +static const char __pyx_k_index[] = "index"; +static const char __pyx_k_int64[] = "int64"; +static const char __pyx_k_numpy[] = "numpy"; +static const char __pyx_k_range[] = "range"; +static const char __pyx_k_shape[] = "shape"; +static const char __pyx_k_sizes[] = "sizes"; +static const char __pyx_k_start[] = "start"; +static const char __pyx_k_state[] = "state"; +static const char __pyx_k_torch[] = "torch"; +static const char __pyx_k_zeros[] = "zeros"; +static const char __pyx_k_cumsum[] = "cumsum"; +static const char __pyx_k_dict_2[] = "_dict"; +static const char __pyx_k_enable[] = "enable"; +static const char __pyx_k_encode[] = "encode"; +static const char __pyx_k_format[] = "format"; +static const char __pyx_k_import[] = "__import__"; +static const char __pyx_k_name_2[] = "__name__"; +static const char __pyx_k_pickle[] = "pickle"; +static const char __pyx_k_reduce[] = "__reduce__"; +static const char __pyx_k_struct[] = "struct"; +static const char __pyx_k_unpack[] = "unpack"; +static const char __pyx_k_update[] = "update"; +static const char __pyx_k_disable[] = "disable"; +static const char __pyx_k_fortran[] = "fortran"; +static const char __pyx_k_memview[] = "memview"; +static const char __pyx_k_reshape[] = "reshape"; +static const char __pyx_k_Ellipsis[] = "Ellipsis"; +static const char __pyx_k_Sequence[] = "Sequence"; +static const char __pyx_k_complete[] = "complete"; +static const char __pyx_k_fromiter[] = "fromiter"; +static const char __pyx_k_getstate[] = "__getstate__"; +static const char __pyx_k_itemsize[] = "itemsize"; +static const char __pyx_k_pyx_type[] = "__pyx_type"; +static const char __pyx_k_register[] = "register"; +static const char __pyx_k_setstate[] = "__setstate__"; +static const char __pyx_k_TypeError[] = "TypeError"; +static const char __pyx_k_enumerate[] = "enumerate"; +static const char __pyx_k_isenabled[] = "isenabled"; +static const char __pyx_k_itertools[] = "itertools"; +static const char __pyx_k_pyx_state[] = "__pyx_state"; +static const char __pyx_k_reduce_ex[] = "__reduce_ex__"; +static const char __pyx_k_IndexError[] = "IndexError"; +static const char __pyx_k_ValueError[] = "ValueError"; +static const char __pyx_k_block_size[] = "block_size"; +static const char __pyx_k_break_mode[] = "break_mode"; +static const char __pyx_k_pyx_result[] = "__pyx_result"; +static const char __pyx_k_pyx_vtable[] = "__pyx_vtable__"; +static const char __pyx_k_ImportError[] = "ImportError"; +static const char __pyx_k_MemoryError[] = "MemoryError"; +static const char __pyx_k_PickleError[] = "PickleError"; +static const char __pyx_k_collections[] = "collections"; +static const char __pyx_k_complete_doc[] = "complete_doc"; +static const char __pyx_k_initializing[] = "_initializing"; +static const char __pyx_k_is_coroutine[] = "_is_coroutine"; +static const char __pyx_k_pyx_checksum[] = "__pyx_checksum"; +static const char __pyx_k_stringsource[] = ""; +static const char __pyx_k_use_setstate[] = "use_setstate"; +static const char __pyx_k_version_info[] = "version_info"; +static const char __pyx_k_class_getitem[] = "__class_getitem__"; +static const char __pyx_k_from_iterable[] = "from_iterable"; +static const char __pyx_k_reduce_cython[] = "__reduce_cython__"; +static const char __pyx_k_slice_indices[] = "slice_indices"; +static const char __pyx_k_AssertionError[] = "AssertionError"; +static const char __pyx_k_DatasetSearcher[] = "DatasetSearcher"; +static const char __pyx_k_View_MemoryView[] = "View.MemoryView"; +static const char __pyx_k_allocate_buffer[] = "allocate_buffer"; +static const char __pyx_k_collections_abc[] = "collections.abc"; +static const char __pyx_k_dtype_is_object[] = "dtype_is_object"; +static const char __pyx_k_pyx_PickleError[] = "__pyx_PickleError"; +static const char __pyx_k_setstate_cython[] = "__setstate_cython__"; +static const char __pyx_k_document_sep_len[] = "document_sep_len"; +static const char __pyx_k_pyx_unpickle_Enum[] = "__pyx_unpickle_Enum"; +static const char __pyx_k_Invalid_break_mode[] = "Invalid break_mode: "; +static const char __pyx_k_asyncio_coroutines[] = "asyncio.coroutines"; +static const char __pyx_k_cline_in_traceback[] = "cline_in_traceback"; +static const char __pyx_k_strided_and_direct[] = ""; +static const char __pyx_k_strided_and_indirect[] = ""; +static const char __pyx_k_Invalid_shape_in_axis[] = "Invalid shape in axis "; +static const char __pyx_k_contiguous_and_direct[] = ""; +static const char __pyx_k_Cannot_index_with_type[] = "Cannot index with type '"; +static const char __pyx_k_MemoryView_of_r_object[] = ""; +static const char __pyx_k_get_slice_indices_fast[] = "_get_slice_indices_fast"; +static const char __pyx_k_MemoryView_of_r_at_0x_x[] = ""; +static const char __pyx_k_contiguous_and_indirect[] = ""; +static const char __pyx_k_Dimension_d_is_not_direct[] = "Dimension %d is not direct"; +static const char __pyx_k_Index_out_of_bounds_axis_d[] = "Index out of bounds (axis %d)"; +static const char __pyx_k_Step_may_not_be_zero_axis_d[] = "Step may not be zero (axis %d)"; +static const char __pyx_k_itemsize_0_for_cython_array[] = "itemsize <= 0 for cython.array"; +static const char __pyx_k_pyx_unpickle_DatasetSearcher[] = "__pyx_unpickle_DatasetSearcher"; +static const char __pyx_k_unable_to_allocate_array_data[] = "unable to allocate array data."; +static const char __pyx_k_strided_and_direct_or_indirect[] = ""; +static const char __pyx_k_DatasetSearcher___reduce_cython[] = "DatasetSearcher.__reduce_cython__"; +static const char __pyx_k_get_block_to_dataset_index_fast[] = "_get_block_to_dataset_index_fast"; +static const char __pyx_k_All_dimensions_preceding_dimensi[] = "All dimensions preceding dimension %d must be indexed and not sliced"; +static const char __pyx_k_Buffer_view_does_not_expose_stri[] = "Buffer view does not expose strides"; +static const char __pyx_k_Can_only_create_a_buffer_that_is[] = "Can only create a buffer that is contiguous in memory."; +static const char __pyx_k_Cannot_assign_to_read_only_memor[] = "Cannot assign to read-only memoryview"; +static const char __pyx_k_Cannot_create_writable_memory_vi[] = "Cannot create writable memory view from read-only memoryview"; +static const char __pyx_k_Cannot_transpose_memoryview_with[] = "Cannot transpose memoryview with indirect dimensions"; +static const char __pyx_k_DatasetSearcher___setstate_cytho[] = "DatasetSearcher.__setstate_cython__"; +static const char __pyx_k_Empty_shape_tuple_for_cython_arr[] = "Empty shape tuple for cython.array"; +static const char __pyx_k_Incompatible_checksums_0x_x_vs_0[] = "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))"; +static const char __pyx_k_Indirect_dimensions_not_supporte[] = "Indirect dimensions not supported"; +static const char __pyx_k_Invalid_mode_expected_c_or_fortr[] = "Invalid mode, expected 'c' or 'fortran', got "; +static const char __pyx_k_Out_of_bounds_on_buffer_access_a[] = "Out of bounds on buffer access (axis "; +static const char __pyx_k_Unable_to_convert_item_to_object[] = "Unable to convert item to object"; +static const char __pyx_k_fairseq_data_token_block_utils_f[] = "fairseq/data/token_block_utils_fast.pyx"; +static const char __pyx_k_got_differing_extents_in_dimensi[] = "got differing extents in dimension "; +static const char __pyx_k_no_default___reduce___due_to_non[] = "no default __reduce__ due to non-trivial __cinit__"; +static const char __pyx_k_numpy__core_multiarray_failed_to[] = "numpy._core.multiarray failed to import"; +static const char __pyx_k_numpy__core_umath_failed_to_impo[] = "numpy._core.umath failed to import"; +static const char __pyx_k_unable_to_allocate_shape_and_str[] = "unable to allocate shape and strides."; +static const char __pyx_k_Incompatible_checksums_0x_x_vs_0_2[] = "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))"; +static const char __pyx_k_fairseq_data_token_block_utils_f_2[] = "fairseq.data.token_block_utils_fast"; +/* #### Code section: decls ### */ +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_shape, Py_ssize_t __pyx_v_itemsize, PyObject *__pyx_v_format, PyObject *__pyx_v_mode, int __pyx_v_allocate_buffer); /* proto */ +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(struct __pyx_array_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /* proto */ +static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struct __pyx_array_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_5array_7memview___get__(struct __pyx_array_obj *__pyx_v_self); /* proto */ +static Py_ssize_t __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(struct __pyx_array_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_attr); /* proto */ +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item); /* proto */ +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_12__setitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value); /* proto */ +static PyObject *__pyx_pf___pyx_array___reduce_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_array_2__setstate_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state); /* proto */ +static int __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v_name); /* proto */ +static PyObject *__pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr__(struct __pyx_MemviewEnum_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_MemviewEnum___reduce_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_MemviewEnum_2__setstate_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v___pyx_state); /* proto */ +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj, int __pyx_v_flags, int __pyx_v_dtype_is_object); /* proto */ +static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__dealloc__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4__getitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index); /* proto */ +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value); /* proto */ +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbuffer__(struct __pyx_memoryview_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_10__len__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12__repr__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14__str__(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16is_c_contig(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18is_f_contig(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20copy(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22copy_fortran(struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryview___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryview_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state); /* proto */ +static void __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewslice___dealloc__(struct __pyx_memoryviewslice_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryviewslice___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf___pyx_memoryviewslice_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state); /* proto */ +static PyObject *__pyx_pf_15View_dot_MemoryView___pyx_unpickle_Enum(CYTHON_UNUSED PyObject *__pyx_self, PyObject *__pyx_v___pyx_type, long __pyx_v___pyx_checksum, PyObject *__pyx_v___pyx_state); /* proto */ +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast__get_slice_indices_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_sizes, PyObject *__pyx_v_break_mode, int __pyx_v_block_size, int __pyx_v_document_sep_len); /* proto */ +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_2_get_block_to_dataset_index_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_sizes, PyArrayObject *__pyx_v_slice_indices); /* proto */ +static int __pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher___init__(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, __Pyx_memviewslice __pyx_v_sizes); /* proto */ +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_2__reduce_cython__(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self); /* proto */ +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_4__setstate_cython__(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, PyObject *__pyx_v___pyx_state); /* proto */ +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_4__pyx_unpickle_DatasetSearcher(CYTHON_UNUSED PyObject *__pyx_self, PyObject *__pyx_v___pyx_type, long __pyx_v___pyx_checksum, PyObject *__pyx_v___pyx_state); /* proto */ +static PyObject *__pyx_tp_new_7fairseq_4data_22token_block_utils_fast_DatasetSearcher(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +static PyObject *__pyx_tp_new_array(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +static PyObject *__pyx_tp_new_Enum(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +static PyObject *__pyx_tp_new_memoryview(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +static PyObject *__pyx_tp_new__memoryviewslice(PyTypeObject *t, PyObject *a, PyObject *k); /*proto*/ +/* #### Code section: late_includes ### */ +/* #### Code section: module_state ### */ +typedef struct { + PyObject *__pyx_d; + PyObject *__pyx_b; + PyObject *__pyx_cython_runtime; + PyObject *__pyx_empty_tuple; + PyObject *__pyx_empty_bytes; + PyObject *__pyx_empty_unicode; + #ifdef __Pyx_CyFunction_USED + PyTypeObject *__pyx_CyFunctionType; + #endif + #ifdef __Pyx_FusedFunction_USED + PyTypeObject *__pyx_FusedFunctionType; + #endif + #ifdef __Pyx_Generator_USED + PyTypeObject *__pyx_GeneratorType; + #endif + #ifdef __Pyx_IterableCoroutine_USED + PyTypeObject *__pyx_IterableCoroutineType; + #endif + #ifdef __Pyx_Coroutine_USED + PyTypeObject *__pyx_CoroutineAwaitType; + #endif + #ifdef __Pyx_Coroutine_USED + PyTypeObject *__pyx_CoroutineType; + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + PyTypeObject *__pyx_ptype_7cpython_4type_type; + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + #if CYTHON_USE_MODULE_STATE + #endif + PyTypeObject *__pyx_ptype_5numpy_dtype; + PyTypeObject *__pyx_ptype_5numpy_flatiter; + PyTypeObject *__pyx_ptype_5numpy_broadcast; + PyTypeObject *__pyx_ptype_5numpy_ndarray; + PyTypeObject *__pyx_ptype_5numpy_generic; + PyTypeObject *__pyx_ptype_5numpy_number; + PyTypeObject *__pyx_ptype_5numpy_integer; + PyTypeObject *__pyx_ptype_5numpy_signedinteger; + PyTypeObject *__pyx_ptype_5numpy_unsignedinteger; + PyTypeObject *__pyx_ptype_5numpy_inexact; + PyTypeObject *__pyx_ptype_5numpy_floating; + PyTypeObject *__pyx_ptype_5numpy_complexfloating; + PyTypeObject *__pyx_ptype_5numpy_flexible; + PyTypeObject *__pyx_ptype_5numpy_character; + PyTypeObject *__pyx_ptype_5numpy_ufunc; + #if CYTHON_USE_MODULE_STATE + PyObject *__pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; + PyObject *__pyx_type___pyx_array; + PyObject *__pyx_type___pyx_MemviewEnum; + PyObject *__pyx_type___pyx_memoryview; + PyObject *__pyx_type___pyx_memoryviewslice; + #endif + PyTypeObject *__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; + PyTypeObject *__pyx_array_type; + PyTypeObject *__pyx_MemviewEnum_type; + PyTypeObject *__pyx_memoryview_type; + PyTypeObject *__pyx_memoryviewslice_type; + PyObject *__pyx_kp_u_; + PyObject *__pyx_n_s_ASCII; + PyObject *__pyx_kp_s_All_dimensions_preceding_dimensi; + PyObject *__pyx_n_s_AssertionError; + PyObject *__pyx_kp_s_Buffer_view_does_not_expose_stri; + PyObject *__pyx_kp_s_Can_only_create_a_buffer_that_is; + PyObject *__pyx_kp_s_Cannot_assign_to_read_only_memor; + PyObject *__pyx_kp_s_Cannot_create_writable_memory_vi; + PyObject *__pyx_kp_u_Cannot_index_with_type; + PyObject *__pyx_kp_s_Cannot_transpose_memoryview_with; + PyObject *__pyx_n_s_DTYPE; + PyObject *__pyx_n_s_DatasetSearcher; + PyObject *__pyx_n_s_DatasetSearcher___reduce_cython; + PyObject *__pyx_n_s_DatasetSearcher___setstate_cytho; + PyObject *__pyx_kp_s_Dimension_d_is_not_direct; + PyObject *__pyx_n_s_Ellipsis; + PyObject *__pyx_kp_s_Empty_shape_tuple_for_cython_arr; + PyObject *__pyx_n_s_ImportError; + PyObject *__pyx_kp_s_Incompatible_checksums_0x_x_vs_0; + PyObject *__pyx_kp_s_Incompatible_checksums_0x_x_vs_0_2; + PyObject *__pyx_n_s_IndexError; + PyObject *__pyx_kp_s_Index_out_of_bounds_axis_d; + PyObject *__pyx_kp_s_Indirect_dimensions_not_supporte; + PyObject *__pyx_kp_u_Invalid_break_mode; + PyObject *__pyx_kp_u_Invalid_mode_expected_c_or_fortr; + PyObject *__pyx_kp_u_Invalid_shape_in_axis; + PyObject *__pyx_n_s_MemoryError; + PyObject *__pyx_kp_s_MemoryView_of_r_at_0x_x; + PyObject *__pyx_kp_s_MemoryView_of_r_object; + PyObject *__pyx_n_b_O; + PyObject *__pyx_kp_u_Out_of_bounds_on_buffer_access_a; + PyObject *__pyx_n_s_PickleError; + PyObject *__pyx_n_s_Sequence; + PyObject *__pyx_kp_s_Step_may_not_be_zero_axis_d; + PyObject *__pyx_n_s_TypeError; + PyObject *__pyx_kp_s_Unable_to_convert_item_to_object; + PyObject *__pyx_n_s_ValueError; + PyObject *__pyx_n_s_View_MemoryView; + PyObject *__pyx_kp_u__2; + PyObject *__pyx_n_s__3; + PyObject *__pyx_n_s__35; + PyObject *__pyx_kp_u__6; + PyObject *__pyx_kp_u__7; + PyObject *__pyx_n_s_abc; + PyObject *__pyx_n_s_allocate_buffer; + PyObject *__pyx_kp_u_and; + PyObject *__pyx_n_s_asyncio_coroutines; + PyObject *__pyx_n_s_axis; + PyObject *__pyx_n_s_base; + PyObject *__pyx_n_s_block_size; + PyObject *__pyx_n_s_break_mode; + PyObject *__pyx_n_s_c; + PyObject *__pyx_n_u_c; + PyObject *__pyx_n_s_chain; + PyObject *__pyx_n_s_class; + PyObject *__pyx_n_s_class_getitem; + PyObject *__pyx_n_s_cline_in_traceback; + PyObject *__pyx_n_s_collections; + PyObject *__pyx_kp_s_collections_abc; + PyObject *__pyx_n_u_complete; + PyObject *__pyx_n_u_complete_doc; + PyObject *__pyx_kp_s_contiguous_and_direct; + PyObject *__pyx_kp_s_contiguous_and_indirect; + PyObject *__pyx_n_s_count; + PyObject *__pyx_n_s_cumsum; + PyObject *__pyx_n_s_dict; + PyObject *__pyx_n_s_dict_2; + PyObject *__pyx_kp_u_disable; + PyObject *__pyx_n_s_document_sep_len; + PyObject *__pyx_n_s_dtype; + PyObject *__pyx_n_s_dtype_is_object; + PyObject *__pyx_kp_u_enable; + PyObject *__pyx_n_s_encode; + PyObject *__pyx_n_s_enumerate; + PyObject *__pyx_n_u_eos; + PyObject *__pyx_n_s_error; + PyObject *__pyx_kp_s_fairseq_data_token_block_utils_f; + PyObject *__pyx_n_s_fairseq_data_token_block_utils_f_2; + PyObject *__pyx_n_s_flags; + PyObject *__pyx_n_s_format; + PyObject *__pyx_n_s_fortran; + PyObject *__pyx_n_u_fortran; + PyObject *__pyx_n_s_from_iterable; + PyObject *__pyx_n_s_fromiter; + PyObject *__pyx_kp_u_gc; + PyObject *__pyx_n_s_get_block_to_dataset_index_fast; + PyObject *__pyx_n_s_get_slice_indices_fast; + PyObject *__pyx_n_s_getstate; + PyObject *__pyx_kp_u_got; + PyObject *__pyx_kp_u_got_differing_extents_in_dimensi; + PyObject *__pyx_n_s_id; + PyObject *__pyx_n_s_import; + PyObject *__pyx_n_s_index; + PyObject *__pyx_n_s_initializing; + PyObject *__pyx_n_s_int64; + PyObject *__pyx_n_s_is_coroutine; + PyObject *__pyx_kp_u_isenabled; + PyObject *__pyx_n_s_itemsize; + PyObject *__pyx_kp_s_itemsize_0_for_cython_array; + PyObject *__pyx_n_s_itertools; + PyObject *__pyx_n_s_main; + PyObject *__pyx_n_s_memview; + PyObject *__pyx_n_s_mode; + PyObject *__pyx_n_s_name; + PyObject *__pyx_n_s_name_2; + PyObject *__pyx_n_s_ndim; + PyObject *__pyx_n_s_new; + PyObject *__pyx_kp_s_no_default___reduce___due_to_non; + PyObject *__pyx_n_u_none; + PyObject *__pyx_n_s_np; + PyObject *__pyx_n_s_numpy; + PyObject *__pyx_kp_u_numpy__core_multiarray_failed_to; + PyObject *__pyx_kp_u_numpy__core_umath_failed_to_impo; + PyObject *__pyx_n_s_obj; + PyObject *__pyx_n_s_pack; + PyObject *__pyx_n_s_pickle; + PyObject *__pyx_n_s_pyx_PickleError; + PyObject *__pyx_n_s_pyx_checksum; + PyObject *__pyx_n_s_pyx_result; + PyObject *__pyx_n_s_pyx_state; + PyObject *__pyx_n_s_pyx_type; + PyObject *__pyx_n_s_pyx_unpickle_DatasetSearcher; + PyObject *__pyx_n_s_pyx_unpickle_Enum; + PyObject *__pyx_n_s_pyx_vtable; + PyObject *__pyx_n_s_range; + PyObject *__pyx_n_s_reduce; + PyObject *__pyx_n_s_reduce_cython; + PyObject *__pyx_n_s_reduce_ex; + PyObject *__pyx_n_s_register; + PyObject *__pyx_n_s_reshape; + PyObject *__pyx_n_s_self; + PyObject *__pyx_n_s_setstate; + PyObject *__pyx_n_s_setstate_cython; + PyObject *__pyx_n_s_shape; + PyObject *__pyx_n_s_size; + PyObject *__pyx_n_s_sizes; + PyObject *__pyx_n_s_slice_indices; + PyObject *__pyx_n_s_spec; + PyObject *__pyx_n_s_start; + PyObject *__pyx_n_s_state; + PyObject *__pyx_n_s_step; + PyObject *__pyx_n_s_stop; + PyObject *__pyx_kp_s_strided_and_direct; + PyObject *__pyx_kp_s_strided_and_direct_or_indirect; + PyObject *__pyx_kp_s_strided_and_indirect; + PyObject *__pyx_kp_s_stringsource; + PyObject *__pyx_n_s_struct; + PyObject *__pyx_n_s_sum; + PyObject *__pyx_n_s_sys; + PyObject *__pyx_n_s_test; + PyObject *__pyx_n_s_torch; + PyObject *__pyx_kp_s_unable_to_allocate_array_data; + PyObject *__pyx_kp_s_unable_to_allocate_shape_and_str; + PyObject *__pyx_n_s_unpack; + PyObject *__pyx_n_s_update; + PyObject *__pyx_n_s_use_setstate; + PyObject *__pyx_n_s_version_info; + PyObject *__pyx_n_s_zeros; + PyObject *__pyx_int_0; + PyObject *__pyx_int_1; + PyObject *__pyx_int_2; + PyObject *__pyx_int_3; + PyObject *__pyx_int_48422178; + PyObject *__pyx_int_107161605; + PyObject *__pyx_int_112105877; + PyObject *__pyx_int_136983863; + PyObject *__pyx_int_147225413; + PyObject *__pyx_int_184977713; + PyObject *__pyx_int_neg_1; + PyObject *__pyx_slice__5; + PyObject *__pyx_tuple__4; + PyObject *__pyx_tuple__8; + PyObject *__pyx_tuple__9; + PyObject *__pyx_slice__11; + PyObject *__pyx_tuple__10; + PyObject *__pyx_tuple__12; + PyObject *__pyx_tuple__13; + PyObject *__pyx_tuple__14; + PyObject *__pyx_tuple__15; + PyObject *__pyx_tuple__16; + PyObject *__pyx_tuple__17; + PyObject *__pyx_tuple__18; + PyObject *__pyx_tuple__19; + PyObject *__pyx_tuple__20; + PyObject *__pyx_tuple__21; + PyObject *__pyx_tuple__22; + PyObject *__pyx_tuple__23; + PyObject *__pyx_tuple__24; + PyObject *__pyx_tuple__26; + PyObject *__pyx_tuple__28; + PyObject *__pyx_tuple__30; + PyObject *__pyx_tuple__32; + PyObject *__pyx_codeobj__25; + PyObject *__pyx_codeobj__27; + PyObject *__pyx_codeobj__29; + PyObject *__pyx_codeobj__31; + PyObject *__pyx_codeobj__33; + PyObject *__pyx_codeobj__34; +} __pyx_mstate; + +#if CYTHON_USE_MODULE_STATE +#ifdef __cplusplus +namespace { + extern struct PyModuleDef __pyx_moduledef; +} /* anonymous namespace */ +#else +static struct PyModuleDef __pyx_moduledef; +#endif + +#define __pyx_mstate(o) ((__pyx_mstate *)__Pyx_PyModule_GetState(o)) + +#define __pyx_mstate_global (__pyx_mstate(PyState_FindModule(&__pyx_moduledef))) + +#define __pyx_m (PyState_FindModule(&__pyx_moduledef)) +#else +static __pyx_mstate __pyx_mstate_global_static = +#ifdef __cplusplus + {}; +#else + {0}; +#endif +static __pyx_mstate *__pyx_mstate_global = &__pyx_mstate_global_static; +#endif +/* #### Code section: module_state_clear ### */ +#if CYTHON_USE_MODULE_STATE +static int __pyx_m_clear(PyObject *m) { + __pyx_mstate *clear_module_state = __pyx_mstate(m); + if (!clear_module_state) return 0; + Py_CLEAR(clear_module_state->__pyx_d); + Py_CLEAR(clear_module_state->__pyx_b); + Py_CLEAR(clear_module_state->__pyx_cython_runtime); + Py_CLEAR(clear_module_state->__pyx_empty_tuple); + Py_CLEAR(clear_module_state->__pyx_empty_bytes); + Py_CLEAR(clear_module_state->__pyx_empty_unicode); + #ifdef __Pyx_CyFunction_USED + Py_CLEAR(clear_module_state->__pyx_CyFunctionType); + #endif + #ifdef __Pyx_FusedFunction_USED + Py_CLEAR(clear_module_state->__pyx_FusedFunctionType); + #endif + Py_CLEAR(clear_module_state->__pyx_ptype_7cpython_4type_type); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_dtype); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_flatiter); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_broadcast); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_ndarray); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_generic); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_number); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_integer); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_signedinteger); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_unsignedinteger); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_inexact); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_floating); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_complexfloating); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_flexible); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_character); + Py_CLEAR(clear_module_state->__pyx_ptype_5numpy_ufunc); + Py_CLEAR(clear_module_state->__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher); + Py_CLEAR(clear_module_state->__pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher); + Py_CLEAR(clear_module_state->__pyx_array_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_array); + Py_CLEAR(clear_module_state->__pyx_MemviewEnum_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_MemviewEnum); + Py_CLEAR(clear_module_state->__pyx_memoryview_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_memoryview); + Py_CLEAR(clear_module_state->__pyx_memoryviewslice_type); + Py_CLEAR(clear_module_state->__pyx_type___pyx_memoryviewslice); + Py_CLEAR(clear_module_state->__pyx_kp_u_); + Py_CLEAR(clear_module_state->__pyx_n_s_ASCII); + Py_CLEAR(clear_module_state->__pyx_kp_s_All_dimensions_preceding_dimensi); + Py_CLEAR(clear_module_state->__pyx_n_s_AssertionError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Buffer_view_does_not_expose_stri); + Py_CLEAR(clear_module_state->__pyx_kp_s_Can_only_create_a_buffer_that_is); + Py_CLEAR(clear_module_state->__pyx_kp_s_Cannot_assign_to_read_only_memor); + Py_CLEAR(clear_module_state->__pyx_kp_s_Cannot_create_writable_memory_vi); + Py_CLEAR(clear_module_state->__pyx_kp_u_Cannot_index_with_type); + Py_CLEAR(clear_module_state->__pyx_kp_s_Cannot_transpose_memoryview_with); + Py_CLEAR(clear_module_state->__pyx_n_s_DTYPE); + Py_CLEAR(clear_module_state->__pyx_n_s_DatasetSearcher); + Py_CLEAR(clear_module_state->__pyx_n_s_DatasetSearcher___reduce_cython); + Py_CLEAR(clear_module_state->__pyx_n_s_DatasetSearcher___setstate_cytho); + Py_CLEAR(clear_module_state->__pyx_kp_s_Dimension_d_is_not_direct); + Py_CLEAR(clear_module_state->__pyx_n_s_Ellipsis); + Py_CLEAR(clear_module_state->__pyx_kp_s_Empty_shape_tuple_for_cython_arr); + Py_CLEAR(clear_module_state->__pyx_n_s_ImportError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0); + Py_CLEAR(clear_module_state->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0_2); + Py_CLEAR(clear_module_state->__pyx_n_s_IndexError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Index_out_of_bounds_axis_d); + Py_CLEAR(clear_module_state->__pyx_kp_s_Indirect_dimensions_not_supporte); + Py_CLEAR(clear_module_state->__pyx_kp_u_Invalid_break_mode); + Py_CLEAR(clear_module_state->__pyx_kp_u_Invalid_mode_expected_c_or_fortr); + Py_CLEAR(clear_module_state->__pyx_kp_u_Invalid_shape_in_axis); + Py_CLEAR(clear_module_state->__pyx_n_s_MemoryError); + Py_CLEAR(clear_module_state->__pyx_kp_s_MemoryView_of_r_at_0x_x); + Py_CLEAR(clear_module_state->__pyx_kp_s_MemoryView_of_r_object); + Py_CLEAR(clear_module_state->__pyx_n_b_O); + Py_CLEAR(clear_module_state->__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + Py_CLEAR(clear_module_state->__pyx_n_s_PickleError); + Py_CLEAR(clear_module_state->__pyx_n_s_Sequence); + Py_CLEAR(clear_module_state->__pyx_kp_s_Step_may_not_be_zero_axis_d); + Py_CLEAR(clear_module_state->__pyx_n_s_TypeError); + Py_CLEAR(clear_module_state->__pyx_kp_s_Unable_to_convert_item_to_object); + Py_CLEAR(clear_module_state->__pyx_n_s_ValueError); + Py_CLEAR(clear_module_state->__pyx_n_s_View_MemoryView); + Py_CLEAR(clear_module_state->__pyx_kp_u__2); + Py_CLEAR(clear_module_state->__pyx_n_s__3); + Py_CLEAR(clear_module_state->__pyx_n_s__35); + Py_CLEAR(clear_module_state->__pyx_kp_u__6); + Py_CLEAR(clear_module_state->__pyx_kp_u__7); + Py_CLEAR(clear_module_state->__pyx_n_s_abc); + Py_CLEAR(clear_module_state->__pyx_n_s_allocate_buffer); + Py_CLEAR(clear_module_state->__pyx_kp_u_and); + Py_CLEAR(clear_module_state->__pyx_n_s_asyncio_coroutines); + Py_CLEAR(clear_module_state->__pyx_n_s_axis); + Py_CLEAR(clear_module_state->__pyx_n_s_base); + Py_CLEAR(clear_module_state->__pyx_n_s_block_size); + Py_CLEAR(clear_module_state->__pyx_n_s_break_mode); + Py_CLEAR(clear_module_state->__pyx_n_s_c); + Py_CLEAR(clear_module_state->__pyx_n_u_c); + Py_CLEAR(clear_module_state->__pyx_n_s_chain); + Py_CLEAR(clear_module_state->__pyx_n_s_class); + Py_CLEAR(clear_module_state->__pyx_n_s_class_getitem); + Py_CLEAR(clear_module_state->__pyx_n_s_cline_in_traceback); + Py_CLEAR(clear_module_state->__pyx_n_s_collections); + Py_CLEAR(clear_module_state->__pyx_kp_s_collections_abc); + Py_CLEAR(clear_module_state->__pyx_n_u_complete); + Py_CLEAR(clear_module_state->__pyx_n_u_complete_doc); + Py_CLEAR(clear_module_state->__pyx_kp_s_contiguous_and_direct); + Py_CLEAR(clear_module_state->__pyx_kp_s_contiguous_and_indirect); + Py_CLEAR(clear_module_state->__pyx_n_s_count); + Py_CLEAR(clear_module_state->__pyx_n_s_cumsum); + Py_CLEAR(clear_module_state->__pyx_n_s_dict); + Py_CLEAR(clear_module_state->__pyx_n_s_dict_2); + Py_CLEAR(clear_module_state->__pyx_kp_u_disable); + Py_CLEAR(clear_module_state->__pyx_n_s_document_sep_len); + Py_CLEAR(clear_module_state->__pyx_n_s_dtype); + Py_CLEAR(clear_module_state->__pyx_n_s_dtype_is_object); + Py_CLEAR(clear_module_state->__pyx_kp_u_enable); + Py_CLEAR(clear_module_state->__pyx_n_s_encode); + Py_CLEAR(clear_module_state->__pyx_n_s_enumerate); + Py_CLEAR(clear_module_state->__pyx_n_u_eos); + Py_CLEAR(clear_module_state->__pyx_n_s_error); + Py_CLEAR(clear_module_state->__pyx_kp_s_fairseq_data_token_block_utils_f); + Py_CLEAR(clear_module_state->__pyx_n_s_fairseq_data_token_block_utils_f_2); + Py_CLEAR(clear_module_state->__pyx_n_s_flags); + Py_CLEAR(clear_module_state->__pyx_n_s_format); + Py_CLEAR(clear_module_state->__pyx_n_s_fortran); + Py_CLEAR(clear_module_state->__pyx_n_u_fortran); + Py_CLEAR(clear_module_state->__pyx_n_s_from_iterable); + Py_CLEAR(clear_module_state->__pyx_n_s_fromiter); + Py_CLEAR(clear_module_state->__pyx_kp_u_gc); + Py_CLEAR(clear_module_state->__pyx_n_s_get_block_to_dataset_index_fast); + Py_CLEAR(clear_module_state->__pyx_n_s_get_slice_indices_fast); + Py_CLEAR(clear_module_state->__pyx_n_s_getstate); + Py_CLEAR(clear_module_state->__pyx_kp_u_got); + Py_CLEAR(clear_module_state->__pyx_kp_u_got_differing_extents_in_dimensi); + Py_CLEAR(clear_module_state->__pyx_n_s_id); + Py_CLEAR(clear_module_state->__pyx_n_s_import); + Py_CLEAR(clear_module_state->__pyx_n_s_index); + Py_CLEAR(clear_module_state->__pyx_n_s_initializing); + Py_CLEAR(clear_module_state->__pyx_n_s_int64); + Py_CLEAR(clear_module_state->__pyx_n_s_is_coroutine); + Py_CLEAR(clear_module_state->__pyx_kp_u_isenabled); + Py_CLEAR(clear_module_state->__pyx_n_s_itemsize); + Py_CLEAR(clear_module_state->__pyx_kp_s_itemsize_0_for_cython_array); + Py_CLEAR(clear_module_state->__pyx_n_s_itertools); + Py_CLEAR(clear_module_state->__pyx_n_s_main); + Py_CLEAR(clear_module_state->__pyx_n_s_memview); + Py_CLEAR(clear_module_state->__pyx_n_s_mode); + Py_CLEAR(clear_module_state->__pyx_n_s_name); + Py_CLEAR(clear_module_state->__pyx_n_s_name_2); + Py_CLEAR(clear_module_state->__pyx_n_s_ndim); + Py_CLEAR(clear_module_state->__pyx_n_s_new); + Py_CLEAR(clear_module_state->__pyx_kp_s_no_default___reduce___due_to_non); + Py_CLEAR(clear_module_state->__pyx_n_u_none); + Py_CLEAR(clear_module_state->__pyx_n_s_np); + Py_CLEAR(clear_module_state->__pyx_n_s_numpy); + Py_CLEAR(clear_module_state->__pyx_kp_u_numpy__core_multiarray_failed_to); + Py_CLEAR(clear_module_state->__pyx_kp_u_numpy__core_umath_failed_to_impo); + Py_CLEAR(clear_module_state->__pyx_n_s_obj); + Py_CLEAR(clear_module_state->__pyx_n_s_pack); + Py_CLEAR(clear_module_state->__pyx_n_s_pickle); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_PickleError); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_checksum); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_result); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_state); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_type); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_unpickle_DatasetSearcher); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_unpickle_Enum); + Py_CLEAR(clear_module_state->__pyx_n_s_pyx_vtable); + Py_CLEAR(clear_module_state->__pyx_n_s_range); + Py_CLEAR(clear_module_state->__pyx_n_s_reduce); + Py_CLEAR(clear_module_state->__pyx_n_s_reduce_cython); + Py_CLEAR(clear_module_state->__pyx_n_s_reduce_ex); + Py_CLEAR(clear_module_state->__pyx_n_s_register); + Py_CLEAR(clear_module_state->__pyx_n_s_reshape); + Py_CLEAR(clear_module_state->__pyx_n_s_self); + Py_CLEAR(clear_module_state->__pyx_n_s_setstate); + Py_CLEAR(clear_module_state->__pyx_n_s_setstate_cython); + Py_CLEAR(clear_module_state->__pyx_n_s_shape); + Py_CLEAR(clear_module_state->__pyx_n_s_size); + Py_CLEAR(clear_module_state->__pyx_n_s_sizes); + Py_CLEAR(clear_module_state->__pyx_n_s_slice_indices); + Py_CLEAR(clear_module_state->__pyx_n_s_spec); + Py_CLEAR(clear_module_state->__pyx_n_s_start); + Py_CLEAR(clear_module_state->__pyx_n_s_state); + Py_CLEAR(clear_module_state->__pyx_n_s_step); + Py_CLEAR(clear_module_state->__pyx_n_s_stop); + Py_CLEAR(clear_module_state->__pyx_kp_s_strided_and_direct); + Py_CLEAR(clear_module_state->__pyx_kp_s_strided_and_direct_or_indirect); + Py_CLEAR(clear_module_state->__pyx_kp_s_strided_and_indirect); + Py_CLEAR(clear_module_state->__pyx_kp_s_stringsource); + Py_CLEAR(clear_module_state->__pyx_n_s_struct); + Py_CLEAR(clear_module_state->__pyx_n_s_sum); + Py_CLEAR(clear_module_state->__pyx_n_s_sys); + Py_CLEAR(clear_module_state->__pyx_n_s_test); + Py_CLEAR(clear_module_state->__pyx_n_s_torch); + Py_CLEAR(clear_module_state->__pyx_kp_s_unable_to_allocate_array_data); + Py_CLEAR(clear_module_state->__pyx_kp_s_unable_to_allocate_shape_and_str); + Py_CLEAR(clear_module_state->__pyx_n_s_unpack); + Py_CLEAR(clear_module_state->__pyx_n_s_update); + Py_CLEAR(clear_module_state->__pyx_n_s_use_setstate); + Py_CLEAR(clear_module_state->__pyx_n_s_version_info); + Py_CLEAR(clear_module_state->__pyx_n_s_zeros); + Py_CLEAR(clear_module_state->__pyx_int_0); + Py_CLEAR(clear_module_state->__pyx_int_1); + Py_CLEAR(clear_module_state->__pyx_int_2); + Py_CLEAR(clear_module_state->__pyx_int_3); + Py_CLEAR(clear_module_state->__pyx_int_48422178); + Py_CLEAR(clear_module_state->__pyx_int_107161605); + Py_CLEAR(clear_module_state->__pyx_int_112105877); + Py_CLEAR(clear_module_state->__pyx_int_136983863); + Py_CLEAR(clear_module_state->__pyx_int_147225413); + Py_CLEAR(clear_module_state->__pyx_int_184977713); + Py_CLEAR(clear_module_state->__pyx_int_neg_1); + Py_CLEAR(clear_module_state->__pyx_slice__5); + Py_CLEAR(clear_module_state->__pyx_tuple__4); + Py_CLEAR(clear_module_state->__pyx_tuple__8); + Py_CLEAR(clear_module_state->__pyx_tuple__9); + Py_CLEAR(clear_module_state->__pyx_slice__11); + Py_CLEAR(clear_module_state->__pyx_tuple__10); + Py_CLEAR(clear_module_state->__pyx_tuple__12); + Py_CLEAR(clear_module_state->__pyx_tuple__13); + Py_CLEAR(clear_module_state->__pyx_tuple__14); + Py_CLEAR(clear_module_state->__pyx_tuple__15); + Py_CLEAR(clear_module_state->__pyx_tuple__16); + Py_CLEAR(clear_module_state->__pyx_tuple__17); + Py_CLEAR(clear_module_state->__pyx_tuple__18); + Py_CLEAR(clear_module_state->__pyx_tuple__19); + Py_CLEAR(clear_module_state->__pyx_tuple__20); + Py_CLEAR(clear_module_state->__pyx_tuple__21); + Py_CLEAR(clear_module_state->__pyx_tuple__22); + Py_CLEAR(clear_module_state->__pyx_tuple__23); + Py_CLEAR(clear_module_state->__pyx_tuple__24); + Py_CLEAR(clear_module_state->__pyx_tuple__26); + Py_CLEAR(clear_module_state->__pyx_tuple__28); + Py_CLEAR(clear_module_state->__pyx_tuple__30); + Py_CLEAR(clear_module_state->__pyx_tuple__32); + Py_CLEAR(clear_module_state->__pyx_codeobj__25); + Py_CLEAR(clear_module_state->__pyx_codeobj__27); + Py_CLEAR(clear_module_state->__pyx_codeobj__29); + Py_CLEAR(clear_module_state->__pyx_codeobj__31); + Py_CLEAR(clear_module_state->__pyx_codeobj__33); + Py_CLEAR(clear_module_state->__pyx_codeobj__34); + return 0; +} +#endif +/* #### Code section: module_state_traverse ### */ +#if CYTHON_USE_MODULE_STATE +static int __pyx_m_traverse(PyObject *m, visitproc visit, void *arg) { + __pyx_mstate *traverse_module_state = __pyx_mstate(m); + if (!traverse_module_state) return 0; + Py_VISIT(traverse_module_state->__pyx_d); + Py_VISIT(traverse_module_state->__pyx_b); + Py_VISIT(traverse_module_state->__pyx_cython_runtime); + Py_VISIT(traverse_module_state->__pyx_empty_tuple); + Py_VISIT(traverse_module_state->__pyx_empty_bytes); + Py_VISIT(traverse_module_state->__pyx_empty_unicode); + #ifdef __Pyx_CyFunction_USED + Py_VISIT(traverse_module_state->__pyx_CyFunctionType); + #endif + #ifdef __Pyx_FusedFunction_USED + Py_VISIT(traverse_module_state->__pyx_FusedFunctionType); + #endif + Py_VISIT(traverse_module_state->__pyx_ptype_7cpython_4type_type); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_dtype); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_flatiter); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_broadcast); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_ndarray); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_generic); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_number); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_integer); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_signedinteger); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_unsignedinteger); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_inexact); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_floating); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_complexfloating); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_flexible); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_character); + Py_VISIT(traverse_module_state->__pyx_ptype_5numpy_ufunc); + Py_VISIT(traverse_module_state->__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher); + Py_VISIT(traverse_module_state->__pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher); + Py_VISIT(traverse_module_state->__pyx_array_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_array); + Py_VISIT(traverse_module_state->__pyx_MemviewEnum_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_MemviewEnum); + Py_VISIT(traverse_module_state->__pyx_memoryview_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_memoryview); + Py_VISIT(traverse_module_state->__pyx_memoryviewslice_type); + Py_VISIT(traverse_module_state->__pyx_type___pyx_memoryviewslice); + Py_VISIT(traverse_module_state->__pyx_kp_u_); + Py_VISIT(traverse_module_state->__pyx_n_s_ASCII); + Py_VISIT(traverse_module_state->__pyx_kp_s_All_dimensions_preceding_dimensi); + Py_VISIT(traverse_module_state->__pyx_n_s_AssertionError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Buffer_view_does_not_expose_stri); + Py_VISIT(traverse_module_state->__pyx_kp_s_Can_only_create_a_buffer_that_is); + Py_VISIT(traverse_module_state->__pyx_kp_s_Cannot_assign_to_read_only_memor); + Py_VISIT(traverse_module_state->__pyx_kp_s_Cannot_create_writable_memory_vi); + Py_VISIT(traverse_module_state->__pyx_kp_u_Cannot_index_with_type); + Py_VISIT(traverse_module_state->__pyx_kp_s_Cannot_transpose_memoryview_with); + Py_VISIT(traverse_module_state->__pyx_n_s_DTYPE); + Py_VISIT(traverse_module_state->__pyx_n_s_DatasetSearcher); + Py_VISIT(traverse_module_state->__pyx_n_s_DatasetSearcher___reduce_cython); + Py_VISIT(traverse_module_state->__pyx_n_s_DatasetSearcher___setstate_cytho); + Py_VISIT(traverse_module_state->__pyx_kp_s_Dimension_d_is_not_direct); + Py_VISIT(traverse_module_state->__pyx_n_s_Ellipsis); + Py_VISIT(traverse_module_state->__pyx_kp_s_Empty_shape_tuple_for_cython_arr); + Py_VISIT(traverse_module_state->__pyx_n_s_ImportError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0); + Py_VISIT(traverse_module_state->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0_2); + Py_VISIT(traverse_module_state->__pyx_n_s_IndexError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Index_out_of_bounds_axis_d); + Py_VISIT(traverse_module_state->__pyx_kp_s_Indirect_dimensions_not_supporte); + Py_VISIT(traverse_module_state->__pyx_kp_u_Invalid_break_mode); + Py_VISIT(traverse_module_state->__pyx_kp_u_Invalid_mode_expected_c_or_fortr); + Py_VISIT(traverse_module_state->__pyx_kp_u_Invalid_shape_in_axis); + Py_VISIT(traverse_module_state->__pyx_n_s_MemoryError); + Py_VISIT(traverse_module_state->__pyx_kp_s_MemoryView_of_r_at_0x_x); + Py_VISIT(traverse_module_state->__pyx_kp_s_MemoryView_of_r_object); + Py_VISIT(traverse_module_state->__pyx_n_b_O); + Py_VISIT(traverse_module_state->__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + Py_VISIT(traverse_module_state->__pyx_n_s_PickleError); + Py_VISIT(traverse_module_state->__pyx_n_s_Sequence); + Py_VISIT(traverse_module_state->__pyx_kp_s_Step_may_not_be_zero_axis_d); + Py_VISIT(traverse_module_state->__pyx_n_s_TypeError); + Py_VISIT(traverse_module_state->__pyx_kp_s_Unable_to_convert_item_to_object); + Py_VISIT(traverse_module_state->__pyx_n_s_ValueError); + Py_VISIT(traverse_module_state->__pyx_n_s_View_MemoryView); + Py_VISIT(traverse_module_state->__pyx_kp_u__2); + Py_VISIT(traverse_module_state->__pyx_n_s__3); + Py_VISIT(traverse_module_state->__pyx_n_s__35); + Py_VISIT(traverse_module_state->__pyx_kp_u__6); + Py_VISIT(traverse_module_state->__pyx_kp_u__7); + Py_VISIT(traverse_module_state->__pyx_n_s_abc); + Py_VISIT(traverse_module_state->__pyx_n_s_allocate_buffer); + Py_VISIT(traverse_module_state->__pyx_kp_u_and); + Py_VISIT(traverse_module_state->__pyx_n_s_asyncio_coroutines); + Py_VISIT(traverse_module_state->__pyx_n_s_axis); + Py_VISIT(traverse_module_state->__pyx_n_s_base); + Py_VISIT(traverse_module_state->__pyx_n_s_block_size); + Py_VISIT(traverse_module_state->__pyx_n_s_break_mode); + Py_VISIT(traverse_module_state->__pyx_n_s_c); + Py_VISIT(traverse_module_state->__pyx_n_u_c); + Py_VISIT(traverse_module_state->__pyx_n_s_chain); + Py_VISIT(traverse_module_state->__pyx_n_s_class); + Py_VISIT(traverse_module_state->__pyx_n_s_class_getitem); + Py_VISIT(traverse_module_state->__pyx_n_s_cline_in_traceback); + Py_VISIT(traverse_module_state->__pyx_n_s_collections); + Py_VISIT(traverse_module_state->__pyx_kp_s_collections_abc); + Py_VISIT(traverse_module_state->__pyx_n_u_complete); + Py_VISIT(traverse_module_state->__pyx_n_u_complete_doc); + Py_VISIT(traverse_module_state->__pyx_kp_s_contiguous_and_direct); + Py_VISIT(traverse_module_state->__pyx_kp_s_contiguous_and_indirect); + Py_VISIT(traverse_module_state->__pyx_n_s_count); + Py_VISIT(traverse_module_state->__pyx_n_s_cumsum); + Py_VISIT(traverse_module_state->__pyx_n_s_dict); + Py_VISIT(traverse_module_state->__pyx_n_s_dict_2); + Py_VISIT(traverse_module_state->__pyx_kp_u_disable); + Py_VISIT(traverse_module_state->__pyx_n_s_document_sep_len); + Py_VISIT(traverse_module_state->__pyx_n_s_dtype); + Py_VISIT(traverse_module_state->__pyx_n_s_dtype_is_object); + Py_VISIT(traverse_module_state->__pyx_kp_u_enable); + Py_VISIT(traverse_module_state->__pyx_n_s_encode); + Py_VISIT(traverse_module_state->__pyx_n_s_enumerate); + Py_VISIT(traverse_module_state->__pyx_n_u_eos); + Py_VISIT(traverse_module_state->__pyx_n_s_error); + Py_VISIT(traverse_module_state->__pyx_kp_s_fairseq_data_token_block_utils_f); + Py_VISIT(traverse_module_state->__pyx_n_s_fairseq_data_token_block_utils_f_2); + Py_VISIT(traverse_module_state->__pyx_n_s_flags); + Py_VISIT(traverse_module_state->__pyx_n_s_format); + Py_VISIT(traverse_module_state->__pyx_n_s_fortran); + Py_VISIT(traverse_module_state->__pyx_n_u_fortran); + Py_VISIT(traverse_module_state->__pyx_n_s_from_iterable); + Py_VISIT(traverse_module_state->__pyx_n_s_fromiter); + Py_VISIT(traverse_module_state->__pyx_kp_u_gc); + Py_VISIT(traverse_module_state->__pyx_n_s_get_block_to_dataset_index_fast); + Py_VISIT(traverse_module_state->__pyx_n_s_get_slice_indices_fast); + Py_VISIT(traverse_module_state->__pyx_n_s_getstate); + Py_VISIT(traverse_module_state->__pyx_kp_u_got); + Py_VISIT(traverse_module_state->__pyx_kp_u_got_differing_extents_in_dimensi); + Py_VISIT(traverse_module_state->__pyx_n_s_id); + Py_VISIT(traverse_module_state->__pyx_n_s_import); + Py_VISIT(traverse_module_state->__pyx_n_s_index); + Py_VISIT(traverse_module_state->__pyx_n_s_initializing); + Py_VISIT(traverse_module_state->__pyx_n_s_int64); + Py_VISIT(traverse_module_state->__pyx_n_s_is_coroutine); + Py_VISIT(traverse_module_state->__pyx_kp_u_isenabled); + Py_VISIT(traverse_module_state->__pyx_n_s_itemsize); + Py_VISIT(traverse_module_state->__pyx_kp_s_itemsize_0_for_cython_array); + Py_VISIT(traverse_module_state->__pyx_n_s_itertools); + Py_VISIT(traverse_module_state->__pyx_n_s_main); + Py_VISIT(traverse_module_state->__pyx_n_s_memview); + Py_VISIT(traverse_module_state->__pyx_n_s_mode); + Py_VISIT(traverse_module_state->__pyx_n_s_name); + Py_VISIT(traverse_module_state->__pyx_n_s_name_2); + Py_VISIT(traverse_module_state->__pyx_n_s_ndim); + Py_VISIT(traverse_module_state->__pyx_n_s_new); + Py_VISIT(traverse_module_state->__pyx_kp_s_no_default___reduce___due_to_non); + Py_VISIT(traverse_module_state->__pyx_n_u_none); + Py_VISIT(traverse_module_state->__pyx_n_s_np); + Py_VISIT(traverse_module_state->__pyx_n_s_numpy); + Py_VISIT(traverse_module_state->__pyx_kp_u_numpy__core_multiarray_failed_to); + Py_VISIT(traverse_module_state->__pyx_kp_u_numpy__core_umath_failed_to_impo); + Py_VISIT(traverse_module_state->__pyx_n_s_obj); + Py_VISIT(traverse_module_state->__pyx_n_s_pack); + Py_VISIT(traverse_module_state->__pyx_n_s_pickle); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_PickleError); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_checksum); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_result); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_state); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_type); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_unpickle_DatasetSearcher); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_unpickle_Enum); + Py_VISIT(traverse_module_state->__pyx_n_s_pyx_vtable); + Py_VISIT(traverse_module_state->__pyx_n_s_range); + Py_VISIT(traverse_module_state->__pyx_n_s_reduce); + Py_VISIT(traverse_module_state->__pyx_n_s_reduce_cython); + Py_VISIT(traverse_module_state->__pyx_n_s_reduce_ex); + Py_VISIT(traverse_module_state->__pyx_n_s_register); + Py_VISIT(traverse_module_state->__pyx_n_s_reshape); + Py_VISIT(traverse_module_state->__pyx_n_s_self); + Py_VISIT(traverse_module_state->__pyx_n_s_setstate); + Py_VISIT(traverse_module_state->__pyx_n_s_setstate_cython); + Py_VISIT(traverse_module_state->__pyx_n_s_shape); + Py_VISIT(traverse_module_state->__pyx_n_s_size); + Py_VISIT(traverse_module_state->__pyx_n_s_sizes); + Py_VISIT(traverse_module_state->__pyx_n_s_slice_indices); + Py_VISIT(traverse_module_state->__pyx_n_s_spec); + Py_VISIT(traverse_module_state->__pyx_n_s_start); + Py_VISIT(traverse_module_state->__pyx_n_s_state); + Py_VISIT(traverse_module_state->__pyx_n_s_step); + Py_VISIT(traverse_module_state->__pyx_n_s_stop); + Py_VISIT(traverse_module_state->__pyx_kp_s_strided_and_direct); + Py_VISIT(traverse_module_state->__pyx_kp_s_strided_and_direct_or_indirect); + Py_VISIT(traverse_module_state->__pyx_kp_s_strided_and_indirect); + Py_VISIT(traverse_module_state->__pyx_kp_s_stringsource); + Py_VISIT(traverse_module_state->__pyx_n_s_struct); + Py_VISIT(traverse_module_state->__pyx_n_s_sum); + Py_VISIT(traverse_module_state->__pyx_n_s_sys); + Py_VISIT(traverse_module_state->__pyx_n_s_test); + Py_VISIT(traverse_module_state->__pyx_n_s_torch); + Py_VISIT(traverse_module_state->__pyx_kp_s_unable_to_allocate_array_data); + Py_VISIT(traverse_module_state->__pyx_kp_s_unable_to_allocate_shape_and_str); + Py_VISIT(traverse_module_state->__pyx_n_s_unpack); + Py_VISIT(traverse_module_state->__pyx_n_s_update); + Py_VISIT(traverse_module_state->__pyx_n_s_use_setstate); + Py_VISIT(traverse_module_state->__pyx_n_s_version_info); + Py_VISIT(traverse_module_state->__pyx_n_s_zeros); + Py_VISIT(traverse_module_state->__pyx_int_0); + Py_VISIT(traverse_module_state->__pyx_int_1); + Py_VISIT(traverse_module_state->__pyx_int_2); + Py_VISIT(traverse_module_state->__pyx_int_3); + Py_VISIT(traverse_module_state->__pyx_int_48422178); + Py_VISIT(traverse_module_state->__pyx_int_107161605); + Py_VISIT(traverse_module_state->__pyx_int_112105877); + Py_VISIT(traverse_module_state->__pyx_int_136983863); + Py_VISIT(traverse_module_state->__pyx_int_147225413); + Py_VISIT(traverse_module_state->__pyx_int_184977713); + Py_VISIT(traverse_module_state->__pyx_int_neg_1); + Py_VISIT(traverse_module_state->__pyx_slice__5); + Py_VISIT(traverse_module_state->__pyx_tuple__4); + Py_VISIT(traverse_module_state->__pyx_tuple__8); + Py_VISIT(traverse_module_state->__pyx_tuple__9); + Py_VISIT(traverse_module_state->__pyx_slice__11); + Py_VISIT(traverse_module_state->__pyx_tuple__10); + Py_VISIT(traverse_module_state->__pyx_tuple__12); + Py_VISIT(traverse_module_state->__pyx_tuple__13); + Py_VISIT(traverse_module_state->__pyx_tuple__14); + Py_VISIT(traverse_module_state->__pyx_tuple__15); + Py_VISIT(traverse_module_state->__pyx_tuple__16); + Py_VISIT(traverse_module_state->__pyx_tuple__17); + Py_VISIT(traverse_module_state->__pyx_tuple__18); + Py_VISIT(traverse_module_state->__pyx_tuple__19); + Py_VISIT(traverse_module_state->__pyx_tuple__20); + Py_VISIT(traverse_module_state->__pyx_tuple__21); + Py_VISIT(traverse_module_state->__pyx_tuple__22); + Py_VISIT(traverse_module_state->__pyx_tuple__23); + Py_VISIT(traverse_module_state->__pyx_tuple__24); + Py_VISIT(traverse_module_state->__pyx_tuple__26); + Py_VISIT(traverse_module_state->__pyx_tuple__28); + Py_VISIT(traverse_module_state->__pyx_tuple__30); + Py_VISIT(traverse_module_state->__pyx_tuple__32); + Py_VISIT(traverse_module_state->__pyx_codeobj__25); + Py_VISIT(traverse_module_state->__pyx_codeobj__27); + Py_VISIT(traverse_module_state->__pyx_codeobj__29); + Py_VISIT(traverse_module_state->__pyx_codeobj__31); + Py_VISIT(traverse_module_state->__pyx_codeobj__33); + Py_VISIT(traverse_module_state->__pyx_codeobj__34); + return 0; +} +#endif +/* #### Code section: module_state_defines ### */ +#define __pyx_d __pyx_mstate_global->__pyx_d +#define __pyx_b __pyx_mstate_global->__pyx_b +#define __pyx_cython_runtime __pyx_mstate_global->__pyx_cython_runtime +#define __pyx_empty_tuple __pyx_mstate_global->__pyx_empty_tuple +#define __pyx_empty_bytes __pyx_mstate_global->__pyx_empty_bytes +#define __pyx_empty_unicode __pyx_mstate_global->__pyx_empty_unicode +#ifdef __Pyx_CyFunction_USED +#define __pyx_CyFunctionType __pyx_mstate_global->__pyx_CyFunctionType +#endif +#ifdef __Pyx_FusedFunction_USED +#define __pyx_FusedFunctionType __pyx_mstate_global->__pyx_FusedFunctionType +#endif +#ifdef __Pyx_Generator_USED +#define __pyx_GeneratorType __pyx_mstate_global->__pyx_GeneratorType +#endif +#ifdef __Pyx_IterableCoroutine_USED +#define __pyx_IterableCoroutineType __pyx_mstate_global->__pyx_IterableCoroutineType +#endif +#ifdef __Pyx_Coroutine_USED +#define __pyx_CoroutineAwaitType __pyx_mstate_global->__pyx_CoroutineAwaitType +#endif +#ifdef __Pyx_Coroutine_USED +#define __pyx_CoroutineType __pyx_mstate_global->__pyx_CoroutineType +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#define __pyx_ptype_7cpython_4type_type __pyx_mstate_global->__pyx_ptype_7cpython_4type_type +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#if CYTHON_USE_MODULE_STATE +#endif +#define __pyx_ptype_5numpy_dtype __pyx_mstate_global->__pyx_ptype_5numpy_dtype +#define __pyx_ptype_5numpy_flatiter __pyx_mstate_global->__pyx_ptype_5numpy_flatiter +#define __pyx_ptype_5numpy_broadcast __pyx_mstate_global->__pyx_ptype_5numpy_broadcast +#define __pyx_ptype_5numpy_ndarray __pyx_mstate_global->__pyx_ptype_5numpy_ndarray +#define __pyx_ptype_5numpy_generic __pyx_mstate_global->__pyx_ptype_5numpy_generic +#define __pyx_ptype_5numpy_number __pyx_mstate_global->__pyx_ptype_5numpy_number +#define __pyx_ptype_5numpy_integer __pyx_mstate_global->__pyx_ptype_5numpy_integer +#define __pyx_ptype_5numpy_signedinteger __pyx_mstate_global->__pyx_ptype_5numpy_signedinteger +#define __pyx_ptype_5numpy_unsignedinteger __pyx_mstate_global->__pyx_ptype_5numpy_unsignedinteger +#define __pyx_ptype_5numpy_inexact __pyx_mstate_global->__pyx_ptype_5numpy_inexact +#define __pyx_ptype_5numpy_floating __pyx_mstate_global->__pyx_ptype_5numpy_floating +#define __pyx_ptype_5numpy_complexfloating __pyx_mstate_global->__pyx_ptype_5numpy_complexfloating +#define __pyx_ptype_5numpy_flexible __pyx_mstate_global->__pyx_ptype_5numpy_flexible +#define __pyx_ptype_5numpy_character __pyx_mstate_global->__pyx_ptype_5numpy_character +#define __pyx_ptype_5numpy_ufunc __pyx_mstate_global->__pyx_ptype_5numpy_ufunc +#if CYTHON_USE_MODULE_STATE +#define __pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher __pyx_mstate_global->__pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher +#define __pyx_type___pyx_array __pyx_mstate_global->__pyx_type___pyx_array +#define __pyx_type___pyx_MemviewEnum __pyx_mstate_global->__pyx_type___pyx_MemviewEnum +#define __pyx_type___pyx_memoryview __pyx_mstate_global->__pyx_type___pyx_memoryview +#define __pyx_type___pyx_memoryviewslice __pyx_mstate_global->__pyx_type___pyx_memoryviewslice +#endif +#define __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher __pyx_mstate_global->__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher +#define __pyx_array_type __pyx_mstate_global->__pyx_array_type +#define __pyx_MemviewEnum_type __pyx_mstate_global->__pyx_MemviewEnum_type +#define __pyx_memoryview_type __pyx_mstate_global->__pyx_memoryview_type +#define __pyx_memoryviewslice_type __pyx_mstate_global->__pyx_memoryviewslice_type +#define __pyx_kp_u_ __pyx_mstate_global->__pyx_kp_u_ +#define __pyx_n_s_ASCII __pyx_mstate_global->__pyx_n_s_ASCII +#define __pyx_kp_s_All_dimensions_preceding_dimensi __pyx_mstate_global->__pyx_kp_s_All_dimensions_preceding_dimensi +#define __pyx_n_s_AssertionError __pyx_mstate_global->__pyx_n_s_AssertionError +#define __pyx_kp_s_Buffer_view_does_not_expose_stri __pyx_mstate_global->__pyx_kp_s_Buffer_view_does_not_expose_stri +#define __pyx_kp_s_Can_only_create_a_buffer_that_is __pyx_mstate_global->__pyx_kp_s_Can_only_create_a_buffer_that_is +#define __pyx_kp_s_Cannot_assign_to_read_only_memor __pyx_mstate_global->__pyx_kp_s_Cannot_assign_to_read_only_memor +#define __pyx_kp_s_Cannot_create_writable_memory_vi __pyx_mstate_global->__pyx_kp_s_Cannot_create_writable_memory_vi +#define __pyx_kp_u_Cannot_index_with_type __pyx_mstate_global->__pyx_kp_u_Cannot_index_with_type +#define __pyx_kp_s_Cannot_transpose_memoryview_with __pyx_mstate_global->__pyx_kp_s_Cannot_transpose_memoryview_with +#define __pyx_n_s_DTYPE __pyx_mstate_global->__pyx_n_s_DTYPE +#define __pyx_n_s_DatasetSearcher __pyx_mstate_global->__pyx_n_s_DatasetSearcher +#define __pyx_n_s_DatasetSearcher___reduce_cython __pyx_mstate_global->__pyx_n_s_DatasetSearcher___reduce_cython +#define __pyx_n_s_DatasetSearcher___setstate_cytho __pyx_mstate_global->__pyx_n_s_DatasetSearcher___setstate_cytho +#define __pyx_kp_s_Dimension_d_is_not_direct __pyx_mstate_global->__pyx_kp_s_Dimension_d_is_not_direct +#define __pyx_n_s_Ellipsis __pyx_mstate_global->__pyx_n_s_Ellipsis +#define __pyx_kp_s_Empty_shape_tuple_for_cython_arr __pyx_mstate_global->__pyx_kp_s_Empty_shape_tuple_for_cython_arr +#define __pyx_n_s_ImportError __pyx_mstate_global->__pyx_n_s_ImportError +#define __pyx_kp_s_Incompatible_checksums_0x_x_vs_0 __pyx_mstate_global->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0 +#define __pyx_kp_s_Incompatible_checksums_0x_x_vs_0_2 __pyx_mstate_global->__pyx_kp_s_Incompatible_checksums_0x_x_vs_0_2 +#define __pyx_n_s_IndexError __pyx_mstate_global->__pyx_n_s_IndexError +#define __pyx_kp_s_Index_out_of_bounds_axis_d __pyx_mstate_global->__pyx_kp_s_Index_out_of_bounds_axis_d +#define __pyx_kp_s_Indirect_dimensions_not_supporte __pyx_mstate_global->__pyx_kp_s_Indirect_dimensions_not_supporte +#define __pyx_kp_u_Invalid_break_mode __pyx_mstate_global->__pyx_kp_u_Invalid_break_mode +#define __pyx_kp_u_Invalid_mode_expected_c_or_fortr __pyx_mstate_global->__pyx_kp_u_Invalid_mode_expected_c_or_fortr +#define __pyx_kp_u_Invalid_shape_in_axis __pyx_mstate_global->__pyx_kp_u_Invalid_shape_in_axis +#define __pyx_n_s_MemoryError __pyx_mstate_global->__pyx_n_s_MemoryError +#define __pyx_kp_s_MemoryView_of_r_at_0x_x __pyx_mstate_global->__pyx_kp_s_MemoryView_of_r_at_0x_x +#define __pyx_kp_s_MemoryView_of_r_object __pyx_mstate_global->__pyx_kp_s_MemoryView_of_r_object +#define __pyx_n_b_O __pyx_mstate_global->__pyx_n_b_O +#define __pyx_kp_u_Out_of_bounds_on_buffer_access_a __pyx_mstate_global->__pyx_kp_u_Out_of_bounds_on_buffer_access_a +#define __pyx_n_s_PickleError __pyx_mstate_global->__pyx_n_s_PickleError +#define __pyx_n_s_Sequence __pyx_mstate_global->__pyx_n_s_Sequence +#define __pyx_kp_s_Step_may_not_be_zero_axis_d __pyx_mstate_global->__pyx_kp_s_Step_may_not_be_zero_axis_d +#define __pyx_n_s_TypeError __pyx_mstate_global->__pyx_n_s_TypeError +#define __pyx_kp_s_Unable_to_convert_item_to_object __pyx_mstate_global->__pyx_kp_s_Unable_to_convert_item_to_object +#define __pyx_n_s_ValueError __pyx_mstate_global->__pyx_n_s_ValueError +#define __pyx_n_s_View_MemoryView __pyx_mstate_global->__pyx_n_s_View_MemoryView +#define __pyx_kp_u__2 __pyx_mstate_global->__pyx_kp_u__2 +#define __pyx_n_s__3 __pyx_mstate_global->__pyx_n_s__3 +#define __pyx_n_s__35 __pyx_mstate_global->__pyx_n_s__35 +#define __pyx_kp_u__6 __pyx_mstate_global->__pyx_kp_u__6 +#define __pyx_kp_u__7 __pyx_mstate_global->__pyx_kp_u__7 +#define __pyx_n_s_abc __pyx_mstate_global->__pyx_n_s_abc +#define __pyx_n_s_allocate_buffer __pyx_mstate_global->__pyx_n_s_allocate_buffer +#define __pyx_kp_u_and __pyx_mstate_global->__pyx_kp_u_and +#define __pyx_n_s_asyncio_coroutines __pyx_mstate_global->__pyx_n_s_asyncio_coroutines +#define __pyx_n_s_axis __pyx_mstate_global->__pyx_n_s_axis +#define __pyx_n_s_base __pyx_mstate_global->__pyx_n_s_base +#define __pyx_n_s_block_size __pyx_mstate_global->__pyx_n_s_block_size +#define __pyx_n_s_break_mode __pyx_mstate_global->__pyx_n_s_break_mode +#define __pyx_n_s_c __pyx_mstate_global->__pyx_n_s_c +#define __pyx_n_u_c __pyx_mstate_global->__pyx_n_u_c +#define __pyx_n_s_chain __pyx_mstate_global->__pyx_n_s_chain +#define __pyx_n_s_class __pyx_mstate_global->__pyx_n_s_class +#define __pyx_n_s_class_getitem __pyx_mstate_global->__pyx_n_s_class_getitem +#define __pyx_n_s_cline_in_traceback __pyx_mstate_global->__pyx_n_s_cline_in_traceback +#define __pyx_n_s_collections __pyx_mstate_global->__pyx_n_s_collections +#define __pyx_kp_s_collections_abc __pyx_mstate_global->__pyx_kp_s_collections_abc +#define __pyx_n_u_complete __pyx_mstate_global->__pyx_n_u_complete +#define __pyx_n_u_complete_doc __pyx_mstate_global->__pyx_n_u_complete_doc +#define __pyx_kp_s_contiguous_and_direct __pyx_mstate_global->__pyx_kp_s_contiguous_and_direct +#define __pyx_kp_s_contiguous_and_indirect __pyx_mstate_global->__pyx_kp_s_contiguous_and_indirect +#define __pyx_n_s_count __pyx_mstate_global->__pyx_n_s_count +#define __pyx_n_s_cumsum __pyx_mstate_global->__pyx_n_s_cumsum +#define __pyx_n_s_dict __pyx_mstate_global->__pyx_n_s_dict +#define __pyx_n_s_dict_2 __pyx_mstate_global->__pyx_n_s_dict_2 +#define __pyx_kp_u_disable __pyx_mstate_global->__pyx_kp_u_disable +#define __pyx_n_s_document_sep_len __pyx_mstate_global->__pyx_n_s_document_sep_len +#define __pyx_n_s_dtype __pyx_mstate_global->__pyx_n_s_dtype +#define __pyx_n_s_dtype_is_object __pyx_mstate_global->__pyx_n_s_dtype_is_object +#define __pyx_kp_u_enable __pyx_mstate_global->__pyx_kp_u_enable +#define __pyx_n_s_encode __pyx_mstate_global->__pyx_n_s_encode +#define __pyx_n_s_enumerate __pyx_mstate_global->__pyx_n_s_enumerate +#define __pyx_n_u_eos __pyx_mstate_global->__pyx_n_u_eos +#define __pyx_n_s_error __pyx_mstate_global->__pyx_n_s_error +#define __pyx_kp_s_fairseq_data_token_block_utils_f __pyx_mstate_global->__pyx_kp_s_fairseq_data_token_block_utils_f +#define __pyx_n_s_fairseq_data_token_block_utils_f_2 __pyx_mstate_global->__pyx_n_s_fairseq_data_token_block_utils_f_2 +#define __pyx_n_s_flags __pyx_mstate_global->__pyx_n_s_flags +#define __pyx_n_s_format __pyx_mstate_global->__pyx_n_s_format +#define __pyx_n_s_fortran __pyx_mstate_global->__pyx_n_s_fortran +#define __pyx_n_u_fortran __pyx_mstate_global->__pyx_n_u_fortran +#define __pyx_n_s_from_iterable __pyx_mstate_global->__pyx_n_s_from_iterable +#define __pyx_n_s_fromiter __pyx_mstate_global->__pyx_n_s_fromiter +#define __pyx_kp_u_gc __pyx_mstate_global->__pyx_kp_u_gc +#define __pyx_n_s_get_block_to_dataset_index_fast __pyx_mstate_global->__pyx_n_s_get_block_to_dataset_index_fast +#define __pyx_n_s_get_slice_indices_fast __pyx_mstate_global->__pyx_n_s_get_slice_indices_fast +#define __pyx_n_s_getstate __pyx_mstate_global->__pyx_n_s_getstate +#define __pyx_kp_u_got __pyx_mstate_global->__pyx_kp_u_got +#define __pyx_kp_u_got_differing_extents_in_dimensi __pyx_mstate_global->__pyx_kp_u_got_differing_extents_in_dimensi +#define __pyx_n_s_id __pyx_mstate_global->__pyx_n_s_id +#define __pyx_n_s_import __pyx_mstate_global->__pyx_n_s_import +#define __pyx_n_s_index __pyx_mstate_global->__pyx_n_s_index +#define __pyx_n_s_initializing __pyx_mstate_global->__pyx_n_s_initializing +#define __pyx_n_s_int64 __pyx_mstate_global->__pyx_n_s_int64 +#define __pyx_n_s_is_coroutine __pyx_mstate_global->__pyx_n_s_is_coroutine +#define __pyx_kp_u_isenabled __pyx_mstate_global->__pyx_kp_u_isenabled +#define __pyx_n_s_itemsize __pyx_mstate_global->__pyx_n_s_itemsize +#define __pyx_kp_s_itemsize_0_for_cython_array __pyx_mstate_global->__pyx_kp_s_itemsize_0_for_cython_array +#define __pyx_n_s_itertools __pyx_mstate_global->__pyx_n_s_itertools +#define __pyx_n_s_main __pyx_mstate_global->__pyx_n_s_main +#define __pyx_n_s_memview __pyx_mstate_global->__pyx_n_s_memview +#define __pyx_n_s_mode __pyx_mstate_global->__pyx_n_s_mode +#define __pyx_n_s_name __pyx_mstate_global->__pyx_n_s_name +#define __pyx_n_s_name_2 __pyx_mstate_global->__pyx_n_s_name_2 +#define __pyx_n_s_ndim __pyx_mstate_global->__pyx_n_s_ndim +#define __pyx_n_s_new __pyx_mstate_global->__pyx_n_s_new +#define __pyx_kp_s_no_default___reduce___due_to_non __pyx_mstate_global->__pyx_kp_s_no_default___reduce___due_to_non +#define __pyx_n_u_none __pyx_mstate_global->__pyx_n_u_none +#define __pyx_n_s_np __pyx_mstate_global->__pyx_n_s_np +#define __pyx_n_s_numpy __pyx_mstate_global->__pyx_n_s_numpy +#define __pyx_kp_u_numpy__core_multiarray_failed_to __pyx_mstate_global->__pyx_kp_u_numpy__core_multiarray_failed_to +#define __pyx_kp_u_numpy__core_umath_failed_to_impo __pyx_mstate_global->__pyx_kp_u_numpy__core_umath_failed_to_impo +#define __pyx_n_s_obj __pyx_mstate_global->__pyx_n_s_obj +#define __pyx_n_s_pack __pyx_mstate_global->__pyx_n_s_pack +#define __pyx_n_s_pickle __pyx_mstate_global->__pyx_n_s_pickle +#define __pyx_n_s_pyx_PickleError __pyx_mstate_global->__pyx_n_s_pyx_PickleError +#define __pyx_n_s_pyx_checksum __pyx_mstate_global->__pyx_n_s_pyx_checksum +#define __pyx_n_s_pyx_result __pyx_mstate_global->__pyx_n_s_pyx_result +#define __pyx_n_s_pyx_state __pyx_mstate_global->__pyx_n_s_pyx_state +#define __pyx_n_s_pyx_type __pyx_mstate_global->__pyx_n_s_pyx_type +#define __pyx_n_s_pyx_unpickle_DatasetSearcher __pyx_mstate_global->__pyx_n_s_pyx_unpickle_DatasetSearcher +#define __pyx_n_s_pyx_unpickle_Enum __pyx_mstate_global->__pyx_n_s_pyx_unpickle_Enum +#define __pyx_n_s_pyx_vtable __pyx_mstate_global->__pyx_n_s_pyx_vtable +#define __pyx_n_s_range __pyx_mstate_global->__pyx_n_s_range +#define __pyx_n_s_reduce __pyx_mstate_global->__pyx_n_s_reduce +#define __pyx_n_s_reduce_cython __pyx_mstate_global->__pyx_n_s_reduce_cython +#define __pyx_n_s_reduce_ex __pyx_mstate_global->__pyx_n_s_reduce_ex +#define __pyx_n_s_register __pyx_mstate_global->__pyx_n_s_register +#define __pyx_n_s_reshape __pyx_mstate_global->__pyx_n_s_reshape +#define __pyx_n_s_self __pyx_mstate_global->__pyx_n_s_self +#define __pyx_n_s_setstate __pyx_mstate_global->__pyx_n_s_setstate +#define __pyx_n_s_setstate_cython __pyx_mstate_global->__pyx_n_s_setstate_cython +#define __pyx_n_s_shape __pyx_mstate_global->__pyx_n_s_shape +#define __pyx_n_s_size __pyx_mstate_global->__pyx_n_s_size +#define __pyx_n_s_sizes __pyx_mstate_global->__pyx_n_s_sizes +#define __pyx_n_s_slice_indices __pyx_mstate_global->__pyx_n_s_slice_indices +#define __pyx_n_s_spec __pyx_mstate_global->__pyx_n_s_spec +#define __pyx_n_s_start __pyx_mstate_global->__pyx_n_s_start +#define __pyx_n_s_state __pyx_mstate_global->__pyx_n_s_state +#define __pyx_n_s_step __pyx_mstate_global->__pyx_n_s_step +#define __pyx_n_s_stop __pyx_mstate_global->__pyx_n_s_stop +#define __pyx_kp_s_strided_and_direct __pyx_mstate_global->__pyx_kp_s_strided_and_direct +#define __pyx_kp_s_strided_and_direct_or_indirect __pyx_mstate_global->__pyx_kp_s_strided_and_direct_or_indirect +#define __pyx_kp_s_strided_and_indirect __pyx_mstate_global->__pyx_kp_s_strided_and_indirect +#define __pyx_kp_s_stringsource __pyx_mstate_global->__pyx_kp_s_stringsource +#define __pyx_n_s_struct __pyx_mstate_global->__pyx_n_s_struct +#define __pyx_n_s_sum __pyx_mstate_global->__pyx_n_s_sum +#define __pyx_n_s_sys __pyx_mstate_global->__pyx_n_s_sys +#define __pyx_n_s_test __pyx_mstate_global->__pyx_n_s_test +#define __pyx_n_s_torch __pyx_mstate_global->__pyx_n_s_torch +#define __pyx_kp_s_unable_to_allocate_array_data __pyx_mstate_global->__pyx_kp_s_unable_to_allocate_array_data +#define __pyx_kp_s_unable_to_allocate_shape_and_str __pyx_mstate_global->__pyx_kp_s_unable_to_allocate_shape_and_str +#define __pyx_n_s_unpack __pyx_mstate_global->__pyx_n_s_unpack +#define __pyx_n_s_update __pyx_mstate_global->__pyx_n_s_update +#define __pyx_n_s_use_setstate __pyx_mstate_global->__pyx_n_s_use_setstate +#define __pyx_n_s_version_info __pyx_mstate_global->__pyx_n_s_version_info +#define __pyx_n_s_zeros __pyx_mstate_global->__pyx_n_s_zeros +#define __pyx_int_0 __pyx_mstate_global->__pyx_int_0 +#define __pyx_int_1 __pyx_mstate_global->__pyx_int_1 +#define __pyx_int_2 __pyx_mstate_global->__pyx_int_2 +#define __pyx_int_3 __pyx_mstate_global->__pyx_int_3 +#define __pyx_int_48422178 __pyx_mstate_global->__pyx_int_48422178 +#define __pyx_int_107161605 __pyx_mstate_global->__pyx_int_107161605 +#define __pyx_int_112105877 __pyx_mstate_global->__pyx_int_112105877 +#define __pyx_int_136983863 __pyx_mstate_global->__pyx_int_136983863 +#define __pyx_int_147225413 __pyx_mstate_global->__pyx_int_147225413 +#define __pyx_int_184977713 __pyx_mstate_global->__pyx_int_184977713 +#define __pyx_int_neg_1 __pyx_mstate_global->__pyx_int_neg_1 +#define __pyx_slice__5 __pyx_mstate_global->__pyx_slice__5 +#define __pyx_tuple__4 __pyx_mstate_global->__pyx_tuple__4 +#define __pyx_tuple__8 __pyx_mstate_global->__pyx_tuple__8 +#define __pyx_tuple__9 __pyx_mstate_global->__pyx_tuple__9 +#define __pyx_slice__11 __pyx_mstate_global->__pyx_slice__11 +#define __pyx_tuple__10 __pyx_mstate_global->__pyx_tuple__10 +#define __pyx_tuple__12 __pyx_mstate_global->__pyx_tuple__12 +#define __pyx_tuple__13 __pyx_mstate_global->__pyx_tuple__13 +#define __pyx_tuple__14 __pyx_mstate_global->__pyx_tuple__14 +#define __pyx_tuple__15 __pyx_mstate_global->__pyx_tuple__15 +#define __pyx_tuple__16 __pyx_mstate_global->__pyx_tuple__16 +#define __pyx_tuple__17 __pyx_mstate_global->__pyx_tuple__17 +#define __pyx_tuple__18 __pyx_mstate_global->__pyx_tuple__18 +#define __pyx_tuple__19 __pyx_mstate_global->__pyx_tuple__19 +#define __pyx_tuple__20 __pyx_mstate_global->__pyx_tuple__20 +#define __pyx_tuple__21 __pyx_mstate_global->__pyx_tuple__21 +#define __pyx_tuple__22 __pyx_mstate_global->__pyx_tuple__22 +#define __pyx_tuple__23 __pyx_mstate_global->__pyx_tuple__23 +#define __pyx_tuple__24 __pyx_mstate_global->__pyx_tuple__24 +#define __pyx_tuple__26 __pyx_mstate_global->__pyx_tuple__26 +#define __pyx_tuple__28 __pyx_mstate_global->__pyx_tuple__28 +#define __pyx_tuple__30 __pyx_mstate_global->__pyx_tuple__30 +#define __pyx_tuple__32 __pyx_mstate_global->__pyx_tuple__32 +#define __pyx_codeobj__25 __pyx_mstate_global->__pyx_codeobj__25 +#define __pyx_codeobj__27 __pyx_mstate_global->__pyx_codeobj__27 +#define __pyx_codeobj__29 __pyx_mstate_global->__pyx_codeobj__29 +#define __pyx_codeobj__31 __pyx_mstate_global->__pyx_codeobj__31 +#define __pyx_codeobj__33 __pyx_mstate_global->__pyx_codeobj__33 +#define __pyx_codeobj__34 __pyx_mstate_global->__pyx_codeobj__34 +/* #### Code section: module_code ### */ + +/* "View.MemoryView":131 + * cdef bint dtype_is_object + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, # <<<<<<<<<<<<<< + * mode="c", bint allocate_buffer=True): + * + */ + +/* Python wrapper */ +static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/ +static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds) { + PyObject *__pyx_v_shape = 0; + Py_ssize_t __pyx_v_itemsize; + PyObject *__pyx_v_format = 0; + PyObject *__pyx_v_mode = 0; + int __pyx_v_allocate_buffer; + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[5] = {0,0,0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__cinit__ (wrapper)", 0); + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return -1; + #endif + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_shape,&__pyx_n_s_itemsize,&__pyx_n_s_format,&__pyx_n_s_mode,&__pyx_n_s_allocate_buffer,0}; + values[3] = __Pyx_Arg_NewRef_VARARGS(((PyObject *)__pyx_n_s_c)); + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 5: values[4] = __Pyx_Arg_VARARGS(__pyx_args, 4); + CYTHON_FALLTHROUGH; + case 4: values[3] = __Pyx_Arg_VARARGS(__pyx_args, 3); + CYTHON_FALLTHROUGH; + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_VARARGS(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_shape)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_itemsize)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 1); __PYX_ERR(1, 131, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_format)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 2); __PYX_ERR(1, 131, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 3: + if (kw_args > 0) { + PyObject* value = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_mode); + if (value) { values[3] = __Pyx_Arg_NewRef_VARARGS(value); kw_args--; } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 4: + if (kw_args > 0) { + PyObject* value = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_allocate_buffer); + if (value) { values[4] = __Pyx_Arg_NewRef_VARARGS(value); kw_args--; } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__cinit__") < 0)) __PYX_ERR(1, 131, __pyx_L3_error) + } + } else { + switch (__pyx_nargs) { + case 5: values[4] = __Pyx_Arg_VARARGS(__pyx_args, 4); + CYTHON_FALLTHROUGH; + case 4: values[3] = __Pyx_Arg_VARARGS(__pyx_args, 3); + CYTHON_FALLTHROUGH; + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + break; + default: goto __pyx_L5_argtuple_error; + } + } + __pyx_v_shape = ((PyObject*)values[0]); + __pyx_v_itemsize = __Pyx_PyIndex_AsSsize_t(values[1]); if (unlikely((__pyx_v_itemsize == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 131, __pyx_L3_error) + __pyx_v_format = values[2]; + __pyx_v_mode = values[3]; + if (values[4]) { + __pyx_v_allocate_buffer = __Pyx_PyObject_IsTrue(values[4]); if (unlikely((__pyx_v_allocate_buffer == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 132, __pyx_L3_error) + } else { + + /* "View.MemoryView":132 + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, + * mode="c", bint allocate_buffer=True): # <<<<<<<<<<<<<< + * + * cdef int idx + */ + __pyx_v_allocate_buffer = ((int)1); + } + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, __pyx_nargs); __PYX_ERR(1, 131, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.array.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return -1; + __pyx_L4_argument_unpacking_done:; + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_shape), (&PyTuple_Type), 1, "shape", 1))) __PYX_ERR(1, 131, __pyx_L1_error) + if (unlikely(((PyObject *)__pyx_v_format) == Py_None)) { + PyErr_Format(PyExc_TypeError, "Argument '%.200s' must not be None", "format"); __PYX_ERR(1, 131, __pyx_L1_error) + } + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(((struct __pyx_array_obj *)__pyx_v_self), __pyx_v_shape, __pyx_v_itemsize, __pyx_v_format, __pyx_v_mode, __pyx_v_allocate_buffer); + + /* "View.MemoryView":131 + * cdef bint dtype_is_object + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, # <<<<<<<<<<<<<< + * mode="c", bint allocate_buffer=True): + * + */ + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __pyx_r = -1; + __pyx_L0:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_shape, Py_ssize_t __pyx_v_itemsize, PyObject *__pyx_v_format, PyObject *__pyx_v_mode, int __pyx_v_allocate_buffer) { + int __pyx_v_idx; + Py_ssize_t __pyx_v_dim; + char __pyx_v_order; + int __pyx_r; + __Pyx_RefNannyDeclarations + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + unsigned int __pyx_t_7; + char *__pyx_t_8; + int __pyx_t_9; + Py_ssize_t __pyx_t_10; + Py_UCS4 __pyx_t_11; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__cinit__", 0); + __Pyx_INCREF(__pyx_v_format); + + /* "View.MemoryView":137 + * cdef Py_ssize_t dim + * + * self.ndim = len(shape) # <<<<<<<<<<<<<< + * self.itemsize = itemsize + * + */ + if (unlikely(__pyx_v_shape == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(1, 137, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_PyTuple_GET_SIZE(__pyx_v_shape); if (unlikely(__pyx_t_1 == ((Py_ssize_t)-1))) __PYX_ERR(1, 137, __pyx_L1_error) + __pyx_v_self->ndim = ((int)__pyx_t_1); + + /* "View.MemoryView":138 + * + * self.ndim = len(shape) + * self.itemsize = itemsize # <<<<<<<<<<<<<< + * + * if not self.ndim: + */ + __pyx_v_self->itemsize = __pyx_v_itemsize; + + /* "View.MemoryView":140 + * self.itemsize = itemsize + * + * if not self.ndim: # <<<<<<<<<<<<<< + * raise ValueError, "Empty shape tuple for cython.array" + * + */ + __pyx_t_2 = (!(__pyx_v_self->ndim != 0)); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":141 + * + * if not self.ndim: + * raise ValueError, "Empty shape tuple for cython.array" # <<<<<<<<<<<<<< + * + * if itemsize <= 0: + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Empty_shape_tuple_for_cython_arr, 0, 0); + __PYX_ERR(1, 141, __pyx_L1_error) + + /* "View.MemoryView":140 + * self.itemsize = itemsize + * + * if not self.ndim: # <<<<<<<<<<<<<< + * raise ValueError, "Empty shape tuple for cython.array" + * + */ + } + + /* "View.MemoryView":143 + * raise ValueError, "Empty shape tuple for cython.array" + * + * if itemsize <= 0: # <<<<<<<<<<<<<< + * raise ValueError, "itemsize <= 0 for cython.array" + * + */ + __pyx_t_2 = (__pyx_v_itemsize <= 0); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":144 + * + * if itemsize <= 0: + * raise ValueError, "itemsize <= 0 for cython.array" # <<<<<<<<<<<<<< + * + * if not isinstance(format, bytes): + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_itemsize_0_for_cython_array, 0, 0); + __PYX_ERR(1, 144, __pyx_L1_error) + + /* "View.MemoryView":143 + * raise ValueError, "Empty shape tuple for cython.array" + * + * if itemsize <= 0: # <<<<<<<<<<<<<< + * raise ValueError, "itemsize <= 0 for cython.array" + * + */ + } + + /* "View.MemoryView":146 + * raise ValueError, "itemsize <= 0 for cython.array" + * + * if not isinstance(format, bytes): # <<<<<<<<<<<<<< + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string + */ + __pyx_t_2 = PyBytes_Check(__pyx_v_format); + __pyx_t_3 = (!__pyx_t_2); + if (__pyx_t_3) { + + /* "View.MemoryView":147 + * + * if not isinstance(format, bytes): + * format = format.encode('ASCII') # <<<<<<<<<<<<<< + * self._format = format # keep a reference to the byte string + * self.format = self._format + */ + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_format, __pyx_n_s_encode); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 147, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_6 = NULL; + __pyx_t_7 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_5))) { + __pyx_t_6 = PyMethod_GET_SELF(__pyx_t_5); + if (likely(__pyx_t_6)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_5); + __Pyx_INCREF(__pyx_t_6); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_5, function); + __pyx_t_7 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_6, __pyx_n_s_ASCII}; + __pyx_t_4 = __Pyx_PyObject_FastCall(__pyx_t_5, __pyx_callargs+1-__pyx_t_7, 1+__pyx_t_7); + __Pyx_XDECREF(__pyx_t_6); __pyx_t_6 = 0; + if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 147, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + __Pyx_DECREF_SET(__pyx_v_format, __pyx_t_4); + __pyx_t_4 = 0; + + /* "View.MemoryView":146 + * raise ValueError, "itemsize <= 0 for cython.array" + * + * if not isinstance(format, bytes): # <<<<<<<<<<<<<< + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string + */ + } + + /* "View.MemoryView":148 + * if not isinstance(format, bytes): + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string # <<<<<<<<<<<<<< + * self.format = self._format + * + */ + if (!(likely(PyBytes_CheckExact(__pyx_v_format))||((__pyx_v_format) == Py_None) || __Pyx_RaiseUnexpectedTypeError("bytes", __pyx_v_format))) __PYX_ERR(1, 148, __pyx_L1_error) + __pyx_t_4 = __pyx_v_format; + __Pyx_INCREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + __Pyx_GOTREF(__pyx_v_self->_format); + __Pyx_DECREF(__pyx_v_self->_format); + __pyx_v_self->_format = ((PyObject*)__pyx_t_4); + __pyx_t_4 = 0; + + /* "View.MemoryView":149 + * format = format.encode('ASCII') + * self._format = format # keep a reference to the byte string + * self.format = self._format # <<<<<<<<<<<<<< + * + * + */ + if (unlikely(__pyx_v_self->_format == Py_None)) { + PyErr_SetString(PyExc_TypeError, "expected bytes, NoneType found"); + __PYX_ERR(1, 149, __pyx_L1_error) + } + __pyx_t_8 = __Pyx_PyBytes_AsWritableString(__pyx_v_self->_format); if (unlikely((!__pyx_t_8) && PyErr_Occurred())) __PYX_ERR(1, 149, __pyx_L1_error) + __pyx_v_self->format = __pyx_t_8; + + /* "View.MemoryView":152 + * + * + * self._shape = PyObject_Malloc(sizeof(Py_ssize_t)*self.ndim*2) # <<<<<<<<<<<<<< + * self._strides = self._shape + self.ndim + * + */ + __pyx_v_self->_shape = ((Py_ssize_t *)PyObject_Malloc((((sizeof(Py_ssize_t)) * __pyx_v_self->ndim) * 2))); + + /* "View.MemoryView":153 + * + * self._shape = PyObject_Malloc(sizeof(Py_ssize_t)*self.ndim*2) + * self._strides = self._shape + self.ndim # <<<<<<<<<<<<<< + * + * if not self._shape: + */ + __pyx_v_self->_strides = (__pyx_v_self->_shape + __pyx_v_self->ndim); + + /* "View.MemoryView":155 + * self._strides = self._shape + self.ndim + * + * if not self._shape: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate shape and strides." + * + */ + __pyx_t_3 = (!(__pyx_v_self->_shape != 0)); + if (unlikely(__pyx_t_3)) { + + /* "View.MemoryView":156 + * + * if not self._shape: + * raise MemoryError, "unable to allocate shape and strides." # <<<<<<<<<<<<<< + * + * + */ + __Pyx_Raise(__pyx_builtin_MemoryError, __pyx_kp_s_unable_to_allocate_shape_and_str, 0, 0); + __PYX_ERR(1, 156, __pyx_L1_error) + + /* "View.MemoryView":155 + * self._strides = self._shape + self.ndim + * + * if not self._shape: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate shape and strides." + * + */ + } + + /* "View.MemoryView":159 + * + * + * for idx, dim in enumerate(shape): # <<<<<<<<<<<<<< + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + */ + __pyx_t_9 = 0; + __pyx_t_4 = __pyx_v_shape; __Pyx_INCREF(__pyx_t_4); + __pyx_t_1 = 0; + for (;;) { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_4); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 159, __pyx_L1_error) + #endif + if (__pyx_t_1 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_4, __pyx_t_1); __Pyx_INCREF(__pyx_t_5); __pyx_t_1++; if (unlikely((0 < 0))) __PYX_ERR(1, 159, __pyx_L1_error) + #else + __pyx_t_5 = __Pyx_PySequence_ITEM(__pyx_t_4, __pyx_t_1); __pyx_t_1++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 159, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + #endif + __pyx_t_10 = __Pyx_PyIndex_AsSsize_t(__pyx_t_5); if (unlikely((__pyx_t_10 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 159, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __pyx_v_dim = __pyx_t_10; + __pyx_v_idx = __pyx_t_9; + __pyx_t_9 = (__pyx_t_9 + 1); + + /* "View.MemoryView":160 + * + * for idx, dim in enumerate(shape): + * if dim <= 0: # <<<<<<<<<<<<<< + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + * self._shape[idx] = dim + */ + __pyx_t_3 = (__pyx_v_dim <= 0); + if (unlikely(__pyx_t_3)) { + + /* "View.MemoryView":161 + * for idx, dim in enumerate(shape): + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." # <<<<<<<<<<<<<< + * self._shape[idx] = dim + * + */ + __pyx_t_5 = PyTuple_New(5); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_10 = 0; + __pyx_t_11 = 127; + __Pyx_INCREF(__pyx_kp_u_Invalid_shape_in_axis); + __pyx_t_10 += 22; + __Pyx_GIVEREF(__pyx_kp_u_Invalid_shape_in_axis); + PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_kp_u_Invalid_shape_in_axis); + __pyx_t_6 = __Pyx_PyUnicode_From_int(__pyx_v_idx, 0, ' ', 'd'); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __pyx_t_10 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_6); + __Pyx_GIVEREF(__pyx_t_6); + PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_t_6); + __pyx_t_6 = 0; + __Pyx_INCREF(__pyx_kp_u_); + __pyx_t_10 += 2; + __Pyx_GIVEREF(__pyx_kp_u_); + PyTuple_SET_ITEM(__pyx_t_5, 2, __pyx_kp_u_); + __pyx_t_6 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_dim, 0, ' ', 'd'); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __pyx_t_10 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_6); + __Pyx_GIVEREF(__pyx_t_6); + PyTuple_SET_ITEM(__pyx_t_5, 3, __pyx_t_6); + __pyx_t_6 = 0; + __Pyx_INCREF(__pyx_kp_u__2); + __pyx_t_10 += 1; + __Pyx_GIVEREF(__pyx_kp_u__2); + PyTuple_SET_ITEM(__pyx_t_5, 4, __pyx_kp_u__2); + __pyx_t_6 = __Pyx_PyUnicode_Join(__pyx_t_5, 5, __pyx_t_10, __pyx_t_11); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 161, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_t_6, 0, 0); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __PYX_ERR(1, 161, __pyx_L1_error) + + /* "View.MemoryView":160 + * + * for idx, dim in enumerate(shape): + * if dim <= 0: # <<<<<<<<<<<<<< + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + * self._shape[idx] = dim + */ + } + + /* "View.MemoryView":162 + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + * self._shape[idx] = dim # <<<<<<<<<<<<<< + * + * cdef char order + */ + (__pyx_v_self->_shape[__pyx_v_idx]) = __pyx_v_dim; + + /* "View.MemoryView":159 + * + * + * for idx, dim in enumerate(shape): # <<<<<<<<<<<<<< + * if dim <= 0: + * raise ValueError, f"Invalid shape in axis {idx}: {dim}." + */ + } + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "View.MemoryView":165 + * + * cdef char order + * if mode == 'c': # <<<<<<<<<<<<<< + * order = b'C' + * self.mode = u'c' + */ + __pyx_t_3 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_c, Py_EQ)); if (unlikely((__pyx_t_3 < 0))) __PYX_ERR(1, 165, __pyx_L1_error) + if (__pyx_t_3) { + + /* "View.MemoryView":166 + * cdef char order + * if mode == 'c': + * order = b'C' # <<<<<<<<<<<<<< + * self.mode = u'c' + * elif mode == 'fortran': + */ + __pyx_v_order = 'C'; + + /* "View.MemoryView":167 + * if mode == 'c': + * order = b'C' + * self.mode = u'c' # <<<<<<<<<<<<<< + * elif mode == 'fortran': + * order = b'F' + */ + __Pyx_INCREF(__pyx_n_u_c); + __Pyx_GIVEREF(__pyx_n_u_c); + __Pyx_GOTREF(__pyx_v_self->mode); + __Pyx_DECREF(__pyx_v_self->mode); + __pyx_v_self->mode = __pyx_n_u_c; + + /* "View.MemoryView":165 + * + * cdef char order + * if mode == 'c': # <<<<<<<<<<<<<< + * order = b'C' + * self.mode = u'c' + */ + goto __pyx_L11; + } + + /* "View.MemoryView":168 + * order = b'C' + * self.mode = u'c' + * elif mode == 'fortran': # <<<<<<<<<<<<<< + * order = b'F' + * self.mode = u'fortran' + */ + __pyx_t_3 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_fortran, Py_EQ)); if (unlikely((__pyx_t_3 < 0))) __PYX_ERR(1, 168, __pyx_L1_error) + if (likely(__pyx_t_3)) { + + /* "View.MemoryView":169 + * self.mode = u'c' + * elif mode == 'fortran': + * order = b'F' # <<<<<<<<<<<<<< + * self.mode = u'fortran' + * else: + */ + __pyx_v_order = 'F'; + + /* "View.MemoryView":170 + * elif mode == 'fortran': + * order = b'F' + * self.mode = u'fortran' # <<<<<<<<<<<<<< + * else: + * raise ValueError, f"Invalid mode, expected 'c' or 'fortran', got {mode}" + */ + __Pyx_INCREF(__pyx_n_u_fortran); + __Pyx_GIVEREF(__pyx_n_u_fortran); + __Pyx_GOTREF(__pyx_v_self->mode); + __Pyx_DECREF(__pyx_v_self->mode); + __pyx_v_self->mode = __pyx_n_u_fortran; + + /* "View.MemoryView":168 + * order = b'C' + * self.mode = u'c' + * elif mode == 'fortran': # <<<<<<<<<<<<<< + * order = b'F' + * self.mode = u'fortran' + */ + goto __pyx_L11; + } + + /* "View.MemoryView":172 + * self.mode = u'fortran' + * else: + * raise ValueError, f"Invalid mode, expected 'c' or 'fortran', got {mode}" # <<<<<<<<<<<<<< + * + * self.len = fill_contig_strides_array(self._shape, self._strides, itemsize, self.ndim, order) + */ + /*else*/ { + __pyx_t_4 = __Pyx_PyObject_FormatSimple(__pyx_v_mode, __pyx_empty_unicode); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 172, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_6 = __Pyx_PyUnicode_Concat(__pyx_kp_u_Invalid_mode_expected_c_or_fortr, __pyx_t_4); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 172, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_t_6, 0, 0); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __PYX_ERR(1, 172, __pyx_L1_error) + } + __pyx_L11:; + + /* "View.MemoryView":174 + * raise ValueError, f"Invalid mode, expected 'c' or 'fortran', got {mode}" + * + * self.len = fill_contig_strides_array(self._shape, self._strides, itemsize, self.ndim, order) # <<<<<<<<<<<<<< + * + * self.free_data = allocate_buffer + */ + __pyx_v_self->len = __pyx_fill_contig_strides_array(__pyx_v_self->_shape, __pyx_v_self->_strides, __pyx_v_itemsize, __pyx_v_self->ndim, __pyx_v_order); + + /* "View.MemoryView":176 + * self.len = fill_contig_strides_array(self._shape, self._strides, itemsize, self.ndim, order) + * + * self.free_data = allocate_buffer # <<<<<<<<<<<<<< + * self.dtype_is_object = format == b'O' + * + */ + __pyx_v_self->free_data = __pyx_v_allocate_buffer; + + /* "View.MemoryView":177 + * + * self.free_data = allocate_buffer + * self.dtype_is_object = format == b'O' # <<<<<<<<<<<<<< + * + * if allocate_buffer: + */ + __pyx_t_6 = PyObject_RichCompare(__pyx_v_format, __pyx_n_b_O, Py_EQ); __Pyx_XGOTREF(__pyx_t_6); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 177, __pyx_L1_error) + __pyx_t_3 = __Pyx_PyObject_IsTrue(__pyx_t_6); if (unlikely((__pyx_t_3 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 177, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __pyx_v_self->dtype_is_object = __pyx_t_3; + + /* "View.MemoryView":179 + * self.dtype_is_object = format == b'O' + * + * if allocate_buffer: # <<<<<<<<<<<<<< + * _allocate_buffer(self) + * + */ + if (__pyx_v_allocate_buffer) { + + /* "View.MemoryView":180 + * + * if allocate_buffer: + * _allocate_buffer(self) # <<<<<<<<<<<<<< + * + * @cname('getbuffer') + */ + __pyx_t_9 = __pyx_array_allocate_buffer(__pyx_v_self); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 180, __pyx_L1_error) + + /* "View.MemoryView":179 + * self.dtype_is_object = format == b'O' + * + * if allocate_buffer: # <<<<<<<<<<<<<< + * _allocate_buffer(self) + * + */ + } + + /* "View.MemoryView":131 + * cdef bint dtype_is_object + * + * def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None, # <<<<<<<<<<<<<< + * mode="c", bint allocate_buffer=True): + * + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_AddTraceback("View.MemoryView.array.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_format); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":182 + * _allocate_buffer(self) + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + */ + +/* Python wrapper */ +CYTHON_UNUSED static int __pyx_array_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +CYTHON_UNUSED static int __pyx_array_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getbuffer__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(((struct __pyx_array_obj *)__pyx_v_self), ((Py_buffer *)__pyx_v_info), ((int)__pyx_v_flags)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(struct __pyx_array_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + int __pyx_v_bufmode; + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + char *__pyx_t_2; + Py_ssize_t __pyx_t_3; + int __pyx_t_4; + Py_ssize_t *__pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + if (unlikely(__pyx_v_info == NULL)) { + PyErr_SetString(PyExc_BufferError, "PyObject_GetBuffer: view==NULL argument is obsolete"); + return -1; + } + __Pyx_RefNannySetupContext("__getbuffer__", 0); + __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(__pyx_v_info->obj); + + /* "View.MemoryView":184 + * @cname('getbuffer') + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 # <<<<<<<<<<<<<< + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": + */ + __pyx_v_bufmode = -1; + + /* "View.MemoryView":185 + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): # <<<<<<<<<<<<<< + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + */ + __pyx_t_1 = ((__pyx_v_flags & ((PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS) | PyBUF_ANY_CONTIGUOUS)) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":186 + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": # <<<<<<<<<<<<<< + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": + */ + __pyx_t_1 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_c, Py_EQ)); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 186, __pyx_L1_error) + if (__pyx_t_1) { + + /* "View.MemoryView":187 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS # <<<<<<<<<<<<<< + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + */ + __pyx_v_bufmode = (PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS); + + /* "View.MemoryView":186 + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): + * if self.mode == u"c": # <<<<<<<<<<<<<< + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": + */ + goto __pyx_L4; + } + + /* "View.MemoryView":188 + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": # <<<<<<<<<<<<<< + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): + */ + __pyx_t_1 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_fortran, Py_EQ)); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 188, __pyx_L1_error) + if (__pyx_t_1) { + + /* "View.MemoryView":189 + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS # <<<<<<<<<<<<<< + * if not (flags & bufmode): + * raise ValueError, "Can only create a buffer that is contiguous in memory." + */ + __pyx_v_bufmode = (PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS); + + /* "View.MemoryView":188 + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * elif self.mode == u"fortran": # <<<<<<<<<<<<<< + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): + */ + } + __pyx_L4:; + + /* "View.MemoryView":190 + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): # <<<<<<<<<<<<<< + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data + */ + __pyx_t_1 = (!((__pyx_v_flags & __pyx_v_bufmode) != 0)); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":191 + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): + * raise ValueError, "Can only create a buffer that is contiguous in memory." # <<<<<<<<<<<<<< + * info.buf = self.data + * info.len = self.len + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Can_only_create_a_buffer_that_is, 0, 0); + __PYX_ERR(1, 191, __pyx_L1_error) + + /* "View.MemoryView":190 + * elif self.mode == u"fortran": + * bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + * if not (flags & bufmode): # <<<<<<<<<<<<<< + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data + */ + } + + /* "View.MemoryView":185 + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + * if flags & (PyBUF_C_CONTIGUOUS | PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS): # <<<<<<<<<<<<<< + * if self.mode == u"c": + * bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS + */ + } + + /* "View.MemoryView":192 + * if not (flags & bufmode): + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data # <<<<<<<<<<<<<< + * info.len = self.len + * + */ + __pyx_t_2 = __pyx_v_self->data; + __pyx_v_info->buf = __pyx_t_2; + + /* "View.MemoryView":193 + * raise ValueError, "Can only create a buffer that is contiguous in memory." + * info.buf = self.data + * info.len = self.len # <<<<<<<<<<<<<< + * + * if flags & PyBUF_STRIDES: + */ + __pyx_t_3 = __pyx_v_self->len; + __pyx_v_info->len = __pyx_t_3; + + /* "View.MemoryView":195 + * info.len = self.len + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.ndim = self.ndim + * info.shape = self._shape + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_STRIDES) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":196 + * + * if flags & PyBUF_STRIDES: + * info.ndim = self.ndim # <<<<<<<<<<<<<< + * info.shape = self._shape + * info.strides = self._strides + */ + __pyx_t_4 = __pyx_v_self->ndim; + __pyx_v_info->ndim = __pyx_t_4; + + /* "View.MemoryView":197 + * if flags & PyBUF_STRIDES: + * info.ndim = self.ndim + * info.shape = self._shape # <<<<<<<<<<<<<< + * info.strides = self._strides + * else: + */ + __pyx_t_5 = __pyx_v_self->_shape; + __pyx_v_info->shape = __pyx_t_5; + + /* "View.MemoryView":198 + * info.ndim = self.ndim + * info.shape = self._shape + * info.strides = self._strides # <<<<<<<<<<<<<< + * else: + * info.ndim = 1 + */ + __pyx_t_5 = __pyx_v_self->_strides; + __pyx_v_info->strides = __pyx_t_5; + + /* "View.MemoryView":195 + * info.len = self.len + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.ndim = self.ndim + * info.shape = self._shape + */ + goto __pyx_L6; + } + + /* "View.MemoryView":200 + * info.strides = self._strides + * else: + * info.ndim = 1 # <<<<<<<<<<<<<< + * info.shape = &self.len if flags & PyBUF_ND else NULL + * info.strides = NULL + */ + /*else*/ { + __pyx_v_info->ndim = 1; + + /* "View.MemoryView":201 + * else: + * info.ndim = 1 + * info.shape = &self.len if flags & PyBUF_ND else NULL # <<<<<<<<<<<<<< + * info.strides = NULL + * + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_ND) != 0); + if (__pyx_t_1) { + __pyx_t_5 = (&__pyx_v_self->len); + } else { + __pyx_t_5 = NULL; + } + __pyx_v_info->shape = __pyx_t_5; + + /* "View.MemoryView":202 + * info.ndim = 1 + * info.shape = &self.len if flags & PyBUF_ND else NULL + * info.strides = NULL # <<<<<<<<<<<<<< + * + * info.suboffsets = NULL + */ + __pyx_v_info->strides = NULL; + } + __pyx_L6:; + + /* "View.MemoryView":204 + * info.strides = NULL + * + * info.suboffsets = NULL # <<<<<<<<<<<<<< + * info.itemsize = self.itemsize + * info.readonly = 0 + */ + __pyx_v_info->suboffsets = NULL; + + /* "View.MemoryView":205 + * + * info.suboffsets = NULL + * info.itemsize = self.itemsize # <<<<<<<<<<<<<< + * info.readonly = 0 + * info.format = self.format if flags & PyBUF_FORMAT else NULL + */ + __pyx_t_3 = __pyx_v_self->itemsize; + __pyx_v_info->itemsize = __pyx_t_3; + + /* "View.MemoryView":206 + * info.suboffsets = NULL + * info.itemsize = self.itemsize + * info.readonly = 0 # <<<<<<<<<<<<<< + * info.format = self.format if flags & PyBUF_FORMAT else NULL + * info.obj = self + */ + __pyx_v_info->readonly = 0; + + /* "View.MemoryView":207 + * info.itemsize = self.itemsize + * info.readonly = 0 + * info.format = self.format if flags & PyBUF_FORMAT else NULL # <<<<<<<<<<<<<< + * info.obj = self + * + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0); + if (__pyx_t_1) { + __pyx_t_2 = __pyx_v_self->format; + } else { + __pyx_t_2 = NULL; + } + __pyx_v_info->format = __pyx_t_2; + + /* "View.MemoryView":208 + * info.readonly = 0 + * info.format = self.format if flags & PyBUF_FORMAT else NULL + * info.obj = self # <<<<<<<<<<<<<< + * + * def __dealloc__(array self): + */ + __Pyx_INCREF((PyObject *)__pyx_v_self); + __Pyx_GIVEREF((PyObject *)__pyx_v_self); + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); + __pyx_v_info->obj = ((PyObject *)__pyx_v_self); + + /* "View.MemoryView":182 + * _allocate_buffer(self) + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * cdef int bufmode = -1 + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.array.__getbuffer__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + if (__pyx_v_info->obj != NULL) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + goto __pyx_L2; + __pyx_L0:; + if (__pyx_v_info->obj == Py_None) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + __pyx_L2:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":210 + * info.obj = self + * + * def __dealloc__(array self): # <<<<<<<<<<<<<< + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + */ + +/* Python wrapper */ +static void __pyx_array___dealloc__(PyObject *__pyx_v_self); /*proto*/ +static void __pyx_array___dealloc__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__dealloc__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); +} + +static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struct __pyx_array_obj *__pyx_v_self) { + int __pyx_t_1; + int __pyx_t_2; + + /* "View.MemoryView":211 + * + * def __dealloc__(array self): + * if self.callback_free_data != NULL: # <<<<<<<<<<<<<< + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + */ + __pyx_t_1 = (__pyx_v_self->callback_free_data != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":212 + * def __dealloc__(array self): + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) # <<<<<<<<<<<<<< + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: + */ + __pyx_v_self->callback_free_data(__pyx_v_self->data); + + /* "View.MemoryView":211 + * + * def __dealloc__(array self): + * if self.callback_free_data != NULL: # <<<<<<<<<<<<<< + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":213 + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: # <<<<<<<<<<<<<< + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + */ + if (__pyx_v_self->free_data) { + } else { + __pyx_t_1 = __pyx_v_self->free_data; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_self->data != NULL); + __pyx_t_1 = __pyx_t_2; + __pyx_L4_bool_binop_done:; + if (__pyx_t_1) { + + /* "View.MemoryView":214 + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) + */ + if (__pyx_v_self->dtype_is_object) { + + /* "View.MemoryView":215 + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) # <<<<<<<<<<<<<< + * free(self.data) + * PyObject_Free(self._shape) + */ + __pyx_memoryview_refcount_objects_in_slice(__pyx_v_self->data, __pyx_v_self->_shape, __pyx_v_self->_strides, __pyx_v_self->ndim, 0); + + /* "View.MemoryView":214 + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) + */ + } + + /* "View.MemoryView":216 + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) # <<<<<<<<<<<<<< + * PyObject_Free(self._shape) + * + */ + free(__pyx_v_self->data); + + /* "View.MemoryView":213 + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + * elif self.free_data and self.data is not NULL: # <<<<<<<<<<<<<< + * if self.dtype_is_object: + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + */ + } + __pyx_L3:; + + /* "View.MemoryView":217 + * refcount_objects_in_slice(self.data, self._shape, self._strides, self.ndim, inc=False) + * free(self.data) + * PyObject_Free(self._shape) # <<<<<<<<<<<<<< + * + * @property + */ + PyObject_Free(__pyx_v_self->_shape); + + /* "View.MemoryView":210 + * info.obj = self + * + * def __dealloc__(array self): # <<<<<<<<<<<<<< + * if self.callback_free_data != NULL: + * self.callback_free_data(self.data) + */ + + /* function exit code */ +} + +/* "View.MemoryView":219 + * PyObject_Free(self._shape) + * + * @property # <<<<<<<<<<<<<< + * def memview(self): + * return self.get_memview() + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_5array_7memview_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_5array_7memview_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_5array_7memview___get__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_5array_7memview___get__(struct __pyx_array_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":221 + * @property + * def memview(self): + * return self.get_memview() # <<<<<<<<<<<<<< + * + * @cname('get_memview') + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = ((struct __pyx_vtabstruct_array *)__pyx_v_self->__pyx_vtab)->get_memview(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 221, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":219 + * PyObject_Free(self._shape) + * + * @property # <<<<<<<<<<<<<< + * def memview(self): + * return self.get_memview() + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.array.memview.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":224 + * + * @cname('get_memview') + * cdef get_memview(self): # <<<<<<<<<<<<<< + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE + * return memoryview(self, flags, self.dtype_is_object) + */ + +static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self) { + int __pyx_v_flags; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("get_memview", 1); + + /* "View.MemoryView":225 + * @cname('get_memview') + * cdef get_memview(self): + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE # <<<<<<<<<<<<<< + * return memoryview(self, flags, self.dtype_is_object) + * + */ + __pyx_v_flags = ((PyBUF_ANY_CONTIGUOUS | PyBUF_FORMAT) | PyBUF_WRITABLE); + + /* "View.MemoryView":226 + * cdef get_memview(self): + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE + * return memoryview(self, flags, self.dtype_is_object) # <<<<<<<<<<<<<< + * + * def __len__(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF((PyObject *)__pyx_v_self); + __Pyx_GIVEREF((PyObject *)__pyx_v_self); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, ((PyObject *)__pyx_v_self))) __PYX_ERR(1, 226, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_1)) __PYX_ERR(1, 226, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2)) __PYX_ERR(1, 226, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 226, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":224 + * + * @cname('get_memview') + * cdef get_memview(self): # <<<<<<<<<<<<<< + * flags = PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE + * return memoryview(self, flags, self.dtype_is_object) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.array.get_memview", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":228 + * return memoryview(self, flags, self.dtype_is_object) + * + * def __len__(self): # <<<<<<<<<<<<<< + * return self._shape[0] + * + */ + +/* Python wrapper */ +static Py_ssize_t __pyx_array___len__(PyObject *__pyx_v_self); /*proto*/ +static Py_ssize_t __pyx_array___len__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + Py_ssize_t __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__len__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static Py_ssize_t __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(struct __pyx_array_obj *__pyx_v_self) { + Py_ssize_t __pyx_r; + + /* "View.MemoryView":229 + * + * def __len__(self): + * return self._shape[0] # <<<<<<<<<<<<<< + * + * def __getattr__(self, attr): + */ + __pyx_r = (__pyx_v_self->_shape[0]); + goto __pyx_L0; + + /* "View.MemoryView":228 + * return memoryview(self, flags, self.dtype_is_object) + * + * def __len__(self): # <<<<<<<<<<<<<< + * return self._shape[0] + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":231 + * return self._shape[0] + * + * def __getattr__(self, attr): # <<<<<<<<<<<<<< + * return getattr(self.memview, attr) + * + */ + +/* Python wrapper */ +static PyObject *__pyx_array___getattr__(PyObject *__pyx_v_self, PyObject *__pyx_v_attr); /*proto*/ +static PyObject *__pyx_array___getattr__(PyObject *__pyx_v_self, PyObject *__pyx_v_attr) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getattr__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(((struct __pyx_array_obj *)__pyx_v_self), ((PyObject *)__pyx_v_attr)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_attr) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__getattr__", 1); + + /* "View.MemoryView":232 + * + * def __getattr__(self, attr): + * return getattr(self.memview, attr) # <<<<<<<<<<<<<< + * + * def __getitem__(self, item): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 232, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_GetAttr(__pyx_t_1, __pyx_v_attr); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 232, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":231 + * return self._shape[0] + * + * def __getattr__(self, attr): # <<<<<<<<<<<<<< + * return getattr(self.memview, attr) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.array.__getattr__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":234 + * return getattr(self.memview, attr) + * + * def __getitem__(self, item): # <<<<<<<<<<<<<< + * return self.memview[item] + * + */ + +/* Python wrapper */ +static PyObject *__pyx_array___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item); /*proto*/ +static PyObject *__pyx_array___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__(((struct __pyx_array_obj *)__pyx_v_self), ((PyObject *)__pyx_v_item)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__getitem__", 1); + + /* "View.MemoryView":235 + * + * def __getitem__(self, item): + * return self.memview[item] # <<<<<<<<<<<<<< + * + * def __setitem__(self, item, value): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 235, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetItem(__pyx_t_1, __pyx_v_item); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 235, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":234 + * return getattr(self.memview, attr) + * + * def __getitem__(self, item): # <<<<<<<<<<<<<< + * return self.memview[item] + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.array.__getitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":237 + * return self.memview[item] + * + * def __setitem__(self, item, value): # <<<<<<<<<<<<<< + * self.memview[item] = value + * + */ + +/* Python wrapper */ +static int __pyx_array___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value); /*proto*/ +static int __pyx_array___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array_12__setitem__(((struct __pyx_array_obj *)__pyx_v_self), ((PyObject *)__pyx_v_item), ((PyObject *)__pyx_v_value)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_12__setitem__(struct __pyx_array_obj *__pyx_v_self, PyObject *__pyx_v_item, PyObject *__pyx_v_value) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setitem__", 1); + + /* "View.MemoryView":238 + * + * def __setitem__(self, item, value): + * self.memview[item] = value # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 238, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + if (unlikely((PyObject_SetItem(__pyx_t_1, __pyx_v_item, __pyx_v_value) < 0))) __PYX_ERR(1, 238, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "View.MemoryView":237 + * return self.memview[item] + * + * def __setitem__(self, item, value): # <<<<<<<<<<<<<< + * self.memview[item] = value + * + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.array.__setitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_array_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_array_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_array___reduce_cython__(((struct __pyx_array_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_array___reduce_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":2 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 2, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.array.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_array_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_array_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + CYTHON_UNUSED PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 3, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 3, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 3, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.array.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_array_2__setstate_cython__(((struct __pyx_array_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_array_2__setstate_cython__(CYTHON_UNUSED struct __pyx_array_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":4 + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 4, __pyx_L1_error) + + /* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.array.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":248 + * + * @cname("__pyx_array_allocate_buffer") + * cdef int _allocate_buffer(array self) except -1: # <<<<<<<<<<<<<< + * + * + */ + +static int __pyx_array_allocate_buffer(struct __pyx_array_obj *__pyx_v_self) { + Py_ssize_t __pyx_v_i; + PyObject **__pyx_v_p; + int __pyx_r; + int __pyx_t_1; + Py_ssize_t __pyx_t_2; + Py_ssize_t __pyx_t_3; + Py_ssize_t __pyx_t_4; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + + /* "View.MemoryView":254 + * cdef PyObject **p + * + * self.free_data = True # <<<<<<<<<<<<<< + * self.data = malloc(self.len) + * if not self.data: + */ + __pyx_v_self->free_data = 1; + + /* "View.MemoryView":255 + * + * self.free_data = True + * self.data = malloc(self.len) # <<<<<<<<<<<<<< + * if not self.data: + * raise MemoryError, "unable to allocate array data." + */ + __pyx_v_self->data = ((char *)malloc(__pyx_v_self->len)); + + /* "View.MemoryView":256 + * self.free_data = True + * self.data = malloc(self.len) + * if not self.data: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate array data." + * + */ + __pyx_t_1 = (!(__pyx_v_self->data != 0)); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":257 + * self.data = malloc(self.len) + * if not self.data: + * raise MemoryError, "unable to allocate array data." # <<<<<<<<<<<<<< + * + * if self.dtype_is_object: + */ + __Pyx_Raise(__pyx_builtin_MemoryError, __pyx_kp_s_unable_to_allocate_array_data, 0, 0); + __PYX_ERR(1, 257, __pyx_L1_error) + + /* "View.MemoryView":256 + * self.free_data = True + * self.data = malloc(self.len) + * if not self.data: # <<<<<<<<<<<<<< + * raise MemoryError, "unable to allocate array data." + * + */ + } + + /* "View.MemoryView":259 + * raise MemoryError, "unable to allocate array data." + * + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * p = self.data + * for i in range(self.len // self.itemsize): + */ + if (__pyx_v_self->dtype_is_object) { + + /* "View.MemoryView":260 + * + * if self.dtype_is_object: + * p = self.data # <<<<<<<<<<<<<< + * for i in range(self.len // self.itemsize): + * p[i] = Py_None + */ + __pyx_v_p = ((PyObject **)__pyx_v_self->data); + + /* "View.MemoryView":261 + * if self.dtype_is_object: + * p = self.data + * for i in range(self.len // self.itemsize): # <<<<<<<<<<<<<< + * p[i] = Py_None + * Py_INCREF(Py_None) + */ + if (unlikely(__pyx_v_self->itemsize == 0)) { + PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero"); + __PYX_ERR(1, 261, __pyx_L1_error) + } + else if (sizeof(Py_ssize_t) == sizeof(long) && (!(((Py_ssize_t)-1) > 0)) && unlikely(__pyx_v_self->itemsize == (Py_ssize_t)-1) && unlikely(__Pyx_UNARY_NEG_WOULD_OVERFLOW(__pyx_v_self->len))) { + PyErr_SetString(PyExc_OverflowError, "value too large to perform division"); + __PYX_ERR(1, 261, __pyx_L1_error) + } + __pyx_t_2 = __Pyx_div_Py_ssize_t(__pyx_v_self->len, __pyx_v_self->itemsize); + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":262 + * p = self.data + * for i in range(self.len // self.itemsize): + * p[i] = Py_None # <<<<<<<<<<<<<< + * Py_INCREF(Py_None) + * return 0 + */ + (__pyx_v_p[__pyx_v_i]) = Py_None; + + /* "View.MemoryView":263 + * for i in range(self.len // self.itemsize): + * p[i] = Py_None + * Py_INCREF(Py_None) # <<<<<<<<<<<<<< + * return 0 + * + */ + Py_INCREF(Py_None); + } + + /* "View.MemoryView":259 + * raise MemoryError, "unable to allocate array data." + * + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * p = self.data + * for i in range(self.len // self.itemsize): + */ + } + + /* "View.MemoryView":264 + * p[i] = Py_None + * Py_INCREF(Py_None) + * return 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":248 + * + * @cname("__pyx_array_allocate_buffer") + * cdef int _allocate_buffer(array self) except -1: # <<<<<<<<<<<<<< + * + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._allocate_buffer", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":268 + * + * @cname("__pyx_array_new") + * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format, char *c_mode, char *buf): # <<<<<<<<<<<<<< + * cdef array result + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + */ + +static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize_t __pyx_v_itemsize, char *__pyx_v_format, char *__pyx_v_c_mode, char *__pyx_v_buf) { + struct __pyx_array_obj *__pyx_v_result = 0; + PyObject *__pyx_v_mode = 0; + struct __pyx_array_obj *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("array_cwrapper", 1); + + /* "View.MemoryView":270 + * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format, char *c_mode, char *buf): + * cdef array result + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. # <<<<<<<<<<<<<< + * + * if buf is NULL: + */ + __pyx_t_2 = ((__pyx_v_c_mode[0]) == 'f'); + if (__pyx_t_2) { + __Pyx_INCREF(__pyx_n_s_fortran); + __pyx_t_1 = __pyx_n_s_fortran; + } else { + __Pyx_INCREF(__pyx_n_s_c); + __pyx_t_1 = __pyx_n_s_c; + } + __pyx_v_mode = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":272 + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + * + * if buf is NULL: # <<<<<<<<<<<<<< + * result = array.__new__(array, shape, itemsize, format, mode) + * else: + */ + __pyx_t_2 = (__pyx_v_buf == NULL); + if (__pyx_t_2) { + + /* "View.MemoryView":273 + * + * if buf is NULL: + * result = array.__new__(array, shape, itemsize, format, mode) # <<<<<<<<<<<<<< + * else: + * result = array.__new__(array, shape, itemsize, format, mode, allocate_buffer=False) + */ + __pyx_t_1 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = PyTuple_New(4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_INCREF(__pyx_v_shape); + __Pyx_GIVEREF(__pyx_v_shape); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_v_shape)) __PYX_ERR(1, 273, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 1, __pyx_t_1)) __PYX_ERR(1, 273, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 2, __pyx_t_3)) __PYX_ERR(1, 273, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_mode); + __Pyx_GIVEREF(__pyx_v_mode); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 3, __pyx_v_mode)) __PYX_ERR(1, 273, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_3 = 0; + __pyx_t_3 = ((PyObject *)__pyx_tp_new_array(((PyTypeObject *)__pyx_array_type), __pyx_t_4, NULL)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 273, __pyx_L1_error) + __Pyx_GOTREF((PyObject *)__pyx_t_3); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_v_result = ((struct __pyx_array_obj *)__pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":272 + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + * + * if buf is NULL: # <<<<<<<<<<<<<< + * result = array.__new__(array, shape, itemsize, format, mode) + * else: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":275 + * result = array.__new__(array, shape, itemsize, format, mode) + * else: + * result = array.__new__(array, shape, itemsize, format, mode, allocate_buffer=False) # <<<<<<<<<<<<<< + * result.data = buf + * + */ + /*else*/ { + __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_1 = PyTuple_New(4); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_v_shape); + __Pyx_GIVEREF(__pyx_v_shape); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_shape)) __PYX_ERR(1, 275, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_t_3)) __PYX_ERR(1, 275, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_4); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 2, __pyx_t_4)) __PYX_ERR(1, 275, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_mode); + __Pyx_GIVEREF(__pyx_v_mode); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 3, __pyx_v_mode)) __PYX_ERR(1, 275, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_4 = 0; + __pyx_t_4 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + if (PyDict_SetItem(__pyx_t_4, __pyx_n_s_allocate_buffer, Py_False) < 0) __PYX_ERR(1, 275, __pyx_L1_error) + __pyx_t_3 = ((PyObject *)__pyx_tp_new_array(((PyTypeObject *)__pyx_array_type), __pyx_t_1, __pyx_t_4)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 275, __pyx_L1_error) + __Pyx_GOTREF((PyObject *)__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_v_result = ((struct __pyx_array_obj *)__pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":276 + * else: + * result = array.__new__(array, shape, itemsize, format, mode, allocate_buffer=False) + * result.data = buf # <<<<<<<<<<<<<< + * + * return result + */ + __pyx_v_result->data = __pyx_v_buf; + } + __pyx_L3:; + + /* "View.MemoryView":278 + * result.data = buf + * + * return result # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = __pyx_v_result; + goto __pyx_L0; + + /* "View.MemoryView":268 + * + * @cname("__pyx_array_new") + * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format, char *c_mode, char *buf): # <<<<<<<<<<<<<< + * cdef array result + * cdef str mode = "fortran" if c_mode[0] == b'f' else "c" # this often comes from a constant C string. + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.array_cwrapper", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XDECREF(__pyx_v_mode); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":304 + * cdef class Enum(object): + * cdef object name + * def __init__(self, name): # <<<<<<<<<<<<<< + * self.name = name + * def __repr__(self): + */ + +/* Python wrapper */ +static int __pyx_MemviewEnum___init__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/ +static int __pyx_MemviewEnum___init__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds) { + PyObject *__pyx_v_name = 0; + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__init__ (wrapper)", 0); + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return -1; + #endif + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_name,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_VARARGS(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_name)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 304, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__init__") < 0)) __PYX_ERR(1, 304, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + } + __pyx_v_name = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__init__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 304, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.Enum.__init__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return -1; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self), __pyx_v_name); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v_name) { + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__init__", 1); + + /* "View.MemoryView":305 + * cdef object name + * def __init__(self, name): + * self.name = name # <<<<<<<<<<<<<< + * def __repr__(self): + * return self.name + */ + __Pyx_INCREF(__pyx_v_name); + __Pyx_GIVEREF(__pyx_v_name); + __Pyx_GOTREF(__pyx_v_self->name); + __Pyx_DECREF(__pyx_v_self->name); + __pyx_v_self->name = __pyx_v_name; + + /* "View.MemoryView":304 + * cdef class Enum(object): + * cdef object name + * def __init__(self, name): # <<<<<<<<<<<<<< + * self.name = name + * def __repr__(self): + */ + + /* function exit code */ + __pyx_r = 0; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":306 + * def __init__(self, name): + * self.name = name + * def __repr__(self): # <<<<<<<<<<<<<< + * return self.name + * + */ + +/* Python wrapper */ +static PyObject *__pyx_MemviewEnum___repr__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_MemviewEnum___repr__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__repr__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr__(struct __pyx_MemviewEnum_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__repr__", 1); + + /* "View.MemoryView":307 + * self.name = name + * def __repr__(self): + * return self.name # <<<<<<<<<<<<<< + * + * cdef generic = Enum("") + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->name); + __pyx_r = __pyx_v_self->name; + goto __pyx_L0; + + /* "View.MemoryView":306 + * def __init__(self, name): + * self.name = name + * def __repr__(self): # <<<<<<<<<<<<<< + * return self.name + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_MemviewEnum_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_MemviewEnum_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_MemviewEnum___reduce_cython__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_MemviewEnum___reduce_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self) { + PyObject *__pyx_v_state = 0; + PyObject *__pyx_v__dict = 0; + int __pyx_v_use_setstate; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":5 + * cdef object _dict + * cdef bint use_setstate + * state = (self.name,) # <<<<<<<<<<<<<< + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: + */ + __pyx_t_1 = PyTuple_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_v_self->name); + __Pyx_GIVEREF(__pyx_v_self->name); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_self->name)) __PYX_ERR(1, 5, __pyx_L1_error); + __pyx_v_state = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "(tree fragment)":6 + * cdef bint use_setstate + * state = (self.name,) + * _dict = getattr(self, '__dict__', None) # <<<<<<<<<<<<<< + * if _dict is not None: + * state += (_dict,) + */ + __pyx_t_1 = __Pyx_GetAttr3(((PyObject *)__pyx_v_self), __pyx_n_s_dict, Py_None); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v__dict = __pyx_t_1; + __pyx_t_1 = 0; + + /* "(tree fragment)":7 + * state = (self.name,) + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: # <<<<<<<<<<<<<< + * state += (_dict,) + * use_setstate = True + */ + __pyx_t_2 = (__pyx_v__dict != Py_None); + if (__pyx_t_2) { + + /* "(tree fragment)":8 + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: + * state += (_dict,) # <<<<<<<<<<<<<< + * use_setstate = True + * else: + */ + __pyx_t_1 = PyTuple_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 8, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_v__dict); + __Pyx_GIVEREF(__pyx_v__dict); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v__dict)) __PYX_ERR(1, 8, __pyx_L1_error); + __pyx_t_3 = PyNumber_InPlaceAdd(__pyx_v_state, __pyx_t_1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 8, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF_SET(__pyx_v_state, ((PyObject*)__pyx_t_3)); + __pyx_t_3 = 0; + + /* "(tree fragment)":9 + * if _dict is not None: + * state += (_dict,) + * use_setstate = True # <<<<<<<<<<<<<< + * else: + * use_setstate = self.name is not None + */ + __pyx_v_use_setstate = 1; + + /* "(tree fragment)":7 + * state = (self.name,) + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: # <<<<<<<<<<<<<< + * state += (_dict,) + * use_setstate = True + */ + goto __pyx_L3; + } + + /* "(tree fragment)":11 + * use_setstate = True + * else: + * use_setstate = self.name is not None # <<<<<<<<<<<<<< + * if use_setstate: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + */ + /*else*/ { + __pyx_t_2 = (__pyx_v_self->name != Py_None); + __pyx_v_use_setstate = __pyx_t_2; + } + __pyx_L3:; + + /* "(tree fragment)":12 + * else: + * use_setstate = self.name is not None + * if use_setstate: # <<<<<<<<<<<<<< + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + * else: + */ + if (__pyx_v_use_setstate) { + + /* "(tree fragment)":13 + * use_setstate = self.name is not None + * if use_setstate: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state # <<<<<<<<<<<<<< + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_GetModuleGlobalName(__pyx_t_3, __pyx_n_s_pyx_unpickle_Enum); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 = PyTuple_New(3); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + __Pyx_GIVEREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, ((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))))) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_136983863); + __Pyx_GIVEREF(__pyx_int_136983863); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_136983863)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(Py_None); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 2, Py_None)) __PYX_ERR(1, 13, __pyx_L1_error); + __pyx_t_4 = PyTuple_New(3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_t_3)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 1, __pyx_t_1)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_state); + __Pyx_GIVEREF(__pyx_v_state); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 2, __pyx_v_state)) __PYX_ERR(1, 13, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_1 = 0; + __pyx_r = __pyx_t_4; + __pyx_t_4 = 0; + goto __pyx_L0; + + /* "(tree fragment)":12 + * else: + * use_setstate = self.name is not None + * if use_setstate: # <<<<<<<<<<<<<< + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + * else: + */ + } + + /* "(tree fragment)":15 + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, None), state + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * __pyx_unpickle_Enum__set_state(self, __pyx_state) + */ + /*else*/ { + __Pyx_XDECREF(__pyx_r); + __Pyx_GetModuleGlobalName(__pyx_t_4, __pyx_n_s_pyx_unpickle_Enum); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_1 = PyTuple_New(3); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + __Pyx_GIVEREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, ((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))))) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_136983863); + __Pyx_GIVEREF(__pyx_int_136983863); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_int_136983863)) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_state); + __Pyx_GIVEREF(__pyx_v_state); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 2, __pyx_v_state)) __PYX_ERR(1, 15, __pyx_L1_error); + __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_4); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_4)) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_1)) __PYX_ERR(1, 15, __pyx_L1_error); + __pyx_t_4 = 0; + __pyx_t_1 = 0; + __pyx_r = __pyx_t_3; + __pyx_t_3 = 0; + goto __pyx_L0; + } + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.Enum.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_state); + __Pyx_XDECREF(__pyx_v__dict); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":16 + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state(self, __pyx_state) + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_MemviewEnum_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_MemviewEnum_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 16, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 16, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 16, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.Enum.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_MemviewEnum_2__setstate_cython__(((struct __pyx_MemviewEnum_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_MemviewEnum_2__setstate_cython__(struct __pyx_MemviewEnum_obj *__pyx_v_self, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":17 + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + * def __setstate_cython__(self, __pyx_state): + * __pyx_unpickle_Enum__set_state(self, __pyx_state) # <<<<<<<<<<<<<< + */ + if (!(likely(PyTuple_CheckExact(__pyx_v___pyx_state))||((__pyx_v___pyx_state) == Py_None) || __Pyx_RaiseUnexpectedTypeError("tuple", __pyx_v___pyx_state))) __PYX_ERR(1, 17, __pyx_L1_error) + __pyx_t_1 = __pyx_unpickle_Enum__set_state(__pyx_v_self, ((PyObject*)__pyx_v___pyx_state)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 17, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":16 + * else: + * return __pyx_unpickle_Enum, (type(self), 0x82a3537, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state(self, __pyx_state) + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.Enum.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":349 + * cdef __Pyx_TypeInfo *typeinfo + * + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): # <<<<<<<<<<<<<< + * self.obj = obj + * self.flags = flags + */ + +/* Python wrapper */ +static int __pyx_memoryview___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/ +static int __pyx_memoryview___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds) { + PyObject *__pyx_v_obj = 0; + int __pyx_v_flags; + int __pyx_v_dtype_is_object; + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[3] = {0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__cinit__ (wrapper)", 0); + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return -1; + #endif + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_obj,&__pyx_n_s_flags,&__pyx_n_s_dtype_is_object,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_VARARGS(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_obj)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_flags)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, 1); __PYX_ERR(1, 349, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (kw_args > 0) { + PyObject* value = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_dtype_is_object); + if (value) { values[2] = __Pyx_Arg_NewRef_VARARGS(value); kw_args--; } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__cinit__") < 0)) __PYX_ERR(1, 349, __pyx_L3_error) + } + } else { + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_VARARGS(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_VARARGS(__pyx_args, 1); + values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + break; + default: goto __pyx_L5_argtuple_error; + } + } + __pyx_v_obj = values[0]; + __pyx_v_flags = __Pyx_PyInt_As_int(values[1]); if (unlikely((__pyx_v_flags == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + if (values[2]) { + __pyx_v_dtype_is_object = __Pyx_PyObject_IsTrue(values[2]); if (unlikely((__pyx_v_dtype_is_object == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 349, __pyx_L3_error) + } else { + __pyx_v_dtype_is_object = ((int)0); + } + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, __pyx_nargs); __PYX_ERR(1, 349, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.memoryview.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return -1; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit__(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_obj, __pyx_v_flags, __pyx_v_dtype_is_object); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj, int __pyx_v_flags, int __pyx_v_dtype_is_object) { + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + Py_intptr_t __pyx_t_4; + size_t __pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__cinit__", 1); + + /* "View.MemoryView":350 + * + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): + * self.obj = obj # <<<<<<<<<<<<<< + * self.flags = flags + * if type(self) is memoryview or obj is not None: + */ + __Pyx_INCREF(__pyx_v_obj); + __Pyx_GIVEREF(__pyx_v_obj); + __Pyx_GOTREF(__pyx_v_self->obj); + __Pyx_DECREF(__pyx_v_self->obj); + __pyx_v_self->obj = __pyx_v_obj; + + /* "View.MemoryView":351 + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): + * self.obj = obj + * self.flags = flags # <<<<<<<<<<<<<< + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) + */ + __pyx_v_self->flags = __pyx_v_flags; + + /* "View.MemoryView":352 + * self.obj = obj + * self.flags = flags + * if type(self) is memoryview or obj is not None: # <<<<<<<<<<<<<< + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: + */ + __pyx_t_2 = (((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))) == ((PyObject *)__pyx_memoryview_type)); + if (!__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_obj != Py_None); + __pyx_t_1 = __pyx_t_2; + __pyx_L4_bool_binop_done:; + if (__pyx_t_1) { + + /* "View.MemoryView":353 + * self.flags = flags + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) # <<<<<<<<<<<<<< + * if self.view.obj == NULL: + * (<__pyx_buffer *> &self.view).obj = Py_None + */ + __pyx_t_3 = __Pyx_GetBuffer(__pyx_v_obj, (&__pyx_v_self->view), __pyx_v_flags); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 353, __pyx_L1_error) + + /* "View.MemoryView":354 + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: # <<<<<<<<<<<<<< + * (<__pyx_buffer *> &self.view).obj = Py_None + * Py_INCREF(Py_None) + */ + __pyx_t_1 = (((PyObject *)__pyx_v_self->view.obj) == NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":355 + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: + * (<__pyx_buffer *> &self.view).obj = Py_None # <<<<<<<<<<<<<< + * Py_INCREF(Py_None) + * + */ + ((Py_buffer *)(&__pyx_v_self->view))->obj = Py_None; + + /* "View.MemoryView":356 + * if self.view.obj == NULL: + * (<__pyx_buffer *> &self.view).obj = Py_None + * Py_INCREF(Py_None) # <<<<<<<<<<<<<< + * + * if not __PYX_CYTHON_ATOMICS_ENABLED(): + */ + Py_INCREF(Py_None); + + /* "View.MemoryView":354 + * if type(self) is memoryview or obj is not None: + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: # <<<<<<<<<<<<<< + * (<__pyx_buffer *> &self.view).obj = Py_None + * Py_INCREF(Py_None) + */ + } + + /* "View.MemoryView":352 + * self.obj = obj + * self.flags = flags + * if type(self) is memoryview or obj is not None: # <<<<<<<<<<<<<< + * __Pyx_GetBuffer(obj, &self.view, flags) + * if self.view.obj == NULL: + */ + } + + /* "View.MemoryView":358 + * Py_INCREF(Py_None) + * + * if not __PYX_CYTHON_ATOMICS_ENABLED(): # <<<<<<<<<<<<<< + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: + */ + __pyx_t_1 = (!__PYX_CYTHON_ATOMICS_ENABLED()); + if (__pyx_t_1) { + + /* "View.MemoryView":360 + * if not __PYX_CYTHON_ATOMICS_ENABLED(): + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: # <<<<<<<<<<<<<< + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + */ + __pyx_t_1 = (__pyx_memoryview_thread_locks_used < 8); + if (__pyx_t_1) { + + /* "View.MemoryView":361 + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: + */ + __pyx_v_self->lock = (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]); + + /* "View.MemoryView":362 + * if __pyx_memoryview_thread_locks_used < 8: + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 # <<<<<<<<<<<<<< + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() + */ + __pyx_memoryview_thread_locks_used = (__pyx_memoryview_thread_locks_used + 1); + + /* "View.MemoryView":360 + * if not __PYX_CYTHON_ATOMICS_ENABLED(): + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: # <<<<<<<<<<<<<< + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + */ + } + + /* "View.MemoryView":363 + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: # <<<<<<<<<<<<<< + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: + */ + __pyx_t_1 = (__pyx_v_self->lock == NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":364 + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() # <<<<<<<<<<<<<< + * if self.lock is NULL: + * raise MemoryError + */ + __pyx_v_self->lock = PyThread_allocate_lock(); + + /* "View.MemoryView":365 + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + __pyx_t_1 = (__pyx_v_self->lock == NULL); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":366 + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: + * raise MemoryError # <<<<<<<<<<<<<< + * + * if flags & PyBUF_FORMAT: + */ + PyErr_NoMemory(); __PYX_ERR(1, 366, __pyx_L1_error) + + /* "View.MemoryView":365 + * if self.lock is NULL: + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + } + + /* "View.MemoryView":363 + * self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] + * __pyx_memoryview_thread_locks_used += 1 + * if self.lock is NULL: # <<<<<<<<<<<<<< + * self.lock = PyThread_allocate_lock() + * if self.lock is NULL: + */ + } + + /* "View.MemoryView":358 + * Py_INCREF(Py_None) + * + * if not __PYX_CYTHON_ATOMICS_ENABLED(): # <<<<<<<<<<<<<< + * global __pyx_memoryview_thread_locks_used + * if __pyx_memoryview_thread_locks_used < 8: + */ + } + + /* "View.MemoryView":368 + * raise MemoryError + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":369 + * + * if flags & PyBUF_FORMAT: + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') # <<<<<<<<<<<<<< + * else: + * self.dtype_is_object = dtype_is_object + */ + __pyx_t_2 = ((__pyx_v_self->view.format[0]) == 'O'); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L12_bool_binop_done; + } + __pyx_t_2 = ((__pyx_v_self->view.format[1]) == '\x00'); + __pyx_t_1 = __pyx_t_2; + __pyx_L12_bool_binop_done:; + __pyx_v_self->dtype_is_object = __pyx_t_1; + + /* "View.MemoryView":368 + * raise MemoryError + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') + * else: + */ + goto __pyx_L11; + } + + /* "View.MemoryView":371 + * self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0') + * else: + * self.dtype_is_object = dtype_is_object # <<<<<<<<<<<<<< + * + * assert (&self.acquisition_count) % sizeof(__pyx_atomic_int_type) == 0 + */ + /*else*/ { + __pyx_v_self->dtype_is_object = __pyx_v_dtype_is_object; + } + __pyx_L11:; + + /* "View.MemoryView":373 + * self.dtype_is_object = dtype_is_object + * + * assert (&self.acquisition_count) % sizeof(__pyx_atomic_int_type) == 0 # <<<<<<<<<<<<<< + * self.typeinfo = NULL + * + */ + #ifndef CYTHON_WITHOUT_ASSERTIONS + if (unlikely(__pyx_assertions_enabled())) { + __pyx_t_4 = ((Py_intptr_t)((void *)(&__pyx_v_self->acquisition_count))); + __pyx_t_5 = (sizeof(__pyx_atomic_int_type)); + if (unlikely(__pyx_t_5 == 0)) { + PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero"); + __PYX_ERR(1, 373, __pyx_L1_error) + } + __pyx_t_1 = ((__pyx_t_4 % __pyx_t_5) == 0); + if (unlikely(!__pyx_t_1)) { + __Pyx_Raise(__pyx_builtin_AssertionError, 0, 0, 0); + __PYX_ERR(1, 373, __pyx_L1_error) + } + } + #else + if ((1)); else __PYX_ERR(1, 373, __pyx_L1_error) + #endif + + /* "View.MemoryView":374 + * + * assert (&self.acquisition_count) % sizeof(__pyx_atomic_int_type) == 0 + * self.typeinfo = NULL # <<<<<<<<<<<<<< + * + * def __dealloc__(memoryview self): + */ + __pyx_v_self->typeinfo = NULL; + + /* "View.MemoryView":349 + * cdef __Pyx_TypeInfo *typeinfo + * + * def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False): # <<<<<<<<<<<<<< + * self.obj = obj + * self.flags = flags + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":376 + * self.typeinfo = NULL + * + * def __dealloc__(memoryview self): # <<<<<<<<<<<<<< + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + */ + +/* Python wrapper */ +static void __pyx_memoryview___dealloc__(PyObject *__pyx_v_self); /*proto*/ +static void __pyx_memoryview___dealloc__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__dealloc__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__dealloc__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); +} + +static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__dealloc__(struct __pyx_memoryview_obj *__pyx_v_self) { + int __pyx_v_i; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + PyThread_type_lock __pyx_t_5; + PyThread_type_lock __pyx_t_6; + + /* "View.MemoryView":377 + * + * def __dealloc__(memoryview self): + * if self.obj is not None: # <<<<<<<<<<<<<< + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + */ + __pyx_t_1 = (__pyx_v_self->obj != Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":378 + * def __dealloc__(memoryview self): + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) # <<<<<<<<<<<<<< + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + * + */ + __Pyx_ReleaseBuffer((&__pyx_v_self->view)); + + /* "View.MemoryView":377 + * + * def __dealloc__(memoryview self): + * if self.obj is not None: # <<<<<<<<<<<<<< + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":379 + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: # <<<<<<<<<<<<<< + * + * (<__pyx_buffer *> &self.view).obj = NULL + */ + __pyx_t_1 = (((Py_buffer *)(&__pyx_v_self->view))->obj == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":381 + * elif (<__pyx_buffer *> &self.view).obj == Py_None: + * + * (<__pyx_buffer *> &self.view).obj = NULL # <<<<<<<<<<<<<< + * Py_DECREF(Py_None) + * + */ + ((Py_buffer *)(&__pyx_v_self->view))->obj = NULL; + + /* "View.MemoryView":382 + * + * (<__pyx_buffer *> &self.view).obj = NULL + * Py_DECREF(Py_None) # <<<<<<<<<<<<<< + * + * cdef int i + */ + Py_DECREF(Py_None); + + /* "View.MemoryView":379 + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + * elif (<__pyx_buffer *> &self.view).obj == Py_None: # <<<<<<<<<<<<<< + * + * (<__pyx_buffer *> &self.view).obj = NULL + */ + } + __pyx_L3:; + + /* "View.MemoryView":386 + * cdef int i + * global __pyx_memoryview_thread_locks_used + * if self.lock != NULL: # <<<<<<<<<<<<<< + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: + */ + __pyx_t_1 = (__pyx_v_self->lock != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":387 + * global __pyx_memoryview_thread_locks_used + * if self.lock != NULL: + * for i in range(__pyx_memoryview_thread_locks_used): # <<<<<<<<<<<<<< + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 + */ + __pyx_t_2 = __pyx_memoryview_thread_locks_used; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":388 + * if self.lock != NULL: + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: + */ + __pyx_t_1 = ((__pyx_memoryview_thread_locks[__pyx_v_i]) == __pyx_v_self->lock); + if (__pyx_t_1) { + + /* "View.MemoryView":389 + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 # <<<<<<<<<<<<<< + * if i != __pyx_memoryview_thread_locks_used: + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + */ + __pyx_memoryview_thread_locks_used = (__pyx_memoryview_thread_locks_used - 1); + + /* "View.MemoryView":390 + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + */ + __pyx_t_1 = (__pyx_v_i != __pyx_memoryview_thread_locks_used); + if (__pyx_t_1) { + + /* "View.MemoryView":392 + * if i != __pyx_memoryview_thread_locks_used: + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) # <<<<<<<<<<<<<< + * break + * else: + */ + __pyx_t_5 = (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]); + __pyx_t_6 = (__pyx_memoryview_thread_locks[__pyx_v_i]); + + /* "View.MemoryView":391 + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + * break + */ + (__pyx_memoryview_thread_locks[__pyx_v_i]) = __pyx_t_5; + (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]) = __pyx_t_6; + + /* "View.MemoryView":390 + * if __pyx_memoryview_thread_locks[i] is self.lock: + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + */ + } + + /* "View.MemoryView":393 + * __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = ( + * __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i]) + * break # <<<<<<<<<<<<<< + * else: + * PyThread_free_lock(self.lock) + */ + goto __pyx_L6_break; + + /* "View.MemoryView":388 + * if self.lock != NULL: + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: # <<<<<<<<<<<<<< + * __pyx_memoryview_thread_locks_used -= 1 + * if i != __pyx_memoryview_thread_locks_used: + */ + } + } + /*else*/ { + + /* "View.MemoryView":395 + * break + * else: + * PyThread_free_lock(self.lock) # <<<<<<<<<<<<<< + * + * cdef char *get_item_pointer(memoryview self, object index) except NULL: + */ + PyThread_free_lock(__pyx_v_self->lock); + } + __pyx_L6_break:; + + /* "View.MemoryView":386 + * cdef int i + * global __pyx_memoryview_thread_locks_used + * if self.lock != NULL: # <<<<<<<<<<<<<< + * for i in range(__pyx_memoryview_thread_locks_used): + * if __pyx_memoryview_thread_locks[i] is self.lock: + */ + } + + /* "View.MemoryView":376 + * self.typeinfo = NULL + * + * def __dealloc__(memoryview self): # <<<<<<<<<<<<<< + * if self.obj is not None: + * __Pyx_ReleaseBuffer(&self.view) + */ + + /* function exit code */ +} + +/* "View.MemoryView":397 + * PyThread_free_lock(self.lock) + * + * cdef char *get_item_pointer(memoryview self, object index) except NULL: # <<<<<<<<<<<<<< + * cdef Py_ssize_t dim + * cdef char *itemp = self.view.buf + */ + +static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index) { + Py_ssize_t __pyx_v_dim; + char *__pyx_v_itemp; + PyObject *__pyx_v_idx = NULL; + char *__pyx_r; + __Pyx_RefNannyDeclarations + Py_ssize_t __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + Py_ssize_t __pyx_t_3; + PyObject *(*__pyx_t_4)(PyObject *); + PyObject *__pyx_t_5 = NULL; + Py_ssize_t __pyx_t_6; + char *__pyx_t_7; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("get_item_pointer", 1); + + /* "View.MemoryView":399 + * cdef char *get_item_pointer(memoryview self, object index) except NULL: + * cdef Py_ssize_t dim + * cdef char *itemp = self.view.buf # <<<<<<<<<<<<<< + * + * for dim, idx in enumerate(index): + */ + __pyx_v_itemp = ((char *)__pyx_v_self->view.buf); + + /* "View.MemoryView":401 + * cdef char *itemp = self.view.buf + * + * for dim, idx in enumerate(index): # <<<<<<<<<<<<<< + * itemp = pybuffer_index(&self.view, itemp, idx, dim) + * + */ + __pyx_t_1 = 0; + if (likely(PyList_CheckExact(__pyx_v_index)) || PyTuple_CheckExact(__pyx_v_index)) { + __pyx_t_2 = __pyx_v_index; __Pyx_INCREF(__pyx_t_2); + __pyx_t_3 = 0; + __pyx_t_4 = NULL; + } else { + __pyx_t_3 = -1; __pyx_t_2 = PyObject_GetIter(__pyx_v_index); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 401, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = __Pyx_PyObject_GetIterNextFunc(__pyx_t_2); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 401, __pyx_L1_error) + } + for (;;) { + if (likely(!__pyx_t_4)) { + if (likely(PyList_CheckExact(__pyx_t_2))) { + { + Py_ssize_t __pyx_temp = __Pyx_PyList_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #endif + if (__pyx_t_3 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_5 = PyList_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely((0 < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #else + __pyx_t_5 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 401, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + #endif + } else { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #endif + if (__pyx_t_3 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely((0 < 0))) __PYX_ERR(1, 401, __pyx_L1_error) + #else + __pyx_t_5 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 401, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + #endif + } + } else { + __pyx_t_5 = __pyx_t_4(__pyx_t_2); + if (unlikely(!__pyx_t_5)) { + PyObject* exc_type = PyErr_Occurred(); + if (exc_type) { + if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear(); + else __PYX_ERR(1, 401, __pyx_L1_error) + } + break; + } + __Pyx_GOTREF(__pyx_t_5); + } + __Pyx_XDECREF_SET(__pyx_v_idx, __pyx_t_5); + __pyx_t_5 = 0; + __pyx_v_dim = __pyx_t_1; + __pyx_t_1 = (__pyx_t_1 + 1); + + /* "View.MemoryView":402 + * + * for dim, idx in enumerate(index): + * itemp = pybuffer_index(&self.view, itemp, idx, dim) # <<<<<<<<<<<<<< + * + * return itemp + */ + __pyx_t_6 = __Pyx_PyIndex_AsSsize_t(__pyx_v_idx); if (unlikely((__pyx_t_6 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 402, __pyx_L1_error) + __pyx_t_7 = __pyx_pybuffer_index((&__pyx_v_self->view), __pyx_v_itemp, __pyx_t_6, __pyx_v_dim); if (unlikely(__pyx_t_7 == ((char *)NULL))) __PYX_ERR(1, 402, __pyx_L1_error) + __pyx_v_itemp = __pyx_t_7; + + /* "View.MemoryView":401 + * cdef char *itemp = self.view.buf + * + * for dim, idx in enumerate(index): # <<<<<<<<<<<<<< + * itemp = pybuffer_index(&self.view, itemp, idx, dim) + * + */ + } + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":404 + * itemp = pybuffer_index(&self.view, itemp, idx, dim) + * + * return itemp # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = __pyx_v_itemp; + goto __pyx_L0; + + /* "View.MemoryView":397 + * PyThread_free_lock(self.lock) + * + * cdef char *get_item_pointer(memoryview self, object index) except NULL: # <<<<<<<<<<<<<< + * cdef Py_ssize_t dim + * cdef char *itemp = self.view.buf + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.memoryview.get_item_pointer", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_idx); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":407 + * + * + * def __getitem__(memoryview self, object index): # <<<<<<<<<<<<<< + * if index is Ellipsis: + * return self + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index); /*proto*/ +static PyObject *__pyx_memoryview___getitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4__getitem__(((struct __pyx_memoryview_obj *)__pyx_v_self), ((PyObject *)__pyx_v_index)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4__getitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index) { + PyObject *__pyx_v_have_slices = NULL; + PyObject *__pyx_v_indices = NULL; + char *__pyx_v_itemp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + char *__pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__getitem__", 1); + + /* "View.MemoryView":408 + * + * def __getitem__(memoryview self, object index): + * if index is Ellipsis: # <<<<<<<<<<<<<< + * return self + * + */ + __pyx_t_1 = (__pyx_v_index == __pyx_builtin_Ellipsis); + if (__pyx_t_1) { + + /* "View.MemoryView":409 + * def __getitem__(memoryview self, object index): + * if index is Ellipsis: + * return self # <<<<<<<<<<<<<< + * + * have_slices, indices = _unellipsify(index, self.view.ndim) + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_self); + __pyx_r = ((PyObject *)__pyx_v_self); + goto __pyx_L0; + + /* "View.MemoryView":408 + * + * def __getitem__(memoryview self, object index): + * if index is Ellipsis: # <<<<<<<<<<<<<< + * return self + * + */ + } + + /* "View.MemoryView":411 + * return self + * + * have_slices, indices = _unellipsify(index, self.view.ndim) # <<<<<<<<<<<<<< + * + * cdef char *itemp + */ + __pyx_t_2 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 411, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + if (likely(__pyx_t_2 != Py_None)) { + PyObject* sequence = __pyx_t_2; + Py_ssize_t size = __Pyx_PySequence_SIZE(sequence); + if (unlikely(size != 2)) { + if (size > 2) __Pyx_RaiseTooManyValuesError(2); + else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size); + __PYX_ERR(1, 411, __pyx_L1_error) + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_3 = PyTuple_GET_ITEM(sequence, 0); + __pyx_t_4 = PyTuple_GET_ITEM(sequence, 1); + __Pyx_INCREF(__pyx_t_3); + __Pyx_INCREF(__pyx_t_4); + #else + __pyx_t_3 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 411, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 411, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + #endif + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + } else { + __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 411, __pyx_L1_error) + } + __pyx_v_have_slices = __pyx_t_3; + __pyx_t_3 = 0; + __pyx_v_indices = __pyx_t_4; + __pyx_t_4 = 0; + + /* "View.MemoryView":414 + * + * cdef char *itemp + * if have_slices: # <<<<<<<<<<<<<< + * return memview_slice(self, indices) + * else: + */ + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 414, __pyx_L1_error) + if (__pyx_t_1) { + + /* "View.MemoryView":415 + * cdef char *itemp + * if have_slices: + * return memview_slice(self, indices) # <<<<<<<<<<<<<< + * else: + * itemp = self.get_item_pointer(indices) + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = ((PyObject *)__pyx_memview_slice(__pyx_v_self, __pyx_v_indices)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 415, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":414 + * + * cdef char *itemp + * if have_slices: # <<<<<<<<<<<<<< + * return memview_slice(self, indices) + * else: + */ + } + + /* "View.MemoryView":417 + * return memview_slice(self, indices) + * else: + * itemp = self.get_item_pointer(indices) # <<<<<<<<<<<<<< + * return self.convert_item_to_object(itemp) + * + */ + /*else*/ { + __pyx_t_5 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_indices); if (unlikely(__pyx_t_5 == ((char *)NULL))) __PYX_ERR(1, 417, __pyx_L1_error) + __pyx_v_itemp = __pyx_t_5; + + /* "View.MemoryView":418 + * else: + * itemp = self.get_item_pointer(indices) + * return self.convert_item_to_object(itemp) # <<<<<<<<<<<<<< + * + * def __setitem__(memoryview self, object index, object value): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->convert_item_to_object(__pyx_v_self, __pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 418, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + } + + /* "View.MemoryView":407 + * + * + * def __getitem__(memoryview self, object index): # <<<<<<<<<<<<<< + * if index is Ellipsis: + * return self + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.memoryview.__getitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_have_slices); + __Pyx_XDECREF(__pyx_v_indices); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":420 + * return self.convert_item_to_object(itemp) + * + * def __setitem__(memoryview self, object index, object value): # <<<<<<<<<<<<<< + * if self.view.readonly: + * raise TypeError, "Cannot assign to read-only memoryview" + */ + +/* Python wrapper */ +static int __pyx_memoryview___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value); /*proto*/ +static int __pyx_memoryview___setitem__(PyObject *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setitem__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setitem__(((struct __pyx_memoryview_obj *)__pyx_v_self), ((PyObject *)__pyx_v_index), ((PyObject *)__pyx_v_value)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setitem__(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value) { + PyObject *__pyx_v_have_slices = NULL; + PyObject *__pyx_v_obj = NULL; + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setitem__", 0); + __Pyx_INCREF(__pyx_v_index); + + /* "View.MemoryView":421 + * + * def __setitem__(memoryview self, object index, object value): + * if self.view.readonly: # <<<<<<<<<<<<<< + * raise TypeError, "Cannot assign to read-only memoryview" + * + */ + if (unlikely(__pyx_v_self->view.readonly)) { + + /* "View.MemoryView":422 + * def __setitem__(memoryview self, object index, object value): + * if self.view.readonly: + * raise TypeError, "Cannot assign to read-only memoryview" # <<<<<<<<<<<<<< + * + * have_slices, index = _unellipsify(index, self.view.ndim) + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_Cannot_assign_to_read_only_memor, 0, 0); + __PYX_ERR(1, 422, __pyx_L1_error) + + /* "View.MemoryView":421 + * + * def __setitem__(memoryview self, object index, object value): + * if self.view.readonly: # <<<<<<<<<<<<<< + * raise TypeError, "Cannot assign to read-only memoryview" + * + */ + } + + /* "View.MemoryView":424 + * raise TypeError, "Cannot assign to read-only memoryview" + * + * have_slices, index = _unellipsify(index, self.view.ndim) # <<<<<<<<<<<<<< + * + * if have_slices: + */ + __pyx_t_1 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 424, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + if (likely(__pyx_t_1 != Py_None)) { + PyObject* sequence = __pyx_t_1; + Py_ssize_t size = __Pyx_PySequence_SIZE(sequence); + if (unlikely(size != 2)) { + if (size > 2) __Pyx_RaiseTooManyValuesError(2); + else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size); + __PYX_ERR(1, 424, __pyx_L1_error) + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_2 = PyTuple_GET_ITEM(sequence, 0); + __pyx_t_3 = PyTuple_GET_ITEM(sequence, 1); + __Pyx_INCREF(__pyx_t_2); + __Pyx_INCREF(__pyx_t_3); + #else + __pyx_t_2 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 424, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 424, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + #endif + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + } else { + __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 424, __pyx_L1_error) + } + __pyx_v_have_slices = __pyx_t_2; + __pyx_t_2 = 0; + __Pyx_DECREF_SET(__pyx_v_index, __pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":426 + * have_slices, index = _unellipsify(index, self.view.ndim) + * + * if have_slices: # <<<<<<<<<<<<<< + * obj = self.is_slice(value) + * if obj is not None: + */ + __pyx_t_4 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely((__pyx_t_4 < 0))) __PYX_ERR(1, 426, __pyx_L1_error) + if (__pyx_t_4) { + + /* "View.MemoryView":427 + * + * if have_slices: + * obj = self.is_slice(value) # <<<<<<<<<<<<<< + * if obj is not None: + * self.setitem_slice_assignment(self[index], obj) + */ + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->is_slice(__pyx_v_self, __pyx_v_value); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 427, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_obj = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":428 + * if have_slices: + * obj = self.is_slice(value) + * if obj is not None: # <<<<<<<<<<<<<< + * self.setitem_slice_assignment(self[index], obj) + * else: + */ + __pyx_t_4 = (__pyx_v_obj != Py_None); + if (__pyx_t_4) { + + /* "View.MemoryView":429 + * obj = self.is_slice(value) + * if obj is not None: + * self.setitem_slice_assignment(self[index], obj) # <<<<<<<<<<<<<< + * else: + * self.setitem_slice_assign_scalar(self[index], value) + */ + __pyx_t_1 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 429, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assignment(__pyx_v_self, __pyx_t_1, __pyx_v_obj); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 429, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + + /* "View.MemoryView":428 + * if have_slices: + * obj = self.is_slice(value) + * if obj is not None: # <<<<<<<<<<<<<< + * self.setitem_slice_assignment(self[index], obj) + * else: + */ + goto __pyx_L5; + } + + /* "View.MemoryView":431 + * self.setitem_slice_assignment(self[index], obj) + * else: + * self.setitem_slice_assign_scalar(self[index], value) # <<<<<<<<<<<<<< + * else: + * self.setitem_indexed(index, value) + */ + /*else*/ { + __pyx_t_3 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 431, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_memoryview_type))))) __PYX_ERR(1, 431, __pyx_L1_error) + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assign_scalar(__pyx_v_self, ((struct __pyx_memoryview_obj *)__pyx_t_3), __pyx_v_value); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 431, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + } + __pyx_L5:; + + /* "View.MemoryView":426 + * have_slices, index = _unellipsify(index, self.view.ndim) + * + * if have_slices: # <<<<<<<<<<<<<< + * obj = self.is_slice(value) + * if obj is not None: + */ + goto __pyx_L4; + } + + /* "View.MemoryView":433 + * self.setitem_slice_assign_scalar(self[index], value) + * else: + * self.setitem_indexed(index, value) # <<<<<<<<<<<<<< + * + * cdef is_slice(self, obj): + */ + /*else*/ { + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_indexed(__pyx_v_self, __pyx_v_index, __pyx_v_value); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 433, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + } + __pyx_L4:; + + /* "View.MemoryView":420 + * return self.convert_item_to_object(itemp) + * + * def __setitem__(memoryview self, object index, object value): # <<<<<<<<<<<<<< + * if self.view.readonly: + * raise TypeError, "Cannot assign to read-only memoryview" + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.__setitem__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_have_slices); + __Pyx_XDECREF(__pyx_v_obj); + __Pyx_XDECREF(__pyx_v_index); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":435 + * self.setitem_indexed(index, value) + * + * cdef is_slice(self, obj): # <<<<<<<<<<<<<< + * if not isinstance(obj, memoryview): + * try: + */ + +static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_obj) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_t_9; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("is_slice", 0); + __Pyx_INCREF(__pyx_v_obj); + + /* "View.MemoryView":436 + * + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): # <<<<<<<<<<<<<< + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + */ + __pyx_t_1 = __Pyx_TypeCheck(__pyx_v_obj, __pyx_memoryview_type); + __pyx_t_2 = (!__pyx_t_1); + if (__pyx_t_2) { + + /* "View.MemoryView":437 + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): + * try: # <<<<<<<<<<<<<< + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_3, &__pyx_t_4, &__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_4); + __Pyx_XGOTREF(__pyx_t_5); + /*try:*/ { + + /* "View.MemoryView":438 + * if not isinstance(obj, memoryview): + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, # <<<<<<<<<<<<<< + * self.dtype_is_object) + * except TypeError: + */ + __pyx_t_6 = __Pyx_PyInt_From_int(((__pyx_v_self->flags & (~PyBUF_WRITABLE)) | PyBUF_ANY_CONTIGUOUS)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 438, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_6); + + /* "View.MemoryView":439 + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) # <<<<<<<<<<<<<< + * except TypeError: + * return None + */ + __pyx_t_7 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 439, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_7); + + /* "View.MemoryView":438 + * if not isinstance(obj, memoryview): + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, # <<<<<<<<<<<<<< + * self.dtype_is_object) + * except TypeError: + */ + __pyx_t_8 = PyTuple_New(3); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 438, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_INCREF(__pyx_v_obj); + __Pyx_GIVEREF(__pyx_v_obj); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_8, 0, __pyx_v_obj)) __PYX_ERR(1, 438, __pyx_L4_error); + __Pyx_GIVEREF(__pyx_t_6); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_8, 1, __pyx_t_6)) __PYX_ERR(1, 438, __pyx_L4_error); + __Pyx_GIVEREF(__pyx_t_7); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_8, 2, __pyx_t_7)) __PYX_ERR(1, 438, __pyx_L4_error); + __pyx_t_6 = 0; + __pyx_t_7 = 0; + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_8, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 438, __pyx_L4_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __Pyx_DECREF_SET(__pyx_v_obj, __pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":437 + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): + * try: # <<<<<<<<<<<<<< + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + */ + } + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + goto __pyx_L9_try_end; + __pyx_L4_error:; + __Pyx_XDECREF(__pyx_t_6); __pyx_t_6 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_XDECREF(__pyx_t_8); __pyx_t_8 = 0; + + /* "View.MemoryView":440 + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + * except TypeError: # <<<<<<<<<<<<<< + * return None + * + */ + __pyx_t_9 = __Pyx_PyErr_ExceptionMatches(__pyx_builtin_TypeError); + if (__pyx_t_9) { + __Pyx_AddTraceback("View.MemoryView.memoryview.is_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_7, &__pyx_t_8, &__pyx_t_6) < 0) __PYX_ERR(1, 440, __pyx_L6_except_error) + __Pyx_XGOTREF(__pyx_t_7); + __Pyx_XGOTREF(__pyx_t_8); + __Pyx_XGOTREF(__pyx_t_6); + + /* "View.MemoryView":441 + * self.dtype_is_object) + * except TypeError: + * return None # <<<<<<<<<<<<<< + * + * return obj + */ + __Pyx_XDECREF(__pyx_r); + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L7_except_return; + } + goto __pyx_L6_except_error; + + /* "View.MemoryView":437 + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): + * try: # <<<<<<<<<<<<<< + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + * self.dtype_is_object) + */ + __pyx_L6_except_error:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_XGIVEREF(__pyx_t_5); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_4, __pyx_t_5); + goto __pyx_L1_error; + __pyx_L7_except_return:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_XGIVEREF(__pyx_t_5); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_4, __pyx_t_5); + goto __pyx_L0; + __pyx_L9_try_end:; + } + + /* "View.MemoryView":436 + * + * cdef is_slice(self, obj): + * if not isinstance(obj, memoryview): # <<<<<<<<<<<<<< + * try: + * obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS, + */ + } + + /* "View.MemoryView":443 + * return None + * + * return obj # <<<<<<<<<<<<<< + * + * cdef setitem_slice_assignment(self, dst, src): + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_obj); + __pyx_r = __pyx_v_obj; + goto __pyx_L0; + + /* "View.MemoryView":435 + * self.setitem_indexed(index, value) + * + * cdef is_slice(self, obj): # <<<<<<<<<<<<<< + * if not isinstance(obj, memoryview): + * try: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("View.MemoryView.memoryview.is_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_obj); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":445 + * return obj + * + * cdef setitem_slice_assignment(self, dst, src): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice dst_slice + * cdef __Pyx_memviewslice src_slice + */ + +static PyObject *__pyx_memoryview_setitem_slice_assignment(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_dst, PyObject *__pyx_v_src) { + __Pyx_memviewslice __pyx_v_dst_slice; + __Pyx_memviewslice __pyx_v_src_slice; + __Pyx_memviewslice __pyx_v_msrc; + __Pyx_memviewslice __pyx_v_mdst; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_t_3; + int __pyx_t_4; + int __pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("setitem_slice_assignment", 1); + + /* "View.MemoryView":448 + * cdef __Pyx_memviewslice dst_slice + * cdef __Pyx_memviewslice src_slice + * cdef __Pyx_memviewslice msrc = get_slice_from_memview(src, &src_slice)[0] # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice mdst = get_slice_from_memview(dst, &dst_slice)[0] + * + */ + if (!(likely(((__pyx_v_src) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_src, __pyx_memoryview_type))))) __PYX_ERR(1, 448, __pyx_L1_error) + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_src), (&__pyx_v_src_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 448, __pyx_L1_error) + __pyx_v_msrc = (__pyx_t_1[0]); + + /* "View.MemoryView":449 + * cdef __Pyx_memviewslice src_slice + * cdef __Pyx_memviewslice msrc = get_slice_from_memview(src, &src_slice)[0] + * cdef __Pyx_memviewslice mdst = get_slice_from_memview(dst, &dst_slice)[0] # <<<<<<<<<<<<<< + * + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) + */ + if (!(likely(((__pyx_v_dst) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_dst, __pyx_memoryview_type))))) __PYX_ERR(1, 449, __pyx_L1_error) + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_dst), (&__pyx_v_dst_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 449, __pyx_L1_error) + __pyx_v_mdst = (__pyx_t_1[0]); + + /* "View.MemoryView":451 + * cdef __Pyx_memviewslice mdst = get_slice_from_memview(dst, &dst_slice)[0] + * + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) # <<<<<<<<<<<<<< + * + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): + */ + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_v_src, __pyx_n_s_ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = __Pyx_PyInt_As_int(__pyx_t_2); if (unlikely((__pyx_t_3 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_v_dst, __pyx_n_s_ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_t_2); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 451, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_5 = __pyx_memoryview_copy_contents(__pyx_v_msrc, __pyx_v_mdst, __pyx_t_3, __pyx_t_4, __pyx_v_self->dtype_is_object); if (unlikely(__pyx_t_5 == ((int)-1))) __PYX_ERR(1, 451, __pyx_L1_error) + + /* "View.MemoryView":445 + * return obj + * + * cdef setitem_slice_assignment(self, dst, src): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice dst_slice + * cdef __Pyx_memviewslice src_slice + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.setitem_slice_assignment", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":453 + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) + * + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): # <<<<<<<<<<<<<< + * cdef int array[128] + * cdef void *tmp = NULL + */ + +static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memoryview_obj *__pyx_v_self, struct __pyx_memoryview_obj *__pyx_v_dst, PyObject *__pyx_v_value) { + int __pyx_v_array[0x80]; + void *__pyx_v_tmp; + void *__pyx_v_item; + __Pyx_memviewslice *__pyx_v_dst_slice; + __Pyx_memviewslice __pyx_v_tmp_slice; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + int __pyx_t_5; + char const *__pyx_t_6; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + PyObject *__pyx_t_9 = NULL; + PyObject *__pyx_t_10 = NULL; + PyObject *__pyx_t_11 = NULL; + PyObject *__pyx_t_12 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("setitem_slice_assign_scalar", 1); + + /* "View.MemoryView":455 + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): + * cdef int array[128] + * cdef void *tmp = NULL # <<<<<<<<<<<<<< + * cdef void *item + * + */ + __pyx_v_tmp = NULL; + + /* "View.MemoryView":460 + * cdef __Pyx_memviewslice *dst_slice + * cdef __Pyx_memviewslice tmp_slice + * dst_slice = get_slice_from_memview(dst, &tmp_slice) # <<<<<<<<<<<<<< + * + * if self.view.itemsize > sizeof(array): + */ + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_dst, (&__pyx_v_tmp_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 460, __pyx_L1_error) + __pyx_v_dst_slice = __pyx_t_1; + + /* "View.MemoryView":462 + * dst_slice = get_slice_from_memview(dst, &tmp_slice) + * + * if self.view.itemsize > sizeof(array): # <<<<<<<<<<<<<< + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: + */ + __pyx_t_2 = (((size_t)__pyx_v_self->view.itemsize) > (sizeof(__pyx_v_array))); + if (__pyx_t_2) { + + /* "View.MemoryView":463 + * + * if self.view.itemsize > sizeof(array): + * tmp = PyMem_Malloc(self.view.itemsize) # <<<<<<<<<<<<<< + * if tmp == NULL: + * raise MemoryError + */ + __pyx_v_tmp = PyMem_Malloc(__pyx_v_self->view.itemsize); + + /* "View.MemoryView":464 + * if self.view.itemsize > sizeof(array): + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * item = tmp + */ + __pyx_t_2 = (__pyx_v_tmp == NULL); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":465 + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: + * raise MemoryError # <<<<<<<<<<<<<< + * item = tmp + * else: + */ + PyErr_NoMemory(); __PYX_ERR(1, 465, __pyx_L1_error) + + /* "View.MemoryView":464 + * if self.view.itemsize > sizeof(array): + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: # <<<<<<<<<<<<<< + * raise MemoryError + * item = tmp + */ + } + + /* "View.MemoryView":466 + * if tmp == NULL: + * raise MemoryError + * item = tmp # <<<<<<<<<<<<<< + * else: + * item = array + */ + __pyx_v_item = __pyx_v_tmp; + + /* "View.MemoryView":462 + * dst_slice = get_slice_from_memview(dst, &tmp_slice) + * + * if self.view.itemsize > sizeof(array): # <<<<<<<<<<<<<< + * tmp = PyMem_Malloc(self.view.itemsize) + * if tmp == NULL: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":468 + * item = tmp + * else: + * item = array # <<<<<<<<<<<<<< + * + * try: + */ + /*else*/ { + __pyx_v_item = ((void *)__pyx_v_array); + } + __pyx_L3:; + + /* "View.MemoryView":470 + * item = array + * + * try: # <<<<<<<<<<<<<< + * if self.dtype_is_object: + * ( item)[0] = value + */ + /*try:*/ { + + /* "View.MemoryView":471 + * + * try: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * ( item)[0] = value + * else: + */ + if (__pyx_v_self->dtype_is_object) { + + /* "View.MemoryView":472 + * try: + * if self.dtype_is_object: + * ( item)[0] = value # <<<<<<<<<<<<<< + * else: + * self.assign_item_from_object( item, value) + */ + (((PyObject **)__pyx_v_item)[0]) = ((PyObject *)__pyx_v_value); + + /* "View.MemoryView":471 + * + * try: + * if self.dtype_is_object: # <<<<<<<<<<<<<< + * ( item)[0] = value + * else: + */ + goto __pyx_L8; + } + + /* "View.MemoryView":474 + * ( item)[0] = value + * else: + * self.assign_item_from_object( item, value) # <<<<<<<<<<<<<< + * + * + */ + /*else*/ { + __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, ((char *)__pyx_v_item), __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 474, __pyx_L6_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + __pyx_L8:; + + /* "View.MemoryView":478 + * + * + * if self.view.suboffsets != NULL: # <<<<<<<<<<<<<< + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, + */ + __pyx_t_2 = (__pyx_v_self->view.suboffsets != NULL); + if (__pyx_t_2) { + + /* "View.MemoryView":479 + * + * if self.view.suboffsets != NULL: + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) # <<<<<<<<<<<<<< + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, + * item, self.dtype_is_object) + */ + __pyx_t_4 = assert_direct_dimensions(__pyx_v_self->view.suboffsets, __pyx_v_self->view.ndim); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 479, __pyx_L6_error) + + /* "View.MemoryView":478 + * + * + * if self.view.suboffsets != NULL: # <<<<<<<<<<<<<< + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, + */ + } + + /* "View.MemoryView":480 + * if self.view.suboffsets != NULL: + * assert_direct_dimensions(self.view.suboffsets, self.view.ndim) + * slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize, # <<<<<<<<<<<<<< + * item, self.dtype_is_object) + * finally: + */ + __pyx_memoryview_slice_assign_scalar(__pyx_v_dst_slice, __pyx_v_dst->view.ndim, __pyx_v_self->view.itemsize, __pyx_v_item, __pyx_v_self->dtype_is_object); + } + + /* "View.MemoryView":483 + * item, self.dtype_is_object) + * finally: + * PyMem_Free(tmp) # <<<<<<<<<<<<<< + * + * cdef setitem_indexed(self, index, value): + */ + /*finally:*/ { + /*normal exit:*/{ + PyMem_Free(__pyx_v_tmp); + goto __pyx_L7; + } + __pyx_L6_error:; + /*exception exit:*/{ + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __pyx_t_7 = 0; __pyx_t_8 = 0; __pyx_t_9 = 0; __pyx_t_10 = 0; __pyx_t_11 = 0; __pyx_t_12 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + if (PY_MAJOR_VERSION >= 3) __Pyx_ExceptionSwap(&__pyx_t_10, &__pyx_t_11, &__pyx_t_12); + if ((PY_MAJOR_VERSION < 3) || unlikely(__Pyx_GetException(&__pyx_t_7, &__pyx_t_8, &__pyx_t_9) < 0)) __Pyx_ErrFetch(&__pyx_t_7, &__pyx_t_8, &__pyx_t_9); + __Pyx_XGOTREF(__pyx_t_7); + __Pyx_XGOTREF(__pyx_t_8); + __Pyx_XGOTREF(__pyx_t_9); + __Pyx_XGOTREF(__pyx_t_10); + __Pyx_XGOTREF(__pyx_t_11); + __Pyx_XGOTREF(__pyx_t_12); + __pyx_t_4 = __pyx_lineno; __pyx_t_5 = __pyx_clineno; __pyx_t_6 = __pyx_filename; + { + PyMem_Free(__pyx_v_tmp); + } + if (PY_MAJOR_VERSION >= 3) { + __Pyx_XGIVEREF(__pyx_t_10); + __Pyx_XGIVEREF(__pyx_t_11); + __Pyx_XGIVEREF(__pyx_t_12); + __Pyx_ExceptionReset(__pyx_t_10, __pyx_t_11, __pyx_t_12); + } + __Pyx_XGIVEREF(__pyx_t_7); + __Pyx_XGIVEREF(__pyx_t_8); + __Pyx_XGIVEREF(__pyx_t_9); + __Pyx_ErrRestore(__pyx_t_7, __pyx_t_8, __pyx_t_9); + __pyx_t_7 = 0; __pyx_t_8 = 0; __pyx_t_9 = 0; __pyx_t_10 = 0; __pyx_t_11 = 0; __pyx_t_12 = 0; + __pyx_lineno = __pyx_t_4; __pyx_clineno = __pyx_t_5; __pyx_filename = __pyx_t_6; + goto __pyx_L1_error; + } + __pyx_L7:; + } + + /* "View.MemoryView":453 + * memoryview_copy_contents(msrc, mdst, src.ndim, dst.ndim, self.dtype_is_object) + * + * cdef setitem_slice_assign_scalar(self, memoryview dst, value): # <<<<<<<<<<<<<< + * cdef int array[128] + * cdef void *tmp = NULL + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.setitem_slice_assign_scalar", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":485 + * PyMem_Free(tmp) + * + * cdef setitem_indexed(self, index, value): # <<<<<<<<<<<<<< + * cdef char *itemp = self.get_item_pointer(index) + * self.assign_item_from_object(itemp, value) + */ + +static PyObject *__pyx_memoryview_setitem_indexed(struct __pyx_memoryview_obj *__pyx_v_self, PyObject *__pyx_v_index, PyObject *__pyx_v_value) { + char *__pyx_v_itemp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + char *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("setitem_indexed", 1); + + /* "View.MemoryView":486 + * + * cdef setitem_indexed(self, index, value): + * cdef char *itemp = self.get_item_pointer(index) # <<<<<<<<<<<<<< + * self.assign_item_from_object(itemp, value) + * + */ + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_index); if (unlikely(__pyx_t_1 == ((char *)NULL))) __PYX_ERR(1, 486, __pyx_L1_error) + __pyx_v_itemp = __pyx_t_1; + + /* "View.MemoryView":487 + * cdef setitem_indexed(self, index, value): + * cdef char *itemp = self.get_item_pointer(index) + * self.assign_item_from_object(itemp, value) # <<<<<<<<<<<<<< + * + * cdef convert_item_to_object(self, char *itemp): + */ + __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 487, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":485 + * PyMem_Free(tmp) + * + * cdef setitem_indexed(self, index, value): # <<<<<<<<<<<<<< + * cdef char *itemp = self.get_item_pointer(index) + * self.assign_item_from_object(itemp, value) + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.setitem_indexed", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":489 + * self.assign_item_from_object(itemp, value) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + +static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp) { + PyObject *__pyx_v_struct = NULL; + PyObject *__pyx_v_bytesitem = 0; + PyObject *__pyx_v_result = NULL; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + unsigned int __pyx_t_8; + Py_ssize_t __pyx_t_9; + int __pyx_t_10; + int __pyx_t_11; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("convert_item_to_object", 1); + + /* "View.MemoryView":492 + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + * import struct # <<<<<<<<<<<<<< + * cdef bytes bytesitem + * + */ + __pyx_t_1 = __Pyx_ImportDottedModule(__pyx_n_s_struct, NULL); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 492, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_struct = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":495 + * cdef bytes bytesitem + * + * bytesitem = itemp[:self.view.itemsize] # <<<<<<<<<<<<<< + * try: + * result = struct.unpack(self.view.format, bytesitem) + */ + __pyx_t_1 = __Pyx_PyBytes_FromStringAndSize(__pyx_v_itemp + 0, __pyx_v_self->view.itemsize - 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 495, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_bytesitem = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":496 + * + * bytesitem = itemp[:self.view.itemsize] + * try: # <<<<<<<<<<<<<< + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_2, &__pyx_t_3, &__pyx_t_4); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_4); + /*try:*/ { + + /* "View.MemoryView":497 + * bytesitem = itemp[:self.view.itemsize] + * try: + * result = struct.unpack(self.view.format, bytesitem) # <<<<<<<<<<<<<< + * except struct.error: + * raise ValueError, "Unable to convert item to object" + */ + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_unpack); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 497, __pyx_L3_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_6 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 497, __pyx_L3_error) + __Pyx_GOTREF(__pyx_t_6); + __pyx_t_7 = NULL; + __pyx_t_8 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_5))) { + __pyx_t_7 = PyMethod_GET_SELF(__pyx_t_5); + if (likely(__pyx_t_7)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_5); + __Pyx_INCREF(__pyx_t_7); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_5, function); + __pyx_t_8 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[3] = {__pyx_t_7, __pyx_t_6, __pyx_v_bytesitem}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_5, __pyx_callargs+1-__pyx_t_8, 2+__pyx_t_8); + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 497, __pyx_L3_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + __pyx_v_result = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":496 + * + * bytesitem = itemp[:self.view.itemsize] + * try: # <<<<<<<<<<<<<< + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + */ + } + + /* "View.MemoryView":501 + * raise ValueError, "Unable to convert item to object" + * else: + * if len(self.view.format) == 1: # <<<<<<<<<<<<<< + * return result[0] + * return result + */ + /*else:*/ { + __pyx_t_9 = __Pyx_ssize_strlen(__pyx_v_self->view.format); if (unlikely(__pyx_t_9 == ((Py_ssize_t)-1))) __PYX_ERR(1, 501, __pyx_L5_except_error) + __pyx_t_10 = (__pyx_t_9 == 1); + if (__pyx_t_10) { + + /* "View.MemoryView":502 + * else: + * if len(self.view.format) == 1: + * return result[0] # <<<<<<<<<<<<<< + * return result + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_GetItemInt(__pyx_v_result, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 502, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L6_except_return; + + /* "View.MemoryView":501 + * raise ValueError, "Unable to convert item to object" + * else: + * if len(self.view.format) == 1: # <<<<<<<<<<<<<< + * return result[0] + * return result + */ + } + + /* "View.MemoryView":503 + * if len(self.view.format) == 1: + * return result[0] + * return result # <<<<<<<<<<<<<< + * + * cdef assign_item_from_object(self, char *itemp, object value): + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_result); + __pyx_r = __pyx_v_result; + goto __pyx_L6_except_return; + } + __pyx_L3_error:; + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_6); __pyx_t_6 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":498 + * try: + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: # <<<<<<<<<<<<<< + * raise ValueError, "Unable to convert item to object" + * else: + */ + __Pyx_ErrFetch(&__pyx_t_1, &__pyx_t_5, &__pyx_t_6); + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_error); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 498, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_11 = __Pyx_PyErr_GivenExceptionMatches(__pyx_t_1, __pyx_t_7); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_ErrRestore(__pyx_t_1, __pyx_t_5, __pyx_t_6); + __pyx_t_1 = 0; __pyx_t_5 = 0; __pyx_t_6 = 0; + if (__pyx_t_11) { + __Pyx_AddTraceback("View.MemoryView.memoryview.convert_item_to_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_6, &__pyx_t_5, &__pyx_t_1) < 0) __PYX_ERR(1, 498, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_1); + + /* "View.MemoryView":499 + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + * raise ValueError, "Unable to convert item to object" # <<<<<<<<<<<<<< + * else: + * if len(self.view.format) == 1: + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Unable_to_convert_item_to_object, 0, 0); + __PYX_ERR(1, 499, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "View.MemoryView":496 + * + * bytesitem = itemp[:self.view.itemsize] + * try: # <<<<<<<<<<<<<< + * result = struct.unpack(self.view.format, bytesitem) + * except struct.error: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_ExceptionReset(__pyx_t_2, __pyx_t_3, __pyx_t_4); + goto __pyx_L1_error; + __pyx_L6_except_return:; + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_4); + __Pyx_ExceptionReset(__pyx_t_2, __pyx_t_3, __pyx_t_4); + goto __pyx_L0; + } + + /* "View.MemoryView":489 + * self.assign_item_from_object(itemp, value) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_AddTraceback("View.MemoryView.memoryview.convert_item_to_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_struct); + __Pyx_XDECREF(__pyx_v_bytesitem); + __Pyx_XDECREF(__pyx_v_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":505 + * return result + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + +static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryview_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value) { + PyObject *__pyx_v_struct = NULL; + char __pyx_v_c; + PyObject *__pyx_v_bytesvalue = 0; + Py_ssize_t __pyx_v_i; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + unsigned int __pyx_t_6; + Py_ssize_t __pyx_t_7; + PyObject *__pyx_t_8 = NULL; + char *__pyx_t_9; + char *__pyx_t_10; + char *__pyx_t_11; + char *__pyx_t_12; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("assign_item_from_object", 1); + + /* "View.MemoryView":508 + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + * import struct # <<<<<<<<<<<<<< + * cdef char c + * cdef bytes bytesvalue + */ + __pyx_t_1 = __Pyx_ImportDottedModule(__pyx_n_s_struct, NULL); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 508, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_struct = __pyx_t_1; + __pyx_t_1 = 0; + + /* "View.MemoryView":513 + * cdef Py_ssize_t i + * + * if isinstance(value, tuple): # <<<<<<<<<<<<<< + * bytesvalue = struct.pack(self.view.format, *value) + * else: + */ + __pyx_t_2 = PyTuple_Check(__pyx_v_value); + if (__pyx_t_2) { + + /* "View.MemoryView":514 + * + * if isinstance(value, tuple): + * bytesvalue = struct.pack(self.view.format, *value) # <<<<<<<<<<<<<< + * else: + * bytesvalue = struct.pack(self.view.format, value) + */ + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = PyTuple_New(1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_3 = __Pyx_PySequence_Tuple(__pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_5 = PyNumber_Add(__pyx_t_4, __pyx_t_3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_t_3 = __Pyx_PyObject_Call(__pyx_t_1, __pyx_t_5, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 514, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + if (!(likely(PyBytes_CheckExact(__pyx_t_3))||((__pyx_t_3) == Py_None) || __Pyx_RaiseUnexpectedTypeError("bytes", __pyx_t_3))) __PYX_ERR(1, 514, __pyx_L1_error) + __pyx_v_bytesvalue = ((PyObject*)__pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":513 + * cdef Py_ssize_t i + * + * if isinstance(value, tuple): # <<<<<<<<<<<<<< + * bytesvalue = struct.pack(self.view.format, *value) + * else: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":516 + * bytesvalue = struct.pack(self.view.format, *value) + * else: + * bytesvalue = struct.pack(self.view.format, value) # <<<<<<<<<<<<<< + * + * for i, c in enumerate(bytesvalue): + */ + /*else*/ { + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 516, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_1 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 516, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_4 = NULL; + __pyx_t_6 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_5))) { + __pyx_t_4 = PyMethod_GET_SELF(__pyx_t_5); + if (likely(__pyx_t_4)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_5); + __Pyx_INCREF(__pyx_t_4); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_5, function); + __pyx_t_6 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[3] = {__pyx_t_4, __pyx_t_1, __pyx_v_value}; + __pyx_t_3 = __Pyx_PyObject_FastCall(__pyx_t_5, __pyx_callargs+1-__pyx_t_6, 2+__pyx_t_6); + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 516, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + if (!(likely(PyBytes_CheckExact(__pyx_t_3))||((__pyx_t_3) == Py_None) || __Pyx_RaiseUnexpectedTypeError("bytes", __pyx_t_3))) __PYX_ERR(1, 516, __pyx_L1_error) + __pyx_v_bytesvalue = ((PyObject*)__pyx_t_3); + __pyx_t_3 = 0; + } + __pyx_L3:; + + /* "View.MemoryView":518 + * bytesvalue = struct.pack(self.view.format, value) + * + * for i, c in enumerate(bytesvalue): # <<<<<<<<<<<<<< + * itemp[i] = c + * + */ + __pyx_t_7 = 0; + if (unlikely(__pyx_v_bytesvalue == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' is not iterable"); + __PYX_ERR(1, 518, __pyx_L1_error) + } + __Pyx_INCREF(__pyx_v_bytesvalue); + __pyx_t_8 = __pyx_v_bytesvalue; + __pyx_t_10 = PyBytes_AS_STRING(__pyx_t_8); + __pyx_t_11 = (__pyx_t_10 + PyBytes_GET_SIZE(__pyx_t_8)); + for (__pyx_t_12 = __pyx_t_10; __pyx_t_12 < __pyx_t_11; __pyx_t_12++) { + __pyx_t_9 = __pyx_t_12; + __pyx_v_c = (__pyx_t_9[0]); + + /* "View.MemoryView":519 + * + * for i, c in enumerate(bytesvalue): + * itemp[i] = c # <<<<<<<<<<<<<< + * + * @cname('getbuffer') + */ + __pyx_v_i = __pyx_t_7; + + /* "View.MemoryView":518 + * bytesvalue = struct.pack(self.view.format, value) + * + * for i, c in enumerate(bytesvalue): # <<<<<<<<<<<<<< + * itemp[i] = c + * + */ + __pyx_t_7 = (__pyx_t_7 + 1); + + /* "View.MemoryView":519 + * + * for i, c in enumerate(bytesvalue): + * itemp[i] = c # <<<<<<<<<<<<<< + * + * @cname('getbuffer') + */ + (__pyx_v_itemp[__pyx_v_i]) = __pyx_v_c; + } + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + + /* "View.MemoryView":505 + * return result + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * """Only used if instantiated manually by the user, or if Cython doesn't + * know how to convert the type""" + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("View.MemoryView.memoryview.assign_item_from_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_struct); + __Pyx_XDECREF(__pyx_v_bytesvalue); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":521 + * itemp[i] = c + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: + */ + +/* Python wrapper */ +CYTHON_UNUSED static int __pyx_memoryview_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags); /*proto*/ +CYTHON_UNUSED static int __pyx_memoryview_getbuffer(PyObject *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__getbuffer__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbuffer__(((struct __pyx_memoryview_obj *)__pyx_v_self), ((Py_buffer *)__pyx_v_info), ((int)__pyx_v_flags)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbuffer__(struct __pyx_memoryview_obj *__pyx_v_self, Py_buffer *__pyx_v_info, int __pyx_v_flags) { + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + Py_ssize_t *__pyx_t_3; + char *__pyx_t_4; + void *__pyx_t_5; + int __pyx_t_6; + Py_ssize_t __pyx_t_7; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + if (unlikely(__pyx_v_info == NULL)) { + PyErr_SetString(PyExc_BufferError, "PyObject_GetBuffer: view==NULL argument is obsolete"); + return -1; + } + __Pyx_RefNannySetupContext("__getbuffer__", 0); + __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(__pyx_v_info->obj); + + /* "View.MemoryView":523 + * @cname('getbuffer') + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: # <<<<<<<<<<<<<< + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + */ + __pyx_t_2 = ((__pyx_v_flags & PyBUF_WRITABLE) != 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_1 = __pyx_v_self->view.readonly; + __pyx_L4_bool_binop_done:; + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":524 + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: + * raise ValueError, "Cannot create writable memory view from read-only memoryview" # <<<<<<<<<<<<<< + * + * if flags & PyBUF_ND: + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Cannot_create_writable_memory_vi, 0, 0); + __PYX_ERR(1, 524, __pyx_L1_error) + + /* "View.MemoryView":523 + * @cname('getbuffer') + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: # <<<<<<<<<<<<<< + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + */ + } + + /* "View.MemoryView":526 + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + * if flags & PyBUF_ND: # <<<<<<<<<<<<<< + * info.shape = self.view.shape + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_ND) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":527 + * + * if flags & PyBUF_ND: + * info.shape = self.view.shape # <<<<<<<<<<<<<< + * else: + * info.shape = NULL + */ + __pyx_t_3 = __pyx_v_self->view.shape; + __pyx_v_info->shape = __pyx_t_3; + + /* "View.MemoryView":526 + * raise ValueError, "Cannot create writable memory view from read-only memoryview" + * + * if flags & PyBUF_ND: # <<<<<<<<<<<<<< + * info.shape = self.view.shape + * else: + */ + goto __pyx_L6; + } + + /* "View.MemoryView":529 + * info.shape = self.view.shape + * else: + * info.shape = NULL # <<<<<<<<<<<<<< + * + * if flags & PyBUF_STRIDES: + */ + /*else*/ { + __pyx_v_info->shape = NULL; + } + __pyx_L6:; + + /* "View.MemoryView":531 + * info.shape = NULL + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.strides = self.view.strides + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_STRIDES) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":532 + * + * if flags & PyBUF_STRIDES: + * info.strides = self.view.strides # <<<<<<<<<<<<<< + * else: + * info.strides = NULL + */ + __pyx_t_3 = __pyx_v_self->view.strides; + __pyx_v_info->strides = __pyx_t_3; + + /* "View.MemoryView":531 + * info.shape = NULL + * + * if flags & PyBUF_STRIDES: # <<<<<<<<<<<<<< + * info.strides = self.view.strides + * else: + */ + goto __pyx_L7; + } + + /* "View.MemoryView":534 + * info.strides = self.view.strides + * else: + * info.strides = NULL # <<<<<<<<<<<<<< + * + * if flags & PyBUF_INDIRECT: + */ + /*else*/ { + __pyx_v_info->strides = NULL; + } + __pyx_L7:; + + /* "View.MemoryView":536 + * info.strides = NULL + * + * if flags & PyBUF_INDIRECT: # <<<<<<<<<<<<<< + * info.suboffsets = self.view.suboffsets + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_INDIRECT) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":537 + * + * if flags & PyBUF_INDIRECT: + * info.suboffsets = self.view.suboffsets # <<<<<<<<<<<<<< + * else: + * info.suboffsets = NULL + */ + __pyx_t_3 = __pyx_v_self->view.suboffsets; + __pyx_v_info->suboffsets = __pyx_t_3; + + /* "View.MemoryView":536 + * info.strides = NULL + * + * if flags & PyBUF_INDIRECT: # <<<<<<<<<<<<<< + * info.suboffsets = self.view.suboffsets + * else: + */ + goto __pyx_L8; + } + + /* "View.MemoryView":539 + * info.suboffsets = self.view.suboffsets + * else: + * info.suboffsets = NULL # <<<<<<<<<<<<<< + * + * if flags & PyBUF_FORMAT: + */ + /*else*/ { + __pyx_v_info->suboffsets = NULL; + } + __pyx_L8:; + + /* "View.MemoryView":541 + * info.suboffsets = NULL + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * info.format = self.view.format + * else: + */ + __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":542 + * + * if flags & PyBUF_FORMAT: + * info.format = self.view.format # <<<<<<<<<<<<<< + * else: + * info.format = NULL + */ + __pyx_t_4 = __pyx_v_self->view.format; + __pyx_v_info->format = __pyx_t_4; + + /* "View.MemoryView":541 + * info.suboffsets = NULL + * + * if flags & PyBUF_FORMAT: # <<<<<<<<<<<<<< + * info.format = self.view.format + * else: + */ + goto __pyx_L9; + } + + /* "View.MemoryView":544 + * info.format = self.view.format + * else: + * info.format = NULL # <<<<<<<<<<<<<< + * + * info.buf = self.view.buf + */ + /*else*/ { + __pyx_v_info->format = NULL; + } + __pyx_L9:; + + /* "View.MemoryView":546 + * info.format = NULL + * + * info.buf = self.view.buf # <<<<<<<<<<<<<< + * info.ndim = self.view.ndim + * info.itemsize = self.view.itemsize + */ + __pyx_t_5 = __pyx_v_self->view.buf; + __pyx_v_info->buf = __pyx_t_5; + + /* "View.MemoryView":547 + * + * info.buf = self.view.buf + * info.ndim = self.view.ndim # <<<<<<<<<<<<<< + * info.itemsize = self.view.itemsize + * info.len = self.view.len + */ + __pyx_t_6 = __pyx_v_self->view.ndim; + __pyx_v_info->ndim = __pyx_t_6; + + /* "View.MemoryView":548 + * info.buf = self.view.buf + * info.ndim = self.view.ndim + * info.itemsize = self.view.itemsize # <<<<<<<<<<<<<< + * info.len = self.view.len + * info.readonly = self.view.readonly + */ + __pyx_t_7 = __pyx_v_self->view.itemsize; + __pyx_v_info->itemsize = __pyx_t_7; + + /* "View.MemoryView":549 + * info.ndim = self.view.ndim + * info.itemsize = self.view.itemsize + * info.len = self.view.len # <<<<<<<<<<<<<< + * info.readonly = self.view.readonly + * info.obj = self + */ + __pyx_t_7 = __pyx_v_self->view.len; + __pyx_v_info->len = __pyx_t_7; + + /* "View.MemoryView":550 + * info.itemsize = self.view.itemsize + * info.len = self.view.len + * info.readonly = self.view.readonly # <<<<<<<<<<<<<< + * info.obj = self + * + */ + __pyx_t_1 = __pyx_v_self->view.readonly; + __pyx_v_info->readonly = __pyx_t_1; + + /* "View.MemoryView":551 + * info.len = self.view.len + * info.readonly = self.view.readonly + * info.obj = self # <<<<<<<<<<<<<< + * + * + */ + __Pyx_INCREF((PyObject *)__pyx_v_self); + __Pyx_GIVEREF((PyObject *)__pyx_v_self); + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); + __pyx_v_info->obj = ((PyObject *)__pyx_v_self); + + /* "View.MemoryView":521 + * itemp[i] = c + * + * @cname('getbuffer') # <<<<<<<<<<<<<< + * def __getbuffer__(self, Py_buffer *info, int flags): + * if flags & PyBUF_WRITABLE and self.view.readonly: + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__getbuffer__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + if (__pyx_v_info->obj != NULL) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + goto __pyx_L2; + __pyx_L0:; + if (__pyx_v_info->obj == Py_None) { + __Pyx_GOTREF(__pyx_v_info->obj); + __Pyx_DECREF(__pyx_v_info->obj); __pyx_v_info->obj = 0; + } + __pyx_L2:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":554 + * + * + * @property # <<<<<<<<<<<<<< + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_1T_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_1T_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + struct __pyx_memoryviewslice_obj *__pyx_v_result = 0; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":556 + * @property + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) # <<<<<<<<<<<<<< + * transpose_memslice(&result.from_slice) + * return result + */ + __pyx_t_1 = __pyx_memoryview_copy_object(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 556, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + if (!(likely(((__pyx_t_1) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_1, __pyx_memoryviewslice_type))))) __PYX_ERR(1, 556, __pyx_L1_error) + __pyx_v_result = ((struct __pyx_memoryviewslice_obj *)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":557 + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) + * transpose_memslice(&result.from_slice) # <<<<<<<<<<<<<< + * return result + * + */ + __pyx_t_2 = __pyx_memslice_transpose((&__pyx_v_result->from_slice)); if (unlikely(__pyx_t_2 == ((int)-1))) __PYX_ERR(1, 557, __pyx_L1_error) + + /* "View.MemoryView":558 + * cdef _memoryviewslice result = memoryview_copy(self) + * transpose_memslice(&result.from_slice) + * return result # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = ((PyObject *)__pyx_v_result); + goto __pyx_L0; + + /* "View.MemoryView":554 + * + * + * @property # <<<<<<<<<<<<<< + * def T(self): + * cdef _memoryviewslice result = memoryview_copy(self) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.T.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":560 + * return result + * + * @property # <<<<<<<<<<<<<< + * def base(self): + * return self._get_base() + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4base_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4base_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":562 + * @property + * def base(self): + * return self._get_base() # <<<<<<<<<<<<<< + * + * cdef _get_base(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->_get_base(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 562, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":560 + * return result + * + * @property # <<<<<<<<<<<<<< + * def base(self): + * return self._get_base() + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.base.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":564 + * return self._get_base() + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.obj + * + */ + +static PyObject *__pyx_memoryview__get_base(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("_get_base", 1); + + /* "View.MemoryView":565 + * + * cdef _get_base(self): + * return self.obj # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->obj); + __pyx_r = __pyx_v_self->obj; + goto __pyx_L0; + + /* "View.MemoryView":564 + * return self._get_base() + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.obj + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":567 + * return self.obj + * + * @property # <<<<<<<<<<<<<< + * def shape(self): + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_5shape_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_5shape_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_7genexpr__pyx_v_length; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + PyObject *__pyx_t_5 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":569 + * @property + * def shape(self): + * return tuple([length for length in self.view.shape[:self.view.ndim]]) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + { /* enter inner scope */ + __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = (__pyx_v_self->view.shape + __pyx_v_self->view.ndim); + for (__pyx_t_4 = __pyx_v_self->view.shape; __pyx_t_4 < __pyx_t_3; __pyx_t_4++) { + __pyx_t_2 = __pyx_t_4; + __pyx_7genexpr__pyx_v_length = (__pyx_t_2[0]); + __pyx_t_5 = PyInt_FromSsize_t(__pyx_7genexpr__pyx_v_length); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + if (unlikely(__Pyx_ListComp_Append(__pyx_t_1, (PyObject*)__pyx_t_5))) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + } /* exit inner scope */ + __pyx_t_5 = PyList_AsTuple(((PyObject*)__pyx_t_1)); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 569, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_r = __pyx_t_5; + __pyx_t_5 = 0; + goto __pyx_L0; + + /* "View.MemoryView":567 + * return self.obj + * + * @property # <<<<<<<<<<<<<< + * def shape(self): + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.memoryview.shape.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":571 + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def strides(self): + * if self.view.strides == NULL: + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_7strides_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_7strides_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_8genexpr1__pyx_v_stride; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + Py_ssize_t *__pyx_t_5; + PyObject *__pyx_t_6 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":573 + * @property + * def strides(self): + * if self.view.strides == NULL: # <<<<<<<<<<<<<< + * + * raise ValueError, "Buffer view does not expose strides" + */ + __pyx_t_1 = (__pyx_v_self->view.strides == NULL); + if (unlikely(__pyx_t_1)) { + + /* "View.MemoryView":575 + * if self.view.strides == NULL: + * + * raise ValueError, "Buffer view does not expose strides" # <<<<<<<<<<<<<< + * + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Buffer_view_does_not_expose_stri, 0, 0); + __PYX_ERR(1, 575, __pyx_L1_error) + + /* "View.MemoryView":573 + * @property + * def strides(self): + * if self.view.strides == NULL: # <<<<<<<<<<<<<< + * + * raise ValueError, "Buffer view does not expose strides" + */ + } + + /* "View.MemoryView":577 + * raise ValueError, "Buffer view does not expose strides" + * + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + { /* enter inner scope */ + __pyx_t_2 = PyList_New(0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = (__pyx_v_self->view.strides + __pyx_v_self->view.ndim); + for (__pyx_t_5 = __pyx_v_self->view.strides; __pyx_t_5 < __pyx_t_4; __pyx_t_5++) { + __pyx_t_3 = __pyx_t_5; + __pyx_8genexpr1__pyx_v_stride = (__pyx_t_3[0]); + __pyx_t_6 = PyInt_FromSsize_t(__pyx_8genexpr1__pyx_v_stride); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + if (unlikely(__Pyx_ListComp_Append(__pyx_t_2, (PyObject*)__pyx_t_6))) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + } + } /* exit inner scope */ + __pyx_t_6 = PyList_AsTuple(((PyObject*)__pyx_t_2)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 577, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_6; + __pyx_t_6 = 0; + goto __pyx_L0; + + /* "View.MemoryView":571 + * return tuple([length for length in self.view.shape[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def strides(self): + * if self.view.strides == NULL: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_AddTraceback("View.MemoryView.memoryview.strides.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":579 + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def suboffsets(self): + * if self.view.suboffsets == NULL: + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_10suboffsets_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_10suboffsets_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_8genexpr2__pyx_v_suboffset; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + Py_ssize_t *__pyx_t_5; + PyObject *__pyx_t_6 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":581 + * @property + * def suboffsets(self): + * if self.view.suboffsets == NULL: # <<<<<<<<<<<<<< + * return (-1,) * self.view.ndim + * + */ + __pyx_t_1 = (__pyx_v_self->view.suboffsets == NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":582 + * def suboffsets(self): + * if self.view.suboffsets == NULL: + * return (-1,) * self.view.ndim # <<<<<<<<<<<<<< + * + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __Pyx_PySequence_Multiply(__pyx_tuple__4, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 582, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":581 + * @property + * def suboffsets(self): + * if self.view.suboffsets == NULL: # <<<<<<<<<<<<<< + * return (-1,) * self.view.ndim + * + */ + } + + /* "View.MemoryView":584 + * return (-1,) * self.view.ndim + * + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + { /* enter inner scope */ + __pyx_t_2 = PyList_New(0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_4 = (__pyx_v_self->view.suboffsets + __pyx_v_self->view.ndim); + for (__pyx_t_5 = __pyx_v_self->view.suboffsets; __pyx_t_5 < __pyx_t_4; __pyx_t_5++) { + __pyx_t_3 = __pyx_t_5; + __pyx_8genexpr2__pyx_v_suboffset = (__pyx_t_3[0]); + __pyx_t_6 = PyInt_FromSsize_t(__pyx_8genexpr2__pyx_v_suboffset); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + if (unlikely(__Pyx_ListComp_Append(__pyx_t_2, (PyObject*)__pyx_t_6))) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + } + } /* exit inner scope */ + __pyx_t_6 = PyList_AsTuple(((PyObject*)__pyx_t_2)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 584, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_6; + __pyx_t_6 = 0; + goto __pyx_L0; + + /* "View.MemoryView":579 + * return tuple([stride for stride in self.view.strides[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def suboffsets(self): + * if self.view.suboffsets == NULL: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_AddTraceback("View.MemoryView.memoryview.suboffsets.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":586 + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def ndim(self): + * return self.view.ndim + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4ndim_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4ndim_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":588 + * @property + * def ndim(self): + * return self.view.ndim # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_self->view.ndim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 588, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":586 + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + * + * @property # <<<<<<<<<<<<<< + * def ndim(self): + * return self.view.ndim + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.ndim.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":590 + * return self.view.ndim + * + * @property # <<<<<<<<<<<<<< + * def itemsize(self): + * return self.view.itemsize + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_8itemsize_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_8itemsize_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":592 + * @property + * def itemsize(self): + * return self.view.itemsize # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 592, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":590 + * return self.view.ndim + * + * @property # <<<<<<<<<<<<<< + * def itemsize(self): + * return self.view.itemsize + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview.itemsize.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":594 + * return self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def nbytes(self): + * return self.size * self.view.itemsize + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_6nbytes_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_6nbytes_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":596 + * @property + * def nbytes(self): + * return self.size * self.view.itemsize # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_size); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 596, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 596, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyNumber_Multiply(__pyx_t_1, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 596, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_3; + __pyx_t_3 = 0; + goto __pyx_L0; + + /* "View.MemoryView":594 + * return self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def nbytes(self): + * return self.size * self.view.itemsize + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.nbytes.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":598 + * return self.size * self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def size(self): + * if self._size is None: + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4size_1__get__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_pw_15View_dot_MemoryView_10memoryview_4size_1__get__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__get__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_v_result = NULL; + PyObject *__pyx_v_length = NULL; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + PyObject *__pyx_t_5 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__get__", 1); + + /* "View.MemoryView":600 + * @property + * def size(self): + * if self._size is None: # <<<<<<<<<<<<<< + * result = 1 + * + */ + __pyx_t_1 = (__pyx_v_self->_size == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":601 + * def size(self): + * if self._size is None: + * result = 1 # <<<<<<<<<<<<<< + * + * for length in self.view.shape[:self.view.ndim]: + */ + __Pyx_INCREF(__pyx_int_1); + __pyx_v_result = __pyx_int_1; + + /* "View.MemoryView":603 + * result = 1 + * + * for length in self.view.shape[:self.view.ndim]: # <<<<<<<<<<<<<< + * result *= length + * + */ + __pyx_t_3 = (__pyx_v_self->view.shape + __pyx_v_self->view.ndim); + for (__pyx_t_4 = __pyx_v_self->view.shape; __pyx_t_4 < __pyx_t_3; __pyx_t_4++) { + __pyx_t_2 = __pyx_t_4; + __pyx_t_5 = PyInt_FromSsize_t((__pyx_t_2[0])); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 603, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_XDECREF_SET(__pyx_v_length, __pyx_t_5); + __pyx_t_5 = 0; + + /* "View.MemoryView":604 + * + * for length in self.view.shape[:self.view.ndim]: + * result *= length # <<<<<<<<<<<<<< + * + * self._size = result + */ + __pyx_t_5 = PyNumber_InPlaceMultiply(__pyx_v_result, __pyx_v_length); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 604, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF_SET(__pyx_v_result, __pyx_t_5); + __pyx_t_5 = 0; + } + + /* "View.MemoryView":606 + * result *= length + * + * self._size = result # <<<<<<<<<<<<<< + * + * return self._size + */ + __Pyx_INCREF(__pyx_v_result); + __Pyx_GIVEREF(__pyx_v_result); + __Pyx_GOTREF(__pyx_v_self->_size); + __Pyx_DECREF(__pyx_v_self->_size); + __pyx_v_self->_size = __pyx_v_result; + + /* "View.MemoryView":600 + * @property + * def size(self): + * if self._size is None: # <<<<<<<<<<<<<< + * result = 1 + * + */ + } + + /* "View.MemoryView":608 + * self._size = result + * + * return self._size # <<<<<<<<<<<<<< + * + * def __len__(self): + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->_size); + __pyx_r = __pyx_v_self->_size; + goto __pyx_L0; + + /* "View.MemoryView":598 + * return self.size * self.view.itemsize + * + * @property # <<<<<<<<<<<<<< + * def size(self): + * if self._size is None: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.memoryview.size.__get__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_result); + __Pyx_XDECREF(__pyx_v_length); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":610 + * return self._size + * + * def __len__(self): # <<<<<<<<<<<<<< + * if self.view.ndim >= 1: + * return self.view.shape[0] + */ + +/* Python wrapper */ +static Py_ssize_t __pyx_memoryview___len__(PyObject *__pyx_v_self); /*proto*/ +static Py_ssize_t __pyx_memoryview___len__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + Py_ssize_t __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__len__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_10__len__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_10__len__(struct __pyx_memoryview_obj *__pyx_v_self) { + Py_ssize_t __pyx_r; + int __pyx_t_1; + + /* "View.MemoryView":611 + * + * def __len__(self): + * if self.view.ndim >= 1: # <<<<<<<<<<<<<< + * return self.view.shape[0] + * + */ + __pyx_t_1 = (__pyx_v_self->view.ndim >= 1); + if (__pyx_t_1) { + + /* "View.MemoryView":612 + * def __len__(self): + * if self.view.ndim >= 1: + * return self.view.shape[0] # <<<<<<<<<<<<<< + * + * return 0 + */ + __pyx_r = (__pyx_v_self->view.shape[0]); + goto __pyx_L0; + + /* "View.MemoryView":611 + * + * def __len__(self): + * if self.view.ndim >= 1: # <<<<<<<<<<<<<< + * return self.view.shape[0] + * + */ + } + + /* "View.MemoryView":614 + * return self.view.shape[0] + * + * return 0 # <<<<<<<<<<<<<< + * + * def __repr__(self): + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":610 + * return self._size + * + * def __len__(self): # <<<<<<<<<<<<<< + * if self.view.ndim >= 1: + * return self.view.shape[0] + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":616 + * return 0 + * + * def __repr__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__, + * id(self)) + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview___repr__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_memoryview___repr__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__repr__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12__repr__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12__repr__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__repr__", 1); + + /* "View.MemoryView":617 + * + * def __repr__(self): + * return "" % (self.base.__class__.__name__, # <<<<<<<<<<<<<< + * id(self)) + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":618 + * def __repr__(self): + * return "" % (self.base.__class__.__name__, + * id(self)) # <<<<<<<<<<<<<< + * + * def __str__(self): + */ + __pyx_t_2 = __Pyx_PyObject_CallOneArg(__pyx_builtin_id, ((PyObject *)__pyx_v_self)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 618, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + + /* "View.MemoryView":617 + * + * def __repr__(self): + * return "" % (self.base.__class__.__name__, # <<<<<<<<<<<<<< + * id(self)) + * + */ + __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1)) __PYX_ERR(1, 617, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_2)) __PYX_ERR(1, 617, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_at_0x_x, __pyx_t_3); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 617, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":616 + * return 0 + * + * def __repr__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__, + * id(self)) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview.__repr__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":620 + * id(self)) + * + * def __str__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__,) + * + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview___str__(PyObject *__pyx_v_self); /*proto*/ +static PyObject *__pyx_memoryview___str__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__str__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14__str__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14__str__(struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__str__", 1); + + /* "View.MemoryView":621 + * + * def __str__(self): + * return "" % (self.base.__class__.__name__,) # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_2 = PyTuple_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 0, __pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_object, __pyx_t_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 621, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":620 + * id(self)) + * + * def __str__(self): # <<<<<<<<<<<<<< + * return "" % (self.base.__class__.__name__,) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.__str__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":624 + * + * + * def is_c_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_is_c_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_is_c_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("is_c_contig (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("is_c_contig", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "is_c_contig", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16is_c_contig(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16is_c_contig(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice *__pyx_v_mslice; + __Pyx_memviewslice __pyx_v_tmp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("is_c_contig", 1); + + /* "View.MemoryView":627 + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) # <<<<<<<<<<<<<< + * return slice_is_contig(mslice[0], 'C', self.view.ndim) + * + */ + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 627, __pyx_L1_error) + __pyx_v_mslice = __pyx_t_1; + + /* "View.MemoryView":628 + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) + * return slice_is_contig(mslice[0], 'C', self.view.ndim) # <<<<<<<<<<<<<< + * + * def is_f_contig(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'C', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 628, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":624 + * + * + * def is_c_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.is_c_contig", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":630 + * return slice_is_contig(mslice[0], 'C', self.view.ndim) + * + * def is_f_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_is_f_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_is_f_contig(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("is_f_contig (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("is_f_contig", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "is_f_contig", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18is_f_contig(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18is_f_contig(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice *__pyx_v_mslice; + __Pyx_memviewslice __pyx_v_tmp; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice *__pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("is_f_contig", 1); + + /* "View.MemoryView":633 + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) # <<<<<<<<<<<<<< + * return slice_is_contig(mslice[0], 'F', self.view.ndim) + * + */ + __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 633, __pyx_L1_error) + __pyx_v_mslice = __pyx_t_1; + + /* "View.MemoryView":634 + * cdef __Pyx_memviewslice tmp + * mslice = get_slice_from_memview(self, &tmp) + * return slice_is_contig(mslice[0], 'F', self.view.ndim) # <<<<<<<<<<<<<< + * + * def copy(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'F', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 634, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":630 + * return slice_is_contig(mslice[0], 'C', self.view.ndim) + * + * def is_f_contig(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice *mslice + * cdef __Pyx_memviewslice tmp + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.is_f_contig", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":636 + * return slice_is_contig(mslice[0], 'F', self.view.ndim) + * + * def copy(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice mslice + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_copy(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_copy(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("copy (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("copy", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "copy", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20copy(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20copy(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice __pyx_v_mslice; + int __pyx_v_flags; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("copy", 1); + + /* "View.MemoryView":638 + * def copy(self): + * cdef __Pyx_memviewslice mslice + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS # <<<<<<<<<<<<<< + * + * slice_copy(self, &mslice) + */ + __pyx_v_flags = (__pyx_v_self->flags & (~PyBUF_F_CONTIGUOUS)); + + /* "View.MemoryView":640 + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS + * + * slice_copy(self, &mslice) # <<<<<<<<<<<<<< + * mslice = slice_copy_contig(&mslice, "c", self.view.ndim, + * self.view.itemsize, + */ + __pyx_memoryview_slice_copy(__pyx_v_self, (&__pyx_v_mslice)); + + /* "View.MemoryView":641 + * + * slice_copy(self, &mslice) + * mslice = slice_copy_contig(&mslice, "c", self.view.ndim, # <<<<<<<<<<<<<< + * self.view.itemsize, + * flags|PyBUF_C_CONTIGUOUS, + */ + __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_mslice), ((char *)"c"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_C_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 641, __pyx_L1_error) + __pyx_v_mslice = __pyx_t_1; + + /* "View.MemoryView":646 + * self.dtype_is_object) + * + * return memoryview_copy_from_slice(self, &mslice) # <<<<<<<<<<<<<< + * + * def copy_fortran(self): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_mslice)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 646, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":636 + * return slice_is_contig(mslice[0], 'F', self.view.ndim) + * + * def copy(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice mslice + * cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.copy", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":648 + * return memoryview_copy_from_slice(self, &mslice) + * + * def copy_fortran(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice src, dst + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS + */ + +/* Python wrapper */ +static PyObject *__pyx_memoryview_copy_fortran(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_memoryview_copy_fortran(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("copy_fortran (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("copy_fortran", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "copy_fortran", 0))) return NULL; + __pyx_r = __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22copy_fortran(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22copy_fortran(struct __pyx_memoryview_obj *__pyx_v_self) { + __Pyx_memviewslice __pyx_v_src; + __Pyx_memviewslice __pyx_v_dst; + int __pyx_v_flags; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("copy_fortran", 1); + + /* "View.MemoryView":650 + * def copy_fortran(self): + * cdef __Pyx_memviewslice src, dst + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS # <<<<<<<<<<<<<< + * + * slice_copy(self, &src) + */ + __pyx_v_flags = (__pyx_v_self->flags & (~PyBUF_C_CONTIGUOUS)); + + /* "View.MemoryView":652 + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS + * + * slice_copy(self, &src) # <<<<<<<<<<<<<< + * dst = slice_copy_contig(&src, "fortran", self.view.ndim, + * self.view.itemsize, + */ + __pyx_memoryview_slice_copy(__pyx_v_self, (&__pyx_v_src)); + + /* "View.MemoryView":653 + * + * slice_copy(self, &src) + * dst = slice_copy_contig(&src, "fortran", self.view.ndim, # <<<<<<<<<<<<<< + * self.view.itemsize, + * flags|PyBUF_F_CONTIGUOUS, + */ + __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_src), ((char *)"fortran"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_F_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 653, __pyx_L1_error) + __pyx_v_dst = __pyx_t_1; + + /* "View.MemoryView":658 + * self.dtype_is_object) + * + * return memoryview_copy_from_slice(self, &dst) # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_dst)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 658, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":648 + * return memoryview_copy_from_slice(self, &mslice) + * + * def copy_fortran(self): # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice src, dst + * cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.memoryview.copy_fortran", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryview_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryview_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_memoryview___reduce_cython__(((struct __pyx_memoryview_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryview___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":2 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 2, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryview_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryview_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + CYTHON_UNUSED PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 3, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 3, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 3, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.memoryview.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_memoryview_2__setstate_cython__(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryview_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryview_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":4 + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 4, __pyx_L1_error) + + /* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.memoryview.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":662 + * + * @cname('__pyx_memoryview_new') + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): # <<<<<<<<<<<<<< + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo + */ + +static PyObject *__pyx_memoryview_new(PyObject *__pyx_v_o, int __pyx_v_flags, int __pyx_v_dtype_is_object, __Pyx_TypeInfo *__pyx_v_typeinfo) { + struct __pyx_memoryview_obj *__pyx_v_result = 0; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_cwrapper", 1); + + /* "View.MemoryView":663 + * @cname('__pyx_memoryview_new') + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): + * cdef memoryview result = memoryview(o, flags, dtype_is_object) # <<<<<<<<<<<<<< + * result.typeinfo = typeinfo + * return result + */ + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF(__pyx_v_o); + __Pyx_GIVEREF(__pyx_v_o); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_o)) __PYX_ERR(1, 663, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_1)) __PYX_ERR(1, 663, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2)) __PYX_ERR(1, 663, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 663, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_v_result = ((struct __pyx_memoryview_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":664 + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo # <<<<<<<<<<<<<< + * return result + * + */ + __pyx_v_result->typeinfo = __pyx_v_typeinfo; + + /* "View.MemoryView":665 + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo + * return result # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_check') + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = ((PyObject *)__pyx_v_result); + goto __pyx_L0; + + /* "View.MemoryView":662 + * + * @cname('__pyx_memoryview_new') + * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo): # <<<<<<<<<<<<<< + * cdef memoryview result = memoryview(o, flags, dtype_is_object) + * result.typeinfo = typeinfo + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview_cwrapper", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":668 + * + * @cname('__pyx_memoryview_check') + * cdef inline bint memoryview_check(object o) noexcept: # <<<<<<<<<<<<<< + * return isinstance(o, memoryview) + * + */ + +static CYTHON_INLINE int __pyx_memoryview_check(PyObject *__pyx_v_o) { + int __pyx_r; + int __pyx_t_1; + + /* "View.MemoryView":669 + * @cname('__pyx_memoryview_check') + * cdef inline bint memoryview_check(object o) noexcept: + * return isinstance(o, memoryview) # <<<<<<<<<<<<<< + * + * cdef tuple _unellipsify(object index, int ndim): + */ + __pyx_t_1 = __Pyx_TypeCheck(__pyx_v_o, __pyx_memoryview_type); + __pyx_r = __pyx_t_1; + goto __pyx_L0; + + /* "View.MemoryView":668 + * + * @cname('__pyx_memoryview_check') + * cdef inline bint memoryview_check(object o) noexcept: # <<<<<<<<<<<<<< + * return isinstance(o, memoryview) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":671 + * return isinstance(o, memoryview) + * + * cdef tuple _unellipsify(object index, int ndim): # <<<<<<<<<<<<<< + * """ + * Replace all ellipses with full slices and fill incomplete indices with + */ + +static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) { + Py_ssize_t __pyx_v_idx; + PyObject *__pyx_v_tup = NULL; + PyObject *__pyx_v_result = NULL; + int __pyx_v_have_slices; + int __pyx_v_seen_ellipsis; + PyObject *__pyx_v_item = NULL; + Py_ssize_t __pyx_v_nslices; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + Py_ssize_t __pyx_t_4; + Py_ssize_t __pyx_t_5; + Py_UCS4 __pyx_t_6; + PyObject *__pyx_t_7 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_unellipsify", 1); + + /* "View.MemoryView":677 + * """ + * cdef Py_ssize_t idx + * tup = index if isinstance(index, tuple) else (index,) # <<<<<<<<<<<<<< + * + * result = [slice(None)] * ndim + */ + __pyx_t_2 = PyTuple_Check(__pyx_v_index); + if (__pyx_t_2) { + __Pyx_INCREF(((PyObject*)__pyx_v_index)); + __pyx_t_1 = __pyx_v_index; + } else { + __pyx_t_3 = PyTuple_New(1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 677, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF(__pyx_v_index); + __Pyx_GIVEREF(__pyx_v_index); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_v_index)) __PYX_ERR(1, 677, __pyx_L1_error); + __pyx_t_1 = __pyx_t_3; + __pyx_t_3 = 0; + } + __pyx_v_tup = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":679 + * tup = index if isinstance(index, tuple) else (index,) + * + * result = [slice(None)] * ndim # <<<<<<<<<<<<<< + * have_slices = False + * seen_ellipsis = False + */ + __pyx_t_1 = PyList_New(1 * ((__pyx_v_ndim<0) ? 0:__pyx_v_ndim)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 679, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + { Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < __pyx_v_ndim; __pyx_temp++) { + __Pyx_INCREF(__pyx_slice__5); + __Pyx_GIVEREF(__pyx_slice__5); + if (__Pyx_PyList_SET_ITEM(__pyx_t_1, __pyx_temp, __pyx_slice__5)) __PYX_ERR(1, 679, __pyx_L1_error); + } + } + __pyx_v_result = ((PyObject*)__pyx_t_1); + __pyx_t_1 = 0; + + /* "View.MemoryView":680 + * + * result = [slice(None)] * ndim + * have_slices = False # <<<<<<<<<<<<<< + * seen_ellipsis = False + * idx = 0 + */ + __pyx_v_have_slices = 0; + + /* "View.MemoryView":681 + * result = [slice(None)] * ndim + * have_slices = False + * seen_ellipsis = False # <<<<<<<<<<<<<< + * idx = 0 + * for item in tup: + */ + __pyx_v_seen_ellipsis = 0; + + /* "View.MemoryView":682 + * have_slices = False + * seen_ellipsis = False + * idx = 0 # <<<<<<<<<<<<<< + * for item in tup: + * if item is Ellipsis: + */ + __pyx_v_idx = 0; + + /* "View.MemoryView":683 + * seen_ellipsis = False + * idx = 0 + * for item in tup: # <<<<<<<<<<<<<< + * if item is Ellipsis: + * if not seen_ellipsis: + */ + if (unlikely(__pyx_v_tup == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable"); + __PYX_ERR(1, 683, __pyx_L1_error) + } + __pyx_t_1 = __pyx_v_tup; __Pyx_INCREF(__pyx_t_1); + __pyx_t_4 = 0; + for (;;) { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_1); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 683, __pyx_L1_error) + #endif + if (__pyx_t_4 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_3 = PyTuple_GET_ITEM(__pyx_t_1, __pyx_t_4); __Pyx_INCREF(__pyx_t_3); __pyx_t_4++; if (unlikely((0 < 0))) __PYX_ERR(1, 683, __pyx_L1_error) + #else + __pyx_t_3 = __Pyx_PySequence_ITEM(__pyx_t_1, __pyx_t_4); __pyx_t_4++; if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 683, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + #endif + __Pyx_XDECREF_SET(__pyx_v_item, __pyx_t_3); + __pyx_t_3 = 0; + + /* "View.MemoryView":684 + * idx = 0 + * for item in tup: + * if item is Ellipsis: # <<<<<<<<<<<<<< + * if not seen_ellipsis: + * idx += ndim - len(tup) + */ + __pyx_t_2 = (__pyx_v_item == __pyx_builtin_Ellipsis); + if (__pyx_t_2) { + + /* "View.MemoryView":685 + * for item in tup: + * if item is Ellipsis: + * if not seen_ellipsis: # <<<<<<<<<<<<<< + * idx += ndim - len(tup) + * seen_ellipsis = True + */ + __pyx_t_2 = (!__pyx_v_seen_ellipsis); + if (__pyx_t_2) { + + /* "View.MemoryView":686 + * if item is Ellipsis: + * if not seen_ellipsis: + * idx += ndim - len(tup) # <<<<<<<<<<<<<< + * seen_ellipsis = True + * have_slices = True + */ + if (unlikely(__pyx_v_tup == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(1, 686, __pyx_L1_error) + } + __pyx_t_5 = __Pyx_PyTuple_GET_SIZE(__pyx_v_tup); if (unlikely(__pyx_t_5 == ((Py_ssize_t)-1))) __PYX_ERR(1, 686, __pyx_L1_error) + __pyx_v_idx = (__pyx_v_idx + (__pyx_v_ndim - __pyx_t_5)); + + /* "View.MemoryView":687 + * if not seen_ellipsis: + * idx += ndim - len(tup) + * seen_ellipsis = True # <<<<<<<<<<<<<< + * have_slices = True + * else: + */ + __pyx_v_seen_ellipsis = 1; + + /* "View.MemoryView":685 + * for item in tup: + * if item is Ellipsis: + * if not seen_ellipsis: # <<<<<<<<<<<<<< + * idx += ndim - len(tup) + * seen_ellipsis = True + */ + } + + /* "View.MemoryView":688 + * idx += ndim - len(tup) + * seen_ellipsis = True + * have_slices = True # <<<<<<<<<<<<<< + * else: + * if isinstance(item, slice): + */ + __pyx_v_have_slices = 1; + + /* "View.MemoryView":684 + * idx = 0 + * for item in tup: + * if item is Ellipsis: # <<<<<<<<<<<<<< + * if not seen_ellipsis: + * idx += ndim - len(tup) + */ + goto __pyx_L5; + } + + /* "View.MemoryView":690 + * have_slices = True + * else: + * if isinstance(item, slice): # <<<<<<<<<<<<<< + * have_slices = True + * elif not PyIndex_Check(item): + */ + /*else*/ { + __pyx_t_2 = PySlice_Check(__pyx_v_item); + if (__pyx_t_2) { + + /* "View.MemoryView":691 + * else: + * if isinstance(item, slice): + * have_slices = True # <<<<<<<<<<<<<< + * elif not PyIndex_Check(item): + * raise TypeError, f"Cannot index with type '{type(item)}'" + */ + __pyx_v_have_slices = 1; + + /* "View.MemoryView":690 + * have_slices = True + * else: + * if isinstance(item, slice): # <<<<<<<<<<<<<< + * have_slices = True + * elif not PyIndex_Check(item): + */ + goto __pyx_L7; + } + + /* "View.MemoryView":692 + * if isinstance(item, slice): + * have_slices = True + * elif not PyIndex_Check(item): # <<<<<<<<<<<<<< + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item + */ + __pyx_t_2 = (!(PyIndex_Check(__pyx_v_item) != 0)); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":693 + * have_slices = True + * elif not PyIndex_Check(item): + * raise TypeError, f"Cannot index with type '{type(item)}'" # <<<<<<<<<<<<<< + * result[idx] = item + * idx += 1 + */ + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 693, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_5 = 0; + __pyx_t_6 = 127; + __Pyx_INCREF(__pyx_kp_u_Cannot_index_with_type); + __pyx_t_5 += 24; + __Pyx_GIVEREF(__pyx_kp_u_Cannot_index_with_type); + PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_kp_u_Cannot_index_with_type); + __pyx_t_7 = __Pyx_PyObject_FormatSimple(((PyObject *)Py_TYPE(__pyx_v_item)), __pyx_empty_unicode); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 693, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_6 = (__Pyx_PyUnicode_MAX_CHAR_VALUE(__pyx_t_7) > __pyx_t_6) ? __Pyx_PyUnicode_MAX_CHAR_VALUE(__pyx_t_7) : __pyx_t_6; + __pyx_t_5 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_7); + __pyx_t_7 = 0; + __Pyx_INCREF(__pyx_kp_u__6); + __pyx_t_5 += 1; + __Pyx_GIVEREF(__pyx_kp_u__6); + PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_kp_u__6); + __pyx_t_7 = __Pyx_PyUnicode_Join(__pyx_t_3, 3, __pyx_t_5, __pyx_t_6); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 693, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_t_7, 0, 0); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __PYX_ERR(1, 693, __pyx_L1_error) + + /* "View.MemoryView":692 + * if isinstance(item, slice): + * have_slices = True + * elif not PyIndex_Check(item): # <<<<<<<<<<<<<< + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item + */ + } + __pyx_L7:; + + /* "View.MemoryView":694 + * elif not PyIndex_Check(item): + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item # <<<<<<<<<<<<<< + * idx += 1 + * + */ + if (unlikely((__Pyx_SetItemInt(__pyx_v_result, __pyx_v_idx, __pyx_v_item, Py_ssize_t, 1, PyInt_FromSsize_t, 1, 1, 1) < 0))) __PYX_ERR(1, 694, __pyx_L1_error) + } + __pyx_L5:; + + /* "View.MemoryView":695 + * raise TypeError, f"Cannot index with type '{type(item)}'" + * result[idx] = item + * idx += 1 # <<<<<<<<<<<<<< + * + * nslices = ndim - idx + */ + __pyx_v_idx = (__pyx_v_idx + 1); + + /* "View.MemoryView":683 + * seen_ellipsis = False + * idx = 0 + * for item in tup: # <<<<<<<<<<<<<< + * if item is Ellipsis: + * if not seen_ellipsis: + */ + } + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "View.MemoryView":697 + * idx += 1 + * + * nslices = ndim - idx # <<<<<<<<<<<<<< + * return have_slices or nslices, tuple(result) + * + */ + __pyx_v_nslices = (__pyx_v_ndim - __pyx_v_idx); + + /* "View.MemoryView":698 + * + * nslices = ndim - idx + * return have_slices or nslices, tuple(result) # <<<<<<<<<<<<<< + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + */ + __Pyx_XDECREF(__pyx_r); + if (!__pyx_v_have_slices) { + } else { + __pyx_t_7 = __Pyx_PyBool_FromLong(__pyx_v_have_slices); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_1 = __pyx_t_7; + __pyx_t_7 = 0; + goto __pyx_L9_bool_binop_done; + } + __pyx_t_7 = PyInt_FromSsize_t(__pyx_v_nslices); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_1 = __pyx_t_7; + __pyx_t_7 = 0; + __pyx_L9_bool_binop_done:; + __pyx_t_7 = PyList_AsTuple(__pyx_v_result); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 698, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1)) __PYX_ERR(1, 698, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_7); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_7)) __PYX_ERR(1, 698, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_7 = 0; + __pyx_r = ((PyObject*)__pyx_t_3); + __pyx_t_3 = 0; + goto __pyx_L0; + + /* "View.MemoryView":671 + * return isinstance(o, memoryview) + * + * cdef tuple _unellipsify(object index, int ndim): # <<<<<<<<<<<<<< + * """ + * Replace all ellipses with full slices and fill incomplete indices with + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_AddTraceback("View.MemoryView._unellipsify", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_tup); + __Pyx_XDECREF(__pyx_v_result); + __Pyx_XDECREF(__pyx_v_item); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":700 + * return have_slices or nslices, tuple(result) + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: # <<<<<<<<<<<<<< + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: + */ + +static int assert_direct_dimensions(Py_ssize_t *__pyx_v_suboffsets, int __pyx_v_ndim) { + Py_ssize_t __pyx_v_suboffset; + int __pyx_r; + Py_ssize_t *__pyx_t_1; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + int __pyx_t_4; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + + /* "View.MemoryView":701 + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + * for suboffset in suboffsets[:ndim]: # <<<<<<<<<<<<<< + * if suboffset >= 0: + * raise ValueError, "Indirect dimensions not supported" + */ + __pyx_t_2 = (__pyx_v_suboffsets + __pyx_v_ndim); + for (__pyx_t_3 = __pyx_v_suboffsets; __pyx_t_3 < __pyx_t_2; __pyx_t_3++) { + __pyx_t_1 = __pyx_t_3; + __pyx_v_suboffset = (__pyx_t_1[0]); + + /* "View.MemoryView":702 + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * raise ValueError, "Indirect dimensions not supported" + * return 0 # return type just used as an error flag + */ + __pyx_t_4 = (__pyx_v_suboffset >= 0); + if (unlikely(__pyx_t_4)) { + + /* "View.MemoryView":703 + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: + * raise ValueError, "Indirect dimensions not supported" # <<<<<<<<<<<<<< + * return 0 # return type just used as an error flag + * + */ + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_kp_s_Indirect_dimensions_not_supporte, 0, 0); + __PYX_ERR(1, 703, __pyx_L1_error) + + /* "View.MemoryView":702 + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * raise ValueError, "Indirect dimensions not supported" + * return 0 # return type just used as an error flag + */ + } + } + + /* "View.MemoryView":704 + * if suboffset >= 0: + * raise ValueError, "Indirect dimensions not supported" + * return 0 # return type just used as an error flag # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":700 + * return have_slices or nslices, tuple(result) + * + * cdef int assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim) except -1: # <<<<<<<<<<<<<< + * for suboffset in suboffsets[:ndim]: + * if suboffset >= 0: + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView.assert_direct_dimensions", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":711 + * + * @cname('__pyx_memview_slice') + * cdef memoryview memview_slice(memoryview memview, object indices): # <<<<<<<<<<<<<< + * cdef int new_ndim = 0, suboffset_dim = -1, dim + * cdef bint negative_step + */ + +static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_obj *__pyx_v_memview, PyObject *__pyx_v_indices) { + int __pyx_v_new_ndim; + int __pyx_v_suboffset_dim; + int __pyx_v_dim; + __Pyx_memviewslice __pyx_v_src; + __Pyx_memviewslice __pyx_v_dst; + __Pyx_memviewslice *__pyx_v_p_src; + struct __pyx_memoryviewslice_obj *__pyx_v_memviewsliceobj = 0; + __Pyx_memviewslice *__pyx_v_p_dst; + int *__pyx_v_p_suboffset_dim; + Py_ssize_t __pyx_v_start; + Py_ssize_t __pyx_v_stop; + Py_ssize_t __pyx_v_step; + Py_ssize_t __pyx_v_cindex; + int __pyx_v_have_start; + int __pyx_v_have_stop; + int __pyx_v_have_step; + PyObject *__pyx_v_index = NULL; + struct __pyx_memoryview_obj *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + struct __pyx_memoryview_obj *__pyx_t_3; + char *__pyx_t_4; + int __pyx_t_5; + Py_ssize_t __pyx_t_6; + PyObject *(*__pyx_t_7)(PyObject *); + PyObject *__pyx_t_8 = NULL; + Py_ssize_t __pyx_t_9; + int __pyx_t_10; + Py_ssize_t __pyx_t_11; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memview_slice", 1); + + /* "View.MemoryView":712 + * @cname('__pyx_memview_slice') + * cdef memoryview memview_slice(memoryview memview, object indices): + * cdef int new_ndim = 0, suboffset_dim = -1, dim # <<<<<<<<<<<<<< + * cdef bint negative_step + * cdef __Pyx_memviewslice src, dst + */ + __pyx_v_new_ndim = 0; + __pyx_v_suboffset_dim = -1; + + /* "View.MemoryView":719 + * + * + * memset(&dst, 0, sizeof(dst)) # <<<<<<<<<<<<<< + * + * cdef _memoryviewslice memviewsliceobj + */ + (void)(memset((&__pyx_v_dst), 0, (sizeof(__pyx_v_dst)))); + + /* "View.MemoryView":723 + * cdef _memoryviewslice memviewsliceobj + * + * assert memview.view.ndim > 0 # <<<<<<<<<<<<<< + * + * if isinstance(memview, _memoryviewslice): + */ + #ifndef CYTHON_WITHOUT_ASSERTIONS + if (unlikely(__pyx_assertions_enabled())) { + __pyx_t_1 = (__pyx_v_memview->view.ndim > 0); + if (unlikely(!__pyx_t_1)) { + __Pyx_Raise(__pyx_builtin_AssertionError, 0, 0, 0); + __PYX_ERR(1, 723, __pyx_L1_error) + } + } + #else + if ((1)); else __PYX_ERR(1, 723, __pyx_L1_error) + #endif + + /* "View.MemoryView":725 + * assert memview.view.ndim > 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * memviewsliceobj = memview + * p_src = &memviewsliceobj.from_slice + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":726 + * + * if isinstance(memview, _memoryviewslice): + * memviewsliceobj = memview # <<<<<<<<<<<<<< + * p_src = &memviewsliceobj.from_slice + * else: + */ + if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 726, __pyx_L1_error) + __pyx_t_2 = ((PyObject *)__pyx_v_memview); + __Pyx_INCREF(__pyx_t_2); + __pyx_v_memviewsliceobj = ((struct __pyx_memoryviewslice_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":727 + * if isinstance(memview, _memoryviewslice): + * memviewsliceobj = memview + * p_src = &memviewsliceobj.from_slice # <<<<<<<<<<<<<< + * else: + * slice_copy(memview, &src) + */ + __pyx_v_p_src = (&__pyx_v_memviewsliceobj->from_slice); + + /* "View.MemoryView":725 + * assert memview.view.ndim > 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * memviewsliceobj = memview + * p_src = &memviewsliceobj.from_slice + */ + goto __pyx_L3; + } + + /* "View.MemoryView":729 + * p_src = &memviewsliceobj.from_slice + * else: + * slice_copy(memview, &src) # <<<<<<<<<<<<<< + * p_src = &src + * + */ + /*else*/ { + __pyx_memoryview_slice_copy(__pyx_v_memview, (&__pyx_v_src)); + + /* "View.MemoryView":730 + * else: + * slice_copy(memview, &src) + * p_src = &src # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_p_src = (&__pyx_v_src); + } + __pyx_L3:; + + /* "View.MemoryView":736 + * + * + * dst.memview = p_src.memview # <<<<<<<<<<<<<< + * dst.data = p_src.data + * + */ + __pyx_t_3 = __pyx_v_p_src->memview; + __pyx_v_dst.memview = __pyx_t_3; + + /* "View.MemoryView":737 + * + * dst.memview = p_src.memview + * dst.data = p_src.data # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_4 = __pyx_v_p_src->data; + __pyx_v_dst.data = __pyx_t_4; + + /* "View.MemoryView":742 + * + * + * cdef __Pyx_memviewslice *p_dst = &dst # <<<<<<<<<<<<<< + * cdef int *p_suboffset_dim = &suboffset_dim + * cdef Py_ssize_t start, stop, step, cindex + */ + __pyx_v_p_dst = (&__pyx_v_dst); + + /* "View.MemoryView":743 + * + * cdef __Pyx_memviewslice *p_dst = &dst + * cdef int *p_suboffset_dim = &suboffset_dim # <<<<<<<<<<<<<< + * cdef Py_ssize_t start, stop, step, cindex + * cdef bint have_start, have_stop, have_step + */ + __pyx_v_p_suboffset_dim = (&__pyx_v_suboffset_dim); + + /* "View.MemoryView":747 + * cdef bint have_start, have_stop, have_step + * + * for dim, index in enumerate(indices): # <<<<<<<<<<<<<< + * if PyIndex_Check(index): + * cindex = index + */ + __pyx_t_5 = 0; + if (likely(PyList_CheckExact(__pyx_v_indices)) || PyTuple_CheckExact(__pyx_v_indices)) { + __pyx_t_2 = __pyx_v_indices; __Pyx_INCREF(__pyx_t_2); + __pyx_t_6 = 0; + __pyx_t_7 = NULL; + } else { + __pyx_t_6 = -1; __pyx_t_2 = PyObject_GetIter(__pyx_v_indices); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 747, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_7 = __Pyx_PyObject_GetIterNextFunc(__pyx_t_2); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 747, __pyx_L1_error) + } + for (;;) { + if (likely(!__pyx_t_7)) { + if (likely(PyList_CheckExact(__pyx_t_2))) { + { + Py_ssize_t __pyx_temp = __Pyx_PyList_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #endif + if (__pyx_t_6 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_8 = PyList_GET_ITEM(__pyx_t_2, __pyx_t_6); __Pyx_INCREF(__pyx_t_8); __pyx_t_6++; if (unlikely((0 < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #else + __pyx_t_8 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_6); __pyx_t_6++; if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 747, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + #endif + } else { + { + Py_ssize_t __pyx_temp = __Pyx_PyTuple_GET_SIZE(__pyx_t_2); + #if !CYTHON_ASSUME_SAFE_MACROS + if (unlikely((__pyx_temp < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #endif + if (__pyx_t_6 >= __pyx_temp) break; + } + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + __pyx_t_8 = PyTuple_GET_ITEM(__pyx_t_2, __pyx_t_6); __Pyx_INCREF(__pyx_t_8); __pyx_t_6++; if (unlikely((0 < 0))) __PYX_ERR(1, 747, __pyx_L1_error) + #else + __pyx_t_8 = __Pyx_PySequence_ITEM(__pyx_t_2, __pyx_t_6); __pyx_t_6++; if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 747, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + #endif + } + } else { + __pyx_t_8 = __pyx_t_7(__pyx_t_2); + if (unlikely(!__pyx_t_8)) { + PyObject* exc_type = PyErr_Occurred(); + if (exc_type) { + if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear(); + else __PYX_ERR(1, 747, __pyx_L1_error) + } + break; + } + __Pyx_GOTREF(__pyx_t_8); + } + __Pyx_XDECREF_SET(__pyx_v_index, __pyx_t_8); + __pyx_t_8 = 0; + __pyx_v_dim = __pyx_t_5; + __pyx_t_5 = (__pyx_t_5 + 1); + + /* "View.MemoryView":748 + * + * for dim, index in enumerate(indices): + * if PyIndex_Check(index): # <<<<<<<<<<<<<< + * cindex = index + * slice_memviewslice( + */ + __pyx_t_1 = (PyIndex_Check(__pyx_v_index) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":749 + * for dim, index in enumerate(indices): + * if PyIndex_Check(index): + * cindex = index # <<<<<<<<<<<<<< + * slice_memviewslice( + * p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim], + */ + __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_v_index); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 749, __pyx_L1_error) + __pyx_v_cindex = __pyx_t_9; + + /* "View.MemoryView":750 + * if PyIndex_Check(index): + * cindex = index + * slice_memviewslice( # <<<<<<<<<<<<<< + * p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim], + * dim, new_ndim, p_suboffset_dim, + */ + __pyx_t_10 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_v_cindex, 0, 0, 0, 0, 0, 0); if (unlikely(__pyx_t_10 == ((int)-1))) __PYX_ERR(1, 750, __pyx_L1_error) + + /* "View.MemoryView":748 + * + * for dim, index in enumerate(indices): + * if PyIndex_Check(index): # <<<<<<<<<<<<<< + * cindex = index + * slice_memviewslice( + */ + goto __pyx_L6; + } + + /* "View.MemoryView":756 + * 0, 0, 0, # have_{start,stop,step} + * False) + * elif index is None: # <<<<<<<<<<<<<< + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 + */ + __pyx_t_1 = (__pyx_v_index == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":757 + * False) + * elif index is None: + * p_dst.shape[new_ndim] = 1 # <<<<<<<<<<<<<< + * p_dst.strides[new_ndim] = 0 + * p_dst.suboffsets[new_ndim] = -1 + */ + (__pyx_v_p_dst->shape[__pyx_v_new_ndim]) = 1; + + /* "View.MemoryView":758 + * elif index is None: + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 # <<<<<<<<<<<<<< + * p_dst.suboffsets[new_ndim] = -1 + * new_ndim += 1 + */ + (__pyx_v_p_dst->strides[__pyx_v_new_ndim]) = 0; + + /* "View.MemoryView":759 + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 + * p_dst.suboffsets[new_ndim] = -1 # <<<<<<<<<<<<<< + * new_ndim += 1 + * else: + */ + (__pyx_v_p_dst->suboffsets[__pyx_v_new_ndim]) = -1L; + + /* "View.MemoryView":760 + * p_dst.strides[new_ndim] = 0 + * p_dst.suboffsets[new_ndim] = -1 + * new_ndim += 1 # <<<<<<<<<<<<<< + * else: + * start = index.start or 0 + */ + __pyx_v_new_ndim = (__pyx_v_new_ndim + 1); + + /* "View.MemoryView":756 + * 0, 0, 0, # have_{start,stop,step} + * False) + * elif index is None: # <<<<<<<<<<<<<< + * p_dst.shape[new_ndim] = 1 + * p_dst.strides[new_ndim] = 0 + */ + goto __pyx_L6; + } + + /* "View.MemoryView":762 + * new_ndim += 1 + * else: + * start = index.start or 0 # <<<<<<<<<<<<<< + * stop = index.stop or 0 + * step = index.step or 0 + */ + /*else*/ { + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 762, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_8); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 762, __pyx_L1_error) + if (!__pyx_t_1) { + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + } else { + __pyx_t_11 = __Pyx_PyIndex_AsSsize_t(__pyx_t_8); if (unlikely((__pyx_t_11 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 762, __pyx_L1_error) + __pyx_t_9 = __pyx_t_11; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L7_bool_binop_done; + } + __pyx_t_9 = 0; + __pyx_L7_bool_binop_done:; + __pyx_v_start = __pyx_t_9; + + /* "View.MemoryView":763 + * else: + * start = index.start or 0 + * stop = index.stop or 0 # <<<<<<<<<<<<<< + * step = index.step or 0 + * + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 763, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_8); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 763, __pyx_L1_error) + if (!__pyx_t_1) { + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + } else { + __pyx_t_11 = __Pyx_PyIndex_AsSsize_t(__pyx_t_8); if (unlikely((__pyx_t_11 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 763, __pyx_L1_error) + __pyx_t_9 = __pyx_t_11; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L9_bool_binop_done; + } + __pyx_t_9 = 0; + __pyx_L9_bool_binop_done:; + __pyx_v_stop = __pyx_t_9; + + /* "View.MemoryView":764 + * start = index.start or 0 + * stop = index.stop or 0 + * step = index.step or 0 # <<<<<<<<<<<<<< + * + * have_start = index.start is not None + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 764, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_8); if (unlikely((__pyx_t_1 < 0))) __PYX_ERR(1, 764, __pyx_L1_error) + if (!__pyx_t_1) { + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + } else { + __pyx_t_11 = __Pyx_PyIndex_AsSsize_t(__pyx_t_8); if (unlikely((__pyx_t_11 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 764, __pyx_L1_error) + __pyx_t_9 = __pyx_t_11; + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + goto __pyx_L11_bool_binop_done; + } + __pyx_t_9 = 0; + __pyx_L11_bool_binop_done:; + __pyx_v_step = __pyx_t_9; + + /* "View.MemoryView":766 + * step = index.step or 0 + * + * have_start = index.start is not None # <<<<<<<<<<<<<< + * have_stop = index.stop is not None + * have_step = index.step is not None + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 766, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = (__pyx_t_8 != Py_None); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __pyx_v_have_start = __pyx_t_1; + + /* "View.MemoryView":767 + * + * have_start = index.start is not None + * have_stop = index.stop is not None # <<<<<<<<<<<<<< + * have_step = index.step is not None + * + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 767, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = (__pyx_t_8 != Py_None); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __pyx_v_have_stop = __pyx_t_1; + + /* "View.MemoryView":768 + * have_start = index.start is not None + * have_stop = index.stop is not None + * have_step = index.step is not None # <<<<<<<<<<<<<< + * + * slice_memviewslice( + */ + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 768, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __pyx_t_1 = (__pyx_t_8 != Py_None); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __pyx_v_have_step = __pyx_t_1; + + /* "View.MemoryView":770 + * have_step = index.step is not None + * + * slice_memviewslice( # <<<<<<<<<<<<<< + * p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim], + * dim, new_ndim, p_suboffset_dim, + */ + __pyx_t_10 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_v_start, __pyx_v_stop, __pyx_v_step, __pyx_v_have_start, __pyx_v_have_stop, __pyx_v_have_step, 1); if (unlikely(__pyx_t_10 == ((int)-1))) __PYX_ERR(1, 770, __pyx_L1_error) + + /* "View.MemoryView":776 + * have_start, have_stop, have_step, + * True) + * new_ndim += 1 # <<<<<<<<<<<<<< + * + * if isinstance(memview, _memoryviewslice): + */ + __pyx_v_new_ndim = (__pyx_v_new_ndim + 1); + } + __pyx_L6:; + + /* "View.MemoryView":747 + * cdef bint have_start, have_stop, have_step + * + * for dim, index in enumerate(indices): # <<<<<<<<<<<<<< + * if PyIndex_Check(index): + * cindex = index + */ + } + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "View.MemoryView":778 + * new_ndim += 1 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":779 + * + * if isinstance(memview, _memoryviewslice): + * return memoryview_fromslice(dst, new_ndim, # <<<<<<<<<<<<<< + * memviewsliceobj.to_object_func, + * memviewsliceobj.to_dtype_func, + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + + /* "View.MemoryView":780 + * if isinstance(memview, _memoryviewslice): + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, # <<<<<<<<<<<<<< + * memviewsliceobj.to_dtype_func, + * memview.dtype_is_object) + */ + if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 780, __pyx_L1_error) } + + /* "View.MemoryView":781 + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, + * memviewsliceobj.to_dtype_func, # <<<<<<<<<<<<<< + * memview.dtype_is_object) + * else: + */ + if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 781, __pyx_L1_error) } + + /* "View.MemoryView":779 + * + * if isinstance(memview, _memoryviewslice): + * return memoryview_fromslice(dst, new_ndim, # <<<<<<<<<<<<<< + * memviewsliceobj.to_object_func, + * memviewsliceobj.to_dtype_func, + */ + __pyx_t_2 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, __pyx_v_memviewsliceobj->to_object_func, __pyx_v_memviewsliceobj->to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 779, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + if (!(likely(((__pyx_t_2) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_2, __pyx_memoryview_type))))) __PYX_ERR(1, 779, __pyx_L1_error) + __pyx_r = ((struct __pyx_memoryview_obj *)__pyx_t_2); + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":778 + * new_ndim += 1 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * return memoryview_fromslice(dst, new_ndim, + * memviewsliceobj.to_object_func, + */ + } + + /* "View.MemoryView":784 + * memview.dtype_is_object) + * else: + * return memoryview_fromslice(dst, new_ndim, NULL, NULL, # <<<<<<<<<<<<<< + * memview.dtype_is_object) + * + */ + /*else*/ { + __Pyx_XDECREF((PyObject *)__pyx_r); + + /* "View.MemoryView":785 + * else: + * return memoryview_fromslice(dst, new_ndim, NULL, NULL, + * memview.dtype_is_object) # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, NULL, NULL, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 784, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + + /* "View.MemoryView":784 + * memview.dtype_is_object) + * else: + * return memoryview_fromslice(dst, new_ndim, NULL, NULL, # <<<<<<<<<<<<<< + * memview.dtype_is_object) + * + */ + if (!(likely(((__pyx_t_2) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_2, __pyx_memoryview_type))))) __PYX_ERR(1, 784, __pyx_L1_error) + __pyx_r = ((struct __pyx_memoryview_obj *)__pyx_t_2); + __pyx_t_2 = 0; + goto __pyx_L0; + } + + /* "View.MemoryView":711 + * + * @cname('__pyx_memview_slice') + * cdef memoryview memview_slice(memoryview memview, object indices): # <<<<<<<<<<<<<< + * cdef int new_ndim = 0, suboffset_dim = -1, dim + * cdef bint negative_step + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("View.MemoryView.memview_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_memviewsliceobj); + __Pyx_XDECREF(__pyx_v_index); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":793 + * + * @cname('__pyx_memoryview_slice_memviewslice') + * cdef int slice_memviewslice( # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * Py_ssize_t shape, Py_ssize_t stride, Py_ssize_t suboffset, + */ + +static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst, Py_ssize_t __pyx_v_shape, Py_ssize_t __pyx_v_stride, Py_ssize_t __pyx_v_suboffset, int __pyx_v_dim, int __pyx_v_new_ndim, int *__pyx_v_suboffset_dim, Py_ssize_t __pyx_v_start, Py_ssize_t __pyx_v_stop, Py_ssize_t __pyx_v_step, int __pyx_v_have_start, int __pyx_v_have_stop, int __pyx_v_have_step, int __pyx_v_is_slice) { + Py_ssize_t __pyx_v_new_shape; + int __pyx_v_negative_step; + int __pyx_r; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":813 + * cdef bint negative_step + * + * if not is_slice: # <<<<<<<<<<<<<< + * + * if start < 0: + */ + __pyx_t_1 = (!__pyx_v_is_slice); + if (__pyx_t_1) { + + /* "View.MemoryView":815 + * if not is_slice: + * + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if not 0 <= start < shape: + */ + __pyx_t_1 = (__pyx_v_start < 0); + if (__pyx_t_1) { + + /* "View.MemoryView":816 + * + * if start < 0: + * start += shape # <<<<<<<<<<<<<< + * if not 0 <= start < shape: + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) + */ + __pyx_v_start = (__pyx_v_start + __pyx_v_shape); + + /* "View.MemoryView":815 + * if not is_slice: + * + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if not 0 <= start < shape: + */ + } + + /* "View.MemoryView":817 + * if start < 0: + * start += shape + * if not 0 <= start < shape: # <<<<<<<<<<<<<< + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) + * else: + */ + __pyx_t_1 = (0 <= __pyx_v_start); + if (__pyx_t_1) { + __pyx_t_1 = (__pyx_v_start < __pyx_v_shape); + } + __pyx_t_2 = (!__pyx_t_1); + if (__pyx_t_2) { + + /* "View.MemoryView":818 + * start += shape + * if not 0 <= start < shape: + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) # <<<<<<<<<<<<<< + * else: + * + */ + __pyx_t_3 = __pyx_memoryview_err_dim(PyExc_IndexError, __pyx_kp_s_Index_out_of_bounds_axis_d, __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 818, __pyx_L1_error) + + /* "View.MemoryView":817 + * if start < 0: + * start += shape + * if not 0 <= start < shape: # <<<<<<<<<<<<<< + * _err_dim(PyExc_IndexError, "Index out of bounds (axis %d)", dim) + * else: + */ + } + + /* "View.MemoryView":813 + * cdef bint negative_step + * + * if not is_slice: # <<<<<<<<<<<<<< + * + * if start < 0: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":821 + * else: + * + * if have_step: # <<<<<<<<<<<<<< + * negative_step = step < 0 + * if step == 0: + */ + /*else*/ { + __pyx_t_2 = (__pyx_v_have_step != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":822 + * + * if have_step: + * negative_step = step < 0 # <<<<<<<<<<<<<< + * if step == 0: + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + */ + __pyx_v_negative_step = (__pyx_v_step < 0); + + /* "View.MemoryView":823 + * if have_step: + * negative_step = step < 0 + * if step == 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + * else: + */ + __pyx_t_2 = (__pyx_v_step == 0); + if (__pyx_t_2) { + + /* "View.MemoryView":824 + * negative_step = step < 0 + * if step == 0: + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) # <<<<<<<<<<<<<< + * else: + * negative_step = False + */ + __pyx_t_3 = __pyx_memoryview_err_dim(PyExc_ValueError, __pyx_kp_s_Step_may_not_be_zero_axis_d, __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 824, __pyx_L1_error) + + /* "View.MemoryView":823 + * if have_step: + * negative_step = step < 0 + * if step == 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + * else: + */ + } + + /* "View.MemoryView":821 + * else: + * + * if have_step: # <<<<<<<<<<<<<< + * negative_step = step < 0 + * if step == 0: + */ + goto __pyx_L6; + } + + /* "View.MemoryView":826 + * _err_dim(PyExc_ValueError, "Step may not be zero (axis %d)", dim) + * else: + * negative_step = False # <<<<<<<<<<<<<< + * step = 1 + * + */ + /*else*/ { + __pyx_v_negative_step = 0; + + /* "View.MemoryView":827 + * else: + * negative_step = False + * step = 1 # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_step = 1; + } + __pyx_L6:; + + /* "View.MemoryView":830 + * + * + * if have_start: # <<<<<<<<<<<<<< + * if start < 0: + * start += shape + */ + __pyx_t_2 = (__pyx_v_have_start != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":831 + * + * if have_start: + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if start < 0: + */ + __pyx_t_2 = (__pyx_v_start < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":832 + * if have_start: + * if start < 0: + * start += shape # <<<<<<<<<<<<<< + * if start < 0: + * start = 0 + */ + __pyx_v_start = (__pyx_v_start + __pyx_v_shape); + + /* "View.MemoryView":833 + * if start < 0: + * start += shape + * if start < 0: # <<<<<<<<<<<<<< + * start = 0 + * elif start >= shape: + */ + __pyx_t_2 = (__pyx_v_start < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":834 + * start += shape + * if start < 0: + * start = 0 # <<<<<<<<<<<<<< + * elif start >= shape: + * if negative_step: + */ + __pyx_v_start = 0; + + /* "View.MemoryView":833 + * if start < 0: + * start += shape + * if start < 0: # <<<<<<<<<<<<<< + * start = 0 + * elif start >= shape: + */ + } + + /* "View.MemoryView":831 + * + * if have_start: + * if start < 0: # <<<<<<<<<<<<<< + * start += shape + * if start < 0: + */ + goto __pyx_L9; + } + + /* "View.MemoryView":835 + * if start < 0: + * start = 0 + * elif start >= shape: # <<<<<<<<<<<<<< + * if negative_step: + * start = shape - 1 + */ + __pyx_t_2 = (__pyx_v_start >= __pyx_v_shape); + if (__pyx_t_2) { + + /* "View.MemoryView":836 + * start = 0 + * elif start >= shape: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + if (__pyx_v_negative_step) { + + /* "View.MemoryView":837 + * elif start >= shape: + * if negative_step: + * start = shape - 1 # <<<<<<<<<<<<<< + * else: + * start = shape + */ + __pyx_v_start = (__pyx_v_shape - 1); + + /* "View.MemoryView":836 + * start = 0 + * elif start >= shape: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + goto __pyx_L11; + } + + /* "View.MemoryView":839 + * start = shape - 1 + * else: + * start = shape # <<<<<<<<<<<<<< + * else: + * if negative_step: + */ + /*else*/ { + __pyx_v_start = __pyx_v_shape; + } + __pyx_L11:; + + /* "View.MemoryView":835 + * if start < 0: + * start = 0 + * elif start >= shape: # <<<<<<<<<<<<<< + * if negative_step: + * start = shape - 1 + */ + } + __pyx_L9:; + + /* "View.MemoryView":830 + * + * + * if have_start: # <<<<<<<<<<<<<< + * if start < 0: + * start += shape + */ + goto __pyx_L8; + } + + /* "View.MemoryView":841 + * start = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + /*else*/ { + if (__pyx_v_negative_step) { + + /* "View.MemoryView":842 + * else: + * if negative_step: + * start = shape - 1 # <<<<<<<<<<<<<< + * else: + * start = 0 + */ + __pyx_v_start = (__pyx_v_shape - 1); + + /* "View.MemoryView":841 + * start = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * start = shape - 1 + * else: + */ + goto __pyx_L12; + } + + /* "View.MemoryView":844 + * start = shape - 1 + * else: + * start = 0 # <<<<<<<<<<<<<< + * + * if have_stop: + */ + /*else*/ { + __pyx_v_start = 0; + } + __pyx_L12:; + } + __pyx_L8:; + + /* "View.MemoryView":846 + * start = 0 + * + * if have_stop: # <<<<<<<<<<<<<< + * if stop < 0: + * stop += shape + */ + __pyx_t_2 = (__pyx_v_have_stop != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":847 + * + * if have_stop: + * if stop < 0: # <<<<<<<<<<<<<< + * stop += shape + * if stop < 0: + */ + __pyx_t_2 = (__pyx_v_stop < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":848 + * if have_stop: + * if stop < 0: + * stop += shape # <<<<<<<<<<<<<< + * if stop < 0: + * stop = 0 + */ + __pyx_v_stop = (__pyx_v_stop + __pyx_v_shape); + + /* "View.MemoryView":849 + * if stop < 0: + * stop += shape + * if stop < 0: # <<<<<<<<<<<<<< + * stop = 0 + * elif stop > shape: + */ + __pyx_t_2 = (__pyx_v_stop < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":850 + * stop += shape + * if stop < 0: + * stop = 0 # <<<<<<<<<<<<<< + * elif stop > shape: + * stop = shape + */ + __pyx_v_stop = 0; + + /* "View.MemoryView":849 + * if stop < 0: + * stop += shape + * if stop < 0: # <<<<<<<<<<<<<< + * stop = 0 + * elif stop > shape: + */ + } + + /* "View.MemoryView":847 + * + * if have_stop: + * if stop < 0: # <<<<<<<<<<<<<< + * stop += shape + * if stop < 0: + */ + goto __pyx_L14; + } + + /* "View.MemoryView":851 + * if stop < 0: + * stop = 0 + * elif stop > shape: # <<<<<<<<<<<<<< + * stop = shape + * else: + */ + __pyx_t_2 = (__pyx_v_stop > __pyx_v_shape); + if (__pyx_t_2) { + + /* "View.MemoryView":852 + * stop = 0 + * elif stop > shape: + * stop = shape # <<<<<<<<<<<<<< + * else: + * if negative_step: + */ + __pyx_v_stop = __pyx_v_shape; + + /* "View.MemoryView":851 + * if stop < 0: + * stop = 0 + * elif stop > shape: # <<<<<<<<<<<<<< + * stop = shape + * else: + */ + } + __pyx_L14:; + + /* "View.MemoryView":846 + * start = 0 + * + * if have_stop: # <<<<<<<<<<<<<< + * if stop < 0: + * stop += shape + */ + goto __pyx_L13; + } + + /* "View.MemoryView":854 + * stop = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * stop = -1 + * else: + */ + /*else*/ { + if (__pyx_v_negative_step) { + + /* "View.MemoryView":855 + * else: + * if negative_step: + * stop = -1 # <<<<<<<<<<<<<< + * else: + * stop = shape + */ + __pyx_v_stop = -1L; + + /* "View.MemoryView":854 + * stop = shape + * else: + * if negative_step: # <<<<<<<<<<<<<< + * stop = -1 + * else: + */ + goto __pyx_L16; + } + + /* "View.MemoryView":857 + * stop = -1 + * else: + * stop = shape # <<<<<<<<<<<<<< + * + * + */ + /*else*/ { + __pyx_v_stop = __pyx_v_shape; + } + __pyx_L16:; + } + __pyx_L13:; + + /* "View.MemoryView":861 + * + * with cython.cdivision(True): + * new_shape = (stop - start) // step # <<<<<<<<<<<<<< + * + * if (stop - start) - step * new_shape: + */ + __pyx_v_new_shape = ((__pyx_v_stop - __pyx_v_start) / __pyx_v_step); + + /* "View.MemoryView":863 + * new_shape = (stop - start) // step + * + * if (stop - start) - step * new_shape: # <<<<<<<<<<<<<< + * new_shape += 1 + * + */ + __pyx_t_2 = (((__pyx_v_stop - __pyx_v_start) - (__pyx_v_step * __pyx_v_new_shape)) != 0); + if (__pyx_t_2) { + + /* "View.MemoryView":864 + * + * if (stop - start) - step * new_shape: + * new_shape += 1 # <<<<<<<<<<<<<< + * + * if new_shape < 0: + */ + __pyx_v_new_shape = (__pyx_v_new_shape + 1); + + /* "View.MemoryView":863 + * new_shape = (stop - start) // step + * + * if (stop - start) - step * new_shape: # <<<<<<<<<<<<<< + * new_shape += 1 + * + */ + } + + /* "View.MemoryView":866 + * new_shape += 1 + * + * if new_shape < 0: # <<<<<<<<<<<<<< + * new_shape = 0 + * + */ + __pyx_t_2 = (__pyx_v_new_shape < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":867 + * + * if new_shape < 0: + * new_shape = 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_new_shape = 0; + + /* "View.MemoryView":866 + * new_shape += 1 + * + * if new_shape < 0: # <<<<<<<<<<<<<< + * new_shape = 0 + * + */ + } + + /* "View.MemoryView":870 + * + * + * dst.strides[new_ndim] = stride * step # <<<<<<<<<<<<<< + * dst.shape[new_ndim] = new_shape + * dst.suboffsets[new_ndim] = suboffset + */ + (__pyx_v_dst->strides[__pyx_v_new_ndim]) = (__pyx_v_stride * __pyx_v_step); + + /* "View.MemoryView":871 + * + * dst.strides[new_ndim] = stride * step + * dst.shape[new_ndim] = new_shape # <<<<<<<<<<<<<< + * dst.suboffsets[new_ndim] = suboffset + * + */ + (__pyx_v_dst->shape[__pyx_v_new_ndim]) = __pyx_v_new_shape; + + /* "View.MemoryView":872 + * dst.strides[new_ndim] = stride * step + * dst.shape[new_ndim] = new_shape + * dst.suboffsets[new_ndim] = suboffset # <<<<<<<<<<<<<< + * + * + */ + (__pyx_v_dst->suboffsets[__pyx_v_new_ndim]) = __pyx_v_suboffset; + } + __pyx_L3:; + + /* "View.MemoryView":875 + * + * + * if suboffset_dim[0] < 0: # <<<<<<<<<<<<<< + * dst.data += start * stride + * else: + */ + __pyx_t_2 = ((__pyx_v_suboffset_dim[0]) < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":876 + * + * if suboffset_dim[0] < 0: + * dst.data += start * stride # <<<<<<<<<<<<<< + * else: + * dst.suboffsets[suboffset_dim[0]] += start * stride + */ + __pyx_v_dst->data = (__pyx_v_dst->data + (__pyx_v_start * __pyx_v_stride)); + + /* "View.MemoryView":875 + * + * + * if suboffset_dim[0] < 0: # <<<<<<<<<<<<<< + * dst.data += start * stride + * else: + */ + goto __pyx_L19; + } + + /* "View.MemoryView":878 + * dst.data += start * stride + * else: + * dst.suboffsets[suboffset_dim[0]] += start * stride # <<<<<<<<<<<<<< + * + * if suboffset >= 0: + */ + /*else*/ { + __pyx_t_3 = (__pyx_v_suboffset_dim[0]); + (__pyx_v_dst->suboffsets[__pyx_t_3]) = ((__pyx_v_dst->suboffsets[__pyx_t_3]) + (__pyx_v_start * __pyx_v_stride)); + } + __pyx_L19:; + + /* "View.MemoryView":880 + * dst.suboffsets[suboffset_dim[0]] += start * stride + * + * if suboffset >= 0: # <<<<<<<<<<<<<< + * if not is_slice: + * if new_ndim == 0: + */ + __pyx_t_2 = (__pyx_v_suboffset >= 0); + if (__pyx_t_2) { + + /* "View.MemoryView":881 + * + * if suboffset >= 0: + * if not is_slice: # <<<<<<<<<<<<<< + * if new_ndim == 0: + * dst.data = ( dst.data)[0] + suboffset + */ + __pyx_t_2 = (!__pyx_v_is_slice); + if (__pyx_t_2) { + + /* "View.MemoryView":882 + * if suboffset >= 0: + * if not is_slice: + * if new_ndim == 0: # <<<<<<<<<<<<<< + * dst.data = ( dst.data)[0] + suboffset + * else: + */ + __pyx_t_2 = (__pyx_v_new_ndim == 0); + if (__pyx_t_2) { + + /* "View.MemoryView":883 + * if not is_slice: + * if new_ndim == 0: + * dst.data = ( dst.data)[0] + suboffset # <<<<<<<<<<<<<< + * else: + * _err_dim(PyExc_IndexError, "All dimensions preceding dimension %d " + */ + __pyx_v_dst->data = ((((char **)__pyx_v_dst->data)[0]) + __pyx_v_suboffset); + + /* "View.MemoryView":882 + * if suboffset >= 0: + * if not is_slice: + * if new_ndim == 0: # <<<<<<<<<<<<<< + * dst.data = ( dst.data)[0] + suboffset + * else: + */ + goto __pyx_L22; + } + + /* "View.MemoryView":885 + * dst.data = ( dst.data)[0] + suboffset + * else: + * _err_dim(PyExc_IndexError, "All dimensions preceding dimension %d " # <<<<<<<<<<<<<< + * "must be indexed and not sliced", dim) + * else: + */ + /*else*/ { + + /* "View.MemoryView":886 + * else: + * _err_dim(PyExc_IndexError, "All dimensions preceding dimension %d " + * "must be indexed and not sliced", dim) # <<<<<<<<<<<<<< + * else: + * suboffset_dim[0] = new_ndim + */ + __pyx_t_3 = __pyx_memoryview_err_dim(PyExc_IndexError, __pyx_kp_s_All_dimensions_preceding_dimensi, __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 885, __pyx_L1_error) + } + __pyx_L22:; + + /* "View.MemoryView":881 + * + * if suboffset >= 0: + * if not is_slice: # <<<<<<<<<<<<<< + * if new_ndim == 0: + * dst.data = ( dst.data)[0] + suboffset + */ + goto __pyx_L21; + } + + /* "View.MemoryView":888 + * "must be indexed and not sliced", dim) + * else: + * suboffset_dim[0] = new_ndim # <<<<<<<<<<<<<< + * + * return 0 + */ + /*else*/ { + (__pyx_v_suboffset_dim[0]) = __pyx_v_new_ndim; + } + __pyx_L21:; + + /* "View.MemoryView":880 + * dst.suboffsets[suboffset_dim[0]] += start * stride + * + * if suboffset >= 0: # <<<<<<<<<<<<<< + * if not is_slice: + * if new_ndim == 0: + */ + } + + /* "View.MemoryView":890 + * suboffset_dim[0] = new_ndim + * + * return 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":793 + * + * @cname('__pyx_memoryview_slice_memviewslice') + * cdef int slice_memviewslice( # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * Py_ssize_t shape, Py_ssize_t stride, Py_ssize_t suboffset, + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.slice_memviewslice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":896 + * + * @cname('__pyx_pybuffer_index') + * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index, # <<<<<<<<<<<<<< + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 + */ + +static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, Py_ssize_t __pyx_v_index, Py_ssize_t __pyx_v_dim) { + Py_ssize_t __pyx_v_shape; + Py_ssize_t __pyx_v_stride; + Py_ssize_t __pyx_v_suboffset; + Py_ssize_t __pyx_v_itemsize; + char *__pyx_v_resultp; + char *__pyx_r; + __Pyx_RefNannyDeclarations + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + Py_UCS4 __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("pybuffer_index", 1); + + /* "View.MemoryView":898 + * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index, + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 # <<<<<<<<<<<<<< + * cdef Py_ssize_t itemsize = view.itemsize + * cdef char *resultp + */ + __pyx_v_suboffset = -1L; + + /* "View.MemoryView":899 + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 + * cdef Py_ssize_t itemsize = view.itemsize # <<<<<<<<<<<<<< + * cdef char *resultp + * + */ + __pyx_t_1 = __pyx_v_view->itemsize; + __pyx_v_itemsize = __pyx_t_1; + + /* "View.MemoryView":902 + * cdef char *resultp + * + * if view.ndim == 0: # <<<<<<<<<<<<<< + * shape = view.len // itemsize + * stride = itemsize + */ + __pyx_t_2 = (__pyx_v_view->ndim == 0); + if (__pyx_t_2) { + + /* "View.MemoryView":903 + * + * if view.ndim == 0: + * shape = view.len // itemsize # <<<<<<<<<<<<<< + * stride = itemsize + * else: + */ + if (unlikely(__pyx_v_itemsize == 0)) { + PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero"); + __PYX_ERR(1, 903, __pyx_L1_error) + } + else if (sizeof(Py_ssize_t) == sizeof(long) && (!(((Py_ssize_t)-1) > 0)) && unlikely(__pyx_v_itemsize == (Py_ssize_t)-1) && unlikely(__Pyx_UNARY_NEG_WOULD_OVERFLOW(__pyx_v_view->len))) { + PyErr_SetString(PyExc_OverflowError, "value too large to perform division"); + __PYX_ERR(1, 903, __pyx_L1_error) + } + __pyx_v_shape = __Pyx_div_Py_ssize_t(__pyx_v_view->len, __pyx_v_itemsize); + + /* "View.MemoryView":904 + * if view.ndim == 0: + * shape = view.len // itemsize + * stride = itemsize # <<<<<<<<<<<<<< + * else: + * shape = view.shape[dim] + */ + __pyx_v_stride = __pyx_v_itemsize; + + /* "View.MemoryView":902 + * cdef char *resultp + * + * if view.ndim == 0: # <<<<<<<<<<<<<< + * shape = view.len // itemsize + * stride = itemsize + */ + goto __pyx_L3; + } + + /* "View.MemoryView":906 + * stride = itemsize + * else: + * shape = view.shape[dim] # <<<<<<<<<<<<<< + * stride = view.strides[dim] + * if view.suboffsets != NULL: + */ + /*else*/ { + __pyx_v_shape = (__pyx_v_view->shape[__pyx_v_dim]); + + /* "View.MemoryView":907 + * else: + * shape = view.shape[dim] + * stride = view.strides[dim] # <<<<<<<<<<<<<< + * if view.suboffsets != NULL: + * suboffset = view.suboffsets[dim] + */ + __pyx_v_stride = (__pyx_v_view->strides[__pyx_v_dim]); + + /* "View.MemoryView":908 + * shape = view.shape[dim] + * stride = view.strides[dim] + * if view.suboffsets != NULL: # <<<<<<<<<<<<<< + * suboffset = view.suboffsets[dim] + * + */ + __pyx_t_2 = (__pyx_v_view->suboffsets != NULL); + if (__pyx_t_2) { + + /* "View.MemoryView":909 + * stride = view.strides[dim] + * if view.suboffsets != NULL: + * suboffset = view.suboffsets[dim] # <<<<<<<<<<<<<< + * + * if index < 0: + */ + __pyx_v_suboffset = (__pyx_v_view->suboffsets[__pyx_v_dim]); + + /* "View.MemoryView":908 + * shape = view.shape[dim] + * stride = view.strides[dim] + * if view.suboffsets != NULL: # <<<<<<<<<<<<<< + * suboffset = view.suboffsets[dim] + * + */ + } + } + __pyx_L3:; + + /* "View.MemoryView":911 + * suboffset = view.suboffsets[dim] + * + * if index < 0: # <<<<<<<<<<<<<< + * index += view.shape[dim] + * if index < 0: + */ + __pyx_t_2 = (__pyx_v_index < 0); + if (__pyx_t_2) { + + /* "View.MemoryView":912 + * + * if index < 0: + * index += view.shape[dim] # <<<<<<<<<<<<<< + * if index < 0: + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + */ + __pyx_v_index = (__pyx_v_index + (__pyx_v_view->shape[__pyx_v_dim])); + + /* "View.MemoryView":913 + * if index < 0: + * index += view.shape[dim] + * if index < 0: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + __pyx_t_2 = (__pyx_v_index < 0); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":914 + * index += view.shape[dim] + * if index < 0: + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" # <<<<<<<<<<<<<< + * + * if index >= shape: + */ + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 914, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 = 0; + __pyx_t_4 = 127; + __Pyx_INCREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_1 += 37; + __Pyx_GIVEREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_5 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_dim, 0, ' ', 'd'); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 914, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_1 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_5); + __Pyx_GIVEREF(__pyx_t_5); + PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_5); + __pyx_t_5 = 0; + __Pyx_INCREF(__pyx_kp_u__7); + __pyx_t_1 += 1; + __Pyx_GIVEREF(__pyx_kp_u__7); + PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_kp_u__7); + __pyx_t_5 = __Pyx_PyUnicode_Join(__pyx_t_3, 3, __pyx_t_1, __pyx_t_4); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 914, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_Raise(__pyx_builtin_IndexError, __pyx_t_5, 0, 0); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __PYX_ERR(1, 914, __pyx_L1_error) + + /* "View.MemoryView":913 + * if index < 0: + * index += view.shape[dim] + * if index < 0: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + } + + /* "View.MemoryView":911 + * suboffset = view.suboffsets[dim] + * + * if index < 0: # <<<<<<<<<<<<<< + * index += view.shape[dim] + * if index < 0: + */ + } + + /* "View.MemoryView":916 + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + * if index >= shape: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + __pyx_t_2 = (__pyx_v_index >= __pyx_v_shape); + if (unlikely(__pyx_t_2)) { + + /* "View.MemoryView":917 + * + * if index >= shape: + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" # <<<<<<<<<<<<<< + * + * resultp = bufp + index * stride + */ + __pyx_t_5 = PyTuple_New(3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 917, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_1 = 0; + __pyx_t_4 = 127; + __Pyx_INCREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_1 += 37; + __Pyx_GIVEREF(__pyx_kp_u_Out_of_bounds_on_buffer_access_a); + PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_kp_u_Out_of_bounds_on_buffer_access_a); + __pyx_t_3 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_dim, 0, ' ', 'd'); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 917, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_3); + PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_t_3); + __pyx_t_3 = 0; + __Pyx_INCREF(__pyx_kp_u__7); + __pyx_t_1 += 1; + __Pyx_GIVEREF(__pyx_kp_u__7); + PyTuple_SET_ITEM(__pyx_t_5, 2, __pyx_kp_u__7); + __pyx_t_3 = __Pyx_PyUnicode_Join(__pyx_t_5, 3, __pyx_t_1, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 917, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_Raise(__pyx_builtin_IndexError, __pyx_t_3, 0, 0); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __PYX_ERR(1, 917, __pyx_L1_error) + + /* "View.MemoryView":916 + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + * if index >= shape: # <<<<<<<<<<<<<< + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + */ + } + + /* "View.MemoryView":919 + * raise IndexError, f"Out of bounds on buffer access (axis {dim})" + * + * resultp = bufp + index * stride # <<<<<<<<<<<<<< + * if suboffset >= 0: + * resultp = ( resultp)[0] + suboffset + */ + __pyx_v_resultp = (__pyx_v_bufp + (__pyx_v_index * __pyx_v_stride)); + + /* "View.MemoryView":920 + * + * resultp = bufp + index * stride + * if suboffset >= 0: # <<<<<<<<<<<<<< + * resultp = ( resultp)[0] + suboffset + * + */ + __pyx_t_2 = (__pyx_v_suboffset >= 0); + if (__pyx_t_2) { + + /* "View.MemoryView":921 + * resultp = bufp + index * stride + * if suboffset >= 0: + * resultp = ( resultp)[0] + suboffset # <<<<<<<<<<<<<< + * + * return resultp + */ + __pyx_v_resultp = ((((char **)__pyx_v_resultp)[0]) + __pyx_v_suboffset); + + /* "View.MemoryView":920 + * + * resultp = bufp + index * stride + * if suboffset >= 0: # <<<<<<<<<<<<<< + * resultp = ( resultp)[0] + suboffset + * + */ + } + + /* "View.MemoryView":923 + * resultp = ( resultp)[0] + suboffset + * + * return resultp # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = __pyx_v_resultp; + goto __pyx_L0; + + /* "View.MemoryView":896 + * + * @cname('__pyx_pybuffer_index') + * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index, # <<<<<<<<<<<<<< + * Py_ssize_t dim) except NULL: + * cdef Py_ssize_t shape, stride, suboffset = -1 + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("View.MemoryView.pybuffer_index", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":929 + * + * @cname('__pyx_memslice_transpose') + * cdef int transpose_memslice(__Pyx_memviewslice *memslice) except -1 nogil: # <<<<<<<<<<<<<< + * cdef int ndim = memslice.memview.view.ndim + * + */ + +static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) { + int __pyx_v_ndim; + Py_ssize_t *__pyx_v_shape; + Py_ssize_t *__pyx_v_strides; + int __pyx_v_i; + int __pyx_v_j; + int __pyx_r; + int __pyx_t_1; + Py_ssize_t *__pyx_t_2; + long __pyx_t_3; + long __pyx_t_4; + Py_ssize_t __pyx_t_5; + Py_ssize_t __pyx_t_6; + int __pyx_t_7; + int __pyx_t_8; + int __pyx_t_9; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":930 + * @cname('__pyx_memslice_transpose') + * cdef int transpose_memslice(__Pyx_memviewslice *memslice) except -1 nogil: + * cdef int ndim = memslice.memview.view.ndim # <<<<<<<<<<<<<< + * + * cdef Py_ssize_t *shape = memslice.shape + */ + __pyx_t_1 = __pyx_v_memslice->memview->view.ndim; + __pyx_v_ndim = __pyx_t_1; + + /* "View.MemoryView":932 + * cdef int ndim = memslice.memview.view.ndim + * + * cdef Py_ssize_t *shape = memslice.shape # <<<<<<<<<<<<<< + * cdef Py_ssize_t *strides = memslice.strides + * + */ + __pyx_t_2 = __pyx_v_memslice->shape; + __pyx_v_shape = __pyx_t_2; + + /* "View.MemoryView":933 + * + * cdef Py_ssize_t *shape = memslice.shape + * cdef Py_ssize_t *strides = memslice.strides # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = __pyx_v_memslice->strides; + __pyx_v_strides = __pyx_t_2; + + /* "View.MemoryView":937 + * + * cdef int i, j + * for i in range(ndim // 2): # <<<<<<<<<<<<<< + * j = ndim - 1 - i + * strides[i], strides[j] = strides[j], strides[i] + */ + __pyx_t_3 = __Pyx_div_long(__pyx_v_ndim, 2); + __pyx_t_4 = __pyx_t_3; + for (__pyx_t_1 = 0; __pyx_t_1 < __pyx_t_4; __pyx_t_1+=1) { + __pyx_v_i = __pyx_t_1; + + /* "View.MemoryView":938 + * cdef int i, j + * for i in range(ndim // 2): + * j = ndim - 1 - i # <<<<<<<<<<<<<< + * strides[i], strides[j] = strides[j], strides[i] + * shape[i], shape[j] = shape[j], shape[i] + */ + __pyx_v_j = ((__pyx_v_ndim - 1) - __pyx_v_i); + + /* "View.MemoryView":939 + * for i in range(ndim // 2): + * j = ndim - 1 - i + * strides[i], strides[j] = strides[j], strides[i] # <<<<<<<<<<<<<< + * shape[i], shape[j] = shape[j], shape[i] + * + */ + __pyx_t_5 = (__pyx_v_strides[__pyx_v_j]); + __pyx_t_6 = (__pyx_v_strides[__pyx_v_i]); + (__pyx_v_strides[__pyx_v_i]) = __pyx_t_5; + (__pyx_v_strides[__pyx_v_j]) = __pyx_t_6; + + /* "View.MemoryView":940 + * j = ndim - 1 - i + * strides[i], strides[j] = strides[j], strides[i] + * shape[i], shape[j] = shape[j], shape[i] # <<<<<<<<<<<<<< + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: + */ + __pyx_t_6 = (__pyx_v_shape[__pyx_v_j]); + __pyx_t_5 = (__pyx_v_shape[__pyx_v_i]); + (__pyx_v_shape[__pyx_v_i]) = __pyx_t_6; + (__pyx_v_shape[__pyx_v_j]) = __pyx_t_5; + + /* "View.MemoryView":942 + * shape[i], shape[j] = shape[j], shape[i] + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: # <<<<<<<<<<<<<< + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") + * + */ + __pyx_t_8 = ((__pyx_v_memslice->suboffsets[__pyx_v_i]) >= 0); + if (!__pyx_t_8) { + } else { + __pyx_t_7 = __pyx_t_8; + goto __pyx_L6_bool_binop_done; + } + __pyx_t_8 = ((__pyx_v_memslice->suboffsets[__pyx_v_j]) >= 0); + __pyx_t_7 = __pyx_t_8; + __pyx_L6_bool_binop_done:; + if (__pyx_t_7) { + + /* "View.MemoryView":943 + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") # <<<<<<<<<<<<<< + * + * return 0 + */ + __pyx_t_9 = __pyx_memoryview_err(PyExc_ValueError, __pyx_kp_s_Cannot_transpose_memoryview_with); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 943, __pyx_L1_error) + + /* "View.MemoryView":942 + * shape[i], shape[j] = shape[j], shape[i] + * + * if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0: # <<<<<<<<<<<<<< + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") + * + */ + } + } + + /* "View.MemoryView":945 + * _err(PyExc_ValueError, "Cannot transpose memoryview with indirect dimensions") + * + * return 0 # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":929 + * + * @cname('__pyx_memslice_transpose') + * cdef int transpose_memslice(__Pyx_memviewslice *memslice) except -1 nogil: # <<<<<<<<<<<<<< + * cdef int ndim = memslice.memview.view.ndim + * + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.transpose_memslice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":963 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * def __dealloc__(self): # <<<<<<<<<<<<<< + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + */ + +/* Python wrapper */ +static void __pyx_memoryviewslice___dealloc__(PyObject *__pyx_v_self); /*proto*/ +static void __pyx_memoryviewslice___dealloc__(PyObject *__pyx_v_self) { + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__dealloc__ (wrapper)", 0); + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewslice___dealloc__(((struct __pyx_memoryviewslice_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); +} + +static void __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewslice___dealloc__(struct __pyx_memoryviewslice_obj *__pyx_v_self) { + + /* "View.MemoryView":964 + * + * def __dealloc__(self): + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) # <<<<<<<<<<<<<< + * + * cdef convert_item_to_object(self, char *itemp): + */ + __PYX_XCLEAR_MEMVIEW((&__pyx_v_self->from_slice), 1); + + /* "View.MemoryView":963 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * def __dealloc__(self): # <<<<<<<<<<<<<< + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + */ + + /* function exit code */ +} + +/* "View.MemoryView":966 + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * if self.to_object_func != NULL: + * return self.to_object_func(itemp) + */ + +static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("convert_item_to_object", 1); + + /* "View.MemoryView":967 + * + * cdef convert_item_to_object(self, char *itemp): + * if self.to_object_func != NULL: # <<<<<<<<<<<<<< + * return self.to_object_func(itemp) + * else: + */ + __pyx_t_1 = (__pyx_v_self->to_object_func != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":968 + * cdef convert_item_to_object(self, char *itemp): + * if self.to_object_func != NULL: + * return self.to_object_func(itemp) # <<<<<<<<<<<<<< + * else: + * return memoryview.convert_item_to_object(self, itemp) + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_v_self->to_object_func(__pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 968, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + + /* "View.MemoryView":967 + * + * cdef convert_item_to_object(self, char *itemp): + * if self.to_object_func != NULL: # <<<<<<<<<<<<<< + * return self.to_object_func(itemp) + * else: + */ + } + + /* "View.MemoryView":970 + * return self.to_object_func(itemp) + * else: + * return memoryview.convert_item_to_object(self, itemp) # <<<<<<<<<<<<<< + * + * cdef assign_item_from_object(self, char *itemp, object value): + */ + /*else*/ { + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_memoryview_convert_item_to_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 970, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_r = __pyx_t_2; + __pyx_t_2 = 0; + goto __pyx_L0; + } + + /* "View.MemoryView":966 + * __PYX_XCLEAR_MEMVIEW(&self.from_slice, 1) + * + * cdef convert_item_to_object(self, char *itemp): # <<<<<<<<<<<<<< + * if self.to_object_func != NULL: + * return self.to_object_func(itemp) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.convert_item_to_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":972 + * return memoryview.convert_item_to_object(self, itemp) + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * if self.to_dtype_func != NULL: + * self.to_dtype_func(itemp, value) + */ + +static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memoryviewslice_obj *__pyx_v_self, char *__pyx_v_itemp, PyObject *__pyx_v_value) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("assign_item_from_object", 1); + + /* "View.MemoryView":973 + * + * cdef assign_item_from_object(self, char *itemp, object value): + * if self.to_dtype_func != NULL: # <<<<<<<<<<<<<< + * self.to_dtype_func(itemp, value) + * else: + */ + __pyx_t_1 = (__pyx_v_self->to_dtype_func != NULL); + if (__pyx_t_1) { + + /* "View.MemoryView":974 + * cdef assign_item_from_object(self, char *itemp, object value): + * if self.to_dtype_func != NULL: + * self.to_dtype_func(itemp, value) # <<<<<<<<<<<<<< + * else: + * memoryview.assign_item_from_object(self, itemp, value) + */ + __pyx_t_2 = __pyx_v_self->to_dtype_func(__pyx_v_itemp, __pyx_v_value); if (unlikely(__pyx_t_2 == ((int)0))) __PYX_ERR(1, 974, __pyx_L1_error) + + /* "View.MemoryView":973 + * + * cdef assign_item_from_object(self, char *itemp, object value): + * if self.to_dtype_func != NULL: # <<<<<<<<<<<<<< + * self.to_dtype_func(itemp, value) + * else: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":976 + * self.to_dtype_func(itemp, value) + * else: + * memoryview.assign_item_from_object(self, itemp, value) # <<<<<<<<<<<<<< + * + * cdef _get_base(self): + */ + /*else*/ { + __pyx_t_3 = __pyx_memoryview_assign_item_from_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 976, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + __pyx_L3:; + + /* "View.MemoryView":972 + * return memoryview.convert_item_to_object(self, itemp) + * + * cdef assign_item_from_object(self, char *itemp, object value): # <<<<<<<<<<<<<< + * if self.to_dtype_func != NULL: + * self.to_dtype_func(itemp, value) + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.assign_item_from_object", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":978 + * memoryview.assign_item_from_object(self, itemp, value) + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.from_object + * + */ + +static PyObject *__pyx_memoryviewslice__get_base(struct __pyx_memoryviewslice_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("_get_base", 1); + + /* "View.MemoryView":979 + * + * cdef _get_base(self): + * return self.from_object # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v_self->from_object); + __pyx_r = __pyx_v_self->from_object; + goto __pyx_L0; + + /* "View.MemoryView":978 + * memoryview.assign_item_from_object(self, itemp, value) + * + * cdef _get_base(self): # <<<<<<<<<<<<<< + * return self.from_object + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryviewslice_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryviewslice_1__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf___pyx_memoryviewslice___reduce_cython__(((struct __pyx_memoryviewslice_obj *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryviewslice___reduce_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":2 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 2, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + +/* Python wrapper */ +static PyObject *__pyx_pw___pyx_memoryviewslice_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyObject *__pyx_pw___pyx_memoryviewslice_3__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + CYTHON_UNUSED PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 3, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 3, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 3, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf___pyx_memoryviewslice_2__setstate_cython__(((struct __pyx_memoryviewslice_obj *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf___pyx_memoryviewslice_2__setstate_cython__(CYTHON_UNUSED struct __pyx_memoryviewslice_obj *__pyx_v_self, CYTHON_UNUSED PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":4 + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" # <<<<<<<<<<<<<< + */ + __Pyx_Raise(__pyx_builtin_TypeError, __pyx_kp_s_no_default___reduce___due_to_non, 0, 0); + __PYX_ERR(1, 4, __pyx_L1_error) + + /* "(tree fragment)":3 + * def __reduce_cython__(self): + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * raise TypeError, "no default __reduce__ due to non-trivial __cinit__" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._memoryviewslice.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":999 + * + * @cname('__pyx_memoryview_fromslice') + * cdef memoryview_fromslice(__Pyx_memviewslice memviewslice, # <<<<<<<<<<<<<< + * int ndim, + * object (*to_object_func)(char *), + */ + +static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewslice, int __pyx_v_ndim, PyObject *(*__pyx_v_to_object_func)(char *), int (*__pyx_v_to_dtype_func)(char *, PyObject *), int __pyx_v_dtype_is_object) { + struct __pyx_memoryviewslice_obj *__pyx_v_result = 0; + Py_ssize_t __pyx_v_suboffset; + PyObject *__pyx_v_length = NULL; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + __Pyx_TypeInfo *__pyx_t_4; + Py_buffer __pyx_t_5; + Py_ssize_t *__pyx_t_6; + Py_ssize_t *__pyx_t_7; + Py_ssize_t *__pyx_t_8; + Py_ssize_t __pyx_t_9; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_fromslice", 1); + + /* "View.MemoryView":1007 + * cdef _memoryviewslice result + * + * if memviewslice.memview == Py_None: # <<<<<<<<<<<<<< + * return None + * + */ + __pyx_t_1 = (((PyObject *)__pyx_v_memviewslice.memview) == Py_None); + if (__pyx_t_1) { + + /* "View.MemoryView":1008 + * + * if memviewslice.memview == Py_None: + * return None # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + + /* "View.MemoryView":1007 + * cdef _memoryviewslice result + * + * if memviewslice.memview == Py_None: # <<<<<<<<<<<<<< + * return None + * + */ + } + + /* "View.MemoryView":1013 + * + * + * result = _memoryviewslice.__new__(_memoryviewslice, None, 0, dtype_is_object) # <<<<<<<<<<<<<< + * + * result.from_slice = memviewslice + */ + __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1013, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(Py_None); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, Py_None)) __PYX_ERR(1, 1013, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_0); + __Pyx_GIVEREF(__pyx_int_0); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_int_0)) __PYX_ERR(1, 1013, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error); + __pyx_t_2 = 0; + __pyx_t_2 = ((PyObject *)__pyx_tp_new__memoryviewslice(((PyTypeObject *)__pyx_memoryviewslice_type), __pyx_t_3, NULL)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error) + __Pyx_GOTREF((PyObject *)__pyx_t_2); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_v_result = ((struct __pyx_memoryviewslice_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":1015 + * result = _memoryviewslice.__new__(_memoryviewslice, None, 0, dtype_is_object) + * + * result.from_slice = memviewslice # <<<<<<<<<<<<<< + * __PYX_INC_MEMVIEW(&memviewslice, 1) + * + */ + __pyx_v_result->from_slice = __pyx_v_memviewslice; + + /* "View.MemoryView":1016 + * + * result.from_slice = memviewslice + * __PYX_INC_MEMVIEW(&memviewslice, 1) # <<<<<<<<<<<<<< + * + * result.from_object = ( memviewslice.memview)._get_base() + */ + __PYX_INC_MEMVIEW((&__pyx_v_memviewslice), 1); + + /* "View.MemoryView":1018 + * __PYX_INC_MEMVIEW(&memviewslice, 1) + * + * result.from_object = ( memviewslice.memview)._get_base() # <<<<<<<<<<<<<< + * result.typeinfo = memviewslice.memview.typeinfo + * + */ + __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)((struct __pyx_memoryview_obj *)__pyx_v_memviewslice.memview)->__pyx_vtab)->_get_base(((struct __pyx_memoryview_obj *)__pyx_v_memviewslice.memview)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1018, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_2); + __Pyx_GOTREF(__pyx_v_result->from_object); + __Pyx_DECREF(__pyx_v_result->from_object); + __pyx_v_result->from_object = __pyx_t_2; + __pyx_t_2 = 0; + + /* "View.MemoryView":1019 + * + * result.from_object = ( memviewslice.memview)._get_base() + * result.typeinfo = memviewslice.memview.typeinfo # <<<<<<<<<<<<<< + * + * result.view = memviewslice.memview.view + */ + __pyx_t_4 = __pyx_v_memviewslice.memview->typeinfo; + __pyx_v_result->__pyx_base.typeinfo = __pyx_t_4; + + /* "View.MemoryView":1021 + * result.typeinfo = memviewslice.memview.typeinfo + * + * result.view = memviewslice.memview.view # <<<<<<<<<<<<<< + * result.view.buf = memviewslice.data + * result.view.ndim = ndim + */ + __pyx_t_5 = __pyx_v_memviewslice.memview->view; + __pyx_v_result->__pyx_base.view = __pyx_t_5; + + /* "View.MemoryView":1022 + * + * result.view = memviewslice.memview.view + * result.view.buf = memviewslice.data # <<<<<<<<<<<<<< + * result.view.ndim = ndim + * (<__pyx_buffer *> &result.view).obj = Py_None + */ + __pyx_v_result->__pyx_base.view.buf = ((void *)__pyx_v_memviewslice.data); + + /* "View.MemoryView":1023 + * result.view = memviewslice.memview.view + * result.view.buf = memviewslice.data + * result.view.ndim = ndim # <<<<<<<<<<<<<< + * (<__pyx_buffer *> &result.view).obj = Py_None + * Py_INCREF(Py_None) + */ + __pyx_v_result->__pyx_base.view.ndim = __pyx_v_ndim; + + /* "View.MemoryView":1024 + * result.view.buf = memviewslice.data + * result.view.ndim = ndim + * (<__pyx_buffer *> &result.view).obj = Py_None # <<<<<<<<<<<<<< + * Py_INCREF(Py_None) + * + */ + ((Py_buffer *)(&__pyx_v_result->__pyx_base.view))->obj = Py_None; + + /* "View.MemoryView":1025 + * result.view.ndim = ndim + * (<__pyx_buffer *> &result.view).obj = Py_None + * Py_INCREF(Py_None) # <<<<<<<<<<<<<< + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: + */ + Py_INCREF(Py_None); + + /* "View.MemoryView":1027 + * Py_INCREF(Py_None) + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: # <<<<<<<<<<<<<< + * result.flags = PyBUF_RECORDS + * else: + */ + __pyx_t_1 = ((((struct __pyx_memoryview_obj *)__pyx_v_memviewslice.memview)->flags & PyBUF_WRITABLE) != 0); + if (__pyx_t_1) { + + /* "View.MemoryView":1028 + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: + * result.flags = PyBUF_RECORDS # <<<<<<<<<<<<<< + * else: + * result.flags = PyBUF_RECORDS_RO + */ + __pyx_v_result->__pyx_base.flags = PyBUF_RECORDS; + + /* "View.MemoryView":1027 + * Py_INCREF(Py_None) + * + * if (memviewslice.memview).flags & PyBUF_WRITABLE: # <<<<<<<<<<<<<< + * result.flags = PyBUF_RECORDS + * else: + */ + goto __pyx_L4; + } + + /* "View.MemoryView":1030 + * result.flags = PyBUF_RECORDS + * else: + * result.flags = PyBUF_RECORDS_RO # <<<<<<<<<<<<<< + * + * result.view.shape = result.from_slice.shape + */ + /*else*/ { + __pyx_v_result->__pyx_base.flags = PyBUF_RECORDS_RO; + } + __pyx_L4:; + + /* "View.MemoryView":1032 + * result.flags = PyBUF_RECORDS_RO + * + * result.view.shape = result.from_slice.shape # <<<<<<<<<<<<<< + * result.view.strides = result.from_slice.strides + * + */ + __pyx_v_result->__pyx_base.view.shape = ((Py_ssize_t *)__pyx_v_result->from_slice.shape); + + /* "View.MemoryView":1033 + * + * result.view.shape = result.from_slice.shape + * result.view.strides = result.from_slice.strides # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_result->__pyx_base.view.strides = ((Py_ssize_t *)__pyx_v_result->from_slice.strides); + + /* "View.MemoryView":1036 + * + * + * result.view.suboffsets = NULL # <<<<<<<<<<<<<< + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: + */ + __pyx_v_result->__pyx_base.view.suboffsets = NULL; + + /* "View.MemoryView":1037 + * + * result.view.suboffsets = NULL + * for suboffset in result.from_slice.suboffsets[:ndim]: # <<<<<<<<<<<<<< + * if suboffset >= 0: + * result.view.suboffsets = result.from_slice.suboffsets + */ + __pyx_t_7 = (__pyx_v_result->from_slice.suboffsets + __pyx_v_ndim); + for (__pyx_t_8 = __pyx_v_result->from_slice.suboffsets; __pyx_t_8 < __pyx_t_7; __pyx_t_8++) { + __pyx_t_6 = __pyx_t_8; + __pyx_v_suboffset = (__pyx_t_6[0]); + + /* "View.MemoryView":1038 + * result.view.suboffsets = NULL + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * result.view.suboffsets = result.from_slice.suboffsets + * break + */ + __pyx_t_1 = (__pyx_v_suboffset >= 0); + if (__pyx_t_1) { + + /* "View.MemoryView":1039 + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: + * result.view.suboffsets = result.from_slice.suboffsets # <<<<<<<<<<<<<< + * break + * + */ + __pyx_v_result->__pyx_base.view.suboffsets = ((Py_ssize_t *)__pyx_v_result->from_slice.suboffsets); + + /* "View.MemoryView":1040 + * if suboffset >= 0: + * result.view.suboffsets = result.from_slice.suboffsets + * break # <<<<<<<<<<<<<< + * + * result.view.len = result.view.itemsize + */ + goto __pyx_L6_break; + + /* "View.MemoryView":1038 + * result.view.suboffsets = NULL + * for suboffset in result.from_slice.suboffsets[:ndim]: + * if suboffset >= 0: # <<<<<<<<<<<<<< + * result.view.suboffsets = result.from_slice.suboffsets + * break + */ + } + } + __pyx_L6_break:; + + /* "View.MemoryView":1042 + * break + * + * result.view.len = result.view.itemsize # <<<<<<<<<<<<<< + * for length in result.view.shape[:ndim]: + * result.view.len *= length + */ + __pyx_t_9 = __pyx_v_result->__pyx_base.view.itemsize; + __pyx_v_result->__pyx_base.view.len = __pyx_t_9; + + /* "View.MemoryView":1043 + * + * result.view.len = result.view.itemsize + * for length in result.view.shape[:ndim]: # <<<<<<<<<<<<<< + * result.view.len *= length + * + */ + __pyx_t_7 = (__pyx_v_result->__pyx_base.view.shape + __pyx_v_ndim); + for (__pyx_t_8 = __pyx_v_result->__pyx_base.view.shape; __pyx_t_8 < __pyx_t_7; __pyx_t_8++) { + __pyx_t_6 = __pyx_t_8; + __pyx_t_2 = PyInt_FromSsize_t((__pyx_t_6[0])); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1043, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_XDECREF_SET(__pyx_v_length, __pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":1044 + * result.view.len = result.view.itemsize + * for length in result.view.shape[:ndim]: + * result.view.len *= length # <<<<<<<<<<<<<< + * + * result.to_object_func = to_object_func + */ + __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_result->__pyx_base.view.len); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1044, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = PyNumber_InPlaceMultiply(__pyx_t_2, __pyx_v_length); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1044, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_t_3); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 1044, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __pyx_v_result->__pyx_base.view.len = __pyx_t_9; + } + + /* "View.MemoryView":1046 + * result.view.len *= length + * + * result.to_object_func = to_object_func # <<<<<<<<<<<<<< + * result.to_dtype_func = to_dtype_func + * + */ + __pyx_v_result->to_object_func = __pyx_v_to_object_func; + + /* "View.MemoryView":1047 + * + * result.to_object_func = to_object_func + * result.to_dtype_func = to_dtype_func # <<<<<<<<<<<<<< + * + * return result + */ + __pyx_v_result->to_dtype_func = __pyx_v_to_dtype_func; + + /* "View.MemoryView":1049 + * result.to_dtype_func = to_dtype_func + * + * return result # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_get_slice_from_memoryview') + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_result); + __pyx_r = ((PyObject *)__pyx_v_result); + goto __pyx_L0; + + /* "View.MemoryView":999 + * + * @cname('__pyx_memoryview_fromslice') + * cdef memoryview_fromslice(__Pyx_memviewslice memviewslice, # <<<<<<<<<<<<<< + * int ndim, + * object (*to_object_func)(char *), + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_AddTraceback("View.MemoryView.memoryview_fromslice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_result); + __Pyx_XDECREF(__pyx_v_length); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1052 + * + * @cname('__pyx_memoryview_get_slice_from_memoryview') + * cdef __Pyx_memviewslice *get_slice_from_memview(memoryview memview, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + */ + +static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __pyx_memoryview_obj *__pyx_v_memview, __Pyx_memviewslice *__pyx_v_mslice) { + struct __pyx_memoryviewslice_obj *__pyx_v_obj = 0; + __Pyx_memviewslice *__pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("get_slice_from_memview", 1); + + /* "View.MemoryView":1055 + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * obj = memview + * return &obj.from_slice + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":1056 + * cdef _memoryviewslice obj + * if isinstance(memview, _memoryviewslice): + * obj = memview # <<<<<<<<<<<<<< + * return &obj.from_slice + * else: + */ + if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 1056, __pyx_L1_error) + __pyx_t_2 = ((PyObject *)__pyx_v_memview); + __Pyx_INCREF(__pyx_t_2); + __pyx_v_obj = ((struct __pyx_memoryviewslice_obj *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "View.MemoryView":1057 + * if isinstance(memview, _memoryviewslice): + * obj = memview + * return &obj.from_slice # <<<<<<<<<<<<<< + * else: + * slice_copy(memview, mslice) + */ + __pyx_r = (&__pyx_v_obj->from_slice); + goto __pyx_L0; + + /* "View.MemoryView":1055 + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * obj = memview + * return &obj.from_slice + */ + } + + /* "View.MemoryView":1059 + * return &obj.from_slice + * else: + * slice_copy(memview, mslice) # <<<<<<<<<<<<<< + * return mslice + * + */ + /*else*/ { + __pyx_memoryview_slice_copy(__pyx_v_memview, __pyx_v_mslice); + + /* "View.MemoryView":1060 + * else: + * slice_copy(memview, mslice) + * return mslice # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_slice_copy') + */ + __pyx_r = __pyx_v_mslice; + goto __pyx_L0; + } + + /* "View.MemoryView":1052 + * + * @cname('__pyx_memoryview_get_slice_from_memoryview') + * cdef __Pyx_memviewslice *get_slice_from_memview(memoryview memview, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *mslice) except NULL: + * cdef _memoryviewslice obj + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView.get_slice_from_memview", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF((PyObject *)__pyx_v_obj); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1063 + * + * @cname('__pyx_memoryview_slice_copy') + * cdef void slice_copy(memoryview memview, __Pyx_memviewslice *dst) noexcept: # <<<<<<<<<<<<<< + * cdef int dim + * cdef (Py_ssize_t*) shape, strides, suboffsets + */ + +static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_memview, __Pyx_memviewslice *__pyx_v_dst) { + int __pyx_v_dim; + Py_ssize_t *__pyx_v_shape; + Py_ssize_t *__pyx_v_strides; + Py_ssize_t *__pyx_v_suboffsets; + Py_ssize_t *__pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + Py_ssize_t __pyx_t_5; + int __pyx_t_6; + + /* "View.MemoryView":1067 + * cdef (Py_ssize_t*) shape, strides, suboffsets + * + * shape = memview.view.shape # <<<<<<<<<<<<<< + * strides = memview.view.strides + * suboffsets = memview.view.suboffsets + */ + __pyx_t_1 = __pyx_v_memview->view.shape; + __pyx_v_shape = __pyx_t_1; + + /* "View.MemoryView":1068 + * + * shape = memview.view.shape + * strides = memview.view.strides # <<<<<<<<<<<<<< + * suboffsets = memview.view.suboffsets + * + */ + __pyx_t_1 = __pyx_v_memview->view.strides; + __pyx_v_strides = __pyx_t_1; + + /* "View.MemoryView":1069 + * shape = memview.view.shape + * strides = memview.view.strides + * suboffsets = memview.view.suboffsets # <<<<<<<<<<<<<< + * + * dst.memview = <__pyx_memoryview *> memview + */ + __pyx_t_1 = __pyx_v_memview->view.suboffsets; + __pyx_v_suboffsets = __pyx_t_1; + + /* "View.MemoryView":1071 + * suboffsets = memview.view.suboffsets + * + * dst.memview = <__pyx_memoryview *> memview # <<<<<<<<<<<<<< + * dst.data = memview.view.buf + * + */ + __pyx_v_dst->memview = ((struct __pyx_memoryview_obj *)__pyx_v_memview); + + /* "View.MemoryView":1072 + * + * dst.memview = <__pyx_memoryview *> memview + * dst.data = memview.view.buf # <<<<<<<<<<<<<< + * + * for dim in range(memview.view.ndim): + */ + __pyx_v_dst->data = ((char *)__pyx_v_memview->view.buf); + + /* "View.MemoryView":1074 + * dst.data = memview.view.buf + * + * for dim in range(memview.view.ndim): # <<<<<<<<<<<<<< + * dst.shape[dim] = shape[dim] + * dst.strides[dim] = strides[dim] + */ + __pyx_t_2 = __pyx_v_memview->view.ndim; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_dim = __pyx_t_4; + + /* "View.MemoryView":1075 + * + * for dim in range(memview.view.ndim): + * dst.shape[dim] = shape[dim] # <<<<<<<<<<<<<< + * dst.strides[dim] = strides[dim] + * dst.suboffsets[dim] = suboffsets[dim] if suboffsets else -1 + */ + (__pyx_v_dst->shape[__pyx_v_dim]) = (__pyx_v_shape[__pyx_v_dim]); + + /* "View.MemoryView":1076 + * for dim in range(memview.view.ndim): + * dst.shape[dim] = shape[dim] + * dst.strides[dim] = strides[dim] # <<<<<<<<<<<<<< + * dst.suboffsets[dim] = suboffsets[dim] if suboffsets else -1 + * + */ + (__pyx_v_dst->strides[__pyx_v_dim]) = (__pyx_v_strides[__pyx_v_dim]); + + /* "View.MemoryView":1077 + * dst.shape[dim] = shape[dim] + * dst.strides[dim] = strides[dim] + * dst.suboffsets[dim] = suboffsets[dim] if suboffsets else -1 # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_copy_object') + */ + __pyx_t_6 = (__pyx_v_suboffsets != 0); + if (__pyx_t_6) { + __pyx_t_5 = (__pyx_v_suboffsets[__pyx_v_dim]); + } else { + __pyx_t_5 = -1L; + } + (__pyx_v_dst->suboffsets[__pyx_v_dim]) = __pyx_t_5; + } + + /* "View.MemoryView":1063 + * + * @cname('__pyx_memoryview_slice_copy') + * cdef void slice_copy(memoryview memview, __Pyx_memviewslice *dst) noexcept: # <<<<<<<<<<<<<< + * cdef int dim + * cdef (Py_ssize_t*) shape, strides, suboffsets + */ + + /* function exit code */ +} + +/* "View.MemoryView":1080 + * + * @cname('__pyx_memoryview_copy_object') + * cdef memoryview_copy(memoryview memview): # <<<<<<<<<<<<<< + * "Create a new memoryview object" + * cdef __Pyx_memviewslice memviewslice + */ + +static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *__pyx_v_memview) { + __Pyx_memviewslice __pyx_v_memviewslice; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_copy", 1); + + /* "View.MemoryView":1083 + * "Create a new memoryview object" + * cdef __Pyx_memviewslice memviewslice + * slice_copy(memview, &memviewslice) # <<<<<<<<<<<<<< + * return memoryview_copy_from_slice(memview, &memviewslice) + * + */ + __pyx_memoryview_slice_copy(__pyx_v_memview, (&__pyx_v_memviewslice)); + + /* "View.MemoryView":1084 + * cdef __Pyx_memviewslice memviewslice + * slice_copy(memview, &memviewslice) + * return memoryview_copy_from_slice(memview, &memviewslice) # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_copy_object_from_slice') + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = __pyx_memoryview_copy_object_from_slice(__pyx_v_memview, (&__pyx_v_memviewslice)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1084, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "View.MemoryView":1080 + * + * @cname('__pyx_memoryview_copy_object') + * cdef memoryview_copy(memoryview memview): # <<<<<<<<<<<<<< + * "Create a new memoryview object" + * cdef __Pyx_memviewslice memviewslice + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("View.MemoryView.memoryview_copy", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1087 + * + * @cname('__pyx_memoryview_copy_object_from_slice') + * cdef memoryview_copy_from_slice(memoryview memview, __Pyx_memviewslice *memviewslice): # <<<<<<<<<<<<<< + * """ + * Create a new memoryview object from a given memoryview object and slice. + */ + +static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview_obj *__pyx_v_memview, __Pyx_memviewslice *__pyx_v_memviewslice) { + PyObject *(*__pyx_v_to_object_func)(char *); + int (*__pyx_v_to_dtype_func)(char *, PyObject *); + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *(*__pyx_t_2)(char *); + int (*__pyx_t_3)(char *, PyObject *); + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("memoryview_copy_from_slice", 1); + + /* "View.MemoryView":1094 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * to_object_func = (<_memoryviewslice> memview).to_object_func + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + */ + __pyx_t_1 = __Pyx_TypeCheck(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type); + if (__pyx_t_1) { + + /* "View.MemoryView":1095 + * + * if isinstance(memview, _memoryviewslice): + * to_object_func = (<_memoryviewslice> memview).to_object_func # <<<<<<<<<<<<<< + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + * else: + */ + __pyx_t_2 = ((struct __pyx_memoryviewslice_obj *)__pyx_v_memview)->to_object_func; + __pyx_v_to_object_func = __pyx_t_2; + + /* "View.MemoryView":1096 + * if isinstance(memview, _memoryviewslice): + * to_object_func = (<_memoryviewslice> memview).to_object_func + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func # <<<<<<<<<<<<<< + * else: + * to_object_func = NULL + */ + __pyx_t_3 = ((struct __pyx_memoryviewslice_obj *)__pyx_v_memview)->to_dtype_func; + __pyx_v_to_dtype_func = __pyx_t_3; + + /* "View.MemoryView":1094 + * cdef int (*to_dtype_func)(char *, object) except 0 + * + * if isinstance(memview, _memoryviewslice): # <<<<<<<<<<<<<< + * to_object_func = (<_memoryviewslice> memview).to_object_func + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1098 + * to_dtype_func = (<_memoryviewslice> memview).to_dtype_func + * else: + * to_object_func = NULL # <<<<<<<<<<<<<< + * to_dtype_func = NULL + * + */ + /*else*/ { + __pyx_v_to_object_func = NULL; + + /* "View.MemoryView":1099 + * else: + * to_object_func = NULL + * to_dtype_func = NULL # <<<<<<<<<<<<<< + * + * return memoryview_fromslice(memviewslice[0], memview.view.ndim, + */ + __pyx_v_to_dtype_func = NULL; + } + __pyx_L3:; + + /* "View.MemoryView":1101 + * to_dtype_func = NULL + * + * return memoryview_fromslice(memviewslice[0], memview.view.ndim, # <<<<<<<<<<<<<< + * to_object_func, to_dtype_func, + * memview.dtype_is_object) + */ + __Pyx_XDECREF(__pyx_r); + + /* "View.MemoryView":1103 + * return memoryview_fromslice(memviewslice[0], memview.view.ndim, + * to_object_func, to_dtype_func, + * memview.dtype_is_object) # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_4 = __pyx_memoryview_fromslice((__pyx_v_memviewslice[0]), __pyx_v_memview->view.ndim, __pyx_v_to_object_func, __pyx_v_to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1101, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_r = __pyx_t_4; + __pyx_t_4 = 0; + goto __pyx_L0; + + /* "View.MemoryView":1087 + * + * @cname('__pyx_memoryview_copy_object_from_slice') + * cdef memoryview_copy_from_slice(memoryview memview, __Pyx_memviewslice *memviewslice): # <<<<<<<<<<<<<< + * """ + * Create a new memoryview object from a given memoryview object and slice. + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.memoryview_copy_from_slice", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "View.MemoryView":1109 + * + * + * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) noexcept nogil: # <<<<<<<<<<<<<< + * return -arg if arg < 0 else arg + * + */ + +static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) { + Py_ssize_t __pyx_r; + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + + /* "View.MemoryView":1110 + * + * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) noexcept nogil: + * return -arg if arg < 0 else arg # <<<<<<<<<<<<<< + * + * @cname('__pyx_get_best_slice_order') + */ + __pyx_t_2 = (__pyx_v_arg < 0); + if (__pyx_t_2) { + __pyx_t_1 = (-__pyx_v_arg); + } else { + __pyx_t_1 = __pyx_v_arg; + } + __pyx_r = __pyx_t_1; + goto __pyx_L0; + + /* "View.MemoryView":1109 + * + * + * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) noexcept nogil: # <<<<<<<<<<<<<< + * return -arg if arg < 0 else arg + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1113 + * + * @cname('__pyx_get_best_slice_order') + * cdef char get_best_order(__Pyx_memviewslice *mslice, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * Figure out the best memory access order for a given slice. + */ + +static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int __pyx_v_ndim) { + int __pyx_v_i; + Py_ssize_t __pyx_v_c_stride; + Py_ssize_t __pyx_v_f_stride; + char __pyx_r; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + + /* "View.MemoryView":1118 + * """ + * cdef int i + * cdef Py_ssize_t c_stride = 0 # <<<<<<<<<<<<<< + * cdef Py_ssize_t f_stride = 0 + * + */ + __pyx_v_c_stride = 0; + + /* "View.MemoryView":1119 + * cdef int i + * cdef Py_ssize_t c_stride = 0 + * cdef Py_ssize_t f_stride = 0 # <<<<<<<<<<<<<< + * + * for i in range(ndim - 1, -1, -1): + */ + __pyx_v_f_stride = 0; + + /* "View.MemoryView":1121 + * cdef Py_ssize_t f_stride = 0 + * + * for i in range(ndim - 1, -1, -1): # <<<<<<<<<<<<<< + * if mslice.shape[i] > 1: + * c_stride = mslice.strides[i] + */ + for (__pyx_t_1 = (__pyx_v_ndim - 1); __pyx_t_1 > -1; __pyx_t_1-=1) { + __pyx_v_i = __pyx_t_1; + + /* "View.MemoryView":1122 + * + * for i in range(ndim - 1, -1, -1): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * c_stride = mslice.strides[i] + * break + */ + __pyx_t_2 = ((__pyx_v_mslice->shape[__pyx_v_i]) > 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1123 + * for i in range(ndim - 1, -1, -1): + * if mslice.shape[i] > 1: + * c_stride = mslice.strides[i] # <<<<<<<<<<<<<< + * break + * + */ + __pyx_v_c_stride = (__pyx_v_mslice->strides[__pyx_v_i]); + + /* "View.MemoryView":1124 + * if mslice.shape[i] > 1: + * c_stride = mslice.strides[i] + * break # <<<<<<<<<<<<<< + * + * for i in range(ndim): + */ + goto __pyx_L4_break; + + /* "View.MemoryView":1122 + * + * for i in range(ndim - 1, -1, -1): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * c_stride = mslice.strides[i] + * break + */ + } + } + __pyx_L4_break:; + + /* "View.MemoryView":1126 + * break + * + * for i in range(ndim): # <<<<<<<<<<<<<< + * if mslice.shape[i] > 1: + * f_stride = mslice.strides[i] + */ + __pyx_t_1 = __pyx_v_ndim; + __pyx_t_3 = __pyx_t_1; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1127 + * + * for i in range(ndim): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * f_stride = mslice.strides[i] + * break + */ + __pyx_t_2 = ((__pyx_v_mslice->shape[__pyx_v_i]) > 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1128 + * for i in range(ndim): + * if mslice.shape[i] > 1: + * f_stride = mslice.strides[i] # <<<<<<<<<<<<<< + * break + * + */ + __pyx_v_f_stride = (__pyx_v_mslice->strides[__pyx_v_i]); + + /* "View.MemoryView":1129 + * if mslice.shape[i] > 1: + * f_stride = mslice.strides[i] + * break # <<<<<<<<<<<<<< + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): + */ + goto __pyx_L7_break; + + /* "View.MemoryView":1127 + * + * for i in range(ndim): + * if mslice.shape[i] > 1: # <<<<<<<<<<<<<< + * f_stride = mslice.strides[i] + * break + */ + } + } + __pyx_L7_break:; + + /* "View.MemoryView":1131 + * break + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): # <<<<<<<<<<<<<< + * return 'C' + * else: + */ + __pyx_t_2 = (abs_py_ssize_t(__pyx_v_c_stride) <= abs_py_ssize_t(__pyx_v_f_stride)); + if (__pyx_t_2) { + + /* "View.MemoryView":1132 + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): + * return 'C' # <<<<<<<<<<<<<< + * else: + * return 'F' + */ + __pyx_r = 'C'; + goto __pyx_L0; + + /* "View.MemoryView":1131 + * break + * + * if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride): # <<<<<<<<<<<<<< + * return 'C' + * else: + */ + } + + /* "View.MemoryView":1134 + * return 'C' + * else: + * return 'F' # <<<<<<<<<<<<<< + * + * @cython.cdivision(True) + */ + /*else*/ { + __pyx_r = 'F'; + goto __pyx_L0; + } + + /* "View.MemoryView":1113 + * + * @cname('__pyx_get_best_slice_order') + * cdef char get_best_order(__Pyx_memviewslice *mslice, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * Figure out the best memory access order for a given slice. + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1137 + * + * @cython.cdivision(True) + * cdef void _copy_strided_to_strided(char *src_data, Py_ssize_t *src_strides, # <<<<<<<<<<<<<< + * char *dst_data, Py_ssize_t *dst_strides, + * Py_ssize_t *src_shape, Py_ssize_t *dst_shape, + */ + +static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v_src_strides, char *__pyx_v_dst_data, Py_ssize_t *__pyx_v_dst_strides, Py_ssize_t *__pyx_v_src_shape, Py_ssize_t *__pyx_v_dst_shape, int __pyx_v_ndim, size_t __pyx_v_itemsize) { + CYTHON_UNUSED Py_ssize_t __pyx_v_i; + CYTHON_UNUSED Py_ssize_t __pyx_v_src_extent; + Py_ssize_t __pyx_v_dst_extent; + Py_ssize_t __pyx_v_src_stride; + Py_ssize_t __pyx_v_dst_stride; + int __pyx_t_1; + int __pyx_t_2; + Py_ssize_t __pyx_t_3; + Py_ssize_t __pyx_t_4; + Py_ssize_t __pyx_t_5; + + /* "View.MemoryView":1144 + * + * cdef Py_ssize_t i + * cdef Py_ssize_t src_extent = src_shape[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t dst_extent = dst_shape[0] + * cdef Py_ssize_t src_stride = src_strides[0] + */ + __pyx_v_src_extent = (__pyx_v_src_shape[0]); + + /* "View.MemoryView":1145 + * cdef Py_ssize_t i + * cdef Py_ssize_t src_extent = src_shape[0] + * cdef Py_ssize_t dst_extent = dst_shape[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t src_stride = src_strides[0] + * cdef Py_ssize_t dst_stride = dst_strides[0] + */ + __pyx_v_dst_extent = (__pyx_v_dst_shape[0]); + + /* "View.MemoryView":1146 + * cdef Py_ssize_t src_extent = src_shape[0] + * cdef Py_ssize_t dst_extent = dst_shape[0] + * cdef Py_ssize_t src_stride = src_strides[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t dst_stride = dst_strides[0] + * + */ + __pyx_v_src_stride = (__pyx_v_src_strides[0]); + + /* "View.MemoryView":1147 + * cdef Py_ssize_t dst_extent = dst_shape[0] + * cdef Py_ssize_t src_stride = src_strides[0] + * cdef Py_ssize_t dst_stride = dst_strides[0] # <<<<<<<<<<<<<< + * + * if ndim == 1: + */ + __pyx_v_dst_stride = (__pyx_v_dst_strides[0]); + + /* "View.MemoryView":1149 + * cdef Py_ssize_t dst_stride = dst_strides[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): + */ + __pyx_t_1 = (__pyx_v_ndim == 1); + if (__pyx_t_1) { + + /* "View.MemoryView":1150 + * + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and # <<<<<<<<<<<<<< + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) + */ + __pyx_t_2 = (__pyx_v_src_stride > 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L5_bool_binop_done; + } + __pyx_t_2 = (__pyx_v_dst_stride > 0); + if (__pyx_t_2) { + } else { + __pyx_t_1 = __pyx_t_2; + goto __pyx_L5_bool_binop_done; + } + + /* "View.MemoryView":1151 + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): # <<<<<<<<<<<<<< + * memcpy(dst_data, src_data, itemsize * dst_extent) + * else: + */ + __pyx_t_2 = (((size_t)__pyx_v_src_stride) == __pyx_v_itemsize); + if (__pyx_t_2) { + __pyx_t_2 = (__pyx_v_itemsize == ((size_t)__pyx_v_dst_stride)); + } + __pyx_t_1 = __pyx_t_2; + __pyx_L5_bool_binop_done:; + + /* "View.MemoryView":1150 + * + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and # <<<<<<<<<<<<<< + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) + */ + if (__pyx_t_1) { + + /* "View.MemoryView":1152 + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) # <<<<<<<<<<<<<< + * else: + * for i in range(dst_extent): + */ + (void)(memcpy(__pyx_v_dst_data, __pyx_v_src_data, (__pyx_v_itemsize * __pyx_v_dst_extent))); + + /* "View.MemoryView":1150 + * + * if ndim == 1: + * if (src_stride > 0 and dst_stride > 0 and # <<<<<<<<<<<<<< + * src_stride == itemsize == dst_stride): + * memcpy(dst_data, src_data, itemsize * dst_extent) + */ + goto __pyx_L4; + } + + /* "View.MemoryView":1154 + * memcpy(dst_data, src_data, itemsize * dst_extent) + * else: + * for i in range(dst_extent): # <<<<<<<<<<<<<< + * memcpy(dst_data, src_data, itemsize) + * src_data += src_stride + */ + /*else*/ { + __pyx_t_3 = __pyx_v_dst_extent; + __pyx_t_4 = __pyx_t_3; + for (__pyx_t_5 = 0; __pyx_t_5 < __pyx_t_4; __pyx_t_5+=1) { + __pyx_v_i = __pyx_t_5; + + /* "View.MemoryView":1155 + * else: + * for i in range(dst_extent): + * memcpy(dst_data, src_data, itemsize) # <<<<<<<<<<<<<< + * src_data += src_stride + * dst_data += dst_stride + */ + (void)(memcpy(__pyx_v_dst_data, __pyx_v_src_data, __pyx_v_itemsize)); + + /* "View.MemoryView":1156 + * for i in range(dst_extent): + * memcpy(dst_data, src_data, itemsize) + * src_data += src_stride # <<<<<<<<<<<<<< + * dst_data += dst_stride + * else: + */ + __pyx_v_src_data = (__pyx_v_src_data + __pyx_v_src_stride); + + /* "View.MemoryView":1157 + * memcpy(dst_data, src_data, itemsize) + * src_data += src_stride + * dst_data += dst_stride # <<<<<<<<<<<<<< + * else: + * for i in range(dst_extent): + */ + __pyx_v_dst_data = (__pyx_v_dst_data + __pyx_v_dst_stride); + } + } + __pyx_L4:; + + /* "View.MemoryView":1149 + * cdef Py_ssize_t dst_stride = dst_strides[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * if (src_stride > 0 and dst_stride > 0 and + * src_stride == itemsize == dst_stride): + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1159 + * dst_data += dst_stride + * else: + * for i in range(dst_extent): # <<<<<<<<<<<<<< + * _copy_strided_to_strided(src_data, src_strides + 1, + * dst_data, dst_strides + 1, + */ + /*else*/ { + __pyx_t_3 = __pyx_v_dst_extent; + __pyx_t_4 = __pyx_t_3; + for (__pyx_t_5 = 0; __pyx_t_5 < __pyx_t_4; __pyx_t_5+=1) { + __pyx_v_i = __pyx_t_5; + + /* "View.MemoryView":1160 + * else: + * for i in range(dst_extent): + * _copy_strided_to_strided(src_data, src_strides + 1, # <<<<<<<<<<<<<< + * dst_data, dst_strides + 1, + * src_shape + 1, dst_shape + 1, + */ + _copy_strided_to_strided(__pyx_v_src_data, (__pyx_v_src_strides + 1), __pyx_v_dst_data, (__pyx_v_dst_strides + 1), (__pyx_v_src_shape + 1), (__pyx_v_dst_shape + 1), (__pyx_v_ndim - 1), __pyx_v_itemsize); + + /* "View.MemoryView":1164 + * src_shape + 1, dst_shape + 1, + * ndim - 1, itemsize) + * src_data += src_stride # <<<<<<<<<<<<<< + * dst_data += dst_stride + * + */ + __pyx_v_src_data = (__pyx_v_src_data + __pyx_v_src_stride); + + /* "View.MemoryView":1165 + * ndim - 1, itemsize) + * src_data += src_stride + * dst_data += dst_stride # <<<<<<<<<<<<<< + * + * cdef void copy_strided_to_strided(__Pyx_memviewslice *src, + */ + __pyx_v_dst_data = (__pyx_v_dst_data + __pyx_v_dst_stride); + } + } + __pyx_L3:; + + /* "View.MemoryView":1137 + * + * @cython.cdivision(True) + * cdef void _copy_strided_to_strided(char *src_data, Py_ssize_t *src_strides, # <<<<<<<<<<<<<< + * char *dst_data, Py_ssize_t *dst_strides, + * Py_ssize_t *src_shape, Py_ssize_t *dst_shape, + */ + + /* function exit code */ +} + +/* "View.MemoryView":1167 + * dst_data += dst_stride + * + * cdef void copy_strided_to_strided(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * int ndim, size_t itemsize) noexcept nogil: + */ + +static void copy_strided_to_strided(__Pyx_memviewslice *__pyx_v_src, __Pyx_memviewslice *__pyx_v_dst, int __pyx_v_ndim, size_t __pyx_v_itemsize) { + + /* "View.MemoryView":1170 + * __Pyx_memviewslice *dst, + * int ndim, size_t itemsize) noexcept nogil: + * _copy_strided_to_strided(src.data, src.strides, dst.data, dst.strides, # <<<<<<<<<<<<<< + * src.shape, dst.shape, ndim, itemsize) + * + */ + _copy_strided_to_strided(__pyx_v_src->data, __pyx_v_src->strides, __pyx_v_dst->data, __pyx_v_dst->strides, __pyx_v_src->shape, __pyx_v_dst->shape, __pyx_v_ndim, __pyx_v_itemsize); + + /* "View.MemoryView":1167 + * dst_data += dst_stride + * + * cdef void copy_strided_to_strided(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *dst, + * int ndim, size_t itemsize) noexcept nogil: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1174 + * + * @cname('__pyx_memoryview_slice_get_size') + * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * "Return the size of the memory occupied by the slice in number of bytes" + * cdef Py_ssize_t shape, size = src.memview.view.itemsize + */ + +static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_src, int __pyx_v_ndim) { + Py_ssize_t __pyx_v_shape; + Py_ssize_t __pyx_v_size; + Py_ssize_t __pyx_r; + Py_ssize_t __pyx_t_1; + Py_ssize_t *__pyx_t_2; + Py_ssize_t *__pyx_t_3; + Py_ssize_t *__pyx_t_4; + + /* "View.MemoryView":1176 + * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) noexcept nogil: + * "Return the size of the memory occupied by the slice in number of bytes" + * cdef Py_ssize_t shape, size = src.memview.view.itemsize # <<<<<<<<<<<<<< + * + * for shape in src.shape[:ndim]: + */ + __pyx_t_1 = __pyx_v_src->memview->view.itemsize; + __pyx_v_size = __pyx_t_1; + + /* "View.MemoryView":1178 + * cdef Py_ssize_t shape, size = src.memview.view.itemsize + * + * for shape in src.shape[:ndim]: # <<<<<<<<<<<<<< + * size *= shape + * + */ + __pyx_t_3 = (__pyx_v_src->shape + __pyx_v_ndim); + for (__pyx_t_4 = __pyx_v_src->shape; __pyx_t_4 < __pyx_t_3; __pyx_t_4++) { + __pyx_t_2 = __pyx_t_4; + __pyx_v_shape = (__pyx_t_2[0]); + + /* "View.MemoryView":1179 + * + * for shape in src.shape[:ndim]: + * size *= shape # <<<<<<<<<<<<<< + * + * return size + */ + __pyx_v_size = (__pyx_v_size * __pyx_v_shape); + } + + /* "View.MemoryView":1181 + * size *= shape + * + * return size # <<<<<<<<<<<<<< + * + * @cname('__pyx_fill_contig_strides_array') + */ + __pyx_r = __pyx_v_size; + goto __pyx_L0; + + /* "View.MemoryView":1174 + * + * @cname('__pyx_memoryview_slice_get_size') + * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) noexcept nogil: # <<<<<<<<<<<<<< + * "Return the size of the memory occupied by the slice in number of bytes" + * cdef Py_ssize_t shape, size = src.memview.view.itemsize + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1184 + * + * @cname('__pyx_fill_contig_strides_array') + * cdef Py_ssize_t fill_contig_strides_array( # <<<<<<<<<<<<<< + * Py_ssize_t *shape, Py_ssize_t *strides, Py_ssize_t stride, + * int ndim, char order) noexcept nogil: + */ + +static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, Py_ssize_t __pyx_v_stride, int __pyx_v_ndim, char __pyx_v_order) { + int __pyx_v_idx; + Py_ssize_t __pyx_r; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + + /* "View.MemoryView":1193 + * cdef int idx + * + * if order == 'F': # <<<<<<<<<<<<<< + * for idx in range(ndim): + * strides[idx] = stride + */ + __pyx_t_1 = (__pyx_v_order == 'F'); + if (__pyx_t_1) { + + /* "View.MemoryView":1194 + * + * if order == 'F': + * for idx in range(ndim): # <<<<<<<<<<<<<< + * strides[idx] = stride + * stride *= shape[idx] + */ + __pyx_t_2 = __pyx_v_ndim; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_idx = __pyx_t_4; + + /* "View.MemoryView":1195 + * if order == 'F': + * for idx in range(ndim): + * strides[idx] = stride # <<<<<<<<<<<<<< + * stride *= shape[idx] + * else: + */ + (__pyx_v_strides[__pyx_v_idx]) = __pyx_v_stride; + + /* "View.MemoryView":1196 + * for idx in range(ndim): + * strides[idx] = stride + * stride *= shape[idx] # <<<<<<<<<<<<<< + * else: + * for idx in range(ndim - 1, -1, -1): + */ + __pyx_v_stride = (__pyx_v_stride * (__pyx_v_shape[__pyx_v_idx])); + } + + /* "View.MemoryView":1193 + * cdef int idx + * + * if order == 'F': # <<<<<<<<<<<<<< + * for idx in range(ndim): + * strides[idx] = stride + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1198 + * stride *= shape[idx] + * else: + * for idx in range(ndim - 1, -1, -1): # <<<<<<<<<<<<<< + * strides[idx] = stride + * stride *= shape[idx] + */ + /*else*/ { + for (__pyx_t_2 = (__pyx_v_ndim - 1); __pyx_t_2 > -1; __pyx_t_2-=1) { + __pyx_v_idx = __pyx_t_2; + + /* "View.MemoryView":1199 + * else: + * for idx in range(ndim - 1, -1, -1): + * strides[idx] = stride # <<<<<<<<<<<<<< + * stride *= shape[idx] + * + */ + (__pyx_v_strides[__pyx_v_idx]) = __pyx_v_stride; + + /* "View.MemoryView":1200 + * for idx in range(ndim - 1, -1, -1): + * strides[idx] = stride + * stride *= shape[idx] # <<<<<<<<<<<<<< + * + * return stride + */ + __pyx_v_stride = (__pyx_v_stride * (__pyx_v_shape[__pyx_v_idx])); + } + } + __pyx_L3:; + + /* "View.MemoryView":1202 + * stride *= shape[idx] + * + * return stride # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_copy_data_to_temp') + */ + __pyx_r = __pyx_v_stride; + goto __pyx_L0; + + /* "View.MemoryView":1184 + * + * @cname('__pyx_fill_contig_strides_array') + * cdef Py_ssize_t fill_contig_strides_array( # <<<<<<<<<<<<<< + * Py_ssize_t *shape, Py_ssize_t *strides, Py_ssize_t stride, + * int ndim, char order) noexcept nogil: + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1205 + * + * @cname('__pyx_memoryview_copy_data_to_temp') + * cdef void *copy_data_to_temp(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *tmpslice, + * char order, + */ + +static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src, __Pyx_memviewslice *__pyx_v_tmpslice, char __pyx_v_order, int __pyx_v_ndim) { + int __pyx_v_i; + void *__pyx_v_result; + size_t __pyx_v_itemsize; + size_t __pyx_v_size; + void *__pyx_r; + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + struct __pyx_memoryview_obj *__pyx_t_4; + int __pyx_t_5; + int __pyx_t_6; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":1216 + * cdef void *result + * + * cdef size_t itemsize = src.memview.view.itemsize # <<<<<<<<<<<<<< + * cdef size_t size = slice_get_size(src, ndim) + * + */ + __pyx_t_1 = __pyx_v_src->memview->view.itemsize; + __pyx_v_itemsize = __pyx_t_1; + + /* "View.MemoryView":1217 + * + * cdef size_t itemsize = src.memview.view.itemsize + * cdef size_t size = slice_get_size(src, ndim) # <<<<<<<<<<<<<< + * + * result = malloc(size) + */ + __pyx_v_size = __pyx_memoryview_slice_get_size(__pyx_v_src, __pyx_v_ndim); + + /* "View.MemoryView":1219 + * cdef size_t size = slice_get_size(src, ndim) + * + * result = malloc(size) # <<<<<<<<<<<<<< + * if not result: + * _err_no_memory() + */ + __pyx_v_result = malloc(__pyx_v_size); + + /* "View.MemoryView":1220 + * + * result = malloc(size) + * if not result: # <<<<<<<<<<<<<< + * _err_no_memory() + * + */ + __pyx_t_2 = (!(__pyx_v_result != 0)); + if (__pyx_t_2) { + + /* "View.MemoryView":1221 + * result = malloc(size) + * if not result: + * _err_no_memory() # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_3 = __pyx_memoryview_err_no_memory(); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 1221, __pyx_L1_error) + + /* "View.MemoryView":1220 + * + * result = malloc(size) + * if not result: # <<<<<<<<<<<<<< + * _err_no_memory() + * + */ + } + + /* "View.MemoryView":1224 + * + * + * tmpslice.data = result # <<<<<<<<<<<<<< + * tmpslice.memview = src.memview + * for i in range(ndim): + */ + __pyx_v_tmpslice->data = ((char *)__pyx_v_result); + + /* "View.MemoryView":1225 + * + * tmpslice.data = result + * tmpslice.memview = src.memview # <<<<<<<<<<<<<< + * for i in range(ndim): + * tmpslice.shape[i] = src.shape[i] + */ + __pyx_t_4 = __pyx_v_src->memview; + __pyx_v_tmpslice->memview = __pyx_t_4; + + /* "View.MemoryView":1226 + * tmpslice.data = result + * tmpslice.memview = src.memview + * for i in range(ndim): # <<<<<<<<<<<<<< + * tmpslice.shape[i] = src.shape[i] + * tmpslice.suboffsets[i] = -1 + */ + __pyx_t_3 = __pyx_v_ndim; + __pyx_t_5 = __pyx_t_3; + for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) { + __pyx_v_i = __pyx_t_6; + + /* "View.MemoryView":1227 + * tmpslice.memview = src.memview + * for i in range(ndim): + * tmpslice.shape[i] = src.shape[i] # <<<<<<<<<<<<<< + * tmpslice.suboffsets[i] = -1 + * + */ + (__pyx_v_tmpslice->shape[__pyx_v_i]) = (__pyx_v_src->shape[__pyx_v_i]); + + /* "View.MemoryView":1228 + * for i in range(ndim): + * tmpslice.shape[i] = src.shape[i] + * tmpslice.suboffsets[i] = -1 # <<<<<<<<<<<<<< + * + * fill_contig_strides_array(&tmpslice.shape[0], &tmpslice.strides[0], itemsize, ndim, order) + */ + (__pyx_v_tmpslice->suboffsets[__pyx_v_i]) = -1L; + } + + /* "View.MemoryView":1230 + * tmpslice.suboffsets[i] = -1 + * + * fill_contig_strides_array(&tmpslice.shape[0], &tmpslice.strides[0], itemsize, ndim, order) # <<<<<<<<<<<<<< + * + * + */ + (void)(__pyx_fill_contig_strides_array((&(__pyx_v_tmpslice->shape[0])), (&(__pyx_v_tmpslice->strides[0])), __pyx_v_itemsize, __pyx_v_ndim, __pyx_v_order)); + + /* "View.MemoryView":1233 + * + * + * for i in range(ndim): # <<<<<<<<<<<<<< + * if tmpslice.shape[i] == 1: + * tmpslice.strides[i] = 0 + */ + __pyx_t_3 = __pyx_v_ndim; + __pyx_t_5 = __pyx_t_3; + for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) { + __pyx_v_i = __pyx_t_6; + + /* "View.MemoryView":1234 + * + * for i in range(ndim): + * if tmpslice.shape[i] == 1: # <<<<<<<<<<<<<< + * tmpslice.strides[i] = 0 + * + */ + __pyx_t_2 = ((__pyx_v_tmpslice->shape[__pyx_v_i]) == 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1235 + * for i in range(ndim): + * if tmpslice.shape[i] == 1: + * tmpslice.strides[i] = 0 # <<<<<<<<<<<<<< + * + * if slice_is_contig(src[0], order, ndim): + */ + (__pyx_v_tmpslice->strides[__pyx_v_i]) = 0; + + /* "View.MemoryView":1234 + * + * for i in range(ndim): + * if tmpslice.shape[i] == 1: # <<<<<<<<<<<<<< + * tmpslice.strides[i] = 0 + * + */ + } + } + + /* "View.MemoryView":1237 + * tmpslice.strides[i] = 0 + * + * if slice_is_contig(src[0], order, ndim): # <<<<<<<<<<<<<< + * memcpy(result, src.data, size) + * else: + */ + __pyx_t_2 = __pyx_memviewslice_is_contig((__pyx_v_src[0]), __pyx_v_order, __pyx_v_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1238 + * + * if slice_is_contig(src[0], order, ndim): + * memcpy(result, src.data, size) # <<<<<<<<<<<<<< + * else: + * copy_strided_to_strided(src, tmpslice, ndim, itemsize) + */ + (void)(memcpy(__pyx_v_result, __pyx_v_src->data, __pyx_v_size)); + + /* "View.MemoryView":1237 + * tmpslice.strides[i] = 0 + * + * if slice_is_contig(src[0], order, ndim): # <<<<<<<<<<<<<< + * memcpy(result, src.data, size) + * else: + */ + goto __pyx_L9; + } + + /* "View.MemoryView":1240 + * memcpy(result, src.data, size) + * else: + * copy_strided_to_strided(src, tmpslice, ndim, itemsize) # <<<<<<<<<<<<<< + * + * return result + */ + /*else*/ { + copy_strided_to_strided(__pyx_v_src, __pyx_v_tmpslice, __pyx_v_ndim, __pyx_v_itemsize); + } + __pyx_L9:; + + /* "View.MemoryView":1242 + * copy_strided_to_strided(src, tmpslice, ndim, itemsize) + * + * return result # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = __pyx_v_result; + goto __pyx_L0; + + /* "View.MemoryView":1205 + * + * @cname('__pyx_memoryview_copy_data_to_temp') + * cdef void *copy_data_to_temp(__Pyx_memviewslice *src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice *tmpslice, + * char order, + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.copy_data_to_temp", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1247 + * + * @cname('__pyx_memoryview_err_extents') + * cdef int _err_extents(int i, Py_ssize_t extent1, # <<<<<<<<<<<<<< + * Py_ssize_t extent2) except -1 with gil: + * raise ValueError, f"got differing extents in dimension {i} (got {extent1} and {extent2})" + */ + +static int __pyx_memoryview_err_extents(int __pyx_v_i, Py_ssize_t __pyx_v_extent1, Py_ssize_t __pyx_v_extent2) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + Py_ssize_t __pyx_t_2; + Py_UCS4 __pyx_t_3; + PyObject *__pyx_t_4 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_RefNannySetupContext("_err_extents", 0); + + /* "View.MemoryView":1249 + * cdef int _err_extents(int i, Py_ssize_t extent1, + * Py_ssize_t extent2) except -1 with gil: + * raise ValueError, f"got differing extents in dimension {i} (got {extent1} and {extent2})" # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_err_dim') + */ + __pyx_t_1 = PyTuple_New(7); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = 0; + __pyx_t_3 = 127; + __Pyx_INCREF(__pyx_kp_u_got_differing_extents_in_dimensi); + __pyx_t_2 += 35; + __Pyx_GIVEREF(__pyx_kp_u_got_differing_extents_in_dimensi); + PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_kp_u_got_differing_extents_in_dimensi); + __pyx_t_4 = __Pyx_PyUnicode_From_int(__pyx_v_i, 0, ' ', 'd'); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_2 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_t_4); + __pyx_t_4 = 0; + __Pyx_INCREF(__pyx_kp_u_got); + __pyx_t_2 += 6; + __Pyx_GIVEREF(__pyx_kp_u_got); + PyTuple_SET_ITEM(__pyx_t_1, 2, __pyx_kp_u_got); + __pyx_t_4 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_extent1, 0, ' ', 'd'); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_2 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + PyTuple_SET_ITEM(__pyx_t_1, 3, __pyx_t_4); + __pyx_t_4 = 0; + __Pyx_INCREF(__pyx_kp_u_and); + __pyx_t_2 += 5; + __Pyx_GIVEREF(__pyx_kp_u_and); + PyTuple_SET_ITEM(__pyx_t_1, 4, __pyx_kp_u_and); + __pyx_t_4 = __Pyx_PyUnicode_From_Py_ssize_t(__pyx_v_extent2, 0, ' ', 'd'); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_2 += __Pyx_PyUnicode_GET_LENGTH(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + PyTuple_SET_ITEM(__pyx_t_1, 5, __pyx_t_4); + __pyx_t_4 = 0; + __Pyx_INCREF(__pyx_kp_u__7); + __pyx_t_2 += 1; + __Pyx_GIVEREF(__pyx_kp_u__7); + PyTuple_SET_ITEM(__pyx_t_1, 6, __pyx_kp_u__7); + __pyx_t_4 = __Pyx_PyUnicode_Join(__pyx_t_1, 7, __pyx_t_2, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1249, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_Raise(__pyx_builtin_ValueError, __pyx_t_4, 0, 0); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __PYX_ERR(1, 1249, __pyx_L1_error) + + /* "View.MemoryView":1247 + * + * @cname('__pyx_memoryview_err_extents') + * cdef int _err_extents(int i, Py_ssize_t extent1, # <<<<<<<<<<<<<< + * Py_ssize_t extent2) except -1 with gil: + * raise ValueError, f"got differing extents in dimension {i} (got {extent1} and {extent2})" + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView._err_extents", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __Pyx_RefNannyFinishContext(); + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1252 + * + * @cname('__pyx_memoryview_err_dim') + * cdef int _err_dim(PyObject *error, str msg, int dim) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg % dim + * + */ + +static int __pyx_memoryview_err_dim(PyObject *__pyx_v_error, PyObject *__pyx_v_msg, int __pyx_v_dim) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_RefNannySetupContext("_err_dim", 0); + __Pyx_INCREF(__pyx_v_msg); + + /* "View.MemoryView":1253 + * @cname('__pyx_memoryview_err_dim') + * cdef int _err_dim(PyObject *error, str msg, int dim) except -1 with gil: + * raise error, msg % dim # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_err') + */ + __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_dim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1253, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyString_FormatSafe(__pyx_v_msg, __pyx_t_1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1253, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_Raise(((PyObject *)__pyx_v_error), __pyx_t_2, 0, 0); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __PYX_ERR(1, 1253, __pyx_L1_error) + + /* "View.MemoryView":1252 + * + * @cname('__pyx_memoryview_err_dim') + * cdef int _err_dim(PyObject *error, str msg, int dim) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg % dim + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("View.MemoryView._err_dim", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __Pyx_XDECREF(__pyx_v_msg); + __Pyx_RefNannyFinishContext(); + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1256 + * + * @cname('__pyx_memoryview_err') + * cdef int _err(PyObject *error, str msg) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg + * + */ + +static int __pyx_memoryview_err(PyObject *__pyx_v_error, PyObject *__pyx_v_msg) { + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_RefNannySetupContext("_err", 0); + __Pyx_INCREF(__pyx_v_msg); + + /* "View.MemoryView":1257 + * @cname('__pyx_memoryview_err') + * cdef int _err(PyObject *error, str msg) except -1 with gil: + * raise error, msg # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_err_no_memory') + */ + __Pyx_Raise(((PyObject *)__pyx_v_error), __pyx_v_msg, 0, 0); + __PYX_ERR(1, 1257, __pyx_L1_error) + + /* "View.MemoryView":1256 + * + * @cname('__pyx_memoryview_err') + * cdef int _err(PyObject *error, str msg) except -1 with gil: # <<<<<<<<<<<<<< + * raise error, msg + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._err", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __Pyx_XDECREF(__pyx_v_msg); + __Pyx_RefNannyFinishContext(); + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1260 + * + * @cname('__pyx_memoryview_err_no_memory') + * cdef int _err_no_memory() except -1 with gil: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + +static int __pyx_memoryview_err_no_memory(void) { + int __pyx_r; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + + /* "View.MemoryView":1261 + * @cname('__pyx_memoryview_err_no_memory') + * cdef int _err_no_memory() except -1 with gil: + * raise MemoryError # <<<<<<<<<<<<<< + * + * + */ + PyErr_NoMemory(); __PYX_ERR(1, 1261, __pyx_L1_error) + + /* "View.MemoryView":1260 + * + * @cname('__pyx_memoryview_err_no_memory') + * cdef int _err_no_memory() except -1 with gil: # <<<<<<<<<<<<<< + * raise MemoryError + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_AddTraceback("View.MemoryView._err_no_memory", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + return __pyx_r; +} + +/* "View.MemoryView":1265 + * + * @cname('__pyx_memoryview_copy_contents') + * cdef int memoryview_copy_contents(__Pyx_memviewslice src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice dst, + * int src_ndim, int dst_ndim, + */ + +static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_memviewslice __pyx_v_dst, int __pyx_v_src_ndim, int __pyx_v_dst_ndim, int __pyx_v_dtype_is_object) { + void *__pyx_v_tmpdata; + size_t __pyx_v_itemsize; + int __pyx_v_i; + char __pyx_v_order; + int __pyx_v_broadcasting; + int __pyx_v_direct_copy; + __Pyx_memviewslice __pyx_v_tmp; + int __pyx_v_ndim; + int __pyx_r; + Py_ssize_t __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + int __pyx_t_4; + int __pyx_t_5; + int __pyx_t_6; + void *__pyx_t_7; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save; + #endif + + /* "View.MemoryView":1273 + * Check for overlapping memory and verify the shapes. + * """ + * cdef void *tmpdata = NULL # <<<<<<<<<<<<<< + * cdef size_t itemsize = src.memview.view.itemsize + * cdef int i + */ + __pyx_v_tmpdata = NULL; + + /* "View.MemoryView":1274 + * """ + * cdef void *tmpdata = NULL + * cdef size_t itemsize = src.memview.view.itemsize # <<<<<<<<<<<<<< + * cdef int i + * cdef char order = get_best_order(&src, src_ndim) + */ + __pyx_t_1 = __pyx_v_src.memview->view.itemsize; + __pyx_v_itemsize = __pyx_t_1; + + /* "View.MemoryView":1276 + * cdef size_t itemsize = src.memview.view.itemsize + * cdef int i + * cdef char order = get_best_order(&src, src_ndim) # <<<<<<<<<<<<<< + * cdef bint broadcasting = False + * cdef bint direct_copy = False + */ + __pyx_v_order = __pyx_get_best_slice_order((&__pyx_v_src), __pyx_v_src_ndim); + + /* "View.MemoryView":1277 + * cdef int i + * cdef char order = get_best_order(&src, src_ndim) + * cdef bint broadcasting = False # <<<<<<<<<<<<<< + * cdef bint direct_copy = False + * cdef __Pyx_memviewslice tmp + */ + __pyx_v_broadcasting = 0; + + /* "View.MemoryView":1278 + * cdef char order = get_best_order(&src, src_ndim) + * cdef bint broadcasting = False + * cdef bint direct_copy = False # <<<<<<<<<<<<<< + * cdef __Pyx_memviewslice tmp + * + */ + __pyx_v_direct_copy = 0; + + /* "View.MemoryView":1281 + * cdef __Pyx_memviewslice tmp + * + * if src_ndim < dst_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: + */ + __pyx_t_2 = (__pyx_v_src_ndim < __pyx_v_dst_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1282 + * + * if src_ndim < dst_ndim: + * broadcast_leading(&src, src_ndim, dst_ndim) # <<<<<<<<<<<<<< + * elif dst_ndim < src_ndim: + * broadcast_leading(&dst, dst_ndim, src_ndim) + */ + __pyx_memoryview_broadcast_leading((&__pyx_v_src), __pyx_v_src_ndim, __pyx_v_dst_ndim); + + /* "View.MemoryView":1281 + * cdef __Pyx_memviewslice tmp + * + * if src_ndim < dst_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1283 + * if src_ndim < dst_ndim: + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&dst, dst_ndim, src_ndim) + * + */ + __pyx_t_2 = (__pyx_v_dst_ndim < __pyx_v_src_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1284 + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: + * broadcast_leading(&dst, dst_ndim, src_ndim) # <<<<<<<<<<<<<< + * + * cdef int ndim = max(src_ndim, dst_ndim) + */ + __pyx_memoryview_broadcast_leading((&__pyx_v_dst), __pyx_v_dst_ndim, __pyx_v_src_ndim); + + /* "View.MemoryView":1283 + * if src_ndim < dst_ndim: + * broadcast_leading(&src, src_ndim, dst_ndim) + * elif dst_ndim < src_ndim: # <<<<<<<<<<<<<< + * broadcast_leading(&dst, dst_ndim, src_ndim) + * + */ + } + __pyx_L3:; + + /* "View.MemoryView":1286 + * broadcast_leading(&dst, dst_ndim, src_ndim) + * + * cdef int ndim = max(src_ndim, dst_ndim) # <<<<<<<<<<<<<< + * + * for i in range(ndim): + */ + __pyx_t_3 = __pyx_v_dst_ndim; + __pyx_t_4 = __pyx_v_src_ndim; + __pyx_t_2 = (__pyx_t_3 > __pyx_t_4); + if (__pyx_t_2) { + __pyx_t_5 = __pyx_t_3; + } else { + __pyx_t_5 = __pyx_t_4; + } + __pyx_v_ndim = __pyx_t_5; + + /* "View.MemoryView":1288 + * cdef int ndim = max(src_ndim, dst_ndim) + * + * for i in range(ndim): # <<<<<<<<<<<<<< + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: + */ + __pyx_t_5 = __pyx_v_ndim; + __pyx_t_3 = __pyx_t_5; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1289 + * + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: # <<<<<<<<<<<<<< + * if src.shape[i] == 1: + * broadcasting = True + */ + __pyx_t_2 = ((__pyx_v_src.shape[__pyx_v_i]) != (__pyx_v_dst.shape[__pyx_v_i])); + if (__pyx_t_2) { + + /* "View.MemoryView":1290 + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: # <<<<<<<<<<<<<< + * broadcasting = True + * src.strides[i] = 0 + */ + __pyx_t_2 = ((__pyx_v_src.shape[__pyx_v_i]) == 1); + if (__pyx_t_2) { + + /* "View.MemoryView":1291 + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: + * broadcasting = True # <<<<<<<<<<<<<< + * src.strides[i] = 0 + * else: + */ + __pyx_v_broadcasting = 1; + + /* "View.MemoryView":1292 + * if src.shape[i] == 1: + * broadcasting = True + * src.strides[i] = 0 # <<<<<<<<<<<<<< + * else: + * _err_extents(i, dst.shape[i], src.shape[i]) + */ + (__pyx_v_src.strides[__pyx_v_i]) = 0; + + /* "View.MemoryView":1290 + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: + * if src.shape[i] == 1: # <<<<<<<<<<<<<< + * broadcasting = True + * src.strides[i] = 0 + */ + goto __pyx_L7; + } + + /* "View.MemoryView":1294 + * src.strides[i] = 0 + * else: + * _err_extents(i, dst.shape[i], src.shape[i]) # <<<<<<<<<<<<<< + * + * if src.suboffsets[i] >= 0: + */ + /*else*/ { + __pyx_t_6 = __pyx_memoryview_err_extents(__pyx_v_i, (__pyx_v_dst.shape[__pyx_v_i]), (__pyx_v_src.shape[__pyx_v_i])); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1294, __pyx_L1_error) + } + __pyx_L7:; + + /* "View.MemoryView":1289 + * + * for i in range(ndim): + * if src.shape[i] != dst.shape[i]: # <<<<<<<<<<<<<< + * if src.shape[i] == 1: + * broadcasting = True + */ + } + + /* "View.MemoryView":1296 + * _err_extents(i, dst.shape[i], src.shape[i]) + * + * if src.suboffsets[i] >= 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + */ + __pyx_t_2 = ((__pyx_v_src.suboffsets[__pyx_v_i]) >= 0); + if (__pyx_t_2) { + + /* "View.MemoryView":1297 + * + * if src.suboffsets[i] >= 0: + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) # <<<<<<<<<<<<<< + * + * if slices_overlap(&src, &dst, ndim, itemsize): + */ + __pyx_t_6 = __pyx_memoryview_err_dim(PyExc_ValueError, __pyx_kp_s_Dimension_d_is_not_direct, __pyx_v_i); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1297, __pyx_L1_error) + + /* "View.MemoryView":1296 + * _err_extents(i, dst.shape[i], src.shape[i]) + * + * if src.suboffsets[i] >= 0: # <<<<<<<<<<<<<< + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + */ + } + } + + /* "View.MemoryView":1299 + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + * if slices_overlap(&src, &dst, ndim, itemsize): # <<<<<<<<<<<<<< + * + * if not slice_is_contig(src, order, ndim): + */ + __pyx_t_2 = __pyx_slices_overlap((&__pyx_v_src), (&__pyx_v_dst), __pyx_v_ndim, __pyx_v_itemsize); + if (__pyx_t_2) { + + /* "View.MemoryView":1301 + * if slices_overlap(&src, &dst, ndim, itemsize): + * + * if not slice_is_contig(src, order, ndim): # <<<<<<<<<<<<<< + * order = get_best_order(&dst, ndim) + * + */ + __pyx_t_2 = (!__pyx_memviewslice_is_contig(__pyx_v_src, __pyx_v_order, __pyx_v_ndim)); + if (__pyx_t_2) { + + /* "View.MemoryView":1302 + * + * if not slice_is_contig(src, order, ndim): + * order = get_best_order(&dst, ndim) # <<<<<<<<<<<<<< + * + * tmpdata = copy_data_to_temp(&src, &tmp, order, ndim) + */ + __pyx_v_order = __pyx_get_best_slice_order((&__pyx_v_dst), __pyx_v_ndim); + + /* "View.MemoryView":1301 + * if slices_overlap(&src, &dst, ndim, itemsize): + * + * if not slice_is_contig(src, order, ndim): # <<<<<<<<<<<<<< + * order = get_best_order(&dst, ndim) + * + */ + } + + /* "View.MemoryView":1304 + * order = get_best_order(&dst, ndim) + * + * tmpdata = copy_data_to_temp(&src, &tmp, order, ndim) # <<<<<<<<<<<<<< + * src = tmp + * + */ + __pyx_t_7 = __pyx_memoryview_copy_data_to_temp((&__pyx_v_src), (&__pyx_v_tmp), __pyx_v_order, __pyx_v_ndim); if (unlikely(__pyx_t_7 == ((void *)NULL))) __PYX_ERR(1, 1304, __pyx_L1_error) + __pyx_v_tmpdata = __pyx_t_7; + + /* "View.MemoryView":1305 + * + * tmpdata = copy_data_to_temp(&src, &tmp, order, ndim) + * src = tmp # <<<<<<<<<<<<<< + * + * if not broadcasting: + */ + __pyx_v_src = __pyx_v_tmp; + + /* "View.MemoryView":1299 + * _err_dim(PyExc_ValueError, "Dimension %d is not direct", i) + * + * if slices_overlap(&src, &dst, ndim, itemsize): # <<<<<<<<<<<<<< + * + * if not slice_is_contig(src, order, ndim): + */ + } + + /* "View.MemoryView":1307 + * src = tmp + * + * if not broadcasting: # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = (!__pyx_v_broadcasting); + if (__pyx_t_2) { + + /* "View.MemoryView":1310 + * + * + * if slice_is_contig(src, 'C', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): + */ + __pyx_t_2 = __pyx_memviewslice_is_contig(__pyx_v_src, 'C', __pyx_v_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1311 + * + * if slice_is_contig(src, 'C', ndim): + * direct_copy = slice_is_contig(dst, 'C', ndim) # <<<<<<<<<<<<<< + * elif slice_is_contig(src, 'F', ndim): + * direct_copy = slice_is_contig(dst, 'F', ndim) + */ + __pyx_v_direct_copy = __pyx_memviewslice_is_contig(__pyx_v_dst, 'C', __pyx_v_ndim); + + /* "View.MemoryView":1310 + * + * + * if slice_is_contig(src, 'C', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): + */ + goto __pyx_L12; + } + + /* "View.MemoryView":1312 + * if slice_is_contig(src, 'C', ndim): + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + */ + __pyx_t_2 = __pyx_memviewslice_is_contig(__pyx_v_src, 'F', __pyx_v_ndim); + if (__pyx_t_2) { + + /* "View.MemoryView":1313 + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): + * direct_copy = slice_is_contig(dst, 'F', ndim) # <<<<<<<<<<<<<< + * + * if direct_copy: + */ + __pyx_v_direct_copy = __pyx_memviewslice_is_contig(__pyx_v_dst, 'F', __pyx_v_ndim); + + /* "View.MemoryView":1312 + * if slice_is_contig(src, 'C', ndim): + * direct_copy = slice_is_contig(dst, 'C', ndim) + * elif slice_is_contig(src, 'F', ndim): # <<<<<<<<<<<<<< + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + */ + } + __pyx_L12:; + + /* "View.MemoryView":1315 + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + * if direct_copy: # <<<<<<<<<<<<<< + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + */ + if (__pyx_v_direct_copy) { + + /* "View.MemoryView":1317 + * if direct_copy: + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) # <<<<<<<<<<<<<< + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 0); + + /* "View.MemoryView":1318 + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) # <<<<<<<<<<<<<< + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * free(tmpdata) + */ + (void)(memcpy(__pyx_v_dst.data, __pyx_v_src.data, __pyx_memoryview_slice_get_size((&__pyx_v_src), __pyx_v_ndim))); + + /* "View.MemoryView":1319 + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) # <<<<<<<<<<<<<< + * free(tmpdata) + * return 0 + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 1); + + /* "View.MemoryView":1320 + * memcpy(dst.data, src.data, slice_get_size(&src, ndim)) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * free(tmpdata) # <<<<<<<<<<<<<< + * return 0 + * + */ + free(__pyx_v_tmpdata); + + /* "View.MemoryView":1321 + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * free(tmpdata) + * return 0 # <<<<<<<<<<<<<< + * + * if order == 'F' == get_best_order(&dst, ndim): + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":1315 + * direct_copy = slice_is_contig(dst, 'F', ndim) + * + * if direct_copy: # <<<<<<<<<<<<<< + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + */ + } + + /* "View.MemoryView":1307 + * src = tmp + * + * if not broadcasting: # <<<<<<<<<<<<<< + * + * + */ + } + + /* "View.MemoryView":1323 + * return 0 + * + * if order == 'F' == get_best_order(&dst, ndim): # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_2 = (__pyx_v_order == 'F'); + if (__pyx_t_2) { + __pyx_t_2 = ('F' == __pyx_get_best_slice_order((&__pyx_v_dst), __pyx_v_ndim)); + } + if (__pyx_t_2) { + + /* "View.MemoryView":1326 + * + * + * transpose_memslice(&src) # <<<<<<<<<<<<<< + * transpose_memslice(&dst) + * + */ + __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_src)); if (unlikely(__pyx_t_5 == ((int)-1))) __PYX_ERR(1, 1326, __pyx_L1_error) + + /* "View.MemoryView":1327 + * + * transpose_memslice(&src) + * transpose_memslice(&dst) # <<<<<<<<<<<<<< + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + */ + __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_dst)); if (unlikely(__pyx_t_5 == ((int)-1))) __PYX_ERR(1, 1327, __pyx_L1_error) + + /* "View.MemoryView":1323 + * return 0 + * + * if order == 'F' == get_best_order(&dst, ndim): # <<<<<<<<<<<<<< + * + * + */ + } + + /* "View.MemoryView":1329 + * transpose_memslice(&dst) + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) # <<<<<<<<<<<<<< + * copy_strided_to_strided(&src, &dst, ndim, itemsize) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 0); + + /* "View.MemoryView":1330 + * + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * copy_strided_to_strided(&src, &dst, ndim, itemsize) # <<<<<<<<<<<<<< + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * + */ + copy_strided_to_strided((&__pyx_v_src), (&__pyx_v_dst), __pyx_v_ndim, __pyx_v_itemsize); + + /* "View.MemoryView":1331 + * refcount_copying(&dst, dtype_is_object, ndim, inc=False) + * copy_strided_to_strided(&src, &dst, ndim, itemsize) + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) # <<<<<<<<<<<<<< + * + * free(tmpdata) + */ + __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 1); + + /* "View.MemoryView":1333 + * refcount_copying(&dst, dtype_is_object, ndim, inc=True) + * + * free(tmpdata) # <<<<<<<<<<<<<< + * return 0 + * + */ + free(__pyx_v_tmpdata); + + /* "View.MemoryView":1334 + * + * free(tmpdata) + * return 0 # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_broadcast_leading') + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "View.MemoryView":1265 + * + * @cname('__pyx_memoryview_copy_contents') + * cdef int memoryview_copy_contents(__Pyx_memviewslice src, # <<<<<<<<<<<<<< + * __Pyx_memviewslice dst, + * int src_ndim, int dst_ndim, + */ + + /* function exit code */ + __pyx_L1_error:; + #ifdef WITH_THREAD + __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + __Pyx_AddTraceback("View.MemoryView.memoryview_copy_contents", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif + __pyx_L0:; + return __pyx_r; +} + +/* "View.MemoryView":1337 + * + * @cname('__pyx_memoryview_broadcast_leading') + * cdef void broadcast_leading(__Pyx_memviewslice *mslice, # <<<<<<<<<<<<<< + * int ndim, + * int ndim_other) noexcept nogil: + */ + +static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslice, int __pyx_v_ndim, int __pyx_v_ndim_other) { + int __pyx_v_i; + int __pyx_v_offset; + int __pyx_t_1; + int __pyx_t_2; + int __pyx_t_3; + + /* "View.MemoryView":1341 + * int ndim_other) noexcept nogil: + * cdef int i + * cdef int offset = ndim_other - ndim # <<<<<<<<<<<<<< + * + * for i in range(ndim - 1, -1, -1): + */ + __pyx_v_offset = (__pyx_v_ndim_other - __pyx_v_ndim); + + /* "View.MemoryView":1343 + * cdef int offset = ndim_other - ndim + * + * for i in range(ndim - 1, -1, -1): # <<<<<<<<<<<<<< + * mslice.shape[i + offset] = mslice.shape[i] + * mslice.strides[i + offset] = mslice.strides[i] + */ + for (__pyx_t_1 = (__pyx_v_ndim - 1); __pyx_t_1 > -1; __pyx_t_1-=1) { + __pyx_v_i = __pyx_t_1; + + /* "View.MemoryView":1344 + * + * for i in range(ndim - 1, -1, -1): + * mslice.shape[i + offset] = mslice.shape[i] # <<<<<<<<<<<<<< + * mslice.strides[i + offset] = mslice.strides[i] + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] + */ + (__pyx_v_mslice->shape[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->shape[__pyx_v_i]); + + /* "View.MemoryView":1345 + * for i in range(ndim - 1, -1, -1): + * mslice.shape[i + offset] = mslice.shape[i] + * mslice.strides[i + offset] = mslice.strides[i] # <<<<<<<<<<<<<< + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] + * + */ + (__pyx_v_mslice->strides[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->strides[__pyx_v_i]); + + /* "View.MemoryView":1346 + * mslice.shape[i + offset] = mslice.shape[i] + * mslice.strides[i + offset] = mslice.strides[i] + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] # <<<<<<<<<<<<<< + * + * for i in range(offset): + */ + (__pyx_v_mslice->suboffsets[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->suboffsets[__pyx_v_i]); + } + + /* "View.MemoryView":1348 + * mslice.suboffsets[i + offset] = mslice.suboffsets[i] + * + * for i in range(offset): # <<<<<<<<<<<<<< + * mslice.shape[i] = 1 + * mslice.strides[i] = mslice.strides[0] + */ + __pyx_t_1 = __pyx_v_offset; + __pyx_t_2 = __pyx_t_1; + for (__pyx_t_3 = 0; __pyx_t_3 < __pyx_t_2; __pyx_t_3+=1) { + __pyx_v_i = __pyx_t_3; + + /* "View.MemoryView":1349 + * + * for i in range(offset): + * mslice.shape[i] = 1 # <<<<<<<<<<<<<< + * mslice.strides[i] = mslice.strides[0] + * mslice.suboffsets[i] = -1 + */ + (__pyx_v_mslice->shape[__pyx_v_i]) = 1; + + /* "View.MemoryView":1350 + * for i in range(offset): + * mslice.shape[i] = 1 + * mslice.strides[i] = mslice.strides[0] # <<<<<<<<<<<<<< + * mslice.suboffsets[i] = -1 + * + */ + (__pyx_v_mslice->strides[__pyx_v_i]) = (__pyx_v_mslice->strides[0]); + + /* "View.MemoryView":1351 + * mslice.shape[i] = 1 + * mslice.strides[i] = mslice.strides[0] + * mslice.suboffsets[i] = -1 # <<<<<<<<<<<<<< + * + * + */ + (__pyx_v_mslice->suboffsets[__pyx_v_i]) = -1L; + } + + /* "View.MemoryView":1337 + * + * @cname('__pyx_memoryview_broadcast_leading') + * cdef void broadcast_leading(__Pyx_memviewslice *mslice, # <<<<<<<<<<<<<< + * int ndim, + * int ndim_other) noexcept nogil: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1359 + * + * @cname('__pyx_memoryview_refcount_copying') + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: # <<<<<<<<<<<<<< + * + * if dtype_is_object: + */ + +static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *__pyx_v_dst, int __pyx_v_dtype_is_object, int __pyx_v_ndim, int __pyx_v_inc) { + + /* "View.MemoryView":1361 + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: + * + * if dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice_with_gil(dst.data, dst.shape, dst.strides, ndim, inc) + * + */ + if (__pyx_v_dtype_is_object) { + + /* "View.MemoryView":1362 + * + * if dtype_is_object: + * refcount_objects_in_slice_with_gil(dst.data, dst.shape, dst.strides, ndim, inc) # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil') + */ + __pyx_memoryview_refcount_objects_in_slice_with_gil(__pyx_v_dst->data, __pyx_v_dst->shape, __pyx_v_dst->strides, __pyx_v_ndim, __pyx_v_inc); + + /* "View.MemoryView":1361 + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: + * + * if dtype_is_object: # <<<<<<<<<<<<<< + * refcount_objects_in_slice_with_gil(dst.data, dst.shape, dst.strides, ndim, inc) + * + */ + } + + /* "View.MemoryView":1359 + * + * @cname('__pyx_memoryview_refcount_copying') + * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object, int ndim, bint inc) noexcept nogil: # <<<<<<<<<<<<<< + * + * if dtype_is_object: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1365 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil') + * cdef void refcount_objects_in_slice_with_gil(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * bint inc) noexcept with gil: + */ + +static void __pyx_memoryview_refcount_objects_in_slice_with_gil(char *__pyx_v_data, Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, int __pyx_v_ndim, int __pyx_v_inc) { + #ifdef WITH_THREAD + PyGILState_STATE __pyx_gilstate_save = __Pyx_PyGILState_Ensure(); + #endif + + /* "View.MemoryView":1368 + * Py_ssize_t *strides, int ndim, + * bint inc) noexcept with gil: + * refcount_objects_in_slice(data, shape, strides, ndim, inc) # <<<<<<<<<<<<<< + * + * @cname('__pyx_memoryview_refcount_objects_in_slice') + */ + __pyx_memoryview_refcount_objects_in_slice(__pyx_v_data, __pyx_v_shape, __pyx_v_strides, __pyx_v_ndim, __pyx_v_inc); + + /* "View.MemoryView":1365 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil') + * cdef void refcount_objects_in_slice_with_gil(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * bint inc) noexcept with gil: + */ + + /* function exit code */ + #ifdef WITH_THREAD + __Pyx_PyGILState_Release(__pyx_gilstate_save); + #endif +} + +/* "View.MemoryView":1371 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice') + * cdef void refcount_objects_in_slice(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, bint inc) noexcept: + * cdef Py_ssize_t i + */ + +static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, int __pyx_v_ndim, int __pyx_v_inc) { + CYTHON_UNUSED Py_ssize_t __pyx_v_i; + Py_ssize_t __pyx_v_stride; + Py_ssize_t __pyx_t_1; + Py_ssize_t __pyx_t_2; + Py_ssize_t __pyx_t_3; + int __pyx_t_4; + + /* "View.MemoryView":1374 + * Py_ssize_t *strides, int ndim, bint inc) noexcept: + * cdef Py_ssize_t i + * cdef Py_ssize_t stride = strides[0] # <<<<<<<<<<<<<< + * + * for i in range(shape[0]): + */ + __pyx_v_stride = (__pyx_v_strides[0]); + + /* "View.MemoryView":1376 + * cdef Py_ssize_t stride = strides[0] + * + * for i in range(shape[0]): # <<<<<<<<<<<<<< + * if ndim == 1: + * if inc: + */ + __pyx_t_1 = (__pyx_v_shape[0]); + __pyx_t_2 = __pyx_t_1; + for (__pyx_t_3 = 0; __pyx_t_3 < __pyx_t_2; __pyx_t_3+=1) { + __pyx_v_i = __pyx_t_3; + + /* "View.MemoryView":1377 + * + * for i in range(shape[0]): + * if ndim == 1: # <<<<<<<<<<<<<< + * if inc: + * Py_INCREF(( data)[0]) + */ + __pyx_t_4 = (__pyx_v_ndim == 1); + if (__pyx_t_4) { + + /* "View.MemoryView":1378 + * for i in range(shape[0]): + * if ndim == 1: + * if inc: # <<<<<<<<<<<<<< + * Py_INCREF(( data)[0]) + * else: + */ + if (__pyx_v_inc) { + + /* "View.MemoryView":1379 + * if ndim == 1: + * if inc: + * Py_INCREF(( data)[0]) # <<<<<<<<<<<<<< + * else: + * Py_DECREF(( data)[0]) + */ + Py_INCREF((((PyObject **)__pyx_v_data)[0])); + + /* "View.MemoryView":1378 + * for i in range(shape[0]): + * if ndim == 1: + * if inc: # <<<<<<<<<<<<<< + * Py_INCREF(( data)[0]) + * else: + */ + goto __pyx_L6; + } + + /* "View.MemoryView":1381 + * Py_INCREF(( data)[0]) + * else: + * Py_DECREF(( data)[0]) # <<<<<<<<<<<<<< + * else: + * refcount_objects_in_slice(data, shape + 1, strides + 1, ndim - 1, inc) + */ + /*else*/ { + Py_DECREF((((PyObject **)__pyx_v_data)[0])); + } + __pyx_L6:; + + /* "View.MemoryView":1377 + * + * for i in range(shape[0]): + * if ndim == 1: # <<<<<<<<<<<<<< + * if inc: + * Py_INCREF(( data)[0]) + */ + goto __pyx_L5; + } + + /* "View.MemoryView":1383 + * Py_DECREF(( data)[0]) + * else: + * refcount_objects_in_slice(data, shape + 1, strides + 1, ndim - 1, inc) # <<<<<<<<<<<<<< + * + * data += stride + */ + /*else*/ { + __pyx_memoryview_refcount_objects_in_slice(__pyx_v_data, (__pyx_v_shape + 1), (__pyx_v_strides + 1), (__pyx_v_ndim - 1), __pyx_v_inc); + } + __pyx_L5:; + + /* "View.MemoryView":1385 + * refcount_objects_in_slice(data, shape + 1, strides + 1, ndim - 1, inc) + * + * data += stride # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_data = (__pyx_v_data + __pyx_v_stride); + } + + /* "View.MemoryView":1371 + * + * @cname('__pyx_memoryview_refcount_objects_in_slice') + * cdef void refcount_objects_in_slice(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, bint inc) noexcept: + * cdef Py_ssize_t i + */ + + /* function exit code */ +} + +/* "View.MemoryView":1391 + * + * @cname('__pyx_memoryview_slice_assign_scalar') + * cdef void slice_assign_scalar(__Pyx_memviewslice *dst, int ndim, # <<<<<<<<<<<<<< + * size_t itemsize, void *item, + * bint dtype_is_object) noexcept nogil: + */ + +static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *__pyx_v_dst, int __pyx_v_ndim, size_t __pyx_v_itemsize, void *__pyx_v_item, int __pyx_v_dtype_is_object) { + + /* "View.MemoryView":1394 + * size_t itemsize, void *item, + * bint dtype_is_object) noexcept nogil: + * refcount_copying(dst, dtype_is_object, ndim, inc=False) # <<<<<<<<<<<<<< + * _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim, itemsize, item) + * refcount_copying(dst, dtype_is_object, ndim, inc=True) + */ + __pyx_memoryview_refcount_copying(__pyx_v_dst, __pyx_v_dtype_is_object, __pyx_v_ndim, 0); + + /* "View.MemoryView":1395 + * bint dtype_is_object) noexcept nogil: + * refcount_copying(dst, dtype_is_object, ndim, inc=False) + * _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim, itemsize, item) # <<<<<<<<<<<<<< + * refcount_copying(dst, dtype_is_object, ndim, inc=True) + * + */ + __pyx_memoryview__slice_assign_scalar(__pyx_v_dst->data, __pyx_v_dst->shape, __pyx_v_dst->strides, __pyx_v_ndim, __pyx_v_itemsize, __pyx_v_item); + + /* "View.MemoryView":1396 + * refcount_copying(dst, dtype_is_object, ndim, inc=False) + * _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim, itemsize, item) + * refcount_copying(dst, dtype_is_object, ndim, inc=True) # <<<<<<<<<<<<<< + * + * + */ + __pyx_memoryview_refcount_copying(__pyx_v_dst, __pyx_v_dtype_is_object, __pyx_v_ndim, 1); + + /* "View.MemoryView":1391 + * + * @cname('__pyx_memoryview_slice_assign_scalar') + * cdef void slice_assign_scalar(__Pyx_memviewslice *dst, int ndim, # <<<<<<<<<<<<<< + * size_t itemsize, void *item, + * bint dtype_is_object) noexcept nogil: + */ + + /* function exit code */ +} + +/* "View.MemoryView":1400 + * + * @cname('__pyx_memoryview__slice_assign_scalar') + * cdef void _slice_assign_scalar(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * size_t itemsize, void *item) noexcept nogil: + */ + +static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t *__pyx_v_shape, Py_ssize_t *__pyx_v_strides, int __pyx_v_ndim, size_t __pyx_v_itemsize, void *__pyx_v_item) { + CYTHON_UNUSED Py_ssize_t __pyx_v_i; + Py_ssize_t __pyx_v_stride; + Py_ssize_t __pyx_v_extent; + int __pyx_t_1; + Py_ssize_t __pyx_t_2; + Py_ssize_t __pyx_t_3; + Py_ssize_t __pyx_t_4; + + /* "View.MemoryView":1404 + * size_t itemsize, void *item) noexcept nogil: + * cdef Py_ssize_t i + * cdef Py_ssize_t stride = strides[0] # <<<<<<<<<<<<<< + * cdef Py_ssize_t extent = shape[0] + * + */ + __pyx_v_stride = (__pyx_v_strides[0]); + + /* "View.MemoryView":1405 + * cdef Py_ssize_t i + * cdef Py_ssize_t stride = strides[0] + * cdef Py_ssize_t extent = shape[0] # <<<<<<<<<<<<<< + * + * if ndim == 1: + */ + __pyx_v_extent = (__pyx_v_shape[0]); + + /* "View.MemoryView":1407 + * cdef Py_ssize_t extent = shape[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * for i in range(extent): + * memcpy(data, item, itemsize) + */ + __pyx_t_1 = (__pyx_v_ndim == 1); + if (__pyx_t_1) { + + /* "View.MemoryView":1408 + * + * if ndim == 1: + * for i in range(extent): # <<<<<<<<<<<<<< + * memcpy(data, item, itemsize) + * data += stride + */ + __pyx_t_2 = __pyx_v_extent; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1409 + * if ndim == 1: + * for i in range(extent): + * memcpy(data, item, itemsize) # <<<<<<<<<<<<<< + * data += stride + * else: + */ + (void)(memcpy(__pyx_v_data, __pyx_v_item, __pyx_v_itemsize)); + + /* "View.MemoryView":1410 + * for i in range(extent): + * memcpy(data, item, itemsize) + * data += stride # <<<<<<<<<<<<<< + * else: + * for i in range(extent): + */ + __pyx_v_data = (__pyx_v_data + __pyx_v_stride); + } + + /* "View.MemoryView":1407 + * cdef Py_ssize_t extent = shape[0] + * + * if ndim == 1: # <<<<<<<<<<<<<< + * for i in range(extent): + * memcpy(data, item, itemsize) + */ + goto __pyx_L3; + } + + /* "View.MemoryView":1412 + * data += stride + * else: + * for i in range(extent): # <<<<<<<<<<<<<< + * _slice_assign_scalar(data, shape + 1, strides + 1, ndim - 1, itemsize, item) + * data += stride + */ + /*else*/ { + __pyx_t_2 = __pyx_v_extent; + __pyx_t_3 = __pyx_t_2; + for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) { + __pyx_v_i = __pyx_t_4; + + /* "View.MemoryView":1413 + * else: + * for i in range(extent): + * _slice_assign_scalar(data, shape + 1, strides + 1, ndim - 1, itemsize, item) # <<<<<<<<<<<<<< + * data += stride + * + */ + __pyx_memoryview__slice_assign_scalar(__pyx_v_data, (__pyx_v_shape + 1), (__pyx_v_strides + 1), (__pyx_v_ndim - 1), __pyx_v_itemsize, __pyx_v_item); + + /* "View.MemoryView":1414 + * for i in range(extent): + * _slice_assign_scalar(data, shape + 1, strides + 1, ndim - 1, itemsize, item) + * data += stride # <<<<<<<<<<<<<< + * + * + */ + __pyx_v_data = (__pyx_v_data + __pyx_v_stride); + } + } + __pyx_L3:; + + /* "View.MemoryView":1400 + * + * @cname('__pyx_memoryview__slice_assign_scalar') + * cdef void _slice_assign_scalar(char *data, Py_ssize_t *shape, # <<<<<<<<<<<<<< + * Py_ssize_t *strides, int ndim, + * size_t itemsize, void *item) noexcept nogil: + */ + + /* function exit code */ +} + +/* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_15View_dot_MemoryView_1__pyx_unpickle_Enum(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_15View_dot_MemoryView_1__pyx_unpickle_Enum = {"__pyx_unpickle_Enum", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_15View_dot_MemoryView_1__pyx_unpickle_Enum, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_15View_dot_MemoryView_1__pyx_unpickle_Enum(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyObject *__pyx_v___pyx_type = 0; + long __pyx_v___pyx_checksum; + PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[3] = {0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__pyx_unpickle_Enum (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_type,&__pyx_n_s_pyx_checksum,&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_type)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_checksum)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_Enum", 1, 3, 3, 1); __PYX_ERR(1, 1, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_Enum", 1, 3, 3, 2); __PYX_ERR(1, 1, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__pyx_unpickle_Enum") < 0)) __PYX_ERR(1, 1, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 3)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + } + __pyx_v___pyx_type = values[0]; + __pyx_v___pyx_checksum = __Pyx_PyInt_As_long(values[1]); if (unlikely((__pyx_v___pyx_checksum == (long)-1) && PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + __pyx_v___pyx_state = values[2]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_Enum", 1, 3, 3, __pyx_nargs); __PYX_ERR(1, 1, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("View.MemoryView.__pyx_unpickle_Enum", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf_15View_dot_MemoryView___pyx_unpickle_Enum(__pyx_self, __pyx_v___pyx_type, __pyx_v___pyx_checksum, __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_15View_dot_MemoryView___pyx_unpickle_Enum(CYTHON_UNUSED PyObject *__pyx_self, PyObject *__pyx_v___pyx_type, long __pyx_v___pyx_checksum, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_v___pyx_PickleError = 0; + PyObject *__pyx_v___pyx_result = 0; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + unsigned int __pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__pyx_unpickle_Enum", 1); + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + */ + __pyx_t_1 = __Pyx_PyInt_From_long(__pyx_v___pyx_checksum); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = (__Pyx_PySequence_ContainsTF(__pyx_t_1, __pyx_tuple__8, Py_NE)); if (unlikely((__pyx_t_2 < 0))) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + if (__pyx_t_2) { + + /* "(tree fragment)":5 + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): + * from pickle import PickleError as __pyx_PickleError # <<<<<<<<<<<<<< + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) + */ + __pyx_t_1 = PyList_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_n_s_PickleError); + __Pyx_GIVEREF(__pyx_n_s_PickleError); + if (__Pyx_PyList_SET_ITEM(__pyx_t_1, 0, __pyx_n_s_PickleError)) __PYX_ERR(1, 5, __pyx_L1_error); + __pyx_t_3 = __Pyx_Import(__pyx_n_s_pickle, __pyx_t_1, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_ImportFrom(__pyx_t_3, __pyx_n_s_PickleError); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_t_1); + __pyx_v___pyx_PickleError = __pyx_t_1; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + + /* "(tree fragment)":6 + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum # <<<<<<<<<<<<<< + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: + */ + __pyx_t_3 = __Pyx_PyInt_From_long(__pyx_v___pyx_checksum); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 = __Pyx_PyString_Format(__pyx_kp_s_Incompatible_checksums_0x_x_vs_0, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_Raise(__pyx_v___pyx_PickleError, __pyx_t_1, 0, 0); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __PYX_ERR(1, 6, __pyx_L1_error) + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + */ + } + + /* "(tree fragment)":7 + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) # <<<<<<<<<<<<<< + * if __pyx_state is not None: + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + */ + __pyx_t_3 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_MemviewEnum_type), __pyx_n_s_new); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = NULL; + __pyx_t_5 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_3))) { + __pyx_t_4 = PyMethod_GET_SELF(__pyx_t_3); + if (likely(__pyx_t_4)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_3); + __Pyx_INCREF(__pyx_t_4); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_3, function); + __pyx_t_5 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_4, __pyx_v___pyx_type}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_3, __pyx_callargs+1-__pyx_t_5, 1+__pyx_t_5); + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + __pyx_v___pyx_result = __pyx_t_1; + __pyx_t_1 = 0; + + /* "(tree fragment)":8 + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + */ + __pyx_t_2 = (__pyx_v___pyx_state != Py_None); + if (__pyx_t_2) { + + /* "(tree fragment)":9 + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) # <<<<<<<<<<<<<< + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + */ + if (!(likely(PyTuple_CheckExact(__pyx_v___pyx_state))||((__pyx_v___pyx_state) == Py_None) || __Pyx_RaiseUnexpectedTypeError("tuple", __pyx_v___pyx_state))) __PYX_ERR(1, 9, __pyx_L1_error) + __pyx_t_1 = __pyx_unpickle_Enum__set_state(((struct __pyx_MemviewEnum_obj *)__pyx_v___pyx_result), ((PyObject*)__pyx_v___pyx_state)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":8 + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + * __pyx_result = Enum.__new__(__pyx_type) + * if __pyx_state is not None: # <<<<<<<<<<<<<< + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + */ + } + + /* "(tree fragment)":10 + * if __pyx_state is not None: + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result # <<<<<<<<<<<<<< + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v___pyx_result); + __pyx_r = __pyx_v___pyx_result; + goto __pyx_L0; + + /* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("View.MemoryView.__pyx_unpickle_Enum", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v___pyx_PickleError); + __Pyx_XDECREF(__pyx_v___pyx_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":11 + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): # <<<<<<<<<<<<<< + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + */ + +static PyObject *__pyx_unpickle_Enum__set_state(struct __pyx_MemviewEnum_obj *__pyx_v___pyx_result, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + Py_ssize_t __pyx_t_3; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + unsigned int __pyx_t_8; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__pyx_unpickle_Enum__set_state", 1); + + /* "(tree fragment)":12 + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] # <<<<<<<<<<<<<< + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + * __pyx_result.__dict__.update(__pyx_state[1]) + */ + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 12, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_GIVEREF(__pyx_t_1); + __Pyx_GOTREF(__pyx_v___pyx_result->name); + __Pyx_DECREF(__pyx_v___pyx_result->name); + __pyx_v___pyx_result->name = __pyx_t_1; + __pyx_t_1 = 0; + + /* "(tree fragment)":13 + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): # <<<<<<<<<<<<<< + * __pyx_result.__dict__.update(__pyx_state[1]) + */ + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(1, 13, __pyx_L1_error) + } + __pyx_t_3 = __Pyx_PyTuple_GET_SIZE(__pyx_v___pyx_state); if (unlikely(__pyx_t_3 == ((Py_ssize_t)-1))) __PYX_ERR(1, 13, __pyx_L1_error) + __pyx_t_4 = (__pyx_t_3 > 1); + if (__pyx_t_4) { + } else { + __pyx_t_2 = __pyx_t_4; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_4 = __Pyx_HasAttr(((PyObject *)__pyx_v___pyx_result), __pyx_n_s_dict); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 13, __pyx_L1_error) + __pyx_t_2 = __pyx_t_4; + __pyx_L4_bool_binop_done:; + if (__pyx_t_2) { + + /* "(tree fragment)":14 + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + * __pyx_result.__dict__.update(__pyx_state[1]) # <<<<<<<<<<<<<< + */ + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v___pyx_result), __pyx_n_s_dict); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_6 = __Pyx_PyObject_GetAttrStr(__pyx_t_5, __pyx_n_s_update); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 14, __pyx_L1_error) + } + __pyx_t_5 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 1, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_7 = NULL; + __pyx_t_8 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_6))) { + __pyx_t_7 = PyMethod_GET_SELF(__pyx_t_6); + if (likely(__pyx_t_7)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_6); + __Pyx_INCREF(__pyx_t_7); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_6, function); + __pyx_t_8 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_7, __pyx_t_5}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_6, __pyx_callargs+1-__pyx_t_8, 1+__pyx_t_8); + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + } + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":13 + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): # <<<<<<<<<<<<<< + * __pyx_result.__dict__.update(__pyx_state[1]) + */ + } + + /* "(tree fragment)":11 + * __pyx_unpickle_Enum__set_state( __pyx_result, __pyx_state) + * return __pyx_result + * cdef __pyx_unpickle_Enum__set_state(Enum __pyx_result, tuple __pyx_state): # <<<<<<<<<<<<<< + * __pyx_result.name = __pyx_state[0] + * if len(__pyx_state) > 1 and hasattr(__pyx_result, '__dict__'): + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_AddTraceback("View.MemoryView.__pyx_unpickle_Enum__set_state", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":287 + * + * @property + * cdef inline npy_intp itemsize(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ELSIZE(self) + * + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_8itemsize_itemsize(PyArray_Descr *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":288 + * @property + * cdef inline npy_intp itemsize(self) noexcept nogil: + * return PyDataType_ELSIZE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyDataType_ELSIZE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":287 + * + * @property + * cdef inline npy_intp itemsize(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ELSIZE(self) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":291 + * + * @property + * cdef inline npy_intp alignment(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ALIGNMENT(self) + * + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_5dtype_9alignment_alignment(PyArray_Descr *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":292 + * @property + * cdef inline npy_intp alignment(self) noexcept nogil: + * return PyDataType_ALIGNMENT(self) # <<<<<<<<<<<<<< + * + * # Use fields/names with care as they may be NULL. You must check + */ + __pyx_r = PyDataType_ALIGNMENT(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":291 + * + * @property + * cdef inline npy_intp alignment(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_ALIGNMENT(self) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":297 + * # for this using PyDataType_HASFIELDS. + * @property + * cdef inline object fields(self): # <<<<<<<<<<<<<< + * return PyDataType_FIELDS(self) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_6fields_fields(PyArray_Descr *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1; + __Pyx_RefNannySetupContext("fields", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":298 + * @property + * cdef inline object fields(self): + * return PyDataType_FIELDS(self) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyDataType_FIELDS(__pyx_v_self); + __Pyx_INCREF(((PyObject *)__pyx_t_1)); + __pyx_r = ((PyObject *)__pyx_t_1); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":297 + * # for this using PyDataType_HASFIELDS. + * @property + * cdef inline object fields(self): # <<<<<<<<<<<<<< + * return PyDataType_FIELDS(self) + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":301 + * + * @property + * cdef inline tuple names(self): # <<<<<<<<<<<<<< + * return PyDataType_NAMES(self) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_5dtype_5names_names(PyArray_Descr *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1; + __Pyx_RefNannySetupContext("names", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":302 + * @property + * cdef inline tuple names(self): + * return PyDataType_NAMES(self) # <<<<<<<<<<<<<< + * + * # Use PyDataType_HASSUBARRAY to test whether this field is + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyDataType_NAMES(__pyx_v_self); + __Pyx_INCREF(((PyObject*)__pyx_t_1)); + __pyx_r = ((PyObject*)__pyx_t_1); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":301 + * + * @property + * cdef inline tuple names(self): # <<<<<<<<<<<<<< + * return PyDataType_NAMES(self) + * + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":308 + * # this field via the inline helper method PyDataType_SHAPE. + * @property + * cdef inline PyArray_ArrayDescr* subarray(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_SUBARRAY(self) + * + */ + +static CYTHON_INLINE PyArray_ArrayDescr *__pyx_f_5numpy_5dtype_8subarray_subarray(PyArray_Descr *__pyx_v_self) { + PyArray_ArrayDescr *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":309 + * @property + * cdef inline PyArray_ArrayDescr* subarray(self) noexcept nogil: + * return PyDataType_SUBARRAY(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyDataType_SUBARRAY(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":308 + * # this field via the inline helper method PyDataType_SHAPE. + * @property + * cdef inline PyArray_ArrayDescr* subarray(self) noexcept nogil: # <<<<<<<<<<<<<< + * return PyDataType_SUBARRAY(self) + * + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":312 + * + * @property + * cdef inline npy_uint64 flags(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The data types flags.""" + * return PyDataType_FLAGS(self) + */ + +static CYTHON_INLINE npy_uint64 __pyx_f_5numpy_5dtype_5flags_flags(PyArray_Descr *__pyx_v_self) { + npy_uint64 __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":314 + * cdef inline npy_uint64 flags(self) noexcept nogil: + * """The data types flags.""" + * return PyDataType_FLAGS(self) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyDataType_FLAGS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":312 + * + * @property + * cdef inline npy_uint64 flags(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The data types flags.""" + * return PyDataType_FLAGS(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":324 + * + * @property + * cdef inline int numiter(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of arrays that need to be broadcast to the same shape.""" + * return PyArray_MultiIter_NUMITER(self) + */ + +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_7numiter_numiter(PyArrayMultiIterObject *__pyx_v_self) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":326 + * cdef inline int numiter(self) noexcept nogil: + * """The number of arrays that need to be broadcast to the same shape.""" + * return PyArray_MultiIter_NUMITER(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_NUMITER(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":324 + * + * @property + * cdef inline int numiter(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of arrays that need to be broadcast to the same shape.""" + * return PyArray_MultiIter_NUMITER(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":329 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The total broadcasted size.""" + * return PyArray_MultiIter_SIZE(self) + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_4size_size(PyArrayMultiIterObject *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":331 + * cdef inline npy_intp size(self) noexcept nogil: + * """The total broadcasted size.""" + * return PyArray_MultiIter_SIZE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_SIZE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":329 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The total broadcasted size.""" + * return PyArray_MultiIter_SIZE(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":334 + * + * @property + * cdef inline npy_intp index(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The current (1-d) index into the broadcasted result.""" + * return PyArray_MultiIter_INDEX(self) + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_9broadcast_5index_index(PyArrayMultiIterObject *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":336 + * cdef inline npy_intp index(self) noexcept nogil: + * """The current (1-d) index into the broadcasted result.""" + * return PyArray_MultiIter_INDEX(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_INDEX(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":334 + * + * @property + * cdef inline npy_intp index(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The current (1-d) index into the broadcasted result.""" + * return PyArray_MultiIter_INDEX(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":339 + * + * @property + * cdef inline int nd(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of dimensions in the broadcasted result.""" + * return PyArray_MultiIter_NDIM(self) + */ + +static CYTHON_INLINE int __pyx_f_5numpy_9broadcast_2nd_nd(PyArrayMultiIterObject *__pyx_v_self) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":341 + * cdef inline int nd(self) noexcept nogil: + * """The number of dimensions in the broadcasted result.""" + * return PyArray_MultiIter_NDIM(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_NDIM(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":339 + * + * @property + * cdef inline int nd(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The number of dimensions in the broadcasted result.""" + * return PyArray_MultiIter_NDIM(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":344 + * + * @property + * cdef inline npy_intp* dimensions(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The shape of the broadcasted result.""" + * return PyArray_MultiIter_DIMS(self) + */ + +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_9broadcast_10dimensions_dimensions(PyArrayMultiIterObject *__pyx_v_self) { + npy_intp *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":346 + * cdef inline npy_intp* dimensions(self) noexcept nogil: + * """The shape of the broadcasted result.""" + * return PyArray_MultiIter_DIMS(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_MultiIter_DIMS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":344 + * + * @property + * cdef inline npy_intp* dimensions(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The shape of the broadcasted result.""" + * return PyArray_MultiIter_DIMS(self) + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":349 + * + * @property + * cdef inline void** iters(self) noexcept nogil: # <<<<<<<<<<<<<< + * """An array of iterator objects that holds the iterators for the arrays to be broadcast together. + * On return, the iterators are adjusted for broadcasting.""" + */ + +static CYTHON_INLINE void **__pyx_f_5numpy_9broadcast_5iters_iters(PyArrayMultiIterObject *__pyx_v_self) { + void **__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":352 + * """An array of iterator objects that holds the iterators for the arrays to be broadcast together. + * On return, the iterators are adjusted for broadcasting.""" + * return PyArray_MultiIter_ITERS(self) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyArray_MultiIter_ITERS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":349 + * + * @property + * cdef inline void** iters(self) noexcept nogil: # <<<<<<<<<<<<<< + * """An array of iterator objects that holds the iterators for the arrays to be broadcast together. + * On return, the iterators are adjusted for broadcasting.""" + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":367 + * + * @property + * cdef inline PyObject* base(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a borrowed reference to the object owning the data/memory. + * """ + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_7ndarray_4base_base(PyArrayObject *__pyx_v_self) { + PyObject *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":370 + * """Returns a borrowed reference to the object owning the data/memory. + * """ + * return PyArray_BASE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_BASE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":367 + * + * @property + * cdef inline PyObject* base(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a borrowed reference to the object owning the data/memory. + * """ + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":373 + * + * @property + * cdef inline dtype descr(self): # <<<<<<<<<<<<<< + * """Returns an owned reference to the dtype of the array. + * """ + */ + +static CYTHON_INLINE PyArray_Descr *__pyx_f_5numpy_7ndarray_5descr_descr(PyArrayObject *__pyx_v_self) { + PyArray_Descr *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyArray_Descr *__pyx_t_1; + __Pyx_RefNannySetupContext("descr", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":376 + * """Returns an owned reference to the dtype of the array. + * """ + * return PyArray_DESCR(self) # <<<<<<<<<<<<<< + * + * @property + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __pyx_t_1 = PyArray_DESCR(__pyx_v_self); + __Pyx_INCREF((PyObject *)((PyArray_Descr *)__pyx_t_1)); + __pyx_r = ((PyArray_Descr *)__pyx_t_1); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":373 + * + * @property + * cdef inline dtype descr(self): # <<<<<<<<<<<<<< + * """Returns an owned reference to the dtype of the array. + * """ + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":379 + * + * @property + * cdef inline int ndim(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the number of dimensions in the array. + * """ + */ + +static CYTHON_INLINE int __pyx_f_5numpy_7ndarray_4ndim_ndim(PyArrayObject *__pyx_v_self) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":382 + * """Returns the number of dimensions in the array. + * """ + * return PyArray_NDIM(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_NDIM(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":379 + * + * @property + * cdef inline int ndim(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the number of dimensions in the array. + * """ + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":385 + * + * @property + * cdef inline npy_intp *shape(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the dimensions/shape of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_5shape_shape(PyArrayObject *__pyx_v_self) { + npy_intp *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":390 + * Can return NULL for 0-dimensional arrays. + * """ + * return PyArray_DIMS(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_DIMS(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":385 + * + * @property + * cdef inline npy_intp *shape(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the dimensions/shape of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":393 + * + * @property + * cdef inline npy_intp *strides(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the strides of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + +static CYTHON_INLINE npy_intp *__pyx_f_5numpy_7ndarray_7strides_strides(PyArrayObject *__pyx_v_self) { + npy_intp *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":397 + * The number of elements matches the number of dimensions of the array (ndim). + * """ + * return PyArray_STRIDES(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_STRIDES(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":393 + * + * @property + * cdef inline npy_intp *strides(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns a pointer to the strides of the array. + * The number of elements matches the number of dimensions of the array (ndim). + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":400 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the total size (in number of elements) of the array. + * """ + */ + +static CYTHON_INLINE npy_intp __pyx_f_5numpy_7ndarray_4size_size(PyArrayObject *__pyx_v_self) { + npy_intp __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":403 + * """Returns the total size (in number of elements) of the array. + * """ + * return PyArray_SIZE(self) # <<<<<<<<<<<<<< + * + * @property + */ + __pyx_r = PyArray_SIZE(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":400 + * + * @property + * cdef inline npy_intp size(self) noexcept nogil: # <<<<<<<<<<<<<< + * """Returns the total size (in number of elements) of the array. + * """ + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":406 + * + * @property + * cdef inline char* data(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The pointer to the data buffer as a char*. + * This is provided for legacy reasons to avoid direct struct field access. + */ + +static CYTHON_INLINE char *__pyx_f_5numpy_7ndarray_4data_data(PyArrayObject *__pyx_v_self) { + char *__pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":412 + * of `PyArray_DATA()` instead, which returns a 'void*'. + * """ + * return PyArray_BYTES(self) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyArray_BYTES(__pyx_v_self); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":406 + * + * @property + * cdef inline char* data(self) noexcept nogil: # <<<<<<<<<<<<<< + * """The pointer to the data buffer as a char*. + * This is provided for legacy reasons to avoid direct struct field access. + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":824 + * ctypedef long double complex clongdouble_t + * + * cdef inline object PyArray_MultiIterNew1(a): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(1, a) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew1(PyObject *__pyx_v_a) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew1", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":825 + * + * cdef inline object PyArray_MultiIterNew1(a): + * return PyArray_MultiIterNew(1, a) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew2(a, b): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(1, ((void *)__pyx_v_a)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 825, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":824 + * ctypedef long double complex clongdouble_t + * + * cdef inline object PyArray_MultiIterNew1(a): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(1, a) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew1", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":827 + * return PyArray_MultiIterNew(1, a) + * + * cdef inline object PyArray_MultiIterNew2(a, b): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(2, a, b) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew2(PyObject *__pyx_v_a, PyObject *__pyx_v_b) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew2", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":828 + * + * cdef inline object PyArray_MultiIterNew2(a, b): + * return PyArray_MultiIterNew(2, a, b) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(2, ((void *)__pyx_v_a), ((void *)__pyx_v_b)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 828, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":827 + * return PyArray_MultiIterNew(1, a) + * + * cdef inline object PyArray_MultiIterNew2(a, b): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(2, a, b) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew2", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":830 + * return PyArray_MultiIterNew(2, a, b) + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(3, a, b, c) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew3(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew3", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":831 + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): + * return PyArray_MultiIterNew(3, a, b, c) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(3, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 831, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":830 + * return PyArray_MultiIterNew(2, a, b) + * + * cdef inline object PyArray_MultiIterNew3(a, b, c): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(3, a, b, c) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew3", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":833 + * return PyArray_MultiIterNew(3, a, b, c) + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(4, a, b, c, d) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew4(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew4", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":834 + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): + * return PyArray_MultiIterNew(4, a, b, c, d) # <<<<<<<<<<<<<< + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(4, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 834, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":833 + * return PyArray_MultiIterNew(3, a, b, c) + * + * cdef inline object PyArray_MultiIterNew4(a, b, c, d): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(4, a, b, c, d) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew4", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":836 + * return PyArray_MultiIterNew(4, a, b, c, d) + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyArray_MultiIterNew5(PyObject *__pyx_v_a, PyObject *__pyx_v_b, PyObject *__pyx_v_c, PyObject *__pyx_v_d, PyObject *__pyx_v_e) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("PyArray_MultiIterNew5", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":837 + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): + * return PyArray_MultiIterNew(5, a, b, c, d, e) # <<<<<<<<<<<<<< + * + * cdef inline tuple PyDataType_SHAPE(dtype d): + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = PyArray_MultiIterNew(5, ((void *)__pyx_v_a), ((void *)__pyx_v_b), ((void *)__pyx_v_c), ((void *)__pyx_v_d), ((void *)__pyx_v_e)); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 837, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":836 + * return PyArray_MultiIterNew(4, a, b, c, d) + * + * cdef inline object PyArray_MultiIterNew5(a, b, c, d, e): # <<<<<<<<<<<<<< + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("numpy.PyArray_MultiIterNew5", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":839 + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + * cdef inline tuple PyDataType_SHAPE(dtype d): # <<<<<<<<<<<<<< + * if PyDataType_HASSUBARRAY(d): + * return d.subarray.shape + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_PyDataType_SHAPE(PyArray_Descr *__pyx_v_d) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2; + __Pyx_RefNannySetupContext("PyDataType_SHAPE", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":840 + * + * cdef inline tuple PyDataType_SHAPE(dtype d): + * if PyDataType_HASSUBARRAY(d): # <<<<<<<<<<<<<< + * return d.subarray.shape + * else: + */ + __pyx_t_1 = PyDataType_HASSUBARRAY(__pyx_v_d); + if (__pyx_t_1) { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":841 + * cdef inline tuple PyDataType_SHAPE(dtype d): + * if PyDataType_HASSUBARRAY(d): + * return d.subarray.shape # <<<<<<<<<<<<<< + * else: + * return () + */ + __Pyx_XDECREF(__pyx_r); + __pyx_t_2 = __pyx_f_5numpy_5dtype_8subarray_subarray(__pyx_v_d)->shape; + __Pyx_INCREF(((PyObject*)__pyx_t_2)); + __pyx_r = ((PyObject*)__pyx_t_2); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":840 + * + * cdef inline tuple PyDataType_SHAPE(dtype d): + * if PyDataType_HASSUBARRAY(d): # <<<<<<<<<<<<<< + * return d.subarray.shape + * else: + */ + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":843 + * return d.subarray.shape + * else: + * return () # <<<<<<<<<<<<<< + * + * + */ + /*else*/ { + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_empty_tuple); + __pyx_r = __pyx_empty_tuple; + goto __pyx_L0; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":839 + * return PyArray_MultiIterNew(5, a, b, c, d, e) + * + * cdef inline tuple PyDataType_SHAPE(dtype d): # <<<<<<<<<<<<<< + * if PyDataType_HASSUBARRAY(d): + * return d.subarray.shape + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1027 + * int _import_umath() except -1 + * + * cdef inline void set_array_base(ndarray arr, object base) except *: # <<<<<<<<<<<<<< + * Py_INCREF(base) # important to do this before stealing the reference below! + * PyArray_SetBaseObject(arr, base) + */ + +static CYTHON_INLINE void __pyx_f_5numpy_set_array_base(PyArrayObject *__pyx_v_arr, PyObject *__pyx_v_base) { + int __pyx_t_1; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1028 + * + * cdef inline void set_array_base(ndarray arr, object base) except *: + * Py_INCREF(base) # important to do this before stealing the reference below! # <<<<<<<<<<<<<< + * PyArray_SetBaseObject(arr, base) + * + */ + Py_INCREF(__pyx_v_base); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1029 + * cdef inline void set_array_base(ndarray arr, object base) except *: + * Py_INCREF(base) # important to do this before stealing the reference below! + * PyArray_SetBaseObject(arr, base) # <<<<<<<<<<<<<< + * + * cdef inline object get_array_base(ndarray arr): + */ + __pyx_t_1 = PyArray_SetBaseObject(__pyx_v_arr, __pyx_v_base); if (unlikely(__pyx_t_1 == ((int)-1))) __PYX_ERR(2, 1029, __pyx_L1_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1027 + * int _import_umath() except -1 + * + * cdef inline void set_array_base(ndarray arr, object base) except *: # <<<<<<<<<<<<<< + * Py_INCREF(base) # important to do this before stealing the reference below! + * PyArray_SetBaseObject(arr, base) + */ + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("numpy.set_array_base", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_L0:; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1031 + * PyArray_SetBaseObject(arr, base) + * + * cdef inline object get_array_base(ndarray arr): # <<<<<<<<<<<<<< + * base = PyArray_BASE(arr) + * if base is NULL: + */ + +static CYTHON_INLINE PyObject *__pyx_f_5numpy_get_array_base(PyArrayObject *__pyx_v_arr) { + PyObject *__pyx_v_base; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + __Pyx_RefNannySetupContext("get_array_base", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1032 + * + * cdef inline object get_array_base(ndarray arr): + * base = PyArray_BASE(arr) # <<<<<<<<<<<<<< + * if base is NULL: + * return None + */ + __pyx_v_base = PyArray_BASE(__pyx_v_arr); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1033 + * cdef inline object get_array_base(ndarray arr): + * base = PyArray_BASE(arr) + * if base is NULL: # <<<<<<<<<<<<<< + * return None + * return base + */ + __pyx_t_1 = (__pyx_v_base == NULL); + if (__pyx_t_1) { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1034 + * base = PyArray_BASE(arr) + * if base is NULL: + * return None # <<<<<<<<<<<<<< + * return base + * + */ + __Pyx_XDECREF(__pyx_r); + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1033 + * cdef inline object get_array_base(ndarray arr): + * base = PyArray_BASE(arr) + * if base is NULL: # <<<<<<<<<<<<<< + * return None + * return base + */ + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1035 + * if base is NULL: + * return None + * return base # <<<<<<<<<<<<<< + * + * # Versions of the import_* functions which are more suitable for + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(((PyObject *)__pyx_v_base)); + __pyx_r = ((PyObject *)__pyx_v_base); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1031 + * PyArray_SetBaseObject(arr, base) + * + * cdef inline object get_array_base(ndarray arr): # <<<<<<<<<<<<<< + * base = PyArray_BASE(arr) + * if base is NULL: + */ + + /* function exit code */ + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1039 + * # Versions of the import_* functions which are more suitable for + * # Cython code. + * cdef inline int import_array() except -1: # <<<<<<<<<<<<<< + * try: + * __pyx_import_array() + */ + +static CYTHON_INLINE int __pyx_f_5numpy_import_array(void) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("import_array", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1040 + * # Cython code. + * cdef inline int import_array() except -1: + * try: # <<<<<<<<<<<<<< + * __pyx_import_array() + * except Exception: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1041 + * cdef inline int import_array() except -1: + * try: + * __pyx_import_array() # <<<<<<<<<<<<<< + * except Exception: + * raise ImportError("numpy._core.multiarray failed to import") + */ + __pyx_t_4 = _import_array(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(2, 1041, __pyx_L3_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1040 + * # Cython code. + * cdef inline int import_array() except -1: + * try: # <<<<<<<<<<<<<< + * __pyx_import_array() + * except Exception: + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L8_try_end; + __pyx_L3_error:; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1042 + * try: + * __pyx_import_array() + * except Exception: # <<<<<<<<<<<<<< + * raise ImportError("numpy._core.multiarray failed to import") + * + */ + __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0]))); + if (__pyx_t_4) { + __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(2, 1042, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_7); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1043 + * __pyx_import_array() + * except Exception: + * raise ImportError("numpy._core.multiarray failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_umath() except -1: + */ + __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__9, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(2, 1043, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_Raise(__pyx_t_8, 0, 0, 0); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __PYX_ERR(2, 1043, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1040 + * # Cython code. + * cdef inline int import_array() except -1: + * try: # <<<<<<<<<<<<<< + * __pyx_import_array() + * except Exception: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L8_try_end:; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1039 + * # Versions of the import_* functions which are more suitable for + * # Cython code. + * cdef inline int import_array() except -1: # <<<<<<<<<<<<<< + * try: + * __pyx_import_array() + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("numpy.import_array", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1045 + * raise ImportError("numpy._core.multiarray failed to import") + * + * cdef inline int import_umath() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + +static CYTHON_INLINE int __pyx_f_5numpy_import_umath(void) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("import_umath", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1046 + * + * cdef inline int import_umath() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1047 + * cdef inline int import_umath() except -1: + * try: + * _import_umath() # <<<<<<<<<<<<<< + * except Exception: + * raise ImportError("numpy._core.umath failed to import") + */ + __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(2, 1047, __pyx_L3_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1046 + * + * cdef inline int import_umath() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L8_try_end; + __pyx_L3_error:; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1048 + * try: + * _import_umath() + * except Exception: # <<<<<<<<<<<<<< + * raise ImportError("numpy._core.umath failed to import") + * + */ + __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0]))); + if (__pyx_t_4) { + __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(2, 1048, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_7); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1049 + * _import_umath() + * except Exception: + * raise ImportError("numpy._core.umath failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_ufunc() except -1: + */ + __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__10, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(2, 1049, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_Raise(__pyx_t_8, 0, 0, 0); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __PYX_ERR(2, 1049, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1046 + * + * cdef inline int import_umath() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L8_try_end:; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1045 + * raise ImportError("numpy._core.multiarray failed to import") + * + * cdef inline int import_umath() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("numpy.import_umath", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1051 + * raise ImportError("numpy._core.umath failed to import") + * + * cdef inline int import_ufunc() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + +static CYTHON_INLINE int __pyx_f_5numpy_import_ufunc(void) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + int __pyx_t_4; + PyObject *__pyx_t_5 = NULL; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("import_ufunc", 1); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1052 + * + * cdef inline int import_ufunc() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1053 + * cdef inline int import_ufunc() except -1: + * try: + * _import_umath() # <<<<<<<<<<<<<< + * except Exception: + * raise ImportError("numpy._core.umath failed to import") + */ + __pyx_t_4 = _import_umath(); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(2, 1053, __pyx_L3_error) + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1052 + * + * cdef inline int import_ufunc() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L8_try_end; + __pyx_L3_error:; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1054 + * try: + * _import_umath() + * except Exception: # <<<<<<<<<<<<<< + * raise ImportError("numpy._core.umath failed to import") + * + */ + __pyx_t_4 = __Pyx_PyErr_ExceptionMatches(((PyObject *)(&((PyTypeObject*)PyExc_Exception)[0]))); + if (__pyx_t_4) { + __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_6, &__pyx_t_7) < 0) __PYX_ERR(2, 1054, __pyx_L5_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_6); + __Pyx_XGOTREF(__pyx_t_7); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1055 + * _import_umath() + * except Exception: + * raise ImportError("numpy._core.umath failed to import") # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_8 = __Pyx_PyObject_Call(__pyx_builtin_ImportError, __pyx_tuple__10, NULL); if (unlikely(!__pyx_t_8)) __PYX_ERR(2, 1055, __pyx_L5_except_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_Raise(__pyx_t_8, 0, 0, 0); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + __PYX_ERR(2, 1055, __pyx_L5_except_error) + } + goto __pyx_L5_except_error; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1052 + * + * cdef inline int import_ufunc() except -1: + * try: # <<<<<<<<<<<<<< + * _import_umath() + * except Exception: + */ + __pyx_L5_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L8_try_end:; + } + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1051 + * raise ImportError("numpy._core.umath failed to import") + * + * cdef inline int import_ufunc() except -1: # <<<<<<<<<<<<<< + * try: + * _import_umath() + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_6); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_AddTraceback("numpy.import_ufunc", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1058 + * + * + * cdef inline bint is_timedelta64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.timedelta64)` + */ + +static CYTHON_INLINE int __pyx_f_5numpy_is_timedelta64_object(PyObject *__pyx_v_obj) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1070 + * bool + * """ + * return PyObject_TypeCheck(obj, &PyTimedeltaArrType_Type) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyObject_TypeCheck(__pyx_v_obj, (&PyTimedeltaArrType_Type)); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1058 + * + * + * cdef inline bint is_timedelta64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.timedelta64)` + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1073 + * + * + * cdef inline bint is_datetime64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.datetime64)` + */ + +static CYTHON_INLINE int __pyx_f_5numpy_is_datetime64_object(PyObject *__pyx_v_obj) { + int __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1085 + * bool + * """ + * return PyObject_TypeCheck(obj, &PyDatetimeArrType_Type) # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = PyObject_TypeCheck(__pyx_v_obj, (&PyDatetimeArrType_Type)); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1073 + * + * + * cdef inline bint is_datetime64_object(object obj) noexcept: # <<<<<<<<<<<<<< + * """ + * Cython equivalent of `isinstance(obj, np.datetime64)` + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1088 + * + * + * cdef inline npy_datetime get_datetime64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy datetime64 object + */ + +static CYTHON_INLINE npy_datetime __pyx_f_5numpy_get_datetime64_value(PyObject *__pyx_v_obj) { + npy_datetime __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1095 + * also needed. That can be found using `get_datetime64_unit`. + * """ + * return (obj).obval # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = ((PyDatetimeScalarObject *)__pyx_v_obj)->obval; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1088 + * + * + * cdef inline npy_datetime get_datetime64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy datetime64 object + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1098 + * + * + * cdef inline npy_timedelta get_timedelta64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy timedelta64 object + */ + +static CYTHON_INLINE npy_timedelta __pyx_f_5numpy_get_timedelta64_value(PyObject *__pyx_v_obj) { + npy_timedelta __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1102 + * returns the int64 value underlying scalar numpy timedelta64 object + * """ + * return (obj).obval # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = ((PyTimedeltaScalarObject *)__pyx_v_obj)->obval; + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1098 + * + * + * cdef inline npy_timedelta get_timedelta64_value(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the int64 value underlying scalar numpy timedelta64 object + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1105 + * + * + * cdef inline NPY_DATETIMEUNIT get_datetime64_unit(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the unit part of the dtype for a numpy datetime64 object. + */ + +static CYTHON_INLINE NPY_DATETIMEUNIT __pyx_f_5numpy_get_datetime64_unit(PyObject *__pyx_v_obj) { + NPY_DATETIMEUNIT __pyx_r; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1109 + * returns the unit part of the dtype for a numpy datetime64 object. + * """ + * return (obj).obmeta.base # <<<<<<<<<<<<<< + * + * + */ + __pyx_r = ((NPY_DATETIMEUNIT)((PyDatetimeScalarObject *)__pyx_v_obj)->obmeta.base); + goto __pyx_L0; + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1105 + * + * + * cdef inline NPY_DATETIMEUNIT get_datetime64_unit(object obj) noexcept nogil: # <<<<<<<<<<<<<< + * """ + * returns the unit part of the dtype for a numpy datetime64 object. + */ + + /* function exit code */ + __pyx_L0:; + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":22 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_none_mode(np.ndarray[DTYPE_t, ndim=1] sizes, int block_size): # <<<<<<<<<<<<<< + * cdef DTYPE_t total_size = sizes.sum() + * cdef DTYPE_t length = ceil(total_size / block_size) + */ + +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__get_slice_indices_none_mode(PyArrayObject *__pyx_v_sizes, int __pyx_v_block_size) { + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_total_size; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_length; + PyArrayObject *__pyx_v_slice_indices = 0; + __Pyx_memviewslice __pyx_v_slice_indices_view = { 0, 0, { 0 }, { 0 }, { 0 } }; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_i; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_start; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_end; + __Pyx_LocalBuf_ND __pyx_pybuffernd_sizes; + __Pyx_Buffer __pyx_pybuffer_sizes; + __Pyx_LocalBuf_ND __pyx_pybuffernd_slice_indices; + __Pyx_Buffer __pyx_pybuffer_slice_indices; + PyArrayObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + unsigned int __pyx_t_4; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_5; + PyObject *__pyx_t_6 = NULL; + PyArrayObject *__pyx_t_7 = NULL; + __Pyx_memviewslice __pyx_t_8 = { 0, 0, { 0 }, { 0 }, { 0 } }; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_9; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_10; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_11; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_12; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_13; + int __pyx_t_14; + Py_ssize_t __pyx_t_15; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_get_slice_indices_none_mode", 1); + __pyx_pybuffer_slice_indices.pybuffer.buf = NULL; + __pyx_pybuffer_slice_indices.refcount = 0; + __pyx_pybuffernd_slice_indices.data = NULL; + __pyx_pybuffernd_slice_indices.rcbuffer = &__pyx_pybuffer_slice_indices; + __pyx_pybuffer_sizes.pybuffer.buf = NULL; + __pyx_pybuffer_sizes.refcount = 0; + __pyx_pybuffernd_sizes.data = NULL; + __pyx_pybuffernd_sizes.rcbuffer = &__pyx_pybuffer_sizes; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer, (PyObject*)__pyx_v_sizes, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 22, __pyx_L1_error) + } + __pyx_pybuffernd_sizes.diminfo[0].strides = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_sizes.diminfo[0].shape = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.shape[0]; + + /* "fairseq/data/token_block_utils_fast.pyx":23 + * @cython.nonecheck(False) + * cdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_none_mode(np.ndarray[DTYPE_t, ndim=1] sizes, int block_size): + * cdef DTYPE_t total_size = sizes.sum() # <<<<<<<<<<<<<< + * cdef DTYPE_t length = ceil(total_size / block_size) + * cdef np.ndarray[DTYPE_t, ndim=2] slice_indices = np.zeros([length, 2], dtype=DTYPE) + */ + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_sizes), __pyx_n_s_sum); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 23, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = NULL; + __pyx_t_4 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_2))) { + __pyx_t_3 = PyMethod_GET_SELF(__pyx_t_2); + if (likely(__pyx_t_3)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_2); + __Pyx_INCREF(__pyx_t_3); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_2, function); + __pyx_t_4 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_3, NULL}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_2, __pyx_callargs+1-__pyx_t_4, 0+__pyx_t_4); + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 23, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + } + __pyx_t_5 = __Pyx_PyInt_As_npy_int64(__pyx_t_1); if (unlikely((__pyx_t_5 == ((npy_int64)-1)) && PyErr_Occurred())) __PYX_ERR(0, 23, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_v_total_size = __pyx_t_5; + + /* "fairseq/data/token_block_utils_fast.pyx":24 + * cdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_none_mode(np.ndarray[DTYPE_t, ndim=1] sizes, int block_size): + * cdef DTYPE_t total_size = sizes.sum() + * cdef DTYPE_t length = ceil(total_size / block_size) # <<<<<<<<<<<<<< + * cdef np.ndarray[DTYPE_t, ndim=2] slice_indices = np.zeros([length, 2], dtype=DTYPE) + * cdef DTYPE_t[:, :] slice_indices_view = slice_indices + */ + if (unlikely(((double)__pyx_v_block_size) == 0)) { + PyErr_SetString(PyExc_ZeroDivisionError, "float division"); + __PYX_ERR(0, 24, __pyx_L1_error) + } + __pyx_v_length = ((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t)ceil((((double)__pyx_v_total_size) / ((double)__pyx_v_block_size)))); + + /* "fairseq/data/token_block_utils_fast.pyx":25 + * cdef DTYPE_t total_size = sizes.sum() + * cdef DTYPE_t length = ceil(total_size / block_size) + * cdef np.ndarray[DTYPE_t, ndim=2] slice_indices = np.zeros([length, 2], dtype=DTYPE) # <<<<<<<<<<<<<< + * cdef DTYPE_t[:, :] slice_indices_view = slice_indices + * cdef DTYPE_t i + */ + __Pyx_GetModuleGlobalName(__pyx_t_1, __pyx_n_s_np); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_zeros); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_PyInt_From_npy_int64(__pyx_v_length); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = PyList_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyList_SET_ITEM(__pyx_t_3, 0, __pyx_t_1)) __PYX_ERR(0, 25, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_2); + __Pyx_GIVEREF(__pyx_int_2); + if (__Pyx_PyList_SET_ITEM(__pyx_t_3, 1, __pyx_int_2)) __PYX_ERR(0, 25, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_1 = PyTuple_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_3)) __PYX_ERR(0, 25, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_3 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GetModuleGlobalName(__pyx_t_6, __pyx_n_s_DTYPE); if (unlikely(!__pyx_t_6)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + if (PyDict_SetItem(__pyx_t_3, __pyx_n_s_dtype, __pyx_t_6) < 0) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0; + __pyx_t_6 = __Pyx_PyObject_Call(__pyx_t_2, __pyx_t_1, __pyx_t_3); if (unlikely(!__pyx_t_6)) __PYX_ERR(0, 25, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_6); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + if (!(likely(((__pyx_t_6) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_6, __pyx_ptype_5numpy_ndarray))))) __PYX_ERR(0, 25, __pyx_L1_error) + __pyx_t_7 = ((PyArrayObject *)__pyx_t_6); + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_t_7, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) { + __pyx_v_slice_indices = ((PyArrayObject *)Py_None); __Pyx_INCREF(Py_None); __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.buf = NULL; + __PYX_ERR(0, 25, __pyx_L1_error) + } else {__pyx_pybuffernd_slice_indices.diminfo[0].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_slice_indices.diminfo[0].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_slice_indices.diminfo[1].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_slice_indices.diminfo[1].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[1]; + } + } + __pyx_t_7 = 0; + __pyx_v_slice_indices = ((PyArrayObject *)__pyx_t_6); + __pyx_t_6 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":26 + * cdef DTYPE_t length = ceil(total_size / block_size) + * cdef np.ndarray[DTYPE_t, ndim=2] slice_indices = np.zeros([length, 2], dtype=DTYPE) + * cdef DTYPE_t[:, :] slice_indices_view = slice_indices # <<<<<<<<<<<<<< + * cdef DTYPE_t i + * cdef DTYPE_t start + */ + __pyx_t_8 = __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(((PyObject *)__pyx_v_slice_indices), PyBUF_WRITABLE); if (unlikely(!__pyx_t_8.memview)) __PYX_ERR(0, 26, __pyx_L1_error) + __pyx_v_slice_indices_view = __pyx_t_8; + __pyx_t_8.memview = NULL; + __pyx_t_8.data = NULL; + + /* "fairseq/data/token_block_utils_fast.pyx":30 + * cdef DTYPE_t start + * cdef DTYPE_t end + * for i in range(length): # <<<<<<<<<<<<<< + * start = i * block_size + * end = min(start + block_size, total_size) + */ + __pyx_t_5 = __pyx_v_length; + __pyx_t_9 = __pyx_t_5; + for (__pyx_t_10 = 0; __pyx_t_10 < __pyx_t_9; __pyx_t_10+=1) { + __pyx_v_i = __pyx_t_10; + + /* "fairseq/data/token_block_utils_fast.pyx":31 + * cdef DTYPE_t end + * for i in range(length): + * start = i * block_size # <<<<<<<<<<<<<< + * end = min(start + block_size, total_size) + * slice_indices_view[i][0] = start + */ + __pyx_v_start = (__pyx_v_i * __pyx_v_block_size); + + /* "fairseq/data/token_block_utils_fast.pyx":32 + * for i in range(length): + * start = i * block_size + * end = min(start + block_size, total_size) # <<<<<<<<<<<<<< + * slice_indices_view[i][0] = start + * slice_indices_view[i][1] = end + */ + __pyx_t_11 = __pyx_v_total_size; + __pyx_t_12 = (__pyx_v_start + __pyx_v_block_size); + __pyx_t_14 = (__pyx_t_11 < __pyx_t_12); + if (__pyx_t_14) { + __pyx_t_13 = __pyx_t_11; + } else { + __pyx_t_13 = __pyx_t_12; + } + __pyx_v_end = __pyx_t_13; + + /* "fairseq/data/token_block_utils_fast.pyx":33 + * start = i * block_size + * end = min(start + block_size, total_size) + * slice_indices_view[i][0] = start # <<<<<<<<<<<<<< + * slice_indices_view[i][1] = end + * return slice_indices + */ + __pyx_t_13 = __pyx_v_i; + __pyx_t_15 = 0; + *((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_slice_indices_view.data + __pyx_t_13 * __pyx_v_slice_indices_view.strides[0]) ) + __pyx_t_15 * __pyx_v_slice_indices_view.strides[1]) )) = __pyx_v_start; + + /* "fairseq/data/token_block_utils_fast.pyx":34 + * end = min(start + block_size, total_size) + * slice_indices_view[i][0] = start + * slice_indices_view[i][1] = end # <<<<<<<<<<<<<< + * return slice_indices + * + */ + __pyx_t_13 = __pyx_v_i; + __pyx_t_15 = 1; + *((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_slice_indices_view.data + __pyx_t_13 * __pyx_v_slice_indices_view.strides[0]) ) + __pyx_t_15 * __pyx_v_slice_indices_view.strides[1]) )) = __pyx_v_end; + } + + /* "fairseq/data/token_block_utils_fast.pyx":35 + * slice_indices_view[i][0] = start + * slice_indices_view[i][1] = end + * return slice_indices # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_slice_indices); + __pyx_r = ((PyArrayObject *)__pyx_v_slice_indices); + goto __pyx_L0; + + /* "fairseq/data/token_block_utils_fast.pyx":22 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_none_mode(np.ndarray[DTYPE_t, ndim=1] sizes, int block_size): # <<<<<<<<<<<<<< + * cdef DTYPE_t total_size = sizes.sum() + * cdef DTYPE_t length = ceil(total_size / block_size) + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_6); + __PYX_XCLEAR_MEMVIEW(&__pyx_t_8, 1); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._get_slice_indices_none_mode", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XDECREF((PyObject *)__pyx_v_slice_indices); + __PYX_XCLEAR_MEMVIEW(&__pyx_v_slice_indices_view, 1); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":38 + * + * + * cdef np.ndarray[DTYPE_t, ndim=2] _fast_convert_to_np_array(list list_of_list): # <<<<<<<<<<<<<< + * """ + * Faster function to convert DTYPE_t list of list. + */ + +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__fast_convert_to_np_array(PyObject *__pyx_v_list_of_list) { + PyArrayObject *__pyx_v_flat = 0; + __Pyx_LocalBuf_ND __pyx_pybuffernd_flat; + __Pyx_Buffer __pyx_pybuffer_flat; + PyArrayObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + unsigned int __pyx_t_6; + PyArrayObject *__pyx_t_7 = NULL; + Py_ssize_t __pyx_t_8; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_fast_convert_to_np_array", 1); + __pyx_pybuffer_flat.pybuffer.buf = NULL; + __pyx_pybuffer_flat.refcount = 0; + __pyx_pybuffernd_flat.data = NULL; + __pyx_pybuffernd_flat.rcbuffer = &__pyx_pybuffer_flat; + + /* "fairseq/data/token_block_utils_fast.pyx":43 + * Only fast when there are huge number of rows and low number of columns. + * """ + * cdef np.ndarray[DTYPE_t, ndim=1] flat = np.fromiter(chain.from_iterable(list_of_list), DTYPE, -1) # <<<<<<<<<<<<<< + * return flat.reshape((len(list_of_list), -1)) + * + */ + __Pyx_GetModuleGlobalName(__pyx_t_2, __pyx_n_s_np); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 43, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_fromiter); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 43, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_GetModuleGlobalName(__pyx_t_4, __pyx_n_s_chain); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 43, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_from_iterable); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 43, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_t_4 = NULL; + __pyx_t_6 = 0; + #if CYTHON_UNPACK_METHODS + if (unlikely(PyMethod_Check(__pyx_t_5))) { + __pyx_t_4 = PyMethod_GET_SELF(__pyx_t_5); + if (likely(__pyx_t_4)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_5); + __Pyx_INCREF(__pyx_t_4); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_5, function); + __pyx_t_6 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_4, __pyx_v_list_of_list}; + __pyx_t_2 = __Pyx_PyObject_FastCall(__pyx_t_5, __pyx_callargs+1-__pyx_t_6, 1+__pyx_t_6); + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 43, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + } + __Pyx_GetModuleGlobalName(__pyx_t_5, __pyx_n_s_DTYPE); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 43, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_4 = NULL; + __pyx_t_6 = 0; + #if CYTHON_UNPACK_METHODS + if (unlikely(PyMethod_Check(__pyx_t_3))) { + __pyx_t_4 = PyMethod_GET_SELF(__pyx_t_3); + if (likely(__pyx_t_4)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_3); + __Pyx_INCREF(__pyx_t_4); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_3, function); + __pyx_t_6 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[4] = {__pyx_t_4, __pyx_t_2, __pyx_t_5, __pyx_int_neg_1}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_3, __pyx_callargs+1-__pyx_t_6, 3+__pyx_t_6); + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 43, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + if (!(likely(((__pyx_t_1) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_1, __pyx_ptype_5numpy_ndarray))))) __PYX_ERR(0, 43, __pyx_L1_error) + __pyx_t_7 = ((PyArrayObject *)__pyx_t_1); + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_flat.rcbuffer->pybuffer, (PyObject*)__pyx_t_7, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) { + __pyx_v_flat = ((PyArrayObject *)Py_None); __Pyx_INCREF(Py_None); __pyx_pybuffernd_flat.rcbuffer->pybuffer.buf = NULL; + __PYX_ERR(0, 43, __pyx_L1_error) + } else {__pyx_pybuffernd_flat.diminfo[0].strides = __pyx_pybuffernd_flat.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_flat.diminfo[0].shape = __pyx_pybuffernd_flat.rcbuffer->pybuffer.shape[0]; + } + } + __pyx_t_7 = 0; + __pyx_v_flat = ((PyArrayObject *)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":44 + * """ + * cdef np.ndarray[DTYPE_t, ndim=1] flat = np.fromiter(chain.from_iterable(list_of_list), DTYPE, -1) + * return flat.reshape((len(list_of_list), -1)) # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __pyx_t_3 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_flat), __pyx_n_s_reshape); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 44, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + if (unlikely(__pyx_v_list_of_list == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(0, 44, __pyx_L1_error) + } + __pyx_t_8 = __Pyx_PyList_GET_SIZE(__pyx_v_list_of_list); if (unlikely(__pyx_t_8 == ((Py_ssize_t)-1))) __PYX_ERR(0, 44, __pyx_L1_error) + __pyx_t_5 = PyInt_FromSsize_t(__pyx_t_8); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 44, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_t_2 = PyTuple_New(2); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 44, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_5); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 0, __pyx_t_5)) __PYX_ERR(0, 44, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_neg_1); + __Pyx_GIVEREF(__pyx_int_neg_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 1, __pyx_int_neg_1)) __PYX_ERR(0, 44, __pyx_L1_error); + __pyx_t_5 = 0; + __pyx_t_5 = NULL; + __pyx_t_6 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_3))) { + __pyx_t_5 = PyMethod_GET_SELF(__pyx_t_3); + if (likely(__pyx_t_5)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_3); + __Pyx_INCREF(__pyx_t_5); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_3, function); + __pyx_t_6 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_5, __pyx_t_2}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_3, __pyx_callargs+1-__pyx_t_6, 1+__pyx_t_6); + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 44, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + if (!(likely(((__pyx_t_1) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_1, __pyx_ptype_5numpy_ndarray))))) __PYX_ERR(0, 44, __pyx_L1_error) + __pyx_r = ((PyArrayObject *)__pyx_t_1); + __pyx_t_1 = 0; + goto __pyx_L0; + + /* "fairseq/data/token_block_utils_fast.pyx":38 + * + * + * cdef np.ndarray[DTYPE_t, ndim=2] _fast_convert_to_np_array(list list_of_list): # <<<<<<<<<<<<<< + * """ + * Faster function to convert DTYPE_t list of list. + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_flat.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._fast_convert_to_np_array", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_flat.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XDECREF((PyObject *)__pyx_v_flat); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":50 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): # <<<<<<<<<<<<<< + * cdef DTYPE_t tok_idx = 0 + * cdef DTYPE_t sz_idx = 0 + */ + +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_1_get_slice_indices_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__get_slice_indices_fast(PyArrayObject *__pyx_v_sizes, PyObject *__pyx_v_break_mode, int __pyx_v_block_size, int __pyx_v_document_sep_len, CYTHON_UNUSED int __pyx_skip_dispatch) { + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_tok_idx; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_sz_idx; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_curr_size; + CYTHON_UNUSED __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_i; + __Pyx_memviewslice __pyx_v_sizes_view = { 0, 0, { 0 }, { 0 }, { 0 } }; + PyArrayObject *__pyx_v_slice_indices = 0; + PyObject *__pyx_v_slice_indices_list = 0; + PyObject *__pyx_v_cumsum = NULL; + __Pyx_LocalBuf_ND __pyx_pybuffernd_sizes; + __Pyx_Buffer __pyx_pybuffer_sizes; + __Pyx_LocalBuf_ND __pyx_pybuffernd_slice_indices; + __Pyx_Buffer __pyx_pybuffer_slice_indices; + PyArrayObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_memviewslice __pyx_t_1 = { 0, 0, { 0 }, { 0 }, { 0 } }; + PyObject *__pyx_t_2 = NULL; + int __pyx_t_3; + int __pyx_t_4; + int __pyx_t_5; + PyObject *__pyx_t_6 = NULL; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + Py_ssize_t __pyx_t_9; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_10; + PyObject *__pyx_t_11 = NULL; + PyObject *__pyx_t_12 = NULL; + int __pyx_t_13; + PyObject *__pyx_t_14 = NULL; + PyArrayObject *__pyx_t_15 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_get_slice_indices_fast", 1); + __pyx_pybuffer_slice_indices.pybuffer.buf = NULL; + __pyx_pybuffer_slice_indices.refcount = 0; + __pyx_pybuffernd_slice_indices.data = NULL; + __pyx_pybuffernd_slice_indices.rcbuffer = &__pyx_pybuffer_slice_indices; + __pyx_pybuffer_sizes.pybuffer.buf = NULL; + __pyx_pybuffer_sizes.refcount = 0; + __pyx_pybuffernd_sizes.data = NULL; + __pyx_pybuffernd_sizes.rcbuffer = &__pyx_pybuffer_sizes; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer, (PyObject*)__pyx_v_sizes, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 50, __pyx_L1_error) + } + __pyx_pybuffernd_sizes.diminfo[0].strides = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_sizes.diminfo[0].shape = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.shape[0]; + + /* "fairseq/data/token_block_utils_fast.pyx":51 + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): + * cdef DTYPE_t tok_idx = 0 # <<<<<<<<<<<<<< + * cdef DTYPE_t sz_idx = 0 + * cdef DTYPE_t curr_size = 0 + */ + __pyx_v_tok_idx = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":52 + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): + * cdef DTYPE_t tok_idx = 0 + * cdef DTYPE_t sz_idx = 0 # <<<<<<<<<<<<<< + * cdef DTYPE_t curr_size = 0 + * cdef DTYPE_t i = 0 + */ + __pyx_v_sz_idx = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":53 + * cdef DTYPE_t tok_idx = 0 + * cdef DTYPE_t sz_idx = 0 + * cdef DTYPE_t curr_size = 0 # <<<<<<<<<<<<<< + * cdef DTYPE_t i = 0 + * cdef DTYPE_t length + */ + __pyx_v_curr_size = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":54 + * cdef DTYPE_t sz_idx = 0 + * cdef DTYPE_t curr_size = 0 + * cdef DTYPE_t i = 0 # <<<<<<<<<<<<<< + * cdef DTYPE_t length + * cdef DTYPE_t total_size + */ + __pyx_v_i = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":57 + * cdef DTYPE_t length + * cdef DTYPE_t total_size + * cdef DTYPE_t[:] sizes_view = sizes # <<<<<<<<<<<<<< + * cdef np.ndarray[DTYPE_t, ndim=2] slice_indices + * cdef list slice_indices_list = [] + */ + __pyx_t_1 = __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(((PyObject *)__pyx_v_sizes), PyBUF_WRITABLE); if (unlikely(!__pyx_t_1.memview)) __PYX_ERR(0, 57, __pyx_L1_error) + __pyx_v_sizes_view = __pyx_t_1; + __pyx_t_1.memview = NULL; + __pyx_t_1.data = NULL; + + /* "fairseq/data/token_block_utils_fast.pyx":59 + * cdef DTYPE_t[:] sizes_view = sizes + * cdef np.ndarray[DTYPE_t, ndim=2] slice_indices + * cdef list slice_indices_list = [] # <<<<<<<<<<<<<< + * + * if break_mode is None or break_mode == 'none': + */ + __pyx_t_2 = PyList_New(0); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 59, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_v_slice_indices_list = ((PyObject*)__pyx_t_2); + __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":61 + * cdef list slice_indices_list = [] + * + * if break_mode is None or break_mode == 'none': # <<<<<<<<<<<<<< + * slice_indices = _get_slice_indices_none_mode(sizes, block_size) + * elif break_mode == 'complete': + */ + __pyx_t_4 = (__pyx_v_break_mode == ((PyObject*)Py_None)); + if (!__pyx_t_4) { + } else { + __pyx_t_3 = __pyx_t_4; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_4 = (__Pyx_PyUnicode_Equals(__pyx_v_break_mode, __pyx_n_u_none, Py_EQ)); if (unlikely((__pyx_t_4 < 0))) __PYX_ERR(0, 61, __pyx_L1_error) + __pyx_t_3 = __pyx_t_4; + __pyx_L4_bool_binop_done:; + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":62 + * + * if break_mode is None or break_mode == 'none': + * slice_indices = _get_slice_indices_none_mode(sizes, block_size) # <<<<<<<<<<<<<< + * elif break_mode == 'complete': + * while sz_idx < len(sizes_view): + */ + __pyx_t_2 = ((PyObject *)__pyx_f_7fairseq_4data_22token_block_utils_fast__get_slice_indices_none_mode(((PyArrayObject *)__pyx_v_sizes), __pyx_v_block_size)); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 62, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_t_5 = __Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)((PyArrayObject *)__pyx_t_2), &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack); + if (unlikely(__pyx_t_5 < 0)) { + PyErr_Fetch(&__pyx_t_6, &__pyx_t_7, &__pyx_t_8); + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_slice_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) { + Py_XDECREF(__pyx_t_6); Py_XDECREF(__pyx_t_7); Py_XDECREF(__pyx_t_8); + __Pyx_RaiseBufferFallbackError(); + } else { + PyErr_Restore(__pyx_t_6, __pyx_t_7, __pyx_t_8); + } + __pyx_t_6 = __pyx_t_7 = __pyx_t_8 = 0; + } + __pyx_pybuffernd_slice_indices.diminfo[0].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_slice_indices.diminfo[0].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_slice_indices.diminfo[1].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_slice_indices.diminfo[1].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[1]; + if (unlikely((__pyx_t_5 < 0))) __PYX_ERR(0, 62, __pyx_L1_error) + } + __pyx_v_slice_indices = ((PyArrayObject *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":61 + * cdef list slice_indices_list = [] + * + * if break_mode is None or break_mode == 'none': # <<<<<<<<<<<<<< + * slice_indices = _get_slice_indices_none_mode(sizes, block_size) + * elif break_mode == 'complete': + */ + goto __pyx_L3; + } + + /* "fairseq/data/token_block_utils_fast.pyx":63 + * if break_mode is None or break_mode == 'none': + * slice_indices = _get_slice_indices_none_mode(sizes, block_size) + * elif break_mode == 'complete': # <<<<<<<<<<<<<< + * while sz_idx < len(sizes_view): + * if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: + */ + __pyx_t_3 = (__Pyx_PyUnicode_Equals(__pyx_v_break_mode, __pyx_n_u_complete, Py_EQ)); if (unlikely((__pyx_t_3 < 0))) __PYX_ERR(0, 63, __pyx_L1_error) + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":64 + * slice_indices = _get_slice_indices_none_mode(sizes, block_size) + * elif break_mode == 'complete': + * while sz_idx < len(sizes_view): # <<<<<<<<<<<<<< + * if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: + * curr_size += sizes_view[sz_idx] + */ + while (1) { + __pyx_t_9 = __Pyx_MemoryView_Len(__pyx_v_sizes_view); + __pyx_t_3 = (__pyx_v_sz_idx < __pyx_t_9); + if (!__pyx_t_3) break; + + /* "fairseq/data/token_block_utils_fast.pyx":65 + * elif break_mode == 'complete': + * while sz_idx < len(sizes_view): + * if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: # <<<<<<<<<<<<<< + * curr_size += sizes_view[sz_idx] + * sz_idx += 1 + */ + __pyx_t_10 = __pyx_v_sz_idx; + __pyx_t_4 = ((__pyx_v_curr_size + (*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_sizes_view.data + __pyx_t_10 * __pyx_v_sizes_view.strides[0]) )))) <= __pyx_v_block_size); + if (!__pyx_t_4) { + } else { + __pyx_t_3 = __pyx_t_4; + goto __pyx_L9_bool_binop_done; + } + __pyx_t_4 = (__pyx_v_curr_size == 0); + __pyx_t_3 = __pyx_t_4; + __pyx_L9_bool_binop_done:; + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":66 + * while sz_idx < len(sizes_view): + * if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: + * curr_size += sizes_view[sz_idx] # <<<<<<<<<<<<<< + * sz_idx += 1 + * else: + */ + __pyx_t_10 = __pyx_v_sz_idx; + __pyx_v_curr_size = (__pyx_v_curr_size + (*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_sizes_view.data + __pyx_t_10 * __pyx_v_sizes_view.strides[0]) )))); + + /* "fairseq/data/token_block_utils_fast.pyx":67 + * if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: + * curr_size += sizes_view[sz_idx] + * sz_idx += 1 # <<<<<<<<<<<<<< + * else: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + */ + __pyx_v_sz_idx = (__pyx_v_sz_idx + 1); + + /* "fairseq/data/token_block_utils_fast.pyx":65 + * elif break_mode == 'complete': + * while sz_idx < len(sizes_view): + * if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: # <<<<<<<<<<<<<< + * curr_size += sizes_view[sz_idx] + * sz_idx += 1 + */ + goto __pyx_L8; + } + + /* "fairseq/data/token_block_utils_fast.pyx":69 + * sz_idx += 1 + * else: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) # <<<<<<<<<<<<<< + * tok_idx += curr_size + * curr_size = 0 + */ + /*else*/ { + __pyx_t_2 = __Pyx_PyInt_From_npy_int64(__pyx_v_tok_idx); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 69, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_11 = __Pyx_PyInt_From_npy_int64((__pyx_v_tok_idx + __pyx_v_curr_size)); if (unlikely(!__pyx_t_11)) __PYX_ERR(0, 69, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_11); + __pyx_t_12 = PyTuple_New(2); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 69, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_12, 0, __pyx_t_2)) __PYX_ERR(0, 69, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_11); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_12, 1, __pyx_t_11)) __PYX_ERR(0, 69, __pyx_L1_error); + __pyx_t_2 = 0; + __pyx_t_11 = 0; + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_slice_indices_list, __pyx_t_12); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 69, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_12); __pyx_t_12 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":70 + * else: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * tok_idx += curr_size # <<<<<<<<<<<<<< + * curr_size = 0 + * if curr_size > 0: + */ + __pyx_v_tok_idx = (__pyx_v_tok_idx + __pyx_v_curr_size); + + /* "fairseq/data/token_block_utils_fast.pyx":71 + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * tok_idx += curr_size + * curr_size = 0 # <<<<<<<<<<<<<< + * if curr_size > 0: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + */ + __pyx_v_curr_size = 0; + } + __pyx_L8:; + } + + /* "fairseq/data/token_block_utils_fast.pyx":72 + * tok_idx += curr_size + * curr_size = 0 + * if curr_size > 0: # <<<<<<<<<<<<<< + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + */ + __pyx_t_3 = (__pyx_v_curr_size > 0); + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":73 + * curr_size = 0 + * if curr_size > 0: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) # <<<<<<<<<<<<<< + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'complete_doc': + */ + __pyx_t_12 = __Pyx_PyInt_From_npy_int64(__pyx_v_tok_idx); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 73, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __pyx_t_11 = __Pyx_PyInt_From_npy_int64((__pyx_v_tok_idx + __pyx_v_curr_size)); if (unlikely(!__pyx_t_11)) __PYX_ERR(0, 73, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_11); + __pyx_t_2 = PyTuple_New(2); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 73, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_12); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 0, __pyx_t_12)) __PYX_ERR(0, 73, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_11); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 1, __pyx_t_11)) __PYX_ERR(0, 73, __pyx_L1_error); + __pyx_t_12 = 0; + __pyx_t_11 = 0; + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_slice_indices_list, __pyx_t_2); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 73, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":72 + * tok_idx += curr_size + * curr_size = 0 + * if curr_size > 0: # <<<<<<<<<<<<<< + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + */ + } + + /* "fairseq/data/token_block_utils_fast.pyx":74 + * if curr_size > 0: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) # <<<<<<<<<<<<<< + * elif break_mode == 'complete_doc': + * while sz_idx < len(sizes_view): + */ + __pyx_t_2 = ((PyObject *)__pyx_f_7fairseq_4data_22token_block_utils_fast__fast_convert_to_np_array(__pyx_v_slice_indices_list)); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 74, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_t_5 = __Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)((PyArrayObject *)__pyx_t_2), &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack); + if (unlikely(__pyx_t_5 < 0)) { + PyErr_Fetch(&__pyx_t_8, &__pyx_t_7, &__pyx_t_6); + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_slice_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) { + Py_XDECREF(__pyx_t_8); Py_XDECREF(__pyx_t_7); Py_XDECREF(__pyx_t_6); + __Pyx_RaiseBufferFallbackError(); + } else { + PyErr_Restore(__pyx_t_8, __pyx_t_7, __pyx_t_6); + } + __pyx_t_8 = __pyx_t_7 = __pyx_t_6 = 0; + } + __pyx_pybuffernd_slice_indices.diminfo[0].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_slice_indices.diminfo[0].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_slice_indices.diminfo[1].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_slice_indices.diminfo[1].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[1]; + if (unlikely((__pyx_t_5 < 0))) __PYX_ERR(0, 74, __pyx_L1_error) + } + __pyx_v_slice_indices = ((PyArrayObject *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":63 + * if break_mode is None or break_mode == 'none': + * slice_indices = _get_slice_indices_none_mode(sizes, block_size) + * elif break_mode == 'complete': # <<<<<<<<<<<<<< + * while sz_idx < len(sizes_view): + * if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: + */ + goto __pyx_L3; + } + + /* "fairseq/data/token_block_utils_fast.pyx":75 + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'complete_doc': # <<<<<<<<<<<<<< + * while sz_idx < len(sizes_view): + * if ( + */ + __pyx_t_3 = (__Pyx_PyUnicode_Equals(__pyx_v_break_mode, __pyx_n_u_complete_doc, Py_EQ)); if (unlikely((__pyx_t_3 < 0))) __PYX_ERR(0, 75, __pyx_L1_error) + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":76 + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'complete_doc': + * while sz_idx < len(sizes_view): # <<<<<<<<<<<<<< + * if ( + * (curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0) + */ + while (1) { + __pyx_t_9 = __Pyx_MemoryView_Len(__pyx_v_sizes_view); + __pyx_t_3 = (__pyx_v_sz_idx < __pyx_t_9); + if (!__pyx_t_3) break; + + /* "fairseq/data/token_block_utils_fast.pyx":78 + * while sz_idx < len(sizes_view): + * if ( + * (curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0) # <<<<<<<<<<<<<< + * # an empty sentence indicates end-of-document: + * and sizes_view[sz_idx] != document_sep_len + */ + __pyx_t_10 = __pyx_v_sz_idx; + __pyx_t_4 = ((__pyx_v_curr_size + (*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_sizes_view.data + __pyx_t_10 * __pyx_v_sizes_view.strides[0]) )))) <= __pyx_v_block_size); + if (!__pyx_t_4) { + } else { + goto __pyx_L16_next_and; + } + __pyx_t_4 = (__pyx_v_curr_size == 0); + if (__pyx_t_4) { + } else { + __pyx_t_3 = __pyx_t_4; + goto __pyx_L15_bool_binop_done; + } + __pyx_L16_next_and:; + + /* "fairseq/data/token_block_utils_fast.pyx":80 + * (curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0) + * # an empty sentence indicates end-of-document: + * and sizes_view[sz_idx] != document_sep_len # <<<<<<<<<<<<<< + * ): + * curr_size += sizes_view[sz_idx] + */ + __pyx_t_10 = __pyx_v_sz_idx; + __pyx_t_4 = ((*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_sizes_view.data + __pyx_t_10 * __pyx_v_sizes_view.strides[0]) ))) != __pyx_v_document_sep_len); + __pyx_t_3 = __pyx_t_4; + __pyx_L15_bool_binop_done:; + + /* "fairseq/data/token_block_utils_fast.pyx":77 + * elif break_mode == 'complete_doc': + * while sz_idx < len(sizes_view): + * if ( # <<<<<<<<<<<<<< + * (curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0) + * # an empty sentence indicates end-of-document: + */ + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":82 + * and sizes_view[sz_idx] != document_sep_len + * ): + * curr_size += sizes_view[sz_idx] # <<<<<<<<<<<<<< + * sz_idx += 1 + * else: + */ + __pyx_t_10 = __pyx_v_sz_idx; + __pyx_v_curr_size = (__pyx_v_curr_size + (*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_sizes_view.data + __pyx_t_10 * __pyx_v_sizes_view.strides[0]) )))); + + /* "fairseq/data/token_block_utils_fast.pyx":83 + * ): + * curr_size += sizes_view[sz_idx] + * sz_idx += 1 # <<<<<<<<<<<<<< + * else: + * # Only keep non-empty documents. + */ + __pyx_v_sz_idx = (__pyx_v_sz_idx + 1); + + /* "fairseq/data/token_block_utils_fast.pyx":77 + * elif break_mode == 'complete_doc': + * while sz_idx < len(sizes_view): + * if ( # <<<<<<<<<<<<<< + * (curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0) + * # an empty sentence indicates end-of-document: + */ + goto __pyx_L14; + } + + /* "fairseq/data/token_block_utils_fast.pyx":86 + * else: + * # Only keep non-empty documents. + * if curr_size > 1: # <<<<<<<<<<<<<< + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * tok_idx += curr_size + */ + /*else*/ { + __pyx_t_3 = (__pyx_v_curr_size > 1); + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":87 + * # Only keep non-empty documents. + * if curr_size > 1: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) # <<<<<<<<<<<<<< + * tok_idx += curr_size + * curr_size = 0 + */ + __pyx_t_2 = __Pyx_PyInt_From_npy_int64(__pyx_v_tok_idx); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 87, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_11 = __Pyx_PyInt_From_npy_int64((__pyx_v_tok_idx + __pyx_v_curr_size)); if (unlikely(!__pyx_t_11)) __PYX_ERR(0, 87, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_11); + __pyx_t_12 = PyTuple_New(2); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 87, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_12, 0, __pyx_t_2)) __PYX_ERR(0, 87, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_11); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_12, 1, __pyx_t_11)) __PYX_ERR(0, 87, __pyx_L1_error); + __pyx_t_2 = 0; + __pyx_t_11 = 0; + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_slice_indices_list, __pyx_t_12); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 87, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_12); __pyx_t_12 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":86 + * else: + * # Only keep non-empty documents. + * if curr_size > 1: # <<<<<<<<<<<<<< + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * tok_idx += curr_size + */ + } + + /* "fairseq/data/token_block_utils_fast.pyx":88 + * if curr_size > 1: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * tok_idx += curr_size # <<<<<<<<<<<<<< + * curr_size = 0 + * if sizes_view[sz_idx] == document_sep_len: + */ + __pyx_v_tok_idx = (__pyx_v_tok_idx + __pyx_v_curr_size); + + /* "fairseq/data/token_block_utils_fast.pyx":89 + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * tok_idx += curr_size + * curr_size = 0 # <<<<<<<<<<<<<< + * if sizes_view[sz_idx] == document_sep_len: + * tok_idx += sizes_view[sz_idx] + */ + __pyx_v_curr_size = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":90 + * tok_idx += curr_size + * curr_size = 0 + * if sizes_view[sz_idx] == document_sep_len: # <<<<<<<<<<<<<< + * tok_idx += sizes_view[sz_idx] + * sz_idx += 1 + */ + __pyx_t_10 = __pyx_v_sz_idx; + __pyx_t_3 = ((*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_sizes_view.data + __pyx_t_10 * __pyx_v_sizes_view.strides[0]) ))) == __pyx_v_document_sep_len); + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":91 + * curr_size = 0 + * if sizes_view[sz_idx] == document_sep_len: + * tok_idx += sizes_view[sz_idx] # <<<<<<<<<<<<<< + * sz_idx += 1 + * if curr_size > 1: + */ + __pyx_t_10 = __pyx_v_sz_idx; + __pyx_v_tok_idx = (__pyx_v_tok_idx + (*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_sizes_view.data + __pyx_t_10 * __pyx_v_sizes_view.strides[0]) )))); + + /* "fairseq/data/token_block_utils_fast.pyx":92 + * if sizes_view[sz_idx] == document_sep_len: + * tok_idx += sizes_view[sz_idx] + * sz_idx += 1 # <<<<<<<<<<<<<< + * if curr_size > 1: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + */ + __pyx_v_sz_idx = (__pyx_v_sz_idx + 1); + + /* "fairseq/data/token_block_utils_fast.pyx":90 + * tok_idx += curr_size + * curr_size = 0 + * if sizes_view[sz_idx] == document_sep_len: # <<<<<<<<<<<<<< + * tok_idx += sizes_view[sz_idx] + * sz_idx += 1 + */ + } + } + __pyx_L14:; + } + + /* "fairseq/data/token_block_utils_fast.pyx":93 + * tok_idx += sizes_view[sz_idx] + * sz_idx += 1 + * if curr_size > 1: # <<<<<<<<<<<<<< + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + */ + __pyx_t_3 = (__pyx_v_curr_size > 1); + if (__pyx_t_3) { + + /* "fairseq/data/token_block_utils_fast.pyx":94 + * sz_idx += 1 + * if curr_size > 1: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) # <<<<<<<<<<<<<< + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'eos': + */ + __pyx_t_12 = __Pyx_PyInt_From_npy_int64(__pyx_v_tok_idx); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 94, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __pyx_t_11 = __Pyx_PyInt_From_npy_int64((__pyx_v_tok_idx + __pyx_v_curr_size)); if (unlikely(!__pyx_t_11)) __PYX_ERR(0, 94, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_11); + __pyx_t_2 = PyTuple_New(2); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 94, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_12); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 0, __pyx_t_12)) __PYX_ERR(0, 94, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_11); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 1, __pyx_t_11)) __PYX_ERR(0, 94, __pyx_L1_error); + __pyx_t_12 = 0; + __pyx_t_11 = 0; + __pyx_t_13 = __Pyx_PyList_Append(__pyx_v_slice_indices_list, __pyx_t_2); if (unlikely(__pyx_t_13 == ((int)-1))) __PYX_ERR(0, 94, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":93 + * tok_idx += sizes_view[sz_idx] + * sz_idx += 1 + * if curr_size > 1: # <<<<<<<<<<<<<< + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + */ + } + + /* "fairseq/data/token_block_utils_fast.pyx":95 + * if curr_size > 1: + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) # <<<<<<<<<<<<<< + * elif break_mode == 'eos': + * slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + */ + __pyx_t_2 = ((PyObject *)__pyx_f_7fairseq_4data_22token_block_utils_fast__fast_convert_to_np_array(__pyx_v_slice_indices_list)); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 95, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_t_5 = __Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)((PyArrayObject *)__pyx_t_2), &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack); + if (unlikely(__pyx_t_5 < 0)) { + PyErr_Fetch(&__pyx_t_6, &__pyx_t_7, &__pyx_t_8); + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_slice_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) { + Py_XDECREF(__pyx_t_6); Py_XDECREF(__pyx_t_7); Py_XDECREF(__pyx_t_8); + __Pyx_RaiseBufferFallbackError(); + } else { + PyErr_Restore(__pyx_t_6, __pyx_t_7, __pyx_t_8); + } + __pyx_t_6 = __pyx_t_7 = __pyx_t_8 = 0; + } + __pyx_pybuffernd_slice_indices.diminfo[0].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_slice_indices.diminfo[0].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_slice_indices.diminfo[1].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_slice_indices.diminfo[1].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[1]; + if (unlikely((__pyx_t_5 < 0))) __PYX_ERR(0, 95, __pyx_L1_error) + } + __pyx_v_slice_indices = ((PyArrayObject *)__pyx_t_2); + __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":75 + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'complete_doc': # <<<<<<<<<<<<<< + * while sz_idx < len(sizes_view): + * if ( + */ + goto __pyx_L3; + } + + /* "fairseq/data/token_block_utils_fast.pyx":96 + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'eos': # <<<<<<<<<<<<<< + * slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + * cumsum = sizes.cumsum(axis=0) + */ + __pyx_t_3 = (__Pyx_PyUnicode_Equals(__pyx_v_break_mode, __pyx_n_u_eos, Py_EQ)); if (unlikely((__pyx_t_3 < 0))) __PYX_ERR(0, 96, __pyx_L1_error) + if (likely(__pyx_t_3)) { + + /* "fairseq/data/token_block_utils_fast.pyx":97 + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'eos': + * slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) # <<<<<<<<<<<<<< + * cumsum = sizes.cumsum(axis=0) + * slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] + */ + __Pyx_GetModuleGlobalName(__pyx_t_2, __pyx_n_s_np); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_11 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_zeros); if (unlikely(!__pyx_t_11)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_11); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_9 = PyObject_Length(((PyObject *)__pyx_v_sizes)); if (unlikely(__pyx_t_9 == ((Py_ssize_t)-1))) __PYX_ERR(0, 97, __pyx_L1_error) + __pyx_t_2 = PyInt_FromSsize_t(__pyx_t_9); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_12 = PyTuple_New(2); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_12, 0, __pyx_t_2)) __PYX_ERR(0, 97, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_2); + __Pyx_GIVEREF(__pyx_int_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_12, 1, __pyx_int_2)) __PYX_ERR(0, 97, __pyx_L1_error); + __pyx_t_2 = 0; + __pyx_t_2 = PyTuple_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_GIVEREF(__pyx_t_12); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_2, 0, __pyx_t_12)) __PYX_ERR(0, 97, __pyx_L1_error); + __pyx_t_12 = 0; + __pyx_t_12 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __Pyx_GetModuleGlobalName(__pyx_t_14, __pyx_n_s_DTYPE); if (unlikely(!__pyx_t_14)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_14); + if (PyDict_SetItem(__pyx_t_12, __pyx_n_s_dtype, __pyx_t_14) < 0) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_14); __pyx_t_14 = 0; + __pyx_t_14 = __Pyx_PyObject_Call(__pyx_t_11, __pyx_t_2, __pyx_t_12); if (unlikely(!__pyx_t_14)) __PYX_ERR(0, 97, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_14); + __Pyx_DECREF(__pyx_t_11); __pyx_t_11 = 0; + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_DECREF(__pyx_t_12); __pyx_t_12 = 0; + if (!(likely(((__pyx_t_14) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_14, __pyx_ptype_5numpy_ndarray))))) __PYX_ERR(0, 97, __pyx_L1_error) + __pyx_t_15 = ((PyArrayObject *)__pyx_t_14); + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_t_5 = __Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_t_15, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack); + if (unlikely(__pyx_t_5 < 0)) { + PyErr_Fetch(&__pyx_t_8, &__pyx_t_7, &__pyx_t_6); + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_slice_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) { + Py_XDECREF(__pyx_t_8); Py_XDECREF(__pyx_t_7); Py_XDECREF(__pyx_t_6); + __Pyx_RaiseBufferFallbackError(); + } else { + PyErr_Restore(__pyx_t_8, __pyx_t_7, __pyx_t_6); + } + __pyx_t_8 = __pyx_t_7 = __pyx_t_6 = 0; + } + __pyx_pybuffernd_slice_indices.diminfo[0].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_slice_indices.diminfo[0].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_slice_indices.diminfo[1].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_slice_indices.diminfo[1].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[1]; + if (unlikely((__pyx_t_5 < 0))) __PYX_ERR(0, 97, __pyx_L1_error) + } + __pyx_t_15 = 0; + __pyx_v_slice_indices = ((PyArrayObject *)__pyx_t_14); + __pyx_t_14 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":98 + * elif break_mode == 'eos': + * slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + * cumsum = sizes.cumsum(axis=0) # <<<<<<<<<<<<<< + * slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] + * slice_indices[:, 1] = cumsum + */ + __pyx_t_14 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_sizes), __pyx_n_s_cumsum); if (unlikely(!__pyx_t_14)) __PYX_ERR(0, 98, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_14); + __pyx_t_12 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 98, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + if (PyDict_SetItem(__pyx_t_12, __pyx_n_s_axis, __pyx_int_0) < 0) __PYX_ERR(0, 98, __pyx_L1_error) + __pyx_t_2 = __Pyx_PyObject_Call(__pyx_t_14, __pyx_empty_tuple, __pyx_t_12); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 98, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_14); __pyx_t_14 = 0; + __Pyx_DECREF(__pyx_t_12); __pyx_t_12 = 0; + __pyx_v_cumsum = __pyx_t_2; + __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":99 + * slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + * cumsum = sizes.cumsum(axis=0) + * slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] # <<<<<<<<<<<<<< + * slice_indices[:, 1] = cumsum + * else: + */ + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_v_cumsum, __pyx_n_s_shape); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 99, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_12 = __Pyx_GetItemInt(__pyx_t_2, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 0); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 99, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __pyx_t_2 = __Pyx_PyInt_SubtractObjC(__pyx_t_12, __pyx_int_1, 1, 0, 0); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 99, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_12); __pyx_t_12 = 0; + __pyx_t_12 = __Pyx_PyObject_GetSlice(__pyx_v_cumsum, 0, 0, NULL, &__pyx_t_2, NULL, 0, 0, 0); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 99, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + if (unlikely((PyObject_SetItem(((PyObject *)__pyx_v_slice_indices), __pyx_tuple__12, __pyx_t_12) < 0))) __PYX_ERR(0, 99, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_12); __pyx_t_12 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":100 + * cumsum = sizes.cumsum(axis=0) + * slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] + * slice_indices[:, 1] = cumsum # <<<<<<<<<<<<<< + * else: + * raise ValueError('Invalid break_mode: ' + break_mode) + */ + if (unlikely((PyObject_SetItem(((PyObject *)__pyx_v_slice_indices), __pyx_tuple__13, __pyx_v_cumsum) < 0))) __PYX_ERR(0, 100, __pyx_L1_error) + + /* "fairseq/data/token_block_utils_fast.pyx":96 + * slice_indices_list.append((tok_idx, tok_idx + curr_size)) + * slice_indices = _fast_convert_to_np_array(slice_indices_list) + * elif break_mode == 'eos': # <<<<<<<<<<<<<< + * slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + * cumsum = sizes.cumsum(axis=0) + */ + goto __pyx_L3; + } + + /* "fairseq/data/token_block_utils_fast.pyx":102 + * slice_indices[:, 1] = cumsum + * else: + * raise ValueError('Invalid break_mode: ' + break_mode) # <<<<<<<<<<<<<< + * return slice_indices + * + */ + /*else*/ { + __pyx_t_12 = __Pyx_PyUnicode_ConcatSafe(__pyx_kp_u_Invalid_break_mode, __pyx_v_break_mode); if (unlikely(!__pyx_t_12)) __PYX_ERR(0, 102, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_12); + __pyx_t_2 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_12); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 102, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_12); __pyx_t_12 = 0; + __Pyx_Raise(__pyx_t_2, 0, 0, 0); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __PYX_ERR(0, 102, __pyx_L1_error) + } + __pyx_L3:; + + /* "fairseq/data/token_block_utils_fast.pyx":103 + * else: + * raise ValueError('Invalid break_mode: ' + break_mode) + * return slice_indices # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_slice_indices); + __pyx_r = ((PyArrayObject *)__pyx_v_slice_indices); + goto __pyx_L0; + + /* "fairseq/data/token_block_utils_fast.pyx":50 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): # <<<<<<<<<<<<<< + * cdef DTYPE_t tok_idx = 0 + * cdef DTYPE_t sz_idx = 0 + */ + + /* function exit code */ + __pyx_L1_error:; + __PYX_XCLEAR_MEMVIEW(&__pyx_t_1, 1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_11); + __Pyx_XDECREF(__pyx_t_12); + __Pyx_XDECREF(__pyx_t_14); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._get_slice_indices_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_L2:; + __PYX_XCLEAR_MEMVIEW(&__pyx_v_sizes_view, 1); + __Pyx_XDECREF((PyObject *)__pyx_v_slice_indices); + __Pyx_XDECREF(__pyx_v_slice_indices_list); + __Pyx_XDECREF(__pyx_v_cumsum); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* Python wrapper */ +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_1_get_slice_indices_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_7fairseq_4data_22token_block_utils_fast_1_get_slice_indices_fast = {"_get_slice_indices_fast", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_22token_block_utils_fast_1_get_slice_indices_fast, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_1_get_slice_indices_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyArrayObject *__pyx_v_sizes = 0; + PyObject *__pyx_v_break_mode = 0; + int __pyx_v_block_size; + int __pyx_v_document_sep_len; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[4] = {0,0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("_get_slice_indices_fast (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_sizes,&__pyx_n_s_break_mode,&__pyx_n_s_block_size,&__pyx_n_s_document_sep_len,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 4: values[3] = __Pyx_Arg_FASTCALL(__pyx_args, 3); + CYTHON_FALLTHROUGH; + case 3: values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_sizes)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 50, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_break_mode)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 50, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("_get_slice_indices_fast", 1, 4, 4, 1); __PYX_ERR(0, 50, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_block_size)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 50, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("_get_slice_indices_fast", 1, 4, 4, 2); __PYX_ERR(0, 50, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 3: + if (likely((values[3] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_document_sep_len)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[3]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 50, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("_get_slice_indices_fast", 1, 4, 4, 3); __PYX_ERR(0, 50, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "_get_slice_indices_fast") < 0)) __PYX_ERR(0, 50, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 4)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + values[3] = __Pyx_Arg_FASTCALL(__pyx_args, 3); + } + __pyx_v_sizes = ((PyArrayObject *)values[0]); + __pyx_v_break_mode = ((PyObject*)values[1]); + __pyx_v_block_size = __Pyx_PyInt_As_int(values[2]); if (unlikely((__pyx_v_block_size == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 50, __pyx_L3_error) + __pyx_v_document_sep_len = __Pyx_PyInt_As_int(values[3]); if (unlikely((__pyx_v_document_sep_len == (int)-1) && PyErr_Occurred())) __PYX_ERR(0, 50, __pyx_L3_error) + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("_get_slice_indices_fast", 1, 4, 4, __pyx_nargs); __PYX_ERR(0, 50, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._get_slice_indices_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_sizes), __pyx_ptype_5numpy_ndarray, 1, "sizes", 0))) __PYX_ERR(0, 50, __pyx_L1_error) + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_break_mode), (&PyUnicode_Type), 1, "break_mode", 1))) __PYX_ERR(0, 50, __pyx_L1_error) + __pyx_r = __pyx_pf_7fairseq_4data_22token_block_utils_fast__get_slice_indices_fast(__pyx_self, __pyx_v_sizes, __pyx_v_break_mode, __pyx_v_block_size, __pyx_v_document_sep_len); + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __pyx_r = NULL; + __pyx_L0:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast__get_slice_indices_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_sizes, PyObject *__pyx_v_break_mode, int __pyx_v_block_size, int __pyx_v_document_sep_len) { + __Pyx_LocalBuf_ND __pyx_pybuffernd_sizes; + __Pyx_Buffer __pyx_pybuffer_sizes; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_get_slice_indices_fast", 1); + __pyx_pybuffer_sizes.pybuffer.buf = NULL; + __pyx_pybuffer_sizes.refcount = 0; + __pyx_pybuffernd_sizes.data = NULL; + __pyx_pybuffernd_sizes.rcbuffer = &__pyx_pybuffer_sizes; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer, (PyObject*)__pyx_v_sizes, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 50, __pyx_L1_error) + } + __pyx_pybuffernd_sizes.diminfo[0].strides = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_sizes.diminfo[0].shape = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.shape[0]; + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = ((PyObject *)__pyx_f_7fairseq_4data_22token_block_utils_fast__get_slice_indices_fast(__pyx_v_sizes, __pyx_v_break_mode, __pyx_v_block_size, __pyx_v_document_sep_len, 0)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 50, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._get_slice_indices_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":109 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_block_to_dataset_index_fast(np.ndarray[DTYPE_t, ndim=1] sizes, np.ndarray[DTYPE_t, ndim=2] slice_indices): # <<<<<<<<<<<<<< + * cdef DTYPE_t start_ds_idx + * cdef DTYPE_t start_offset + */ + +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_3_get_block_to_dataset_index_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyArrayObject *__pyx_f_7fairseq_4data_22token_block_utils_fast__get_block_to_dataset_index_fast(PyArrayObject *__pyx_v_sizes, PyArrayObject *__pyx_v_slice_indices, CYTHON_UNUSED int __pyx_skip_dispatch) { + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_start_ds_idx; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_start_offset; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_end_ds_idx; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_i; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_s; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_e; + struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_ds = 0; + PyArrayObject *__pyx_v_block_to_dataset_index = 0; + __Pyx_memviewslice __pyx_v_block_to_dataset_index_view = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_memviewslice __pyx_v_slice_indices_view = { 0, 0, { 0 }, { 0 }, { 0 } }; + Py_ssize_t __pyx_v_x_max; + __Pyx_LocalBuf_ND __pyx_pybuffernd_block_to_dataset_index; + __Pyx_Buffer __pyx_pybuffer_block_to_dataset_index; + __Pyx_LocalBuf_ND __pyx_pybuffernd_sizes; + __Pyx_Buffer __pyx_pybuffer_sizes; + __Pyx_LocalBuf_ND __pyx_pybuffernd_slice_indices; + __Pyx_Buffer __pyx_pybuffer_slice_indices; + PyArrayObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + Py_ssize_t __pyx_t_3; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + PyArrayObject *__pyx_t_6 = NULL; + __Pyx_memviewslice __pyx_t_7 = { 0, 0, { 0 }, { 0 }, { 0 } }; + Py_ssize_t __pyx_t_8; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_9; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_10; + Py_ssize_t __pyx_t_11; + int __pyx_t_12; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_get_block_to_dataset_index_fast", 1); + __pyx_pybuffer_block_to_dataset_index.pybuffer.buf = NULL; + __pyx_pybuffer_block_to_dataset_index.refcount = 0; + __pyx_pybuffernd_block_to_dataset_index.data = NULL; + __pyx_pybuffernd_block_to_dataset_index.rcbuffer = &__pyx_pybuffer_block_to_dataset_index; + __pyx_pybuffer_sizes.pybuffer.buf = NULL; + __pyx_pybuffer_sizes.refcount = 0; + __pyx_pybuffernd_sizes.data = NULL; + __pyx_pybuffernd_sizes.rcbuffer = &__pyx_pybuffer_sizes; + __pyx_pybuffer_slice_indices.pybuffer.buf = NULL; + __pyx_pybuffer_slice_indices.refcount = 0; + __pyx_pybuffernd_slice_indices.data = NULL; + __pyx_pybuffernd_slice_indices.rcbuffer = &__pyx_pybuffer_slice_indices; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer, (PyObject*)__pyx_v_sizes, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 109, __pyx_L1_error) + } + __pyx_pybuffernd_sizes.diminfo[0].strides = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_sizes.diminfo[0].shape = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.shape[0]; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_slice_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 109, __pyx_L1_error) + } + __pyx_pybuffernd_slice_indices.diminfo[0].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_slice_indices.diminfo[0].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_slice_indices.diminfo[1].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_slice_indices.diminfo[1].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[1]; + + /* "fairseq/data/token_block_utils_fast.pyx":116 + * cdef DTYPE_t s + * cdef DTYPE_t e + * cdef DatasetSearcher ds = DatasetSearcher(sizes) # <<<<<<<<<<<<<< + * cdef np.ndarray[DTYPE_t, ndim=2] block_to_dataset_index = np.zeros([len(slice_indices), 3], dtype=DTYPE) + * cdef DTYPE_t[:, :] block_to_dataset_index_view = block_to_dataset_index + */ + __pyx_t_1 = __Pyx_PyObject_CallOneArg(((PyObject *)__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher), ((PyObject *)__pyx_v_sizes)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 116, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_v_ds = ((struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_t_1); + __pyx_t_1 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":117 + * cdef DTYPE_t e + * cdef DatasetSearcher ds = DatasetSearcher(sizes) + * cdef np.ndarray[DTYPE_t, ndim=2] block_to_dataset_index = np.zeros([len(slice_indices), 3], dtype=DTYPE) # <<<<<<<<<<<<<< + * cdef DTYPE_t[:, :] block_to_dataset_index_view = block_to_dataset_index + * cdef DTYPE_t[:, :] slice_indices_view = slice_indices + */ + __Pyx_GetModuleGlobalName(__pyx_t_1, __pyx_n_s_np); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_zeros); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_3 = PyObject_Length(((PyObject *)__pyx_v_slice_indices)); if (unlikely(__pyx_t_3 == ((Py_ssize_t)-1))) __PYX_ERR(0, 117, __pyx_L1_error) + __pyx_t_1 = PyInt_FromSsize_t(__pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_4 = PyList_New(2); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyList_SET_ITEM(__pyx_t_4, 0, __pyx_t_1)) __PYX_ERR(0, 117, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_3); + __Pyx_GIVEREF(__pyx_int_3); + if (__Pyx_PyList_SET_ITEM(__pyx_t_4, 1, __pyx_int_3)) __PYX_ERR(0, 117, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_1 = PyTuple_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_GIVEREF(__pyx_t_4); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_t_4)) __PYX_ERR(0, 117, __pyx_L1_error); + __pyx_t_4 = 0; + __pyx_t_4 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_GetModuleGlobalName(__pyx_t_5, __pyx_n_s_DTYPE); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + if (PyDict_SetItem(__pyx_t_4, __pyx_n_s_dtype, __pyx_t_5) < 0) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __pyx_t_5 = __Pyx_PyObject_Call(__pyx_t_2, __pyx_t_1, __pyx_t_4); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 117, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + if (!(likely(((__pyx_t_5) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_5, __pyx_ptype_5numpy_ndarray))))) __PYX_ERR(0, 117, __pyx_L1_error) + __pyx_t_6 = ((PyArrayObject *)__pyx_t_5); + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer, (PyObject*)__pyx_t_6, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) { + __pyx_v_block_to_dataset_index = ((PyArrayObject *)Py_None); __Pyx_INCREF(Py_None); __pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer.buf = NULL; + __PYX_ERR(0, 117, __pyx_L1_error) + } else {__pyx_pybuffernd_block_to_dataset_index.diminfo[0].strides = __pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_block_to_dataset_index.diminfo[0].shape = __pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_block_to_dataset_index.diminfo[1].strides = __pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_block_to_dataset_index.diminfo[1].shape = __pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer.shape[1]; + } + } + __pyx_t_6 = 0; + __pyx_v_block_to_dataset_index = ((PyArrayObject *)__pyx_t_5); + __pyx_t_5 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":118 + * cdef DatasetSearcher ds = DatasetSearcher(sizes) + * cdef np.ndarray[DTYPE_t, ndim=2] block_to_dataset_index = np.zeros([len(slice_indices), 3], dtype=DTYPE) + * cdef DTYPE_t[:, :] block_to_dataset_index_view = block_to_dataset_index # <<<<<<<<<<<<<< + * cdef DTYPE_t[:, :] slice_indices_view = slice_indices + * cdef Py_ssize_t x_max = slice_indices.shape[0] + */ + __pyx_t_7 = __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(((PyObject *)__pyx_v_block_to_dataset_index), PyBUF_WRITABLE); if (unlikely(!__pyx_t_7.memview)) __PYX_ERR(0, 118, __pyx_L1_error) + __pyx_v_block_to_dataset_index_view = __pyx_t_7; + __pyx_t_7.memview = NULL; + __pyx_t_7.data = NULL; + + /* "fairseq/data/token_block_utils_fast.pyx":119 + * cdef np.ndarray[DTYPE_t, ndim=2] block_to_dataset_index = np.zeros([len(slice_indices), 3], dtype=DTYPE) + * cdef DTYPE_t[:, :] block_to_dataset_index_view = block_to_dataset_index + * cdef DTYPE_t[:, :] slice_indices_view = slice_indices # <<<<<<<<<<<<<< + * cdef Py_ssize_t x_max = slice_indices.shape[0] + * + */ + __pyx_t_7 = __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(((PyObject *)__pyx_v_slice_indices), PyBUF_WRITABLE); if (unlikely(!__pyx_t_7.memview)) __PYX_ERR(0, 119, __pyx_L1_error) + __pyx_v_slice_indices_view = __pyx_t_7; + __pyx_t_7.memview = NULL; + __pyx_t_7.data = NULL; + + /* "fairseq/data/token_block_utils_fast.pyx":120 + * cdef DTYPE_t[:, :] block_to_dataset_index_view = block_to_dataset_index + * cdef DTYPE_t[:, :] slice_indices_view = slice_indices + * cdef Py_ssize_t x_max = slice_indices.shape[0] # <<<<<<<<<<<<<< + * + * for i in range(x_max): + */ + __pyx_v_x_max = (__pyx_f_5numpy_7ndarray_5shape_shape(((PyArrayObject *)__pyx_v_slice_indices))[0]); + + /* "fairseq/data/token_block_utils_fast.pyx":122 + * cdef Py_ssize_t x_max = slice_indices.shape[0] + * + * for i in range(x_max): # <<<<<<<<<<<<<< + * s = slice_indices_view[i][0] + * e = slice_indices_view[i][1] + */ + __pyx_t_3 = __pyx_v_x_max; + __pyx_t_8 = __pyx_t_3; + for (__pyx_t_9 = 0; __pyx_t_9 < __pyx_t_8; __pyx_t_9+=1) { + __pyx_v_i = __pyx_t_9; + + /* "fairseq/data/token_block_utils_fast.pyx":123 + * + * for i in range(x_max): + * s = slice_indices_view[i][0] # <<<<<<<<<<<<<< + * e = slice_indices_view[i][1] + * ds.seek(s) + */ + __pyx_t_10 = __pyx_v_i; + __pyx_t_11 = 0; + __pyx_v_s = (*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_slice_indices_view.data + __pyx_t_10 * __pyx_v_slice_indices_view.strides[0]) ) + __pyx_t_11 * __pyx_v_slice_indices_view.strides[1]) ))); + + /* "fairseq/data/token_block_utils_fast.pyx":124 + * for i in range(x_max): + * s = slice_indices_view[i][0] + * e = slice_indices_view[i][1] # <<<<<<<<<<<<<< + * ds.seek(s) + * start_ds_idx = ds.current_index + */ + __pyx_t_10 = __pyx_v_i; + __pyx_t_11 = 1; + __pyx_v_e = (*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_slice_indices_view.data + __pyx_t_10 * __pyx_v_slice_indices_view.strides[0]) ) + __pyx_t_11 * __pyx_v_slice_indices_view.strides[1]) ))); + + /* "fairseq/data/token_block_utils_fast.pyx":125 + * s = slice_indices_view[i][0] + * e = slice_indices_view[i][1] + * ds.seek(s) # <<<<<<<<<<<<<< + * start_ds_idx = ds.current_index + * start_offset = ds.current_offset + */ + __pyx_t_5 = ((struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_ds->__pyx_vtab)->seek(__pyx_v_ds, __pyx_v_s); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 125, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":126 + * e = slice_indices_view[i][1] + * ds.seek(s) + * start_ds_idx = ds.current_index # <<<<<<<<<<<<<< + * start_offset = ds.current_offset + * if e <= s: + */ + __pyx_t_10 = __pyx_v_ds->current_index; + __pyx_v_start_ds_idx = __pyx_t_10; + + /* "fairseq/data/token_block_utils_fast.pyx":127 + * ds.seek(s) + * start_ds_idx = ds.current_index + * start_offset = ds.current_offset # <<<<<<<<<<<<<< + * if e <= s: + * end_ds_idx = start_ds_idx + */ + __pyx_t_10 = __pyx_v_ds->current_offset; + __pyx_v_start_offset = __pyx_t_10; + + /* "fairseq/data/token_block_utils_fast.pyx":128 + * start_ds_idx = ds.current_index + * start_offset = ds.current_offset + * if e <= s: # <<<<<<<<<<<<<< + * end_ds_idx = start_ds_idx + * else: + */ + __pyx_t_12 = (__pyx_v_e <= __pyx_v_s); + if (__pyx_t_12) { + + /* "fairseq/data/token_block_utils_fast.pyx":129 + * start_offset = ds.current_offset + * if e <= s: + * end_ds_idx = start_ds_idx # <<<<<<<<<<<<<< + * else: + * ds.seek(e - 1) + */ + __pyx_v_end_ds_idx = __pyx_v_start_ds_idx; + + /* "fairseq/data/token_block_utils_fast.pyx":128 + * start_ds_idx = ds.current_index + * start_offset = ds.current_offset + * if e <= s: # <<<<<<<<<<<<<< + * end_ds_idx = start_ds_idx + * else: + */ + goto __pyx_L5; + } + + /* "fairseq/data/token_block_utils_fast.pyx":131 + * end_ds_idx = start_ds_idx + * else: + * ds.seek(e - 1) # <<<<<<<<<<<<<< + * end_ds_idx = ds.current_index + * block_to_dataset_index_view[i][0] = start_ds_idx # starting index in dataset + */ + /*else*/ { + __pyx_t_5 = ((struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_ds->__pyx_vtab)->seek(__pyx_v_ds, (__pyx_v_e - 1)); if (unlikely(!__pyx_t_5)) __PYX_ERR(0, 131, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":132 + * else: + * ds.seek(e - 1) + * end_ds_idx = ds.current_index # <<<<<<<<<<<<<< + * block_to_dataset_index_view[i][0] = start_ds_idx # starting index in dataset + * block_to_dataset_index_view[i][1] = start_offset # starting offset within starting index + */ + __pyx_t_10 = __pyx_v_ds->current_index; + __pyx_v_end_ds_idx = __pyx_t_10; + } + __pyx_L5:; + + /* "fairseq/data/token_block_utils_fast.pyx":133 + * ds.seek(e - 1) + * end_ds_idx = ds.current_index + * block_to_dataset_index_view[i][0] = start_ds_idx # starting index in dataset # <<<<<<<<<<<<<< + * block_to_dataset_index_view[i][1] = start_offset # starting offset within starting index + * block_to_dataset_index_view[i][2] = end_ds_idx # ending index in dataset + */ + __pyx_t_10 = __pyx_v_i; + __pyx_t_11 = 0; + *((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_block_to_dataset_index_view.data + __pyx_t_10 * __pyx_v_block_to_dataset_index_view.strides[0]) ) + __pyx_t_11 * __pyx_v_block_to_dataset_index_view.strides[1]) )) = __pyx_v_start_ds_idx; + + /* "fairseq/data/token_block_utils_fast.pyx":134 + * end_ds_idx = ds.current_index + * block_to_dataset_index_view[i][0] = start_ds_idx # starting index in dataset + * block_to_dataset_index_view[i][1] = start_offset # starting offset within starting index # <<<<<<<<<<<<<< + * block_to_dataset_index_view[i][2] = end_ds_idx # ending index in dataset + * return block_to_dataset_index + */ + __pyx_t_10 = __pyx_v_i; + __pyx_t_11 = 1; + *((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_block_to_dataset_index_view.data + __pyx_t_10 * __pyx_v_block_to_dataset_index_view.strides[0]) ) + __pyx_t_11 * __pyx_v_block_to_dataset_index_view.strides[1]) )) = __pyx_v_start_offset; + + /* "fairseq/data/token_block_utils_fast.pyx":135 + * block_to_dataset_index_view[i][0] = start_ds_idx # starting index in dataset + * block_to_dataset_index_view[i][1] = start_offset # starting offset within starting index + * block_to_dataset_index_view[i][2] = end_ds_idx # ending index in dataset # <<<<<<<<<<<<<< + * return block_to_dataset_index + * + */ + __pyx_t_10 = __pyx_v_i; + __pyx_t_11 = 2; + *((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=1 */ (( /* dim=0 */ (__pyx_v_block_to_dataset_index_view.data + __pyx_t_10 * __pyx_v_block_to_dataset_index_view.strides[0]) ) + __pyx_t_11 * __pyx_v_block_to_dataset_index_view.strides[1]) )) = __pyx_v_end_ds_idx; + } + + /* "fairseq/data/token_block_utils_fast.pyx":136 + * block_to_dataset_index_view[i][1] = start_offset # starting offset within starting index + * block_to_dataset_index_view[i][2] = end_ds_idx # ending index in dataset + * return block_to_dataset_index # <<<<<<<<<<<<<< + * + * + */ + __Pyx_XDECREF((PyObject *)__pyx_r); + __Pyx_INCREF((PyObject *)__pyx_v_block_to_dataset_index); + __pyx_r = ((PyArrayObject *)__pyx_v_block_to_dataset_index); + goto __pyx_L0; + + /* "fairseq/data/token_block_utils_fast.pyx":109 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_block_to_dataset_index_fast(np.ndarray[DTYPE_t, ndim=1] sizes, np.ndarray[DTYPE_t, ndim=2] slice_indices): # <<<<<<<<<<<<<< + * cdef DTYPE_t start_ds_idx + * cdef DTYPE_t start_offset + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __PYX_XCLEAR_MEMVIEW(&__pyx_t_7, 1); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._get_block_to_dataset_index_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_block_to_dataset_index.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XDECREF((PyObject *)__pyx_v_ds); + __Pyx_XDECREF((PyObject *)__pyx_v_block_to_dataset_index); + __PYX_XCLEAR_MEMVIEW(&__pyx_v_block_to_dataset_index_view, 1); + __PYX_XCLEAR_MEMVIEW(&__pyx_v_slice_indices_view, 1); + __Pyx_XGIVEREF((PyObject *)__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* Python wrapper */ +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_3_get_block_to_dataset_index_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_7fairseq_4data_22token_block_utils_fast_3_get_block_to_dataset_index_fast = {"_get_block_to_dataset_index_fast", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_22token_block_utils_fast_3_get_block_to_dataset_index_fast, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_3_get_block_to_dataset_index_fast(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyArrayObject *__pyx_v_sizes = 0; + PyArrayObject *__pyx_v_slice_indices = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[2] = {0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("_get_block_to_dataset_index_fast (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_sizes,&__pyx_n_s_slice_indices,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 2: values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_sizes)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 109, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_slice_indices)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 109, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("_get_block_to_dataset_index_fast", 1, 2, 2, 1); __PYX_ERR(0, 109, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "_get_block_to_dataset_index_fast") < 0)) __PYX_ERR(0, 109, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 2)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + } + __pyx_v_sizes = ((PyArrayObject *)values[0]); + __pyx_v_slice_indices = ((PyArrayObject *)values[1]); + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("_get_block_to_dataset_index_fast", 1, 2, 2, __pyx_nargs); __PYX_ERR(0, 109, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._get_block_to_dataset_index_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_sizes), __pyx_ptype_5numpy_ndarray, 1, "sizes", 0))) __PYX_ERR(0, 109, __pyx_L1_error) + if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_slice_indices), __pyx_ptype_5numpy_ndarray, 1, "slice_indices", 0))) __PYX_ERR(0, 109, __pyx_L1_error) + __pyx_r = __pyx_pf_7fairseq_4data_22token_block_utils_fast_2_get_block_to_dataset_index_fast(__pyx_self, __pyx_v_sizes, __pyx_v_slice_indices); + + /* function exit code */ + goto __pyx_L0; + __pyx_L1_error:; + __pyx_r = NULL; + __pyx_L0:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_2_get_block_to_dataset_index_fast(CYTHON_UNUSED PyObject *__pyx_self, PyArrayObject *__pyx_v_sizes, PyArrayObject *__pyx_v_slice_indices) { + __Pyx_LocalBuf_ND __pyx_pybuffernd_sizes; + __Pyx_Buffer __pyx_pybuffer_sizes; + __Pyx_LocalBuf_ND __pyx_pybuffernd_slice_indices; + __Pyx_Buffer __pyx_pybuffer_slice_indices; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("_get_block_to_dataset_index_fast", 1); + __pyx_pybuffer_sizes.pybuffer.buf = NULL; + __pyx_pybuffer_sizes.refcount = 0; + __pyx_pybuffernd_sizes.data = NULL; + __pyx_pybuffernd_sizes.rcbuffer = &__pyx_pybuffer_sizes; + __pyx_pybuffer_slice_indices.pybuffer.buf = NULL; + __pyx_pybuffer_slice_indices.refcount = 0; + __pyx_pybuffernd_slice_indices.data = NULL; + __pyx_pybuffernd_slice_indices.rcbuffer = &__pyx_pybuffer_slice_indices; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer, (PyObject*)__pyx_v_sizes, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) __PYX_ERR(0, 109, __pyx_L1_error) + } + __pyx_pybuffernd_sizes.diminfo[0].strides = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_sizes.diminfo[0].shape = __pyx_pybuffernd_sizes.rcbuffer->pybuffer.shape[0]; + { + __Pyx_BufFmt_StackElem __pyx_stack[1]; + if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer, (PyObject*)__pyx_v_slice_indices, &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, PyBUF_FORMAT| PyBUF_STRIDES, 2, 0, __pyx_stack) == -1)) __PYX_ERR(0, 109, __pyx_L1_error) + } + __pyx_pybuffernd_slice_indices.diminfo[0].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_slice_indices.diminfo[0].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[0]; __pyx_pybuffernd_slice_indices.diminfo[1].strides = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.strides[1]; __pyx_pybuffernd_slice_indices.diminfo[1].shape = __pyx_pybuffernd_slice_indices.rcbuffer->pybuffer.shape[1]; + __Pyx_XDECREF(__pyx_r); + __pyx_t_1 = ((PyObject *)__pyx_f_7fairseq_4data_22token_block_utils_fast__get_block_to_dataset_index_fast(__pyx_v_sizes, __pyx_v_slice_indices, 0)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 109, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_r = __pyx_t_1; + __pyx_t_1 = 0; + goto __pyx_L0; + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + { PyObject *__pyx_type, *__pyx_value, *__pyx_tb; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&__pyx_type, &__pyx_value, &__pyx_tb); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __Pyx_ErrRestore(__pyx_type, __pyx_value, __pyx_tb);} + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast._get_block_to_dataset_index_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + goto __pyx_L2; + __pyx_L0:; + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_sizes.rcbuffer->pybuffer); + __Pyx_SafeReleaseBuffer(&__pyx_pybuffernd_slice_indices.rcbuffer->pybuffer); + __pyx_L2:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":147 + * cdef DTYPE_t[:] sizes + * + * def __init__(self, DTYPE_t[:] sizes): # <<<<<<<<<<<<<< + * self.sizes = sizes + * self.reset() + */ + +/* Python wrapper */ +static int __pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_1__init__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds); /*proto*/ +static int __pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_1__init__(PyObject *__pyx_v_self, PyObject *__pyx_args, PyObject *__pyx_kwds) { + __Pyx_memviewslice __pyx_v_sizes = { 0, 0, { 0 }, { 0 }, { 0 } }; + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + int __pyx_r; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__init__ (wrapper)", 0); + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return -1; + #endif + __pyx_kwvalues = __Pyx_KwValues_VARARGS(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_sizes,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_VARARGS(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_VARARGS(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_sizes)) != 0)) { + (void)__Pyx_Arg_NewRef_VARARGS(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 147, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__init__") < 0)) __PYX_ERR(0, 147, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_VARARGS(__pyx_args, 0); + } + __pyx_v_sizes = __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(values[0], PyBUF_WRITABLE); if (unlikely(!__pyx_v_sizes.memview)) __PYX_ERR(0, 147, __pyx_L3_error) + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__init__", 1, 1, 1, __pyx_nargs); __PYX_ERR(0, 147, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __PYX_XCLEAR_MEMVIEW(&__pyx_v_sizes, 1); + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.DatasetSearcher.__init__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return -1; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher___init__(((struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_self), __pyx_v_sizes); + + /* function exit code */ + __PYX_XCLEAR_MEMVIEW(&__pyx_v_sizes, 1); + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_VARARGS(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static int __pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher___init__(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, __Pyx_memviewslice __pyx_v_sizes) { + int __pyx_r; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__init__", 1); + + /* "fairseq/data/token_block_utils_fast.pyx":148 + * + * def __init__(self, DTYPE_t[:] sizes): + * self.sizes = sizes # <<<<<<<<<<<<<< + * self.reset() + * + */ + __PYX_XCLEAR_MEMVIEW(&__pyx_v_self->sizes, 0); + __PYX_INC_MEMVIEW(&__pyx_v_sizes, 1); + __pyx_v_self->sizes = __pyx_v_sizes; + + /* "fairseq/data/token_block_utils_fast.pyx":149 + * def __init__(self, DTYPE_t[:] sizes): + * self.sizes = sizes + * self.reset() # <<<<<<<<<<<<<< + * + * cdef reset(self): + */ + __pyx_t_1 = ((struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_self->__pyx_vtab)->reset(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 149, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":147 + * cdef DTYPE_t[:] sizes + * + * def __init__(self, DTYPE_t[:] sizes): # <<<<<<<<<<<<<< + * self.sizes = sizes + * self.reset() + */ + + /* function exit code */ + __pyx_r = 0; + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.DatasetSearcher.__init__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = -1; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":151 + * self.reset() + * + * cdef reset(self): # <<<<<<<<<<<<<< + * self.current_offset = 0 # offset within current index in underlying dataset + * self.current_i = 0 # "flat" index + */ + +static PyObject *__pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_reset(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("reset", 1); + + /* "fairseq/data/token_block_utils_fast.pyx":152 + * + * cdef reset(self): + * self.current_offset = 0 # offset within current index in underlying dataset # <<<<<<<<<<<<<< + * self.current_i = 0 # "flat" index + * self.current_index = 0 # index in underlying dataset + */ + __pyx_v_self->current_offset = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":153 + * cdef reset(self): + * self.current_offset = 0 # offset within current index in underlying dataset + * self.current_i = 0 # "flat" index # <<<<<<<<<<<<<< + * self.current_index = 0 # index in underlying dataset + * + */ + __pyx_v_self->current_i = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":154 + * self.current_offset = 0 # offset within current index in underlying dataset + * self.current_i = 0 # "flat" index + * self.current_index = 0 # index in underlying dataset # <<<<<<<<<<<<<< + * + * @cython.boundscheck(False) + */ + __pyx_v_self->current_index = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":151 + * self.reset() + * + * cdef reset(self): # <<<<<<<<<<<<<< + * self.current_offset = 0 # offset within current index in underlying dataset + * self.current_i = 0 # "flat" index + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":159 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cdef int step(self, DTYPE_t i): # <<<<<<<<<<<<<< + * cdef DTYPE_t to_consume + * cdef DTYPE_t remaining + */ + +static int __pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_step(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_i) { + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_to_consume; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_remaining; + int __pyx_r; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + PyObject *__pyx_t_2 = NULL; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_3; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("step", 1); + + /* "fairseq/data/token_block_utils_fast.pyx":162 + * cdef DTYPE_t to_consume + * cdef DTYPE_t remaining + * if i < self.current_i: # <<<<<<<<<<<<<< + * self.reset() + * if i > self.current_i: + */ + __pyx_t_1 = (__pyx_v_i < __pyx_v_self->current_i); + if (__pyx_t_1) { + + /* "fairseq/data/token_block_utils_fast.pyx":163 + * cdef DTYPE_t remaining + * if i < self.current_i: + * self.reset() # <<<<<<<<<<<<<< + * if i > self.current_i: + * to_consume = i - self.current_i + */ + __pyx_t_2 = ((struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_self->__pyx_vtab)->reset(__pyx_v_self); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 163, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":162 + * cdef DTYPE_t to_consume + * cdef DTYPE_t remaining + * if i < self.current_i: # <<<<<<<<<<<<<< + * self.reset() + * if i > self.current_i: + */ + } + + /* "fairseq/data/token_block_utils_fast.pyx":164 + * if i < self.current_i: + * self.reset() + * if i > self.current_i: # <<<<<<<<<<<<<< + * to_consume = i - self.current_i + * remaining = self.sizes[self.current_index] - self.current_offset + */ + __pyx_t_1 = (__pyx_v_i > __pyx_v_self->current_i); + if (__pyx_t_1) { + + /* "fairseq/data/token_block_utils_fast.pyx":165 + * self.reset() + * if i > self.current_i: + * to_consume = i - self.current_i # <<<<<<<<<<<<<< + * remaining = self.sizes[self.current_index] - self.current_offset + * if remaining > to_consume: + */ + __pyx_v_to_consume = (__pyx_v_i - __pyx_v_self->current_i); + + /* "fairseq/data/token_block_utils_fast.pyx":166 + * if i > self.current_i: + * to_consume = i - self.current_i + * remaining = self.sizes[self.current_index] - self.current_offset # <<<<<<<<<<<<<< + * if remaining > to_consume: + * self.current_offset += to_consume + */ + if (unlikely(!__pyx_v_self->sizes.memview)) {PyErr_SetString(PyExc_AttributeError,"Memoryview is not initialized");__PYX_ERR(0, 166, __pyx_L1_error)} + __pyx_t_3 = __pyx_v_self->current_index; + __pyx_v_remaining = ((*((__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) ( /* dim=0 */ (__pyx_v_self->sizes.data + __pyx_t_3 * __pyx_v_self->sizes.strides[0]) ))) - __pyx_v_self->current_offset); + + /* "fairseq/data/token_block_utils_fast.pyx":167 + * to_consume = i - self.current_i + * remaining = self.sizes[self.current_index] - self.current_offset + * if remaining > to_consume: # <<<<<<<<<<<<<< + * self.current_offset += to_consume + * self.current_i += to_consume + */ + __pyx_t_1 = (__pyx_v_remaining > __pyx_v_to_consume); + if (__pyx_t_1) { + + /* "fairseq/data/token_block_utils_fast.pyx":168 + * remaining = self.sizes[self.current_index] - self.current_offset + * if remaining > to_consume: + * self.current_offset += to_consume # <<<<<<<<<<<<<< + * self.current_i += to_consume + * else: + */ + __pyx_v_self->current_offset = (__pyx_v_self->current_offset + __pyx_v_to_consume); + + /* "fairseq/data/token_block_utils_fast.pyx":169 + * if remaining > to_consume: + * self.current_offset += to_consume + * self.current_i += to_consume # <<<<<<<<<<<<<< + * else: + * assert remaining > 0 + */ + __pyx_v_self->current_i = (__pyx_v_self->current_i + __pyx_v_to_consume); + + /* "fairseq/data/token_block_utils_fast.pyx":167 + * to_consume = i - self.current_i + * remaining = self.sizes[self.current_index] - self.current_offset + * if remaining > to_consume: # <<<<<<<<<<<<<< + * self.current_offset += to_consume + * self.current_i += to_consume + */ + goto __pyx_L5; + } + + /* "fairseq/data/token_block_utils_fast.pyx":171 + * self.current_i += to_consume + * else: + * assert remaining > 0 # <<<<<<<<<<<<<< + * self.current_i += remaining + * self.current_index += 1 + */ + /*else*/ { + #ifndef CYTHON_WITHOUT_ASSERTIONS + if (unlikely(__pyx_assertions_enabled())) { + __pyx_t_1 = (__pyx_v_remaining > 0); + if (unlikely(!__pyx_t_1)) { + __Pyx_Raise(__pyx_builtin_AssertionError, 0, 0, 0); + __PYX_ERR(0, 171, __pyx_L1_error) + } + } + #else + if ((1)); else __PYX_ERR(0, 171, __pyx_L1_error) + #endif + + /* "fairseq/data/token_block_utils_fast.pyx":172 + * else: + * assert remaining > 0 + * self.current_i += remaining # <<<<<<<<<<<<<< + * self.current_index += 1 + * self.current_offset = 0 + */ + __pyx_v_self->current_i = (__pyx_v_self->current_i + __pyx_v_remaining); + + /* "fairseq/data/token_block_utils_fast.pyx":173 + * assert remaining > 0 + * self.current_i += remaining + * self.current_index += 1 # <<<<<<<<<<<<<< + * self.current_offset = 0 + * return 1 + */ + __pyx_v_self->current_index = (__pyx_v_self->current_index + 1); + + /* "fairseq/data/token_block_utils_fast.pyx":174 + * self.current_i += remaining + * self.current_index += 1 + * self.current_offset = 0 # <<<<<<<<<<<<<< + * return 1 + * return 0 + */ + __pyx_v_self->current_offset = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":175 + * self.current_index += 1 + * self.current_offset = 0 + * return 1 # <<<<<<<<<<<<<< + * return 0 + * + */ + __pyx_r = 1; + goto __pyx_L0; + } + __pyx_L5:; + + /* "fairseq/data/token_block_utils_fast.pyx":164 + * if i < self.current_i: + * self.reset() + * if i > self.current_i: # <<<<<<<<<<<<<< + * to_consume = i - self.current_i + * remaining = self.sizes[self.current_index] - self.current_offset + */ + } + + /* "fairseq/data/token_block_utils_fast.pyx":176 + * self.current_offset = 0 + * return 1 + * return 0 # <<<<<<<<<<<<<< + * + * @cython.boundscheck(False) + */ + __pyx_r = 0; + goto __pyx_L0; + + /* "fairseq/data/token_block_utils_fast.pyx":159 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cdef int step(self, DTYPE_t i): # <<<<<<<<<<<<<< + * cdef DTYPE_t to_consume + * cdef DTYPE_t remaining + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_2); + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.DatasetSearcher.step", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "fairseq/data/token_block_utils_fast.pyx":181 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cdef seek(self, DTYPE_t i): # <<<<<<<<<<<<<< + * cdef int not_done = 1 + * while not_done == 1: + */ + +static PyObject *__pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_seek(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_v_i) { + int __pyx_v_not_done; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + int __pyx_t_1; + int __pyx_t_2; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("seek", 1); + + /* "fairseq/data/token_block_utils_fast.pyx":182 + * @cython.nonecheck(False) + * cdef seek(self, DTYPE_t i): + * cdef int not_done = 1 # <<<<<<<<<<<<<< + * while not_done == 1: + * not_done = self.step(i) + */ + __pyx_v_not_done = 1; + + /* "fairseq/data/token_block_utils_fast.pyx":183 + * cdef seek(self, DTYPE_t i): + * cdef int not_done = 1 + * while not_done == 1: # <<<<<<<<<<<<<< + * not_done = self.step(i) + * assert self.current_i == i + */ + while (1) { + __pyx_t_1 = (__pyx_v_not_done == 1); + if (!__pyx_t_1) break; + + /* "fairseq/data/token_block_utils_fast.pyx":184 + * cdef int not_done = 1 + * while not_done == 1: + * not_done = self.step(i) # <<<<<<<<<<<<<< + * assert self.current_i == i + */ + __pyx_t_2 = ((struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_self->__pyx_vtab)->step(__pyx_v_self, __pyx_v_i); if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 184, __pyx_L1_error) + __pyx_v_not_done = __pyx_t_2; + } + + /* "fairseq/data/token_block_utils_fast.pyx":185 + * while not_done == 1: + * not_done = self.step(i) + * assert self.current_i == i # <<<<<<<<<<<<<< + */ + #ifndef CYTHON_WITHOUT_ASSERTIONS + if (unlikely(__pyx_assertions_enabled())) { + __pyx_t_1 = (__pyx_v_self->current_i == __pyx_v_i); + if (unlikely(!__pyx_t_1)) { + __Pyx_Raise(__pyx_builtin_AssertionError, 0, 0, 0); + __PYX_ERR(0, 185, __pyx_L1_error) + } + } + #else + if ((1)); else __PYX_ERR(0, 185, __pyx_L1_error) + #endif + + /* "fairseq/data/token_block_utils_fast.pyx":181 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cdef seek(self, DTYPE_t i): # <<<<<<<<<<<<<< + * cdef int not_done = 1 + * while not_done == 1: + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.DatasetSearcher.seek", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_3__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_3__reduce_cython__ = {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_3__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_3__reduce_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__reduce_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + if (unlikely(__pyx_nargs > 0)) { + __Pyx_RaiseArgtupleInvalid("__reduce_cython__", 1, 0, 0, __pyx_nargs); return NULL;} + if (unlikely(__pyx_kwds) && __Pyx_NumKwargs_FASTCALL(__pyx_kwds) && unlikely(!__Pyx_CheckKeywordStrings(__pyx_kwds, "__reduce_cython__", 0))) return NULL; + __pyx_r = __pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_2__reduce_cython__(((struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_self)); + + /* function exit code */ + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_2__reduce_cython__(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self) { + PyObject *__pyx_v_state = 0; + PyObject *__pyx_v__dict = 0; + int __pyx_v_use_setstate; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + int __pyx_t_6; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__reduce_cython__", 1); + + /* "(tree fragment)":5 + * cdef object _dict + * cdef bint use_setstate + * state = (self.current_i, self.current_index, self.current_offset, self.sizes) # <<<<<<<<<<<<<< + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: + */ + __pyx_t_1 = __Pyx_PyInt_From_npy_int64(__pyx_v_self->current_i); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyInt_From_npy_int64(__pyx_v_self->current_index); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_2); + __pyx_t_3 = __Pyx_PyInt_From_npy_int64(__pyx_v_self->current_offset); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + if (unlikely(!__pyx_v_self->sizes.memview)) {PyErr_SetString(PyExc_AttributeError,"Memoryview is not initialized");__PYX_ERR(1, 5, __pyx_L1_error)} + __pyx_t_4 = __pyx_memoryview_fromslice(__pyx_v_self->sizes, 1, (PyObject *(*)(char *)) __pyx_memview_get_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, (int (*)(char *, PyObject *)) __pyx_memview_set_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, 0);; if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = PyTuple_New(4); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_GIVEREF(__pyx_t_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_2); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_t_2)) __PYX_ERR(1, 5, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 2, __pyx_t_3)) __PYX_ERR(1, 5, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_4); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 3, __pyx_t_4)) __PYX_ERR(1, 5, __pyx_L1_error); + __pyx_t_1 = 0; + __pyx_t_2 = 0; + __pyx_t_3 = 0; + __pyx_t_4 = 0; + __pyx_v_state = ((PyObject*)__pyx_t_5); + __pyx_t_5 = 0; + + /* "(tree fragment)":6 + * cdef bint use_setstate + * state = (self.current_i, self.current_index, self.current_offset, self.sizes) + * _dict = getattr(self, '__dict__', None) # <<<<<<<<<<<<<< + * if _dict is not None: + * state += (_dict,) + */ + __pyx_t_5 = __Pyx_GetAttr3(((PyObject *)__pyx_v_self), __pyx_n_s_dict, Py_None); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __pyx_v__dict = __pyx_t_5; + __pyx_t_5 = 0; + + /* "(tree fragment)":7 + * state = (self.current_i, self.current_index, self.current_offset, self.sizes) + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: # <<<<<<<<<<<<<< + * state += (_dict,) + * use_setstate = True + */ + __pyx_t_6 = (__pyx_v__dict != Py_None); + if (__pyx_t_6) { + + /* "(tree fragment)":8 + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: + * state += (_dict,) # <<<<<<<<<<<<<< + * use_setstate = True + * else: + */ + __pyx_t_5 = PyTuple_New(1); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 8, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_INCREF(__pyx_v__dict); + __Pyx_GIVEREF(__pyx_v__dict); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_v__dict)) __PYX_ERR(1, 8, __pyx_L1_error); + __pyx_t_4 = PyNumber_InPlaceAdd(__pyx_v_state, __pyx_t_5); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 8, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_DECREF_SET(__pyx_v_state, ((PyObject*)__pyx_t_4)); + __pyx_t_4 = 0; + + /* "(tree fragment)":9 + * if _dict is not None: + * state += (_dict,) + * use_setstate = True # <<<<<<<<<<<<<< + * else: + * use_setstate = False + */ + __pyx_v_use_setstate = 1; + + /* "(tree fragment)":7 + * state = (self.current_i, self.current_index, self.current_offset, self.sizes) + * _dict = getattr(self, '__dict__', None) + * if _dict is not None: # <<<<<<<<<<<<<< + * state += (_dict,) + * use_setstate = True + */ + goto __pyx_L3; + } + + /* "(tree fragment)":11 + * use_setstate = True + * else: + * use_setstate = False # <<<<<<<<<<<<<< + * if use_setstate: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, None), state + */ + /*else*/ { + __pyx_v_use_setstate = 0; + } + __pyx_L3:; + + /* "(tree fragment)":12 + * else: + * use_setstate = False + * if use_setstate: # <<<<<<<<<<<<<< + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, None), state + * else: + */ + if (__pyx_v_use_setstate) { + + /* "(tree fragment)":13 + * use_setstate = False + * if use_setstate: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, None), state # <<<<<<<<<<<<<< + * else: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, state) + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_GetModuleGlobalName(__pyx_t_4, __pyx_n_s_pyx_unpickle_DatasetSearcher); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = PyTuple_New(3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_INCREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + __Pyx_GIVEREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 0, ((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))))) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_147225413); + __Pyx_GIVEREF(__pyx_int_147225413); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_147225413)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(Py_None); + __Pyx_GIVEREF(Py_None); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 2, Py_None)) __PYX_ERR(1, 13, __pyx_L1_error); + __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 13, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_GIVEREF(__pyx_t_4); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_4)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_5); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_5)) __PYX_ERR(1, 13, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_state); + __Pyx_GIVEREF(__pyx_v_state); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_v_state)) __PYX_ERR(1, 13, __pyx_L1_error); + __pyx_t_4 = 0; + __pyx_t_5 = 0; + __pyx_r = __pyx_t_3; + __pyx_t_3 = 0; + goto __pyx_L0; + + /* "(tree fragment)":12 + * else: + * use_setstate = False + * if use_setstate: # <<<<<<<<<<<<<< + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, None), state + * else: + */ + } + + /* "(tree fragment)":15 + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, None), state + * else: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, state) # <<<<<<<<<<<<<< + * def __setstate_cython__(self, __pyx_state): + * __pyx_unpickle_DatasetSearcher__set_state(self, __pyx_state) + */ + /*else*/ { + __Pyx_XDECREF(__pyx_r); + __Pyx_GetModuleGlobalName(__pyx_t_3, __pyx_n_s_pyx_unpickle_DatasetSearcher); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_5 = PyTuple_New(3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_INCREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + __Pyx_GIVEREF(((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self)))); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 0, ((PyObject *)Py_TYPE(((PyObject *)__pyx_v_self))))) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_INCREF(__pyx_int_147225413); + __Pyx_GIVEREF(__pyx_int_147225413); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 1, __pyx_int_147225413)) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_INCREF(__pyx_v_state); + __Pyx_GIVEREF(__pyx_v_state); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_5, 2, __pyx_v_state)) __PYX_ERR(1, 15, __pyx_L1_error); + __pyx_t_4 = PyTuple_New(2); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_GIVEREF(__pyx_t_3); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_t_3)) __PYX_ERR(1, 15, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_t_5); + if (__Pyx_PyTuple_SET_ITEM(__pyx_t_4, 1, __pyx_t_5)) __PYX_ERR(1, 15, __pyx_L1_error); + __pyx_t_3 = 0; + __pyx_t_5 = 0; + __pyx_r = __pyx_t_4; + __pyx_t_4 = 0; + goto __pyx_L0; + } + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_2); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.DatasetSearcher.__reduce_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v_state); + __Pyx_XDECREF(__pyx_v__dict); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":16 + * else: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_DatasetSearcher__set_state(self, __pyx_state) + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_5__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_5__setstate_cython__ = {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_5__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_5__setstate_cython__(PyObject *__pyx_v_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[1] = {0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__setstate_cython__ (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 16, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__setstate_cython__") < 0)) __PYX_ERR(1, 16, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 1)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + } + __pyx_v___pyx_state = values[0]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__setstate_cython__", 1, 1, 1, __pyx_nargs); __PYX_ERR(1, 16, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.DatasetSearcher.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_4__setstate_cython__(((struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v_self), __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_4__setstate_cython__(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v_self, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__setstate_cython__", 1); + + /* "(tree fragment)":17 + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, state) + * def __setstate_cython__(self, __pyx_state): + * __pyx_unpickle_DatasetSearcher__set_state(self, __pyx_state) # <<<<<<<<<<<<<< + */ + if (!(likely(PyTuple_CheckExact(__pyx_v___pyx_state))||((__pyx_v___pyx_state) == Py_None) || __Pyx_RaiseUnexpectedTypeError("tuple", __pyx_v___pyx_state))) __PYX_ERR(1, 17, __pyx_L1_error) + __pyx_t_1 = __pyx_f_7fairseq_4data_22token_block_utils_fast___pyx_unpickle_DatasetSearcher__set_state(__pyx_v_self, ((PyObject*)__pyx_v___pyx_state)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 17, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":16 + * else: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_DatasetSearcher__set_state(self, __pyx_state) + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.DatasetSearcher.__setstate_cython__", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":1 + * def __pyx_unpickle_DatasetSearcher(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + +/* Python wrapper */ +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_5__pyx_unpickle_DatasetSearcher(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +); /*proto*/ +static PyMethodDef __pyx_mdef_7fairseq_4data_22token_block_utils_fast_5__pyx_unpickle_DatasetSearcher = {"__pyx_unpickle_DatasetSearcher", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_22token_block_utils_fast_5__pyx_unpickle_DatasetSearcher, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}; +static PyObject *__pyx_pw_7fairseq_4data_22token_block_utils_fast_5__pyx_unpickle_DatasetSearcher(PyObject *__pyx_self, +#if CYTHON_METH_FASTCALL +PyObject *const *__pyx_args, Py_ssize_t __pyx_nargs, PyObject *__pyx_kwds +#else +PyObject *__pyx_args, PyObject *__pyx_kwds +#endif +) { + PyObject *__pyx_v___pyx_type = 0; + long __pyx_v___pyx_checksum; + PyObject *__pyx_v___pyx_state = 0; + #if !CYTHON_METH_FASTCALL + CYTHON_UNUSED Py_ssize_t __pyx_nargs; + #endif + CYTHON_UNUSED PyObject *const *__pyx_kwvalues; + PyObject* values[3] = {0,0,0}; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + PyObject *__pyx_r = 0; + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__pyx_unpickle_DatasetSearcher (wrapper)", 0); + #if !CYTHON_METH_FASTCALL + #if CYTHON_ASSUME_SAFE_MACROS + __pyx_nargs = PyTuple_GET_SIZE(__pyx_args); + #else + __pyx_nargs = PyTuple_Size(__pyx_args); if (unlikely(__pyx_nargs < 0)) return NULL; + #endif + #endif + __pyx_kwvalues = __Pyx_KwValues_FASTCALL(__pyx_args, __pyx_nargs); + { + PyObject **__pyx_pyargnames[] = {&__pyx_n_s_pyx_type,&__pyx_n_s_pyx_checksum,&__pyx_n_s_pyx_state,0}; + if (__pyx_kwds) { + Py_ssize_t kw_args; + switch (__pyx_nargs) { + case 3: values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + CYTHON_FALLTHROUGH; + case 2: values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + CYTHON_FALLTHROUGH; + case 1: values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + CYTHON_FALLTHROUGH; + case 0: break; + default: goto __pyx_L5_argtuple_error; + } + kw_args = __Pyx_NumKwargs_FASTCALL(__pyx_kwds); + switch (__pyx_nargs) { + case 0: + if (likely((values[0] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_type)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[0]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else goto __pyx_L5_argtuple_error; + CYTHON_FALLTHROUGH; + case 1: + if (likely((values[1] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_checksum)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[1]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_DatasetSearcher", 1, 3, 3, 1); __PYX_ERR(1, 1, __pyx_L3_error) + } + CYTHON_FALLTHROUGH; + case 2: + if (likely((values[2] = __Pyx_GetKwValue_FASTCALL(__pyx_kwds, __pyx_kwvalues, __pyx_n_s_pyx_state)) != 0)) { + (void)__Pyx_Arg_NewRef_FASTCALL(values[2]); + kw_args--; + } + else if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + else { + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_DatasetSearcher", 1, 3, 3, 2); __PYX_ERR(1, 1, __pyx_L3_error) + } + } + if (unlikely(kw_args > 0)) { + const Py_ssize_t kwd_pos_args = __pyx_nargs; + if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_kwvalues, __pyx_pyargnames, 0, values + 0, kwd_pos_args, "__pyx_unpickle_DatasetSearcher") < 0)) __PYX_ERR(1, 1, __pyx_L3_error) + } + } else if (unlikely(__pyx_nargs != 3)) { + goto __pyx_L5_argtuple_error; + } else { + values[0] = __Pyx_Arg_FASTCALL(__pyx_args, 0); + values[1] = __Pyx_Arg_FASTCALL(__pyx_args, 1); + values[2] = __Pyx_Arg_FASTCALL(__pyx_args, 2); + } + __pyx_v___pyx_type = values[0]; + __pyx_v___pyx_checksum = __Pyx_PyInt_As_long(values[1]); if (unlikely((__pyx_v___pyx_checksum == (long)-1) && PyErr_Occurred())) __PYX_ERR(1, 1, __pyx_L3_error) + __pyx_v___pyx_state = values[2]; + } + goto __pyx_L6_skip; + __pyx_L5_argtuple_error:; + __Pyx_RaiseArgtupleInvalid("__pyx_unpickle_DatasetSearcher", 1, 3, 3, __pyx_nargs); __PYX_ERR(1, 1, __pyx_L3_error) + __pyx_L6_skip:; + goto __pyx_L4_argument_unpacking_done; + __pyx_L3_error:; + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.__pyx_unpickle_DatasetSearcher", __pyx_clineno, __pyx_lineno, __pyx_filename); + __Pyx_RefNannyFinishContext(); + return NULL; + __pyx_L4_argument_unpacking_done:; + __pyx_r = __pyx_pf_7fairseq_4data_22token_block_utils_fast_4__pyx_unpickle_DatasetSearcher(__pyx_self, __pyx_v___pyx_type, __pyx_v___pyx_checksum, __pyx_v___pyx_state); + + /* function exit code */ + { + Py_ssize_t __pyx_temp; + for (__pyx_temp=0; __pyx_temp < (Py_ssize_t)(sizeof(values)/sizeof(values[0])); ++__pyx_temp) { + __Pyx_Arg_XDECREF_FASTCALL(values[__pyx_temp]); + } + } + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +static PyObject *__pyx_pf_7fairseq_4data_22token_block_utils_fast_4__pyx_unpickle_DatasetSearcher(CYTHON_UNUSED PyObject *__pyx_self, PyObject *__pyx_v___pyx_type, long __pyx_v___pyx_checksum, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_v___pyx_PickleError = 0; + PyObject *__pyx_v___pyx_result = 0; + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_t_2; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + unsigned int __pyx_t_5; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__pyx_unpickle_DatasetSearcher", 1); + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x8c67b45, 0x2e2dd22, 0x6632805): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum + */ + __pyx_t_1 = __Pyx_PyInt_From_long(__pyx_v___pyx_checksum); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = (__Pyx_PySequence_ContainsTF(__pyx_t_1, __pyx_tuple__14, Py_NE)); if (unlikely((__pyx_t_2 < 0))) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + if (__pyx_t_2) { + + /* "(tree fragment)":5 + * cdef object __pyx_result + * if __pyx_checksum not in (0x8c67b45, 0x2e2dd22, 0x6632805): + * from pickle import PickleError as __pyx_PickleError # <<<<<<<<<<<<<< + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum + * __pyx_result = DatasetSearcher.__new__(__pyx_type) + */ + __pyx_t_1 = PyList_New(1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_n_s_PickleError); + __Pyx_GIVEREF(__pyx_n_s_PickleError); + if (__Pyx_PyList_SET_ITEM(__pyx_t_1, 0, __pyx_n_s_PickleError)) __PYX_ERR(1, 5, __pyx_L1_error); + __pyx_t_3 = __Pyx_Import(__pyx_n_s_pickle, __pyx_t_1, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = __Pyx_ImportFrom(__pyx_t_3, __pyx_n_s_PickleError); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 5, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_INCREF(__pyx_t_1); + __pyx_v___pyx_PickleError = __pyx_t_1; + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + + /* "(tree fragment)":6 + * if __pyx_checksum not in (0x8c67b45, 0x2e2dd22, 0x6632805): + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum # <<<<<<<<<<<<<< + * __pyx_result = DatasetSearcher.__new__(__pyx_type) + * if __pyx_state is not None: + */ + __pyx_t_3 = __Pyx_PyInt_From_long(__pyx_v___pyx_checksum); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_1 = __Pyx_PyString_Format(__pyx_kp_s_Incompatible_checksums_0x_x_vs_0_2, __pyx_t_3); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 6, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_Raise(__pyx_v___pyx_PickleError, __pyx_t_1, 0, 0); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __PYX_ERR(1, 6, __pyx_L1_error) + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x8c67b45, 0x2e2dd22, 0x6632805): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum + */ + } + + /* "(tree fragment)":7 + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum + * __pyx_result = DatasetSearcher.__new__(__pyx_type) # <<<<<<<<<<<<<< + * if __pyx_state is not None: + * __pyx_unpickle_DatasetSearcher__set_state( __pyx_result, __pyx_state) + */ + __pyx_t_3 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher), __pyx_n_s_new); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_3); + __pyx_t_4 = NULL; + __pyx_t_5 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_3))) { + __pyx_t_4 = PyMethod_GET_SELF(__pyx_t_3); + if (likely(__pyx_t_4)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_3); + __Pyx_INCREF(__pyx_t_4); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_3, function); + __pyx_t_5 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_4, __pyx_v___pyx_type}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_3, __pyx_callargs+1-__pyx_t_5, 1+__pyx_t_5); + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0; + } + __pyx_v___pyx_result = __pyx_t_1; + __pyx_t_1 = 0; + + /* "(tree fragment)":8 + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum + * __pyx_result = DatasetSearcher.__new__(__pyx_type) + * if __pyx_state is not None: # <<<<<<<<<<<<<< + * __pyx_unpickle_DatasetSearcher__set_state( __pyx_result, __pyx_state) + * return __pyx_result + */ + __pyx_t_2 = (__pyx_v___pyx_state != Py_None); + if (__pyx_t_2) { + + /* "(tree fragment)":9 + * __pyx_result = DatasetSearcher.__new__(__pyx_type) + * if __pyx_state is not None: + * __pyx_unpickle_DatasetSearcher__set_state( __pyx_result, __pyx_state) # <<<<<<<<<<<<<< + * return __pyx_result + * cdef __pyx_unpickle_DatasetSearcher__set_state(DatasetSearcher __pyx_result, tuple __pyx_state): + */ + if (!(likely(PyTuple_CheckExact(__pyx_v___pyx_state))||((__pyx_v___pyx_state) == Py_None) || __Pyx_RaiseUnexpectedTypeError("tuple", __pyx_v___pyx_state))) __PYX_ERR(1, 9, __pyx_L1_error) + __pyx_t_1 = __pyx_f_7fairseq_4data_22token_block_utils_fast___pyx_unpickle_DatasetSearcher__set_state(((struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)__pyx_v___pyx_result), ((PyObject*)__pyx_v___pyx_state)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":8 + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum + * __pyx_result = DatasetSearcher.__new__(__pyx_type) + * if __pyx_state is not None: # <<<<<<<<<<<<<< + * __pyx_unpickle_DatasetSearcher__set_state( __pyx_result, __pyx_state) + * return __pyx_result + */ + } + + /* "(tree fragment)":10 + * if __pyx_state is not None: + * __pyx_unpickle_DatasetSearcher__set_state( __pyx_result, __pyx_state) + * return __pyx_result # <<<<<<<<<<<<<< + * cdef __pyx_unpickle_DatasetSearcher__set_state(DatasetSearcher __pyx_result, tuple __pyx_state): + * __pyx_result.current_i = __pyx_state[0]; __pyx_result.current_index = __pyx_state[1]; __pyx_result.current_offset = __pyx_state[2]; __pyx_result.sizes = __pyx_state[3] + */ + __Pyx_XDECREF(__pyx_r); + __Pyx_INCREF(__pyx_v___pyx_result); + __pyx_r = __pyx_v___pyx_result; + goto __pyx_L0; + + /* "(tree fragment)":1 + * def __pyx_unpickle_DatasetSearcher(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + + /* function exit code */ + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_XDECREF(__pyx_t_3); + __Pyx_XDECREF(__pyx_t_4); + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.__pyx_unpickle_DatasetSearcher", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = NULL; + __pyx_L0:; + __Pyx_XDECREF(__pyx_v___pyx_PickleError); + __Pyx_XDECREF(__pyx_v___pyx_result); + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} + +/* "(tree fragment)":11 + * __pyx_unpickle_DatasetSearcher__set_state( __pyx_result, __pyx_state) + * return __pyx_result + * cdef __pyx_unpickle_DatasetSearcher__set_state(DatasetSearcher __pyx_result, tuple __pyx_state): # <<<<<<<<<<<<<< + * __pyx_result.current_i = __pyx_state[0]; __pyx_result.current_index = __pyx_state[1]; __pyx_result.current_offset = __pyx_state[2]; __pyx_result.sizes = __pyx_state[3] + * if len(__pyx_state) > 4 and hasattr(__pyx_result, '__dict__'): + */ + +static PyObject *__pyx_f_7fairseq_4data_22token_block_utils_fast___pyx_unpickle_DatasetSearcher__set_state(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *__pyx_v___pyx_result, PyObject *__pyx_v___pyx_state) { + PyObject *__pyx_r = NULL; + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t __pyx_t_2; + __Pyx_memviewslice __pyx_t_3 = { 0, 0, { 0 }, { 0 }, { 0 } }; + int __pyx_t_4; + Py_ssize_t __pyx_t_5; + int __pyx_t_6; + PyObject *__pyx_t_7 = NULL; + PyObject *__pyx_t_8 = NULL; + PyObject *__pyx_t_9 = NULL; + unsigned int __pyx_t_10; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__pyx_unpickle_DatasetSearcher__set_state", 1); + + /* "(tree fragment)":12 + * return __pyx_result + * cdef __pyx_unpickle_DatasetSearcher__set_state(DatasetSearcher __pyx_result, tuple __pyx_state): + * __pyx_result.current_i = __pyx_state[0]; __pyx_result.current_index = __pyx_state[1]; __pyx_result.current_offset = __pyx_state[2]; __pyx_result.sizes = __pyx_state[3] # <<<<<<<<<<<<<< + * if len(__pyx_state) > 4 and hasattr(__pyx_result, '__dict__'): + * __pyx_result.__dict__.update(__pyx_state[4]) + */ + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 12, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyInt_As_npy_int64(__pyx_t_1); if (unlikely((__pyx_t_2 == ((npy_int64)-1)) && PyErr_Occurred())) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_v___pyx_result->current_i = __pyx_t_2; + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 12, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 1, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyInt_As_npy_int64(__pyx_t_1); if (unlikely((__pyx_t_2 == ((npy_int64)-1)) && PyErr_Occurred())) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_v___pyx_result->current_index = __pyx_t_2; + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 12, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 2, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_2 = __Pyx_PyInt_As_npy_int64(__pyx_t_1); if (unlikely((__pyx_t_2 == ((npy_int64)-1)) && PyErr_Occurred())) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_v___pyx_result->current_offset = __pyx_t_2; + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 12, __pyx_L1_error) + } + __pyx_t_1 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 3, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_t_3 = __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(__pyx_t_1, PyBUF_WRITABLE); if (unlikely(!__pyx_t_3.memview)) __PYX_ERR(1, 12, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __PYX_XCLEAR_MEMVIEW(&__pyx_v___pyx_result->sizes, 0); + __pyx_v___pyx_result->sizes = __pyx_t_3; + __pyx_t_3.memview = NULL; + __pyx_t_3.data = NULL; + + /* "(tree fragment)":13 + * cdef __pyx_unpickle_DatasetSearcher__set_state(DatasetSearcher __pyx_result, tuple __pyx_state): + * __pyx_result.current_i = __pyx_state[0]; __pyx_result.current_index = __pyx_state[1]; __pyx_result.current_offset = __pyx_state[2]; __pyx_result.sizes = __pyx_state[3] + * if len(__pyx_state) > 4 and hasattr(__pyx_result, '__dict__'): # <<<<<<<<<<<<<< + * __pyx_result.__dict__.update(__pyx_state[4]) + */ + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()"); + __PYX_ERR(1, 13, __pyx_L1_error) + } + __pyx_t_5 = __Pyx_PyTuple_GET_SIZE(__pyx_v___pyx_state); if (unlikely(__pyx_t_5 == ((Py_ssize_t)-1))) __PYX_ERR(1, 13, __pyx_L1_error) + __pyx_t_6 = (__pyx_t_5 > 4); + if (__pyx_t_6) { + } else { + __pyx_t_4 = __pyx_t_6; + goto __pyx_L4_bool_binop_done; + } + __pyx_t_6 = __Pyx_HasAttr(((PyObject *)__pyx_v___pyx_result), __pyx_n_s_dict); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 13, __pyx_L1_error) + __pyx_t_4 = __pyx_t_6; + __pyx_L4_bool_binop_done:; + if (__pyx_t_4) { + + /* "(tree fragment)":14 + * __pyx_result.current_i = __pyx_state[0]; __pyx_result.current_index = __pyx_state[1]; __pyx_result.current_offset = __pyx_state[2]; __pyx_result.sizes = __pyx_state[3] + * if len(__pyx_state) > 4 and hasattr(__pyx_result, '__dict__'): + * __pyx_result.__dict__.update(__pyx_state[4]) # <<<<<<<<<<<<<< + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v___pyx_result), __pyx_n_s_dict); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_8 = __Pyx_PyObject_GetAttrStr(__pyx_t_7, __pyx_n_s_update); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_8); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + if (unlikely(__pyx_v___pyx_state == Py_None)) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not subscriptable"); + __PYX_ERR(1, 14, __pyx_L1_error) + } + __pyx_t_7 = __Pyx_GetItemInt_Tuple(__pyx_v___pyx_state, 4, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_9 = NULL; + __pyx_t_10 = 0; + #if CYTHON_UNPACK_METHODS + if (likely(PyMethod_Check(__pyx_t_8))) { + __pyx_t_9 = PyMethod_GET_SELF(__pyx_t_8); + if (likely(__pyx_t_9)) { + PyObject* function = PyMethod_GET_FUNCTION(__pyx_t_8); + __Pyx_INCREF(__pyx_t_9); + __Pyx_INCREF(function); + __Pyx_DECREF_SET(__pyx_t_8, function); + __pyx_t_10 = 1; + } + } + #endif + { + PyObject *__pyx_callargs[2] = {__pyx_t_9, __pyx_t_7}; + __pyx_t_1 = __Pyx_PyObject_FastCall(__pyx_t_8, __pyx_callargs+1-__pyx_t_10, 1+__pyx_t_10); + __Pyx_XDECREF(__pyx_t_9); __pyx_t_9 = 0; + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 14, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0; + } + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + + /* "(tree fragment)":13 + * cdef __pyx_unpickle_DatasetSearcher__set_state(DatasetSearcher __pyx_result, tuple __pyx_state): + * __pyx_result.current_i = __pyx_state[0]; __pyx_result.current_index = __pyx_state[1]; __pyx_result.current_offset = __pyx_state[2]; __pyx_result.sizes = __pyx_state[3] + * if len(__pyx_state) > 4 and hasattr(__pyx_result, '__dict__'): # <<<<<<<<<<<<<< + * __pyx_result.__dict__.update(__pyx_state[4]) + */ + } + + /* "(tree fragment)":11 + * __pyx_unpickle_DatasetSearcher__set_state( __pyx_result, __pyx_state) + * return __pyx_result + * cdef __pyx_unpickle_DatasetSearcher__set_state(DatasetSearcher __pyx_result, tuple __pyx_state): # <<<<<<<<<<<<<< + * __pyx_result.current_i = __pyx_state[0]; __pyx_result.current_index = __pyx_state[1]; __pyx_result.current_offset = __pyx_state[2]; __pyx_result.sizes = __pyx_state[3] + * if len(__pyx_state) > 4 and hasattr(__pyx_result, '__dict__'): + */ + + /* function exit code */ + __pyx_r = Py_None; __Pyx_INCREF(Py_None); + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __PYX_XCLEAR_MEMVIEW(&__pyx_t_3, 1); + __Pyx_XDECREF(__pyx_t_7); + __Pyx_XDECREF(__pyx_t_8); + __Pyx_XDECREF(__pyx_t_9); + __Pyx_AddTraceback("fairseq.data.token_block_utils_fast.__pyx_unpickle_DatasetSearcher__set_state", __pyx_clineno, __pyx_lineno, __pyx_filename); + __pyx_r = 0; + __pyx_L0:; + __Pyx_XGIVEREF(__pyx_r); + __Pyx_RefNannyFinishContext(); + return __pyx_r; +} +static struct __pyx_vtabstruct_7fairseq_4data_22token_block_utils_fast_DatasetSearcher __pyx_vtable_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; + +static PyObject *__pyx_tp_new_7fairseq_4data_22token_block_utils_fast_DatasetSearcher(PyTypeObject *t, CYTHON_UNUSED PyObject *a, CYTHON_UNUSED PyObject *k) { + struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *p; + PyObject *o; + #if CYTHON_COMPILING_IN_LIMITED_API + allocfunc alloc_func = (allocfunc)PyType_GetSlot(t, Py_tp_alloc); + o = alloc_func(t, 0); + #else + if (likely(!__Pyx_PyType_HasFeature(t, Py_TPFLAGS_IS_ABSTRACT))) { + o = (*t->tp_alloc)(t, 0); + } else { + o = (PyObject *) PyBaseObject_Type.tp_new(t, __pyx_empty_tuple, 0); + } + if (unlikely(!o)) return 0; + #endif + p = ((struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)o); + p->__pyx_vtab = __pyx_vtabptr_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; + p->sizes.data = NULL; + p->sizes.memview = NULL; + return o; +} + +static void __pyx_tp_dealloc_7fairseq_4data_22token_block_utils_fast_DatasetSearcher(PyObject *o) { + struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *p = (struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && (!PyType_IS_GC(Py_TYPE(o)) || !__Pyx_PyObject_GC_IsFinalized(o))) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc_7fairseq_4data_22token_block_utils_fast_DatasetSearcher) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + __PYX_XCLEAR_MEMVIEW(&p->sizes, 1); + p->sizes.memview = NULL; p->sizes.data = NULL; + #if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + (*Py_TYPE(o)->tp_free)(o); + #else + { + freefunc tp_free = (freefunc)PyType_GetSlot(Py_TYPE(o), Py_tp_free); + if (tp_free) tp_free(o); + } + #endif +} + +static PyMethodDef __pyx_methods_7fairseq_4data_22token_block_utils_fast_DatasetSearcher[] = { + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_3__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_5__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +static PyType_Slot __pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc_7fairseq_4data_22token_block_utils_fast_DatasetSearcher}, + {Py_tp_doc, (void *)PyDoc_STR("Helper for mapping \"flat\" indices to indices and offsets in an\n underlying dataset.")}, + {Py_tp_methods, (void *)__pyx_methods_7fairseq_4data_22token_block_utils_fast_DatasetSearcher}, + {Py_tp_init, (void *)__pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_1__init__}, + {Py_tp_new, (void *)__pyx_tp_new_7fairseq_4data_22token_block_utils_fast_DatasetSearcher}, + {0, 0}, +}; +static PyType_Spec __pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher_spec = { + "fairseq.data.token_block_utils_fast.DatasetSearcher", + sizeof(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE, + __pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher_slots, +}; +#else + +static PyTypeObject __pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.token_block_utils_fast.""DatasetSearcher", /*tp_name*/ + sizeof(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc_7fairseq_4data_22token_block_utils_fast_DatasetSearcher, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + 0, /*tp_repr*/ + 0, /*tp_as_number*/ + 0, /*tp_as_sequence*/ + 0, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + 0, /*tp_str*/ + 0, /*tp_getattro*/ + 0, /*tp_setattro*/ + 0, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE, /*tp_flags*/ + PyDoc_STR("Helper for mapping \"flat\" indices to indices and offsets in an\n underlying dataset."), /*tp_doc*/ + 0, /*tp_traverse*/ + 0, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods_7fairseq_4data_22token_block_utils_fast_DatasetSearcher, /*tp_methods*/ + 0, /*tp_members*/ + 0, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + __pyx_pw_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_1__init__, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new_7fairseq_4data_22token_block_utils_fast_DatasetSearcher, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif +static struct __pyx_vtabstruct_array __pyx_vtable_array; + +static PyObject *__pyx_tp_new_array(PyTypeObject *t, PyObject *a, PyObject *k) { + struct __pyx_array_obj *p; + PyObject *o; + #if CYTHON_COMPILING_IN_LIMITED_API + allocfunc alloc_func = (allocfunc)PyType_GetSlot(t, Py_tp_alloc); + o = alloc_func(t, 0); + #else + if (likely(!__Pyx_PyType_HasFeature(t, Py_TPFLAGS_IS_ABSTRACT))) { + o = (*t->tp_alloc)(t, 0); + } else { + o = (PyObject *) PyBaseObject_Type.tp_new(t, __pyx_empty_tuple, 0); + } + if (unlikely(!o)) return 0; + #endif + p = ((struct __pyx_array_obj *)o); + p->__pyx_vtab = __pyx_vtabptr_array; + p->mode = ((PyObject*)Py_None); Py_INCREF(Py_None); + p->_format = ((PyObject*)Py_None); Py_INCREF(Py_None); + if (unlikely(__pyx_array___cinit__(o, a, k) < 0)) goto bad; + return o; + bad: + Py_DECREF(o); o = 0; + return NULL; +} + +static void __pyx_tp_dealloc_array(PyObject *o) { + struct __pyx_array_obj *p = (struct __pyx_array_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && (!PyType_IS_GC(Py_TYPE(o)) || !__Pyx_PyObject_GC_IsFinalized(o))) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc_array) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + { + PyObject *etype, *eval, *etb; + PyErr_Fetch(&etype, &eval, &etb); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1); + __pyx_array___dealloc__(o); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1); + PyErr_Restore(etype, eval, etb); + } + Py_CLEAR(p->mode); + Py_CLEAR(p->_format); + #if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + (*Py_TYPE(o)->tp_free)(o); + #else + { + freefunc tp_free = (freefunc)PyType_GetSlot(Py_TYPE(o), Py_tp_free); + if (tp_free) tp_free(o); + } + #endif +} +static PyObject *__pyx_sq_item_array(PyObject *o, Py_ssize_t i) { + PyObject *r; + PyObject *x = PyInt_FromSsize_t(i); if(!x) return 0; + r = Py_TYPE(o)->tp_as_mapping->mp_subscript(o, x); + Py_DECREF(x); + return r; +} + +static int __pyx_mp_ass_subscript_array(PyObject *o, PyObject *i, PyObject *v) { + if (v) { + return __pyx_array___setitem__(o, i, v); + } + else { + __Pyx_TypeName o_type_name; + o_type_name = __Pyx_PyType_GetName(Py_TYPE(o)); + PyErr_Format(PyExc_NotImplementedError, + "Subscript deletion not supported by " __Pyx_FMT_TYPENAME, o_type_name); + __Pyx_DECREF_TypeName(o_type_name); + return -1; + } +} + +static PyObject *__pyx_tp_getattro_array(PyObject *o, PyObject *n) { + PyObject *v = __Pyx_PyObject_GenericGetAttr(o, n); + if (!v && PyErr_ExceptionMatches(PyExc_AttributeError)) { + PyErr_Clear(); + v = __pyx_array___getattr__(o, n); + } + return v; +} + +static PyObject *__pyx_getprop___pyx_array_memview(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_5array_7memview_1__get__(o); +} + +static PyMethodDef __pyx_methods_array[] = { + {"__getattr__", (PyCFunction)__pyx_array___getattr__, METH_O|METH_COEXIST, 0}, + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_array_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_array_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; + +static struct PyGetSetDef __pyx_getsets_array[] = { + {(char *)"memview", __pyx_getprop___pyx_array_memview, 0, (char *)0, 0}, + {0, 0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +#if !CYTHON_COMPILING_IN_LIMITED_API + +static PyBufferProcs __pyx_tp_as_buffer_array = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_array_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; +#endif +static PyType_Slot __pyx_type___pyx_array_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc_array}, + {Py_sq_length, (void *)__pyx_array___len__}, + {Py_sq_item, (void *)__pyx_sq_item_array}, + {Py_mp_length, (void *)__pyx_array___len__}, + {Py_mp_subscript, (void *)__pyx_array___getitem__}, + {Py_mp_ass_subscript, (void *)__pyx_mp_ass_subscript_array}, + {Py_tp_getattro, (void *)__pyx_tp_getattro_array}, + #if defined(Py_bf_getbuffer) + {Py_bf_getbuffer, (void *)__pyx_array_getbuffer}, + #endif + {Py_tp_methods, (void *)__pyx_methods_array}, + {Py_tp_getset, (void *)__pyx_getsets_array}, + {Py_tp_new, (void *)__pyx_tp_new_array}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_array_spec = { + "fairseq.data.token_block_utils_fast.array", + sizeof(struct __pyx_array_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_SEQUENCE, + __pyx_type___pyx_array_slots, +}; +#else + +static PySequenceMethods __pyx_tp_as_sequence_array = { + __pyx_array___len__, /*sq_length*/ + 0, /*sq_concat*/ + 0, /*sq_repeat*/ + __pyx_sq_item_array, /*sq_item*/ + 0, /*sq_slice*/ + 0, /*sq_ass_item*/ + 0, /*sq_ass_slice*/ + 0, /*sq_contains*/ + 0, /*sq_inplace_concat*/ + 0, /*sq_inplace_repeat*/ +}; + +static PyMappingMethods __pyx_tp_as_mapping_array = { + __pyx_array___len__, /*mp_length*/ + __pyx_array___getitem__, /*mp_subscript*/ + __pyx_mp_ass_subscript_array, /*mp_ass_subscript*/ +}; + +static PyBufferProcs __pyx_tp_as_buffer_array = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_array_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; + +static PyTypeObject __pyx_type___pyx_array = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.token_block_utils_fast.""array", /*tp_name*/ + sizeof(struct __pyx_array_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc_array, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + 0, /*tp_repr*/ + 0, /*tp_as_number*/ + &__pyx_tp_as_sequence_array, /*tp_as_sequence*/ + &__pyx_tp_as_mapping_array, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + 0, /*tp_str*/ + __pyx_tp_getattro_array, /*tp_getattro*/ + 0, /*tp_setattro*/ + &__pyx_tp_as_buffer_array, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_SEQUENCE, /*tp_flags*/ + 0, /*tp_doc*/ + 0, /*tp_traverse*/ + 0, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods_array, /*tp_methods*/ + 0, /*tp_members*/ + __pyx_getsets_array, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + 0, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new_array, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif + +static PyObject *__pyx_tp_new_Enum(PyTypeObject *t, CYTHON_UNUSED PyObject *a, CYTHON_UNUSED PyObject *k) { + struct __pyx_MemviewEnum_obj *p; + PyObject *o; + #if CYTHON_COMPILING_IN_LIMITED_API + allocfunc alloc_func = (allocfunc)PyType_GetSlot(t, Py_tp_alloc); + o = alloc_func(t, 0); + #else + if (likely(!__Pyx_PyType_HasFeature(t, Py_TPFLAGS_IS_ABSTRACT))) { + o = (*t->tp_alloc)(t, 0); + } else { + o = (PyObject *) PyBaseObject_Type.tp_new(t, __pyx_empty_tuple, 0); + } + if (unlikely(!o)) return 0; + #endif + p = ((struct __pyx_MemviewEnum_obj *)o); + p->name = Py_None; Py_INCREF(Py_None); + return o; +} + +static void __pyx_tp_dealloc_Enum(PyObject *o) { + struct __pyx_MemviewEnum_obj *p = (struct __pyx_MemviewEnum_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && !__Pyx_PyObject_GC_IsFinalized(o)) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc_Enum) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + PyObject_GC_UnTrack(o); + Py_CLEAR(p->name); + #if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + (*Py_TYPE(o)->tp_free)(o); + #else + { + freefunc tp_free = (freefunc)PyType_GetSlot(Py_TYPE(o), Py_tp_free); + if (tp_free) tp_free(o); + } + #endif +} + +static int __pyx_tp_traverse_Enum(PyObject *o, visitproc v, void *a) { + int e; + struct __pyx_MemviewEnum_obj *p = (struct __pyx_MemviewEnum_obj *)o; + if (p->name) { + e = (*v)(p->name, a); if (e) return e; + } + return 0; +} + +static int __pyx_tp_clear_Enum(PyObject *o) { + PyObject* tmp; + struct __pyx_MemviewEnum_obj *p = (struct __pyx_MemviewEnum_obj *)o; + tmp = ((PyObject*)p->name); + p->name = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + return 0; +} + +static PyObject *__pyx_specialmethod___pyx_MemviewEnum___repr__(PyObject *self, CYTHON_UNUSED PyObject *arg) { + return __pyx_MemviewEnum___repr__(self); +} + +static PyMethodDef __pyx_methods_Enum[] = { + {"__repr__", (PyCFunction)__pyx_specialmethod___pyx_MemviewEnum___repr__, METH_NOARGS|METH_COEXIST, 0}, + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_MemviewEnum_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_MemviewEnum_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +static PyType_Slot __pyx_type___pyx_MemviewEnum_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc_Enum}, + {Py_tp_repr, (void *)__pyx_MemviewEnum___repr__}, + {Py_tp_traverse, (void *)__pyx_tp_traverse_Enum}, + {Py_tp_clear, (void *)__pyx_tp_clear_Enum}, + {Py_tp_methods, (void *)__pyx_methods_Enum}, + {Py_tp_init, (void *)__pyx_MemviewEnum___init__}, + {Py_tp_new, (void *)__pyx_tp_new_Enum}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_MemviewEnum_spec = { + "fairseq.data.token_block_utils_fast.Enum", + sizeof(struct __pyx_MemviewEnum_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, + __pyx_type___pyx_MemviewEnum_slots, +}; +#else + +static PyTypeObject __pyx_type___pyx_MemviewEnum = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.token_block_utils_fast.""Enum", /*tp_name*/ + sizeof(struct __pyx_MemviewEnum_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc_Enum, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + __pyx_MemviewEnum___repr__, /*tp_repr*/ + 0, /*tp_as_number*/ + 0, /*tp_as_sequence*/ + 0, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + 0, /*tp_str*/ + 0, /*tp_getattro*/ + 0, /*tp_setattro*/ + 0, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, /*tp_flags*/ + 0, /*tp_doc*/ + __pyx_tp_traverse_Enum, /*tp_traverse*/ + __pyx_tp_clear_Enum, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods_Enum, /*tp_methods*/ + 0, /*tp_members*/ + 0, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + __pyx_MemviewEnum___init__, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new_Enum, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif +static struct __pyx_vtabstruct_memoryview __pyx_vtable_memoryview; + +static PyObject *__pyx_tp_new_memoryview(PyTypeObject *t, PyObject *a, PyObject *k) { + struct __pyx_memoryview_obj *p; + PyObject *o; + #if CYTHON_COMPILING_IN_LIMITED_API + allocfunc alloc_func = (allocfunc)PyType_GetSlot(t, Py_tp_alloc); + o = alloc_func(t, 0); + #else + if (likely(!__Pyx_PyType_HasFeature(t, Py_TPFLAGS_IS_ABSTRACT))) { + o = (*t->tp_alloc)(t, 0); + } else { + o = (PyObject *) PyBaseObject_Type.tp_new(t, __pyx_empty_tuple, 0); + } + if (unlikely(!o)) return 0; + #endif + p = ((struct __pyx_memoryview_obj *)o); + p->__pyx_vtab = __pyx_vtabptr_memoryview; + p->obj = Py_None; Py_INCREF(Py_None); + p->_size = Py_None; Py_INCREF(Py_None); + p->_array_interface = Py_None; Py_INCREF(Py_None); + p->view.obj = NULL; + if (unlikely(__pyx_memoryview___cinit__(o, a, k) < 0)) goto bad; + return o; + bad: + Py_DECREF(o); o = 0; + return NULL; +} + +static void __pyx_tp_dealloc_memoryview(PyObject *o) { + struct __pyx_memoryview_obj *p = (struct __pyx_memoryview_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && !__Pyx_PyObject_GC_IsFinalized(o)) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc_memoryview) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + PyObject_GC_UnTrack(o); + { + PyObject *etype, *eval, *etb; + PyErr_Fetch(&etype, &eval, &etb); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1); + __pyx_memoryview___dealloc__(o); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1); + PyErr_Restore(etype, eval, etb); + } + Py_CLEAR(p->obj); + Py_CLEAR(p->_size); + Py_CLEAR(p->_array_interface); + #if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + (*Py_TYPE(o)->tp_free)(o); + #else + { + freefunc tp_free = (freefunc)PyType_GetSlot(Py_TYPE(o), Py_tp_free); + if (tp_free) tp_free(o); + } + #endif +} + +static int __pyx_tp_traverse_memoryview(PyObject *o, visitproc v, void *a) { + int e; + struct __pyx_memoryview_obj *p = (struct __pyx_memoryview_obj *)o; + if (p->obj) { + e = (*v)(p->obj, a); if (e) return e; + } + if (p->_size) { + e = (*v)(p->_size, a); if (e) return e; + } + if (p->_array_interface) { + e = (*v)(p->_array_interface, a); if (e) return e; + } + if (p->view.obj) { + e = (*v)(p->view.obj, a); if (e) return e; + } + return 0; +} + +static int __pyx_tp_clear_memoryview(PyObject *o) { + PyObject* tmp; + struct __pyx_memoryview_obj *p = (struct __pyx_memoryview_obj *)o; + tmp = ((PyObject*)p->obj); + p->obj = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + tmp = ((PyObject*)p->_size); + p->_size = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + tmp = ((PyObject*)p->_array_interface); + p->_array_interface = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + Py_CLEAR(p->view.obj); + return 0; +} +static PyObject *__pyx_sq_item_memoryview(PyObject *o, Py_ssize_t i) { + PyObject *r; + PyObject *x = PyInt_FromSsize_t(i); if(!x) return 0; + r = Py_TYPE(o)->tp_as_mapping->mp_subscript(o, x); + Py_DECREF(x); + return r; +} + +static int __pyx_mp_ass_subscript_memoryview(PyObject *o, PyObject *i, PyObject *v) { + if (v) { + return __pyx_memoryview___setitem__(o, i, v); + } + else { + __Pyx_TypeName o_type_name; + o_type_name = __Pyx_PyType_GetName(Py_TYPE(o)); + PyErr_Format(PyExc_NotImplementedError, + "Subscript deletion not supported by " __Pyx_FMT_TYPENAME, o_type_name); + __Pyx_DECREF_TypeName(o_type_name); + return -1; + } +} + +static PyObject *__pyx_getprop___pyx_memoryview_T(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_1T_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_base(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_4base_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_shape(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_5shape_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_strides(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_7strides_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_suboffsets(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_10suboffsets_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_ndim(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_4ndim_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_itemsize(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_8itemsize_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_nbytes(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_6nbytes_1__get__(o); +} + +static PyObject *__pyx_getprop___pyx_memoryview_size(PyObject *o, CYTHON_UNUSED void *x) { + return __pyx_pw_15View_dot_MemoryView_10memoryview_4size_1__get__(o); +} + +static PyObject *__pyx_specialmethod___pyx_memoryview___repr__(PyObject *self, CYTHON_UNUSED PyObject *arg) { + return __pyx_memoryview___repr__(self); +} + +static PyMethodDef __pyx_methods_memoryview[] = { + {"__repr__", (PyCFunction)__pyx_specialmethod___pyx_memoryview___repr__, METH_NOARGS|METH_COEXIST, 0}, + {"is_c_contig", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_is_c_contig, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"is_f_contig", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_is_f_contig, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"copy", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_copy, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"copy_fortran", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_memoryview_copy_fortran, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryview_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryview_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; + +static struct PyGetSetDef __pyx_getsets_memoryview[] = { + {(char *)"T", __pyx_getprop___pyx_memoryview_T, 0, (char *)0, 0}, + {(char *)"base", __pyx_getprop___pyx_memoryview_base, 0, (char *)0, 0}, + {(char *)"shape", __pyx_getprop___pyx_memoryview_shape, 0, (char *)0, 0}, + {(char *)"strides", __pyx_getprop___pyx_memoryview_strides, 0, (char *)0, 0}, + {(char *)"suboffsets", __pyx_getprop___pyx_memoryview_suboffsets, 0, (char *)0, 0}, + {(char *)"ndim", __pyx_getprop___pyx_memoryview_ndim, 0, (char *)0, 0}, + {(char *)"itemsize", __pyx_getprop___pyx_memoryview_itemsize, 0, (char *)0, 0}, + {(char *)"nbytes", __pyx_getprop___pyx_memoryview_nbytes, 0, (char *)0, 0}, + {(char *)"size", __pyx_getprop___pyx_memoryview_size, 0, (char *)0, 0}, + {0, 0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +#if !CYTHON_COMPILING_IN_LIMITED_API + +static PyBufferProcs __pyx_tp_as_buffer_memoryview = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_memoryview_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; +#endif +static PyType_Slot __pyx_type___pyx_memoryview_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc_memoryview}, + {Py_tp_repr, (void *)__pyx_memoryview___repr__}, + {Py_sq_length, (void *)__pyx_memoryview___len__}, + {Py_sq_item, (void *)__pyx_sq_item_memoryview}, + {Py_mp_length, (void *)__pyx_memoryview___len__}, + {Py_mp_subscript, (void *)__pyx_memoryview___getitem__}, + {Py_mp_ass_subscript, (void *)__pyx_mp_ass_subscript_memoryview}, + {Py_tp_str, (void *)__pyx_memoryview___str__}, + #if defined(Py_bf_getbuffer) + {Py_bf_getbuffer, (void *)__pyx_memoryview_getbuffer}, + #endif + {Py_tp_traverse, (void *)__pyx_tp_traverse_memoryview}, + {Py_tp_clear, (void *)__pyx_tp_clear_memoryview}, + {Py_tp_methods, (void *)__pyx_methods_memoryview}, + {Py_tp_getset, (void *)__pyx_getsets_memoryview}, + {Py_tp_new, (void *)__pyx_tp_new_memoryview}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_memoryview_spec = { + "fairseq.data.token_block_utils_fast.memoryview", + sizeof(struct __pyx_memoryview_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, + __pyx_type___pyx_memoryview_slots, +}; +#else + +static PySequenceMethods __pyx_tp_as_sequence_memoryview = { + __pyx_memoryview___len__, /*sq_length*/ + 0, /*sq_concat*/ + 0, /*sq_repeat*/ + __pyx_sq_item_memoryview, /*sq_item*/ + 0, /*sq_slice*/ + 0, /*sq_ass_item*/ + 0, /*sq_ass_slice*/ + 0, /*sq_contains*/ + 0, /*sq_inplace_concat*/ + 0, /*sq_inplace_repeat*/ +}; + +static PyMappingMethods __pyx_tp_as_mapping_memoryview = { + __pyx_memoryview___len__, /*mp_length*/ + __pyx_memoryview___getitem__, /*mp_subscript*/ + __pyx_mp_ass_subscript_memoryview, /*mp_ass_subscript*/ +}; + +static PyBufferProcs __pyx_tp_as_buffer_memoryview = { + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getreadbuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getwritebuffer*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getsegcount*/ + #endif + #if PY_MAJOR_VERSION < 3 + 0, /*bf_getcharbuffer*/ + #endif + __pyx_memoryview_getbuffer, /*bf_getbuffer*/ + 0, /*bf_releasebuffer*/ +}; + +static PyTypeObject __pyx_type___pyx_memoryview = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.token_block_utils_fast.""memoryview", /*tp_name*/ + sizeof(struct __pyx_memoryview_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc_memoryview, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + __pyx_memoryview___repr__, /*tp_repr*/ + 0, /*tp_as_number*/ + &__pyx_tp_as_sequence_memoryview, /*tp_as_sequence*/ + &__pyx_tp_as_mapping_memoryview, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + __pyx_memoryview___str__, /*tp_str*/ + 0, /*tp_getattro*/ + 0, /*tp_setattro*/ + &__pyx_tp_as_buffer_memoryview, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC, /*tp_flags*/ + 0, /*tp_doc*/ + __pyx_tp_traverse_memoryview, /*tp_traverse*/ + __pyx_tp_clear_memoryview, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods_memoryview, /*tp_methods*/ + 0, /*tp_members*/ + __pyx_getsets_memoryview, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + 0, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new_memoryview, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif +static struct __pyx_vtabstruct__memoryviewslice __pyx_vtable__memoryviewslice; + +static PyObject *__pyx_tp_new__memoryviewslice(PyTypeObject *t, PyObject *a, PyObject *k) { + struct __pyx_memoryviewslice_obj *p; + PyObject *o = __pyx_tp_new_memoryview(t, a, k); + if (unlikely(!o)) return 0; + p = ((struct __pyx_memoryviewslice_obj *)o); + p->__pyx_base.__pyx_vtab = (struct __pyx_vtabstruct_memoryview*)__pyx_vtabptr__memoryviewslice; + new((void*)&(p->from_slice)) __Pyx_memviewslice(); + p->from_object = Py_None; Py_INCREF(Py_None); + p->from_slice.memview = NULL; + return o; +} + +static void __pyx_tp_dealloc__memoryviewslice(PyObject *o) { + struct __pyx_memoryviewslice_obj *p = (struct __pyx_memoryviewslice_obj *)o; + #if CYTHON_USE_TP_FINALIZE + if (unlikely((PY_VERSION_HEX >= 0x03080000 || __Pyx_PyType_HasFeature(Py_TYPE(o), Py_TPFLAGS_HAVE_FINALIZE)) && __Pyx_PyObject_GetSlot(o, tp_finalize, destructor)) && !__Pyx_PyObject_GC_IsFinalized(o)) { + if (__Pyx_PyObject_GetSlot(o, tp_dealloc, destructor) == __pyx_tp_dealloc__memoryviewslice) { + if (PyObject_CallFinalizerFromDealloc(o)) return; + } + } + #endif + PyObject_GC_UnTrack(o); + { + PyObject *etype, *eval, *etb; + PyErr_Fetch(&etype, &eval, &etb); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1); + __pyx_memoryviewslice___dealloc__(o); + __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1); + PyErr_Restore(etype, eval, etb); + } + __Pyx_call_destructor(p->from_slice); + Py_CLEAR(p->from_object); + PyObject_GC_Track(o); + __pyx_tp_dealloc_memoryview(o); +} + +static int __pyx_tp_traverse__memoryviewslice(PyObject *o, visitproc v, void *a) { + int e; + struct __pyx_memoryviewslice_obj *p = (struct __pyx_memoryviewslice_obj *)o; + e = __pyx_tp_traverse_memoryview(o, v, a); if (e) return e; + if (p->from_object) { + e = (*v)(p->from_object, a); if (e) return e; + } + return 0; +} + +static int __pyx_tp_clear__memoryviewslice(PyObject *o) { + PyObject* tmp; + struct __pyx_memoryviewslice_obj *p = (struct __pyx_memoryviewslice_obj *)o; + __pyx_tp_clear_memoryview(o); + tmp = ((PyObject*)p->from_object); + p->from_object = Py_None; Py_INCREF(Py_None); + Py_XDECREF(tmp); + __PYX_XCLEAR_MEMVIEW(&p->from_slice, 1); + return 0; +} + +static PyMethodDef __pyx_methods__memoryviewslice[] = { + {"__reduce_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryviewslice_1__reduce_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {"__setstate_cython__", (PyCFunction)(void*)(__Pyx_PyCFunction_FastCallWithKeywords)__pyx_pw___pyx_memoryviewslice_3__setstate_cython__, __Pyx_METH_FASTCALL|METH_KEYWORDS, 0}, + {0, 0, 0, 0} +}; +#if CYTHON_USE_TYPE_SPECS +static PyType_Slot __pyx_type___pyx_memoryviewslice_slots[] = { + {Py_tp_dealloc, (void *)__pyx_tp_dealloc__memoryviewslice}, + {Py_tp_doc, (void *)PyDoc_STR("Internal class for passing memoryview slices to Python")}, + {Py_tp_traverse, (void *)__pyx_tp_traverse__memoryviewslice}, + {Py_tp_clear, (void *)__pyx_tp_clear__memoryviewslice}, + {Py_tp_methods, (void *)__pyx_methods__memoryviewslice}, + {Py_tp_new, (void *)__pyx_tp_new__memoryviewslice}, + {0, 0}, +}; +static PyType_Spec __pyx_type___pyx_memoryviewslice_spec = { + "fairseq.data.token_block_utils_fast._memoryviewslice", + sizeof(struct __pyx_memoryviewslice_obj), + 0, + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC|Py_TPFLAGS_SEQUENCE, + __pyx_type___pyx_memoryviewslice_slots, +}; +#else + +static PyTypeObject __pyx_type___pyx_memoryviewslice = { + PyVarObject_HEAD_INIT(0, 0) + "fairseq.data.token_block_utils_fast.""_memoryviewslice", /*tp_name*/ + sizeof(struct __pyx_memoryviewslice_obj), /*tp_basicsize*/ + 0, /*tp_itemsize*/ + __pyx_tp_dealloc__memoryviewslice, /*tp_dealloc*/ + #if PY_VERSION_HEX < 0x030800b4 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030800b4 + 0, /*tp_vectorcall_offset*/ + #endif + 0, /*tp_getattr*/ + 0, /*tp_setattr*/ + #if PY_MAJOR_VERSION < 3 + 0, /*tp_compare*/ + #endif + #if PY_MAJOR_VERSION >= 3 + 0, /*tp_as_async*/ + #endif + #if CYTHON_COMPILING_IN_PYPY || 0 + __pyx_memoryview___repr__, /*tp_repr*/ + #else + 0, /*tp_repr*/ + #endif + 0, /*tp_as_number*/ + 0, /*tp_as_sequence*/ + 0, /*tp_as_mapping*/ + 0, /*tp_hash*/ + 0, /*tp_call*/ + #if CYTHON_COMPILING_IN_PYPY || 0 + __pyx_memoryview___str__, /*tp_str*/ + #else + 0, /*tp_str*/ + #endif + 0, /*tp_getattro*/ + 0, /*tp_setattro*/ + 0, /*tp_as_buffer*/ + Py_TPFLAGS_DEFAULT|Py_TPFLAGS_HAVE_VERSION_TAG|Py_TPFLAGS_CHECKTYPES|Py_TPFLAGS_HAVE_NEWBUFFER|Py_TPFLAGS_BASETYPE|Py_TPFLAGS_HAVE_GC|Py_TPFLAGS_SEQUENCE, /*tp_flags*/ + PyDoc_STR("Internal class for passing memoryview slices to Python"), /*tp_doc*/ + __pyx_tp_traverse__memoryviewslice, /*tp_traverse*/ + __pyx_tp_clear__memoryviewslice, /*tp_clear*/ + 0, /*tp_richcompare*/ + 0, /*tp_weaklistoffset*/ + 0, /*tp_iter*/ + 0, /*tp_iternext*/ + __pyx_methods__memoryviewslice, /*tp_methods*/ + 0, /*tp_members*/ + 0, /*tp_getset*/ + 0, /*tp_base*/ + 0, /*tp_dict*/ + 0, /*tp_descr_get*/ + 0, /*tp_descr_set*/ + #if !CYTHON_USE_TYPE_SPECS + 0, /*tp_dictoffset*/ + #endif + 0, /*tp_init*/ + 0, /*tp_alloc*/ + __pyx_tp_new__memoryviewslice, /*tp_new*/ + 0, /*tp_free*/ + 0, /*tp_is_gc*/ + 0, /*tp_bases*/ + 0, /*tp_mro*/ + 0, /*tp_cache*/ + 0, /*tp_subclasses*/ + 0, /*tp_weaklist*/ + 0, /*tp_del*/ + 0, /*tp_version_tag*/ + #if PY_VERSION_HEX >= 0x030400a1 + #if CYTHON_USE_TP_FINALIZE + 0, /*tp_finalize*/ + #else + NULL, /*tp_finalize*/ + #endif + #endif + #if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, /*tp_vectorcall*/ + #endif + #if __PYX_NEED_TP_PRINT_SLOT == 1 + 0, /*tp_print*/ + #endif + #if PY_VERSION_HEX >= 0x030C0000 + 0, /*tp_watched*/ + #endif + #if PY_VERSION_HEX >= 0x030d00A4 + 0, /*tp_versions_used*/ + #endif + #if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, /*tp_pypy_flags*/ + #endif +}; +#endif + +static PyMethodDef __pyx_methods[] = { + {0, 0, 0, 0} +}; +#ifndef CYTHON_SMALL_CODE +#if defined(__clang__) + #define CYTHON_SMALL_CODE +#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3)) + #define CYTHON_SMALL_CODE __attribute__((cold)) +#else + #define CYTHON_SMALL_CODE +#endif +#endif +/* #### Code section: pystring_table ### */ + +static int __Pyx_CreateStringTabAndInitStrings(void) { + __Pyx_StringTabEntry __pyx_string_tab[] = { + {&__pyx_kp_u_, __pyx_k_, sizeof(__pyx_k_), 0, 1, 0, 0}, + {&__pyx_n_s_ASCII, __pyx_k_ASCII, sizeof(__pyx_k_ASCII), 0, 0, 1, 1}, + {&__pyx_kp_s_All_dimensions_preceding_dimensi, __pyx_k_All_dimensions_preceding_dimensi, sizeof(__pyx_k_All_dimensions_preceding_dimensi), 0, 0, 1, 0}, + {&__pyx_n_s_AssertionError, __pyx_k_AssertionError, sizeof(__pyx_k_AssertionError), 0, 0, 1, 1}, + {&__pyx_kp_s_Buffer_view_does_not_expose_stri, __pyx_k_Buffer_view_does_not_expose_stri, sizeof(__pyx_k_Buffer_view_does_not_expose_stri), 0, 0, 1, 0}, + {&__pyx_kp_s_Can_only_create_a_buffer_that_is, __pyx_k_Can_only_create_a_buffer_that_is, sizeof(__pyx_k_Can_only_create_a_buffer_that_is), 0, 0, 1, 0}, + {&__pyx_kp_s_Cannot_assign_to_read_only_memor, __pyx_k_Cannot_assign_to_read_only_memor, sizeof(__pyx_k_Cannot_assign_to_read_only_memor), 0, 0, 1, 0}, + {&__pyx_kp_s_Cannot_create_writable_memory_vi, __pyx_k_Cannot_create_writable_memory_vi, sizeof(__pyx_k_Cannot_create_writable_memory_vi), 0, 0, 1, 0}, + {&__pyx_kp_u_Cannot_index_with_type, __pyx_k_Cannot_index_with_type, sizeof(__pyx_k_Cannot_index_with_type), 0, 1, 0, 0}, + {&__pyx_kp_s_Cannot_transpose_memoryview_with, __pyx_k_Cannot_transpose_memoryview_with, sizeof(__pyx_k_Cannot_transpose_memoryview_with), 0, 0, 1, 0}, + {&__pyx_n_s_DTYPE, __pyx_k_DTYPE, sizeof(__pyx_k_DTYPE), 0, 0, 1, 1}, + {&__pyx_n_s_DatasetSearcher, __pyx_k_DatasetSearcher, sizeof(__pyx_k_DatasetSearcher), 0, 0, 1, 1}, + {&__pyx_n_s_DatasetSearcher___reduce_cython, __pyx_k_DatasetSearcher___reduce_cython, sizeof(__pyx_k_DatasetSearcher___reduce_cython), 0, 0, 1, 1}, + {&__pyx_n_s_DatasetSearcher___setstate_cytho, __pyx_k_DatasetSearcher___setstate_cytho, sizeof(__pyx_k_DatasetSearcher___setstate_cytho), 0, 0, 1, 1}, + {&__pyx_kp_s_Dimension_d_is_not_direct, __pyx_k_Dimension_d_is_not_direct, sizeof(__pyx_k_Dimension_d_is_not_direct), 0, 0, 1, 0}, + {&__pyx_n_s_Ellipsis, __pyx_k_Ellipsis, sizeof(__pyx_k_Ellipsis), 0, 0, 1, 1}, + {&__pyx_kp_s_Empty_shape_tuple_for_cython_arr, __pyx_k_Empty_shape_tuple_for_cython_arr, sizeof(__pyx_k_Empty_shape_tuple_for_cython_arr), 0, 0, 1, 0}, + {&__pyx_n_s_ImportError, __pyx_k_ImportError, sizeof(__pyx_k_ImportError), 0, 0, 1, 1}, + {&__pyx_kp_s_Incompatible_checksums_0x_x_vs_0, __pyx_k_Incompatible_checksums_0x_x_vs_0, sizeof(__pyx_k_Incompatible_checksums_0x_x_vs_0), 0, 0, 1, 0}, + {&__pyx_kp_s_Incompatible_checksums_0x_x_vs_0_2, __pyx_k_Incompatible_checksums_0x_x_vs_0_2, sizeof(__pyx_k_Incompatible_checksums_0x_x_vs_0_2), 0, 0, 1, 0}, + {&__pyx_n_s_IndexError, __pyx_k_IndexError, sizeof(__pyx_k_IndexError), 0, 0, 1, 1}, + {&__pyx_kp_s_Index_out_of_bounds_axis_d, __pyx_k_Index_out_of_bounds_axis_d, sizeof(__pyx_k_Index_out_of_bounds_axis_d), 0, 0, 1, 0}, + {&__pyx_kp_s_Indirect_dimensions_not_supporte, __pyx_k_Indirect_dimensions_not_supporte, sizeof(__pyx_k_Indirect_dimensions_not_supporte), 0, 0, 1, 0}, + {&__pyx_kp_u_Invalid_break_mode, __pyx_k_Invalid_break_mode, sizeof(__pyx_k_Invalid_break_mode), 0, 1, 0, 0}, + {&__pyx_kp_u_Invalid_mode_expected_c_or_fortr, __pyx_k_Invalid_mode_expected_c_or_fortr, sizeof(__pyx_k_Invalid_mode_expected_c_or_fortr), 0, 1, 0, 0}, + {&__pyx_kp_u_Invalid_shape_in_axis, __pyx_k_Invalid_shape_in_axis, sizeof(__pyx_k_Invalid_shape_in_axis), 0, 1, 0, 0}, + {&__pyx_n_s_MemoryError, __pyx_k_MemoryError, sizeof(__pyx_k_MemoryError), 0, 0, 1, 1}, + {&__pyx_kp_s_MemoryView_of_r_at_0x_x, __pyx_k_MemoryView_of_r_at_0x_x, sizeof(__pyx_k_MemoryView_of_r_at_0x_x), 0, 0, 1, 0}, + {&__pyx_kp_s_MemoryView_of_r_object, __pyx_k_MemoryView_of_r_object, sizeof(__pyx_k_MemoryView_of_r_object), 0, 0, 1, 0}, + {&__pyx_n_b_O, __pyx_k_O, sizeof(__pyx_k_O), 0, 0, 0, 1}, + {&__pyx_kp_u_Out_of_bounds_on_buffer_access_a, __pyx_k_Out_of_bounds_on_buffer_access_a, sizeof(__pyx_k_Out_of_bounds_on_buffer_access_a), 0, 1, 0, 0}, + {&__pyx_n_s_PickleError, __pyx_k_PickleError, sizeof(__pyx_k_PickleError), 0, 0, 1, 1}, + {&__pyx_n_s_Sequence, __pyx_k_Sequence, sizeof(__pyx_k_Sequence), 0, 0, 1, 1}, + {&__pyx_kp_s_Step_may_not_be_zero_axis_d, __pyx_k_Step_may_not_be_zero_axis_d, sizeof(__pyx_k_Step_may_not_be_zero_axis_d), 0, 0, 1, 0}, + {&__pyx_n_s_TypeError, __pyx_k_TypeError, sizeof(__pyx_k_TypeError), 0, 0, 1, 1}, + {&__pyx_kp_s_Unable_to_convert_item_to_object, __pyx_k_Unable_to_convert_item_to_object, sizeof(__pyx_k_Unable_to_convert_item_to_object), 0, 0, 1, 0}, + {&__pyx_n_s_ValueError, __pyx_k_ValueError, sizeof(__pyx_k_ValueError), 0, 0, 1, 1}, + {&__pyx_n_s_View_MemoryView, __pyx_k_View_MemoryView, sizeof(__pyx_k_View_MemoryView), 0, 0, 1, 1}, + {&__pyx_kp_u__2, __pyx_k__2, sizeof(__pyx_k__2), 0, 1, 0, 0}, + {&__pyx_n_s__3, __pyx_k__3, sizeof(__pyx_k__3), 0, 0, 1, 1}, + {&__pyx_n_s__35, __pyx_k__35, sizeof(__pyx_k__35), 0, 0, 1, 1}, + {&__pyx_kp_u__6, __pyx_k__6, sizeof(__pyx_k__6), 0, 1, 0, 0}, + {&__pyx_kp_u__7, __pyx_k__7, sizeof(__pyx_k__7), 0, 1, 0, 0}, + {&__pyx_n_s_abc, __pyx_k_abc, sizeof(__pyx_k_abc), 0, 0, 1, 1}, + {&__pyx_n_s_allocate_buffer, __pyx_k_allocate_buffer, sizeof(__pyx_k_allocate_buffer), 0, 0, 1, 1}, + {&__pyx_kp_u_and, __pyx_k_and, sizeof(__pyx_k_and), 0, 1, 0, 0}, + {&__pyx_n_s_asyncio_coroutines, __pyx_k_asyncio_coroutines, sizeof(__pyx_k_asyncio_coroutines), 0, 0, 1, 1}, + {&__pyx_n_s_axis, __pyx_k_axis, sizeof(__pyx_k_axis), 0, 0, 1, 1}, + {&__pyx_n_s_base, __pyx_k_base, sizeof(__pyx_k_base), 0, 0, 1, 1}, + {&__pyx_n_s_block_size, __pyx_k_block_size, sizeof(__pyx_k_block_size), 0, 0, 1, 1}, + {&__pyx_n_s_break_mode, __pyx_k_break_mode, sizeof(__pyx_k_break_mode), 0, 0, 1, 1}, + {&__pyx_n_s_c, __pyx_k_c, sizeof(__pyx_k_c), 0, 0, 1, 1}, + {&__pyx_n_u_c, __pyx_k_c, sizeof(__pyx_k_c), 0, 1, 0, 1}, + {&__pyx_n_s_chain, __pyx_k_chain, sizeof(__pyx_k_chain), 0, 0, 1, 1}, + {&__pyx_n_s_class, __pyx_k_class, sizeof(__pyx_k_class), 0, 0, 1, 1}, + {&__pyx_n_s_class_getitem, __pyx_k_class_getitem, sizeof(__pyx_k_class_getitem), 0, 0, 1, 1}, + {&__pyx_n_s_cline_in_traceback, __pyx_k_cline_in_traceback, sizeof(__pyx_k_cline_in_traceback), 0, 0, 1, 1}, + {&__pyx_n_s_collections, __pyx_k_collections, sizeof(__pyx_k_collections), 0, 0, 1, 1}, + {&__pyx_kp_s_collections_abc, __pyx_k_collections_abc, sizeof(__pyx_k_collections_abc), 0, 0, 1, 0}, + {&__pyx_n_u_complete, __pyx_k_complete, sizeof(__pyx_k_complete), 0, 1, 0, 1}, + {&__pyx_n_u_complete_doc, __pyx_k_complete_doc, sizeof(__pyx_k_complete_doc), 0, 1, 0, 1}, + {&__pyx_kp_s_contiguous_and_direct, __pyx_k_contiguous_and_direct, sizeof(__pyx_k_contiguous_and_direct), 0, 0, 1, 0}, + {&__pyx_kp_s_contiguous_and_indirect, __pyx_k_contiguous_and_indirect, sizeof(__pyx_k_contiguous_and_indirect), 0, 0, 1, 0}, + {&__pyx_n_s_count, __pyx_k_count, sizeof(__pyx_k_count), 0, 0, 1, 1}, + {&__pyx_n_s_cumsum, __pyx_k_cumsum, sizeof(__pyx_k_cumsum), 0, 0, 1, 1}, + {&__pyx_n_s_dict, __pyx_k_dict, sizeof(__pyx_k_dict), 0, 0, 1, 1}, + {&__pyx_n_s_dict_2, __pyx_k_dict_2, sizeof(__pyx_k_dict_2), 0, 0, 1, 1}, + {&__pyx_kp_u_disable, __pyx_k_disable, sizeof(__pyx_k_disable), 0, 1, 0, 0}, + {&__pyx_n_s_document_sep_len, __pyx_k_document_sep_len, sizeof(__pyx_k_document_sep_len), 0, 0, 1, 1}, + {&__pyx_n_s_dtype, __pyx_k_dtype, sizeof(__pyx_k_dtype), 0, 0, 1, 1}, + {&__pyx_n_s_dtype_is_object, __pyx_k_dtype_is_object, sizeof(__pyx_k_dtype_is_object), 0, 0, 1, 1}, + {&__pyx_kp_u_enable, __pyx_k_enable, sizeof(__pyx_k_enable), 0, 1, 0, 0}, + {&__pyx_n_s_encode, __pyx_k_encode, sizeof(__pyx_k_encode), 0, 0, 1, 1}, + {&__pyx_n_s_enumerate, __pyx_k_enumerate, sizeof(__pyx_k_enumerate), 0, 0, 1, 1}, + {&__pyx_n_u_eos, __pyx_k_eos, sizeof(__pyx_k_eos), 0, 1, 0, 1}, + {&__pyx_n_s_error, __pyx_k_error, sizeof(__pyx_k_error), 0, 0, 1, 1}, + {&__pyx_kp_s_fairseq_data_token_block_utils_f, __pyx_k_fairseq_data_token_block_utils_f, sizeof(__pyx_k_fairseq_data_token_block_utils_f), 0, 0, 1, 0}, + {&__pyx_n_s_fairseq_data_token_block_utils_f_2, __pyx_k_fairseq_data_token_block_utils_f_2, sizeof(__pyx_k_fairseq_data_token_block_utils_f_2), 0, 0, 1, 1}, + {&__pyx_n_s_flags, __pyx_k_flags, sizeof(__pyx_k_flags), 0, 0, 1, 1}, + {&__pyx_n_s_format, __pyx_k_format, sizeof(__pyx_k_format), 0, 0, 1, 1}, + {&__pyx_n_s_fortran, __pyx_k_fortran, sizeof(__pyx_k_fortran), 0, 0, 1, 1}, + {&__pyx_n_u_fortran, __pyx_k_fortran, sizeof(__pyx_k_fortran), 0, 1, 0, 1}, + {&__pyx_n_s_from_iterable, __pyx_k_from_iterable, sizeof(__pyx_k_from_iterable), 0, 0, 1, 1}, + {&__pyx_n_s_fromiter, __pyx_k_fromiter, sizeof(__pyx_k_fromiter), 0, 0, 1, 1}, + {&__pyx_kp_u_gc, __pyx_k_gc, sizeof(__pyx_k_gc), 0, 1, 0, 0}, + {&__pyx_n_s_get_block_to_dataset_index_fast, __pyx_k_get_block_to_dataset_index_fast, sizeof(__pyx_k_get_block_to_dataset_index_fast), 0, 0, 1, 1}, + {&__pyx_n_s_get_slice_indices_fast, __pyx_k_get_slice_indices_fast, sizeof(__pyx_k_get_slice_indices_fast), 0, 0, 1, 1}, + {&__pyx_n_s_getstate, __pyx_k_getstate, sizeof(__pyx_k_getstate), 0, 0, 1, 1}, + {&__pyx_kp_u_got, __pyx_k_got, sizeof(__pyx_k_got), 0, 1, 0, 0}, + {&__pyx_kp_u_got_differing_extents_in_dimensi, __pyx_k_got_differing_extents_in_dimensi, sizeof(__pyx_k_got_differing_extents_in_dimensi), 0, 1, 0, 0}, + {&__pyx_n_s_id, __pyx_k_id, sizeof(__pyx_k_id), 0, 0, 1, 1}, + {&__pyx_n_s_import, __pyx_k_import, sizeof(__pyx_k_import), 0, 0, 1, 1}, + {&__pyx_n_s_index, __pyx_k_index, sizeof(__pyx_k_index), 0, 0, 1, 1}, + {&__pyx_n_s_initializing, __pyx_k_initializing, sizeof(__pyx_k_initializing), 0, 0, 1, 1}, + {&__pyx_n_s_int64, __pyx_k_int64, sizeof(__pyx_k_int64), 0, 0, 1, 1}, + {&__pyx_n_s_is_coroutine, __pyx_k_is_coroutine, sizeof(__pyx_k_is_coroutine), 0, 0, 1, 1}, + {&__pyx_kp_u_isenabled, __pyx_k_isenabled, sizeof(__pyx_k_isenabled), 0, 1, 0, 0}, + {&__pyx_n_s_itemsize, __pyx_k_itemsize, sizeof(__pyx_k_itemsize), 0, 0, 1, 1}, + {&__pyx_kp_s_itemsize_0_for_cython_array, __pyx_k_itemsize_0_for_cython_array, sizeof(__pyx_k_itemsize_0_for_cython_array), 0, 0, 1, 0}, + {&__pyx_n_s_itertools, __pyx_k_itertools, sizeof(__pyx_k_itertools), 0, 0, 1, 1}, + {&__pyx_n_s_main, __pyx_k_main, sizeof(__pyx_k_main), 0, 0, 1, 1}, + {&__pyx_n_s_memview, __pyx_k_memview, sizeof(__pyx_k_memview), 0, 0, 1, 1}, + {&__pyx_n_s_mode, __pyx_k_mode, sizeof(__pyx_k_mode), 0, 0, 1, 1}, + {&__pyx_n_s_name, __pyx_k_name, sizeof(__pyx_k_name), 0, 0, 1, 1}, + {&__pyx_n_s_name_2, __pyx_k_name_2, sizeof(__pyx_k_name_2), 0, 0, 1, 1}, + {&__pyx_n_s_ndim, __pyx_k_ndim, sizeof(__pyx_k_ndim), 0, 0, 1, 1}, + {&__pyx_n_s_new, __pyx_k_new, sizeof(__pyx_k_new), 0, 0, 1, 1}, + {&__pyx_kp_s_no_default___reduce___due_to_non, __pyx_k_no_default___reduce___due_to_non, sizeof(__pyx_k_no_default___reduce___due_to_non), 0, 0, 1, 0}, + {&__pyx_n_u_none, __pyx_k_none, sizeof(__pyx_k_none), 0, 1, 0, 1}, + {&__pyx_n_s_np, __pyx_k_np, sizeof(__pyx_k_np), 0, 0, 1, 1}, + {&__pyx_n_s_numpy, __pyx_k_numpy, sizeof(__pyx_k_numpy), 0, 0, 1, 1}, + {&__pyx_kp_u_numpy__core_multiarray_failed_to, __pyx_k_numpy__core_multiarray_failed_to, sizeof(__pyx_k_numpy__core_multiarray_failed_to), 0, 1, 0, 0}, + {&__pyx_kp_u_numpy__core_umath_failed_to_impo, __pyx_k_numpy__core_umath_failed_to_impo, sizeof(__pyx_k_numpy__core_umath_failed_to_impo), 0, 1, 0, 0}, + {&__pyx_n_s_obj, __pyx_k_obj, sizeof(__pyx_k_obj), 0, 0, 1, 1}, + {&__pyx_n_s_pack, __pyx_k_pack, sizeof(__pyx_k_pack), 0, 0, 1, 1}, + {&__pyx_n_s_pickle, __pyx_k_pickle, sizeof(__pyx_k_pickle), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_PickleError, __pyx_k_pyx_PickleError, sizeof(__pyx_k_pyx_PickleError), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_checksum, __pyx_k_pyx_checksum, sizeof(__pyx_k_pyx_checksum), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_result, __pyx_k_pyx_result, sizeof(__pyx_k_pyx_result), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_state, __pyx_k_pyx_state, sizeof(__pyx_k_pyx_state), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_type, __pyx_k_pyx_type, sizeof(__pyx_k_pyx_type), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_unpickle_DatasetSearcher, __pyx_k_pyx_unpickle_DatasetSearcher, sizeof(__pyx_k_pyx_unpickle_DatasetSearcher), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_unpickle_Enum, __pyx_k_pyx_unpickle_Enum, sizeof(__pyx_k_pyx_unpickle_Enum), 0, 0, 1, 1}, + {&__pyx_n_s_pyx_vtable, __pyx_k_pyx_vtable, sizeof(__pyx_k_pyx_vtable), 0, 0, 1, 1}, + {&__pyx_n_s_range, __pyx_k_range, sizeof(__pyx_k_range), 0, 0, 1, 1}, + {&__pyx_n_s_reduce, __pyx_k_reduce, sizeof(__pyx_k_reduce), 0, 0, 1, 1}, + {&__pyx_n_s_reduce_cython, __pyx_k_reduce_cython, sizeof(__pyx_k_reduce_cython), 0, 0, 1, 1}, + {&__pyx_n_s_reduce_ex, __pyx_k_reduce_ex, sizeof(__pyx_k_reduce_ex), 0, 0, 1, 1}, + {&__pyx_n_s_register, __pyx_k_register, sizeof(__pyx_k_register), 0, 0, 1, 1}, + {&__pyx_n_s_reshape, __pyx_k_reshape, sizeof(__pyx_k_reshape), 0, 0, 1, 1}, + {&__pyx_n_s_self, __pyx_k_self, sizeof(__pyx_k_self), 0, 0, 1, 1}, + {&__pyx_n_s_setstate, __pyx_k_setstate, sizeof(__pyx_k_setstate), 0, 0, 1, 1}, + {&__pyx_n_s_setstate_cython, __pyx_k_setstate_cython, sizeof(__pyx_k_setstate_cython), 0, 0, 1, 1}, + {&__pyx_n_s_shape, __pyx_k_shape, sizeof(__pyx_k_shape), 0, 0, 1, 1}, + {&__pyx_n_s_size, __pyx_k_size, sizeof(__pyx_k_size), 0, 0, 1, 1}, + {&__pyx_n_s_sizes, __pyx_k_sizes, sizeof(__pyx_k_sizes), 0, 0, 1, 1}, + {&__pyx_n_s_slice_indices, __pyx_k_slice_indices, sizeof(__pyx_k_slice_indices), 0, 0, 1, 1}, + {&__pyx_n_s_spec, __pyx_k_spec, sizeof(__pyx_k_spec), 0, 0, 1, 1}, + {&__pyx_n_s_start, __pyx_k_start, sizeof(__pyx_k_start), 0, 0, 1, 1}, + {&__pyx_n_s_state, __pyx_k_state, sizeof(__pyx_k_state), 0, 0, 1, 1}, + {&__pyx_n_s_step, __pyx_k_step, sizeof(__pyx_k_step), 0, 0, 1, 1}, + {&__pyx_n_s_stop, __pyx_k_stop, sizeof(__pyx_k_stop), 0, 0, 1, 1}, + {&__pyx_kp_s_strided_and_direct, __pyx_k_strided_and_direct, sizeof(__pyx_k_strided_and_direct), 0, 0, 1, 0}, + {&__pyx_kp_s_strided_and_direct_or_indirect, __pyx_k_strided_and_direct_or_indirect, sizeof(__pyx_k_strided_and_direct_or_indirect), 0, 0, 1, 0}, + {&__pyx_kp_s_strided_and_indirect, __pyx_k_strided_and_indirect, sizeof(__pyx_k_strided_and_indirect), 0, 0, 1, 0}, + {&__pyx_kp_s_stringsource, __pyx_k_stringsource, sizeof(__pyx_k_stringsource), 0, 0, 1, 0}, + {&__pyx_n_s_struct, __pyx_k_struct, sizeof(__pyx_k_struct), 0, 0, 1, 1}, + {&__pyx_n_s_sum, __pyx_k_sum, sizeof(__pyx_k_sum), 0, 0, 1, 1}, + {&__pyx_n_s_sys, __pyx_k_sys, sizeof(__pyx_k_sys), 0, 0, 1, 1}, + {&__pyx_n_s_test, __pyx_k_test, sizeof(__pyx_k_test), 0, 0, 1, 1}, + {&__pyx_n_s_torch, __pyx_k_torch, sizeof(__pyx_k_torch), 0, 0, 1, 1}, + {&__pyx_kp_s_unable_to_allocate_array_data, __pyx_k_unable_to_allocate_array_data, sizeof(__pyx_k_unable_to_allocate_array_data), 0, 0, 1, 0}, + {&__pyx_kp_s_unable_to_allocate_shape_and_str, __pyx_k_unable_to_allocate_shape_and_str, sizeof(__pyx_k_unable_to_allocate_shape_and_str), 0, 0, 1, 0}, + {&__pyx_n_s_unpack, __pyx_k_unpack, sizeof(__pyx_k_unpack), 0, 0, 1, 1}, + {&__pyx_n_s_update, __pyx_k_update, sizeof(__pyx_k_update), 0, 0, 1, 1}, + {&__pyx_n_s_use_setstate, __pyx_k_use_setstate, sizeof(__pyx_k_use_setstate), 0, 0, 1, 1}, + {&__pyx_n_s_version_info, __pyx_k_version_info, sizeof(__pyx_k_version_info), 0, 0, 1, 1}, + {&__pyx_n_s_zeros, __pyx_k_zeros, sizeof(__pyx_k_zeros), 0, 0, 1, 1}, + {0, 0, 0, 0, 0, 0, 0} + }; + return __Pyx_InitStrings(__pyx_string_tab); +} +/* #### Code section: cached_builtins ### */ +static CYTHON_SMALL_CODE int __Pyx_InitCachedBuiltins(void) { + __pyx_builtin_range = __Pyx_GetBuiltinName(__pyx_n_s_range); if (!__pyx_builtin_range) __PYX_ERR(0, 30, __pyx_L1_error) + __pyx_builtin_ValueError = __Pyx_GetBuiltinName(__pyx_n_s_ValueError); if (!__pyx_builtin_ValueError) __PYX_ERR(0, 102, __pyx_L1_error) + __pyx_builtin_AssertionError = __Pyx_GetBuiltinName(__pyx_n_s_AssertionError); if (!__pyx_builtin_AssertionError) __PYX_ERR(0, 171, __pyx_L1_error) + __pyx_builtin___import__ = __Pyx_GetBuiltinName(__pyx_n_s_import); if (!__pyx_builtin___import__) __PYX_ERR(1, 100, __pyx_L1_error) + __pyx_builtin_MemoryError = __Pyx_GetBuiltinName(__pyx_n_s_MemoryError); if (!__pyx_builtin_MemoryError) __PYX_ERR(1, 156, __pyx_L1_error) + __pyx_builtin_enumerate = __Pyx_GetBuiltinName(__pyx_n_s_enumerate); if (!__pyx_builtin_enumerate) __PYX_ERR(1, 159, __pyx_L1_error) + __pyx_builtin_TypeError = __Pyx_GetBuiltinName(__pyx_n_s_TypeError); if (!__pyx_builtin_TypeError) __PYX_ERR(1, 2, __pyx_L1_error) + __pyx_builtin_Ellipsis = __Pyx_GetBuiltinName(__pyx_n_s_Ellipsis); if (!__pyx_builtin_Ellipsis) __PYX_ERR(1, 408, __pyx_L1_error) + __pyx_builtin_id = __Pyx_GetBuiltinName(__pyx_n_s_id); if (!__pyx_builtin_id) __PYX_ERR(1, 618, __pyx_L1_error) + __pyx_builtin_IndexError = __Pyx_GetBuiltinName(__pyx_n_s_IndexError); if (!__pyx_builtin_IndexError) __PYX_ERR(1, 914, __pyx_L1_error) + __pyx_builtin_ImportError = __Pyx_GetBuiltinName(__pyx_n_s_ImportError); if (!__pyx_builtin_ImportError) __PYX_ERR(2, 1043, __pyx_L1_error) + return 0; + __pyx_L1_error:; + return -1; +} +/* #### Code section: cached_constants ### */ + +static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_InitCachedConstants", 0); + + /* "View.MemoryView":582 + * def suboffsets(self): + * if self.view.suboffsets == NULL: + * return (-1,) * self.view.ndim # <<<<<<<<<<<<<< + * + * return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]]) + */ + __pyx_tuple__4 = PyTuple_New(1); if (unlikely(!__pyx_tuple__4)) __PYX_ERR(1, 582, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__4); + __Pyx_INCREF(__pyx_int_neg_1); + __Pyx_GIVEREF(__pyx_int_neg_1); + if (__Pyx_PyTuple_SET_ITEM(__pyx_tuple__4, 0, __pyx_int_neg_1)) __PYX_ERR(1, 582, __pyx_L1_error); + __Pyx_GIVEREF(__pyx_tuple__4); + + /* "View.MemoryView":679 + * tup = index if isinstance(index, tuple) else (index,) + * + * result = [slice(None)] * ndim # <<<<<<<<<<<<<< + * have_slices = False + * seen_ellipsis = False + */ + __pyx_slice__5 = PySlice_New(Py_None, Py_None, Py_None); if (unlikely(!__pyx_slice__5)) __PYX_ERR(1, 679, __pyx_L1_error) + __Pyx_GOTREF(__pyx_slice__5); + __Pyx_GIVEREF(__pyx_slice__5); + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x82a3537, 0x6ae9995, 0xb068931): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x82a3537, 0x6ae9995, 0xb068931) = (name))" % __pyx_checksum + */ + __pyx_tuple__8 = PyTuple_Pack(3, __pyx_int_136983863, __pyx_int_112105877, __pyx_int_184977713); if (unlikely(!__pyx_tuple__8)) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__8); + __Pyx_GIVEREF(__pyx_tuple__8); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1043 + * __pyx_import_array() + * except Exception: + * raise ImportError("numpy._core.multiarray failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_umath() except -1: + */ + __pyx_tuple__9 = PyTuple_Pack(1, __pyx_kp_u_numpy__core_multiarray_failed_to); if (unlikely(!__pyx_tuple__9)) __PYX_ERR(2, 1043, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__9); + __Pyx_GIVEREF(__pyx_tuple__9); + + /* "../../../../tmp/pip-build-env-7z019gw3/normal/lib/python3.10/site-packages/numpy/__init__.cython-30.pxd":1049 + * _import_umath() + * except Exception: + * raise ImportError("numpy._core.umath failed to import") # <<<<<<<<<<<<<< + * + * cdef inline int import_ufunc() except -1: + */ + __pyx_tuple__10 = PyTuple_Pack(1, __pyx_kp_u_numpy__core_umath_failed_to_impo); if (unlikely(!__pyx_tuple__10)) __PYX_ERR(2, 1049, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__10); + __Pyx_GIVEREF(__pyx_tuple__10); + + /* "fairseq/data/token_block_utils_fast.pyx":99 + * slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + * cumsum = sizes.cumsum(axis=0) + * slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] # <<<<<<<<<<<<<< + * slice_indices[:, 1] = cumsum + * else: + */ + __pyx_slice__11 = PySlice_New(__pyx_int_1, Py_None, Py_None); if (unlikely(!__pyx_slice__11)) __PYX_ERR(0, 99, __pyx_L1_error) + __Pyx_GOTREF(__pyx_slice__11); + __Pyx_GIVEREF(__pyx_slice__11); + __pyx_tuple__12 = PyTuple_Pack(2, __pyx_slice__11, __pyx_int_0); if (unlikely(!__pyx_tuple__12)) __PYX_ERR(0, 99, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__12); + __Pyx_GIVEREF(__pyx_tuple__12); + + /* "fairseq/data/token_block_utils_fast.pyx":100 + * cumsum = sizes.cumsum(axis=0) + * slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] + * slice_indices[:, 1] = cumsum # <<<<<<<<<<<<<< + * else: + * raise ValueError('Invalid break_mode: ' + break_mode) + */ + __pyx_tuple__13 = PyTuple_Pack(2, __pyx_slice__5, __pyx_int_1); if (unlikely(!__pyx_tuple__13)) __PYX_ERR(0, 100, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__13); + __Pyx_GIVEREF(__pyx_tuple__13); + + /* "(tree fragment)":4 + * cdef object __pyx_PickleError + * cdef object __pyx_result + * if __pyx_checksum not in (0x8c67b45, 0x2e2dd22, 0x6632805): # <<<<<<<<<<<<<< + * from pickle import PickleError as __pyx_PickleError + * raise __pyx_PickleError, "Incompatible checksums (0x%x vs (0x8c67b45, 0x2e2dd22, 0x6632805) = (current_i, current_index, current_offset, sizes))" % __pyx_checksum + */ + __pyx_tuple__14 = PyTuple_Pack(3, __pyx_int_147225413, __pyx_int_48422178, __pyx_int_107161605); if (unlikely(!__pyx_tuple__14)) __PYX_ERR(1, 4, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__14); + __Pyx_GIVEREF(__pyx_tuple__14); + + /* "View.MemoryView":100 + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: + * if __import__("sys").version_info >= (3, 3): # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + */ + __pyx_tuple__15 = PyTuple_Pack(1, __pyx_n_s_sys); if (unlikely(!__pyx_tuple__15)) __PYX_ERR(1, 100, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__15); + __Pyx_GIVEREF(__pyx_tuple__15); + __pyx_tuple__16 = PyTuple_Pack(2, __pyx_int_3, __pyx_int_3); if (unlikely(!__pyx_tuple__16)) __PYX_ERR(1, 100, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__16); + __Pyx_GIVEREF(__pyx_tuple__16); + + /* "View.MemoryView":101 + * try: + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence # <<<<<<<<<<<<<< + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence + */ + __pyx_tuple__17 = PyTuple_Pack(1, __pyx_kp_s_collections_abc); if (unlikely(!__pyx_tuple__17)) __PYX_ERR(1, 101, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__17); + __Pyx_GIVEREF(__pyx_tuple__17); + + /* "View.MemoryView":103 + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence # <<<<<<<<<<<<<< + * except: + * + */ + __pyx_tuple__18 = PyTuple_Pack(1, __pyx_n_s_collections); if (unlikely(!__pyx_tuple__18)) __PYX_ERR(1, 103, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__18); + __Pyx_GIVEREF(__pyx_tuple__18); + + /* "View.MemoryView":309 + * return self.name + * + * cdef generic = Enum("") # <<<<<<<<<<<<<< + * cdef strided = Enum("") # default + * cdef indirect = Enum("") + */ + __pyx_tuple__19 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct_or_indirect); if (unlikely(!__pyx_tuple__19)) __PYX_ERR(1, 309, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__19); + __Pyx_GIVEREF(__pyx_tuple__19); + + /* "View.MemoryView":310 + * + * cdef generic = Enum("") + * cdef strided = Enum("") # default # <<<<<<<<<<<<<< + * cdef indirect = Enum("") + * + */ + __pyx_tuple__20 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct); if (unlikely(!__pyx_tuple__20)) __PYX_ERR(1, 310, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__20); + __Pyx_GIVEREF(__pyx_tuple__20); + + /* "View.MemoryView":311 + * cdef generic = Enum("") + * cdef strided = Enum("") # default + * cdef indirect = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_tuple__21 = PyTuple_Pack(1, __pyx_kp_s_strided_and_indirect); if (unlikely(!__pyx_tuple__21)) __PYX_ERR(1, 311, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__21); + __Pyx_GIVEREF(__pyx_tuple__21); + + /* "View.MemoryView":314 + * + * + * cdef contiguous = Enum("") # <<<<<<<<<<<<<< + * cdef indirect_contiguous = Enum("") + * + */ + __pyx_tuple__22 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_direct); if (unlikely(!__pyx_tuple__22)) __PYX_ERR(1, 314, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__22); + __Pyx_GIVEREF(__pyx_tuple__22); + + /* "View.MemoryView":315 + * + * cdef contiguous = Enum("") + * cdef indirect_contiguous = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_tuple__23 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_indirect); if (unlikely(!__pyx_tuple__23)) __PYX_ERR(1, 315, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__23); + __Pyx_GIVEREF(__pyx_tuple__23); + + /* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + __pyx_tuple__24 = PyTuple_Pack(5, __pyx_n_s_pyx_type, __pyx_n_s_pyx_checksum, __pyx_n_s_pyx_state, __pyx_n_s_pyx_PickleError, __pyx_n_s_pyx_result); if (unlikely(!__pyx_tuple__24)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__24); + __Pyx_GIVEREF(__pyx_tuple__24); + __pyx_codeobj__25 = (PyObject*)__Pyx_PyCode_New(3, 0, 0, 5, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__24, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_stringsource, __pyx_n_s_pyx_unpickle_Enum, 1, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__25)) __PYX_ERR(1, 1, __pyx_L1_error) + + /* "fairseq/data/token_block_utils_fast.pyx":50 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): # <<<<<<<<<<<<<< + * cdef DTYPE_t tok_idx = 0 + * cdef DTYPE_t sz_idx = 0 + */ + __pyx_tuple__26 = PyTuple_Pack(4, __pyx_n_s_sizes, __pyx_n_s_break_mode, __pyx_n_s_block_size, __pyx_n_s_document_sep_len); if (unlikely(!__pyx_tuple__26)) __PYX_ERR(0, 50, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__26); + __Pyx_GIVEREF(__pyx_tuple__26); + __pyx_codeobj__27 = (PyObject*)__Pyx_PyCode_New(4, 0, 0, 4, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__26, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_fairseq_data_token_block_utils_f, __pyx_n_s_get_slice_indices_fast, 50, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__27)) __PYX_ERR(0, 50, __pyx_L1_error) + + /* "fairseq/data/token_block_utils_fast.pyx":109 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_block_to_dataset_index_fast(np.ndarray[DTYPE_t, ndim=1] sizes, np.ndarray[DTYPE_t, ndim=2] slice_indices): # <<<<<<<<<<<<<< + * cdef DTYPE_t start_ds_idx + * cdef DTYPE_t start_offset + */ + __pyx_tuple__28 = PyTuple_Pack(2, __pyx_n_s_sizes, __pyx_n_s_slice_indices); if (unlikely(!__pyx_tuple__28)) __PYX_ERR(0, 109, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__28); + __Pyx_GIVEREF(__pyx_tuple__28); + __pyx_codeobj__29 = (PyObject*)__Pyx_PyCode_New(2, 0, 0, 2, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__28, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_fairseq_data_token_block_utils_f, __pyx_n_s_get_block_to_dataset_index_fast, 109, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__29)) __PYX_ERR(0, 109, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + __pyx_tuple__30 = PyTuple_Pack(4, __pyx_n_s_self, __pyx_n_s_state, __pyx_n_s_dict_2, __pyx_n_s_use_setstate); if (unlikely(!__pyx_tuple__30)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__30); + __Pyx_GIVEREF(__pyx_tuple__30); + __pyx_codeobj__31 = (PyObject*)__Pyx_PyCode_New(1, 0, 0, 4, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__30, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_stringsource, __pyx_n_s_reduce_cython, 1, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__31)) __PYX_ERR(1, 1, __pyx_L1_error) + + /* "(tree fragment)":16 + * else: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_DatasetSearcher__set_state(self, __pyx_state) + */ + __pyx_tuple__32 = PyTuple_Pack(2, __pyx_n_s_self, __pyx_n_s_pyx_state); if (unlikely(!__pyx_tuple__32)) __PYX_ERR(1, 16, __pyx_L1_error) + __Pyx_GOTREF(__pyx_tuple__32); + __Pyx_GIVEREF(__pyx_tuple__32); + __pyx_codeobj__33 = (PyObject*)__Pyx_PyCode_New(2, 0, 0, 2, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__32, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_stringsource, __pyx_n_s_setstate_cython, 16, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__33)) __PYX_ERR(1, 16, __pyx_L1_error) + + /* "(tree fragment)":1 + * def __pyx_unpickle_DatasetSearcher(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + __pyx_codeobj__34 = (PyObject*)__Pyx_PyCode_New(3, 0, 0, 5, 0, CO_OPTIMIZED|CO_NEWLOCALS, __pyx_empty_bytes, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_tuple__24, __pyx_empty_tuple, __pyx_empty_tuple, __pyx_kp_s_stringsource, __pyx_n_s_pyx_unpickle_DatasetSearcher, 1, __pyx_empty_bytes); if (unlikely(!__pyx_codeobj__34)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_RefNannyFinishContext(); + return 0; + __pyx_L1_error:; + __Pyx_RefNannyFinishContext(); + return -1; +} +/* #### Code section: init_constants ### */ + +static CYTHON_SMALL_CODE int __Pyx_InitConstants(void) { + if (__Pyx_CreateStringTabAndInitStrings() < 0) __PYX_ERR(0, 1, __pyx_L1_error); + __pyx_int_0 = PyInt_FromLong(0); if (unlikely(!__pyx_int_0)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_1 = PyInt_FromLong(1); if (unlikely(!__pyx_int_1)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_2 = PyInt_FromLong(2); if (unlikely(!__pyx_int_2)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_3 = PyInt_FromLong(3); if (unlikely(!__pyx_int_3)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_48422178 = PyInt_FromLong(48422178L); if (unlikely(!__pyx_int_48422178)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_107161605 = PyInt_FromLong(107161605L); if (unlikely(!__pyx_int_107161605)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_112105877 = PyInt_FromLong(112105877L); if (unlikely(!__pyx_int_112105877)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_136983863 = PyInt_FromLong(136983863L); if (unlikely(!__pyx_int_136983863)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_147225413 = PyInt_FromLong(147225413L); if (unlikely(!__pyx_int_147225413)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_184977713 = PyInt_FromLong(184977713L); if (unlikely(!__pyx_int_184977713)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_int_neg_1 = PyInt_FromLong(-1); if (unlikely(!__pyx_int_neg_1)) __PYX_ERR(0, 1, __pyx_L1_error) + return 0; + __pyx_L1_error:; + return -1; +} +/* #### Code section: init_globals ### */ + +static CYTHON_SMALL_CODE int __Pyx_InitGlobals(void) { + /* AssertionsEnabled.init */ + if (likely(__Pyx_init_assertions_enabled() == 0)); else + +if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 1, __pyx_L1_error) + + /* NumpyImportArray.init */ + /* + * Cython has automatically inserted a call to _import_array since + * you didn't include one when you cimported numpy. To disable this + * add the line + * numpy._import_array + */ +#ifdef NPY_FEATURE_VERSION +#ifndef NO_IMPORT_ARRAY +if (unlikely(_import_array() == -1)) { + PyErr_SetString(PyExc_ImportError, "numpy.core.multiarray failed to import " + "(auto-generated because you didn't call 'numpy.import_array()' after cimporting numpy; " + "use 'numpy._import_array' to disable if you are certain you don't need it)."); +} +#endif +#endif + +if (unlikely(PyErr_Occurred())) __PYX_ERR(0, 1, __pyx_L1_error) + + return 0; + __pyx_L1_error:; + return -1; +} +/* #### Code section: init_module ### */ + +static CYTHON_SMALL_CODE int __Pyx_modinit_global_init_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_variable_export_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_function_export_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_type_init_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_type_import_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_variable_import_code(void); /*proto*/ +static CYTHON_SMALL_CODE int __Pyx_modinit_function_import_code(void); /*proto*/ + +static int __Pyx_modinit_global_init_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_global_init_code", 0); + /*--- Global init code ---*/ + __pyx_collections_abc_Sequence = Py_None; Py_INCREF(Py_None); + generic = Py_None; Py_INCREF(Py_None); + strided = Py_None; Py_INCREF(Py_None); + indirect = Py_None; Py_INCREF(Py_None); + contiguous = Py_None; Py_INCREF(Py_None); + indirect_contiguous = Py_None; Py_INCREF(Py_None); + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_variable_export_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_variable_export_code", 0); + /*--- Variable export code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_function_export_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_function_export_code", 0); + /*--- Function export code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_type_init_code(void) { + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__Pyx_modinit_type_init_code", 0); + /*--- Type init code ---*/ + __pyx_vtabptr_7fairseq_4data_22token_block_utils_fast_DatasetSearcher = &__pyx_vtable_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; + __pyx_vtable_7fairseq_4data_22token_block_utils_fast_DatasetSearcher.reset = (PyObject *(*)(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *))__pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_reset; + __pyx_vtable_7fairseq_4data_22token_block_utils_fast_DatasetSearcher.step = (int (*)(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t))__pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_step; + __pyx_vtable_7fairseq_4data_22token_block_utils_fast_DatasetSearcher.seek = (PyObject *(*)(struct __pyx_obj_7fairseq_4data_22token_block_utils_fast_DatasetSearcher *, __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t))__pyx_f_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_seek; + #if CYTHON_USE_TYPE_SPECS + __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher_spec, NULL); if (unlikely(!__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher)) __PYX_ERR(0, 139, __pyx_L1_error) + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher_spec, __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher) < 0) __PYX_ERR(0, 139, __pyx_L1_error) + #else + __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher = &__pyx_type_7fairseq_4data_22token_block_utils_fast_DatasetSearcher; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher) < 0) __PYX_ERR(0, 139, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher->tp_print = 0; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher->tp_dictoffset && __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher->tp_getattro == PyObject_GenericGetAttr)) { + __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher->tp_getattro = __Pyx_PyObject_GenericGetAttr; + } + #endif + if (__Pyx_SetVtable(__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher, __pyx_vtabptr_7fairseq_4data_22token_block_utils_fast_DatasetSearcher) < 0) __PYX_ERR(0, 139, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_MergeVtables(__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher) < 0) __PYX_ERR(0, 139, __pyx_L1_error) + #endif + if (PyObject_SetAttr(__pyx_m, __pyx_n_s_DatasetSearcher, (PyObject *) __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher) < 0) __PYX_ERR(0, 139, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher) < 0) __PYX_ERR(0, 139, __pyx_L1_error) + #endif + __pyx_vtabptr_array = &__pyx_vtable_array; + __pyx_vtable_array.get_memview = (PyObject *(*)(struct __pyx_array_obj *))__pyx_array_get_memview; + #if CYTHON_USE_TYPE_SPECS + __pyx_array_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_array_spec, NULL); if (unlikely(!__pyx_array_type)) __PYX_ERR(1, 114, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + __pyx_array_type->tp_as_buffer = &__pyx_tp_as_buffer_array; + if (!__pyx_array_type->tp_as_buffer->bf_releasebuffer && __pyx_array_type->tp_base->tp_as_buffer && __pyx_array_type->tp_base->tp_as_buffer->bf_releasebuffer) { + __pyx_array_type->tp_as_buffer->bf_releasebuffer = __pyx_array_type->tp_base->tp_as_buffer->bf_releasebuffer; + } + #elif defined(Py_bf_getbuffer) && defined(Py_bf_releasebuffer) + /* PY_VERSION_HEX >= 0x03090000 || Py_LIMITED_API >= 0x030B0000 */ + #elif defined(_MSC_VER) + #pragma message ("The buffer protocol is not supported in the Limited C-API < 3.11.") + #else + #warning "The buffer protocol is not supported in the Limited C-API < 3.11." + #endif + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_array_spec, __pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #else + __pyx_array_type = &__pyx_type___pyx_array; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_array_type->tp_print = 0; + #endif + if (__Pyx_SetVtable(__pyx_array_type, __pyx_vtabptr_array) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_MergeVtables(__pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_array_type) < 0) __PYX_ERR(1, 114, __pyx_L1_error) + #endif + #if CYTHON_USE_TYPE_SPECS + __pyx_MemviewEnum_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_MemviewEnum_spec, NULL); if (unlikely(!__pyx_MemviewEnum_type)) __PYX_ERR(1, 302, __pyx_L1_error) + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_MemviewEnum_spec, __pyx_MemviewEnum_type) < 0) __PYX_ERR(1, 302, __pyx_L1_error) + #else + __pyx_MemviewEnum_type = &__pyx_type___pyx_MemviewEnum; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_MemviewEnum_type) < 0) __PYX_ERR(1, 302, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_MemviewEnum_type->tp_print = 0; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_MemviewEnum_type->tp_dictoffset && __pyx_MemviewEnum_type->tp_getattro == PyObject_GenericGetAttr)) { + __pyx_MemviewEnum_type->tp_getattro = __Pyx_PyObject_GenericGetAttr; + } + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_MemviewEnum_type) < 0) __PYX_ERR(1, 302, __pyx_L1_error) + #endif + __pyx_vtabptr_memoryview = &__pyx_vtable_memoryview; + __pyx_vtable_memoryview.get_item_pointer = (char *(*)(struct __pyx_memoryview_obj *, PyObject *))__pyx_memoryview_get_item_pointer; + __pyx_vtable_memoryview.is_slice = (PyObject *(*)(struct __pyx_memoryview_obj *, PyObject *))__pyx_memoryview_is_slice; + __pyx_vtable_memoryview.setitem_slice_assignment = (PyObject *(*)(struct __pyx_memoryview_obj *, PyObject *, PyObject *))__pyx_memoryview_setitem_slice_assignment; + __pyx_vtable_memoryview.setitem_slice_assign_scalar = (PyObject *(*)(struct __pyx_memoryview_obj *, struct __pyx_memoryview_obj *, PyObject *))__pyx_memoryview_setitem_slice_assign_scalar; + __pyx_vtable_memoryview.setitem_indexed = (PyObject *(*)(struct __pyx_memoryview_obj *, PyObject *, PyObject *))__pyx_memoryview_setitem_indexed; + __pyx_vtable_memoryview.convert_item_to_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *))__pyx_memoryview_convert_item_to_object; + __pyx_vtable_memoryview.assign_item_from_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *, PyObject *))__pyx_memoryview_assign_item_from_object; + __pyx_vtable_memoryview._get_base = (PyObject *(*)(struct __pyx_memoryview_obj *))__pyx_memoryview__get_base; + #if CYTHON_USE_TYPE_SPECS + __pyx_memoryview_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_memoryview_spec, NULL); if (unlikely(!__pyx_memoryview_type)) __PYX_ERR(1, 337, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + __pyx_memoryview_type->tp_as_buffer = &__pyx_tp_as_buffer_memoryview; + if (!__pyx_memoryview_type->tp_as_buffer->bf_releasebuffer && __pyx_memoryview_type->tp_base->tp_as_buffer && __pyx_memoryview_type->tp_base->tp_as_buffer->bf_releasebuffer) { + __pyx_memoryview_type->tp_as_buffer->bf_releasebuffer = __pyx_memoryview_type->tp_base->tp_as_buffer->bf_releasebuffer; + } + #elif defined(Py_bf_getbuffer) && defined(Py_bf_releasebuffer) + /* PY_VERSION_HEX >= 0x03090000 || Py_LIMITED_API >= 0x030B0000 */ + #elif defined(_MSC_VER) + #pragma message ("The buffer protocol is not supported in the Limited C-API < 3.11.") + #else + #warning "The buffer protocol is not supported in the Limited C-API < 3.11." + #endif + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_memoryview_spec, __pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #else + __pyx_memoryview_type = &__pyx_type___pyx_memoryview; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_memoryview_type->tp_print = 0; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_memoryview_type->tp_dictoffset && __pyx_memoryview_type->tp_getattro == PyObject_GenericGetAttr)) { + __pyx_memoryview_type->tp_getattro = __Pyx_PyObject_GenericGetAttr; + } + #endif + if (__Pyx_SetVtable(__pyx_memoryview_type, __pyx_vtabptr_memoryview) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_MergeVtables(__pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_memoryview_type) < 0) __PYX_ERR(1, 337, __pyx_L1_error) + #endif + __pyx_vtabptr__memoryviewslice = &__pyx_vtable__memoryviewslice; + __pyx_vtable__memoryviewslice.__pyx_base = *__pyx_vtabptr_memoryview; + __pyx_vtable__memoryviewslice.__pyx_base.convert_item_to_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *))__pyx_memoryviewslice_convert_item_to_object; + __pyx_vtable__memoryviewslice.__pyx_base.assign_item_from_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *, PyObject *))__pyx_memoryviewslice_assign_item_from_object; + __pyx_vtable__memoryviewslice.__pyx_base._get_base = (PyObject *(*)(struct __pyx_memoryview_obj *))__pyx_memoryviewslice__get_base; + #if CYTHON_USE_TYPE_SPECS + __pyx_t_1 = PyTuple_Pack(1, (PyObject *)__pyx_memoryview_type); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 952, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_memoryviewslice_type = (PyTypeObject *) __Pyx_PyType_FromModuleAndSpec(__pyx_m, &__pyx_type___pyx_memoryviewslice_spec, __pyx_t_1); + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + if (unlikely(!__pyx_memoryviewslice_type)) __PYX_ERR(1, 952, __pyx_L1_error) + if (__Pyx_fix_up_extension_type_from_spec(&__pyx_type___pyx_memoryviewslice_spec, __pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #else + __pyx_memoryviewslice_type = &__pyx_type___pyx_memoryviewslice; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + __pyx_memoryviewslice_type->tp_base = __pyx_memoryview_type; + #endif + #if !CYTHON_USE_TYPE_SPECS + if (__Pyx_PyType_Ready(__pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #endif + #if PY_MAJOR_VERSION < 3 + __pyx_memoryviewslice_type->tp_print = 0; + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_memoryviewslice_type->tp_dictoffset && __pyx_memoryviewslice_type->tp_getattro == PyObject_GenericGetAttr)) { + __pyx_memoryviewslice_type->tp_getattro = __Pyx_PyObject_GenericGetAttr; + } + #endif + if (__Pyx_SetVtable(__pyx_memoryviewslice_type, __pyx_vtabptr__memoryviewslice) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_MergeVtables(__pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #endif + #if !CYTHON_COMPILING_IN_LIMITED_API + if (__Pyx_setup_reduce((PyObject *) __pyx_memoryviewslice_type) < 0) __PYX_ERR(1, 952, __pyx_L1_error) + #endif + __Pyx_RefNannyFinishContext(); + return 0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_RefNannyFinishContext(); + return -1; +} + +static int __Pyx_modinit_type_import_code(void) { + __Pyx_RefNannyDeclarations + PyObject *__pyx_t_1 = NULL; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannySetupContext("__Pyx_modinit_type_import_code", 0); + /*--- Type import code ---*/ + __pyx_t_1 = PyImport_ImportModule(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_t_1)) __PYX_ERR(3, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_ptype_7cpython_4type_type = __Pyx_ImportType_3_0_12(__pyx_t_1, __Pyx_BUILTIN_MODULE_NAME, "type", + #if defined(PYPY_VERSION_NUM) && PYPY_VERSION_NUM < 0x050B0000 + sizeof(PyTypeObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyTypeObject), + #elif CYTHON_COMPILING_IN_LIMITED_API + sizeof(PyTypeObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyTypeObject), + #else + sizeof(PyHeapTypeObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyHeapTypeObject), + #endif + __Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_7cpython_4type_type) __PYX_ERR(3, 9, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __pyx_t_1 = PyImport_ImportModule("numpy"); if (unlikely(!__pyx_t_1)) __PYX_ERR(2, 272, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_1); + __pyx_ptype_5numpy_dtype = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "dtype", sizeof(PyArray_Descr), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArray_Descr),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_dtype) __PYX_ERR(2, 272, __pyx_L1_error) + __pyx_ptype_5numpy_flatiter = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "flatiter", sizeof(PyArrayIterObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArrayIterObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_flatiter) __PYX_ERR(2, 317, __pyx_L1_error) + __pyx_ptype_5numpy_broadcast = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "broadcast", sizeof(PyArrayMultiIterObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArrayMultiIterObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_broadcast) __PYX_ERR(2, 321, __pyx_L1_error) + __pyx_ptype_5numpy_ndarray = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "ndarray", sizeof(PyArrayObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyArrayObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_ndarray) __PYX_ERR(2, 360, __pyx_L1_error) + __pyx_ptype_5numpy_generic = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "generic", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_generic) __PYX_ERR(2, 865, __pyx_L1_error) + __pyx_ptype_5numpy_number = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "number", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_number) __PYX_ERR(2, 867, __pyx_L1_error) + __pyx_ptype_5numpy_integer = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "integer", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_integer) __PYX_ERR(2, 869, __pyx_L1_error) + __pyx_ptype_5numpy_signedinteger = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "signedinteger", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_signedinteger) __PYX_ERR(2, 871, __pyx_L1_error) + __pyx_ptype_5numpy_unsignedinteger = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "unsignedinteger", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_unsignedinteger) __PYX_ERR(2, 873, __pyx_L1_error) + __pyx_ptype_5numpy_inexact = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "inexact", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_inexact) __PYX_ERR(2, 875, __pyx_L1_error) + __pyx_ptype_5numpy_floating = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "floating", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_floating) __PYX_ERR(2, 877, __pyx_L1_error) + __pyx_ptype_5numpy_complexfloating = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "complexfloating", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_complexfloating) __PYX_ERR(2, 879, __pyx_L1_error) + __pyx_ptype_5numpy_flexible = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "flexible", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_flexible) __PYX_ERR(2, 881, __pyx_L1_error) + __pyx_ptype_5numpy_character = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "character", sizeof(PyObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyObject),__Pyx_ImportType_CheckSize_Warn_3_0_12); if (!__pyx_ptype_5numpy_character) __PYX_ERR(2, 883, __pyx_L1_error) + __pyx_ptype_5numpy_ufunc = __Pyx_ImportType_3_0_12(__pyx_t_1, "numpy", "ufunc", sizeof(PyUFuncObject), __PYX_GET_STRUCT_ALIGNMENT_3_0_12(PyUFuncObject),__Pyx_ImportType_CheckSize_Ignore_3_0_12); if (!__pyx_ptype_5numpy_ufunc) __PYX_ERR(2, 947, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_RefNannyFinishContext(); + return 0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_1); + __Pyx_RefNannyFinishContext(); + return -1; +} + +static int __Pyx_modinit_variable_import_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_variable_import_code", 0); + /*--- Variable import code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + +static int __Pyx_modinit_function_import_code(void) { + __Pyx_RefNannyDeclarations + __Pyx_RefNannySetupContext("__Pyx_modinit_function_import_code", 0); + /*--- Function import code ---*/ + __Pyx_RefNannyFinishContext(); + return 0; +} + + +#if PY_MAJOR_VERSION >= 3 +#if CYTHON_PEP489_MULTI_PHASE_INIT +static PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def); /*proto*/ +static int __pyx_pymod_exec_token_block_utils_fast(PyObject* module); /*proto*/ +static PyModuleDef_Slot __pyx_moduledef_slots[] = { + {Py_mod_create, (void*)__pyx_pymod_create}, + {Py_mod_exec, (void*)__pyx_pymod_exec_token_block_utils_fast}, + {0, NULL} +}; +#endif + +#ifdef __cplusplus +namespace { + struct PyModuleDef __pyx_moduledef = + #else + static struct PyModuleDef __pyx_moduledef = + #endif + { + PyModuleDef_HEAD_INIT, + "token_block_utils_fast", + 0, /* m_doc */ + #if CYTHON_PEP489_MULTI_PHASE_INIT + 0, /* m_size */ + #elif CYTHON_USE_MODULE_STATE + sizeof(__pyx_mstate), /* m_size */ + #else + -1, /* m_size */ + #endif + __pyx_methods /* m_methods */, + #if CYTHON_PEP489_MULTI_PHASE_INIT + __pyx_moduledef_slots, /* m_slots */ + #else + NULL, /* m_reload */ + #endif + #if CYTHON_USE_MODULE_STATE + __pyx_m_traverse, /* m_traverse */ + __pyx_m_clear, /* m_clear */ + NULL /* m_free */ + #else + NULL, /* m_traverse */ + NULL, /* m_clear */ + NULL /* m_free */ + #endif + }; + #ifdef __cplusplus +} /* anonymous namespace */ +#endif +#endif + +#ifndef CYTHON_NO_PYINIT_EXPORT +#define __Pyx_PyMODINIT_FUNC PyMODINIT_FUNC +#elif PY_MAJOR_VERSION < 3 +#ifdef __cplusplus +#define __Pyx_PyMODINIT_FUNC extern "C" void +#else +#define __Pyx_PyMODINIT_FUNC void +#endif +#else +#ifdef __cplusplus +#define __Pyx_PyMODINIT_FUNC extern "C" PyObject * +#else +#define __Pyx_PyMODINIT_FUNC PyObject * +#endif +#endif + + +#if PY_MAJOR_VERSION < 3 +__Pyx_PyMODINIT_FUNC inittoken_block_utils_fast(void) CYTHON_SMALL_CODE; /*proto*/ +__Pyx_PyMODINIT_FUNC inittoken_block_utils_fast(void) +#else +__Pyx_PyMODINIT_FUNC PyInit_token_block_utils_fast(void) CYTHON_SMALL_CODE; /*proto*/ +__Pyx_PyMODINIT_FUNC PyInit_token_block_utils_fast(void) +#if CYTHON_PEP489_MULTI_PHASE_INIT +{ + return PyModuleDef_Init(&__pyx_moduledef); +} +static CYTHON_SMALL_CODE int __Pyx_check_single_interpreter(void) { + #if PY_VERSION_HEX >= 0x030700A1 + static PY_INT64_T main_interpreter_id = -1; + PY_INT64_T current_id = PyInterpreterState_GetID(PyThreadState_Get()->interp); + if (main_interpreter_id == -1) { + main_interpreter_id = current_id; + return (unlikely(current_id == -1)) ? -1 : 0; + } else if (unlikely(main_interpreter_id != current_id)) + #else + static PyInterpreterState *main_interpreter = NULL; + PyInterpreterState *current_interpreter = PyThreadState_Get()->interp; + if (!main_interpreter) { + main_interpreter = current_interpreter; + } else if (unlikely(main_interpreter != current_interpreter)) + #endif + { + PyErr_SetString( + PyExc_ImportError, + "Interpreter change detected - this module can only be loaded into one interpreter per process."); + return -1; + } + return 0; +} +#if CYTHON_COMPILING_IN_LIMITED_API +static CYTHON_SMALL_CODE int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *module, const char* from_name, const char* to_name, int allow_none) +#else +static CYTHON_SMALL_CODE int __Pyx_copy_spec_to_module(PyObject *spec, PyObject *moddict, const char* from_name, const char* to_name, int allow_none) +#endif +{ + PyObject *value = PyObject_GetAttrString(spec, from_name); + int result = 0; + if (likely(value)) { + if (allow_none || value != Py_None) { +#if CYTHON_COMPILING_IN_LIMITED_API + result = PyModule_AddObject(module, to_name, value); +#else + result = PyDict_SetItemString(moddict, to_name, value); +#endif + } + Py_DECREF(value); + } else if (PyErr_ExceptionMatches(PyExc_AttributeError)) { + PyErr_Clear(); + } else { + result = -1; + } + return result; +} +static CYTHON_SMALL_CODE PyObject* __pyx_pymod_create(PyObject *spec, PyModuleDef *def) { + PyObject *module = NULL, *moddict, *modname; + CYTHON_UNUSED_VAR(def); + if (__Pyx_check_single_interpreter()) + return NULL; + if (__pyx_m) + return __Pyx_NewRef(__pyx_m); + modname = PyObject_GetAttrString(spec, "name"); + if (unlikely(!modname)) goto bad; + module = PyModule_NewObject(modname); + Py_DECREF(modname); + if (unlikely(!module)) goto bad; +#if CYTHON_COMPILING_IN_LIMITED_API + moddict = module; +#else + moddict = PyModule_GetDict(module); + if (unlikely(!moddict)) goto bad; +#endif + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "loader", "__loader__", 1) < 0)) goto bad; + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "origin", "__file__", 1) < 0)) goto bad; + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "parent", "__package__", 1) < 0)) goto bad; + if (unlikely(__Pyx_copy_spec_to_module(spec, moddict, "submodule_search_locations", "__path__", 0) < 0)) goto bad; + return module; +bad: + Py_XDECREF(module); + return NULL; +} + + +static CYTHON_SMALL_CODE int __pyx_pymod_exec_token_block_utils_fast(PyObject *__pyx_pyinit_module) +#endif +#endif +{ + int stringtab_initialized = 0; + #if CYTHON_USE_MODULE_STATE + int pystate_addmodule_run = 0; + #endif + PyObject *__pyx_t_1 = NULL; + PyObject *__pyx_t_2 = NULL; + PyObject *__pyx_t_3 = NULL; + PyObject *__pyx_t_4 = NULL; + PyObject *__pyx_t_5 = NULL; + int __pyx_t_6; + PyObject *__pyx_t_7 = NULL; + static PyThread_type_lock __pyx_t_8[8]; + int __pyx_lineno = 0; + const char *__pyx_filename = NULL; + int __pyx_clineno = 0; + __Pyx_RefNannyDeclarations + #if CYTHON_PEP489_MULTI_PHASE_INIT + if (__pyx_m) { + if (__pyx_m == __pyx_pyinit_module) return 0; + PyErr_SetString(PyExc_RuntimeError, "Module 'token_block_utils_fast' has already been imported. Re-initialisation is not supported."); + return -1; + } + #elif PY_MAJOR_VERSION >= 3 + if (__pyx_m) return __Pyx_NewRef(__pyx_m); + #endif + /*--- Module creation code ---*/ + #if CYTHON_PEP489_MULTI_PHASE_INIT + __pyx_m = __pyx_pyinit_module; + Py_INCREF(__pyx_m); + #else + #if PY_MAJOR_VERSION < 3 + __pyx_m = Py_InitModule4("token_block_utils_fast", __pyx_methods, 0, 0, PYTHON_API_VERSION); Py_XINCREF(__pyx_m); + if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error) + #elif CYTHON_USE_MODULE_STATE + __pyx_t_1 = PyModule_Create(&__pyx_moduledef); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 1, __pyx_L1_error) + { + int add_module_result = PyState_AddModule(__pyx_t_1, &__pyx_moduledef); + __pyx_t_1 = 0; /* transfer ownership from __pyx_t_1 to "token_block_utils_fast" pseudovariable */ + if (unlikely((add_module_result < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + pystate_addmodule_run = 1; + } + #else + __pyx_m = PyModule_Create(&__pyx_moduledef); + if (unlikely(!__pyx_m)) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #endif + CYTHON_UNUSED_VAR(__pyx_t_1); + __pyx_d = PyModule_GetDict(__pyx_m); if (unlikely(!__pyx_d)) __PYX_ERR(0, 1, __pyx_L1_error) + Py_INCREF(__pyx_d); + __pyx_b = __Pyx_PyImport_AddModuleRef(__Pyx_BUILTIN_MODULE_NAME); if (unlikely(!__pyx_b)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_cython_runtime = __Pyx_PyImport_AddModuleRef((const char *) "cython_runtime"); if (unlikely(!__pyx_cython_runtime)) __PYX_ERR(0, 1, __pyx_L1_error) + if (PyObject_SetAttrString(__pyx_m, "__builtins__", __pyx_b) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #if CYTHON_REFNANNY +__Pyx_RefNanny = __Pyx_RefNannyImportAPI("refnanny"); +if (!__Pyx_RefNanny) { + PyErr_Clear(); + __Pyx_RefNanny = __Pyx_RefNannyImportAPI("Cython.Runtime.refnanny"); + if (!__Pyx_RefNanny) + Py_FatalError("failed to import 'refnanny' module"); +} +#endif + __Pyx_RefNannySetupContext("__Pyx_PyMODINIT_FUNC PyInit_token_block_utils_fast(void)", 0); + if (__Pyx_check_binary_version(__PYX_LIMITED_VERSION_HEX, __Pyx_get_runtime_version(), CYTHON_COMPILING_IN_LIMITED_API) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #ifdef __Pxy_PyFrame_Initialize_Offsets + __Pxy_PyFrame_Initialize_Offsets(); + #endif + __pyx_empty_tuple = PyTuple_New(0); if (unlikely(!__pyx_empty_tuple)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_empty_bytes = PyBytes_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_bytes)) __PYX_ERR(0, 1, __pyx_L1_error) + __pyx_empty_unicode = PyUnicode_FromStringAndSize("", 0); if (unlikely(!__pyx_empty_unicode)) __PYX_ERR(0, 1, __pyx_L1_error) + #ifdef __Pyx_CyFunction_USED + if (__pyx_CyFunction_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_FusedFunction_USED + if (__pyx_FusedFunction_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_Coroutine_USED + if (__pyx_Coroutine_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_Generator_USED + if (__pyx_Generator_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_AsyncGen_USED + if (__pyx_AsyncGen_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + #ifdef __Pyx_StopAsyncIteration_USED + if (__pyx_StopAsyncIteration_init(__pyx_m) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + /*--- Library function declarations ---*/ + /*--- Threads initialization code ---*/ + #if defined(WITH_THREAD) && PY_VERSION_HEX < 0x030700F0 && defined(__PYX_FORCE_INIT_THREADS) && __PYX_FORCE_INIT_THREADS + PyEval_InitThreads(); + #endif + /*--- Initialize various global constants etc. ---*/ + if (__Pyx_InitConstants() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + stringtab_initialized = 1; + if (__Pyx_InitGlobals() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #if PY_MAJOR_VERSION < 3 && (__PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT) + if (__Pyx_init_sys_getdefaultencoding_params() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + if (__pyx_module_is_main_fairseq__data__token_block_utils_fast) { + if (PyObject_SetAttr(__pyx_m, __pyx_n_s_name_2, __pyx_n_s_main) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + } + #if PY_MAJOR_VERSION >= 3 + { + PyObject *modules = PyImport_GetModuleDict(); if (unlikely(!modules)) __PYX_ERR(0, 1, __pyx_L1_error) + if (!PyDict_GetItemString(modules, "fairseq.data.token_block_utils_fast")) { + if (unlikely((PyDict_SetItemString(modules, "fairseq.data.token_block_utils_fast", __pyx_m) < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + } + } + #endif + /*--- Builtin init code ---*/ + if (__Pyx_InitCachedBuiltins() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + /*--- Constants init code ---*/ + if (__Pyx_InitCachedConstants() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + /*--- Global type/function init code ---*/ + (void)__Pyx_modinit_global_init_code(); + (void)__Pyx_modinit_variable_export_code(); + (void)__Pyx_modinit_function_export_code(); + if (unlikely((__Pyx_modinit_type_init_code() < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + if (unlikely((__Pyx_modinit_type_import_code() < 0))) __PYX_ERR(0, 1, __pyx_L1_error) + (void)__Pyx_modinit_variable_import_code(); + (void)__Pyx_modinit_function_import_code(); + /*--- Execution code ---*/ + #if defined(__Pyx_Generator_USED) || defined(__Pyx_Coroutine_USED) + if (__Pyx_patch_abc() < 0) __PYX_ERR(0, 1, __pyx_L1_error) + #endif + + /* "View.MemoryView":99 + * + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: # <<<<<<<<<<<<<< + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "View.MemoryView":100 + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: + * if __import__("sys").version_info >= (3, 3): # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + */ + __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin___import__, __pyx_tuple__15, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_version_info); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_t_4 = PyObject_RichCompare(__pyx_t_5, __pyx_tuple__16, Py_GE); __Pyx_XGOTREF(__pyx_t_4); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_t_4); if (unlikely((__pyx_t_6 < 0))) __PYX_ERR(1, 100, __pyx_L2_error) + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + if (__pyx_t_6) { + + /* "View.MemoryView":101 + * try: + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence # <<<<<<<<<<<<<< + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence + */ + __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin___import__, __pyx_tuple__17, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 101, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_abc); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 101, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __pyx_t_4 = __Pyx_PyObject_GetAttrStr(__pyx_t_5, __pyx_n_s_Sequence); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 101, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XGOTREF(__pyx_collections_abc_Sequence); + __Pyx_DECREF_SET(__pyx_collections_abc_Sequence, __pyx_t_4); + __Pyx_GIVEREF(__pyx_t_4); + __pyx_t_4 = 0; + + /* "View.MemoryView":100 + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: + * if __import__("sys").version_info >= (3, 3): # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + */ + goto __pyx_L8; + } + + /* "View.MemoryView":103 + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence # <<<<<<<<<<<<<< + * except: + * + */ + /*else*/ { + __pyx_t_4 = __Pyx_PyObject_Call(__pyx_builtin___import__, __pyx_tuple__18, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 103, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_Sequence); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 103, __pyx_L2_error) + __Pyx_GOTREF(__pyx_t_5); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XGOTREF(__pyx_collections_abc_Sequence); + __Pyx_DECREF_SET(__pyx_collections_abc_Sequence, __pyx_t_5); + __Pyx_GIVEREF(__pyx_t_5); + __pyx_t_5 = 0; + } + __pyx_L8:; + + /* "View.MemoryView":99 + * + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: # <<<<<<<<<<<<<< + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L7_try_end; + __pyx_L2_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + + /* "View.MemoryView":104 + * else: + * __pyx_collections_abc_Sequence = __import__("collections").Sequence + * except: # <<<<<<<<<<<<<< + * + * __pyx_collections_abc_Sequence = None + */ + /*except:*/ { + __Pyx_AddTraceback("View.MemoryView", __pyx_clineno, __pyx_lineno, __pyx_filename); + if (__Pyx_GetException(&__pyx_t_5, &__pyx_t_4, &__pyx_t_7) < 0) __PYX_ERR(1, 104, __pyx_L4_except_error) + __Pyx_XGOTREF(__pyx_t_5); + __Pyx_XGOTREF(__pyx_t_4); + __Pyx_XGOTREF(__pyx_t_7); + + /* "View.MemoryView":106 + * except: + * + * __pyx_collections_abc_Sequence = None # <<<<<<<<<<<<<< + * + * + */ + __Pyx_INCREF(Py_None); + __Pyx_XGOTREF(__pyx_collections_abc_Sequence); + __Pyx_DECREF_SET(__pyx_collections_abc_Sequence, Py_None); + __Pyx_GIVEREF(Py_None); + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + goto __pyx_L3_exception_handled; + } + + /* "View.MemoryView":99 + * + * cdef object __pyx_collections_abc_Sequence "__pyx_collections_abc_Sequence" + * try: # <<<<<<<<<<<<<< + * if __import__("sys").version_info >= (3, 3): + * __pyx_collections_abc_Sequence = __import__("collections.abc").abc.Sequence + */ + __pyx_L4_except_error:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + goto __pyx_L1_error; + __pyx_L3_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + __pyx_L7_try_end:; + } + + /* "View.MemoryView":241 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_3, &__pyx_t_2, &__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_1); + /*try:*/ { + + /* "View.MemoryView":242 + * + * try: + * count = __pyx_collections_abc_Sequence.count # <<<<<<<<<<<<<< + * index = __pyx_collections_abc_Sequence.index + * except: + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_count); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 242, __pyx_L11_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_array_type, __pyx_n_s_count, __pyx_t_7) < 0) __PYX_ERR(1, 242, __pyx_L11_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_array_type); + + /* "View.MemoryView":243 + * try: + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index # <<<<<<<<<<<<<< + * except: + * pass + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_index); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 243, __pyx_L11_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_array_type, __pyx_n_s_index, __pyx_t_7) < 0) __PYX_ERR(1, 243, __pyx_L11_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_array_type); + + /* "View.MemoryView":241 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + } + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + goto __pyx_L16_try_end; + __pyx_L11_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":244 + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + * except: # <<<<<<<<<<<<<< + * pass + * + */ + /*except:*/ { + __Pyx_ErrRestore(0,0,0); + goto __pyx_L12_exception_handled; + } + __pyx_L12_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_2, __pyx_t_1); + __pyx_L16_try_end:; + } + + /* "View.MemoryView":309 + * return self.name + * + * cdef generic = Enum("") # <<<<<<<<<<<<<< + * cdef strided = Enum("") # default + * cdef indirect = Enum("") + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__19, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 309, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(generic); + __Pyx_DECREF_SET(generic, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":310 + * + * cdef generic = Enum("") + * cdef strided = Enum("") # default # <<<<<<<<<<<<<< + * cdef indirect = Enum("") + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__20, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 310, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(strided); + __Pyx_DECREF_SET(strided, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":311 + * cdef generic = Enum("") + * cdef strided = Enum("") # default + * cdef indirect = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__21, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 311, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(indirect); + __Pyx_DECREF_SET(indirect, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":314 + * + * + * cdef contiguous = Enum("") # <<<<<<<<<<<<<< + * cdef indirect_contiguous = Enum("") + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__22, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 314, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(contiguous); + __Pyx_DECREF_SET(contiguous, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":315 + * + * cdef contiguous = Enum("") + * cdef indirect_contiguous = Enum("") # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__23, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 315, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_XGOTREF(indirect_contiguous); + __Pyx_DECREF_SET(indirect_contiguous, __pyx_t_7); + __Pyx_GIVEREF(__pyx_t_7); + __pyx_t_7 = 0; + + /* "View.MemoryView":323 + * + * + * cdef int __pyx_memoryview_thread_locks_used = 0 # <<<<<<<<<<<<<< + * cdef PyThread_type_lock[8] __pyx_memoryview_thread_locks = [ + * PyThread_allocate_lock(), + */ + __pyx_memoryview_thread_locks_used = 0; + + /* "View.MemoryView":324 + * + * cdef int __pyx_memoryview_thread_locks_used = 0 + * cdef PyThread_type_lock[8] __pyx_memoryview_thread_locks = [ # <<<<<<<<<<<<<< + * PyThread_allocate_lock(), + * PyThread_allocate_lock(), + */ + __pyx_t_8[0] = PyThread_allocate_lock(); + __pyx_t_8[1] = PyThread_allocate_lock(); + __pyx_t_8[2] = PyThread_allocate_lock(); + __pyx_t_8[3] = PyThread_allocate_lock(); + __pyx_t_8[4] = PyThread_allocate_lock(); + __pyx_t_8[5] = PyThread_allocate_lock(); + __pyx_t_8[6] = PyThread_allocate_lock(); + __pyx_t_8[7] = PyThread_allocate_lock(); + memcpy(&(__pyx_memoryview_thread_locks[0]), __pyx_t_8, sizeof(__pyx_memoryview_thread_locks[0]) * (8)); + + /* "View.MemoryView":982 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_1, &__pyx_t_2, &__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_3); + /*try:*/ { + + /* "View.MemoryView":983 + * + * try: + * count = __pyx_collections_abc_Sequence.count # <<<<<<<<<<<<<< + * index = __pyx_collections_abc_Sequence.index + * except: + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_count); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 983, __pyx_L17_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_memoryviewslice_type, __pyx_n_s_count, __pyx_t_7) < 0) __PYX_ERR(1, 983, __pyx_L17_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_memoryviewslice_type); + + /* "View.MemoryView":984 + * try: + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index # <<<<<<<<<<<<<< + * except: + * pass + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_index); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 984, __pyx_L17_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict(__pyx_memoryviewslice_type, __pyx_n_s_index, __pyx_t_7) < 0) __PYX_ERR(1, 984, __pyx_L17_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_memoryviewslice_type); + + /* "View.MemoryView":982 + * + * + * try: # <<<<<<<<<<<<<< + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + */ + } + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + goto __pyx_L22_try_end; + __pyx_L17_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":985 + * count = __pyx_collections_abc_Sequence.count + * index = __pyx_collections_abc_Sequence.index + * except: # <<<<<<<<<<<<<< + * pass + * + */ + /*except:*/ { + __Pyx_ErrRestore(0,0,0); + goto __pyx_L18_exception_handled; + } + __pyx_L18_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_ExceptionReset(__pyx_t_1, __pyx_t_2, __pyx_t_3); + __pyx_L22_try_end:; + } + + /* "View.MemoryView":988 + * pass + * + * try: # <<<<<<<<<<<<<< + * if __pyx_collections_abc_Sequence: + * + */ + { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ExceptionSave(&__pyx_t_3, &__pyx_t_2, &__pyx_t_1); + __Pyx_XGOTREF(__pyx_t_3); + __Pyx_XGOTREF(__pyx_t_2); + __Pyx_XGOTREF(__pyx_t_1); + /*try:*/ { + + /* "View.MemoryView":989 + * + * try: + * if __pyx_collections_abc_Sequence: # <<<<<<<<<<<<<< + * + * + */ + __pyx_t_6 = __Pyx_PyObject_IsTrue(__pyx_collections_abc_Sequence); if (unlikely((__pyx_t_6 < 0))) __PYX_ERR(1, 989, __pyx_L23_error) + if (__pyx_t_6) { + + /* "View.MemoryView":993 + * + * + * __pyx_collections_abc_Sequence.register(_memoryviewslice) # <<<<<<<<<<<<<< + * __pyx_collections_abc_Sequence.register(array) + * except: + */ + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_register); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 993, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_7); + __pyx_t_4 = __Pyx_PyObject_CallOneArg(__pyx_t_7, ((PyObject *)__pyx_memoryviewslice_type)); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 993, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "View.MemoryView":994 + * + * __pyx_collections_abc_Sequence.register(_memoryviewslice) + * __pyx_collections_abc_Sequence.register(array) # <<<<<<<<<<<<<< + * except: + * pass # ignore failure, it's a minor issue + */ + __pyx_t_4 = __Pyx_PyObject_GetAttrStr(__pyx_collections_abc_Sequence, __pyx_n_s_register); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 994, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_7 = __Pyx_PyObject_CallOneArg(__pyx_t_4, ((PyObject *)__pyx_array_type)); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 994, __pyx_L23_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":989 + * + * try: + * if __pyx_collections_abc_Sequence: # <<<<<<<<<<<<<< + * + * + */ + } + + /* "View.MemoryView":988 + * pass + * + * try: # <<<<<<<<<<<<<< + * if __pyx_collections_abc_Sequence: + * + */ + } + __Pyx_XDECREF(__pyx_t_3); __pyx_t_3 = 0; + __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0; + __Pyx_XDECREF(__pyx_t_1); __pyx_t_1 = 0; + goto __pyx_L28_try_end; + __pyx_L23_error:; + __Pyx_XDECREF(__pyx_t_4); __pyx_t_4 = 0; + __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0; + __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "View.MemoryView":995 + * __pyx_collections_abc_Sequence.register(_memoryviewslice) + * __pyx_collections_abc_Sequence.register(array) + * except: # <<<<<<<<<<<<<< + * pass # ignore failure, it's a minor issue + * + */ + /*except:*/ { + __Pyx_ErrRestore(0,0,0); + goto __pyx_L24_exception_handled; + } + __pyx_L24_exception_handled:; + __Pyx_XGIVEREF(__pyx_t_3); + __Pyx_XGIVEREF(__pyx_t_2); + __Pyx_XGIVEREF(__pyx_t_1); + __Pyx_ExceptionReset(__pyx_t_3, __pyx_t_2, __pyx_t_1); + __pyx_L28_try_end:; + } + + /* "(tree fragment)":1 + * def __pyx_unpickle_Enum(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + __pyx_t_7 = PyCFunction_NewEx(&__pyx_mdef_15View_dot_MemoryView_1__pyx_unpickle_Enum, NULL, __pyx_n_s_View_MemoryView); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_pyx_unpickle_Enum, __pyx_t_7) < 0) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":7 + * # LICENSE file in the root directory of this source tree. + * + * import numpy as np # <<<<<<<<<<<<<< + * import torch + * from itertools import chain + */ + __pyx_t_7 = __Pyx_ImportDottedModule(__pyx_n_s_numpy, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 7, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_np, __pyx_t_7) < 0) __PYX_ERR(0, 7, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":8 + * + * import numpy as np + * import torch # <<<<<<<<<<<<<< + * from itertools import chain + * from libc.math cimport ceil + */ + __pyx_t_7 = __Pyx_ImportDottedModule(__pyx_n_s_torch, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 8, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_torch, __pyx_t_7) < 0) __PYX_ERR(0, 8, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":9 + * import numpy as np + * import torch + * from itertools import chain # <<<<<<<<<<<<<< + * from libc.math cimport ceil + * + */ + __pyx_t_7 = PyList_New(1); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_INCREF(__pyx_n_s_chain); + __Pyx_GIVEREF(__pyx_n_s_chain); + if (__Pyx_PyList_SET_ITEM(__pyx_t_7, 0, __pyx_n_s_chain)) __PYX_ERR(0, 9, __pyx_L1_error); + __pyx_t_4 = __Pyx_Import(__pyx_n_s_itertools, __pyx_t_7, 0); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __pyx_t_7 = __Pyx_ImportFrom(__pyx_t_4, __pyx_n_s_chain); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 9, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_chain, __pyx_t_7) < 0) __PYX_ERR(0, 9, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":15 + * cimport numpy as np + * + * DTYPE = np.int64 # <<<<<<<<<<<<<< + * ctypedef np.int64_t DTYPE_t + * + */ + __Pyx_GetModuleGlobalName(__pyx_t_4, __pyx_n_s_np); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_4); + __pyx_t_7 = __Pyx_PyObject_GetAttrStr(__pyx_t_4, __pyx_n_s_int64); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 15, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0; + if (PyDict_SetItem(__pyx_d, __pyx_n_s_DTYPE, __pyx_t_7) < 0) __PYX_ERR(0, 15, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":50 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): # <<<<<<<<<<<<<< + * cdef DTYPE_t tok_idx = 0 + * cdef DTYPE_t sz_idx = 0 + */ + __pyx_t_7 = __Pyx_CyFunction_New(&__pyx_mdef_7fairseq_4data_22token_block_utils_fast_1_get_slice_indices_fast, 0, __pyx_n_s_get_slice_indices_fast, NULL, __pyx_n_s_fairseq_data_token_block_utils_f_2, __pyx_d, ((PyObject *)__pyx_codeobj__27)); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 50, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_get_slice_indices_fast, __pyx_t_7) < 0) __PYX_ERR(0, 50, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":109 + * @cython.wraparound(False) + * @cython.nonecheck(False) + * cpdef np.ndarray[DTYPE_t, ndim=2] _get_block_to_dataset_index_fast(np.ndarray[DTYPE_t, ndim=1] sizes, np.ndarray[DTYPE_t, ndim=2] slice_indices): # <<<<<<<<<<<<<< + * cdef DTYPE_t start_ds_idx + * cdef DTYPE_t start_offset + */ + __pyx_t_7 = __Pyx_CyFunction_New(&__pyx_mdef_7fairseq_4data_22token_block_utils_fast_3_get_block_to_dataset_index_fast, 0, __pyx_n_s_get_block_to_dataset_index_fast, NULL, __pyx_n_s_fairseq_data_token_block_utils_f_2, __pyx_d, ((PyObject *)__pyx_codeobj__29)); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 109, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_get_block_to_dataset_index_fast, __pyx_t_7) < 0) __PYX_ERR(0, 109, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "(tree fragment)":1 + * def __reduce_cython__(self): # <<<<<<<<<<<<<< + * cdef tuple state + * cdef object _dict + */ + __pyx_t_7 = __Pyx_CyFunction_New(&__pyx_mdef_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_3__reduce_cython__, __Pyx_CYFUNCTION_CCLASS, __pyx_n_s_DatasetSearcher___reduce_cython, NULL, __pyx_n_s_fairseq_data_token_block_utils_f_2, __pyx_d, ((PyObject *)__pyx_codeobj__31)); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict((PyObject *)__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher, __pyx_n_s_reduce_cython, __pyx_t_7) < 0) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher); + + /* "(tree fragment)":16 + * else: + * return __pyx_unpickle_DatasetSearcher, (type(self), 0x8c67b45, state) + * def __setstate_cython__(self, __pyx_state): # <<<<<<<<<<<<<< + * __pyx_unpickle_DatasetSearcher__set_state(self, __pyx_state) + */ + __pyx_t_7 = __Pyx_CyFunction_New(&__pyx_mdef_7fairseq_4data_22token_block_utils_fast_15DatasetSearcher_5__setstate_cython__, __Pyx_CYFUNCTION_CCLASS, __pyx_n_s_DatasetSearcher___setstate_cytho, NULL, __pyx_n_s_fairseq_data_token_block_utils_f_2, __pyx_d, ((PyObject *)__pyx_codeobj__33)); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 16, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (__Pyx_SetItemOnTypeDict((PyObject *)__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher, __pyx_n_s_setstate_cython, __pyx_t_7) < 0) __PYX_ERR(1, 16, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + PyType_Modified(__pyx_ptype_7fairseq_4data_22token_block_utils_fast_DatasetSearcher); + + /* "(tree fragment)":1 + * def __pyx_unpickle_DatasetSearcher(__pyx_type, long __pyx_checksum, __pyx_state): # <<<<<<<<<<<<<< + * cdef object __pyx_PickleError + * cdef object __pyx_result + */ + __pyx_t_7 = __Pyx_CyFunction_New(&__pyx_mdef_7fairseq_4data_22token_block_utils_fast_5__pyx_unpickle_DatasetSearcher, 0, __pyx_n_s_pyx_unpickle_DatasetSearcher, NULL, __pyx_n_s_fairseq_data_token_block_utils_f_2, __pyx_d, ((PyObject *)__pyx_codeobj__34)); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_pyx_unpickle_DatasetSearcher, __pyx_t_7) < 0) __PYX_ERR(1, 1, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /* "fairseq/data/token_block_utils_fast.pyx":1 + * # cython: language_level=3 # <<<<<<<<<<<<<< + * # Copyright (c) Facebook, Inc. and its affiliates. + * # + */ + __pyx_t_7 = __Pyx_PyDict_NewPresized(0); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 1, __pyx_L1_error) + __Pyx_GOTREF(__pyx_t_7); + if (PyDict_SetItem(__pyx_d, __pyx_n_s_test, __pyx_t_7) < 0) __PYX_ERR(0, 1, __pyx_L1_error) + __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0; + + /*--- Wrapped vars code ---*/ + + goto __pyx_L0; + __pyx_L1_error:; + __Pyx_XDECREF(__pyx_t_4); + __Pyx_XDECREF(__pyx_t_5); + __Pyx_XDECREF(__pyx_t_7); + if (__pyx_m) { + if (__pyx_d && stringtab_initialized) { + __Pyx_AddTraceback("init fairseq.data.token_block_utils_fast", __pyx_clineno, __pyx_lineno, __pyx_filename); + } + #if !CYTHON_USE_MODULE_STATE + Py_CLEAR(__pyx_m); + #else + Py_DECREF(__pyx_m); + if (pystate_addmodule_run) { + PyObject *tp, *value, *tb; + PyErr_Fetch(&tp, &value, &tb); + PyState_RemoveModule(&__pyx_moduledef); + PyErr_Restore(tp, value, tb); + } + #endif + } else if (!PyErr_Occurred()) { + PyErr_SetString(PyExc_ImportError, "init fairseq.data.token_block_utils_fast"); + } + __pyx_L0:; + __Pyx_RefNannyFinishContext(); + #if CYTHON_PEP489_MULTI_PHASE_INIT + return (__pyx_m != NULL) ? 0 : -1; + #elif PY_MAJOR_VERSION >= 3 + return __pyx_m; + #else + return; + #endif +} +/* #### Code section: cleanup_globals ### */ +/* #### Code section: cleanup_module ### */ +/* #### Code section: main_method ### */ +/* #### Code section: utility_code_pragmas ### */ +#ifdef _MSC_VER +#pragma warning( push ) +/* Warning 4127: conditional expression is constant + * Cython uses constant conditional expressions to allow in inline functions to be optimized at + * compile-time, so this warning is not useful + */ +#pragma warning( disable : 4127 ) +#endif + + + +/* #### Code section: utility_code_def ### */ + +/* --- Runtime support code --- */ +/* Refnanny */ +#if CYTHON_REFNANNY +static __Pyx_RefNannyAPIStruct *__Pyx_RefNannyImportAPI(const char *modname) { + PyObject *m = NULL, *p = NULL; + void *r = NULL; + m = PyImport_ImportModule(modname); + if (!m) goto end; + p = PyObject_GetAttrString(m, "RefNannyAPI"); + if (!p) goto end; + r = PyLong_AsVoidPtr(p); +end: + Py_XDECREF(p); + Py_XDECREF(m); + return (__Pyx_RefNannyAPIStruct *)r; +} +#endif + +/* PyErrExceptionMatches */ +#if CYTHON_FAST_THREAD_STATE +static int __Pyx_PyErr_ExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) { + Py_ssize_t i, n; + n = PyTuple_GET_SIZE(tuple); +#if PY_MAJOR_VERSION >= 3 + for (i=0; i= 0x030C00A6 + PyObject *current_exception = tstate->current_exception; + if (unlikely(!current_exception)) return 0; + exc_type = (PyObject*) Py_TYPE(current_exception); + if (exc_type == err) return 1; +#else + exc_type = tstate->curexc_type; + if (exc_type == err) return 1; + if (unlikely(!exc_type)) return 0; +#endif + #if CYTHON_AVOID_BORROWED_REFS + Py_INCREF(exc_type); + #endif + if (unlikely(PyTuple_Check(err))) { + result = __Pyx_PyErr_ExceptionMatchesTuple(exc_type, err); + } else { + result = __Pyx_PyErr_GivenExceptionMatches(exc_type, err); + } + #if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(exc_type); + #endif + return result; +} +#endif + +/* PyErrFetchRestore */ +#if CYTHON_FAST_THREAD_STATE +static CYTHON_INLINE void __Pyx_ErrRestoreInState(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) { +#if PY_VERSION_HEX >= 0x030C00A6 + PyObject *tmp_value; + assert(type == NULL || (value != NULL && type == (PyObject*) Py_TYPE(value))); + if (value) { + #if CYTHON_COMPILING_IN_CPYTHON + if (unlikely(((PyBaseExceptionObject*) value)->traceback != tb)) + #endif + PyException_SetTraceback(value, tb); + } + tmp_value = tstate->current_exception; + tstate->current_exception = value; + Py_XDECREF(tmp_value); + Py_XDECREF(type); + Py_XDECREF(tb); +#else + PyObject *tmp_type, *tmp_value, *tmp_tb; + tmp_type = tstate->curexc_type; + tmp_value = tstate->curexc_value; + tmp_tb = tstate->curexc_traceback; + tstate->curexc_type = type; + tstate->curexc_value = value; + tstate->curexc_traceback = tb; + Py_XDECREF(tmp_type); + Py_XDECREF(tmp_value); + Py_XDECREF(tmp_tb); +#endif +} +static CYTHON_INLINE void __Pyx_ErrFetchInState(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) { +#if PY_VERSION_HEX >= 0x030C00A6 + PyObject* exc_value; + exc_value = tstate->current_exception; + tstate->current_exception = 0; + *value = exc_value; + *type = NULL; + *tb = NULL; + if (exc_value) { + *type = (PyObject*) Py_TYPE(exc_value); + Py_INCREF(*type); + #if CYTHON_COMPILING_IN_CPYTHON + *tb = ((PyBaseExceptionObject*) exc_value)->traceback; + Py_XINCREF(*tb); + #else + *tb = PyException_GetTraceback(exc_value); + #endif + } +#else + *type = tstate->curexc_type; + *value = tstate->curexc_value; + *tb = tstate->curexc_traceback; + tstate->curexc_type = 0; + tstate->curexc_value = 0; + tstate->curexc_traceback = 0; +#endif +} +#endif + +/* PyObjectGetAttrStr */ +#if CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStr(PyObject* obj, PyObject* attr_name) { + PyTypeObject* tp = Py_TYPE(obj); + if (likely(tp->tp_getattro)) + return tp->tp_getattro(obj, attr_name); +#if PY_MAJOR_VERSION < 3 + if (likely(tp->tp_getattr)) + return tp->tp_getattr(obj, PyString_AS_STRING(attr_name)); +#endif + return PyObject_GetAttr(obj, attr_name); +} +#endif + +/* PyObjectGetAttrStrNoError */ +#if __PYX_LIMITED_VERSION_HEX < 0x030d00A1 +static void __Pyx_PyObject_GetAttrStr_ClearAttributeError(void) { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + if (likely(__Pyx_PyErr_ExceptionMatches(PyExc_AttributeError))) + __Pyx_PyErr_Clear(); +} +#endif +static CYTHON_INLINE PyObject* __Pyx_PyObject_GetAttrStrNoError(PyObject* obj, PyObject* attr_name) { + PyObject *result; +#if __PYX_LIMITED_VERSION_HEX >= 0x030d00A1 + (void) PyObject_GetOptionalAttr(obj, attr_name, &result); + return result; +#else +#if CYTHON_COMPILING_IN_CPYTHON && CYTHON_USE_TYPE_SLOTS && PY_VERSION_HEX >= 0x030700B1 + PyTypeObject* tp = Py_TYPE(obj); + if (likely(tp->tp_getattro == PyObject_GenericGetAttr)) { + return _PyObject_GenericGetAttrWithDict(obj, attr_name, NULL, 1); + } +#endif + result = __Pyx_PyObject_GetAttrStr(obj, attr_name); + if (unlikely(!result)) { + __Pyx_PyObject_GetAttrStr_ClearAttributeError(); + } + return result; +#endif +} + +/* GetBuiltinName */ +static PyObject *__Pyx_GetBuiltinName(PyObject *name) { + PyObject* result = __Pyx_PyObject_GetAttrStrNoError(__pyx_b, name); + if (unlikely(!result) && !PyErr_Occurred()) { + PyErr_Format(PyExc_NameError, +#if PY_MAJOR_VERSION >= 3 + "name '%U' is not defined", name); +#else + "name '%.200s' is not defined", PyString_AS_STRING(name)); +#endif + } + return result; +} + +/* TupleAndListFromArray */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE void __Pyx_copy_object_array(PyObject *const *CYTHON_RESTRICT src, PyObject** CYTHON_RESTRICT dest, Py_ssize_t length) { + PyObject *v; + Py_ssize_t i; + for (i = 0; i < length; i++) { + v = dest[i] = src[i]; + Py_INCREF(v); + } +} +static CYTHON_INLINE PyObject * +__Pyx_PyTuple_FromArray(PyObject *const *src, Py_ssize_t n) +{ + PyObject *res; + if (n <= 0) { + Py_INCREF(__pyx_empty_tuple); + return __pyx_empty_tuple; + } + res = PyTuple_New(n); + if (unlikely(res == NULL)) return NULL; + __Pyx_copy_object_array(src, ((PyTupleObject*)res)->ob_item, n); + return res; +} +static CYTHON_INLINE PyObject * +__Pyx_PyList_FromArray(PyObject *const *src, Py_ssize_t n) +{ + PyObject *res; + if (n <= 0) { + return PyList_New(0); + } + res = PyList_New(n); + if (unlikely(res == NULL)) return NULL; + __Pyx_copy_object_array(src, ((PyListObject*)res)->ob_item, n); + return res; +} +#endif + +/* BytesEquals */ +static CYTHON_INLINE int __Pyx_PyBytes_Equals(PyObject* s1, PyObject* s2, int equals) { +#if CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API + return PyObject_RichCompareBool(s1, s2, equals); +#else + if (s1 == s2) { + return (equals == Py_EQ); + } else if (PyBytes_CheckExact(s1) & PyBytes_CheckExact(s2)) { + const char *ps1, *ps2; + Py_ssize_t length = PyBytes_GET_SIZE(s1); + if (length != PyBytes_GET_SIZE(s2)) + return (equals == Py_NE); + ps1 = PyBytes_AS_STRING(s1); + ps2 = PyBytes_AS_STRING(s2); + if (ps1[0] != ps2[0]) { + return (equals == Py_NE); + } else if (length == 1) { + return (equals == Py_EQ); + } else { + int result; +#if CYTHON_USE_UNICODE_INTERNALS && (PY_VERSION_HEX < 0x030B0000) + Py_hash_t hash1, hash2; + hash1 = ((PyBytesObject*)s1)->ob_shash; + hash2 = ((PyBytesObject*)s2)->ob_shash; + if (hash1 != hash2 && hash1 != -1 && hash2 != -1) { + return (equals == Py_NE); + } +#endif + result = memcmp(ps1, ps2, (size_t)length); + return (equals == Py_EQ) ? (result == 0) : (result != 0); + } + } else if ((s1 == Py_None) & PyBytes_CheckExact(s2)) { + return (equals == Py_NE); + } else if ((s2 == Py_None) & PyBytes_CheckExact(s1)) { + return (equals == Py_NE); + } else { + int result; + PyObject* py_result = PyObject_RichCompare(s1, s2, equals); + if (!py_result) + return -1; + result = __Pyx_PyObject_IsTrue(py_result); + Py_DECREF(py_result); + return result; + } +#endif +} + +/* UnicodeEquals */ +static CYTHON_INLINE int __Pyx_PyUnicode_Equals(PyObject* s1, PyObject* s2, int equals) { +#if CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API + return PyObject_RichCompareBool(s1, s2, equals); +#else +#if PY_MAJOR_VERSION < 3 + PyObject* owned_ref = NULL; +#endif + int s1_is_unicode, s2_is_unicode; + if (s1 == s2) { + goto return_eq; + } + s1_is_unicode = PyUnicode_CheckExact(s1); + s2_is_unicode = PyUnicode_CheckExact(s2); +#if PY_MAJOR_VERSION < 3 + if ((s1_is_unicode & (!s2_is_unicode)) && PyString_CheckExact(s2)) { + owned_ref = PyUnicode_FromObject(s2); + if (unlikely(!owned_ref)) + return -1; + s2 = owned_ref; + s2_is_unicode = 1; + } else if ((s2_is_unicode & (!s1_is_unicode)) && PyString_CheckExact(s1)) { + owned_ref = PyUnicode_FromObject(s1); + if (unlikely(!owned_ref)) + return -1; + s1 = owned_ref; + s1_is_unicode = 1; + } else if (((!s2_is_unicode) & (!s1_is_unicode))) { + return __Pyx_PyBytes_Equals(s1, s2, equals); + } +#endif + if (s1_is_unicode & s2_is_unicode) { + Py_ssize_t length; + int kind; + void *data1, *data2; + if (unlikely(__Pyx_PyUnicode_READY(s1) < 0) || unlikely(__Pyx_PyUnicode_READY(s2) < 0)) + return -1; + length = __Pyx_PyUnicode_GET_LENGTH(s1); + if (length != __Pyx_PyUnicode_GET_LENGTH(s2)) { + goto return_ne; + } +#if CYTHON_USE_UNICODE_INTERNALS + { + Py_hash_t hash1, hash2; + #if CYTHON_PEP393_ENABLED + hash1 = ((PyASCIIObject*)s1)->hash; + hash2 = ((PyASCIIObject*)s2)->hash; + #else + hash1 = ((PyUnicodeObject*)s1)->hash; + hash2 = ((PyUnicodeObject*)s2)->hash; + #endif + if (hash1 != hash2 && hash1 != -1 && hash2 != -1) { + goto return_ne; + } + } +#endif + kind = __Pyx_PyUnicode_KIND(s1); + if (kind != __Pyx_PyUnicode_KIND(s2)) { + goto return_ne; + } + data1 = __Pyx_PyUnicode_DATA(s1); + data2 = __Pyx_PyUnicode_DATA(s2); + if (__Pyx_PyUnicode_READ(kind, data1, 0) != __Pyx_PyUnicode_READ(kind, data2, 0)) { + goto return_ne; + } else if (length == 1) { + goto return_eq; + } else { + int result = memcmp(data1, data2, (size_t)(length * kind)); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + return (equals == Py_EQ) ? (result == 0) : (result != 0); + } + } else if ((s1 == Py_None) & s2_is_unicode) { + goto return_ne; + } else if ((s2 == Py_None) & s1_is_unicode) { + goto return_ne; + } else { + int result; + PyObject* py_result = PyObject_RichCompare(s1, s2, equals); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + if (!py_result) + return -1; + result = __Pyx_PyObject_IsTrue(py_result); + Py_DECREF(py_result); + return result; + } +return_eq: + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + return (equals == Py_EQ); +return_ne: + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(owned_ref); + #endif + return (equals == Py_NE); +#endif +} + +/* fastcall */ +#if CYTHON_METH_FASTCALL +static CYTHON_INLINE PyObject * __Pyx_GetKwValue_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues, PyObject *s) +{ + Py_ssize_t i, n = PyTuple_GET_SIZE(kwnames); + for (i = 0; i < n; i++) + { + if (s == PyTuple_GET_ITEM(kwnames, i)) return kwvalues[i]; + } + for (i = 0; i < n; i++) + { + int eq = __Pyx_PyUnicode_Equals(s, PyTuple_GET_ITEM(kwnames, i), Py_EQ); + if (unlikely(eq != 0)) { + if (unlikely(eq < 0)) return NULL; + return kwvalues[i]; + } + } + return NULL; +} +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030d0000 +CYTHON_UNUSED static PyObject *__Pyx_KwargsAsDict_FASTCALL(PyObject *kwnames, PyObject *const *kwvalues) { + Py_ssize_t i, nkwargs = PyTuple_GET_SIZE(kwnames); + PyObject *dict; + dict = PyDict_New(); + if (unlikely(!dict)) + return NULL; + for (i=0; i= 3 + "%s() got multiple values for keyword argument '%U'", func_name, kw_name); + #else + "%s() got multiple values for keyword argument '%s'", func_name, + PyString_AsString(kw_name)); + #endif +} + +/* ParseKeywords */ +static int __Pyx_ParseOptionalKeywords( + PyObject *kwds, + PyObject *const *kwvalues, + PyObject **argnames[], + PyObject *kwds2, + PyObject *values[], + Py_ssize_t num_pos_args, + const char* function_name) +{ + PyObject *key = 0, *value = 0; + Py_ssize_t pos = 0; + PyObject*** name; + PyObject*** first_kw_arg = argnames + num_pos_args; + int kwds_is_tuple = CYTHON_METH_FASTCALL && likely(PyTuple_Check(kwds)); + while (1) { + Py_XDECREF(key); key = NULL; + Py_XDECREF(value); value = NULL; + if (kwds_is_tuple) { + Py_ssize_t size; +#if CYTHON_ASSUME_SAFE_MACROS + size = PyTuple_GET_SIZE(kwds); +#else + size = PyTuple_Size(kwds); + if (size < 0) goto bad; +#endif + if (pos >= size) break; +#if CYTHON_AVOID_BORROWED_REFS + key = __Pyx_PySequence_ITEM(kwds, pos); + if (!key) goto bad; +#elif CYTHON_ASSUME_SAFE_MACROS + key = PyTuple_GET_ITEM(kwds, pos); +#else + key = PyTuple_GetItem(kwds, pos); + if (!key) goto bad; +#endif + value = kwvalues[pos]; + pos++; + } + else + { + if (!PyDict_Next(kwds, &pos, &key, &value)) break; +#if CYTHON_AVOID_BORROWED_REFS + Py_INCREF(key); +#endif + } + name = first_kw_arg; + while (*name && (**name != key)) name++; + if (*name) { + values[name-argnames] = value; +#if CYTHON_AVOID_BORROWED_REFS + Py_INCREF(value); + Py_DECREF(key); +#endif + key = NULL; + value = NULL; + continue; + } +#if !CYTHON_AVOID_BORROWED_REFS + Py_INCREF(key); +#endif + Py_INCREF(value); + name = first_kw_arg; + #if PY_MAJOR_VERSION < 3 + if (likely(PyString_Check(key))) { + while (*name) { + if ((CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**name) == PyString_GET_SIZE(key)) + && _PyString_Eq(**name, key)) { + values[name-argnames] = value; +#if CYTHON_AVOID_BORROWED_REFS + value = NULL; +#endif + break; + } + name++; + } + if (*name) continue; + else { + PyObject*** argname = argnames; + while (argname != first_kw_arg) { + if ((**argname == key) || ( + (CYTHON_COMPILING_IN_PYPY || PyString_GET_SIZE(**argname) == PyString_GET_SIZE(key)) + && _PyString_Eq(**argname, key))) { + goto arg_passed_twice; + } + argname++; + } + } + } else + #endif + if (likely(PyUnicode_Check(key))) { + while (*name) { + int cmp = ( + #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3 + (__Pyx_PyUnicode_GET_LENGTH(**name) != __Pyx_PyUnicode_GET_LENGTH(key)) ? 1 : + #endif + PyUnicode_Compare(**name, key) + ); + if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad; + if (cmp == 0) { + values[name-argnames] = value; +#if CYTHON_AVOID_BORROWED_REFS + value = NULL; +#endif + break; + } + name++; + } + if (*name) continue; + else { + PyObject*** argname = argnames; + while (argname != first_kw_arg) { + int cmp = (**argname == key) ? 0 : + #if !CYTHON_COMPILING_IN_PYPY && PY_MAJOR_VERSION >= 3 + (__Pyx_PyUnicode_GET_LENGTH(**argname) != __Pyx_PyUnicode_GET_LENGTH(key)) ? 1 : + #endif + PyUnicode_Compare(**argname, key); + if (cmp < 0 && unlikely(PyErr_Occurred())) goto bad; + if (cmp == 0) goto arg_passed_twice; + argname++; + } + } + } else + goto invalid_keyword_type; + if (kwds2) { + if (unlikely(PyDict_SetItem(kwds2, key, value))) goto bad; + } else { + goto invalid_keyword; + } + } + Py_XDECREF(key); + Py_XDECREF(value); + return 0; +arg_passed_twice: + __Pyx_RaiseDoubleKeywordsError(function_name, key); + goto bad; +invalid_keyword_type: + PyErr_Format(PyExc_TypeError, + "%.200s() keywords must be strings", function_name); + goto bad; +invalid_keyword: + #if PY_MAJOR_VERSION < 3 + PyErr_Format(PyExc_TypeError, + "%.200s() got an unexpected keyword argument '%.200s'", + function_name, PyString_AsString(key)); + #else + PyErr_Format(PyExc_TypeError, + "%s() got an unexpected keyword argument '%U'", + function_name, key); + #endif +bad: + Py_XDECREF(key); + Py_XDECREF(value); + return -1; +} + +/* ArgTypeTest */ +static int __Pyx__ArgTypeTest(PyObject *obj, PyTypeObject *type, const char *name, int exact) +{ + __Pyx_TypeName type_name; + __Pyx_TypeName obj_type_name; + if (unlikely(!type)) { + PyErr_SetString(PyExc_SystemError, "Missing type object"); + return 0; + } + else if (exact) { + #if PY_MAJOR_VERSION == 2 + if ((type == &PyBaseString_Type) && likely(__Pyx_PyBaseString_CheckExact(obj))) return 1; + #endif + } + else { + if (likely(__Pyx_TypeCheck(obj, type))) return 1; + } + type_name = __Pyx_PyType_GetName(type); + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, + "Argument '%.200s' has incorrect type (expected " __Pyx_FMT_TYPENAME + ", got " __Pyx_FMT_TYPENAME ")", name, type_name, obj_type_name); + __Pyx_DECREF_TypeName(type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return 0; +} + +/* RaiseException */ +#if PY_MAJOR_VERSION < 3 +static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause) { + __Pyx_PyThreadState_declare + CYTHON_UNUSED_VAR(cause); + Py_XINCREF(type); + if (!value || value == Py_None) + value = NULL; + else + Py_INCREF(value); + if (!tb || tb == Py_None) + tb = NULL; + else { + Py_INCREF(tb); + if (!PyTraceBack_Check(tb)) { + PyErr_SetString(PyExc_TypeError, + "raise: arg 3 must be a traceback or None"); + goto raise_error; + } + } + if (PyType_Check(type)) { +#if CYTHON_COMPILING_IN_PYPY + if (!value) { + Py_INCREF(Py_None); + value = Py_None; + } +#endif + PyErr_NormalizeException(&type, &value, &tb); + } else { + if (value) { + PyErr_SetString(PyExc_TypeError, + "instance exception may not have a separate value"); + goto raise_error; + } + value = type; + type = (PyObject*) Py_TYPE(type); + Py_INCREF(type); + if (!PyType_IsSubtype((PyTypeObject *)type, (PyTypeObject *)PyExc_BaseException)) { + PyErr_SetString(PyExc_TypeError, + "raise: exception class must be a subclass of BaseException"); + goto raise_error; + } + } + __Pyx_PyThreadState_assign + __Pyx_ErrRestore(type, value, tb); + return; +raise_error: + Py_XDECREF(value); + Py_XDECREF(type); + Py_XDECREF(tb); + return; +} +#else +static void __Pyx_Raise(PyObject *type, PyObject *value, PyObject *tb, PyObject *cause) { + PyObject* owned_instance = NULL; + if (tb == Py_None) { + tb = 0; + } else if (tb && !PyTraceBack_Check(tb)) { + PyErr_SetString(PyExc_TypeError, + "raise: arg 3 must be a traceback or None"); + goto bad; + } + if (value == Py_None) + value = 0; + if (PyExceptionInstance_Check(type)) { + if (value) { + PyErr_SetString(PyExc_TypeError, + "instance exception may not have a separate value"); + goto bad; + } + value = type; + type = (PyObject*) Py_TYPE(value); + } else if (PyExceptionClass_Check(type)) { + PyObject *instance_class = NULL; + if (value && PyExceptionInstance_Check(value)) { + instance_class = (PyObject*) Py_TYPE(value); + if (instance_class != type) { + int is_subclass = PyObject_IsSubclass(instance_class, type); + if (!is_subclass) { + instance_class = NULL; + } else if (unlikely(is_subclass == -1)) { + goto bad; + } else { + type = instance_class; + } + } + } + if (!instance_class) { + PyObject *args; + if (!value) + args = PyTuple_New(0); + else if (PyTuple_Check(value)) { + Py_INCREF(value); + args = value; + } else + args = PyTuple_Pack(1, value); + if (!args) + goto bad; + owned_instance = PyObject_Call(type, args, NULL); + Py_DECREF(args); + if (!owned_instance) + goto bad; + value = owned_instance; + if (!PyExceptionInstance_Check(value)) { + PyErr_Format(PyExc_TypeError, + "calling %R should have returned an instance of " + "BaseException, not %R", + type, Py_TYPE(value)); + goto bad; + } + } + } else { + PyErr_SetString(PyExc_TypeError, + "raise: exception class must be a subclass of BaseException"); + goto bad; + } + if (cause) { + PyObject *fixed_cause; + if (cause == Py_None) { + fixed_cause = NULL; + } else if (PyExceptionClass_Check(cause)) { + fixed_cause = PyObject_CallObject(cause, NULL); + if (fixed_cause == NULL) + goto bad; + } else if (PyExceptionInstance_Check(cause)) { + fixed_cause = cause; + Py_INCREF(fixed_cause); + } else { + PyErr_SetString(PyExc_TypeError, + "exception causes must derive from " + "BaseException"); + goto bad; + } + PyException_SetCause(value, fixed_cause); + } + PyErr_SetObject(type, value); + if (tb) { + #if PY_VERSION_HEX >= 0x030C00A6 + PyException_SetTraceback(value, tb); + #elif CYTHON_FAST_THREAD_STATE + PyThreadState *tstate = __Pyx_PyThreadState_Current; + PyObject* tmp_tb = tstate->curexc_traceback; + if (tb != tmp_tb) { + Py_INCREF(tb); + tstate->curexc_traceback = tb; + Py_XDECREF(tmp_tb); + } +#else + PyObject *tmp_type, *tmp_value, *tmp_tb; + PyErr_Fetch(&tmp_type, &tmp_value, &tmp_tb); + Py_INCREF(tb); + PyErr_Restore(tmp_type, tmp_value, tb); + Py_XDECREF(tmp_tb); +#endif + } +bad: + Py_XDECREF(owned_instance); + return; +} +#endif + +/* PyFunctionFastCall */ +#if CYTHON_FAST_PYCALL && !CYTHON_VECTORCALL +static PyObject* __Pyx_PyFunction_FastCallNoKw(PyCodeObject *co, PyObject **args, Py_ssize_t na, + PyObject *globals) { + PyFrameObject *f; + PyThreadState *tstate = __Pyx_PyThreadState_Current; + PyObject **fastlocals; + Py_ssize_t i; + PyObject *result; + assert(globals != NULL); + /* XXX Perhaps we should create a specialized + PyFrame_New() that doesn't take locals, but does + take builtins without sanity checking them. + */ + assert(tstate != NULL); + f = PyFrame_New(tstate, co, globals, NULL); + if (f == NULL) { + return NULL; + } + fastlocals = __Pyx_PyFrame_GetLocalsplus(f); + for (i = 0; i < na; i++) { + Py_INCREF(*args); + fastlocals[i] = *args++; + } + result = PyEval_EvalFrameEx(f,0); + ++tstate->recursion_depth; + Py_DECREF(f); + --tstate->recursion_depth; + return result; +} +static PyObject *__Pyx_PyFunction_FastCallDict(PyObject *func, PyObject **args, Py_ssize_t nargs, PyObject *kwargs) { + PyCodeObject *co = (PyCodeObject *)PyFunction_GET_CODE(func); + PyObject *globals = PyFunction_GET_GLOBALS(func); + PyObject *argdefs = PyFunction_GET_DEFAULTS(func); + PyObject *closure; +#if PY_MAJOR_VERSION >= 3 + PyObject *kwdefs; +#endif + PyObject *kwtuple, **k; + PyObject **d; + Py_ssize_t nd; + Py_ssize_t nk; + PyObject *result; + assert(kwargs == NULL || PyDict_Check(kwargs)); + nk = kwargs ? PyDict_Size(kwargs) : 0; + #if PY_MAJOR_VERSION < 3 + if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object"))) { + return NULL; + } + #else + if (unlikely(Py_EnterRecursiveCall(" while calling a Python object"))) { + return NULL; + } + #endif + if ( +#if PY_MAJOR_VERSION >= 3 + co->co_kwonlyargcount == 0 && +#endif + likely(kwargs == NULL || nk == 0) && + co->co_flags == (CO_OPTIMIZED | CO_NEWLOCALS | CO_NOFREE)) { + if (argdefs == NULL && co->co_argcount == nargs) { + result = __Pyx_PyFunction_FastCallNoKw(co, args, nargs, globals); + goto done; + } + else if (nargs == 0 && argdefs != NULL + && co->co_argcount == Py_SIZE(argdefs)) { + /* function called with no arguments, but all parameters have + a default value: use default values as arguments .*/ + args = &PyTuple_GET_ITEM(argdefs, 0); + result =__Pyx_PyFunction_FastCallNoKw(co, args, Py_SIZE(argdefs), globals); + goto done; + } + } + if (kwargs != NULL) { + Py_ssize_t pos, i; + kwtuple = PyTuple_New(2 * nk); + if (kwtuple == NULL) { + result = NULL; + goto done; + } + k = &PyTuple_GET_ITEM(kwtuple, 0); + pos = i = 0; + while (PyDict_Next(kwargs, &pos, &k[i], &k[i+1])) { + Py_INCREF(k[i]); + Py_INCREF(k[i+1]); + i += 2; + } + nk = i / 2; + } + else { + kwtuple = NULL; + k = NULL; + } + closure = PyFunction_GET_CLOSURE(func); +#if PY_MAJOR_VERSION >= 3 + kwdefs = PyFunction_GET_KW_DEFAULTS(func); +#endif + if (argdefs != NULL) { + d = &PyTuple_GET_ITEM(argdefs, 0); + nd = Py_SIZE(argdefs); + } + else { + d = NULL; + nd = 0; + } +#if PY_MAJOR_VERSION >= 3 + result = PyEval_EvalCodeEx((PyObject*)co, globals, (PyObject *)NULL, + args, (int)nargs, + k, (int)nk, + d, (int)nd, kwdefs, closure); +#else + result = PyEval_EvalCodeEx(co, globals, (PyObject *)NULL, + args, (int)nargs, + k, (int)nk, + d, (int)nd, closure); +#endif + Py_XDECREF(kwtuple); +done: + Py_LeaveRecursiveCall(); + return result; +} +#endif + +/* PyObjectCall */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw) { + PyObject *result; + ternaryfunc call = Py_TYPE(func)->tp_call; + if (unlikely(!call)) + return PyObject_Call(func, arg, kw); + #if PY_MAJOR_VERSION < 3 + if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object"))) + return NULL; + #else + if (unlikely(Py_EnterRecursiveCall(" while calling a Python object"))) + return NULL; + #endif + result = (*call)(func, arg, kw); + Py_LeaveRecursiveCall(); + if (unlikely(!result) && unlikely(!PyErr_Occurred())) { + PyErr_SetString( + PyExc_SystemError, + "NULL result without error in PyObject_Call"); + } + return result; +} +#endif + +/* PyObjectCallMethO */ +#if CYTHON_COMPILING_IN_CPYTHON +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallMethO(PyObject *func, PyObject *arg) { + PyObject *self, *result; + PyCFunction cfunc; + cfunc = __Pyx_CyOrPyCFunction_GET_FUNCTION(func); + self = __Pyx_CyOrPyCFunction_GET_SELF(func); + #if PY_MAJOR_VERSION < 3 + if (unlikely(Py_EnterRecursiveCall((char*)" while calling a Python object"))) + return NULL; + #else + if (unlikely(Py_EnterRecursiveCall(" while calling a Python object"))) + return NULL; + #endif + result = cfunc(self, arg); + Py_LeaveRecursiveCall(); + if (unlikely(!result) && unlikely(!PyErr_Occurred())) { + PyErr_SetString( + PyExc_SystemError, + "NULL result without error in PyObject_Call"); + } + return result; +} +#endif + +/* PyObjectFastCall */ +#if PY_VERSION_HEX < 0x03090000 || CYTHON_COMPILING_IN_LIMITED_API +static PyObject* __Pyx_PyObject_FastCall_fallback(PyObject *func, PyObject **args, size_t nargs, PyObject *kwargs) { + PyObject *argstuple; + PyObject *result = 0; + size_t i; + argstuple = PyTuple_New((Py_ssize_t)nargs); + if (unlikely(!argstuple)) return NULL; + for (i = 0; i < nargs; i++) { + Py_INCREF(args[i]); + if (__Pyx_PyTuple_SET_ITEM(argstuple, (Py_ssize_t)i, args[i]) < 0) goto bad; + } + result = __Pyx_PyObject_Call(func, argstuple, kwargs); + bad: + Py_DECREF(argstuple); + return result; +} +#endif +static CYTHON_INLINE PyObject* __Pyx_PyObject_FastCallDict(PyObject *func, PyObject **args, size_t _nargs, PyObject *kwargs) { + Py_ssize_t nargs = __Pyx_PyVectorcall_NARGS(_nargs); +#if CYTHON_COMPILING_IN_CPYTHON + if (nargs == 0 && kwargs == NULL) { + if (__Pyx_CyOrPyCFunction_Check(func) && likely( __Pyx_CyOrPyCFunction_GET_FLAGS(func) & METH_NOARGS)) + return __Pyx_PyObject_CallMethO(func, NULL); + } + else if (nargs == 1 && kwargs == NULL) { + if (__Pyx_CyOrPyCFunction_Check(func) && likely( __Pyx_CyOrPyCFunction_GET_FLAGS(func) & METH_O)) + return __Pyx_PyObject_CallMethO(func, args[0]); + } +#endif + #if PY_VERSION_HEX < 0x030800B1 + #if CYTHON_FAST_PYCCALL + if (PyCFunction_Check(func)) { + if (kwargs) { + return _PyCFunction_FastCallDict(func, args, nargs, kwargs); + } else { + return _PyCFunction_FastCallKeywords(func, args, nargs, NULL); + } + } + #if PY_VERSION_HEX >= 0x030700A1 + if (!kwargs && __Pyx_IS_TYPE(func, &PyMethodDescr_Type)) { + return _PyMethodDescr_FastCallKeywords(func, args, nargs, NULL); + } + #endif + #endif + #if CYTHON_FAST_PYCALL + if (PyFunction_Check(func)) { + return __Pyx_PyFunction_FastCallDict(func, args, nargs, kwargs); + } + #endif + #endif + if (kwargs == NULL) { + #if CYTHON_VECTORCALL + #if PY_VERSION_HEX < 0x03090000 + vectorcallfunc f = _PyVectorcall_Function(func); + #else + vectorcallfunc f = PyVectorcall_Function(func); + #endif + if (f) { + return f(func, args, (size_t)nargs, NULL); + } + #elif defined(__Pyx_CyFunction_USED) && CYTHON_BACKPORT_VECTORCALL + if (__Pyx_CyFunction_CheckExact(func)) { + __pyx_vectorcallfunc f = __Pyx_CyFunction_func_vectorcall(func); + if (f) return f(func, args, (size_t)nargs, NULL); + } + #endif + } + if (nargs == 0) { + return __Pyx_PyObject_Call(func, __pyx_empty_tuple, kwargs); + } + #if PY_VERSION_HEX >= 0x03090000 && !CYTHON_COMPILING_IN_LIMITED_API + return PyObject_VectorcallDict(func, args, (size_t)nargs, kwargs); + #else + return __Pyx_PyObject_FastCall_fallback(func, args, (size_t)nargs, kwargs); + #endif +} + +/* RaiseUnexpectedTypeError */ +static int +__Pyx_RaiseUnexpectedTypeError(const char *expected, PyObject *obj) +{ + __Pyx_TypeName obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, "Expected %s, got " __Pyx_FMT_TYPENAME, + expected, obj_type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return 0; +} + +/* CIntToDigits */ +static const char DIGIT_PAIRS_10[2*10*10+1] = { + "00010203040506070809" + "10111213141516171819" + "20212223242526272829" + "30313233343536373839" + "40414243444546474849" + "50515253545556575859" + "60616263646566676869" + "70717273747576777879" + "80818283848586878889" + "90919293949596979899" +}; +static const char DIGIT_PAIRS_8[2*8*8+1] = { + "0001020304050607" + "1011121314151617" + "2021222324252627" + "3031323334353637" + "4041424344454647" + "5051525354555657" + "6061626364656667" + "7071727374757677" +}; +static const char DIGITS_HEX[2*16+1] = { + "0123456789abcdef" + "0123456789ABCDEF" +}; + +/* BuildPyUnicode */ +static PyObject* __Pyx_PyUnicode_BuildFromAscii(Py_ssize_t ulength, char* chars, int clength, + int prepend_sign, char padding_char) { + PyObject *uval; + Py_ssize_t uoffset = ulength - clength; +#if CYTHON_USE_UNICODE_INTERNALS + Py_ssize_t i; +#if CYTHON_PEP393_ENABLED + void *udata; + uval = PyUnicode_New(ulength, 127); + if (unlikely(!uval)) return NULL; + udata = PyUnicode_DATA(uval); +#else + Py_UNICODE *udata; + uval = PyUnicode_FromUnicode(NULL, ulength); + if (unlikely(!uval)) return NULL; + udata = PyUnicode_AS_UNICODE(uval); +#endif + if (uoffset > 0) { + i = 0; + if (prepend_sign) { + __Pyx_PyUnicode_WRITE(PyUnicode_1BYTE_KIND, udata, 0, '-'); + i++; + } + for (; i < uoffset; i++) { + __Pyx_PyUnicode_WRITE(PyUnicode_1BYTE_KIND, udata, i, padding_char); + } + } + for (i=0; i < clength; i++) { + __Pyx_PyUnicode_WRITE(PyUnicode_1BYTE_KIND, udata, uoffset+i, chars[i]); + } +#else + { + PyObject *sign = NULL, *padding = NULL; + uval = NULL; + if (uoffset > 0) { + prepend_sign = !!prepend_sign; + if (uoffset > prepend_sign) { + padding = PyUnicode_FromOrdinal(padding_char); + if (likely(padding) && uoffset > prepend_sign + 1) { + PyObject *tmp; + PyObject *repeat = PyInt_FromSsize_t(uoffset - prepend_sign); + if (unlikely(!repeat)) goto done_or_error; + tmp = PyNumber_Multiply(padding, repeat); + Py_DECREF(repeat); + Py_DECREF(padding); + padding = tmp; + } + if (unlikely(!padding)) goto done_or_error; + } + if (prepend_sign) { + sign = PyUnicode_FromOrdinal('-'); + if (unlikely(!sign)) goto done_or_error; + } + } + uval = PyUnicode_DecodeASCII(chars, clength, NULL); + if (likely(uval) && padding) { + PyObject *tmp = PyNumber_Add(padding, uval); + Py_DECREF(uval); + uval = tmp; + } + if (likely(uval) && sign) { + PyObject *tmp = PyNumber_Add(sign, uval); + Py_DECREF(uval); + uval = tmp; + } +done_or_error: + Py_XDECREF(padding); + Py_XDECREF(sign); + } +#endif + return uval; +} + +/* CIntToPyUnicode */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_int(int value, Py_ssize_t width, char padding_char, char format_char) { + char digits[sizeof(int)*3+2]; + char *dpos, *end = digits + sizeof(int)*3+2; + const char *hex_digits = DIGITS_HEX; + Py_ssize_t length, ulength; + int prepend_sign, last_one_off; + int remaining; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const int neg_one = (int) -1, const_zero = (int) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (format_char == 'X') { + hex_digits += 16; + format_char = 'x'; + } + remaining = value; + last_one_off = 0; + dpos = end; + do { + int digit_pos; + switch (format_char) { + case 'o': + digit_pos = abs((int)(remaining % (8*8))); + remaining = (int) (remaining / (8*8)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_8 + digit_pos * 2, 2); + last_one_off = (digit_pos < 8); + break; + case 'd': + digit_pos = abs((int)(remaining % (10*10))); + remaining = (int) (remaining / (10*10)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_10 + digit_pos * 2, 2); + last_one_off = (digit_pos < 10); + break; + case 'x': + *(--dpos) = hex_digits[abs((int)(remaining % 16))]; + remaining = (int) (remaining / 16); + break; + default: + assert(0); + break; + } + } while (unlikely(remaining != 0)); + assert(!last_one_off || *dpos == '0'); + dpos += last_one_off; + length = end - dpos; + ulength = length; + prepend_sign = 0; + if (!is_unsigned && value <= neg_one) { + if (padding_char == ' ' || width <= length + 1) { + *(--dpos) = '-'; + ++length; + } else { + prepend_sign = 1; + } + ++ulength; + } + if (width > ulength) { + ulength = width; + } + if (ulength == 1) { + return PyUnicode_FromOrdinal(*dpos); + } + return __Pyx_PyUnicode_BuildFromAscii(ulength, dpos, (int) length, prepend_sign, padding_char); +} + +/* CIntToPyUnicode */ +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_From_Py_ssize_t(Py_ssize_t value, Py_ssize_t width, char padding_char, char format_char) { + char digits[sizeof(Py_ssize_t)*3+2]; + char *dpos, *end = digits + sizeof(Py_ssize_t)*3+2; + const char *hex_digits = DIGITS_HEX; + Py_ssize_t length, ulength; + int prepend_sign, last_one_off; + Py_ssize_t remaining; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const Py_ssize_t neg_one = (Py_ssize_t) -1, const_zero = (Py_ssize_t) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (format_char == 'X') { + hex_digits += 16; + format_char = 'x'; + } + remaining = value; + last_one_off = 0; + dpos = end; + do { + int digit_pos; + switch (format_char) { + case 'o': + digit_pos = abs((int)(remaining % (8*8))); + remaining = (Py_ssize_t) (remaining / (8*8)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_8 + digit_pos * 2, 2); + last_one_off = (digit_pos < 8); + break; + case 'd': + digit_pos = abs((int)(remaining % (10*10))); + remaining = (Py_ssize_t) (remaining / (10*10)); + dpos -= 2; + memcpy(dpos, DIGIT_PAIRS_10 + digit_pos * 2, 2); + last_one_off = (digit_pos < 10); + break; + case 'x': + *(--dpos) = hex_digits[abs((int)(remaining % 16))]; + remaining = (Py_ssize_t) (remaining / 16); + break; + default: + assert(0); + break; + } + } while (unlikely(remaining != 0)); + assert(!last_one_off || *dpos == '0'); + dpos += last_one_off; + length = end - dpos; + ulength = length; + prepend_sign = 0; + if (!is_unsigned && value <= neg_one) { + if (padding_char == ' ' || width <= length + 1) { + *(--dpos) = '-'; + ++length; + } else { + prepend_sign = 1; + } + ++ulength; + } + if (width > ulength) { + ulength = width; + } + if (ulength == 1) { + return PyUnicode_FromOrdinal(*dpos); + } + return __Pyx_PyUnicode_BuildFromAscii(ulength, dpos, (int) length, prepend_sign, padding_char); +} + +/* JoinPyUnicode */ +static PyObject* __Pyx_PyUnicode_Join(PyObject* value_tuple, Py_ssize_t value_count, Py_ssize_t result_ulength, + Py_UCS4 max_char) { +#if CYTHON_USE_UNICODE_INTERNALS && CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + PyObject *result_uval; + int result_ukind, kind_shift; + Py_ssize_t i, char_pos; + void *result_udata; + CYTHON_MAYBE_UNUSED_VAR(max_char); +#if CYTHON_PEP393_ENABLED + result_uval = PyUnicode_New(result_ulength, max_char); + if (unlikely(!result_uval)) return NULL; + result_ukind = (max_char <= 255) ? PyUnicode_1BYTE_KIND : (max_char <= 65535) ? PyUnicode_2BYTE_KIND : PyUnicode_4BYTE_KIND; + kind_shift = (result_ukind == PyUnicode_4BYTE_KIND) ? 2 : result_ukind - 1; + result_udata = PyUnicode_DATA(result_uval); +#else + result_uval = PyUnicode_FromUnicode(NULL, result_ulength); + if (unlikely(!result_uval)) return NULL; + result_ukind = sizeof(Py_UNICODE); + kind_shift = (result_ukind == 4) ? 2 : result_ukind - 1; + result_udata = PyUnicode_AS_UNICODE(result_uval); +#endif + assert(kind_shift == 2 || kind_shift == 1 || kind_shift == 0); + char_pos = 0; + for (i=0; i < value_count; i++) { + int ukind; + Py_ssize_t ulength; + void *udata; + PyObject *uval = PyTuple_GET_ITEM(value_tuple, i); + if (unlikely(__Pyx_PyUnicode_READY(uval))) + goto bad; + ulength = __Pyx_PyUnicode_GET_LENGTH(uval); + if (unlikely(!ulength)) + continue; + if (unlikely((PY_SSIZE_T_MAX >> kind_shift) - ulength < char_pos)) + goto overflow; + ukind = __Pyx_PyUnicode_KIND(uval); + udata = __Pyx_PyUnicode_DATA(uval); + if (!CYTHON_PEP393_ENABLED || ukind == result_ukind) { + memcpy((char *)result_udata + (char_pos << kind_shift), udata, (size_t) (ulength << kind_shift)); + } else { + #if PY_VERSION_HEX >= 0x030d0000 + if (unlikely(PyUnicode_CopyCharacters(result_uval, char_pos, uval, 0, ulength) < 0)) goto bad; + #elif CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030300F0 || defined(_PyUnicode_FastCopyCharacters) + _PyUnicode_FastCopyCharacters(result_uval, char_pos, uval, 0, ulength); + #else + Py_ssize_t j; + for (j=0; j < ulength; j++) { + Py_UCS4 uchar = __Pyx_PyUnicode_READ(ukind, udata, j); + __Pyx_PyUnicode_WRITE(result_ukind, result_udata, char_pos+j, uchar); + } + #endif + } + char_pos += ulength; + } + return result_uval; +overflow: + PyErr_SetString(PyExc_OverflowError, "join() result is too long for a Python string"); +bad: + Py_DECREF(result_uval); + return NULL; +#else + CYTHON_UNUSED_VAR(max_char); + CYTHON_UNUSED_VAR(result_ulength); + CYTHON_UNUSED_VAR(value_count); + return PyUnicode_Join(__pyx_empty_unicode, value_tuple); +#endif +} + +/* GetAttr */ +static CYTHON_INLINE PyObject *__Pyx_GetAttr(PyObject *o, PyObject *n) { +#if CYTHON_USE_TYPE_SLOTS +#if PY_MAJOR_VERSION >= 3 + if (likely(PyUnicode_Check(n))) +#else + if (likely(PyString_Check(n))) +#endif + return __Pyx_PyObject_GetAttrStr(o, n); +#endif + return PyObject_GetAttr(o, n); +} + +/* GetItemInt */ +static PyObject *__Pyx_GetItemInt_Generic(PyObject *o, PyObject* j) { + PyObject *r; + if (unlikely(!j)) return NULL; + r = PyObject_GetItem(o, j); + Py_DECREF(j); + return r; +} +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_List_Fast(PyObject *o, Py_ssize_t i, + CYTHON_NCP_UNUSED int wraparound, + CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + Py_ssize_t wrapped_i = i; + if (wraparound & unlikely(i < 0)) { + wrapped_i += PyList_GET_SIZE(o); + } + if ((!boundscheck) || likely(__Pyx_is_valid_index(wrapped_i, PyList_GET_SIZE(o)))) { + PyObject *r = PyList_GET_ITEM(o, wrapped_i); + Py_INCREF(r); + return r; + } + return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i)); +#else + return PySequence_GetItem(o, i); +#endif +} +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Tuple_Fast(PyObject *o, Py_ssize_t i, + CYTHON_NCP_UNUSED int wraparound, + CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + Py_ssize_t wrapped_i = i; + if (wraparound & unlikely(i < 0)) { + wrapped_i += PyTuple_GET_SIZE(o); + } + if ((!boundscheck) || likely(__Pyx_is_valid_index(wrapped_i, PyTuple_GET_SIZE(o)))) { + PyObject *r = PyTuple_GET_ITEM(o, wrapped_i); + Py_INCREF(r); + return r; + } + return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i)); +#else + return PySequence_GetItem(o, i); +#endif +} +static CYTHON_INLINE PyObject *__Pyx_GetItemInt_Fast(PyObject *o, Py_ssize_t i, int is_list, + CYTHON_NCP_UNUSED int wraparound, + CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS && CYTHON_USE_TYPE_SLOTS + if (is_list || PyList_CheckExact(o)) { + Py_ssize_t n = ((!wraparound) | likely(i >= 0)) ? i : i + PyList_GET_SIZE(o); + if ((!boundscheck) || (likely(__Pyx_is_valid_index(n, PyList_GET_SIZE(o))))) { + PyObject *r = PyList_GET_ITEM(o, n); + Py_INCREF(r); + return r; + } + } + else if (PyTuple_CheckExact(o)) { + Py_ssize_t n = ((!wraparound) | likely(i >= 0)) ? i : i + PyTuple_GET_SIZE(o); + if ((!boundscheck) || likely(__Pyx_is_valid_index(n, PyTuple_GET_SIZE(o)))) { + PyObject *r = PyTuple_GET_ITEM(o, n); + Py_INCREF(r); + return r; + } + } else { + PyMappingMethods *mm = Py_TYPE(o)->tp_as_mapping; + PySequenceMethods *sm = Py_TYPE(o)->tp_as_sequence; + if (mm && mm->mp_subscript) { + PyObject *r, *key = PyInt_FromSsize_t(i); + if (unlikely(!key)) return NULL; + r = mm->mp_subscript(o, key); + Py_DECREF(key); + return r; + } + if (likely(sm && sm->sq_item)) { + if (wraparound && unlikely(i < 0) && likely(sm->sq_length)) { + Py_ssize_t l = sm->sq_length(o); + if (likely(l >= 0)) { + i += l; + } else { + if (!PyErr_ExceptionMatches(PyExc_OverflowError)) + return NULL; + PyErr_Clear(); + } + } + return sm->sq_item(o, i); + } + } +#else + if (is_list || !PyMapping_Check(o)) { + return PySequence_GetItem(o, i); + } +#endif + return __Pyx_GetItemInt_Generic(o, PyInt_FromSsize_t(i)); +} + +/* PyObjectCallOneArg */ +static CYTHON_INLINE PyObject* __Pyx_PyObject_CallOneArg(PyObject *func, PyObject *arg) { + PyObject *args[2] = {NULL, arg}; + return __Pyx_PyObject_FastCall(func, args+1, 1 | __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET); +} + +/* ObjectGetItem */ +#if CYTHON_USE_TYPE_SLOTS +static PyObject *__Pyx_PyObject_GetIndex(PyObject *obj, PyObject *index) { + PyObject *runerr = NULL; + Py_ssize_t key_value; + key_value = __Pyx_PyIndex_AsSsize_t(index); + if (likely(key_value != -1 || !(runerr = PyErr_Occurred()))) { + return __Pyx_GetItemInt_Fast(obj, key_value, 0, 1, 1); + } + if (PyErr_GivenExceptionMatches(runerr, PyExc_OverflowError)) { + __Pyx_TypeName index_type_name = __Pyx_PyType_GetName(Py_TYPE(index)); + PyErr_Clear(); + PyErr_Format(PyExc_IndexError, + "cannot fit '" __Pyx_FMT_TYPENAME "' into an index-sized integer", index_type_name); + __Pyx_DECREF_TypeName(index_type_name); + } + return NULL; +} +static PyObject *__Pyx_PyObject_GetItem_Slow(PyObject *obj, PyObject *key) { + __Pyx_TypeName obj_type_name; + if (likely(PyType_Check(obj))) { + PyObject *meth = __Pyx_PyObject_GetAttrStrNoError(obj, __pyx_n_s_class_getitem); + if (!meth) { + PyErr_Clear(); + } else { + PyObject *result = __Pyx_PyObject_CallOneArg(meth, key); + Py_DECREF(meth); + return result; + } + } + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, + "'" __Pyx_FMT_TYPENAME "' object is not subscriptable", obj_type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return NULL; +} +static PyObject *__Pyx_PyObject_GetItem(PyObject *obj, PyObject *key) { + PyTypeObject *tp = Py_TYPE(obj); + PyMappingMethods *mm = tp->tp_as_mapping; + PySequenceMethods *sm = tp->tp_as_sequence; + if (likely(mm && mm->mp_subscript)) { + return mm->mp_subscript(obj, key); + } + if (likely(sm && sm->sq_item)) { + return __Pyx_PyObject_GetIndex(obj, key); + } + return __Pyx_PyObject_GetItem_Slow(obj, key); +} +#endif + +/* KeywordStringCheck */ +static int __Pyx_CheckKeywordStrings( + PyObject *kw, + const char* function_name, + int kw_allowed) +{ + PyObject* key = 0; + Py_ssize_t pos = 0; +#if CYTHON_COMPILING_IN_PYPY + if (!kw_allowed && PyDict_Next(kw, &pos, &key, 0)) + goto invalid_keyword; + return 1; +#else + if (CYTHON_METH_FASTCALL && likely(PyTuple_Check(kw))) { + Py_ssize_t kwsize; +#if CYTHON_ASSUME_SAFE_MACROS + kwsize = PyTuple_GET_SIZE(kw); +#else + kwsize = PyTuple_Size(kw); + if (kwsize < 0) return 0; +#endif + if (unlikely(kwsize == 0)) + return 1; + if (!kw_allowed) { +#if CYTHON_ASSUME_SAFE_MACROS + key = PyTuple_GET_ITEM(kw, 0); +#else + key = PyTuple_GetItem(kw, pos); + if (!key) return 0; +#endif + goto invalid_keyword; + } +#if PY_VERSION_HEX < 0x03090000 + for (pos = 0; pos < kwsize; pos++) { +#if CYTHON_ASSUME_SAFE_MACROS + key = PyTuple_GET_ITEM(kw, pos); +#else + key = PyTuple_GetItem(kw, pos); + if (!key) return 0; +#endif + if (unlikely(!PyUnicode_Check(key))) + goto invalid_keyword_type; + } +#endif + return 1; + } + while (PyDict_Next(kw, &pos, &key, 0)) { + #if PY_MAJOR_VERSION < 3 + if (unlikely(!PyString_Check(key))) + #endif + if (unlikely(!PyUnicode_Check(key))) + goto invalid_keyword_type; + } + if (!kw_allowed && unlikely(key)) + goto invalid_keyword; + return 1; +invalid_keyword_type: + PyErr_Format(PyExc_TypeError, + "%.200s() keywords must be strings", function_name); + return 0; +#endif +invalid_keyword: + #if PY_MAJOR_VERSION < 3 + PyErr_Format(PyExc_TypeError, + "%.200s() got an unexpected keyword argument '%.200s'", + function_name, PyString_AsString(key)); + #else + PyErr_Format(PyExc_TypeError, + "%s() got an unexpected keyword argument '%U'", + function_name, key); + #endif + return 0; +} + +/* DivInt[Py_ssize_t] */ +static CYTHON_INLINE Py_ssize_t __Pyx_div_Py_ssize_t(Py_ssize_t a, Py_ssize_t b) { + Py_ssize_t q = a / b; + Py_ssize_t r = a - q*b; + q -= ((r != 0) & ((r ^ b) < 0)); + return q; +} + +/* GetAttr3 */ +#if __PYX_LIMITED_VERSION_HEX < 0x030d00A1 +static PyObject *__Pyx_GetAttr3Default(PyObject *d) { + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + if (unlikely(!__Pyx_PyErr_ExceptionMatches(PyExc_AttributeError))) + return NULL; + __Pyx_PyErr_Clear(); + Py_INCREF(d); + return d; +} +#endif +static CYTHON_INLINE PyObject *__Pyx_GetAttr3(PyObject *o, PyObject *n, PyObject *d) { + PyObject *r; +#if __PYX_LIMITED_VERSION_HEX >= 0x030d00A1 + int res = PyObject_GetOptionalAttr(o, n, &r); + return (res != 0) ? r : __Pyx_NewRef(d); +#else + #if CYTHON_USE_TYPE_SLOTS + if (likely(PyString_Check(n))) { + r = __Pyx_PyObject_GetAttrStrNoError(o, n); + if (unlikely(!r) && likely(!PyErr_Occurred())) { + r = __Pyx_NewRef(d); + } + return r; + } + #endif + r = PyObject_GetAttr(o, n); + return (likely(r)) ? r : __Pyx_GetAttr3Default(d); +#endif +} + +/* PyDictVersioning */ +#if CYTHON_USE_DICT_VERSIONS && CYTHON_USE_TYPE_SLOTS +static CYTHON_INLINE PY_UINT64_T __Pyx_get_tp_dict_version(PyObject *obj) { + PyObject *dict = Py_TYPE(obj)->tp_dict; + return likely(dict) ? __PYX_GET_DICT_VERSION(dict) : 0; +} +static CYTHON_INLINE PY_UINT64_T __Pyx_get_object_dict_version(PyObject *obj) { + PyObject **dictptr = NULL; + Py_ssize_t offset = Py_TYPE(obj)->tp_dictoffset; + if (offset) { +#if CYTHON_COMPILING_IN_CPYTHON + dictptr = (likely(offset > 0)) ? (PyObject **) ((char *)obj + offset) : _PyObject_GetDictPtr(obj); +#else + dictptr = _PyObject_GetDictPtr(obj); +#endif + } + return (dictptr && *dictptr) ? __PYX_GET_DICT_VERSION(*dictptr) : 0; +} +static CYTHON_INLINE int __Pyx_object_dict_version_matches(PyObject* obj, PY_UINT64_T tp_dict_version, PY_UINT64_T obj_dict_version) { + PyObject *dict = Py_TYPE(obj)->tp_dict; + if (unlikely(!dict) || unlikely(tp_dict_version != __PYX_GET_DICT_VERSION(dict))) + return 0; + return obj_dict_version == __Pyx_get_object_dict_version(obj); +} +#endif + +/* GetModuleGlobalName */ +#if CYTHON_USE_DICT_VERSIONS +static PyObject *__Pyx__GetModuleGlobalName(PyObject *name, PY_UINT64_T *dict_version, PyObject **dict_cached_value) +#else +static CYTHON_INLINE PyObject *__Pyx__GetModuleGlobalName(PyObject *name) +#endif +{ + PyObject *result; +#if !CYTHON_AVOID_BORROWED_REFS +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030500A1 && PY_VERSION_HEX < 0x030d0000 + result = _PyDict_GetItem_KnownHash(__pyx_d, name, ((PyASCIIObject *) name)->hash); + __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version) + if (likely(result)) { + return __Pyx_NewRef(result); + } else if (unlikely(PyErr_Occurred())) { + return NULL; + } +#elif CYTHON_COMPILING_IN_LIMITED_API + if (unlikely(!__pyx_m)) { + return NULL; + } + result = PyObject_GetAttr(__pyx_m, name); + if (likely(result)) { + return result; + } +#else + result = PyDict_GetItem(__pyx_d, name); + __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version) + if (likely(result)) { + return __Pyx_NewRef(result); + } +#endif +#else + result = PyObject_GetItem(__pyx_d, name); + __PYX_UPDATE_DICT_CACHE(__pyx_d, result, *dict_cached_value, *dict_version) + if (likely(result)) { + return __Pyx_NewRef(result); + } + PyErr_Clear(); +#endif + return __Pyx_GetBuiltinName(name); +} + +/* RaiseTooManyValuesToUnpack */ +static CYTHON_INLINE void __Pyx_RaiseTooManyValuesError(Py_ssize_t expected) { + PyErr_Format(PyExc_ValueError, + "too many values to unpack (expected %" CYTHON_FORMAT_SSIZE_T "d)", expected); +} + +/* RaiseNeedMoreValuesToUnpack */ +static CYTHON_INLINE void __Pyx_RaiseNeedMoreValuesError(Py_ssize_t index) { + PyErr_Format(PyExc_ValueError, + "need more than %" CYTHON_FORMAT_SSIZE_T "d value%.1s to unpack", + index, (index == 1) ? "" : "s"); +} + +/* RaiseNoneIterError */ +static CYTHON_INLINE void __Pyx_RaiseNoneNotIterableError(void) { + PyErr_SetString(PyExc_TypeError, "'NoneType' object is not iterable"); +} + +/* ExtTypeTest */ +static CYTHON_INLINE int __Pyx_TypeTest(PyObject *obj, PyTypeObject *type) { + __Pyx_TypeName obj_type_name; + __Pyx_TypeName type_name; + if (unlikely(!type)) { + PyErr_SetString(PyExc_SystemError, "Missing type object"); + return 0; + } + if (likely(__Pyx_TypeCheck(obj, type))) + return 1; + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + type_name = __Pyx_PyType_GetName(type); + PyErr_Format(PyExc_TypeError, + "Cannot convert " __Pyx_FMT_TYPENAME " to " __Pyx_FMT_TYPENAME, + obj_type_name, type_name); + __Pyx_DECREF_TypeName(obj_type_name); + __Pyx_DECREF_TypeName(type_name); + return 0; +} + +/* GetTopmostException */ +#if CYTHON_USE_EXC_INFO_STACK && CYTHON_FAST_THREAD_STATE +static _PyErr_StackItem * +__Pyx_PyErr_GetTopmostException(PyThreadState *tstate) +{ + _PyErr_StackItem *exc_info = tstate->exc_info; + while ((exc_info->exc_value == NULL || exc_info->exc_value == Py_None) && + exc_info->previous_item != NULL) + { + exc_info = exc_info->previous_item; + } + return exc_info; +} +#endif + +/* SaveResetException */ +#if CYTHON_FAST_THREAD_STATE +static CYTHON_INLINE void __Pyx__ExceptionSave(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) { + #if CYTHON_USE_EXC_INFO_STACK && PY_VERSION_HEX >= 0x030B00a4 + _PyErr_StackItem *exc_info = __Pyx_PyErr_GetTopmostException(tstate); + PyObject *exc_value = exc_info->exc_value; + if (exc_value == NULL || exc_value == Py_None) { + *value = NULL; + *type = NULL; + *tb = NULL; + } else { + *value = exc_value; + Py_INCREF(*value); + *type = (PyObject*) Py_TYPE(exc_value); + Py_INCREF(*type); + *tb = PyException_GetTraceback(exc_value); + } + #elif CYTHON_USE_EXC_INFO_STACK + _PyErr_StackItem *exc_info = __Pyx_PyErr_GetTopmostException(tstate); + *type = exc_info->exc_type; + *value = exc_info->exc_value; + *tb = exc_info->exc_traceback; + Py_XINCREF(*type); + Py_XINCREF(*value); + Py_XINCREF(*tb); + #else + *type = tstate->exc_type; + *value = tstate->exc_value; + *tb = tstate->exc_traceback; + Py_XINCREF(*type); + Py_XINCREF(*value); + Py_XINCREF(*tb); + #endif +} +static CYTHON_INLINE void __Pyx__ExceptionReset(PyThreadState *tstate, PyObject *type, PyObject *value, PyObject *tb) { + #if CYTHON_USE_EXC_INFO_STACK && PY_VERSION_HEX >= 0x030B00a4 + _PyErr_StackItem *exc_info = tstate->exc_info; + PyObject *tmp_value = exc_info->exc_value; + exc_info->exc_value = value; + Py_XDECREF(tmp_value); + Py_XDECREF(type); + Py_XDECREF(tb); + #else + PyObject *tmp_type, *tmp_value, *tmp_tb; + #if CYTHON_USE_EXC_INFO_STACK + _PyErr_StackItem *exc_info = tstate->exc_info; + tmp_type = exc_info->exc_type; + tmp_value = exc_info->exc_value; + tmp_tb = exc_info->exc_traceback; + exc_info->exc_type = type; + exc_info->exc_value = value; + exc_info->exc_traceback = tb; + #else + tmp_type = tstate->exc_type; + tmp_value = tstate->exc_value; + tmp_tb = tstate->exc_traceback; + tstate->exc_type = type; + tstate->exc_value = value; + tstate->exc_traceback = tb; + #endif + Py_XDECREF(tmp_type); + Py_XDECREF(tmp_value); + Py_XDECREF(tmp_tb); + #endif +} +#endif + +/* GetException */ +#if CYTHON_FAST_THREAD_STATE +static int __Pyx__GetException(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) +#else +static int __Pyx_GetException(PyObject **type, PyObject **value, PyObject **tb) +#endif +{ + PyObject *local_type = NULL, *local_value, *local_tb = NULL; +#if CYTHON_FAST_THREAD_STATE + PyObject *tmp_type, *tmp_value, *tmp_tb; + #if PY_VERSION_HEX >= 0x030C00A6 + local_value = tstate->current_exception; + tstate->current_exception = 0; + if (likely(local_value)) { + local_type = (PyObject*) Py_TYPE(local_value); + Py_INCREF(local_type); + local_tb = PyException_GetTraceback(local_value); + } + #else + local_type = tstate->curexc_type; + local_value = tstate->curexc_value; + local_tb = tstate->curexc_traceback; + tstate->curexc_type = 0; + tstate->curexc_value = 0; + tstate->curexc_traceback = 0; + #endif +#else + PyErr_Fetch(&local_type, &local_value, &local_tb); +#endif + PyErr_NormalizeException(&local_type, &local_value, &local_tb); +#if CYTHON_FAST_THREAD_STATE && PY_VERSION_HEX >= 0x030C00A6 + if (unlikely(tstate->current_exception)) +#elif CYTHON_FAST_THREAD_STATE + if (unlikely(tstate->curexc_type)) +#else + if (unlikely(PyErr_Occurred())) +#endif + goto bad; + #if PY_MAJOR_VERSION >= 3 + if (local_tb) { + if (unlikely(PyException_SetTraceback(local_value, local_tb) < 0)) + goto bad; + } + #endif + Py_XINCREF(local_tb); + Py_XINCREF(local_type); + Py_XINCREF(local_value); + *type = local_type; + *value = local_value; + *tb = local_tb; +#if CYTHON_FAST_THREAD_STATE + #if CYTHON_USE_EXC_INFO_STACK + { + _PyErr_StackItem *exc_info = tstate->exc_info; + #if PY_VERSION_HEX >= 0x030B00a4 + tmp_value = exc_info->exc_value; + exc_info->exc_value = local_value; + tmp_type = NULL; + tmp_tb = NULL; + Py_XDECREF(local_type); + Py_XDECREF(local_tb); + #else + tmp_type = exc_info->exc_type; + tmp_value = exc_info->exc_value; + tmp_tb = exc_info->exc_traceback; + exc_info->exc_type = local_type; + exc_info->exc_value = local_value; + exc_info->exc_traceback = local_tb; + #endif + } + #else + tmp_type = tstate->exc_type; + tmp_value = tstate->exc_value; + tmp_tb = tstate->exc_traceback; + tstate->exc_type = local_type; + tstate->exc_value = local_value; + tstate->exc_traceback = local_tb; + #endif + Py_XDECREF(tmp_type); + Py_XDECREF(tmp_value); + Py_XDECREF(tmp_tb); +#else + PyErr_SetExcInfo(local_type, local_value, local_tb); +#endif + return 0; +bad: + *type = 0; + *value = 0; + *tb = 0; + Py_XDECREF(local_type); + Py_XDECREF(local_value); + Py_XDECREF(local_tb); + return -1; +} + +/* SwapException */ +#if CYTHON_FAST_THREAD_STATE +static CYTHON_INLINE void __Pyx__ExceptionSwap(PyThreadState *tstate, PyObject **type, PyObject **value, PyObject **tb) { + PyObject *tmp_type, *tmp_value, *tmp_tb; + #if CYTHON_USE_EXC_INFO_STACK && PY_VERSION_HEX >= 0x030B00a4 + _PyErr_StackItem *exc_info = tstate->exc_info; + tmp_value = exc_info->exc_value; + exc_info->exc_value = *value; + if (tmp_value == NULL || tmp_value == Py_None) { + Py_XDECREF(tmp_value); + tmp_value = NULL; + tmp_type = NULL; + tmp_tb = NULL; + } else { + tmp_type = (PyObject*) Py_TYPE(tmp_value); + Py_INCREF(tmp_type); + #if CYTHON_COMPILING_IN_CPYTHON + tmp_tb = ((PyBaseExceptionObject*) tmp_value)->traceback; + Py_XINCREF(tmp_tb); + #else + tmp_tb = PyException_GetTraceback(tmp_value); + #endif + } + #elif CYTHON_USE_EXC_INFO_STACK + _PyErr_StackItem *exc_info = tstate->exc_info; + tmp_type = exc_info->exc_type; + tmp_value = exc_info->exc_value; + tmp_tb = exc_info->exc_traceback; + exc_info->exc_type = *type; + exc_info->exc_value = *value; + exc_info->exc_traceback = *tb; + #else + tmp_type = tstate->exc_type; + tmp_value = tstate->exc_value; + tmp_tb = tstate->exc_traceback; + tstate->exc_type = *type; + tstate->exc_value = *value; + tstate->exc_traceback = *tb; + #endif + *type = tmp_type; + *value = tmp_value; + *tb = tmp_tb; +} +#else +static CYTHON_INLINE void __Pyx_ExceptionSwap(PyObject **type, PyObject **value, PyObject **tb) { + PyObject *tmp_type, *tmp_value, *tmp_tb; + PyErr_GetExcInfo(&tmp_type, &tmp_value, &tmp_tb); + PyErr_SetExcInfo(*type, *value, *tb); + *type = tmp_type; + *value = tmp_value; + *tb = tmp_tb; +} +#endif + +/* Import */ +static PyObject *__Pyx_Import(PyObject *name, PyObject *from_list, int level) { + PyObject *module = 0; + PyObject *empty_dict = 0; + PyObject *empty_list = 0; + #if PY_MAJOR_VERSION < 3 + PyObject *py_import; + py_import = __Pyx_PyObject_GetAttrStr(__pyx_b, __pyx_n_s_import); + if (unlikely(!py_import)) + goto bad; + if (!from_list) { + empty_list = PyList_New(0); + if (unlikely(!empty_list)) + goto bad; + from_list = empty_list; + } + #endif + empty_dict = PyDict_New(); + if (unlikely(!empty_dict)) + goto bad; + { + #if PY_MAJOR_VERSION >= 3 + if (level == -1) { + if (strchr(__Pyx_MODULE_NAME, '.') != NULL) { + module = PyImport_ImportModuleLevelObject( + name, __pyx_d, empty_dict, from_list, 1); + if (unlikely(!module)) { + if (unlikely(!PyErr_ExceptionMatches(PyExc_ImportError))) + goto bad; + PyErr_Clear(); + } + } + level = 0; + } + #endif + if (!module) { + #if PY_MAJOR_VERSION < 3 + PyObject *py_level = PyInt_FromLong(level); + if (unlikely(!py_level)) + goto bad; + module = PyObject_CallFunctionObjArgs(py_import, + name, __pyx_d, empty_dict, from_list, py_level, (PyObject *)NULL); + Py_DECREF(py_level); + #else + module = PyImport_ImportModuleLevelObject( + name, __pyx_d, empty_dict, from_list, level); + #endif + } + } +bad: + Py_XDECREF(empty_dict); + Py_XDECREF(empty_list); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(py_import); + #endif + return module; +} + +/* ImportDottedModule */ +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx__ImportDottedModule_Error(PyObject *name, PyObject *parts_tuple, Py_ssize_t count) { + PyObject *partial_name = NULL, *slice = NULL, *sep = NULL; + if (unlikely(PyErr_Occurred())) { + PyErr_Clear(); + } + if (likely(PyTuple_GET_SIZE(parts_tuple) == count)) { + partial_name = name; + } else { + slice = PySequence_GetSlice(parts_tuple, 0, count); + if (unlikely(!slice)) + goto bad; + sep = PyUnicode_FromStringAndSize(".", 1); + if (unlikely(!sep)) + goto bad; + partial_name = PyUnicode_Join(sep, slice); + } + PyErr_Format( +#if PY_MAJOR_VERSION < 3 + PyExc_ImportError, + "No module named '%s'", PyString_AS_STRING(partial_name)); +#else +#if PY_VERSION_HEX >= 0x030600B1 + PyExc_ModuleNotFoundError, +#else + PyExc_ImportError, +#endif + "No module named '%U'", partial_name); +#endif +bad: + Py_XDECREF(sep); + Py_XDECREF(slice); + Py_XDECREF(partial_name); + return NULL; +} +#endif +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx__ImportDottedModule_Lookup(PyObject *name) { + PyObject *imported_module; +#if PY_VERSION_HEX < 0x030700A1 || (CYTHON_COMPILING_IN_PYPY && PYPY_VERSION_NUM < 0x07030400) + PyObject *modules = PyImport_GetModuleDict(); + if (unlikely(!modules)) + return NULL; + imported_module = __Pyx_PyDict_GetItemStr(modules, name); + Py_XINCREF(imported_module); +#else + imported_module = PyImport_GetModule(name); +#endif + return imported_module; +} +#endif +#if PY_MAJOR_VERSION >= 3 +static PyObject *__Pyx_ImportDottedModule_WalkParts(PyObject *module, PyObject *name, PyObject *parts_tuple) { + Py_ssize_t i, nparts; + nparts = PyTuple_GET_SIZE(parts_tuple); + for (i=1; i < nparts && module; i++) { + PyObject *part, *submodule; +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + part = PyTuple_GET_ITEM(parts_tuple, i); +#else + part = PySequence_ITEM(parts_tuple, i); +#endif + submodule = __Pyx_PyObject_GetAttrStrNoError(module, part); +#if !(CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS) + Py_DECREF(part); +#endif + Py_DECREF(module); + module = submodule; + } + if (unlikely(!module)) { + return __Pyx__ImportDottedModule_Error(name, parts_tuple, i); + } + return module; +} +#endif +static PyObject *__Pyx__ImportDottedModule(PyObject *name, PyObject *parts_tuple) { +#if PY_MAJOR_VERSION < 3 + PyObject *module, *from_list, *star = __pyx_n_s__3; + CYTHON_UNUSED_VAR(parts_tuple); + from_list = PyList_New(1); + if (unlikely(!from_list)) + return NULL; + Py_INCREF(star); + PyList_SET_ITEM(from_list, 0, star); + module = __Pyx_Import(name, from_list, 0); + Py_DECREF(from_list); + return module; +#else + PyObject *imported_module; + PyObject *module = __Pyx_Import(name, NULL, 0); + if (!parts_tuple || unlikely(!module)) + return module; + imported_module = __Pyx__ImportDottedModule_Lookup(name); + if (likely(imported_module)) { + Py_DECREF(module); + return imported_module; + } + PyErr_Clear(); + return __Pyx_ImportDottedModule_WalkParts(module, name, parts_tuple); +#endif +} +static PyObject *__Pyx_ImportDottedModule(PyObject *name, PyObject *parts_tuple) { +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX >= 0x030400B1 + PyObject *module = __Pyx__ImportDottedModule_Lookup(name); + if (likely(module)) { + PyObject *spec = __Pyx_PyObject_GetAttrStrNoError(module, __pyx_n_s_spec); + if (likely(spec)) { + PyObject *unsafe = __Pyx_PyObject_GetAttrStrNoError(spec, __pyx_n_s_initializing); + if (likely(!unsafe || !__Pyx_PyObject_IsTrue(unsafe))) { + Py_DECREF(spec); + spec = NULL; + } + Py_XDECREF(unsafe); + } + if (likely(!spec)) { + PyErr_Clear(); + return module; + } + Py_DECREF(spec); + Py_DECREF(module); + } else if (PyErr_Occurred()) { + PyErr_Clear(); + } +#endif + return __Pyx__ImportDottedModule(name, parts_tuple); +} + +/* FastTypeChecks */ +#if CYTHON_COMPILING_IN_CPYTHON +static int __Pyx_InBases(PyTypeObject *a, PyTypeObject *b) { + while (a) { + a = __Pyx_PyType_GetSlot(a, tp_base, PyTypeObject*); + if (a == b) + return 1; + } + return b == &PyBaseObject_Type; +} +static CYTHON_INLINE int __Pyx_IsSubtype(PyTypeObject *a, PyTypeObject *b) { + PyObject *mro; + if (a == b) return 1; + mro = a->tp_mro; + if (likely(mro)) { + Py_ssize_t i, n; + n = PyTuple_GET_SIZE(mro); + for (i = 0; i < n; i++) { + if (PyTuple_GET_ITEM(mro, i) == (PyObject *)b) + return 1; + } + return 0; + } + return __Pyx_InBases(a, b); +} +static CYTHON_INLINE int __Pyx_IsAnySubtype2(PyTypeObject *cls, PyTypeObject *a, PyTypeObject *b) { + PyObject *mro; + if (cls == a || cls == b) return 1; + mro = cls->tp_mro; + if (likely(mro)) { + Py_ssize_t i, n; + n = PyTuple_GET_SIZE(mro); + for (i = 0; i < n; i++) { + PyObject *base = PyTuple_GET_ITEM(mro, i); + if (base == (PyObject *)a || base == (PyObject *)b) + return 1; + } + return 0; + } + return __Pyx_InBases(cls, a) || __Pyx_InBases(cls, b); +} +#if PY_MAJOR_VERSION == 2 +static int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject* exc_type2) { + PyObject *exception, *value, *tb; + int res; + __Pyx_PyThreadState_declare + __Pyx_PyThreadState_assign + __Pyx_ErrFetch(&exception, &value, &tb); + res = exc_type1 ? PyObject_IsSubclass(err, exc_type1) : 0; + if (unlikely(res == -1)) { + PyErr_WriteUnraisable(err); + res = 0; + } + if (!res) { + res = PyObject_IsSubclass(err, exc_type2); + if (unlikely(res == -1)) { + PyErr_WriteUnraisable(err); + res = 0; + } + } + __Pyx_ErrRestore(exception, value, tb); + return res; +} +#else +static CYTHON_INLINE int __Pyx_inner_PyErr_GivenExceptionMatches2(PyObject *err, PyObject* exc_type1, PyObject *exc_type2) { + if (exc_type1) { + return __Pyx_IsAnySubtype2((PyTypeObject*)err, (PyTypeObject*)exc_type1, (PyTypeObject*)exc_type2); + } else { + return __Pyx_IsSubtype((PyTypeObject*)err, (PyTypeObject*)exc_type2); + } +} +#endif +static int __Pyx_PyErr_GivenExceptionMatchesTuple(PyObject *exc_type, PyObject *tuple) { + Py_ssize_t i, n; + assert(PyExceptionClass_Check(exc_type)); + n = PyTuple_GET_SIZE(tuple); +#if PY_MAJOR_VERSION >= 3 + for (i=0; itp_as_sequence && type->tp_as_sequence->sq_repeat)) { + return type->tp_as_sequence->sq_repeat(seq, mul); + } else +#endif + { + return __Pyx_PySequence_Multiply_Generic(seq, mul); + } +} + +/* SetItemInt */ +static int __Pyx_SetItemInt_Generic(PyObject *o, PyObject *j, PyObject *v) { + int r; + if (unlikely(!j)) return -1; + r = PyObject_SetItem(o, j, v); + Py_DECREF(j); + return r; +} +static CYTHON_INLINE int __Pyx_SetItemInt_Fast(PyObject *o, Py_ssize_t i, PyObject *v, int is_list, + CYTHON_NCP_UNUSED int wraparound, CYTHON_NCP_UNUSED int boundscheck) { +#if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS && CYTHON_USE_TYPE_SLOTS + if (is_list || PyList_CheckExact(o)) { + Py_ssize_t n = (!wraparound) ? i : ((likely(i >= 0)) ? i : i + PyList_GET_SIZE(o)); + if ((!boundscheck) || likely(__Pyx_is_valid_index(n, PyList_GET_SIZE(o)))) { + PyObject* old = PyList_GET_ITEM(o, n); + Py_INCREF(v); + PyList_SET_ITEM(o, n, v); + Py_DECREF(old); + return 1; + } + } else { + PyMappingMethods *mm = Py_TYPE(o)->tp_as_mapping; + PySequenceMethods *sm = Py_TYPE(o)->tp_as_sequence; + if (mm && mm->mp_ass_subscript) { + int r; + PyObject *key = PyInt_FromSsize_t(i); + if (unlikely(!key)) return -1; + r = mm->mp_ass_subscript(o, key, v); + Py_DECREF(key); + return r; + } + if (likely(sm && sm->sq_ass_item)) { + if (wraparound && unlikely(i < 0) && likely(sm->sq_length)) { + Py_ssize_t l = sm->sq_length(o); + if (likely(l >= 0)) { + i += l; + } else { + if (!PyErr_ExceptionMatches(PyExc_OverflowError)) + return -1; + PyErr_Clear(); + } + } + return sm->sq_ass_item(o, i, v); + } + } +#else + if (is_list || !PyMapping_Check(o)) + { + return PySequence_SetItem(o, i, v); + } +#endif + return __Pyx_SetItemInt_Generic(o, PyInt_FromSsize_t(i), v); +} + +/* RaiseUnboundLocalError */ +static CYTHON_INLINE void __Pyx_RaiseUnboundLocalError(const char *varname) { + PyErr_Format(PyExc_UnboundLocalError, "local variable '%s' referenced before assignment", varname); +} + +/* DivInt[long] */ +static CYTHON_INLINE long __Pyx_div_long(long a, long b) { + long q = a / b; + long r = a - q*b; + q -= ((r != 0) & ((r ^ b) < 0)); + return q; +} + +/* ImportFrom */ +static PyObject* __Pyx_ImportFrom(PyObject* module, PyObject* name) { + PyObject* value = __Pyx_PyObject_GetAttrStr(module, name); + if (unlikely(!value) && PyErr_ExceptionMatches(PyExc_AttributeError)) { + const char* module_name_str = 0; + PyObject* module_name = 0; + PyObject* module_dot = 0; + PyObject* full_name = 0; + PyErr_Clear(); + module_name_str = PyModule_GetName(module); + if (unlikely(!module_name_str)) { goto modbad; } + module_name = PyUnicode_FromString(module_name_str); + if (unlikely(!module_name)) { goto modbad; } + module_dot = PyUnicode_Concat(module_name, __pyx_kp_u__2); + if (unlikely(!module_dot)) { goto modbad; } + full_name = PyUnicode_Concat(module_dot, name); + if (unlikely(!full_name)) { goto modbad; } + #if PY_VERSION_HEX < 0x030700A1 || (CYTHON_COMPILING_IN_PYPY && PYPY_VERSION_NUM < 0x07030400) + { + PyObject *modules = PyImport_GetModuleDict(); + if (unlikely(!modules)) + goto modbad; + value = PyObject_GetItem(modules, full_name); + } + #else + value = PyImport_GetModule(full_name); + #endif + modbad: + Py_XDECREF(full_name); + Py_XDECREF(module_dot); + Py_XDECREF(module_name); + } + if (unlikely(!value)) { + PyErr_Format(PyExc_ImportError, + #if PY_MAJOR_VERSION < 3 + "cannot import name %.230s", PyString_AS_STRING(name)); + #else + "cannot import name %S", name); + #endif + } + return value; +} + +/* HasAttr */ +static CYTHON_INLINE int __Pyx_HasAttr(PyObject *o, PyObject *n) { + PyObject *r; + if (unlikely(!__Pyx_PyBaseString_Check(n))) { + PyErr_SetString(PyExc_TypeError, + "hasattr(): attribute name must be string"); + return -1; + } + r = __Pyx_GetAttr(o, n); + if (!r) { + PyErr_Clear(); + return 0; + } else { + Py_DECREF(r); + return 1; + } +} + +/* IsLittleEndian */ +static CYTHON_INLINE int __Pyx_Is_Little_Endian(void) +{ + union { + uint32_t u32; + uint8_t u8[4]; + } S; + S.u32 = 0x01020304; + return S.u8[0] == 4; +} + +/* BufferFormatCheck */ +static void __Pyx_BufFmt_Init(__Pyx_BufFmt_Context* ctx, + __Pyx_BufFmt_StackElem* stack, + __Pyx_TypeInfo* type) { + stack[0].field = &ctx->root; + stack[0].parent_offset = 0; + ctx->root.type = type; + ctx->root.name = "buffer dtype"; + ctx->root.offset = 0; + ctx->head = stack; + ctx->head->field = &ctx->root; + ctx->fmt_offset = 0; + ctx->head->parent_offset = 0; + ctx->new_packmode = '@'; + ctx->enc_packmode = '@'; + ctx->new_count = 1; + ctx->enc_count = 0; + ctx->enc_type = 0; + ctx->is_complex = 0; + ctx->is_valid_array = 0; + ctx->struct_alignment = 0; + while (type->typegroup == 'S') { + ++ctx->head; + ctx->head->field = type->fields; + ctx->head->parent_offset = 0; + type = type->fields->type; + } +} +static int __Pyx_BufFmt_ParseNumber(const char** ts) { + int count; + const char* t = *ts; + if (*t < '0' || *t > '9') { + return -1; + } else { + count = *t++ - '0'; + while (*t >= '0' && *t <= '9') { + count *= 10; + count += *t++ - '0'; + } + } + *ts = t; + return count; +} +static int __Pyx_BufFmt_ExpectNumber(const char **ts) { + int number = __Pyx_BufFmt_ParseNumber(ts); + if (number == -1) + PyErr_Format(PyExc_ValueError,\ + "Does not understand character buffer dtype format string ('%c')", **ts); + return number; +} +static void __Pyx_BufFmt_RaiseUnexpectedChar(char ch) { + PyErr_Format(PyExc_ValueError, + "Unexpected format string character: '%c'", ch); +} +static const char* __Pyx_BufFmt_DescribeTypeChar(char ch, int is_complex) { + switch (ch) { + case '?': return "'bool'"; + case 'c': return "'char'"; + case 'b': return "'signed char'"; + case 'B': return "'unsigned char'"; + case 'h': return "'short'"; + case 'H': return "'unsigned short'"; + case 'i': return "'int'"; + case 'I': return "'unsigned int'"; + case 'l': return "'long'"; + case 'L': return "'unsigned long'"; + case 'q': return "'long long'"; + case 'Q': return "'unsigned long long'"; + case 'f': return (is_complex ? "'complex float'" : "'float'"); + case 'd': return (is_complex ? "'complex double'" : "'double'"); + case 'g': return (is_complex ? "'complex long double'" : "'long double'"); + case 'T': return "a struct"; + case 'O': return "Python object"; + case 'P': return "a pointer"; + case 's': case 'p': return "a string"; + case 0: return "end"; + default: return "unparsable format string"; + } +} +static size_t __Pyx_BufFmt_TypeCharToStandardSize(char ch, int is_complex) { + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return 2; + case 'i': case 'I': case 'l': case 'L': return 4; + case 'q': case 'Q': return 8; + case 'f': return (is_complex ? 8 : 4); + case 'd': return (is_complex ? 16 : 8); + case 'g': { + PyErr_SetString(PyExc_ValueError, "Python does not define a standard format string size for long double ('g').."); + return 0; + } + case 'O': case 'P': return sizeof(void*); + default: + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } +} +static size_t __Pyx_BufFmt_TypeCharToNativeSize(char ch, int is_complex) { + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return sizeof(short); + case 'i': case 'I': return sizeof(int); + case 'l': case 'L': return sizeof(long); + #ifdef HAVE_LONG_LONG + case 'q': case 'Q': return sizeof(PY_LONG_LONG); + #endif + case 'f': return sizeof(float) * (is_complex ? 2 : 1); + case 'd': return sizeof(double) * (is_complex ? 2 : 1); + case 'g': return sizeof(long double) * (is_complex ? 2 : 1); + case 'O': case 'P': return sizeof(void*); + default: { + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } + } +} +typedef struct { char c; short x; } __Pyx_st_short; +typedef struct { char c; int x; } __Pyx_st_int; +typedef struct { char c; long x; } __Pyx_st_long; +typedef struct { char c; float x; } __Pyx_st_float; +typedef struct { char c; double x; } __Pyx_st_double; +typedef struct { char c; long double x; } __Pyx_st_longdouble; +typedef struct { char c; void *x; } __Pyx_st_void_p; +#ifdef HAVE_LONG_LONG +typedef struct { char c; PY_LONG_LONG x; } __Pyx_st_longlong; +#endif +static size_t __Pyx_BufFmt_TypeCharToAlignment(char ch, int is_complex) { + CYTHON_UNUSED_VAR(is_complex); + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return sizeof(__Pyx_st_short) - sizeof(short); + case 'i': case 'I': return sizeof(__Pyx_st_int) - sizeof(int); + case 'l': case 'L': return sizeof(__Pyx_st_long) - sizeof(long); +#ifdef HAVE_LONG_LONG + case 'q': case 'Q': return sizeof(__Pyx_st_longlong) - sizeof(PY_LONG_LONG); +#endif + case 'f': return sizeof(__Pyx_st_float) - sizeof(float); + case 'd': return sizeof(__Pyx_st_double) - sizeof(double); + case 'g': return sizeof(__Pyx_st_longdouble) - sizeof(long double); + case 'P': case 'O': return sizeof(__Pyx_st_void_p) - sizeof(void*); + default: + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } +} +/* These are for computing the padding at the end of the struct to align + on the first member of the struct. This will probably the same as above, + but we don't have any guarantees. + */ +typedef struct { short x; char c; } __Pyx_pad_short; +typedef struct { int x; char c; } __Pyx_pad_int; +typedef struct { long x; char c; } __Pyx_pad_long; +typedef struct { float x; char c; } __Pyx_pad_float; +typedef struct { double x; char c; } __Pyx_pad_double; +typedef struct { long double x; char c; } __Pyx_pad_longdouble; +typedef struct { void *x; char c; } __Pyx_pad_void_p; +#ifdef HAVE_LONG_LONG +typedef struct { PY_LONG_LONG x; char c; } __Pyx_pad_longlong; +#endif +static size_t __Pyx_BufFmt_TypeCharToPadding(char ch, int is_complex) { + CYTHON_UNUSED_VAR(is_complex); + switch (ch) { + case '?': case 'c': case 'b': case 'B': case 's': case 'p': return 1; + case 'h': case 'H': return sizeof(__Pyx_pad_short) - sizeof(short); + case 'i': case 'I': return sizeof(__Pyx_pad_int) - sizeof(int); + case 'l': case 'L': return sizeof(__Pyx_pad_long) - sizeof(long); +#ifdef HAVE_LONG_LONG + case 'q': case 'Q': return sizeof(__Pyx_pad_longlong) - sizeof(PY_LONG_LONG); +#endif + case 'f': return sizeof(__Pyx_pad_float) - sizeof(float); + case 'd': return sizeof(__Pyx_pad_double) - sizeof(double); + case 'g': return sizeof(__Pyx_pad_longdouble) - sizeof(long double); + case 'P': case 'O': return sizeof(__Pyx_pad_void_p) - sizeof(void*); + default: + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } +} +static char __Pyx_BufFmt_TypeCharToGroup(char ch, int is_complex) { + switch (ch) { + case 'c': + return 'H'; + case 'b': case 'h': case 'i': + case 'l': case 'q': case 's': case 'p': + return 'I'; + case '?': case 'B': case 'H': case 'I': case 'L': case 'Q': + return 'U'; + case 'f': case 'd': case 'g': + return (is_complex ? 'C' : 'R'); + case 'O': + return 'O'; + case 'P': + return 'P'; + default: { + __Pyx_BufFmt_RaiseUnexpectedChar(ch); + return 0; + } + } +} +static void __Pyx_BufFmt_RaiseExpected(__Pyx_BufFmt_Context* ctx) { + if (ctx->head == NULL || ctx->head->field == &ctx->root) { + const char* expected; + const char* quote; + if (ctx->head == NULL) { + expected = "end"; + quote = ""; + } else { + expected = ctx->head->field->type->name; + quote = "'"; + } + PyErr_Format(PyExc_ValueError, + "Buffer dtype mismatch, expected %s%s%s but got %s", + quote, expected, quote, + __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex)); + } else { + __Pyx_StructField* field = ctx->head->field; + __Pyx_StructField* parent = (ctx->head - 1)->field; + PyErr_Format(PyExc_ValueError, + "Buffer dtype mismatch, expected '%s' but got %s in '%s.%s'", + field->type->name, __Pyx_BufFmt_DescribeTypeChar(ctx->enc_type, ctx->is_complex), + parent->type->name, field->name); + } +} +static int __Pyx_BufFmt_ProcessTypeChunk(__Pyx_BufFmt_Context* ctx) { + char group; + size_t size, offset, arraysize = 1; + if (ctx->enc_type == 0) return 0; + if (ctx->head->field->type->arraysize[0]) { + int i, ndim = 0; + if (ctx->enc_type == 's' || ctx->enc_type == 'p') { + ctx->is_valid_array = ctx->head->field->type->ndim == 1; + ndim = 1; + if (ctx->enc_count != ctx->head->field->type->arraysize[0]) { + PyErr_Format(PyExc_ValueError, + "Expected a dimension of size %zu, got %zu", + ctx->head->field->type->arraysize[0], ctx->enc_count); + return -1; + } + } + if (!ctx->is_valid_array) { + PyErr_Format(PyExc_ValueError, "Expected %d dimensions, got %d", + ctx->head->field->type->ndim, ndim); + return -1; + } + for (i = 0; i < ctx->head->field->type->ndim; i++) { + arraysize *= ctx->head->field->type->arraysize[i]; + } + ctx->is_valid_array = 0; + ctx->enc_count = 1; + } + group = __Pyx_BufFmt_TypeCharToGroup(ctx->enc_type, ctx->is_complex); + do { + __Pyx_StructField* field = ctx->head->field; + __Pyx_TypeInfo* type = field->type; + if (ctx->enc_packmode == '@' || ctx->enc_packmode == '^') { + size = __Pyx_BufFmt_TypeCharToNativeSize(ctx->enc_type, ctx->is_complex); + } else { + size = __Pyx_BufFmt_TypeCharToStandardSize(ctx->enc_type, ctx->is_complex); + } + if (ctx->enc_packmode == '@') { + size_t align_at = __Pyx_BufFmt_TypeCharToAlignment(ctx->enc_type, ctx->is_complex); + size_t align_mod_offset; + if (align_at == 0) return -1; + align_mod_offset = ctx->fmt_offset % align_at; + if (align_mod_offset > 0) ctx->fmt_offset += align_at - align_mod_offset; + if (ctx->struct_alignment == 0) + ctx->struct_alignment = __Pyx_BufFmt_TypeCharToPadding(ctx->enc_type, + ctx->is_complex); + } + if (type->size != size || type->typegroup != group) { + if (type->typegroup == 'C' && type->fields != NULL) { + size_t parent_offset = ctx->head->parent_offset + field->offset; + ++ctx->head; + ctx->head->field = type->fields; + ctx->head->parent_offset = parent_offset; + continue; + } + if ((type->typegroup == 'H' || group == 'H') && type->size == size) { + } else { + __Pyx_BufFmt_RaiseExpected(ctx); + return -1; + } + } + offset = ctx->head->parent_offset + field->offset; + if (ctx->fmt_offset != offset) { + PyErr_Format(PyExc_ValueError, + "Buffer dtype mismatch; next field is at offset %" CYTHON_FORMAT_SSIZE_T "d but %" CYTHON_FORMAT_SSIZE_T "d expected", + (Py_ssize_t)ctx->fmt_offset, (Py_ssize_t)offset); + return -1; + } + ctx->fmt_offset += size; + if (arraysize) + ctx->fmt_offset += (arraysize - 1) * size; + --ctx->enc_count; + while (1) { + if (field == &ctx->root) { + ctx->head = NULL; + if (ctx->enc_count != 0) { + __Pyx_BufFmt_RaiseExpected(ctx); + return -1; + } + break; + } + ctx->head->field = ++field; + if (field->type == NULL) { + --ctx->head; + field = ctx->head->field; + continue; + } else if (field->type->typegroup == 'S') { + size_t parent_offset = ctx->head->parent_offset + field->offset; + if (field->type->fields->type == NULL) continue; + field = field->type->fields; + ++ctx->head; + ctx->head->field = field; + ctx->head->parent_offset = parent_offset; + break; + } else { + break; + } + } + } while (ctx->enc_count); + ctx->enc_type = 0; + ctx->is_complex = 0; + return 0; +} +static int +__pyx_buffmt_parse_array(__Pyx_BufFmt_Context* ctx, const char** tsp) +{ + const char *ts = *tsp; + int i = 0, number, ndim; + ++ts; + if (ctx->new_count != 1) { + PyErr_SetString(PyExc_ValueError, + "Cannot handle repeated arrays in format string"); + return -1; + } + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return -1; + ndim = ctx->head->field->type->ndim; + while (*ts && *ts != ')') { + switch (*ts) { + case ' ': case '\f': case '\r': case '\n': case '\t': case '\v': continue; + default: break; + } + number = __Pyx_BufFmt_ExpectNumber(&ts); + if (number == -1) return -1; + if (i < ndim && (size_t) number != ctx->head->field->type->arraysize[i]) { + PyErr_Format(PyExc_ValueError, + "Expected a dimension of size %zu, got %d", + ctx->head->field->type->arraysize[i], number); + return -1; + } + if (*ts != ',' && *ts != ')') { + PyErr_Format(PyExc_ValueError, + "Expected a comma in format string, got '%c'", *ts); + return -1; + } + if (*ts == ',') ts++; + i++; + } + if (i != ndim) { + PyErr_Format(PyExc_ValueError, "Expected %d dimension(s), got %d", + ctx->head->field->type->ndim, i); + return -1; + } + if (!*ts) { + PyErr_SetString(PyExc_ValueError, + "Unexpected end of format string, expected ')'"); + return -1; + } + ctx->is_valid_array = 1; + ctx->new_count = 1; + *tsp = ++ts; + return 0; +} +static const char* __Pyx_BufFmt_CheckString(__Pyx_BufFmt_Context* ctx, const char* ts) { + int got_Z = 0; + while (1) { + switch(*ts) { + case 0: + if (ctx->enc_type != 0 && ctx->head == NULL) { + __Pyx_BufFmt_RaiseExpected(ctx); + return NULL; + } + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + if (ctx->head != NULL) { + __Pyx_BufFmt_RaiseExpected(ctx); + return NULL; + } + return ts; + case ' ': + case '\r': + case '\n': + ++ts; + break; + case '<': + if (!__Pyx_Is_Little_Endian()) { + PyErr_SetString(PyExc_ValueError, "Little-endian buffer not supported on big-endian compiler"); + return NULL; + } + ctx->new_packmode = '='; + ++ts; + break; + case '>': + case '!': + if (__Pyx_Is_Little_Endian()) { + PyErr_SetString(PyExc_ValueError, "Big-endian buffer not supported on little-endian compiler"); + return NULL; + } + ctx->new_packmode = '='; + ++ts; + break; + case '=': + case '@': + case '^': + ctx->new_packmode = *ts++; + break; + case 'T': + { + const char* ts_after_sub; + size_t i, struct_count = ctx->new_count; + size_t struct_alignment = ctx->struct_alignment; + ctx->new_count = 1; + ++ts; + if (*ts != '{') { + PyErr_SetString(PyExc_ValueError, "Buffer acquisition: Expected '{' after 'T'"); + return NULL; + } + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->enc_type = 0; + ctx->enc_count = 0; + ctx->struct_alignment = 0; + ++ts; + ts_after_sub = ts; + for (i = 0; i != struct_count; ++i) { + ts_after_sub = __Pyx_BufFmt_CheckString(ctx, ts); + if (!ts_after_sub) return NULL; + } + ts = ts_after_sub; + if (struct_alignment) ctx->struct_alignment = struct_alignment; + } + break; + case '}': + { + size_t alignment = ctx->struct_alignment; + ++ts; + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->enc_type = 0; + if (alignment && ctx->fmt_offset % alignment) { + ctx->fmt_offset += alignment - (ctx->fmt_offset % alignment); + } + } + return ts; + case 'x': + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->fmt_offset += ctx->new_count; + ctx->new_count = 1; + ctx->enc_count = 0; + ctx->enc_type = 0; + ctx->enc_packmode = ctx->new_packmode; + ++ts; + break; + case 'Z': + got_Z = 1; + ++ts; + if (*ts != 'f' && *ts != 'd' && *ts != 'g') { + __Pyx_BufFmt_RaiseUnexpectedChar('Z'); + return NULL; + } + CYTHON_FALLTHROUGH; + case '?': case 'c': case 'b': case 'B': case 'h': case 'H': case 'i': case 'I': + case 'l': case 'L': case 'q': case 'Q': + case 'f': case 'd': case 'g': + case 'O': case 'p': + if ((ctx->enc_type == *ts) && (got_Z == ctx->is_complex) && + (ctx->enc_packmode == ctx->new_packmode) && (!ctx->is_valid_array)) { + ctx->enc_count += ctx->new_count; + ctx->new_count = 1; + got_Z = 0; + ++ts; + break; + } + CYTHON_FALLTHROUGH; + case 's': + if (__Pyx_BufFmt_ProcessTypeChunk(ctx) == -1) return NULL; + ctx->enc_count = ctx->new_count; + ctx->enc_packmode = ctx->new_packmode; + ctx->enc_type = *ts; + ctx->is_complex = got_Z; + ++ts; + ctx->new_count = 1; + got_Z = 0; + break; + case ':': + ++ts; + while(*ts != ':') ++ts; + ++ts; + break; + case '(': + if (__pyx_buffmt_parse_array(ctx, &ts) < 0) return NULL; + break; + default: + { + int number = __Pyx_BufFmt_ExpectNumber(&ts); + if (number == -1) return NULL; + ctx->new_count = (size_t)number; + } + } + } +} + +/* BufferGetAndValidate */ + static CYTHON_INLINE void __Pyx_SafeReleaseBuffer(Py_buffer* info) { + if (unlikely(info->buf == NULL)) return; + if (info->suboffsets == __Pyx_minusones) info->suboffsets = NULL; + __Pyx_ReleaseBuffer(info); +} +static void __Pyx_ZeroBuffer(Py_buffer* buf) { + buf->buf = NULL; + buf->obj = NULL; + buf->strides = __Pyx_zeros; + buf->shape = __Pyx_zeros; + buf->suboffsets = __Pyx_minusones; +} +static int __Pyx__GetBufferAndValidate( + Py_buffer* buf, PyObject* obj, __Pyx_TypeInfo* dtype, int flags, + int nd, int cast, __Pyx_BufFmt_StackElem* stack) +{ + buf->buf = NULL; + if (unlikely(__Pyx_GetBuffer(obj, buf, flags) == -1)) { + __Pyx_ZeroBuffer(buf); + return -1; + } + if (unlikely(buf->ndim != nd)) { + PyErr_Format(PyExc_ValueError, + "Buffer has wrong number of dimensions (expected %d, got %d)", + nd, buf->ndim); + goto fail; + } + if (!cast) { + __Pyx_BufFmt_Context ctx; + __Pyx_BufFmt_Init(&ctx, stack, dtype); + if (!__Pyx_BufFmt_CheckString(&ctx, buf->format)) goto fail; + } + if (unlikely((size_t)buf->itemsize != dtype->size)) { + PyErr_Format(PyExc_ValueError, + "Item size of buffer (%" CYTHON_FORMAT_SSIZE_T "d byte%s) does not match size of '%s' (%" CYTHON_FORMAT_SSIZE_T "d byte%s)", + buf->itemsize, (buf->itemsize > 1) ? "s" : "", + dtype->name, (Py_ssize_t)dtype->size, (dtype->size > 1) ? "s" : ""); + goto fail; + } + if (buf->suboffsets == NULL) buf->suboffsets = __Pyx_minusones; + return 0; +fail:; + __Pyx_SafeReleaseBuffer(buf); + return -1; +} + +/* BufferFallbackError */ + static void __Pyx_RaiseBufferFallbackError(void) { + PyErr_SetString(PyExc_ValueError, + "Buffer acquisition failed on assignment; and then reacquiring the old buffer failed too!"); +} + +/* PyIntBinop */ + #if !CYTHON_COMPILING_IN_PYPY +static PyObject* __Pyx_PyInt_SubtractObjC(PyObject *op1, PyObject *op2, long intval, int inplace, int zerodivision_check) { + CYTHON_MAYBE_UNUSED_VAR(intval); + CYTHON_MAYBE_UNUSED_VAR(inplace); + CYTHON_UNUSED_VAR(zerodivision_check); + #if PY_MAJOR_VERSION < 3 + if (likely(PyInt_CheckExact(op1))) { + const long b = intval; + long x; + long a = PyInt_AS_LONG(op1); + + x = (long)((unsigned long)a - (unsigned long)b); + if (likely((x^a) >= 0 || (x^~b) >= 0)) + return PyInt_FromLong(x); + return PyLong_Type.tp_as_number->nb_subtract(op1, op2); + } + #endif + #if CYTHON_USE_PYLONG_INTERNALS + if (likely(PyLong_CheckExact(op1))) { + const long b = intval; + long a, x; +#ifdef HAVE_LONG_LONG + const PY_LONG_LONG llb = intval; + PY_LONG_LONG lla, llx; +#endif + if (unlikely(__Pyx_PyLong_IsZero(op1))) { + return PyLong_FromLong(-intval); + } + if (likely(__Pyx_PyLong_IsCompact(op1))) { + a = __Pyx_PyLong_CompactValue(op1); + } else { + const digit* digits = __Pyx_PyLong_Digits(op1); + const Py_ssize_t size = __Pyx_PyLong_SignedDigitCount(op1); + switch (size) { + case -2: + if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) { + a = -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])); + break; + #ifdef HAVE_LONG_LONG + } else if (8 * sizeof(PY_LONG_LONG) - 1 > 2 * PyLong_SHIFT) { + lla = -(PY_LONG_LONG) (((((unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0])); + goto long_long; + #endif + } + CYTHON_FALLTHROUGH; + case 2: + if (8 * sizeof(long) - 1 > 2 * PyLong_SHIFT) { + a = (long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])); + break; + #ifdef HAVE_LONG_LONG + } else if (8 * sizeof(PY_LONG_LONG) - 1 > 2 * PyLong_SHIFT) { + lla = (PY_LONG_LONG) (((((unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0])); + goto long_long; + #endif + } + CYTHON_FALLTHROUGH; + case -3: + if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) { + a = -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])); + break; + #ifdef HAVE_LONG_LONG + } else if (8 * sizeof(PY_LONG_LONG) - 1 > 3 * PyLong_SHIFT) { + lla = -(PY_LONG_LONG) (((((((unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0])); + goto long_long; + #endif + } + CYTHON_FALLTHROUGH; + case 3: + if (8 * sizeof(long) - 1 > 3 * PyLong_SHIFT) { + a = (long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])); + break; + #ifdef HAVE_LONG_LONG + } else if (8 * sizeof(PY_LONG_LONG) - 1 > 3 * PyLong_SHIFT) { + lla = (PY_LONG_LONG) (((((((unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0])); + goto long_long; + #endif + } + CYTHON_FALLTHROUGH; + case -4: + if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) { + a = -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])); + break; + #ifdef HAVE_LONG_LONG + } else if (8 * sizeof(PY_LONG_LONG) - 1 > 4 * PyLong_SHIFT) { + lla = -(PY_LONG_LONG) (((((((((unsigned PY_LONG_LONG)digits[3]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0])); + goto long_long; + #endif + } + CYTHON_FALLTHROUGH; + case 4: + if (8 * sizeof(long) - 1 > 4 * PyLong_SHIFT) { + a = (long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0])); + break; + #ifdef HAVE_LONG_LONG + } else if (8 * sizeof(PY_LONG_LONG) - 1 > 4 * PyLong_SHIFT) { + lla = (PY_LONG_LONG) (((((((((unsigned PY_LONG_LONG)digits[3]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[2]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[1]) << PyLong_SHIFT) | (unsigned PY_LONG_LONG)digits[0])); + goto long_long; + #endif + } + CYTHON_FALLTHROUGH; + default: return PyLong_Type.tp_as_number->nb_subtract(op1, op2); + } + } + x = a - b; + return PyLong_FromLong(x); +#ifdef HAVE_LONG_LONG + long_long: + llx = lla - llb; + return PyLong_FromLongLong(llx); +#endif + + + } + #endif + if (PyFloat_CheckExact(op1)) { + const long b = intval; +#if CYTHON_COMPILING_IN_LIMITED_API + double a = __pyx_PyFloat_AsDouble(op1); +#else + double a = PyFloat_AS_DOUBLE(op1); +#endif + double result; + + PyFPE_START_PROTECT("subtract", return NULL) + result = ((double)a) - (double)b; + PyFPE_END_PROTECT(result) + return PyFloat_FromDouble(result); + } + return (inplace ? PyNumber_InPlaceSubtract : PyNumber_Subtract)(op1, op2); +} +#endif + +/* SliceObject */ + static CYTHON_INLINE PyObject* __Pyx_PyObject_GetSlice(PyObject* obj, + Py_ssize_t cstart, Py_ssize_t cstop, + PyObject** _py_start, PyObject** _py_stop, PyObject** _py_slice, + int has_cstart, int has_cstop, int wraparound) { + __Pyx_TypeName obj_type_name; +#if CYTHON_USE_TYPE_SLOTS + PyMappingMethods* mp; +#if PY_MAJOR_VERSION < 3 + PySequenceMethods* ms = Py_TYPE(obj)->tp_as_sequence; + if (likely(ms && ms->sq_slice)) { + if (!has_cstart) { + if (_py_start && (*_py_start != Py_None)) { + cstart = __Pyx_PyIndex_AsSsize_t(*_py_start); + if ((cstart == (Py_ssize_t)-1) && PyErr_Occurred()) goto bad; + } else + cstart = 0; + } + if (!has_cstop) { + if (_py_stop && (*_py_stop != Py_None)) { + cstop = __Pyx_PyIndex_AsSsize_t(*_py_stop); + if ((cstop == (Py_ssize_t)-1) && PyErr_Occurred()) goto bad; + } else + cstop = PY_SSIZE_T_MAX; + } + if (wraparound && unlikely((cstart < 0) | (cstop < 0)) && likely(ms->sq_length)) { + Py_ssize_t l = ms->sq_length(obj); + if (likely(l >= 0)) { + if (cstop < 0) { + cstop += l; + if (cstop < 0) cstop = 0; + } + if (cstart < 0) { + cstart += l; + if (cstart < 0) cstart = 0; + } + } else { + if (!PyErr_ExceptionMatches(PyExc_OverflowError)) + goto bad; + PyErr_Clear(); + } + } + return ms->sq_slice(obj, cstart, cstop); + } +#else + CYTHON_UNUSED_VAR(wraparound); +#endif + mp = Py_TYPE(obj)->tp_as_mapping; + if (likely(mp && mp->mp_subscript)) +#else + CYTHON_UNUSED_VAR(wraparound); +#endif + { + PyObject* result; + PyObject *py_slice, *py_start, *py_stop; + if (_py_slice) { + py_slice = *_py_slice; + } else { + PyObject* owned_start = NULL; + PyObject* owned_stop = NULL; + if (_py_start) { + py_start = *_py_start; + } else { + if (has_cstart) { + owned_start = py_start = PyInt_FromSsize_t(cstart); + if (unlikely(!py_start)) goto bad; + } else + py_start = Py_None; + } + if (_py_stop) { + py_stop = *_py_stop; + } else { + if (has_cstop) { + owned_stop = py_stop = PyInt_FromSsize_t(cstop); + if (unlikely(!py_stop)) { + Py_XDECREF(owned_start); + goto bad; + } + } else + py_stop = Py_None; + } + py_slice = PySlice_New(py_start, py_stop, Py_None); + Py_XDECREF(owned_start); + Py_XDECREF(owned_stop); + if (unlikely(!py_slice)) goto bad; + } +#if CYTHON_USE_TYPE_SLOTS + result = mp->mp_subscript(obj, py_slice); +#else + result = PyObject_GetItem(obj, py_slice); +#endif + if (!_py_slice) { + Py_DECREF(py_slice); + } + return result; + } + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, + "'" __Pyx_FMT_TYPENAME "' object is unsliceable", obj_type_name); + __Pyx_DECREF_TypeName(obj_type_name); +bad: + return NULL; +} + +/* PyObject_GenericGetAttrNoDict */ + #if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static PyObject *__Pyx_RaiseGenericGetAttributeError(PyTypeObject *tp, PyObject *attr_name) { + __Pyx_TypeName type_name = __Pyx_PyType_GetName(tp); + PyErr_Format(PyExc_AttributeError, +#if PY_MAJOR_VERSION >= 3 + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%U'", + type_name, attr_name); +#else + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%.400s'", + type_name, PyString_AS_STRING(attr_name)); +#endif + __Pyx_DECREF_TypeName(type_name); + return NULL; +} +static CYTHON_INLINE PyObject* __Pyx_PyObject_GenericGetAttrNoDict(PyObject* obj, PyObject* attr_name) { + PyObject *descr; + PyTypeObject *tp = Py_TYPE(obj); + if (unlikely(!PyString_Check(attr_name))) { + return PyObject_GenericGetAttr(obj, attr_name); + } + assert(!tp->tp_dictoffset); + descr = _PyType_Lookup(tp, attr_name); + if (unlikely(!descr)) { + return __Pyx_RaiseGenericGetAttributeError(tp, attr_name); + } + Py_INCREF(descr); + #if PY_MAJOR_VERSION < 3 + if (likely(PyType_HasFeature(Py_TYPE(descr), Py_TPFLAGS_HAVE_CLASS))) + #endif + { + descrgetfunc f = Py_TYPE(descr)->tp_descr_get; + if (unlikely(f)) { + PyObject *res = f(descr, obj, (PyObject *)tp); + Py_DECREF(descr); + return res; + } + } + return descr; +} +#endif + +/* PyObject_GenericGetAttr */ + #if CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP && PY_VERSION_HEX < 0x03070000 +static PyObject* __Pyx_PyObject_GenericGetAttr(PyObject* obj, PyObject* attr_name) { + if (unlikely(Py_TYPE(obj)->tp_dictoffset)) { + return PyObject_GenericGetAttr(obj, attr_name); + } + return __Pyx_PyObject_GenericGetAttrNoDict(obj, attr_name); +} +#endif + +/* FixUpExtensionType */ + #if CYTHON_USE_TYPE_SPECS +static int __Pyx_fix_up_extension_type_from_spec(PyType_Spec *spec, PyTypeObject *type) { +#if PY_VERSION_HEX > 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + CYTHON_UNUSED_VAR(spec); + CYTHON_UNUSED_VAR(type); +#else + const PyType_Slot *slot = spec->slots; + while (slot && slot->slot && slot->slot != Py_tp_members) + slot++; + if (slot && slot->slot == Py_tp_members) { + int changed = 0; +#if !(PY_VERSION_HEX <= 0x030900b1 && CYTHON_COMPILING_IN_CPYTHON) + const +#endif + PyMemberDef *memb = (PyMemberDef*) slot->pfunc; + while (memb && memb->name) { + if (memb->name[0] == '_' && memb->name[1] == '_') { +#if PY_VERSION_HEX < 0x030900b1 + if (strcmp(memb->name, "__weaklistoffset__") == 0) { + assert(memb->type == T_PYSSIZET); + assert(memb->flags == READONLY); + type->tp_weaklistoffset = memb->offset; + changed = 1; + } + else if (strcmp(memb->name, "__dictoffset__") == 0) { + assert(memb->type == T_PYSSIZET); + assert(memb->flags == READONLY); + type->tp_dictoffset = memb->offset; + changed = 1; + } +#if CYTHON_METH_FASTCALL + else if (strcmp(memb->name, "__vectorcalloffset__") == 0) { + assert(memb->type == T_PYSSIZET); + assert(memb->flags == READONLY); +#if PY_VERSION_HEX >= 0x030800b4 + type->tp_vectorcall_offset = memb->offset; +#else + type->tp_print = (printfunc) memb->offset; +#endif + changed = 1; + } +#endif +#else + if ((0)); +#endif +#if PY_VERSION_HEX <= 0x030900b1 && CYTHON_COMPILING_IN_CPYTHON + else if (strcmp(memb->name, "__module__") == 0) { + PyObject *descr; + assert(memb->type == T_OBJECT); + assert(memb->flags == 0 || memb->flags == READONLY); + descr = PyDescr_NewMember(type, memb); + if (unlikely(!descr)) + return -1; + if (unlikely(PyDict_SetItem(type->tp_dict, PyDescr_NAME(descr), descr) < 0)) { + Py_DECREF(descr); + return -1; + } + Py_DECREF(descr); + changed = 1; + } +#endif + } + memb++; + } + if (changed) + PyType_Modified(type); + } +#endif + return 0; +} +#endif + +/* PyObjectCallNoArg */ + static CYTHON_INLINE PyObject* __Pyx_PyObject_CallNoArg(PyObject *func) { + PyObject *arg[2] = {NULL, NULL}; + return __Pyx_PyObject_FastCall(func, arg + 1, 0 | __Pyx_PY_VECTORCALL_ARGUMENTS_OFFSET); +} + +/* PyObjectGetMethod */ + static int __Pyx_PyObject_GetMethod(PyObject *obj, PyObject *name, PyObject **method) { + PyObject *attr; +#if CYTHON_UNPACK_METHODS && CYTHON_COMPILING_IN_CPYTHON && CYTHON_USE_PYTYPE_LOOKUP + __Pyx_TypeName type_name; + PyTypeObject *tp = Py_TYPE(obj); + PyObject *descr; + descrgetfunc f = NULL; + PyObject **dictptr, *dict; + int meth_found = 0; + assert (*method == NULL); + if (unlikely(tp->tp_getattro != PyObject_GenericGetAttr)) { + attr = __Pyx_PyObject_GetAttrStr(obj, name); + goto try_unpack; + } + if (unlikely(tp->tp_dict == NULL) && unlikely(PyType_Ready(tp) < 0)) { + return 0; + } + descr = _PyType_Lookup(tp, name); + if (likely(descr != NULL)) { + Py_INCREF(descr); +#if defined(Py_TPFLAGS_METHOD_DESCRIPTOR) && Py_TPFLAGS_METHOD_DESCRIPTOR + if (__Pyx_PyType_HasFeature(Py_TYPE(descr), Py_TPFLAGS_METHOD_DESCRIPTOR)) +#elif PY_MAJOR_VERSION >= 3 + #ifdef __Pyx_CyFunction_USED + if (likely(PyFunction_Check(descr) || __Pyx_IS_TYPE(descr, &PyMethodDescr_Type) || __Pyx_CyFunction_Check(descr))) + #else + if (likely(PyFunction_Check(descr) || __Pyx_IS_TYPE(descr, &PyMethodDescr_Type))) + #endif +#else + #ifdef __Pyx_CyFunction_USED + if (likely(PyFunction_Check(descr) || __Pyx_CyFunction_Check(descr))) + #else + if (likely(PyFunction_Check(descr))) + #endif +#endif + { + meth_found = 1; + } else { + f = Py_TYPE(descr)->tp_descr_get; + if (f != NULL && PyDescr_IsData(descr)) { + attr = f(descr, obj, (PyObject *)Py_TYPE(obj)); + Py_DECREF(descr); + goto try_unpack; + } + } + } + dictptr = _PyObject_GetDictPtr(obj); + if (dictptr != NULL && (dict = *dictptr) != NULL) { + Py_INCREF(dict); + attr = __Pyx_PyDict_GetItemStr(dict, name); + if (attr != NULL) { + Py_INCREF(attr); + Py_DECREF(dict); + Py_XDECREF(descr); + goto try_unpack; + } + Py_DECREF(dict); + } + if (meth_found) { + *method = descr; + return 1; + } + if (f != NULL) { + attr = f(descr, obj, (PyObject *)Py_TYPE(obj)); + Py_DECREF(descr); + goto try_unpack; + } + if (likely(descr != NULL)) { + *method = descr; + return 0; + } + type_name = __Pyx_PyType_GetName(tp); + PyErr_Format(PyExc_AttributeError, +#if PY_MAJOR_VERSION >= 3 + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%U'", + type_name, name); +#else + "'" __Pyx_FMT_TYPENAME "' object has no attribute '%.400s'", + type_name, PyString_AS_STRING(name)); +#endif + __Pyx_DECREF_TypeName(type_name); + return 0; +#else + attr = __Pyx_PyObject_GetAttrStr(obj, name); + goto try_unpack; +#endif +try_unpack: +#if CYTHON_UNPACK_METHODS + if (likely(attr) && PyMethod_Check(attr) && likely(PyMethod_GET_SELF(attr) == obj)) { + PyObject *function = PyMethod_GET_FUNCTION(attr); + Py_INCREF(function); + Py_DECREF(attr); + *method = function; + return 1; + } +#endif + *method = attr; + return 0; +} + +/* PyObjectCallMethod0 */ + static PyObject* __Pyx_PyObject_CallMethod0(PyObject* obj, PyObject* method_name) { + PyObject *method = NULL, *result = NULL; + int is_method = __Pyx_PyObject_GetMethod(obj, method_name, &method); + if (likely(is_method)) { + result = __Pyx_PyObject_CallOneArg(method, obj); + Py_DECREF(method); + return result; + } + if (unlikely(!method)) goto bad; + result = __Pyx_PyObject_CallNoArg(method); + Py_DECREF(method); +bad: + return result; +} + +/* ValidateBasesTuple */ + #if CYTHON_COMPILING_IN_CPYTHON || CYTHON_COMPILING_IN_LIMITED_API || CYTHON_USE_TYPE_SPECS +static int __Pyx_validate_bases_tuple(const char *type_name, Py_ssize_t dictoffset, PyObject *bases) { + Py_ssize_t i, n; +#if CYTHON_ASSUME_SAFE_MACROS + n = PyTuple_GET_SIZE(bases); +#else + n = PyTuple_Size(bases); + if (n < 0) return -1; +#endif + for (i = 1; i < n; i++) + { +#if CYTHON_AVOID_BORROWED_REFS + PyObject *b0 = PySequence_GetItem(bases, i); + if (!b0) return -1; +#elif CYTHON_ASSUME_SAFE_MACROS + PyObject *b0 = PyTuple_GET_ITEM(bases, i); +#else + PyObject *b0 = PyTuple_GetItem(bases, i); + if (!b0) return -1; +#endif + PyTypeObject *b; +#if PY_MAJOR_VERSION < 3 + if (PyClass_Check(b0)) + { + PyErr_Format(PyExc_TypeError, "base class '%.200s' is an old-style class", + PyString_AS_STRING(((PyClassObject*)b0)->cl_name)); +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + return -1; + } +#endif + b = (PyTypeObject*) b0; + if (!__Pyx_PyType_HasFeature(b, Py_TPFLAGS_HEAPTYPE)) + { + __Pyx_TypeName b_name = __Pyx_PyType_GetName(b); + PyErr_Format(PyExc_TypeError, + "base class '" __Pyx_FMT_TYPENAME "' is not a heap type", b_name); + __Pyx_DECREF_TypeName(b_name); +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + return -1; + } + if (dictoffset == 0) + { + Py_ssize_t b_dictoffset = 0; +#if CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY + b_dictoffset = b->tp_dictoffset; +#else + PyObject *py_b_dictoffset = PyObject_GetAttrString((PyObject*)b, "__dictoffset__"); + if (!py_b_dictoffset) goto dictoffset_return; + b_dictoffset = PyLong_AsSsize_t(py_b_dictoffset); + Py_DECREF(py_b_dictoffset); + if (b_dictoffset == -1 && PyErr_Occurred()) goto dictoffset_return; +#endif + if (b_dictoffset) { + { + __Pyx_TypeName b_name = __Pyx_PyType_GetName(b); + PyErr_Format(PyExc_TypeError, + "extension type '%.200s' has no __dict__ slot, " + "but base type '" __Pyx_FMT_TYPENAME "' has: " + "either add 'cdef dict __dict__' to the extension type " + "or add '__slots__ = [...]' to the base type", + type_name, b_name); + __Pyx_DECREF_TypeName(b_name); + } +#if !(CYTHON_USE_TYPE_SLOTS || CYTHON_COMPILING_IN_PYPY) + dictoffset_return: +#endif +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + return -1; + } + } +#if CYTHON_AVOID_BORROWED_REFS + Py_DECREF(b0); +#endif + } + return 0; +} +#endif + +/* PyType_Ready */ + static int __Pyx_PyType_Ready(PyTypeObject *t) { +#if CYTHON_USE_TYPE_SPECS || !(CYTHON_COMPILING_IN_CPYTHON || CYTHON_COMPILING_IN_LIMITED_API) || defined(PYSTON_MAJOR_VERSION) + (void)__Pyx_PyObject_CallMethod0; +#if CYTHON_USE_TYPE_SPECS + (void)__Pyx_validate_bases_tuple; +#endif + return PyType_Ready(t); +#else + int r; + PyObject *bases = __Pyx_PyType_GetSlot(t, tp_bases, PyObject*); + if (bases && unlikely(__Pyx_validate_bases_tuple(t->tp_name, t->tp_dictoffset, bases) == -1)) + return -1; +#if PY_VERSION_HEX >= 0x03050000 && !defined(PYSTON_MAJOR_VERSION) + { + int gc_was_enabled; + #if PY_VERSION_HEX >= 0x030A00b1 + gc_was_enabled = PyGC_Disable(); + (void)__Pyx_PyObject_CallMethod0; + #else + PyObject *ret, *py_status; + PyObject *gc = NULL; + #if PY_VERSION_HEX >= 0x030700a1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM+0 >= 0x07030400) + gc = PyImport_GetModule(__pyx_kp_u_gc); + #endif + if (unlikely(!gc)) gc = PyImport_Import(__pyx_kp_u_gc); + if (unlikely(!gc)) return -1; + py_status = __Pyx_PyObject_CallMethod0(gc, __pyx_kp_u_isenabled); + if (unlikely(!py_status)) { + Py_DECREF(gc); + return -1; + } + gc_was_enabled = __Pyx_PyObject_IsTrue(py_status); + Py_DECREF(py_status); + if (gc_was_enabled > 0) { + ret = __Pyx_PyObject_CallMethod0(gc, __pyx_kp_u_disable); + if (unlikely(!ret)) { + Py_DECREF(gc); + return -1; + } + Py_DECREF(ret); + } else if (unlikely(gc_was_enabled == -1)) { + Py_DECREF(gc); + return -1; + } + #endif + t->tp_flags |= Py_TPFLAGS_HEAPTYPE; +#if PY_VERSION_HEX >= 0x030A0000 + t->tp_flags |= Py_TPFLAGS_IMMUTABLETYPE; +#endif +#else + (void)__Pyx_PyObject_CallMethod0; +#endif + r = PyType_Ready(t); +#if PY_VERSION_HEX >= 0x03050000 && !defined(PYSTON_MAJOR_VERSION) + t->tp_flags &= ~Py_TPFLAGS_HEAPTYPE; + #if PY_VERSION_HEX >= 0x030A00b1 + if (gc_was_enabled) + PyGC_Enable(); + #else + if (gc_was_enabled) { + PyObject *tp, *v, *tb; + PyErr_Fetch(&tp, &v, &tb); + ret = __Pyx_PyObject_CallMethod0(gc, __pyx_kp_u_enable); + if (likely(ret || r == -1)) { + Py_XDECREF(ret); + PyErr_Restore(tp, v, tb); + } else { + Py_XDECREF(tp); + Py_XDECREF(v); + Py_XDECREF(tb); + r = -1; + } + } + Py_DECREF(gc); + #endif + } +#endif + return r; +#endif +} + +/* SetVTable */ + static int __Pyx_SetVtable(PyTypeObject *type, void *vtable) { + PyObject *ob = PyCapsule_New(vtable, 0, 0); + if (unlikely(!ob)) + goto bad; +#if CYTHON_COMPILING_IN_LIMITED_API + if (unlikely(PyObject_SetAttr((PyObject *) type, __pyx_n_s_pyx_vtable, ob) < 0)) +#else + if (unlikely(PyDict_SetItem(type->tp_dict, __pyx_n_s_pyx_vtable, ob) < 0)) +#endif + goto bad; + Py_DECREF(ob); + return 0; +bad: + Py_XDECREF(ob); + return -1; +} + +/* GetVTable */ + static void* __Pyx_GetVtable(PyTypeObject *type) { + void* ptr; +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *ob = PyObject_GetAttr((PyObject *)type, __pyx_n_s_pyx_vtable); +#else + PyObject *ob = PyObject_GetItem(type->tp_dict, __pyx_n_s_pyx_vtable); +#endif + if (!ob) + goto bad; + ptr = PyCapsule_GetPointer(ob, 0); + if (!ptr && !PyErr_Occurred()) + PyErr_SetString(PyExc_RuntimeError, "invalid vtable found for imported type"); + Py_DECREF(ob); + return ptr; +bad: + Py_XDECREF(ob); + return NULL; +} + +/* MergeVTables */ + #if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_MergeVtables(PyTypeObject *type) { + int i; + void** base_vtables; + __Pyx_TypeName tp_base_name; + __Pyx_TypeName base_name; + void* unknown = (void*)-1; + PyObject* bases = type->tp_bases; + int base_depth = 0; + { + PyTypeObject* base = type->tp_base; + while (base) { + base_depth += 1; + base = base->tp_base; + } + } + base_vtables = (void**) malloc(sizeof(void*) * (size_t)(base_depth + 1)); + base_vtables[0] = unknown; + for (i = 1; i < PyTuple_GET_SIZE(bases); i++) { + void* base_vtable = __Pyx_GetVtable(((PyTypeObject*)PyTuple_GET_ITEM(bases, i))); + if (base_vtable != NULL) { + int j; + PyTypeObject* base = type->tp_base; + for (j = 0; j < base_depth; j++) { + if (base_vtables[j] == unknown) { + base_vtables[j] = __Pyx_GetVtable(base); + base_vtables[j + 1] = unknown; + } + if (base_vtables[j] == base_vtable) { + break; + } else if (base_vtables[j] == NULL) { + goto bad; + } + base = base->tp_base; + } + } + } + PyErr_Clear(); + free(base_vtables); + return 0; +bad: + tp_base_name = __Pyx_PyType_GetName(type->tp_base); + base_name = __Pyx_PyType_GetName((PyTypeObject*)PyTuple_GET_ITEM(bases, i)); + PyErr_Format(PyExc_TypeError, + "multiple bases have vtable conflict: '" __Pyx_FMT_TYPENAME "' and '" __Pyx_FMT_TYPENAME "'", tp_base_name, base_name); + __Pyx_DECREF_TypeName(tp_base_name); + __Pyx_DECREF_TypeName(base_name); + free(base_vtables); + return -1; +} +#endif + +/* SetupReduce */ + #if !CYTHON_COMPILING_IN_LIMITED_API +static int __Pyx_setup_reduce_is_named(PyObject* meth, PyObject* name) { + int ret; + PyObject *name_attr; + name_attr = __Pyx_PyObject_GetAttrStrNoError(meth, __pyx_n_s_name_2); + if (likely(name_attr)) { + ret = PyObject_RichCompareBool(name_attr, name, Py_EQ); + } else { + ret = -1; + } + if (unlikely(ret < 0)) { + PyErr_Clear(); + ret = 0; + } + Py_XDECREF(name_attr); + return ret; +} +static int __Pyx_setup_reduce(PyObject* type_obj) { + int ret = 0; + PyObject *object_reduce = NULL; + PyObject *object_getstate = NULL; + PyObject *object_reduce_ex = NULL; + PyObject *reduce = NULL; + PyObject *reduce_ex = NULL; + PyObject *reduce_cython = NULL; + PyObject *setstate = NULL; + PyObject *setstate_cython = NULL; + PyObject *getstate = NULL; +#if CYTHON_USE_PYTYPE_LOOKUP + getstate = _PyType_Lookup((PyTypeObject*)type_obj, __pyx_n_s_getstate); +#else + getstate = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_getstate); + if (!getstate && PyErr_Occurred()) { + goto __PYX_BAD; + } +#endif + if (getstate) { +#if CYTHON_USE_PYTYPE_LOOKUP + object_getstate = _PyType_Lookup(&PyBaseObject_Type, __pyx_n_s_getstate); +#else + object_getstate = __Pyx_PyObject_GetAttrStrNoError((PyObject*)&PyBaseObject_Type, __pyx_n_s_getstate); + if (!object_getstate && PyErr_Occurred()) { + goto __PYX_BAD; + } +#endif + if (object_getstate != getstate) { + goto __PYX_GOOD; + } + } +#if CYTHON_USE_PYTYPE_LOOKUP + object_reduce_ex = _PyType_Lookup(&PyBaseObject_Type, __pyx_n_s_reduce_ex); if (!object_reduce_ex) goto __PYX_BAD; +#else + object_reduce_ex = __Pyx_PyObject_GetAttrStr((PyObject*)&PyBaseObject_Type, __pyx_n_s_reduce_ex); if (!object_reduce_ex) goto __PYX_BAD; +#endif + reduce_ex = __Pyx_PyObject_GetAttrStr(type_obj, __pyx_n_s_reduce_ex); if (unlikely(!reduce_ex)) goto __PYX_BAD; + if (reduce_ex == object_reduce_ex) { +#if CYTHON_USE_PYTYPE_LOOKUP + object_reduce = _PyType_Lookup(&PyBaseObject_Type, __pyx_n_s_reduce); if (!object_reduce) goto __PYX_BAD; +#else + object_reduce = __Pyx_PyObject_GetAttrStr((PyObject*)&PyBaseObject_Type, __pyx_n_s_reduce); if (!object_reduce) goto __PYX_BAD; +#endif + reduce = __Pyx_PyObject_GetAttrStr(type_obj, __pyx_n_s_reduce); if (unlikely(!reduce)) goto __PYX_BAD; + if (reduce == object_reduce || __Pyx_setup_reduce_is_named(reduce, __pyx_n_s_reduce_cython)) { + reduce_cython = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_reduce_cython); + if (likely(reduce_cython)) { + ret = PyDict_SetItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_reduce, reduce_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + ret = PyDict_DelItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_reduce_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + } else if (reduce == object_reduce || PyErr_Occurred()) { + goto __PYX_BAD; + } + setstate = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_setstate); + if (!setstate) PyErr_Clear(); + if (!setstate || __Pyx_setup_reduce_is_named(setstate, __pyx_n_s_setstate_cython)) { + setstate_cython = __Pyx_PyObject_GetAttrStrNoError(type_obj, __pyx_n_s_setstate_cython); + if (likely(setstate_cython)) { + ret = PyDict_SetItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_setstate, setstate_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + ret = PyDict_DelItem(((PyTypeObject*)type_obj)->tp_dict, __pyx_n_s_setstate_cython); if (unlikely(ret < 0)) goto __PYX_BAD; + } else if (!setstate || PyErr_Occurred()) { + goto __PYX_BAD; + } + } + PyType_Modified((PyTypeObject*)type_obj); + } + } + goto __PYX_GOOD; +__PYX_BAD: + if (!PyErr_Occurred()) { + __Pyx_TypeName type_obj_name = + __Pyx_PyType_GetName((PyTypeObject*)type_obj); + PyErr_Format(PyExc_RuntimeError, + "Unable to initialize pickling for " __Pyx_FMT_TYPENAME, type_obj_name); + __Pyx_DECREF_TypeName(type_obj_name); + } + ret = -1; +__PYX_GOOD: +#if !CYTHON_USE_PYTYPE_LOOKUP + Py_XDECREF(object_reduce); + Py_XDECREF(object_reduce_ex); + Py_XDECREF(object_getstate); + Py_XDECREF(getstate); +#endif + Py_XDECREF(reduce); + Py_XDECREF(reduce_ex); + Py_XDECREF(reduce_cython); + Py_XDECREF(setstate); + Py_XDECREF(setstate_cython); + return ret; +} +#endif + +/* TypeImport */ + #ifndef __PYX_HAVE_RT_ImportType_3_0_12 +#define __PYX_HAVE_RT_ImportType_3_0_12 +static PyTypeObject *__Pyx_ImportType_3_0_12(PyObject *module, const char *module_name, const char *class_name, + size_t size, size_t alignment, enum __Pyx_ImportType_CheckSize_3_0_12 check_size) +{ + PyObject *result = 0; + char warning[200]; + Py_ssize_t basicsize; + Py_ssize_t itemsize; +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *py_basicsize; + PyObject *py_itemsize; +#endif + result = PyObject_GetAttrString(module, class_name); + if (!result) + goto bad; + if (!PyType_Check(result)) { + PyErr_Format(PyExc_TypeError, + "%.200s.%.200s is not a type object", + module_name, class_name); + goto bad; + } +#if !CYTHON_COMPILING_IN_LIMITED_API + basicsize = ((PyTypeObject *)result)->tp_basicsize; + itemsize = ((PyTypeObject *)result)->tp_itemsize; +#else + py_basicsize = PyObject_GetAttrString(result, "__basicsize__"); + if (!py_basicsize) + goto bad; + basicsize = PyLong_AsSsize_t(py_basicsize); + Py_DECREF(py_basicsize); + py_basicsize = 0; + if (basicsize == (Py_ssize_t)-1 && PyErr_Occurred()) + goto bad; + py_itemsize = PyObject_GetAttrString(result, "__itemsize__"); + if (!py_itemsize) + goto bad; + itemsize = PyLong_AsSsize_t(py_itemsize); + Py_DECREF(py_itemsize); + py_itemsize = 0; + if (itemsize == (Py_ssize_t)-1 && PyErr_Occurred()) + goto bad; +#endif + if (itemsize) { + if (size % alignment) { + alignment = size % alignment; + } + if (itemsize < (Py_ssize_t)alignment) + itemsize = (Py_ssize_t)alignment; + } + if ((size_t)(basicsize + itemsize) < size) { + PyErr_Format(PyExc_ValueError, + "%.200s.%.200s size changed, may indicate binary incompatibility. " + "Expected %zd from C header, got %zd from PyObject", + module_name, class_name, size, basicsize+itemsize); + goto bad; + } + if (check_size == __Pyx_ImportType_CheckSize_Error_3_0_12 && + ((size_t)basicsize > size || (size_t)(basicsize + itemsize) < size)) { + PyErr_Format(PyExc_ValueError, + "%.200s.%.200s size changed, may indicate binary incompatibility. " + "Expected %zd from C header, got %zd-%zd from PyObject", + module_name, class_name, size, basicsize, basicsize+itemsize); + goto bad; + } + else if (check_size == __Pyx_ImportType_CheckSize_Warn_3_0_12 && (size_t)basicsize > size) { + PyOS_snprintf(warning, sizeof(warning), + "%s.%s size changed, may indicate binary incompatibility. " + "Expected %zd from C header, got %zd from PyObject", + module_name, class_name, size, basicsize); + if (PyErr_WarnEx(NULL, warning, 0) < 0) goto bad; + } + return (PyTypeObject *)result; +bad: + Py_XDECREF(result); + return NULL; +} +#endif + +/* FetchSharedCythonModule */ + static PyObject *__Pyx_FetchSharedCythonABIModule(void) { + return __Pyx_PyImport_AddModuleRef((char*) __PYX_ABI_MODULE_NAME); +} + +/* FetchCommonType */ + static int __Pyx_VerifyCachedType(PyObject *cached_type, + const char *name, + Py_ssize_t basicsize, + Py_ssize_t expected_basicsize) { + if (!PyType_Check(cached_type)) { + PyErr_Format(PyExc_TypeError, + "Shared Cython type %.200s is not a type object", name); + return -1; + } + if (basicsize != expected_basicsize) { + PyErr_Format(PyExc_TypeError, + "Shared Cython type %.200s has the wrong size, try recompiling", + name); + return -1; + } + return 0; +} +#if !CYTHON_USE_TYPE_SPECS +static PyTypeObject* __Pyx_FetchCommonType(PyTypeObject* type) { + PyObject* abi_module; + const char* object_name; + PyTypeObject *cached_type = NULL; + abi_module = __Pyx_FetchSharedCythonABIModule(); + if (!abi_module) return NULL; + object_name = strrchr(type->tp_name, '.'); + object_name = object_name ? object_name+1 : type->tp_name; + cached_type = (PyTypeObject*) PyObject_GetAttrString(abi_module, object_name); + if (cached_type) { + if (__Pyx_VerifyCachedType( + (PyObject *)cached_type, + object_name, + cached_type->tp_basicsize, + type->tp_basicsize) < 0) { + goto bad; + } + goto done; + } + if (!PyErr_ExceptionMatches(PyExc_AttributeError)) goto bad; + PyErr_Clear(); + if (PyType_Ready(type) < 0) goto bad; + if (PyObject_SetAttrString(abi_module, object_name, (PyObject *)type) < 0) + goto bad; + Py_INCREF(type); + cached_type = type; +done: + Py_DECREF(abi_module); + return cached_type; +bad: + Py_XDECREF(cached_type); + cached_type = NULL; + goto done; +} +#else +static PyTypeObject *__Pyx_FetchCommonTypeFromSpec(PyObject *module, PyType_Spec *spec, PyObject *bases) { + PyObject *abi_module, *cached_type = NULL; + const char* object_name = strrchr(spec->name, '.'); + object_name = object_name ? object_name+1 : spec->name; + abi_module = __Pyx_FetchSharedCythonABIModule(); + if (!abi_module) return NULL; + cached_type = PyObject_GetAttrString(abi_module, object_name); + if (cached_type) { + Py_ssize_t basicsize; +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *py_basicsize; + py_basicsize = PyObject_GetAttrString(cached_type, "__basicsize__"); + if (unlikely(!py_basicsize)) goto bad; + basicsize = PyLong_AsSsize_t(py_basicsize); + Py_DECREF(py_basicsize); + py_basicsize = 0; + if (unlikely(basicsize == (Py_ssize_t)-1) && PyErr_Occurred()) goto bad; +#else + basicsize = likely(PyType_Check(cached_type)) ? ((PyTypeObject*) cached_type)->tp_basicsize : -1; +#endif + if (__Pyx_VerifyCachedType( + cached_type, + object_name, + basicsize, + spec->basicsize) < 0) { + goto bad; + } + goto done; + } + if (!PyErr_ExceptionMatches(PyExc_AttributeError)) goto bad; + PyErr_Clear(); + CYTHON_UNUSED_VAR(module); + cached_type = __Pyx_PyType_FromModuleAndSpec(abi_module, spec, bases); + if (unlikely(!cached_type)) goto bad; + if (unlikely(__Pyx_fix_up_extension_type_from_spec(spec, (PyTypeObject *) cached_type) < 0)) goto bad; + if (PyObject_SetAttrString(abi_module, object_name, cached_type) < 0) goto bad; +done: + Py_DECREF(abi_module); + assert(cached_type == NULL || PyType_Check(cached_type)); + return (PyTypeObject *) cached_type; +bad: + Py_XDECREF(cached_type); + cached_type = NULL; + goto done; +} +#endif + +/* PyVectorcallFastCallDict */ + #if CYTHON_METH_FASTCALL +static PyObject *__Pyx_PyVectorcall_FastCallDict_kw(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw) +{ + PyObject *res = NULL; + PyObject *kwnames; + PyObject **newargs; + PyObject **kwvalues; + Py_ssize_t i, pos; + size_t j; + PyObject *key, *value; + unsigned long keys_are_strings; + Py_ssize_t nkw = PyDict_GET_SIZE(kw); + newargs = (PyObject **)PyMem_Malloc((nargs + (size_t)nkw) * sizeof(args[0])); + if (unlikely(newargs == NULL)) { + PyErr_NoMemory(); + return NULL; + } + for (j = 0; j < nargs; j++) newargs[j] = args[j]; + kwnames = PyTuple_New(nkw); + if (unlikely(kwnames == NULL)) { + PyMem_Free(newargs); + return NULL; + } + kwvalues = newargs + nargs; + pos = i = 0; + keys_are_strings = Py_TPFLAGS_UNICODE_SUBCLASS; + while (PyDict_Next(kw, &pos, &key, &value)) { + keys_are_strings &= Py_TYPE(key)->tp_flags; + Py_INCREF(key); + Py_INCREF(value); + PyTuple_SET_ITEM(kwnames, i, key); + kwvalues[i] = value; + i++; + } + if (unlikely(!keys_are_strings)) { + PyErr_SetString(PyExc_TypeError, "keywords must be strings"); + goto cleanup; + } + res = vc(func, newargs, nargs, kwnames); +cleanup: + Py_DECREF(kwnames); + for (i = 0; i < nkw; i++) + Py_DECREF(kwvalues[i]); + PyMem_Free(newargs); + return res; +} +static CYTHON_INLINE PyObject *__Pyx_PyVectorcall_FastCallDict(PyObject *func, __pyx_vectorcallfunc vc, PyObject *const *args, size_t nargs, PyObject *kw) +{ + if (likely(kw == NULL) || PyDict_GET_SIZE(kw) == 0) { + return vc(func, args, nargs, NULL); + } + return __Pyx_PyVectorcall_FastCallDict_kw(func, vc, args, nargs, kw); +} +#endif + +/* CythonFunctionShared */ + #if CYTHON_COMPILING_IN_LIMITED_API +static CYTHON_INLINE int __Pyx__IsSameCyOrCFunction(PyObject *func, void *cfunc) { + if (__Pyx_CyFunction_Check(func)) { + return PyCFunction_GetFunction(((__pyx_CyFunctionObject*)func)->func) == (PyCFunction) cfunc; + } else if (PyCFunction_Check(func)) { + return PyCFunction_GetFunction(func) == (PyCFunction) cfunc; + } + return 0; +} +#else +static CYTHON_INLINE int __Pyx__IsSameCyOrCFunction(PyObject *func, void *cfunc) { + return __Pyx_CyOrPyCFunction_Check(func) && __Pyx_CyOrPyCFunction_GET_FUNCTION(func) == (PyCFunction) cfunc; +} +#endif +static CYTHON_INLINE void __Pyx__CyFunction_SetClassObj(__pyx_CyFunctionObject* f, PyObject* classobj) { +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + __Pyx_Py_XDECREF_SET( + __Pyx_CyFunction_GetClassObj(f), + ((classobj) ? __Pyx_NewRef(classobj) : NULL)); +#else + __Pyx_Py_XDECREF_SET( + ((PyCMethodObject *) (f))->mm_class, + (PyTypeObject*)((classobj) ? __Pyx_NewRef(classobj) : NULL)); +#endif +} +static PyObject * +__Pyx_CyFunction_get_doc(__pyx_CyFunctionObject *op, void *closure) +{ + CYTHON_UNUSED_VAR(closure); + if (unlikely(op->func_doc == NULL)) { +#if CYTHON_COMPILING_IN_LIMITED_API + op->func_doc = PyObject_GetAttrString(op->func, "__doc__"); + if (unlikely(!op->func_doc)) return NULL; +#else + if (((PyCFunctionObject*)op)->m_ml->ml_doc) { +#if PY_MAJOR_VERSION >= 3 + op->func_doc = PyUnicode_FromString(((PyCFunctionObject*)op)->m_ml->ml_doc); +#else + op->func_doc = PyString_FromString(((PyCFunctionObject*)op)->m_ml->ml_doc); +#endif + if (unlikely(op->func_doc == NULL)) + return NULL; + } else { + Py_INCREF(Py_None); + return Py_None; + } +#endif + } + Py_INCREF(op->func_doc); + return op->func_doc; +} +static int +__Pyx_CyFunction_set_doc(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (value == NULL) { + value = Py_None; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_doc, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_name(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (unlikely(op->func_name == NULL)) { +#if CYTHON_COMPILING_IN_LIMITED_API + op->func_name = PyObject_GetAttrString(op->func, "__name__"); +#elif PY_MAJOR_VERSION >= 3 + op->func_name = PyUnicode_InternFromString(((PyCFunctionObject*)op)->m_ml->ml_name); +#else + op->func_name = PyString_InternFromString(((PyCFunctionObject*)op)->m_ml->ml_name); +#endif + if (unlikely(op->func_name == NULL)) + return NULL; + } + Py_INCREF(op->func_name); + return op->func_name; +} +static int +__Pyx_CyFunction_set_name(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); +#if PY_MAJOR_VERSION >= 3 + if (unlikely(value == NULL || !PyUnicode_Check(value))) +#else + if (unlikely(value == NULL || !PyString_Check(value))) +#endif + { + PyErr_SetString(PyExc_TypeError, + "__name__ must be set to a string object"); + return -1; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_name, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_qualname(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + Py_INCREF(op->func_qualname); + return op->func_qualname; +} +static int +__Pyx_CyFunction_set_qualname(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); +#if PY_MAJOR_VERSION >= 3 + if (unlikely(value == NULL || !PyUnicode_Check(value))) +#else + if (unlikely(value == NULL || !PyString_Check(value))) +#endif + { + PyErr_SetString(PyExc_TypeError, + "__qualname__ must be set to a string object"); + return -1; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_qualname, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_dict(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (unlikely(op->func_dict == NULL)) { + op->func_dict = PyDict_New(); + if (unlikely(op->func_dict == NULL)) + return NULL; + } + Py_INCREF(op->func_dict); + return op->func_dict; +} +static int +__Pyx_CyFunction_set_dict(__pyx_CyFunctionObject *op, PyObject *value, void *context) +{ + CYTHON_UNUSED_VAR(context); + if (unlikely(value == NULL)) { + PyErr_SetString(PyExc_TypeError, + "function's dictionary may not be deleted"); + return -1; + } + if (unlikely(!PyDict_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "setting function's dictionary to a non-dict"); + return -1; + } + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->func_dict, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_globals(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(context); + Py_INCREF(op->func_globals); + return op->func_globals; +} +static PyObject * +__Pyx_CyFunction_get_closure(__pyx_CyFunctionObject *op, void *context) +{ + CYTHON_UNUSED_VAR(op); + CYTHON_UNUSED_VAR(context); + Py_INCREF(Py_None); + return Py_None; +} +static PyObject * +__Pyx_CyFunction_get_code(__pyx_CyFunctionObject *op, void *context) +{ + PyObject* result = (op->func_code) ? op->func_code : Py_None; + CYTHON_UNUSED_VAR(context); + Py_INCREF(result); + return result; +} +static int +__Pyx_CyFunction_init_defaults(__pyx_CyFunctionObject *op) { + int result = 0; + PyObject *res = op->defaults_getter((PyObject *) op); + if (unlikely(!res)) + return -1; + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + op->defaults_tuple = PyTuple_GET_ITEM(res, 0); + Py_INCREF(op->defaults_tuple); + op->defaults_kwdict = PyTuple_GET_ITEM(res, 1); + Py_INCREF(op->defaults_kwdict); + #else + op->defaults_tuple = __Pyx_PySequence_ITEM(res, 0); + if (unlikely(!op->defaults_tuple)) result = -1; + else { + op->defaults_kwdict = __Pyx_PySequence_ITEM(res, 1); + if (unlikely(!op->defaults_kwdict)) result = -1; + } + #endif + Py_DECREF(res); + return result; +} +static int +__Pyx_CyFunction_set_defaults(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + if (!value) { + value = Py_None; + } else if (unlikely(value != Py_None && !PyTuple_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "__defaults__ must be set to a tuple object"); + return -1; + } + PyErr_WarnEx(PyExc_RuntimeWarning, "changes to cyfunction.__defaults__ will not " + "currently affect the values used in function calls", 1); + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->defaults_tuple, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_defaults(__pyx_CyFunctionObject *op, void *context) { + PyObject* result = op->defaults_tuple; + CYTHON_UNUSED_VAR(context); + if (unlikely(!result)) { + if (op->defaults_getter) { + if (unlikely(__Pyx_CyFunction_init_defaults(op) < 0)) return NULL; + result = op->defaults_tuple; + } else { + result = Py_None; + } + } + Py_INCREF(result); + return result; +} +static int +__Pyx_CyFunction_set_kwdefaults(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + if (!value) { + value = Py_None; + } else if (unlikely(value != Py_None && !PyDict_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "__kwdefaults__ must be set to a dict object"); + return -1; + } + PyErr_WarnEx(PyExc_RuntimeWarning, "changes to cyfunction.__kwdefaults__ will not " + "currently affect the values used in function calls", 1); + Py_INCREF(value); + __Pyx_Py_XDECREF_SET(op->defaults_kwdict, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_kwdefaults(__pyx_CyFunctionObject *op, void *context) { + PyObject* result = op->defaults_kwdict; + CYTHON_UNUSED_VAR(context); + if (unlikely(!result)) { + if (op->defaults_getter) { + if (unlikely(__Pyx_CyFunction_init_defaults(op) < 0)) return NULL; + result = op->defaults_kwdict; + } else { + result = Py_None; + } + } + Py_INCREF(result); + return result; +} +static int +__Pyx_CyFunction_set_annotations(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + if (!value || value == Py_None) { + value = NULL; + } else if (unlikely(!PyDict_Check(value))) { + PyErr_SetString(PyExc_TypeError, + "__annotations__ must be set to a dict object"); + return -1; + } + Py_XINCREF(value); + __Pyx_Py_XDECREF_SET(op->func_annotations, value); + return 0; +} +static PyObject * +__Pyx_CyFunction_get_annotations(__pyx_CyFunctionObject *op, void *context) { + PyObject* result = op->func_annotations; + CYTHON_UNUSED_VAR(context); + if (unlikely(!result)) { + result = PyDict_New(); + if (unlikely(!result)) return NULL; + op->func_annotations = result; + } + Py_INCREF(result); + return result; +} +static PyObject * +__Pyx_CyFunction_get_is_coroutine(__pyx_CyFunctionObject *op, void *context) { + int is_coroutine; + CYTHON_UNUSED_VAR(context); + if (op->func_is_coroutine) { + return __Pyx_NewRef(op->func_is_coroutine); + } + is_coroutine = op->flags & __Pyx_CYFUNCTION_COROUTINE; +#if PY_VERSION_HEX >= 0x03050000 + if (is_coroutine) { + PyObject *module, *fromlist, *marker = __pyx_n_s_is_coroutine; + fromlist = PyList_New(1); + if (unlikely(!fromlist)) return NULL; + Py_INCREF(marker); +#if CYTHON_ASSUME_SAFE_MACROS + PyList_SET_ITEM(fromlist, 0, marker); +#else + if (unlikely(PyList_SetItem(fromlist, 0, marker) < 0)) { + Py_DECREF(marker); + Py_DECREF(fromlist); + return NULL; + } +#endif + module = PyImport_ImportModuleLevelObject(__pyx_n_s_asyncio_coroutines, NULL, NULL, fromlist, 0); + Py_DECREF(fromlist); + if (unlikely(!module)) goto ignore; + op->func_is_coroutine = __Pyx_PyObject_GetAttrStr(module, marker); + Py_DECREF(module); + if (likely(op->func_is_coroutine)) { + return __Pyx_NewRef(op->func_is_coroutine); + } +ignore: + PyErr_Clear(); + } +#endif + op->func_is_coroutine = __Pyx_PyBool_FromLong(is_coroutine); + return __Pyx_NewRef(op->func_is_coroutine); +} +#if CYTHON_COMPILING_IN_LIMITED_API +static PyObject * +__Pyx_CyFunction_get_module(__pyx_CyFunctionObject *op, void *context) { + CYTHON_UNUSED_VAR(context); + return PyObject_GetAttrString(op->func, "__module__"); +} +static int +__Pyx_CyFunction_set_module(__pyx_CyFunctionObject *op, PyObject* value, void *context) { + CYTHON_UNUSED_VAR(context); + return PyObject_SetAttrString(op->func, "__module__", value); +} +#endif +static PyGetSetDef __pyx_CyFunction_getsets[] = { + {(char *) "func_doc", (getter)__Pyx_CyFunction_get_doc, (setter)__Pyx_CyFunction_set_doc, 0, 0}, + {(char *) "__doc__", (getter)__Pyx_CyFunction_get_doc, (setter)__Pyx_CyFunction_set_doc, 0, 0}, + {(char *) "func_name", (getter)__Pyx_CyFunction_get_name, (setter)__Pyx_CyFunction_set_name, 0, 0}, + {(char *) "__name__", (getter)__Pyx_CyFunction_get_name, (setter)__Pyx_CyFunction_set_name, 0, 0}, + {(char *) "__qualname__", (getter)__Pyx_CyFunction_get_qualname, (setter)__Pyx_CyFunction_set_qualname, 0, 0}, + {(char *) "func_dict", (getter)__Pyx_CyFunction_get_dict, (setter)__Pyx_CyFunction_set_dict, 0, 0}, + {(char *) "__dict__", (getter)__Pyx_CyFunction_get_dict, (setter)__Pyx_CyFunction_set_dict, 0, 0}, + {(char *) "func_globals", (getter)__Pyx_CyFunction_get_globals, 0, 0, 0}, + {(char *) "__globals__", (getter)__Pyx_CyFunction_get_globals, 0, 0, 0}, + {(char *) "func_closure", (getter)__Pyx_CyFunction_get_closure, 0, 0, 0}, + {(char *) "__closure__", (getter)__Pyx_CyFunction_get_closure, 0, 0, 0}, + {(char *) "func_code", (getter)__Pyx_CyFunction_get_code, 0, 0, 0}, + {(char *) "__code__", (getter)__Pyx_CyFunction_get_code, 0, 0, 0}, + {(char *) "func_defaults", (getter)__Pyx_CyFunction_get_defaults, (setter)__Pyx_CyFunction_set_defaults, 0, 0}, + {(char *) "__defaults__", (getter)__Pyx_CyFunction_get_defaults, (setter)__Pyx_CyFunction_set_defaults, 0, 0}, + {(char *) "__kwdefaults__", (getter)__Pyx_CyFunction_get_kwdefaults, (setter)__Pyx_CyFunction_set_kwdefaults, 0, 0}, + {(char *) "__annotations__", (getter)__Pyx_CyFunction_get_annotations, (setter)__Pyx_CyFunction_set_annotations, 0, 0}, + {(char *) "_is_coroutine", (getter)__Pyx_CyFunction_get_is_coroutine, 0, 0, 0}, +#if CYTHON_COMPILING_IN_LIMITED_API + {"__module__", (getter)__Pyx_CyFunction_get_module, (setter)__Pyx_CyFunction_set_module, 0, 0}, +#endif + {0, 0, 0, 0, 0} +}; +static PyMemberDef __pyx_CyFunction_members[] = { +#if !CYTHON_COMPILING_IN_LIMITED_API + {(char *) "__module__", T_OBJECT, offsetof(PyCFunctionObject, m_module), 0, 0}, +#endif +#if CYTHON_USE_TYPE_SPECS + {(char *) "__dictoffset__", T_PYSSIZET, offsetof(__pyx_CyFunctionObject, func_dict), READONLY, 0}, +#if CYTHON_METH_FASTCALL +#if CYTHON_BACKPORT_VECTORCALL + {(char *) "__vectorcalloffset__", T_PYSSIZET, offsetof(__pyx_CyFunctionObject, func_vectorcall), READONLY, 0}, +#else +#if !CYTHON_COMPILING_IN_LIMITED_API + {(char *) "__vectorcalloffset__", T_PYSSIZET, offsetof(PyCFunctionObject, vectorcall), READONLY, 0}, +#endif +#endif +#endif +#if PY_VERSION_HEX < 0x030500A0 || CYTHON_COMPILING_IN_LIMITED_API + {(char *) "__weaklistoffset__", T_PYSSIZET, offsetof(__pyx_CyFunctionObject, func_weakreflist), READONLY, 0}, +#else + {(char *) "__weaklistoffset__", T_PYSSIZET, offsetof(PyCFunctionObject, m_weakreflist), READONLY, 0}, +#endif +#endif + {0, 0, 0, 0, 0} +}; +static PyObject * +__Pyx_CyFunction_reduce(__pyx_CyFunctionObject *m, PyObject *args) +{ + CYTHON_UNUSED_VAR(args); +#if PY_MAJOR_VERSION >= 3 + Py_INCREF(m->func_qualname); + return m->func_qualname; +#else + return PyString_FromString(((PyCFunctionObject*)m)->m_ml->ml_name); +#endif +} +static PyMethodDef __pyx_CyFunction_methods[] = { + {"__reduce__", (PyCFunction)__Pyx_CyFunction_reduce, METH_VARARGS, 0}, + {0, 0, 0, 0} +}; +#if PY_VERSION_HEX < 0x030500A0 || CYTHON_COMPILING_IN_LIMITED_API +#define __Pyx_CyFunction_weakreflist(cyfunc) ((cyfunc)->func_weakreflist) +#else +#define __Pyx_CyFunction_weakreflist(cyfunc) (((PyCFunctionObject*)cyfunc)->m_weakreflist) +#endif +static PyObject *__Pyx_CyFunction_Init(__pyx_CyFunctionObject *op, PyMethodDef *ml, int flags, PyObject* qualname, + PyObject *closure, PyObject *module, PyObject* globals, PyObject* code) { +#if !CYTHON_COMPILING_IN_LIMITED_API + PyCFunctionObject *cf = (PyCFunctionObject*) op; +#endif + if (unlikely(op == NULL)) + return NULL; +#if CYTHON_COMPILING_IN_LIMITED_API + op->func = PyCFunction_NewEx(ml, (PyObject*)op, module); + if (unlikely(!op->func)) return NULL; +#endif + op->flags = flags; + __Pyx_CyFunction_weakreflist(op) = NULL; +#if !CYTHON_COMPILING_IN_LIMITED_API + cf->m_ml = ml; + cf->m_self = (PyObject *) op; +#endif + Py_XINCREF(closure); + op->func_closure = closure; +#if !CYTHON_COMPILING_IN_LIMITED_API + Py_XINCREF(module); + cf->m_module = module; +#endif + op->func_dict = NULL; + op->func_name = NULL; + Py_INCREF(qualname); + op->func_qualname = qualname; + op->func_doc = NULL; +#if PY_VERSION_HEX < 0x030900B1 || CYTHON_COMPILING_IN_LIMITED_API + op->func_classobj = NULL; +#else + ((PyCMethodObject*)op)->mm_class = NULL; +#endif + op->func_globals = globals; + Py_INCREF(op->func_globals); + Py_XINCREF(code); + op->func_code = code; + op->defaults_pyobjects = 0; + op->defaults_size = 0; + op->defaults = NULL; + op->defaults_tuple = NULL; + op->defaults_kwdict = NULL; + op->defaults_getter = NULL; + op->func_annotations = NULL; + op->func_is_coroutine = NULL; +#if CYTHON_METH_FASTCALL + switch (ml->ml_flags & (METH_VARARGS | METH_FASTCALL | METH_NOARGS | METH_O | METH_KEYWORDS | METH_METHOD)) { + case METH_NOARGS: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_NOARGS; + break; + case METH_O: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_O; + break; + case METH_METHOD | METH_FASTCALL | METH_KEYWORDS: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS_METHOD; + break; + case METH_FASTCALL | METH_KEYWORDS: + __Pyx_CyFunction_func_vectorcall(op) = __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS; + break; + case METH_VARARGS | METH_KEYWORDS: + __Pyx_CyFunction_func_vectorcall(op) = NULL; + break; + default: + PyErr_SetString(PyExc_SystemError, "Bad call flags for CyFunction"); + Py_DECREF(op); + return NULL; + } +#endif + return (PyObject *) op; +} +static int +__Pyx_CyFunction_clear(__pyx_CyFunctionObject *m) +{ + Py_CLEAR(m->func_closure); +#if CYTHON_COMPILING_IN_LIMITED_API + Py_CLEAR(m->func); +#else + Py_CLEAR(((PyCFunctionObject*)m)->m_module); +#endif + Py_CLEAR(m->func_dict); + Py_CLEAR(m->func_name); + Py_CLEAR(m->func_qualname); + Py_CLEAR(m->func_doc); + Py_CLEAR(m->func_globals); + Py_CLEAR(m->func_code); +#if !CYTHON_COMPILING_IN_LIMITED_API +#if PY_VERSION_HEX < 0x030900B1 + Py_CLEAR(__Pyx_CyFunction_GetClassObj(m)); +#else + { + PyObject *cls = (PyObject*) ((PyCMethodObject *) (m))->mm_class; + ((PyCMethodObject *) (m))->mm_class = NULL; + Py_XDECREF(cls); + } +#endif +#endif + Py_CLEAR(m->defaults_tuple); + Py_CLEAR(m->defaults_kwdict); + Py_CLEAR(m->func_annotations); + Py_CLEAR(m->func_is_coroutine); + if (m->defaults) { + PyObject **pydefaults = __Pyx_CyFunction_Defaults(PyObject *, m); + int i; + for (i = 0; i < m->defaults_pyobjects; i++) + Py_XDECREF(pydefaults[i]); + PyObject_Free(m->defaults); + m->defaults = NULL; + } + return 0; +} +static void __Pyx__CyFunction_dealloc(__pyx_CyFunctionObject *m) +{ + if (__Pyx_CyFunction_weakreflist(m) != NULL) + PyObject_ClearWeakRefs((PyObject *) m); + __Pyx_CyFunction_clear(m); + __Pyx_PyHeapTypeObject_GC_Del(m); +} +static void __Pyx_CyFunction_dealloc(__pyx_CyFunctionObject *m) +{ + PyObject_GC_UnTrack(m); + __Pyx__CyFunction_dealloc(m); +} +static int __Pyx_CyFunction_traverse(__pyx_CyFunctionObject *m, visitproc visit, void *arg) +{ + Py_VISIT(m->func_closure); +#if CYTHON_COMPILING_IN_LIMITED_API + Py_VISIT(m->func); +#else + Py_VISIT(((PyCFunctionObject*)m)->m_module); +#endif + Py_VISIT(m->func_dict); + Py_VISIT(m->func_name); + Py_VISIT(m->func_qualname); + Py_VISIT(m->func_doc); + Py_VISIT(m->func_globals); + Py_VISIT(m->func_code); +#if !CYTHON_COMPILING_IN_LIMITED_API + Py_VISIT(__Pyx_CyFunction_GetClassObj(m)); +#endif + Py_VISIT(m->defaults_tuple); + Py_VISIT(m->defaults_kwdict); + Py_VISIT(m->func_is_coroutine); + if (m->defaults) { + PyObject **pydefaults = __Pyx_CyFunction_Defaults(PyObject *, m); + int i; + for (i = 0; i < m->defaults_pyobjects; i++) + Py_VISIT(pydefaults[i]); + } + return 0; +} +static PyObject* +__Pyx_CyFunction_repr(__pyx_CyFunctionObject *op) +{ +#if PY_MAJOR_VERSION >= 3 + return PyUnicode_FromFormat("", + op->func_qualname, (void *)op); +#else + return PyString_FromFormat("", + PyString_AsString(op->func_qualname), (void *)op); +#endif +} +static PyObject * __Pyx_CyFunction_CallMethod(PyObject *func, PyObject *self, PyObject *arg, PyObject *kw) { +#if CYTHON_COMPILING_IN_LIMITED_API + PyObject *f = ((__pyx_CyFunctionObject*)func)->func; + PyObject *py_name = NULL; + PyCFunction meth; + int flags; + meth = PyCFunction_GetFunction(f); + if (unlikely(!meth)) return NULL; + flags = PyCFunction_GetFlags(f); + if (unlikely(flags < 0)) return NULL; +#else + PyCFunctionObject* f = (PyCFunctionObject*)func; + PyCFunction meth = f->m_ml->ml_meth; + int flags = f->m_ml->ml_flags; +#endif + Py_ssize_t size; + switch (flags & (METH_VARARGS | METH_KEYWORDS | METH_NOARGS | METH_O)) { + case METH_VARARGS: + if (likely(kw == NULL || PyDict_Size(kw) == 0)) + return (*meth)(self, arg); + break; + case METH_VARARGS | METH_KEYWORDS: + return (*(PyCFunctionWithKeywords)(void*)meth)(self, arg, kw); + case METH_NOARGS: + if (likely(kw == NULL || PyDict_Size(kw) == 0)) { +#if CYTHON_ASSUME_SAFE_MACROS + size = PyTuple_GET_SIZE(arg); +#else + size = PyTuple_Size(arg); + if (unlikely(size < 0)) return NULL; +#endif + if (likely(size == 0)) + return (*meth)(self, NULL); +#if CYTHON_COMPILING_IN_LIMITED_API + py_name = __Pyx_CyFunction_get_name((__pyx_CyFunctionObject*)func, NULL); + if (!py_name) return NULL; + PyErr_Format(PyExc_TypeError, + "%.200S() takes no arguments (%" CYTHON_FORMAT_SSIZE_T "d given)", + py_name, size); + Py_DECREF(py_name); +#else + PyErr_Format(PyExc_TypeError, + "%.200s() takes no arguments (%" CYTHON_FORMAT_SSIZE_T "d given)", + f->m_ml->ml_name, size); +#endif + return NULL; + } + break; + case METH_O: + if (likely(kw == NULL || PyDict_Size(kw) == 0)) { +#if CYTHON_ASSUME_SAFE_MACROS + size = PyTuple_GET_SIZE(arg); +#else + size = PyTuple_Size(arg); + if (unlikely(size < 0)) return NULL; +#endif + if (likely(size == 1)) { + PyObject *result, *arg0; + #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS + arg0 = PyTuple_GET_ITEM(arg, 0); + #else + arg0 = __Pyx_PySequence_ITEM(arg, 0); if (unlikely(!arg0)) return NULL; + #endif + result = (*meth)(self, arg0); + #if !(CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS) + Py_DECREF(arg0); + #endif + return result; + } +#if CYTHON_COMPILING_IN_LIMITED_API + py_name = __Pyx_CyFunction_get_name((__pyx_CyFunctionObject*)func, NULL); + if (!py_name) return NULL; + PyErr_Format(PyExc_TypeError, + "%.200S() takes exactly one argument (%" CYTHON_FORMAT_SSIZE_T "d given)", + py_name, size); + Py_DECREF(py_name); +#else + PyErr_Format(PyExc_TypeError, + "%.200s() takes exactly one argument (%" CYTHON_FORMAT_SSIZE_T "d given)", + f->m_ml->ml_name, size); +#endif + return NULL; + } + break; + default: + PyErr_SetString(PyExc_SystemError, "Bad call flags for CyFunction"); + return NULL; + } +#if CYTHON_COMPILING_IN_LIMITED_API + py_name = __Pyx_CyFunction_get_name((__pyx_CyFunctionObject*)func, NULL); + if (!py_name) return NULL; + PyErr_Format(PyExc_TypeError, "%.200S() takes no keyword arguments", + py_name); + Py_DECREF(py_name); +#else + PyErr_Format(PyExc_TypeError, "%.200s() takes no keyword arguments", + f->m_ml->ml_name); +#endif + return NULL; +} +static CYTHON_INLINE PyObject *__Pyx_CyFunction_Call(PyObject *func, PyObject *arg, PyObject *kw) { + PyObject *self, *result; +#if CYTHON_COMPILING_IN_LIMITED_API + self = PyCFunction_GetSelf(((__pyx_CyFunctionObject*)func)->func); + if (unlikely(!self) && PyErr_Occurred()) return NULL; +#else + self = ((PyCFunctionObject*)func)->m_self; +#endif + result = __Pyx_CyFunction_CallMethod(func, self, arg, kw); + return result; +} +static PyObject *__Pyx_CyFunction_CallAsMethod(PyObject *func, PyObject *args, PyObject *kw) { + PyObject *result; + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *) func; +#if CYTHON_METH_FASTCALL + __pyx_vectorcallfunc vc = __Pyx_CyFunction_func_vectorcall(cyfunc); + if (vc) { +#if CYTHON_ASSUME_SAFE_MACROS + return __Pyx_PyVectorcall_FastCallDict(func, vc, &PyTuple_GET_ITEM(args, 0), (size_t)PyTuple_GET_SIZE(args), kw); +#else + (void) &__Pyx_PyVectorcall_FastCallDict; + return PyVectorcall_Call(func, args, kw); +#endif + } +#endif + if ((cyfunc->flags & __Pyx_CYFUNCTION_CCLASS) && !(cyfunc->flags & __Pyx_CYFUNCTION_STATICMETHOD)) { + Py_ssize_t argc; + PyObject *new_args; + PyObject *self; +#if CYTHON_ASSUME_SAFE_MACROS + argc = PyTuple_GET_SIZE(args); +#else + argc = PyTuple_Size(args); + if (unlikely(!argc) < 0) return NULL; +#endif + new_args = PyTuple_GetSlice(args, 1, argc); + if (unlikely(!new_args)) + return NULL; + self = PyTuple_GetItem(args, 0); + if (unlikely(!self)) { + Py_DECREF(new_args); +#if PY_MAJOR_VERSION > 2 + PyErr_Format(PyExc_TypeError, + "unbound method %.200S() needs an argument", + cyfunc->func_qualname); +#else + PyErr_SetString(PyExc_TypeError, + "unbound method needs an argument"); +#endif + return NULL; + } + result = __Pyx_CyFunction_CallMethod(func, self, new_args, kw); + Py_DECREF(new_args); + } else { + result = __Pyx_CyFunction_Call(func, args, kw); + } + return result; +} +#if CYTHON_METH_FASTCALL +static CYTHON_INLINE int __Pyx_CyFunction_Vectorcall_CheckArgs(__pyx_CyFunctionObject *cyfunc, Py_ssize_t nargs, PyObject *kwnames) +{ + int ret = 0; + if ((cyfunc->flags & __Pyx_CYFUNCTION_CCLASS) && !(cyfunc->flags & __Pyx_CYFUNCTION_STATICMETHOD)) { + if (unlikely(nargs < 1)) { + PyErr_Format(PyExc_TypeError, "%.200s() needs an argument", + ((PyCFunctionObject*)cyfunc)->m_ml->ml_name); + return -1; + } + ret = 1; + } + if (unlikely(kwnames) && unlikely(PyTuple_GET_SIZE(kwnames))) { + PyErr_Format(PyExc_TypeError, + "%.200s() takes no keyword arguments", ((PyCFunctionObject*)cyfunc)->m_ml->ml_name); + return -1; + } + return ret; +} +static PyObject * __Pyx_CyFunction_Vectorcall_NOARGS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, kwnames)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + if (unlikely(nargs != 0)) { + PyErr_Format(PyExc_TypeError, + "%.200s() takes no arguments (%" CYTHON_FORMAT_SSIZE_T "d given)", + def->ml_name, nargs); + return NULL; + } + return def->ml_meth(self, NULL); +} +static PyObject * __Pyx_CyFunction_Vectorcall_O(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, kwnames)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + if (unlikely(nargs != 1)) { + PyErr_Format(PyExc_TypeError, + "%.200s() takes exactly one argument (%" CYTHON_FORMAT_SSIZE_T "d given)", + def->ml_name, nargs); + return NULL; + } + return def->ml_meth(self, args[0]); +} +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, NULL)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + return ((__Pyx_PyCFunctionFastWithKeywords)(void(*)(void))def->ml_meth)(self, args, nargs, kwnames); +} +static PyObject * __Pyx_CyFunction_Vectorcall_FASTCALL_KEYWORDS_METHOD(PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames) +{ + __pyx_CyFunctionObject *cyfunc = (__pyx_CyFunctionObject *)func; + PyMethodDef* def = ((PyCFunctionObject*)cyfunc)->m_ml; + PyTypeObject *cls = (PyTypeObject *) __Pyx_CyFunction_GetClassObj(cyfunc); +#if CYTHON_BACKPORT_VECTORCALL + Py_ssize_t nargs = (Py_ssize_t)nargsf; +#else + Py_ssize_t nargs = PyVectorcall_NARGS(nargsf); +#endif + PyObject *self; + switch (__Pyx_CyFunction_Vectorcall_CheckArgs(cyfunc, nargs, NULL)) { + case 1: + self = args[0]; + args += 1; + nargs -= 1; + break; + case 0: + self = ((PyCFunctionObject*)cyfunc)->m_self; + break; + default: + return NULL; + } + return ((__Pyx_PyCMethod)(void(*)(void))def->ml_meth)(self, cls, args, (size_t)nargs, kwnames); +} +#endif +#if CYTHON_USE_TYPE_SPECS +static PyType_Slot __pyx_CyFunctionType_slots[] = { + {Py_tp_dealloc, (void *)__Pyx_CyFunction_dealloc}, + {Py_tp_repr, (void *)__Pyx_CyFunction_repr}, + {Py_tp_call, (void *)__Pyx_CyFunction_CallAsMethod}, + {Py_tp_traverse, (void *)__Pyx_CyFunction_traverse}, + {Py_tp_clear, (void *)__Pyx_CyFunction_clear}, + {Py_tp_methods, (void *)__pyx_CyFunction_methods}, + {Py_tp_members, (void *)__pyx_CyFunction_members}, + {Py_tp_getset, (void *)__pyx_CyFunction_getsets}, + {Py_tp_descr_get, (void *)__Pyx_PyMethod_New}, + {0, 0}, +}; +static PyType_Spec __pyx_CyFunctionType_spec = { + __PYX_TYPE_MODULE_PREFIX "cython_function_or_method", + sizeof(__pyx_CyFunctionObject), + 0, +#ifdef Py_TPFLAGS_METHOD_DESCRIPTOR + Py_TPFLAGS_METHOD_DESCRIPTOR | +#endif +#if (defined(_Py_TPFLAGS_HAVE_VECTORCALL) && CYTHON_METH_FASTCALL) + _Py_TPFLAGS_HAVE_VECTORCALL | +#endif + Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HAVE_GC | Py_TPFLAGS_BASETYPE, + __pyx_CyFunctionType_slots +}; +#else +static PyTypeObject __pyx_CyFunctionType_type = { + PyVarObject_HEAD_INIT(0, 0) + __PYX_TYPE_MODULE_PREFIX "cython_function_or_method", + sizeof(__pyx_CyFunctionObject), + 0, + (destructor) __Pyx_CyFunction_dealloc, +#if !CYTHON_METH_FASTCALL + 0, +#elif CYTHON_BACKPORT_VECTORCALL + (printfunc)offsetof(__pyx_CyFunctionObject, func_vectorcall), +#else + offsetof(PyCFunctionObject, vectorcall), +#endif + 0, + 0, +#if PY_MAJOR_VERSION < 3 + 0, +#else + 0, +#endif + (reprfunc) __Pyx_CyFunction_repr, + 0, + 0, + 0, + 0, + __Pyx_CyFunction_CallAsMethod, + 0, + 0, + 0, + 0, +#ifdef Py_TPFLAGS_METHOD_DESCRIPTOR + Py_TPFLAGS_METHOD_DESCRIPTOR | +#endif +#if defined(_Py_TPFLAGS_HAVE_VECTORCALL) && CYTHON_METH_FASTCALL + _Py_TPFLAGS_HAVE_VECTORCALL | +#endif + Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HAVE_GC | Py_TPFLAGS_BASETYPE, + 0, + (traverseproc) __Pyx_CyFunction_traverse, + (inquiry) __Pyx_CyFunction_clear, + 0, +#if PY_VERSION_HEX < 0x030500A0 + offsetof(__pyx_CyFunctionObject, func_weakreflist), +#else + offsetof(PyCFunctionObject, m_weakreflist), +#endif + 0, + 0, + __pyx_CyFunction_methods, + __pyx_CyFunction_members, + __pyx_CyFunction_getsets, + 0, + 0, + __Pyx_PyMethod_New, + 0, + offsetof(__pyx_CyFunctionObject, func_dict), + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, + 0, +#if PY_VERSION_HEX >= 0x030400a1 + 0, +#endif +#if PY_VERSION_HEX >= 0x030800b1 && (!CYTHON_COMPILING_IN_PYPY || PYPY_VERSION_NUM >= 0x07030800) + 0, +#endif +#if __PYX_NEED_TP_PRINT_SLOT + 0, +#endif +#if PY_VERSION_HEX >= 0x030C0000 + 0, +#endif +#if PY_VERSION_HEX >= 0x030d00A4 + 0, +#endif +#if CYTHON_COMPILING_IN_PYPY && PY_VERSION_HEX >= 0x03090000 && PY_VERSION_HEX < 0x030a0000 + 0, +#endif +}; +#endif +static int __pyx_CyFunction_init(PyObject *module) { +#if CYTHON_USE_TYPE_SPECS + __pyx_CyFunctionType = __Pyx_FetchCommonTypeFromSpec(module, &__pyx_CyFunctionType_spec, NULL); +#else + CYTHON_UNUSED_VAR(module); + __pyx_CyFunctionType = __Pyx_FetchCommonType(&__pyx_CyFunctionType_type); +#endif + if (unlikely(__pyx_CyFunctionType == NULL)) { + return -1; + } + return 0; +} +static CYTHON_INLINE void *__Pyx_CyFunction_InitDefaults(PyObject *func, size_t size, int pyobjects) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->defaults = PyObject_Malloc(size); + if (unlikely(!m->defaults)) + return PyErr_NoMemory(); + memset(m->defaults, 0, size); + m->defaults_pyobjects = pyobjects; + m->defaults_size = size; + return m->defaults; +} +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsTuple(PyObject *func, PyObject *tuple) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->defaults_tuple = tuple; + Py_INCREF(tuple); +} +static CYTHON_INLINE void __Pyx_CyFunction_SetDefaultsKwDict(PyObject *func, PyObject *dict) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->defaults_kwdict = dict; + Py_INCREF(dict); +} +static CYTHON_INLINE void __Pyx_CyFunction_SetAnnotationsDict(PyObject *func, PyObject *dict) { + __pyx_CyFunctionObject *m = (__pyx_CyFunctionObject *) func; + m->func_annotations = dict; + Py_INCREF(dict); +} + +/* CythonFunction */ + static PyObject *__Pyx_CyFunction_New(PyMethodDef *ml, int flags, PyObject* qualname, + PyObject *closure, PyObject *module, PyObject* globals, PyObject* code) { + PyObject *op = __Pyx_CyFunction_Init( + PyObject_GC_New(__pyx_CyFunctionObject, __pyx_CyFunctionType), + ml, flags, qualname, closure, module, globals, code + ); + if (likely(op)) { + PyObject_GC_Track(op); + } + return op; +} + +/* CLineInTraceback */ + #ifndef CYTHON_CLINE_IN_TRACEBACK +static int __Pyx_CLineForTraceback(PyThreadState *tstate, int c_line) { + PyObject *use_cline; + PyObject *ptype, *pvalue, *ptraceback; +#if CYTHON_COMPILING_IN_CPYTHON + PyObject **cython_runtime_dict; +#endif + CYTHON_MAYBE_UNUSED_VAR(tstate); + if (unlikely(!__pyx_cython_runtime)) { + return c_line; + } + __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback); +#if CYTHON_COMPILING_IN_CPYTHON + cython_runtime_dict = _PyObject_GetDictPtr(__pyx_cython_runtime); + if (likely(cython_runtime_dict)) { + __PYX_PY_DICT_LOOKUP_IF_MODIFIED( + use_cline, *cython_runtime_dict, + __Pyx_PyDict_GetItemStr(*cython_runtime_dict, __pyx_n_s_cline_in_traceback)) + } else +#endif + { + PyObject *use_cline_obj = __Pyx_PyObject_GetAttrStrNoError(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback); + if (use_cline_obj) { + use_cline = PyObject_Not(use_cline_obj) ? Py_False : Py_True; + Py_DECREF(use_cline_obj); + } else { + PyErr_Clear(); + use_cline = NULL; + } + } + if (!use_cline) { + c_line = 0; + (void) PyObject_SetAttr(__pyx_cython_runtime, __pyx_n_s_cline_in_traceback, Py_False); + } + else if (use_cline == Py_False || (use_cline != Py_True && PyObject_Not(use_cline) != 0)) { + c_line = 0; + } + __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback); + return c_line; +} +#endif + +/* CodeObjectCache */ + #if !CYTHON_COMPILING_IN_LIMITED_API +static int __pyx_bisect_code_objects(__Pyx_CodeObjectCacheEntry* entries, int count, int code_line) { + int start = 0, mid = 0, end = count - 1; + if (end >= 0 && code_line > entries[end].code_line) { + return count; + } + while (start < end) { + mid = start + (end - start) / 2; + if (code_line < entries[mid].code_line) { + end = mid; + } else if (code_line > entries[mid].code_line) { + start = mid + 1; + } else { + return mid; + } + } + if (code_line <= entries[mid].code_line) { + return mid; + } else { + return mid + 1; + } +} +static PyCodeObject *__pyx_find_code_object(int code_line) { + PyCodeObject* code_object; + int pos; + if (unlikely(!code_line) || unlikely(!__pyx_code_cache.entries)) { + return NULL; + } + pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line); + if (unlikely(pos >= __pyx_code_cache.count) || unlikely(__pyx_code_cache.entries[pos].code_line != code_line)) { + return NULL; + } + code_object = __pyx_code_cache.entries[pos].code_object; + Py_INCREF(code_object); + return code_object; +} +static void __pyx_insert_code_object(int code_line, PyCodeObject* code_object) { + int pos, i; + __Pyx_CodeObjectCacheEntry* entries = __pyx_code_cache.entries; + if (unlikely(!code_line)) { + return; + } + if (unlikely(!entries)) { + entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Malloc(64*sizeof(__Pyx_CodeObjectCacheEntry)); + if (likely(entries)) { + __pyx_code_cache.entries = entries; + __pyx_code_cache.max_count = 64; + __pyx_code_cache.count = 1; + entries[0].code_line = code_line; + entries[0].code_object = code_object; + Py_INCREF(code_object); + } + return; + } + pos = __pyx_bisect_code_objects(__pyx_code_cache.entries, __pyx_code_cache.count, code_line); + if ((pos < __pyx_code_cache.count) && unlikely(__pyx_code_cache.entries[pos].code_line == code_line)) { + PyCodeObject* tmp = entries[pos].code_object; + entries[pos].code_object = code_object; + Py_DECREF(tmp); + return; + } + if (__pyx_code_cache.count == __pyx_code_cache.max_count) { + int new_max = __pyx_code_cache.max_count + 64; + entries = (__Pyx_CodeObjectCacheEntry*)PyMem_Realloc( + __pyx_code_cache.entries, ((size_t)new_max) * sizeof(__Pyx_CodeObjectCacheEntry)); + if (unlikely(!entries)) { + return; + } + __pyx_code_cache.entries = entries; + __pyx_code_cache.max_count = new_max; + } + for (i=__pyx_code_cache.count; i>pos; i--) { + entries[i] = entries[i-1]; + } + entries[pos].code_line = code_line; + entries[pos].code_object = code_object; + __pyx_code_cache.count++; + Py_INCREF(code_object); +} +#endif + +/* AddTraceback */ + #include "compile.h" +#include "frameobject.h" +#include "traceback.h" +#if PY_VERSION_HEX >= 0x030b00a6 && !CYTHON_COMPILING_IN_LIMITED_API && !defined(PYPY_VERSION) + #ifndef Py_BUILD_CORE + #define Py_BUILD_CORE 1 + #endif + #include "internal/pycore_frame.h" +#endif +#if CYTHON_COMPILING_IN_LIMITED_API +static PyObject *__Pyx_PyCode_Replace_For_AddTraceback(PyObject *code, PyObject *scratch_dict, + PyObject *firstlineno, PyObject *name) { + PyObject *replace = NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "co_firstlineno", firstlineno))) return NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "co_name", name))) return NULL; + replace = PyObject_GetAttrString(code, "replace"); + if (likely(replace)) { + PyObject *result; + result = PyObject_Call(replace, __pyx_empty_tuple, scratch_dict); + Py_DECREF(replace); + return result; + } + PyErr_Clear(); + #if __PYX_LIMITED_VERSION_HEX < 0x030780000 + { + PyObject *compiled = NULL, *result = NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "code", code))) return NULL; + if (unlikely(PyDict_SetItemString(scratch_dict, "type", (PyObject*)(&PyType_Type)))) return NULL; + compiled = Py_CompileString( + "out = type(code)(\n" + " code.co_argcount, code.co_kwonlyargcount, code.co_nlocals, code.co_stacksize,\n" + " code.co_flags, code.co_code, code.co_consts, code.co_names,\n" + " code.co_varnames, code.co_filename, co_name, co_firstlineno,\n" + " code.co_lnotab)\n", "", Py_file_input); + if (!compiled) return NULL; + result = PyEval_EvalCode(compiled, scratch_dict, scratch_dict); + Py_DECREF(compiled); + if (!result) PyErr_Print(); + Py_DECREF(result); + result = PyDict_GetItemString(scratch_dict, "out"); + if (result) Py_INCREF(result); + return result; + } + #else + return NULL; + #endif +} +static void __Pyx_AddTraceback(const char *funcname, int c_line, + int py_line, const char *filename) { + PyObject *code_object = NULL, *py_py_line = NULL, *py_funcname = NULL, *dict = NULL; + PyObject *replace = NULL, *getframe = NULL, *frame = NULL; + PyObject *exc_type, *exc_value, *exc_traceback; + int success = 0; + if (c_line) { + (void) __pyx_cfilenm; + (void) __Pyx_CLineForTraceback(__Pyx_PyThreadState_Current, c_line); + } + PyErr_Fetch(&exc_type, &exc_value, &exc_traceback); + code_object = Py_CompileString("_getframe()", filename, Py_eval_input); + if (unlikely(!code_object)) goto bad; + py_py_line = PyLong_FromLong(py_line); + if (unlikely(!py_py_line)) goto bad; + py_funcname = PyUnicode_FromString(funcname); + if (unlikely(!py_funcname)) goto bad; + dict = PyDict_New(); + if (unlikely(!dict)) goto bad; + { + PyObject *old_code_object = code_object; + code_object = __Pyx_PyCode_Replace_For_AddTraceback(code_object, dict, py_py_line, py_funcname); + Py_DECREF(old_code_object); + } + if (unlikely(!code_object)) goto bad; + getframe = PySys_GetObject("_getframe"); + if (unlikely(!getframe)) goto bad; + if (unlikely(PyDict_SetItemString(dict, "_getframe", getframe))) goto bad; + frame = PyEval_EvalCode(code_object, dict, dict); + if (unlikely(!frame) || frame == Py_None) goto bad; + success = 1; + bad: + PyErr_Restore(exc_type, exc_value, exc_traceback); + Py_XDECREF(code_object); + Py_XDECREF(py_py_line); + Py_XDECREF(py_funcname); + Py_XDECREF(dict); + Py_XDECREF(replace); + if (success) { + PyTraceBack_Here( + (struct _frame*)frame); + } + Py_XDECREF(frame); +} +#else +static PyCodeObject* __Pyx_CreateCodeObjectForTraceback( + const char *funcname, int c_line, + int py_line, const char *filename) { + PyCodeObject *py_code = NULL; + PyObject *py_funcname = NULL; + #if PY_MAJOR_VERSION < 3 + PyObject *py_srcfile = NULL; + py_srcfile = PyString_FromString(filename); + if (!py_srcfile) goto bad; + #endif + if (c_line) { + #if PY_MAJOR_VERSION < 3 + py_funcname = PyString_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line); + if (!py_funcname) goto bad; + #else + py_funcname = PyUnicode_FromFormat( "%s (%s:%d)", funcname, __pyx_cfilenm, c_line); + if (!py_funcname) goto bad; + funcname = PyUnicode_AsUTF8(py_funcname); + if (!funcname) goto bad; + #endif + } + else { + #if PY_MAJOR_VERSION < 3 + py_funcname = PyString_FromString(funcname); + if (!py_funcname) goto bad; + #endif + } + #if PY_MAJOR_VERSION < 3 + py_code = __Pyx_PyCode_New( + 0, + 0, + 0, + 0, + 0, + 0, + __pyx_empty_bytes, /*PyObject *code,*/ + __pyx_empty_tuple, /*PyObject *consts,*/ + __pyx_empty_tuple, /*PyObject *names,*/ + __pyx_empty_tuple, /*PyObject *varnames,*/ + __pyx_empty_tuple, /*PyObject *freevars,*/ + __pyx_empty_tuple, /*PyObject *cellvars,*/ + py_srcfile, /*PyObject *filename,*/ + py_funcname, /*PyObject *name,*/ + py_line, + __pyx_empty_bytes /*PyObject *lnotab*/ + ); + Py_DECREF(py_srcfile); + #else + py_code = PyCode_NewEmpty(filename, funcname, py_line); + #endif + Py_XDECREF(py_funcname); + return py_code; +bad: + Py_XDECREF(py_funcname); + #if PY_MAJOR_VERSION < 3 + Py_XDECREF(py_srcfile); + #endif + return NULL; +} +static void __Pyx_AddTraceback(const char *funcname, int c_line, + int py_line, const char *filename) { + PyCodeObject *py_code = 0; + PyFrameObject *py_frame = 0; + PyThreadState *tstate = __Pyx_PyThreadState_Current; + PyObject *ptype, *pvalue, *ptraceback; + if (c_line) { + c_line = __Pyx_CLineForTraceback(tstate, c_line); + } + py_code = __pyx_find_code_object(c_line ? -c_line : py_line); + if (!py_code) { + __Pyx_ErrFetchInState(tstate, &ptype, &pvalue, &ptraceback); + py_code = __Pyx_CreateCodeObjectForTraceback( + funcname, c_line, py_line, filename); + if (!py_code) { + /* If the code object creation fails, then we should clear the + fetched exception references and propagate the new exception */ + Py_XDECREF(ptype); + Py_XDECREF(pvalue); + Py_XDECREF(ptraceback); + goto bad; + } + __Pyx_ErrRestoreInState(tstate, ptype, pvalue, ptraceback); + __pyx_insert_code_object(c_line ? -c_line : py_line, py_code); + } + py_frame = PyFrame_New( + tstate, /*PyThreadState *tstate,*/ + py_code, /*PyCodeObject *code,*/ + __pyx_d, /*PyObject *globals,*/ + 0 /*PyObject *locals*/ + ); + if (!py_frame) goto bad; + __Pyx_PyFrame_SetLineNumber(py_frame, py_line); + PyTraceBack_Here(py_frame); +bad: + Py_XDECREF(py_code); + Py_XDECREF(py_frame); +} +#endif + +#if PY_MAJOR_VERSION < 3 +static int __Pyx_GetBuffer(PyObject *obj, Py_buffer *view, int flags) { + __Pyx_TypeName obj_type_name; + if (PyObject_CheckBuffer(obj)) return PyObject_GetBuffer(obj, view, flags); + if (__Pyx_TypeCheck(obj, __pyx_array_type)) return __pyx_array_getbuffer(obj, view, flags); + if (__Pyx_TypeCheck(obj, __pyx_memoryview_type)) return __pyx_memoryview_getbuffer(obj, view, flags); + obj_type_name = __Pyx_PyType_GetName(Py_TYPE(obj)); + PyErr_Format(PyExc_TypeError, + "'" __Pyx_FMT_TYPENAME "' does not have the buffer interface", + obj_type_name); + __Pyx_DECREF_TypeName(obj_type_name); + return -1; +} +static void __Pyx_ReleaseBuffer(Py_buffer *view) { + PyObject *obj = view->obj; + if (!obj) return; + if (PyObject_CheckBuffer(obj)) { + PyBuffer_Release(view); + return; + } + if ((0)) {} + view->obj = NULL; + Py_DECREF(obj); +} +#endif + + + /* MemviewSliceIsContig */ + static int +__pyx_memviewslice_is_contig(const __Pyx_memviewslice mvs, char order, int ndim) +{ + int i, index, step, start; + Py_ssize_t itemsize = mvs.memview->view.itemsize; + if (order == 'F') { + step = 1; + start = 0; + } else { + step = -1; + start = ndim - 1; + } + for (i = 0; i < ndim; i++) { + index = start + step * i; + if (mvs.suboffsets[index] >= 0 || mvs.strides[index] != itemsize) + return 0; + itemsize *= mvs.shape[index]; + } + return 1; +} + +/* OverlappingSlices */ + static void +__pyx_get_array_memory_extents(__Pyx_memviewslice *slice, + void **out_start, void **out_end, + int ndim, size_t itemsize) +{ + char *start, *end; + int i; + start = end = slice->data; + for (i = 0; i < ndim; i++) { + Py_ssize_t stride = slice->strides[i]; + Py_ssize_t extent = slice->shape[i]; + if (extent == 0) { + *out_start = *out_end = start; + return; + } else { + if (stride > 0) + end += stride * (extent - 1); + else + start += stride * (extent - 1); + } + } + *out_start = start; + *out_end = end + itemsize; +} +static int +__pyx_slices_overlap(__Pyx_memviewslice *slice1, + __Pyx_memviewslice *slice2, + int ndim, size_t itemsize) +{ + void *start1, *end1, *start2, *end2; + __pyx_get_array_memory_extents(slice1, &start1, &end1, ndim, itemsize); + __pyx_get_array_memory_extents(slice2, &start2, &end2, ndim, itemsize); + return (start1 < end2) && (start2 < end1); +} + +/* CIntFromPyVerify */ + #define __PYX_VERIFY_RETURN_INT(target_type, func_type, func_value)\ + __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 0) +#define __PYX_VERIFY_RETURN_INT_EXC(target_type, func_type, func_value)\ + __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, 1) +#define __PYX__VERIFY_RETURN_INT(target_type, func_type, func_value, exc)\ + {\ + func_type value = func_value;\ + if (sizeof(target_type) < sizeof(func_type)) {\ + if (unlikely(value != (func_type) (target_type) value)) {\ + func_type zero = 0;\ + if (exc && unlikely(value == (func_type)-1 && PyErr_Occurred()))\ + return (target_type) -1;\ + if (is_unsigned && unlikely(value < zero))\ + goto raise_neg_overflow;\ + else\ + goto raise_overflow;\ + }\ + }\ + return (target_type) value;\ + } + +/* MemviewDtypeToObject */ + static CYTHON_INLINE PyObject *__pyx_memview_get_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(const char *itemp) { + return (PyObject *) __Pyx_PyInt_From_npy_int64(*(__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) itemp); +} +static CYTHON_INLINE int __pyx_memview_set_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(const char *itemp, PyObject *obj) { + __pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t value = __Pyx_PyInt_As_npy_int64(obj); + if (unlikely((value == ((npy_int64)-1)) && PyErr_Occurred())) + return 0; + *(__pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t *) itemp = value; + return 1; +} + +/* TypeInfoCompare */ + static int +__pyx_typeinfo_cmp(__Pyx_TypeInfo *a, __Pyx_TypeInfo *b) +{ + int i; + if (!a || !b) + return 0; + if (a == b) + return 1; + if (a->size != b->size || a->typegroup != b->typegroup || + a->is_unsigned != b->is_unsigned || a->ndim != b->ndim) { + if (a->typegroup == 'H' || b->typegroup == 'H') { + return a->size == b->size; + } else { + return 0; + } + } + if (a->ndim) { + for (i = 0; i < a->ndim; i++) + if (a->arraysize[i] != b->arraysize[i]) + return 0; + } + if (a->typegroup == 'S') { + if (a->flags != b->flags) + return 0; + if (a->fields || b->fields) { + if (!(a->fields && b->fields)) + return 0; + for (i = 0; a->fields[i].type && b->fields[i].type; i++) { + __Pyx_StructField *field_a = a->fields + i; + __Pyx_StructField *field_b = b->fields + i; + if (field_a->offset != field_b->offset || + !__pyx_typeinfo_cmp(field_a->type, field_b->type)) + return 0; + } + return !a->fields[i].type && !b->fields[i].type; + } + } + return 1; +} + +/* MemviewSliceValidateAndInit */ + static int +__pyx_check_strides(Py_buffer *buf, int dim, int ndim, int spec) +{ + if (buf->shape[dim] <= 1) + return 1; + if (buf->strides) { + if (spec & __Pyx_MEMVIEW_CONTIG) { + if (spec & (__Pyx_MEMVIEW_PTR|__Pyx_MEMVIEW_FULL)) { + if (unlikely(buf->strides[dim] != sizeof(void *))) { + PyErr_Format(PyExc_ValueError, + "Buffer is not indirectly contiguous " + "in dimension %d.", dim); + goto fail; + } + } else if (unlikely(buf->strides[dim] != buf->itemsize)) { + PyErr_SetString(PyExc_ValueError, + "Buffer and memoryview are not contiguous " + "in the same dimension."); + goto fail; + } + } + if (spec & __Pyx_MEMVIEW_FOLLOW) { + Py_ssize_t stride = buf->strides[dim]; + if (stride < 0) + stride = -stride; + if (unlikely(stride < buf->itemsize)) { + PyErr_SetString(PyExc_ValueError, + "Buffer and memoryview are not contiguous " + "in the same dimension."); + goto fail; + } + } + } else { + if (unlikely(spec & __Pyx_MEMVIEW_CONTIG && dim != ndim - 1)) { + PyErr_Format(PyExc_ValueError, + "C-contiguous buffer is not contiguous in " + "dimension %d", dim); + goto fail; + } else if (unlikely(spec & (__Pyx_MEMVIEW_PTR))) { + PyErr_Format(PyExc_ValueError, + "C-contiguous buffer is not indirect in " + "dimension %d", dim); + goto fail; + } else if (unlikely(buf->suboffsets)) { + PyErr_SetString(PyExc_ValueError, + "Buffer exposes suboffsets but no strides"); + goto fail; + } + } + return 1; +fail: + return 0; +} +static int +__pyx_check_suboffsets(Py_buffer *buf, int dim, int ndim, int spec) +{ + CYTHON_UNUSED_VAR(ndim); + if (spec & __Pyx_MEMVIEW_DIRECT) { + if (unlikely(buf->suboffsets && buf->suboffsets[dim] >= 0)) { + PyErr_Format(PyExc_ValueError, + "Buffer not compatible with direct access " + "in dimension %d.", dim); + goto fail; + } + } + if (spec & __Pyx_MEMVIEW_PTR) { + if (unlikely(!buf->suboffsets || (buf->suboffsets[dim] < 0))) { + PyErr_Format(PyExc_ValueError, + "Buffer is not indirectly accessible " + "in dimension %d.", dim); + goto fail; + } + } + return 1; +fail: + return 0; +} +static int +__pyx_verify_contig(Py_buffer *buf, int ndim, int c_or_f_flag) +{ + int i; + if (c_or_f_flag & __Pyx_IS_F_CONTIG) { + Py_ssize_t stride = 1; + for (i = 0; i < ndim; i++) { + if (unlikely(stride * buf->itemsize != buf->strides[i] && buf->shape[i] > 1)) { + PyErr_SetString(PyExc_ValueError, + "Buffer not fortran contiguous."); + goto fail; + } + stride = stride * buf->shape[i]; + } + } else if (c_or_f_flag & __Pyx_IS_C_CONTIG) { + Py_ssize_t stride = 1; + for (i = ndim - 1; i >- 1; i--) { + if (unlikely(stride * buf->itemsize != buf->strides[i] && buf->shape[i] > 1)) { + PyErr_SetString(PyExc_ValueError, + "Buffer not C contiguous."); + goto fail; + } + stride = stride * buf->shape[i]; + } + } + return 1; +fail: + return 0; +} +static int __Pyx_ValidateAndInit_memviewslice( + int *axes_specs, + int c_or_f_flag, + int buf_flags, + int ndim, + __Pyx_TypeInfo *dtype, + __Pyx_BufFmt_StackElem stack[], + __Pyx_memviewslice *memviewslice, + PyObject *original_obj) +{ + struct __pyx_memoryview_obj *memview, *new_memview; + __Pyx_RefNannyDeclarations + Py_buffer *buf; + int i, spec = 0, retval = -1; + __Pyx_BufFmt_Context ctx; + int from_memoryview = __pyx_memoryview_check(original_obj); + __Pyx_RefNannySetupContext("ValidateAndInit_memviewslice", 0); + if (from_memoryview && __pyx_typeinfo_cmp(dtype, ((struct __pyx_memoryview_obj *) + original_obj)->typeinfo)) { + memview = (struct __pyx_memoryview_obj *) original_obj; + new_memview = NULL; + } else { + memview = (struct __pyx_memoryview_obj *) __pyx_memoryview_new( + original_obj, buf_flags, 0, dtype); + new_memview = memview; + if (unlikely(!memview)) + goto fail; + } + buf = &memview->view; + if (unlikely(buf->ndim != ndim)) { + PyErr_Format(PyExc_ValueError, + "Buffer has wrong number of dimensions (expected %d, got %d)", + ndim, buf->ndim); + goto fail; + } + if (new_memview) { + __Pyx_BufFmt_Init(&ctx, stack, dtype); + if (unlikely(!__Pyx_BufFmt_CheckString(&ctx, buf->format))) goto fail; + } + if (unlikely((unsigned) buf->itemsize != dtype->size)) { + PyErr_Format(PyExc_ValueError, + "Item size of buffer (%" CYTHON_FORMAT_SSIZE_T "u byte%s) " + "does not match size of '%s' (%" CYTHON_FORMAT_SSIZE_T "u byte%s)", + buf->itemsize, + (buf->itemsize > 1) ? "s" : "", + dtype->name, + dtype->size, + (dtype->size > 1) ? "s" : ""); + goto fail; + } + if (buf->len > 0) { + for (i = 0; i < ndim; i++) { + spec = axes_specs[i]; + if (unlikely(!__pyx_check_strides(buf, i, ndim, spec))) + goto fail; + if (unlikely(!__pyx_check_suboffsets(buf, i, ndim, spec))) + goto fail; + } + if (unlikely(buf->strides && !__pyx_verify_contig(buf, ndim, c_or_f_flag))) + goto fail; + } + if (unlikely(__Pyx_init_memviewslice(memview, ndim, memviewslice, + new_memview != NULL) == -1)) { + goto fail; + } + retval = 0; + goto no_fail; +fail: + Py_XDECREF(new_memview); + retval = -1; +no_fail: + __Pyx_RefNannyFinishContext(); + return retval; +} + +/* ObjectToMemviewSlice */ + static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_ds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(PyObject *obj, int writable_flag) { + __Pyx_memviewslice result = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_BufFmt_StackElem stack[1]; + int axes_specs[] = { (__Pyx_MEMVIEW_DIRECT | __Pyx_MEMVIEW_STRIDED) }; + int retcode; + if (obj == Py_None) { + result.memview = (struct __pyx_memoryview_obj *) Py_None; + return result; + } + retcode = __Pyx_ValidateAndInit_memviewslice(axes_specs, 0, + PyBUF_RECORDS_RO | writable_flag, 1, + &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, stack, + &result, obj); + if (unlikely(retcode == -1)) + goto __pyx_fail; + return result; +__pyx_fail: + result.memview = NULL; + result.data = NULL; + return result; +} + +/* ObjectToMemviewSlice */ + static CYTHON_INLINE __Pyx_memviewslice __Pyx_PyObject_to_MemoryviewSlice_dsds_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t(PyObject *obj, int writable_flag) { + __Pyx_memviewslice result = { 0, 0, { 0 }, { 0 }, { 0 } }; + __Pyx_BufFmt_StackElem stack[1]; + int axes_specs[] = { (__Pyx_MEMVIEW_DIRECT | __Pyx_MEMVIEW_STRIDED), (__Pyx_MEMVIEW_DIRECT | __Pyx_MEMVIEW_STRIDED) }; + int retcode; + if (obj == Py_None) { + result.memview = (struct __pyx_memoryview_obj *) Py_None; + return result; + } + retcode = __Pyx_ValidateAndInit_memviewslice(axes_specs, 0, + PyBUF_RECORDS_RO | writable_flag, 2, + &__Pyx_TypeInfo_nn___pyx_t_7fairseq_4data_22token_block_utils_fast_DTYPE_t, stack, + &result, obj); + if (unlikely(retcode == -1)) + goto __pyx_fail; + return result; +__pyx_fail: + result.memview = NULL; + result.data = NULL; + return result; +} + +/* Declarations */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) { + return ::std::complex< float >(x, y); + } + #else + static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) { + return x + y*(__pyx_t_float_complex)_Complex_I; + } + #endif +#else + static CYTHON_INLINE __pyx_t_float_complex __pyx_t_float_complex_from_parts(float x, float y) { + __pyx_t_float_complex z; + z.real = x; + z.imag = y; + return z; + } +#endif + +/* Arithmetic */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) +#else + static CYTHON_INLINE int __Pyx_c_eq_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + return (a.real == b.real) && (a.imag == b.imag); + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_sum_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + z.real = a.real + b.real; + z.imag = a.imag + b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_diff_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + z.real = a.real - b.real; + z.imag = a.imag - b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_prod_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + z.real = a.real * b.real - a.imag * b.imag; + z.imag = a.real * b.imag + a.imag * b.real; + return z; + } + #if 1 + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + if (b.imag == 0) { + return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real); + } else if (fabsf(b.real) >= fabsf(b.imag)) { + if (b.real == 0 && b.imag == 0) { + return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.imag); + } else { + float r = b.imag / b.real; + float s = (float)(1.0) / (b.real + b.imag * r); + return __pyx_t_float_complex_from_parts( + (a.real + a.imag * r) * s, (a.imag - a.real * r) * s); + } + } else { + float r = b.real / b.imag; + float s = (float)(1.0) / (b.imag + b.real * r); + return __pyx_t_float_complex_from_parts( + (a.real * r + a.imag) * s, (a.imag * r - a.real) * s); + } + } + #else + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_quot_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + if (b.imag == 0) { + return __pyx_t_float_complex_from_parts(a.real / b.real, a.imag / b.real); + } else { + float denom = b.real * b.real + b.imag * b.imag; + return __pyx_t_float_complex_from_parts( + (a.real * b.real + a.imag * b.imag) / denom, + (a.imag * b.real - a.real * b.imag) / denom); + } + } + #endif + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_neg_float(__pyx_t_float_complex a) { + __pyx_t_float_complex z; + z.real = -a.real; + z.imag = -a.imag; + return z; + } + static CYTHON_INLINE int __Pyx_c_is_zero_float(__pyx_t_float_complex a) { + return (a.real == 0) && (a.imag == 0); + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_conj_float(__pyx_t_float_complex a) { + __pyx_t_float_complex z; + z.real = a.real; + z.imag = -a.imag; + return z; + } + #if 1 + static CYTHON_INLINE float __Pyx_c_abs_float(__pyx_t_float_complex z) { + #if !defined(HAVE_HYPOT) || defined(_MSC_VER) + return sqrtf(z.real*z.real + z.imag*z.imag); + #else + return hypotf(z.real, z.imag); + #endif + } + static CYTHON_INLINE __pyx_t_float_complex __Pyx_c_pow_float(__pyx_t_float_complex a, __pyx_t_float_complex b) { + __pyx_t_float_complex z; + float r, lnr, theta, z_r, z_theta; + if (b.imag == 0 && b.real == (int)b.real) { + if (b.real < 0) { + float denom = a.real * a.real + a.imag * a.imag; + a.real = a.real / denom; + a.imag = -a.imag / denom; + b.real = -b.real; + } + switch ((int)b.real) { + case 0: + z.real = 1; + z.imag = 0; + return z; + case 1: + return a; + case 2: + return __Pyx_c_prod_float(a, a); + case 3: + z = __Pyx_c_prod_float(a, a); + return __Pyx_c_prod_float(z, a); + case 4: + z = __Pyx_c_prod_float(a, a); + return __Pyx_c_prod_float(z, z); + } + } + if (a.imag == 0) { + if (a.real == 0) { + return a; + } else if ((b.imag == 0) && (a.real >= 0)) { + z.real = powf(a.real, b.real); + z.imag = 0; + return z; + } else if (a.real > 0) { + r = a.real; + theta = 0; + } else { + r = -a.real; + theta = atan2f(0.0, -1.0); + } + } else { + r = __Pyx_c_abs_float(a); + theta = atan2f(a.imag, a.real); + } + lnr = logf(r); + z_r = expf(lnr * b.real - theta * b.imag); + z_theta = theta * b.real + lnr * b.imag; + z.real = z_r * cosf(z_theta); + z.imag = z_r * sinf(z_theta); + return z; + } + #endif +#endif + +/* Declarations */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) { + return ::std::complex< double >(x, y); + } + #else + static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) { + return x + y*(__pyx_t_double_complex)_Complex_I; + } + #endif +#else + static CYTHON_INLINE __pyx_t_double_complex __pyx_t_double_complex_from_parts(double x, double y) { + __pyx_t_double_complex z; + z.real = x; + z.imag = y; + return z; + } +#endif + +/* Arithmetic */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) +#else + static CYTHON_INLINE int __Pyx_c_eq_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + return (a.real == b.real) && (a.imag == b.imag); + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_sum_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + z.real = a.real + b.real; + z.imag = a.imag + b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_diff_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + z.real = a.real - b.real; + z.imag = a.imag - b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_prod_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + z.real = a.real * b.real - a.imag * b.imag; + z.imag = a.real * b.imag + a.imag * b.real; + return z; + } + #if 1 + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + if (b.imag == 0) { + return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else if (fabs(b.real) >= fabs(b.imag)) { + if (b.real == 0 && b.imag == 0) { + return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.imag); + } else { + double r = b.imag / b.real; + double s = (double)(1.0) / (b.real + b.imag * r); + return __pyx_t_double_complex_from_parts( + (a.real + a.imag * r) * s, (a.imag - a.real * r) * s); + } + } else { + double r = b.real / b.imag; + double s = (double)(1.0) / (b.imag + b.real * r); + return __pyx_t_double_complex_from_parts( + (a.real * r + a.imag) * s, (a.imag * r - a.real) * s); + } + } + #else + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_quot_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + if (b.imag == 0) { + return __pyx_t_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else { + double denom = b.real * b.real + b.imag * b.imag; + return __pyx_t_double_complex_from_parts( + (a.real * b.real + a.imag * b.imag) / denom, + (a.imag * b.real - a.real * b.imag) / denom); + } + } + #endif + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_neg_double(__pyx_t_double_complex a) { + __pyx_t_double_complex z; + z.real = -a.real; + z.imag = -a.imag; + return z; + } + static CYTHON_INLINE int __Pyx_c_is_zero_double(__pyx_t_double_complex a) { + return (a.real == 0) && (a.imag == 0); + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_conj_double(__pyx_t_double_complex a) { + __pyx_t_double_complex z; + z.real = a.real; + z.imag = -a.imag; + return z; + } + #if 1 + static CYTHON_INLINE double __Pyx_c_abs_double(__pyx_t_double_complex z) { + #if !defined(HAVE_HYPOT) || defined(_MSC_VER) + return sqrt(z.real*z.real + z.imag*z.imag); + #else + return hypot(z.real, z.imag); + #endif + } + static CYTHON_INLINE __pyx_t_double_complex __Pyx_c_pow_double(__pyx_t_double_complex a, __pyx_t_double_complex b) { + __pyx_t_double_complex z; + double r, lnr, theta, z_r, z_theta; + if (b.imag == 0 && b.real == (int)b.real) { + if (b.real < 0) { + double denom = a.real * a.real + a.imag * a.imag; + a.real = a.real / denom; + a.imag = -a.imag / denom; + b.real = -b.real; + } + switch ((int)b.real) { + case 0: + z.real = 1; + z.imag = 0; + return z; + case 1: + return a; + case 2: + return __Pyx_c_prod_double(a, a); + case 3: + z = __Pyx_c_prod_double(a, a); + return __Pyx_c_prod_double(z, a); + case 4: + z = __Pyx_c_prod_double(a, a); + return __Pyx_c_prod_double(z, z); + } + } + if (a.imag == 0) { + if (a.real == 0) { + return a; + } else if ((b.imag == 0) && (a.real >= 0)) { + z.real = pow(a.real, b.real); + z.imag = 0; + return z; + } else if (a.real > 0) { + r = a.real; + theta = 0; + } else { + r = -a.real; + theta = atan2(0.0, -1.0); + } + } else { + r = __Pyx_c_abs_double(a); + theta = atan2(a.imag, a.real); + } + lnr = log(r); + z_r = exp(lnr * b.real - theta * b.imag); + z_theta = theta * b.real + lnr * b.imag; + z.real = z_r * cos(z_theta); + z.imag = z_r * sin(z_theta); + return z; + } + #endif +#endif + +/* Declarations */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) + #ifdef __cplusplus + static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double x, long double y) { + return ::std::complex< long double >(x, y); + } + #else + static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double x, long double y) { + return x + y*(__pyx_t_long_double_complex)_Complex_I; + } + #endif +#else + static CYTHON_INLINE __pyx_t_long_double_complex __pyx_t_long_double_complex_from_parts(long double x, long double y) { + __pyx_t_long_double_complex z; + z.real = x; + z.imag = y; + return z; + } +#endif + +/* Arithmetic */ + #if CYTHON_CCOMPLEX && (1) && (!0 || __cplusplus) +#else + static CYTHON_INLINE int __Pyx_c_eq_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + return (a.real == b.real) && (a.imag == b.imag); + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_sum_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + z.real = a.real + b.real; + z.imag = a.imag + b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_diff_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + z.real = a.real - b.real; + z.imag = a.imag - b.imag; + return z; + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_prod_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + z.real = a.real * b.real - a.imag * b.imag; + z.imag = a.real * b.imag + a.imag * b.real; + return z; + } + #if 1 + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_quot_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + if (b.imag == 0) { + return __pyx_t_long_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else if (fabsl(b.real) >= fabsl(b.imag)) { + if (b.real == 0 && b.imag == 0) { + return __pyx_t_long_double_complex_from_parts(a.real / b.real, a.imag / b.imag); + } else { + long double r = b.imag / b.real; + long double s = (long double)(1.0) / (b.real + b.imag * r); + return __pyx_t_long_double_complex_from_parts( + (a.real + a.imag * r) * s, (a.imag - a.real * r) * s); + } + } else { + long double r = b.real / b.imag; + long double s = (long double)(1.0) / (b.imag + b.real * r); + return __pyx_t_long_double_complex_from_parts( + (a.real * r + a.imag) * s, (a.imag * r - a.real) * s); + } + } + #else + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_quot_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + if (b.imag == 0) { + return __pyx_t_long_double_complex_from_parts(a.real / b.real, a.imag / b.real); + } else { + long double denom = b.real * b.real + b.imag * b.imag; + return __pyx_t_long_double_complex_from_parts( + (a.real * b.real + a.imag * b.imag) / denom, + (a.imag * b.real - a.real * b.imag) / denom); + } + } + #endif + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_neg_long__double(__pyx_t_long_double_complex a) { + __pyx_t_long_double_complex z; + z.real = -a.real; + z.imag = -a.imag; + return z; + } + static CYTHON_INLINE int __Pyx_c_is_zero_long__double(__pyx_t_long_double_complex a) { + return (a.real == 0) && (a.imag == 0); + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_conj_long__double(__pyx_t_long_double_complex a) { + __pyx_t_long_double_complex z; + z.real = a.real; + z.imag = -a.imag; + return z; + } + #if 1 + static CYTHON_INLINE long double __Pyx_c_abs_long__double(__pyx_t_long_double_complex z) { + #if !defined(HAVE_HYPOT) || defined(_MSC_VER) + return sqrtl(z.real*z.real + z.imag*z.imag); + #else + return hypotl(z.real, z.imag); + #endif + } + static CYTHON_INLINE __pyx_t_long_double_complex __Pyx_c_pow_long__double(__pyx_t_long_double_complex a, __pyx_t_long_double_complex b) { + __pyx_t_long_double_complex z; + long double r, lnr, theta, z_r, z_theta; + if (b.imag == 0 && b.real == (int)b.real) { + if (b.real < 0) { + long double denom = a.real * a.real + a.imag * a.imag; + a.real = a.real / denom; + a.imag = -a.imag / denom; + b.real = -b.real; + } + switch ((int)b.real) { + case 0: + z.real = 1; + z.imag = 0; + return z; + case 1: + return a; + case 2: + return __Pyx_c_prod_long__double(a, a); + case 3: + z = __Pyx_c_prod_long__double(a, a); + return __Pyx_c_prod_long__double(z, a); + case 4: + z = __Pyx_c_prod_long__double(a, a); + return __Pyx_c_prod_long__double(z, z); + } + } + if (a.imag == 0) { + if (a.real == 0) { + return a; + } else if ((b.imag == 0) && (a.real >= 0)) { + z.real = powl(a.real, b.real); + z.imag = 0; + return z; + } else if (a.real > 0) { + r = a.real; + theta = 0; + } else { + r = -a.real; + theta = atan2l(0.0, -1.0); + } + } else { + r = __Pyx_c_abs_long__double(a); + theta = atan2l(a.imag, a.real); + } + lnr = logl(r); + z_r = expl(lnr * b.real - theta * b.imag); + z_theta = theta * b.real + lnr * b.imag; + z.real = z_r * cosl(z_theta); + z.imag = z_r * sinl(z_theta); + return z; + } + #endif +#endif + +/* MemviewSliceCopyTemplate */ + static __Pyx_memviewslice +__pyx_memoryview_copy_new_contig(const __Pyx_memviewslice *from_mvs, + const char *mode, int ndim, + size_t sizeof_dtype, int contig_flag, + int dtype_is_object) +{ + __Pyx_RefNannyDeclarations + int i; + __Pyx_memviewslice new_mvs = { 0, 0, { 0 }, { 0 }, { 0 } }; + struct __pyx_memoryview_obj *from_memview = from_mvs->memview; + Py_buffer *buf = &from_memview->view; + PyObject *shape_tuple = NULL; + PyObject *temp_int = NULL; + struct __pyx_array_obj *array_obj = NULL; + struct __pyx_memoryview_obj *memview_obj = NULL; + __Pyx_RefNannySetupContext("__pyx_memoryview_copy_new_contig", 0); + for (i = 0; i < ndim; i++) { + if (unlikely(from_mvs->suboffsets[i] >= 0)) { + PyErr_Format(PyExc_ValueError, "Cannot copy memoryview slice with " + "indirect dimensions (axis %d)", i); + goto fail; + } + } + shape_tuple = PyTuple_New(ndim); + if (unlikely(!shape_tuple)) { + goto fail; + } + __Pyx_GOTREF(shape_tuple); + for(i = 0; i < ndim; i++) { + temp_int = PyInt_FromSsize_t(from_mvs->shape[i]); + if(unlikely(!temp_int)) { + goto fail; + } else { + PyTuple_SET_ITEM(shape_tuple, i, temp_int); + temp_int = NULL; + } + } + array_obj = __pyx_array_new(shape_tuple, sizeof_dtype, buf->format, (char *) mode, NULL); + if (unlikely(!array_obj)) { + goto fail; + } + __Pyx_GOTREF(array_obj); + memview_obj = (struct __pyx_memoryview_obj *) __pyx_memoryview_new( + (PyObject *) array_obj, contig_flag, + dtype_is_object, + from_mvs->memview->typeinfo); + if (unlikely(!memview_obj)) + goto fail; + if (unlikely(__Pyx_init_memviewslice(memview_obj, ndim, &new_mvs, 1) < 0)) + goto fail; + if (unlikely(__pyx_memoryview_copy_contents(*from_mvs, new_mvs, ndim, ndim, + dtype_is_object) < 0)) + goto fail; + goto no_fail; +fail: + __Pyx_XDECREF(new_mvs.memview); + new_mvs.memview = NULL; + new_mvs.data = NULL; +no_fail: + __Pyx_XDECREF(shape_tuple); + __Pyx_XDECREF(temp_int); + __Pyx_XDECREF(array_obj); + __Pyx_RefNannyFinishContext(); + return new_mvs; +} + +/* MemviewSliceInit */ + static int +__Pyx_init_memviewslice(struct __pyx_memoryview_obj *memview, + int ndim, + __Pyx_memviewslice *memviewslice, + int memview_is_new_reference) +{ + __Pyx_RefNannyDeclarations + int i, retval=-1; + Py_buffer *buf = &memview->view; + __Pyx_RefNannySetupContext("init_memviewslice", 0); + if (unlikely(memviewslice->memview || memviewslice->data)) { + PyErr_SetString(PyExc_ValueError, + "memviewslice is already initialized!"); + goto fail; + } + if (buf->strides) { + for (i = 0; i < ndim; i++) { + memviewslice->strides[i] = buf->strides[i]; + } + } else { + Py_ssize_t stride = buf->itemsize; + for (i = ndim - 1; i >= 0; i--) { + memviewslice->strides[i] = stride; + stride *= buf->shape[i]; + } + } + for (i = 0; i < ndim; i++) { + memviewslice->shape[i] = buf->shape[i]; + if (buf->suboffsets) { + memviewslice->suboffsets[i] = buf->suboffsets[i]; + } else { + memviewslice->suboffsets[i] = -1; + } + } + memviewslice->memview = memview; + memviewslice->data = (char *)buf->buf; + if (__pyx_add_acquisition_count(memview) == 0 && !memview_is_new_reference) { + Py_INCREF(memview); + } + retval = 0; + goto no_fail; +fail: + memviewslice->memview = 0; + memviewslice->data = 0; + retval = -1; +no_fail: + __Pyx_RefNannyFinishContext(); + return retval; +} +#ifndef Py_NO_RETURN +#define Py_NO_RETURN +#endif +static void __pyx_fatalerror(const char *fmt, ...) Py_NO_RETURN { + va_list vargs; + char msg[200]; +#if PY_VERSION_HEX >= 0x030A0000 || defined(HAVE_STDARG_PROTOTYPES) + va_start(vargs, fmt); +#else + va_start(vargs); +#endif + vsnprintf(msg, 200, fmt, vargs); + va_end(vargs); + Py_FatalError(msg); +} +static CYTHON_INLINE int +__pyx_add_acquisition_count_locked(__pyx_atomic_int_type *acquisition_count, + PyThread_type_lock lock) +{ + int result; + PyThread_acquire_lock(lock, 1); + result = (*acquisition_count)++; + PyThread_release_lock(lock); + return result; +} +static CYTHON_INLINE int +__pyx_sub_acquisition_count_locked(__pyx_atomic_int_type *acquisition_count, + PyThread_type_lock lock) +{ + int result; + PyThread_acquire_lock(lock, 1); + result = (*acquisition_count)--; + PyThread_release_lock(lock); + return result; +} +static CYTHON_INLINE void +__Pyx_INC_MEMVIEW(__Pyx_memviewslice *memslice, int have_gil, int lineno) +{ + __pyx_nonatomic_int_type old_acquisition_count; + struct __pyx_memoryview_obj *memview = memslice->memview; + if (unlikely(!memview || (PyObject *) memview == Py_None)) { + return; + } + old_acquisition_count = __pyx_add_acquisition_count(memview); + if (unlikely(old_acquisition_count <= 0)) { + if (likely(old_acquisition_count == 0)) { + if (have_gil) { + Py_INCREF((PyObject *) memview); + } else { + PyGILState_STATE _gilstate = PyGILState_Ensure(); + Py_INCREF((PyObject *) memview); + PyGILState_Release(_gilstate); + } + } else { + __pyx_fatalerror("Acquisition count is %d (line %d)", + old_acquisition_count+1, lineno); + } + } +} +static CYTHON_INLINE void __Pyx_XCLEAR_MEMVIEW(__Pyx_memviewslice *memslice, + int have_gil, int lineno) { + __pyx_nonatomic_int_type old_acquisition_count; + struct __pyx_memoryview_obj *memview = memslice->memview; + if (unlikely(!memview || (PyObject *) memview == Py_None)) { + memslice->memview = NULL; + return; + } + old_acquisition_count = __pyx_sub_acquisition_count(memview); + memslice->data = NULL; + if (likely(old_acquisition_count > 1)) { + memslice->memview = NULL; + } else if (likely(old_acquisition_count == 1)) { + if (have_gil) { + Py_CLEAR(memslice->memview); + } else { + PyGILState_STATE _gilstate = PyGILState_Ensure(); + Py_CLEAR(memslice->memview); + PyGILState_Release(_gilstate); + } + } else { + __pyx_fatalerror("Acquisition count is %d (line %d)", + old_acquisition_count-1, lineno); + } +} + +/* CIntToPy */ + static CYTHON_INLINE PyObject* __Pyx_PyInt_From_npy_int64(npy_int64 value) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const npy_int64 neg_one = (npy_int64) -1, const_zero = (npy_int64) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (is_unsigned) { + if (sizeof(npy_int64) < sizeof(long)) { + return PyInt_FromLong((long) value); + } else if (sizeof(npy_int64) <= sizeof(unsigned long)) { + return PyLong_FromUnsignedLong((unsigned long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(npy_int64) <= sizeof(unsigned PY_LONG_LONG)) { + return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value); +#endif + } + } else { + if (sizeof(npy_int64) <= sizeof(long)) { + return PyInt_FromLong((long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(npy_int64) <= sizeof(PY_LONG_LONG)) { + return PyLong_FromLongLong((PY_LONG_LONG) value); +#endif + } + } + { + unsigned char *bytes = (unsigned char *)&value; +#if !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX >= 0x030d00A4 + if (is_unsigned) { + return PyLong_FromUnsignedNativeBytes(bytes, sizeof(value), -1); + } else { + return PyLong_FromNativeBytes(bytes, sizeof(value), -1); + } +#elif !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX < 0x030d0000 + int one = 1; int little = (int)*(unsigned char *)&one; + return _PyLong_FromByteArray(bytes, sizeof(npy_int64), + little, !is_unsigned); +#else + int one = 1; int little = (int)*(unsigned char *)&one; + PyObject *from_bytes, *result = NULL; + PyObject *py_bytes = NULL, *arg_tuple = NULL, *kwds = NULL, *order_str = NULL; + from_bytes = PyObject_GetAttrString((PyObject*)&PyLong_Type, "from_bytes"); + if (!from_bytes) return NULL; + py_bytes = PyBytes_FromStringAndSize((char*)bytes, sizeof(npy_int64)); + if (!py_bytes) goto limited_bad; + order_str = PyUnicode_FromString(little ? "little" : "big"); + if (!order_str) goto limited_bad; + arg_tuple = PyTuple_Pack(2, py_bytes, order_str); + if (!arg_tuple) goto limited_bad; + if (!is_unsigned) { + kwds = PyDict_New(); + if (!kwds) goto limited_bad; + if (PyDict_SetItemString(kwds, "signed", __Pyx_NewRef(Py_True))) goto limited_bad; + } + result = PyObject_Call(from_bytes, arg_tuple, kwds); + limited_bad: + Py_XDECREF(kwds); + Py_XDECREF(arg_tuple); + Py_XDECREF(order_str); + Py_XDECREF(py_bytes); + Py_XDECREF(from_bytes); + return result; +#endif + } +} + +/* CIntFromPy */ + static CYTHON_INLINE npy_int64 __Pyx_PyInt_As_npy_int64(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const npy_int64 neg_one = (npy_int64) -1, const_zero = (npy_int64) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(npy_int64) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (npy_int64) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + npy_int64 val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (npy_int64) -1; + val = __Pyx_PyInt_As_npy_int64(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(npy_int64, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(npy_int64) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) >= 2 * PyLong_SHIFT)) { + return (npy_int64) (((((npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(npy_int64) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) >= 3 * PyLong_SHIFT)) { + return (npy_int64) (((((((npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(npy_int64) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) >= 4 * PyLong_SHIFT)) { + return (npy_int64) (((((((((npy_int64)digits[3]) << PyLong_SHIFT) | (npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (npy_int64) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(npy_int64) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(npy_int64) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(npy_int64, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(npy_int64) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 2 * PyLong_SHIFT)) { + return (npy_int64) (((npy_int64)-1)*(((((npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(npy_int64) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 2 * PyLong_SHIFT)) { + return (npy_int64) ((((((npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(npy_int64) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 3 * PyLong_SHIFT)) { + return (npy_int64) (((npy_int64)-1)*(((((((npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(npy_int64) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 3 * PyLong_SHIFT)) { + return (npy_int64) ((((((((npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(npy_int64) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 4 * PyLong_SHIFT)) { + return (npy_int64) (((npy_int64)-1)*(((((((((npy_int64)digits[3]) << PyLong_SHIFT) | (npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(npy_int64) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(npy_int64, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(npy_int64) - 1 > 4 * PyLong_SHIFT)) { + return (npy_int64) ((((((((((npy_int64)digits[3]) << PyLong_SHIFT) | (npy_int64)digits[2]) << PyLong_SHIFT) | (npy_int64)digits[1]) << PyLong_SHIFT) | (npy_int64)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(npy_int64) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(npy_int64) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(npy_int64, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + npy_int64 val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (npy_int64) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (npy_int64) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (npy_int64) -1; + } else { + stepval = v; + } + v = NULL; + val = (npy_int64) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(npy_int64) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((npy_int64) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(npy_int64) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((npy_int64) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((npy_int64) 1) << (sizeof(npy_int64) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (npy_int64) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to npy_int64"); + return (npy_int64) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to npy_int64"); + return (npy_int64) -1; +} + +/* CIntFromPy */ + static CYTHON_INLINE int __Pyx_PyInt_As_int(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const int neg_one = (int) -1, const_zero = (int) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(int) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(int, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (int) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + int val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (int) -1; + val = __Pyx_PyInt_As_int(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(int, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(int) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) >= 2 * PyLong_SHIFT)) { + return (int) (((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(int) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) >= 3 * PyLong_SHIFT)) { + return (int) (((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(int) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) >= 4 * PyLong_SHIFT)) { + return (int) (((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (int) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(int) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(int, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(int) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(int, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(int, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(int) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 2 * PyLong_SHIFT)) { + return (int) (((int)-1)*(((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(int) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 2 * PyLong_SHIFT)) { + return (int) ((((((int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(int) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 3 * PyLong_SHIFT)) { + return (int) (((int)-1)*(((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(int) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 3 * PyLong_SHIFT)) { + return (int) ((((((((int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(int) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 4 * PyLong_SHIFT)) { + return (int) (((int)-1)*(((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(int) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(int, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(int) - 1 > 4 * PyLong_SHIFT)) { + return (int) ((((((((((int)digits[3]) << PyLong_SHIFT) | (int)digits[2]) << PyLong_SHIFT) | (int)digits[1]) << PyLong_SHIFT) | (int)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(int) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(int, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(int) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(int, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + int val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (int) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (int) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (int) -1; + } else { + stepval = v; + } + v = NULL; + val = (int) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(int) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((int) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(int) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((int) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((int) 1) << (sizeof(int) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (int) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to int"); + return (int) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to int"); + return (int) -1; +} + +/* CIntFromPy */ + static CYTHON_INLINE long __Pyx_PyInt_As_long(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const long neg_one = (long) -1, const_zero = (long) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(long) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(long, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (long) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + long val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (long) -1; + val = __Pyx_PyInt_As_long(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(long, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(long) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) >= 2 * PyLong_SHIFT)) { + return (long) (((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(long) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) >= 3 * PyLong_SHIFT)) { + return (long) (((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(long) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) >= 4 * PyLong_SHIFT)) { + return (long) (((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (long) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(long) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(long, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(long) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(long, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(long, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(long) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 2 * PyLong_SHIFT)) { + return (long) (((long)-1)*(((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(long) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 2 * PyLong_SHIFT)) { + return (long) ((((((long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(long) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 3 * PyLong_SHIFT)) { + return (long) (((long)-1)*(((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(long) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 3 * PyLong_SHIFT)) { + return (long) ((((((((long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(long) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 4 * PyLong_SHIFT)) { + return (long) (((long)-1)*(((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(long) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(long, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(long) - 1 > 4 * PyLong_SHIFT)) { + return (long) ((((((((((long)digits[3]) << PyLong_SHIFT) | (long)digits[2]) << PyLong_SHIFT) | (long)digits[1]) << PyLong_SHIFT) | (long)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(long) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(long, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(long) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(long, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + long val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (long) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (long) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (long) -1; + } else { + stepval = v; + } + v = NULL; + val = (long) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(long) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((long) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(long) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((long) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((long) 1) << (sizeof(long) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (long) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to long"); + return (long) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to long"); + return (long) -1; +} + +/* CIntToPy */ + static CYTHON_INLINE PyObject* __Pyx_PyInt_From_long(long value) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const long neg_one = (long) -1, const_zero = (long) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (is_unsigned) { + if (sizeof(long) < sizeof(long)) { + return PyInt_FromLong((long) value); + } else if (sizeof(long) <= sizeof(unsigned long)) { + return PyLong_FromUnsignedLong((unsigned long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(long) <= sizeof(unsigned PY_LONG_LONG)) { + return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value); +#endif + } + } else { + if (sizeof(long) <= sizeof(long)) { + return PyInt_FromLong((long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(long) <= sizeof(PY_LONG_LONG)) { + return PyLong_FromLongLong((PY_LONG_LONG) value); +#endif + } + } + { + unsigned char *bytes = (unsigned char *)&value; +#if !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX >= 0x030d00A4 + if (is_unsigned) { + return PyLong_FromUnsignedNativeBytes(bytes, sizeof(value), -1); + } else { + return PyLong_FromNativeBytes(bytes, sizeof(value), -1); + } +#elif !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX < 0x030d0000 + int one = 1; int little = (int)*(unsigned char *)&one; + return _PyLong_FromByteArray(bytes, sizeof(long), + little, !is_unsigned); +#else + int one = 1; int little = (int)*(unsigned char *)&one; + PyObject *from_bytes, *result = NULL; + PyObject *py_bytes = NULL, *arg_tuple = NULL, *kwds = NULL, *order_str = NULL; + from_bytes = PyObject_GetAttrString((PyObject*)&PyLong_Type, "from_bytes"); + if (!from_bytes) return NULL; + py_bytes = PyBytes_FromStringAndSize((char*)bytes, sizeof(long)); + if (!py_bytes) goto limited_bad; + order_str = PyUnicode_FromString(little ? "little" : "big"); + if (!order_str) goto limited_bad; + arg_tuple = PyTuple_Pack(2, py_bytes, order_str); + if (!arg_tuple) goto limited_bad; + if (!is_unsigned) { + kwds = PyDict_New(); + if (!kwds) goto limited_bad; + if (PyDict_SetItemString(kwds, "signed", __Pyx_NewRef(Py_True))) goto limited_bad; + } + result = PyObject_Call(from_bytes, arg_tuple, kwds); + limited_bad: + Py_XDECREF(kwds); + Py_XDECREF(arg_tuple); + Py_XDECREF(order_str); + Py_XDECREF(py_bytes); + Py_XDECREF(from_bytes); + return result; +#endif + } +} + +/* CIntToPy */ + static CYTHON_INLINE PyObject* __Pyx_PyInt_From_int(int value) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const int neg_one = (int) -1, const_zero = (int) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; + if (is_unsigned) { + if (sizeof(int) < sizeof(long)) { + return PyInt_FromLong((long) value); + } else if (sizeof(int) <= sizeof(unsigned long)) { + return PyLong_FromUnsignedLong((unsigned long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(int) <= sizeof(unsigned PY_LONG_LONG)) { + return PyLong_FromUnsignedLongLong((unsigned PY_LONG_LONG) value); +#endif + } + } else { + if (sizeof(int) <= sizeof(long)) { + return PyInt_FromLong((long) value); +#ifdef HAVE_LONG_LONG + } else if (sizeof(int) <= sizeof(PY_LONG_LONG)) { + return PyLong_FromLongLong((PY_LONG_LONG) value); +#endif + } + } + { + unsigned char *bytes = (unsigned char *)&value; +#if !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX >= 0x030d00A4 + if (is_unsigned) { + return PyLong_FromUnsignedNativeBytes(bytes, sizeof(value), -1); + } else { + return PyLong_FromNativeBytes(bytes, sizeof(value), -1); + } +#elif !CYTHON_COMPILING_IN_LIMITED_API && PY_VERSION_HEX < 0x030d0000 + int one = 1; int little = (int)*(unsigned char *)&one; + return _PyLong_FromByteArray(bytes, sizeof(int), + little, !is_unsigned); +#else + int one = 1; int little = (int)*(unsigned char *)&one; + PyObject *from_bytes, *result = NULL; + PyObject *py_bytes = NULL, *arg_tuple = NULL, *kwds = NULL, *order_str = NULL; + from_bytes = PyObject_GetAttrString((PyObject*)&PyLong_Type, "from_bytes"); + if (!from_bytes) return NULL; + py_bytes = PyBytes_FromStringAndSize((char*)bytes, sizeof(int)); + if (!py_bytes) goto limited_bad; + order_str = PyUnicode_FromString(little ? "little" : "big"); + if (!order_str) goto limited_bad; + arg_tuple = PyTuple_Pack(2, py_bytes, order_str); + if (!arg_tuple) goto limited_bad; + if (!is_unsigned) { + kwds = PyDict_New(); + if (!kwds) goto limited_bad; + if (PyDict_SetItemString(kwds, "signed", __Pyx_NewRef(Py_True))) goto limited_bad; + } + result = PyObject_Call(from_bytes, arg_tuple, kwds); + limited_bad: + Py_XDECREF(kwds); + Py_XDECREF(arg_tuple); + Py_XDECREF(order_str); + Py_XDECREF(py_bytes); + Py_XDECREF(from_bytes); + return result; +#endif + } +} + +/* CIntFromPy */ + static CYTHON_INLINE char __Pyx_PyInt_As_char(PyObject *x) { +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" +#endif + const char neg_one = (char) -1, const_zero = (char) 0; +#ifdef __Pyx_HAS_GCC_DIAGNOSTIC +#pragma GCC diagnostic pop +#endif + const int is_unsigned = neg_one > const_zero; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x))) { + if ((sizeof(char) < sizeof(long))) { + __PYX_VERIFY_RETURN_INT(char, long, PyInt_AS_LONG(x)) + } else { + long val = PyInt_AS_LONG(x); + if (is_unsigned && unlikely(val < 0)) { + goto raise_neg_overflow; + } + return (char) val; + } + } +#endif + if (unlikely(!PyLong_Check(x))) { + char val; + PyObject *tmp = __Pyx_PyNumber_IntOrLong(x); + if (!tmp) return (char) -1; + val = __Pyx_PyInt_As_char(tmp); + Py_DECREF(tmp); + return val; + } + if (is_unsigned) { +#if CYTHON_USE_PYLONG_INTERNALS + if (unlikely(__Pyx_PyLong_IsNeg(x))) { + goto raise_neg_overflow; + } else if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(char, __Pyx_compact_upylong, __Pyx_PyLong_CompactValueUnsigned(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_DigitCount(x)) { + case 2: + if ((8 * sizeof(char) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) >= 2 * PyLong_SHIFT)) { + return (char) (((((char)digits[1]) << PyLong_SHIFT) | (char)digits[0])); + } + } + break; + case 3: + if ((8 * sizeof(char) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) >= 3 * PyLong_SHIFT)) { + return (char) (((((((char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0])); + } + } + break; + case 4: + if ((8 * sizeof(char) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) >= 4 * PyLong_SHIFT)) { + return (char) (((((((((char)digits[3]) << PyLong_SHIFT) | (char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0])); + } + } + break; + } + } +#endif +#if CYTHON_COMPILING_IN_CPYTHON && PY_VERSION_HEX < 0x030C00A7 + if (unlikely(Py_SIZE(x) < 0)) { + goto raise_neg_overflow; + } +#else + { + int result = PyObject_RichCompareBool(x, Py_False, Py_LT); + if (unlikely(result < 0)) + return (char) -1; + if (unlikely(result == 1)) + goto raise_neg_overflow; + } +#endif + if ((sizeof(char) <= sizeof(unsigned long))) { + __PYX_VERIFY_RETURN_INT_EXC(char, unsigned long, PyLong_AsUnsignedLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(char) <= sizeof(unsigned PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(char, unsigned PY_LONG_LONG, PyLong_AsUnsignedLongLong(x)) +#endif + } + } else { +#if CYTHON_USE_PYLONG_INTERNALS + if (__Pyx_PyLong_IsCompact(x)) { + __PYX_VERIFY_RETURN_INT(char, __Pyx_compact_pylong, __Pyx_PyLong_CompactValue(x)) + } else { + const digit* digits = __Pyx_PyLong_Digits(x); + assert(__Pyx_PyLong_DigitCount(x) > 1); + switch (__Pyx_PyLong_SignedDigitCount(x)) { + case -2: + if ((8 * sizeof(char) - 1 > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, long, -(long) (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 2 * PyLong_SHIFT)) { + return (char) (((char)-1)*(((((char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case 2: + if ((8 * sizeof(char) > 1 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 2 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 2 * PyLong_SHIFT)) { + return (char) ((((((char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case -3: + if ((8 * sizeof(char) - 1 > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, long, -(long) (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 3 * PyLong_SHIFT)) { + return (char) (((char)-1)*(((((((char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case 3: + if ((8 * sizeof(char) > 2 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 3 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 3 * PyLong_SHIFT)) { + return (char) ((((((((char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case -4: + if ((8 * sizeof(char) - 1 > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, long, -(long) (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 4 * PyLong_SHIFT)) { + return (char) (((char)-1)*(((((((((char)digits[3]) << PyLong_SHIFT) | (char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + case 4: + if ((8 * sizeof(char) > 3 * PyLong_SHIFT)) { + if ((8 * sizeof(unsigned long) > 4 * PyLong_SHIFT)) { + __PYX_VERIFY_RETURN_INT(char, unsigned long, (((((((((unsigned long)digits[3]) << PyLong_SHIFT) | (unsigned long)digits[2]) << PyLong_SHIFT) | (unsigned long)digits[1]) << PyLong_SHIFT) | (unsigned long)digits[0]))) + } else if ((8 * sizeof(char) - 1 > 4 * PyLong_SHIFT)) { + return (char) ((((((((((char)digits[3]) << PyLong_SHIFT) | (char)digits[2]) << PyLong_SHIFT) | (char)digits[1]) << PyLong_SHIFT) | (char)digits[0]))); + } + } + break; + } + } +#endif + if ((sizeof(char) <= sizeof(long))) { + __PYX_VERIFY_RETURN_INT_EXC(char, long, PyLong_AsLong(x)) +#ifdef HAVE_LONG_LONG + } else if ((sizeof(char) <= sizeof(PY_LONG_LONG))) { + __PYX_VERIFY_RETURN_INT_EXC(char, PY_LONG_LONG, PyLong_AsLongLong(x)) +#endif + } + } + { + char val; + int ret = -1; +#if PY_VERSION_HEX >= 0x030d00A6 && !CYTHON_COMPILING_IN_LIMITED_API + Py_ssize_t bytes_copied = PyLong_AsNativeBytes( + x, &val, sizeof(val), Py_ASNATIVEBYTES_NATIVE_ENDIAN | (is_unsigned ? Py_ASNATIVEBYTES_UNSIGNED_BUFFER | Py_ASNATIVEBYTES_REJECT_NEGATIVE : 0)); + if (unlikely(bytes_copied == -1)) { + } else if (unlikely(bytes_copied > (Py_ssize_t) sizeof(val))) { + goto raise_overflow; + } else { + ret = 0; + } +#elif PY_VERSION_HEX < 0x030d0000 && !(CYTHON_COMPILING_IN_PYPY || CYTHON_COMPILING_IN_LIMITED_API) || defined(_PyLong_AsByteArray) + int one = 1; int is_little = (int)*(unsigned char *)&one; + unsigned char *bytes = (unsigned char *)&val; + ret = _PyLong_AsByteArray((PyLongObject *)x, + bytes, sizeof(val), + is_little, !is_unsigned); +#else + PyObject *v; + PyObject *stepval = NULL, *mask = NULL, *shift = NULL; + int bits, remaining_bits, is_negative = 0; + int chunk_size = (sizeof(long) < 8) ? 30 : 62; + if (likely(PyLong_CheckExact(x))) { + v = __Pyx_NewRef(x); + } else { + v = PyNumber_Long(x); + if (unlikely(!v)) return (char) -1; + assert(PyLong_CheckExact(v)); + } + { + int result = PyObject_RichCompareBool(v, Py_False, Py_LT); + if (unlikely(result < 0)) { + Py_DECREF(v); + return (char) -1; + } + is_negative = result == 1; + } + if (is_unsigned && unlikely(is_negative)) { + Py_DECREF(v); + goto raise_neg_overflow; + } else if (is_negative) { + stepval = PyNumber_Invert(v); + Py_DECREF(v); + if (unlikely(!stepval)) + return (char) -1; + } else { + stepval = v; + } + v = NULL; + val = (char) 0; + mask = PyLong_FromLong((1L << chunk_size) - 1); if (unlikely(!mask)) goto done; + shift = PyLong_FromLong(chunk_size); if (unlikely(!shift)) goto done; + for (bits = 0; bits < (int) sizeof(char) * 8 - chunk_size; bits += chunk_size) { + PyObject *tmp, *digit; + long idigit; + digit = PyNumber_And(stepval, mask); + if (unlikely(!digit)) goto done; + idigit = PyLong_AsLong(digit); + Py_DECREF(digit); + if (unlikely(idigit < 0)) goto done; + val |= ((char) idigit) << bits; + tmp = PyNumber_Rshift(stepval, shift); + if (unlikely(!tmp)) goto done; + Py_DECREF(stepval); stepval = tmp; + } + Py_DECREF(shift); shift = NULL; + Py_DECREF(mask); mask = NULL; + { + long idigit = PyLong_AsLong(stepval); + if (unlikely(idigit < 0)) goto done; + remaining_bits = ((int) sizeof(char) * 8) - bits - (is_unsigned ? 0 : 1); + if (unlikely(idigit >= (1L << remaining_bits))) + goto raise_overflow; + val |= ((char) idigit) << bits; + } + if (!is_unsigned) { + if (unlikely(val & (((char) 1) << (sizeof(char) * 8 - 1)))) + goto raise_overflow; + if (is_negative) + val = ~val; + } + ret = 0; + done: + Py_XDECREF(shift); + Py_XDECREF(mask); + Py_XDECREF(stepval); +#endif + if (unlikely(ret)) + return (char) -1; + return val; + } +raise_overflow: + PyErr_SetString(PyExc_OverflowError, + "value too large to convert to char"); + return (char) -1; +raise_neg_overflow: + PyErr_SetString(PyExc_OverflowError, + "can't convert negative value to char"); + return (char) -1; +} + +/* FormatTypeName */ + #if CYTHON_COMPILING_IN_LIMITED_API +static __Pyx_TypeName +__Pyx_PyType_GetName(PyTypeObject* tp) +{ + PyObject *name = __Pyx_PyObject_GetAttrStr((PyObject *)tp, + __pyx_n_s_name_2); + if (unlikely(name == NULL) || unlikely(!PyUnicode_Check(name))) { + PyErr_Clear(); + Py_XDECREF(name); + name = __Pyx_NewRef(__pyx_n_s__35); + } + return name; +} +#endif + +/* CheckBinaryVersion */ + static unsigned long __Pyx_get_runtime_version(void) { +#if __PYX_LIMITED_VERSION_HEX >= 0x030B00A4 + return Py_Version & ~0xFFUL; +#else + const char* rt_version = Py_GetVersion(); + unsigned long version = 0; + unsigned long factor = 0x01000000UL; + unsigned int digit = 0; + int i = 0; + while (factor) { + while ('0' <= rt_version[i] && rt_version[i] <= '9') { + digit = digit * 10 + (unsigned int) (rt_version[i] - '0'); + ++i; + } + version += factor * digit; + if (rt_version[i] != '.') + break; + digit = 0; + factor >>= 8; + ++i; + } + return version; +#endif +} +static int __Pyx_check_binary_version(unsigned long ct_version, unsigned long rt_version, int allow_newer) { + const unsigned long MAJOR_MINOR = 0xFFFF0000UL; + if ((rt_version & MAJOR_MINOR) == (ct_version & MAJOR_MINOR)) + return 0; + if (likely(allow_newer && (rt_version & MAJOR_MINOR) > (ct_version & MAJOR_MINOR))) + return 1; + { + char message[200]; + PyOS_snprintf(message, sizeof(message), + "compile time Python version %d.%d " + "of module '%.100s' " + "%s " + "runtime version %d.%d", + (int) (ct_version >> 24), (int) ((ct_version >> 16) & 0xFF), + __Pyx_MODULE_NAME, + (allow_newer) ? "was newer than" : "does not match", + (int) (rt_version >> 24), (int) ((rt_version >> 16) & 0xFF) + ); + return PyErr_WarnEx(NULL, message, 1); + } +} + +/* InitStrings */ + #if PY_MAJOR_VERSION >= 3 +static int __Pyx_InitString(__Pyx_StringTabEntry t, PyObject **str) { + if (t.is_unicode | t.is_str) { + if (t.intern) { + *str = PyUnicode_InternFromString(t.s); + } else if (t.encoding) { + *str = PyUnicode_Decode(t.s, t.n - 1, t.encoding, NULL); + } else { + *str = PyUnicode_FromStringAndSize(t.s, t.n - 1); + } + } else { + *str = PyBytes_FromStringAndSize(t.s, t.n - 1); + } + if (!*str) + return -1; + if (PyObject_Hash(*str) == -1) + return -1; + return 0; +} +#endif +static int __Pyx_InitStrings(__Pyx_StringTabEntry *t) { + while (t->p) { + #if PY_MAJOR_VERSION >= 3 + __Pyx_InitString(*t, t->p); + #else + if (t->is_unicode) { + *t->p = PyUnicode_DecodeUTF8(t->s, t->n - 1, NULL); + } else if (t->intern) { + *t->p = PyString_InternFromString(t->s); + } else { + *t->p = PyString_FromStringAndSize(t->s, t->n - 1); + } + if (!*t->p) + return -1; + if (PyObject_Hash(*t->p) == -1) + return -1; + #endif + ++t; + } + return 0; +} + +#include +static CYTHON_INLINE Py_ssize_t __Pyx_ssize_strlen(const char *s) { + size_t len = strlen(s); + if (unlikely(len > (size_t) PY_SSIZE_T_MAX)) { + PyErr_SetString(PyExc_OverflowError, "byte string is too long"); + return -1; + } + return (Py_ssize_t) len; +} +static CYTHON_INLINE PyObject* __Pyx_PyUnicode_FromString(const char* c_str) { + Py_ssize_t len = __Pyx_ssize_strlen(c_str); + if (unlikely(len < 0)) return NULL; + return __Pyx_PyUnicode_FromStringAndSize(c_str, len); +} +static CYTHON_INLINE PyObject* __Pyx_PyByteArray_FromString(const char* c_str) { + Py_ssize_t len = __Pyx_ssize_strlen(c_str); + if (unlikely(len < 0)) return NULL; + return PyByteArray_FromStringAndSize(c_str, len); +} +static CYTHON_INLINE const char* __Pyx_PyObject_AsString(PyObject* o) { + Py_ssize_t ignore; + return __Pyx_PyObject_AsStringAndSize(o, &ignore); +} +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT +#if !CYTHON_PEP393_ENABLED +static const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) { + char* defenc_c; + PyObject* defenc = _PyUnicode_AsDefaultEncodedString(o, NULL); + if (!defenc) return NULL; + defenc_c = PyBytes_AS_STRING(defenc); +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII + { + char* end = defenc_c + PyBytes_GET_SIZE(defenc); + char* c; + for (c = defenc_c; c < end; c++) { + if ((unsigned char) (*c) >= 128) { + PyUnicode_AsASCIIString(o); + return NULL; + } + } + } +#endif + *length = PyBytes_GET_SIZE(defenc); + return defenc_c; +} +#else +static CYTHON_INLINE const char* __Pyx_PyUnicode_AsStringAndSize(PyObject* o, Py_ssize_t *length) { + if (unlikely(__Pyx_PyUnicode_READY(o) == -1)) return NULL; +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII + if (likely(PyUnicode_IS_ASCII(o))) { + *length = PyUnicode_GET_LENGTH(o); + return PyUnicode_AsUTF8(o); + } else { + PyUnicode_AsASCIIString(o); + return NULL; + } +#else + return PyUnicode_AsUTF8AndSize(o, length); +#endif +} +#endif +#endif +static CYTHON_INLINE const char* __Pyx_PyObject_AsStringAndSize(PyObject* o, Py_ssize_t *length) { +#if __PYX_DEFAULT_STRING_ENCODING_IS_ASCII || __PYX_DEFAULT_STRING_ENCODING_IS_DEFAULT + if ( +#if PY_MAJOR_VERSION < 3 && __PYX_DEFAULT_STRING_ENCODING_IS_ASCII + __Pyx_sys_getdefaultencoding_not_ascii && +#endif + PyUnicode_Check(o)) { + return __Pyx_PyUnicode_AsStringAndSize(o, length); + } else +#endif +#if (!CYTHON_COMPILING_IN_PYPY && !CYTHON_COMPILING_IN_LIMITED_API) || (defined(PyByteArray_AS_STRING) && defined(PyByteArray_GET_SIZE)) + if (PyByteArray_Check(o)) { + *length = PyByteArray_GET_SIZE(o); + return PyByteArray_AS_STRING(o); + } else +#endif + { + char* result; + int r = PyBytes_AsStringAndSize(o, &result, length); + if (unlikely(r < 0)) { + return NULL; + } else { + return result; + } + } +} +static CYTHON_INLINE int __Pyx_PyObject_IsTrue(PyObject* x) { + int is_true = x == Py_True; + if (is_true | (x == Py_False) | (x == Py_None)) return is_true; + else return PyObject_IsTrue(x); +} +static CYTHON_INLINE int __Pyx_PyObject_IsTrueAndDecref(PyObject* x) { + int retval; + if (unlikely(!x)) return -1; + retval = __Pyx_PyObject_IsTrue(x); + Py_DECREF(x); + return retval; +} +static PyObject* __Pyx_PyNumber_IntOrLongWrongResultType(PyObject* result, const char* type_name) { + __Pyx_TypeName result_type_name = __Pyx_PyType_GetName(Py_TYPE(result)); +#if PY_MAJOR_VERSION >= 3 + if (PyLong_Check(result)) { + if (PyErr_WarnFormat(PyExc_DeprecationWarning, 1, + "__int__ returned non-int (type " __Pyx_FMT_TYPENAME "). " + "The ability to return an instance of a strict subclass of int is deprecated, " + "and may be removed in a future version of Python.", + result_type_name)) { + __Pyx_DECREF_TypeName(result_type_name); + Py_DECREF(result); + return NULL; + } + __Pyx_DECREF_TypeName(result_type_name); + return result; + } +#endif + PyErr_Format(PyExc_TypeError, + "__%.4s__ returned non-%.4s (type " __Pyx_FMT_TYPENAME ")", + type_name, type_name, result_type_name); + __Pyx_DECREF_TypeName(result_type_name); + Py_DECREF(result); + return NULL; +} +static CYTHON_INLINE PyObject* __Pyx_PyNumber_IntOrLong(PyObject* x) { +#if CYTHON_USE_TYPE_SLOTS + PyNumberMethods *m; +#endif + const char *name = NULL; + PyObject *res = NULL; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_Check(x) || PyLong_Check(x))) +#else + if (likely(PyLong_Check(x))) +#endif + return __Pyx_NewRef(x); +#if CYTHON_USE_TYPE_SLOTS + m = Py_TYPE(x)->tp_as_number; + #if PY_MAJOR_VERSION < 3 + if (m && m->nb_int) { + name = "int"; + res = m->nb_int(x); + } + else if (m && m->nb_long) { + name = "long"; + res = m->nb_long(x); + } + #else + if (likely(m && m->nb_int)) { + name = "int"; + res = m->nb_int(x); + } + #endif +#else + if (!PyBytes_CheckExact(x) && !PyUnicode_CheckExact(x)) { + res = PyNumber_Int(x); + } +#endif + if (likely(res)) { +#if PY_MAJOR_VERSION < 3 + if (unlikely(!PyInt_Check(res) && !PyLong_Check(res))) { +#else + if (unlikely(!PyLong_CheckExact(res))) { +#endif + return __Pyx_PyNumber_IntOrLongWrongResultType(res, name); + } + } + else if (!PyErr_Occurred()) { + PyErr_SetString(PyExc_TypeError, + "an integer is required"); + } + return res; +} +static CYTHON_INLINE Py_ssize_t __Pyx_PyIndex_AsSsize_t(PyObject* b) { + Py_ssize_t ival; + PyObject *x; +#if PY_MAJOR_VERSION < 3 + if (likely(PyInt_CheckExact(b))) { + if (sizeof(Py_ssize_t) >= sizeof(long)) + return PyInt_AS_LONG(b); + else + return PyInt_AsSsize_t(b); + } +#endif + if (likely(PyLong_CheckExact(b))) { + #if CYTHON_USE_PYLONG_INTERNALS + if (likely(__Pyx_PyLong_IsCompact(b))) { + return __Pyx_PyLong_CompactValue(b); + } else { + const digit* digits = __Pyx_PyLong_Digits(b); + const Py_ssize_t size = __Pyx_PyLong_SignedDigitCount(b); + switch (size) { + case 2: + if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) { + return (Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case -2: + if (8 * sizeof(Py_ssize_t) > 2 * PyLong_SHIFT) { + return -(Py_ssize_t) (((((size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case 3: + if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) { + return (Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case -3: + if (8 * sizeof(Py_ssize_t) > 3 * PyLong_SHIFT) { + return -(Py_ssize_t) (((((((size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case 4: + if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) { + return (Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + case -4: + if (8 * sizeof(Py_ssize_t) > 4 * PyLong_SHIFT) { + return -(Py_ssize_t) (((((((((size_t)digits[3]) << PyLong_SHIFT) | (size_t)digits[2]) << PyLong_SHIFT) | (size_t)digits[1]) << PyLong_SHIFT) | (size_t)digits[0])); + } + break; + } + } + #endif + return PyLong_AsSsize_t(b); + } + x = PyNumber_Index(b); + if (!x) return -1; + ival = PyInt_AsSsize_t(x); + Py_DECREF(x); + return ival; +} +static CYTHON_INLINE Py_hash_t __Pyx_PyIndex_AsHash_t(PyObject* o) { + if (sizeof(Py_hash_t) == sizeof(Py_ssize_t)) { + return (Py_hash_t) __Pyx_PyIndex_AsSsize_t(o); +#if PY_MAJOR_VERSION < 3 + } else if (likely(PyInt_CheckExact(o))) { + return PyInt_AS_LONG(o); +#endif + } else { + Py_ssize_t ival; + PyObject *x; + x = PyNumber_Index(o); + if (!x) return -1; + ival = PyInt_AsLong(x); + Py_DECREF(x); + return ival; + } +} +static CYTHON_INLINE PyObject * __Pyx_PyBool_FromLong(long b) { + return b ? __Pyx_NewRef(Py_True) : __Pyx_NewRef(Py_False); +} +static CYTHON_INLINE PyObject * __Pyx_PyInt_FromSize_t(size_t ival) { + return PyInt_FromSize_t(ival); +} + + +/* #### Code section: utility_code_pragmas_end ### */ +#ifdef _MSC_VER +#pragma warning( pop ) +#endif + + + +/* #### Code section: end ### */ +#endif /* Py_PYTHON_H */ diff --git a/fairseq/data/token_block_utils_fast.cpython-310-x86_64-linux-gnu.so b/fairseq/data/token_block_utils_fast.cpython-310-x86_64-linux-gnu.so new file mode 100644 index 0000000000000000000000000000000000000000..2b4f9ea5610821a608fdb09e64c4618d3d88e1cf --- /dev/null +++ b/fairseq/data/token_block_utils_fast.cpython-310-x86_64-linux-gnu.so @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:105d14894186afdca051690cd766de67eb6156fa362d0356b978a58d830a33c0 +size 287008 diff --git a/fairseq/data/token_block_utils_fast.pyx b/fairseq/data/token_block_utils_fast.pyx new file mode 100644 index 0000000000000000000000000000000000000000..5563b973e9da2b9560e7660a370f9ef30a1d7ce6 --- /dev/null +++ b/fairseq/data/token_block_utils_fast.pyx @@ -0,0 +1,185 @@ +# cython: language_level=3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +from itertools import chain +from libc.math cimport ceil + +cimport cython +cimport numpy as np + +DTYPE = np.int64 +ctypedef np.int64_t DTYPE_t + + +@cython.boundscheck(False) +@cython.wraparound(False) +@cython.nonecheck(False) +cdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_none_mode(np.ndarray[DTYPE_t, ndim=1] sizes, int block_size): + cdef DTYPE_t total_size = sizes.sum() + cdef DTYPE_t length = ceil(total_size / block_size) + cdef np.ndarray[DTYPE_t, ndim=2] slice_indices = np.zeros([length, 2], dtype=DTYPE) + cdef DTYPE_t[:, :] slice_indices_view = slice_indices + cdef DTYPE_t i + cdef DTYPE_t start + cdef DTYPE_t end + for i in range(length): + start = i * block_size + end = min(start + block_size, total_size) + slice_indices_view[i][0] = start + slice_indices_view[i][1] = end + return slice_indices + + +cdef np.ndarray[DTYPE_t, ndim=2] _fast_convert_to_np_array(list list_of_list): + """ + Faster function to convert DTYPE_t list of list. + Only fast when there are huge number of rows and low number of columns. + """ + cdef np.ndarray[DTYPE_t, ndim=1] flat = np.fromiter(chain.from_iterable(list_of_list), DTYPE, -1) + return flat.reshape((len(list_of_list), -1)) + + +@cython.boundscheck(False) +@cython.wraparound(False) +@cython.nonecheck(False) +cpdef np.ndarray[DTYPE_t, ndim=2] _get_slice_indices_fast(np.ndarray[DTYPE_t, ndim=1] sizes, str break_mode, int block_size, int document_sep_len): + cdef DTYPE_t tok_idx = 0 + cdef DTYPE_t sz_idx = 0 + cdef DTYPE_t curr_size = 0 + cdef DTYPE_t i = 0 + cdef DTYPE_t length + cdef DTYPE_t total_size + cdef DTYPE_t[:] sizes_view = sizes + cdef np.ndarray[DTYPE_t, ndim=2] slice_indices + cdef list slice_indices_list = [] + + if break_mode is None or break_mode == 'none': + slice_indices = _get_slice_indices_none_mode(sizes, block_size) + elif break_mode == 'complete': + while sz_idx < len(sizes_view): + if curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0: + curr_size += sizes_view[sz_idx] + sz_idx += 1 + else: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + tok_idx += curr_size + curr_size = 0 + if curr_size > 0: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + slice_indices = _fast_convert_to_np_array(slice_indices_list) + elif break_mode == 'complete_doc': + while sz_idx < len(sizes_view): + if ( + (curr_size + sizes_view[sz_idx] <= block_size or curr_size == 0) + # an empty sentence indicates end-of-document: + and sizes_view[sz_idx] != document_sep_len + ): + curr_size += sizes_view[sz_idx] + sz_idx += 1 + else: + # Only keep non-empty documents. + if curr_size > 1: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + tok_idx += curr_size + curr_size = 0 + if sizes_view[sz_idx] == document_sep_len: + tok_idx += sizes_view[sz_idx] + sz_idx += 1 + if curr_size > 1: + slice_indices_list.append((tok_idx, tok_idx + curr_size)) + slice_indices = _fast_convert_to_np_array(slice_indices_list) + elif break_mode == 'eos': + slice_indices = np.zeros((len(sizes), 2), dtype=DTYPE) + cumsum = sizes.cumsum(axis=0) + slice_indices[1:, 0] = cumsum[:cumsum.shape[0] - 1] + slice_indices[:, 1] = cumsum + else: + raise ValueError('Invalid break_mode: ' + break_mode) + return slice_indices + + +@cython.boundscheck(False) +@cython.wraparound(False) +@cython.nonecheck(False) +cpdef np.ndarray[DTYPE_t, ndim=2] _get_block_to_dataset_index_fast(np.ndarray[DTYPE_t, ndim=1] sizes, np.ndarray[DTYPE_t, ndim=2] slice_indices): + cdef DTYPE_t start_ds_idx + cdef DTYPE_t start_offset + cdef DTYPE_t end_ds_idx + cdef DTYPE_t i + cdef DTYPE_t s + cdef DTYPE_t e + cdef DatasetSearcher ds = DatasetSearcher(sizes) + cdef np.ndarray[DTYPE_t, ndim=2] block_to_dataset_index = np.zeros([len(slice_indices), 3], dtype=DTYPE) + cdef DTYPE_t[:, :] block_to_dataset_index_view = block_to_dataset_index + cdef DTYPE_t[:, :] slice_indices_view = slice_indices + cdef Py_ssize_t x_max = slice_indices.shape[0] + + for i in range(x_max): + s = slice_indices_view[i][0] + e = slice_indices_view[i][1] + ds.seek(s) + start_ds_idx = ds.current_index + start_offset = ds.current_offset + if e <= s: + end_ds_idx = start_ds_idx + else: + ds.seek(e - 1) + end_ds_idx = ds.current_index + block_to_dataset_index_view[i][0] = start_ds_idx # starting index in dataset + block_to_dataset_index_view[i][1] = start_offset # starting offset within starting index + block_to_dataset_index_view[i][2] = end_ds_idx # ending index in dataset + return block_to_dataset_index + + +cdef class DatasetSearcher(object): + """Helper for mapping "flat" indices to indices and offsets in an + underlying dataset.""" + cdef DTYPE_t current_i + cdef DTYPE_t current_offset + cdef DTYPE_t current_index + cdef DTYPE_t[:] sizes + + def __init__(self, DTYPE_t[:] sizes): + self.sizes = sizes + self.reset() + + cdef reset(self): + self.current_offset = 0 # offset within current index in underlying dataset + self.current_i = 0 # "flat" index + self.current_index = 0 # index in underlying dataset + + @cython.boundscheck(False) + @cython.wraparound(False) + @cython.nonecheck(False) + cdef int step(self, DTYPE_t i): + cdef DTYPE_t to_consume + cdef DTYPE_t remaining + if i < self.current_i: + self.reset() + if i > self.current_i: + to_consume = i - self.current_i + remaining = self.sizes[self.current_index] - self.current_offset + if remaining > to_consume: + self.current_offset += to_consume + self.current_i += to_consume + else: + assert remaining > 0 + self.current_i += remaining + self.current_index += 1 + self.current_offset = 0 + return 1 + return 0 + + @cython.boundscheck(False) + @cython.wraparound(False) + @cython.nonecheck(False) + cdef seek(self, DTYPE_t i): + cdef int not_done = 1 + while not_done == 1: + not_done = self.step(i) + assert self.current_i == i diff --git a/fairseq/data/transform_eos_dataset.py b/fairseq/data/transform_eos_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..4ce5ad811bbcd01c1d46cea6acd5206e730f4f76 --- /dev/null +++ b/fairseq/data/transform_eos_dataset.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from . import FairseqDataset + + +class TransformEosDataset(FairseqDataset): + """A :class:`~fairseq.data.FairseqDataset` wrapper that appends/prepends/strips EOS. + + Note that the transformation is applied in :func:`collater`. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset to wrap + eos (int): index of the end-of-sentence symbol + append_eos_to_src (bool, optional): append EOS to the end of src + remove_eos_from_src (bool, optional): remove EOS from the end of src + append_eos_to_tgt (bool, optional): append EOS to the end of tgt + remove_eos_from_tgt (bool, optional): remove EOS from the end of tgt + """ + + def __init__( + self, + dataset, + eos, + append_eos_to_src=False, + remove_eos_from_src=False, + append_eos_to_tgt=False, + remove_eos_from_tgt=False, + has_target=True, + ): + if not isinstance(dataset, FairseqDataset): + raise ValueError('dataset must be an instance of FairseqDataset') + if append_eos_to_src and remove_eos_from_src: + raise ValueError('cannot combine append_eos_to_src and remove_eos_from_src') + if append_eos_to_tgt and remove_eos_from_tgt: + raise ValueError('cannot combine append_eos_to_tgt and remove_eos_from_tgt') + + self.dataset = dataset + self.eos = torch.LongTensor([eos]) + self.append_eos_to_src = append_eos_to_src + self.remove_eos_from_src = remove_eos_from_src + self.append_eos_to_tgt = append_eos_to_tgt + self.remove_eos_from_tgt = remove_eos_from_tgt + self.has_target = has_target + + # precompute how we should adjust the reported sizes + self._src_delta = 0 + self._src_delta += 1 if append_eos_to_src else 0 + self._src_delta -= 1 if remove_eos_from_src else 0 + self._tgt_delta = 0 + self._tgt_delta += 1 if append_eos_to_tgt else 0 + self._tgt_delta -= 1 if remove_eos_from_tgt else 0 + + self._checked_src = False + self._checked_tgt = False + + def _check_src(self, src, expect_eos): + if not self._checked_src: + assert (src[-1] == self.eos[0]) == expect_eos + self._checked_src = True + + def _check_tgt(self, tgt, expect_eos): + if self.has_target and not self._checked_tgt: + assert (tgt[-1] == self.eos[0]) == expect_eos + self._checked_tgt = True + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples): + + def transform(item): + if self.append_eos_to_src: + self.eos = self.eos.to(device=item['source'].device) + self._check_src(item['source'], expect_eos=False) + item['source'] = torch.cat([item['source'], self.eos]) + if self.remove_eos_from_src: + self.eos = self.eos.to(device=item['source'].device) + self._check_src(item['source'], expect_eos=True) + item['source'] = item['source'][:-1] + if self.append_eos_to_tgt: + self.eos = self.eos.to(device=item['target'].device) + self._check_tgt(item['target'], expect_eos=False) + item['target'] = torch.cat([item['target'], self.eos]) + if self.remove_eos_from_tgt: + self.eos = self.eos.to(device=item['target'].device) + self._check_tgt(item['target'], expect_eos=True) + item['target'] = item['target'][:-1] + return item + + samples = list(map(transform, samples)) + return self.dataset.collater(samples) + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + if self.has_target: + src_len, tgt_len = self.dataset.size(index) + return (src_len + self._src_delta, tgt_len + self._tgt_delta) + else: + return self.dataset.size(index) + + def ordered_indices(self): + # NOTE: we assume that the ordering does not change based on the + # addition or removal of eos + return self.dataset.ordered_indices() + + @property + def supports_prefetch(self): + return getattr(self.dataset, 'supports_prefetch', False) + + def prefetch(self, indices): + return self.dataset.prefetch(indices) diff --git a/fairseq/data/transform_eos_lang_pair_dataset.py b/fairseq/data/transform_eos_lang_pair_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..55137ca55cc442d760a21c93c215bd74d3cea868 --- /dev/null +++ b/fairseq/data/transform_eos_lang_pair_dataset.py @@ -0,0 +1,89 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from . import FairseqDataset +import torch +from typing import Optional + + +class TransformEosLangPairDataset(FairseqDataset): + """A :class:`~fairseq.data.FairseqDataset` wrapper that transform bos on + collated samples of language pair dataset. + + Note that the transformation is applied in :func:`collater`. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset that collates sample into + LanguagePairDataset schema + src_eos (int): original source end-of-sentence symbol index to be replaced + new_src_eos (int, optional): new end-of-sentence symbol index to replace source eos symbol + tgt_bos (int, optional): original target beginning-of-sentence symbol index to be replaced + new_tgt_bos (int, optional): new beginning-of-sentence symbol index to replace at the + beginning of 'prev_output_tokens' + """ + + def __init__( + self, + dataset: FairseqDataset, + src_eos: int, + new_src_eos: Optional[int] = None, + tgt_bos: Optional[int] = None, + new_tgt_bos: Optional[int] = None, + ): + self.dataset = dataset + self.src_eos = src_eos + self.new_src_eos = new_src_eos + self.tgt_bos = tgt_bos + self.new_tgt_bos = new_tgt_bos + + def __getitem__(self, index): + return self.dataset[index] + + def __len__(self): + return len(self.dataset) + + def collater(self, samples, **extra_args): + samples = self.dataset.collater(samples, **extra_args) + + if self.new_src_eos is not None: + if self.dataset.left_pad_source: + assert(samples['net_input']['src_tokens'][:, -1] != self.src_eos).sum() == 0 + samples['net_input']['src_tokens'][:, -1] = self.new_src_eos + else: + eos_idx = samples['net_input']['src_lengths'] - 1 + assert( + samples['net_input']['src_tokens'][torch.arange(eos_idx.size(0)), eos_idx] != self.src_eos + ).sum() == 0 + eos_idx = eos_idx.resize_(len(samples['net_input']['src_lengths']), 1) + samples['net_input']['src_tokens'].scatter_(1, eos_idx, self.new_src_eos) + + if self.new_tgt_bos is not None and 'prev_output_tokens' in samples['net_input']: + if self.dataset.left_pad_target: + # TODO: support different padding direction on target side + raise NotImplementedError( + 'TransformEosLangPairDataset does not implement --left-pad-target True option' + ) + else: + assert (samples['net_input']['prev_output_tokens'][:, 0] != self.tgt_bos).sum() == 0 + samples['net_input']['prev_output_tokens'][:, 0] = self.new_tgt_bos + + return samples + + def num_tokens(self, index): + return self.dataset.num_tokens(index) + + def size(self, index): + return self.dataset.size(index) + + def ordered_indices(self): + return self.dataset.ordered_indices() + + @property + def supports_prefetch(self): + return getattr(self.dataset, 'supports_prefetch', False) + + def prefetch(self, indices): + return self.dataset.prefetch(indices) diff --git a/fairseq/distributed_utils.py b/fairseq/distributed_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7ee89adce9e4bee192253fb247e879a064036bf5 --- /dev/null +++ b/fairseq/distributed_utils.py @@ -0,0 +1,322 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import pickle +import random +import socket +import struct +import subprocess +import warnings +from collections import OrderedDict +from typing import Any, Dict, Mapping + +import torch +import torch.distributed as dist + +from fairseq import utils + + +logger = logging.getLogger(__name__) + + +def is_master(args): + return args.distributed_rank == 0 + + +def infer_init_method(args, force_distributed=False): + if args.distributed_init_method is not None or getattr(args, 'tpu', False): + return + + # support torch.distributed.launch + if all(key in os.environ for key in [ + 'MASTER_ADDR', 'MASTER_PORT', 'WORLD_SIZE', 'RANK' + ]): + args.distributed_init_method = 'env://' + args.distributed_world_size = int(os.environ['WORLD_SIZE']) + args.distributed_rank = int(os.environ['RANK']) + # processes are created by torch.distributed.launch + args.distributed_no_spawn = True + + # we can determine the init method automatically for Slurm + elif args.distributed_port > 0: + node_list = os.environ.get('SLURM_STEP_NODELIST') + if node_list is None: + node_list = os.environ.get('SLURM_JOB_NODELIST') + if node_list is not None: + try: + hostnames = subprocess.check_output(['scontrol', 'show', 'hostnames', node_list]) + args.distributed_init_method = 'tcp://{host}:{port}'.format( + host=hostnames.split()[0].decode('utf-8'), + port=args.distributed_port, + ) + nnodes = int(os.environ.get('SLURM_NNODES')) + ntasks_per_node = os.environ.get('SLURM_NTASKS_PER_NODE') + if ntasks_per_node is not None: + ntasks_per_node = int(ntasks_per_node) + else: + ntasks = int(os.environ.get('SLURM_NTASKS')) + nnodes = int(os.environ.get('SLURM_NNODES')) + assert ntasks % nnodes == 0 + ntasks_per_node = int(ntasks / nnodes) + if ntasks_per_node == 1: + assert args.distributed_world_size % nnodes == 0 + gpus_per_node = args.distributed_world_size // nnodes + node_id = int(os.environ.get('SLURM_NODEID')) + args.distributed_rank = node_id * gpus_per_node + else: + assert ntasks_per_node == args.distributed_world_size // nnodes + args.distributed_no_spawn = True + args.distributed_rank = int(os.environ.get('SLURM_PROCID')) + args.device_id = int(os.environ.get('SLURM_LOCALID')) + except subprocess.CalledProcessError as e: # scontrol failed + raise e + except FileNotFoundError: # Slurm is not installed + pass + + elif args.distributed_world_size > 1 or force_distributed: + # fallback for single node with multiple GPUs + assert args.distributed_world_size <= torch.cuda.device_count() + port = random.randint(10000, 20000) + args.distributed_init_method = 'tcp://localhost:{port}'.format(port=port) + + +def distributed_init(args): + if not getattr(args, 'tpu', False): + if torch.distributed.is_initialized(): + warnings.warn('Distributed is already initialized, cannot initialize twice!') + else: + logger.info('distributed init (rank {}): {}'.format( + args.distributed_rank, args.distributed_init_method, + )) + dist.init_process_group( + backend=args.distributed_backend, + init_method=args.distributed_init_method, + world_size=args.distributed_world_size, + rank=args.distributed_rank, + ) + logger.info('initialized host {} as rank {}'.format( + socket.gethostname(), args.distributed_rank, + )) + + # perform a dummy all-reduce to initialize the NCCL communicator + if torch.cuda.is_available(): + dist.all_reduce(torch.zeros(1).cuda()) + + args.distributed_rank = torch.distributed.get_rank() + else: + import torch_xla.core.xla_model as xm + assert xm.xrt_world_size() == args.distributed_world_size + args.device_id = xm.get_local_ordinal() + args.distributed_rank = xm.get_ordinal() + xm.rendezvous('distributed_init') # wait for all workers + xm.mark_step() + + if is_master(args): + logging.getLogger().setLevel(logging.INFO) + else: + logging.getLogger().setLevel(logging.WARNING) + + if args.model_parallel_size > 1: + try: + from fairseq.model_parallel.megatron.mpu import ( + get_model_parallel_rank, + initialize_model_parallel, + model_parallel_cuda_manual_seed, + ) + except ImportError: + raise ImportError( + '\n\nPlease install the megatron submodule:' + '\n\n git submodule update --init ' + 'fairseq/model_parallel/megatron' + ) + initialize_model_parallel(args.model_parallel_size) + model_parallel_cuda_manual_seed(args.seed) + model_part_number = get_model_parallel_rank() + args.checkpoint_suffix += '-model_part-{0}'.format(model_part_number) + return args.distributed_rank + + +def distributed_main(i, main, args, kwargs): + args.device_id = i + if torch.cuda.is_available() and not args.cpu and not getattr(args, "tpu", False): + torch.cuda.set_device(args.device_id) + if args.distributed_rank is None: # torch.multiprocessing.spawn + args.distributed_rank = kwargs.pop('start_rank', 0) + i + + args.distributed_rank = distributed_init(args) + + after_distributed_init_fn = kwargs.pop('after_distributed_init_fn', None) + if after_distributed_init_fn: + args = after_distributed_init_fn(args) + + main(args, **kwargs) + + +def call_main(args, main, **kwargs): + if args.distributed_init_method is None: + infer_init_method(args) + + if args.distributed_init_method is not None: + # distributed training + if not args.distributed_no_spawn: + start_rank = args.distributed_rank + args.distributed_rank = None # assign automatically + kwargs['start_rank'] = start_rank + torch.multiprocessing.spawn( + fn=distributed_main, + args=(main, args, kwargs), + nprocs=min( + torch.cuda.device_count(), + args.distributed_world_size, + ), + ) + else: + distributed_main(args.device_id, main, args, kwargs) + elif getattr(args, "tpu", False): + import torch_xla.distributed.xla_multiprocessing as xmp + torch.multiprocessing.set_sharing_strategy("file_system") + xmp.spawn( + fn=distributed_main, + args=(main, args, kwargs), + nprocs=8, # use all 8 TPU cores + ) + else: + # single GPU main + main(args, **kwargs) + + +def get_rank(): + return dist.get_rank() + + +def get_world_size(): + return dist.get_world_size() + + +def get_default_group(): + return dist.group.WORLD + + +def all_reduce(tensor, group=None): + if isinstance(group, tuple) and group[0] == 'tpu': + import torch_xla.core.xla_model as xm + return xm.all_reduce('sum', [tensor], groups=group[1]) + else: + if group is None: + group = get_default_group() + return dist.all_reduce(tensor, group=group) + + +def all_gather_list(data, group=None, max_size=16384): + """Gathers arbitrary data from all nodes into a list. + + Similar to :func:`~torch.distributed.all_gather` but for arbitrary Python + data. Note that *data* must be picklable. + + Args: + data (Any): data from the local worker to be gathered on other workers + group (optional): group of the collective + max_size (int, optional): maximum size of the data to be gathered + across workers + """ + rank = get_rank() + world_size = get_world_size() + + buffer_size = max_size * world_size + if not hasattr(all_gather_list, '_buffer') or \ + all_gather_list._buffer.numel() < buffer_size: + all_gather_list._buffer = torch.cuda.ByteTensor(buffer_size) + all_gather_list._cpu_buffer = torch.ByteTensor(max_size).pin_memory() + buffer = all_gather_list._buffer + buffer.zero_() + cpu_buffer = all_gather_list._cpu_buffer + + data = utils.move_to_cpu(data) + enc = pickle.dumps(data) + enc_size = len(enc) + header_size = 4 # size of header that contains the length of the encoded data + size = header_size + enc_size + if size > max_size: + raise ValueError('encoded data size ({}) exceeds max_size ({})'.format(size, max_size)) + + header = struct.pack(">I", enc_size) + cpu_buffer[:size] = torch.ByteTensor(list(header + enc)) + start = rank * max_size + buffer[start:start + size].copy_(cpu_buffer[:size]) + + all_reduce(buffer, group=group) + + buffer = buffer.cpu() + try: + result = [] + for i in range(world_size): + out_buffer = buffer[i * max_size:(i + 1) * max_size] + enc_size, = struct.unpack(">I", bytes(out_buffer[:header_size].tolist())) + if enc_size > 0: + result.append(pickle.loads(bytes(out_buffer[header_size:header_size + enc_size].tolist()))) + return result + except pickle.UnpicklingError: + raise Exception( + 'Unable to unpickle data from other workers. all_gather_list requires all ' + 'workers to enter the function together, so this error usually indicates ' + 'that the workers have fallen out of sync somehow. Workers can fall out of ' + 'sync if one of them runs out of memory, or if there are other conditions ' + 'in your training script that can cause one worker to finish an epoch ' + 'while other workers are still iterating over their portions of the data. ' + 'Try rerunning with --ddp-backend=no_c10d and see if that helps.' + ) + + +def all_reduce_dict( + data: Mapping[str, Any], + device, + group=None, +) -> Dict[str, Any]: + """ + AllReduce a dictionary of values across workers. We separately + reduce items that are already on the device and items on CPU for + better performance. + + Args: + data (Mapping[str, Any]): dictionary of data to all-reduce, but + cannot be a nested dictionary + device (torch.device): device for the reduction + group (optional): group of the collective + """ + data_keys = list(data.keys()) + + # We want to separately reduce items that are already on the + # device and items on CPU for performance reasons. + cpu_data = OrderedDict() + device_data = OrderedDict() + for k in data_keys: + t = data[k] + if not torch.is_tensor(t): + cpu_data[k] = torch.tensor(t, dtype=torch.double) + elif t.device.type != device.type: + cpu_data[k] = t.to(dtype=torch.double) + else: + device_data[k] = t.to(dtype=torch.double) + + def _all_reduce_dict(data: OrderedDict): + if len(data) == 0: + return data + buf = torch.stack(list(data.values())).to(device=device) + all_reduce(buf, group=group) + return {k: buf[i] for i, k in enumerate(data)} + + cpu_data = _all_reduce_dict(cpu_data) + device_data = _all_reduce_dict(device_data) + + def get_from_stack(key): + if key in cpu_data: + return cpu_data[key] + elif key in device_data: + return device_data[key] + raise KeyError + + return OrderedDict([(key, get_from_stack(key)) for key in data_keys]) diff --git a/fairseq/file_io.py b/fairseq/file_io.py new file mode 100644 index 0000000000000000000000000000000000000000..b57373f8b51ab23394296346c3e2a6a97a891f78 --- /dev/null +++ b/fairseq/file_io.py @@ -0,0 +1,106 @@ +#!/usr/bin/env python3 + +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +import shutil +from typing import List, Optional + + +try: + from fvcore.common.file_io import PathManager as FVCorePathManager + +except ImportError: + FVCorePathManager = None + + +class PathManager: + """ + Wrapper for insulating OSS I/O (using Python builtin operations) from + fvcore's PathManager abstraction (for transparently handling various + internal backends). + """ + + @staticmethod + def open( + path: str, + mode: str = "r", + buffering: int = -1, + encoding: Optional[str] = None, + errors: Optional[str] = None, + newline: Optional[str] = None, + ): + if FVCorePathManager: + return FVCorePathManager.open( + path=path, + mode=mode, + buffering=buffering, + encoding=encoding, + errors=errors, + newline=newline, + ) + return open( + path, + mode=mode, + buffering=buffering, + encoding=encoding, + errors=errors, + newline=newline, + ) + + @staticmethod + def copy(src_path: str, dst_path: str, overwrite: bool = False) -> bool: + if FVCorePathManager: + return FVCorePathManager.copy( + src_path=src_path, dst_path=dst_path, overwrite=overwrite + ) + return shutil.copyfile(src_path, dst_path) + + @staticmethod + def get_local_path(path: str, **kwargs) -> str: + if FVCorePathManager: + return FVCorePathManager.get_local_path(path, **kwargs) + return path + + @staticmethod + def exists(path: str) -> bool: + if FVCorePathManager: + return FVCorePathManager.exists(path) + return os.path.exists(path) + + @staticmethod + def isfile(path: str) -> bool: + if FVCorePathManager: + return FVCorePathManager.isfile(path) + return os.path.isfile(path) + + @staticmethod + def ls(path: str) -> List[str]: + if FVCorePathManager: + return FVCorePathManager.ls(path) + return os.listdir(path) + + @staticmethod + def mkdirs(path: str) -> None: + if FVCorePathManager: + return FVCorePathManager.mkdirs(path) + os.makedirs(path, exist_ok=True) + + @staticmethod + def rm(path: str) -> None: + if FVCorePathManager: + return FVCorePathManager.rm(path) + os.remove(path) + + @staticmethod + def chmod(path: str, mode: int) -> None: + if "manifold" not in path: + os.chmod(path, mode) + + @staticmethod + def register_handler(handler) -> None: + if FVCorePathManager: + return FVCorePathManager.register_handler(handler=handler) diff --git a/fairseq/file_utils.py b/fairseq/file_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..62278b367d659dc1c565428f903832d3a882848f --- /dev/null +++ b/fairseq/file_utils.py @@ -0,0 +1,331 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Utilities for working with the local dataset cache. +This file is adapted from `AllenNLP `_. +and `huggingface `_. +""" + +import fnmatch +from functools import wraps, partial +from hashlib import sha256 +from io import open +import json +import logging +import os +import shutil +import tarfile +import tempfile + + +try: + from torch.hub import _get_torch_home + torch_cache_home = _get_torch_home() +except ImportError: + torch_cache_home = os.path.expanduser( + os.getenv('TORCH_HOME', os.path.join( + os.getenv('XDG_CACHE_HOME', '~/.cache'), 'torch'))) +default_cache_path = os.path.join(torch_cache_home, 'pytorch_fairseq') + +try: + from urllib.parse import urlparse +except ImportError: + from urlparse import urlparse + +try: + from pathlib import Path + PYTORCH_FAIRSEQ_CACHE = Path( + os.getenv('PYTORCH_FAIRSEQ_CACHE', default_cache_path)) +except (AttributeError, ImportError): + PYTORCH_FAIRSEQ_CACHE = os.getenv( + 'PYTORCH_FAIRSEQ_CACHE', default_cache_path) + +CONFIG_NAME = "config.json" +WEIGHTS_NAME = "pytorch_model.bin" + +logger = logging.getLogger(__name__) # pylint: disable=invalid-name + + +def load_archive_file(archive_file): + # redirect to the cache, if necessary + try: + resolved_archive_file = cached_path(archive_file, cache_dir=None) + except EnvironmentError: + logger.info( + "Archive name '{}' was not found in archive name list. " + "We assumed '{}' was a path or URL but couldn't find any file " + "associated to this path or URL.".format( + archive_file, + archive_file, + ) + ) + return None + + if resolved_archive_file == archive_file: + logger.info("loading archive file {}".format(archive_file)) + else: + logger.info("loading archive file {} from cache at {}".format( + archive_file, resolved_archive_file)) + + # Extract archive to temp dir and replace .tar.bz2 if necessary + tempdir = None + if not os.path.isdir(resolved_archive_file): + tempdir = tempfile.mkdtemp() + logger.info("extracting archive file {} to temp dir {}".format( + resolved_archive_file, tempdir)) + ext = os.path.splitext(archive_file)[1][1:] + with tarfile.open(resolved_archive_file, 'r:' + ext) as archive: + top_dir = os.path.commonprefix(archive.getnames()) + archive.extractall(tempdir) + os.remove(resolved_archive_file) + shutil.move(os.path.join(tempdir, top_dir), resolved_archive_file) + shutil.rmtree(tempdir) + + return resolved_archive_file + + +def url_to_filename(url, etag=None): + """ + Convert `url` into a hashed filename in a repeatable way. + If `etag` is specified, append its hash to the URL's, delimited + by a period. + """ + url_bytes = url.encode('utf-8') + url_hash = sha256(url_bytes) + filename = url_hash.hexdigest() + + if etag: + etag_bytes = etag.encode('utf-8') + etag_hash = sha256(etag_bytes) + filename += '.' + etag_hash.hexdigest() + + return filename + + +def filename_to_url(filename, cache_dir=None): + """ + Return the url and etag (which may be ``None``) stored for `filename`. + Raise ``EnvironmentError`` if `filename` or its stored metadata do not exist. + """ + if cache_dir is None: + cache_dir = PYTORCH_FAIRSEQ_CACHE + if isinstance(cache_dir, Path): + cache_dir = str(cache_dir) + + cache_path = os.path.join(cache_dir, filename) + if not os.path.exists(cache_path): + raise EnvironmentError("file {} not found".format(cache_path)) + + meta_path = cache_path + '.json' + if not os.path.exists(meta_path): + raise EnvironmentError("file {} not found".format(meta_path)) + + with open(meta_path, encoding="utf-8") as meta_file: + metadata = json.load(meta_file) + url = metadata['url'] + etag = metadata['etag'] + + return url, etag + + +def cached_path(url_or_filename, cache_dir=None): + """ + Given something that might be a URL (or might be a local path), + determine which. If it's a URL, download the file and cache it, and + return the path to the cached file. If it's already a local path, + make sure the file exists and then return the path. + """ + if cache_dir is None: + cache_dir = PYTORCH_FAIRSEQ_CACHE + if isinstance(url_or_filename, Path): + url_or_filename = str(url_or_filename) + if isinstance(cache_dir, Path): + cache_dir = str(cache_dir) + + parsed = urlparse(url_or_filename) + + if parsed.scheme in ('http', 'https', 's3'): + # URL, so get it from the cache (downloading if necessary) + return get_from_cache(url_or_filename, cache_dir) + elif os.path.exists(url_or_filename): + # File, and it exists. + return url_or_filename + elif parsed.scheme == '': + # File, but it doesn't exist. + raise EnvironmentError("file {} not found".format(url_or_filename)) + else: + # Something unknown + raise ValueError("unable to parse {} as a URL or as a local path".format(url_or_filename)) + + +def split_s3_path(url): + """Split a full s3 path into the bucket name and path.""" + parsed = urlparse(url) + if not parsed.netloc or not parsed.path: + raise ValueError("bad s3 path {}".format(url)) + bucket_name = parsed.netloc + s3_path = parsed.path + # Remove '/' at beginning of path. + if s3_path.startswith("/"): + s3_path = s3_path[1:] + return bucket_name, s3_path + + +def s3_request(func): + """ + Wrapper function for s3 requests in order to create more helpful error + messages. + """ + + @wraps(func) + def wrapper(url, *args, **kwargs): + from botocore.exceptions import ClientError + try: + return func(url, *args, **kwargs) + except ClientError as exc: + if int(exc.response["Error"]["Code"]) == 404: + raise EnvironmentError("file {} not found".format(url)) + else: + raise + + return wrapper + + +@s3_request +def s3_etag(url): + """Check ETag on S3 object.""" + import boto3 + s3_resource = boto3.resource("s3") + bucket_name, s3_path = split_s3_path(url) + s3_object = s3_resource.Object(bucket_name, s3_path) + return s3_object.e_tag + + +@s3_request +def s3_get(url, temp_file): + """Pull a file directly from S3.""" + import boto3 + s3_resource = boto3.resource("s3") + bucket_name, s3_path = split_s3_path(url) + s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file) + + +def request_wrap_timeout(func, url): + import requests + for attempt, timeout in enumerate([10, 20, 40, 60, 60]): + try: + return func(timeout=timeout) + except requests.exceptions.Timeout as e: + logger.warning("Request for %s timed-out (attempt %d). Retrying with a timeout of %d secs", + url, attempt, timeout, exc_info=e) + continue + raise RuntimeError(f"Unable to fetch file {url}") + + +def http_get(url, temp_file): + import requests + from tqdm import tqdm + + req = request_wrap_timeout(partial(requests.get, url, stream=True), url) + content_length = req.headers.get('Content-Length') + total = int(content_length) if content_length is not None else None + progress = tqdm(unit="B", total=total) + for chunk in req.iter_content(chunk_size=1024): + if chunk: # filter out keep-alive new chunks + progress.update(len(chunk)) + temp_file.write(chunk) + progress.close() + + +def get_from_cache(url, cache_dir=None): + """ + Given a URL, look for the corresponding dataset in the local cache. + If it's not there, download it. Then return the path to the cached file. + """ + if cache_dir is None: + cache_dir = PYTORCH_FAIRSEQ_CACHE + if isinstance(cache_dir, Path): + cache_dir = str(cache_dir) + + if not os.path.exists(cache_dir): + os.makedirs(cache_dir) + + # Get eTag to add to filename, if it exists. + if url.startswith("s3://"): + etag = s3_etag(url) + else: + try: + import requests + response = request_wrap_timeout(partial(requests.head, url, allow_redirects=True), url) + if response.status_code != 200: + etag = None + else: + etag = response.headers.get("ETag") + except EnvironmentError: + etag = None + + filename = url_to_filename(url, etag) + + # get cache path to put the file + cache_path = os.path.join(cache_dir, filename) + + # If we don't have a connection (etag is None) and can't identify the file + # try to get the last downloaded one + if not os.path.exists(cache_path) and etag is None: + matching_files = fnmatch.filter(os.listdir(cache_dir), filename + '.*') + matching_files = list(filter(lambda s: not s.endswith('.json'), matching_files)) + if matching_files: + cache_path = os.path.join(cache_dir, matching_files[-1]) + + if not os.path.exists(cache_path): + # Download to temporary file, then copy to cache dir once finished. + # Otherwise you get corrupt cache entries if the download gets interrupted. + with tempfile.NamedTemporaryFile() as temp_file: + logger.info("%s not found in cache, downloading to %s", url, temp_file.name) + + # GET file object + if url.startswith("s3://"): + s3_get(url, temp_file) + else: + http_get(url, temp_file) + + # we are copying the file before closing it, so flush to avoid truncation + temp_file.flush() + # shutil.copyfileobj() starts at the current position, so go to the start + temp_file.seek(0) + + logger.info("copying %s to cache at %s", temp_file.name, cache_path) + with open(cache_path, 'wb') as cache_file: + shutil.copyfileobj(temp_file, cache_file) + + logger.info("creating metadata file for %s", cache_path) + meta = {'url': url, 'etag': etag} + meta_path = cache_path + '.json' + with open(meta_path, 'w') as meta_file: + output_string = json.dumps(meta) + meta_file.write(output_string) + + logger.info("removing temp file %s", temp_file.name) + + return cache_path + + +def read_set_from_file(filename): + ''' + Extract a de-duped collection (set) of text from a file. + Expected file format is one item per line. + ''' + collection = set() + with open(filename, 'r', encoding='utf-8') as file_: + for line in file_: + collection.add(line.rstrip()) + return collection + + +def get_file_extension(path, dot=True, lower=True): + ext = os.path.splitext(path)[1] + ext = ext if dot else ext[1:] + return ext.lower() if lower else ext diff --git a/fairseq/hub_utils.py b/fairseq/hub_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..e040a8c3f348e6865317e45e4fb20a6aae9cc16f --- /dev/null +++ b/fairseq/hub_utils.py @@ -0,0 +1,269 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import copy +import logging +import os +from typing import List, Dict, Iterator, Tuple, Any + +import torch +from torch import nn + +from fairseq import utils +from fairseq.data import encoders + + +logger = logging.getLogger(__name__) + + +def from_pretrained( + model_name_or_path, + checkpoint_file='model.pt', + data_name_or_path='.', + archive_map=None, + **kwargs +): + from fairseq import checkpoint_utils, file_utils + + if archive_map is not None: + if model_name_or_path in archive_map: + model_name_or_path = archive_map[model_name_or_path] + if data_name_or_path is not None and data_name_or_path in archive_map: + data_name_or_path = archive_map[data_name_or_path] + + # allow archive_map to set default arg_overrides (e.g., tokenizer, bpe) + # for each model + if isinstance(model_name_or_path, dict): + for k, v in model_name_or_path.items(): + if k == 'checkpoint_file': + checkpoint_file = v + elif ( + k != 'path' + # only set kwargs that don't already have overrides + and k not in kwargs + ): + kwargs[k] = v + model_name_or_path = model_name_or_path['path'] + + model_path = file_utils.load_archive_file(model_name_or_path) + + # convenience hack for loading data and BPE codes from model archive + if data_name_or_path.startswith('.'): + kwargs['data'] = os.path.abspath(os.path.join(model_path, data_name_or_path)) + else: + kwargs['data'] = file_utils.load_archive_file(data_name_or_path) + for file, arg in { + 'code': 'bpe_codes', + 'bpecodes': 'bpe_codes', + 'sentencepiece.bpe.model': 'sentencepiece_model', + }.items(): + path = os.path.join(model_path, file) + if os.path.exists(path): + kwargs[arg] = path + + if 'user_dir' in kwargs: + utils.import_user_module(argparse.Namespace(user_dir=kwargs['user_dir'])) + + models, args, task = checkpoint_utils.load_model_ensemble_and_task( + [os.path.join(model_path, cpt) for cpt in checkpoint_file.split(os.pathsep)], + arg_overrides=kwargs, + ) + + return { + 'args': args, + 'task': task, + 'models': models, + } + + +class GeneratorHubInterface(nn.Module): + """ + PyTorch Hub interface for generating sequences from a pre-trained + translation or language model. + """ + + def __init__(self, args, task, models): + super().__init__() + self.args = args + self.task = task + self.models = nn.ModuleList(models) + self.src_dict = task.source_dictionary + self.tgt_dict = task.target_dictionary + + # optimize model for generation + for model in self.models: + model.prepare_for_inference_(args) + + # Load alignment dictionary for unknown word replacement + # (None if no unknown word replacement, empty if no path to align dictionary) + self.align_dict = utils.load_align_dict(getattr(args, 'replace_unk', None)) + + self.tokenizer = encoders.build_tokenizer(args) + self.bpe = encoders.build_bpe(args) + + self.max_positions = utils.resolve_max_positions( + self.task.max_positions(), *[model.max_positions() for model in models] + ) + + # this is useful for determining the device + self.register_buffer('_float_tensor', torch.tensor([0], dtype=torch.float)) + + @property + def device(self): + return self._float_tensor.device + + def translate(self, sentences: List[str], beam: int = 5, verbose: bool = False, **kwargs) -> List[str]: + return self.sample(sentences, beam, verbose, **kwargs) + + def sample(self, sentences: List[str], beam: int = 1, verbose: bool = False, **kwargs) -> List[str]: + if isinstance(sentences, str): + return self.sample([sentences], beam=beam, verbose=verbose, **kwargs)[0] + tokenized_sentences = [self.encode(sentence) for sentence in sentences] + batched_hypos = self.generate(tokenized_sentences, beam, verbose, **kwargs) + return [self.decode(hypos[0]['tokens']) for hypos in batched_hypos] + + def score(self, sentences: List[str], **kwargs): + if isinstance(sentences, str): + return self.score([sentences], **kwargs)[0] + # NOTE: this doesn't support translation tasks currently + tokenized_sentences = [self.encode(sentence) for sentence in sentences] + return [hypos[0] for hypos in self.generate(tokenized_sentences, score_reference=True, **kwargs)] + + def generate( + self, + tokenized_sentences: List[torch.LongTensor], + beam: int = 5, + verbose: bool = False, + skip_invalid_size_inputs=False, + inference_step_args=None, + **kwargs + ) -> List[List[Dict[str, torch.Tensor]]]: + if torch.is_tensor(tokenized_sentences) and tokenized_sentences.dim() == 1: + return self.generate( + tokenized_sentences.unsqueeze(0), beam=beam, verbose=verbose, **kwargs + )[0] + + # build generator using current args as well as any kwargs + gen_args = copy.copy(self.args) + gen_args.beam = beam + for k, v in kwargs.items(): + setattr(gen_args, k, v) + generator = self.task.build_generator(self.models, gen_args) + + inference_step_args = inference_step_args or {} + results = [] + for batch in self._build_batches(tokenized_sentences, skip_invalid_size_inputs): + batch = utils.apply_to_sample(lambda t: t.to(self.device), batch) + translations = self.task.inference_step( + generator, self.models, batch, **inference_step_args + ) + for id, hypos in zip(batch["id"].tolist(), translations): + results.append((id, hypos)) + + # sort output to match input order + outputs = [hypos for _, hypos in sorted(results, key=lambda x: x[0])] + + if verbose: + + def getarg(name, default): + return getattr(gen_args, name, getattr(self.args, name, default)) + + for source_tokens, target_hypotheses in zip(tokenized_sentences, outputs): + src_str_with_unk = self.string(source_tokens) + logger.info('S\t{}'.format(src_str_with_unk)) + for hypo in target_hypotheses: + hypo_str = self.decode(hypo['tokens']) + logger.info('H\t{}\t{}'.format(hypo['score'], hypo_str)) + logger.info('P\t{}'.format( + ' '.join(map(lambda x: '{:.4f}'.format(x), hypo['positional_scores'].tolist())) + )) + if hypo['alignment'] is not None and getarg('print_alignment', False): + logger.info('A\t{}'.format( + ' '.join(['{}-{}'.format(src_idx, tgt_idx) for src_idx, tgt_idx in hypo['alignment']]) + )) + return outputs + + def encode(self, sentence: str) -> torch.LongTensor: + sentence = self.tokenize(sentence) + sentence = self.apply_bpe(sentence) + return self.binarize(sentence) + + def decode(self, tokens: torch.LongTensor) -> str: + sentence = self.string(tokens) + sentence = self.remove_bpe(sentence) + return self.detokenize(sentence) + + def tokenize(self, sentence: str) -> str: + if self.tokenizer is not None: + sentence = self.tokenizer.encode(sentence) + return sentence + + def detokenize(self, sentence: str) -> str: + if self.tokenizer is not None: + sentence = self.tokenizer.decode(sentence) + return sentence + + def apply_bpe(self, sentence: str) -> str: + if self.bpe is not None: + sentence = self.bpe.encode(sentence) + return sentence + + def remove_bpe(self, sentence: str) -> str: + if self.bpe is not None: + sentence = self.bpe.decode(sentence) + return sentence + + def binarize(self, sentence: str) -> torch.LongTensor: + return self.src_dict.encode_line(sentence, add_if_not_exist=False).long() + + def string(self, tokens: torch.LongTensor) -> str: + return self.tgt_dict.string(tokens) + + def _build_batches( + self, tokens: List[List[int]], skip_invalid_size_inputs: bool + ) -> Iterator[Dict[str, Any]]: + lengths = torch.LongTensor([t.numel() for t in tokens]) + batch_iterator = self.task.get_batch_iterator( + dataset=self.task.build_dataset_for_inference(tokens, lengths), + max_tokens=self.args.max_tokens, + max_sentences=self.args.max_sentences, + max_positions=self.max_positions, + ignore_invalid_inputs=skip_invalid_size_inputs, + ).next_epoch_itr(shuffle=False) + return batch_iterator + + +class BPEHubInterface(object): + """PyTorch Hub interface for Byte-Pair Encoding (BPE).""" + + def __init__(self, bpe, **kwargs): + super().__init__() + args = argparse.Namespace(bpe=bpe, **kwargs) + self.bpe = encoders.build_bpe(args) + assert self.bpe is not None + + def encode(self, sentence: str) -> str: + return self.bpe.encode(sentence) + + def decode(self, sentence: str) -> str: + return self.bpe.decode(sentence) + + +class TokenizerHubInterface(object): + """PyTorch Hub interface for tokenization.""" + + def __init__(self, tokenizer, **kwargs): + super().__init__() + args = argparse.Namespace(tokenizer=tokenizer, **kwargs) + self.tokenizer = encoders.build_tokenizer(args) + assert self.tokenizer is not None + + def encode(self, sentence: str) -> str: + return self.tokenizer.encode(sentence) + + def decode(self, sentence: str) -> str: + return self.tokenizer.decode(sentence) diff --git a/fairseq/incremental_decoding_utils.py b/fairseq/incremental_decoding_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..91128e8879673b57eab26b00680155a567e12907 --- /dev/null +++ b/fairseq/incremental_decoding_utils.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional +import uuid + +from torch import Tensor + + +class FairseqIncrementalState(object): + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.init_incremental_state() + + def init_incremental_state(self): + self._incremental_state_id = str(uuid.uuid4()) + + def _get_full_incremental_state_key(self, key: str) -> str: + return "{}.{}".format(self._incremental_state_id, key) + + def get_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, + ) -> Optional[Dict[str, Optional[Tensor]]]: + """Helper for getting incremental state for an nn.Module.""" + full_key = self._get_full_incremental_state_key(key) + if incremental_state is None or full_key not in incremental_state: + return None + return incremental_state[full_key] + + def set_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, + value: Dict[str, Optional[Tensor]], + ) -> Optional[Dict[str, Dict[str, Optional[Tensor]]]]: + """Helper for setting incremental state for an nn.Module.""" + if incremental_state is not None: + full_key = self._get_full_incremental_state_key(key) + incremental_state[full_key] = value + return incremental_state + + +def with_incremental_state(cls): + cls.__bases__ = (FairseqIncrementalState,) + tuple(b for b in cls.__bases__ if b != FairseqIncrementalState) + return cls diff --git a/fairseq/iterative_refinement_generator.py b/fairseq/iterative_refinement_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..c7a267d2589e3fa03404b778634b99dea7809426 --- /dev/null +++ b/fairseq/iterative_refinement_generator.py @@ -0,0 +1,315 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import namedtuple + +import torch +import numpy as np + +from fairseq import utils + + +DecoderOut = namedtuple('IterativeRefinementDecoderOut', [ + 'output_tokens', + 'output_scores', + 'attn', + 'step', + 'max_step', + 'history' +]) + + +class IterativeRefinementGenerator(object): + def __init__( + self, + tgt_dict, + models=None, + eos_penalty=0.0, + max_iter=10, + max_ratio=2, + beam_size=1, + decoding_format=None, + retain_dropout=False, + adaptive=True, + retain_history=False, + reranking=False, + ): + """ + Generates translations based on iterative refinement. + + Args: + tgt_dict: target dictionary + eos_penalty: if > 0.0, it penalized early-stopping in decoding + max_iter: maximum number of refinement iterations + max_ratio: generate sequences of maximum length ax, where x is the source length + decoding_format: decoding mode in {'unigram', 'ensemble', 'vote', 'dp', 'bs'} + retain_dropout: retaining dropout in the inference + adaptive: decoding with early stop + """ + self.bos = tgt_dict.bos() + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() + self.vocab_size = len(tgt_dict) + self.eos_penalty = eos_penalty + self.max_iter = max_iter + self.max_ratio = max_ratio + self.beam_size = beam_size + self.reranking = reranking + self.decoding_format = decoding_format + self.retain_dropout = retain_dropout + self.retain_history = retain_history + self.adaptive = adaptive + self.models = models + + def generate_batched_itr( + self, + data_itr, + maxlen_a=None, + maxlen_b=None, + cuda=False, + timer=None, + prefix_size=0, + ): + """Iterate over a batched dataset and yield individual translations. + + Args: + maxlen_a/b: generate sequences of maximum length ax + b, + where x is the source sentence length. + cuda: use GPU for generation + timer: StopwatchMeter for timing generations. + """ + + for sample in data_itr: + if "net_input" not in sample: + continue + if timer is not None: + timer.start() + with torch.no_grad(): + hypos = self.generate( + self.models, + sample, + prefix_tokens=sample["target"][:, :prefix_size] + if prefix_size > 0 + else None, + ) + if timer is not None: + timer.stop(sample["ntokens"]) + for i, id in enumerate(sample["id"]): + # remove padding + src = utils.strip_pad(sample["net_input"]["src_tokens"][i, :], self.pad) + ref = utils.strip_pad(sample["target"][i, :], self.pad) + yield id, src, ref, hypos[i] + + + @torch.no_grad() + def generate(self, models, sample, prefix_tokens=None): + + # TODO: iterative refinement generator does not support ensemble for now. + if not self.retain_dropout: + for model in models: + model.eval() + + model, reranker = models[0], None + if self.reranking: + assert len(models) > 1, "Assuming the last checkpoint is the reranker" + assert self.beam_size > 1, "Reranking requires multiple translation for each example" + + reranker = models[-1] + models = models[:-1] + + if len(models) > 1 and hasattr(model, 'enable_ensemble'): + assert model.allow_ensemble, "{} does not support ensembling".format(model.__class__.__name__) + model.enable_ensemble(models) + + # TODO: better encoder inputs? + src_tokens = sample["net_input"]["src_tokens"] + src_lengths = sample["net_input"]["src_lengths"] + bsz, src_len = src_tokens.size() + + # initialize + encoder_out = model.forward_encoder([src_tokens, src_lengths]) + prev_decoder_out = model.initialize_output_tokens(encoder_out, src_tokens) + + if self.beam_size > 1: + assert model.allow_length_beam, \ + "{} does not support decoding with length beam.".format(model.__class__.__name__) + + # regenerate data based on length-beam + length_beam_order = utils.new_arange(src_tokens, self.beam_size, bsz).t().reshape(-1) + encoder_out = model.encoder.reorder_encoder_out(encoder_out, length_beam_order) + prev_decoder_out = model.regenerate_length_beam(prev_decoder_out, self.beam_size) + bsz = bsz * self.beam_size + + sent_idxs = torch.arange(bsz) + prev_output_tokens = prev_decoder_out.output_tokens.clone() + + if self.retain_history: + prev_decoder_out = prev_decoder_out._replace(history=[prev_output_tokens]) + + finalized = [[] for _ in range(bsz)] + + def is_a_loop(x, y, s, a): + b, l_x, l_y = x.size(0), x.size(1), y.size(1) + if l_x > l_y: + y = torch.cat([y, x.new_zeros(b, l_x - l_y).fill_(self.pad)], 1) + s = torch.cat([s, s.new_zeros(b, l_x - l_y)], 1) + if a is not None: + a = torch.cat([a, a.new_zeros(b, l_x - l_y, a.size(2))], 1) + elif l_x < l_y: + x = torch.cat([x, y.new_zeros(b, l_y - l_x).fill_(self.pad)], 1) + return (x == y).all(1), y, s, a + + def finalized_hypos(step, prev_out_token, prev_out_score, prev_out_attn): + cutoff = prev_out_token.ne(self.pad) + tokens = prev_out_token[cutoff] + if prev_out_score is None: + scores, score = None, None + else: + scores = prev_out_score[cutoff] + score = scores.mean() + + if prev_out_attn is None: + hypo_attn, alignment = None, None + else: + hypo_attn = prev_out_attn[cutoff] + alignment = hypo_attn.max(dim=1)[1] + return { + "steps": step, + "tokens": tokens, + "positional_scores": scores, + "score": score, + "hypo_attn": hypo_attn, + "alignment": alignment, + } + + for step in range(self.max_iter + 1): + + decoder_options = { + "eos_penalty": self.eos_penalty, + "max_ratio": self.max_ratio, + "decoding_format": self.decoding_format, + } + prev_decoder_out = prev_decoder_out._replace( + step=step, + max_step=self.max_iter + 1, + ) + + decoder_out = model.forward_decoder( + prev_decoder_out, encoder_out, **decoder_options + ) + + if self.adaptive: + # terminate if there is a loop + terminated, out_tokens, out_scores, out_attn = is_a_loop( + prev_output_tokens, decoder_out.output_tokens, decoder_out.output_scores, decoder_out.attn + ) + decoder_out = decoder_out._replace( + output_tokens=out_tokens, + output_scores=out_scores, + attn=out_attn, + ) + + else: + terminated = decoder_out.output_tokens.new_zeros(decoder_out.output_tokens.size(0)).bool() + + if step == self.max_iter: # reach last iteration, terminate + terminated.fill_(1) + + # collect finalized sentences + finalized_idxs = sent_idxs[terminated] + finalized_tokens = decoder_out.output_tokens[terminated] + finalized_scores = decoder_out.output_scores[terminated] + finalized_attn = ( + None if (decoder_out.attn is None or decoder_out.attn.size(0) == 0) else decoder_out.attn[terminated] + ) + + if self.retain_history: + finalized_history_tokens = [h[terminated] for h in decoder_out.history] + + for i in range(finalized_idxs.size(0)): + finalized[finalized_idxs[i]] = [ + finalized_hypos( + step, + finalized_tokens[i], + finalized_scores[i], + None if finalized_attn is None else finalized_attn[i], + ) + ] + + if self.retain_history: + finalized[finalized_idxs[i]][0]['history'] = [] + for j in range(len(finalized_history_tokens)): + finalized[finalized_idxs[i]][0]['history'].append( + finalized_hypos( + step, + finalized_history_tokens[j][i], + None, None + ) + ) + + # check if all terminated + if terminated.sum() == terminated.size(0): + break + + # for next step + not_terminated = ~terminated + prev_decoder_out = decoder_out._replace( + output_tokens=decoder_out.output_tokens[not_terminated], + output_scores=decoder_out.output_scores[not_terminated], + attn=decoder_out.attn[not_terminated] + if (decoder_out.attn is not None and decoder_out.attn.size(0) > 0) + else None, + history=[h[not_terminated] for h in decoder_out.history] + if decoder_out.history is not None + else None, + ) + encoder_out = model.encoder.reorder_encoder_out(encoder_out, not_terminated.nonzero().squeeze()) + sent_idxs = sent_idxs[not_terminated] + prev_output_tokens = prev_decoder_out.output_tokens.clone() + + if self.beam_size > 1: + if reranker is not None: + finalized = self.rerank( + reranker, finalized, [src_tokens, src_lengths], self.beam_size + ) + + # aggregate information from length beam + finalized = [ + finalized[np.argmax( + [finalized[self.beam_size * i + j][0]['score'] for j in range(self.beam_size)] + ) + self.beam_size * i] for i in range(len(finalized) // self.beam_size) + ] + + return finalized + + def rerank(self, reranker, finalized, encoder_input, beam_size): + + def rebuild_batch(finalized): + finalized_tokens = [f[0]['tokens'] for f in finalized] + finalized_maxlen = max(f.size(0) for f in finalized_tokens) + final_output_tokens = finalized_tokens[0].new_zeros(len(finalized_tokens), finalized_maxlen).fill_(self.pad) + for i, f in enumerate(finalized_tokens): + final_output_tokens[i, :f.size(0)] = f + return final_output_tokens + + final_output_tokens = rebuild_batch(finalized) + final_output_tokens[:, 0] = self.eos # autoregressive model assumes starting with EOS + + reranker_encoder_out = reranker.encoder(*encoder_input) + length_beam_order = utils.new_arange( + final_output_tokens, beam_size, reranker_encoder_out.encoder_out.size(1)).t().reshape(-1) + reranker_encoder_out = reranker.encoder.reorder_encoder_out(reranker_encoder_out, length_beam_order) + reranking_scores = reranker.get_normalized_probs( + reranker.decoder(final_output_tokens[:, :-1], reranker_encoder_out), True, None) + reranking_scores = reranking_scores.gather(2, final_output_tokens[:, 1:, None]) + reranking_masks = final_output_tokens[:, 1:].ne(self.pad) + reranking_scores = reranking_scores[:, :, 0].masked_fill_(~reranking_masks, 0).sum(1) + reranking_scores = reranking_scores / reranking_masks.sum(1).type_as(reranking_scores) + + for i in range(len(finalized)): + finalized[i][0]['score'] = reranking_scores[i] + + return finalized diff --git a/fairseq/legacy_distributed_data_parallel.py b/fairseq/legacy_distributed_data_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..9832f2c97a43dda2cf01ce5f467cdc60aa439207 --- /dev/null +++ b/fairseq/legacy_distributed_data_parallel.py @@ -0,0 +1,170 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +A modified version of the legacy DistributedDataParallel module that uses c10d +communication primitives. This version is simpler than the latest PyTorch +version and is useful for debugging. Notably it does not overlap gradient +communication with the backward pass, which makes it slower but more robust +than the PyTorch version. + +This version also supports the *no_sync* context manager, which allows faster +training with `--update-freq`. +""" + +from collections import OrderedDict +from contextlib import contextmanager +import copy + +import torch +from torch import nn +from torch.autograd import Variable + +from . import distributed_utils + + +class LegacyDistributedDataParallel(nn.Module): + """Implements distributed data parallelism at the module level. + + A simplified version of :class:`torch.nn.parallel.DistributedDataParallel`. + This version uses a c10d process group for communication and does not + broadcast buffers. + + Args: + module (~torch.nn.Module): module to be parallelized + world_size (int): number of parallel workers + process_group (optional): the c10d process group to be used for + distributed data all-reduction. If None, the default process group + will be used. + buffer_size (int, optional): number of elements to buffer before + performing all-reduce (default: 256M). + """ + + def __init__(self, module, world_size, process_group=None, buffer_size=2**28): + super().__init__() + + self.module = module + self.world_size = world_size + self.process_group = process_group + + # Never use a bigger buffer than the number of model params + self.buffer_size = min(buffer_size, sum(p.numel() for p in module.parameters())) + self.buffer = None + + # We can also forcibly accumulate grads locally and only do the + # all-reduce at some later time + self.accumulate_grads = False + + # make per-device lists of parameters + paramlists = OrderedDict() + for param in self.module.parameters(): + device = param.device + if paramlists.get(device) is None: + paramlists[device] = [] + paramlists[device] += [param] + self.per_device_params = list(paramlists.values()) + + + def __getstate__(self): + attrs = copy.copy(self.__dict__) + return attrs + + def __setstate__(self, state): + super().__setstate__(state) + + @contextmanager + def no_sync(self): + """A context manager to disable gradient synchronization.""" + old_accumulate_grads = self.accumulate_grads + self.accumulate_grads = True + yield + self.accumulate_grads = old_accumulate_grads + + def forward(self, *inputs, **kwargs): + return self.module(*inputs, **kwargs) + + def all_reduce(self): + """ + This function must be called explicitly after backward to reduce + gradients. There is no automatic hook like c10d. + """ + + def all_reduce_params(params): + buffer = self.buffer + nonzero_buffer = False + if len(params) > 1: + offset = 0 + for p in params: + sz = p.numel() + if p.grad is not None: + buffer[offset:offset+sz].copy_(p.grad.data.view(-1)) + nonzero_buffer = True + else: + buffer[offset:offset+sz].zero_() + offset += sz + else: + # we only have a single grad to all-reduce + p = params[0] + if p.grad is not None: + buffer = p.grad.data + nonzero_buffer = True + elif p.numel() <= self.buffer.numel(): + buffer = buffer[:p.numel()] + buffer.zero_() + else: + buffer = torch.zeros_like(p) + + if nonzero_buffer: + buffer.div_(self.world_size) + + distributed_utils.all_reduce(buffer, self.process_group) + + # copy all-reduced grads back into their original place + offset = 0 + for p in params: + sz = p.numel() + if p.grad is not None: + p.grad.data.copy_(buffer[offset:offset+sz].view_as(p)) + else: + p.grad = buffer[offset:offset+sz].view_as(p).clone() + offset += sz + + def reduction_fn(): + # This function only needs to be called once + if self.accumulate_grads: + return + + if self.buffer is None: + self.buffer = next(self.module.parameters()).new(self.buffer_size) + + for params in self.per_device_params: + # All-reduce the gradients in buckets + offset = 0 + buffered_params = [] + for param in params: + if not param.requires_grad: + continue + if param.grad is None: + param.grad = torch.zeros_like(param) + if param.grad.requires_grad: + raise RuntimeError("DistributedDataParallel only works " + "with gradients that don't require " + "grad") + sz = param.numel() + if sz > self.buffer.numel(): + # all-reduce big params directly + all_reduce_params([param]) + else: + if offset + sz > self.buffer.numel(): + all_reduce_params(buffered_params) + offset = 0 + buffered_params.clear() + buffered_params.append(param) + offset += sz + + if len(buffered_params) > 0: + all_reduce_params(buffered_params) + + reduction_fn() diff --git a/fairseq/libbleu.cpython-310-x86_64-linux-gnu.so b/fairseq/libbleu.cpython-310-x86_64-linux-gnu.so new file mode 100644 index 0000000000000000000000000000000000000000..c4b4489d1ea9f43ec7fa2c50ec0ff337eb50ac4c Binary files /dev/null and b/fairseq/libbleu.cpython-310-x86_64-linux-gnu.so differ diff --git a/fairseq/logging/__init__.py b/fairseq/logging/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fairseq/logging/__pycache__/__init__.cpython-310.pyc b/fairseq/logging/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..24e0482907e608a5f9b1dafd593fb91bc2b84567 Binary files /dev/null and b/fairseq/logging/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/logging/__pycache__/meters.cpython-310.pyc b/fairseq/logging/__pycache__/meters.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ff6282099c6a88842b8eec01843fa948ac992729 Binary files /dev/null and b/fairseq/logging/__pycache__/meters.cpython-310.pyc differ diff --git a/fairseq/logging/__pycache__/metrics.cpython-310.pyc b/fairseq/logging/__pycache__/metrics.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..930920f6fde1973d76d4d1173fa123fa100e008e Binary files /dev/null and b/fairseq/logging/__pycache__/metrics.cpython-310.pyc differ diff --git a/fairseq/logging/__pycache__/progress_bar.cpython-310.pyc b/fairseq/logging/__pycache__/progress_bar.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c4b3e21466ae4436f98af143e095b9dc4d58964a Binary files /dev/null and b/fairseq/logging/__pycache__/progress_bar.cpython-310.pyc differ diff --git a/fairseq/logging/meters.py b/fairseq/logging/meters.py new file mode 100644 index 0000000000000000000000000000000000000000..78e6d4d224d8a490a02ecdff1487186b830ecb3b --- /dev/null +++ b/fairseq/logging/meters.py @@ -0,0 +1,286 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import bisect +from collections import OrderedDict +import time +from typing import Dict, Optional + +try: + import torch + + def type_as(a, b): + if torch.is_tensor(a) and torch.is_tensor(b): + return a.to(b) + else: + return a +except ImportError: + torch = None + + def type_as(a, b): + return a + + +try: + import numpy as np +except ImportError: + np = None + + +class Meter(object): + """Base class for Meters.""" + + def __init__(self): + pass + + def state_dict(self): + return {} + + def load_state_dict(self, state_dict): + pass + + def reset(self): + raise NotImplementedError + + @property + def smoothed_value(self) -> float: + """Smoothed value used for logging.""" + raise NotImplementedError + + +def safe_round(number, ndigits): + if hasattr(number, '__round__'): + return round(number, ndigits) + elif torch is not None and torch.is_tensor(number) and number.numel() == 1: + return safe_round(number.item(), ndigits) + elif np is not None and np.ndim(number) == 0 and hasattr(number, 'item'): + return safe_round(number.item(), ndigits) + else: + return number + + +class AverageMeter(Meter): + """Computes and stores the average and current value""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.reset() + + def reset(self): + self.val = None # most recent update + self.sum = 0 # sum from all updates + self.count = 0 # total n from all updates + + def update(self, val, n=1): + if val is not None: + self.val = val + if n > 0: + self.sum = type_as(self.sum, val) + (val * n) + self.count = type_as(self.count, n) + n + + def state_dict(self): + return { + 'val': self.val, + 'sum': self.sum, + 'count': self.count, + 'round': self.round, + } + + def load_state_dict(self, state_dict): + self.val = state_dict['val'] + self.sum = state_dict['sum'] + self.count = state_dict['count'] + self.round = state_dict.get('round', None) + + @property + def avg(self): + return self.sum / self.count if self.count > 0 else self.val + + @property + def smoothed_value(self) -> float: + val = self.avg + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class TimeMeter(Meter): + """Computes the average occurrence of some event per second""" + + def __init__( + self, + init: int = 0, + n: int = 0, + round: Optional[int] = None, + ): + self.round = round + self.reset(init, n) + + def reset(self, init=0, n=0): + self.init = init + self.start = time.perf_counter() + self.n = n + self.i = 0 + + def update(self, val=1): + self.n = type_as(self.n, val) + val + self.i += 1 + + def state_dict(self): + return { + 'init': self.elapsed_time, + 'n': self.n, + 'round': self.round, + } + + def load_state_dict(self, state_dict): + if 'start' in state_dict: + # backwards compatibility for old state_dicts + self.reset(init=state_dict['init']) + else: + self.reset(init=state_dict['init'], n=state_dict['n']) + self.round = state_dict.get('round', None) + + @property + def avg(self): + return self.n / self.elapsed_time + + @property + def elapsed_time(self): + return self.init + (time.perf_counter() - self.start) + + @property + def smoothed_value(self) -> float: + val = self.avg + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class StopwatchMeter(Meter): + """Computes the sum/avg duration of some event in seconds""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.sum = 0 + self.n = 0 + self.start_time = None + + def start(self): + self.start_time = time.perf_counter() + + def stop(self, n=1, prehook=None): + if self.start_time is not None: + if prehook is not None: + prehook() + delta = time.perf_counter() - self.start_time + self.sum = self.sum + delta + self.n = type_as(self.n, n) + n + + def reset(self): + self.sum = 0 # cumulative time during which stopwatch was active + self.n = 0 # total n across all start/stop + self.start() + + def state_dict(self): + return { + 'sum': self.sum, + 'n': self.n, + 'round': self.round, + } + + def load_state_dict(self, state_dict): + self.sum = state_dict['sum'] + self.n = state_dict['n'] + self.start_time = None + self.round = state_dict.get('round', None) + + @property + def avg(self): + return self.sum / self.n if self.n > 0 else self.sum + + @property + def elapsed_time(self): + if self.start_time is None: + return 0. + return time.perf_counter() - self.start_time + + @property + def smoothed_value(self) -> float: + val = self.avg if self.sum > 0 else self.elapsed_time + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class MetersDict(OrderedDict): + """A sorted dictionary of :class:`Meters`. + + Meters are sorted according to a priority that is given when the + meter is first added to the dictionary. + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.priorities = [] + + def __setitem__(self, key, value): + assert key not in self, "MetersDict doesn't support reassignment" + priority, value = value + bisect.insort(self.priorities, (priority, len(self.priorities), key)) + super().__setitem__(key, value) + for _, _, key in self.priorities: # reorder dict to match priorities + self.move_to_end(key) + + def add_meter(self, key, meter, priority): + self.__setitem__(key, (priority, meter)) + + def state_dict(self): + return [ + (pri, key, self[key].__class__.__name__, self[key].state_dict()) + for pri, _, key in self.priorities + # can't serialize DerivedMeter instances + if not isinstance(self[key], MetersDict._DerivedMeter) + ] + + def load_state_dict(self, state_dict): + self.clear() + self.priorities.clear() + for pri, key, meter_cls, meter_state in state_dict: + meter = globals()[meter_cls]() + meter.load_state_dict(meter_state) + self.add_meter(key, meter, pri) + + def get_smoothed_value(self, key: str) -> float: + """Get a single smoothed value.""" + meter = self[key] + if isinstance(meter, MetersDict._DerivedMeter): + return meter.fn(self) + else: + return meter.smoothed_value + + def get_smoothed_values(self) -> Dict[str, float]: + """Get all smoothed values.""" + return OrderedDict([ + (key, self.get_smoothed_value(key)) + for key in self.keys() + if not key.startswith("_") + ]) + + def reset(self): + """Reset Meter instances.""" + for meter in self.values(): + if isinstance(meter, MetersDict._DerivedMeter): + continue + meter.reset() + + class _DerivedMeter(Meter): + """A Meter whose values are derived from other Meters.""" + + def __init__(self, fn): + self.fn = fn + + def reset(self): + pass diff --git a/fairseq/logging/metrics.py b/fairseq/logging/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..6ca1d201e02c57846d0ec810dbd175fd5fd8023d --- /dev/null +++ b/fairseq/logging/metrics.py @@ -0,0 +1,291 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +A standalone module for aggregating metrics. + +Metrics can be logged from anywhere using the `log_*` functions defined +in this module. The logged values will be aggregated dynamically based +on the aggregation context in which the logging occurs. See the +:func:`aggregate` context manager for more details. +""" + +from collections import defaultdict, OrderedDict +import contextlib +import time +from typing import Callable, Dict, List, Optional +import uuid + +from .meters import * + + +# Aggregation contexts are considered "active" when inside the scope +# created by the :func:`aggregate` context manager. +_aggregators = OrderedDict() +_active_aggregators = OrderedDict() +_active_aggregators_cnt = defaultdict(lambda: 0) + + +def reset() -> None: + """Reset all metrics aggregators.""" + _aggregators.clear() + _active_aggregators.clear() + _active_aggregators_cnt.clear() + + # The "default" aggregator observes all logged values. + _aggregators["default"] = MetersDict() + _active_aggregators["default"] = _aggregators["default"] + _active_aggregators_cnt["default"] = 1 + + +reset() + + +@contextlib.contextmanager +def aggregate(name: Optional[str] = None, new_root: bool = False): + """Context manager to aggregate metrics under a given name. + + Aggregations can be nested. If *new_root* is ``False``, then logged + metrics will be recorded along the entire stack of nested + aggregators, including a global "default" aggregator. If *new_root* + is ``True``, then this aggregator will be the root of a new + aggregation stack, thus bypassing any parent aggregators. + + Note that aggregation contexts are uniquely identified by their + *name* (e.g., train, valid). Creating a context with an existing + name will reuse the corresponding :class:`MetersDict` instance. + If no name is given, then a temporary aggregator will be created. + + Usage:: + + with metrics.aggregate("train"): + for step, batch in enumerate(epoch): + with metrics.aggregate("train_inner") as agg: + metrics.log_scalar("loss", get_loss(batch)) + if step % log_interval == 0: + print(agg.get_smoothed_value("loss")) + agg.reset() + print(metrics.get_smoothed_values("train")["loss"]) + + Args: + name (str): name of the aggregation. Defaults to a + random/temporary name if not given explicitly. + new_root (bool): make this aggregation the root of a new + aggregation stack. + """ + if name is None: + # generate a temporary name + name = str(uuid.uuid4()) + assert name not in _aggregators + agg = MetersDict() + else: + assert name != "default" + agg = _aggregators.setdefault(name, MetersDict()) + + if new_root: + backup_aggregators = _active_aggregators.copy() + _active_aggregators.clear() + backup_aggregators_cnt = _active_aggregators_cnt.copy() + _active_aggregators_cnt.clear() + + _active_aggregators[name] = agg + _active_aggregators_cnt[name] += 1 + + yield agg + + _active_aggregators_cnt[name] -= 1 + if _active_aggregators_cnt[name] == 0 and name in _active_aggregators: + del _active_aggregators[name] + + if new_root: + _active_aggregators.clear() + _active_aggregators.update(backup_aggregators) + _active_aggregators_cnt.clear() + _active_aggregators_cnt.update(backup_aggregators_cnt) + + +def get_active_aggregators() -> List[MetersDict]: + return list(_active_aggregators.values()) + + +def log_scalar( + key: str, + value: float, + weight: float = 1, + priority: int = 10, + round: Optional[int] = None, +): + """Log a scalar value. + + Args: + key (str): name of the field to log + value (float): value to log + weight (float): weight that this value contributes to the average. + A weight of 0 will always log the latest value. + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, AverageMeter(round=round), priority) + agg[key].update(value, weight) + + +def log_derived(key: str, fn: Callable[[MetersDict], float], priority: int = 20): + """Log a scalar value derived from other meters. + + Args: + key (str): name of the field to log + fn (Callable[[MetersDict], float]): function that takes a single + argument *meters* and returns the derived value + priority (int): smaller values are logged earlier in the output + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, MetersDict._DerivedMeter(fn), priority) + + +def log_speed( + key: str, + value: float, + priority: int = 30, + round: Optional[int] = None, +): + """Log the rate of some quantity per second. + + Args: + key (str): name of the field to log + value (float): value to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, TimeMeter(round=round), priority) + agg[key].reset() # reset meter on the first call + else: + agg[key].update(value) + + +def log_start_time(key: str, priority: int = 40, round: Optional[int] = None): + """Log the duration of some event in seconds. + + The duration will be computed once :func:`log_stop_time` is called. + + Args: + key (str): name of the field to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, StopwatchMeter(round=round), priority) + agg[key].start() + + +def log_stop_time(key: str, weight: float = 0., prehook=None): + """Log the duration of some event in seconds. + + The duration will be computed since :func:`log_start_time` was called. + Set weight > 0 to report the average time instead of the sum. + + Args: + key (str): name of the field to log + weight (float): weight that this time contributes to the average + prehook (function, no arguments): will be called before the timer + is stopped. For example, use prehook=torch.cuda.synchronize to + make sure all gpu operations are done before timer is stopped. + """ + for agg in get_active_aggregators(): + if key in agg: + agg[key].stop(weight, prehook) + + +def log_custom( + new_meter_fn: Callable[[], Meter], + key: str, + *args, + priority: int = 50, + **kwargs, +): + """Log using a custom Meter. + + Any extra *args* or *kwargs* will be passed through to the Meter's + *update* method. + + Args: + new_meter_fn (Callable[[], Meter]): function that returns a new + Meter instance + key (str): name of the field to log + priority (int): smaller values are logged earlier in the output + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, new_meter_fn(), priority) + agg[key].update(*args, **kwargs) + + +def reset_meter(name: str, key: str) -> None: + """Reset Meter instance aggregated under a given *name* and *key*.""" + meter = get_meter(name, key) + if meter is not None: + meter.reset() + + +def reset_meters(name: str) -> None: + """Reset Meter instances aggregated under a given *name*.""" + meters = get_meters(name) + if meters is not None: + meters.reset() + + +def get_meter(name: str, key: str) -> Meter: + """Get a single Meter instance aggregated under *name* and *key*. + + Returns: + Meter or None if no metrics have been logged under *name* and *key*. + """ + if name not in _aggregators: + return None + return _aggregators[name].get(key, None) + + +def get_meters(name: str) -> MetersDict: + """Get Meter instances aggregated under a given *name*. + + Returns: + MetersDict or None if no metrics have been logged under *name*. + """ + return _aggregators.get(name, None) + + +def get_smoothed_value(name: str, key: str) -> float: + """Get a single smoothed value. + + Raises: + KeyError: if no metrics have been logged under *name* and *key*. + """ + return _aggregators[name].get_smoothed_value(key) + + +def get_smoothed_values(name: str) -> Dict[str, float]: + """Get smoothed values aggregated under a given *name*. + + Raises: + KeyError: if no metrics have been logged under *name*. + """ + return _aggregators[name].get_smoothed_values() + + +def state_dict(): + return OrderedDict([ + (name, agg.state_dict()) + for name, agg in _aggregators.items() + ]) + + +def load_state_dict(state_dict): + for name, agg_state in state_dict.items(): + _aggregators[name] = MetersDict() + _aggregators[name].load_state_dict(agg_state) diff --git a/fairseq/logging/progress_bar.py b/fairseq/logging/progress_bar.py new file mode 100644 index 0000000000000000000000000000000000000000..97e4162ea0a55f9b0d1caa95c082616a966c9d10 --- /dev/null +++ b/fairseq/logging/progress_bar.py @@ -0,0 +1,359 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Wrapper around various loggers and progress bars (e.g., tqdm). +""" + +import atexit +import json +import logging +import os +import sys +from collections import OrderedDict +from contextlib import contextmanager +from numbers import Number +from typing import Optional + +import torch + +from .meters import AverageMeter, StopwatchMeter, TimeMeter + + +logger = logging.getLogger(__name__) + + +def progress_bar( + iterator, + log_format: Optional[str] = None, + log_interval: int = 100, + epoch: Optional[int] = None, + prefix: Optional[str] = None, + tensorboard_logdir: Optional[str] = None, + default_log_format: str = 'tqdm', +): + if log_format is None: + log_format = default_log_format + if log_format == 'tqdm' and not sys.stderr.isatty(): + log_format = 'simple' + + if log_format == 'json': + bar = JsonProgressBar(iterator, epoch, prefix, log_interval) + elif log_format == 'none': + bar = NoopProgressBar(iterator, epoch, prefix) + elif log_format == 'simple': + bar = SimpleProgressBar(iterator, epoch, prefix, log_interval) + elif log_format == 'tqdm': + bar = TqdmProgressBar(iterator, epoch, prefix) + else: + raise ValueError('Unknown log format: {}'.format(log_format)) + + if tensorboard_logdir: + try: + # [FB only] custom wrapper for TensorBoard + import palaas # noqa + from .fb_tbmf_wrapper import FbTbmfWrapper + bar = FbTbmfWrapper(bar, log_interval) + except ImportError: + bar = TensorboardProgressBarWrapper(bar, tensorboard_logdir) + + return bar + + +def build_progress_bar( + args, + iterator, + epoch: Optional[int] = None, + prefix: Optional[str] = None, + default: str = 'tqdm', + no_progress_bar: str = 'none', +): + """Legacy wrapper that takes an argparse.Namespace.""" + if getattr(args, 'no_progress_bar', False): + default = no_progress_bar + if getattr(args, 'distributed_rank', 0) == 0: + tensorboard_logdir = getattr(args, 'tensorboard_logdir', None) + else: + tensorboard_logdir = None + return progress_bar( + iterator, + log_format=args.log_format, + log_interval=args.log_interval, + epoch=epoch, + prefix=prefix, + tensorboard_logdir=tensorboard_logdir, + default_log_format=default, + ) + + +def format_stat(stat): + if isinstance(stat, Number): + stat = '{:g}'.format(stat) + elif isinstance(stat, AverageMeter): + stat = '{:.3f}'.format(stat.avg) + elif isinstance(stat, TimeMeter): + stat = '{:g}'.format(round(stat.avg)) + elif isinstance(stat, StopwatchMeter): + stat = '{:g}'.format(round(stat.sum)) + elif torch.is_tensor(stat): + stat = stat.tolist() + return stat + + +class BaseProgressBar(object): + """Abstract class for progress bars.""" + def __init__(self, iterable, epoch=None, prefix=None): + self.iterable = iterable + self.n = getattr(iterable, 'n', 0) + self.epoch = epoch + self.prefix = '' + if epoch is not None: + self.prefix += 'epoch {:03d}'.format(epoch) + if prefix is not None: + self.prefix += ' | {}'.format(prefix) + + def __len__(self): + return len(self.iterable) + + def __enter__(self): + return self + + def __exit__(self, *exc): + return False + + def __iter__(self): + raise NotImplementedError + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + raise NotImplementedError + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + raise NotImplementedError + + def _str_commas(self, stats): + return ', '.join(key + '=' + stats[key].strip() + for key in stats.keys()) + + def _str_pipes(self, stats): + return ' | '.join(key + ' ' + stats[key].strip() + for key in stats.keys()) + + def _format_stats(self, stats): + postfix = OrderedDict(stats) + # Preprocess stats according to datatype + for key in postfix.keys(): + postfix[key] = str(format_stat(postfix[key])) + return postfix + + +@contextmanager +def rename_logger(logger, new_name): + old_name = logger.name + if new_name is not None: + logger.name = new_name + yield logger + logger.name = old_name + + +class JsonProgressBar(BaseProgressBar): + """Log output in JSON format.""" + + def __init__(self, iterable, epoch=None, prefix=None, log_interval=1000): + super().__init__(iterable, epoch, prefix) + self.log_interval = log_interval + self.i = None + self.size = None + + def __iter__(self): + self.size = len(self.iterable) + for i, obj in enumerate(self.iterable, start=self.n): + self.i = i + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + step = step or self.i or 0 + if ( + step > 0 + and self.log_interval is not None + and step % self.log_interval == 0 + ): + update = ( + self.epoch - 1 + (self.i + 1) / float(self.size) + if self.epoch is not None + else None + ) + stats = self._format_stats(stats, epoch=self.epoch, update=update) + with rename_logger(logger, tag): + logger.info(json.dumps(stats)) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self.stats = stats + if tag is not None: + self.stats = OrderedDict([(tag + '_' + k, v) for k, v in self.stats.items()]) + stats = self._format_stats(self.stats, epoch=self.epoch) + with rename_logger(logger, tag): + logger.info(json.dumps(stats)) + + def _format_stats(self, stats, epoch=None, update=None): + postfix = OrderedDict() + if epoch is not None: + postfix['epoch'] = epoch + if update is not None: + postfix['update'] = round(update, 3) + # Preprocess stats according to datatype + for key in stats.keys(): + postfix[key] = format_stat(stats[key]) + return postfix + + +class NoopProgressBar(BaseProgressBar): + """No logging.""" + + def __init__(self, iterable, epoch=None, prefix=None): + super().__init__(iterable, epoch, prefix) + + def __iter__(self): + for obj in self.iterable: + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + pass + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + pass + + +class SimpleProgressBar(BaseProgressBar): + """A minimal logger for non-TTY environments.""" + + def __init__(self, iterable, epoch=None, prefix=None, log_interval=1000): + super().__init__(iterable, epoch, prefix) + self.log_interval = log_interval + self.i = None + self.size = None + + def __iter__(self): + self.size = len(self.iterable) + for i, obj in enumerate(self.iterable, start=self.n): + self.i = i + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + step = step or self.i or 0 + if ( + step > 0 + and self.log_interval is not None + and step % self.log_interval == 0 + ): + stats = self._format_stats(stats) + postfix = self._str_commas(stats) + with rename_logger(logger, tag): + logger.info( + '{}: {:5d} / {:d} {}' + .format(self.prefix, self.i + 1, self.size, postfix) + ) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + postfix = self._str_pipes(self._format_stats(stats)) + with rename_logger(logger, tag): + logger.info('{} | {}'.format(self.prefix, postfix)) + + +class TqdmProgressBar(BaseProgressBar): + """Log to tqdm.""" + + def __init__(self, iterable, epoch=None, prefix=None): + super().__init__(iterable, epoch, prefix) + from tqdm import tqdm + self.tqdm = tqdm( + iterable, + self.prefix, + leave=False, + disable=(logger.getEffectiveLevel() > logging.INFO), + ) + + def __iter__(self): + return iter(self.tqdm) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + self.tqdm.set_postfix(self._format_stats(stats), refresh=False) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + postfix = self._str_pipes(self._format_stats(stats)) + with rename_logger(logger, tag): + logger.info('{} | {}'.format(self.prefix, postfix)) + + +try: + _tensorboard_writers = {} + from tensorboardX import SummaryWriter +except ImportError: + SummaryWriter = None + + +def _close_writers(): + for w in _tensorboard_writers.values(): + w.close() + + +atexit.register(_close_writers) + + +class TensorboardProgressBarWrapper(BaseProgressBar): + """Log to tensorboard.""" + + def __init__(self, wrapped_bar, tensorboard_logdir): + self.wrapped_bar = wrapped_bar + self.tensorboard_logdir = tensorboard_logdir + + if SummaryWriter is None: + logger.warning( + "tensorboard not found, please install with: pip install tensorboardX" + ) + + def _writer(self, key): + if SummaryWriter is None: + return None + _writers = _tensorboard_writers + if key not in _writers: + _writers[key] = SummaryWriter(os.path.join(self.tensorboard_logdir, key)) + _writers[key].add_text('sys.argv', " ".join(sys.argv)) + return _writers[key] + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to tensorboard.""" + self._log_to_tensorboard(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self._log_to_tensorboard(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def _log_to_tensorboard(self, stats, tag=None, step=None): + writer = self._writer(tag or '') + if writer is None: + return + if step is None: + step = stats['num_updates'] + for key in stats.keys() - {'num_updates'}: + if isinstance(stats[key], AverageMeter): + writer.add_scalar(key, stats[key].val, step) + elif isinstance(stats[key], Number): + writer.add_scalar(key, stats[key], step) + writer.flush() diff --git a/fairseq/model_parallel/__init__.py b/fairseq/model_parallel/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..cc563db40b9441c7d1471041572cf029a8eb3919 --- /dev/null +++ b/fairseq/model_parallel/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import criterions, modules, models # noqa diff --git a/fairseq/model_parallel/__pycache__/__init__.cpython-310.pyc b/fairseq/model_parallel/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7ff2937fde134d7b71c37b679aa1981ec462e939 Binary files /dev/null and b/fairseq/model_parallel/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/model_parallel/criterions/__init__.py b/fairseq/model_parallel/criterions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b74de559824a03eb9511c83f9f61d183a0506597 --- /dev/null +++ b/fairseq/model_parallel/criterions/__init__.py @@ -0,0 +1,14 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the criterions/ directory +for file in os.listdir(os.path.dirname(__file__)): + if file.endswith('.py') and not file.startswith('_'): + module = file[:file.find('.py')] + importlib.import_module('fairseq.model_parallel.criterions.' + module) diff --git a/fairseq/model_parallel/criterions/__pycache__/__init__.cpython-310.pyc b/fairseq/model_parallel/criterions/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fba83c5152c4e2899c12ce44088162d88c9a767c Binary files /dev/null and b/fairseq/model_parallel/criterions/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/model_parallel/criterions/__pycache__/vocab_parallel_cross_entropy.cpython-310.pyc b/fairseq/model_parallel/criterions/__pycache__/vocab_parallel_cross_entropy.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ab62c05af3d464eaa427208c234c2db6bd8bbe72 Binary files /dev/null and b/fairseq/model_parallel/criterions/__pycache__/vocab_parallel_cross_entropy.cpython-310.pyc differ diff --git a/fairseq/model_parallel/criterions/vocab_parallel_cross_entropy.py b/fairseq/model_parallel/criterions/vocab_parallel_cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..eab8f9af4e1c81f4c87ccd67695a8031fc1db1de --- /dev/null +++ b/fairseq/model_parallel/criterions/vocab_parallel_cross_entropy.py @@ -0,0 +1,74 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion + +try: + from fairseq.model_parallel.megatron.mpu.cross_entropy import vocab_parallel_cross_entropy + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +@register_criterion('vocab_parallel_cross_entropy') +class VocabParallelCrossEntropyCriterion(FairseqCriterion): + + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + if not has_megatron_submodule: + raise ImportError( + '\n\nPlease install the megatron submodule:' + '\n\n git submodule update --init ' + 'fairseq/model_parallel/megatron' + ) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample['net_input']) + target = sample['target'] + + loss = vocab_parallel_cross_entropy(net_output[0].float(), target) + loss = (loss * (target != self.padding_idx)).sum() + sample_size = sample['target'].size(0) if self.sentence_avg else sample['ntokens'] + logging_output = { + 'loss': utils.item(loss.data) if reduce else loss.data, + 'ntokens': sample['ntokens'], + 'nsentences': sample['target'].size(0), + 'sample_size': sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get('loss', 0) for log in logging_outputs) + ntokens = sum(log.get('ntokens', 0) for log in logging_outputs) + sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) + + metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) + if sample_size != ntokens: + metrics.log_scalar('nll_loss', loss_sum / ntokens / math.log(2), ntokens, round=3) + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['nll_loss'].avg)) + else: + metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg)) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/fairseq/model_parallel/megatron_trainer.py b/fairseq/model_parallel/megatron_trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..d1142a993ca05e42371f5092d8119b996ca49ef5 --- /dev/null +++ b/fairseq/model_parallel/megatron_trainer.py @@ -0,0 +1,63 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Train a network across multiple GPUs. +""" + +from fairseq import distributed_utils +from fairseq.trainer import Trainer + +try: + from fairseq.model_parallel.megatron.mpu import ( + get_data_parallel_group, + get_data_parallel_rank, + get_data_parallel_world_size, + get_model_parallel_group, + get_model_parallel_src_rank, + ) + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +class MegatronTrainer(Trainer): + """Main class for model parallel with data parallel training. + """ + def __init__(self, args, task, model, criterion): + if not has_megatron_submodule: + raise ImportError( + '\n\nPlease install the megatron submodule:' + '\n\n git submodule update --init ' + 'fairseq/model_parallel/megatron' + ) + super().__init__(args, task, model, criterion) + + @property + def data_parallel_world_size(self): + return get_data_parallel_world_size() + + @property + def data_parallel_process_group(self): + return get_data_parallel_group() + + @property + def data_parallel_rank(self): + return get_data_parallel_rank() + + @property + def is_data_parallel_master(self): + return get_model_parallel_src_rank() == 0 + + def clip_grad_norm(self, clip_norm): + def _aggregate_model_parallel_grad_norm(total_norm): + total_norm = total_norm ** 2 + distributed_utils.all_reduce(total_norm, group=get_model_parallel_group()) + total_norm = total_norm ** 0.5 + return total_norm + return self.optimizer.clip_grad_norm( + clip_norm, + aggregate_norm_fn=_aggregate_model_parallel_grad_norm, + ) diff --git a/fairseq/model_parallel/models/__init__.py b/fairseq/model_parallel/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a3207981adcb59289d59d18961b961c819e52013 --- /dev/null +++ b/fairseq/model_parallel/models/__init__.py @@ -0,0 +1,16 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the models/ directory +models_dir = os.path.dirname(__file__) +for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if not file.startswith('_') and not file.startswith('.') and (file.endswith('.py') or os.path.isdir(path)): + model_name = file[:file.find('.py')] if file.endswith('.py') else file + module = importlib.import_module('fairseq.model_parallel.models.' + model_name) diff --git a/fairseq/model_parallel/models/__pycache__/__init__.cpython-310.pyc b/fairseq/model_parallel/models/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eb7c6d40fc5ba40ad6cb13499661dd747be55cdf Binary files /dev/null and b/fairseq/model_parallel/models/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/model_parallel/models/__pycache__/transformer.cpython-310.pyc b/fairseq/model_parallel/models/__pycache__/transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..669af6407599406e90831a11fbd31f3a075587be Binary files /dev/null and b/fairseq/model_parallel/models/__pycache__/transformer.cpython-310.pyc differ diff --git a/fairseq/model_parallel/models/__pycache__/transformer_lm.cpython-310.pyc b/fairseq/model_parallel/models/__pycache__/transformer_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9cad6d405663be6a943a5972d801d927773d8537 Binary files /dev/null and b/fairseq/model_parallel/models/__pycache__/transformer_lm.cpython-310.pyc differ diff --git a/fairseq/model_parallel/models/roberta/__init__.py b/fairseq/model_parallel/models/roberta/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..117827c3e9c176477f33e3a6fd7fe19a922411a2 --- /dev/null +++ b/fairseq/model_parallel/models/roberta/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .model import * # noqa diff --git a/fairseq/model_parallel/models/roberta/__pycache__/__init__.cpython-310.pyc b/fairseq/model_parallel/models/roberta/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bc734ba288d5004fe10b83a3ac7ed9f513764c17 Binary files /dev/null and b/fairseq/model_parallel/models/roberta/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/model_parallel/models/roberta/__pycache__/model.cpython-310.pyc b/fairseq/model_parallel/models/roberta/__pycache__/model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a8093469ba2b23258a8c0fc45b6f4d7b14324e7a Binary files /dev/null and b/fairseq/model_parallel/models/roberta/__pycache__/model.cpython-310.pyc differ diff --git a/fairseq/model_parallel/models/roberta/model.py b/fairseq/model_parallel/models/roberta/model.py new file mode 100644 index 0000000000000000000000000000000000000000..e0ae4a2c8ff2a33eda5f804a0fc7802799dd697a --- /dev/null +++ b/fairseq/model_parallel/models/roberta/model.py @@ -0,0 +1,268 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +RoBERTa: A Robustly Optimized BERT Pretraining Approach. +""" + +import logging + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + register_model, + register_model_architecture, +) +from fairseq.models.roberta import ( + RobertaModel, + RobertaEncoder, + RobertaLMHead, + RobertaClassificationHead, +) +from fairseq.modules import ( + LayerNorm, + TransformerSentenceEncoder, +) +from fairseq.model_parallel.modules import ( + ModelParallelTransformerSentenceEncoder, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params +try: + from fairseq.model_parallel.megatron.mpu import ( + copy_to_model_parallel_region, + gather_from_model_parallel_region, + ColumnParallelLinear, + RowParallelLinear, + ) + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + +logger = logging.getLogger(__name__) + + +@register_model('model_parallel_roberta') +class ModelParallelRobertaModel(RobertaModel): + + + def __init__(self, args, encoder): + super().__init__(args, encoder) + + self.classification_heads = nn.ModuleDict() + + @staticmethod + def add_args(parser): + super(ModelParallelRobertaModel, ModelParallelRobertaModel).add_args(parser) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present + base_architecture(args) + + if not hasattr(args, 'max_positions'): + args.max_positions = args.tokens_per_sample + + encoder = ModelParallelRobertaEncoder(args, task.source_dictionary) + return cls(args, encoder) + + def forward(self, src_tokens, features_only=False, return_all_hiddens=False, classification_head_name=None, **kwargs): + if classification_head_name is not None: + features_only = True + + x, extra = self.encoder(src_tokens, features_only, return_all_hiddens, **kwargs) + + if classification_head_name is not None: + x = self.classification_heads[classification_head_name](x) + return x, extra + + def register_classification_head(self, name, num_classes=None, inner_dim=None, **kwargs): + """Register a classification head.""" + if name in self.classification_heads: + prev_num_classes = self.classification_heads[name].out_proj.out_features + prev_inner_dim = self.classification_heads[name].dense.out_features + if num_classes != prev_num_classes or inner_dim != prev_inner_dim: + logger.warning( + 're-registering head "{}" with num_classes {} (prev: {}) ' + 'and inner_dim {} (prev: {})'.format( + name, num_classes, prev_num_classes, inner_dim, prev_inner_dim + ) + ) + self.classification_heads[name] = ModelParallelRobertaClassificationHead( + self.args.encoder_embed_dim, + inner_dim or self.args.encoder_embed_dim, + num_classes, + self.args.pooler_activation_fn, + self.args.pooler_dropout, + ) + + +class ModelParallelRobertaLMHead(nn.Module): + """Head for masked language modeling.""" + + def __init__(self, embed_dim, output_dim, activation_fn, weight=None): + super().__init__() + self.dense = ColumnParallelLinear(embed_dim, embed_dim, gather_output=True) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.layer_norm = LayerNorm(embed_dim) + + if weight is None: + weight = nn.Linear(embed_dim, output_dim, bias=False).weight + self.weight = weight + self.bias = nn.Parameter(torch.zeros(output_dim)) + + def forward(self, features, masked_tokens=None, **kwargs): + # Only project the unmasked tokens while training, + # saves both memory and computation + if masked_tokens is not None: + features = features[masked_tokens, :] + + x = self.dense(features) + x = self.activation_fn(x) + x = self.layer_norm(x) + + features = copy_to_model_parallel_region(features) + # project back to size of vocabulary with bias + x = F.linear(x, self.weight) + x = gather_from_model_parallel_region(x).contiguous() + x = x + self.bias + return x + + +class ModelParallelRobertaClassificationHead(nn.Module): + """Head for sentence-level classification tasks.""" + + def __init__(self, input_dim, inner_dim, num_classes, activation_fn, pooler_dropout): + super().__init__() + self.dense = ColumnParallelLinear(input_dim, inner_dim, gather_output=True) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.dropout = nn.Dropout(p=pooler_dropout) + self.out_proj = nn.Linear(inner_dim, num_classes) + + def forward(self, features, **kwargs): + x = features[:, 0, :] # take token (equiv. to [CLS]) + x = self.dropout(x) + x = self.dense(x) + x = self.activation_fn(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +class ModelParallelRobertaEncoder(FairseqEncoder): + """RoBERTa encoder. + + Implements the :class:`~fairseq.models.FairseqDecoder` interface required + by :class:`~fairseq.models.FairseqLanguageModel`. + """ + + def __init__(self, args, dictionary): + super().__init__(dictionary) + self.args = args + + # RoBERTa is a sentence encoder model, so users will intuitively trim + # encoder layers. However, the implementation uses the fairseq decoder, + # so we fix here. + if args.encoder_layers_to_keep: + args.encoder_layers = len(args.encoder_layers_to_keep.split(",")) + args.decoder_layers_to_keep = args.encoder_layers_to_keep + args.encoder_layers_to_keep = None + + self.sentence_encoder = ModelParallelTransformerSentenceEncoder( + padding_idx=dictionary.pad(), + vocab_size=len(dictionary), + num_encoder_layers=args.encoder_layers, + embedding_dim=args.encoder_embed_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=args.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + layerdrop=args.encoder_layerdrop, + max_seq_len=args.max_positions, + num_segments=0, + encoder_normalize_before=False, + apply_bert_init=False, + activation_fn=args.activation_fn, + ) + self.lm_head = ModelParallelRobertaLMHead( + embed_dim=args.encoder_embed_dim, + output_dim=len(dictionary), + activation_fn=args.activation_fn, + weight=self.sentence_encoder.embed_tokens.weight, + ) + + def forward(self, src_tokens, features_only=False, return_all_hiddens=False, masked_tokens=None, **unused): + """ + Args: + src_tokens (LongTensor): input tokens of shape `(batch, src_len)` + features_only (bool, optional): skip LM head and just return + features. If True, the output will be of shape + `(batch, src_len, embed_dim)`. + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + + Returns: + tuple: + - the LM output of shape `(batch, src_len, vocab)` + - a dictionary of additional data, where 'inner_states' + is a list of hidden states. Note that the hidden + states have shape `(src_len, batch, vocab)`. + """ + x, extra = self.extract_features(src_tokens, return_all_hiddens=return_all_hiddens) + if not features_only: + x = self.output_layer(x, masked_tokens=masked_tokens) + return x, extra + + def extract_features(self, src_tokens, return_all_hiddens=False, **unused): + inner_states, _ = self.sentence_encoder( + src_tokens, + last_state_only=not return_all_hiddens, + ) + features = inner_states[-1].transpose(0, 1) # T x B x C -> B x T x C + return features, {'inner_states': inner_states if return_all_hiddens else None} + + def output_layer(self, features, masked_tokens=None, **unused): + return self.lm_head(features, masked_tokens) + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.args.max_positions + + +@register_model_architecture('model_parallel_roberta', 'model_parallel_roberta') +def base_architecture(args): + args.encoder_layers = getattr(args, 'encoder_layers', 12) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 768) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 3072) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 12) + + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + args.pooler_activation_fn = getattr(args, 'pooler_activation_fn', 'tanh') + + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_dropout = getattr(args, 'activation_dropout', 0.0) + args.pooler_dropout = getattr(args, 'pooler_dropout', 0.0) + args.encoder_layers_to_keep = getattr(args, 'encoder_layers_to_keep', None) + args.encoder_layerdrop = getattr(args, 'encoder_layerdrop', 0.0) + + +@register_model_architecture('model_parallel_roberta', 'model_parallel_roberta_base') +def roberta_base_architecture(args): + base_architecture(args) + + +@register_model_architecture('model_parallel_roberta', 'model_parallel_roberta_large') +def roberta_large_architecture(args): + args.encoder_layers = getattr(args, 'encoder_layers', 24) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1024) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4096) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 16) + base_architecture(args) diff --git a/fairseq/model_parallel/models/transformer.py b/fairseq/model_parallel/models/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..f5756ad8982631213d5b54d889e866d3585bac89 --- /dev/null +++ b/fairseq/model_parallel/models/transformer.py @@ -0,0 +1,112 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch.nn as nn +import torch.nn.functional as F + +from fairseq.models import ( + register_model, +) + +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + TransformerModel, +) + +from fairseq.model_parallel.modules import ( + ModelParallelTransformerDecoderLayer, + ModelParallelTransformerEncoderLayer, +) + +try: + from fairseq.model_parallel.megatron.mpu import ( + copy_to_model_parallel_region, + gather_from_model_parallel_region, + VocabParallelEmbedding, + ) + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +logger = logging.getLogger(__name__) + + +@register_model('model_parallel_transformer') +class ModelParallelTransformerModel(TransformerModel): + """ + Model parallel Transformer model. + """ + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + if not has_megatron_submodule: + raise ImportError( + '\n\nPlease install the megatron submodule:' + '\n\n git submodule update --init ' + 'fairseq/model_parallel/megatron' + ) + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + + def _vocab_init(tensor, **kwargs): + nn.init.normal_(tensor, mean=0, std=num_embeddings ** -0.5) + nn.init.constant_(tensor[1], 0) + emb = VocabParallelEmbedding(num_embeddings, embed_dim, padding_idx, init_method=_vocab_init) + # if provided, load from preloaded dictionaries + if path: + raise NotImplementedError("Loading of embedding from path is not supported for model parallel") + return emb + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return ModelParallelTransformerEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return ModelParallelTransformerDecoder( + args, + tgt_dict, + embed_tokens, + no_encoder_attn=getattr(args, 'no_cross_attention', False), + ) + + +class ModelParallelTransformerEncoder(TransformerEncoder): + """ + Model parallel Transformer encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`ModelParallelTransformerEncoderLayer`. + """ + + def build_encoder_layer(self, args): + return ModelParallelTransformerEncoderLayer(args) + + +class ModelParallelTransformerDecoder(TransformerDecoder): + """ + Model Parallel Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`ModelParallelTransformerDecoderLayer`. + """ + + def build_decoder_layer(self, args, no_encoder_attn=False): + return ModelParallelTransformerDecoderLayer(args, no_encoder_attn) + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + if not self.share_input_output_embed: + raise NotImplementedError( + 'Model parallel training currently requires --share-decoder-input-output-embed' + ) + + features = copy_to_model_parallel_region(features) + + # project back to size of vocabulary + x = self.output_projection(features) + + if getattr(self.args, 'criterion') != 'vocab_parallel_cross_entropy': + x = gather_from_model_parallel_region(x).contiguous() + return x diff --git a/fairseq/model_parallel/models/transformer_lm.py b/fairseq/model_parallel/models/transformer_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..81bc93bc0a975309bfe8b97ac9be1f119ce28117 --- /dev/null +++ b/fairseq/model_parallel/models/transformer_lm.py @@ -0,0 +1,88 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer_lm import ( + base_lm_architecture, + TransformerLanguageModel, +) +from fairseq.model_parallel.models.transformer import ModelParallelTransformerDecoder +try: + from fairseq.model_parallel.megatron.mpu import VocabParallelEmbedding + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model('model_parallel_transformer_lm') +class ModelParallelTransformerLanguageModel(TransformerLanguageModel): + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + if not has_megatron_submodule: + raise ImportError( + '\n\nPlease install the megatron submodule:' + '\n\n git submodule update --init ' + 'fairseq/model_parallel/megatron' + ) + + # make sure all arguments are present in older models + base_lm_architecture(args) + + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, 'max_target_positions', None) is None: + args.max_target_positions = getattr(args, 'tokens_per_sample', DEFAULT_MAX_TARGET_POSITIONS) + + if args.character_embeddings: + raise NotImplementedError("Character embeddings is not supported for model parallel") + elif args.adaptive_input: + raise NotImplementedError("Adaptive input is not supported for model parallel") + else: + embed_tokens = cls.build_embedding(args, task.source_dictionary, args.decoder_input_dim) + + decoder = ModelParallelTransformerDecoder( + args, task.target_dictionary, embed_tokens, no_encoder_attn=True, + ) + return cls(decoder) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + def _vocab_init(tensor, **kwargs): + nn.init.normal_(tensor, mean=0, std=embed_dim ** -0.5) + nn.init.constant_(tensor[1], 0) + embed_tokens = VocabParallelEmbedding(len(dictionary), embed_dim, dictionary.pad(), init_method=_vocab_init) + return embed_tokens + + +@register_model_architecture('model_parallel_transformer_lm', 'transformer_lm_megatron') +def transformer_lm_megatron(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 3072) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 3072 * 4) + args.decoder_layers = getattr(args, 'decoder_layers', 72) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 32) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + base_lm_architecture(args) + + +@register_model_architecture('model_parallel_transformer_lm', 'transformer_lm_megatron_11b') +def transformer_lm_megatron_11b(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 3072) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 3072 * 6) + args.decoder_layers = getattr(args, 'decoder_layers', 72) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 32) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + base_lm_architecture(args) diff --git a/fairseq/model_parallel/modules/__init__.py b/fairseq/model_parallel/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5c9431f92be6bc609bd999e77abf70197960cd48 --- /dev/null +++ b/fairseq/model_parallel/modules/__init__.py @@ -0,0 +1,17 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .multihead_attention import ModelParallelMultiheadAttention +from .transformer_layer import ModelParallelTransformerEncoderLayer, ModelParallelTransformerDecoderLayer +from .transformer_sentence_encoder_layer import ModelParallelTransformerSentenceEncoderLayer +from .transformer_sentence_encoder import ModelParallelTransformerSentenceEncoder + +__all__ = [ + 'ModelParallelMultiheadAttention', + 'ModelParallelTransformerEncoderLayer', + 'ModelParallelTransformerDecoderLayer', + 'ModelParallelTransformerSentenceEncoder', + 'ModelParallelTransformerSentenceEncoderLayer', +] diff --git a/fairseq/model_parallel/modules/__pycache__/__init__.cpython-310.pyc b/fairseq/model_parallel/modules/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e88d51da6986a3e9dd01e24f3a0151434b5c77e3 Binary files /dev/null and b/fairseq/model_parallel/modules/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/model_parallel/modules/__pycache__/multihead_attention.cpython-310.pyc b/fairseq/model_parallel/modules/__pycache__/multihead_attention.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..676b7075d92ca623d4155bbcb0ac288eb797dd28 Binary files /dev/null and b/fairseq/model_parallel/modules/__pycache__/multihead_attention.cpython-310.pyc differ diff --git a/fairseq/model_parallel/modules/__pycache__/transformer_layer.cpython-310.pyc b/fairseq/model_parallel/modules/__pycache__/transformer_layer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..231680fa552a46cd62de920363105d0821d8c63b Binary files /dev/null and b/fairseq/model_parallel/modules/__pycache__/transformer_layer.cpython-310.pyc differ diff --git a/fairseq/model_parallel/modules/__pycache__/transformer_sentence_encoder.cpython-310.pyc b/fairseq/model_parallel/modules/__pycache__/transformer_sentence_encoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f02fa78442c3ad10f20bec72618d34bdd645a259 Binary files /dev/null and b/fairseq/model_parallel/modules/__pycache__/transformer_sentence_encoder.cpython-310.pyc differ diff --git a/fairseq/model_parallel/modules/__pycache__/transformer_sentence_encoder_layer.cpython-310.pyc b/fairseq/model_parallel/modules/__pycache__/transformer_sentence_encoder_layer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4a8361919bf68d0d299d2de95c6994bb2396c491 Binary files /dev/null and b/fairseq/model_parallel/modules/__pycache__/transformer_sentence_encoder_layer.cpython-310.pyc differ diff --git a/fairseq/model_parallel/modules/multihead_attention.py b/fairseq/model_parallel/modules/multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..e92a3f6a71c7225b293a941020749046d2c628d5 --- /dev/null +++ b/fairseq/model_parallel/modules/multihead_attention.py @@ -0,0 +1,312 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional, Tuple + +import torch +import torch.nn.functional as F +from fairseq import utils +from torch import Tensor, nn +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout + +try: + from fairseq.model_parallel.megatron.mpu import ( + get_cuda_rng_tracker, + get_model_parallel_world_size, + ColumnParallelLinear, + RowParallelLinear, + ) + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +@with_incremental_state +class ModelParallelMultiheadAttention(nn.Module): + """Model parallel Multi-headed attention. + This performs the Multi-headed attention over multiple gpus. + + See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + self_attention=False, + encoder_decoder_attention=False, + ): + super().__init__() + if not has_megatron_submodule: + raise ImportError( + '\n\nPlease install the megatron submodule:' + '\n\n git submodule update --init ' + 'fairseq/model_parallel/megatron' + ) + self.embed_dim = embed_dim + self.kdim = kdim if kdim is not None else embed_dim + self.vdim = vdim if vdim is not None else embed_dim + self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim + + self.model_parallel_size = get_model_parallel_world_size() + + self.num_heads_partition = num_heads // self.model_parallel_size + assert ( + self.num_heads_partition * self.model_parallel_size == num_heads + ), "Number of heads must be divisble by model parallel size" + + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + self.scaling = self.head_dim ** -0.5 + + self.self_attention = self_attention + self.encoder_decoder_attention = encoder_decoder_attention + + assert not self.self_attention or self.qkv_same_dim, ( + "Self-attention requires query, key and value to be of the same size" + ) + + self.k_proj = ColumnParallelLinear(self.kdim, embed_dim, bias=bias, gather_output=False) + self.v_proj = ColumnParallelLinear(self.vdim, embed_dim, bias=bias, gather_output=False) + self.q_proj = ColumnParallelLinear(embed_dim, embed_dim, bias=bias, gather_output=False) + self.out_proj = RowParallelLinear(embed_dim, embed_dim, bias=bias, input_is_parallel=True) + + def forward( + self, + query, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + static_kv: bool = False, + attn_mask: Optional[Tensor] = None, + **unused_kwargs, + ) -> Tuple[Tensor, Optional[Tensor]]: + """Input shape: Time x Batch x Channel + + Args: + key_padding_mask (ByteTensor, optional): mask to exclude + keys that are pads, of shape `(batch, src_len)`, where + padding elements are indicated by 1s. + attn_mask (ByteTensor, optional): typically used to + implement causal attention, where the mask prevents the + attention from looking forward in time (default: None). + """ + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + + if incremental_state is not None: + saved_state = self._get_input_buffer(incremental_state) + if saved_state is not None and "prev_key" in saved_state: + # previous time steps are cached - no need to recompute + # key and value if they are static + if static_kv: + assert self.encoder_decoder_attention and not self.self_attention + key = value = None + else: + saved_state = None + + if self.self_attention: + q = self.q_proj(query) + k = self.k_proj(query) + v = self.v_proj(query) + elif self.encoder_decoder_attention: + # encoder-decoder attention + q = self.q_proj(query) + if key is None: + assert value is None + k = v = None + else: + k = self.k_proj(key) + v = self.v_proj(key) + + else: + assert key is not None and value is not None + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + q *= self.scaling + + + q = ( + q.contiguous() + .view(tgt_len, bsz * self.num_heads_partition, self.head_dim) + .transpose(0, 1) + ) + if k is not None: + k = ( + k.contiguous() + .view(-1, bsz * self.num_heads_partition, self.head_dim) + .transpose(0, 1) + ) + if v is not None: + v = ( + v.contiguous() + .view(-1, bsz * self.num_heads_partition, self.head_dim) + .transpose(0, 1) + ) + + if saved_state is not None: + # saved states are stored with shape (bsz, num_heads_partition, seq_len, head_dim) + if "prev_key" in saved_state: + _prev_key = saved_state["prev_key"] + assert _prev_key is not None + prev_key = _prev_key.view(bsz * self.num_heads_partition, -1, self.head_dim) + if static_kv: + k = prev_key + else: + assert k is not None + k = torch.cat([prev_key, k], dim=1) + if "prev_value" in saved_state: + _prev_value = saved_state["prev_value"] + assert _prev_value is not None + prev_value = _prev_value.view(bsz * self.num_heads_partition, -1, self.head_dim) + if static_kv: + v = prev_value + else: + assert v is not None + v = torch.cat([prev_value, v], dim=1) + prev_key_padding_mask: Optional[Tensor] = None + if "prev_key_padding_mask" in saved_state: + prev_key_padding_mask = saved_state["prev_key_padding_mask"] + assert k is not None and v is not None + key_padding_mask = ModelParallelMultiheadAttention._append_prev_key_padding_mask( + key_padding_mask=key_padding_mask, + prev_key_padding_mask=prev_key_padding_mask, + batch_size=bsz, + src_len=k.size(1), + static_kv=static_kv, + ) + + saved_state["prev_key"] = k.view(bsz, self.num_heads_partition, -1, self.head_dim) + saved_state["prev_value"] = v.view(bsz, self.num_heads_partition, -1, self.head_dim) + saved_state["prev_key_padding_mask"] = key_padding_mask + # In this branch incremental_state is never None + assert incremental_state is not None + incremental_state = self._set_input_buffer(incremental_state, saved_state) + assert k is not None + src_len = k.size(1) + + # This is part of a workaround to get around fork/join parallelism + # not supporting Optional types. + if key_padding_mask is not None and key_padding_mask.dim() == 0: + key_padding_mask = None + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == src_len + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + + assert list(attn_weights.size()) == [bsz * self.num_heads_partition, tgt_len, src_len] + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + attn_weights += attn_mask + + if key_padding_mask is not None: + # don't attend to padding symbols + attn_weights = attn_weights.view(bsz, self.num_heads_partition, tgt_len, src_len) + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), float("-inf") + ) + attn_weights = attn_weights.view(bsz * self.num_heads_partition, tgt_len, src_len) + + attn_weights_float = utils.softmax( + attn_weights, dim=-1 + ) + attn_weights = attn_weights_float.type_as(attn_weights) + + with get_cuda_rng_tracker().fork(): + attn_probs = self.dropout_module(attn_weights) + + assert v is not None + attn = torch.bmm(attn_probs, v) + assert list(attn.size()) == [bsz * self.num_heads_partition, tgt_len, self.head_dim] + embed_dim_partition = embed_dim // self.model_parallel_size + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim_partition) + attn = self.out_proj(attn) + # return attn_weights None to keep the return type same as single gpu multihead attention + # This will be deprecated. + attn_weights: Optional[Tensor] = None + + return attn, attn_weights + + @staticmethod + def _append_prev_key_padding_mask( + key_padding_mask: Optional[Tensor], + prev_key_padding_mask: Optional[Tensor], + batch_size: int, + src_len: int, + static_kv: bool, + ) -> Optional[Tensor]: + # saved key padding masks have shape (bsz, seq_len) + if prev_key_padding_mask is not None and static_kv: + new_key_padding_mask = prev_key_padding_mask + elif prev_key_padding_mask is not None and key_padding_mask is not None: + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1 + ) + # During incremental decoding, as the padding token enters and + # leaves the frame, there will be a time when prev or current + # is None + elif prev_key_padding_mask is not None: + + filler = torch.zeros(batch_size, src_len - prev_key_padding_mask.size(1)) + if prev_key_padding_mask.is_cuda: + filler = filler.cuda() + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), filler.float()], dim=1 + ) + elif key_padding_mask is not None: + filler = torch.zeros(batch_size, src_len - key_padding_mask.size(1)) + if key_padding_mask.is_cuda: + filler = filler.cuda() + new_key_padding_mask = torch.cat( + [filler.float(), key_padding_mask.float()], dim=1 + ) + else: + new_key_padding_mask = prev_key_padding_mask + return new_key_padding_mask + + def reorder_incremental_state( + self, incremental_state: Dict[str, Dict[str, Optional[Tensor]]], new_order + ): + """Reorder buffered internal state (for incremental generation).""" + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + for k in input_buffer.keys(): + if input_buffer[k] is not None: + input_buffer[k] = input_buffer[k].index_select(0, new_order) + incremental_state = self._set_input_buffer(incremental_state, input_buffer) + return incremental_state + + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ) -> Dict[str, Optional[Tensor]]: + result = self.get_incremental_state(incremental_state, "attn_state") + if result is not None: + return result + else: + empty_result: Dict[str, Optional[Tensor]] = {} + return empty_result + + def _set_input_buffer( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + buffer: Dict[str, Optional[Tensor]], + ): + return self.set_incremental_state(incremental_state, "attn_state", buffer) diff --git a/fairseq/model_parallel/modules/transformer_layer.py b/fairseq/model_parallel/modules/transformer_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..30b23d518c4c2f0815be7229d96c0689d2867e81 --- /dev/null +++ b/fairseq/model_parallel/modules/transformer_layer.py @@ -0,0 +1,79 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.modules import ( + TransformerEncoderLayer, + TransformerDecoderLayer, +) + +from fairseq.model_parallel.modules import ModelParallelMultiheadAttention + +try: + from fairseq.model_parallel.megatron.mpu import ( + ColumnParallelLinear, + RowParallelLinear, + ) + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +class ModelParallelTransformerEncoderLayer(TransformerEncoderLayer): + """Encoder layer block over multiple gpus. + + See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. + """ + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return ColumnParallelLinear(input_dim, output_dim, gather_output=False) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return RowParallelLinear(input_dim, output_dim, input_is_parallel=True) + + def build_self_attention(self, embed_dim, args, **unused_kwargs): + return ModelParallelMultiheadAttention( + embed_dim, + args.encoder_attention_heads, + dropout=args.attention_dropout, + self_attention=True, + ) + + +class ModelParallelTransformerDecoderLayer(TransformerDecoderLayer): + """Decoder layer block. + + See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. + """ + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return ColumnParallelLinear(input_dim, output_dim, gather_output=False) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + if q_noise > 0: + raise NotImplementedError + return RowParallelLinear(input_dim, output_dim, input_is_parallel=True) + + def build_self_attention(self, embed_dim, args, **unused_kwargs): + return ModelParallelMultiheadAttention( + embed_dim=embed_dim, + num_heads=args.decoder_attention_heads, + dropout=args.attention_dropout, + self_attention=not getattr(args, "cross_self_attention", False), + ) + + def build_encoder_attention(self, embed_dim, args, **unused_kwargs): + return ModelParallelMultiheadAttention( + embed_dim=embed_dim, + num_heads=args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + ) diff --git a/fairseq/model_parallel/modules/transformer_sentence_encoder.py b/fairseq/model_parallel/modules/transformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..101eca7bd4d143195ec89ffb1cfdd7b05060fcf1 --- /dev/null +++ b/fairseq/model_parallel/modules/transformer_sentence_encoder.py @@ -0,0 +1,64 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.modules import ( + LayerNorm, + MultiheadAttention, + PositionalEmbedding, + TransformerSentenceEncoder, +) + +from fairseq.model_parallel.modules import ( + ModelParallelTransformerSentenceEncoderLayer, +) + +try: + from fairseq.model_parallel.megatron.mpu import ( + copy_to_model_parallel_region, + gather_from_model_parallel_region, + VocabParallelEmbedding, + ) + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + +import random + + +class ModelParallelTransformerSentenceEncoder(TransformerSentenceEncoder): + """ + Implementation for a Model Parallel Bi-directional Transformer based + Sentence Encoder used in BERT/XLM style pre-trained models. + """ + def build_embedding(self, vocab_size, embedding_dim, padding_idx): + return VocabParallelEmbedding(vocab_size, embedding_dim, padding_idx) + + def build_transformer_sentence_encoder_layer( + self, + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + activation_fn, + export, + **unused, + ): + return ModelParallelTransformerSentenceEncoderLayer( + embedding_dim=embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + ) diff --git a/fairseq/model_parallel/modules/transformer_sentence_encoder_layer.py b/fairseq/model_parallel/modules/transformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..0e1ea2b7d757c041d6f4d0ba00b20478e3d72edc --- /dev/null +++ b/fairseq/model_parallel/modules/transformer_sentence_encoder_layer.py @@ -0,0 +1,79 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F + +from fairseq import utils +from fairseq.modules import ( + TransformerSentenceEncoderLayer +) +from fairseq.model_parallel.modules import ModelParallelMultiheadAttention +try: + from fairseq.model_parallel.megatron.mpu import ( + ColumnParallelLinear, + RowParallelLinear, + ) + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +class ModelParallelTransformerSentenceEncoderLayer(TransformerSentenceEncoderLayer): + """ + Implements a Model Parallel Transformer Encoder Layer used in + BERT/XLM style pre-trained models. + """ + def build_fc1(self, input_dim, output_dim, **unused): + return ColumnParallelLinear(input_dim, output_dim, gather_output=False) + + def build_fc2(self, input_dim, output_dim, **unused): + return RowParallelLinear(input_dim, output_dim, input_is_parallel=True) + + def build_self_attention( + self, + embed_dim, + num_attention_heads, + dropout, + **kwargs, + ): + return ModelParallelMultiheadAttention( + embed_dim, + num_attention_heads, + dropout=dropout, + self_attention=True + ) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer imlementation. + """ + residual = x + x = self.self_attn_layer_norm(x) + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = F.dropout(x, p=self.dropout, training=self.training) + x = residual + x + + residual = x + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = F.dropout(x, p=self.activation_dropout, training=self.training) + x = self.fc2(x) + x = F.dropout(x, p=self.dropout, training=self.training) + x = residual + x + return x, None diff --git a/fairseq/models/__init__.py b/fairseq/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f7d8eaafadf8f39a3e79501d86809aed6f5b414b --- /dev/null +++ b/fairseq/models/__init__.py @@ -0,0 +1,141 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import importlib +import os + +from .fairseq_decoder import FairseqDecoder +from .fairseq_encoder import FairseqEncoder +from .fairseq_incremental_decoder import FairseqIncrementalDecoder +from .fairseq_model import ( + BaseFairseqModel, + FairseqEncoderModel, + FairseqEncoderDecoderModel, + FairseqLanguageModel, + FairseqModel, + FairseqMultiModel, +) + +from .composite_encoder import CompositeEncoder +from .distributed_fairseq_model import DistributedFairseqModel + + +MODEL_REGISTRY = {} +ARCH_MODEL_REGISTRY = {} +ARCH_MODEL_INV_REGISTRY = {} +ARCH_CONFIG_REGISTRY = {} + + +__all__ = [ + 'BaseFairseqModel', + 'CompositeEncoder', + 'DistributedFairseqModel', + 'FairseqDecoder', + 'FairseqEncoder', + 'FairseqEncoderDecoderModel', + 'FairseqEncoderModel', + 'FairseqIncrementalDecoder', + 'FairseqLanguageModel', + 'FairseqModel', + 'FairseqMultiModel', +] + + +def build_model(args, task): + return ARCH_MODEL_REGISTRY[args.arch].build_model(args, task) + + +def register_model(name): + """ + New model types can be added to fairseq with the :func:`register_model` + function decorator. + + For example:: + + @register_model('lstm') + class LSTM(FairseqEncoderDecoderModel): + (...) + + .. note:: All models must implement the :class:`BaseFairseqModel` interface. + Typically you will extend :class:`FairseqEncoderDecoderModel` for + sequence-to-sequence tasks or :class:`FairseqLanguageModel` for + language modeling tasks. + + Args: + name (str): the name of the model + """ + + def register_model_cls(cls): + if name in MODEL_REGISTRY: + raise ValueError('Cannot register duplicate model ({})'.format(name)) + if not issubclass(cls, BaseFairseqModel): + raise ValueError('Model ({}: {}) must extend BaseFairseqModel'.format(name, cls.__name__)) + MODEL_REGISTRY[name] = cls + return cls + + return register_model_cls + + +def register_model_architecture(model_name, arch_name): + """ + New model architectures can be added to fairseq with the + :func:`register_model_architecture` function decorator. After registration, + model architectures can be selected with the ``--arch`` command-line + argument. + + For example:: + + @register_model_architecture('lstm', 'lstm_luong_wmt_en_de') + def lstm_luong_wmt_en_de(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1000) + (...) + + The decorated function should take a single argument *args*, which is a + :class:`argparse.Namespace` of arguments parsed from the command-line. The + decorated function should modify these arguments in-place to match the + desired architecture. + + Args: + model_name (str): the name of the Model (Model must already be + registered) + arch_name (str): the name of the model architecture (``--arch``) + """ + + def register_model_arch_fn(fn): + if model_name not in MODEL_REGISTRY: + raise ValueError('Cannot register model architecture for unknown model type ({})'.format(model_name)) + if arch_name in ARCH_MODEL_REGISTRY: + raise ValueError('Cannot register duplicate model architecture ({})'.format(arch_name)) + if not callable(fn): + raise ValueError('Model architecture must be callable ({})'.format(arch_name)) + ARCH_MODEL_REGISTRY[arch_name] = MODEL_REGISTRY[model_name] + ARCH_MODEL_INV_REGISTRY.setdefault(model_name, []).append(arch_name) + ARCH_CONFIG_REGISTRY[arch_name] = fn + return fn + + return register_model_arch_fn + + +# automatically import any Python files in the models/ directory +models_dir = os.path.dirname(__file__) +for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if ( + not file.startswith('_') + and not file.startswith('.') + and (file.endswith('.py') or os.path.isdir(path)) + ): + model_name = file[:file.find('.py')] if file.endswith('.py') else file + module = importlib.import_module('fairseq.models.' + model_name) + + # extra `model_parser` for sphinx + if model_name in MODEL_REGISTRY: + parser = argparse.ArgumentParser(add_help=False) + group_archs = parser.add_argument_group('Named architectures') + group_archs.add_argument('--arch', choices=ARCH_MODEL_INV_REGISTRY[model_name]) + group_args = parser.add_argument_group('Additional command-line arguments') + MODEL_REGISTRY[model_name].add_args(group_args) + globals()[model_name + '_parser'] = parser diff --git a/fairseq/models/__pycache__/__init__.cpython-310.pyc b/fairseq/models/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..16b3681e3bc9575399d60db0ff55d1b3ed73f2f0 Binary files /dev/null and b/fairseq/models/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/composite_encoder.cpython-310.pyc b/fairseq/models/__pycache__/composite_encoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e7ebad2d7c651120d3acd3709c52f76aab4224ea Binary files /dev/null and b/fairseq/models/__pycache__/composite_encoder.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/distributed_fairseq_model.cpython-310.pyc b/fairseq/models/__pycache__/distributed_fairseq_model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2a969f70aed607de317193a7feaf770b1e37e49e Binary files /dev/null and b/fairseq/models/__pycache__/distributed_fairseq_model.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/fairseq_decoder.cpython-310.pyc b/fairseq/models/__pycache__/fairseq_decoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..33a7b844878ade67071ede8a4fcf4dabaa8f05bb Binary files /dev/null and b/fairseq/models/__pycache__/fairseq_decoder.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/fairseq_encoder.cpython-310.pyc b/fairseq/models/__pycache__/fairseq_encoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4e48ed38968989defa68a0ba427b2b720292683b Binary files /dev/null and b/fairseq/models/__pycache__/fairseq_encoder.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/fairseq_incremental_decoder.cpython-310.pyc b/fairseq/models/__pycache__/fairseq_incremental_decoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5dc4f2486441b4931ee09178ecc40d0e577aa5fc Binary files /dev/null and b/fairseq/models/__pycache__/fairseq_incremental_decoder.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/fairseq_model.cpython-310.pyc b/fairseq/models/__pycache__/fairseq_model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f606b4714e3b9bccc5e8d32ebe3a855d91299909 Binary files /dev/null and b/fairseq/models/__pycache__/fairseq_model.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/fconv.cpython-310.pyc b/fairseq/models/__pycache__/fconv.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8e841ca9b9e89aa26c5730e66ef1ef5cedfe374b Binary files /dev/null and b/fairseq/models/__pycache__/fconv.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/fconv_lm.cpython-310.pyc b/fairseq/models/__pycache__/fconv_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..47c5c5c5f58cac62a664707fca4d33be6d66d17a Binary files /dev/null and b/fairseq/models/__pycache__/fconv_lm.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/fconv_self_att.cpython-310.pyc b/fairseq/models/__pycache__/fconv_self_att.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2c3611a19caf045c014abeb4e54de171735acc8c Binary files /dev/null and b/fairseq/models/__pycache__/fconv_self_att.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/lightconv.cpython-310.pyc b/fairseq/models/__pycache__/lightconv.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cf0c35473484632803abee533df83ea60cdcdbb3 Binary files /dev/null and b/fairseq/models/__pycache__/lightconv.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/lightconv_lm.cpython-310.pyc b/fairseq/models/__pycache__/lightconv_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6f2ded58ffaa8906236eb70958e3d5dec6c5a8e0 Binary files /dev/null and b/fairseq/models/__pycache__/lightconv_lm.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/lstm.cpython-310.pyc b/fairseq/models/__pycache__/lstm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0f009f56398a0a982f1ac509e946026b27175bf2 Binary files /dev/null and b/fairseq/models/__pycache__/lstm.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/lstm_lm.cpython-310.pyc b/fairseq/models/__pycache__/lstm_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..196c61b8fd0a6197e7ee76e8fde861f96a38c1da Binary files /dev/null and b/fairseq/models/__pycache__/lstm_lm.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/masked_lm.cpython-310.pyc b/fairseq/models/__pycache__/masked_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..83b9a5635a1e4109c6ffac4e638021a1664ab928 Binary files /dev/null and b/fairseq/models/__pycache__/masked_lm.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/model_utils.cpython-310.pyc b/fairseq/models/__pycache__/model_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f261c19ee2d6b4917db2e70bcbed59adeac20ab6 Binary files /dev/null and b/fairseq/models/__pycache__/model_utils.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/multilingual_transformer.cpython-310.pyc b/fairseq/models/__pycache__/multilingual_transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e669ce5b860dde61f385f4ac12bdd3a735f6530f Binary files /dev/null and b/fairseq/models/__pycache__/multilingual_transformer.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/transformer.cpython-310.pyc b/fairseq/models/__pycache__/transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..93fd92e0916756734d53679ae15c4eb9aa8640df Binary files /dev/null and b/fairseq/models/__pycache__/transformer.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/transformer_align.cpython-310.pyc b/fairseq/models/__pycache__/transformer_align.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..83eb5091aa5da58c49500b72fc5813c3c9381138 Binary files /dev/null and b/fairseq/models/__pycache__/transformer_align.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/transformer_from_pretrained_xlm.cpython-310.pyc b/fairseq/models/__pycache__/transformer_from_pretrained_xlm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a84183f493fc8a791c8aa4513617d4aafe479814 Binary files /dev/null and b/fairseq/models/__pycache__/transformer_from_pretrained_xlm.cpython-310.pyc differ diff --git a/fairseq/models/__pycache__/transformer_lm.cpython-310.pyc b/fairseq/models/__pycache__/transformer_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3a5b1280469b6846ccde2757b3911af2974e07e9 Binary files /dev/null and b/fairseq/models/__pycache__/transformer_lm.cpython-310.pyc differ diff --git a/fairseq/models/bart/__init__.py b/fairseq/models/bart/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a701923f7e5a2a8aa9b75e5580ddea22907f53ee --- /dev/null +++ b/fairseq/models/bart/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .hub_interface import * # noqa +from .model import * # noqa diff --git a/fairseq/models/bart/__pycache__/__init__.cpython-310.pyc b/fairseq/models/bart/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e0a70e92a3d3067df8f0f49b373909b7fc621bbf Binary files /dev/null and b/fairseq/models/bart/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/models/bart/__pycache__/hub_interface.cpython-310.pyc b/fairseq/models/bart/__pycache__/hub_interface.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b9824277d744697b3583a8104b6885f5d4563216 Binary files /dev/null and b/fairseq/models/bart/__pycache__/hub_interface.cpython-310.pyc differ diff --git a/fairseq/models/bart/__pycache__/model.cpython-310.pyc b/fairseq/models/bart/__pycache__/model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e7e08a4d573a65722d86723ae4c05bbb7b4a55d0 Binary files /dev/null and b/fairseq/models/bart/__pycache__/model.cpython-310.pyc differ diff --git a/fairseq/models/bart/hub_interface.py b/fairseq/models/bart/hub_interface.py new file mode 100644 index 0000000000000000000000000000000000000000..48c59cb91df3a693a68db1bd1ab7004f23b8dcfa --- /dev/null +++ b/fairseq/models/bart/hub_interface.py @@ -0,0 +1,186 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import copy +import logging + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F + +from typing import List + +from fairseq import utils +from fairseq.data import encoders + + +logger = logging.getLogger(__name__) + + +class BARTHubInterface(nn.Module): + """A simple PyTorch Hub interface to BART. + + Usage: https://github.com/pytorch/fairseq/tree/master/examples/bart + """ + + def __init__(self, args, task, model): + super().__init__() + self.args = args + self.task = task + self.model = model + + self.bpe = encoders.build_bpe(args) + + self.max_positions = min(utils.resolve_max_positions( + self.task.max_positions(), + self.model.max_positions(), + )) + + # this is useful for determining the device + self.register_buffer('_float_tensor', torch.tensor([0], dtype=torch.float)) + + @property + def device(self): + return self._float_tensor.device + + def encode(self, sentence: str, *addl_sentences, no_separator=True) -> torch.LongTensor: + """ + BPE-encode a sentence (or multiple sentences). + + Every sequence begins with a beginning-of-sentence (``) symbol. + Every sentence ends with an end-of-sentence (``). + + Example (single sentence): ` a b c ` + Example (sentence pair): ` d e f 1 2 3 ` + + The BPE encoding follows GPT-2. One subtle detail is that the GPT-2 BPE + requires leading spaces. For example:: + + >>> bart.encode('Hello world').tolist() + [0, 31414, 232, 2] + >>> bart.encode(' world').tolist() + [0, 232, 2] + >>> bart.encode('world').tolist() + [0, 8331, 2] + """ + tokens = self.bpe.encode(sentence) + if len(tokens.split(' ')) > self.max_positions - 2: + tokens = ' '.join(tokens.split(' ')[:self.max_positions - 2]) + bpe_sentence = ' ' + tokens + ' ' + for s in addl_sentences: + bpe_sentence += (' ' if not no_separator else '') + bpe_sentence += ' ' + self.bpe.encode(s) + ' ' + tokens = self.task.source_dictionary.encode_line(bpe_sentence, append_eos=False) + return tokens.long() + + def decode(self, tokens: torch.LongTensor): + assert tokens.dim() == 1 + tokens = tokens.cpu().numpy() + if tokens[0] == self.task.source_dictionary.bos(): + tokens = tokens[1:] # remove + eos_mask = (tokens == self.task.source_dictionary.eos()) + doc_mask = eos_mask[1:] & eos_mask[:-1] + sentences = np.split(tokens, doc_mask.nonzero()[0] + 1) + sentences = [self.bpe.decode(self.task.source_dictionary.string(s)) for s in sentences] + if len(sentences) == 1: + return sentences[0] + return sentences + + def _build_sample(self, src_tokens: List[torch.LongTensor]): + # assert torch.is_tensor(src_tokens) + dataset = self.task.build_dataset_for_inference( + src_tokens, + [x.numel() for x in src_tokens], + ) + sample = dataset.collater(dataset) + sample = utils.apply_to_sample( + lambda tensor: tensor.to(self.device), + sample + ) + return sample + + def sample(self, sentences: List[str], beam: int = 1, verbose: bool = False, **kwargs) -> str: + input = [self.encode(sentence) for sentence in sentences] + hypos = self.generate(input, beam, verbose, **kwargs) + return [self.decode(x['tokens']) for x in hypos] + + def generate(self, tokens: List[torch.LongTensor], beam: int = 5, verbose: bool = False, **kwargs) -> torch.LongTensor: + sample = self._build_sample(tokens) + + # build generator using current args as well as any kwargs + gen_args = copy.copy(self.args) + gen_args.beam = beam + for k, v in kwargs.items(): + setattr(gen_args, k, v) + generator = self.task.build_generator([self.model], gen_args) + translations = self.task.inference_step( + generator, + [self.model], + sample, + prefix_tokens=sample['net_input']['src_tokens'].new_zeros((len(tokens), 1)).fill_(self.task.source_dictionary.bos()), + ) + + if verbose: + src_str_with_unk = self.string(tokens) + logger.info('S\t{}'.format(src_str_with_unk)) + + def getarg(name, default): + return getattr(gen_args, name, getattr(self.args, name, default)) + + # Process top predictions + hypos = [x[0] for x in translations] + hypos = [v for _, v in sorted(zip(sample['id'].tolist(), hypos))] + return hypos + + def extract_features(self, tokens: torch.LongTensor, return_all_hiddens: bool = False) -> torch.Tensor: + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + if tokens.size(-1) > min(self.model.max_positions()): + raise ValueError('tokens exceeds maximum length: {} > {}'.format( + tokens.size(-1), self.model.max_positions() + )) + tokens.to(device=self.device), + prev_output_tokens = tokens.clone() + + prev_output_tokens[:, 0] = tokens.gather( + 1, + (tokens.ne(self.task.source_dictionary.pad()).sum(dim=1)- 1).unsqueeze(-1), + ).squeeze() + + prev_output_tokens[:, 1:] = tokens[:, :-1] + features, extra = self.model( + src_tokens=tokens, + src_lengths=None, + prev_output_tokens=prev_output_tokens, + features_only=True, + return_all_hiddens=return_all_hiddens, + ) + if return_all_hiddens: + # convert from T x B x C -> B x T x C + inner_states = extra['inner_states'] + return [inner_state.transpose(0, 1) for inner_state in inner_states] + else: + return features # just the last layer's features + + def register_classification_head( + self, name: str, num_classes: int = None, embedding_size: int = None, **kwargs + ): + self.model.register_classification_head( + name, num_classes=num_classes, embedding_size=embedding_size, **kwargs + ) + + def predict(self, head: str, tokens: torch.LongTensor, return_logits: bool = False): + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + features = self.extract_features(tokens.to(device=self.device)) + sentence_representation = features[ + tokens.eq(self.task.source_dictionary.eos()), : + ].view(features.size(0), -1, features.size(-1))[:, -1, :] + + logits = self.model.classification_heads[head](sentence_representation) + if return_logits: + return logits + return F.log_softmax(logits, dim=-1) diff --git a/fairseq/models/bart/model.py b/fairseq/models/bart/model.py new file mode 100644 index 0000000000000000000000000000000000000000..62c495cb640dfd262ef2726144e38a9658b3409c --- /dev/null +++ b/fairseq/models/bart/model.py @@ -0,0 +1,323 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +BART: Denoising Sequence-to-Sequence Pre-training for +Natural Language Generation, Translation, and Comprehension +""" + +import logging + +import torch +import torch.nn as nn + +from fairseq import utils +from fairseq.models import ( + register_model, + register_model_architecture, +) +from fairseq.models.transformer import TransformerModel +from fairseq.modules.transformer_sentence_encoder import init_bert_params + +from .hub_interface import BARTHubInterface + + +logger = logging.getLogger(__name__) + + +@register_model('bart') +class BARTModel(TransformerModel): + + @classmethod + def hub_models(cls): + return { + 'bart.base': 'http://dl.fbaipublicfiles.com/fairseq/models/bart.base.tar.gz', + 'bart.large': 'http://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz', + 'bart.large.mnli': 'http://dl.fbaipublicfiles.com/fairseq/models/bart.large.mnli.tar.gz', + 'bart.large.cnn': 'http://dl.fbaipublicfiles.com/fairseq/models/bart.large.cnn.tar.gz', + 'bart.large.xsum': 'http://dl.fbaipublicfiles.com/fairseq/models/bart.large.xsum.tar.gz', + } + + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + + # We follow BERT's random weight initialization + self.apply(init_bert_params) + + self.classification_heads = nn.ModuleDict() + + @staticmethod + def add_args(parser): + super(BARTModel, BARTModel).add_args(parser) + parser.add_argument( + '--pooler-dropout', type=float, metavar='D', + help='dropout probability in the masked_lm pooler layers' + ) + parser.add_argument( + '--pooler-activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use for pooler layer' + ) + + @property + def supported_targets(self): + return {'self'} + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, + features_only=False, classification_head_name=None, **kwargs + ): + if classification_head_name is not None: + features_only = True + + encoder_out = self.encoder( + src_tokens, + src_lengths=src_lengths, + **kwargs, + ) + x, extra = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + features_only=features_only, + **kwargs, + ) + + if classification_head_name is not None: + sentence_representation = x[ + src_tokens.eq(self.encoder.dictionary.eos()), : + ].view(x.size(0), -1, x.size(-1))[:, -1, :] + x = self.classification_heads[classification_head_name]( + sentence_representation + ) + return x, extra + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file='model.pt', + data_name_or_path='.', + bpe='gpt2', + **kwargs, + ): + from fairseq import hub_utils + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + return BARTHubInterface(x['args'], x['task'], x['models'][0]) + + def register_classification_head(self, name, num_classes=None, inner_dim=None, **kwargs): + """Register a classification head.""" + logger.info("Registering classification head: {0}".format(name)) + if name in self.classification_heads: + prev_num_classes = self.classification_heads[name].out_proj.out_features + prev_inner_dim = self.classification_heads[name].dense.out_features + if num_classes != prev_num_classes or inner_dim != prev_inner_dim: + logger.warning( + 're-registering head "{}" with num_classes {} (prev: {}) ' + 'and inner_dim {} (prev: {})'.format( + name, num_classes, prev_num_classes, inner_dim, prev_inner_dim + ) + ) + self.classification_heads[name] = BARTClassificationHead( + self.args.encoder_embed_dim, + inner_dim or self.args.encoder_embed_dim, + num_classes, + self.args.pooler_activation_fn, + self.args.pooler_dropout, + ) + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + + prefix = name + '.' if name != '' else '' + current_head_names = [] if not hasattr(self, 'classification_heads') else \ + self.classification_heads.keys() + + # Handle new classification heads present in the state dict. + keys_to_delete = [] + for k in state_dict.keys(): + if not k.startswith(prefix + 'classification_heads.'): + continue + + head_name = k[len(prefix + 'classification_heads.'):].split('.')[0] + num_classes = state_dict[prefix + 'classification_heads.' + head_name + '.out_proj.weight'].size(0) + inner_dim = state_dict[prefix + 'classification_heads.' + head_name + '.dense.weight'].size(0) + + if getattr(self.args, 'load_checkpoint_heads', False): + if head_name not in current_head_names: + self.register_classification_head(head_name, num_classes, inner_dim) + else: + if head_name not in current_head_names: + logger.warning( + 'deleting classification head ({}) from checkpoint ' + 'not present in current model: {}'.format(head_name, k) + ) + keys_to_delete.append(k) + elif ( + num_classes != self.classification_heads[head_name].out_proj.out_features + or inner_dim != self.classification_heads[head_name].dense.out_features + ): + logger.warning( + 'deleting classification head ({}) from checkpoint ' + 'with different dimensions than current model: {}'.format(head_name, k) + ) + keys_to_delete.append(k) + for k in keys_to_delete: + del state_dict[k] + + def truncate_emb(key): + if key in state_dict: + state_dict[key] = state_dict[key][:-1, :] + + # When finetuning on translation task, remove last row of + # embedding matrix that corresponds to mask_idx token. + loaded_dict_size = state_dict['encoder.embed_tokens.weight'].size(0) + if loaded_dict_size == len(self.encoder.dictionary) + 1 and '' not in self.encoder.dictionary: + truncate_emb('encoder.embed_tokens.weight') + truncate_emb('decoder.embed_tokens.weight') + truncate_emb('encoder.output_projection.weight') + truncate_emb('decoder.output_projection.weight') + + # When continued pretraining on new set of languages for mbart, + # add extra lang embeddings at the end of embed_tokens. + # Note: newly added languages are assumed to have been added at the end. + if self.args.task == 'multilingual_denoising' and loaded_dict_size < len(self.encoder.dictionary): + logger.info( + "Adding extra language embeddings not found in pretrained model for "\ + "continued pretraining of MBART on new set of languages." + ) + loaded_mask_token_embedding = state_dict['encoder.embed_tokens.weight'][-1, :] + + num_langids_to_add = len(self.encoder.dictionary) - loaded_dict_size + embed_dim = state_dict['encoder.embed_tokens.weight'].size(1) + + new_lang_embed_to_add = torch.zeros(num_langids_to_add, embed_dim) + nn.init.normal_( + new_lang_embed_to_add, + mean=0, + std=embed_dim ** -0.5 + ) + new_lang_embed_to_add = new_lang_embed_to_add.to( + dtype=state_dict['encoder.embed_tokens.weight'].dtype, + ) + + state_dict['encoder.embed_tokens.weight'] = torch.cat([ + state_dict['encoder.embed_tokens.weight'][:loaded_dict_size-1, :], + new_lang_embed_to_add, + loaded_mask_token_embedding.unsqueeze(0)] + ) + state_dict['decoder.embed_tokens.weight'] = torch.cat([ + state_dict['decoder.embed_tokens.weight'][:loaded_dict_size-1, :], + new_lang_embed_to_add, + loaded_mask_token_embedding.unsqueeze(0)] + ) + + # Copy any newly-added classification heads into the state dict + # with their current weights. + if hasattr(self, 'classification_heads'): + cur_state = self.classification_heads.state_dict() + for k, v in cur_state.items(): + if prefix + 'classification_heads.' + k not in state_dict: + logger.info('Overwriting', prefix + 'classification_heads.' + k) + state_dict[prefix + 'classification_heads.' + k] = v + + +class BARTClassificationHead(nn.Module): + """Head for sentence-level classification tasks.""" + + def __init__( + self, + input_dim, + inner_dim, + num_classes, + activation_fn, + pooler_dropout, + ): + super().__init__() + self.dense = nn.Linear(input_dim, inner_dim) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.dropout = nn.Dropout(p=pooler_dropout) + self.out_proj = nn.Linear(inner_dim, num_classes) + + def forward(self, features, **kwargs): + x = features + x = self.dropout(x) + x = self.dense(x) + x = self.activation_fn(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +@register_model_architecture('bart', 'bart_large') +def bart_large_architecture(args): + args.encoder_embed_path = getattr(args, 'encoder_embed_path', None) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1024) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4*1024) + args.encoder_layers = getattr(args, 'encoder_layers', 12) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 16) + args.encoder_normalize_before = getattr(args, 'encoder_normalize_before', False) + args.encoder_learned_pos = getattr(args, 'encoder_learned_pos', True) + args.decoder_embed_path = getattr(args, 'decoder_embed_path', None) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', args.encoder_ffn_embed_dim) + args.decoder_layers = getattr(args, 'decoder_layers', 12) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 16) + args.decoder_normalize_before = getattr(args, 'decoder_normalize_before', False) + args.decoder_learned_pos = getattr(args, 'decoder_learned_pos', True) + args.attention_dropout = getattr(args, 'attention_dropout', 0.) + args.relu_dropout = getattr(args, 'relu_dropout', 0.) + args.dropout = getattr(args, 'dropout', 0.1) + args.max_target_positions = getattr(args, 'max_target_positions', 1024) + args.max_source_positions = getattr(args, 'max_source_positions', 1024) + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', None) + args.adaptive_softmax_dropout = getattr(args, 'adaptive_softmax_dropout', 0) + args.share_decoder_input_output_embed = getattr(args, 'share_decoder_input_output_embed', True) + args.share_all_embeddings = getattr(args, 'share_all_embeddings', True) + + args.decoder_output_dim = getattr(args, 'decoder_output_dim', args.decoder_embed_dim) + args.decoder_input_dim = getattr(args, 'decoder_input_dim', args.decoder_embed_dim) + + args.no_scale_embedding = getattr(args, 'no_scale_embedding', True) + args.layernorm_embedding = getattr(args, 'layernorm_embedding', True) + + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + args.pooler_activation_fn = getattr(args, 'pooler_activation_fn', 'tanh') + args.pooler_dropout = getattr(args, 'pooler_dropout', 0.0) + + +@register_model_architecture('bart', 'bart_base') +def bart_base_architecture(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 768) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4*768) + args.encoder_layers = getattr(args, 'encoder_layers', 6) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 12) + args.decoder_layers = getattr(args, 'decoder_layers', 6) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 12) + bart_large_architecture(args) + + +@register_model_architecture('bart', 'mbart_large') +def mbart_large_architecture(args): + args.no_scale_embedding = getattr(args, 'no_scale_embedding', False) + bart_large_architecture(args) + + +@register_model_architecture('bart', 'mbart_base') +def mbart_base_architecture(args): + args.no_scale_embedding = getattr(args, 'no_scale_embedding', False) + bart_base_architecture(args) + + +@register_model_architecture('bart', 'mbart_base_wmt20') +def mbart_base_wmt20_architecture(args): + args.layernorm_embedding = getattr(args, 'layernorm_embedding', False) + mbart_base_architecture(args) diff --git a/fairseq/models/composite_encoder.py b/fairseq/models/composite_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..afef248cdcc5e657350ce37c8ba434bc01d70558 --- /dev/null +++ b/fairseq/models/composite_encoder.py @@ -0,0 +1,55 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.models import FairseqEncoder + + +class CompositeEncoder(FairseqEncoder): + """ + A wrapper around a dictionary of :class:`FairseqEncoder` objects. + + We run forward on each encoder and return a dictionary of outputs. The first + encoder's dictionary is used for initialization. + + Args: + encoders (dict): a dictionary of :class:`FairseqEncoder` objects. + """ + + def __init__(self, encoders): + super().__init__(next(iter(encoders.values())).dictionary) + self.encoders = encoders + for key in self.encoders: + self.add_module(key, self.encoders[key]) + + def forward(self, src_tokens, src_lengths): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of shape + `(batch)` + + Returns: + dict: + the outputs from each Encoder + """ + encoder_out = {} + for key in self.encoders: + encoder_out[key] = self.encoders[key](src_tokens, src_lengths) + return encoder_out + + def reorder_encoder_out(self, encoder_out, new_order): + """Reorder encoder output according to new_order.""" + for key in self.encoders: + encoder_out[key] = self.encoders[key].reorder_encoder_out(encoder_out[key], new_order) + return encoder_out + + def max_positions(self): + return min(self.encoders[key].max_positions() for key in self.encoders) + + def upgrade_state_dict(self, state_dict): + for key in self.encoders: + self.encoders[key].upgrade_state_dict(state_dict) + return state_dict diff --git a/fairseq/models/distributed_fairseq_model.py b/fairseq/models/distributed_fairseq_model.py new file mode 100644 index 0000000000000000000000000000000000000000..dd74bf1f1344f0fbafa6b13bbe8f43b343561100 --- /dev/null +++ b/fairseq/models/distributed_fairseq_model.py @@ -0,0 +1,105 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import inspect + +import torch.nn as nn + +from fairseq.legacy_distributed_data_parallel import LegacyDistributedDataParallel +from fairseq.models import BaseFairseqModel + + +_GOSSIP_DISABLED = False +try: + import gossip +except ImportError: + _GOSSIP_DISABLED = True + + +def DistributedFairseqModel(args, model, process_group=None): + """ + Wrap a *model* to support distributed data parallel training. + + This is similar to the built-in DistributedDataParallel, but allows + additional configuration of the DistributedDataParallel class to + use, and also provides easier access to the wrapped model by + forwarding requests for missing attributes to the wrapped model. + + Args: + args (argparse.Namespace): fairseq args + model (BaseFairseqModel): model to wrap + """ + # determine which DDP class to extend + assert isinstance(model, nn.Module) + if args.distributed_wrapper == 'DDP' and args.ddp_backend == 'c10d': + ddp_class = nn.parallel.DistributedDataParallel + init_kwargs = dict( + module=model, + device_ids=[args.device_id], + output_device=args.device_id, + broadcast_buffers=args.broadcast_buffers, + bucket_cap_mb=args.bucket_cap_mb, + process_group=process_group, + ) + # Maintain backward compatibility + if 'check_reduction' in inspect.getargspec(ddp_class)[0]: + init_kwargs['check_reduction'] = True + if 'find_unused_parameters' in inspect.getargspec(ddp_class)[0]: + init_kwargs['find_unused_parameters'] = args.find_unused_parameters + elif args.distributed_wrapper == 'DDP' and args.ddp_backend == 'no_c10d': + ddp_class = LegacyDistributedDataParallel + init_kwargs = dict( + module=model, + world_size=args.distributed_world_size, + buffer_size=2**28, + process_group=process_group, + ) + elif args.distributed_wrapper == 'SlowMo': + if _GOSSIP_DISABLED: + raise ImportError( + 'Cannot find gossip library. Please install from: ' + 'github.com/facebookresearch/stochastic_gradient_push' + ) + ddp_class = gossip.GossipDataParallel + + # The values of slowmo_momentum below were obtained by tuning on the + # En-De 16 dataset by training the transformer_wmt_en_de_large model + if args.slowmo_momentum is None: + if args.distributed_world_size <= 16: + args.slowmo_momentum = 0.0 + elif args.distributed_world_size <= 32: + args.slowmo_momentum = 0.2 + elif args.distributed_world_size <= 64: + args.slowmo_momentum = 0.5 + else: + args.slowmo_momentum = 0.6 + + init_kwargs = dict( + module=model, + device_ids=[args.device_id], + output_device=args.device_id, + broadcast_buffers=args.broadcast_buffers, + nprocs_per_node=args.nprocs_per_node, + slowmo_momentum=args.slowmo_momentum, + localsgd=(args.slowmo_algorithm == 'LocalSGD'), + localsgd_frequency=args.localsgd_frequency + ) + else: + raise ValueError('Unknown --ddp-backend: ' + args.ddp_backend) + + class _DistributedFairseqModel(ddp_class): + """Extend DistributedDataParallel to check for missing + attributes in the wrapped module.""" + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def __getattr__(self, name): + wrapped_module = super().__getattr__('module') + if hasattr(wrapped_module, name): + return getattr(wrapped_module, name) + return super().__getattr__(name) + + return _DistributedFairseqModel(**init_kwargs) diff --git a/fairseq/models/fairseq_decoder.py b/fairseq/models/fairseq_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..fb6c52dc7ffd95c63e0b43512db398cbb8b91582 --- /dev/null +++ b/fairseq/models/fairseq_decoder.py @@ -0,0 +1,90 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional, Tuple + +import torch.nn as nn +from fairseq import utils +from torch import Tensor + + +class FairseqDecoder(nn.Module): + """Base class for decoders.""" + + def __init__(self, dictionary): + super().__init__() + self.dictionary = dictionary + self.onnx_trace = False + + def forward(self, prev_output_tokens, encoder_out=None, **kwargs): + """ + Args: + prev_output_tokens (LongTensor): shifted output tokens of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (dict, optional): output from the encoder, used for + encoder-side attention + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + x, extra = self.extract_features( + prev_output_tokens, encoder_out=encoder_out, **kwargs + ) + x = self.output_layer(x) + return x, extra + + def extract_features(self, prev_output_tokens, encoder_out=None, **kwargs): + """ + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + raise NotImplementedError + + def output_layer(self, features, **kwargs): + """ + Project features to the default output size, e.g., vocabulary size. + + Args: + features (Tensor): features returned by *extract_features*. + """ + raise NotImplementedError + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + + if hasattr(self, "adaptive_softmax") and self.adaptive_softmax is not None: + if sample is not None: + assert "target" in sample + target = sample["target"] + else: + target = None + out = self.adaptive_softmax.get_log_prob(net_output[0], target=target) + return out.exp_() if not log_probs else out + + logits = net_output[0] + if log_probs: + return utils.log_softmax(logits, dim=-1, onnx_trace=self.onnx_trace) + else: + return utils.softmax(logits, dim=-1, onnx_trace=self.onnx_trace) + + def max_positions(self): + """Maximum input length supported by the decoder.""" + return 1e6 # an arbitrary large number + + def upgrade_state_dict(self, state_dict): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + def prepare_for_onnx_export_(self): + self.onnx_trace = True diff --git a/fairseq/models/fairseq_encoder.py b/fairseq/models/fairseq_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..7ddc0fba01ffd6ce0fa7b2bdb2761ea01392bbcf --- /dev/null +++ b/fairseq/models/fairseq_encoder.py @@ -0,0 +1,91 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +from typing import Dict, List, NamedTuple, Optional +from torch import Tensor + +EncoderOut = NamedTuple( + "EncoderOut", + [ + ("encoder_out", Tensor), # T x B x C + ("encoder_padding_mask", Optional[Tensor]), # B x T + ("encoder_embedding", Optional[Tensor]), # B x T x C + ("encoder_states", Optional[List[Tensor]]), # List[T x B x C] + ("src_tokens", Optional[Tensor]), # B x T + ("src_lengths", Optional[Tensor]), # B x 1 + ], +) + + +class FairseqEncoder(nn.Module): + """Base class for encoders.""" + + def __init__(self, dictionary): + super().__init__() + self.dictionary = dictionary + + def forward(self, src_tokens, src_lengths=None, **kwargs): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of shape + `(batch)` + """ + raise NotImplementedError + + def forward_torchscript(self, net_input: Dict[str, Tensor]): + """A TorchScript-compatible version of forward. + + Encoders which use additional arguments may want to override + this method for TorchScript compatibility. + """ + if torch.jit.is_scripting(): + return self.forward( + src_tokens=net_input["src_tokens"], + src_lengths=net_input["src_lengths"], + ) + else: + return self.forward_non_torchscript(net_input) + + @torch.jit.unused + def forward_non_torchscript(self, net_input: Dict[str, Tensor]): + encoder_input = { + k: v + for k, v in net_input.items() + if k != "prev_output_tokens" + } + return self.forward(**encoder_input) + + def reorder_encoder_out(self, encoder_out, new_order): + """ + Reorder encoder output according to `new_order`. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + `encoder_out` rearranged according to `new_order` + """ + raise NotImplementedError + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return 1e6 # an arbitrary large number + + def upgrade_state_dict(self, state_dict): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + def set_num_updates(self, num_updates): + """State from trainer to pass along to model at every update.""" + + def _apply(m): + if hasattr(m, 'set_num_updates') and m != self: + m.set_num_updates(num_updates) + self.apply(_apply) diff --git a/fairseq/models/fairseq_incremental_decoder.py b/fairseq/models/fairseq_incremental_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..68e583fea8c368f01a493fab41dc275cb18c76e2 --- /dev/null +++ b/fairseq/models/fairseq_incremental_decoder.py @@ -0,0 +1,112 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import Dict, Optional + +from torch import Tensor + +from fairseq.models import FairseqDecoder +from fairseq.incremental_decoding_utils import with_incremental_state + + +logger = logging.getLogger(__name__) + + +@with_incremental_state +class FairseqIncrementalDecoder(FairseqDecoder): + """Base class for incremental decoders. + + Incremental decoding is a special mode at inference time where the Model + only receives a single timestep of input corresponding to the previous + output token (for teacher forcing) and must produce the next output + *incrementally*. Thus the model must cache any long-term state that is + needed about the sequence, e.g., hidden states, convolutional states, etc. + + Compared to the standard :class:`FairseqDecoder` interface, the incremental + decoder interface allows :func:`forward` functions to take an extra keyword + argument (*incremental_state*) that can be used to cache state across + time-steps. + + The :class:`FairseqIncrementalDecoder` interface also defines the + :func:`reorder_incremental_state` method, which is used during beam search + to select and reorder the incremental state based on the selection of beams. + + To learn more about how incremental decoding works, refer to `this blog + `_. + """ + + def __init__(self, dictionary): + super().__init__(dictionary) + + def forward(self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs): + """ + Args: + prev_output_tokens (LongTensor): shifted output tokens of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (dict, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict, optional): dictionary used for storing + state during :ref:`Incremental decoding` + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + raise NotImplementedError + + def extract_features(self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs): + """ + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + raise NotImplementedError + + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + """Reorder incremental state. + + This will be called when the order of the input has changed from the + previous time step. A typical use case is beam search, where the input + order changes between time steps based on the selection of beams. + """ + pass + + def reorder_incremental_state_scripting( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + """Main entry point for reordering the incremental state. + + Due to limitations in TorchScript, we call this function in + :class:`fairseq.sequence_generator.SequenceGenerator` instead of + calling :func:`reorder_incremental_state` directly. + """ + for module in self.modules(): + if hasattr(module, 'reorder_incremental_state'): + result = module.reorder_incremental_state(incremental_state, new_order) + if result is not None: + incremental_state = result + + def set_beam_size(self, beam_size): + """Sets the beam size in the decoder and all children.""" + if getattr(self, '_beam_size', -1) != beam_size: + seen = set() + + def apply_set_beam_size(module): + if module != self and hasattr(module, 'set_beam_size') \ + and module not in seen: + seen.add(module) + module.set_beam_size(beam_size) + + self.apply(apply_set_beam_size) + self._beam_size = beam_size diff --git a/fairseq/models/fairseq_model.py b/fairseq/models/fairseq_model.py new file mode 100644 index 0000000000000000000000000000000000000000..5cf6cba1189d94da32c4ea262e1eabaf9c82e34d --- /dev/null +++ b/fairseq/models/fairseq_model.py @@ -0,0 +1,540 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Base classes for various fairseq models. +""" + +import logging +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.checkpoint_utils import prune_state_dict +from fairseq.data import Dictionary +from fairseq.models import FairseqDecoder, FairseqEncoder +from torch import Tensor + + +logger = logging.getLogger(__name__) + + +class BaseFairseqModel(nn.Module): + """Base class for fairseq models.""" + + def __init__(self): + super().__init__() + self._is_generation_fast = False + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + pass + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + raise NotImplementedError("Model must implement the build_model method") + + def get_targets(self, sample, net_output): + """Get targets from either the sample or the net's output.""" + return sample["target"] + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + return self.get_normalized_probs_scriptable(net_output, log_probs, sample) + + # TorchScript doesn't support super() method so that the scriptable Subclass + # can't access the base class model in Torchscript. + # Current workaround is to add a helper function with different name and + # call the helper function from scriptable Subclass. + def get_normalized_probs_scriptable( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Scriptable helper function for get_normalized_probs in ~BaseFairseqModel""" + if hasattr(self, "decoder"): + return self.decoder.get_normalized_probs(net_output, log_probs, sample) + elif torch.is_tensor(net_output): + logits = net_output.float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + raise NotImplementedError + + def extract_features(self, *args, **kwargs): + """Similar to *forward* but only return features.""" + return self(*args, **kwargs) + + def max_positions(self): + """Maximum length supported by the model.""" + return None + + def load_state_dict(self, state_dict, strict=True, args=None): + """Copies parameters and buffers from *state_dict* into this module and + its descendants. + + Overrides the method in :class:`nn.Module`. Compared with that method + this additionally "upgrades" *state_dicts* from old checkpoints. + """ + self.upgrade_state_dict(state_dict) + new_state_dict = prune_state_dict(state_dict, args) + return super().load_state_dict(new_state_dict, strict) + + def upgrade_state_dict(self, state_dict): + """Upgrade old state dicts to work with newer code.""" + self.upgrade_state_dict_named(state_dict, "") + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade old state dicts to work with newer code. + + Args: + state_dict (dict): state dictionary to upgrade, in place + name (str): the state dict key corresponding to the current module + """ + assert state_dict is not None + + def do_upgrade(m, prefix): + if len(prefix) > 0: + prefix += "." + + for n, c in m.named_children(): + name = prefix + n + if hasattr(c, "upgrade_state_dict_named"): + c.upgrade_state_dict_named(state_dict, name) + elif hasattr(c, "upgrade_state_dict"): + c.upgrade_state_dict(state_dict) + do_upgrade(c, name) + + do_upgrade(self, name) + + def set_num_updates(self, num_updates): + """State from trainer to pass along to model at every update.""" + + def _apply(m): + if hasattr(m, 'set_num_updates') and m != self: + m.set_num_updates(num_updates) + self.apply(_apply) + + def prepare_for_inference_(self, args): + """Prepare model for inference.""" + kwargs = {} + kwargs['beamable_mm_beam_size'] = ( + None if getattr(args, 'no_beamable_mm', False) + else getattr(args, 'beam', 5) + ) + kwargs['need_attn'] = getattr(args, 'print_alignment', False) + if hasattr(args, 'retain_dropout'): + kwargs['retain_dropout'] = args.retain_dropout + kwargs['retain_dropout_modules'] = getattr( + args, 'retain_dropout_modules', None + ) + self.make_generation_fast_(**kwargs) + + def make_generation_fast_(self, **kwargs): + """ + Legacy entry point to optimize model for faster generation. + Prefer prepare_for_inference_. + """ + if self._is_generation_fast: + return # only apply once + self._is_generation_fast = True + + # remove weight norm from all modules in the network + def apply_remove_weight_norm(module): + try: + nn.utils.remove_weight_norm(module) + except ValueError: # this module didn't have weight norm + return + + self.apply(apply_remove_weight_norm) + + def apply_make_generation_fast_(module, prefix): + if len(prefix) > 0: + prefix += "." + + base_func = BaseFairseqModel.make_generation_fast_ + for n, m in module.named_modules(): + if ( + m != self + and hasattr(m, "make_generation_fast_") + # don't call this implementation again, e.g., if + # children modules also inherit from BaseFairseqModel + and m.make_generation_fast_.__func__ is not base_func + ): + name = prefix + n + m.make_generation_fast_(name=name, **kwargs) + + apply_make_generation_fast_(self, "") + + def train(mode=True): + if mode: + raise RuntimeError("cannot train after make_generation_fast") + + # this model should no longer be used for training + self.eval() + self.train = train + + def prepare_for_onnx_export_(self, **kwargs): + """Make model exportable via ONNX trace.""" + seen = set() + + def apply_prepare_for_onnx_export_(module): + if ( + module != self + and hasattr(module, "prepare_for_onnx_export_") + and module not in seen + ): + seen.add(module) + module.prepare_for_onnx_export_(**kwargs) + + self.apply(apply_prepare_for_onnx_export_) + + def prepare_for_tpu_(self, **kwargs): + """Optionally modify model for use on TPUs.""" + seen = set() + + def apply_prepare_for_tpu_(module): + if ( + module != self + and hasattr(module, "prepare_for_tpu_") + and module not in seen + ): + seen.add(module) + module.prepare_for_tpu_(**kwargs) + + self.apply(apply_prepare_for_tpu_) + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + **kwargs, + ): + """ + Load a :class:`~fairseq.models.FairseqModel` from a pre-trained model + file. Downloads and caches the pre-trained model file if needed. + + The base implementation returns a + :class:`~fairseq.hub_utils.GeneratorHubInterface`, which can be used to + generate translations or sample from language models. The underlying + :class:`~fairseq.models.FairseqModel` can be accessed via the + *generator.models* attribute. + + Other models may override this to implement custom hub interfaces. + + Args: + model_name_or_path (str): either the name of a pre-trained model to + load or a path/URL to a pre-trained model state dict + checkpoint_file (str, optional): colon-separated list of checkpoint + files in the model archive to ensemble (default: 'model.pt') + data_name_or_path (str, optional): point args.data to the archive + at the given path/URL. Can start with '.' or './' to reuse the + model archive path. + """ + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + **kwargs, + ) + logger.info(x["args"]) + return hub_utils.GeneratorHubInterface(x["args"], x["task"], x["models"]) + + @classmethod + def hub_models(cls): + return {} + + +class FairseqEncoderDecoderModel(BaseFairseqModel): + """Base class for encoder-decoder models. + + Args: + encoder (FairseqEncoder): the encoder + decoder (FairseqDecoder): the decoder + """ + + def __init__(self, encoder, decoder): + super().__init__() + + self.encoder = encoder + self.decoder = decoder + assert isinstance(self.encoder, FairseqEncoder) + assert isinstance(self.decoder, FairseqDecoder) + + def forward(self, src_tokens, src_lengths, prev_output_tokens, **kwargs): + """ + Run the forward pass for an encoder-decoder model. + + First feed a batch of source tokens through the encoder. Then, feed the + encoder output and previous decoder outputs (i.e., teacher forcing) to + the decoder to produce the next outputs:: + + encoder_out = self.encoder(src_tokens, src_lengths) + return self.decoder(prev_output_tokens, encoder_out) + + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): source sentence lengths of shape `(batch)` + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + decoder_out = self.decoder( + prev_output_tokens, encoder_out=encoder_out, **kwargs + ) + return decoder_out + + def forward_decoder(self, prev_output_tokens, **kwargs): + return self.decoder(prev_output_tokens, **kwargs) + + def extract_features(self, src_tokens, src_lengths, prev_output_tokens, **kwargs): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + features = self.decoder.extract_features( + prev_output_tokens, encoder_out=encoder_out, **kwargs + ) + return features + + def output_layer(self, features, **kwargs): + """Project features to the default output size (typically vocabulary size).""" + return self.decoder.output_layer(features, **kwargs) + + def max_positions(self): + """Maximum length supported by the model.""" + return (self.encoder.max_positions(), self.decoder.max_positions()) + + def max_decoder_positions(self): + """Maximum length supported by the decoder.""" + return self.decoder.max_positions() + + +class FairseqModel(FairseqEncoderDecoderModel): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + utils.deprecation_warning( + "FairseqModel is deprecated, please use FairseqEncoderDecoderModel " + "or BaseFairseqModel instead", + stacklevel=4, + ) + + +class FairseqMultiModel(BaseFairseqModel): + """Base class for combining multiple encoder-decoder models.""" + + def __init__(self, encoders, decoders): + super().__init__() + assert encoders.keys() == decoders.keys() + self.keys = list(encoders.keys()) + for key in self.keys: + assert isinstance(encoders[key], FairseqEncoder) + assert isinstance(decoders[key], FairseqDecoder) + + self.models = nn.ModuleDict( + { + key: FairseqEncoderDecoderModel(encoders[key], decoders[key]) + for key in self.keys + } + ) + + @staticmethod + def build_shared_embeddings( + dicts: Dict[str, Dictionary], + langs: List[str], + embed_dim: int, + build_embedding: callable, + pretrained_embed_path: Optional[str] = None, + ): + """ + Helper function to build shared embeddings for a set of languages after + checking that all dicts corresponding to those languages are equivalent. + + Args: + dicts: Dict of lang_id to its corresponding Dictionary + langs: languages that we want to share embeddings for + embed_dim: embedding dimension + build_embedding: callable function to actually build the embedding + pretrained_embed_path: Optional path to load pretrained embeddings + """ + shared_dict = dicts[langs[0]] + if any(dicts[lang] != shared_dict for lang in langs): + raise ValueError( + "--share-*-embeddings requires a joined dictionary: " + "--share-encoder-embeddings requires a joined source " + "dictionary, --share-decoder-embeddings requires a joined " + "target dictionary, and --share-all-embeddings requires a " + "joint source + target dictionary." + ) + return build_embedding(shared_dict, embed_dim, pretrained_embed_path) + + def forward(self, src_tokens, src_lengths, prev_output_tokens, **kwargs): + raise NotImplementedError + + def max_positions(self): + """Maximum length supported by the model.""" + return { + key: ( + self.models[key].encoder.max_positions(), + self.models[key].decoder.max_positions(), + ) + for key in self.keys + } + + def max_decoder_positions(self): + """Maximum length supported by the decoder.""" + return min(model.decoder.max_positions() for model in self.models.values()) + + @property + def encoder(self): + return self.models[self.keys[0]].encoder + + @property + def decoder(self): + return self.models[self.keys[0]].decoder + + def forward_decoder(self, prev_output_tokens, **kwargs): + return self.decoder(prev_output_tokens, **kwargs) + + def load_state_dict(self, state_dict, strict=True, args=None): + """Copies parameters and buffers from *state_dict* into this module and + its descendants. + + Overrides the method in :class:`nn.Module`. Compared with that method + this additionally "upgrades" *state_dicts* from old checkpoints. + """ + self.upgrade_state_dict(state_dict) + new_state_dict = prune_state_dict(state_dict, args) + return super().load_state_dict(new_state_dict, strict) + + +class FairseqLanguageModel(BaseFairseqModel): + """Base class for decoder-only models. + + Args: + decoder (FairseqDecoder): the decoder + """ + + def __init__(self, decoder): + super().__init__() + self.decoder = decoder + assert isinstance(self.decoder, FairseqDecoder) + + def forward(self, src_tokens, **kwargs): + """ + Run the forward pass for a decoder-only model. + + Feeds a batch of tokens through the decoder to predict the next tokens. + + Args: + src_tokens (LongTensor): tokens on which to condition the decoder, + of shape `(batch, tgt_len)` + src_lengths (LongTensor): source sentence lengths of shape `(batch)` + + Returns: + tuple: + - the decoder's output of shape `(batch, seq_len, vocab)` + - a dictionary with any model-specific outputs + """ + return self.decoder(src_tokens, **kwargs) + + def forward_decoder(self, prev_output_tokens, **kwargs): + return self.decoder(prev_output_tokens, **kwargs) + + def extract_features(self, src_tokens, **kwargs): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, seq_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + return self.decoder.extract_features(src_tokens, **kwargs) + + def output_layer(self, features, **kwargs): + """Project features to the default output size (typically vocabulary size).""" + return self.decoder.output_layer(features, **kwargs) + + def max_positions(self): + """Maximum length supported by the model.""" + return self.decoder.max_positions() + + def max_decoder_positions(self): + """Maximum length supported by the decoder.""" + return self.decoder.max_positions() + + @property + def supported_targets(self): + return {"future"} + + +class FairseqEncoderModel(BaseFairseqModel): + """Base class for encoder-only models. + + Args: + encoder (FairseqEncoder): the encoder + """ + + def __init__(self, encoder): + super().__init__() + self.encoder = encoder + assert isinstance(self.encoder, FairseqEncoder) + + def forward(self, src_tokens, src_lengths, **kwargs): + """ + Run the forward pass for a encoder-only model. + + Feeds a batch of tokens through the encoder to generate features. + + Args: + src_tokens (LongTensor): input tokens of shape `(batch, src_len)` + src_lengths (LongTensor): source sentence lengths of shape `(batch)` + + Returns: + the encoder's output, typically of shape `(batch, src_len, features)` + """ + return self.encoder(src_tokens, src_lengths, **kwargs) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + """Get normalized probabilities (or log probs) from a net's output.""" + encoder_out = net_output["encoder_out"] + if torch.is_tensor(encoder_out): + logits = encoder_out.float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + raise NotImplementedError + + def max_positions(self): + """Maximum length supported by the model.""" + return self.encoder.max_positions() diff --git a/fairseq/models/fconv.py b/fairseq/models/fconv.py new file mode 100644 index 0000000000000000000000000000000000000000..c60a2f4e5f5d4b9596146299a8baec1b031ef7f5 --- /dev/null +++ b/fairseq/models/fconv.py @@ -0,0 +1,672 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqIncrementalDecoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + AdaptiveSoftmax, BeamableMM, FairseqDropout, GradMultiply, LearnedPositionalEmbedding, + LinearizedConvolution, +) + + +@register_model('fconv') +class FConvModel(FairseqEncoderDecoderModel): + """ + A fully convolutional model, i.e. a convolutional encoder and a + convolutional decoder, as described in `"Convolutional Sequence to Sequence + Learning" (Gehring et al., 2017) `_. + + Args: + encoder (FConvEncoder): the encoder + decoder (FConvDecoder): the decoder + + The Convolutional model provides the following named architectures and + command-line arguments: + + .. argparse:: + :ref: fairseq.models.fconv_parser + :prog: + """ + + @classmethod + def hub_models(cls): + + def moses_subword(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'subword_nmt', + } + + return { + 'conv.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2'), + 'conv.wmt14.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-de.fconv-py.tar.bz2'), + 'conv.wmt17.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt17.v2.en-de.fconv-py.tar.bz2'), + } + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + self.encoder.num_attention_layers = sum(layer is not None for layer in decoder.attention) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-layers', type=str, metavar='EXPR', + help='encoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-layers', type=str, metavar='EXPR', + help='decoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='EXPR', + help='decoder attention [True, ...]') + parser.add_argument('--share-input-output-embed', action='store_true', + help='share input and output embeddings (requires' + ' --decoder-out-embed-dim and --decoder-embed-dim' + ' to be equal)') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure that all args are properly defaulted (in case there are any new ones) + base_architecture(args) + + encoder_embed_dict = None + if args.encoder_embed_path: + encoder_embed_dict = utils.parse_embedding(args.encoder_embed_path) + utils.print_embed_overlap(encoder_embed_dict, task.source_dictionary) + + decoder_embed_dict = None + if args.decoder_embed_path: + decoder_embed_dict = utils.parse_embedding(args.decoder_embed_path) + utils.print_embed_overlap(decoder_embed_dict, task.target_dictionary) + + encoder = FConvEncoder( + dictionary=task.source_dictionary, + embed_dim=args.encoder_embed_dim, + embed_dict=encoder_embed_dict, + convolutions=eval(args.encoder_layers), + dropout=args.dropout, + max_positions=args.max_source_positions, + ) + decoder = FConvDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + embed_dict=decoder_embed_dict, + convolutions=eval(args.decoder_layers), + out_embed_dim=args.decoder_out_embed_dim, + attention=eval(args.decoder_attention), + dropout=args.dropout, + max_positions=args.max_target_positions, + share_embed=args.share_input_output_embed, + ) + return FConvModel(encoder, decoder) + + +class FConvEncoder(FairseqEncoder): + """ + Convolutional encoder consisting of `len(convolutions)` layers. + + Args: + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_dim (int, optional): embedding dimension + embed_dict (str, optional): filename from which to load pre-trained + embeddings + max_positions (int, optional): maximum supported input sequence length + convolutions (list, optional): the convolutional layer structure. Each + list item `i` corresponds to convolutional layer `i`. Layers are + given as ``(out_channels, kernel_width, [residual])``. Residual + connections are added between layers when ``residual=1`` (which is + the default behavior). + dropout (float, optional): dropout to be applied before each conv layer + """ + + def __init__( + self, dictionary, embed_dim=512, embed_dict=None, max_positions=1024, + convolutions=((512, 3),) * 20, dropout=0.1, + ): + super().__init__(dictionary) + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.num_attention_layers = None + + num_embeddings = len(dictionary) + self.padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, self.padding_idx) + if embed_dict: + self.embed_tokens = utils.load_embedding(embed_dict, self.dictionary, self.embed_tokens) + + self.embed_positions = PositionalEmbedding( + max_positions, + embed_dim, + self.padding_idx, + ) + + convolutions = extend_conv_spec(convolutions) + in_channels = convolutions[0][0] + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.residuals = [] + + layer_in_channels = [in_channels] + for _, (out_channels, kernel_size, residual) in enumerate(convolutions): + if residual == 0: + residual_dim = out_channels + else: + residual_dim = layer_in_channels[-residual] + self.projections.append(Linear(residual_dim, out_channels) + if residual_dim != out_channels else None) + if kernel_size % 2 == 1: + padding = kernel_size // 2 + else: + padding = 0 + self.convolutions.append( + ConvTBC(in_channels, out_channels * 2, kernel_size, + dropout=dropout, padding=padding) + ) + self.residuals.append(residual) + in_channels = out_channels + layer_in_channels.append(out_channels) + self.fc2 = Linear(in_channels, embed_dim) + + def forward(self, src_tokens, src_lengths): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of shape + `(batch)` + + Returns: + dict: + - **encoder_out** (tuple): a tuple with two elements, where the + first element is the last encoder layer's output and the + second element is the same quantity summed with the input + embedding (used for attention). The shape of both tensors is + `(batch, src_len, embed_dim)`. + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + """ + # embed tokens and positions + x = self.embed_tokens(src_tokens) + self.embed_positions(src_tokens) + x = self.dropout_module(x) + input_embedding = x + + # project to size of convolution + x = self.fc1(x) + + # used to mask padding in input + encoder_padding_mask = src_tokens.eq(self.padding_idx).t() # -> T x B + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + residuals = [x] + # temporal convolutions + for proj, conv, res_layer in zip(self.projections, self.convolutions, self.residuals): + if res_layer > 0: + residual = residuals[-res_layer] + residual = residual if proj is None else proj(residual) + else: + residual = None + + if encoder_padding_mask is not None: + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + x = self.dropout_module(x) + if conv.kernel_size[0] % 2 == 1: + # padding is implicit in the conv + x = conv(x) + else: + padding_l = (conv.kernel_size[0] - 1) // 2 + padding_r = conv.kernel_size[0] // 2 + x = F.pad(x, (0, 0, 0, 0, padding_l, padding_r)) + x = conv(x) + x = F.glu(x, dim=2) + + if residual is not None: + x = (x + residual) * math.sqrt(0.5) + residuals.append(x) + + # T x B x C -> B x T x C + x = x.transpose(1, 0) + + # project back to size of embedding + x = self.fc2(x) + + if encoder_padding_mask is not None: + encoder_padding_mask = encoder_padding_mask.t() # -> B x T + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + # scale gradients (this only affects backward, not forward) + x = GradMultiply.apply(x, 1.0 / (2.0 * self.num_attention_layers)) + + # add output to input embedding for attention + y = (x + input_embedding) * math.sqrt(0.5) + + return { + 'encoder_out': (x, y), + 'encoder_padding_mask': encoder_padding_mask, # B x T + } + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out['encoder_out'] is not None: + encoder_out['encoder_out'] = ( + encoder_out['encoder_out'][0].index_select(0, new_order), + encoder_out['encoder_out'][1].index_select(0, new_order), + ) + if encoder_out['encoder_padding_mask'] is not None: + encoder_out['encoder_padding_mask'] = \ + encoder_out['encoder_padding_mask'].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return self.embed_positions.max_positions + + +class AttentionLayer(nn.Module): + def __init__(self, conv_channels, embed_dim, bmm=None): + super().__init__() + # projects from output of convolution to embedding dimension + self.in_projection = Linear(conv_channels, embed_dim) + # projects from embedding dimension to convolution size + self.out_projection = Linear(embed_dim, conv_channels) + + self.bmm = bmm if bmm is not None else torch.bmm + + def forward(self, x, target_embedding, encoder_out, encoder_padding_mask): + residual = x + + # attention + x = (self.in_projection(x) + target_embedding) * math.sqrt(0.5) + x = self.bmm(x, encoder_out[0]) + + # don't attend over padding + if encoder_padding_mask is not None: + x = x.float().masked_fill( + encoder_padding_mask.unsqueeze(1), + float('-inf') + ).type_as(x) # FP16 support: cast to float and back + + # softmax over last dim + sz = x.size() + x = F.softmax(x.view(sz[0] * sz[1], sz[2]), dim=1) + x = x.view(sz) + attn_scores = x + + x = self.bmm(x, encoder_out[1]) + + # scale attention output (respecting potentially different lengths) + s = encoder_out[1].size(1) + if encoder_padding_mask is None: + x = x * (s * math.sqrt(1.0 / s)) + else: + s = s - encoder_padding_mask.type_as(x).sum(dim=1, keepdim=True) # exclude padding + s = s.unsqueeze(-1) + x = x * (s * s.rsqrt()) + + # project back + x = (self.out_projection(x) + residual) * math.sqrt(0.5) + return x, attn_scores + + def make_generation_fast_(self, beamable_mm_beam_size=None, **kwargs): + """Replace torch.bmm with BeamableMM.""" + if beamable_mm_beam_size is not None: + del self.bmm + self.add_module('bmm', BeamableMM(beamable_mm_beam_size)) + + +class FConvDecoder(FairseqIncrementalDecoder): + """Convolutional decoder""" + + def __init__( + self, dictionary, embed_dim=512, embed_dict=None, out_embed_dim=256, + max_positions=1024, convolutions=((512, 3),) * 20, attention=True, + dropout=0.1, share_embed=False, positional_embeddings=True, + adaptive_softmax_cutoff=None, adaptive_softmax_dropout=0., + ): + super().__init__(dictionary) + self.register_buffer('version', torch.Tensor([2])) + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.need_attn = True + + convolutions = extend_conv_spec(convolutions) + in_channels = convolutions[0][0] + if isinstance(attention, bool): + # expand True into [True, True, ...] and do the same with False + attention = [attention] * len(convolutions) + if not isinstance(attention, list) or len(attention) != len(convolutions): + raise ValueError('Attention is expected to be a list of booleans of ' + 'length equal to the number of layers.') + + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + if embed_dict: + self.embed_tokens = utils.load_embedding(embed_dict, self.dictionary, self.embed_tokens) + + self.embed_positions = PositionalEmbedding( + max_positions, + embed_dim, + padding_idx, + ) if positional_embeddings else None + + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.attention = nn.ModuleList() + self.residuals = [] + + layer_in_channels = [in_channels] + for i, (out_channels, kernel_size, residual) in enumerate(convolutions): + if residual == 0: + residual_dim = out_channels + else: + residual_dim = layer_in_channels[-residual] + self.projections.append(Linear(residual_dim, out_channels) + if residual_dim != out_channels else None) + self.convolutions.append( + LinearizedConv1d(in_channels, out_channels * 2, kernel_size, + padding=(kernel_size - 1), dropout=dropout) + ) + self.attention.append(AttentionLayer(out_channels, embed_dim) + if attention[i] else None) + self.residuals.append(residual) + in_channels = out_channels + layer_in_channels.append(out_channels) + + self.adaptive_softmax = None + self.fc2 = self.fc3 = None + + if adaptive_softmax_cutoff is not None: + assert not share_embed + self.adaptive_softmax = AdaptiveSoftmax(num_embeddings, in_channels, adaptive_softmax_cutoff, + dropout=adaptive_softmax_dropout) + else: + self.fc2 = Linear(in_channels, out_embed_dim) + if share_embed: + assert out_embed_dim == embed_dim, \ + "Shared embed weights implies same dimensions " \ + " out_embed_dim={} vs embed_dim={}".format(out_embed_dim, embed_dim) + self.fc3 = nn.Linear(out_embed_dim, num_embeddings) + self.fc3.weight = self.embed_tokens.weight + else: + self.fc3 = Linear(out_embed_dim, num_embeddings, dropout=dropout) + + def forward(self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused): + if encoder_out is not None: + encoder_padding_mask = encoder_out['encoder_padding_mask'] + encoder_out = encoder_out['encoder_out'] + + # split and transpose encoder outputs + encoder_a, encoder_b = self._split_encoder_out(encoder_out, incremental_state) + + if self.embed_positions is not None: + pos_embed = self.embed_positions(prev_output_tokens, incremental_state) + else: + pos_embed = 0 + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + x = self._embed_tokens(prev_output_tokens, incremental_state) + + # embed tokens and combine with positional embeddings + x += pos_embed + x = self.dropout_module(x) + target_embedding = x + + # project to size of convolution + x = self.fc1(x) + + # B x T x C -> T x B x C + x = self._transpose_if_training(x, incremental_state) + + # temporal convolutions + avg_attn_scores = None + num_attn_layers = len(self.attention) + residuals = [x] + for proj, conv, attention, res_layer in zip(self.projections, self.convolutions, self.attention, + self.residuals): + if res_layer > 0: + residual = residuals[-res_layer] + residual = residual if proj is None else proj(residual) + else: + residual = None + + x = self.dropout_module(x) + x = conv(x, incremental_state) + x = F.glu(x, dim=2) + + # attention + if attention is not None: + x = self._transpose_if_training(x, incremental_state) + + x, attn_scores = attention(x, target_embedding, (encoder_a, encoder_b), encoder_padding_mask) + + if not self.training and self.need_attn: + attn_scores = attn_scores / num_attn_layers + if avg_attn_scores is None: + avg_attn_scores = attn_scores + else: + avg_attn_scores.add_(attn_scores) + + x = self._transpose_if_training(x, incremental_state) + + # residual + if residual is not None: + x = (x + residual) * math.sqrt(0.5) + residuals.append(x) + + # T x B x C -> B x T x C + x = self._transpose_if_training(x, incremental_state) + + # project back to size of vocabulary if not using adaptive softmax + if self.fc2 is not None and self.fc3 is not None: + x = self.fc2(x) + x = self.dropout_module(x) + x = self.fc3(x) + + return x, avg_attn_scores + + def reorder_incremental_state(self, incremental_state, new_order): + super().reorder_incremental_state(incremental_state, new_order) + encoder_out = utils.get_incremental_state(self, incremental_state, 'encoder_out') + if encoder_out is not None: + encoder_out = tuple(eo.index_select(0, new_order) for eo in encoder_out) + utils.set_incremental_state(self, incremental_state, 'encoder_out', encoder_out) + + def max_positions(self): + """Maximum output length supported by the decoder.""" + return self.embed_positions.max_positions if self.embed_positions is not None else float('inf') + + def upgrade_state_dict(self, state_dict): + if utils.item(state_dict.get('decoder.version', torch.Tensor([1]))[0]) < 2: + # old models use incorrect weight norm dimension + for i, conv in enumerate(self.convolutions): + # reconfigure weight norm + nn.utils.remove_weight_norm(conv) + self.convolutions[i] = nn.utils.weight_norm(conv, dim=0) + state_dict['decoder.version'] = torch.Tensor([1]) + return state_dict + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + def _embed_tokens(self, tokens, incremental_state): + if incremental_state is not None: + # keep only the last token for incremental forward pass + tokens = tokens[:, -1:] + return self.embed_tokens(tokens) + + def _split_encoder_out(self, encoder_out, incremental_state): + """Split and transpose encoder outputs. + + This is cached when doing incremental inference. + """ + cached_result = utils.get_incremental_state(self, incremental_state, 'encoder_out') + if cached_result is not None: + return cached_result + + # transpose only once to speed up attention layers + encoder_a, encoder_b = encoder_out + encoder_a = encoder_a.transpose(1, 2).contiguous() + result = (encoder_a, encoder_b) + + if incremental_state is not None: + utils.set_incremental_state(self, incremental_state, 'encoder_out', result) + return result + + def _transpose_if_training(self, x, incremental_state): + if incremental_state is None: + x = x.transpose(0, 1) + return x + + +def extend_conv_spec(convolutions): + """ + Extends convolutional spec that is a list of tuples of 2 or 3 parameters + (kernel size, dim size and optionally how many layers behind to look for residual) + to default the residual propagation param if it is not specified + """ + extended = [] + for spec in convolutions: + if len(spec) == 3: + extended.append(spec) + elif len(spec) == 2: + extended.append(spec + (1,)) + else: + raise Exception('invalid number of parameters in convolution spec ' + str(spec) + '. expected 2 or 3') + return tuple(extended) + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, 0, 0.1) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def PositionalEmbedding(num_embeddings, embedding_dim, padding_idx): + m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) + nn.init.normal_(m.weight, 0, 0.1) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, dropout=0.): + """Weight-normalized Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features) + nn.init.normal_(m.weight, mean=0, std=math.sqrt((1 - dropout) / in_features)) + nn.init.constant_(m.bias, 0) + return nn.utils.weight_norm(m) + + +def LinearizedConv1d(in_channels, out_channels, kernel_size, dropout=0., **kwargs): + """Weight-normalized Conv1d layer optimized for decoding""" + m = LinearizedConvolution(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + nn.init.normal_(m.weight, mean=0, std=std) + nn.init.constant_(m.bias, 0) + return nn.utils.weight_norm(m, dim=2) + + +def ConvTBC(in_channels, out_channels, kernel_size, dropout=0., **kwargs): + """Weight-normalized Conv1d layer""" + from fairseq.modules import ConvTBC + m = ConvTBC(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + nn.init.normal_(m.weight, mean=0, std=std) + nn.init.constant_(m.bias, 0) + return nn.utils.weight_norm(m, dim=2) + + +@register_model_architecture('fconv', 'fconv') +def base_architecture(args): + args.dropout = getattr(args, 'dropout', 0.1) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 512) + args.encoder_embed_path = getattr(args, 'encoder_embed_path', None) + args.encoder_layers = getattr(args, 'encoder_layers', '[(512, 3)] * 20') + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_embed_path = getattr(args, 'decoder_embed_path', None) + args.decoder_layers = getattr(args, 'decoder_layers', '[(512, 3)] * 20') + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 256) + args.decoder_attention = getattr(args, 'decoder_attention', 'True') + args.share_input_output_embed = getattr(args, 'share_input_output_embed', False) + + +@register_model_architecture('fconv', 'fconv_iwslt_de_en') +def fconv_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 256) + args.encoder_layers = getattr(args, 'encoder_layers', '[(256, 3)] * 4') + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 256) + args.decoder_layers = getattr(args, 'decoder_layers', '[(256, 3)] * 3') + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 256) + base_architecture(args) + + +@register_model_architecture('fconv', 'fconv_wmt_en_ro') +def fconv_wmt_en_ro(args): + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 512) + base_architecture(args) + + +@register_model_architecture('fconv', 'fconv_wmt_en_de') +def fconv_wmt_en_de(args): + convs = '[(512, 3)] * 9' # first 9 layers have 512 units + convs += ' + [(1024, 3)] * 4' # next 4 layers have 1024 units + convs += ' + [(2048, 1)] * 2' # final 2 layers use 1x1 convolutions + + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 768) + args.encoder_layers = getattr(args, 'encoder_layers', convs) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 768) + args.decoder_layers = getattr(args, 'decoder_layers', convs) + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 512) + base_architecture(args) + + +@register_model_architecture('fconv', 'fconv_wmt_en_fr') +def fconv_wmt_en_fr(args): + convs = '[(512, 3)] * 6' # first 6 layers have 512 units + convs += ' + [(768, 3)] * 4' # next 4 layers have 768 units + convs += ' + [(1024, 3)] * 3' # next 3 layers have 1024 units + convs += ' + [(2048, 1)] * 1' # next 1 layer uses 1x1 convolutions + convs += ' + [(4096, 1)] * 1' # final 1 layer uses 1x1 convolutions + + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 768) + args.encoder_layers = getattr(args, 'encoder_layers', convs) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 768) + args.decoder_layers = getattr(args, 'decoder_layers', convs) + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 512) + base_architecture(args) diff --git a/fairseq/models/fconv_lm.py b/fairseq/models/fconv_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..f2320b170059be09a815f20aa99fcecd56bece22 --- /dev/null +++ b/fairseq/models/fconv_lm.py @@ -0,0 +1,104 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import options +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.fconv import FConvDecoder + + +@register_model('fconv_lm') +class FConvLanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-layers', type=str, metavar='EXPR', + help='decoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion') + parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D', + help='sets adaptive softmax dropout for the tail projections') + parser.add_argument('--decoder-attention', type=str, metavar='EXPR', + help='decoder attention [True, ...]') + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure all arguments are present in older models + base_lm_architecture(args) + + if hasattr(args, 'max_target_positions') and not hasattr(args, 'tokens_per_sample'): + args.tokens_per_sample = args.max_target_positions + + decoder = FConvDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + convolutions=eval(args.decoder_layers), + out_embed_dim=args.decoder_embed_dim, + attention=eval(args.decoder_attention), + dropout=args.dropout, + max_positions=args.tokens_per_sample, + share_embed=False, + positional_embeddings=False, + adaptive_softmax_cutoff=( + options.eval_str_list(args.adaptive_softmax_cutoff, type=int) + if args.criterion == 'adaptive_loss' else None + ), + adaptive_softmax_dropout=args.adaptive_softmax_dropout, + ) + return FConvLanguageModel(decoder) + + +@register_model_architecture('fconv_lm', 'fconv_lm') +def base_lm_architecture(args): + args.dropout = getattr(args, 'dropout', 0.1) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 128) + args.decoder_layers = getattr(args, 'decoder_layers', '[(1268, 4)] * 13') + args.decoder_attention = getattr(args, 'decoder_attention', 'False') + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', None) + args.adaptive_softmax_dropout = getattr(args, 'adaptive_softmax_dropout', 0) + + +@register_model_architecture('fconv_lm', 'fconv_lm_dauphin_wikitext103') +def fconv_lm_dauphin_wikitext103(args): + layers = '[(850, 6)] * 3' + layers += ' + [(850, 1)] * 1' + layers += ' + [(850, 5)] * 4' + layers += ' + [(850, 1)] * 1' + layers += ' + [(850, 4)] * 3' + layers += ' + [(1024, 4)] * 1' + layers += ' + [(2048, 4)] * 1' + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 280) + args.decoder_layers = getattr(args, 'decoder_layers', layers) + args.decoder_attention = getattr(args, 'decoder_attention', 'False') + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', '10000,20000,200000') + base_lm_architecture(args) + + +@register_model_architecture('fconv_lm', 'fconv_lm_dauphin_gbw') +def fconv_lm_dauphin_gbw(args): + layers = '[(512, 5)]' + layers += ' + [(128, 1, 0), (128, 5, 0), (512, 1, 3)] * 3' + layers += ' + [(512, 1, 0), (512, 5, 0), (1024, 1, 3)] * 3' + layers += ' + [(1024, 1, 0), (1024, 5, 0), (2048, 1, 3)] * 6' + layers += ' + [(1024, 1, 0), (1024, 5, 0), (4096, 1, 3)]' + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 128) + args.decoder_layers = getattr(args, 'decoder_layers', layers) + args.decoder_attention = getattr(args, 'decoder_attention', 'False') + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', '10000,50000,200000') + base_lm_architecture(args) diff --git a/fairseq/models/fconv_self_att.py b/fairseq/models/fconv_self_att.py new file mode 100644 index 0000000000000000000000000000000000000000..c3582da96f5ad92d3eec2a1f1a70a95463e55f3b --- /dev/null +++ b/fairseq/models/fconv_self_att.py @@ -0,0 +1,589 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +import os + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import checkpoint_utils +from fairseq.models import ( + CompositeEncoder, + FairseqDecoder, + FairseqEncoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + FairseqDropout, + DownsampledMultiHeadAttention, + GradMultiply, + LayerNorm, + LearnedPositionalEmbedding, + LinearizedConvolution, +) +from fairseq.incremental_decoding_utils import with_incremental_state + +logger = logging.getLogger(__name__) + + +@register_model('fconv_self_att') +class FConvModelSelfAtt(FairseqEncoderDecoderModel): + + @classmethod + def hub_models(cls): + return { + 'conv.stories.pretrained': { + 'path': 'https://dl.fbaipublicfiles.com/fairseq/models/stories_checkpoint.tar.gz', + 'checkpoint_file': 'pretrained_checkpoint.pt', + 'tokenizer': 'nltk', + }, + 'conv.stories': { + 'path': 'https://dl.fbaipublicfiles.com/fairseq/models/stories_checkpoint.tar.gz', + 'checkpoint_file': 'fusion_checkpoint.pt', + 'tokenizer': 'nltk', + 'pretrained': 'True', + 'pretrained_checkpoint': './pretrained_checkpoint.pt', + }, + # Test set containing dictionaries + 'data.stories': 'https://dl.fbaipublicfiles.com/fairseq/data/stories_test.tar.bz2', + } + + def __init__(self, encoder, decoder, pretrained_encoder=None): + super().__init__(encoder, decoder) + self.encoder.num_attention_layers = sum(layer is not None for layer in decoder.attention) + self.pretrained_encoder = pretrained_encoder + if self.pretrained_encoder is None: + encoders = {'encoder': encoder} + else: + encoders = {'encoder': encoder, 'pretrained': self.pretrained_encoder} + # for fusion model, CompositeEncoder contains both pretrained and training encoders + # these are forwarded and then combined in the decoder + self.encoder = CompositeEncoder(encoders) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-layers', type=str, metavar='EXPR', + help='encoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-layers', type=str, metavar='EXPR', + help='decoder layers [(dim, kernel_size), ...]') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='EXPR', + help='decoder attention [True, ...]') + parser.add_argument('--self-attention', type=str, metavar='EXPR', + help='decoder self-attention layers, ex: [True] + [False]*5') + parser.add_argument('--multihead-attention-nheads', type=int, + help='Number of heads to use in attention') + parser.add_argument('--multihead-self-attention-nheads', type=int, + help='Number of heads to use in self-attention') + parser.add_argument('--encoder-attention', type=str, metavar='EXPR', + help='encoder attention [True, ...]') + parser.add_argument('--encoder-attention-nheads', type=int, + help='Number of heads to use in encoder attention') + parser.add_argument('--project-input', type=str, metavar='EXPR', + help='Use projections in self-attention [True, ...]') + parser.add_argument('--gated-attention', type=str, metavar='EXPR', + help='Use GLU layers in self-attention projections [True, ...]') + parser.add_argument('--downsample', type=str, metavar='EXPR', + help='Use downsampling in self-attention [True, ...]') + parser.add_argument('--pretrained-checkpoint', metavar='DIR', + help='path to load checkpoint from pretrained model') + parser.add_argument('--pretrained', type=str, metavar='EXPR', + help='use pretrained model when training [True, ...]') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + trained_encoder, trained_decoder = None, None + pretrained = eval(args.pretrained) + if pretrained: + logger.info('loading pretrained model') + if not os.path.exists(args.pretrained_checkpoint): + new_pretrained_checkpoint = os.path.join(args.data, args.pretrained_checkpoint) + if os.path.exists(new_pretrained_checkpoint): + args.pretrained_checkpoint = new_pretrained_checkpoint + trained_model = checkpoint_utils.load_model_ensemble( + filenames=[args.pretrained_checkpoint], + task=task, + )[0][0] + trained_decoder = list(trained_model.children())[1] + trained_encoder = list(trained_model.children())[0] + + # freeze pretrained model + for param in trained_decoder.parameters(): + param.requires_grad = False + for param in trained_encoder.parameters(): + param.requires_grad = False + + encoder = FConvEncoder( + task.source_dictionary, + embed_dim=args.encoder_embed_dim, + convolutions=eval(args.encoder_layers), + dropout=args.dropout, + max_positions=args.max_source_positions, + attention=eval(args.encoder_attention), + attention_nheads=args.encoder_attention_nheads, + ) + + decoder = FConvDecoder( + task.target_dictionary, + embed_dim=args.decoder_embed_dim, + convolutions=eval(args.decoder_layers), + out_embed_dim=args.decoder_out_embed_dim, + attention=eval(args.decoder_attention), + dropout=args.dropout, + max_positions=args.max_target_positions, + selfattention=eval(args.self_attention), + attention_nheads=args.multihead_attention_nheads, + selfattention_nheads=args.multihead_self_attention_nheads, + project_input=eval(args.project_input), + gated_attention=eval(args.gated_attention), + downsample=eval(args.downsample), + pretrained=pretrained, + trained_decoder=trained_decoder, + ) + model = FConvModelSelfAtt(encoder, decoder, trained_encoder) + + return model + + @property + def pretrained(self): + return self.pretrained_encoder is not None + + +class FConvEncoder(FairseqEncoder): + """Convolutional encoder""" + def __init__( + self, dictionary, embed_dim=512, max_positions=1024, + convolutions=((512, 3),) * 20, dropout=0.1, attention=False, + attention_nheads=1, + ): + super().__init__(dictionary) + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.num_attention_layers = None + + num_embeddings = len(dictionary) + self.padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, self.padding_idx) + self.embed_positions = PositionalEmbedding( + max_positions, + embed_dim, + self.padding_idx, + ) + + def expand_bool_array(val): + if isinstance(val, bool): + # expand True into [True, True, ...] and do the same with False + return [val] * len(convolutions) + return val + + attention = expand_bool_array(attention) + + in_channels = convolutions[0][0] + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.attention = nn.ModuleList() + self.attproj = nn.ModuleList() + for i, (out_channels, kernel_size) in enumerate(convolutions): + self.projections.append( + Linear(in_channels, out_channels) if in_channels != out_channels else None + ) + self.convolutions.append( + ConvTBC(in_channels, out_channels * 2, kernel_size, dropout=dropout) + ) + + self.attention.append( + SelfAttention(out_channels, embed_dim, attention_nheads) if attention[i] else None + ) + in_channels = out_channels + + self.fc2 = Linear(in_channels, embed_dim) + + def forward(self, src_tokens, src_lengths): + # embed tokens and positions + x = self.embed_tokens(src_tokens) + self.embed_positions(src_tokens) + x = self.dropout_module(x) + input_embedding = x.transpose(0, 1) + + # project to size of convolution + x = self.fc1(x) + + encoder_padding_mask = src_tokens.eq(self.padding_idx).t() # -> T x B + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # temporal convolutions + for proj, conv, attention in zip(self.projections, self.convolutions, self.attention): + residual = x if proj is None else proj(x) + + if encoder_padding_mask is not None: + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + x = self.dropout_module(x) + padding_l = (conv.kernel_size[0] - 1) // 2 + padding_r = conv.kernel_size[0] // 2 + x = F.pad(x, (0, 0, 0, 0, padding_l, padding_r)) + x = conv(x) + x = F.glu(x, dim=2) + if attention is not None: + x = attention(x) + x = (x + residual) * math.sqrt(0.5) + + # T x B x C -> B x T x C + x = x.transpose(1, 0) + + # project back to size of embedding + x = self.fc2(x) + + if encoder_padding_mask is not None: + encoder_padding_mask = encoder_padding_mask.t() # -> B x T + x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0) + + # scale gradients (this only affects backward, not forward) + x = GradMultiply.apply(x, 1.0 / (2.0 * self.num_attention_layers)) + + # add output to input embedding for attention + y = (x + input_embedding.transpose(0, 1)) * math.sqrt(0.5) + + return { + 'encoder_out': (x, y), + 'encoder_padding_mask': encoder_padding_mask, # B x T + } + + def reorder_encoder_out(self, encoder_out, new_order): + encoder_out['encoder_out'] = tuple( + eo.index_select(0, new_order) for eo in encoder_out['encoder_out'] + ) + + if encoder_out['encoder_padding_mask'] is not None: + encoder_out['encoder_padding_mask'] = \ + encoder_out['encoder_padding_mask'].index_select(0, new_order) + + if 'pretrained' in encoder_out: + encoder_out['pretrained']['encoder_out'] = tuple( + eo.index_select(0, new_order) + for eo in encoder_out['pretrained']['encoder_out'] + ) + + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return self.embed_positions.max_positions + + +@with_incremental_state +class FConvDecoder(FairseqDecoder): + """Convolutional decoder""" + def __init__( + self, dictionary, embed_dim=512, out_embed_dim=256, max_positions=1024, + convolutions=((512, 3),) * 8, attention=True, dropout=0.1, + selfattention=False, attention_nheads=1, selfattention_nheads=1, + project_input=False, gated_attention=False, downsample=False, + pretrained=False, trained_decoder=None, + ): + super().__init__(dictionary) + self.register_buffer('version', torch.Tensor([2])) + self.pretrained = pretrained + self.pretrained_decoder = trained_decoder + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.need_attn = True + in_channels = convolutions[0][0] + + def expand_bool_array(val): + if isinstance(val, bool): + # expand True into [True, True, ...] and do the same with False + return [val] * len(convolutions) + return val + + attention = expand_bool_array(attention) + selfattention = expand_bool_array(selfattention) + + if not isinstance(attention, list) or len(attention) != len(convolutions): + raise ValueError('Attention is expected to be a list of booleans of ' + 'length equal to the number of layers.') + + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + self.embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + + self.embed_positions = PositionalEmbedding( + max_positions, + embed_dim, + padding_idx, + ) + + self.fc1 = Linear(embed_dim, in_channels, dropout=dropout) + self.projections = nn.ModuleList() + self.convolutions = nn.ModuleList() + self.attention = nn.ModuleList() + self.selfattention = nn.ModuleList() + self.attproj = nn.ModuleList() + for i, (out_channels, kernel_size) in enumerate(convolutions): + self.projections.append( + Linear(in_channels, out_channels) if in_channels != out_channels else None + ) + self.convolutions.append( + LinearizedConv1d( + in_channels, out_channels * 2, kernel_size, + padding=(kernel_size - 1), dropout=dropout, + ) + ) + + self.attention.append( + DownsampledMultiHeadAttention( + out_channels, embed_dim, attention_nheads, + project_input=project_input, gated=False, downsample=False, + ) if attention[i] else None + ) + + self.attproj.append( + Linear(out_channels, embed_dim, dropout=dropout) if attention[i] else None + ) + self.selfattention.append( + SelfAttention( + out_channels, embed_dim, selfattention_nheads, + project_input=project_input, gated=gated_attention, + downsample=downsample, + ) if selfattention[i] else None + ) + in_channels = out_channels + + self.fc2 = Linear(in_channels, out_embed_dim) + self.fc3 = Linear(out_embed_dim, num_embeddings, dropout=dropout) + + # model fusion + if self.pretrained: + # independent gates are learned from the concatenated input + self.gate1 = nn.Sequential(Linear(out_embed_dim*2, out_embed_dim), nn.Sigmoid()) + self.gate2 = nn.Sequential(Linear(out_embed_dim*2, out_embed_dim), nn.Sigmoid()) + # pretrained and trained models are joined + self.joining = nn.Sequential( + Linear(out_embed_dim*2, out_embed_dim*2), + LayerNorm(out_embed_dim*2), + nn.GLU(), + Linear(out_embed_dim, out_embed_dim*2), + LayerNorm(out_embed_dim*2), + nn.GLU(), + Linear(out_embed_dim, out_embed_dim), + LayerNorm(out_embed_dim) + ) + # pretrained model contains an output layer that is nhid -> vocab size + # but the models are combined in their hidden state + # the hook stores the output of the pretrained model forward + self.pretrained_outputs = {} + + def save_output(): + def hook(a, b, output): + self.pretrained_outputs["out"] = output + return hook + + self.pretrained_decoder.fc2.register_forward_hook(save_output()) + + def forward(self, prev_output_tokens, encoder_out): + trained_encoder_out = encoder_out['pretrained'] if self.pretrained else None + encoder_out = encoder_out['encoder']['encoder_out'] + + encoder_a, encoder_b = self._split_encoder_out(encoder_out) + + # embed positions + positions = self.embed_positions(prev_output_tokens) + + # embed tokens and positions + x = self.embed_tokens(prev_output_tokens) + positions + x = self.dropout_module(x) + target_embedding = x.transpose(0, 1) + + # project to size of convolution + x = self.fc1(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # temporal convolutions + avg_attn_scores = None + for proj, conv, attention, selfattention, attproj in zip( + self.projections, self.convolutions, self.attention, self.selfattention, self.attproj + ): + residual = x if proj is None else proj(x) + + x = self.dropout_module(x) + x = conv(x) + x = F.glu(x, dim=2) + + # attention + if attention is not None: + r = x + x, attn_scores = attention(attproj(x) + target_embedding, encoder_a, encoder_b) + x = x + r + if not self.training and self.need_attn: + if avg_attn_scores is None: + avg_attn_scores = attn_scores + else: + avg_attn_scores.add_(attn_scores) + + if selfattention is not None: + x = selfattention(x) + + x = (x + residual) * math.sqrt(0.5) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + # project back to size of vocabulary + x = self.fc2(x) + x = self.dropout_module(x) + if not self.pretrained: + x = self.fc3(x) + + # fusion gating + if self.pretrained: + trained_x, _ = self.pretrained_decoder.forward(prev_output_tokens, trained_encoder_out) + y = torch.cat([x, self.pretrained_outputs["out"]], dim=-1) + gate1 = self.gate1(y) + gate2 = self.gate2(y) + gated_x1 = gate1 * x + gated_x2 = gate2 * self.pretrained_outputs["out"] + fusion = torch.cat([gated_x1, gated_x2], dim=-1) + fusion = self.joining(fusion) + fusion_output = self.fc3(fusion) + return fusion_output, avg_attn_scores + else: + return x, avg_attn_scores + + def max_positions(self): + """Maximum output length supported by the decoder.""" + return self.embed_positions.max_positions + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + def _split_encoder_out(self, encoder_out): + """Split and transpose encoder outputs.""" + # transpose only once to speed up attention layers + encoder_a, encoder_b = encoder_out + encoder_a = encoder_a.transpose(0, 1).contiguous() + encoder_b = encoder_b.transpose(0, 1).contiguous() + result = (encoder_a, encoder_b) + return result + + +class SelfAttention(nn.Module): + + def __init__(self, out_channels, embed_dim, num_heads, project_input=False, gated=False, downsample=False): + super().__init__() + self.attention = DownsampledMultiHeadAttention( + out_channels, embed_dim, num_heads, dropout=0, bias=True, + project_input=project_input, gated=gated, downsample=downsample, + ) + self.in_proj_q = Linear(out_channels, embed_dim) + self.in_proj_k = Linear(out_channels, embed_dim) + self.in_proj_v = Linear(out_channels, embed_dim) + self.ln = LayerNorm(out_channels) + + def forward(self, x): + residual = x + query = self.in_proj_q(x) + key = self.in_proj_k(x) + value = self.in_proj_v(x) + x, _ = self.attention(query, key, value, mask_future_timesteps=True, use_scalar_bias=True) + return self.ln(x + residual) + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + m.weight.data.normal_(0, 0.1) + return m + + +def PositionalEmbedding(num_embeddings, embedding_dim, padding_idx): + m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) + m.weight.data.normal_(0, 0.1) + return m + + +def Linear(in_features, out_features, dropout=0.): + """Weight-normalized Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features) + m.weight.data.normal_(mean=0, std=math.sqrt((1 - dropout) / in_features)) + m.bias.data.zero_() + return m + + +def LinearizedConv1d(in_channels, out_channels, kernel_size, dropout=0., **kwargs): + """Weight-normalized Conv1d layer optimized for decoding""" + m = LinearizedConvolution(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + m.weight.data.normal_(mean=0, std=std) + m.bias.data.zero_() + return m + + +def ConvTBC(in_channels, out_channels, kernel_size, dropout=0., **kwargs): + """Weight-normalized Conv1d layer""" + from fairseq.modules import ConvTBC + m = ConvTBC(in_channels, out_channels, kernel_size, **kwargs) + std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels)) + m.weight.data.normal_(mean=0, std=std) + m.bias.data.zero_() + return m + + +@register_model_architecture('fconv_self_att', 'fconv_self_att') +def base_architecture(args): + args.dropout = getattr(args, 'dropout', 0.1) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 512) + args.encoder_layers = getattr(args, 'encoder_layers', '[(512, 3)] * 3') + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_layers = getattr(args, 'decoder_layers', '[(512, 3)] * 8') + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 256) + args.decoder_attention = getattr(args, 'decoder_attention', 'True') + args.self_attention = getattr(args, 'self_attention', 'False') + args.encoder_attention = getattr(args, 'encoder_attention', 'False') + args.multihead_attention_nheads = getattr(args, 'multihead_attention_nheads', 1) + args.multihead_self_attention_nheads = getattr(args, 'multihead_self_attention_nheads', 1) + args.encoder_attention_nheads = getattr(args, 'encoder_attention_nheads', 1) + args.project_input = getattr(args, 'project_input', 'False') + args.gated_attention = getattr(args, 'gated_attention', 'False') + args.downsample = getattr(args, 'downsample', 'False') + args.pretrained_checkpoint = getattr(args, 'pretrained_checkpoint', '') + args.pretrained = getattr(args, 'pretrained', 'False') + + +@register_model_architecture('fconv_self_att', 'fconv_self_att_wp') +def fconv_self_att_wp(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 256) + args.encoder_layers = getattr(args, 'encoder_layers', '[(128, 3)] * 2 + [(512,3)] * 1') + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 256) + args.decoder_layers = getattr(args, 'decoder_layers', '[(512, 4)] * 4 + [(768, 4)] * 2 + [(1024, 4)] * 1') + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 256) + args.self_attention = getattr(args, 'self_attention', 'True') + args.multihead_self_attention_nheads = getattr(args, 'multihead_self_attention_nheads', 4) + args.project_input = getattr(args, 'project_input', 'True') + args.gated_attention = getattr(args, 'gated_attention', 'True') + args.downsample = getattr(args, 'downsample', 'True') + base_architecture(args) diff --git a/fairseq/models/huggingface/__init__.py b/fairseq/models/huggingface/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..633315f54d9b7548d381d3271c19aecd5ef1d042 --- /dev/null +++ b/fairseq/models/huggingface/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the models/huggingface/ directory +models_dir = os.path.dirname(__file__) +for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if ( + not file.startswith('_') + and not file.startswith('.') + and (file.endswith('.py') or os.path.isdir(path)) + ): + model_name = file[:file.find('.py')] if file.endswith('.py') else file + module = importlib.import_module('fairseq.models.huggingface.' + model_name) diff --git a/fairseq/models/huggingface/__pycache__/__init__.cpython-310.pyc b/fairseq/models/huggingface/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0f8508d69896ee63318b8dea27634e6aafdbe172 Binary files /dev/null and b/fairseq/models/huggingface/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/models/huggingface/__pycache__/hf_gpt2.cpython-310.pyc b/fairseq/models/huggingface/__pycache__/hf_gpt2.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..124e87882360b5d9ade0a95a2a6c3d3e5aa201e2 Binary files /dev/null and b/fairseq/models/huggingface/__pycache__/hf_gpt2.cpython-310.pyc differ diff --git a/fairseq/models/huggingface/hf_gpt2.py b/fairseq/models/huggingface/hf_gpt2.py new file mode 100644 index 0000000000000000000000000000000000000000..6a03406ef6fccedd86ef3325dec47de5218456f3 --- /dev/null +++ b/fairseq/models/huggingface/hf_gpt2.py @@ -0,0 +1,199 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import sys +from typing import Dict, List, Optional + +import torch +from fairseq.models import ( + FairseqIncrementalDecoder, + FairseqLanguageModel, + register_model, + register_model_architecture, +) + +try: + # Prepend the transformers submodule to the path, so that + # it's prioritized over other installations. This allows + # making local changes in the submodule. + sys.path.insert( + 0, os.path.join(os.path.dirname(__file__), 'transformers', 'src') + ) + from transformers import AutoModel, GPT2Config, GPT2LMHeadModel + has_hf = True +except ImportError: + has_hf = False + + +logger = logging.getLogger(__name__) + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model('hf_gpt2') +class HuggingFaceGPT2LanguageModel(FairseqLanguageModel): + + def __init__(self, decoder): + super().__init__(decoder) + if not has_hf: + raise ImportError( + '\n\nPlease install huggingface/transformers with:' + '\n\n pip install transformers' + '\n\nOr to make local edits, install the submodule:' + '\n\n git submodule update --init ' + 'fairseq/models/huggingface/transformers' + ) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--embed-dim', type=int, metavar='N', + help='embedding dimension') + parser.add_argument('--num-attention-heads', type=int, metavar='N', + help='num attention heads') + parser.add_argument('--num-layers', type=int, metavar='N', + help='num layers') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability for all fully connected layers ' + 'in the embeddings, encoder, and pooler') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + default_architecture(args) + return cls(HuggingFaceGPT2Decoder(args, task)) + + +class HuggingFaceGPT2Decoder(FairseqIncrementalDecoder): + + def __init__(self, args, task): + super().__init__(task.target_dictionary) + + try: + # Prepend the transformers submodule to the path, so that + # it's prioritized over other installations. This allows + # making local changes in the submodule. + sys.path.insert( + 0, os.path.join(os.path.dirname(__file__), 'transformers', 'src') + ) + from transformers import GPT2Config, GPT2LMHeadModel + except ImportError: + raise ImportError( + '\n\nPlease install huggingface/transformers with:' + '\n\n pip install transformers' + '\n\nOr to make local edits, install the submodule:' + '\n\n git submodule update --init ' + 'fairseq/models/huggingface/transformers' + ) + + config = GPT2Config( + vocab_size=len(task.target_dictionary), + n_positions=args.max_target_positions + 1, + n_ctx=args.max_target_positions, + n_embd=args.embed_dim, + n_layer=args.num_layers, + n_head=args.num_attention_heads, + resid_pdrop=args.dropout, + embd_pdrop=args.dropout, + attn_pdrop=args.attention_dropout, + layer_norm_epsilon=1e-6, + ) + self.model = GPT2LMHeadModel(config) + + # set zero embedding for padding symbol + self.pad_idx = task.target_dictionary.pad() + self.model.transformer.wte.weight.data[self.pad_idx].zero_() + self.model.transformer.wpe.weight.data[0].zero_() + + def forward( + self, + prev_output_tokens, + src_lengths=None, + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + encoder_out=None, + ): + features = self.extract_features(prev_output_tokens, incremental_state) + lm_logits = self.model.lm_head(features) + return (lm_logits, ) + + def extract_features( + self, + prev_output_tokens, + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + ): + if incremental_state: + past = self.get_incremental_state("past") + else: + past = None + + # don't attend to padding symbols + attention_mask = prev_output_tokens.ne(self.pad_idx).int() + + # set position ids to exclude padding symbols + position_ids = attention_mask * ( + torch.arange(1, 1 + prev_output_tokens.size(1)) + .to(prev_output_tokens) + .repeat(prev_output_tokens.size(0), 1) + ) + + outputs = self.model.transformer( + input_ids=prev_output_tokens, + past=past, + attention_mask=attention_mask, + position_ids=position_ids, + ) + last_hidden_states = outputs[0] + + if incremental_state: + self.set_incremental_state(incremental_state, "past", outputs[1]) + + return last_hidden_states + + def max_positions(self): + return self.model.config.n_positions - 1 + + +@register_model_architecture('hf_gpt2', 'hf_gpt2') +def default_architecture(args): + if getattr(args, 'max_target_positions', None) is None: + args.max_target_positions = getattr( + args, 'tokens_per_sample', DEFAULT_MAX_TARGET_POSITIONS + ) + args.embed_dim = getattr(args, 'embed_dim', 768) + args.num_attention_heads = getattr(args, 'num_attention_heads', 12) + args.num_layers = getattr(args, 'num_layers', 12) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + + +@register_model_architecture('hf_gpt2', 'hf_gpt2_medium') +def hf_gpt2_medium(args): + args.embed_dim = getattr(args, 'embed_dim', 1024) + args.num_attention_heads = getattr(args, 'num_attention_heads', 16) + args.num_layers = getattr(args, 'num_layers', 24) + default_architecture(args) + + +@register_model_architecture('hf_gpt2', 'hf_gpt2_large') +def hf_gpt2_large(args): + args.embed_dim = getattr(args, 'embed_dim', 1280) + args.num_attention_heads = getattr(args, 'num_attention_heads', 20) + args.num_layers = getattr(args, 'num_layers', 36) + default_architecture(args) + + +@register_model_architecture('hf_gpt2', 'hf_gpt2_xl') +def hf_gpt2_xl(args): + args.embed_dim = getattr(args, 'embed_dim', 1600) + args.num_attention_heads = getattr(args, 'num_attention_heads', 25) + args.num_layers = getattr(args, 'num_layers', 48) + default_architecture(args) diff --git a/fairseq/models/lightconv.py b/fairseq/models/lightconv.py new file mode 100644 index 0000000000000000000000000000000000000000..05939e1c758d6ce93b0bb4c1cd83bbc97ac98e8b --- /dev/null +++ b/fairseq/models/lightconv.py @@ -0,0 +1,786 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import options, utils +from fairseq.models import ( + FairseqEncoder, + FairseqIncrementalDecoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + AdaptiveSoftmax, + DynamicConv, + FairseqDropout, + LayerNorm, + PositionalEmbedding, + LightweightConv, + MultiheadAttention, +) + + +@register_model('lightconv') +class LightConvModel(FairseqEncoderDecoderModel): + """ + LightConv and DynamicConv model from `"Pay Less Attention with Lightweight and Dynamic Convolutions" (Wu, et al, 2019) + `_. + To use LightConv please set ``--encoder-conv-type lightweight --decoder-conv-type lightweight`` + To use DynamicConv please set ``--encoder-conv-type dynamic --decoder-conv-type dynamic`` + + Args: + encoder (LightConvEncoder): the encoder + decoder (LightConvDecoder): the decoder + + The LightConv model provides the following named architectures and + command-line arguments: + + .. argparse:: + :ref: fairseq.models.lightconv_parser + :prog: + """ + + @classmethod + def hub_models(cls): + # fmt: off + + def moses_subword(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'subword_nmt', + } + + return { + 'lightconv.no_glu.iwslt14.de-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/iwslt14.de-en.lightconv.tar.gz'), + 'dynamicconv.no_glu.iwslt14.de-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/iwslt14.de-en.dynamicconv.tar.gz'), + 'lightconv.no_glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv.tar.gz'), + 'dynamicconv.no_glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv.tar.gz'), + 'lightconv.glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt16.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv-glu.tar.gz'), + 'lightconv.glu.wmt17.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt17.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt16.en-de.joined-dict.dynamicconv-glu.tar.gz'), + 'lightconv.glu.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt14.en-fr.joined-dict.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt14.en-fr.joined-dict.dynamicconv-glu.tar.gz'), + 'lightconv.glu.wmt17.zh-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt17.zh-en.lightconv-glu.tar.gz'), + 'dynamicconv.glu.wmt17.zh-en': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/dynamicconv/wmt17.zh-en.dynamicconv-glu.tar.gz'), + } + # fmt: on + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + parser.add_argument('--relu-dropout', type=float, metavar='D', + help='dropout probability after ReLU in FFN') + parser.add_argument('--input-dropout', type=float, metavar='D', + help='dropout probability of the inputs') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-conv-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-ffn-embed-dim', type=int, metavar='N', + help='encoder embedding dimension for FFN') + parser.add_argument('--encoder-layers', type=int, metavar='N', + help='num encoder layers') + parser.add_argument('--encoder-attention-heads', type=int, metavar='N', + help='num encoder attention heads or LightConv/DynamicConv heads') + parser.add_argument('--encoder-normalize-before', action='store_true', + help='apply layernorm before each encoder block') + parser.add_argument('--encoder-learned-pos', action='store_true', + help='use learned positional embeddings in the encoder') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-conv-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-ffn-embed-dim', type=int, metavar='N', + help='decoder embedding dimension for FFN') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='num decoder layers') + parser.add_argument('--decoder-attention-heads', type=int, metavar='N', + help='num decoder attention heads or LightConv/DynamicConv heads') + parser.add_argument('--decoder-learned-pos', action='store_true', + help='use learned positional embeddings in the decoder') + parser.add_argument('--decoder-normalize-before', action='store_true', + help='apply layernorm before each decoder block') + parser.add_argument('--share-decoder-input-output-embed', action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--share-all-embeddings', action='store_true', + help='share encoder, decoder and output embeddings' + ' (requires shared dictionary and embed dim)') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion'), + parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D', + help='sets adaptive softmax dropout for the tail projections') + + """LightConv and DynamicConv arguments""" + parser.add_argument('--encoder-kernel-size-list', type=lambda x: options.eval_str_list(x, int), + help='list of kernel size (default: "[3,7,15,31,31,31,31]")') + parser.add_argument('--decoder-kernel-size-list', type=lambda x: options.eval_str_list(x, int), + help='list of kernel size (default: "[3,7,15,31,31,31]")') + parser.add_argument('--encoder-glu', type=options.eval_bool, + help='glu after in proj') + parser.add_argument('--decoder-glu', type=options.eval_bool, + help='glu after in proj') + parser.add_argument('--encoder-conv-type', default='dynamic', type=str, + choices=['dynamic', 'lightweight'], + help='type of convolution') + parser.add_argument('--decoder-conv-type', default='dynamic', type=str, + choices=['dynamic', 'lightweight'], + help='type of convolution') + parser.add_argument('--weight-softmax', default=True, type=options.eval_bool) + parser.add_argument('--weight-dropout', type=float, metavar='D', + help='dropout probability for conv weights') + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if not hasattr(args, 'max_source_positions'): + args.max_source_positions = 1024 + if not hasattr(args, 'max_target_positions'): + args.max_target_positions = 1024 + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + def build_embedding(dictionary, embed_dim, path=None): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + if args.share_all_embeddings: + if src_dict != tgt_dict: + raise RuntimeError('--share-all-embeddings requires a joined dictionary') + if args.encoder_embed_dim != args.decoder_embed_dim: + raise RuntimeError( + '--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim') + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path): + raise RuntimeError('--share-all-embeddings not compatible with --decoder-embed-path') + encoder_embed_tokens = build_embedding( + src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + encoder_embed_tokens = build_embedding( + src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = build_embedding( + tgt_dict, args.decoder_embed_dim, args.decoder_embed_path + ) + + encoder = LightConvEncoder(args, src_dict, encoder_embed_tokens) + decoder = LightConvDecoder(args, tgt_dict, decoder_embed_tokens) + return LightConvModel(encoder, decoder) + + +class LightConvEncoder(FairseqEncoder): + """ + LightConv encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`LightConvEncoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_tokens (torch.nn.Embedding): input embedding + """ + + def __init__(self, args, dictionary, embed_tokens): + super().__init__(dictionary) + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + + embed_dim = embed_tokens.embedding_dim + self.padding_idx = embed_tokens.padding_idx + self.max_source_positions = args.max_source_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) + self.embed_positions = PositionalEmbedding( + args.max_source_positions, embed_dim, self.padding_idx, + learned=args.encoder_learned_pos, + ) if not args.no_token_positional_embeddings else None + + self.layers = nn.ModuleList([]) + self.layers.extend([ + LightConvEncoderLayer(args, kernel_size=args.encoder_kernel_size_list[i]) + for i in range(args.encoder_layers) + ]) + self.register_buffer('version', torch.Tensor([2])) + self.normalize = args.encoder_normalize_before + if self.normalize: + self.layer_norm = LayerNorm(embed_dim) + + def forward(self, src_tokens, **unused): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + """ + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(src_tokens) + if self.embed_positions is not None: + x += self.embed_positions(src_tokens) + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # compute padding mask + encoder_padding_mask = src_tokens.eq(self.padding_idx) + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + # encoder layers + for layer in self.layers: + x = layer(x, encoder_padding_mask) + + if self.normalize: + x = self.layer_norm(x) + + return { + 'encoder_out': x, # T x B x C + 'encoder_padding_mask': encoder_padding_mask, # B x T + } + + def reorder_encoder_out(self, encoder_out, new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + if encoder_out['encoder_out'] is not None: + encoder_out['encoder_out'] = \ + encoder_out['encoder_out'].index_select(1, new_order) + if encoder_out['encoder_padding_mask'] is not None: + encoder_out['encoder_padding_mask'] = \ + encoder_out['encoder_padding_mask'].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + if self.embed_positions is None: + return self.max_source_positions + return min(self.max_source_positions, self.embed_positions.max_positions) + + +class LightConvDecoder(FairseqIncrementalDecoder): + """ + LightConv decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`LightConvDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs. + Default: ``False`` + """ + + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False, final_norm=True): + super().__init__(dictionary) + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + self.share_input_output_embed = args.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = args.decoder_embed_dim + output_embed_dim = args.decoder_output_dim + + padding_idx = embed_tokens.padding_idx + self.max_target_positions = args.max_target_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim + + self.project_in_dim = Linear(input_embed_dim, embed_dim, bias=False) if embed_dim != input_embed_dim else None + + self.embed_positions = PositionalEmbedding( + args.max_target_positions, embed_dim, padding_idx, + learned=args.decoder_learned_pos, + ) if not args.no_token_positional_embeddings else None + + self.layers = nn.ModuleList([]) + self.layers.extend([ + LightConvDecoderLayer(args, no_encoder_attn, kernel_size=args.decoder_kernel_size_list[i]) + for i in range(args.decoder_layers) + ]) + + self.adaptive_softmax = None + + self.project_out_dim = Linear(embed_dim, output_embed_dim, bias=False) \ + if embed_dim != output_embed_dim and not args.tie_adaptive_weights else None + + if args.adaptive_softmax_cutoff is not None: + self.adaptive_softmax = AdaptiveSoftmax( + len(dictionary), + output_embed_dim, + options.eval_str_list(args.adaptive_softmax_cutoff, type=int), + dropout=args.adaptive_softmax_dropout, + adaptive_inputs=embed_tokens if args.tie_adaptive_weights else None, + factor=args.adaptive_softmax_factor, + tie_proj=args.tie_adaptive_proj, + ) + elif not self.share_input_output_embed: + self.embed_out = nn.Parameter(torch.Tensor(len(dictionary), output_embed_dim)) + nn.init.normal_(self.embed_out, mean=0, std=output_embed_dim ** -0.5) + self.register_buffer('version', torch.Tensor([2])) + self.normalize = args.decoder_normalize_before and final_norm + if self.normalize: + self.layer_norm = LayerNorm(embed_dim) + + def forward(self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + + Returns: + tuple: + - the last decoder layer's output of shape `(batch, tgt_len, + vocab)` + - the last decoder layer's attention weights of shape `(batch, + tgt_len, src_len)` + """ + # embed positions + positions = self.embed_positions( + prev_output_tokens, + incremental_state=incremental_state, + ) if self.embed_positions is not None else None + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + + inner_states = [x] + + # decoder layers + for layer in self.layers: + x, attn = layer( + x, + encoder_out['encoder_out'] if encoder_out is not None else None, + encoder_out['encoder_padding_mask'] if encoder_out is not None else None, + incremental_state, + ) + inner_states.append(x) + + if self.normalize: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + if self.adaptive_softmax is None: + # project back to size of vocabulary + if self.share_input_output_embed: + x = F.linear(x, self.embed_tokens.weight) + else: + x = F.linear(x, self.embed_out) + + return x, {'attn': attn, 'inner_states': inner_states} + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if not hasattr(self, '_future_mask') or self._future_mask is None or self._future_mask.device != tensor.device: + self._future_mask = torch.triu(utils.fill_with_neg_inf(tensor.new(dim, dim)), 1) + if self._future_mask.size(0) < dim: + self._future_mask = torch.triu(utils.fill_with_neg_inf(self._future_mask.resize_(dim, dim)), 1) + return self._future_mask[:dim, :dim] + + +class LightConvEncoderLayer(nn.Module): + """Encoder layer block. + + Args: + args (argparse.Namespace): parsed command-line arguments + kernel_size: kernel size of the convolution + """ + + def __init__(self, args, kernel_size=0): + super().__init__() + self.embed_dim = args.encoder_embed_dim + self.conv_dim = args.encoder_conv_dim + padding_l = kernel_size // 2 if kernel_size % 2 == 1 else ((kernel_size - 1) // 2, kernel_size // 2) + + if args.encoder_glu: + self.linear1 = Linear(self.embed_dim, 2*self.conv_dim) + self.act = nn.GLU() + else: + self.linear1 = Linear(self.embed_dim, self.conv_dim) + self.act = None + if args.encoder_conv_type == 'lightweight': + self.conv = LightweightConv(self.conv_dim, kernel_size, padding_l=padding_l, + weight_softmax=args.weight_softmax, + num_heads=args.encoder_attention_heads, + weight_dropout=args.weight_dropout) + elif args.encoder_conv_type == 'dynamic': + self.conv = DynamicConv(self.conv_dim, kernel_size, padding_l=padding_l, + weight_softmax=args.weight_softmax, + num_heads=args.encoder_attention_heads, + weight_dropout=args.weight_dropout) + else: + raise NotImplementedError + self.linear2 = Linear(self.conv_dim, self.embed_dim) + + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + self.relu_dropout_module = FairseqDropout(args.relu_dropout, module_name=self.__class__.__name__) + self.input_dropout_module = FairseqDropout(args.input_dropout, module_name=self.__class__.__name__) + self.normalize_before = args.encoder_normalize_before + self.fc1 = Linear(self.embed_dim, args.encoder_ffn_embed_dim) + self.fc2 = Linear(args.encoder_ffn_embed_dim, self.embed_dim) + self.layer_norms = nn.ModuleList([LayerNorm(self.embed_dim) for _ in range(2)]) + + def forward(self, x, encoder_padding_mask): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, src_len)` where padding elements are indicated by ``1``. + + Returns: + encoded output of shape `(batch, src_len, embed_dim)` + """ + residual = x + x = self.maybe_layer_norm(0, x, before=True) + x = self.input_dropout_module(x) + x = self.linear1(x) + if self.act is not None: + x = self.act(x) + if encoder_padding_mask is not None: + x = x.masked_fill(encoder_padding_mask.transpose(0, 1).unsqueeze(2), 0) + x = self.conv(x) + x = self.linear2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(0, x, after=True) + + residual = x + x = self.maybe_layer_norm(1, x, before=True) + x = F.relu(self.fc1(x)) + x = self.relu_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(1, x, after=True) + return x + + def maybe_layer_norm(self, i, x, before=False, after=False): + assert before ^ after + if after ^ self.normalize_before: + return self.layer_norms[i](x) + else: + return x + + def extra_repr(self): + return 'dropout={}, relu_dropout={}, input_dropout={}, normalize_before={}'.format( + self.dropout_module.p, self.relu_dropout_module.p, self.input_dropout_module.p, self.normalize_before) + + +class LightConvDecoderLayer(nn.Module): + """Decoder layer block. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs. + Default: ``False`` + kernel_size: kernel size of the convolution + """ + + def __init__(self, args, no_encoder_attn=False, kernel_size=0): + super().__init__() + self.embed_dim = args.decoder_embed_dim + self.conv_dim = args.decoder_conv_dim + if args.decoder_glu: + self.linear1 = Linear(self.embed_dim, 2*self.conv_dim) + self.act = nn.GLU() + else: + self.linear1 = Linear(self.embed_dim, self.conv_dim) + self.act = None + if args.decoder_conv_type == 'lightweight': + self.conv = LightweightConv(self.conv_dim, kernel_size, padding_l=kernel_size-1, + weight_softmax=args.weight_softmax, + num_heads=args.decoder_attention_heads, + weight_dropout=args.weight_dropout) + elif args.decoder_conv_type == 'dynamic': + self.conv = DynamicConv(self.conv_dim, kernel_size, padding_l=kernel_size-1, + weight_softmax=args.weight_softmax, + num_heads=args.decoder_attention_heads, + weight_dropout=args.weight_dropout) + else: + raise NotImplementedError + self.linear2 = Linear(self.conv_dim, self.embed_dim) + + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + self.relu_dropout_module = FairseqDropout(args.relu_dropout, module_name=self.__class__.__name__) + self.input_dropout_module = FairseqDropout(args.input_dropout, module_name=self.__class__.__name__) + self.normalize_before = args.decoder_normalize_before + + self.conv_layer_norm = LayerNorm(self.embed_dim) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = MultiheadAttention( + self.embed_dim, args.decoder_attention_heads, + dropout=args.attention_dropout, encoder_decoder_attention=True, + ) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim) + + self.fc1 = Linear(self.embed_dim, args.decoder_ffn_embed_dim) + self.fc2 = Linear(args.decoder_ffn_embed_dim, self.embed_dim) + + self.final_layer_norm = LayerNorm(self.embed_dim) + self.need_attn = True + + def forward(self, x, encoder_out, encoder_padding_mask, incremental_state, + prev_conv_state=None, prev_attn_state=None, conv_mask=None, + conv_padding_mask=None): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, src_len)` where padding elements are indicated by ``1``. + + Returns: + encoded output of shape `(batch, src_len, embed_dim)` + """ + residual = x + x = self.maybe_layer_norm(self.conv_layer_norm, x, before=True) + if prev_conv_state is not None: + if incremental_state is None: + incremental_state = {} + self.conv._set_input_buffer(incremental_state, prev_conv_state) + x = self.input_dropout_module(x) + x = self.linear1(x) + if self.act is not None: + x = self.act(x) + x = self.conv(x, incremental_state=incremental_state) + x = self.linear2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(self.conv_layer_norm, x, after=True) + + attn = None + if self.encoder_attn is not None: + residual = x + x = self.maybe_layer_norm(self.encoder_attn_layer_norm, x, before=True) + if prev_attn_state is not None: + if incremental_state is None: + incremental_state = {} + prev_key, prev_value = prev_attn_state + saved_state = {"prev_key": prev_key, "prev_value": prev_value} + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=(not self.training and self.need_attn), + ) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(self.encoder_attn_layer_norm, x, after=True) + + residual = x + x = self.maybe_layer_norm(self.final_layer_norm, x, before=True) + x = F.relu(self.fc1(x)) + x = self.relu_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + x = self.maybe_layer_norm(self.final_layer_norm, x, after=True) + return x, attn + + def maybe_layer_norm(self, layer_norm, x, before=False, after=False): + assert before ^ after + if after ^ self.normalize_before: + return layer_norm(x) + else: + return x + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + def extra_repr(self): + return 'dropout={}, relu_dropout={}, input_dropout={}, normalize_before={}'.format( + self.dropout_module.p, self.relu_dropout_module.p, self.input_dropout_module.p, self.normalize_before) + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.) + return m + + +@register_model_architecture('lightconv', 'lightconv') +def base_architecture(args): + args.encoder_embed_path = getattr(args, 'encoder_embed_path', None) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 512) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 2048) + args.encoder_layers = getattr(args, 'encoder_layers', 7) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 8) + args.encoder_normalize_before = getattr(args, 'encoder_normalize_before', False) + args.encoder_learned_pos = getattr(args, 'encoder_learned_pos', False) + args.decoder_embed_path = getattr(args, 'decoder_embed_path', None) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', args.encoder_ffn_embed_dim) + args.decoder_layers = getattr(args, 'decoder_layers', 6) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 8) + args.decoder_normalize_before = getattr(args, 'decoder_normalize_before', False) + args.decoder_learned_pos = getattr(args, 'decoder_learned_pos', False) + args.attention_dropout = getattr(args, 'attention_dropout', 0.) + args.relu_dropout = getattr(args, 'relu_dropout', 0.) + args.dropout = getattr(args, 'dropout', 0.1) + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', None) + args.adaptive_softmax_dropout = getattr(args, 'adaptive_softmax_dropout', 0) + args.share_decoder_input_output_embed = getattr(args, 'share_decoder_input_output_embed', False) + args.share_all_embeddings = getattr(args, 'share_all_embeddings', False) + args.no_token_positional_embeddings = getattr(args, 'no_token_positional_embeddings', False) + + args.decoder_output_dim = getattr(args, 'decoder_output_dim', args.decoder_embed_dim) + args.decoder_input_dim = getattr(args, 'decoder_input_dim', args.decoder_embed_dim) + + args.encoder_conv_dim = getattr(args, 'encoder_conv_dim', args.encoder_embed_dim) + args.decoder_conv_dim = getattr(args, 'decoder_conv_dim', args.decoder_embed_dim) + + args.encoder_kernel_size_list = getattr(args, 'encoder_kernel_size_list', [3, 7, 15, 31, 31, 31, 31]) + args.decoder_kernel_size_list = getattr(args, 'decoder_kernel_size_list', [3, 7, 15, 31, 31, 31]) + if len(args.encoder_kernel_size_list) == 1: + args.encoder_kernel_size_list = args.encoder_kernel_size_list * args.encoder_layers + if len(args.decoder_kernel_size_list) == 1: + args.decoder_kernel_size_list = args.decoder_kernel_size_list * args.decoder_layers + assert len(args.encoder_kernel_size_list) == args.encoder_layers, "encoder_kernel_size_list doesn't match encoder_layers" + assert len(args.decoder_kernel_size_list) == args.decoder_layers, "decoder_kernel_size_list doesn't match decoder_layers" + args.encoder_glu = getattr(args, 'encoder_glu', True) + args.decoder_glu = getattr(args, 'decoder_glu', True) + args.input_dropout = getattr(args, 'input_dropout', 0.1) + args.weight_dropout = getattr(args, 'weight_dropout', args.attention_dropout) + + +@register_model_architecture('lightconv', 'lightconv_iwslt_de_en') +def lightconv_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 512) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 1024) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 4) + args.encoder_layers = getattr(args, 'encoder_layers', 7) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 1024) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 4) + args.decoder_layers = getattr(args, 'decoder_layers', 6) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.weight_dropout = getattr(args, 'weight_dropout', 0.1) + args.encoder_glu = getattr(args, 'encoder_glu', False) + args.decoder_glu = getattr(args, 'decoder_glu', False) + args.input_dropout = getattr(args, 'input_dropout', 0.0) + base_architecture(args) + + +@register_model_architecture('lightconv', 'lightconv_wmt_en_de') +def lightconv_wmt_en_de(args): + base_architecture(args) + + +@register_model_architecture('lightconv', 'lightconv_wmt_en_de_big') +def lightconv_wmt_en_de_big(args): + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1024) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4096) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 16) + args.encoder_normalize_before = getattr(args, 'encoder_normalize_before', False) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 1024) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 4096) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 16) + args.dropout = getattr(args, 'dropout', 0.3) + base_architecture(args) + + +@register_model_architecture('lightconv', 'lightconv_wmt_en_fr_big') +def lightconv_wmt_en_fr_big(args): + args.dropout = getattr(args, 'dropout', 0.1) + lightconv_wmt_en_de_big(args) + + +@register_model_architecture('lightconv', 'lightconv_wmt_zh_en_big') +def lightconv_wmt_zh_en_big(args): + args.dropout = getattr(args, 'dropout', 0.2) + args.attention_dropout = getattr(args, 'attention_dropout', 0.2) + args.weight_dropout = getattr(args, 'weight_dropout', 0.2) + lightconv_wmt_en_de_big(args) diff --git a/fairseq/models/lightconv_lm.py b/fairseq/models/lightconv_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..a268ddd8597cdce75613b59befb5095ff6c6eda1 --- /dev/null +++ b/fairseq/models/lightconv_lm.py @@ -0,0 +1,176 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import options +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.lightconv import ( + Embedding, + LightConvDecoder, +) +from fairseq.modules import ( + AdaptiveInput, + CharacterTokenEmbedder, +) + + +@register_model('lightconv_lm') +class LightConvLanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument('--dropout', default=0.1, type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', default=0., type=float, metavar='D', + help='dropout probability for attention weights') + parser.add_argument('--relu-dropout', default=0., type=float, metavar='D', + help='dropout probability after ReLU in FFN') + parser.add_argument('--input-dropout', type=float, metavar='D', + help='dropout probability of the inputs') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-output-dim', type=int, metavar='N', + help='decoder output dimension') + parser.add_argument('--decoder-input-dim', type=int, metavar='N', + help='decoder input dimension') + parser.add_argument('--decoder-ffn-embed-dim', type=int, metavar='N', + help='decoder embedding dimension for FFN') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='num decoder layers') + parser.add_argument('--decoder-attention-heads', type=int, metavar='N', + help='num decoder attention heads or LightConv/DynamicConv heads') + parser.add_argument('--decoder-normalize-before', default=False, action='store_true', + help='apply layernorm before each decoder block') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion') + parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D', + help='sets adaptive softmax dropout for the tail projections') + parser.add_argument('--adaptive-softmax-factor', type=float, metavar='N', + help='adaptive input factor') + parser.add_argument('--no-token-positional-embeddings', default=False, action='store_true', + help='if set, disables positional embeddings (outside self attention)') + parser.add_argument('--share-decoder-input-output-embed', default=False, action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--character-embeddings', default=False, action='store_true', + help='if set, uses character embedding convolutions to produce token embeddings') + parser.add_argument('--character-filters', type=str, metavar='LIST', + default='[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]', + help='size of character embeddings') + parser.add_argument('--character-embedding-dim', type=int, metavar='N', default=4, + help='size of character embeddings') + parser.add_argument('--char-embedder-highway-layers', type=int, metavar='N', default=2, + help='number of highway layers for character token embeddder') + parser.add_argument('--adaptive-input', default=False, action='store_true', + help='if set, uses adaptive input') + parser.add_argument('--adaptive-input-factor', type=float, metavar='N', + help='adaptive input factor') + parser.add_argument('--adaptive-input-cutoff', metavar='EXPR', + help='comma separated list of adaptive input cutoff points.') + parser.add_argument('--tie-adaptive-weights', action='store_true', + help='if set, ties the weights of adaptive softmax and adaptive input') + parser.add_argument('--tie-adaptive-proj', action='store_true', + help='if set, ties the projection weights of adaptive softmax and adaptive input') + parser.add_argument('--decoder-learned-pos', action='store_true', + help='use learned positional embeddings in the decoder') + + """LightConv and DynamicConv arguments""" + parser.add_argument('--decoder-kernel-size-list', type=lambda x: options.eval_str_list(x, int), + help='list of kernel size (default: "[3,7,15,31,31,31]")') + parser.add_argument('--decoder-glu', type=options.eval_bool, + help='glu after in proj') + parser.add_argument('--decoder-conv-type', default='dynamic', type=str, + choices=['dynamic', 'lightweight'], + help='type of convolution') + parser.add_argument('--weight-softmax', default=True, type=options.eval_bool) + parser.add_argument('--weight-dropout', type=float, metavar='D', + help='dropout probability for conv weights') + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_lm_architecture(args) + + if getattr(args, 'max_source_positions', None) is None: + args.max_source_positions = args.tokens_per_sample + if getattr(args, 'max_target_positions', None) is None: + args.max_target_positions = args.tokens_per_sample + + if args.character_embeddings: + embed_tokens = CharacterTokenEmbedder(task.dictionary, eval(args.character_filters), + args.character_embedding_dim, + args.decoder_embed_dim, + args.char_embedder_highway_layers, + ) + elif args.adaptive_input: + embed_tokens = AdaptiveInput(len(task.dictionary), task.dictionary.pad(), args.decoder_input_dim, + args.adaptive_input_factor, args.decoder_embed_dim, + options.eval_str_list(args.adaptive_input_cutoff, type=int)) + else: + embed_tokens = Embedding(len(task.dictionary), args.decoder_input_dim, task.dictionary.pad()) + + if args.tie_adaptive_weights: + assert args.adaptive_input + assert args.adaptive_input_factor == args.adaptive_softmax_factor + assert args.adaptive_softmax_cutoff == args.adaptive_input_cutoff, '{} != {}'.format( + args.adaptive_softmax_cutoff, args.adaptive_input_cutoff) + assert args.decoder_input_dim == args.decoder_output_dim + + decoder = LightConvDecoder(args, task.output_dictionary, embed_tokens, no_encoder_attn=True, final_norm=False) + return LightConvLanguageModel(decoder) + + +@register_model_architecture('lightconv_lm', 'lightconv_lm') +def base_lm_architecture(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 2048) + args.decoder_layers = getattr(args, 'decoder_layers', 6) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 8) + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', None) + args.adaptive_softmax_dropout = getattr(args, 'adaptive_softmax_dropout', 0) + args.adaptive_softmax_factor = getattr(args, 'adaptive_softmax_factor', 4) + args.decoder_learned_pos = getattr(args, 'decoder_learned_pos', False) + + args.character_embeddings = getattr(args, 'character_embeddings', False) + + args.decoder_output_dim = getattr(args, 'decoder_output_dim', args.decoder_embed_dim) + args.decoder_input_dim = getattr(args, 'decoder_input_dim', args.decoder_embed_dim) + args.decoder_conv_dim = getattr(args, 'decoder_conv_dim', args.decoder_embed_dim) + + # The model training is not stable without this + args.decoder_normalize_before = True + + args.adaptive_input = getattr(args, 'adaptive_input', False) + args.adaptive_input_factor = getattr(args, 'adaptive_input_factor', 4) + args.adaptive_input_cutoff = getattr(args, 'adaptive_input_cutoff', None) + + args.tie_adaptive_weights = getattr(args, 'tie_adaptive_weights', False) + args.tie_adaptive_proj = getattr(args, 'tie_adaptive_proj', False) + + args.decoder_kernel_size_list = getattr(args, 'decoder_kernel_size_list', [3, 7, 15, 31, 31, 31]) + if len(args.decoder_kernel_size_list) == 1: + args.decoder_kernel_size_list = args.decoder_kernel_size_list * args.decoder_layers + assert len(args.decoder_kernel_size_list) == args.decoder_layers, "decoder_kernel_size_list doesn't match decoder_layers" + args.decoder_glu = getattr(args, 'decoder_glu', True) + args.input_dropout = getattr(args, 'input_dropout', 0.1) + args.weight_dropout = getattr(args, 'weight_dropout', args.attention_dropout) + + +@register_model_architecture('lightconv_lm', 'lightconv_lm_gbw') +def lightconv_lm_gbw(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 4096) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 16) + base_lm_architecture(args) diff --git a/fairseq/models/lstm.py b/fairseq/models/lstm.py new file mode 100644 index 0000000000000000000000000000000000000000..850428a32d7d82b6f8b2f62d995dae871b2b9556 --- /dev/null +++ b/fairseq/models/lstm.py @@ -0,0 +1,681 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import options, utils +from fairseq.models import ( + FairseqEncoder, + FairseqIncrementalDecoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules import AdaptiveSoftmax, FairseqDropout +from torch import Tensor +from typing import Dict, List, Optional, Tuple + + +DEFAULT_MAX_SOURCE_POSITIONS = 1e5 +DEFAULT_MAX_TARGET_POSITIONS = 1e5 + + +@register_model('lstm') +class LSTMModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-freeze-embed', action='store_true', + help='freeze encoder embeddings') + parser.add_argument('--encoder-hidden-size', type=int, metavar='N', + help='encoder hidden size') + parser.add_argument('--encoder-layers', type=int, metavar='N', + help='number of encoder layers') + parser.add_argument('--encoder-bidirectional', action='store_true', + help='make all layers of encoder bidirectional') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-freeze-embed', action='store_true', + help='freeze decoder embeddings') + parser.add_argument('--decoder-hidden-size', type=int, metavar='N', + help='decoder hidden size') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='number of decoder layers') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='BOOL', + help='decoder attention') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion') + parser.add_argument('--share-decoder-input-output-embed', default=False, + action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--share-all-embeddings', default=False, action='store_true', + help='share encoder, decoder and output embeddings' + ' (requires shared dictionary and embed dim)') + + # Granular dropout settings (if not specified these default to --dropout) + parser.add_argument('--encoder-dropout-in', type=float, metavar='D', + help='dropout probability for encoder input embedding') + parser.add_argument('--encoder-dropout-out', type=float, metavar='D', + help='dropout probability for encoder output') + parser.add_argument('--decoder-dropout-in', type=float, metavar='D', + help='dropout probability for decoder input embedding') + parser.add_argument('--decoder-dropout-out', type=float, metavar='D', + help='dropout probability for decoder output') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure that all args are properly defaulted (in case there are any new ones) + base_architecture(args) + + if args.encoder_layers != args.decoder_layers: + raise ValueError('--encoder-layers must match --decoder-layers') + + max_source_positions = getattr(args, 'max_source_positions', DEFAULT_MAX_SOURCE_POSITIONS) + max_target_positions = getattr(args, 'max_target_positions', DEFAULT_MAX_TARGET_POSITIONS) + + def load_pretrained_embedding_from_file(embed_path, dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + embed_dict = utils.parse_embedding(embed_path) + utils.print_embed_overlap(embed_dict, dictionary) + return utils.load_embedding(embed_dict, dictionary, embed_tokens) + + if args.encoder_embed_path: + pretrained_encoder_embed = load_pretrained_embedding_from_file( + args.encoder_embed_path, task.source_dictionary, args.encoder_embed_dim) + else: + num_embeddings = len(task.source_dictionary) + pretrained_encoder_embed = Embedding( + num_embeddings, args.encoder_embed_dim, task.source_dictionary.pad() + ) + + if args.share_all_embeddings: + # double check all parameters combinations are valid + if task.source_dictionary != task.target_dictionary: + raise ValueError('--share-all-embeddings requires a joint dictionary') + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path): + raise ValueError( + '--share-all-embed not compatible with --decoder-embed-path' + ) + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + '--share-all-embeddings requires --encoder-embed-dim to ' + 'match --decoder-embed-dim' + ) + pretrained_decoder_embed = pretrained_encoder_embed + args.share_decoder_input_output_embed = True + else: + # separate decoder input embeddings + pretrained_decoder_embed = None + if args.decoder_embed_path: + pretrained_decoder_embed = load_pretrained_embedding_from_file( + args.decoder_embed_path, + task.target_dictionary, + args.decoder_embed_dim + ) + # one last double check of parameter combinations + if args.share_decoder_input_output_embed and ( + args.decoder_embed_dim != args.decoder_out_embed_dim): + raise ValueError( + '--share-decoder-input-output-embeddings requires ' + '--decoder-embed-dim to match --decoder-out-embed-dim' + ) + + if args.encoder_freeze_embed: + pretrained_encoder_embed.weight.requires_grad = False + if args.decoder_freeze_embed: + pretrained_decoder_embed.weight.requires_grad = False + + encoder = LSTMEncoder( + dictionary=task.source_dictionary, + embed_dim=args.encoder_embed_dim, + hidden_size=args.encoder_hidden_size, + num_layers=args.encoder_layers, + dropout_in=args.encoder_dropout_in, + dropout_out=args.encoder_dropout_out, + bidirectional=args.encoder_bidirectional, + pretrained_embed=pretrained_encoder_embed, + max_source_positions=max_source_positions, + ) + decoder = LSTMDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + hidden_size=args.decoder_hidden_size, + out_embed_dim=args.decoder_out_embed_dim, + num_layers=args.decoder_layers, + dropout_in=args.decoder_dropout_in, + dropout_out=args.decoder_dropout_out, + attention=options.eval_bool(args.decoder_attention), + encoder_output_units=encoder.output_units, + pretrained_embed=pretrained_decoder_embed, + share_input_output_embed=args.share_decoder_input_output_embed, + adaptive_softmax_cutoff=( + options.eval_str_list(args.adaptive_softmax_cutoff, type=int) + if args.criterion == 'adaptive_loss' else None + ), + max_target_positions=max_target_positions, + residuals=False, + ) + return cls(encoder, decoder) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths) + decoder_out = self.decoder( + prev_output_tokens, encoder_out=encoder_out, incremental_state=incremental_state + ) + return decoder_out + + +class LSTMEncoder(FairseqEncoder): + """LSTM encoder.""" + def __init__( + self, dictionary, embed_dim=512, hidden_size=512, num_layers=1, + dropout_in=0.1, dropout_out=0.1, bidirectional=False, + left_pad=True, pretrained_embed=None, padding_idx=None, + max_source_positions=DEFAULT_MAX_SOURCE_POSITIONS, + ): + super().__init__(dictionary) + self.num_layers = num_layers + self.dropout_in_module = FairseqDropout(dropout_in, module_name=self.__class__.__name__) + self.dropout_out_module = FairseqDropout(dropout_out, module_name=self.__class__.__name__) + self.bidirectional = bidirectional + self.hidden_size = hidden_size + self.max_source_positions = max_source_positions + + num_embeddings = len(dictionary) + self.padding_idx = padding_idx if padding_idx is not None else dictionary.pad() + if pretrained_embed is None: + self.embed_tokens = Embedding(num_embeddings, embed_dim, self.padding_idx) + else: + self.embed_tokens = pretrained_embed + + self.lstm = LSTM( + input_size=embed_dim, + hidden_size=hidden_size, + num_layers=num_layers, + dropout=self.dropout_out_module.p if num_layers > 1 else 0., + bidirectional=bidirectional, + ) + self.left_pad = left_pad + + self.output_units = hidden_size + if bidirectional: + self.output_units *= 2 + + def forward( + self, + src_tokens: Tensor, + src_lengths: Tensor, + enforce_sorted: bool = True, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of + shape `(batch, src_len)` + src_lengths (LongTensor): lengths of each source sentence of + shape `(batch)` + enforce_sorted (bool, optional): if True, `src_tokens` is + expected to contain sequences sorted by length in a + decreasing order. If False, this condition is not + required. Default: True. + """ + if self.left_pad: + # nn.utils.rnn.pack_padded_sequence requires right-padding; + # convert left-padding to right-padding + src_tokens = utils.convert_padding_direction( + src_tokens, + torch.zeros_like(src_tokens).fill_(self.padding_idx), + left_to_right=True, + ) + + bsz, seqlen = src_tokens.size() + + # embed tokens + x = self.embed_tokens(src_tokens) + x = self.dropout_in_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # pack embedded source tokens into a PackedSequence + packed_x = nn.utils.rnn.pack_padded_sequence( + x, src_lengths.data, enforce_sorted=enforce_sorted + ) + + # apply LSTM + if self.bidirectional: + state_size = 2 * self.num_layers, bsz, self.hidden_size + else: + state_size = self.num_layers, bsz, self.hidden_size + h0 = x.new_zeros(*state_size) + c0 = x.new_zeros(*state_size) + packed_outs, (final_hiddens, final_cells) = self.lstm(packed_x, (h0, c0)) + + # unpack outputs and apply dropout + x, _ = nn.utils.rnn.pad_packed_sequence(packed_outs, padding_value=self.padding_idx*1.0) + x = self.dropout_out_module(x) + assert list(x.size()) == [seqlen, bsz, self.output_units] + + if self.bidirectional: + final_hiddens = self.combine_bidir(final_hiddens, bsz) + final_cells = self.combine_bidir(final_cells, bsz) + + encoder_padding_mask = src_tokens.eq(self.padding_idx).t() + + return tuple(( + x, # seq_len x batch x hidden + final_hiddens, # num_layers x batch x num_directions*hidden + final_cells, # num_layers x batch x num_directions*hidden + encoder_padding_mask, # seq_len x batch + )) + + def combine_bidir(self, outs, bsz: int): + out = outs.view(self.num_layers, 2, bsz, -1).transpose(1, 2).contiguous() + return out.view(self.num_layers, bsz, -1) + + def reorder_encoder_out(self, encoder_out, new_order): + return tuple(( + encoder_out[0].index_select(1, new_order), + encoder_out[1].index_select(1, new_order), + encoder_out[2].index_select(1, new_order), + encoder_out[3].index_select(1, new_order), + )) + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return self.max_source_positions + + +class AttentionLayer(nn.Module): + def __init__(self, input_embed_dim, source_embed_dim, output_embed_dim, bias=False): + super().__init__() + + self.input_proj = Linear(input_embed_dim, source_embed_dim, bias=bias) + self.output_proj = Linear(input_embed_dim + source_embed_dim, output_embed_dim, bias=bias) + + def forward(self, input, source_hids, encoder_padding_mask): + # input: bsz x input_embed_dim + # source_hids: srclen x bsz x source_embed_dim + + # x: bsz x source_embed_dim + x = self.input_proj(input) + + # compute attention + attn_scores = (source_hids * x.unsqueeze(0)).sum(dim=2) + + # don't attend over padding + if encoder_padding_mask is not None: + attn_scores = attn_scores.float().masked_fill_( + encoder_padding_mask, + float('-inf') + ).type_as(attn_scores) # FP16 support: cast to float and back + + attn_scores = F.softmax(attn_scores, dim=0) # srclen x bsz + + # sum weighted sources + x = (attn_scores.unsqueeze(2) * source_hids).sum(dim=0) + + x = torch.tanh(self.output_proj(torch.cat((x, input), dim=1))) + return x, attn_scores + + +class LSTMDecoder(FairseqIncrementalDecoder): + """LSTM decoder.""" + def __init__( + self, dictionary, embed_dim=512, hidden_size=512, out_embed_dim=512, + num_layers=1, dropout_in=0.1, dropout_out=0.1, attention=True, + encoder_output_units=512, pretrained_embed=None, + share_input_output_embed=False, adaptive_softmax_cutoff=None, + max_target_positions=DEFAULT_MAX_TARGET_POSITIONS, + residuals=False, + ): + super().__init__(dictionary) + self.dropout_in_module = FairseqDropout(dropout_in, module_name=self.__class__.__name__) + self.dropout_out_module = FairseqDropout(dropout_out, module_name=self.__class__.__name__) + self.hidden_size = hidden_size + self.share_input_output_embed = share_input_output_embed + self.need_attn = True + self.max_target_positions = max_target_positions + self.residuals = residuals + self.num_layers = num_layers + + self.adaptive_softmax = None + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + if pretrained_embed is None: + self.embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + else: + self.embed_tokens = pretrained_embed + + self.encoder_output_units = encoder_output_units + if encoder_output_units != hidden_size and encoder_output_units != 0: + self.encoder_hidden_proj = Linear(encoder_output_units, hidden_size) + self.encoder_cell_proj = Linear(encoder_output_units, hidden_size) + else: + self.encoder_hidden_proj = self.encoder_cell_proj = None + + # disable input feeding if there is no encoder + # input feeding is described in arxiv.org/abs/1508.04025 + input_feed_size = 0 if encoder_output_units == 0 else hidden_size + self.layers = nn.ModuleList([ + LSTMCell( + input_size=input_feed_size + embed_dim if layer == 0 else hidden_size, + hidden_size=hidden_size, + ) + for layer in range(num_layers) + ]) + + if attention: + # TODO make bias configurable + self.attention = AttentionLayer(hidden_size, encoder_output_units, hidden_size, bias=False) + else: + self.attention = None + + if hidden_size != out_embed_dim: + self.additional_fc = Linear(hidden_size, out_embed_dim) + + if adaptive_softmax_cutoff is not None: + # setting adaptive_softmax dropout to dropout_out for now but can be redefined + self.adaptive_softmax = AdaptiveSoftmax( + num_embeddings, hidden_size, adaptive_softmax_cutoff, dropout=dropout_out, + ) + elif not self.share_input_output_embed: + self.fc_out = Linear(out_embed_dim, num_embeddings, dropout=dropout_out) + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Tuple[Tensor, Tensor, Tensor, Tensor]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + src_lengths: Optional[Tensor] = None, + ): + x, attn_scores = self.extract_features( + prev_output_tokens, encoder_out, incremental_state + ) + return self.output_layer(x), attn_scores + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Tuple[Tensor, Tensor, Tensor, Tensor]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + """ + Similar to *forward* but only return features. + """ + # get outputs from encoder + if encoder_out is not None: + encoder_outs = encoder_out[0] + encoder_hiddens = encoder_out[1] + encoder_cells = encoder_out[2] + encoder_padding_mask = encoder_out[3] + else: + encoder_outs = torch.empty(0) + encoder_hiddens = torch.empty(0) + encoder_cells = torch.empty(0) + encoder_padding_mask = torch.empty(0) + srclen = encoder_outs.size(0) + + if incremental_state is not None and len(incremental_state) > 0: + prev_output_tokens = prev_output_tokens[:, -1:] + + bsz, seqlen = prev_output_tokens.size() + + # embed tokens + x = self.embed_tokens(prev_output_tokens) + x = self.dropout_in_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # initialize previous states (or get from cache during incremental generation) + if incremental_state is not None and len(incremental_state) > 0: + prev_hiddens, prev_cells, input_feed = self.get_cached_state(incremental_state) + elif encoder_out is not None: + # setup recurrent cells + prev_hiddens = [encoder_hiddens[i] for i in range(self.num_layers)] + prev_cells = [encoder_cells[i] for i in range(self.num_layers)] + if self.encoder_hidden_proj is not None: + prev_hiddens = [self.encoder_hidden_proj(y) for y in prev_hiddens] + prev_cells = [self.encoder_cell_proj(y) for y in prev_cells] + input_feed = x.new_zeros(bsz, self.hidden_size) + else: + # setup zero cells, since there is no encoder + zero_state = x.new_zeros(bsz, self.hidden_size) + prev_hiddens = [zero_state for i in range(self.num_layers)] + prev_cells = [zero_state for i in range(self.num_layers)] + input_feed = None + + assert srclen > 0 or self.attention is None, \ + "attention is not supported if there are no encoder outputs" + attn_scores = x.new_zeros(srclen, seqlen, bsz) if self.attention is not None else None + outs = [] + for j in range(seqlen): + # input feeding: concatenate context vector from previous time step + if input_feed is not None: + input = torch.cat((x[j, :, :], input_feed), dim=1) + else: + input = x[j] + + for i, rnn in enumerate(self.layers): + # recurrent cell + hidden, cell = rnn(input, (prev_hiddens[i], prev_cells[i])) + + # hidden state becomes the input to the next layer + input = self.dropout_out_module(hidden) + if self.residuals: + input = input + prev_hiddens[i] + + # save state for next time step + prev_hiddens[i] = hidden + prev_cells[i] = cell + + # apply attention using the last layer's hidden state + if self.attention is not None: + assert attn_scores is not None + out, attn_scores[:, j, :] = self.attention(hidden, encoder_outs, encoder_padding_mask) + else: + out = hidden + out = self.dropout_out_module(out) + + # input feeding + if input_feed is not None: + input_feed = out + + # save final output + outs.append(out) + + # Stack all the necessary tensors together and store + prev_hiddens_tensor = torch.stack(prev_hiddens) + prev_cells_tensor = torch.stack(prev_cells) + cache_state = torch.jit.annotate( + Dict[str, Optional[Tensor]], + { + "prev_hiddens": prev_hiddens_tensor, + "prev_cells": prev_cells_tensor, + "input_feed": input_feed, + } + ) + self.set_incremental_state(incremental_state, 'cached_state', cache_state) + + # collect outputs across time steps + x = torch.cat(outs, dim=0).view(seqlen, bsz, self.hidden_size) + + # T x B x C -> B x T x C + x = x.transpose(1, 0) + + if hasattr(self, 'additional_fc') and self.adaptive_softmax is None: + x = self.additional_fc(x) + x = self.dropout_out_module(x) + # srclen x tgtlen x bsz -> bsz x tgtlen x srclen + if not self.training and self.need_attn and self.attention is not None: + assert attn_scores is not None + attn_scores = attn_scores.transpose(0, 2) + else: + attn_scores = None + return x, attn_scores + + def output_layer(self, x): + """Project features to the vocabulary size.""" + if self.adaptive_softmax is None: + if self.share_input_output_embed: + x = F.linear(x, self.embed_tokens.weight) + else: + x = self.fc_out(x) + return x + + def get_cached_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + ) -> Tuple[List[Tensor], List[Tensor], Optional[Tensor]]: + cached_state = self.get_incremental_state(incremental_state, 'cached_state') + assert cached_state is not None + prev_hiddens_ = cached_state["prev_hiddens"] + assert prev_hiddens_ is not None + prev_cells_ = cached_state["prev_cells"] + assert prev_cells_ is not None + prev_hiddens = [prev_hiddens_[i] for i in range(self.num_layers)] + prev_cells = [prev_cells_[j] for j in range(self.num_layers)] + input_feed = cached_state["input_feed"] # can be None for decoder-only language models + return prev_hiddens, prev_cells, input_feed + + def reorder_incremental_state( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + new_order: Tensor, + ): + if incremental_state is None or len(incremental_state) == 0: + return + prev_hiddens, prev_cells, input_feed = self.get_cached_state(incremental_state) + prev_hiddens = [p.index_select(0, new_order) for p in prev_hiddens] + prev_cells = [p.index_select(0, new_order) for p in prev_cells] + if input_feed is not None: + input_feed = input_feed.index_select(0, new_order) + cached_state_new = torch.jit.annotate( + Dict[str, Optional[Tensor]], + { + "prev_hiddens": torch.stack(prev_hiddens), + "prev_cells": torch.stack(prev_cells), + "input_feed": input_feed, + } + ) + self.set_incremental_state(incremental_state, 'cached_state', cached_state_new), + return + + def max_positions(self): + """Maximum output length supported by the decoder.""" + return self.max_target_positions + + def make_generation_fast_(self, need_attn=False, **kwargs): + self.need_attn = need_attn + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.uniform_(m.weight, -0.1, 0.1) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def LSTM(input_size, hidden_size, **kwargs): + m = nn.LSTM(input_size, hidden_size, **kwargs) + for name, param in m.named_parameters(): + if 'weight' in name or 'bias' in name: + param.data.uniform_(-0.1, 0.1) + return m + + +def LSTMCell(input_size, hidden_size, **kwargs): + m = nn.LSTMCell(input_size, hidden_size, **kwargs) + for name, param in m.named_parameters(): + if 'weight' in name or 'bias' in name: + param.data.uniform_(-0.1, 0.1) + return m + + +def Linear(in_features, out_features, bias=True, dropout=0.): + """Linear layer (input: N x T x C)""" + m = nn.Linear(in_features, out_features, bias=bias) + m.weight.data.uniform_(-0.1, 0.1) + if bias: + m.bias.data.uniform_(-0.1, 0.1) + return m + + +@register_model_architecture('lstm', 'lstm') +def base_architecture(args): + args.dropout = getattr(args, 'dropout', 0.1) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 512) + args.encoder_embed_path = getattr(args, 'encoder_embed_path', None) + args.encoder_freeze_embed = getattr(args, 'encoder_freeze_embed', False) + args.encoder_hidden_size = getattr(args, 'encoder_hidden_size', args.encoder_embed_dim) + args.encoder_layers = getattr(args, 'encoder_layers', 1) + args.encoder_bidirectional = getattr(args, 'encoder_bidirectional', False) + args.encoder_dropout_in = getattr(args, 'encoder_dropout_in', args.dropout) + args.encoder_dropout_out = getattr(args, 'encoder_dropout_out', args.dropout) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_embed_path = getattr(args, 'decoder_embed_path', None) + args.decoder_freeze_embed = getattr(args, 'decoder_freeze_embed', False) + args.decoder_hidden_size = getattr(args, 'decoder_hidden_size', args.decoder_embed_dim) + args.decoder_layers = getattr(args, 'decoder_layers', 1) + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 512) + args.decoder_attention = getattr(args, 'decoder_attention', '1') + args.decoder_dropout_in = getattr(args, 'decoder_dropout_in', args.dropout) + args.decoder_dropout_out = getattr(args, 'decoder_dropout_out', args.dropout) + args.share_decoder_input_output_embed = getattr(args, 'share_decoder_input_output_embed', False) + args.share_all_embeddings = getattr(args, 'share_all_embeddings', False) + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', '10000,50000,200000') + + +@register_model_architecture('lstm', 'lstm_wiseman_iwslt_de_en') +def lstm_wiseman_iwslt_de_en(args): + args.dropout = getattr(args, 'dropout', 0.1) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 256) + args.encoder_dropout_in = getattr(args, 'encoder_dropout_in', 0) + args.encoder_dropout_out = getattr(args, 'encoder_dropout_out', 0) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 256) + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 256) + args.decoder_dropout_in = getattr(args, 'decoder_dropout_in', 0) + args.decoder_dropout_out = getattr(args, 'decoder_dropout_out', args.dropout) + base_architecture(args) + + +@register_model_architecture('lstm', 'lstm_luong_wmt_en_de') +def lstm_luong_wmt_en_de(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1000) + args.encoder_layers = getattr(args, 'encoder_layers', 4) + args.encoder_dropout_out = getattr(args, 'encoder_dropout_out', 0) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 1000) + args.decoder_layers = getattr(args, 'decoder_layers', 4) + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 1000) + args.decoder_dropout_out = getattr(args, 'decoder_dropout_out', 0) + base_architecture(args) diff --git a/fairseq/models/lstm_lm.py b/fairseq/models/lstm_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..9f6758a4bc876ff7c2da3fff0c305ab7ee13a1e9 --- /dev/null +++ b/fairseq/models/lstm_lm.py @@ -0,0 +1,130 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import options, utils +from fairseq.models import ( + FairseqLanguageModel, register_model, register_model_architecture +) +from fairseq.models.lstm import ( + LSTMDecoder, Embedding +) + +DEFAULT_MAX_TARGET_POSITIONS = 1e5 + +@register_model('lstm_lm') +class LSTMLanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-hidden-size', type=int, metavar='N', + help='decoder hidden size') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='number of decoder layers') + parser.add_argument('--decoder-out-embed-dim', type=int, metavar='N', + help='decoder output embedding dimension') + parser.add_argument('--decoder-attention', type=str, metavar='BOOL', + help='decoder attention') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion') + parser.add_argument('--residuals', default=False, + action='store_true', + help='applying residuals between LSTM layers') + + # Granular dropout settings (if not specified these default to --dropout) + parser.add_argument('--decoder-dropout-in', type=float, metavar='D', + help='dropout probability for decoder input embedding') + parser.add_argument('--decoder-dropout-out', type=float, metavar='D', + help='dropout probability for decoder output') + parser.add_argument('--share-decoder-input-output-embed', default=False, + action='store_true', + help='share decoder input and output embeddings') + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if getattr(args, 'max_target_positions', None) is not None: + max_target_positions = args.max_target_positions + else: + max_target_positions = getattr(args, 'tokens_per_sample', DEFAULT_MAX_TARGET_POSITIONS) + + def load_pretrained_embedding_from_file(embed_path, dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + embed_tokens = Embedding(num_embeddings, embed_dim, padding_idx) + embed_dict = utils.parse_embedding(embed_path) + utils.print_embed_overlap(embed_dict, dictionary) + return utils.load_embedding(embed_dict, dictionary, embed_tokens) + + pretrained_decoder_embed = None + if args.decoder_embed_path: + pretrained_decoder_embed = load_pretrained_embedding_from_file( + args.decoder_embed_path, + task.target_dictionary, + args.decoder_embed_dim + ) + + if args.share_decoder_input_output_embed: + # double check all parameters combinations are valid + if task.source_dictionary != task.target_dictionary: + raise ValueError('--share-decoder-input-output-embeddings requires a joint dictionary') + + if args.decoder_embed_dim != args.decoder_out_embed_dim: + raise ValueError( + '--share-decoder-input-output-embeddings requires ' + '--decoder-embed-dim to match --decoder-out-embed-dim' + ) + + decoder = LSTMDecoder( + dictionary=task.dictionary, + embed_dim=args.decoder_embed_dim, + hidden_size=args.decoder_hidden_size, + out_embed_dim=args.decoder_out_embed_dim, + num_layers=args.decoder_layers, + dropout_in=args.decoder_dropout_in, + dropout_out=args.decoder_dropout_out, + attention=False, # decoder-only language model doesn't support attention + encoder_output_units=0, + pretrained_embed=pretrained_decoder_embed, + share_input_output_embed=args.share_decoder_input_output_embed, + adaptive_softmax_cutoff=( + options.eval_str_list(args.adaptive_softmax_cutoff, type=int) + if args.criterion == 'adaptive_loss' else None + ), + max_target_positions=max_target_positions, + residuals=args.residuals + ) + + return cls(decoder) + + +@register_model_architecture('lstm_lm', 'lstm_lm') +def base_architecture(args): + args.dropout = getattr(args, 'dropout', 0.1) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_embed_path = getattr(args, 'decoder_embed_path', None) + args.decoder_hidden_size = getattr(args, 'decoder_hidden_size', args.decoder_embed_dim) + args.decoder_layers = getattr(args, 'decoder_layers', 1) + args.decoder_out_embed_dim = getattr(args, 'decoder_out_embed_dim', 512) + args.decoder_attention = getattr(args, 'decoder_attention', '0') + args.decoder_dropout_in = getattr(args, 'decoder_dropout_in', args.dropout) + args.decoder_dropout_out = getattr(args, 'decoder_dropout_out', args.dropout) + args.share_decoder_input_output_embed = getattr(args, 'share_decoder_input_output_embed', False) + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', '10000,50000,200000') + args.residuals = getattr(args, 'residuals', False) diff --git a/fairseq/models/masked_lm.py b/fairseq/models/masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..35a6323ef214786c3391eaa9647cd7a52c4296e4 --- /dev/null +++ b/fairseq/models/masked_lm.py @@ -0,0 +1,352 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from fairseq.models import ( + FairseqEncoderModel, + FairseqEncoder, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + LayerNorm, + SinusoidalPositionalEmbedding, + TransformerSentenceEncoder, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +logger = logging.getLogger(__name__) + + +@register_model('masked_lm') +class MaskedLMModel(FairseqEncoderModel): + """ + Class for training a Masked Language Model. It also supports an + additional sentence level prediction if the sent-loss argument is set. + """ + def __init__(self, args, encoder): + super().__init__(encoder) + self.args = args + + # if specified then apply bert initialization on the model. We need + # to explictly call this to make sure that the output embeddings + # and projection layers are also correctly initialized + if getattr(args, 'apply_bert_init', False): + self.apply(init_bert_params) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # Arguments related to dropout + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', type=float, + metavar='D', help='dropout probability for' + ' attention weights') + parser.add_argument('--act-dropout', type=float, + metavar='D', help='dropout probability after' + ' activation in FFN') + + # Arguments related to hidden states and self-attention + parser.add_argument('--encoder-ffn-embed-dim', type=int, metavar='N', + help='encoder embedding dimension for FFN') + parser.add_argument('--encoder-layers', type=int, metavar='N', + help='num encoder layers') + parser.add_argument('--encoder-attention-heads', type=int, metavar='N', + help='num encoder attention heads') + + # Arguments related to input and output embeddings + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--share-encoder-input-output-embed', + action='store_true', help='share encoder input' + ' and output embeddings') + parser.add_argument('--encoder-learned-pos', action='store_true', + help='use learned positional embeddings in the encoder') + parser.add_argument('--no-token-positional-embeddings', + action='store_true', + help='if set, disables positional embeddings' + ' (outside self attention)') + parser.add_argument('--num-segment', type=int, metavar='N', + help='num segment in the input') + parser.add_argument('--max-positions', type=int, + help='number of positional embeddings to learn') + + # Arguments related to sentence level prediction + parser.add_argument('--sentence-class-num', type=int, metavar='N', + help='number of classes for sentence task') + parser.add_argument('--sent-loss', action='store_true', help='if set,' + ' calculate sentence level predictions') + + # Arguments related to parameter initialization + parser.add_argument('--apply-bert-init', action='store_true', + help='use custom param initialization for BERT') + + # misc params + parser.add_argument('--activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use') + parser.add_argument('--pooler-activation-fn', + choices=utils.get_available_activation_fns(), + help='Which activation function to use for pooler layer.') + parser.add_argument('--encoder-normalize-before', action='store_true', + help='apply layernorm before each encoder block') + + def forward(self, src_tokens, segment_labels=None, **kwargs): + return self.encoder(src_tokens, segment_labels=segment_labels, **kwargs) + + def max_positions(self): + return self.encoder.max_positions + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure all arguments are present in older models + base_architecture(args) + + if not hasattr(args, 'max_positions'): + args.max_positions = args.tokens_per_sample + + logger.info(args) + + encoder = MaskedLMEncoder(args, task.dictionary) + return cls(args, encoder) + + +class MaskedLMEncoder(FairseqEncoder): + """ + Encoder for Masked Language Modelling. + """ + + def __init__(self, args, dictionary): + super().__init__(dictionary) + + self.padding_idx = dictionary.pad() + self.vocab_size = dictionary.__len__() + self.max_positions = args.max_positions + + self.sentence_encoder = TransformerSentenceEncoder( + padding_idx=self.padding_idx, + vocab_size=self.vocab_size, + num_encoder_layers=args.encoder_layers, + embedding_dim=args.encoder_embed_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=args.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.act_dropout, + max_seq_len=self.max_positions, + num_segments=args.num_segment, + use_position_embeddings=not args.no_token_positional_embeddings, + encoder_normalize_before=args.encoder_normalize_before, + apply_bert_init=args.apply_bert_init, + activation_fn=args.activation_fn, + learned_pos_embedding=args.encoder_learned_pos, + ) + + self.share_input_output_embed = args.share_encoder_input_output_embed + self.embed_out = None + self.sentence_projection_layer = None + self.sentence_out_dim = args.sentence_class_num + self.lm_output_learned_bias = None + + # Remove head is set to true during fine-tuning + self.load_softmax = not getattr(args, 'remove_head', False) + + self.masked_lm_pooler = nn.Linear( + args.encoder_embed_dim, args.encoder_embed_dim + ) + self.pooler_activation = utils.get_activation_fn(args.pooler_activation_fn) + + self.lm_head_transform_weight = nn.Linear(args.encoder_embed_dim, args.encoder_embed_dim) + self.activation_fn = utils.get_activation_fn(args.activation_fn) + self.layer_norm = LayerNorm(args.encoder_embed_dim) + + self.lm_output_learned_bias = None + if self.load_softmax: + self.lm_output_learned_bias = nn.Parameter(torch.zeros(self.vocab_size)) + + if not self.share_input_output_embed: + self.embed_out = nn.Linear( + args.encoder_embed_dim, + self.vocab_size, + bias=False + ) + + if args.sent_loss: + self.sentence_projection_layer = nn.Linear( + args.encoder_embed_dim, + self.sentence_out_dim, + bias=False + ) + + def forward(self, src_tokens, segment_labels=None, masked_tokens=None, **unused): + """ + Forward pass for Masked LM encoder. This first computes the token + embedding using the token embedding matrix, position embeddings (if + specified) and segment embeddings (if specified). + + Here we assume that the sentence representation corresponds to the + output of the classification_token (see bert_task or cross_lingual_lm + task for more details). + Args: + - src_tokens: B x T matrix representing sentences + - segment_labels: B x T matrix representing segment label for tokens + Returns: + - a tuple of the following: + - logits for predictions in format B x T x C to be used in + softmax afterwards + - a dictionary of additional data, where 'pooled_output' contains + the representation for classification_token and 'inner_states' + is a list of internal model states used to compute the + predictions (similar in ELMO). 'sentence_logits' + is the prediction logit for NSP task and is only computed if + this is specified in the input arguments. + """ + + inner_states, sentence_rep = self.sentence_encoder( + src_tokens, + segment_labels=segment_labels, + ) + + x = inner_states[-1].transpose(0, 1) + # project masked tokens only + if masked_tokens is not None: + x = x[masked_tokens, :] + x = self.layer_norm(self.activation_fn(self.lm_head_transform_weight(x))) + + pooled_output = self.pooler_activation(self.masked_lm_pooler(sentence_rep)) + + # project back to size of vocabulary + if self.share_input_output_embed \ + and hasattr(self.sentence_encoder.embed_tokens, 'weight'): + x = F.linear(x, self.sentence_encoder.embed_tokens.weight) + elif self.embed_out is not None: + x = self.embed_out(x) + if self.lm_output_learned_bias is not None: + x = x + self.lm_output_learned_bias + sentence_logits = None + if self.sentence_projection_layer: + sentence_logits = self.sentence_projection_layer(pooled_output) + + return x, { + 'inner_states': inner_states, + 'pooled_output': pooled_output, + 'sentence_logits': sentence_logits + } + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.max_positions + + def upgrade_state_dict_named(self, state_dict, name): + if isinstance( + self.sentence_encoder.embed_positions, + SinusoidalPositionalEmbedding + ): + state_dict[ + name + '.sentence_encoder.embed_positions._float_tensor' + ] = torch.FloatTensor(1) + if not self.load_softmax: + for k in list(state_dict.keys()): + if ( + "embed_out.weight" in k or + "sentence_projection_layer.weight" in k or + "lm_output_learned_bias" in k + ): + del state_dict[k] + return state_dict + + +@register_model_architecture('masked_lm', 'masked_lm') +def base_architecture(args): + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.act_dropout = getattr(args, 'act_dropout', 0.0) + + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4096) + args.encoder_layers = getattr(args, 'encoder_layers', 6) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 8) + + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1024) + args.share_encoder_input_output_embed = getattr(args, 'share_encoder_input_output_embed', False) + args.encoder_learned_pos = getattr(args, 'encoder_learned_pos', False) + args.no_token_positional_embeddings = getattr(args, 'no_token_positional_embeddings', False) + args.num_segment = getattr(args, 'num_segment', 2) + + args.sentence_class_num = getattr(args, 'sentence_class_num', 2) + args.sent_loss = getattr(args, 'sent_loss', False) + + args.apply_bert_init = getattr(args, 'apply_bert_init', False) + + args.activation_fn = getattr(args, 'activation_fn', 'relu') + args.pooler_activation_fn = getattr(args, 'pooler_activation_fn', 'tanh') + args.encoder_normalize_before = getattr(args, 'encoder_normalize_before', False) + + +@register_model_architecture('masked_lm', 'bert_base') +def bert_base_architecture(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 768) + args.share_encoder_input_output_embed = getattr( + args, 'share_encoder_input_output_embed', True) + args.no_token_positional_embeddings = getattr( + args, 'no_token_positional_embeddings', False) + args.encoder_learned_pos = getattr(args, 'encoder_learned_pos', True) + args.num_segment = getattr(args, 'num_segment', 2) + + args.encoder_layers = getattr(args, 'encoder_layers', 12) + + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 12) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 3072) + + args.sentence_class_num = getattr(args, 'sentence_class_num', 2) + args.sent_loss = getattr(args, 'sent_loss', True) + + args.apply_bert_init = getattr(args, 'apply_bert_init', True) + + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + args.pooler_activation_fn = getattr(args, 'pooler_activation_fn', 'tanh') + args.encoder_normalize_before = getattr(args, 'encoder_normalize_before', True) + base_architecture(args) + + +@register_model_architecture('masked_lm', 'bert_large') +def bert_large_architecture(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1024) + args.encoder_layers = getattr(args, 'encoder_layers', 24) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 16) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4096) + bert_base_architecture(args) + + +@register_model_architecture('masked_lm', 'xlm_base') +def xlm_architecture(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1024) + args.share_encoder_input_output_embed = getattr( + args, 'share_encoder_input_output_embed', True) + args.no_token_positional_embeddings = getattr( + args, 'no_token_positional_embeddings', False) + args.encoder_learned_pos = getattr(args, 'encoder_learned_pos', True) + args.num_segment = getattr(args, 'num_segment', 1) + + args.encoder_layers = getattr(args, 'encoder_layers', 6) + + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 8) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4096) + + args.sent_loss = getattr(args, 'sent_loss', False) + + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + args.encoder_normalize_before = getattr(args, 'encoder_normalize_before', False) + args.pooler_activation_fn = getattr(args, 'pooler_activation_fn', 'tanh') + args.apply_bert_init = getattr(args, 'apply_bert_init', True) + base_architecture(args) diff --git a/fairseq/models/model_utils.py b/fairseq/models/model_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..46ec62f77243308f82d862a52fd1a9b2615f9e90 --- /dev/null +++ b/fairseq/models/model_utils.py @@ -0,0 +1,90 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import List, Optional + +import torch +from torch import Tensor + + +@torch.jit.script +def script_skip_tensor_list(x: List[Tensor], mask): + res = [xi[mask] if xi.size(0) == mask.size(0) else xi[:, mask] for xi in x] + outputs = [] + for i, t in enumerate(res): + if t.numel() != 0: + outputs.append(t) + else: + outputs.append(x[i]) + return outputs + + +@torch.jit.script +def script_skip_tensor(x: Tensor, mask): + # None case + if x.size(0) == 0: + return x + res = x[mask] if x.size(0) == mask.size(0) else x[:, mask] + if res.numel() == 0: + return x + else: + return res + + +@torch.jit.script +def expand_2d_or_3d_tensor(x, trg_dim: int, padding_idx: int): + """ + Expand 2D/3D tensor on dim=1 + """ + if x is None: + return None + + assert x.dim() == 2 or x.dim() == 3 + assert trg_dim >= x.size(1), (trg_dim, x.size()) + if trg_dim == x.size(1): + return x + + dims = [x.size(0), trg_dim - x.size(1)] + if x.dim() == 3: + dims.append(x.size(2)) + x = torch.cat([x, torch.zeros(dims).to(x).fill_(padding_idx)], 1) + + return x + + +@torch.jit.script +def coalesce(x: Optional[Tensor], y: Tensor) -> Tensor: + return x if x is not None else y + + +@torch.jit.script +def fill_tensors(x: Optional[Tensor], mask, y: Optional[Tensor], padding_idx: int) -> Optional[Tensor]: + """ + Filling tensor x with y at masked positions (dim=0). + """ + if x is None or x.size()[0] == 0 or y is None: + return x + assert x.dim() == y.dim() and mask.size(0) == x.size(0) + assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2)) + + n_selected = mask.sum() + if n_selected == 0: + return x + assert n_selected == y.size(0) + if n_selected == x.size(0): + return y + + if x.size(1) < y.size(1): + x = expand_2d_or_3d_tensor(x, y.size(1), padding_idx) + x[mask] = y + elif x.size(1) > y.size(1): + x[mask] = torch.tensor(padding_idx).type_as(x) + if x.dim() == 2: + x[mask, :y.size(1)] = y + else: + x[mask, :y.size(1), :] = y + else: + x[mask] = y + return x diff --git a/fairseq/models/multilingual_transformer.py b/fairseq/models/multilingual_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..2f6a837805bd2071b0dbe6db644bdea9a8ee2994 --- /dev/null +++ b/fairseq/models/multilingual_transformer.py @@ -0,0 +1,196 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict + +from fairseq import utils +from fairseq.models import ( + FairseqMultiModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import ( + base_architecture, + Embedding, + TransformerModel, + TransformerEncoder, + TransformerDecoder, +) + + +@register_model('multilingual_transformer') +class MultilingualTransformerModel(FairseqMultiModel): + """Train Transformer models for multiple language pairs simultaneously. + + Requires `--task multilingual_translation`. + + We inherit all arguments from TransformerModel and assume that all language + pairs use a single Transformer architecture. In addition, we provide several + options that are specific to the multilingual setting. + + Args: + --share-encoder-embeddings: share encoder embeddings across all source languages + --share-decoder-embeddings: share decoder embeddings across all target languages + --share-encoders: share all encoder params (incl. embeddings) across all source languages + --share-decoders: share all decoder params (incl. embeddings) across all target languages + """ + + def __init__(self, encoders, decoders): + super().__init__(encoders, decoders) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + TransformerModel.add_args(parser) + parser.add_argument('--share-encoder-embeddings', action='store_true', + help='share encoder embeddings across languages') + parser.add_argument('--share-decoder-embeddings', action='store_true', + help='share decoder embeddings across languages') + parser.add_argument('--share-encoders', action='store_true', + help='share encoders across languages') + parser.add_argument('--share-decoders', action='store_true', + help='share decoders across languages') + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + from fairseq.tasks.multilingual_translation import MultilingualTranslationTask + assert isinstance(task, MultilingualTranslationTask) + + # make sure all arguments are present in older models + base_multilingual_architecture(args) + + if not hasattr(args, 'max_source_positions'): + args.max_source_positions = 1024 + if not hasattr(args, 'max_target_positions'): + args.max_target_positions = 1024 + + src_langs = [lang_pair.split('-')[0] for lang_pair in task.model_lang_pairs] + tgt_langs = [lang_pair.split('-')[1] for lang_pair in task.model_lang_pairs] + + if args.share_encoders: + args.share_encoder_embeddings = True + if args.share_decoders: + args.share_decoder_embeddings = True + + def build_embedding(dictionary, embed_dim, path=None): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + # build shared embeddings (if applicable) + shared_encoder_embed_tokens, shared_decoder_embed_tokens = None, None + if args.share_all_embeddings: + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + '--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim') + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path): + raise ValueError('--share-all-embeddings not compatible with --decoder-embed-path') + shared_encoder_embed_tokens = FairseqMultiModel.build_shared_embeddings( + dicts=task.dicts, + langs=task.langs, + embed_dim=args.encoder_embed_dim, + build_embedding=build_embedding, + pretrained_embed_path=args.encoder_embed_path, + ) + shared_decoder_embed_tokens = shared_encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + if args.share_encoder_embeddings: + shared_encoder_embed_tokens = ( + FairseqMultiModel.build_shared_embeddings( + dicts=task.dicts, + langs=src_langs, + embed_dim=args.encoder_embed_dim, + build_embedding=build_embedding, + pretrained_embed_path=args.encoder_embed_path, + ) + ) + if args.share_decoder_embeddings: + shared_decoder_embed_tokens = ( + FairseqMultiModel.build_shared_embeddings( + dicts=task.dicts, + langs=tgt_langs, + embed_dim=args.decoder_embed_dim, + build_embedding=build_embedding, + pretrained_embed_path=args.decoder_embed_path, + ) + ) + + # encoders/decoders for each language + lang_encoders, lang_decoders = {}, {} + + def get_encoder(lang): + if lang not in lang_encoders: + if shared_encoder_embed_tokens is not None: + encoder_embed_tokens = shared_encoder_embed_tokens + else: + encoder_embed_tokens = build_embedding( + task.dicts[lang], args.encoder_embed_dim, args.encoder_embed_path + ) + lang_encoders[lang] = TransformerEncoder(args, task.dicts[lang], encoder_embed_tokens) + return lang_encoders[lang] + + def get_decoder(lang): + if lang not in lang_decoders: + if shared_decoder_embed_tokens is not None: + decoder_embed_tokens = shared_decoder_embed_tokens + else: + decoder_embed_tokens = build_embedding( + task.dicts[lang], args.decoder_embed_dim, args.decoder_embed_path + ) + lang_decoders[lang] = TransformerDecoder(args, task.dicts[lang], decoder_embed_tokens) + return lang_decoders[lang] + + # shared encoders/decoders (if applicable) + shared_encoder, shared_decoder = None, None + if args.share_encoders: + shared_encoder = get_encoder(src_langs[0]) + if args.share_decoders: + shared_decoder = get_decoder(tgt_langs[0]) + + encoders, decoders = OrderedDict(), OrderedDict() + for lang_pair, src, tgt in zip(task.model_lang_pairs, src_langs, tgt_langs): + encoders[lang_pair] = shared_encoder if shared_encoder is not None else get_encoder(src) + decoders[lang_pair] = shared_decoder if shared_decoder is not None else get_decoder(tgt) + + return MultilingualTransformerModel(encoders, decoders) + + def load_state_dict(self, state_dict, strict=True, args=None): + state_dict_subset = state_dict.copy() + for k, _ in state_dict.items(): + assert k.startswith('models.') + lang_pair = k.split('.')[1] + if lang_pair not in self.models: + del state_dict_subset[k] + super().load_state_dict(state_dict_subset, strict=strict, args=args) + + +@register_model_architecture('multilingual_transformer', 'multilingual_transformer') +def base_multilingual_architecture(args): + base_architecture(args) + args.share_encoder_embeddings = getattr(args, 'share_encoder_embeddings', False) + args.share_decoder_embeddings = getattr(args, 'share_decoder_embeddings', False) + args.share_encoders = getattr(args, 'share_encoders', False) + args.share_decoders = getattr(args, 'share_decoders', False) + + +@register_model_architecture('multilingual_transformer', 'multilingual_transformer_iwslt_de_en') +def multilingual_transformer_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 512) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 1024) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 4) + args.encoder_layers = getattr(args, 'encoder_layers', 6) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 1024) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 4) + args.decoder_layers = getattr(args, 'decoder_layers', 6) + base_multilingual_architecture(args) diff --git a/fairseq/models/nat/__init__.py b/fairseq/models/nat/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b6ca06acb931f2fe53ec0dea0e0201c2289c7409 --- /dev/null +++ b/fairseq/models/nat/__init__.py @@ -0,0 +1,7 @@ +from .fairseq_nat_model import * +from .nonautoregressive_transformer import * +from .nat_crf_transformer import * +from .iterative_nonautoregressive_transformer import * +from .cmlm_transformer import * +from .levenshtein_transformer import * +from .insertion_transformer import * diff --git a/fairseq/models/nat/__pycache__/__init__.cpython-310.pyc b/fairseq/models/nat/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e81ccc2c6214ba7f1114628fa9d98c5986527ccd Binary files /dev/null and b/fairseq/models/nat/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/cmlm_transformer.cpython-310.pyc b/fairseq/models/nat/__pycache__/cmlm_transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..185fb60f672ab3c69853d32c46e990b7180930e8 Binary files /dev/null and b/fairseq/models/nat/__pycache__/cmlm_transformer.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/fairseq_nat_model.cpython-310.pyc b/fairseq/models/nat/__pycache__/fairseq_nat_model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eb5a166ae29430172bb09bde2e27dd01bbe0b79f Binary files /dev/null and b/fairseq/models/nat/__pycache__/fairseq_nat_model.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/insertion_transformer.cpython-310.pyc b/fairseq/models/nat/__pycache__/insertion_transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6489f0c539eebda4fa1df0b7f0472c19d0bf003f Binary files /dev/null and b/fairseq/models/nat/__pycache__/insertion_transformer.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/iterative_nonautoregressive_transformer.cpython-310.pyc b/fairseq/models/nat/__pycache__/iterative_nonautoregressive_transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..14245007117d49ce3e251968a1b9573c3ef65e33 Binary files /dev/null and b/fairseq/models/nat/__pycache__/iterative_nonautoregressive_transformer.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/levenshtein_transformer.cpython-310.pyc b/fairseq/models/nat/__pycache__/levenshtein_transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..feebfd93995d834effa3b09f8cfac48ef7220e68 Binary files /dev/null and b/fairseq/models/nat/__pycache__/levenshtein_transformer.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/levenshtein_utils.cpython-310.pyc b/fairseq/models/nat/__pycache__/levenshtein_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c59b896296a3db6aa648366de7168d99d7ca6a07 Binary files /dev/null and b/fairseq/models/nat/__pycache__/levenshtein_utils.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/nat_crf_transformer.cpython-310.pyc b/fairseq/models/nat/__pycache__/nat_crf_transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0453b8c53cfa8483362f0210361a7ec08dc43fa1 Binary files /dev/null and b/fairseq/models/nat/__pycache__/nat_crf_transformer.cpython-310.pyc differ diff --git a/fairseq/models/nat/__pycache__/nonautoregressive_transformer.cpython-310.pyc b/fairseq/models/nat/__pycache__/nonautoregressive_transformer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9b0ae0ae1e6dcbfdb5ea2a7b69f19e89124da553 Binary files /dev/null and b/fairseq/models/nat/__pycache__/nonautoregressive_transformer.cpython-310.pyc differ diff --git a/fairseq/models/nat/cmlm_transformer.py b/fairseq/models/nat/cmlm_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..86c770569d43d9946db9c74cfa4468c432df5482 --- /dev/null +++ b/fairseq/models/nat/cmlm_transformer.py @@ -0,0 +1,154 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +This file implements: +Ghazvininejad, Marjan, et al. +"Constant-time machine translation with conditional masked language models." +arXiv preprint arXiv:1904.09324 (2019). +""" + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel +from fairseq.utils import new_arange + + +def _skeptical_unmasking(output_scores, output_masks, p): + sorted_index = output_scores.sort(-1)[1] + boundary_len = ( + (output_masks.sum(1, keepdim=True).type_as(output_scores) - 2) * p + ).long() + skeptical_mask = new_arange(output_masks) < boundary_len + return skeptical_mask.scatter(1, sorted_index, skeptical_mask) + + +@register_model("cmlm_transformer") +class CMLMNATransformerModel(NATransformerModel): + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + assert not self.decoder.src_embedding_copy, "do not support embedding copy." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + # length prediction + length_out = self.decoder.forward_length(normalize=False, encoder_out=encoder_out) + length_tgt = self.decoder.forward_length_prediction(length_out, encoder_out, tgt_tokens) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out) + word_ins_mask = prev_output_tokens.eq(self.unk) + + return { + "word_ins": { + "out": word_ins_out, "tgt": tgt_tokens, + "mask": word_ins_mask, "ls": self.args.label_smoothing, + "nll_loss": True + }, + "length": { + "out": length_out, "tgt": length_tgt, + "factor": self.decoder.length_loss_factor + } + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + + step = decoder_out.step + max_step = decoder_out.max_step + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder + output_masks = output_tokens.eq(self.unk) + _scores, _tokens = self.decoder( + normalize=True, + prev_output_tokens=output_tokens, + encoder_out=encoder_out, + ).max(-1) + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + + if history is not None: + history.append(output_tokens.clone()) + + # skeptical decoding (depend on the maximum decoding steps.) + if (step + 1) < max_step: + skeptical_mask = _skeptical_unmasking( + output_scores, output_tokens.ne(self.pad), 1 - (step + 1) / max_step + ) + + output_tokens.masked_fill_(skeptical_mask, self.unk) + output_scores.masked_fill_(skeptical_mask, 0.0) + + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history + ) + + +@register_model_architecture("cmlm_transformer", "cmlm_transformer") +def cmlm_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", True) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.ngram_predictor = getattr(args, "ngram_predictor", 1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + +@register_model_architecture("cmlm_transformer", "cmlm_transformer_wmt_en_de") +def cmlm_wmt_en_de(args): + cmlm_base_architecture(args) diff --git a/fairseq/models/nat/fairseq_nat_model.py b/fairseq/models/nat/fairseq_nat_model.py new file mode 100644 index 0000000000000000000000000000000000000000..d37a234ba9c8658eba8b393c25af98807ea1ec10 --- /dev/null +++ b/fairseq/models/nat/fairseq_nat_model.py @@ -0,0 +1,145 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import torch + +from fairseq.models.transformer import TransformerModel, TransformerEncoder, TransformerDecoder +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +def ensemble_encoder(func): + def wrapper(self, *args, **kwargs): + if self.ensemble_models is None or len(self.ensemble_models) == 1: + return func(self, *args, **kwargs) + encoder_outs = [func(model, *args, **kwargs) for model in self.ensemble_models] + _encoder_out = encoder_outs[0] + + def stack(key): + outs = [getattr(e, key) for e in encoder_outs] + return torch.stack(outs, -1) if outs[0] is not None else None + + return _encoder_out._replace( + encoder_out=stack('encoder_out'), + encoder_embedding=stack('encoder_embedding'), + encoder_states=stack('encoder_states') + ) + return wrapper + + +def ensemble_decoder(func): + def wrapper(self, normalize=False, encoder_out=None, *args, **kwargs): + if self.ensemble_models is None or len(self.ensemble_models) == 1: + return func(self, normalize=normalize, encoder_out=encoder_out, *args, **kwargs) + + action_outs = [ + func(model, normalize=normalize, encoder_out=encoder_out._replace( + encoder_out=encoder_out.encoder_out[:, :, :, i] + ), *args, **kwargs) + for i, model in enumerate(self.ensemble_models) + ] + + if not isinstance(action_outs[0], tuple): # return multiple values + action_outs = [[a] for a in action_outs] + else: + action_outs = [list(a) for a in action_outs] + + ensembled_outs = [] + for i in range(len(action_outs[0])): + if i == 0 and normalize: + ensembled_outs += [ + torch.logsumexp( + torch.stack([a[i] for a in action_outs], -1), + dim=-1) - math.log(len(self.ensemble_models)) + ] + elif action_outs[0][i] is not None: + ensembled_outs += [ + torch.stack([a[i] for a in action_outs], -1) + ] + else: + ensembled_outs += [None] + + if len(ensembled_outs) == 1: + return ensembled_outs[0] + return tuple(ensembled_outs) + return wrapper + + +class FairseqNATModel(TransformerModel): + """ + Abstract class for all nonautoregressive-based models + """ + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + self.tgt_dict = decoder.dictionary + self.bos = decoder.dictionary.bos() + self.eos = decoder.dictionary.eos() + self.pad = decoder.dictionary.pad() + self.unk = decoder.dictionary.unk() + + self.ensemble_models = None + + @property + def allow_length_beam(self): + return False + + @property + def allow_ensemble(self): + return True + + def enable_ensemble(self, models): + self.encoder.ensemble_models = [m.encoder for m in models] + self.decoder.ensemble_models = [m.decoder for m in models] + + @staticmethod + def add_args(parser): + TransformerModel.add_args(parser) + parser.add_argument( + "--apply-bert-init", + action="store_true", + help="use custom param initialization for BERT", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = FairseqNATDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + encoder = FairseqNATEncoder(args, src_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + encoder.apply(init_bert_params) + return encoder + + def forward_encoder(self, encoder_inputs): + return self.encoder(*encoder_inputs) + + def forward_decoder(self, *args, **kwargs): + return NotImplementedError + + def initialize_output_tokens(self, *args, **kwargs): + return NotImplementedError + + def forward(self, *args, **kwargs): + return NotImplementedError + + +class FairseqNATEncoder(TransformerEncoder): + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + self.ensemble_models = None + + @ensemble_encoder + def forward(self, *args, **kwargs): + return super().forward(*args, **kwargs) + + +class FairseqNATDecoder(TransformerDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__(args, dictionary, embed_tokens, no_encoder_attn) + self.ensemble_models = None diff --git a/fairseq/models/nat/insertion_transformer.py b/fairseq/models/nat/insertion_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..a5f3c1abc5a7106aac2430e05bfaa8652162913e --- /dev/null +++ b/fairseq/models/nat/insertion_transformer.py @@ -0,0 +1,280 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn.functional as F + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import ( + LevenshteinTransformerDecoder, + LevenshteinTransformerModel, + FairseqNATModel, + ensemble_decoder +) +from fairseq.models.transformer import Linear +from fairseq.utils import new_arange +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +class NegativeDistanceScore(object): + def __init__(self): + + # pre-compute some values + self.scores = {} + + self.scores[0.5] = self.compute_score_full(50, 0.5) + self.scores[1.0] = self.compute_score_full(50, 1.0) + self.scores[2.0] = self.compute_score_full(50, 2.0) + + def __call__(self, i, L, tau): + if (tau is None) or (tau > 1000): + return 1 / L + + if tau in self.scores: + if L < self.scores[tau].shape[0]: + return self.scores[tau][L - 1, i] + return self.compute_score(L, tau)[i] + + def compute_score(self, L, tau): + s = np.array([-abs(L / 2 - i) / tau for i in range(L)]) + s = np.exp(s - s.max()) + return s / s.sum() + + def compute_score_full(self, L, tau): + s = -abs(np.arange(0, L - 1)[:, None] / 2 - np.arange(L)[None, :]) / tau + s = np.tril(s, 0) + np.triu(s - float("inf"), 1) + s = np.exp(s - s.max(1, keepdims=True)) + return s / s.sum(1, keepdims=True) + + +neg_scorer = NegativeDistanceScore() + + +def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx, vocab_size, tau=None): + try: + from fairseq import libnat + except ImportError as e: + import sys + sys.stderr.write('ERROR: missing libnat. run `pip install --editable .`\n') + raise e + + B = in_tokens.size(0) + T = in_tokens.size(1) + V = vocab_size + + with torch.cuda.device_of(in_tokens): + in_tokens_list = [ + [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + insert_labels = [a[:-1] for a in full_labels] + + # numericalize1 + insert_label_tensors = in_tokens.new_zeros(B * (T - 1) * V).float() + insert_index, insert_labels = zip( + *[ + (w + (j + i * (T - 1)) * V, neg_scorer(k, len(label), tau)) + for i, labels in enumerate(insert_labels) + for j, label in enumerate(labels[1:-1]) + for k, w in enumerate(label) + ] + ) # HACK 1:-1 + insert_index, insert_labels = [ + torch.tensor(list(a), device=in_tokens.device) + for a in [insert_index, insert_labels] + ] + insert_label_tensors.scatter_(0, insert_index.long(), insert_labels) + insert_label_tensors = insert_label_tensors.view(B, T - 1, V) + + return insert_label_tensors + + +def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, padding_idx): + + padding_masks = in_tokens[:, 1:].eq(padding_idx) + word_ins_scores.masked_fill_(padding_masks, 0.0) + word_ins_pred.masked_fill_(padding_masks, padding_idx) + + in_coords = new_arange(in_tokens).type_as(in_scores) + + # shift all padding predictions to infinite + out_coords = (in_coords[:, 1:] - 0.5).masked_fill( + word_ins_pred.eq(padding_idx), float("inf") + ) + out_coords = torch.cat([in_coords, out_coords], 1).sort(-1)[1] + out_tokens = torch.cat([in_tokens, word_ins_pred], 1).gather(1, out_coords) + out_scores = torch.cat([in_scores, word_ins_scores], 1).gather(1, out_coords) + return out_tokens, out_scores + + +@register_model("insertion_transformer") +class InsertionTransformerModel(LevenshteinTransformerModel): + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + parser.add_argument("--label-tau", default=None, type=float) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = InsertionTransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + assert tgt_tokens is not None, "forward function only supports training." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # generate training labels for insertion + word_ins_out = self.decoder.forward_word_ins( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out + ) + + word_ins_tgt = _get_ins_targets( + prev_output_tokens, + tgt_tokens, + self.pad, + self.unk, + len(self.tgt_dict), + tau=self.decoder.label_tau, + ).type_as(word_ins_out) + word_ins_masks = prev_output_tokens[:, 1:].ne(self.pad) + + return { + "word_ins": { + "out": word_ins_out, "tgt": word_ins_tgt, + "mask": word_ins_masks, "ls": self.args.label_smoothing, + "nll_loss": True + } + } + + def forward_decoder( + self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs + ): + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # TODO: decoding for InsertionTransformer + word_ins_score = self.decoder.forward_word_ins( + normalize=True, + prev_output_tokens=output_tokens, + encoder_out=encoder_out + ) + + if eos_penalty > 0.0: + word_ins_score[:, :, self.pad] -= eos_penalty + word_ins_score, word_ins_pred = word_ins_score.max(-1) + output_tokens, output_scores = _apply_ins_words( + output_tokens, output_scores, word_ins_pred, word_ins_score, self.pad + ) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history + ) + + +class InsertionTransformerDecoder(LevenshteinTransformerDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + # use the TransformerDecoder's __init__ + super(LevenshteinTransformerDecoder, self).__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + self.pool_out = Linear(self.output_embed_dim * 2, self.output_embed_dim) + + self.label_tau = getattr(args, "label_tau", None) + + @ensemble_decoder + def forward_word_ins(self, normalize, encoder_out, prev_output_tokens): + features = self.extract_features(prev_output_tokens, encoder_out=encoder_out)[0] + features = self.pool_out( + torch.cat([features[:, :-1, :], features[:, 1:, :]], 2) + ) + decoder_out = self.output_layer(features) + return F.log_softmax(decoder_out, -1) if normalize else decoder_out + + def forward_mask_ins(self, *args, **kwargs): + raise NotImplementedError + + def forward_word_del(self, *args, **kwargs): + raise NotImplementedError + + +@register_model_architecture("insertion_transformer", "insertion_transformer") +def insertion_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # special for insertion transformer + args.label_tau = getattr(args, "label_tau", None) diff --git a/fairseq/models/nat/iterative_nonautoregressive_transformer.py b/fairseq/models/nat/iterative_nonautoregressive_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..dc340c387df8bb13eacf8ae527b855a8f3501d53 --- /dev/null +++ b/fairseq/models/nat/iterative_nonautoregressive_transformer.py @@ -0,0 +1,205 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel + + +def _sequential_poisoning(s, V, beta=0.33, bos=2, eos=3, pad=1): + # s: input batch + # V: vocabulary size + rand_words = torch.randint(low=4, high=V, size=s.size(), device=s.device) + choices = torch.rand(size=s.size(), device=s.device) + choices.masked_fill_((s == pad) | (s == bos) | (s == eos), 1) + + replace = choices < beta / 3 + repeat = (choices >= beta / 3) & (choices < beta * 2 / 3) + swap = (choices >= beta * 2 / 3) & (choices < beta) + safe = choices >= beta + + for i in range(s.size(1) - 1): + rand_word = rand_words[:, i] + next_word = s[:, i + 1] + self_word = s[:, i] + + replace_i = replace[:, i] + swap_i = swap[:, i] & (next_word != 3) + repeat_i = repeat[:, i] & (next_word != 3) + safe_i = safe[:, i] | ((next_word == 3) & (~replace_i)) + + s[:, i] = ( + self_word * (safe_i | repeat_i).long() + + next_word * swap_i.long() + + rand_word * replace_i.long() + ) + s[:, i + 1] = ( + next_word * (safe_i | replace_i).long() + + self_word * (swap_i | repeat_i).long() + ) + return s + + +def gumbel_noise(input, TINY=1e-8): + return input.new_zeros(*input.size()).uniform_().add_( + TINY).log_().neg_().add_(TINY).log_().neg_() + + +@register_model("iterative_nonautoregressive_transformer") +class IterNATransformerModel(NATransformerModel): + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + parser.add_argument("--train-step", type=int, + help="number of refinement iterations during training") + parser.add_argument("--dae-ratio", type=float, + help="the probability of switching to the denoising auto-encoder loss") + parser.add_argument("--stochastic-approx", action="store_true", + help="sampling from the decoder as the inputs for next iteration") + + @classmethod + def build_model(cls, args, task): + model = super().build_model(args, task) + model.train_step = getattr(args, "train_step", 4) + model.dae_ratio = getattr(args, "dae_ratio", 0.5) + model.stochastic_approx = getattr(args, "stochastic_approx", False) + return model + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + B, T = prev_output_tokens.size() + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length(normalize=False, encoder_out=encoder_out) + length_tgt = self.decoder.forward_length_prediction(length_out, encoder_out, tgt_tokens) + + # decoding + word_ins_outs, word_ins_tgts, word_ins_masks = [], [], [] + for t in range(self.train_step): + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + step=t, + ) + word_ins_tgt = tgt_tokens + word_ins_mask = word_ins_tgt.ne(self.pad) + + word_ins_outs.append(word_ins_out) + word_ins_tgts.append(word_ins_tgt) + word_ins_masks.append(word_ins_mask) + + if t < (self.train_step - 1): + # prediction for next iteration + if self.stochastic_approx: + word_ins_prediction = ( + word_ins_out + gumbel_noise(word_ins_out) + ).max(-1)[1] + else: + word_ins_prediction = word_ins_out.max(-1)[1] + + prev_output_tokens = prev_output_tokens.masked_scatter( + word_ins_mask, word_ins_prediction[word_ins_mask] + ) + + if self.dae_ratio > 0: + # we do not perform denoising for the first iteration + corrputed = ( + torch.rand(size=(B,), device=prev_output_tokens.device) + < self.dae_ratio + ) + corrputed_tokens = _sequential_poisoning( + tgt_tokens[corrputed], + len(self.tgt_dict), + 0.33, + self.bos, + self.eos, + self.pad, + ) + prev_output_tokens[corrputed] = corrputed_tokens + + # concat everything + word_ins_out = torch.cat(word_ins_outs, 0) + word_ins_tgt = torch.cat(word_ins_tgts, 0) + word_ins_mask = torch.cat(word_ins_masks, 0) + + return { + "word_ins": { + "out": word_ins_out, "tgt": word_ins_tgt, + "mask": word_ins_mask, "ls": self.args.label_smoothing, + "nll_loss": True + }, + "length": { + "out": length_out, "tgt": length_tgt, + "factor": self.decoder.length_loss_factor + } + } + + +@register_model_architecture( + "iterative_nonautoregressive_transformer", "iterative_nonautoregressive_transformer" +) +def inat_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.ngram_predictor = getattr(args, "ngram_predictor", 1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + args.train_step = getattr(args, "train_step", 4) + args.dae_ratio = getattr(args, "dae_ratio", 0.5) + args.stochastic_approx = getattr(args, "stochastic_approx", False) + + +@register_model_architecture( + "iterative_nonautoregressive_transformer", + "iterative_nonautoregressive_transformer_wmt_en_de", +) +def iter_nat_wmt_en_de(args): + inat_base_architecture(args) diff --git a/fairseq/models/nat/levenshtein_transformer.py b/fairseq/models/nat/levenshtein_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..e1748145c3b37de434e82b49d41504867f487271 --- /dev/null +++ b/fairseq/models/nat/levenshtein_transformer.py @@ -0,0 +1,478 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq.iterative_refinement_generator import DecoderOut +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import ( + Embedding, + TransformerDecoderLayer +) + +from fairseq.models.nat import ( + FairseqNATModel, + FairseqNATDecoder, + ensemble_decoder +) + +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +from .levenshtein_utils import ( + _skip, _skip_encoder_out, _fill, + _get_ins_targets, _get_del_targets, + _apply_ins_masks, _apply_ins_words, _apply_del_words +) + + +@register_model("levenshtein_transformer") +class LevenshteinTransformerModel(FairseqNATModel): + + @property + def allow_length_beam(self): + return False + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + parser.add_argument( + "--early-exit", + default="6,6,6", + type=str, + help="number of decoder layers before word_del, mask_ins, word_ins", + ) + parser.add_argument( + "--no-share-discriminator", + action="store_true", + help="separate parameters for discriminator", + ) + parser.add_argument( + "--no-share-maskpredictor", + action="store_true", + help="separate parameters for mask-predictor", + ) + parser.add_argument( + "--share-discriminator-maskpredictor", + action="store_true", + help="share the parameters for both mask-predictor and discriminator", + ) + parser.add_argument( + "--sampling-for-deletion", + action='store_true', + help='instead of argmax, use sampling to predict the tokens' + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = LevenshteinTransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + assert tgt_tokens is not None, "forward function only supports training." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # generate training labels for insertion + masked_tgt_masks, masked_tgt_tokens, mask_ins_targets = _get_ins_targets( + prev_output_tokens, tgt_tokens, self.pad, self.unk + ) + mask_ins_targets = mask_ins_targets.clamp(min=0, max=255) # for safe prediction + mask_ins_masks = prev_output_tokens[:, 1:].ne(self.pad) + + mask_ins_out, _ = self.decoder.forward_mask_ins( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out + ) + word_ins_out, _ = self.decoder.forward_word_ins( + normalize=False, + prev_output_tokens=masked_tgt_tokens, + encoder_out=encoder_out + ) + + # make online prediction + if self.decoder.sampling_for_deletion: + word_predictions = torch.multinomial( + F.softmax(word_ins_out, -1).view(-1, word_ins_out.size(-1)), 1).view( + word_ins_out.size(0), -1) + else: + word_predictions = F.log_softmax(word_ins_out, dim=-1).max(2)[1] + + word_predictions.masked_scatter_( + ~masked_tgt_masks, tgt_tokens[~masked_tgt_masks] + ) + + # generate training labels for deletion + word_del_targets = _get_del_targets(word_predictions, tgt_tokens, self.pad) + word_del_out, _ = self.decoder.forward_word_del( + normalize=False, + prev_output_tokens=word_predictions, + encoder_out=encoder_out) + word_del_masks = word_predictions.ne(self.pad) + + return { + "mask_ins": { + "out": mask_ins_out, "tgt": mask_ins_targets, + "mask": mask_ins_masks, "ls": 0.01, + }, + "word_ins": { + "out": word_ins_out, "tgt": tgt_tokens, + "mask": masked_tgt_masks, "ls": self.args.label_smoothing, + "nll_loss": True + }, + "word_del": { + "out": word_del_out, "tgt": word_del_targets, + "mask": word_del_masks + } + } + + def forward_decoder( + self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs + ): + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + attn = decoder_out.attn + history = decoder_out.history + + bsz = output_tokens.size(0) + if max_ratio is None: + max_lens = torch.zeros_like(output_tokens).fill_(255) + else: + if encoder_out.encoder_padding_mask is None: + max_src_len = encoder_out.encoder_out.size(0) + src_lens = encoder_out.encoder_out.new(bsz).fill_(max_src_len) + else: + src_lens = (~encoder_out.encoder_padding_mask).sum(1) + max_lens = (src_lens * max_ratio).clamp(min=10).long() + + # delete words + # do not delete tokens if it is + can_del_word = output_tokens.ne(self.pad).sum(1) > 2 + if can_del_word.sum() != 0: # we cannot delete, skip + word_del_score, word_del_attn = self.decoder.forward_word_del( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_del_word), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_del_word) + ) + word_del_pred = word_del_score.max(-1)[1].bool() + + _tokens, _scores, _attn = _apply_del_words( + output_tokens[can_del_word], + output_scores[can_del_word], + word_del_attn, + word_del_pred, + self.pad, + self.bos, + self.eos, + ) + output_tokens = _fill(output_tokens, can_del_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_del_word, _scores, 0) + attn = _fill(attn, can_del_word, _attn, 0.) + + if history is not None: + history.append(output_tokens.clone()) + + # insert placeholders + can_ins_mask = output_tokens.ne(self.pad).sum(1) < max_lens + if can_ins_mask.sum() != 0: + mask_ins_score, _ = self.decoder.forward_mask_ins( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_ins_mask), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_ins_mask) + ) + if eos_penalty > 0.0: + mask_ins_score[:, :, 0] = mask_ins_score[:, :, 0] - eos_penalty + mask_ins_pred = mask_ins_score.max(-1)[1] + mask_ins_pred = torch.min( + mask_ins_pred, max_lens[can_ins_mask, None].expand_as(mask_ins_pred) + ) + + _tokens, _scores = _apply_ins_masks( + output_tokens[can_ins_mask], + output_scores[can_ins_mask], + mask_ins_pred, + self.pad, + self.unk, + self.eos, + ) + output_tokens = _fill(output_tokens, can_ins_mask, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_mask, _scores, 0) + + if history is not None: + history.append(output_tokens.clone()) + + # insert words + can_ins_word = output_tokens.eq(self.unk).sum(1) > 0 + if can_ins_word.sum() != 0: + word_ins_score, word_ins_attn = self.decoder.forward_word_ins( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_ins_word), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_ins_word) + ) + word_ins_score, word_ins_pred = word_ins_score.max(-1) + _tokens, _scores = _apply_ins_words( + output_tokens[can_ins_word], + output_scores[can_ins_word], + word_ins_pred, + word_ins_score, + self.unk, + ) + + output_tokens = _fill(output_tokens, can_ins_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_word, _scores, 0) + attn = _fill(attn, can_ins_word, word_ins_attn, 0.) + + if history is not None: + history.append(output_tokens.clone()) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + attn = None if attn is None else attn[:, :cut_off, :] + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=attn, + history=history + ) + + def initialize_output_tokens(self, encoder_out, src_tokens): + initial_output_tokens = src_tokens.new_zeros(src_tokens.size(0), 2) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens[:, 1] = self.eos + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(encoder_out.encoder_out) + + return DecoderOut( + output_tokens=initial_output_tokens, + output_scores=initial_output_scores, + attn=None, + step=0, + max_step=0, + history=None + ) + + +class LevenshteinTransformerDecoder(FairseqNATDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + self.sampling_for_deletion = getattr(args, "sampling_for_deletion", False) + self.embed_mask_ins = Embedding(256, self.output_embed_dim * 2, None) + self.embed_word_del = Embedding(2, self.output_embed_dim, None) + + # del_word, ins_mask, ins_word + self.early_exit = [int(i) for i in args.early_exit.split(',')] + assert len(self.early_exit) == 3 + + # copy layers for mask-predict/deletion + self.layers_msk = None + if getattr(args, "no_share_maskpredictor", False): + self.layers_msk = nn.ModuleList([ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(self.early_exit[1]) + ]) + self.layers_del = None + if getattr(args, "no_share_discriminator", False): + self.layers_del = nn.ModuleList([ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(self.early_exit[0]) + ]) + + if getattr(args, "share_discriminator_maskpredictor", False): + assert getattr(args, "no_share_discriminator", False), "must set saperate discriminator" + self.layers_msk = self.layers_del + + def extract_features( + self, prev_output_tokens, encoder_out=None, early_exit=None, layers=None, **unused + ): + """ + Similar to *forward* but only return features. + Inputs: + prev_output_tokens: Tensor(B, T) + encoder_out: a dictionary of hidden states and masks + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + the LevenshteinTransformer decoder has full-attention to all generated tokens + """ + # embed positions + positions = ( + self.embed_positions(prev_output_tokens) + if self.embed_positions is not None + else None + ) + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + inner_states = [x] + + # decoder layers + decoder_padding_mask = prev_output_tokens.eq(self.padding_idx) + layers = self.layers if layers is None else layers + early_exit = len(layers) if early_exit is None else early_exit + for _, layer in enumerate(layers[: early_exit]): + x, attn, _ = layer( + x, + encoder_out.encoder_out if encoder_out is not None else None, + encoder_out.encoder_padding_mask if encoder_out is not None else None, + self_attn_mask=None, + self_attn_padding_mask=decoder_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": attn, "inner_states": inner_states} + + @ensemble_decoder + def forward_mask_ins(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[1], layers=self.layers_msk, **unused + ) + features_cat = torch.cat([features[:, :-1, :], features[:, 1:, :]], 2) + decoder_out = F.linear(features_cat, self.embed_mask_ins.weight) + if normalize: + return F.log_softmax(decoder_out, -1), extra['attn'] + return decoder_out, extra['attn'] + + @ensemble_decoder + def forward_word_ins(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[2], layers=self.layers, **unused + ) + decoder_out = self.output_layer(features) + if normalize: + return F.log_softmax(decoder_out, -1), extra['attn'] + return decoder_out, extra['attn'] + + @ensemble_decoder + def forward_word_del(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[0], layers=self.layers_del, **unused + ) + decoder_out = F.linear(features, self.embed_word_del.weight) + if normalize: + return F.log_softmax(decoder_out, -1), extra['attn'] + return decoder_out, extra['attn'] + + +@register_model_architecture("levenshtein_transformer", "levenshtein_transformer") +def levenshtein_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.sampling_for_deletion = getattr(args, "sampling_for_deletion", False) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.early_exit = getattr(args, "early_exit", "6,6,6") + args.no_share_discriminator = getattr(args, "no_share_discriminator", False) + args.no_share_maskpredictor = getattr(args, "no_share_maskpredictor", False) + args.share_discriminator_maskpredictor = getattr(args, "share_discriminator_maskpredictor", False) + args.no_share_last_layer = getattr(args, "no_share_last_layer", False) + + +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_wmt_en_de" +) +def levenshtein_transformer_wmt_en_de(args): + levenshtein_base_architecture(args) + + +# similar parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_vaswani_wmt_en_de_big" +) +def levenshtein_transformer_vaswani_wmt_en_de_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + levenshtein_base_architecture(args) + + +# default parameters used in tensor2tensor implementation +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_wmt_en_de_big" +) +def levenshtein_transformer_wmt_en_de_big_t2t(args): + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + levenshtein_transformer_vaswani_wmt_en_de_big(args) diff --git a/fairseq/models/nat/levenshtein_utils.py b/fairseq/models/nat/levenshtein_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..e29b1fa27cb92f2ce3dd2235d3d6ece910c3ffe5 --- /dev/null +++ b/fairseq/models/nat/levenshtein_utils.py @@ -0,0 +1,284 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.utils import new_arange + + +# -------------- Helper Functions --------------------------------------------------- # + +def load_libnat(): + try: + from fairseq import libnat_cuda + return libnat_cuda, True + + except ImportError as e: + print(str(e) + '... fall back to CPU version') + + try: + from fairseq import libnat + return libnat, False + + except ImportError as e: + import sys + sys.stderr.write("ERROR: missing libnat_cuda. run `python setup.py build_ext --inplace`\n") + raise e + + +def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx): + libnat, use_cuda = load_libnat() + + def _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx): + in_masks = in_tokens.ne(padding_idx) + out_masks = out_tokens.ne(padding_idx) + mask_ins_targets, masked_tgt_masks = libnat.generate_insertion_labels( + out_tokens.int(), libnat.levenshtein_distance( + in_tokens.int(), out_tokens.int(), + in_masks.sum(1).int(), out_masks.sum(1).int() + ) + ) + masked_tgt_masks = masked_tgt_masks.bool() & out_masks + mask_ins_targets = mask_ins_targets.type_as( + in_tokens)[:, 1:in_masks.size(1)].masked_fill_(~in_masks[:, 1:], 0) + masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx) + return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets + + def _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx): + in_seq_len, out_seq_len = in_tokens.size(1), out_tokens.size(1) + + in_tokens_list = [ + [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + mask_inputs = [ + [len(c) if c[0] != padding_idx else 0 for c in a[:-1]] for a in full_labels + ] + + # generate labels + masked_tgt_masks = [] + for mask_input in mask_inputs: + mask_label = [] + for beam_size in mask_input[1:-1]: # HACK 1:-1 + mask_label += [0] + [1 for _ in range(beam_size)] + masked_tgt_masks.append( + mask_label + [0 for _ in range(out_seq_len - len(mask_label))] + ) + mask_ins_targets = [ + mask_input[1:-1] + [0 for _ in range(in_seq_len - 1 - len(mask_input[1:-1]))] + for mask_input in mask_inputs + ] + + # transform to tensor + masked_tgt_masks = torch.tensor( + masked_tgt_masks, device=out_tokens.device + ).bool() + mask_ins_targets = torch.tensor(mask_ins_targets, device=in_tokens.device) + masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx) + return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets + + if use_cuda: + return _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx) + return _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx) + + +def _get_del_targets(in_tokens, out_tokens, padding_idx): + libnat, use_cuda = load_libnat() + + def _get_del_targets_cuda(in_tokens, out_tokens, padding_idx): + in_masks = in_tokens.ne(padding_idx) + out_masks = out_tokens.ne(padding_idx) + + word_del_targets = libnat.generate_deletion_labels( + in_tokens.int(), + libnat.levenshtein_distance( + in_tokens.int(), out_tokens.int(), + in_masks.sum(1).int(), out_masks.sum(1).int() + ) + ) + word_del_targets = word_del_targets.type_as(in_tokens).masked_fill_(~in_masks, 0) + return word_del_targets + + def _get_del_targets_cpu(in_tokens, out_tokens, padding_idx): + out_seq_len = out_tokens.size(1) + with torch.cuda.device_of(in_tokens): + in_tokens_list = [ + [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + word_del_targets = [b[-1] for b in full_labels] + word_del_targets = [ + labels + [0 for _ in range(out_seq_len - len(labels))] + for labels in word_del_targets + ] + + # transform to tensor + word_del_targets = torch.tensor(word_del_targets, device=out_tokens.device) + return word_del_targets + + if use_cuda: + return _get_del_targets_cuda(in_tokens, out_tokens, padding_idx) + return _get_del_targets_cpu(in_tokens, out_tokens, padding_idx) + + +def _apply_ins_masks( + in_tokens, in_scores, mask_ins_pred, padding_idx, unk_idx, eos_idx +): + + in_masks = in_tokens.ne(padding_idx) + in_lengths = in_masks.sum(1) + + # HACK: hacky way to shift all the paddings to eos first. + in_tokens.masked_fill_(~in_masks, eos_idx) + mask_ins_pred.masked_fill_(~in_masks[:, 1:], 0) + + out_lengths = in_lengths + mask_ins_pred.sum(1) + out_max_len = out_lengths.max() + out_masks = ( + new_arange(out_lengths, out_max_len)[None, :] + < out_lengths[:, None] + ) + + reordering = (mask_ins_pred + in_masks[:, 1:].long()).cumsum(1) + out_tokens = ( + in_tokens.new_zeros(in_tokens.size(0), out_max_len) + .fill_(padding_idx) + .masked_fill_(out_masks, unk_idx) + ) + out_tokens[:, 0] = in_tokens[:, 0] + out_tokens.scatter_(1, reordering, in_tokens[:, 1:]) + + out_scores = None + if in_scores is not None: + in_scores.masked_fill_(~in_masks, 0) + out_scores = in_scores.new_zeros(*out_tokens.size()) + out_scores[:, 0] = in_scores[:, 0] + out_scores.scatter_(1, reordering, in_scores[:, 1:]) + + return out_tokens, out_scores + + +def _apply_ins_words( + in_tokens, in_scores, word_ins_pred, word_ins_scores, unk_idx +): + word_ins_masks = in_tokens.eq(unk_idx) + out_tokens = in_tokens.masked_scatter(word_ins_masks, word_ins_pred[word_ins_masks]) + + if in_scores is not None: + out_scores = in_scores.masked_scatter( + word_ins_masks, word_ins_scores[word_ins_masks] + ) + else: + out_scores = None + + return out_tokens, out_scores + + +def _apply_del_words( + in_tokens, in_scores, in_attn, word_del_pred, padding_idx, bos_idx, eos_idx +): + # apply deletion to a tensor + in_masks = in_tokens.ne(padding_idx) + bos_eos_masks = in_tokens.eq(bos_idx) | in_tokens.eq(eos_idx) + + max_len = in_tokens.size(1) + word_del_pred.masked_fill_(~in_masks, 1) + word_del_pred.masked_fill_(bos_eos_masks, 0) + + reordering = ( + new_arange(in_tokens) + .masked_fill_(word_del_pred, max_len) + .sort(1)[1] + ) + + out_tokens = in_tokens.masked_fill(word_del_pred, padding_idx).gather(1, reordering) + + out_scores = None + if in_scores is not None: + out_scores = in_scores.masked_fill(word_del_pred, 0).gather(1, reordering) + + out_attn = None + if in_attn is not None: + _mask = word_del_pred[:, :, None].expand_as(in_attn) + _reordering = reordering[:, :, None].expand_as(in_attn) + out_attn = in_attn.masked_fill(_mask, 0.).gather(1, _reordering) + + return out_tokens, out_scores, out_attn + + +def _skip(x, mask): + """ + Getting sliced (dim=0) tensor by mask. Supporting tensor and list/dict of tensors. + """ + if isinstance(x, int): + return x + + if x is None: + return None + + if isinstance(x, torch.Tensor): + if x.size(0) == mask.size(0): + return x[mask] + elif x.size(1) == mask.size(0): + return x[:, mask] + + if isinstance(x, list): + return [_skip(x_i, mask) for x_i in x] + + if isinstance(x, dict): + return {k: _skip(v, mask) for k, v in x.items()} + + raise NotImplementedError + + +def _skip_encoder_out(encoder, encoder_out, mask): + if not mask.any(): + return encoder_out + else: + return encoder.reorder_encoder_out(encoder_out, mask.nonzero().squeeze()) + + +def _fill(x, mask, y, padding_idx): + """ + Filling tensor x with y at masked positions (dim=0). + """ + if x is None: + return y + assert x.dim() == y.dim() and mask.size(0) == x.size(0) + assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2)) + n_selected = mask.sum() + assert n_selected == y.size(0) + + if n_selected == x.size(0): + return y + + if x.size(1) < y.size(1): + dims = [x.size(0), y.size(1) - x.size(1)] + if x.dim() == 3: + dims.append(x.size(2)) + x = torch.cat([x, x.new_zeros(*dims).fill_(padding_idx)], 1) + x[mask] = y + elif x.size(1) > y.size(1): + x[mask] = padding_idx + if x.dim() == 2: + x[mask, :y.size(1)] = y + else: + x[mask, :y.size(1), :] = y + else: + x[mask] = y + return x diff --git a/fairseq/models/nat/nat_crf_transformer.py b/fairseq/models/nat/nat_crf_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..8dd3a08f72abf81800ff9f5c2c46422e2d865f2d --- /dev/null +++ b/fairseq/models/nat/nat_crf_transformer.py @@ -0,0 +1,107 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.models.nat import NATransformerModel, base_architecture +from fairseq.models import register_model, register_model_architecture +from fairseq.modules import DynamicCRF + + +@register_model("nacrf_transformer") +class NACRFTransformerModel(NATransformerModel): + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + self.crf_layer = DynamicCRF( + num_embedding=len(self.tgt_dict), + low_rank=args.crf_lowrank_approx, + beam_size=args.crf_beam_approx + ) + + @property + def allow_ensemble(self): + return False + + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + parser.add_argument("--crf-lowrank-approx", type=int, + help="the dimension of low-rank approximation of transition") + parser.add_argument("--crf-beam-approx", type=int, + help="the beam size for apporixmating the normalizing factor") + parser.add_argument("--word-ins-loss-factor", type=float, + help="weights on NAT loss used to co-training with CRF loss.") + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length(normalize=False, encoder_out=encoder_out) + length_tgt = self.decoder.forward_length_prediction(length_out, encoder_out, tgt_tokens) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out) + word_ins_tgt, word_ins_mask = tgt_tokens, tgt_tokens.ne(self.pad) + + # compute the log-likelihood of CRF + crf_nll = -self.crf_layer(word_ins_out, word_ins_tgt, word_ins_mask) + crf_nll = (crf_nll / word_ins_mask.type_as(crf_nll).sum(-1)).mean() + + return { + "word_ins": { + "out": word_ins_out, "tgt": word_ins_tgt, + "mask": word_ins_mask, "ls": self.args.label_smoothing, + "nll_loss": True, "factor": self.args.word_ins_loss_factor + }, + "word_crf": { + "loss": crf_nll + }, + "length": { + "out": length_out, "tgt": length_tgt, + "factor": self.decoder.length_loss_factor + } + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder and get emission scores + output_masks = output_tokens.ne(self.pad) + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=output_tokens, + encoder_out=encoder_out + ) + + # run viterbi decoding through CRF + _scores, _tokens = self.crf_layer.forward_decoder(word_ins_out, output_masks) + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history + ) + + +@register_model_architecture("nacrf_transformer", "nacrf_transformer") +def nacrf_base_architecture(args): + args.crf_lowrank_approx = getattr(args, "crf_lowrank_approx", 32) + args.crf_beam_approx = getattr(args, "crf_beam_approx", 64) + args.word_ins_loss_factor = getattr(args, "word_ins_loss_factor", 0.5) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + base_architecture(args) diff --git a/fairseq/models/nat/nonautoregressive_ensembles.py b/fairseq/models/nat/nonautoregressive_ensembles.py new file mode 100644 index 0000000000000000000000000000000000000000..2ed4d956e027c9776dcd00bf94cd14af9ece1c39 --- /dev/null +++ b/fairseq/models/nat/nonautoregressive_ensembles.py @@ -0,0 +1,231 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F + +from fairseq.models.nat import ( + _fill, + _skip, + _skip_encoder_out, + _apply_ins_masks, + _apply_ins_words, + _apply_del_words, +) + + +class _EnsembleModelEncoder(object): + def __init__(self, models): + self.models = models + + def reorder_encoder_out(self, encoder_outs, new_order): + encoder_outs = [ + model.encoder.reorder_encoder_out(encoder_out, new_order) + for model, encoder_out in zip(self.models, encoder_outs) + ] + return encoder_outs + + +class BasicEnsembleModel(torch.nn.Module): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__() + self.models = torch.nn.ModuleList(models) + self.bos = self.models[0].decoder.dictionary.bos() + self.eos = self.models[0].decoder.dictionary.eos() + self.pad = self.models[0].decoder.dictionary.pad() + self.unk = self.models[0].decoder.dictionary.unk() + self.encoder = _EnsembleModelEncoder(self.models) + + def has_encoder(self): + return hasattr(self.models[0], 'encoder') + + def max_decoder_positions(self): + return min(m.max_decoder_positions() for m in self.models) + + @torch.no_grad() + def forward_encoder(self, encoder_input): + if not self.has_encoder(): + return None + return [model.forward_encoder(encoder_input) for model in self.models] + + @torch.no_grad() + def forward_decoder(self, *inputs): + raise NotImplementedError + + def initialize_output_tokens(self, *inputs): + raise NotImplementedError + + +class EnsembleLevT(BasicEnsembleModel): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__(models) + + @torch.no_grad() + def forward_decoder(self, decoder_out, encoder_outs, eos_penalty=0.0, max_ratio=None, **kwargs): + # LevT ensembling + # A pipeline of three steps: deletion, placeholder, and word insertion. + # We need to average scores in each step in a pipeline way because of dependence. + # deletion + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + attn = decoder_out.attn + + bsz = output_tokens.size(0) + if max_ratio is None: + max_lens = output_tokens.new().fill_(255) + else: + if encoder_outs[0].encoder_padding_mask is None: + src_lens = encoder_outs[0].encoder_out.new(bsz).fill_(encoder_outs[0].encoder_out.size(1)) + else: + src_lens = (~encoder_outs[0].encoder_padding_mask).sum(1) + max_lens = (src_lens * max_ratio).clamp(min=10).long() + + # delete words + # do not delete tokens if it is + can_del_word = output_tokens.ne(self.pad).sum(1) > 2 + if can_del_word.sum() != 0: # we cannot delete, skip + output_tokens, output_scores, attn = self.forward_word_del( + encoder_outs, + output_tokens, + output_scores, + attn, + can_del_word, + ) + + # insert placeholders + can_ins_mask = output_tokens.ne(self.pad).sum(1) < max_lens + if can_ins_mask.sum() != 0: + output_tokens, output_scores = self.forward_mask_ins( + encoder_outs, + output_tokens, + output_scores, + can_ins_mask, + eos_penalty, + max_lens, + ) + + # insert words + can_ins_word = output_tokens.eq(self.unk).sum(1) > 0 + if can_ins_word.sum() != 0: + output_tokens, output_scores, attn = self.forward_word_ins( + encoder_outs, + output_tokens, + output_scores, + attn, + can_ins_word, + ) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + attn = None if attn is None else attn[:, :cut_off, :] + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=attn, + history=None + ) + + def forward_word_del(self, encoder_outs, output_tokens, output_scores, attn, can_del_word): + word_del_score_avg = [] + word_del_attn_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + word_del_out, word_del_attn = model.decoder.forward_word_del( + _skip(output_tokens, can_del_word), + _skip_encoder_out(model.encoder, encoder_out, can_del_word), + ) + word_del_score = F.log_softmax(word_del_out, 2) + word_del_score_avg.append(word_del_score) + word_del_attn_avg.append(word_del_attn) + word_del_score_avg = torch.logsumexp(torch.stack(word_del_score_avg, dim=0), dim=0) - math.log(len(self.models)) + word_del_pred = word_del_score_avg.max(-1)[1].bool() + if word_del_attn_avg[0] is not None: + word_del_attn_avg = torch.stack(word_del_attn_avg, dim=0)/len(self.models) + else: + word_del_attn_avg = None + + _tokens, _scores, _attn = _apply_del_words( + output_tokens[can_del_word], + output_scores[can_del_word], + word_del_attn_avg, + word_del_pred, + self.pad, + self.bos, + self.eos, + ) + output_tokens = _fill(output_tokens, can_del_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_del_word, _scores, 0) + attn = _fill(attn, can_del_word, _attn, 0.) + return output_tokens, output_scores, attn + + def forward_mask_ins(self, encoder_outs, output_tokens, output_scores, can_ins_mask, eos_penalty, max_lens): + mask_ins_score_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + mask_ins_out, _ = model.decoder.forward_mask_ins( + _skip(output_tokens, can_ins_mask), + _skip_encoder_out(model.encoder, encoder_out, can_ins_mask), + ) + mask_ins_score = F.log_softmax(mask_ins_out, 2) + if eos_penalty > 0.0: + mask_ins_score[:, :, 0] -= eos_penalty + mask_ins_score_avg.append(mask_ins_score) + mask_ins_score_avg = torch.logsumexp(torch.stack(mask_ins_score_avg, dim=0), dim=0) - math.log(len(self.models)) + mask_ins_pred = mask_ins_score_avg.max(-1)[1] + mask_ins_pred = torch.min( + mask_ins_pred, max_lens[can_ins_mask, None].expand_as(mask_ins_pred) + ) + _tokens, _scores = _apply_ins_masks( + output_tokens[can_ins_mask], + output_scores[can_ins_mask], + mask_ins_pred, + self.pad, + self.unk, + self.eos, + ) + output_tokens = _fill(output_tokens, can_ins_mask, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_mask, _scores, 0) + return output_tokens, output_scores + + def forward_word_ins(self, encoder_outs, output_tokens, output_scores, attn, can_ins_word): + word_ins_score_avg = [] + word_ins_attn_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + word_ins_out, word_ins_attn = model.decoder.forward_word_ins( + _skip(output_tokens, can_ins_word), + _skip_encoder_out(model.encoder, encoder_out, can_ins_word), + ) + word_ins_score = F.log_softmax(word_ins_out, 2) + word_ins_score_avg.append(word_ins_score) + word_ins_attn_avg.append(word_ins_attn) + word_ins_score_avg = torch.logsumexp(torch.stack(word_ins_score_avg, dim=0), dim=0) - math.log(len(self.models)) + if word_ins_attn_avg[0] is not None: + word_ins_attn_avg = torch.stack(word_ins_attn_avg, dim=0)/len(self.models) + else: + word_ins_attn_avg = None + word_ins_score_max, word_ins_pred = word_ins_score_avg.max(-1) + + _tokens, _scores = _apply_ins_words( + output_tokens[can_ins_word], + output_scores[can_ins_word], + word_ins_pred, + word_ins_score_max, + self.unk, + ) + + output_tokens = _fill(output_tokens, can_ins_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_word, _scores, 0) + attn = _fill(attn, can_ins_word, word_ins_attn, 0.) + return output_tokens, output_scores, attn + + def initialize_output_tokens(self, encoder_outs, src_tokens): + # LevT doesn't do length prediction. + return self.models[0].initialize_output_tokens(encoder_outs[0], src_tokens) diff --git a/fairseq/models/nat/nonautoregressive_transformer.py b/fairseq/models/nat/nonautoregressive_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..050755c30801120440d28a917f42877d3726dcf6 --- /dev/null +++ b/fairseq/models/nat/nonautoregressive_transformer.py @@ -0,0 +1,424 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F + +from fairseq import utils +from fairseq.iterative_refinement_generator import DecoderOut +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import Embedding + +from fairseq.models.nat import ( + FairseqNATModel, + FairseqNATDecoder, + ensemble_decoder +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +def _mean_pooling(enc_feats, src_masks): + # enc_feats: T x B x C + # src_masks: B x T or None + if src_masks is None: + enc_feats = enc_feats.mean(0) + else: + src_masks = (~src_masks).transpose(0, 1).type_as(enc_feats) + enc_feats = ( + (enc_feats / src_masks.sum(0)[None, :, None]) * src_masks[:, :, None] + ).sum(0) + return enc_feats + + +def _argmax(x, dim): + return (x == x.max(dim, keepdim=True)[0]).type_as(x) + + +def _uniform_assignment(src_lens, trg_lens): + max_trg_len = trg_lens.max() + steps = (src_lens.float() - 1) / (trg_lens.float() - 1) # step-size + # max_trg_len + index_t = utils.new_arange(trg_lens, max_trg_len).float() + index_t = steps[:, None] * index_t[None, :] # batch_size X max_trg_len + index_t = torch.round(index_t).long().detach() + return index_t + + +@register_model("nonautoregressive_transformer") +class NATransformerModel(FairseqNATModel): + + @property + def allow_length_beam(self): + return True + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + + # length prediction + parser.add_argument("--src-embedding-copy", action="store_true", + help="copy encoder word embeddings as the initial input of the decoder") + parser.add_argument("--pred-length-offset", action="store_true", + help="predicting the length difference between the target and source sentences") + parser.add_argument("--sg-length-pred", action="store_true", + help="stop the gradients back-propagated from the length predictor") + parser.add_argument("--length-loss-factor", type=float, + help="weights on the length prediction loss") + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = NATransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length(normalize=False, encoder_out=encoder_out) + length_tgt = self.decoder.forward_length_prediction(length_out, encoder_out, tgt_tokens) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out) + + return { + "word_ins": { + "out": word_ins_out, "tgt": tgt_tokens, + "mask": tgt_tokens.ne(self.pad), "ls": self.args.label_smoothing, + "nll_loss": True + }, + "length": { + "out": length_out, "tgt": length_tgt, + "factor": self.decoder.length_loss_factor + } + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + step = decoder_out.step + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder + output_masks = output_tokens.ne(self.pad) + _scores, _tokens = self.decoder( + normalize=True, + prev_output_tokens=output_tokens, + encoder_out=encoder_out, + step=step, + ).max(-1) + + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history + ) + + def initialize_output_tokens(self, encoder_out, src_tokens): + # length prediction + length_tgt = self.decoder.forward_length_prediction( + self.decoder.forward_length(normalize=True, encoder_out=encoder_out), + encoder_out=encoder_out + ) + + max_length = length_tgt.clamp_(min=2).max() + idx_length = utils.new_arange(src_tokens, max_length) + + initial_output_tokens = src_tokens.new_zeros( + src_tokens.size(0), max_length + ).fill_(self.pad) + initial_output_tokens.masked_fill_( + idx_length[None, :] < length_tgt[:, None], self.unk + ) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens.scatter_(1, length_tgt[:, None] - 1, self.eos) + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(encoder_out.encoder_out) + + return DecoderOut( + output_tokens=initial_output_tokens, + output_scores=initial_output_scores, + attn=None, + step=0, + max_step=0, + history=None + ) + + def regenerate_length_beam(self, decoder_out, beam_size): + output_tokens = decoder_out.output_tokens + length_tgt = output_tokens.ne(self.pad).sum(1) + length_tgt = length_tgt[:, None] + utils.new_arange(length_tgt, 1, beam_size) - beam_size // 2 + length_tgt = length_tgt.view(-1).clamp_(min=2) + max_length = length_tgt.max() + idx_length = utils.new_arange(length_tgt, max_length) + + initial_output_tokens = output_tokens.new_zeros( + length_tgt.size(0), max_length + ).fill_(self.pad) + initial_output_tokens.masked_fill_( + idx_length[None, :] < length_tgt[:, None], self.unk + ) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens.scatter_(1, length_tgt[:, None] - 1, self.eos) + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(decoder_out.output_scores) + + return decoder_out._replace( + output_tokens=initial_output_tokens, + output_scores=initial_output_scores + ) + + +class NATransformerDecoder(FairseqNATDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + + self.encoder_embed_dim = args.encoder_embed_dim + self.sg_length_pred = getattr(args, "sg_length_pred", False) + self.pred_length_offset = getattr(args, "pred_length_offset", False) + self.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + self.src_embedding_copy = getattr(args, "src_embedding_copy", False) + self.embed_length = Embedding(256, self.encoder_embed_dim, None) + + @ensemble_decoder + def forward(self, normalize, encoder_out, prev_output_tokens, step=0, **unused): + features, _ = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + embedding_copy=(step == 0) & self.src_embedding_copy, + ) + decoder_out = self.output_layer(features) + return F.log_softmax(decoder_out, -1) if normalize else decoder_out + + @ensemble_decoder + def forward_length(self, normalize, encoder_out): + enc_feats = encoder_out.encoder_out # T x B x C + src_masks = encoder_out.encoder_padding_mask # B x T or None + enc_feats = _mean_pooling(enc_feats, src_masks) + if self.sg_length_pred: + enc_feats = enc_feats.detach() + length_out = F.linear(enc_feats, self.embed_length.weight) + return F.log_softmax(length_out, -1) if normalize else length_out + + def extract_features( + self, + prev_output_tokens, + encoder_out=None, + early_exit=None, + embedding_copy=False, + **unused + ): + """ + Similar to *forward* but only return features. + + Inputs: + prev_output_tokens: Tensor(B, T) + encoder_out: a dictionary of hidden states and masks + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + the LevenshteinTransformer decoder has full-attention to all generated tokens + """ + # embedding + if embedding_copy: + src_embd = encoder_out.encoder_embedding + src_mask = encoder_out.encoder_padding_mask + src_mask = ( + ~src_mask + if src_mask is not None + else prev_output_tokens.new_ones(*src_embd.size()[:2]).bool() + ) + + x, decoder_padding_mask = self.forward_embedding( + prev_output_tokens, + self.forward_copying_source( + src_embd, src_mask, prev_output_tokens.ne(self.padding_idx) + ), + ) + + else: + + x, decoder_padding_mask = self.forward_embedding(prev_output_tokens) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + inner_states = [x] + + # decoder layers + for i, layer in enumerate(self.layers): + + # early exit from the decoder. + if (early_exit is not None) and (i >= early_exit): + break + + x, attn, _ = layer( + x, + encoder_out.encoder_out if encoder_out is not None else None, + encoder_out.encoder_padding_mask if encoder_out is not None else None, + self_attn_mask=None, + self_attn_padding_mask=decoder_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": attn, "inner_states": inner_states} + + def forward_embedding(self, prev_output_tokens, states=None): + # embed positions + positions = ( + self.embed_positions(prev_output_tokens) + if self.embed_positions is not None + else None + ) + + # embed tokens and positions + if states is None: + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + if self.project_in_dim is not None: + x = self.project_in_dim(x) + else: + x = states + + if positions is not None: + x += positions + x = self.dropout_module(x) + decoder_padding_mask = prev_output_tokens.eq(self.padding_idx) + return x, decoder_padding_mask + + def forward_copying_source(self, src_embeds, src_masks, tgt_masks): + length_sources = src_masks.sum(1) + length_targets = tgt_masks.sum(1) + mapped_inputs = _uniform_assignment(length_sources, length_targets).masked_fill( + ~tgt_masks, 0 + ) + copied_embedding = torch.gather( + src_embeds, + 1, + mapped_inputs.unsqueeze(-1).expand( + *mapped_inputs.size(), src_embeds.size(-1) + ), + ) + return copied_embedding + + def forward_length_prediction(self, length_out, encoder_out, tgt_tokens=None): + enc_feats = encoder_out.encoder_out # T x B x C + src_masks = encoder_out.encoder_padding_mask # B x T or None + if self.pred_length_offset: + if src_masks is None: + src_lengs = enc_feats.new_ones(enc_feats.size(1)).fill_( + enc_feats.size(0) + ) + else: + src_lengs = (~src_masks).transpose(0, 1).type_as(enc_feats).sum(0) + src_lengs = src_lengs.long() + + if tgt_tokens is not None: + # obtain the length target + tgt_lengs = tgt_tokens.ne(self.padding_idx).sum(1).long() + if self.pred_length_offset: + length_tgt = tgt_lengs - src_lengs + 128 + else: + length_tgt = tgt_lengs + length_tgt = length_tgt.clamp(min=0, max=255) + + else: + # predict the length target (greedy for now) + # TODO: implementing length-beam + pred_lengs = length_out.max(-1)[1] + if self.pred_length_offset: + length_tgt = pred_lengs - 128 + src_lengs + else: + length_tgt = pred_lengs + + return length_tgt + + +@register_model_architecture( + "nonautoregressive_transformer", "nonautoregressive_transformer" +) +def base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + +@register_model_architecture( + "nonautoregressive_transformer", "nonautoregressive_transformer_wmt_en_de" +) +def nonautoregressive_transformer_wmt_en_de(args): + base_architecture(args) diff --git a/fairseq/models/roberta/__init__.py b/fairseq/models/roberta/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..56579e591566e014d99ed5a283ee7135257f054c --- /dev/null +++ b/fairseq/models/roberta/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .hub_interface import * # noqa +from .model import * # noqa +from .model_camembert import * # noqa +from .model_xlmr import * # noqa diff --git a/fairseq/models/roberta/__pycache__/__init__.cpython-310.pyc b/fairseq/models/roberta/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3f0ccdcd0c6862f127dde8b32c8883783b9c076c Binary files /dev/null and b/fairseq/models/roberta/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/models/roberta/__pycache__/hub_interface.cpython-310.pyc b/fairseq/models/roberta/__pycache__/hub_interface.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..01162df29f1fbaa1b64ef1468c9cd91cced9fcde Binary files /dev/null and b/fairseq/models/roberta/__pycache__/hub_interface.cpython-310.pyc differ diff --git a/fairseq/models/roberta/__pycache__/model.cpython-310.pyc b/fairseq/models/roberta/__pycache__/model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a0554d1b4ab3ef0cddb697cdc9b4bb8de83cce92 Binary files /dev/null and b/fairseq/models/roberta/__pycache__/model.cpython-310.pyc differ diff --git a/fairseq/models/roberta/__pycache__/model_camembert.cpython-310.pyc b/fairseq/models/roberta/__pycache__/model_camembert.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a4faeb2e409e3f360079b868c094f9bbcb02b503 Binary files /dev/null and b/fairseq/models/roberta/__pycache__/model_camembert.cpython-310.pyc differ diff --git a/fairseq/models/roberta/__pycache__/model_xlmr.cpython-310.pyc b/fairseq/models/roberta/__pycache__/model_xlmr.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cbf3e7f900c264705adcbf5489d2d6aafa739c6c Binary files /dev/null and b/fairseq/models/roberta/__pycache__/model_xlmr.cpython-310.pyc differ diff --git a/fairseq/models/roberta/alignment_utils.py b/fairseq/models/roberta/alignment_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..45d2e37194c0f66e2b063884d7f3291ae48ece0f --- /dev/null +++ b/fairseq/models/roberta/alignment_utils.py @@ -0,0 +1,115 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import Counter +from typing import List + +import torch + + +def align_bpe_to_words(roberta, bpe_tokens: torch.LongTensor, other_tokens: List[str]): + """ + Helper to align GPT-2 BPE to other tokenization formats (e.g., spaCy). + + Args: + roberta (RobertaHubInterface): RoBERTa instance + bpe_tokens (torch.LongTensor): GPT-2 BPE tokens of shape `(T_bpe)` + other_tokens (List[str]): other tokens of shape `(T_words)` + + Returns: + List[str]: mapping from *other_tokens* to corresponding *bpe_tokens*. + """ + assert bpe_tokens.dim() == 1 + assert bpe_tokens[0] == 0 + + def clean(text): + return text.strip() + + # remove whitespaces to simplify alignment + bpe_tokens = [roberta.task.source_dictionary.string([x]) for x in bpe_tokens] + bpe_tokens = [clean(roberta.bpe.decode(x) if x not in {'', ''} else x) for x in bpe_tokens] + other_tokens = [clean(str(o)) for o in other_tokens] + + # strip leading + bpe_tokens = bpe_tokens[1:] + assert ''.join(bpe_tokens) == ''.join(other_tokens) + + # create alignment from every word to a list of BPE tokens + alignment = [] + bpe_toks = filter(lambda item: item[1] != '', enumerate(bpe_tokens, start=1)) + j, bpe_tok = next(bpe_toks) + for other_tok in other_tokens: + bpe_indices = [] + while True: + if other_tok.startswith(bpe_tok): + bpe_indices.append(j) + other_tok = other_tok[len(bpe_tok):] + try: + j, bpe_tok = next(bpe_toks) + except StopIteration: + j, bpe_tok = None, None + elif bpe_tok.startswith(other_tok): + # other_tok spans multiple BPE tokens + bpe_indices.append(j) + bpe_tok = bpe_tok[len(other_tok):] + other_tok = '' + else: + raise Exception('Cannot align "{}" and "{}"'.format(other_tok, bpe_tok)) + if other_tok == '': + break + assert len(bpe_indices) > 0 + alignment.append(bpe_indices) + assert len(alignment) == len(other_tokens) + + return alignment + + +def align_features_to_words(roberta, features, alignment): + """ + Align given features to words. + + Args: + roberta (RobertaHubInterface): RoBERTa instance + features (torch.Tensor): features to align of shape `(T_bpe x C)` + alignment: alignment between BPE tokens and words returned by + func:`align_bpe_to_words`. + """ + assert features.dim() == 2 + + bpe_counts = Counter(j for bpe_indices in alignment for j in bpe_indices) + assert bpe_counts[0] == 0 # shouldn't be aligned + denom = features.new([bpe_counts.get(j, 1) for j in range(len(features))]) + weighted_features = features / denom.unsqueeze(-1) + + output = [weighted_features[0]] + largest_j = -1 + for bpe_indices in alignment: + output.append(weighted_features[bpe_indices].sum(dim=0)) + largest_j = max(largest_j, *bpe_indices) + for j in range(largest_j + 1, len(features)): + output.append(weighted_features[j]) + output = torch.stack(output) + assert torch.all(torch.abs(output.sum(dim=0) - features.sum(dim=0)) < 1e-4) + return output + + +def spacy_nlp(): + if getattr(spacy_nlp, '_nlp', None) is None: + try: + from spacy.lang.en import English + spacy_nlp._nlp = English() + except ImportError: + raise ImportError('Please install spacy with: pip install spacy') + return spacy_nlp._nlp + + +def spacy_tokenizer(): + if getattr(spacy_tokenizer, '_tokenizer', None) is None: + try: + nlp = spacy_nlp() + spacy_tokenizer._tokenizer = nlp.Defaults.create_tokenizer(nlp) + except ImportError: + raise ImportError('Please install spacy with: pip install spacy') + return spacy_tokenizer._tokenizer diff --git a/fairseq/models/roberta/hub_interface.py b/fairseq/models/roberta/hub_interface.py new file mode 100644 index 0000000000000000000000000000000000000000..baf0bf28b9ea4a3b15e042e9f4f86f1ffd499f7c --- /dev/null +++ b/fairseq/models/roberta/hub_interface.py @@ -0,0 +1,204 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from fairseq.data import encoders + + +class RobertaHubInterface(nn.Module): + """A simple PyTorch Hub interface to RoBERTa. + + Usage: https://github.com/pytorch/fairseq/tree/master/examples/roberta + """ + + def __init__(self, args, task, model): + super().__init__() + self.args = args + self.task = task + self.model = model + + self.bpe = encoders.build_bpe(args) + + # this is useful for determining the device + self.register_buffer('_float_tensor', torch.tensor([0], dtype=torch.float)) + + @property + def device(self): + return self._float_tensor.device + + def encode(self, sentence: str, *addl_sentences, no_separator=False) -> torch.LongTensor: + """ + BPE-encode a sentence (or multiple sentences). + + Every sequence begins with a beginning-of-sentence (``) symbol. + Every sentence ends with an end-of-sentence (``) and we use an + extra end-of-sentence (``) as a separator. + + Example (single sentence): ` a b c ` + Example (sentence pair): ` d e f 1 2 3 ` + + The BPE encoding follows GPT-2. One subtle detail is that the GPT-2 BPE + requires leading spaces. For example:: + + >>> roberta.encode('Hello world').tolist() + [0, 31414, 232, 2] + >>> roberta.encode(' world').tolist() + [0, 232, 2] + >>> roberta.encode('world').tolist() + [0, 8331, 2] + """ + bpe_sentence = ' ' + self.bpe.encode(sentence) + ' ' + for s in addl_sentences: + bpe_sentence += (' ' if not no_separator else '') + bpe_sentence += ' ' + self.bpe.encode(s) + ' ' + tokens = self.task.source_dictionary.encode_line(bpe_sentence, append_eos=False, add_if_not_exist=False) + return tokens.long() + + def decode(self, tokens: torch.LongTensor): + assert tokens.dim() == 1 + tokens = tokens.numpy() + if tokens[0] == self.task.source_dictionary.bos(): + tokens = tokens[1:] # remove + eos_mask = (tokens == self.task.source_dictionary.eos()) + doc_mask = eos_mask[1:] & eos_mask[:-1] + sentences = np.split(tokens, doc_mask.nonzero()[0] + 1) + sentences = [self.bpe.decode(self.task.source_dictionary.string(s)) for s in sentences] + if len(sentences) == 1: + return sentences[0] + return sentences + + def extract_features(self, tokens: torch.LongTensor, return_all_hiddens: bool = False) -> torch.Tensor: + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + if tokens.size(-1) > self.model.max_positions(): + raise ValueError('tokens exceeds maximum length: {} > {}'.format( + tokens.size(-1), self.model.max_positions() + )) + features, extra = self.model( + tokens.to(device=self.device), + features_only=True, + return_all_hiddens=return_all_hiddens, + ) + if return_all_hiddens: + # convert from T x B x C -> B x T x C + inner_states = extra['inner_states'] + return [inner_state.transpose(0, 1) for inner_state in inner_states] + else: + return features # just the last layer's features + + def register_classification_head( + self, name: str, num_classes: int = None, embedding_size: int = None, **kwargs + ): + self.model.register_classification_head( + name, num_classes=num_classes, embedding_size=embedding_size, **kwargs + ) + + def predict(self, head: str, tokens: torch.LongTensor, return_logits: bool = False): + features = self.extract_features(tokens.to(device=self.device)) + logits = self.model.classification_heads[head](features) + if return_logits: + return logits + return F.log_softmax(logits, dim=-1) + + def extract_features_aligned_to_words(self, sentence: str, return_all_hiddens: bool = False) -> torch.Tensor: + """Extract RoBERTa features, aligned to spaCy's word-level tokenizer.""" + from fairseq.models.roberta import alignment_utils + from spacy.tokens import Doc + + nlp = alignment_utils.spacy_nlp() + tokenizer = alignment_utils.spacy_tokenizer() + + # tokenize both with GPT-2 BPE and spaCy + bpe_toks = self.encode(sentence) + spacy_toks = tokenizer(sentence) + spacy_toks_ws = [t.text_with_ws for t in tokenizer(sentence)] + alignment = alignment_utils.align_bpe_to_words(self, bpe_toks, spacy_toks_ws) + + # extract features and align them + features = self.extract_features(bpe_toks, return_all_hiddens=return_all_hiddens) + features = features.squeeze(0) + aligned_feats = alignment_utils.align_features_to_words(self, features, alignment) + + # wrap in spaCy Doc + doc = Doc( + nlp.vocab, + words=[''] + [x.text for x in spacy_toks] + [''], + spaces=[True] + [x.endswith(' ') for x in spacy_toks_ws[:-1]] + [True, False], + ) + assert len(doc) == aligned_feats.size(0) + doc.user_token_hooks['vector'] = lambda token: aligned_feats[token.i] + return doc + + def fill_mask(self, masked_input: str, topk: int = 5): + masked_token = '' + assert masked_token in masked_input and masked_input.count(masked_token) == 1, \ + "Please add one {0} token for the input, eg: 'He is a {0} guy'".format(masked_token) + + text_spans = masked_input.split(masked_token) + text_spans_bpe = (' {0} '.format(masked_token)).join( + [self.bpe.encode(text_span.rstrip()) for text_span in text_spans] + ).strip() + tokens = self.task.source_dictionary.encode_line( + ' ' + text_spans_bpe + ' ', + append_eos=False, + add_if_not_exist=False, + ) + + masked_index = (tokens == self.task.mask_idx).nonzero() + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + + with utils.eval(self.model): + features, extra = self.model( + tokens.long().to(device=self.device), + features_only=False, + return_all_hiddens=False, + ) + logits = features[0, masked_index, :].squeeze() + prob = logits.softmax(dim=0) + values, index = prob.topk(k=topk, dim=0) + topk_predicted_token_bpe = self.task.source_dictionary.string(index) + + topk_filled_outputs = [] + for index, predicted_token_bpe in enumerate(topk_predicted_token_bpe.split(' ')): + predicted_token = self.bpe.decode(predicted_token_bpe) + # Quick hack to fix https://github.com/pytorch/fairseq/issues/1306 + if predicted_token_bpe.startswith('\u2581'): + predicted_token = ' ' + predicted_token + if " {0}".format(masked_token) in masked_input: + topk_filled_outputs.append(( + masked_input.replace( + ' {0}'.format(masked_token), predicted_token + ), + values[index].item(), + predicted_token, + )) + else: + topk_filled_outputs.append(( + masked_input.replace(masked_token, predicted_token), + values[index].item(), + predicted_token, + )) + return topk_filled_outputs + + def disambiguate_pronoun(self, sentence: str) -> bool: + """ + Usage:: + + >>> disambiguate_pronoun('The _trophy_ would not fit in the brown suitcase because [it] was too big.') + True + + >>> disambiguate_pronoun('The trophy would not fit in the brown suitcase because [it] was too big.') + 'The trophy' + """ + assert hasattr(self.task, 'disambiguate_pronoun'), \ + 'roberta.disambiguate_pronoun() requires a model trained with the WSC task.' + with utils.eval(self.model): + return self.task.disambiguate_pronoun(self.model, sentence, use_cuda=self.device.type == 'cuda') diff --git a/fairseq/models/roberta/model.py b/fairseq/models/roberta/model.py new file mode 100644 index 0000000000000000000000000000000000000000..2303fbe26e1882ab38784f4064abb5af8bf099a5 --- /dev/null +++ b/fairseq/models/roberta/model.py @@ -0,0 +1,396 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +RoBERTa: A Robustly Optimized BERT Pretraining Approach. +""" + +import logging + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderModel, + register_model, + register_model_architecture, +) +from fairseq.modules import ( + LayerNorm, + TransformerSentenceEncoder, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ + +from .hub_interface import RobertaHubInterface + + +logger = logging.getLogger(__name__) + + +@register_model('roberta') +class RobertaModel(FairseqEncoderModel): + + @classmethod + def hub_models(cls): + return { + 'roberta.base': 'http://dl.fbaipublicfiles.com/fairseq/models/roberta.base.tar.gz', + 'roberta.large': 'http://dl.fbaipublicfiles.com/fairseq/models/roberta.large.tar.gz', + 'roberta.large.mnli': 'http://dl.fbaipublicfiles.com/fairseq/models/roberta.large.mnli.tar.gz', + 'roberta.large.wsc': 'http://dl.fbaipublicfiles.com/fairseq/models/roberta.large.wsc.tar.gz', + } + + def __init__(self, args, encoder): + super().__init__(encoder) + self.args = args + + # We follow BERT's random weight initialization + self.apply(init_bert_params) + + self.classification_heads = nn.ModuleDict() + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument('--encoder-layers', type=int, metavar='L', + help='num encoder layers') + parser.add_argument('--encoder-embed-dim', type=int, metavar='H', + help='encoder embedding dimension') + parser.add_argument('--encoder-ffn-embed-dim', type=int, metavar='F', + help='encoder embedding dimension for FFN') + parser.add_argument('--encoder-attention-heads', type=int, metavar='A', + help='num encoder attention heads') + parser.add_argument('--activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use') + parser.add_argument('--pooler-activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use for pooler layer') + parser.add_argument('--encoder-normalize-before', action='store_true', + help='apply layernorm before each encoder block') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + parser.add_argument('--activation-dropout', type=float, metavar='D', + help='dropout probability after activation in FFN') + parser.add_argument('--pooler-dropout', type=float, metavar='D', + help='dropout probability in the masked_lm pooler layers') + parser.add_argument('--max-positions', type=int, + help='number of positional embeddings to learn') + parser.add_argument('--load-checkpoint-heads', action='store_true', + help='(re-)register and load heads when loading checkpoints') + # args for "Reducing Transformer Depth on Demand with Structured Dropout" (Fan et al., 2019) + parser.add_argument('--encoder-layerdrop', type=float, metavar='D', default=0, + help='LayerDrop probability for encoder') + parser.add_argument('--encoder-layers-to-keep', default=None, + help='which layers to *keep* when pruning as a comma-separated list') + # args for Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020) + parser.add_argument('--quant-noise-pq', type=float, metavar='D', default=0, + help='iterative PQ quantization noise at training time') + parser.add_argument('--quant-noise-pq-block-size', type=int, metavar='D', default=8, + help='block size of quantization noise at training time') + parser.add_argument('--quant-noise-scalar', type=float, metavar='D', default=0, + help='scalar quantization noise and scalar quantization at training time') + parser.add_argument('--untie-weights-roberta', action='store_true', + help='Untie weights between embeddings and classifiers in RoBERTa') + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present + base_architecture(args) + + if not hasattr(args, 'max_positions'): + args.max_positions = args.tokens_per_sample + + encoder = RobertaEncoder(args, task.source_dictionary) + return cls(args, encoder) + + def forward(self, src_tokens, features_only=False, return_all_hiddens=False, classification_head_name=None, **kwargs): + if classification_head_name is not None: + features_only = True + + x, extra = self.encoder(src_tokens, features_only, return_all_hiddens, **kwargs) + + if classification_head_name is not None: + x = self.classification_heads[classification_head_name](x) + return x, extra + + def get_normalized_probs(self, net_output, log_probs, sample=None): + """Get normalized probabilities (or log probs) from a net's output.""" + logits = net_output[0].float() + if log_probs: + return F.log_softmax(logits, dim=-1) + else: + return F.softmax(logits, dim=-1) + + def register_classification_head(self, name, num_classes=None, inner_dim=None, **kwargs): + """Register a classification head.""" + if name in self.classification_heads: + prev_num_classes = self.classification_heads[name].out_proj.out_features + prev_inner_dim = self.classification_heads[name].dense.out_features + if num_classes != prev_num_classes or inner_dim != prev_inner_dim: + logger.warning( + 're-registering head "{}" with num_classes {} (prev: {}) ' + 'and inner_dim {} (prev: {})'.format( + name, num_classes, prev_num_classes, inner_dim, prev_inner_dim + ) + ) + self.classification_heads[name] = RobertaClassificationHead( + self.args.encoder_embed_dim, + inner_dim or self.args.encoder_embed_dim, + num_classes, + self.args.pooler_activation_fn, + self.args.pooler_dropout, + self.args.quant_noise_pq, + self.args.quant_noise_pq_block_size, + ) + + @property + def supported_targets(self): + return {'self'} + + @classmethod + def from_pretrained(cls, model_name_or_path, checkpoint_file='model.pt', data_name_or_path='.', bpe='gpt2', **kwargs): + from fairseq import hub_utils + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + return RobertaHubInterface(x['args'], x['task'], x['models'][0]) + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + '.' if name != '' else '' + + # rename decoder -> encoder before upgrading children modules + for k in list(state_dict.keys()): + if k.startswith(prefix + 'decoder'): + new_k = prefix + 'encoder' + k[len(prefix + 'decoder'):] + state_dict[new_k] = state_dict[k] + del state_dict[k] + + # upgrade children modules + super().upgrade_state_dict_named(state_dict, name) + + # Handle new classification heads present in the state dict. + current_head_names = ( + [] if not hasattr(self, 'classification_heads') + else self.classification_heads.keys() + ) + keys_to_delete = [] + for k in state_dict.keys(): + if not k.startswith(prefix + 'classification_heads.'): + continue + + head_name = k[len(prefix + 'classification_heads.'):].split('.')[0] + num_classes = state_dict[prefix + 'classification_heads.' + head_name + '.out_proj.weight'].size(0) + inner_dim = state_dict[prefix + 'classification_heads.' + head_name + '.dense.weight'].size(0) + + if getattr(self.args, 'load_checkpoint_heads', False): + if head_name not in current_head_names: + self.register_classification_head(head_name, num_classes, inner_dim) + else: + if head_name not in current_head_names: + logger.warning( + 'deleting classification head ({}) from checkpoint ' + 'not present in current model: {}'.format(head_name, k) + ) + keys_to_delete.append(k) + elif ( + num_classes != self.classification_heads[head_name].out_proj.out_features + or inner_dim != self.classification_heads[head_name].dense.out_features + ): + logger.warning( + 'deleting classification head ({}) from checkpoint ' + 'with different dimensions than current model: {}'.format(head_name, k) + ) + keys_to_delete.append(k) + for k in keys_to_delete: + del state_dict[k] + + # Copy any newly-added classification heads into the state dict + # with their current weights. + if hasattr(self, 'classification_heads'): + cur_state = self.classification_heads.state_dict() + for k, v in cur_state.items(): + if prefix + 'classification_heads.' + k not in state_dict: + logger.info('Overwriting ' + prefix + 'classification_heads.' + k) + state_dict[prefix + 'classification_heads.' + k] = v + + +class RobertaLMHead(nn.Module): + """Head for masked language modeling.""" + + def __init__(self, embed_dim, output_dim, activation_fn, weight=None): + super().__init__() + self.dense = nn.Linear(embed_dim, embed_dim) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.layer_norm = LayerNorm(embed_dim) + + if weight is None: + weight = nn.Linear(embed_dim, output_dim, bias=False).weight + self.weight = weight + self.bias = nn.Parameter(torch.zeros(output_dim)) + + def forward(self, features, masked_tokens=None, **kwargs): + # Only project the masked tokens while training, + # saves both memory and computation + if masked_tokens is not None: + features = features[masked_tokens, :] + + x = self.dense(features) + x = self.activation_fn(x) + x = self.layer_norm(x) + # project back to size of vocabulary with bias + x = F.linear(x, self.weight) + self.bias + return x + + +class RobertaClassificationHead(nn.Module): + """Head for sentence-level classification tasks.""" + + def __init__(self, input_dim, inner_dim, num_classes, activation_fn, pooler_dropout, q_noise=0, qn_block_size=8): + super().__init__() + self.dense = nn.Linear(input_dim, inner_dim) + self.activation_fn = utils.get_activation_fn(activation_fn) + self.dropout = nn.Dropout(p=pooler_dropout) + self.out_proj = apply_quant_noise_( + nn.Linear(inner_dim, num_classes), q_noise, qn_block_size + ) + + def forward(self, features, **kwargs): + x = features[:, 0, :] # take token (equiv. to [CLS]) + x = self.dropout(x) + x = self.dense(x) + x = self.activation_fn(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +class RobertaEncoder(FairseqEncoder): + """RoBERTa encoder.""" + + def __init__(self, args, dictionary): + super().__init__(dictionary) + self.args = args + + if args.encoder_layers_to_keep: + args.encoder_layers = len(args.encoder_layers_to_keep.split(",")) + + self.sentence_encoder = TransformerSentenceEncoder( + padding_idx=dictionary.pad(), + vocab_size=len(dictionary), + num_encoder_layers=args.encoder_layers, + embedding_dim=args.encoder_embed_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=args.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + layerdrop=args.encoder_layerdrop, + max_seq_len=args.max_positions, + num_segments=0, + encoder_normalize_before=True, + apply_bert_init=True, + activation_fn=args.activation_fn, + q_noise=args.quant_noise_pq, + qn_block_size=args.quant_noise_pq_block_size, + ) + args.untie_weights_roberta = getattr(args, 'untie_weights_roberta', False) + + self.lm_head = RobertaLMHead( + embed_dim=args.encoder_embed_dim, + output_dim=len(dictionary), + activation_fn=args.activation_fn, + weight=self.sentence_encoder.embed_tokens.weight if not args.untie_weights_roberta else None, + ) + + def forward(self, src_tokens, features_only=False, return_all_hiddens=False, masked_tokens=None, **unused): + """ + Args: + src_tokens (LongTensor): input tokens of shape `(batch, src_len)` + features_only (bool, optional): skip LM head and just return + features. If True, the output will be of shape + `(batch, src_len, embed_dim)`. + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + + Returns: + tuple: + - the LM output of shape `(batch, src_len, vocab)` + - a dictionary of additional data, where 'inner_states' + is a list of hidden states. Note that the hidden + states have shape `(src_len, batch, vocab)`. + """ + x, extra = self.extract_features(src_tokens, return_all_hiddens=return_all_hiddens) + if not features_only: + x = self.output_layer(x, masked_tokens=masked_tokens) + return x, extra + + def extract_features(self, src_tokens, return_all_hiddens=False, **unused): + inner_states, _ = self.sentence_encoder( + src_tokens, + last_state_only=not return_all_hiddens, + ) + features = inner_states[-1].transpose(0, 1) # T x B x C -> B x T x C + return features, {'inner_states': inner_states if return_all_hiddens else None} + + def output_layer(self, features, masked_tokens=None, **unused): + return self.lm_head(features, masked_tokens) + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.args.max_positions + + +@register_model_architecture('roberta', 'roberta') +def base_architecture(args): + args.encoder_layers = getattr(args, 'encoder_layers', 12) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 768) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 3072) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 12) + + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + args.pooler_activation_fn = getattr(args, 'pooler_activation_fn', 'tanh') + + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_dropout = getattr(args, 'activation_dropout', 0.0) + args.pooler_dropout = getattr(args, 'pooler_dropout', 0.0) + args.encoder_layers_to_keep = getattr(args, 'encoder_layers_to_keep', None) + args.encoder_layerdrop = getattr(args, 'encoder_layerdrop', 0.0) + + +@register_model_architecture('roberta', 'roberta_base') +def roberta_base_architecture(args): + base_architecture(args) + + +@register_model_architecture('roberta', 'roberta_large') +def roberta_large_architecture(args): + args.encoder_layers = getattr(args, 'encoder_layers', 24) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1024) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 4096) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 16) + base_architecture(args) + + +@register_model_architecture('roberta', 'xlm') +def xlm_architecture(args): + args.encoder_layers = getattr(args, 'encoder_layers', 16) + args.encoder_embed_dim = getattr(args, 'encoder_embed_dim', 1280) + args.encoder_ffn_embed_dim = getattr(args, 'encoder_ffn_embed_dim', 1280*4) + args.encoder_attention_heads = getattr(args, 'encoder_attention_heads', 16) + base_architecture(args) diff --git a/fairseq/models/roberta/model_camembert.py b/fairseq/models/roberta/model_camembert.py new file mode 100644 index 0000000000000000000000000000000000000000..eb57d81d8df4f73d74dbdab5d877947cc39c03c1 --- /dev/null +++ b/fairseq/models/roberta/model_camembert.py @@ -0,0 +1,43 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +CamemBERT: a Tasty French Language Model +""" + +from fairseq.models import register_model + +from .hub_interface import RobertaHubInterface +from .model import RobertaModel + + +@register_model('camembert') +class CamembertModel(RobertaModel): + + @classmethod + def hub_models(cls): + return { + 'camembert': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz', + 'camembert.v0': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz', + 'camembert-base': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-base.tar.gz', + 'camembert-large': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-large.tar.gz', + 'camembert-base-ccnet': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet.tar.gz', + 'camembert-base-ccnet-4gb': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-ccnet-4gb.tar.gz', + 'camembert-base-wikipedia-4gb': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-wikipedia-4gb.tar.gz', + 'camembert-base-oscar-4gb': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert-base-oscar-4gb.tar.gz', + } + + @classmethod + def from_pretrained(cls, model_name_or_path, checkpoint_file='model.pt', data_name_or_path='.', bpe='sentencepiece', **kwargs): + from fairseq import hub_utils + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + return RobertaHubInterface(x['args'], x['task'], x['models'][0]) diff --git a/fairseq/models/roberta/model_xlmr.py b/fairseq/models/roberta/model_xlmr.py new file mode 100644 index 0000000000000000000000000000000000000000..fa71a27d12ad237012a65019e3c60669b8837055 --- /dev/null +++ b/fairseq/models/roberta/model_xlmr.py @@ -0,0 +1,37 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Unsupervised Cross-lingual Representation Learning at Scale +""" + +from fairseq.models import register_model + +from .hub_interface import RobertaHubInterface +from .model import RobertaModel + + +@register_model('xlmr') +class XLMRModel(RobertaModel): + + @classmethod + def hub_models(cls): + return { + 'xlmr.base': 'http://dl.fbaipublicfiles.com/fairseq/models/xlmr.base.tar.gz', + 'xlmr.large': 'http://dl.fbaipublicfiles.com/fairseq/models/xlmr.large.tar.gz', + } + + @classmethod + def from_pretrained(cls, model_name_or_path, checkpoint_file='model.pt', data_name_or_path='.', bpe='sentencepiece', **kwargs): + from fairseq import hub_utils + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + return RobertaHubInterface(x['args'], x['task'], x['models'][0]) diff --git a/fairseq/models/transformer.py b/fairseq/models/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..6fd5c2bd05de761ce8c1c8bdcd37dcc4c14c62ec --- /dev/null +++ b/fairseq/models/transformer.py @@ -0,0 +1,998 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Any, Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +from fairseq import options, utils +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) +from fairseq.models.fairseq_encoder import EncoderOut +from fairseq.modules import ( + AdaptiveSoftmax, + FairseqDropout, + LayerDropModuleList, + LayerNorm, + PositionalEmbedding, + SinusoidalPositionalEmbedding, + TransformerDecoderLayer, + TransformerEncoderLayer, +) +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ +from torch import Tensor + + +DEFAULT_MAX_SOURCE_POSITIONS = 1024 +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model("transformer") +class TransformerModel(FairseqEncoderDecoderModel): + """ + Transformer model from `"Attention Is All You Need" (Vaswani, et al, 2017) + `_. + + Args: + encoder (TransformerEncoder): the encoder + decoder (TransformerDecoder): the decoder + + The Transformer model provides the following named architectures and + command-line arguments: + + .. argparse:: + :ref: fairseq.models.transformer_parser + :prog: + """ + + @classmethod + def hub_models(cls): + # fmt: off + + def moses_subword(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'subword_nmt', + } + + def moses_fastbpe(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'fastbpe', + } + + return { + 'transformer.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-fr.joined-dict.transformer.tar.bz2'), + 'transformer.wmt16.en-de': 'https://dl.fbaipublicfiles.com/fairseq/models/wmt16.en-de.joined-dict.transformer.tar.bz2', + 'transformer.wmt18.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz'), + 'transformer.wmt19.en-de': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.ensemble.tar.gz'), + 'transformer.wmt19.en-ru': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.ensemble.tar.gz'), + 'transformer.wmt19.de-en': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.ensemble.tar.gz'), + 'transformer.wmt19.ru-en': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.ensemble.tar.gz'), + 'transformer.wmt19.en-de.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.single_model.tar.gz'), + 'transformer.wmt19.en-ru.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.single_model.tar.gz'), + 'transformer.wmt19.de-en.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.single_model.tar.gz'), + 'transformer.wmt19.ru-en.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.single_model.tar.gz'), + } + # fmt: on + + def __init__(self, args, encoder, decoder): + super().__init__(encoder, decoder) + self.args = args + self.supports_align_args = True + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + parser.add_argument('--activation-dropout', '--relu-dropout', type=float, metavar='D', + help='dropout probability after activation in FFN.') + parser.add_argument('--encoder-embed-path', type=str, metavar='STR', + help='path to pre-trained encoder embedding') + parser.add_argument('--encoder-embed-dim', type=int, metavar='N', + help='encoder embedding dimension') + parser.add_argument('--encoder-ffn-embed-dim', type=int, metavar='N', + help='encoder embedding dimension for FFN') + parser.add_argument('--encoder-layers', type=int, metavar='N', + help='num encoder layers') + parser.add_argument('--encoder-attention-heads', type=int, metavar='N', + help='num encoder attention heads') + parser.add_argument('--encoder-normalize-before', action='store_true', + help='apply layernorm before each encoder block') + parser.add_argument('--encoder-learned-pos', action='store_true', + help='use learned positional embeddings in the encoder') + parser.add_argument('--decoder-embed-path', type=str, metavar='STR', + help='path to pre-trained decoder embedding') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-ffn-embed-dim', type=int, metavar='N', + help='decoder embedding dimension for FFN') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='num decoder layers') + parser.add_argument('--decoder-attention-heads', type=int, metavar='N', + help='num decoder attention heads') + parser.add_argument('--decoder-learned-pos', action='store_true', + help='use learned positional embeddings in the decoder') + parser.add_argument('--decoder-normalize-before', action='store_true', + help='apply layernorm before each decoder block') + parser.add_argument('--decoder-output-dim', type=int, metavar='N', + help='decoder output dimension (extra linear layer ' + 'if different from decoder embed dim') + parser.add_argument('--share-decoder-input-output-embed', action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--share-all-embeddings', action='store_true', + help='share encoder, decoder and output embeddings' + ' (requires shared dictionary and embed dim)') + parser.add_argument('--no-token-positional-embeddings', default=False, action='store_true', + help='if set, disables positional embeddings (outside self attention)') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion'), + parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D', + help='sets adaptive softmax dropout for the tail projections') + parser.add_argument('--layernorm-embedding', action='store_true', + help='add layernorm to embedding') + parser.add_argument('--no-scale-embedding', action='store_true', + help='if True, dont scale embeddings') + # args for "Cross+Self-Attention for Transformer Models" (Peitz et al., 2019) + parser.add_argument('--no-cross-attention', default=False, action='store_true', + help='do not perform cross-attention') + parser.add_argument('--cross-self-attention', default=False, action='store_true', + help='perform cross+self-attention') + # args for "Reducing Transformer Depth on Demand with Structured Dropout" (Fan et al., 2019) + parser.add_argument('--encoder-layerdrop', type=float, metavar='D', default=0, + help='LayerDrop probability for encoder') + parser.add_argument('--decoder-layerdrop', type=float, metavar='D', default=0, + help='LayerDrop probability for decoder') + parser.add_argument('--encoder-layers-to-keep', default=None, + help='which layers to *keep* when pruning as a comma-separated list') + parser.add_argument('--decoder-layers-to-keep', default=None, + help='which layers to *keep* when pruning as a comma-separated list') + # args for Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020) + parser.add_argument('--quant-noise-pq', type=float, metavar='D', default=0, + help='iterative PQ quantization noise at training time') + parser.add_argument('--quant-noise-pq-block-size', type=int, metavar='D', default=8, + help='block size of quantization noise at training time') + parser.add_argument('--quant-noise-scalar', type=float, metavar='D', default=0, + help='scalar quantization noise and scalar quantization at training time') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if args.encoder_layers_to_keep: + args.encoder_layers = len(args.encoder_layers_to_keep.split(",")) + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, "max_source_positions", None) is None: + args.max_source_positions = DEFAULT_MAX_SOURCE_POSITIONS + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = DEFAULT_MAX_TARGET_POSITIONS + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + if args.share_all_embeddings: + if src_dict != tgt_dict: + raise ValueError("--share-all-embeddings requires a joined dictionary") + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise ValueError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + encoder_embed_tokens = cls.build_embedding( + args, src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = encoder_embed_tokens + args.share_decoder_input_output_embed = True + else: + encoder_embed_tokens = cls.build_embedding( + args, src_dict, args.encoder_embed_dim, args.encoder_embed_path + ) + decoder_embed_tokens = cls.build_embedding( + args, tgt_dict, args.decoder_embed_dim, args.decoder_embed_path + ) + + encoder = cls.build_encoder(args, src_dict, encoder_embed_tokens) + decoder = cls.build_decoder(args, tgt_dict, decoder_embed_tokens) + return cls(args, encoder, decoder) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + + emb = Embedding(num_embeddings, embed_dim, padding_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerDecoder( + args, + tgt_dict, + embed_tokens, + no_encoder_attn=getattr(args, "no_cross_attention", False), + ) + + # TorchScript doesn't support optional arguments with variable length (**kwargs). + # Current workaround is to add union of all arguments in child classes. + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + return_all_hiddens: bool = True, + features_only: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + Run the forward pass for an encoder-decoder model. + + Copied from the base class, but without ``**kwargs``, + which are not supported by TorchScript. + """ + encoder_out = self.encoder( + src_tokens, src_lengths=src_lengths, return_all_hiddens=return_all_hiddens + ) + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + features_only=features_only, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + src_lengths=src_lengths, + return_all_hiddens=return_all_hiddens, + ) + return decoder_out + + # Since get_normalized_probs is in the Fairseq Model which is not scriptable, + # I rewrite the get_normalized_probs from Base Class to call the + # helper function in the Base Class. + @torch.jit.export + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + return self.get_normalized_probs_scriptable(net_output, log_probs, sample) + + +class TransformerEncoder(FairseqEncoder): + """ + Transformer encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`TransformerEncoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_tokens (torch.nn.Embedding): input embedding + """ + + def __init__(self, args, dictionary, embed_tokens): + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + self.encoder_layerdrop = args.encoder_layerdrop + + embed_dim = embed_tokens.embedding_dim + self.padding_idx = embed_tokens.padding_idx + self.max_source_positions = args.max_source_positions + + self.embed_tokens = embed_tokens + + self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(embed_dim) + + self.embed_positions = ( + PositionalEmbedding( + args.max_source_positions, + embed_dim, + self.padding_idx, + learned=args.encoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + + if getattr(args, "layernorm_embedding", False): + self.layernorm_embedding = LayerNorm(embed_dim) + else: + self.layernorm_embedding = None + + if not args.adaptive_input and args.quant_noise_pq > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(embed_dim, embed_dim, bias=False), + args.quant_noise_pq, + args.quant_noise_pq_block_size, + ) + else: + self.quant_noise = None + + if self.encoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.encoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [self.build_encoder_layer(args) for i in range(args.encoder_layers)] + ) + self.num_layers = len(self.layers) + + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + def build_encoder_layer(self, args): + return TransformerEncoderLayer(args) + + def forward_embedding(self, src_tokens): + # embed tokens and positions + x = embed = self.embed_scale * self.embed_tokens(src_tokens) + if self.embed_positions is not None: + x = embed + self.embed_positions(src_tokens) + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + x = self.dropout_module(x) + if self.quant_noise is not None: + x = self.quant_noise(x) + return x, embed + + def forward(self, src_tokens, src_lengths, return_all_hiddens: bool = False): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + + Returns: + namedtuple: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + """ + x, encoder_embedding = self.forward_embedding(src_tokens) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # compute padding mask + encoder_padding_mask = src_tokens.eq(self.padding_idx) + + encoder_states = [] if return_all_hiddens else None + + # encoder layers + for layer in self.layers: + x = layer(x, encoder_padding_mask) + if return_all_hiddens: + assert encoder_states is not None + encoder_states.append(x) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + return EncoderOut( + encoder_out=x, # T x B x C + encoder_padding_mask=encoder_padding_mask, # B x T + encoder_embedding=encoder_embedding, # B x T x C + encoder_states=encoder_states, # List[T x B x C] + src_tokens=None, + src_lengths=None, + ) + + @torch.jit.export + def reorder_encoder_out(self, encoder_out: EncoderOut, new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + """ + Since encoder_padding_mask and encoder_embedding are both of type + Optional[Tensor] in EncoderOut, they need to be copied as local + variables for Torchscript Optional refinement + """ + encoder_padding_mask: Optional[Tensor] = encoder_out.encoder_padding_mask + encoder_embedding: Optional[Tensor] = encoder_out.encoder_embedding + + new_encoder_out = ( + encoder_out.encoder_out + if encoder_out.encoder_out is None + else encoder_out.encoder_out.index_select(1, new_order) + ) + new_encoder_padding_mask = ( + encoder_padding_mask + if encoder_padding_mask is None + else encoder_padding_mask.index_select(0, new_order) + ) + new_encoder_embedding = ( + encoder_embedding + if encoder_embedding is None + else encoder_embedding.index_select(0, new_order) + ) + src_tokens = encoder_out.src_tokens + if src_tokens is not None: + src_tokens = src_tokens.index_select(0, new_order) + + src_lengths = encoder_out.src_lengths + if src_lengths is not None: + src_lengths = src_lengths.index_select(0, new_order) + + encoder_states = encoder_out.encoder_states + if encoder_states is not None: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return EncoderOut( + encoder_out=new_encoder_out, # T x B x C + encoder_padding_mask=new_encoder_padding_mask, # B x T + encoder_embedding=new_encoder_embedding, # B x T x C + encoder_states=encoder_states, # List[T x B x C] + src_tokens=src_tokens, # B x T + src_lengths=src_lengths, # B x 1 + ) + + def max_positions(self): + """Maximum input length supported by the encoder.""" + if self.embed_positions is None: + return self.max_source_positions + return min(self.max_source_positions, self.embed_positions.max_positions) + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + if isinstance(self.embed_positions, SinusoidalPositionalEmbedding): + weights_key = "{}.embed_positions.weights".format(name) + if weights_key in state_dict: + print("deleting {0}".format(weights_key)) + del state_dict[weights_key] + state_dict[ + "{}.embed_positions._float_tensor".format(name) + ] = torch.FloatTensor(1) + for i in range(self.num_layers): + # update layer norms + self.layers[i].upgrade_state_dict_named( + state_dict, "{}.layers.{}".format(name, i) + ) + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) < 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + return state_dict + + +class TransformerDecoder(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + self.args = args + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + self._future_mask = torch.empty(0) + + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + self.decoder_layerdrop = args.decoder_layerdrop + self.share_input_output_embed = args.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = args.decoder_embed_dim + self.embed_dim = embed_dim + self.output_embed_dim = args.decoder_output_dim + + self.padding_idx = embed_tokens.padding_idx + self.max_target_positions = args.max_target_positions + + self.embed_tokens = embed_tokens + + self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(embed_dim) + + if not args.adaptive_input and args.quant_noise_pq > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(embed_dim, embed_dim, bias=False), + args.quant_noise_pq, + args.quant_noise_pq_block_size, + ) + else: + self.quant_noise = None + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + args.max_target_positions, + embed_dim, + self.padding_idx, + learned=args.decoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + + if getattr(args, "layernorm_embedding", False): + self.layernorm_embedding = LayerNorm(embed_dim) + else: + self.layernorm_embedding = None + + self.cross_self_attention = getattr(args, "cross_self_attention", False) + + if self.decoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.decoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + self.build_decoder_layer(args, no_encoder_attn) + for _ in range(args.decoder_layers) + ] + ) + self.num_layers = len(self.layers) + + if args.decoder_normalize_before and not getattr( + args, "no_decoder_final_norm", False + ): + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + self.project_out_dim = ( + Linear(embed_dim, self.output_embed_dim, bias=False) + if embed_dim != self.output_embed_dim and not args.tie_adaptive_weights + else None + ) + + self.adaptive_softmax = None + self.output_projection = None + if args.adaptive_softmax_cutoff is not None: + self.adaptive_softmax = AdaptiveSoftmax( + len(dictionary), + self.output_embed_dim, + options.eval_str_list(args.adaptive_softmax_cutoff, type=int), + dropout=args.adaptive_softmax_dropout, + adaptive_inputs=embed_tokens if args.tie_adaptive_weights else None, + factor=args.adaptive_softmax_factor, + tie_proj=args.tie_adaptive_proj, + ) + elif self.share_input_output_embed: + self.output_projection = nn.Linear( + self.embed_tokens.weight.shape[1], + self.embed_tokens.weight.shape[0], + bias=False, + ) + self.output_projection.weight = self.embed_tokens.weight + else: + self.output_projection = nn.Linear( + self.output_embed_dim, len(dictionary), bias=False + ) + nn.init.normal_( + self.output_projection.weight, mean=0, std=self.output_embed_dim ** -0.5 + ) + + def build_decoder_layer(self, args, no_encoder_attn=False): + return TransformerDecoderLayer(args, no_encoder_attn) + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[EncoderOut] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + ) + if not features_only: + x = self.output_layer(x) + return x, extra + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[EncoderOut] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + return self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + + """ + A scriptable subclass of this class has an extract_features method and calls + super().extract_features, but super() is not supported in torchscript. Aa copy of + this function is made to be used in the subclass instead. + """ + + def extract_features_scriptable( + self, + prev_output_tokens, + encoder_out: Optional[EncoderOut] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + Similar to *forward* but only return features. + + Includes several features from "Jointly Learning to Align and + Translate with Transformer Models" (Garg et al., EMNLP 2019). + + Args: + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + alignment_layer (int, optional): return mean alignment over + heads at this layer (default: last layer). + alignment_heads (int, optional): only average alignment over + this many heads (default: all heads). + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + if alignment_layer is None: + alignment_layer = self.num_layers - 1 + + # embed positions + positions = ( + self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + self_attn_padding_mask: Optional[Tensor] = None + if self.cross_self_attention or prev_output_tokens.eq(self.padding_idx).any(): + self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx) + + # decoder layers + attn: Optional[Tensor] = None + inner_states: List[Optional[Tensor]] = [x] + for idx, layer in enumerate(self.layers): + if incremental_state is None and not full_context_alignment: + self_attn_mask = self.buffered_future_mask(x) + else: + self_attn_mask = None + + x, layer_attn, _ = layer( + x, + encoder_out.encoder_out if encoder_out is not None else None, + encoder_out.encoder_padding_mask if encoder_out is not None else None, + incremental_state, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_attn=bool((idx == alignment_layer)), + need_head_weights=bool((idx == alignment_layer)), + ) + inner_states.append(x) + if layer_attn is not None and idx == alignment_layer: + attn = layer_attn.float().to(x) + + if attn is not None: + if alignment_heads is not None: + attn = attn[:alignment_heads] + + # average probabilities over heads + attn = attn.mean(dim=0) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": [attn], "inner_states": inner_states} + + def output_layer(self, features): + """Project features to the vocabulary size.""" + if self.adaptive_softmax is None: + # project back to size of vocabulary + return self.output_projection(features) + else: + return features + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + # self._future_mask.device != tensor.device is not working in TorchScript. This is a workaround. + if ( + self._future_mask.size(0) == 0 + or (not self._future_mask.device == tensor.device) + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(torch.zeros([dim, dim])), 1 + ) + self._future_mask = self._future_mask.to(tensor) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + if isinstance(self.embed_positions, SinusoidalPositionalEmbedding): + weights_key = "{}.embed_positions.weights".format(name) + if weights_key in state_dict: + del state_dict[weights_key] + state_dict[ + "{}.embed_positions._float_tensor".format(name) + ] = torch.FloatTensor(1) + + if f"{name}.output_projection.weight" not in state_dict: + if self.share_input_output_embed: + embed_out_key = f"{name}.embed_tokens.weight" + else: + embed_out_key = f"{name}.embed_out" + if embed_out_key in state_dict: + state_dict[f"{name}.output_projection.weight"] = state_dict[ + embed_out_key + ] + if not self.share_input_output_embed: + del state_dict[embed_out_key] + + for i in range(self.num_layers): + # update layer norms + layer_norm_map = { + "0": "self_attn_layer_norm", + "1": "encoder_attn_layer_norm", + "2": "final_layer_norm", + } + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layers.{}.layer_norms.{}.{}".format(name, i, old, m) + if k in state_dict: + state_dict[ + "{}.layers.{}.{}.{}".format(name, i, new, m) + ] = state_dict[k] + del state_dict[k] + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) <= 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m + + +@register_model_architecture("transformer", "transformer") +def base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.no_cross_attention = getattr(args, "no_cross_attention", False) + args.cross_self_attention = getattr(args, "cross_self_attention", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + + +@register_model_architecture("transformer", "transformer_iwslt_de_en") +def transformer_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.decoder_layers = getattr(args, "decoder_layers", 6) + base_architecture(args) + + +@register_model_architecture("transformer", "transformer_wmt_en_de") +def transformer_wmt_en_de(args): + base_architecture(args) + + +# parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture("transformer", "transformer_vaswani_wmt_en_de_big") +def transformer_vaswani_wmt_en_de_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + base_architecture(args) + + +@register_model_architecture("transformer", "transformer_vaswani_wmt_en_fr_big") +def transformer_vaswani_wmt_en_fr_big(args): + args.dropout = getattr(args, "dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) + + +@register_model_architecture("transformer", "transformer_wmt_en_de_big") +def transformer_wmt_en_de_big(args): + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) + + +# default parameters used in tensor2tensor implementation +@register_model_architecture("transformer", "transformer_wmt_en_de_big_t2t") +def transformer_wmt_en_de_big_t2t(args): + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) diff --git a/fairseq/models/transformer_align.py b/fairseq/models/transformer_align.py new file mode 100644 index 0000000000000000000000000000000000000000..4195ff398264936d579574ae646865bd01dac809 --- /dev/null +++ b/fairseq/models/transformer_align.py @@ -0,0 +1,93 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import ( + base_architecture, + transformer_wmt_en_de_big, + TransformerModel, +) + + +@register_model("transformer_align") +class TransformerAlignModel(TransformerModel): + """ + See "Jointly Learning to Align and Translate with Transformer + Models" (Garg et al., EMNLP 2019). + """ + + def __init__(self, encoder, decoder, args): + super().__init__(args, encoder, decoder) + self.alignment_heads = args.alignment_heads + self.alignment_layer = args.alignment_layer + self.full_context_alignment = args.full_context_alignment + + @staticmethod + def add_args(parser): + # fmt: off + super(TransformerAlignModel, TransformerAlignModel).add_args(parser) + parser.add_argument('--alignment-heads', type=int, metavar='D', + help='Number of cross attention heads per layer to supervised with alignments') + parser.add_argument('--alignment-layer', type=int, metavar='D', + help='Layer number which has to be supervised. 0 corresponding to the bottommost layer.') + parser.add_argument('--full-context-alignment', type=bool, metavar='D', + help='Whether or not alignment is supervised conditioned on the full target context.') + # fmt: on + + @classmethod + def build_model(cls, args, task): + # set any default arguments + transformer_align(args) + + transformer_model = TransformerModel.build_model(args, task) + return TransformerAlignModel( + transformer_model.encoder, transformer_model.decoder, args + ) + + def forward(self, src_tokens, src_lengths, prev_output_tokens): + encoder_out = self.encoder(src_tokens, src_lengths) + return self.forward_decoder(prev_output_tokens, encoder_out) + + def forward_decoder( + self, + prev_output_tokens, + encoder_out=None, + incremental_state=None, + features_only=False, + **extra_args, + ): + attn_args = { + "alignment_layer": self.alignment_layer, + "alignment_heads": self.alignment_heads, + } + decoder_out = self.decoder(prev_output_tokens, encoder_out, **attn_args) + + if self.full_context_alignment: + attn_args["full_context_alignment"] = self.full_context_alignment + _, alignment_out = self.decoder( + prev_output_tokens, + encoder_out, + features_only=True, + **attn_args, + **extra_args, + ) + decoder_out[1]["attn"] = alignment_out["attn"] + + return decoder_out + + +@register_model_architecture("transformer_align", "transformer_align") +def transformer_align(args): + args.alignment_heads = getattr(args, "alignment_heads", 1) + args.alignment_layer = getattr(args, "alignment_layer", 4) + args.full_context_alignment = getattr(args, "full_context_alignment", False) + base_architecture(args) + + +@register_model_architecture("transformer_align", "transformer_wmt_en_de_big_align") +def transformer_wmt_en_de_big_align(args): + args.alignment_heads = getattr(args, "alignment_heads", 1) + args.alignment_layer = getattr(args, "alignment_layer", 4) + transformer_wmt_en_de_big(args) diff --git a/fairseq/models/transformer_from_pretrained_xlm.py b/fairseq/models/transformer_from_pretrained_xlm.py new file mode 100644 index 0000000000000000000000000000000000000000..bd03c8450fa35221e7cb6c10ccad7479bff517ce --- /dev/null +++ b/fairseq/models/transformer_from_pretrained_xlm.py @@ -0,0 +1,155 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +from typing import Any, Dict + +from fairseq import checkpoint_utils +from fairseq.data.legacy.masked_lm_dictionary import MaskedLMDictionary +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + TransformerModel, + base_architecture as transformer_base_architecture, +) + + +@register_model("transformer_from_pretrained_xlm") +class TransformerFromPretrainedXLMModel(TransformerModel): + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + TransformerModel.add_args(parser) + parser.add_argument( + "--pretrained-xlm-checkpoint", + type=str, + metavar="STR", + help="XLM model to use for initializing transformer encoder and/or decoder", + ) + parser.add_argument( + "--init-encoder-only", + action="store_true", + help="if set, don't load the XLM weights and embeddings into decoder", + ) + parser.add_argument( + "--init-decoder-only", + action="store_true", + help="if set, don't load the XLM weights and embeddings into encoder", + ) + + @classmethod + def build_model(self, args, task, cls_dictionary=MaskedLMDictionary): + assert hasattr(args, "pretrained_xlm_checkpoint"), ( + "You must specify a path for --pretrained-xlm-checkpoint to use " + "--arch transformer_from_pretrained_xlm" + ) + assert isinstance(task.source_dictionary, cls_dictionary) and isinstance( + task.target_dictionary, cls_dictionary + ), ( + "You should use a MaskedLMDictionary when using --arch " + "transformer_from_pretrained_xlm because the pretrained XLM model " + "was trained using data binarized with MaskedLMDictionary. " + "For translation, you may want to use --task " + "translation_from_pretrained_xlm" + ) + assert not ( + getattr(args, "init_encoder_only", False) + and getattr(args, "init_decoder_only", False) + ), "Only one of --init-encoder-only and --init-decoder-only can be set." + return super().build_model(args, task) + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return TransformerEncoderFromPretrainedXLM(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerDecoderFromPretrainedXLM(args, tgt_dict, embed_tokens) + + +def upgrade_state_dict_with_xlm_weights( + state_dict: Dict[str, Any], pretrained_xlm_checkpoint: str +) -> Dict[str, Any]: + """ + Load XLM weights into a Transformer encoder or decoder model. + + Args: + state_dict: state dict for either TransformerEncoder or + TransformerDecoder + pretrained_xlm_checkpoint: checkpoint to load XLM weights from + + Raises: + AssertionError: If architecture (num layers, attention heads, etc.) + does not match between the current Transformer encoder or + decoder and the pretrained_xlm_checkpoint + """ + if not os.path.exists(pretrained_xlm_checkpoint): + raise IOError("Model file not found: {}".format(pretrained_xlm_checkpoint)) + + state = checkpoint_utils.load_checkpoint_to_cpu(pretrained_xlm_checkpoint) + xlm_state_dict = state["model"] + for key in xlm_state_dict.keys(): + + for search_key in ["embed_tokens", "embed_positions", "layers"]: + if search_key in key: + subkey = key[key.find(search_key):] + assert subkey in state_dict, ( + "{} Transformer encoder / decoder " + "state_dict does not contain {}. Cannot " + "load {} from pretrained XLM checkpoint " + "{} into Transformer.".format( + str(state_dict.keys()), + subkey, key, pretrained_xlm_checkpoint) + ) + + state_dict[subkey] = xlm_state_dict[key] + return state_dict + + +class TransformerEncoderFromPretrainedXLM(TransformerEncoder): + + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + if getattr(args, 'init_decoder_only', False): + # Don't load XLM weights for encoder if --init-decoder-only + return + + assert hasattr(args, "pretrained_xlm_checkpoint"), ( + "--pretrained-xlm-checkpoint must be specified to load Transformer " + "encoder from pretrained XLM" + ) + xlm_loaded_state_dict = upgrade_state_dict_with_xlm_weights( + state_dict=self.state_dict(), + pretrained_xlm_checkpoint=args.pretrained_xlm_checkpoint, + ) + self.load_state_dict(xlm_loaded_state_dict, strict=True) + + +class TransformerDecoderFromPretrainedXLM(TransformerDecoder): + + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__(args, dictionary, embed_tokens, no_encoder_attn) + if getattr(args, 'init_encoder_only', False): + # Don't load XLM weights for decoder if --init-encoder-only + return + assert hasattr(args, "pretrained_xlm_checkpoint"), ( + "--pretrained-xlm-checkpoint must be specified to load Transformer " + "decoder from pretrained XLM" + ) + + xlm_loaded_state_dict = upgrade_state_dict_with_xlm_weights( + state_dict=self.state_dict(), + pretrained_xlm_checkpoint=args.pretrained_xlm_checkpoint, + ) + self.load_state_dict(xlm_loaded_state_dict, strict=True) + + +@register_model_architecture( + "transformer_from_pretrained_xlm", "transformer_from_pretrained_xlm" +) +def base_architecture(args): + transformer_base_architecture(args) diff --git a/fairseq/models/transformer_lm.py b/fairseq/models/transformer_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..b59363900e3ad85fb9e149f575f562efe4740bc8 --- /dev/null +++ b/fairseq/models/transformer_lm.py @@ -0,0 +1,304 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import options, utils +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import ( + Embedding, + TransformerDecoder, +) +from fairseq.modules import ( + AdaptiveInput, + CharacterTokenEmbedder, +) + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model('transformer_lm') +class TransformerLanguageModel(FairseqLanguageModel): + + @classmethod + def hub_models(cls): + + def moses_fastbpe(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'fastbpe', + } + + return { + 'transformer_lm.gbw.adaptive_huge': 'https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_gbw_huge.tar.bz2', + 'transformer_lm.wiki103.adaptive': 'https://dl.fbaipublicfiles.com/fairseq/models/lm/adaptive_lm_wiki103.v2.tar.bz2', + 'transformer_lm.wmt19.en': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.en.tar.bz2'), + 'transformer_lm.wmt19.de': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.de.tar.bz2'), + 'transformer_lm.wmt19.ru': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/lm/wmt19.ru.tar.bz2'), + } + + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--activation-fn', + choices=utils.get_available_activation_fns(), + help='activation function to use') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + parser.add_argument('--activation-dropout', '--relu-dropout', type=float, metavar='D', + help='dropout probability after activation in FFN.') + parser.add_argument('--decoder-embed-dim', type=int, metavar='N', + help='decoder embedding dimension') + parser.add_argument('--decoder-output-dim', type=int, metavar='N', + help='decoder output dimension') + parser.add_argument('--decoder-input-dim', type=int, metavar='N', + help='decoder input dimension') + parser.add_argument('--decoder-ffn-embed-dim', type=int, metavar='N', + help='decoder embedding dimension for FFN') + parser.add_argument('--decoder-layers', type=int, metavar='N', + help='num decoder layers') + parser.add_argument('--decoder-attention-heads', type=int, metavar='N', + help='num decoder attention heads') + parser.add_argument('--decoder-normalize-before', action='store_true', + help='apply layernorm before each decoder block') + parser.add_argument('--no-decoder-final-norm', action='store_true', + help='don\'t add an extra layernorm after the last decoder block') + parser.add_argument('--adaptive-softmax-cutoff', metavar='EXPR', + help='comma separated list of adaptive softmax cutoff points. ' + 'Must be used with adaptive_loss criterion') + parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D', + help='sets adaptive softmax dropout for the tail projections') + parser.add_argument('--adaptive-softmax-factor', type=float, metavar='N', + help='adaptive input factor') + parser.add_argument('--no-token-positional-embeddings', action='store_true', + help='if set, disables positional embeddings (outside self attention)') + parser.add_argument('--share-decoder-input-output-embed', action='store_true', + help='share decoder input and output embeddings') + parser.add_argument('--character-embeddings', action='store_true', + help='if set, uses character embedding convolutions to produce token embeddings') + parser.add_argument('--character-filters', type=str, metavar='LIST', + default='[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]', + help='size of character embeddings') + parser.add_argument('--character-embedding-dim', default=4, type=int, metavar='N', + help='size of character embeddings') + parser.add_argument('--char-embedder-highway-layers', default=2, type=int, metavar='N', + help='number of highway layers for character token embeddder') + parser.add_argument('--adaptive-input', action='store_true', + help='if set, uses adaptive input') + parser.add_argument('--adaptive-input-factor', type=float, metavar='N', + help='adaptive input factor') + parser.add_argument('--adaptive-input-cutoff', metavar='EXPR', + help='comma separated list of adaptive input cutoff points.') + parser.add_argument('--tie-adaptive-weights', action='store_true', + help='if set, ties the weights of adaptive softmax and adaptive input') + parser.add_argument('--tie-adaptive-proj', action='store_true', + help='if set, ties the projection weights of adaptive softmax and adaptive input') + parser.add_argument('--decoder-learned-pos', action='store_true', + help='use learned positional embeddings in the decoder') + parser.add_argument('--layernorm-embedding', action='store_true', + help='add layernorm to embedding') + parser.add_argument('--no-scale-embedding', action='store_true', + help='if True, dont scale embeddings') + # args for "Reducing Transformer Depth on Demand with Structured Dropout" (Fan et al., 2019) + parser.add_argument('--decoder-layerdrop', type=float, metavar='D', default=0, + help='LayerDrop probability for decoder') + parser.add_argument('--decoder-layers-to-keep', default=None, + help='which layers to *keep* when pruning as a comma-separated list') + # args for Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020) + parser.add_argument('--quant-noise-pq', type=float, metavar='D', default=0, + help='iterative PQ quantization noise at training time') + parser.add_argument('--quant-noise-pq-block-size', type=int, metavar='D', default=8, + help='block size of quantization noise at training time') + parser.add_argument('--quant-noise-scalar', type=float, metavar='D', default=0, + help='scalar quantization noise and scalar quantization at training time') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_lm_architecture(args) + + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, 'max_target_positions', None) is None: + args.max_target_positions = getattr(args, 'tokens_per_sample', DEFAULT_MAX_TARGET_POSITIONS) + + if args.character_embeddings: + embed_tokens = CharacterTokenEmbedder( + task.source_dictionary, eval(args.character_filters), + args.character_embedding_dim, args.decoder_embed_dim, + args.char_embedder_highway_layers, + ) + elif args.adaptive_input: + embed_tokens = AdaptiveInput( + len(task.source_dictionary), task.source_dictionary.pad(), args.decoder_input_dim, + args.adaptive_input_factor, args.decoder_embed_dim, + options.eval_str_list(args.adaptive_input_cutoff, type=int), + args.quant_noise_pq, args.quant_noise_pq_block_size, + ) + else: + embed_tokens = cls.build_embedding(args, task.source_dictionary, args.decoder_input_dim) + + if args.tie_adaptive_weights: + assert args.adaptive_input + assert args.adaptive_input_factor == args.adaptive_softmax_factor + assert args.adaptive_softmax_cutoff == args.adaptive_input_cutoff, '{} != {}'.format( + args.adaptive_softmax_cutoff, args.adaptive_input_cutoff) + assert args.decoder_input_dim == args.decoder_output_dim + + decoder = TransformerDecoder( + args, task.target_dictionary, embed_tokens, no_encoder_attn=True, + ) + return cls(decoder) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + embed_tokens = Embedding(len(dictionary), embed_dim, dictionary.pad()) + return embed_tokens + + +@register_model_architecture('transformer_lm', 'transformer_lm') +def base_lm_architecture(args): + # backward compatibility for older model checkpoints + if hasattr(args, 'no_tie_adaptive_proj'): + # previous models defined --no-tie-adaptive-proj, so use the existence of + # that option to determine if this is an "old" model checkpoint + args.no_decoder_final_norm = True # old models always set this to True + if args.no_tie_adaptive_proj is False: + args.tie_adaptive_proj = True + if hasattr(args, 'decoder_final_norm'): + args.no_decoder_final_norm = not args.decoder_final_norm + + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.0) + + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 2048) + args.decoder_layers = getattr(args, 'decoder_layers', 6) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 8) + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', None) + args.adaptive_softmax_dropout = getattr(args, 'adaptive_softmax_dropout', 0) + args.adaptive_softmax_factor = getattr(args, 'adaptive_softmax_factor', 4) + args.decoder_learned_pos = getattr(args, 'decoder_learned_pos', False) + args.activation_fn = getattr(args, 'activation_fn', 'relu') + + args.add_bos_token = getattr(args, 'add_bos_token', False) + args.no_token_positional_embeddings = getattr(args, 'no_token_positional_embeddings', False) + args.share_decoder_input_output_embed = getattr(args, 'share_decoder_input_output_embed', False) + args.character_embeddings = getattr(args, 'character_embeddings', False) + + args.decoder_output_dim = getattr(args, 'decoder_output_dim', args.decoder_embed_dim) + args.decoder_input_dim = getattr(args, 'decoder_input_dim', args.decoder_embed_dim) + + # Model training is not stable without this + args.decoder_normalize_before = True + args.no_decoder_final_norm = getattr(args, 'no_decoder_final_norm', False) + + args.adaptive_input = getattr(args, 'adaptive_input', False) + args.adaptive_input_factor = getattr(args, 'adaptive_input_factor', 4) + args.adaptive_input_cutoff = getattr(args, 'adaptive_input_cutoff', None) + + args.tie_adaptive_weights = getattr(args, 'tie_adaptive_weights', False) + args.tie_adaptive_proj = getattr(args, 'tie_adaptive_proj', False) + + args.no_scale_embedding = getattr(args, 'no_scale_embedding', False) + args.layernorm_embedding = getattr(args, 'layernorm_embedding', False) + + +@register_model_architecture('transformer_lm', 'transformer_lm_big') +def transformer_lm_big(args): + args.decoder_layers = getattr(args, 'decoder_layers', 12) + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 1024) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 4096) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 16) + base_lm_architecture(args) + + +@register_model_architecture('transformer_lm', 'transformer_lm_wiki103') +@register_model_architecture('transformer_lm', 'transformer_lm_baevski_wiki103') +def transformer_lm_baevski_wiki103(args): + args.decoder_layers = getattr(args, 'decoder_layers', 16) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 8) + args.dropout = getattr(args, 'dropout', 0.3) + args.adaptive_input = getattr(args, 'adaptive_input', True) + args.tie_adaptive_weights = getattr(args, 'tie_adaptive_weights', True) + args.adaptive_input_cutoff = getattr(args, 'adaptive_input_cutoff', '20000,60000') + args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', '20000,60000') + args.adaptive_softmax_dropout = getattr(args, 'adaptive_softmax_dropout', 0.2) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_dropout = getattr(args, 'activation_dropout', 0.1) + args.no_decoder_final_norm = getattr(args, 'no_decoder_final_norm', True) + args.tie_adaptive_proj = getattr(args, 'tie_adaptive_proj', True) + transformer_lm_big(args) + + +@register_model_architecture('transformer_lm', 'transformer_lm_gbw') +@register_model_architecture('transformer_lm', 'transformer_lm_baevski_gbw') +def transformer_lm_baevski_gbw(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 512) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.no_decoder_final_norm = getattr(args, 'no_decoder_final_norm', True) + transformer_lm_big(args) + + +@register_model_architecture('transformer_lm', 'transformer_lm_gpt') +def transformer_lm_gpt(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 768) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 3072) + args.decoder_layers = getattr(args, 'decoder_layers', 12) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 12) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + base_lm_architecture(args) + + +@register_model_architecture('transformer_lm', 'transformer_lm_gpt2_small') +def transformer_lm_gpt2_small(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 1024) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 4096) + args.decoder_layers = getattr(args, 'decoder_layers', 24) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 16) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + base_lm_architecture(args) + + +@register_model_architecture('transformer_lm', 'transformer_lm_gpt2_medium') +def transformer_lm_gpt2_medium(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 1280) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 5120) + args.decoder_layers = getattr(args, 'decoder_layers', 36) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 20) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + base_lm_architecture(args) + + +@register_model_architecture('transformer_lm', 'transformer_lm_gpt2_big') +def transformer_lm_gpt2_big(args): + args.decoder_embed_dim = getattr(args, 'decoder_embed_dim', 1600) + args.decoder_ffn_embed_dim = getattr(args, 'decoder_ffn_embed_dim', 6400) + args.decoder_layers = getattr(args, 'decoder_layers', 48) + args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 25) + args.dropout = getattr(args, 'dropout', 0.1) + args.attention_dropout = getattr(args, 'attention_dropout', 0.1) + args.activation_fn = getattr(args, 'activation_fn', 'gelu') + base_lm_architecture(args) diff --git a/fairseq/models/wav2vec/__init__.py b/fairseq/models/wav2vec/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..06cec18183ca14cd534d14558e8b44e25f3e69d5 --- /dev/null +++ b/fairseq/models/wav2vec/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .wav2vec import * # noqa +from .wav2vec2 import * # noqa +from .wav2vec2_asr import * # noqa diff --git a/fairseq/models/wav2vec/__pycache__/__init__.cpython-310.pyc b/fairseq/models/wav2vec/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0c9f317e1c542342e6f8c25c06a88809f46814b2 Binary files /dev/null and b/fairseq/models/wav2vec/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/models/wav2vec/__pycache__/wav2vec.cpython-310.pyc b/fairseq/models/wav2vec/__pycache__/wav2vec.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fd8055aeb10332ffaffbe5770eb4f99f2386967f Binary files /dev/null and b/fairseq/models/wav2vec/__pycache__/wav2vec.cpython-310.pyc differ diff --git a/fairseq/models/wav2vec/__pycache__/wav2vec2.cpython-310.pyc b/fairseq/models/wav2vec/__pycache__/wav2vec2.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3510ed8ea5f865bb482b102f8c3813b826f5212b Binary files /dev/null and b/fairseq/models/wav2vec/__pycache__/wav2vec2.cpython-310.pyc differ diff --git a/fairseq/models/wav2vec/__pycache__/wav2vec2_asr.cpython-310.pyc b/fairseq/models/wav2vec/__pycache__/wav2vec2_asr.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3b8fe2eebf75c01e20816d66bcc2c1a00ed5e876 Binary files /dev/null and b/fairseq/models/wav2vec/__pycache__/wav2vec2_asr.cpython-310.pyc differ diff --git a/fairseq/models/wav2vec/wav2vec.py b/fairseq/models/wav2vec/wav2vec.py new file mode 100644 index 0000000000000000000000000000000000000000..905df824f3ae0f708fc269cc486d03df78611341 --- /dev/null +++ b/fairseq/models/wav2vec/wav2vec.py @@ -0,0 +1,735 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +import sys + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq.models import BaseFairseqModel, register_model, register_model_architecture +from fairseq.modules import ( + Fp32GroupNorm, + Fp32LayerNorm, + GumbelVectorQuantizer, + KmeansVectorQuantizer, + TransposeLast, +) +from fairseq.utils import buffered_arange + +logger = logging.getLogger(__name__) + + +@register_model("wav2vec") +class Wav2VecModel(BaseFairseqModel): + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--prediction-steps", + type=int, + metavar="N", + help="number of steps ahead to predict", + ) + parser.add_argument( + "--sample-distance", + type=int, + metavar="N", + help="sample distance from target. does not work properly with cross-sampling", + ) + parser.add_argument( + "--cross-sample-negatives", + type=int, + metavar="N", + help="num of cross sampled negatives", + ) + parser.add_argument( + "--num-negatives", type=int, metavar="N", help="number of negative examples" + ) + parser.add_argument( + "--conv-feature-layers", + type=str, + metavar="EXPR", + help="convolutional feature extraction layers [(dim, kernel_size, stride), ...]", + ) + parser.add_argument( + "--conv-aggregator-layers", + type=str, + metavar="EXPR", + help="convolutional feature extraction layers [(dim, kernel_size, stride), ...]", + ) + parser.add_argument( + "--dropout", + type=float, + metavar="D", + help="dropout to apply within the model", + ) + parser.add_argument( + "--dropout-features", + type=float, + metavar="D", + help="dropout to apply to the features", + ) + parser.add_argument( + "--dropout-agg", + type=float, + metavar="D", + help="dropout to apply after aggregation step", + ) + parser.add_argument( + "--encoder", type=str, choices=["cnn"], help="type of encoder to use" + ) + parser.add_argument( + "--aggregator", + type=str, + choices=["cnn", "gru"], + help="type of aggregator to use", + ) + parser.add_argument( + "--gru-dim", type=int, metavar="N", help="GRU dimensionality" + ) + + parser.add_argument( + "--no-conv-bias", + action="store_true", + help="if set, does not learn bias for conv layers", + ) + parser.add_argument( + "--agg-zero-pad", + action="store_true", + help="if set, zero pads in aggregator instead of repl pad", + ) + + parser.add_argument( + "--skip-connections-feat", + action="store_true", + help="if set, adds skip connections to the feature extractor", + ) + parser.add_argument( + "--skip-connections-agg", + action="store_true", + help="if set, adds skip connections to the aggregator", + ) + parser.add_argument( + "--residual-scale", + type=float, + metavar="D", + help="scales residual by sqrt(value)", + ) + + parser.add_argument( + "--log-compression", + action="store_true", + help="if set, adds a log compression to feature extractor", + ) + + parser.add_argument( + "--balanced-classes", + action="store_true", + help="if set, loss is scaled to balance for number of negatives", + ) + + parser.add_argument( + "--project-features", + choices=["none", "same", "new"], + help="if not none, features are projected using the (same or new) aggregator", + ) + + parser.add_argument( + "--non-affine-group-norm", + action="store_true", + help="if set, group norm is not affine", + ) + + parser.add_argument( + "--offset", + help="if set, introduces an offset from target to predictions. " + 'if set to "auto", it is computed automatically from the receptive field', + ) + + parser.add_argument( + "--activation", + type=str, + choices=["relu", "gelu"], + help="which activation function to use", + ) + + parser.add_argument( + "--vq-type", + type=str, + choices=["none", "gumbel", "kmeans"], + help="which type of quantizer to use", + ) + parser.add_argument( + "--vq-vars", + type=int, + metavar="N", + help="if set, project to this many vector quantized variables per group", + ) + parser.add_argument( + "--vq-groups", + type=int, + metavar="N", + help="number of groups of latent variables", + ) + parser.add_argument( + "--vq-dim", + type=int, + metavar="N", + help="uses this dimensionality for quantized vectors", + ) + parser.add_argument( + "--vq-depth", + type=int, + metavar="N", + help="number of layers for vq weight projection", + ) + parser.add_argument( + "--combine-groups", + action="store_true", + help="if set, variables are shared among groups", + ) + parser.add_argument( + "--vq-temp", + type=str, + metavar="TEMP", + help="temperature for latent variable sampling with gumbel softmax. should be a tuple of 3 values (start, end, decay)", + ) + parser.add_argument( + "--vq-gamma", + type=float, + metavar="D", + help="gamma parameter for kmeans style vector quantization", + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_wav2vec_architecture(args) + + model = Wav2VecModel(args) + logger.info(model) + return model + + def __init__(self, args): + super().__init__() + + self.prediction_steps = args.prediction_steps + offset = args.offset + + if args.activation == "relu": + activation = nn.ReLU() + elif args.activation == "gelu": + activation = nn.GELU() + else: + raise Exception("unknown activation " + args.activation) + + if args.encoder == "cnn": + feature_enc_layers = eval(args.conv_feature_layers) + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + log_compression=args.log_compression, + skip_connections=args.skip_connections_feat, + residual_scale=args.residual_scale, + non_affine_group_norm=args.non_affine_group_norm, + activation=activation, + ) + embed = feature_enc_layers[-1][0] + else: + raise Exception("unknown encoder type " + args.encoder) + + self.vector_quantizer = None + if args.vq_type == "gumbel": + self.vector_quantizer = GumbelVectorQuantizer( + dim=embed, + num_vars=args.vq_vars, + temp=eval(args.vq_temp), + groups=args.vq_groups, + combine_groups=args.combine_groups, + vq_dim=args.vq_dim if args.vq_dim > 0 else embed, + time_first=False, + activation=activation, + weight_proj_depth=args.vq_depth, + weight_proj_factor=2, + ) + elif args.vq_type == "kmeans": + self.vector_quantizer = KmeansVectorQuantizer( + dim=embed, + num_vars=args.vq_vars, + groups=args.vq_groups, + combine_groups=args.combine_groups, + vq_dim=args.vq_dim if args.vq_dim > 0 else embed, + time_first=False, + gamma=args.vq_gamma, + ) + else: + assert ( + args.vq_type == "none" or args.vq_type is None + ), "Unknown quantizer type" + + if args.offset == "auto": + assert args.encoder == "cnn" + jin = 0 + rin = 0 + for _, k, stride in feature_enc_layers: + if rin == 0: + rin = k + rin = rin + (k - 1) * jin + if jin == 0: + jin = stride + else: + jin *= stride + offset = math.ceil(rin / jin) + + offset = int(offset) + + def make_aggregator(): + if args.aggregator == "cnn": + agg_layers = eval(args.conv_aggregator_layers) + agg_dim = agg_layers[-1][0] + feature_aggregator = ConvAggegator( + conv_layers=agg_layers, + embed=embed, + dropout=args.dropout, + skip_connections=args.skip_connections_agg, + residual_scale=args.residual_scale, + non_affine_group_norm=args.non_affine_group_norm, + conv_bias=not args.no_conv_bias, + zero_pad=args.agg_zero_pad, + activation=activation, + ) + elif args.aggregator == "gru": + agg_dim = args.gru_dim + feature_aggregator = nn.Sequential( + TransposeLast(), + nn.GRU( + input_size=embed, + hidden_size=agg_dim, + num_layers=1, + dropout=args.dropout, + ), + TransposeLast(deconstruct_idx=0), + ) + else: + raise Exception("unknown aggregator type " + args.aggregator) + + return feature_aggregator, agg_dim + + self.feature_aggregator, agg_dim = make_aggregator() + + self.wav2vec_predictions = Wav2VecPredictionsModel( + in_dim=agg_dim, + out_dim=embed, + prediction_steps=args.prediction_steps, + n_negatives=args.num_negatives, + cross_sample_negatives=args.cross_sample_negatives, + sample_distance=args.sample_distance, + dropout=args.dropout, + offset=offset, + balanced_classes=args.balanced_classes, + infonce=args.infonce, + ) + + self.dropout_feats = nn.Dropout(p=args.dropout_features) + self.dropout_agg = nn.Dropout(p=args.dropout_agg) + + if args.project_features == "none": + self.project_features = None + elif args.project_features == "same": + self.project_features = self.feature_aggregator + elif args.project_features == "new": + self.project_features, _ = make_aggregator() + + def forward(self, source): + result = {} + + features = self.feature_extractor(source) + if self.vector_quantizer: + q_res = self.vector_quantizer(features) + features = q_res["x"] + for k in q_res.keys(): + if k != "x": + result[k] = q_res[k] + + x = self.dropout_feats(features) + x = self.feature_aggregator(x) + x = self.dropout_agg(x) + + if self.project_features is not None: + features = self.project_features(features) + x, targets = self.wav2vec_predictions(x, features) + result["cpc_logits"] = x + result["cpc_targets"] = targets + + return result + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + + def max_positions(self): + """Maximum length supported by the model.""" + return sys.maxsize + + def get_logits(self, net_output): + logits = net_output["cpc_logits"] + return logits + + def get_targets(self, sample, net_output): + t = net_output["cpc_targets"] + if isinstance(t, tuple): + t = t[0] + return t.contiguous() + + def get_target_weights(self, targets, net_output): + targets = net_output["cpc_targets"] + if isinstance(targets, tuple) and targets[-1] is not None: + return targets[-1] + return None + + def get_extra_losses(self, net_output): + loss = None + if "prob_perplexity" in net_output: + loss = net_output["num_vars"] - net_output["prob_perplexity"] + elif "kmeans_loss" in net_output: + loss = net_output["kmeans_loss"] + + return loss + + +def norm_block(is_layer_norm, dim, affine=True): + if is_layer_norm: + mod = nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=affine), + TransposeLast(), + ) + else: + mod = Fp32GroupNorm(1, dim, affine=affine) + + return mod + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers, + dropout, + log_compression, + skip_connections, + residual_scale, + non_affine_group_norm, + activation, + ): + super().__init__() + + def block(n_in, n_out, k, stride): + return nn.Sequential( + nn.Conv1d(n_in, n_out, k, stride=stride, bias=False), + nn.Dropout(p=dropout), + norm_block( + is_layer_norm=False, dim=n_out, affine=not non_affine_group_norm + ), + activation, + ) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for dim, k, stride in conv_layers: + self.conv_layers.append(block(in_d, dim, k, stride)) + in_d = dim + + self.log_compression = log_compression + self.skip_connections = skip_connections + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x): + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + residual = x + x = conv(x) + if self.skip_connections and x.size(1) == residual.size(1): + tsz = x.size(2) + r_tsz = residual.size(2) + residual = residual[..., :: r_tsz // tsz][..., :tsz] + x = (x + residual) * self.residual_scale + + if self.log_compression: + x = x.abs() + x = x + 1 + x = x.log() + + return x + + +class ZeroPad1d(nn.Module): + def __init__(self, pad_left, pad_right): + super().__init__() + self.pad_left = pad_left + self.pad_right = pad_right + + def forward(self, x): + return F.pad(x, (self.pad_left, self.pad_right)) + + +class ConvAggegator(nn.Module): + def __init__( + self, + conv_layers, + embed, + dropout, + skip_connections, + residual_scale, + non_affine_group_norm, + conv_bias, + zero_pad, + activation, + ): + super().__init__() + + def block(n_in, n_out, k, stride): + # padding dims only really make sense for stride = 1 + ka = k // 2 + kb = ka - 1 if k % 2 == 0 else ka + + pad = ( + ZeroPad1d(ka + kb, 0) if zero_pad else nn.ReplicationPad1d((ka + kb, 0)) + ) + + return nn.Sequential( + pad, + nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias), + nn.Dropout(p=dropout), + norm_block(False, n_out, affine=not non_affine_group_norm), + activation, + ) + + in_d = embed + self.conv_layers = nn.ModuleList() + self.residual_proj = nn.ModuleList() + for dim, k, stride in conv_layers: + if in_d != dim and skip_connections: + self.residual_proj.append(nn.Conv1d(in_d, dim, 1, bias=False)) + else: + self.residual_proj.append(None) + + self.conv_layers.append(block(in_d, dim, k, stride)) + in_d = dim + self.conv_layers = nn.Sequential(*self.conv_layers) + self.skip_connections = skip_connections + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x): + for rproj, conv in zip(self.residual_proj, self.conv_layers): + residual = x + x = conv(x) + if self.skip_connections: + if rproj is not None: + residual = rproj(residual) + x = (x + residual) * self.residual_scale + return x + + +class Wav2VecPredictionsModel(nn.Module): + def __init__( + self, + in_dim, + out_dim, + prediction_steps, + n_negatives, + cross_sample_negatives, + sample_distance, + dropout, + offset, + balanced_classes, + infonce, + ): + super().__init__() + + self.n_negatives = n_negatives + self.cross_sample_negatives = cross_sample_negatives + self.sample_distance = sample_distance + self.project_to_steps = nn.ConvTranspose2d( + in_dim, out_dim, (1, prediction_steps) + ) + self.dropout = nn.Dropout(p=dropout) + self.offset = offset + self.balanced_classes = balanced_classes + self.infonce = infonce + + def sample_negatives(self, y): + bsz, fsz, tsz = y.shape + + y = y.transpose(0, 1) # BCT -> CBT + y = y.contiguous().view(fsz, -1) # CBT => C(BxT) + + cross_high = tsz * bsz + high = tsz if self.sample_distance is None else min(tsz, self.sample_distance) + assert high > 1 + + neg_idxs = torch.randint(low=0, high=high, size=(bsz, self.n_negatives * tsz)) + + with torch.no_grad(): + if self.n_negatives > 0: + tszs = ( + buffered_arange(tsz) + .unsqueeze(-1) + .expand(-1, self.n_negatives) + .flatten() + ) + + neg_idxs = torch.randint( + low=0, high=high - 1, size=(bsz, self.n_negatives * tsz) + ) + neg_idxs[neg_idxs >= tszs] += 1 + + if self.cross_sample_negatives > 0: + tszs = ( + buffered_arange(tsz) + .unsqueeze(-1) + .expand(-1, self.cross_sample_negatives) + .flatten() + ) + + cross_neg_idxs = torch.randint( + low=0, + high=cross_high - 1, + size=(bsz, self.cross_sample_negatives * tsz), + ) + cross_neg_idxs[cross_neg_idxs >= tszs] += 1 + + if self.n_negatives > 0: + for i in range(1, bsz): + neg_idxs[i] += i * high + else: + neg_idxs = cross_neg_idxs + + if self.cross_sample_negatives > 0 and self.n_negatives > 0: + neg_idxs = torch.cat([neg_idxs, cross_neg_idxs], dim=1) + + negs = y[..., neg_idxs.view(-1)] + negs = negs.view( + fsz, bsz, self.n_negatives + self.cross_sample_negatives, tsz + ).permute( + 2, 1, 0, 3 + ) # to NxBxCxT + + return negs + + def forward(self, x, y): + + x = x.unsqueeze(-1) + x = self.project_to_steps(x) # BxCxTxS + x = self.dropout(x) + + negatives = self.sample_negatives(y) + y = y.unsqueeze(0) + targets = torch.cat([y, negatives], dim=0) # Copies x B x C x T + + copies = targets.size(0) + bsz, dim, tsz, steps = x.shape + steps = min(steps, tsz - self.offset) + + predictions = x.new( + bsz * copies * (tsz - self.offset + 1) * steps + - ((steps + 1) * steps // 2) * copies * bsz + ) + if self.infonce: + labels = predictions.new_full( + (predictions.shape[0] // copies,), 0, dtype=torch.long + ) + else: + labels = torch.zeros_like(predictions) + weights = ( + torch.full_like(labels, 1 / self.n_negatives) + if self.balanced_classes and not self.infonce + else None + ) + + start = end = 0 + for i in range(steps): + offset = i + self.offset + end = start + (tsz - offset) * bsz * copies + if self.infonce: + predictions[start:end] = torch.einsum( + "bct,nbct->tbn", x[..., :-offset, i], targets[..., offset:] + ).flatten() + else: + pos_num = (end - start) // copies + predictions[start:end] = torch.einsum( + "bct,nbct->nbt", x[..., :-offset, i], targets[..., offset:] + ).flatten() + labels[start : start + pos_num] = 1.0 + if weights is not None: + weights[start : start + pos_num] = 1.0 + start = end + assert end == predictions.numel(), "{} != {}".format(end, predictions.numel()) + + if self.infonce: + predictions = predictions.view(-1, copies) + else: + if weights is not None: + labels = (labels, weights) + + return predictions, labels + + +@register_model_architecture("wav2vec", "wav2vec") +def base_wav2vec_architecture(args): + conv_feature_layers = "[(512, 10, 5)]" + conv_feature_layers += " + [(512, 8, 4)]" + conv_feature_layers += " + [(512, 4, 2)] * 3" + args.conv_feature_layers = getattr(args, "conv_feature_layers", conv_feature_layers) + + args.conv_aggregator_layers = getattr( + args, "conv_aggregator_layers", "[(512, 3, 1)] * 9" + ) + + args.prediction_steps = getattr(args, "prediction_steps", 12) + args.num_negatives = getattr(args, "num_negatives", 1) + args.sample_distance = getattr(args, "sample_distance", None) + args.cross_sample_negatives = getattr(args, "cross_sample_negatives", 0) + + args.dropout = getattr(args, "dropout", 0.0) + args.dropout_features = getattr(args, "dropout_features", 0.0) + args.dropout_agg = getattr(args, "dropout_agg", 0.0) + args.encoder = getattr(args, "encoder", "cnn") + args.aggregator = getattr(args, "aggregator", "cnn") + + args.skip_connections_feat = getattr(args, "skip_connections_feat", False) + args.skip_connections_agg = getattr(args, "skip_connections_agg", False) + args.residual_scale = getattr(args, "residual_scale", 0.5) + + args.gru_dim = getattr(args, "gru_dim", 512) + + args.no_conv_bias = getattr(args, "no_conv_bias", False) + args.agg_zero_pad = getattr(args, "agg_zero_pad", False) + + args.log_compression = getattr(args, "log_compression", False) + + args.balanced_classes = getattr(args, "balanced_classes", False) + args.infonce = getattr(args, "infonce", False) + args.project_features = getattr(args, "project_features", "none") + + args.non_affine_group_norm = getattr(args, "non_affine_group_norm", False) + + args.offset = getattr(args, "offset", "auto") + + args.activation = getattr(args, "activation", "relu") + + args.vq_type = getattr(args, "vq_type", "none") + args.vq_vars = getattr(args, "vq_vars", 320) + args.vq_groups = getattr(args, "vq_groups", 2) + args.vq_dim = getattr(args, "vq_dim", 0) + args.vq_depth = getattr(args, "vq_depth", 1) + args.combine_groups = getattr(args, "combine_groups", False) + args.vq_temp = getattr(args, "vq_temp", "(2.0, 0.5, 0.999995)") + args.vq_gamma = getattr(args, "vq_gamma", 0.25) diff --git a/fairseq/models/wav2vec/wav2vec2.py b/fairseq/models/wav2vec/wav2vec2.py new file mode 100644 index 0000000000000000000000000000000000000000..be6d10c7a2df547c310dd3abbe5f65382ab382b0 --- /dev/null +++ b/fairseq/models/wav2vec/wav2vec2.py @@ -0,0 +1,1017 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +import numpy as np + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from typing import List, Tuple + +from fairseq import utils +from fairseq.data.data_utils import compute_mask_indices +from fairseq.models import BaseFairseqModel, register_model, register_model_architecture +from fairseq.modules import ( + Fp32GroupNorm, + Fp32LayerNorm, + GradMultiply, + GumbelVectorQuantizer, + LayerNorm, + MultiheadAttention, + SamePad, + TransposeLast, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params +from fairseq.utils import buffered_arange + + +@register_model("wav2vec2") +class Wav2Vec2Model(BaseFairseqModel): + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + + parser.add_argument( + "--extractor-mode", + choices=["default", "layer_norm"], + help="mode for feature extractor. default has a single group norm with d groups in the first conv block, whereas layer_norm has layer norms in every block (meant to use with --normalize)", + ) + + parser.add_argument( + "--encoder-layers", + type=int, + metavar="L", + help="num encoder layers in the transformer", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="H", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="F", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="A", + help="num encoder attention heads", + ) + parser.add_argument( + "--activation-fn", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + + parser.add_argument( + "--dropout", + type=float, + metavar="D", + help="dropout probability for the transformer", + ) + + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + + parser.add_argument( + "--activation-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN", + ) + + parser.add_argument( + "--final-dim", + type=int, + metavar="D", + help="project final representations and targets to this many dimensions", + ) + + parser.add_argument( + "--layer-norm-first", + action="store_true", + help="apply layernorm first in the transformer", + ) + + parser.add_argument( + "--encoder-layerdrop", + type=float, + help="probability of dropping a tarnsformer layer", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + metavar="EXPR", + help="convolutional feature extraction layers [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--logit-temp", type=float, help="temperature to divide logits by" + ) + + parser.add_argument( + "--quantize-targets", action="store_true", help="use quantized targets" + ) + + parser.add_argument( + "--quantize-input", action="store_true", help="use quantized inputs" + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + help="multiply feature extractor var grads by this", + ) + + parser.add_argument( + "--latent-vars", + type=int, + metavar="N", + help="number of latent variables V in each group of the codebook", + ) + + parser.add_argument( + "--latent-groups", + type=int, + metavar="N", + help="number of groups G of latent variables in the codebook", + ) + + parser.add_argument( + "--latent-dim", + type=int, + metavar="N", + help="if set, uses this dimensionality for latent variables. otherwise uses final_dim / latent_groups", + ) + + parser.add_argument("--mask-length", type=int, help="mask length") + + parser.add_argument( + "--mask-prob", type=float, help="probability of replacing a token with mask" + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + + parser.add_argument( + "--mask-other", + type=float, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indices", + ) + + parser.add_argument( + "--no-mask-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + help="min space between spans (if no overlap is enabled)", + ) + + parser.add_argument( + "--mask-channel-length", + type=int, + help="repeat the mask indices multiple times", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indices", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + help="min space between spans (if no overlap is enabled)", + ) + + parser.add_argument( + "--dropout-input", + type=float, + metavar="D", + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + metavar="D", + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--num-negatives", type=int, metavar="N", help="number of negative examples" + ) + + parser.add_argument( + "--negatives-from-everywhere", + action="store_true", + help="sample negatives from everywhere, not just masked states", + ) + + parser.add_argument( + "--cross-sample-negatives", + type=int, + metavar="N", + help="num of cross sampled negatives", + ) + + parser.add_argument( + "--codebook-negatives", + type=int, + metavar="N", + help="num of codebook sampled negatives", + ) + + parser.add_argument( + "--conv-pos", + type=int, + metavar="N", + help="number of filters for convolutional positional embeddings", + ) + + parser.add_argument( + "--conv-pos-groups", + type=int, + metavar="N", + help="number of groups for convolutional positional embedding", + ) + + parser.add_argument( + "--latent-temp", + type=str, + metavar="D", + help="temperature for latent variable sampling. can be tuple of 3 values (start, end, decay)", + ) + + parser.add_argument( + "--target-glu", action="store_true", help="adds projection + glu to targets" + ) + + parser.add_argument( + "--conv-bias", action="store_true", help="include bias in conv encoder" + ) + + def __init__(self, args): + super().__init__() + self.args = args + + feature_enc_layers = eval(args.conv_feature_layers) + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=args.extractor_mode, + conv_bias=args.conv_bias, + ) + + self.post_extract_proj = ( + nn.Linear(self.embed, args.encoder_embed_dim) + if self.embed != args.encoder_embed_dim and not args.quantize_input + else None + ) + + self.mask_prob = args.mask_prob + self.mask_selection = args.mask_selection + self.mask_other = args.mask_other + self.mask_length = args.mask_length + self.no_mask_overlap = args.no_mask_overlap + self.mask_min_space = args.mask_min_space + + self.mask_channel_prob = args.mask_channel_prob + self.mask_channel_selection = args.mask_channel_selection + self.mask_channel_other = args.mask_channel_other + self.mask_channel_length = args.mask_channel_length + self.no_mask_channel_overlap = args.no_mask_channel_overlap + self.mask_channel_min_space = args.mask_channel_min_space + + self.dropout_input = nn.Dropout(args.dropout_input) + self.dropout_features = nn.Dropout(args.dropout_features) + + self.feature_grad_mult = args.feature_grad_mult + + self.quantizer = None + self.input_quantizer = None + + self.n_negatives = args.num_negatives + self.cross_sample_negatives = args.cross_sample_negatives + self.codebook_negatives = args.codebook_negatives + self.negatives_from_everywhere = args.negatives_from_everywhere + + self.logit_temp = args.logit_temp + + if args.quantize_input: + vq_dim = args.latent_dim if args.latent_dim > 0 else args.encoder_embed_dim + self.input_quantizer = ( + GumbelVectorQuantizer( + dim=args.encoder_embed_dim, + num_vars=args.latent_vars, + temp=eval(args.latent_temp), + groups=args.latent_groups, + combine_groups=False, + vq_dim=vq_dim, + time_first=True, + ) + if not args.same_quantizer + else self.quantizer + ) + self.project_inp = nn.Linear(vq_dim, args.encoder_embed_dim) + + final_dim = args.final_dim if args.final_dim > 0 else args.encoder_embed_dim + + if args.quantize_targets: + vq_dim = args.latent_dim if args.latent_dim > 0 else final_dim + self.quantizer = GumbelVectorQuantizer( + dim=self.embed, + num_vars=args.latent_vars, + temp=eval(args.latent_temp), + groups=args.latent_groups, + combine_groups=False, + vq_dim=vq_dim, + time_first=True, + ) + self.project_q = nn.Linear(vq_dim, final_dim) + else: + self.project_q = nn.Linear(self.embed, final_dim) + + self.mask_emb = nn.Parameter( + torch.FloatTensor(args.encoder_embed_dim).uniform_() + ) + + self.encoder = TransformerEncoder(args) + self.layer_norm = LayerNorm(self.embed) + + self.target_glu = None + if args.target_glu: + self.target_glu = nn.Sequential( + nn.Linear(final_dim, final_dim * 2), nn.GLU() + ) + + self.final_proj = nn.Linear(args.encoder_embed_dim, final_dim) + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + @classmethod + def build_model(cls, args, task=None): + """Build a new model instance.""" + + # make sure all arguments are present + base_architecture(args) + + return cls(args) + + def apply_mask(self, x, padding_mask): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def sample_negatives(self, y, num): + + if self.n_negatives == 0 and self.cross_sample_negatives == 0: + return y.new(0) + + bsz, tsz, fsz = y.shape + y = y.view(-1, fsz) # BTC => (BxT)C + + cross_high = tsz * bsz + high = tsz + with torch.no_grad(): + assert high > 1, f"{bsz,tsz,fsz}" + + if self.n_negatives > 0: + tszs = ( + buffered_arange(num) + .unsqueeze(-1) + .expand(-1, self.n_negatives) + .flatten() + ) + + neg_idxs = torch.randint( + low=0, high=high - 1, size=(bsz, self.n_negatives * num) + ) + neg_idxs[neg_idxs >= tszs] += 1 + + if self.cross_sample_negatives > 0: + tszs = ( + buffered_arange(num) + .unsqueeze(-1) + .expand(-1, self.cross_sample_negatives) + .flatten() + ) + + cross_neg_idxs = torch.randint( + low=0, + high=cross_high - 1, + size=(bsz, self.cross_sample_negatives * num), + ) + cross_neg_idxs[cross_neg_idxs >= tszs] += 1 + + if self.n_negatives > 0: + for i in range(1, bsz): + neg_idxs[i] += i * high + else: + neg_idxs = cross_neg_idxs + + if self.cross_sample_negatives > 0 and self.n_negatives > 0: + neg_idxs = torch.cat([neg_idxs, cross_neg_idxs], dim=1) + + negs = y[neg_idxs.view(-1)] + negs = negs.view( + bsz, num, self.n_negatives + self.cross_sample_negatives, fsz + ).permute( + 2, 0, 1, 3 + ) # to NxBxTxC + return negs, neg_idxs + + def compute_preds(self, x, y, negatives): + + neg_is_pos = (y == negatives).all(-1) + y = y.unsqueeze(0) + targets = torch.cat([y, negatives], dim=0) + + logits = torch.cosine_similarity(x.float(), targets.float(), dim=-1).type_as(x) + + logits /= self.logit_temp + + if neg_is_pos.any(): + logits[1:][neg_is_pos] = float("-inf") + + return logits + + def forward(self, source, padding_mask=None, mask=True, features_only=False): + + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None: + extra = padding_mask.size(1) % features.size(1) + if extra > 0: + padding_mask = padding_mask[:, :-extra] + padding_mask = padding_mask.view(padding_mask.size(0), features.size(1), -1) + padding_mask = padding_mask.all(-1) + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + num_vars = None + code_ppl = None + prob_ppl = None + curr_temp = None + + if self.input_quantizer: + q = self.input_quantizer(features, produce_targets=False) + features = q["x"] + num_vars = q["num_vars"] + code_ppl = q["code_perplexity"] + prob_ppl = q["prob_perplexity"] + curr_temp = q["temp"] + features = self.project_inp(features) + + if mask: + x, mask_indices = self.apply_mask(features, padding_mask) + if mask_indices is not None: + y = unmasked_features[mask_indices].view(unmasked_features.size(0), -1, unmasked_features.size(-1)) + else: + y = unmasked_features + else: + x = features + y = unmasked_features + mask_indices = None + + x = self.encoder(x, padding_mask=padding_mask) + + if features_only: + return {"x": x, "padding_mask": padding_mask} + + if self.quantizer: + q = self.quantizer(y, produce_targets=False) + y = q["x"] + num_vars = q["num_vars"] + code_ppl = q["code_perplexity"] + prob_ppl = q["prob_perplexity"] + curr_temp = q["temp"] + + y = self.project_q(y) + + if self.negatives_from_everywhere: + neg_cands, *_ = self.quantizer(unmasked_features, produce_targets=False) + negs, _ = self.sample_negatives(neg_cands, y.size(1)) + negs = self.project_q(negs) + + else: + negs, _ = self.sample_negatives(y, y.size(1)) + + if self.codebook_negatives > 0: + cb_negs = self.quantizer.sample_from_codebook( + y.size(0) * y.size(1), self.codebook_negatives + ) + cb_negs = cb_negs.view( + self.codebook_negatives, y.size(0), y.size(1), -1 + ) # order doesnt matter + cb_negs = self.project_q(cb_negs) + negs = torch.cat([negs, cb_negs], dim=0) + else: + y = self.project_q(y) + + if self.negatives_from_everywhere: + negs, _ = self.sample_negatives(unmasked_features, y.size(1)) + negs = self.project_q(negs) + else: + negs, _ = self.sample_negatives(y, y.size(1)) + + x = x[mask_indices].view(x.size(0), -1, x.size(-1)) + + if self.target_glu: + y = self.target_glu(y) + negs = self.target_glu(negs) + + x = self.final_proj(x) + x = self.compute_preds(x, y, negs) + + result = {"x": x, "padding_mask": padding_mask, "features_pen": features_pen} + + if prob_ppl is not None: + result["prob_perplexity"] = prob_ppl + result["code_perplexity"] = code_ppl + result["num_vars"] = num_vars + result["temp"] = curr_temp + + return result + + def quantize(self, x): + assert self.quantizer is not None + x = self.feature_extractor(x) + x = x.transpose(1, 2) + x = self.layer_norm(x) + return self.quantizer.forward_idx(x) + + def extract_features(self, source, padding_mask, mask=False): + res = self.forward(source, padding_mask, mask=mask, features_only=True) + return res["x"], res["padding_mask"] + + def get_logits(self, net_output): + logits = net_output["x"] + logits = logits.transpose(0, 2) + logits = logits.reshape(-1, logits.size(-1)) + return logits + + def get_targets(self, sample, net_output, expand_steps=True): + x = net_output["x"] + return x.new_zeros(x.size(1) * x.size(2), dtype=torch.long) + + def get_extra_losses(self, net_output): + pen = [] + + if "prob_perplexity" in net_output: + pen.append( + (net_output["num_vars"] - net_output["prob_perplexity"]) + / net_output["num_vars"] + ) + + if "features_pen" in net_output: + pen.append(net_output["features_pen"]) + + return pen + + def remove_pretraining_modules(self): + self.quantizer = None + self.project_q = None + self.target_glu = None + self.final_proj = None + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers: List[Tuple[int, int, int]], + dropout: float = 0.0, + mode: str = "default", + conv_bias: bool = False, + ): + super().__init__() + + assert mode in {"default", "layer_norm"} + + def block( + n_in, + n_out, + k, + stride, + is_layer_norm=False, + is_group_norm=False, + conv_bias=False, + ): + def make_conv(): + conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias) + nn.init.kaiming_normal_(conv.weight) + return conv + + assert ( + is_layer_norm and is_group_norm + ) == False, "layer norm and group norm are exclusive" + + if is_layer_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=True), + TransposeLast(), + ), + nn.GELU(), + ) + elif is_group_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + Fp32GroupNorm(dim, dim, affine=True), + nn.GELU(), + ) + else: + return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.GELU()) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for i, cl in enumerate(conv_layers): + assert len(cl) == 3, "invalid conv definition: " + str(cl) + (dim, k, stride) = cl + + self.conv_layers.append( + block( + in_d, + dim, + k, + stride, + is_layer_norm=mode == "layer_norm", + is_group_norm=mode == "default" and i == 0, + conv_bias=conv_bias, + ) + ) + in_d = dim + + def forward(self, x): + + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + x = conv(x) + + return x + + +class TransformerEncoder(nn.Module): + def __init__(self, args): + super().__init__() + + self.dropout = args.dropout + self.embedding_dim = args.encoder_embed_dim + + self.pos_conv = nn.Conv1d( + self.embedding_dim, + self.embedding_dim, + kernel_size=args.conv_pos, + padding=args.conv_pos // 2, + groups=args.conv_pos_groups, + ) + dropout = 0 + std = math.sqrt((4 * (1.0 - dropout)) / (args.conv_pos * self.embedding_dim)) + nn.init.normal_(self.pos_conv.weight, mean=0, std=std) + nn.init.constant_(self.pos_conv.bias, 0) + + self.pos_conv = nn.utils.weight_norm(self.pos_conv, name="weight", dim=2) + self.pos_conv = nn.Sequential(self.pos_conv, SamePad(args.conv_pos), nn.GELU()) + + self.layers = nn.ModuleList( + [ + TransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + ) + for _ in range(args.encoder_layers) + ] + ) + + self.layer_norm_first = args.layer_norm_first + self.layer_norm = LayerNorm(self.embedding_dim) + self.layerdrop = args.encoder_layerdrop + + self.apply(init_bert_params) + + def forward(self, x, padding_mask=None): + x = self.extract_features(x, padding_mask) + + if self.layer_norm_first: + x = self.layer_norm(x) + + return x + + def extract_features(self, x, padding_mask=None): + + x_conv = self.pos_conv(x.transpose(1, 2)) + x_conv = x_conv.transpose(1, 2) + x += x_conv + + if not self.layer_norm_first: + x = self.layer_norm(x) + + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + layer_results = [] + for i, layer in enumerate(self.layers): + dropout_probability = np.random.random() + if not self.training or (dropout_probability > self.layerdrop): + x, z = layer(x, self_attn_padding_mask=padding_mask, need_weights=False) + layer_results.append(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + return x + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.args.max_positions + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + +class TransformerSentenceEncoderLayer(nn.Module): + """ + Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__( + self, + embedding_dim: float = 768, + ffn_embedding_dim: float = 3072, + num_attention_heads: float = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + layer_norm_first: bool = False, + ) -> None: + + super().__init__() + # Initialize parameters + self.embedding_dim = embedding_dim + self.dropout = dropout + self.activation_dropout = activation_dropout + + # Initialize blocks + self.activation_fn = utils.get_activation_fn(activation_fn) + self.self_attn = MultiheadAttention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + self_attention=True, + ) + + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(self.activation_dropout) + self.dropout3 = nn.Dropout(dropout) + + self.layer_norm_first = layer_norm_first + + # layer norm associated with the self attention layer + self.self_attn_layer_norm = LayerNorm(self.embedding_dim) + self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim) + self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim) + + # layer norm associated with the position wise feed-forward NN + self.final_layer_norm = LayerNorm(self.embedding_dim) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + need_weights: bool = False, + att_args=None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer imlementation. + """ + residual = x + + if self.layer_norm_first: + x = self.self_attn_layer_norm(x) + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout1(x) + x = residual + x + + residual = x + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + x = self.dropout3(x) + x = residual + x + else: + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=need_weights, + ) + + x = self.dropout1(x) + x = residual + x + + x = self.self_attn_layer_norm(x) + + residual = x + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + x = self.dropout3(x) + x = residual + x + x = self.final_layer_norm(x) + + return x, attn + + +@register_model_architecture("wav2vec2", "wav2vec2") +def base_architecture(args): + args.extractor_mode = getattr(args, "extractor_mode", "default") + + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 768) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 3072) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 12) + + args.activation_fn = getattr(args, "activation_fn", "gelu") + + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + + args.final_dim = getattr(args, "final_dim", 0) + + args.layer_norm_first = getattr(args, "layer_norm_first", False) + args.encoder_layerdrop = getattr(args, "encoder_layerdrop", 0.0) + + conv_feature_layers = "[(512, 10, 5)]" + conv_feature_layers += " + [(512, 8, 4)]" + conv_feature_layers += " + [(512, 4, 2)] * 3" + conv_feature_layers += " + [(512, 1, 1)]" + args.conv_feature_layers = getattr(args, "conv_feature_layers", conv_feature_layers) + + args.logit_temp = getattr(args, "logit_temp", 0.1) + + args.quantize_targets = getattr(args, "quantize_targets", False) + args.quantize_input = getattr(args, "quantize_input", False) + + args.feature_grad_mult = getattr(args, "feature_grad_mult", 1.0) + + args.latent_vars = getattr(args, "latent_vars", 320) + args.latent_groups = getattr(args, "latent_groups", 2) + args.latent_dim = getattr(args, "latent_dim", 0) + + args.mask_length = getattr(args, "mask_length", 10) + args.mask_prob = getattr(args, "mask_prob", 0.65) + args.mask_selection = getattr(args, "mask_selection", "static") + args.mask_other = getattr(args, "mask_other", 0) + args.no_mask_overlap = getattr(args, "no_mask_overlap", False) + args.mask_min_space = getattr(args, "mask_min_space", 1) + + args.mask_channel_length = getattr(args, "mask_channel_length", 10) + args.mask_channel_prob = getattr(args, "mask_channel_prob", 0) + args.mask_channel_selection = getattr(args, "mask_channel_selection", "static") + args.mask_channel_other = getattr(args, "mask_channel_other", 0) + args.no_mask_channel_overlap = getattr(args, "no_mask_channel_overlap", False) + args.mask_channel_min_space = getattr(args, "mask_channel_min_space", 1) + + args.dropout_input = getattr(args, "dropout_input", 0) + args.dropout_features = getattr(args, "dropout_features", 0) + + args.num_negatives = getattr(args, "num_negatives", 100) + args.negatives_from_everywhere = getattr(args, "negatives_from_everywhere", False) + args.cross_sample_negatives = getattr(args, "cross_sample_negatives", 0) + args.codebook_negatives = getattr(args, "codebook_negatives", 0) + + args.conv_pos = getattr(args, "conv_pos", 128) + args.conv_pos_groups = getattr(args, "conv_pos_groups", 16) + + args.latent_temp = getattr(args, "latent_temp", "(2,0.5,0.999995)") + + args.target_glu = getattr(args, "target_glu", False) + + args.conv_bias = getattr(args, "conv_bias", False) diff --git a/fairseq/models/wav2vec/wav2vec2_asr.py b/fairseq/models/wav2vec/wav2vec2_asr.py new file mode 100644 index 0000000000000000000000000000000000000000..e47e1f700920a91d9364f7b62cb5859ddd4d37ae --- /dev/null +++ b/fairseq/models/wav2vec/wav2vec2_asr.py @@ -0,0 +1,673 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import copy +import math +import numpy as np + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import checkpoint_utils, tasks, utils + +from fairseq.models import ( + FairseqEncoder, + FairseqIncrementalDecoder, + FairseqEncoderDecoderModel, + BaseFairseqModel, + register_model, + register_model_architecture, +) +from fairseq.modules import LayerNorm, PositionalEmbedding, TransformerDecoderLayer + + +def add_common_args(parser): + parser.add_argument("--w2v-path", help="path to wav2vec 2.0 model") + parser.add_argument( + "--no-pretrained-weights", + action="store_true", + help="if true, does not load pretrained weights", + ) + parser.add_argument( + "--dropout-input", + type=float, + metavar="D", + help="dropout to apply to the input (after feat extr)", + ) + parser.add_argument( + "--final-dropout", + type=float, + metavar="D", + help="dropout after transformer and before final projection", + ) + parser.add_argument( + "--apply-mask", action="store_true", help="apply masking during fine-tuning" + ) + parser.add_argument( + "--dropout", + type=float, + metavar="D", + help="dropout probability inside wav2vec 2.0 model", + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights inside wav2vec 2.0 model", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN inside wav2vec 2.0 model", + ) + + parser.add_argument( + "--mask-length", type=int, help="repeat the mask indices multiple times" + ) + + parser.add_argument( + "--mask-prob", type=float, help="probability of replacing a token with mask" + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + + parser.add_argument( + "--mask-other", + type=float, + help="stdev of the mask length in case of 'normal' selection strategy", + ) + + parser.add_argument( + "--no-mask-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-channel-length", type=int, help="repeat the mask indices multiple times" + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + help="stdev of the mask length in case of 'normal' selection strategy", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--freeze-finetune-updates", + default=0, + type=int, + help="dont finetune wav2vec for this many updates", + ) + + parser.add_argument( + "--feature-grad-mult", + default=None, + type=float, + help="reset feature grad mult in wav2vec 2.0 to this", + ) + + parser.add_argument( + "--layerdrop", + default=0.0, + type=float, + help="probability of dropping a layer in wav2vec 2.0", + ) + + +@register_model("wav2vec_ctc") +class Wav2VecCtc(BaseFairseqModel): + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + add_common_args(parser) + + def __init__(self, w2v_encoder, args): + super().__init__() + self.w2v_encoder = w2v_encoder + self.args = args + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + base_architecture(args) + w2v_encoder = Wav2VecEncoder(args, task.target_dictionary) + return cls(w2v_encoder, args) + + def get_normalized_probs(self, net_output, log_probs): + """Get normalized probabilities (or log probs) from a net's output.""" + + logits = net_output["encoder_out"] + if log_probs: + return utils.log_softmax(logits.float(), dim=-1) + else: + return utils.softmax(logits.float(), dim=-1) + + def forward(self, **kwargs): + x = self.w2v_encoder(**kwargs) + return x + + # def max_positions(self): + # return None + + +@register_model("wav2vec_seq2seq") +class TransformerModel(FairseqEncoderDecoderModel): + def __init__(self, args, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + add_common_args(parser) + + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-layerdrop", + type=float, + metavar="D", + help="decoder layerdrop chance", + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads", + ) + parser.add_argument( + "--decoder-learned-pos", + action="store_true", + help="use learned positional embeddings in the decoder", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--no-token-positional-embeddings", + default=False, + action="store_true", + help="if set, disables positional embeddings (outside self attention)", + ) + + parser.add_argument( + "--decoder-dropout", + type=float, + metavar="D", + help="dropout probability in the decoder", + ) + parser.add_argument( + "--decoder-attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights inside the decoder", + ) + parser.add_argument( + "--decoder-activation-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN inside the decoder", + ) + + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if not hasattr(args, "max_source_positions"): + args.max_source_positions = 2048 + if not hasattr(args, "max_target_positions"): + args.max_target_positions = 2048 + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + return emb + + decoder_embed_tokens = build_embedding(tgt_dict, args.decoder_embed_dim) + + encoder = cls.build_encoder(args) + decoder = cls.build_decoder(args, tgt_dict, decoder_embed_tokens) + return TransformerModel(args, encoder, decoder) + + @classmethod + def build_encoder(cls, args): + return Wav2VecEncoder(args) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return TransformerDecoder(args, tgt_dict, embed_tokens) + + def forward(self, **kwargs): + encoder_out = self.encoder(tbc=False, **kwargs) + decoder_out = self.decoder(encoder_out=encoder_out, **kwargs) + return decoder_out + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + +class Wav2VecEncoder(FairseqEncoder): + def __init__(self, args, tgt_dict=None): + self.apply_mask = args.apply_mask + + arg_overrides = { + "dropout": args.dropout, + "activation_dropout": args.activation_dropout, + "dropout_input": args.dropout_input, + "attention_dropout": args.attention_dropout, + "mask_length": args.mask_length, + "mask_prob": args.mask_prob, + "mask_selection": args.mask_selection, + "mask_other": args.mask_other, + "no_mask_overlap": args.no_mask_overlap, + "mask_channel_length": args.mask_channel_length, + "mask_channel_prob": args.mask_channel_prob, + "mask_channel_selection": args.mask_channel_selection, + "mask_channel_other": args.mask_channel_other, + "no_mask_channel_overlap": args.no_mask_channel_overlap, + "encoder_layerdrop": args.layerdrop, + "feature_grad_mult": args.feature_grad_mult, + } + + if getattr(args, "w2v_args", None) is None: + state = checkpoint_utils.load_checkpoint_to_cpu( + args.w2v_path, arg_overrides + ) + w2v_args = state["args"] + else: + state = None + w2v_args = args.w2v_args + + assert args.normalize == w2v_args.normalize, 'Fine-tuning works best when data normalization is the same' + + w2v_args.data = args.data + task = tasks.setup_task(w2v_args) + model = task.build_model(w2v_args) + + if state is not None and not args.no_pretrained_weights: + model.load_state_dict(state["model"], strict=True) + + model.remove_pretraining_modules() + + super().__init__(task.source_dictionary) + + d = w2v_args.encoder_embed_dim + + self.w2v_model = model + + self.final_dropout = nn.Dropout(args.final_dropout) + self.freeze_finetune_updates = args.freeze_finetune_updates + self.num_updates = 0 + + if tgt_dict is not None: + self.proj = Linear(d, len(tgt_dict)) + elif getattr(args, 'decoder_embed_dim', d) != d: + self.proj = Linear(d, args.decoder_embed_dim) + else: + self.proj = None + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + super().set_num_updates(num_updates) + self.num_updates = num_updates + + def forward(self, source, padding_mask, tbc=True, **kwargs): + + w2v_args = { + "source": source, + "padding_mask": padding_mask, + "mask": self.apply_mask and self.training, + } + + ft = self.freeze_finetune_updates <= self.num_updates + + with torch.no_grad() if not ft else contextlib.ExitStack(): + x, padding_mask = self.w2v_model.extract_features(**w2v_args) + + if tbc: + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + x = self.final_dropout(x) + + if self.proj: + x = self.proj(x) + + return { + "encoder_out": x, # T x B x C + "encoder_padding_mask": padding_mask, # B x T + "padding_mask": padding_mask, + } + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return None + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +class TransformerDecoder(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__(dictionary) + + self.dropout = args.decoder_dropout + self.share_input_output_embed = args.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = args.decoder_embed_dim + self.output_embed_dim = args.decoder_embed_dim + args.encoder_embed_dim = embed_dim + + self.layerdrop = args.decoder_layerdrop + + padding_idx = embed_tokens.padding_idx + self.max_target_positions = args.max_target_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + args.max_target_positions, + embed_dim, + padding_idx, + learned=args.decoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + + args = copy.deepcopy(args) + args.dropout = args.decoder_dropout + args.attention_dropout = args.decoder_attention_dropout + args.activation_dropout = args.decoder_activation_dropout + + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(args.decoder_layers) + ] + ) + + if not self.share_input_output_embed: + self.embed_out = nn.Parameter( + torch.Tensor(len(dictionary), self.output_embed_dim) + ) + nn.init.normal_(self.embed_out, mean=0, std=self.output_embed_dim ** -0.5) + + if args.decoder_normalize_before and not getattr( + args, "no_decoder_final_norm", False + ): + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + prev_output_tokens = prev_output_tokens.long() + x, extra = self.extract_features( + prev_output_tokens, encoder_out, incremental_state + ) + x = self.output_layer(x) + return x, extra + + def extract_features( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + + # embed positions + positions = ( + self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + + inner_states = [x] + + # decoder layers + for layer in self.layers: + dropout_probability = np.random.random() + if not self.training or (dropout_probability > self.layerdrop): + x, attn, _ = layer( + x, + encoder_out["encoder_out"] if encoder_out is not None else None, + encoder_out["encoder_padding_mask"] + if encoder_out is not None + else None, + incremental_state, + self_attn_mask=self.buffered_future_mask(x) + if incremental_state is None + else None, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + return x, {"attn": attn, "inner_states": inner_states} + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + # project back to size of vocabulary + if self.share_input_output_embed: + return F.linear(features, self.embed_tokens.weight) + else: + return F.linear(features, self.embed_out) + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m + + +@register_model_architecture("wav2vec_ctc", "wav2vec_ctc") +def base_architecture(args): + args.no_pretrained_weights = getattr(args, "no_pretrained_weights", False) + args.dropout_input = getattr(args, "dropout_input", 0) + args.final_dropout = getattr(args, "final_dropout", 0) + args.apply_mask = getattr(args, "apply_mask", False) + args.dropout = getattr(args, "dropout", 0) + args.attention_dropout = getattr(args, "attention_dropout", 0) + args.activation_dropout = getattr(args, "activation_dropout", 0) + + args.mask_length = getattr(args, "mask_length", 10) + args.mask_prob = getattr(args, "mask_prob", 0.5) + args.mask_selection = getattr(args, "mask_selection", "static") + args.mask_other = getattr(args, "mask_other", 0) + args.no_mask_overlap = getattr(args, "no_mask_overlap", False) + args.mask_channel_length = getattr(args, "mask_channel_length", 10) + args.mask_channel_prob = getattr(args, "mask_channel_prob", 0.5) + args.mask_channel_selection = getattr(args, "mask_channel_selection", "static") + args.mask_channel_other = getattr(args, "mask_channel_other", 0) + args.no_mask_channel_overlap = getattr(args, "no_mask_channel_overlap", False) + + args.freeze_finetune_updates = getattr(args, "freeze_finetune_updates", 0) + args.feature_grad_mult = getattr(args, "feature_grad_mult", 0) + args.layerdrop = getattr(args, "layerdrop", 0.0) + + +@register_model_architecture("wav2vec_seq2seq", "wav2vec_seq2seq") +def seq2seq_architecture(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_layers = getattr(args, "decoder_layers", 10) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.decoder_dropout = getattr(args, "decoder_dropout", 0) + args.decoder_attention_dropout = getattr(args, "decoder_attention_dropout", 0) + args.decoder_activation_dropout = getattr(args, "decoder_activation_dropout", 0) + args.share_decoder_input_output_embed = getattr(args, "share_decoder_input_output_embed", False) + + base_architecture(args) diff --git a/fairseq/modules/__init__.py b/fairseq/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..d526d4a92e706f386d815af5303636fc5f955789 --- /dev/null +++ b/fairseq/modules/__init__.py @@ -0,0 +1,75 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .adaptive_input import AdaptiveInput +from .adaptive_softmax import AdaptiveSoftmax +from .beamable_mm import BeamableMM +from .character_token_embedder import CharacterTokenEmbedder +from .conv_tbc import ConvTBC +from .cross_entropy import cross_entropy +from .downsampled_multihead_attention import DownsampledMultiHeadAttention +from .dynamic_convolution import DynamicConv, DynamicConv1dTBC +from .dynamic_crf_layer import DynamicCRF +from .fairseq_dropout import FairseqDropout +from .fp32_group_norm import Fp32GroupNorm +from .gelu import gelu, gelu_accurate +from .grad_multiply import GradMultiply +from .gumbel_vector_quantizer import GumbelVectorQuantizer +from .kmeans_vector_quantizer import KmeansVectorQuantizer +from .layer_drop import LayerDropModuleList +from .layer_norm import Fp32LayerNorm, LayerNorm +from .learned_positional_embedding import LearnedPositionalEmbedding +from .lightweight_convolution import LightweightConv, LightweightConv1dTBC +from .linearized_convolution import LinearizedConvolution +from .multihead_attention import MultiheadAttention +from .positional_embedding import PositionalEmbedding +from .same_pad import SamePad +from .scalar_bias import ScalarBias +from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding +from .transformer_sentence_encoder_layer import TransformerSentenceEncoderLayer +from .transformer_sentence_encoder import TransformerSentenceEncoder +from .transpose_last import TransposeLast +from .unfold import unfold1d +from .transformer_layer import TransformerDecoderLayer, TransformerEncoderLayer +from .vggblock import VGGBlock + +__all__ = [ + 'AdaptiveInput', + 'AdaptiveSoftmax', + 'BeamableMM', + 'CharacterTokenEmbedder', + 'ConvTBC', + 'cross_entropy', + 'DownsampledMultiHeadAttention', + 'DynamicConv1dTBC', + 'DynamicConv', + 'DynamicCRF', + 'FairseqDropout', + 'Fp32GroupNorm', + 'Fp32LayerNorm', + 'gelu', + 'gelu_accurate', + 'GradMultiply', + 'GumbelVectorQuantizer', + 'KmeansVectorQuantizer', + 'LayerDropModuleList', + 'LayerNorm', + 'LearnedPositionalEmbedding', + 'LightweightConv1dTBC', + 'LightweightConv', + 'LinearizedConvolution', + 'MultiheadAttention', + 'PositionalEmbedding', + 'SamePad', + 'ScalarBias', + 'SinusoidalPositionalEmbedding', + 'TransformerSentenceEncoderLayer', + 'TransformerSentenceEncoder', + 'TransformerDecoderLayer', + 'TransformerEncoderLayer', + 'TransposeLast', + 'VGGBlock', + 'unfold1d', +] diff --git a/fairseq/modules/__pycache__/__init__.cpython-310.pyc b/fairseq/modules/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b7647b49cf378406593fe979970afac6cba61cee Binary files /dev/null and b/fairseq/modules/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/adaptive_input.cpython-310.pyc b/fairseq/modules/__pycache__/adaptive_input.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0ea83f226dd60b971007625a5ab3e9886eb5a04e Binary files /dev/null and b/fairseq/modules/__pycache__/adaptive_input.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/adaptive_softmax.cpython-310.pyc b/fairseq/modules/__pycache__/adaptive_softmax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..45d089a17c1af4e5a7290c0c871b8d2ace7bb21c Binary files /dev/null and b/fairseq/modules/__pycache__/adaptive_softmax.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/beamable_mm.cpython-310.pyc b/fairseq/modules/__pycache__/beamable_mm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f2260aed078d5243054bfa8424c2c040a70adb55 Binary files /dev/null and b/fairseq/modules/__pycache__/beamable_mm.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/character_token_embedder.cpython-310.pyc b/fairseq/modules/__pycache__/character_token_embedder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b83ee96b53db529ee2668eaf9bd385ee6d8d3444 Binary files /dev/null and b/fairseq/modules/__pycache__/character_token_embedder.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/conv_tbc.cpython-310.pyc b/fairseq/modules/__pycache__/conv_tbc.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6d6ca22fdaddf541c5f6cc27a97421003810f4d1 Binary files /dev/null and b/fairseq/modules/__pycache__/conv_tbc.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/cross_entropy.cpython-310.pyc b/fairseq/modules/__pycache__/cross_entropy.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1feb46386f0bc977010f5324d8468225b9059d54 Binary files /dev/null and b/fairseq/modules/__pycache__/cross_entropy.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/downsampled_multihead_attention.cpython-310.pyc b/fairseq/modules/__pycache__/downsampled_multihead_attention.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b169b3d38845383da383e9642d838d7b85d53c87 Binary files /dev/null and b/fairseq/modules/__pycache__/downsampled_multihead_attention.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/dynamic_convolution.cpython-310.pyc b/fairseq/modules/__pycache__/dynamic_convolution.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8c2d3e56b70e60cdf03a5fadb372eab5be8426b4 Binary files /dev/null and b/fairseq/modules/__pycache__/dynamic_convolution.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/dynamic_crf_layer.cpython-310.pyc b/fairseq/modules/__pycache__/dynamic_crf_layer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cd1b970ae395a7b860014dacba319e4ea8142b10 Binary files /dev/null and b/fairseq/modules/__pycache__/dynamic_crf_layer.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/fairseq_dropout.cpython-310.pyc b/fairseq/modules/__pycache__/fairseq_dropout.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0170b7a20d98bd77900f16becb9cf60a4bec7b08 Binary files /dev/null and b/fairseq/modules/__pycache__/fairseq_dropout.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/fp32_group_norm.cpython-310.pyc b/fairseq/modules/__pycache__/fp32_group_norm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..41dd550730c8c90f10cd89c6de32125298f4eb14 Binary files /dev/null and b/fairseq/modules/__pycache__/fp32_group_norm.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/gelu.cpython-310.pyc b/fairseq/modules/__pycache__/gelu.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..405941480b21b80d974ca02b95d0f032f3e727a0 Binary files /dev/null and b/fairseq/modules/__pycache__/gelu.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/grad_multiply.cpython-310.pyc b/fairseq/modules/__pycache__/grad_multiply.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ffe9ea33667d5930ab61c657d98bca13e1748297 Binary files /dev/null and b/fairseq/modules/__pycache__/grad_multiply.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/gumbel_vector_quantizer.cpython-310.pyc b/fairseq/modules/__pycache__/gumbel_vector_quantizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..19263504dc7d3d30f2ffe62dc88160733063243b Binary files /dev/null and b/fairseq/modules/__pycache__/gumbel_vector_quantizer.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/kmeans_vector_quantizer.cpython-310.pyc b/fairseq/modules/__pycache__/kmeans_vector_quantizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f5c3a9c0cad80373e4f5c2db0ce2b90162b3fc0e Binary files /dev/null and b/fairseq/modules/__pycache__/kmeans_vector_quantizer.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/layer_drop.cpython-310.pyc b/fairseq/modules/__pycache__/layer_drop.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a1fc4d693222120b4542e7855cc334647bb34fdc Binary files /dev/null and b/fairseq/modules/__pycache__/layer_drop.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/layer_norm.cpython-310.pyc b/fairseq/modules/__pycache__/layer_norm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..01a31d99764aa6f86a22861e0d7337064c1ebb35 Binary files /dev/null and b/fairseq/modules/__pycache__/layer_norm.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/learned_positional_embedding.cpython-310.pyc b/fairseq/modules/__pycache__/learned_positional_embedding.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8df51332596499ec34c1fadbb3e7b3a2b74a98b3 Binary files /dev/null and b/fairseq/modules/__pycache__/learned_positional_embedding.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/lightweight_convolution.cpython-310.pyc b/fairseq/modules/__pycache__/lightweight_convolution.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eae3ab8b823c400254c4f882fc166e60e55a8c1d Binary files /dev/null and b/fairseq/modules/__pycache__/lightweight_convolution.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/linearized_convolution.cpython-310.pyc b/fairseq/modules/__pycache__/linearized_convolution.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4e4bb6eaafe65009e5f4342df0b48452ddb000fa Binary files /dev/null and b/fairseq/modules/__pycache__/linearized_convolution.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/multihead_attention.cpython-310.pyc b/fairseq/modules/__pycache__/multihead_attention.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..25b27da4cd349c39cb4a32df9583cbe975c5bdbc Binary files /dev/null and b/fairseq/modules/__pycache__/multihead_attention.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/positional_embedding.cpython-310.pyc b/fairseq/modules/__pycache__/positional_embedding.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a1d0abd91491a14d982136090e199c4f551d355a Binary files /dev/null and b/fairseq/modules/__pycache__/positional_embedding.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/quant_noise.cpython-310.pyc b/fairseq/modules/__pycache__/quant_noise.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3bfa36feab29401b7c09abcd257e9af3727b2dea Binary files /dev/null and b/fairseq/modules/__pycache__/quant_noise.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/same_pad.cpython-310.pyc b/fairseq/modules/__pycache__/same_pad.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..63dc1d3862ca5d731a5bc44ca6cf51b3840dc945 Binary files /dev/null and b/fairseq/modules/__pycache__/same_pad.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/scalar_bias.cpython-310.pyc b/fairseq/modules/__pycache__/scalar_bias.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c71437cb379d5f42e07bccbeb714396f5a9ada78 Binary files /dev/null and b/fairseq/modules/__pycache__/scalar_bias.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/sinusoidal_positional_embedding.cpython-310.pyc b/fairseq/modules/__pycache__/sinusoidal_positional_embedding.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bc3a680863f1e7d932cfc85b397887a84328ee4d Binary files /dev/null and b/fairseq/modules/__pycache__/sinusoidal_positional_embedding.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/transformer_layer.cpython-310.pyc b/fairseq/modules/__pycache__/transformer_layer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e0cda59153fdef6be6be508ef84628eaa5c9eeb4 Binary files /dev/null and b/fairseq/modules/__pycache__/transformer_layer.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/transformer_sentence_encoder.cpython-310.pyc b/fairseq/modules/__pycache__/transformer_sentence_encoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8f3cc1839eb1c65842488302c76d08396b85b261 Binary files /dev/null and b/fairseq/modules/__pycache__/transformer_sentence_encoder.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/transformer_sentence_encoder_layer.cpython-310.pyc b/fairseq/modules/__pycache__/transformer_sentence_encoder_layer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..37a64b8aaf1ab2d0750dd554313c49e30a45b78f Binary files /dev/null and b/fairseq/modules/__pycache__/transformer_sentence_encoder_layer.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/transpose_last.cpython-310.pyc b/fairseq/modules/__pycache__/transpose_last.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b6f08e03254f64bb4725adb4315130988d763c62 Binary files /dev/null and b/fairseq/modules/__pycache__/transpose_last.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/unfold.cpython-310.pyc b/fairseq/modules/__pycache__/unfold.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e0c86516990ff72131faf40fa7ece0f3e5348cf7 Binary files /dev/null and b/fairseq/modules/__pycache__/unfold.cpython-310.pyc differ diff --git a/fairseq/modules/__pycache__/vggblock.cpython-310.pyc b/fairseq/modules/__pycache__/vggblock.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e06080e8b4223e14217264fa10ca8b00ce902308 Binary files /dev/null and b/fairseq/modules/__pycache__/vggblock.cpython-310.pyc differ diff --git a/fairseq/modules/adaptive_input.py b/fairseq/modules/adaptive_input.py new file mode 100644 index 0000000000000000000000000000000000000000..4cfe8fca6605f4cb5a4b0134d45c362acd60e67c --- /dev/null +++ b/fairseq/modules/adaptive_input.py @@ -0,0 +1,78 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import torch +from torch import nn +from fairseq.modules.quant_noise import quant_noise + +from typing import List + + +class AdaptiveInput(nn.Module): + + def __init__( + self, + vocab_size: int, + padding_idx: int, + initial_dim: int, + factor: float, + output_dim: int, + cutoff: List[int], + q_noise: float = 0, + qn_block_size: int = 8, + ): + super().__init__() + + if vocab_size > cutoff[-1]: + cutoff = cutoff + [vocab_size] + else: + assert vocab_size == cutoff[ + -1], 'cannot specify cutoff larger than vocab size' + + self.cutoff = cutoff + self.embedding_dim = output_dim + self.padding_idx = padding_idx + + self.embeddings = nn.ModuleList() + for i in range(len(self.cutoff)): + prev = self.cutoff[i - 1] if i > 0 else 0 + size = self.cutoff[i] - prev + dim = int(initial_dim // (factor ** i)) + seq = nn.Sequential( + nn.Embedding(size, dim, self.padding_idx), + quant_noise(nn.Linear(dim, output_dim, bias=False), q_noise, qn_block_size), + ) + + self.embeddings.append(seq) + self.padding_idx = None + self.padding_idx = padding_idx + + def init_weights(m): + if isinstance(m, nn.Embedding): + nn.init.normal_(m.weight, mean=0, std=m.weight.shape[1] ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + elif hasattr(m, 'weight'): + nn.init.xavier_uniform_(m.weight) + + self.apply(init_weights) + + self.register_buffer('_float_tensor', torch.FloatTensor(1)) + + def weights_for_band(self, band: int): + return self.embeddings[band][0].weight, self.embeddings[band][1].weight + + def forward(self, input: torch.Tensor): + result = self._float_tensor.new(input.shape + (self.embedding_dim,)) + for i in range(len(self.cutoff)): + mask = input.lt(self.cutoff[i]) + if i > 0: + mask.mul_(input.ge(self.cutoff[i - 1])) + chunk_input = input[mask] - self.cutoff[i - 1] + else: + chunk_input = input[mask] + if mask.any(): + result[mask] = self.embeddings[i](chunk_input) + return result diff --git a/fairseq/modules/adaptive_softmax.py b/fairseq/modules/adaptive_softmax.py new file mode 100644 index 0000000000000000000000000000000000000000..96f8b75ad34da2bc2d0048babbd74034764938ea --- /dev/null +++ b/fairseq/modules/adaptive_softmax.py @@ -0,0 +1,214 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import operator +import functools + +import torch +import torch.nn.functional as F +from fairseq.modules.quant_noise import quant_noise +from fairseq.modules.fairseq_dropout import FairseqDropout +from torch import nn + + +class TiedLinear(nn.Module): + def __init__(self, weight, transpose): + super().__init__() + self.weight = weight + self.transpose = transpose + + def forward(self, input): + return F.linear(input, self.weight.t() if self.transpose else self.weight) + + +class TiedHeadModule(nn.Module): + def __init__(self, weights, input_dim, num_classes, q_noise, qn_block_size): + super().__init__() + tied_emb, _ = weights + self.num_words, emb_dim = tied_emb.size() + + self.word_proj = quant_noise(TiedLinear(tied_emb, transpose=False), q_noise, qn_block_size) + if input_dim != emb_dim: + self.word_proj = nn.Sequential( + quant_noise(nn.Linear(input_dim, emb_dim, bias=False), q_noise, qn_block_size), + self.word_proj, + ) + + self.class_proj = quant_noise(nn.Linear(input_dim, num_classes, bias=False), q_noise, qn_block_size) + self.out_dim = self.num_words + num_classes + + self.register_buffer('_float_tensor', torch.FloatTensor(1)) + + def forward(self, input): + inp_sz = functools.reduce(operator.mul, input.shape[:-1], 1) + out = self._float_tensor.new(inp_sz, self.out_dim) + out[:, :self.num_words] = self.word_proj(input.view(inp_sz, -1)) + out[:, self.num_words:] = self.class_proj(input.view(inp_sz, -1)) + return out + + +class AdaptiveSoftmax(nn.Module): + """ + This is an implementation of the efficient softmax approximation for + graphical processing units (GPU), described in the paper "Efficient softmax + approximation for GPUs" (http://arxiv.org/abs/1609.04309). + """ + + def __init__(self, vocab_size, input_dim, cutoff, dropout, factor=4., adaptive_inputs=None, tie_proj=False, + q_noise=0, qn_block_size=8): + super().__init__() + + if vocab_size > cutoff[-1]: + cutoff = cutoff + [vocab_size] + else: + assert vocab_size == cutoff[ + -1], 'cannot specify cutoff larger than vocab size' + + output_dim = cutoff[0] + len(cutoff) - 1 + + self.vocab_size = vocab_size + self.cutoff = cutoff + self.dropout_module = FairseqDropout(dropout, module_name=self.__class__.__name__) + self.input_dim = input_dim + self.factor = factor + self.q_noise = q_noise + self.qn_block_size = qn_block_size + + self.lsm = nn.LogSoftmax(dim=1) + + if adaptive_inputs is not None: + self.head = TiedHeadModule(adaptive_inputs.weights_for_band(0), input_dim, len(cutoff) - 1, self.q_noise, self.qn_block_size) + else: + self.head = quant_noise(nn.Linear(input_dim, output_dim, bias=False), self.q_noise, self.qn_block_size) + + self._make_tail(adaptive_inputs, tie_proj) + + def init_weights(m): + if hasattr(m, 'weight') and not isinstance(m, TiedLinear) and not isinstance(m, TiedHeadModule): + nn.init.xavier_uniform_(m.weight) + + self.apply(init_weights) + + self.register_buffer('version', torch.LongTensor([1])) + + def _make_tail(self, adaptive_inputs=None, tie_proj=False): + self.tail = nn.ModuleList() + for i in range(len(self.cutoff) - 1): + dim = int(self.input_dim // self.factor ** (i + 1)) + + tied_emb, tied_proj = adaptive_inputs.weights_for_band(i + 1) \ + if adaptive_inputs is not None else (None, None) + + if tied_proj is not None: + if tie_proj: + proj = quant_noise(TiedLinear(tied_proj, transpose=True), self.q_noise, self.qn_block_size) + else: + proj = quant_noise(nn.Linear(tied_proj.size(0), tied_proj.size(1), bias=False), self.q_noise, self.qn_block_size) + else: + proj = quant_noise(nn.Linear(self.input_dim, dim, bias=False), self.q_noise, self.qn_block_size) + + if tied_emb is None: + out_proj = nn.Linear(dim, self.cutoff[i + 1] - self.cutoff[i], bias=False) + else: + out_proj = TiedLinear(tied_emb, transpose=False) + + m = nn.Sequential( + proj, + nn.Dropout(self.dropout_module.p), + quant_noise(out_proj, self.q_noise, self.qn_block_size), + ) + + self.tail.append(m) + + def upgrade_state_dict_named(self, state_dict, name): + version_name = name + '.version' + if version_name not in state_dict: + raise Exception('This version of the model is no longer supported') + + def adapt_target(self, target): + """ + In order to be efficient, the AdaptiveSoftMax does not compute the + scores for all the word of the vocabulary for all the examples. It is + thus necessary to call the method adapt_target of the AdaptiveSoftMax + layer inside each forward pass. + """ + + target = target.view(-1) + new_target = [target.clone()] + target_idxs = [] + + for i in range(len(self.cutoff) - 1): + mask = target.ge(self.cutoff[i]).mul(target.lt(self.cutoff[i + 1])) + new_target[0][mask] = self.cutoff[0] + i + + if mask.any(): + target_idxs.append(mask.nonzero().squeeze(1)) + new_target.append(target[mask].add(-self.cutoff[i])) + else: + target_idxs.append(None) + new_target.append(None) + + return new_target, target_idxs + + def forward(self, input, target): + """ + Args: + input: (b x t x d) + target: (b x t) + Returns: + 2 lists: output for each cutoff section and new targets by cut off + """ + + input = input.contiguous().view(-1, input.size(-1)) + input = self.dropout_module(input) + + new_target, target_idxs = self.adapt_target(target) + output = [self.head(input)] + + for i in range(len(target_idxs)): + if target_idxs[i] is not None: + output.append(self.tail[i](input.index_select(0, target_idxs[i]))) + else: + output.append(None) + + return output, new_target + + def get_log_prob(self, input, target): + """ + Computes the log probabilities for all the words of the vocabulary, + given a 2D tensor of hidden vectors. + """ + + bsz, length, dim = input.size() + input = input.contiguous().view(-1, dim) + + if target is not None: + _, target_idxs = self.adapt_target(target) + else: + target_idxs = None + + head_y = self.head(input) + log_probs = head_y.new_zeros(input.size(0), self.vocab_size) + + head_sz = self.cutoff[0] + len(self.tail) + log_probs[:, :head_sz] = self.lsm(head_y) + tail_priors = log_probs[:, self.cutoff[0]: head_sz].clone() + + for i in range(len(self.tail)): + start = self.cutoff[i] + end = self.cutoff[i + 1] + + if target_idxs is None: + tail_out = log_probs[:, start:end] + tail_out.copy_(self.tail[i](input)) + log_probs[:, start:end] = self.lsm(tail_out).add_(tail_priors[:, i, None]) + elif target_idxs[i] is not None: + idxs = target_idxs[i] + tail_out = log_probs[idxs, start:end] + tail_out.copy_(self.tail[i](input[idxs])) + log_probs[idxs, start:end] = self.lsm(tail_out).add_(tail_priors[idxs, i, None]) + + log_probs = log_probs.view(bsz, length, -1) + return log_probs diff --git a/fairseq/modules/beamable_mm.py b/fairseq/modules/beamable_mm.py new file mode 100644 index 0000000000000000000000000000000000000000..df77105a946258a58f3615fd51f399454223c0e2 --- /dev/null +++ b/fairseq/modules/beamable_mm.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn + + +class BeamableMM(nn.Module): + """This module provides an optimized MM for beam decoding with attention. + + It leverage the fact that the source-side of the input is replicated beam + times and the target-side of the input is of width one. This layer speeds up + inference by replacing the inputs {(bsz x 1 x nhu), (bsz x sz2 x nhu)} + with smaller inputs {(bsz/beam x beam x nhu), (bsz/beam x sz2 x nhu)}. + """ + def __init__(self, beam_size=None): + super(BeamableMM, self).__init__() + self.beam_size = beam_size + + def forward(self, input1, input2): + if ( + not self.training and # test mode + self.beam_size is not None and # beam size is set + input1.dim() == 3 and # only support batched input + input1.size(1) == 1 # single time step update + ): + bsz, beam = input1.size(0), self.beam_size + + # bsz x 1 x nhu --> bsz/beam x beam x nhu + input1 = input1[:, 0, :].unfold(0, beam, beam).transpose(2, 1) + + # bsz x sz2 x nhu --> bsz/beam x sz2 x nhu + input2 = input2.unfold(0, beam, beam)[:, :, :, 0] + + # use non batched operation if bsz = beam + if input1.size(0) == 1: + output = torch.mm(input1[0, :, :], input2[0, :, :]) + else: + output = input1.bmm(input2) + return output.view(bsz, 1, -1) + else: + return input1.bmm(input2) + + def set_beam_size(self, beam_size): + self.beam_size = beam_size diff --git a/fairseq/modules/character_token_embedder.py b/fairseq/modules/character_token_embedder.py new file mode 100644 index 0000000000000000000000000000000000000000..3abdaf4f28a5affb505a40cf07f5d95f550cceb4 --- /dev/null +++ b/fairseq/modules/character_token_embedder.py @@ -0,0 +1,204 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import List, Tuple + +import torch +from torch import nn +import torch.nn.functional as F + +from fairseq.data import Dictionary + +CHAR_PAD_IDX = 0 +CHAR_EOS_IDX = 257 + + +logger = logging.getLogger(__name__) + + +class CharacterTokenEmbedder(torch.nn.Module): + def __init__( + self, + vocab: Dictionary, + filters: List[Tuple[int, int]], + char_embed_dim: int, + word_embed_dim: int, + highway_layers: int, + max_char_len: int = 50, + char_inputs: bool = False + ): + super(CharacterTokenEmbedder, self).__init__() + + self.onnx_trace = False + self.embedding_dim = word_embed_dim + self.max_char_len = max_char_len + self.char_embeddings = nn.Embedding(257, char_embed_dim, padding_idx=0) + self.symbol_embeddings = nn.Parameter(torch.FloatTensor(2, word_embed_dim)) + self.eos_idx, self.unk_idx = 0, 1 + self.char_inputs = char_inputs + + self.convolutions = nn.ModuleList() + for width, out_c in filters: + self.convolutions.append( + nn.Conv1d(char_embed_dim, out_c, kernel_size=width) + ) + + last_dim = sum(f[1] for f in filters) + + self.highway = Highway(last_dim, highway_layers) if highway_layers > 0 else None + + self.projection = nn.Linear(last_dim, word_embed_dim) + + assert vocab is not None or char_inputs, "vocab must be set if not using char inputs" + self.vocab = None + if vocab is not None: + self.set_vocab(vocab, max_char_len) + + self.reset_parameters() + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def set_vocab(self, vocab, max_char_len): + word_to_char = torch.LongTensor(len(vocab), max_char_len) + + truncated = 0 + for i in range(len(vocab)): + if i < vocab.nspecial: + char_idxs = [0] * max_char_len + else: + chars = vocab[i].encode() + # +1 for padding + char_idxs = [c + 1 for c in chars] + [0] * (max_char_len - len(chars)) + if len(char_idxs) > max_char_len: + truncated += 1 + char_idxs = char_idxs[:max_char_len] + word_to_char[i] = torch.LongTensor(char_idxs) + + if truncated > 0: + logger.info('truncated {} words longer than {} characters'.format(truncated, max_char_len)) + + self.vocab = vocab + self.word_to_char = word_to_char + + @property + def padding_idx(self): + return Dictionary().pad() if self.vocab is None else self.vocab.pad() + + def reset_parameters(self): + nn.init.xavier_normal_(self.char_embeddings.weight) + nn.init.xavier_normal_(self.symbol_embeddings) + nn.init.xavier_uniform_(self.projection.weight) + + nn.init.constant_(self.char_embeddings.weight[self.char_embeddings.padding_idx], 0.) + nn.init.constant_(self.projection.bias, 0.) + + def forward( + self, + input: torch.Tensor, + ): + if self.char_inputs: + chars = input.view(-1, self.max_char_len) + pads = chars[:, 0].eq(CHAR_PAD_IDX) + eos = chars[:, 0].eq(CHAR_EOS_IDX) + if eos.any(): + if self.onnx_trace: + chars = torch.where(eos.unsqueeze(1), chars.new_zeros(1), chars) + else: + chars[eos] = 0 + + unk = None + else: + flat_words = input.view(-1) + chars = self.word_to_char[flat_words.type_as(self.word_to_char)].type_as(input) + pads = flat_words.eq(self.vocab.pad()) + eos = flat_words.eq(self.vocab.eos()) + unk = flat_words.eq(self.vocab.unk()) + + word_embs = self._convolve(chars) + if self.onnx_trace: + if pads.any(): + word_embs = torch.where(pads.unsqueeze(1), word_embs.new_zeros(1), word_embs) + if eos.any(): + word_embs = torch.where(eos.unsqueeze(1), self.symbol_embeddings[self.eos_idx], word_embs) + if unk is not None and unk.any(): + word_embs = torch.where(unk.unsqueeze(1), self.symbol_embeddings[self.unk_idx], word_embs) + else: + if pads.any(): + word_embs[pads] = 0 + if eos.any(): + word_embs[eos] = self.symbol_embeddings[self.eos_idx] + if unk is not None and unk.any(): + word_embs[unk] = self.symbol_embeddings[self.unk_idx] + + return word_embs.view(input.size()[:2] + (-1,)) + + def _convolve( + self, + char_idxs: torch.Tensor, + ): + char_embs = self.char_embeddings(char_idxs) + char_embs = char_embs.transpose(1, 2) # BTC -> BCT + + conv_result = [] + + for conv in self.convolutions: + x = conv(char_embs) + x, _ = torch.max(x, -1) + x = F.relu(x) + conv_result.append(x) + + x = torch.cat(conv_result, dim=-1) + + if self.highway is not None: + x = self.highway(x) + x = self.projection(x) + + return x + + +class Highway(torch.nn.Module): + """ + A `Highway layer `_. + Adopted from the AllenNLP implementation. + """ + + def __init__( + self, + input_dim: int, + num_layers: int = 1 + ): + super(Highway, self).__init__() + self.input_dim = input_dim + self.layers = nn.ModuleList([nn.Linear(input_dim, input_dim * 2) + for _ in range(num_layers)]) + self.activation = nn.ReLU() + + self.reset_parameters() + + def reset_parameters(self): + for layer in self.layers: + # As per comment in AllenNLP: + # We should bias the highway layer to just carry its input forward. We do that by + # setting the bias on `B(x)` to be positive, because that means `g` will be biased to + # be high, so we will carry the input forward. The bias on `B(x)` is the second half + # of the bias vector in each Linear layer. + nn.init.constant_(layer.bias[self.input_dim:], 1) + + nn.init.constant_(layer.bias[:self.input_dim], 0) + nn.init.xavier_normal_(layer.weight) + + def forward( + self, + x: torch.Tensor + ): + for layer in self.layers: + projection = layer(x) + proj_x, gate = projection.chunk(2, dim=-1) + proj_x = self.activation(proj_x) + gate = torch.sigmoid(gate) + x = gate * x + (gate.new_tensor([1]) - gate) * proj_x + return x diff --git a/fairseq/modules/conv_tbc.py b/fairseq/modules/conv_tbc.py new file mode 100644 index 0000000000000000000000000000000000000000..1aa3eff9dca20a07ba84438312823d16e2493c12 --- /dev/null +++ b/fairseq/modules/conv_tbc.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch.nn.modules.utils import _single + + +class ConvTBC(torch.nn.Module): + """1D convolution over an input of shape (time x batch x channel) + + The implementation uses gemm to perform the convolution. This implementation + is faster than cuDNN for small kernel sizes. + """ + def __init__(self, in_channels, out_channels, kernel_size, padding=0): + super(ConvTBC, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _single(kernel_size) + self.padding = _single(padding) + + self.weight = torch.nn.Parameter(torch.Tensor( + self.kernel_size[0], in_channels, out_channels)) + self.bias = torch.nn.Parameter(torch.Tensor(out_channels)) + + def forward(self, input): + return torch.conv_tbc(input.contiguous(), self.weight, self.bias, self.padding[0]) + + def __repr__(self): + s = ('{name}({in_channels}, {out_channels}, kernel_size={kernel_size}' + ', padding={padding}') + if self.bias is None: + s += ', bias=False' + s += ')' + return s.format(name=self.__class__.__name__, **self.__dict__) diff --git a/fairseq/modules/cross_entropy.py b/fairseq/modules/cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..b46143f3af3fca88443b202b3416333eb1f8d6a4 --- /dev/null +++ b/fairseq/modules/cross_entropy.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch +import torch.nn.functional as F + + +logger = logging.getLogger(__name__) + + +def _cross_entropy_pytorch(logits, target, ignore_index=None, reduction='mean'): + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + return F.nll_loss( + lprobs, target, ignore_index=ignore_index, reduction=reduction, + ) + + +try: + import xentropy_cuda + from apex.contrib import xentropy + + logger.info('using fused cross entropy') + + def cross_entropy(logits, target, ignore_index=-100, reduction='mean'): + if logits.device == torch.device('cpu'): + return _cross_entropy_pytorch(logits, target, ignore_index, reduction) + else: + half_to_float = (logits.dtype == torch.half) + losses = xentropy.SoftmaxCrossEntropyLoss.apply( + logits, target, 0.0, ignore_index, half_to_float, + ) + if reduction == 'sum': + return losses.sum() + elif reduction == 'mean': + if ignore_index >= 0: + return losses.sum() / target.ne(ignore_index).sum() + else: + return losses.mean() + elif reduction == 'none': + return losses + else: + raise NotImplementedError + +except ImportError: + + def cross_entropy(logits, target, ignore_index=-100, reduction='mean'): + return _cross_entropy_pytorch(logits, target, ignore_index, reduction) diff --git a/fairseq/modules/cuda_utils.cu b/fairseq/modules/cuda_utils.cu new file mode 100644 index 0000000000000000000000000000000000000000..516f1d92440e9e2c092f122e45d81b45cb135602 --- /dev/null +++ b/fairseq/modules/cuda_utils.cu @@ -0,0 +1,203 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + + +template +constexpr __host__ __device__ auto divUp(U a, V b) -> decltype(a + b) { + return (a + b - 1) / b; +} + + +template +__inline__ __device__ +void zeroSharedMem(scalar_t* data) { + /* + Given an array of length FS + SB, zero out the first padding_l and last + (FS - padding_l) values in the array + */ + + int tid = threadIdx.x; + + if (FS < SB) { + + // zero all if we have enough threads in a block to do all of them + if (tid < padding_l || tid > SB - FS + padding_l - 1) { + data[tid] = scalar_t(0.0); + } + } else { + + // otherwise zero out one block at a time + const int numIterations = divUp(FS, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if (tid + offset < padding_l) { + data[tid + offset] = scalar_t(0.0); + } else if (tid + offset < FS) { + data[SB + tid + offset] = scalar_t(0.0); + } + } + } +} + +template +__inline__ __device__ +scalar_t warpReduce(scalar_t data) { + /* + Reduce an array within each warp. After processing all values in warp will + caontain the sum of all original values in that warp. + + data - pointer to data to reduce + */ + data += __shfl_xor_sync(SHFL_MASK, data, 16); + data += __shfl_xor_sync(SHFL_MASK, data, 8); + data += __shfl_xor_sync(SHFL_MASK, data, 4); + data += __shfl_xor_sync(SHFL_MASK, data, 2); + data += __shfl_xor_sync(SHFL_MASK, data, 1); + return data; +} + +template +__inline__ __device__ +scalar_t blockReduce(scalar_t data) { + /* + Reduce an entire array on the block level. After processing, the + first value in the array will contain the reduced sum. + + data - pointer to data to reduce + */ + + static __shared__ scalar_t warpSum[32]; + const int tid = threadIdx.x; + int wid = tid / 32; + int lane = tid % 32; + + __syncthreads(); + + // reduce each warp then write to shared memory + scalar_t sum = warpReduce(data); + if (lane == 0) { + warpSum[wid] = sum; + } + + __syncthreads(); + + scalar_t v; + // perform final sum of partial warp sums + if (tid < blockDim.x / 32) { + v = warpSum[lane]; + } else { + v = scalar_t(0.0); + } + + if (wid == 0) { + v = warpReduce(v); + } + __syncthreads(); + + return v; +} + +void checkCudaStatus(cudaError_t status, int lineNumber = -1) { + + if (status != cudaSuccess) { + std::cout << cudaGetErrorString(status) + << " at line " << lineNumber << std::endl; + std::cout << "Exiting" << std::endl; + exit(1); + } +} + +template +__device__ +void load_input_to_shared(const scalar_t* input, // global memory + int inputOffset, int sequenceLength, + int iteration, int numIterations, + bool no_prev, scalar_t* output /* shared memory */) { + /* + Load a block size of input into shared memory with + right and left overhang of total size FS. If previously + loaded memory, overlap will be shifted over to reduce + global memory access + + input - pointer to start of channel sequence + inputOffset - how far in the sequence to start loading + sequenceLength - total length of sequence + iteration - which block of sequence we are loading + numIterations - total number of blocks to load + no_prev - whether to load the whole block if the previous block + wasn't loaded + output - shared memory to write input to + */ + + const int tid = threadIdx.x; + + // Load the left "overhang" of input + if (iteration > 0) { + if (padding_l < SB) { + + // load all at once + if (tid < padding_l) { + output[tid] = (no_prev) ? input[inputOffset - padding_l + tid] : output[tid + SB]; + } + } else { + + // load in chunks of size SB + int numIterations = divUp(padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < padding_l) { + output[tid + offset] = (no_prev) ? input[inputOffset - padding_l + tid + offset] : output[tid + offset + SB]; + } + } + } + } + + // Load the right "overhang" of input + if (iteration < (numIterations - 1)) { + const int elementsLeft = sequenceLength - (iteration+1) * SB; + + if ((FS - padding_l) < SB) { + + // load all at once + if (tid < (FS - padding_l)) { + output[padding_l + SB + tid] = (tid < elementsLeft) ? input[inputOffset + SB + tid] : scalar_t(0.0); + } + } else { + + // load in chunks of size SB + int numIterations = divUp(FS - padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < (FS - padding_l)) { + output[padding_l + SB + tid + offset] = ((tid + offset) < elementsLeft) ? input[inputOffset + SB + tid + offset] : scalar_t(0.0); + } + } + } + } + + // We should also clear out the right "overhang" + if (iteration == (numIterations - 1)) { + if ((FS - padding_l) < SB) { + + // clear out all at once + if (tid < (FS - padding_l)) { + output[padding_l + SB + tid] = scalar_t(0.0); + } + } else { + + // clear in chunks of size SB + int numIterations = divUp(FS - padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < (FS - padding_l)) { + output[padding_l + SB + tid + offset] = scalar_t(0.0); + } + } + } + } + output[tid + padding_l] = ((inputOffset + tid) < sequenceLength) ? input[inputOffset + tid] : scalar_t(0.0); +} diff --git a/fairseq/modules/downsampled_multihead_attention.py b/fairseq/modules/downsampled_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..eeaf9bbdd3c3315003c55ae8a743bac0db452a29 --- /dev/null +++ b/fairseq/modules/downsampled_multihead_attention.py @@ -0,0 +1,256 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +# + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.modules.scalar_bias import scalar_bias +from fairseq.modules.fairseq_dropout import FairseqDropout + + +class SingleHeadAttention(nn.Module): + """ + Single-head attention that supports Gating and Downsampling + """ + def __init__( + self, out_channels, embed_dim, head_dim, head_index, dropout=0., + bias=True, project_input=True, gated=False, downsample=False, + num_heads=1, + ): + super().__init__() + self.embed_dim = embed_dim + self.dropout_module = FairseqDropout(dropout, module_name=self.__class__.__name__) + self.head_index = head_index + self.head_dim = head_dim + self.project_input = project_input + self.gated = gated + self.downsample = downsample + self.num_heads = num_heads + self.projection = None + + k_layers = [] + v_layers = [] + if self.downsample: + k_layers.append(Downsample(self.head_index)) + v_layers.append(Downsample(self.head_index)) + out_proj_size = self.head_dim + else: + out_proj_size = self.head_dim * self.num_heads + if self.gated: + k_layers.append(GatedLinear(self.embed_dim, out_proj_size, bias=bias)) + self.in_proj_q = GatedLinear(self.embed_dim, out_proj_size, bias=bias) + v_layers.append(GatedLinear(self.embed_dim, out_proj_size, bias=bias)) + else: + k_layers.append(Linear(self.embed_dim, out_proj_size, bias=bias)) + self.in_proj_q = Linear(self.embed_dim, out_proj_size, bias=bias) + v_layers.append(Linear(self.embed_dim, out_proj_size, bias=bias)) + + self.in_proj_k = nn.Sequential(*k_layers) + self.in_proj_v = nn.Sequential(*v_layers) + + if self.downsample: + self.out_proj = Linear(out_proj_size, self.head_dim, bias=bias) + else: + self.out_proj = Linear(out_proj_size, out_channels, bias=bias) + + self.scaling = self.head_dim**-0.5 + + def forward( + self, query, key, value, mask_future_timesteps=False, + key_padding_mask=None, use_scalar_bias=False, + ): + """Input shape: Time x Batch x Channel + Self-attention can be implemented by passing in the same arguments for + query, key and value. Future timesteps can be masked with the + `mask_future_timesteps` argument. Padding elements can be excluded from + the key by passing a binary ByteTensor (`key_padding_mask`) with shape: + batch x src_len, where padding elements are indicated by 1s. + """ + src_len, bsz, out_channels = key.size() + tgt_len = query.size(0) + assert list(query.size()) == [tgt_len, bsz, out_channels] + assert key.size() == value.size() + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == src_len + + if self.downsample: + size = bsz + else: + size = bsz * self.num_heads + + k = key + v = value + q = query + if self.project_input: + q = self.in_proj_q(q) + k = self.in_proj_k(k) + v = self.in_proj_v(v) + src_len = k.size()[0] + q *= self.scaling + + if not self.downsample: + q = q.view(tgt_len, size, self.head_dim) + k = k.view(src_len, size, self.head_dim) + v = v.view(src_len, size, self.head_dim) + + q = q.transpose(0, 1) + k = k.transpose(0, 1) + v = v.transpose(0, 1) + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + if mask_future_timesteps: + assert query.size() == key.size(), \ + 'mask_future_timesteps only applies to self-attention' + attn_weights *= torch.tril( + attn_weights.data.new([1]).expand(tgt_len, tgt_len).clone(), + diagonal=-1, + )[:, ::self.head_index + 1 if self.downsample else 1].unsqueeze(0) + attn_weights += torch.triu( + attn_weights.data.new([-math.inf]).expand(tgt_len, tgt_len).clone(), + diagonal=0 + )[:, ::self.head_index + 1 if self.downsample else 1].unsqueeze(0) + tgt_size = tgt_len + if use_scalar_bias: + attn_weights = scalar_bias(attn_weights, 2) + v = scalar_bias(v, 1) + tgt_size += 1 + + if key_padding_mask is not None: + # don't attend to padding symbols + if key_padding_mask.max() > 0: + if self.downsample: + attn_weights = attn_weights.view(bsz, 1, tgt_len, src_len) + else: + attn_weights = attn_weights.view(size, self.num_heads, tgt_len, src_len) + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + -math.inf, + ) + attn_weights = attn_weights.view(size, tgt_len, src_len) + attn_weights = F.softmax(attn_weights, dim=-1) + attn_weights = self.dropout_module(attn_weights) + + attn = torch.bmm(attn_weights, v) + if self.downsample: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.head_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.embed_dim) + + attn = self.out_proj(attn) + + return attn, attn_weights + + +class DownsampledMultiHeadAttention(nn.ModuleList): + """ + Multi-headed attention with Gating and Downsampling + """ + def __init__( + self, out_channels, embed_dim, num_heads, dropout=0., bias=True, + project_input=True, gated=False, downsample=False, + ): + self.embed_dim = embed_dim + self.num_heads = num_heads + self.head_dim = embed_dim // num_heads + self.downsample = downsample + self.gated = gated + self.project_input = project_input + assert self.head_dim * num_heads == embed_dim + + if self.downsample: + attention_heads = [] + for index in range(self.num_heads): + attention_heads.append( + SingleHeadAttention( + out_channels, self.embed_dim, self.head_dim, index, + dropout, bias, self.project_input, self.gated, + self.downsample, self.num_heads, + ) + ) + super().__init__(modules=attention_heads) + self.out_proj = Linear(embed_dim, out_channels, bias=bias) + else: + # either we have a list of attention heads, or just one attention head + # if not being downsampled, we can do the heads with one linear layer instead of separate ones + super().__init__() + self.attention_module = SingleHeadAttention( + out_channels, self.embed_dim, self.head_dim, 1, dropout, + bias, self.project_input, self.gated, self.downsample, self.num_heads, + ) + + def forward( + self, query, key, value, mask_future_timesteps=False, + key_padding_mask=None, use_scalar_bias=False, + ): + src_len, bsz, embed_dim = key.size() + tgt_len = query.size(0) + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + assert key.size() == value.size() + + tgt_size = tgt_len + if use_scalar_bias: + tgt_size += 1 + + attn = [] + attn_weights = [] + if self.downsample: + for attention_head_number in range(self.num_heads): + # call the forward of each attention head + _attn, _attn_weight = self[attention_head_number]( + query, key, value, mask_future_timesteps, key_padding_mask, use_scalar_bias, + ) + attn.append(_attn) + attn_weights.append(_attn_weight) + full_attn = torch.cat(attn, dim=2) + full_attn = self.out_proj(full_attn) + return full_attn, attn_weights[0].clone() + else: + _attn, _attn_weight = self.attention_module( + query, key, value, mask_future_timesteps, key_padding_mask, use_scalar_bias, + ) + attn.append(_attn) + attn_weights.append(_attn_weight) + full_attn = torch.cat(attn, dim=2) + full_attn_weights = torch.cat(attn_weights) + full_attn_weights = full_attn_weights.view(bsz, self.num_heads, tgt_size, src_len) + full_attn_weights = full_attn_weights.sum(dim=1) / self.num_heads + return full_attn, full_attn_weights + + +class Downsample(nn.Module): + """ + Selects every nth element, where n is the index + """ + def __init__(self, index): + super().__init__() + self.index = index + + def forward(self, x): + return x[::self.index+1] + + +def Linear(in_features, out_features, dropout=0., bias=True): + """Weight-normalized Linear layer (input: B x T x C)""" + m = nn.Linear(in_features, out_features, bias=bias) + m.weight.data.normal_(mean=0, std=math.sqrt((1 - dropout) / in_features)) + m.bias.data.zero_() + return nn.utils.weight_norm(m) + + +def GatedLinear(in_features, out_features, dropout=0., bias=True): + """Weight-normalized Linear layer (input: B x T x C) with interspersed GLU units""" + return nn.Sequential( + Linear(in_features, out_features*4, dropout, bias), + nn.GLU(), + Linear(out_features*2, out_features*2, dropout, bias), + nn.GLU(), + Linear(out_features, out_features, dropout, bias) + ) diff --git a/fairseq/modules/dynamic_convolution.py b/fairseq/modules/dynamic_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..5a8ecb99a8b6a9ccc114f77896e31ae4386dc7a7 --- /dev/null +++ b/fairseq/modules/dynamic_convolution.py @@ -0,0 +1,245 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from .unfold import unfold1d +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout + + +def DynamicConv(input_size, kernel_size=1, padding_l=None, num_heads=1, + weight_dropout=0., weight_softmax=False, + renorm_padding=False, bias=False, conv_bias=False, + query_size=None, in_proj=False): + if torch.cuda.is_available(): + try: + from fairseq.modules.dynamicconv_layer import DynamicconvLayer + return DynamicconvLayer(input_size, kernel_size=kernel_size, + padding_l=padding_l, num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, bias=bias) + except ImportError as e: + print(e) + return DynamicConv1dTBC(input_size, kernel_size=kernel_size, + padding_l=padding_l, num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, bias=bias) + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.) + return m + + +@with_incremental_state +class DynamicConv1dTBC(nn.Module): + '''Dynamic lightweight convolution taking T x B x C inputs + Args: + input_size: # of channels of the input + kernel_size: convolution channels + padding_l: padding to the left when using "same" padding + num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size) + weight_dropout: the drop rate of the DropConnect to drop the weight + weight_softmax: normalize the weight with softmax before the convolution + renorm_padding: re-normalize the filters to ignore the padded part (only the non-padding parts sum up to 1) + bias: use bias + conv_bias: bias of the convolution + query_size: specified when feeding a different input as the query + in_proj: project the input and generate the filter together + + Shape: + Input: TxBxC, i.e. (timesteps, batch_size, input_size) + Output: TxBxC, i.e. (timesteps, batch_size, input_size) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + ''' + def __init__(self, input_size, kernel_size=1, padding_l=None, num_heads=1, + weight_dropout=0., weight_softmax=False, + renorm_padding=False, bias=False, conv_bias=False, + query_size=None, in_proj=False): + super().__init__() + self.input_size = input_size + self.query_size = input_size if query_size is None else query_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_dropout_module = FairseqDropout(weight_dropout, module_name=self.__class__.__name__) + self.weight_softmax = weight_softmax + self.renorm_padding = renorm_padding + + if in_proj: + self.weight_linear = Linear(self.input_size, self.input_size + num_heads * kernel_size * 1) + else: + self.weight_linear = Linear(self.query_size, num_heads * kernel_size * 1, bias=bias) + if conv_bias: + self.conv_bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.conv_bias = None + self.reset_parameters() + + @property + def in_proj(self): + return self.weight_linear.out_features == self.input_size + self.num_heads * self.kernel_size + + def reset_parameters(self): + self.weight_linear.reset_parameters() + if self.conv_bias is not None: + nn.init.constant_(self.conv_bias, 0.) + + def forward(self, x, incremental_state=None, query=None, unfold=None): + '''Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C + args: + x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size) + incremental_state: A dict to keep the state + unfold: unfold the input or not. If not, we use the matrix trick instead + query: use the specified query to predict the conv filters + ''' + unfold = x.size(0) > 512 if unfold is None else unfold # use unfold mode as default for long sequence to save memory + unfold = unfold or (incremental_state is not None) + assert query is None or not self.in_proj + + if query is None: + query = x + if unfold: + output = self._forward_unfolded(x, incremental_state, query) + else: + output = self._forward_expanded(x, incremental_state, query) + + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + return output + + def _forward_unfolded(self, x, incremental_state, query): + '''The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.''' + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + if self.in_proj: + proj = self.weight_linear(x) + x = proj.narrow(2, 0, self.input_size).contiguous() + weight = proj.narrow(2, self.input_size, H*K).contiguous().view(T*B*H, -1) + else: + weight = self.weight_linear(query).view(T*B*H, -1) + + # renorm_padding is only implemented in _forward_expanded + assert not self.renorm_padding or incremental_state is not None + + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer(incremental_state, x_unfold[:, :, :, -self.kernel_size+1:]) + x_unfold = x_unfold.view(T*B*H, R, -1) + else: + padding_l = self.padding_l + if K > T and padding_l == K-1: + weight = weight.narrow(1, K-T, T) + K, padding_l = T, T-1 + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, K, padding_l, 0) + x_unfold = x_unfold.view(T*B*H, R, K) + + if self.weight_softmax and not self.renorm_padding: + weight = F.softmax(weight, dim=1) + weight = weight.narrow(1, 0, K) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2):] + K = weight.size(1) + + if self.weight_softmax and self.renorm_padding: + weight = F.softmax(weight, dim=1) + + weight = self.weight_dropout_module(weight, inplace=False) + + output = torch.bmm(x_unfold, weight.unsqueeze(2)) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_stat, query): + '''Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + ''' + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + if self.in_proj: + proj = self.weight_linear(x) + x = proj.narrow(2, 0, self.input_size).contiguous() + weight = proj.narrow(2, self.input_size, H*K).contiguous().view(T*B*H, -1) + else: + weight = self.weight_linear(query).view(T*B*H, -1) + + if not self.renorm_padding: + if self.weight_softmax: + weight = F.softmax(weight, dim=1) + weight = self.weight_dropout_module(weight, inplace=False) + weight = weight.narrow(1, 0, K).contiguous() + weight = weight.view(T, B*H, K).transpose(0, 1) + + x = x.view(T, B*H, R).transpose(0, 1) + if self.weight_softmax and self.renorm_padding: + # turn the convolution filters into band matrices + weight_expanded = weight.new(B*H, T, T+K-1).fill_(float('-inf')) + weight_expanded.as_strided((B*H, T, K), (T*(T+K-1), T+K, 1)).copy_(weight) + weight_expanded = weight_expanded.narrow(2, self.padding_l, T) + # normalize the weight over valid positions like self-attention + weight_expanded = F.softmax(weight_expanded, dim=2) + weight_expanded = self.weight_dropout_module(weight_expanded, inplace=False) + else: + P = self.padding_l + # For efficieny, we cut the kernel size and reduce the padding when the kernel is larger than the length + if K > T and P == K-1: + weight = weight.narrow(2, K-T, T) + K, P = T, T-1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B*H, T, T+K-1, requires_grad=False) + weight_expanded.as_strided((B*H, T, K), (T*(T+K-1), T+K, 1)).copy_(weight) + weight_expanded = weight_expanded.narrow(2, P, T) # B*H x T x T + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, 'input_buffer') + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state(self, incremental_state, 'input_buffer', new_buffer) + + def extra_repr(self): + s = '{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, conv_bias={}, renorm_padding={}, in_proj={}'.format( + self.input_size, self.kernel_size, self.padding_l, + self.num_heads, self.weight_softmax, self.conv_bias is not None, self.renorm_padding, + self.in_proj, + ) + + if self.query_size != self.input_size: + s += ', query_size={}'.format(self.query_size) + if self.weight_dropout_module.p > 0.: + s += ', weight_dropout={}'.format(self.weight_dropout_module.p) + return s diff --git a/fairseq/modules/dynamic_crf_layer.py b/fairseq/modules/dynamic_crf_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..6f5acf377261b5be80dc09b9b88e507e9f1c9ff7 --- /dev/null +++ b/fairseq/modules/dynamic_crf_layer.py @@ -0,0 +1,184 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +This file is to re-implemented the low-rank and beam approximation of CRF layer +Proposed by: + +Sun, Zhiqing, et al. +Fast Structured Decoding for Sequence Models +https://arxiv.org/abs/1910.11555 + +The CRF implementation is mainly borrowed from +https://github.com/kmkurn/pytorch-crf/blob/master/torchcrf/__init__.py + +""" + +import numpy as np +import torch +import torch.nn as nn + + +def logsumexp(x, dim=1): + return torch.logsumexp(x.float(), dim=dim).type_as(x) + + +class DynamicCRF(nn.Module): + """Dynamic CRF layer is used to approximate the traditional + Conditional Random Fields (CRF) + $P(y | x) = 1/Z(x) exp(sum_i s(y_i, x) + sum_i t(y_{i-1}, y_i, x))$ + + where in this function, we assume the emition scores (s) are given, + and the transition score is a |V| x |V| matrix $M$ + + in the following two aspects: + (1) it used a low-rank approximation for the transition matrix: + $M = E_1 E_2^T$ + (2) it used a beam to estimate the normalizing factor Z(x) + """ + + def __init__(self, num_embedding, low_rank=32, beam_size=64): + super().__init__() + + self.E1 = nn.Embedding(num_embedding, low_rank) + self.E2 = nn.Embedding(num_embedding, low_rank) + + self.vocb = num_embedding + self.rank = low_rank + self.beam = beam_size + + def extra_repr(self): + return "vocab_size={}, low_rank={}, beam_size={}".format( + self.vocb, self.rank, self.beam) + + def forward(self, emissions, targets, masks, beam=None): + """ + Compute the conditional log-likelihood of a sequence of target tokens given emission scores + + Args: + emissions (`~torch.Tensor`): Emission score are usually the unnormalized decoder output + ``(batch_size, seq_len, vocab_size)``. We assume batch-first + targets (`~torch.LongTensor`): Sequence of target token indices + ``(batch_size, seq_len) + masks (`~torch.ByteTensor`): Mask tensor with the same size as targets + + Returns: + `~torch.Tensor`: approximated log-likelihood + """ + numerator = self._compute_score(emissions, targets, masks) + denominator = self._compute_normalizer(emissions, targets, masks, beam) + return numerator - denominator + + def forward_decoder(self, emissions, masks=None, beam=None): + """ + Find the most likely output sequence using Viterbi algorithm. + + Args: + emissions (`~torch.Tensor`): Emission score are usually the unnormalized decoder output + ``(batch_size, seq_len, vocab_size)``. We assume batch-first + masks (`~torch.ByteTensor`): Mask tensor with the same size as targets + + Returns: + `~torch.LongTensor`: decoded sequence from the CRF model + """ + return self._viterbi_decode(emissions, masks, beam) + + def _compute_score(self, emissions, targets, masks=None): + batch_size, seq_len = targets.size() + emission_scores = emissions.gather(2, targets[:, :, None])[:, :, 0] # B x T + transition_scores = (self.E1(targets[:, :-1]) * self.E2(targets[:, 1:])).sum(2) + + scores = emission_scores + scores[:, 1:] += transition_scores + + if masks is not None: + scores = scores * masks.type_as(scores) + return scores.sum(-1) + + def _compute_normalizer(self, emissions, targets=None, masks=None, beam=None): + # HACK: we include "target" which is a hueristic for training + # HACK: we use a beam of tokens to approximate the normalizing factor (which is bad?) + + beam = beam if beam is not None else self.beam + batch_size, seq_len = emissions.size()[:2] + if targets is not None: + _emissions = emissions.scatter(2, targets[:, :, None], np.float('inf')) + beam_targets = _emissions.topk(beam, 2)[1] + beam_emission_scores = emissions.gather(2, beam_targets) + else: + beam_emission_scores, beam_targets = emissions.topk(beam, 2) + beam_transition_score1 = self.E1(beam_targets[:, :-1]) # B x (T-1) x K x D + beam_transition_score2 = self.E2(beam_targets[:, 1:]) # B x (T-1) x K x D + beam_transition_matrix = torch.bmm( + beam_transition_score1.view(-1, beam, self.rank), + beam_transition_score2.view(-1, beam, self.rank).transpose(1, 2)) + beam_transition_matrix = beam_transition_matrix.view(batch_size, -1, beam, beam) + + # compute the normalizer in the log-space + score = beam_emission_scores[:, 0] # B x K + for i in range(1, seq_len): + next_score = score[:, :, None] + beam_transition_matrix[:, i-1] + next_score = logsumexp(next_score, dim=1) + beam_emission_scores[:, i] + + if masks is not None: + score = torch.where(masks[:, i:i+1], next_score, score) + else: + score = next_score + + # Sum (log-sum-exp) over all possible tags + return logsumexp(score, dim=1) + + def _viterbi_decode(self, emissions, masks=None, beam=None): + # HACK: we use a beam of tokens to approximate the normalizing factor (which is bad?) + + beam = beam if beam is not None else self.beam + batch_size, seq_len = emissions.size()[:2] + beam_emission_scores, beam_targets = emissions.topk(beam, 2) + beam_transition_score1 = self.E1(beam_targets[:, :-1]) # B x (T-1) x K x D + beam_transition_score2 = self.E2(beam_targets[:, 1:]) # B x (T-1) x K x D + beam_transition_matrix = torch.bmm( + beam_transition_score1.view(-1, beam, self.rank), + beam_transition_score2.view(-1, beam, self.rank).transpose(1, 2)) + beam_transition_matrix = beam_transition_matrix.view(batch_size, -1, beam, beam) + + traj_tokens, traj_scores = [], [] + finalized_tokens, finalized_scores = [], [] + + # compute the normalizer in the log-space + score = beam_emission_scores[:, 0] # B x K + dummy = torch.arange(beam, device=score.device).expand(*score.size()).contiguous() + + for i in range(1, seq_len): + traj_scores.append(score) + _score = score[:, :, None] + beam_transition_matrix[:, i-1] + _score, _index = _score.max(dim=1) + _score = _score + beam_emission_scores[:, i] + + if masks is not None: + score = torch.where(masks[:, i: i+1], _score, score) + index = torch.where(masks[:, i: i+1], _index, dummy) + else: + score, index = _score, _index + traj_tokens.append(index) + + # now running the back-tracing and find the best + best_score, best_index = score.max(dim=1) + finalized_tokens.append(best_index[:, None]) + finalized_scores.append(best_score[:, None]) + + for idx, scs in zip(reversed(traj_tokens), reversed(traj_scores)): + previous_index = finalized_tokens[-1] + finalized_tokens.append(idx.gather(1, previous_index)) + finalized_scores.append(scs.gather(1, previous_index)) + + finalized_tokens.reverse() + finalized_tokens = torch.cat(finalized_tokens, 1) + finalized_tokens = beam_targets.gather(2, finalized_tokens[:, :, None])[:, :, 0] + + finalized_scores.reverse() + finalized_scores = torch.cat(finalized_scores, 1) + finalized_scores[:, 1:] = finalized_scores[:, 1:] - finalized_scores[:, :-1] + + return finalized_scores, finalized_tokens diff --git a/fairseq/modules/dynamicconv_layer/__init__.py b/fairseq/modules/dynamicconv_layer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..22dc6f403d2a0ecdb1b9e7e69ed96bd560e93b2c --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .dynamicconv_layer import DynamicconvLayer # noqa diff --git a/fairseq/modules/dynamicconv_layer/cuda_function_gen.py b/fairseq/modules/dynamicconv_layer/cuda_function_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..926d6ca846be37a2b7ac451ca62706763b53013b --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/cuda_function_gen.py @@ -0,0 +1,223 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def gen_forward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + blocks = [32, 64, 128, 256] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" + +std::vector dynamicconv_cuda_forward(at::Tensor input, at::Tensor weight, int padding_l) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = weight.size(1); + const auto filterSize = weight.size(2); + + const auto numFiltersInBlock = numFeatures / numHeads; + const dim3 blocks(minibatch, numFeatures); + + auto output = at::zeros_like(input); + auto stream = at::cuda::getCurrentCUDAStream(); +""" + + switch = """ + switch(filterSize) { +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {pad}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "dynamicconv_forward", ([&] {{ + dynamicconv_forward_kernel<{k}, {b_size}, {pad}, scalar_t> + <<>>( + input.data(), + weight.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + output.data()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping forward pass" << std::endl; + } + break;\n +""" + + end = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping forward pass" << std::endl; + } + + return {output}; +} +""" + + with open("dynamicconv_cuda_forward.cu", 'w') as forward: + forward.write(head) + forward.write(switch) + for k in kernels: + b_size = 32 + for b in blocks: + if b > k: + b_size = b + break + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=b_size, pad=pad)) + forward.write(bad_padding) + forward.write(end) + + +def gen_backward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + thresh = [512, 512, 512, 512, 512, 380, 256, 256] + min_block = [64, 64, 64, 64, 64, 64, 128, 256] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" + +std::vector dynamicconv_cuda_backward(at::Tensor gradOutput, int padding_l, at::Tensor input, at::Tensor weight) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = weight.size(1); + const auto filterSize = weight.size(2); + + const auto numFiltersInBlock = numFeatures / numHeads; + auto numChunks = 1; + + auto gradInput = at::zeros_like(input); + auto gradWeight = at::zeros_like(weight); + auto stream = at::cuda::getCurrentCUDAStream(); + + dim3 blocks(minibatch, numHeads, numChunks); +""" + + sequence_if = """ + if (sequenceLength < {seq}) {{ + switch(filterSize) {{ +""" + + case_k = """ + case {k}: +""" + + chunks_reset = """ + numChunks = int(ceilf(sequenceLength/float({b_size}))); + blocks = dim3(minibatch, numHeads, numChunks); +""" + + main_block = """ + if (padding_l == {p}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(gradOutput.scalar_type(), "dynamicconv_backward", ([&] {{ + dynamicconv_backward_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + gradOutput.data(), + input.data(), + weight.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + gradWeight.data(), + gradInput.data()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping backward pass" << std::endl; + } + break;\n +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping backward pass" << std::endl; + } +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + last_return = """ + } + return {gradInput, gradWeight}; +} +""" + + with open("dynamicconv_cuda_backward.cu", 'w') as backward: + backward.write(head) + for seq in seqs: + backward.write(sequence_if.format(seq=seq)) + for k, t, m in zip(kernels, thresh, min_block): + backward.write(case_k.format(k=k)) + if seq <= t: + b_size = seq + else: + b_size = m + backward.write(chunks_reset.format(b_size=b_size)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=b_size, p=p)) + backward.write(bad_padding) + backward.write(bad_filter) + backward.write(con_else) + backward.write(final_else) + for k, m in zip(kernels, min_block): + backward.write(case_k.format(k=k)) + backward.write(chunks_reset.format(b_size=m)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=m, p=p)) + backward.write(bad_padding) + backward.write(bad_filter) + backward.write(last_return) + + +if __name__ == "__main__": + gen_forward() + gen_backward() diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ebd4df0e9608d769f31eadc6e0b487505f11b279 --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp @@ -0,0 +1,56 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +std::vector dynamicconv_cuda_forward( + at::Tensor input, + at::Tensor filters, + int padding_l); + +std::vector dynamicconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters); + + +#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + +std::vector dynamicconv_forward( + at::Tensor input, + at::Tensor filters, + int padding_l) { + + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return dynamicconv_cuda_forward(input, filters, + padding_l); +} + +std::vector dynamicconv_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + + CHECK_INPUT(gradOutput); + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return dynamicconv_cuda_backward(gradOutput, padding_l, + input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &dynamicconv_forward, "dynamicconv forward (CUDA)"); + m.def("backward", &dynamicconv_backward, "dynamicconv backward (CUDA)"); +} diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh new file mode 100644 index 0000000000000000000000000000000000000000..2196259433aefc88f96cd5bbcae57740a9a8c2dc --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh @@ -0,0 +1,51 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +#include +#include +#include + +#include +#include +#include +#include +#include +#include + +#include +#include +#include + +#define SHFL_MASK 0xffffffff + +template +__global__ +void dynamicconv_forward_kernel(const scalar_t* input, + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* output); + +template +__global__ +void dynamicconv_backward_kernel( + const scalar_t* gradOutput, // B * C * T + const scalar_t* input, // B * C * T + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* gradWeight, + scalar_t* gradInput); // B * H * k * T diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..300d35b6478080a9594a22e335988c321d43127f --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu @@ -0,0 +1,168 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" +#include "dynamicconv_cuda_forward.cu" +#include "dynamicconv_cuda_backward.cu" +#include "../cuda_utils.cu" + +// FS is filter size and kernels are specialized for filter sizes +template +__global__ +void dynamicconv_forward_kernel(const scalar_t* input, + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* output) { + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int head = featureIdx / numFiltersInBlock; + + const int IOOffset = batchIdx * numFeatures * sequenceLength + + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + + scalar_t filter[FS]; + + __shared__ scalar_t tempInput[SB + FS]; + zeroSharedMem(tempInput); + + const int numIterations = divUp(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + __syncthreads(); + const int inputOffset = i * SB; + load_input_to_shared(inputFeature, inputOffset, + sequenceLength, i, + numIterations, false, tempInput); + __syncthreads(); + if (inputOffset + tid < sequenceLength) { + + #pragma unroll + for (int k = 0; k < FS; ++k) { + const int filterOffset = batchIdx * numHeads * FS * sequenceLength + + head * FS * sequenceLength + + k * sequenceLength + + i * SB + tid; + filter[k] = weight[filterOffset]; + } + + scalar_t out = scalar_t(0.0); + #pragma unroll + for (int k = 0; k < FS; ++k) { + out += filter[k] * tempInput[tid + k]; + } + + outputFeature[inputOffset + tid] = out; + + } + } +} + +template +__global__ +void dynamicconv_backward_kernel( + const scalar_t* gradOutput, // B * C * T + const scalar_t* input, // B * C * T + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* gradWeight, + scalar_t* gradInput) { // B * H * k * T + + assert(blockDim.x == SB); + + // each block operates on a single batch and filter head + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int headIdx = blockIdx.y; + const int chunkIdx = blockIdx.z; + + const int numChunks = divUp(sequenceLength, SB); + const int inputOffset = chunkIdx * SB; + + // initialize shared memory for output gradient and input + __shared__ scalar_t tempGradOutput[SB + FS]; + __shared__ scalar_t tempInput[SB + FS]; + const int padding = FS - padding_l - 1; + + zeroSharedMem(tempGradOutput); + zeroSharedMem(tempInput); + + // initialize local filter and weight gradient sum arrays + scalar_t tempGradSum[FS]; + scalar_t bfilter[FS]; + for (int k = 0; k < FS; ++k) { + tempGradSum[k] = scalar_t(0.0); + + int idxOffset = inputOffset + tid + k - padding; + if (idxOffset >= 0 && idxOffset < sequenceLength) { + int bfilterOffset = batchIdx * numHeads * FS * sequenceLength + + headIdx * FS * sequenceLength + + (FS - k - 1) * sequenceLength + + idxOffset; + bfilter[k] = weight[bfilterOffset]; + } else { + bfilter[k] = scalar_t(0.0); + } + } + + + // iterate over filter block + for (int featureIdx = 0; featureIdx < numFiltersInBlock; ++featureIdx) { + __syncthreads(); + + // load input and output gradient for this channel and chunk + const int IOOffset = batchIdx * numFeatures * sequenceLength + + (headIdx * numFiltersInBlock + featureIdx) * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + const scalar_t* gradOutputFeature = &gradOutput[IOOffset]; + scalar_t* gradInputFeature = &gradInput[IOOffset]; + + load_input_to_shared(gradOutputFeature, inputOffset, + sequenceLength, chunkIdx, + numChunks, true, tempGradOutput); + load_input_to_shared(inputFeature, inputOffset, + sequenceLength, chunkIdx, + numChunks, true, tempInput); + __syncthreads(); + + // sum input and weight gradients + scalar_t out = scalar_t(0.0); + #pragma unroll + for (int k = 0; k < FS; ++k) { + tempGradSum[k] += tempInput[tid + k] * tempGradOutput[tid + padding]; + out += bfilter[k] * tempGradOutput[tid + k]; + } + + if (inputOffset + tid < sequenceLength) { + gradInputFeature[inputOffset + tid] = out; + } + } + + const int gradOffset = batchIdx * numHeads * FS * sequenceLength + + headIdx * FS * sequenceLength; + scalar_t *gradWeightFeature = &gradWeight[gradOffset]; + + // write weight gradient + if (inputOffset + tid < sequenceLength) { + for (int k = 0; k < FS; ++k) { + const int outputOffset = k * sequenceLength + inputOffset + tid; + gradWeightFeature[outputOffset] = tempGradSum[k]; + } + } +} diff --git a/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py b/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..52cc1e8118885920912ae8d45cad0855dcf73090 --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py @@ -0,0 +1,216 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch import nn +from torch.autograd import Function +import torch.nn.functional as F + +import dynamicconv_cuda +from fairseq import utils +from fairseq.modules.unfold import unfold1d +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout + + +class dynamicconvFunction(Function): + + @staticmethod + def forward(ctx, x, weights, padding_l): + ctx.padding_l = padding_l + outputs = dynamicconv_cuda.forward(x, weights, padding_l) + variables = [x, weights] + ctx.save_for_backward(*variables) + return outputs[0] + + @staticmethod + def backward(ctx, grad_output): + outputs = dynamicconv_cuda.backward( + grad_output.contiguous(), + ctx.padding_l, + *ctx.saved_tensors) + grad_input, grad_weights = outputs + return grad_input, grad_weights, None + + +@with_incremental_state +class DynamicconvLayer(nn.Module): + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + weight_softmax=False, + num_heads=1, + weight_dropout=0., + bias=False, + renorm_padding=False, + conv_bias=False, + query_size=None, + ): + + super(DynamicconvLayer, self).__init__() + self.input_size = input_size + self.query_size = input_size if query_size is None else query_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_softmax = weight_softmax + self.weight_dropout_module = FairseqDropout(weight_dropout, module_name=self.__class__.__name__) + self.renorm_padding = renorm_padding + self.bias = bias + + self.weight_linear = nn.Linear(input_size, num_heads * kernel_size, bias) + if conv_bias: + self.conv_bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.conv_bias = None + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight_linear.weight) + if self.conv_bias is not None: + nn.init.constant_(self.conv_bias, 0.) + nn.init.constant_(self.weight_linaer.bias, 0.) + + def forward(self, x, incremental_state=None, query=None, unfold=None): + + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + # R = C // H + + # during inference time, incremental BMM is faster + if incremental_state is not None: + unfold = x.size(0) > 512 if unfold is None else unfold # use unfold mode as default for long sequence to save memory + unfold = unfold or (incremental_state is not None) + assert query is None + + if query is None: + query = x + if unfold: + output = self._forward_unfolded(x, incremental_state, query) + else: + output = self._forward_expanded(x, incremental_state, query) + + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + + return output + + # during training time, use CUDA kernel + else: + weight = self.weight_linear(x).view(T, B, H, K) + if self.weight_softmax: + weight = F.softmax(weight, dim=-1) + if self.weight_dropout_module.p: + weight = self.weight_dropout_module(weight) + + weight = weight.permute(1, 2, 3, 0).contiguous() + self.filters = weight + x = x.permute(1, 2, 0).contiguous() + output = dynamicconvFunction.apply(x, weight, self.padding_l).permute(2, 0, 1) + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, 'input_buffer') + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state(self, incremental_state, 'input_buffer', new_buffer) + + def _forward_unfolded(self, x, incremental_state, query): + '''The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.''' + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight_linear(query).view(T*B*H, -1) + + # renorm_padding is only implemented in _forward_expanded + assert not self.renorm_padding or incremental_state is not None + + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer(incremental_state, x_unfold[:, :, :, -self.kernel_size+1:]) + x_unfold = x_unfold.view(T*B*H, R, -1) + else: + padding_l = self.padding_l + if K > T and padding_l == K-1: + weight = weight.narrow(1, K-T, T) + K, padding_l = T, T-1 + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, K, padding_l, 0) + x_unfold = x_unfold.view(T*B*H, R, K) + + if self.weight_softmax and not self.renorm_padding: + weight = F.softmax(weight, dim=1) + weight = weight.narrow(1, 0, K) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2):] + K = weight.size(1) + + if self.weight_softmax and self.renorm_padding: + weight = F.softmax(weight, dim=1) + + weight = self.weight_dropout_module(weight, inplace=False) + + output = torch.bmm(x_unfold, weight.unsqueeze(2)) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_stat, query): + '''Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + ''' + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + weight = self.weight_linear(query).view(T*B*H, -1) + + if not self.renorm_padding: + if self.weight_softmax: + weight = F.softmax(weight, dim=1) + weight = self.weight_dropout_module(weight, inplace=False) + weight = weight.narrow(1, 0, K).contiguous() + weight = weight.view(T, B*H, K).transpose(0, 1) + + x = x.view(T, B*H, R).transpose(0, 1) + if self.weight_softmax and self.renorm_padding: + # turn the convolution filters into band matrices + weight_expanded = weight.new(B*H, T, T+K-1).fill_(float('-inf')) + weight_expanded.as_strided((B*H, T, K), (T*(T+K-1), T+K, 1)).copy_(weight) + weight_expanded = weight_expanded.narrow(2, self.padding_l, T) + # normalize the weight over valid positions like self-attention + weight_expanded = F.softmax(weight_expanded, dim=2) + weight_expanded = self.weight_dropout_module(weight_expanded, inplace=False) + else: + P = self.padding_l + # For efficieny, we cut the kernel size and reduce the padding when the kernel is larger than the length + if K > T and P == K-1: + weight = weight.narrow(2, K-T, T) + K, P = T, T-1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B*H, T, T+K-1, requires_grad=False) + weight_expanded.as_strided((B*H, T, K), (T*(T+K-1), T+K, 1)).copy_(weight) + weight_expanded = weight_expanded.narrow(2, P, T) # B*H x T x T + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output diff --git a/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp b/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8a6af4285da3c40a01383541acf1f455ffc060fb --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp @@ -0,0 +1,35 @@ +#include +#include + +std::vector dynamicconv_cpu_forward( + float* input, + float* filters, + int padding_l); + +std::vector dynamicconv_cpu_backward( + float* gradOutput, + int padding_l, + float* input, + float* filters); + +std::vector dynamicconv_forward( + float* input, + float* filters, + int padding_l) { + + return dynamicconv_cpu_forward(input, filters, padding_l); +} + +std::vector dynamicconv_backward( + float* gradOutput, + int padding_l, + float* input, + float* filters) { + + return dynamicconv_cpu_backward(gradOutput, padding_l, input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &dynamicconv_forward, "dynamicconv forward (CPU)"); + m.def("backward", &dynamicconv_backward, "dynamicconv backward (CPU)"); +} diff --git a/fairseq/modules/dynamicconv_layer/setup.py b/fairseq/modules/dynamicconv_layer/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..4d789c328371409bd82c9f0087efe6cff459f151 --- /dev/null +++ b/fairseq/modules/dynamicconv_layer/setup.py @@ -0,0 +1,23 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from setuptools import setup +from torch.utils.cpp_extension import CUDAExtension, BuildExtension + +setup( + name='dynamicconv_layer', + ext_modules=[ + CUDAExtension( + name='dynamicconv_cuda', + sources=[ + 'dynamicconv_cuda.cpp', + 'dynamicconv_cuda_kernel.cu', + ], + ), + ], + cmdclass={ + 'build_ext': BuildExtension + }) diff --git a/fairseq/modules/fairseq_dropout.py b/fairseq/modules/fairseq_dropout.py new file mode 100644 index 0000000000000000000000000000000000000000..cbfacf477f4a0879a40dd15a641b2c3e86cc4ef7 --- /dev/null +++ b/fairseq/modules/fairseq_dropout.py @@ -0,0 +1,52 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import List, Optional + +import torch.nn as nn +import torch.nn.functional as F + + +logger = logging.getLogger(__name__) + + +class FairseqDropout(nn.Module): + + def __init__(self, p, module_name=None): + super().__init__() + self.p = p + self.module_name = module_name + self.apply_during_inference = False + + def forward(self, x, inplace: bool = False): + if self.training or self.apply_during_inference: + return F.dropout(x, p=self.p, training=True, inplace=inplace) + else: + return x + + def make_generation_fast_( + self, + name: str, + retain_dropout: bool = False, + retain_dropout_modules: Optional[List[str]] = None, + **kwargs + ): + if retain_dropout: + if retain_dropout_modules is not None and self.module_name is None: + logger.warning( + 'Cannot enable dropout during inference for module {} ' + 'because module_name was not set'.format(name) + ) + elif ( + retain_dropout_modules is None # if None, apply to all modules + or self.module_name in retain_dropout_modules + ): + logger.info( + 'Enabling dropout during inference for module: {}'.format(name) + ) + self.apply_during_inference = True + else: + logger.info('Disabling dropout for module: {}'.format(name)) diff --git a/fairseq/modules/fp32_group_norm.py b/fairseq/modules/fp32_group_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..d03aac022e30c8c14a600062d1d86429504ba003 --- /dev/null +++ b/fairseq/modules/fp32_group_norm.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Layer norm done in fp32 (for fp16 training) +""" + +import torch.nn as nn +import torch.nn.functional as F + + +class Fp32GroupNorm(nn.GroupNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.group_norm( + input.float(), + self.num_groups, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) diff --git a/fairseq/modules/gelu.py b/fairseq/modules/gelu.py new file mode 100644 index 0000000000000000000000000000000000000000..a2f1ecff4a3ae3de3eb7d327b9163c46b18a15ed --- /dev/null +++ b/fairseq/modules/gelu.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +See "Gaussian Error Linear Units (GELUs)" by Dan Hendrycks and Kevin Gimpel with +the corresponding GitHub repo: https://github.com/hendrycks/GELUs +""" + +import math + +import torch +import torch.nn as nn + + +def gelu_accurate(x): + if not hasattr(gelu_accurate, "_a"): + gelu_accurate._a = math.sqrt(2 / math.pi) + return ( + 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))) + ) + + +def gelu(x: torch.Tensor) -> torch.Tensor: + return torch.nn.functional.gelu(x.float()).type_as(x) diff --git a/fairseq/modules/grad_multiply.py b/fairseq/modules/grad_multiply.py new file mode 100644 index 0000000000000000000000000000000000000000..08d15f55dfda9c61a1cf8641ea31424fe1d97f57 --- /dev/null +++ b/fairseq/modules/grad_multiply.py @@ -0,0 +1,18 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +class GradMultiply(torch.autograd.Function): + @staticmethod + def forward(ctx, x, scale): + ctx.scale = scale + res = x.new(x) + return res + + @staticmethod + def backward(ctx, grad): + return grad * ctx.scale, None diff --git a/fairseq/modules/gumbel_vector_quantizer.py b/fairseq/modules/gumbel_vector_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..01ddd2298b7541bf4923d8df6d82d22ac5d1aadb --- /dev/null +++ b/fairseq/modules/gumbel_vector_quantizer.py @@ -0,0 +1,198 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class GumbelVectorQuantizer(nn.Module): + def __init__( + self, + dim, + num_vars, + temp, + groups, + combine_groups, + vq_dim, + time_first, + activation=nn.GELU(), + weight_proj_depth=1, + weight_proj_factor=1, + ): + """Vector quantization using gumbel softmax + + Args: + dim: input dimension (channels) + num_vars: number of quantized vectors per group + temp: temperature for training. this should be a tuple of 3 elements: (start, stop, decay factor) + groups: number of groups for vector quantization + combine_groups: whether to use the vectors for all groups + vq_dim: dimensionality of the resulting quantized vector + time_first: if true, expect input in BxTxC format, otherwise in BxCxT + activation: what activation to use (should be a module). this is only used if weight_proj_depth is > 1 + weight_proj_depth: number of layers (with activation in between) to project input before computing logits + weight_proj_factor: this is used only if weight_proj_depth is > 1. scales the inner dimensionality of + projections by this factor + """ + super().__init__() + + self.groups = groups + self.combine_groups = combine_groups + self.input_dim = dim + self.num_vars = num_vars + self.time_first = time_first + + assert ( + vq_dim % groups == 0 + ), f"dim {vq_dim} must be divisible by groups {groups} for concatenation" + + var_dim = vq_dim // groups + num_groups = groups if not combine_groups else 1 + + self.vars = nn.Parameter(torch.FloatTensor(1, num_groups * num_vars, var_dim)) + nn.init.uniform_(self.vars) + + if weight_proj_depth > 1: + + def block(input_dim, output_dim): + return nn.Sequential(nn.Linear(input_dim, output_dim), activation) + + inner_dim = self.input_dim * weight_proj_factor + self.weight_proj = nn.Sequential( + *[ + block(self.input_dim if i == 0 else inner_dim, inner_dim) + for i in range(weight_proj_depth - 1) + ], + nn.Linear(inner_dim, groups * num_vars), + ) + else: + self.weight_proj = nn.Linear(self.input_dim, groups * num_vars) + nn.init.normal_(self.weight_proj.weight, mean=0, std=1) + nn.init.zeros_(self.weight_proj.bias) + + assert len(temp) == 3, temp + + self.max_temp, self.min_temp, self.temp_decay = temp + self.curr_temp = self.max_temp + self.codebook_indices = None + + def set_num_updates(self, num_updates): + self.curr_temp = max( + self.max_temp * self.temp_decay ** num_updates, self.min_temp + ) + def get_codebook_indices(self): + if self.codebook_indices is None: + from itertools import product + + p = [range(self.num_vars)] * self.groups + inds = list(product(*p)) + self.codebook_indices = torch.tensor( + inds, dtype=torch.long, device=self.vars.device + ).flatten() + + if not self.combine_groups: + self.codebook_indices = self.codebook_indices.view( + self.num_vars ** self.groups, -1 + ) + for b in range(1, self.groups): + self.codebook_indices[:, b] += self.num_vars * b + self.codebook_indices = self.codebook_indices.flatten() + return self.codebook_indices + + def codebook(self): + indices = self.get_codebook_indices() + return ( + self.vars.squeeze(0) + .index_select(0, indices) + .view(self.num_vars ** self.groups, -1) + ) + + def sample_from_codebook(self, b, n): + indices = self.get_codebook_indices() + indices = indices.view(-1, self.groups) + cb_size = indices.size(0) + assert ( + n < cb_size + ), f"sample size {n} is greater than size of codebook {cb_size}" + sample_idx = torch.randint(low=0, high=cb_size, size=(b * n,)) + indices = indices[sample_idx] + + z = self.vars.squeeze(0).index_select(0, indices.flatten()).view(b, n, -1) + return z + + def to_codebook_index(self, indices): + res = indices.new_full(indices.shape[:-1], 0) + for i in range(self.groups): + exponent = self.groups - i - 1 + res += indices[..., i] * (self.num_vars ** exponent) + return res + + def forward_idx(self, x): + res = self.forward(x, produce_targets=True) + return res["x"], res["targets"] + + def forward(self, x, produce_targets=False): + + result = {"num_vars": self.num_vars * self.groups} + + if not self.time_first: + x = x.transpose(1, 2) + + bsz, tsz, fsz = x.shape + x = x.reshape(-1, fsz) + x = self.weight_proj(x) + x = x.view(bsz * tsz * self.groups, -1) + + _, k = x.max(-1) + hard_x = ( + x.new_zeros(*x.shape) + .scatter_(-1, k.view(-1, 1), 1.0) + .view(bsz * tsz, self.groups, -1) + ) + hard_probs = torch.mean(hard_x.float(), dim=0) + result["code_perplexity"] = torch.exp( + -torch.sum(hard_probs * torch.log(hard_probs + 1e-7), dim=-1) + ).sum() + + avg_probs = torch.softmax( + x.view(bsz * tsz, self.groups, -1).float(), dim=-1 + ).mean(dim=0) + result["prob_perplexity"] = torch.exp( + -torch.sum(avg_probs * torch.log(avg_probs + 1e-7), dim=-1) + ).sum() + + result["temp"] = self.curr_temp + + if self.training: + x = F.gumbel_softmax(x.float(), tau=self.curr_temp, hard=True).type_as(x) + else: + x = hard_x + + x = x.view(bsz * tsz, -1) + + vars = self.vars + if self.combine_groups: + vars = vars.repeat(1, self.groups, 1) + + if produce_targets: + result["targets"] = ( + x.view(bsz * tsz * self.groups, -1) + .argmax(dim=-1) + .view(bsz, tsz, self.groups) + .detach() + ) + + x = x.unsqueeze(-1) * vars + x = x.view(bsz * tsz, self.groups, self.num_vars, -1) + x = x.sum(-2) + x = x.view(bsz, tsz, -1) + + if not self.time_first: + x = x.transpose(1, 2) # BTC -> BCT + + result["x"] = x + + return result diff --git a/fairseq/modules/kmeans_vector_quantizer.py b/fairseq/modules/kmeans_vector_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..be56e6081bc836b77203236b495e4f6391e56020 --- /dev/null +++ b/fairseq/modules/kmeans_vector_quantizer.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn + +from fairseq.modules import Fp32GroupNorm + + +class KmeansVectorQuantizer(nn.Module): + def __init__( + self, dim, num_vars, groups, combine_groups, vq_dim, time_first, gamma=0.25 + ): + '''Vector quantization using straight pass-through estimator (i.e. kmeans) + + Args: + dim: input dimension (channels) + num_vars: number of quantized vectors per group + groups: number of groups for vector quantization + combine_groups: whether to use the vectors for all groups + vq_dim: dimensionality of the resulting quantized vector + time_first: if true, expect input in BxTxC format, otherwise in BxCxT + gamma: commitment loss coefficient + ''' + super().__init__() + + self.groups = groups + self.combine_groups = combine_groups + self.input_dim = dim + self.num_vars = num_vars + self.vq_dim = vq_dim + self.time_first = time_first + + assert ( + vq_dim % groups == 0 + ), f"dim {vq_dim} must be divisible by groups {groups} for concatenation" + + self.var_dim = vq_dim // groups + num_groups = groups if not combine_groups else 1 + + self.embedding = nn.Parameter( + 0.01 * torch.randn(num_vars, num_groups, self.var_dim) + ) + self.projection = nn.Sequential( + nn.Conv1d(dim, dim, kernel_size=1, groups=groups, bias=False), + Fp32GroupNorm(groups, dim), + ) + self.gamma = gamma + self.mse_mean = nn.MSELoss(reduction="mean") + + def _pass_grad(self, x, y): + """ Manually set gradient for backward pass. + for y = f(x), ensure that during the backward pass, + dL/dy = dL/dx regardless of f(x). + Returns: + y, with the gradient forced to be dL/dy = dL/dx. + """ + + return y.detach() + (x - x.detach()) + + @property + def expand_embedding(self): + if self.combine_groups: + return self.embedding.expand(self.num_vars, self.groups, self.var_dim) + return self.embedding + + def forward_idx(self, x): + res = self.forward(x, produce_targets=True) + return res["x"], res["targets"] + + def forward(self, x, produce_targets=False): + + result = {"num_vars": self.num_vars} + + if self.time_first: + x = x.transpose(1, 2) + + bsz, fsz, tsz = x.shape + + ze = self.projection(x) + ze_ = ze.view(bsz, self.groups, self.var_dim, tsz).permute(0, 3, 1, 2) + d = ( + (ze_.unsqueeze(0) - self.expand_embedding.unsqueeze(1).unsqueeze(1)) + .view(self.num_vars, bsz, tsz, self.groups, -1) + .norm(dim=-1, p=2) + ) + idx = d.argmin(dim=0) + zq = ( + torch.stack( + [ + self.expand_embedding[idx[..., group], group] + for group in range(self.groups) + ], + dim=-2, + ) + .view(bsz, tsz, self.groups * self.var_dim) + .permute(0, 2, 1) + ) + assert ze.shape == zq.shape, (ze.shape, zq.shape) + x = self._pass_grad(ze, zq) + + hard_x = ( + idx.new_zeros(bsz*tsz*self.groups, self.num_vars) + .scatter_(-1, idx.view(-1, 1), 1.0) + .view(bsz * tsz, self.groups, -1) + ) + hard_probs = torch.mean(hard_x.float(), dim=0) + result["code_perplexity"] = torch.exp( + -torch.sum(hard_probs * torch.log(hard_probs + 1e-7), dim=-1) + ).sum() + + if produce_targets: + result["targets"] = idx + + if self.time_first: + x = x.transpose(1, 2) # BCT -> BTC + result["x"] = x + + ze = ze.float() + zq = zq.float() + latent_loss = self.mse_mean(zq, ze.detach()) + commitment_loss = self.mse_mean(ze, zq.detach()) + + result["kmeans_loss"] = latent_loss + self.gamma * commitment_loss + + return result diff --git a/fairseq/modules/layer_drop.py b/fairseq/modules/layer_drop.py new file mode 100644 index 0000000000000000000000000000000000000000..8961d8bcbc492c40c6b30973234416ce5a414f5a --- /dev/null +++ b/fairseq/modules/layer_drop.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +LayerDrop as described in https://arxiv.org/abs/1909.11556. +""" + +import torch +import torch.nn as nn + + +class LayerDropModuleList(nn.ModuleList): + """ + A LayerDrop implementation based on :class:`torch.nn.ModuleList`. + + We refresh the choice of which layers to drop every time we iterate + over the LayerDropModuleList instance. During evaluation we always + iterate over all layers. + + Usage:: + + layers = LayerDropList(p=0.5, modules=[layer1, layer2, layer3]) + for layer in layers: # this might iterate over layers 1 and 3 + x = layer(x) + for layer in layers: # this might iterate over all layers + x = layer(x) + for layer in layers: # this might not iterate over any layers + x = layer(x) + + Args: + p (float): probability of dropping out each layer + modules (iterable, optional): an iterable of modules to add + """ + + def __init__(self, p, modules=None): + super().__init__(modules) + self.p = p + + def __iter__(self): + dropout_probs = torch.empty(len(self)).uniform_() + for i, m in enumerate(super().__iter__()): + if not self.training or (dropout_probs[i] > self.p): + yield m diff --git a/fairseq/modules/layer_norm.py b/fairseq/modules/layer_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..4fee32d4fcfa8ff087765ae29028839e090a0288 --- /dev/null +++ b/fairseq/modules/layer_norm.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +try: + from apex.normalization import FusedLayerNorm as _FusedLayerNorm + + has_fused_layernorm = True + + class FusedLayerNorm(_FusedLayerNorm): + @torch.jit.unused + def forward(self, x): + if not x.is_cuda: + return super().forward(x) + else: + with torch.cuda.device(x.device): + return super().forward(x) + +except ImportError: + has_fused_layernorm = False + + +def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False): + if not export and torch.cuda.is_available() and has_fused_layernorm: + return FusedLayerNorm(normalized_shape, eps, elementwise_affine) + return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) + + +class Fp32LayerNorm(nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.layer_norm( + input.float(), + self.normalized_shape, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) diff --git a/fairseq/modules/learned_positional_embedding.py b/fairseq/modules/learned_positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..378d0f707183dd344dbb9288dda394b11053acf0 --- /dev/null +++ b/fairseq/modules/learned_positional_embedding.py @@ -0,0 +1,61 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from torch import Tensor + + +class LearnedPositionalEmbedding(nn.Embedding): + """ + This module learns positional embeddings up to a fixed maximum size. + Padding ids are ignored by either offsetting based on padding_idx + or by setting padding_idx to None and ensuring that the appropriate + position ids are passed to the forward function. + """ + + def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int): + super().__init__(num_embeddings, embedding_dim, padding_idx) + self.onnx_trace = False + if self.padding_idx is not None: + self.max_positions = self.num_embeddings - self.padding_idx - 1 + else: + self.max_positions = self.num_embeddings + + def forward( + self, + input: Tensor, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + positions: Optional[Tensor] = None, + ): + """Input is expected to be of size [bsz x seqlen].""" + assert (positions is None) or ( + self.padding_idx is None + ), "If positions is pre-computed then padding_idx should not be set." + + if positions is None: + if incremental_state is not None: + # positions is the same for every token when decoding a single step + # Without the int() cast, it doesn't work in some cases when exporting to ONNX + positions = torch.zeros( + (1, 1), device=input.device, dtype=input.dtype + ).fill_(int(self.padding_idx + input.size(1))) + else: + positions = utils.make_positions( + input, self.padding_idx, onnx_trace=self.onnx_trace + ) + return F.embedding( + positions, + self.weight, + self.padding_idx, + self.max_norm, + self.norm_type, + self.scale_grad_by_freq, + self.sparse, + ) diff --git a/fairseq/modules/lightconv_layer/__init__.py b/fairseq/modules/lightconv_layer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3b2a99c1227f827768911e5e22e79f6865ffbfd3 --- /dev/null +++ b/fairseq/modules/lightconv_layer/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .lightconv_layer import LightconvLayer # noqa diff --git a/fairseq/modules/lightconv_layer/cuda_function_gen.py b/fairseq/modules/lightconv_layer/cuda_function_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..afec9e19e7176a19c5e60389cb1bb0250c84de4b --- /dev/null +++ b/fairseq/modules/lightconv_layer/cuda_function_gen.py @@ -0,0 +1,289 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def gen_forward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" + +std::vector lightconv_cuda_forward(at::Tensor input, at::Tensor filters, int padding_l) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = filters.size(0); + const auto filterSize = filters.size(1); + + const auto numFiltersInBlock = numFeatures / numHeads; + + const dim3 blocks(minibatch, numFeatures); + + auto output = at::zeros_like(input); + auto stream = at::cuda::getCurrentCUDAStream(); +""" + + sequence_if = """ + if (sequenceLength <= {seq}) {{ + switch(filterSize) {{ +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {pad}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "lightconv_forward", ([&] {{ + lightconv_forward_kernel<{k}, {b_size}, {pad}, scalar_t> + <<>>( + input.data(), + filters.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + output.data()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping forward pass" << std::endl; + } + break; +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping forward pass" << std::endl; + } +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + final_return = """ + } + + return {output}; +} +""" + + with open("lightconv_cuda_forward.cu", 'w') as forward: + forward.write(head) + for seq in seqs: + forward.write(sequence_if.format(seq=seq)) + for k in kernels: + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=seq, pad=pad)) + forward.write(bad_padding) + forward.write(bad_filter) + forward.write(con_else) + + forward.write(final_else) + for k in kernels: + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=seq, pad=pad)) + forward.write(bad_padding) + forward.write(bad_filter) + forward.write(final_return) + + +def gen_backward(): + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" + +std::vector lightconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + + // gradWrtInput + const int minibatch = input.size(0); + const int numFeatures = input.size(1); + const int sequenceLength = input.size(2); + + const int numHeads = filters.size(0); + const int filterSize = filters.size(1); + + const dim3 gradBlocks(minibatch, numFeatures); + const dim3 weightGradFirstpassShortBlocks(minibatch, numHeads); + const dim3 weightGradSecondpassBlocks(numHeads, filterSize); + + const int numFiltersInBlock = numFeatures / numHeads; + + auto gradInput = at::zeros_like(input); + auto gradFilters = at::zeros_like(filters); + + at::DeviceGuard g(input.device()); + auto stream = at::cuda::getCurrentCUDAStream(); + + switch(filterSize) { +""" + + sequence_if = """ + if (sequenceLength <= {seq}) {{ +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {p}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "lightconv_backward", ([&] {{ + lightconv_grad_wrt_input_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + gradOutput.data(), + filters.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + gradInput.data()); + +""" + + weight_grad_short = """ + at::Tensor tempSumGradFilters = at::zeros({{minibatch, numHeads, filterSize}}, input.options().dtype(at::kFloat)); + lightconv_grad_wrt_weights_firstpass_short_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + input.data(), + gradOutput.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + tempSumGradFilters.data() + ); + + lightconv_grad_wrt_weights_secondpass_short_kernel<{k}, {b_size}, scalar_t> + <<>>( + tempSumGradFilters.data(), + minibatch, + numFiltersInBlock, + gradFilters.data() + ); + }})); + }} else +""" + + weight_grad = """ + at::Tensor tempSumGradFilters = at::zeros({{minibatch, numFeatures, filterSize}}, input.options().dtype(at::kFloat)); + lightconv_grad_wrt_weights_firstpass_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + input.data(), + gradOutput.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + tempSumGradFilters.data() + ); + + lightconv_grad_wrt_weights_secondpass_kernel<{k}, {b_size}, scalar_t> + <<>>( + tempSumGradFilters.data(), + minibatch, + numFiltersInBlock, + gradFilters.data() + ); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping backward pass" << std::endl; + } +""" + + breakout = """ + break; +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping backward pass" << std::endl; +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + last_return = """ + } + return {gradInput, gradFilters}; +} +""" + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + thresh = [32, 32, 64, 128, 256, -1, -1, -1] + max_mem = [-1, -1, -1, -1, -1, 192, 96, 64] + + with open("lightconv_cuda_backward.cu", 'w') as backward: + backward.write(head) + for (k, t, mem) in zip(kernels, thresh, max_mem): + backward.write(case_k.format(k=k)) + for seq in seqs: + if (t == -1 or seq <= t) and (mem == -1 or seq < mem): + backward.write(sequence_if.format(seq=seq)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=seq, p=p)) + backward.write(weight_grad_short.format(k=k, b_size=seq, p=p)) + backward.write(bad_padding) + else: + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=32, p=p)) + backward.write(weight_grad.format(k=k, b_size=32, p=p)) + backward.write(bad_padding) + backward.write(breakout) + break + backward.write(con_else) + backward.write(bad_filter) + backward.write(last_return) + + +if __name__ == "__main__": + gen_forward() + gen_backward() diff --git a/fairseq/modules/lightconv_layer/lightconv_cuda.cpp b/fairseq/modules/lightconv_layer/lightconv_cuda.cpp new file mode 100644 index 0000000000000000000000000000000000000000..4bf6b5ad365d604bd91eda384bb422857b640744 --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_cuda.cpp @@ -0,0 +1,54 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +std::vector lightconv_cuda_forward( + at::Tensor input, + at::Tensor filters, + int padding_l); + +std::vector lightconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters); + + +#define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + +std::vector lightconv_forward( + at::Tensor input, + at::Tensor filters, + int padding_l) { + + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return lightconv_cuda_forward(input, filters, padding_l); +} + +std::vector lightconv_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + + CHECK_INPUT(gradOutput); + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return lightconv_cuda_backward(gradOutput, padding_l, input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &lightconv_forward, "lighconv forward (CUDA)"); + m.def("backward", &lightconv_backward, "lighconv backward (CUDA)"); +} diff --git a/fairseq/modules/lightconv_layer/lightconv_cuda.cuh b/fairseq/modules/lightconv_layer/lightconv_cuda.cuh new file mode 100644 index 0000000000000000000000000000000000000000..3cae57b68fc96872a5047a7a0d081b78456e8fae --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_cuda.cuh @@ -0,0 +1,83 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +#include +#include + +#include +#include +#include +#include +#include +#include + +#include +#include + +#define SHFL_MASK 0xffffffff + +template +__global__ +void lightconv_forward_kernel(const scalar_t* input, + const scalar_t* filters, + int minibatch, int sequenceLength, + int numFeatures, int numFiltersInBlock, + scalar_t* output); + +template +__global__ +void lightconv_grad_wrt_input_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output); + +template +__global__ +void lightconv_grad_wrt_weights_firstpass_short_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + float* output); + +template +__global__ +void lightconv_grad_wrt_weights_secondpass_short_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output); + +template +__global__ +void lightconv_grad_wrt_weights_firstpass_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + float* output); + +template +__global__ +void lightconv_grad_wrt_weights_secondpass_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output); + diff --git a/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu b/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..8ee83a56c89754c2abbe717b269d07ca9e64eef2 --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu @@ -0,0 +1,375 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" +#include "lightconv_cuda_forward.cu" +#include "lightconv_cuda_backward.cu" +#include "../cuda_utils.cu" + +template +__global__ +void lightconv_forward_kernel(const scalar_t* input, + const scalar_t* filters, + int minibatch, int sequenceLength, + int numFeatures, int numFiltersInBlock, + scalar_t* output) { + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + + const int IOOffset = numFeatures * sequenceLength * batchIdx + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + const scalar_t* inputFilter = &filters[filterIdx * FS]; + + assert(blockDim.x == SB); + + scalar_t filter[FS]; + #pragma unroll + for (int i = 0; i < FS; ++i) { + filter[i] = inputFilter[i]; + } + + __shared__ scalar_t temp[SB + FS]; + zeroSharedMem(temp); + + const int numIterations = divUp(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + // Read input into shared memory + const int inputOffset = i * SB; + + load_input_to_shared(inputFeature, inputOffset, sequenceLength, + i, numIterations, (numIterations == 1), temp); + + __syncthreads(); + + scalar_t out = 0; + #pragma unroll + for (int j = 0; j < FS; ++j) { + out += filter[j] * temp[tid + j]; + } + + // Write output + const int outputOffset = inputOffset; + if ((outputOffset + tid) < sequenceLength) { + outputFeature[outputOffset + tid] = out; + } + + __syncthreads(); + } +} + +template +__global__ +void lightconv_grad_wrt_input_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output) { + + // input grad kernel is similar to forward kernel + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + + const int IOOffset = numFeatures * sequenceLength * batchIdx + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + const scalar_t* inputFilter = &filters[filterIdx * FS]; + + assert(blockDim.x == SB); + + scalar_t filter[FS]; + + // The only change is loading the filter in reverse + #pragma unroll + for (int i = 0; i < FS; ++i) { + filter[i] = inputFilter[FS - i - 1]; + } + + __shared__ scalar_t temp[SB + FS]; + const int padding = FS - padding_l - 1; + zeroSharedMem(temp); + + __syncthreads(); + + const int numIterations = divUp(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + // Read input into shared memory + const int inputOffset = i * SB; + + load_input_to_shared(inputFeature, inputOffset, sequenceLength, + i, numIterations, false, temp); + + __syncthreads(); + + scalar_t out = 0; + #pragma unroll + for (int j = 0; j < FS; ++j) { + out += filter[j] * temp[tid + j]; + } + + // Write output + const int outputOffset = inputOffset; + if ((outputOffset + tid) < sequenceLength) { + outputFeature[outputOffset + tid] = out; + } + + __syncthreads(); + } +} + +// This is by far the most expensive kernel in terms of time taken. +// Can be 16x slower than the forward or grad_wrt_input when filter size is 31 +template +__global__ +void lightconv_grad_wrt_weights_firstpass_short_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + float* output) { + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int filterIdx = blockIdx.y; + + const int numIterations = divUp(sequenceLength, SB); + + float* tempOutputGradWeight = &output[filterIdx * FS * minibatch]; + + assert(blockDim.x == SB); + + __shared__ scalar_t tempInput[SB + FS]; + __shared__ scalar_t tempGradInput[SB + FS]; + + // local weight accumulation + float accumWeights[FS]; + + // Initialize memory + for (int i = 0; i < FS; ++i) { + accumWeights[i] = float(0.0); + } + + + // loop over each sequence within filterblock + for (int idxInFilterBlock = 0; idxInFilterBlock < numFiltersInBlock; ++idxInFilterBlock) { + + const int featureOffset = batchIdx * numFeatures * sequenceLength + (filterIdx * numFiltersInBlock + idxInFilterBlock) * sequenceLength; + const scalar_t* inputFeature = &input[featureOffset]; + const scalar_t* gradInputFeature = &gradInput[featureOffset]; + + zeroSharedMem(tempInput); + zeroSharedMem(tempGradInput); + __syncthreads(); + + for (int i = 0; i < numIterations; ++i) { + + const int inputOffset = i * SB; + + load_input_to_shared(inputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempInput); + load_input_to_shared(gradInputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempGradInput); + + __syncthreads(); + + const int gradIndex = (FS/2) + tid; + scalar_t tempGrad = tempGradInput[gradIndex]; + + #pragma unroll + for (int j = 0; j < FS; j++) { + const int inputIndex = tid + j; + accumWeights[j] += tempInput[inputIndex] * tempGrad; + } + + __syncthreads(); + + } + + } + + // Row-major sum + for (int filterWeightIdx = 0; filterWeightIdx < FS; ++filterWeightIdx) { + + float temp; + if (tid < sequenceLength) { + temp = accumWeights[filterWeightIdx]; + } else { + temp = float(0.0); + } + + const int outputOffset = filterWeightIdx * minibatch + batchIdx; + + temp = blockReduce(temp); + + if (tid == 0) { + tempOutputGradWeight[outputOffset] = temp; + } + } +} + +template +__global__ +void lightconv_grad_wrt_weights_secondpass_short_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output) { + + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + + const int filterIdx = blockIdx.x; + const int filterWeightIdx = blockIdx.y; + + const int inputOffset = filterIdx * FS * minibatch + + filterWeightIdx * minibatch; + const float* tempInput = &input[inputOffset]; + + // read into shared memory for reduction + int readIndex = tid; + + float sum = 0.0; + while (readIndex < minibatch) { + sum += tempInput[readIndex]; + readIndex += SB; + } + + float temp = blockReduce(sum); + + if (tid == 0) { + output[blockIdx.x * FS + blockIdx.y] = temp; + } +} + +// This is by far the most expensive kernel in terms of time taken. +// Can be 16x slower than the forward or grad_wrt_input when filter size is 31 +template +__global__ +void lightconv_grad_wrt_weights_firstpass_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + float* output) { + + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + const int idxInFilterBlock = featureIdx % numFiltersInBlock; + + const int numIterations = divUp(sequenceLength, SB); + + float temp; + + __shared__ scalar_t tempInput[SB + FS]; + __shared__ scalar_t tempGradInput[SB + FS]; + zeroSharedMem(tempInput); + zeroSharedMem(tempGradInput); + __syncthreads(); + + float accumWeights[FS]; + + for (int i = 0; i < FS; ++i) { + accumWeights[i] = float(0.0); + } + + const int IOOffset = batchIdx * numFeatures * sequenceLength + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + const scalar_t* gradInputFeature = &gradInput[IOOffset]; + float* tempOutputGradWeight = &output[filterIdx * FS * minibatch * numFiltersInBlock]; + + for (int i = 0; i < numIterations; ++i) { + const int inputOffset = i * SB; + + load_input_to_shared(inputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempInput); + load_input_to_shared(gradInputFeature, inputOffset, sequenceLength, + i, numIterations, false, tempGradInput); + __syncthreads(); + + #pragma unroll + for (int j = 0; j < FS; ++j) { + accumWeights[j] += tempInput[tid + j] * tempGradInput[tid + (FS/2)]; + } + + __syncthreads(); + } + + // Row-major sum + for (int filterWeightIdx = 0; filterWeightIdx < FS; ++filterWeightIdx) { + + // Write to shared memory before reduction + if (tid < sequenceLength) { + temp = accumWeights[filterWeightIdx]; + } else { + temp = float(0.0); + } + + temp = blockReduce(temp); + + const int outputOffset = filterWeightIdx * minibatch * numFiltersInBlock + + batchIdx * numFiltersInBlock + + idxInFilterBlock; + + if (tid == 0) { + tempOutputGradWeight[outputOffset] = temp; + } + } +} + +template +__global__ +void lightconv_grad_wrt_weights_secondpass_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output) { + + assert(blockDim.x == SB); + const int tid = threadIdx.x; + + // What is the id within a minibatch + const int filterIdx = blockIdx.x; + const int filterWeightIdx = blockIdx.y; + + const int inputOffset = filterIdx * FS * minibatch * numFiltersInBlock + + filterWeightIdx * minibatch * numFiltersInBlock; + const float* tempInput = &input[inputOffset]; + + int readIndex = tid; + + float sum = float(0.0); + while (readIndex < (minibatch * numFiltersInBlock)) { + sum += tempInput[readIndex]; + readIndex += SB; + } + + float temp = blockReduce(sum); + + if (tid == 0) { + output[blockIdx.x * FS + blockIdx.y] = temp; + } +} diff --git a/fairseq/modules/lightconv_layer/lightconv_layer.py b/fairseq/modules/lightconv_layer/lightconv_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..9b4c9a951eadb628619a68957258586472d208f4 --- /dev/null +++ b/fairseq/modules/lightconv_layer/lightconv_layer.py @@ -0,0 +1,129 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch import nn +from torch.autograd import Function +import torch.nn.functional as F + +import lightconv_cuda +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout + + +class lightconvFunction(Function): + + @staticmethod + def forward(ctx, x, weights, padding_l): + ctx.padding_l = padding_l + outputs = lightconv_cuda.forward(x, weights, padding_l) + variables = [x, weights] + ctx.save_for_backward(*variables) + return outputs[0] + + @staticmethod + def backward(ctx, grad_output): + outputs = lightconv_cuda.backward( + grad_output.contiguous(), + ctx.padding_l, + *ctx.saved_tensors) + grad_input, grad_weights = outputs + return grad_input, grad_weights, None + + +@with_incremental_state +class LightconvLayer(nn.Module): + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + weight_softmax=False, + num_heads=1, + weight_dropout=0., + bias=False, + ): + super(LightconvLayer, self).__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_softmax = weight_softmax + self.weight_dropout_module = FairseqDropout(weight_dropout, module_name=self.__class__.__name__) + + self.weight = nn.Parameter(torch.Tensor(num_heads, kernel_size)) + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + self.reset_parameters() + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + '.' if name != '' else '' + for k, v in state_dict.items(): + if k.endswith(prefix + 'weight'): + if v.dim() == 3 and v.size(1) == 1: + state_dict[k] = v.squeeze(1) + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.) + + def forward(self, x, incremental_state=None): + + # during inference time, incremental BMM is faster + if incremental_state is not None: + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer(incremental_state, x_unfold[:, :, :, -self.kernel_size+1:]) + x_unfold = x_unfold.view(T*B*H, R, -1) + + weight = self.weight + if self.weight_softmax: + weight = F.softmax(weight.float(), dim=1).type_as(weight) + + weight = weight[:, -x_unfold.size(2):] + + K = weight.size(1) + + weight = weight.view(1, H, K).expand(T*B, H, K).contiguous().view(T*B*H, K, 1) + + weight = self.weight_dropout_module(weight) + output = torch.bmm(x_unfold, weight) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + # during training time, use CUDA kernel + else: + x = x.permute(1, 2, 0).contiguous() + weight = self.weight + if self.weight_softmax: + weight = F.softmax(self.weight, -1) + if self.weight_dropout_module.p: + weight = self.weight_dropout_module(weight) + return lightconvFunction.apply(x, weight, self.padding_l).permute(2, 0, 1) + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, 'input_buffer') + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state(self, incremental_state, 'input_buffer', new_buffer) + + def half(self): + return self._apply(lambda t: t.half() if t.is_floating_point() else t) diff --git a/fairseq/modules/lightconv_layer/setup.py b/fairseq/modules/lightconv_layer/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..0eac1df03cd1eeee8bea56f110c6b9f2f97c0dc5 --- /dev/null +++ b/fairseq/modules/lightconv_layer/setup.py @@ -0,0 +1,20 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from setuptools import setup +from torch.utils.cpp_extension import CUDAExtension, BuildExtension + +setup( + name='lightconv_layer', + ext_modules=[ + CUDAExtension('lightconv_cuda', [ + 'lightconv_cuda.cpp', + 'lightconv_cuda_kernel.cu', + ]), + ], + cmdclass={ + 'build_ext': BuildExtension + }) diff --git a/fairseq/modules/lightweight_convolution.py b/fairseq/modules/lightweight_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..3d4cddb134c0367770380977952d1617efeccf06 --- /dev/null +++ b/fairseq/modules/lightweight_convolution.py @@ -0,0 +1,256 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from fairseq.modules.unfold import unfold1d +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout + + +def LightweightConv(input_size, kernel_size=1, padding_l=None, num_heads=1, + weight_dropout=0., weight_softmax=False, bias=False): + if torch.cuda.is_available(): + try: + from fairseq.modules.lightconv_layer import LightconvLayer + return LightconvLayer(input_size, kernel_size=kernel_size, + padding_l=padding_l, num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, bias=bias) + except ImportError as e: + print(e) + return LightweightConv1dTBC(input_size, kernel_size=kernel_size, + padding_l=padding_l, num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, bias=bias) + + +class LightweightConv1d(nn.Module): + '''Lightweight Convolution assuming the input is BxCxT + This is just an example that explains LightConv clearer than the TBC version. + We don't use this module in the model. + + Args: + input_size: # of channels of the input and output + kernel_size: convolution channels + padding: padding + num_heads: number of heads used. The weight is of shape + `(num_heads, 1, kernel_size)` + weight_softmax: normalize the weight with softmax before the convolution + + Shape: + Input: BxCxT, i.e. (batch_size, input_size, timesteps) + Output: BxCxT, i.e. (batch_size, input_size, timesteps) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + ''' + + def __init__(self, input_size, kernel_size=1, padding=0, num_heads=1, + weight_softmax=False, bias=False, weight_dropout=0.): + super().__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.num_heads = num_heads + self.padding = padding + self.weight_softmax = weight_softmax + self.weight = nn.Parameter(torch.Tensor(num_heads, 1, kernel_size)) + + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + self.weight_dropout_module = FairseqDropout(weight_dropout, module_name=self.__class__.__name__) + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.) + + def forward(self, input): + ''' + input size: B x C x T + output size: B x C x T + ''' + B, C, T = input.size() + H = self.num_heads + + weight = self.weight + if self.weight_softmax: + weight = F.softmax(weight, dim=-1) + + weight = self.weight_dropout_module(weight) + # Merge every C/H entries into the batch dimension (C = self.input_size) + # B x C x T -> (B * C/H) x H x T + # One can also expand the weight to C x 1 x K by a factor of C/H + # and do not reshape the input instead, which is slow though + input = input.view(-1, H, T) + output = F.conv1d(input, weight, padding=self.padding, groups=self.num_heads) + output = output.view(B, C, T) + if self.bias is not None: + output = output + self.bias.view(1, -1, 1) + + return output + + +@with_incremental_state +class LightweightConv1dTBC(nn.Module): + '''Lightweight Convolution assuming the input is TxBxC + Args: + input_size: # of channels of the input + kernel_size: convolution channels + padding_l: padding to the left when using "same" padding + num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size) + weight_dropout: the drop rate of the DropConnect to drop the weight + weight_softmax: normalize the weight with softmax before the convolution + bias: use bias + + Shape: + Input: TxBxC, i.e. (timesteps, batch_size, input_size) + Output: TxBxC, i.e. (timesteps, batch_size, input_size) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + ''' + def __init__(self, input_size, kernel_size=1, padding_l=None, num_heads=1, + weight_dropout=0., weight_softmax=False, bias=False): + super().__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_dropout_module = FairseqDropout(weight_dropout, module_name=self.__class__.__name__) + self.weight_softmax = weight_softmax + + self.weight = nn.Parameter(torch.Tensor(num_heads, 1, kernel_size)) + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + + self.reset_parameters() + self.onnx_trace = False + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.) + + def forward(self, x, incremental_state=None, unfold=False): + '''Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C + args: + x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size) + incremental_state: A dict to keep the state + unfold: unfold the input or not. If not, we use the matrix trick instead + ''' + unfold = unfold or (incremental_state is not None) + + if unfold: + output = self._forward_unfolded(x, incremental_state) + else: + output = self._forward_expanded(x, incremental_state) + + if self.bias is not None: + output = output + self.bias.view(1, 1, -1) + return output + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def _forward_unfolded(self, x, incremental_state): + '''The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.''' + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight.view(H, K) + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer(incremental_state, x_unfold[:, :, :, -self.kernel_size+1:]) + x_unfold = x_unfold.view(T*B*H, R, -1) + else: + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, self.kernel_size, self.padding_l, 0) + x_unfold = x_unfold.view(T*B*H, R, K) + + if self.weight_softmax: + weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as(weight) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2):] + K = weight.size(1) + + weight = weight.view(1, H, K).expand(T*B, H, K).contiguous().view(T*B*H, K, 1) + + weight = self.weight_dropout_module(weight) + output = torch.bmm(x_unfold, weight) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_state): + '''Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + ''' + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight.view(H, K) + if self.weight_softmax: + weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as(weight) + weight = weight.view(1, H, K).expand(T*B, H, K).contiguous() + weight = weight.view(T, B*H, K).transpose(0, 1) + + x = x.view(T, B*H, R).transpose(0, 1) + P = self.padding_l + if K > T and P == K-1: + weight = weight.narrow(2, K-T, T) + K, P = T, T-1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B*H, T, T+K-1, requires_grad=False) + weight_expanded.as_strided((B*H, T, K), (T*(T+K-1), T+K, 1)).copy_(weight) + weight_expanded = weight_expanded.narrow(2, P, T) + weight_expanded = self.weight_dropout_module(weight_expanded) + + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, 'input_buffer') + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state(self, incremental_state, 'input_buffer', new_buffer) + + def extra_repr(self): + s = '{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, bias={}'.format( + self.input_size, self.kernel_size, self.padding_l, + self.num_heads, self.weight_softmax, self.bias is not None + ) + if self.weight_dropout_module.p > 0.: + s += ', weight_dropout={}'.format(self.weight_dropout_module.p) + return s diff --git a/fairseq/modules/linearized_convolution.py b/fairseq/modules/linearized_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..3dd4b151c14bf4cb8968567f2574026a033f2be3 --- /dev/null +++ b/fairseq/modules/linearized_convolution.py @@ -0,0 +1,100 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F + +from fairseq import utils +from .conv_tbc import ConvTBC +from fairseq.incremental_decoding_utils import with_incremental_state + + +@with_incremental_state +class LinearizedConvolution(ConvTBC): + """An optimized version of nn.Conv1d. + + At training time, this module uses ConvTBC, which is an optimized version + of Conv1d. At inference time, it optimizes incremental generation (i.e., + one time step at a time) by replacing the convolutions with linear layers. + Note that the input order changes from training to inference. + """ + + def __init__(self, in_channels, out_channels, kernel_size, **kwargs): + super().__init__(in_channels, out_channels, kernel_size, **kwargs) + self._linearized_weight = None + self.register_backward_hook(self._clear_linearized_weight) + + def state_dict(self, destination=None, prefix='', keep_vars=False): + state = ConvTBC.state_dict(self, destination, prefix, keep_vars=keep_vars) + # don't store redundant _linearized_weight in checkpoints + if prefix + '_linearized_weight' in state: + del state[prefix + '_linearized_weight'] + return state + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + '.' if name != '' else '' + if prefix + '_linearized_weight' in state_dict: + del state_dict[prefix + '_linearized_weight'] + + def forward(self, input, incremental_state=None): + """ + Args: + incremental_state: Used to buffer signal; if not None, then input is + expected to contain a single frame. If the input order changes + between time steps, call reorder_incremental_state. + Input: + Time x Batch x Channel during training + Batch x Time x Channel during inference + """ + if incremental_state is None: + output = super().forward(input) + if self.kernel_size[0] > 1 and self.padding[0] > 0: + # remove future timesteps added by padding + output = output[:-self.padding[0], :, :] + return output + + # reshape weight + weight = self._get_linearized_weight() + kw = self.kernel_size[0] + + bsz = input.size(0) # input: bsz x len x dim + if kw > 1: + input = input.data + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = input.new(bsz, kw, input.size(2)).zero_() + self._set_input_buffer(incremental_state, input_buffer) + else: + # shift buffer + input_buffer[:, :-1, :] = input_buffer[:, 1:, :].clone() + # append next input + input_buffer[:, -1, :] = input[:, -1, :] + input = input_buffer + with torch.no_grad(): + output = F.linear(input.view(bsz, -1), weight, self.bias) + return output.view(bsz, 1, -1) + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(0, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, 'input_buffer') + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state(self, incremental_state, 'input_buffer', new_buffer) + + def _get_linearized_weight(self): + if self._linearized_weight is None: + kw = self.kernel_size[0] + weight = self.weight.transpose(2, 1).transpose(1, 0).contiguous() + assert weight.size() == (self.out_channels, kw, self.in_channels) + self._linearized_weight = torch.nn.Parameter(weight.view(self.out_channels, -1)) + return self._linearized_weight + + def _clear_linearized_weight(self, *args): + self._linearized_weight = None diff --git a/fairseq/modules/multihead_attention.py b/fairseq/modules/multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..e33dd450ee837f6a3150efac6d4c05edc637eb04 --- /dev/null +++ b/fairseq/modules/multihead_attention.py @@ -0,0 +1,477 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, Optional, Tuple + +import torch +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn import Parameter + +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise + + +@with_incremental_state +class MultiheadAttention(nn.Module): + """Multi-headed attention. + + See "Attention Is All You Need" for more details. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + q_noise=0.0, + qn_block_size=8, + ): + super().__init__() + self.embed_dim = embed_dim + self.kdim = kdim if kdim is not None else embed_dim + self.vdim = vdim if vdim is not None else embed_dim + self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim + + self.num_heads = num_heads + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + self.scaling = self.head_dim ** -0.5 + + self.self_attention = self_attention + self.encoder_decoder_attention = encoder_decoder_attention + + assert not self.self_attention or self.qkv_same_dim, ( + "Self-attention requires query, key and " "value to be of the same size" + ) + + self.k_proj = quant_noise(nn.Linear(self.kdim, embed_dim, bias=bias), q_noise, qn_block_size) + self.v_proj = quant_noise(nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size) + self.q_proj = quant_noise(nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size) + + self.out_proj = quant_noise(nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size) + + if add_bias_kv: + self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim)) + self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim)) + else: + self.bias_k = self.bias_v = None + + self.add_zero_attn = add_zero_attn + + self.reset_parameters() + + self.onnx_trace = False + self.tpu = False + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def prepare_for_tpu_(self, **kwargs): + self.tpu = True + + def reset_parameters(self): + if self.qkv_same_dim: + # Empirically observed the convergence to be much better with + # the scaled initialization + nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2)) + else: + nn.init.xavier_uniform_(self.k_proj.weight) + nn.init.xavier_uniform_(self.v_proj.weight) + nn.init.xavier_uniform_(self.q_proj.weight) + + nn.init.xavier_uniform_(self.out_proj.weight) + if self.out_proj.bias is not None: + nn.init.constant_(self.out_proj.bias, 0.) + if self.bias_k is not None: + nn.init.xavier_normal_(self.bias_k) + if self.bias_v is not None: + nn.init.xavier_normal_(self.bias_v) + + def forward( + self, + query, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + need_weights: bool = True, + static_kv: bool = False, + attn_mask: Optional[Tensor] = None, + before_softmax: bool = False, + need_head_weights: bool = False, + ) -> Tuple[Tensor, Optional[Tensor]]: + """Input shape: Time x Batch x Channel + + Args: + key_padding_mask (ByteTensor, optional): mask to exclude + keys that are pads, of shape `(batch, src_len)`, where + padding elements are indicated by 1s. + need_weights (bool, optional): return the attention weights, + averaged over heads (default: False). + attn_mask (ByteTensor, optional): typically used to + implement causal attention, where the mask prevents the + attention from looking forward in time (default: None). + before_softmax (bool, optional): return the raw attention + weights and values before the attention softmax. + need_head_weights (bool, optional): return the attention + weights for each head. Implies *need_weights*. Default: + return the average attention weights over all heads. + """ + if need_head_weights: + need_weights = True + + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + + if ( + not self.onnx_trace + and not self.tpu # don't use PyTorch version on TPUs + and incremental_state is None + and not static_kv + # A workaround for quantization to work. Otherwise JIT compilation + # treats bias in linear module as method. + and not torch.jit.is_scripting() + ): + assert key is not None and value is not None + return F.multi_head_attention_forward( + query, + key, + value, + self.embed_dim, + self.num_heads, + torch.empty([0]), + torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)), + self.bias_k, + self.bias_v, + self.add_zero_attn, + self.dropout_module.p, + self.out_proj.weight, + self.out_proj.bias, + self.training or self.dropout_module.apply_during_inference, + key_padding_mask, + need_weights, + attn_mask, + use_separate_proj_weight=True, + q_proj_weight=self.q_proj.weight, + k_proj_weight=self.k_proj.weight, + v_proj_weight=self.v_proj.weight, + ) + + if incremental_state is not None: + saved_state = self._get_input_buffer(incremental_state) + if saved_state is not None and "prev_key" in saved_state: + # previous time steps are cached - no need to recompute + # key and value if they are static + if static_kv: + assert self.encoder_decoder_attention and not self.self_attention + key = value = None + else: + saved_state = None + + if self.self_attention: + q = self.q_proj(query) + k = self.k_proj(query) + v = self.v_proj(query) + elif self.encoder_decoder_attention: + # encoder-decoder attention + q = self.q_proj(query) + if key is None: + assert value is None + k = v = None + else: + k = self.k_proj(key) + v = self.v_proj(key) + + else: + assert key is not None and value is not None + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + q *= self.scaling + + if self.bias_k is not None: + assert self.bias_v is not None + k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)]) + v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)]) + if attn_mask is not None: + attn_mask = torch.cat( + [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 + ) + if key_padding_mask is not None: + key_padding_mask = torch.cat( + [ + key_padding_mask, + key_padding_mask.new_zeros(key_padding_mask.size(0), 1), + ], + dim=1, + ) + + q = ( + q.contiguous() + .view(tgt_len, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if k is not None: + k = ( + k.contiguous() + .view(-1, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if v is not None: + v = ( + v.contiguous() + .view(-1, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + if saved_state is not None: + # saved states are stored with shape (bsz, num_heads, seq_len, head_dim) + if "prev_key" in saved_state: + _prev_key = saved_state["prev_key"] + assert _prev_key is not None + prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim) + if static_kv: + k = prev_key + else: + assert k is not None + k = torch.cat([prev_key, k], dim=1) + if "prev_value" in saved_state: + _prev_value = saved_state["prev_value"] + assert _prev_value is not None + prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim) + if static_kv: + v = prev_value + else: + assert v is not None + v = torch.cat([prev_value, v], dim=1) + prev_key_padding_mask: Optional[Tensor] = None + if "prev_key_padding_mask" in saved_state: + prev_key_padding_mask = saved_state["prev_key_padding_mask"] + assert k is not None and v is not None + key_padding_mask = MultiheadAttention._append_prev_key_padding_mask( + key_padding_mask=key_padding_mask, + prev_key_padding_mask=prev_key_padding_mask, + batch_size=bsz, + src_len=k.size(1), + static_kv=static_kv, + ) + + saved_state["prev_key"] = k.view(bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_value"] = v.view(bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_key_padding_mask"] = key_padding_mask + # In this branch incremental_state is never None + assert incremental_state is not None + incremental_state = self._set_input_buffer(incremental_state, saved_state) + assert k is not None + src_len = k.size(1) + + # This is part of a workaround to get around fork/join parallelism + # not supporting Optional types. + if key_padding_mask is not None and key_padding_mask.dim() == 0: + key_padding_mask = None + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == src_len + + if self.add_zero_attn: + assert v is not None + src_len += 1 + k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1) + v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1) + if attn_mask is not None: + attn_mask = torch.cat( + [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1 + ) + if key_padding_mask is not None: + key_padding_mask = torch.cat( + [ + key_padding_mask, + torch.zeros(key_padding_mask.size(0), 1).type_as( + key_padding_mask + ), + ], + dim=1, + ) + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + attn_weights = MultiheadAttention.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz) + + assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len] + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + if self.onnx_trace: + attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1) + attn_weights += attn_mask + + if key_padding_mask is not None: + # don't attend to padding symbols + attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + if not self.tpu: + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + float("-inf") + ) + else: + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.masked_fill(key_padding_mask, float('-inf')) + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + if before_softmax: + return attn_weights, v + + attn_weights_float = utils.softmax( + attn_weights, dim=-1, onnx_trace=self.onnx_trace + ) + attn_weights = attn_weights_float.type_as(attn_weights) + attn_probs = self.dropout_module(attn_weights) + + assert v is not None + attn = torch.bmm(attn_probs, v) + assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim] + if self.onnx_trace and attn.size(1) == 1: + # when ONNX tracing a single decoder step (sequence length == 1) + # the transpose is a no-op copy before view, thus unnecessary + attn = attn.contiguous().view(tgt_len, bsz, embed_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) + attn = self.out_proj(attn) + attn_weights: Optional[Tensor] = None + if need_weights: + attn_weights = attn_weights_float.view( + bsz, self.num_heads, tgt_len, src_len + ).transpose(1, 0) + if not need_head_weights: + # average attention weights over heads + attn_weights = attn_weights.mean(dim=0) + + return attn, attn_weights + + @staticmethod + def _append_prev_key_padding_mask( + key_padding_mask: Optional[Tensor], + prev_key_padding_mask: Optional[Tensor], + batch_size: int, + src_len: int, + static_kv: bool, + ) -> Optional[Tensor]: + # saved key padding masks have shape (bsz, seq_len) + if prev_key_padding_mask is not None and static_kv: + new_key_padding_mask = prev_key_padding_mask + elif prev_key_padding_mask is not None and key_padding_mask is not None: + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1 + ) + # During incremental decoding, as the padding token enters and + # leaves the frame, there will be a time when prev or current + # is None + elif prev_key_padding_mask is not None: + filler = torch.zeros( + (batch_size, src_len - prev_key_padding_mask.size(1)), + device=prev_key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), filler.float()], dim=1 + ) + elif key_padding_mask is not None: + filler = torch.zeros( + (batch_size, src_len - key_padding_mask.size(1)), + device=key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [filler.float(), key_padding_mask.float()], dim=1 + ) + else: + new_key_padding_mask = prev_key_padding_mask + return new_key_padding_mask + + @torch.jit.export + def reorder_incremental_state( + self, incremental_state: Dict[str, Dict[str, Optional[Tensor]]], new_order: Tensor + ): + """Reorder buffered internal state (for incremental generation).""" + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + for k in input_buffer.keys(): + input_buffer_k = input_buffer[k] + if input_buffer_k is not None: + if self.encoder_decoder_attention and input_buffer_k.size(0) == new_order.size(0): + break + input_buffer[k] = input_buffer_k.index_select(0, new_order) + incremental_state = self._set_input_buffer(incremental_state, input_buffer) + return incremental_state + + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ) -> Dict[str, Optional[Tensor]]: + result = self.get_incremental_state(incremental_state, "attn_state") + if result is not None: + return result + else: + empty_result: Dict[str, Optional[Tensor]] = {} + return empty_result + + def _set_input_buffer( + self, + incremental_state: Dict[str, Dict[str, Optional[Tensor]]], + buffer: Dict[str, Optional[Tensor]], + ): + return self.set_incremental_state(incremental_state, "attn_state", buffer) + + def apply_sparse_mask(attn_weights, tgt_len: int, src_len: int, bsz: int): + return attn_weights + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + items_to_add = {} + keys_to_remove = [] + for k in state_dict.keys(): + if k.endswith(prefix + "in_proj_weight"): + # in_proj_weight used to be q + k + v with same dimensions + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim] + items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim] + items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :] + + keys_to_remove.append(k) + + k_bias = prefix + "in_proj_bias" + if k_bias in state_dict.keys(): + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim] + items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][ + dim : 2 * dim + ] + items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :] + + keys_to_remove.append(prefix + "in_proj_bias") + + for k in keys_to_remove: + del state_dict[k] + + for key, value in items_to_add.items(): + state_dict[key] = value diff --git a/fairseq/modules/positional_embedding.py b/fairseq/modules/positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..511460fcb711ac6016d635f26beee44ae3d630af --- /dev/null +++ b/fairseq/modules/positional_embedding.py @@ -0,0 +1,32 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +from .learned_positional_embedding import LearnedPositionalEmbedding +from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding + + +def PositionalEmbedding( + num_embeddings: int, + embedding_dim: int, + padding_idx: int, + learned: bool = False, +): + if learned: + # if padding_idx is specified then offset the embedding ids by + # this index and adjust num_embeddings appropriately + # TODO: The right place for this offset would be inside + # LearnedPositionalEmbedding. Move this there for a cleaner implementation. + if padding_idx is not None: + num_embeddings = num_embeddings + padding_idx + 1 + m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) + if padding_idx is not None: + nn.init.constant_(m.weight[padding_idx], 0) + else: + m = SinusoidalPositionalEmbedding( + embedding_dim, padding_idx, init_size=num_embeddings + padding_idx + 1, + ) + return m diff --git a/fairseq/modules/quant_noise.py b/fairseq/modules/quant_noise.py new file mode 100644 index 0000000000000000000000000000000000000000..b38ea263d32337affc612b6ca185254bf59e8bf0 --- /dev/null +++ b/fairseq/modules/quant_noise.py @@ -0,0 +1,90 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn + + +def quant_noise(module, p, block_size): + """ + Wraps modules and applies quantization noise to the weights for + subsequent quantization with Iterative Product Quantization as + described in "Training with Quantization Noise for Extreme Model Compression" + + Args: + - module: nn.Module + - p: amount of Quantization Noise + - block_size: size of the blocks for subsequent quantization with iPQ + + Remarks: + - Module weights must have the right sizes wrt the block size + - Only Linear, Embedding and Conv2d modules are supported for the moment + - For more detail on how to quantize by blocks with convolutional weights, + see "And the Bit Goes Down: Revisiting the Quantization of Neural Networks" + - We implement the simplest form of noise here as stated in the paper + which consists in randomly dropping blocks + """ + + # if no quantization noise, don't register hook + if p <= 0: + return module + + # supported modules + assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d)) + + # test whether module.weight has the right sizes wrt block_size + is_conv = module.weight.ndim == 4 + + # 2D matrix + if not is_conv: + assert module.weight.size(1) % block_size == 0, "Input features must be a multiple of block sizes" + + # 4D matrix + else: + # 1x1 convolutions + if module.kernel_size == (1, 1): + assert module.in_channels % block_size == 0, "Input channels must be a multiple of block sizes" + # regular convolutions + else: + k = module.kernel_size[0] * module.kernel_size[1] + assert k % block_size == 0, "Kernel size must be a multiple of block size" + + def _forward_pre_hook(mod, input): + # no noise for evaluation + if mod.training: + if not is_conv: + # gather weight and sizes + weight = mod.weight + in_features = weight.size(1) + out_features = weight.size(0) + + # split weight matrix into blocks and randomly drop selected blocks + mask = torch.zeros(in_features // block_size * out_features, device=weight.device) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_features) + + else: + # gather weight and sizes + weight = mod.weight + in_channels = mod.in_channels + out_channels = mod.out_channels + + # split weight matrix into blocks and randomly drop selected blocks + if mod.kernel_size == (1, 1): + mask = torch.zeros(int(in_channels // block_size * out_channels), device=weight.device) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels) + else: + mask = torch.zeros(weight.size(0), weight.size(1), device=weight.device) + mask.bernoulli_(p) + mask = mask.unsqueeze(2).unsqueeze(3).repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1]) + + # scale weights and apply mask + mask = mask.to(torch.bool) # x.bool() is not currently supported in TorchScript + s = 1 / (1 - p) + mod.weight.data = s * weight.masked_fill(mask, 0) + + module.register_forward_pre_hook(_forward_pre_hook) + return module diff --git a/fairseq/modules/quantization/__init__.py b/fairseq/modules/quantization/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/fairseq/modules/quantization/pq/__init__.py b/fairseq/modules/quantization/pq/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5b10b51b1b0ca21aaec96344f86a0ab9df0c22f8 --- /dev/null +++ b/fairseq/modules/quantization/pq/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .utils import SizeTracker, quantize_model_ # NOQA diff --git a/fairseq/modules/quantization/pq/em.py b/fairseq/modules/quantization/pq/em.py new file mode 100644 index 0000000000000000000000000000000000000000..420d8afda25da3fb4ad34ddc284d969bf4f09dae --- /dev/null +++ b/fairseq/modules/quantization/pq/em.py @@ -0,0 +1,211 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os +import random +import logging +from collections import Counter + +import torch + + +class EM: + """ + EM algorithm used to quantize the columns of W to minimize + + ||W - W_hat||^2 + + Args: + - W: weight matrix of size (in_features x out_features) + - n_iter: number of k-means iterations + - n_centroids: number of centroids (size of codebook) + - eps: for cluster reassignment when an empty cluster is found + - max_tentatives for cluster reassignment when an empty cluster is found + - verbose: print error after each iteration + + Remarks: + - If one cluster is empty, the most populated cluster is split into + two clusters + - All the relevant dimensions are specified in the code + """ + + def __init__( + self, W, n_centroids=256, n_iter=20, eps=1e-6, max_tentatives=30, verbose=True + ): + self.W = W + self.n_centroids = n_centroids + self.n_iter = n_iter + self.eps = eps + self.max_tentatives = max_tentatives + self.verbose = verbose + self.centroids = torch.Tensor() + self.assignments = torch.Tensor() + self.objective = [] + + def initialize_centroids(self): + """ + Initializes the centroids by sampling random columns from W. + """ + + in_features, out_features = self.W.size() + indices = torch.randint( + low=0, high=out_features, size=(self.n_centroids,) + ).long() + self.centroids = self.W[:, indices].t() # (n_centroids x in_features) + + def step(self, i): + """ + There are two standard steps for each iteration: expectation (E) and + minimization (M). The E-step (assignment) is performed with an exhaustive + search and the M-step (centroid computation) is performed with + the exact solution. + + Args: + - i: step number + + Remarks: + - The E-step heavily uses PyTorch broadcasting to speed up computations + and reduce the memory overhead + """ + + # assignments (E-step) + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + n_empty_clusters = self.resolve_empty_clusters() + + # centroids (M-step) + for k in range(self.n_centroids): + W_k = self.W[:, self.assignments == k] # (in_features x size_of_cluster_k) + self.centroids[k] = W_k.mean(dim=1) # (in_features) + + # book-keeping + obj = (self.centroids[self.assignments].t() - self.W).norm(p=2).item() + self.objective.append(obj) + if self.verbose: + logging.info( + f"Iteration: {i},\t" + f"objective: {obj:.6f},\t" + f"resolved empty clusters: {n_empty_clusters}" + ) + + def resolve_empty_clusters(self): + """ + If one cluster is empty, the most populated cluster is split into + two clusters by shifting the respective centroids. This is done + iteratively for a fixed number of tentatives. + """ + + # empty clusters + counts = Counter(map(lambda x: x.item(), self.assignments)) + empty_clusters = set(range(self.n_centroids)) - set(counts.keys()) + n_empty_clusters = len(empty_clusters) + + tentatives = 0 + while len(empty_clusters) > 0: + # given an empty cluster, find most populated cluster and split it into two + k = random.choice(list(empty_clusters)) + m = counts.most_common(1)[0][0] + e = torch.randn_like(self.centroids[m]) * self.eps + self.centroids[k] = self.centroids[m].clone() + self.centroids[k] += e + self.centroids[m] -= e + + # recompute assignments + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + + # check for empty clusters + counts = Counter(map(lambda x: x.item(), self.assignments)) + empty_clusters = set(range(self.n_centroids)) - set(counts.keys()) + + # increment tentatives + if tentatives == self.max_tentatives: + logging.info( + f"Could not resolve all empty clusters, {len(empty_clusters)} remaining" + ) + raise EmptyClusterResolveError + tentatives += 1 + + return n_empty_clusters + + def compute_distances(self): + """ + For every centroid m, computes + + ||M - m[None, :]||_2 + + Remarks: + - We rely on PyTorch's broadcasting to speed up computations + and reduce the memory overhead + - Without chunking, the sizes in the broadcasting are modified as: + (n_centroids x n_samples x out_features) -> (n_centroids x out_features) + - The broadcasting computation is automatically chunked so that + the tensors fit into the memory of the GPU + """ + + nb_centroids_chunks = 1 + + while True: + try: + return torch.cat( + [ + (self.W[None, :, :] - centroids_c[:, :, None]).norm(p=2, dim=1) + for centroids_c in self.centroids.chunk( + nb_centroids_chunks, dim=0 + ) + ], + dim=0, + ) + except RuntimeError: + nb_centroids_chunks *= 2 + + def assign(self): + """ + Assigns each column of W to its closest centroid, thus essentially + performing the E-step in train(). + + Remarks: + - The function must be called after train() or after loading + centroids using self.load(), otherwise it will return empty tensors + """ + + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + + def save(self, path, layer): + """ + Saves centroids and assignments. + + Args: + - path: folder used to save centroids and assignments + """ + + torch.save(self.centroids, os.path.join(path, "{}_centroids.pth".format(layer))) + torch.save( + self.assignments, os.path.join(path, "{}_assignments.pth".format(layer)) + ) + torch.save(self.objective, os.path.join(path, "{}_objective.pth".format(layer))) + + def load(self, path, layer): + """ + Loads centroids and assignments from a given path + + Args: + - path: folder use to load centroids and assignments + """ + + self.centroids = torch.load( + os.path.join(path, "{}_centroids.pth".format(layer)) + ) + self.assignments = torch.load( + os.path.join(path, "{}_assignments.pth".format(layer)) + ) + self.objective = torch.load( + os.path.join(path, "{}_objective.pth".format(layer)) + ) + + +class EmptyClusterResolveError(Exception): + pass diff --git a/fairseq/modules/quantization/pq/modules/__init__.py b/fairseq/modules/quantization/pq/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f52f6f37a6861a87becb7b269117cb9e24285b9e --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .qconv import PQConv2d # NOQA +from .qlinear import PQLinear # NOQA +from .qemb import PQEmbedding # NOQA diff --git a/fairseq/modules/quantization/pq/modules/qconv.py b/fairseq/modules/quantization/pq/modules/qconv.py new file mode 100644 index 0000000000000000000000000000000000000000..d15ec192e8cda6265a198e583a9bf7fb194dd129 --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/qconv.py @@ -0,0 +1,115 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch.nn.modules.utils import _pair + + +class PQConv2d(nn.Module): + """ + Quantized counterpart of nn.Conv2d module. Stores the centroid, the assignments + and the non-quantized biases. The full weight is re-instantiated at each forward + pass and autograd automatically computes the gradients with respect to the + centroids. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_channels x n_blocks + - bias: the non-quantized bias, must be either torch.Tensor or None + + Remarks: + - We refer the reader to the official documentation of the nn.Conv2d module + for the other arguments and the behavior of the module. + - Performance tests on GPU show that this implementation is 10% slower than + the non-quantized nn.Conv2d module for a standard training loop. + - During the backward, the gradients are averaged by cluster and not summed. + This explains the hook registered to the centroids. + """ + + def __init__( + self, + centroids, + assignments, + bias, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + padding_mode="zeros", + ): + super(PQConv2d, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _pair(kernel_size) + self.stride = _pair(stride) + self.padding = _pair(padding) + self.dilation = _pair(dilation) + self.groups = groups + self.padding_mode = padding_mode + # check compatibility + if in_channels // groups * np.prod(self.kernel_size) % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % out_channels != 0: + raise ValueError("Wrong PQ sizes") + if in_channels % groups != 0: + raise ValueError("in_channels must be divisible by groups") + if out_channels % groups != 0: + raise ValueError("out_channels must be divisible by groups") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + if bias is not None: + self.bias = nn.Parameter(bias) + else: + self.register_parameter("bias", None) + # register hook for averaging gradients per centroids instead of summing + self.centroids.register_hook(lambda x: x / self.counts[:, None]) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_channels, self.block_size) + .permute(1, 0, 2) + .reshape( + self.out_channels, self.in_channels // self.groups, *self.kernel_size + ) + ) + + def forward(self, x): + return F.conv2d( + x, + self.weight, + self.bias, + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + def extra_repr(self): + s = "{in_channels}, {out_channels}, kernel_size={kernel_size}, stride={stride}" + if self.padding != (0,) * len(self.padding): + s += ", padding={padding}" + if self.dilation != (1,) * len(self.dilation): + s += ", dilation={dilation}" + if self.groups != 1: + s += ", groups={groups}" + if self.bias is None: + s += ", bias=False" + if self.padding_mode != "zeros": + s += ", padding_mode={padding_mode}" + s += ", n_centroids={n_centroids}, block_size={block_size}" + return s.format(**self.__dict__) diff --git a/fairseq/modules/quantization/pq/modules/qemb.py b/fairseq/modules/quantization/pq/modules/qemb.py new file mode 100644 index 0000000000000000000000000000000000000000..98d856d04e5876dc7cc19067808f98b3106a2558 --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/qemb.py @@ -0,0 +1,87 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class PQEmbedding(nn.Module): + """ + Quantized counterpart of nn.Embedding module. Stores the centroids and + the assignments. The full weight is re-instantiated at each forward + pass. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_features x n_blocks + - bias: the non-quantized bias + + Remarks: + - We refer the reader to the official documentation of the nn.Embedding module + for the other arguments and the behavior of the module + - Performance tests on GPU show that this implementation is 10% slower than + the non-quantized nn.Embedding module for a standard training loop. + """ + + def __init__(self, centroids, assignments, num_embeddings, embedding_dim, + padding_idx=None, max_norm=None, norm_type=2., + scale_grad_by_freq=False, sparse=False, _weight=None): + super(PQEmbedding, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert padding_idx < self.num_embeddings, 'Padding_idx must be within num_embeddings' + elif padding_idx < 0: + assert padding_idx >= -self.num_embeddings, 'Padding_idx must be within num_embeddings' + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.max_norm = max_norm + self.norm_type = norm_type + self.scale_grad_by_freq = scale_grad_by_freq + self.sparse = sparse + # check compatibility + if self.embedding_dim % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % self.num_embeddings != 0: + raise ValueError("Wrong PQ sizes") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.num_embeddings, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + def forward(self, input): + return F.embedding( + input, self.weight, self.padding_idx, self.max_norm, + self.norm_type, self.scale_grad_by_freq, self.sparse) + + def extra_repr(self): + s = '{num_embeddings}, {embedding_dim}' + if self.padding_idx is not None: + s += ', padding_idx={padding_idx}' + if self.max_norm is not None: + s += ', max_norm={max_norm}' + if self.norm_type != 2: + s += ', norm_type={norm_type}' + if self.scale_grad_by_freq is not False: + s += ', scale_grad_by_freq={scale_grad_by_freq}' + if self.sparse is not False: + s += ', sparse=True' + s += ', n_centroids={n_centroids}, block_size={block_size}' + + return s.format(**self.__dict__) diff --git a/fairseq/modules/quantization/pq/modules/qlinear.py b/fairseq/modules/quantization/pq/modules/qlinear.py new file mode 100644 index 0000000000000000000000000000000000000000..9bdd25a8685bb7c7b32e1f02372aaeb26d8ba53a --- /dev/null +++ b/fairseq/modules/quantization/pq/modules/qlinear.py @@ -0,0 +1,71 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class PQLinear(nn.Module): + """ + Quantized counterpart of nn.Linear module. Stores the centroid, the assignments + and the non-quantized biases. The full weight is re-instantiated at each forward + pass. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_features x n_blocks + - bias: the non-quantized bias + + Remarks: + - We refer the reader to the official documentation of the nn.Linear module + for the other arguments and the behavior of the module + - Performance tests on GPU show that this implementation is 15% slower than + the non-quantized nn.Linear module for a standard training loop. + """ + + def __init__(self, centroids, assignments, bias, in_features, out_features): + super(PQLinear, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.in_features = in_features + self.out_features = out_features + # check compatibility + if self.in_features % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % self.out_features != 0: + raise ValueError("Wrong PQ sizes") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + if bias is not None: + self.bias = nn.Parameter(bias) + else: + self.register_parameter("bias", None) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_features, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + def forward(self, x): + return F.linear( + x, + self.weight, + self.bias, + ) + + def extra_repr(self): + return f"in_features={self.in_features},\ + out_features={self.out_features},\ + n_centroids={self.n_centroids},\ + block_size={self.block_size},\ + bias={self.bias is not None}" diff --git a/fairseq/modules/quantization/pq/pq.py b/fairseq/modules/quantization/pq/pq.py new file mode 100644 index 0000000000000000000000000000000000000000..eddc2eb34602403f10979f54cd23a45bc2f104d5 --- /dev/null +++ b/fairseq/modules/quantization/pq/pq.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .em import EM, EmptyClusterResolveError + + +class PQ(EM): + """ + Quantizes the layer weights W with the standard Product Quantization + technique. This learns a codebook of codewords or centroids of size + block_size from W. For further reference on using PQ to quantize + neural networks, see "And the Bit Goes Down: Revisiting the Quantization + of Neural Networks", Stock et al., ICLR 2020. + + PQ is performed in two steps: + (1) The matrix W (weights or fully-connected or convolutional layer) + is reshaped to (block_size, -1). + - If W is fully-connected (2D), its columns are split into + blocks of size block_size. + - If W is convolutional (4D), its filters are split along the + spatial dimension. + (2) We apply the standard EM/k-means algorithm to the resulting reshaped matrix. + + Args: + - W: weight matrix to quantize of size (in_features x out_features) + - block_size: size of the blocks (subvectors) + - n_centroids: number of centroids + - n_iter: number of k-means iterations + - eps: for cluster reassignment when an empty cluster is found + - max_tentatives for cluster reassignment when an empty cluster is found + - verbose: print information after each iteration + + Remarks: + - block_size be compatible with the shape of W + """ + + def __init__( + self, + W, + block_size, + n_centroids=256, + n_iter=20, + eps=1e-6, + max_tentatives=30, + verbose=True, + ): + self.block_size = block_size + W_reshaped = self._reshape(W) + super(PQ, self).__init__( + W_reshaped, + n_centroids=n_centroids, + n_iter=n_iter, + eps=eps, + max_tentatives=max_tentatives, + verbose=verbose, + ) + + def _reshape(self, W): + """ + Reshapes the matrix W as expained in step (1). + """ + + # fully connected: by convention the weight has size out_features x in_features + if len(W.size()) == 2: + self.out_features, self.in_features = W.size() + assert ( + self.in_features % self.block_size == 0 + ), "Linear: n_blocks must be a multiple of in_features" + return ( + W.reshape(self.out_features, -1, self.block_size) + .permute(2, 1, 0) + .flatten(1, 2) + ) + + # convolutional: we reshape along the spatial dimension + elif len(W.size()) == 4: + self.out_channels, self.in_channels, self.k_h, self.k_w = W.size() + assert ( + self.in_channels * self.k_h * self.k_w + ) % self.block_size == 0, ( + "Conv2d: n_blocks must be a multiple of in_channels * k_h * k_w" + ) + return ( + W.reshape(self.out_channels, -1, self.block_size) + .permute(2, 1, 0) + .flatten(1, 2) + ) + # not implemented + else: + raise NotImplementedError(W.size()) + + def encode(self): + """ + Performs self.n_iter EM steps. + """ + + self.initialize_centroids() + for i in range(self.n_iter): + try: + self.step(i) + except EmptyClusterResolveError: + break + + def decode(self): + """ + Returns the encoded full weight matrix. Must be called after + the encode function. + """ + + # fully connected case + if "k_h" not in self.__dict__: + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_features, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + # convolutional case + else: + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_channels, self.block_size) + .permute(1, 0, 2) + .reshape(self.out_channels, self.in_channels, self.k_h, self.k_w) + ) diff --git a/fairseq/modules/quantization/pq/utils.py b/fairseq/modules/quantization/pq/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..57aaa1b7a39cbdcf2684a464fd16ae60df29fae5 --- /dev/null +++ b/fairseq/modules/quantization/pq/utils.py @@ -0,0 +1,335 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import re +from operator import attrgetter, itemgetter + +import numpy as np +import torch.nn as nn +import torch.distributed as dist + +from .modules import PQConv2d, PQLinear, PQEmbedding +from .pq import PQ + + +def quantize_model_( + model, + size_tracker, + layers_to_quantize, + block_sizes_config, + n_centroids_config, + step=0, + n_iter=15, + eps=1e-6, + max_tentatives=100, + verbose=True, +): + """ + Quantize a model in-place by stages. All the targeted + layers are replaced by their quantized counterpart, + and the model is ready for the finetuning of the + centroids in a standard training loop (no modifications + required). Note that we do not quantize biases. + + Args: + - model: a nn.Module + - size_tracker: useful for tracking quatization statistics + - layers_to_quantize: a list containing regexps for + filtering the layers to quantize at each stage according + to their name (as in model.named_parameters()) + - block_sizes_config: dict like + { + 'Conv2d': ('kernel_size', {'(3, 3)': 9, '(1, 1)': 4}), + 'Linear': ('in_features', {'*': 8}) + } + For instance, all conv2d layers with kernel size 3x3 have + a block size of 9 and all Linear layers are quantized with + a block size of 8, irrespective of their size. + - n_centroids_config: dict like + { + 'Conv2d': ('kernel_size', {'*': 256}), + 'Linear': ('in_features', {'*': 256}) + } + For instance, all conv2d layers are quantized with 256 centroids + - step: the layers to quantize inplace corresponding + to layers_to_quantize[step] + """ + + quantized_layers = get_layers(model, layers_to_quantize[step]) + + for layer in quantized_layers: + + # book-keeping + is_master_process = (not dist.is_initialized()) or (dist.is_initialized() and dist.get_rank() == 0) + verbose = verbose and is_master_process + + # get block size and centroids + module = attrgetter(layer)(model) + block_size = get_param(module, layer, block_sizes_config) + n_centroids = get_param(module, layer, n_centroids_config) + if verbose: + logging.info(f"Quantizing layer {layer} with block size {block_size} and {n_centroids} centroids") + + # quantize layer + weight = module.weight.data.clone() + is_bias = 'bias' in [x[0] for x in module.named_parameters()] + bias = module.bias.data.clone() if is_bias else None + quantizer = PQ( + weight, + block_size, + n_centroids=n_centroids, + n_iter=n_iter, + eps=eps, + max_tentatives=max_tentatives, + verbose=verbose, + ) + + # quantization performed on all GPUs with same seed + quantizer.encode() + centroids = quantizer.centroids.contiguous() + assignments = quantizer.assignments.contiguous() + + # broadcast results to make sure weights are up-to-date + if dist.is_initialized(): + dist.broadcast(centroids, 0) + dist.broadcast(assignments, 0) + + # instantiate the quantized counterpart + if isinstance(module, nn.Linear): + out_features, in_features = map( + lambda k: module.__dict__[k], ["out_features", "in_features"] + ) + quantized_module = PQLinear( + centroids, assignments, bias, in_features, out_features + ) + elif isinstance(module, nn.Embedding): + num_embeddings, embedding_dim = map( + lambda k: module.__dict__[k], ["num_embeddings", "embedding_dim"] + ) + quantized_module = PQEmbedding( + centroids, assignments, num_embeddings, embedding_dim + ) + elif isinstance(module, nn.Conv2d): + out_channels, in_channels, kernel_size = map( + lambda k: module.__dict__[k], + ["out_channels", "in_channels", "kernel_size"], + ) + stride, padding, dilation, groups, padding_mode = map( + lambda k: module.__dict__[k], + ["stride", "padding", "dilation", "groups", "padding_mode"], + ) + + quantized_module = PQConv2d( + centroids, + assignments, + bias, + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + dilation=dilation, + groups=groups, + padding_mode=padding_mode, + ) + else: + raise ValueError(f"Module {module} not yet supported for quantization") + + # replace layer by its quantized counterpart + attrsetter(layer)(model, quantized_module) + + # update statistics + size_tracker.update(weight, block_size, n_centroids) + + # return name of quantized layers + return quantized_layers + + +def get_layers(model, filter_regexp): + """ + Filters out the layers according to a regexp. Note that + we omit biases. + + Args: + - model: a nn.Module + - filter_regexp: a regexp to filter the layers to keep + according to their name in model.named_parameters(). + For instance, the regexp: + + down_layers\\.[123456]\\.(conv[12]|identity\\.conv)) + + is keeping blocks down_layers from 1 to 6, and inside + each block is keeping conv1, conv2 and identity.conv. + + Remarks: + - We add (module\\.)? at the beginning of the regexp to + account for the possible use of nn.parallel.DataParallel + """ + + # get all parameter names + all_layers = map(itemgetter(0), model.named_parameters()) + + # remove biases + all_layers = filter(lambda x: "bias" not in x, all_layers) + + # remove .weight in all other names (or .weight_orig is spectral norm) + all_layers = map(lambda x: x.replace(".weight_orig", ""), all_layers) + all_layers = map(lambda x: x.replace(".weight", ""), all_layers) + + # return filtered layers + filter_regexp = "(module\\.)?" + "(" + filter_regexp + ")" + r = re.compile(filter_regexp) + + return list(filter(r.match, all_layers)) + + +def get_param(module, layer_name, param_config): + """ + Given a quantization configuration, get the right parameter + for the module to be quantized. + + Args: + - module: a nn.Module + - layer_name: the name of the layer + - param_config: a dict like + { + 'Conv2d': ('kernel_size', {'(3, 3)': 9, '(1, 1)': 4}), + 'Linear': ('in_features', {'*': 8}) + } + For instance, all conv2d layers with kernel size 3x3 have + a block size of 9 and all Linear layers are quantized with + a block size of 8, irrespective of their size. + + Remarks: + - if 'fuzzy_name' is passed as a parameter, layers whose layer_name + include 'fuzzy_name' will be assigned the given parameter. + In the following example, conv.expand layers will have a block + size of 9 while conv.reduce will have a block size of 4 and all + other layers will have a block size of 2. + { + 'Conv2d': ('fuzzy_name', {'expand': 9, 'reduce': 4, '*': 2}), + 'Linear': ('fuzzy_name', {'classifier': 8, 'projection': 4}) + } + + """ + + layer_type = module.__class__.__name__ + + if layer_type not in param_config: + raise KeyError(f"Layer type {layer_type} not in config for layer {module}") + + feature, params = param_config[module.__class__.__name__] + + if feature != "fuzzy_name": + feature_value = str(getattr(module, feature)) + if feature_value not in params: + if "*" in params: + feature_value = "*" + else: + raise KeyError( + f"{feature}={feature_value} not in config for layer {module}" + ) + else: + feature_values = [name for name in params if name in layer_name] + if len(feature_values) == 0: + if "*" in params: + feature_value = "*" + else: + raise KeyError( + f"name={layer_name} not in config for {module}" + ) + else: + feature_value = feature_values[0] + + return params[feature_value] + + +class SizeTracker(object): + """ + Class to keep track of the compressed network size with iPQ. + + Args: + - model: a nn.Module + + Remarks: + - The compressed size is the sum of three components + for each layer in the network: + (1) Storing the centroids given by iPQ in fp16 + (2) Storing the assignments of the blocks in int8 + (3) Storing all non-compressed elements such as biases + - This cost in only valid if we use 256 centroids (then + indexing can indeed by done with int8). + """ + + def __init__(self, model): + self.model = model + self.size_non_compressed_model = self.compute_size() + self.size_non_quantized = self.size_non_compressed_model + self.size_index = 0 + self.size_centroids = 0 + self.n_quantized_layers = 0 + + def compute_size(self): + """ + Computes the size of the model (in MB). + """ + + res = 0 + for _, p in self.model.named_parameters(): + res += p.numel() + return res * 4 / 1024 / 1024 + + def update(self, W, block_size, n_centroids): + """ + Updates the running statistics when quantizing a new layer. + """ + + # bits per weights + bits_per_weight = np.log2(n_centroids) / block_size + self.n_quantized_layers += 1 + + # size of indexing the subvectors of size block_size (in MB) + size_index_layer = bits_per_weight * W.numel() / 8 / 1024 / 1024 + self.size_index += size_index_layer + + # size of the centroids stored in float16 (in MB) + size_centroids_layer = n_centroids * block_size * 2 / 1024 / 1024 + self.size_centroids += size_centroids_layer + + # size of non-compressed layers, e.g. LayerNorms or biases (in MB) + size_uncompressed_layer = W.numel() * 4 / 1024 / 1024 + self.size_non_quantized -= size_uncompressed_layer + + def __repr__(self): + size_compressed = ( + self.size_index + self.size_centroids + self.size_non_quantized + ) + compression_ratio = self.size_non_compressed_model / size_compressed # NOQA + return ( + f"Non-compressed model size: {self.size_non_compressed_model:.2f} MB. " + f"After quantizing {self.n_quantized_layers} layers, size " + f"(indexing + centroids + other): {self.size_index:.2f} MB + " + f"{self.size_centroids:.2f} MB + {self.size_non_quantized:.2f} MB = " + f"{size_compressed:.2f} MB, compression ratio: {compression_ratio:.2f}x" + ) + + +def attrsetter(*items): + def resolve_attr(obj, attr): + attrs = attr.split(".") + head = attrs[:-1] + tail = attrs[-1] + + for name in head: + obj = getattr(obj, name) + return obj, tail + + def g(obj, val): + for attr in items: + resolved_obj, resolved_attr = resolve_attr(obj, attr) + setattr(resolved_obj, resolved_attr, val) + + return g diff --git a/fairseq/modules/quantization/quantization_options.py b/fairseq/modules/quantization/quantization_options.py new file mode 100644 index 0000000000000000000000000000000000000000..b46d682c0edaeaaf2a230e51d50da2a32d4bda98 --- /dev/null +++ b/fairseq/modules/quantization/quantization_options.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def parse_config_yaml(yaml_data): + # Initialize to default options. + quantization_options = { + "n_centroids": { + "Linear": ["in_features", {"*": 256}], + "Embedding": ["embedding_dim", {"*": 256}], + }, + "block_sizes": { + "Linear": ["fuzzy_name", {"fc": 8, "attn": 4, "emb": 4}], + "Embedding": ["fuzzy_name", {"emb": 8}], + }, + "layers_to_quantize": [ + "decoder\\.layers\\.\\d+\\.fc[12]", + "decoder\\.embed_tokens\\.embeddings\\.[012]\\.[01]", + "decoder\\.layers\\.\\d+\\.self_attn\\.(k_proj|v_proj|q_proj|out_proj)", + ], + } + + if "n_centroids" in yaml_data: + quantization_options["n_centroids"] = { + layer: convert_yaml_to_tuple(layer_data) + for layer, layer_data in yaml_data["n_centroids"].items() + } + if "block_sizes" in yaml_data: + quantization_options["block_sizes"] = { + layer: convert_yaml_to_tuple(layer_data) + for layer, layer_data in yaml_data["block_sizes"].items() + } + if "layers_to_quantize" in yaml_data: + quantization_options["layers_to_quantize"] = yaml_data["layers_to_quantize"] + + return quantization_options + + +def convert_yaml_to_tuple(yaml_dictionary): + """Converts a yaml dictionary with two keys: `key` and `value` into a two + argument tuple of those values.""" + return (yaml_dictionary["key"], yaml_dictionary["value"]) diff --git a/fairseq/modules/quantization/scalar/__init__.py b/fairseq/modules/quantization/scalar/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..143834f3d036780eb6844c82f0c6f2d10cfe2f61 --- /dev/null +++ b/fairseq/modules/quantization/scalar/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .utils import quantize_model_ # NOQA diff --git a/fairseq/modules/quantization/scalar/modules/__init__.py b/fairseq/modules/quantization/scalar/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ead4669611f95d3616ccf1cf2adba631281680e2 --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .qconv import IntConv2d # NOQA +from .qlinear import IntLinear # NOQA +from .qemb import IntEmbedding # NOQA +from .qact import ActivationQuantizer # NOQA diff --git a/fairseq/modules/quantization/scalar/modules/qact.py b/fairseq/modules/quantization/scalar/modules/qact.py new file mode 100644 index 0000000000000000000000000000000000000000..a9f79011c15a463d63eeca880904d68b751ef167 --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qact.py @@ -0,0 +1,80 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from ..ops import emulate_int + + +class ActivationQuantizer: + """ + Fake scalar quantization of the activations using a forward hook. + + Args: + - module. a nn.Module for which we quantize the *post-activations* + - p: proportion of activations to quantize, set by default to 1 + - update_step: to recompute quantization parameters + - bits: number of bits for quantization + - method: choose among {"tensor", "histogram", "channel"} + - clamp_threshold: to prevent gradients overflow + + Remarks: + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - For the list of quantization methods and number of bits, see ops.py + - To remove the hook from the module, simply call self.handle.remove() + - At test time, the activations are fully quantized + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - The activations are hard-clamped in [-clamp_threshold, clamp_threshold] + to prevent overflow during the backward pass + """ + def __init__(self, module, p=1, update_step=1000, bits=8, + method="histogram", clamp_threshold=5): + self.module = module + self.p = p + self.update_step = update_step + self.counter = 0 + self.bits = bits + self.method = method + self.clamp_threshold = clamp_threshold + self.handle = None + self.register_hook() + + def register_hook(self): + # forward hook + def quantize_hook(module, x, y): + + # update parameters every 1000 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.module.training else 1 + + # quantize activations + y_q, self.scale, self.zero_point = emulate_int( + y.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(y) + mask.bernoulli_(1 - p) + noise = (y_q - y).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = - self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + return torch.clamp(y, clamp_low.item(), clamp_high.item()) + noise.detach() + + # register hook + self.handle = self.module.register_forward_hook(quantize_hook) diff --git a/fairseq/modules/quantization/scalar/modules/qconv.py b/fairseq/modules/quantization/scalar/modules/qconv.py new file mode 100644 index 0000000000000000000000000000000000000000..d718c9b90d7610223a6ad8211edda6f0c31cfd41 --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qconv.py @@ -0,0 +1,146 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from torch.nn.modules.conv import _ConvNd +from torch.nn.modules.utils import _pair + +from ..ops import emulate_int + + +class IntConv2d(_ConvNd): + """ + Quantized counterpart of the nn.Conv2d module that applies QuantNoise during training. + + Args: + - standard nn.Conv2d parameters + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-thgourh estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + bias=True, + padding_mode="zeros", + p=0, + bits=8, + method="histogram", + update_step=1000, + ): + kernel_size = _pair(kernel_size) + stride = _pair(stride) + padding = _pair(padding) + dilation = _pair(dilation) + super(IntConv2d, self).__init__( + in_channels, + out_channels, + kernel_size, + stride, + padding, + dilation, + False, + _pair(0), + groups, + bias, + padding_mode, + ) + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def _conv_forward(self, input, weight): + if self.padding_mode != "zeros": + return F.conv2d( + F.pad(input, self._padding_repeated_twice, mode=self.padding_mode), + weight, + self.bias, + self.stride, + _pair(0), + self.dilation, + self.groups, + ) + return F.conv2d( + input, + weight, + self.bias, + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 100 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = - self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + weight = torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + noise.detach() + + # return output + output = self._conv_forward(input, weight) + return output + + def extra_repr(self): + return ( + "in_channels={}, out_channels={}, kernel_size={}, stride={}, " + "padding={}, dilation={}, groups={}, bias={}, quant_noise={}, " + "bits={}, method={}".format( + self.in_channels, + self.out_channels, + self.kernel_size, + self.stride, + self.padding, + self.dilation, + self.groups, + self.bias is not None, + self.p, + self.bits, + self.method, + ) + ) diff --git a/fairseq/modules/quantization/scalar/modules/qemb.py b/fairseq/modules/quantization/scalar/modules/qemb.py new file mode 100644 index 0000000000000000000000000000000000000000..835b2782a756e83d3990d321fe0e43eadfd071d2 --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qemb.py @@ -0,0 +1,132 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ..ops import emulate_int + + +class IntEmbedding(nn.Module): + """ + Quantized counterpart of the nn.Embedding module that applies QuantNoise during training. + + Args: + - num_embeddings: number of tokens + - embedding_dim: embedding dimension + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + num_embeddings, + embedding_dim, + padding_idx=None, + max_norm=None, + norm_type=2., + scale_grad_by_freq=False, + sparse=False, + _weight=None, + p=0, + update_step=1000, + bits=8, + method="histogram", + ): + super(IntEmbedding, self).__init__() + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert padding_idx < self.num_embeddings, 'Padding_idx must be within num_embeddings' + elif padding_idx < 0: + assert padding_idx >= -self.num_embeddings, 'Padding_idx must be within num_embeddings' + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.max_norm = max_norm + self.norm_type = norm_type + self.scale_grad_by_freq = scale_grad_by_freq + if _weight is None: + self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim)) + self.reset_parameters() + else: + assert list(_weight.shape) == [num_embeddings, embedding_dim], \ + 'Shape of weight does not match num_embeddings and embedding_dim' + self.weight = nn.Parameter(_weight) + self.sparse = sparse + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def reset_parameters(self): + nn.init.normal_(self.weight) + if self.padding_idx is not None: + with torch.no_grad(): + self.weight[self.padding_idx].fill_(0) + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 1000 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = - self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + weight = torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + noise.detach() + + # return output + output = F.embedding( + input, weight, self.padding_idx, self.max_norm, + self.norm_type, self.scale_grad_by_freq, self.sparse) + return output + + def extra_repr(self): + s = '{num_embeddings}, {embedding_dim}' + if self.padding_idx is not None: + s += ', padding_idx={padding_idx}' + if self.max_norm is not None: + s += ', max_norm={max_norm}' + if self.norm_type != 2: + s += ', norm_type={norm_type}' + if self.scale_grad_by_freq is not False: + s += ', scale_grad_by_freq={scale_grad_by_freq}' + if self.sparse is not False: + s += ', sparse=True' + s += 'quant_noise={p}, bits={bits}, method={method}' + return s.format(**self.__dict__) diff --git a/fairseq/modules/quantization/scalar/modules/qlinear.py b/fairseq/modules/quantization/scalar/modules/qlinear.py new file mode 100644 index 0000000000000000000000000000000000000000..2d4b27dc6cfbeab0115272b80a0629fa6784e258 --- /dev/null +++ b/fairseq/modules/quantization/scalar/modules/qlinear.py @@ -0,0 +1,110 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ..ops import emulate_int + + +class IntLinear(nn.Module): + """ + Quantized counterpart of the nn.Linear module that applies QuantNoise during training. + + Args: + - in_features: input features + - out_features: output features + - bias: bias or not + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick. + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + in_features, + out_features, + bias=True, + p=0, + update_step=3000, + bits=8, + method="histogram", + ): + super(IntLinear, self).__init__() + self.in_features = int(in_features) + self.out_features = int(out_features) + self.weight = torch.nn.Parameter(torch.Tensor(out_features, in_features)) + self.chosen_bias = bias + if self.chosen_bias: + self.bias = torch.nn.Parameter(torch.Tensor(out_features)) + else: + self.register_parameter("bias", None) + self.reset_parameters() + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.chosen_bias: + nn.init.constant_(self.bias, 0.0) + return + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 100 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = - self.scale * self.zero_point + clamp_high = self.scale * (2 ** self.bits - 1 - self.zero_point) + weight = torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + noise.detach() + + # return output + output = F.linear(input, weight, self.bias) + return output + + def extra_repr(self): + return "in_features={}, out_features={}, bias={}, quant_noise={}, bits={}, method={}".format( + self.in_features, + self.out_features, + self.bias is not None, + self.p, + self.bits, + self.method, + ) diff --git a/fairseq/modules/quantization/scalar/ops.py b/fairseq/modules/quantization/scalar/ops.py new file mode 100644 index 0000000000000000000000000000000000000000..90bc737cc840821504d1a894070482ac952dced4 --- /dev/null +++ b/fairseq/modules/quantization/scalar/ops.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +def emulate_int(w, bits, method, scale=None, zero_point=None): + q = globals()[f"emulate_int{bits}_{method}"] + return q(w, scale=scale, zero_point=zero_point) + + +def quantize(w, scale, zero_point): + return (torch.clamp(torch.round(w / scale + zero_point), 0, 255) - zero_point) * scale + + +def emulate_int8_histogram(w, scale=None, zero_point=None): + if scale is None: + obs = torch.quantization.observer.HistogramObserver() + _ = obs(w.float()) + scale, zero_point = obs.calculate_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point), scale, zero_point + + +def emulate_int8_channel(w, scale=None, zero_point=None): + if scale is None: + obs = torch.quantization.observer.PerChannelMinMaxObserver( + ch_axis=-1, qscheme=torch.per_channel_symmetric + ) + _ = obs(w) + scale, zero_point, ch_axis = obs.get_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point), scale, zero_point + + +def emulate_int8_tensor(w, scale=None, zero_point=None): + if scale is None: + obs = torch.quantization.observer.MinMaxObserver() + _ = obs(w) + scale, zero_point = obs.calculate_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point), scale, zero_point diff --git a/fairseq/modules/quantization/scalar/utils.py b/fairseq/modules/quantization/scalar/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4071f7b80aa101ad60702e3a187c85880c37a62b --- /dev/null +++ b/fairseq/modules/quantization/scalar/utils.py @@ -0,0 +1,67 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from operator import attrgetter + +import torch.nn as nn +import torch.distributed as dist + +from ..pq.utils import get_layers, attrsetter +from .modules import IntConv2d, IntLinear, IntEmbedding, ActivationQuantizer + + +MAPPING = {nn.Linear: IntLinear, nn.Embedding: IntEmbedding, nn.Conv2d: IntConv2d} + + +def quantize_model_(model, p=0.2, bits=8, update_step=3000): + """ + Replaces all modules with their scalar quantized counterpart and + registers hooks to quantize the post-ativations of those modules. + + Args: + - model: a nn.Module + - p: amount of noise (0 for no noise, 1 to quantize all the weights/activations) + - bits: number of bits + - update_step: update quantization parameters every update_step steps + """ + + # quantize all layers + quantized_layers = get_layers(model, "(.*?)") + + for layer in quantized_layers: + + # book-keeping + is_master_process = (not dist.is_initialized()) or (dist.is_initialized() and dist.get_rank() == 0) + + # recover module + module = attrgetter(layer)(model) + if is_master_process: + logging.info(f"Quantizing layer {layer} with bits={bits} and QuantNoise={p}") + + # quantization params + q_params = {"p": p, "update_step": update_step, "bits": bits, "method": "histogram", "counter": 0} + + # instantiate the quantized counterpart + if isinstance(module, tuple(MAPPING.keys())): + QuantizedModule = MAPPING[module.__class__] + quantized_module = QuantizedModule.__new__(QuantizedModule) + params = module.__dict__ + params.update(q_params) + quantized_module.__dict__.update(params) + + else: + if is_master_process: + logging.info(f"Module {module} not yet supported for quantization") + continue + + # activation quantization + a_q = ActivationQuantizer(quantized_module, p=0, bits=bits, method="histogram") + + # replace layer by its quantized counterpart + attrsetter(layer)(model, quantized_module) + + # return name of quantized layers + return quantized_layers diff --git a/fairseq/modules/same_pad.py b/fairseq/modules/same_pad.py new file mode 100644 index 0000000000000000000000000000000000000000..b46f94d6357888bde46035d8fcd57ceff5d24a88 --- /dev/null +++ b/fairseq/modules/same_pad.py @@ -0,0 +1,18 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from torch import nn + + +class SamePad(nn.Module): + def __init__(self, kernel_size): + super().__init__() + self.remove = kernel_size % 2 == 0 + + def forward(self, x): + if self.remove: + x = x[:, :, :-1] + return x diff --git a/fairseq/modules/scalar_bias.py b/fairseq/modules/scalar_bias.py new file mode 100644 index 0000000000000000000000000000000000000000..c96247c75914fabb8a2b7ff731bb82b588f72690 --- /dev/null +++ b/fairseq/modules/scalar_bias.py @@ -0,0 +1,31 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +# + +import torch + + +class ScalarBias(torch.autograd.Function): + """ + Adds a vector of scalars, used in self-attention mechanism to allow + the model to optionally attend to this vector instead of the past + """ + + @staticmethod + def forward(ctx, input, dim, bias_init): + size = list(input.size()) + size[dim] += 1 + output = input.new(*size).fill_(bias_init) + output.narrow(dim, 1, size[dim] - 1).copy_(input) + ctx.dim = dim + return output + + @staticmethod + def backward(ctx, grad): + return grad.narrow(ctx.dim, 1, grad.size(ctx.dim) - 1), None, None + + +def scalar_bias(input, dim, bias_init=0): + return ScalarBias.apply(input, dim, bias_init) diff --git a/fairseq/modules/sinusoidal_positional_embedding.py b/fairseq/modules/sinusoidal_positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..857830faf7cb64950021947e2c5babcb906c48d3 --- /dev/null +++ b/fairseq/modules/sinusoidal_positional_embedding.py @@ -0,0 +1,105 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Any, Optional + +import torch +import torch.onnx.operators +from fairseq import utils +from torch import Tensor, nn + + +class SinusoidalPositionalEmbedding(nn.Module): + """This module produces sinusoidal positional embeddings of any length. + + Padding symbols are ignored. + """ + + def __init__(self, embedding_dim, padding_idx, init_size=1024): + super().__init__() + self.embedding_dim = embedding_dim + self.padding_idx = padding_idx + self.weights = SinusoidalPositionalEmbedding.get_embedding( + init_size, embedding_dim, padding_idx + ) + self.onnx_trace = False + self.register_buffer("_float_tensor", torch.FloatTensor(1)) + self.max_positions = int(1e5) + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + @staticmethod + def get_embedding( + num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None + ): + """Build sinusoidal embeddings. + + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb) + emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze( + 1 + ) * emb.unsqueeze(0) + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view( + num_embeddings, -1 + ) + if embedding_dim % 2 == 1: + # zero pad + emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1) + if padding_idx is not None: + emb[padding_idx, :] = 0 + return emb + + def forward( + self, + input, + incremental_state: Optional[Any] = None, + timestep: Optional[Tensor] = None, + positions: Optional[Any] = None, + ): + """Input is expected to be of size [bsz x seqlen].""" + bspair = torch.onnx.operators.shape_as_tensor(input) + bsz, seq_len = bspair[0], bspair[1] + max_pos = self.padding_idx + 1 + seq_len + if self.weights is None or max_pos > self.weights.size(0): + # recompute/expand embeddings if needed + self.weights = SinusoidalPositionalEmbedding.get_embedding( + max_pos, self.embedding_dim, self.padding_idx + ) + self.weights = self.weights.to(self._float_tensor) + + if incremental_state is not None: + # positions is the same for every token when decoding a single step + pos = timestep.view(-1)[0] + 1 if timestep is not None else seq_len + if self.onnx_trace: + return ( + self.weights.index_select(index=self.padding_idx + pos, dim=0) + .unsqueeze(1) + .repeat(bsz, 1, 1) + ) + return self.weights[self.padding_idx + pos, :].expand(bsz, 1, -1) + + positions = utils.make_positions( + input, self.padding_idx, onnx_trace=self.onnx_trace + ) + if self.onnx_trace: + flat_embeddings = self.weights.detach().index_select(0, positions.view(-1)) + embedding_shape = torch.cat( + (bsz.view(1), seq_len.view(1), torch.tensor([-1], dtype=torch.long)) + ) + embeddings = torch.onnx.operators.reshape_from_tensor_shape( + flat_embeddings, embedding_shape + ) + return embeddings + return ( + self.weights.index_select(0, positions.view(-1)) + .view(bsz, seq_len, -1) + .detach() + ) diff --git a/fairseq/modules/sparse_multihead_attention.py b/fairseq/modules/sparse_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..61430195c253fc0346814f7ba92e2a588553e99d --- /dev/null +++ b/fairseq/modules/sparse_multihead_attention.py @@ -0,0 +1,104 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import torch +from .multihead_attention import MultiheadAttention + + +class SparseMultiheadAttention(MultiheadAttention): + """ Sparse Multi-Headed Attention. + + "Generating Long Sequences with Sparse Transformers". Implements + fixed factorized self attention, where l=stride and c=expressivity. + A(1) includes all words in the stride window and A(2) takes a summary of c + words from the end of each stride window. + If is_bidirectional=False, we do not include any words past the current word, + as in the paper. + """ + + def __init__(self, embed_dim, num_heads, kdim=None, vdim=None, dropout=0., bias=True, + add_bias_kv=False, add_zero_attn=False, self_attention=False, + encoder_decoder_attention=False, stride=32, expressivity=8, is_bidirectional=True): + + super().__init__( + embed_dim, num_heads, kdim, vdim, dropout, bias, add_bias_kv, + add_zero_attn, self_attention, encoder_decoder_attention + ) + + self.is_bidirectional = is_bidirectional + self.stride = stride + self.expressivity = expressivity + assert(self.stride > 0 and self.stride >= self.expressivity) + + # Used for Ai(2) calculations - beginning of [l-c, l] range + def compute_checkpoint(self, word_index): + if word_index % self.stride == 0 and word_index != 0: + checkpoint_index = word_index - self.expressivity + else: + checkpoint_index = ( + math.floor(word_index / self.stride) * self.stride + + self.stride - self.expressivity + ) + return checkpoint_index + + # Computes Ai(2) + def compute_subset_summaries(self, absolute_max): + checkpoint_index = self.compute_checkpoint(0) + subset_two = set() + while checkpoint_index <= absolute_max-1: + summary = set(range(checkpoint_index, min( + checkpoint_index+self.expressivity+1, absolute_max) + )) + subset_two = subset_two.union(summary) + checkpoint_index = self.compute_checkpoint(checkpoint_index+self.stride) + return subset_two + + # Sparse Transformer Fixed Attention Pattern: https://arxiv.org/pdf/1904.10509.pdf + def compute_fixed_attention_subset(self, word_index, tgt_len): + # +1s account for range function; [min, max) -> [min, max] + if not self.is_bidirectional: + absolute_max = word_index + 1 + else: + absolute_max = tgt_len + + # Subset 1 - whole window + rounded_index = math.floor((word_index + self.stride) / self.stride) * self.stride + if word_index % self.stride == 0 and word_index != 0: + subset_one = set(range(word_index-self.stride, min(absolute_max, word_index+1))) + else: + subset_one = set(range(max(0, rounded_index - self.stride), min( + absolute_max, rounded_index+1)) + ) + + # Subset 2 - summary per window + # If bidirectional, subset 2 is the same for every index + subset_two = set() + if not self.is_bidirectional: + subset_two = self.compute_subset_summaries(absolute_max) + + return subset_one.union(subset_two) + + # Compute sparse mask - if bidirectional, can pre-compute and store + def buffered_sparse_mask(self, tensor, tgt_len, src_len): + assert(tgt_len > self.stride) + sparse_mask = torch.empty((tgt_len, src_len)).float().fill_(float('-inf')) + + # If bidirectional, subset 2 is the same for every index + subset_summaries = set() + if self.is_bidirectional: + subset_summaries = self.compute_subset_summaries(tgt_len) + + for i in range(tgt_len): + fixed_attention_subset = self.compute_fixed_attention_subset(i, tgt_len) + fixed_attention_subset = fixed_attention_subset.union(subset_summaries) + included_word_indices = torch.LongTensor(list(fixed_attention_subset)) + sparse_mask[i].index_fill_(0, included_word_indices, 0) + return sparse_mask.type_as(tensor) + + def apply_sparse_mask(self, attn_weights, tgt_len, src_len, bsz): + sparse_mask = self.buffered_sparse_mask(attn_weights, tgt_len, src_len) + sparse_mask = sparse_mask.unsqueeze(0).expand(bsz * self.num_heads, tgt_len, src_len) + attn_weights += sparse_mask diff --git a/fairseq/modules/sparse_transformer_sentence_encoder.py b/fairseq/modules/sparse_transformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..3d50d5a88289d8d24530af300d63cbae829f110f --- /dev/null +++ b/fairseq/modules/sparse_transformer_sentence_encoder.py @@ -0,0 +1,79 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +from fairseq.modules import TransformerSentenceEncoder +from fairseq.modules.sparse_transformer_sentence_encoder_layer import SparseTransformerSentenceEncoderLayer + + +class SparseTransformerSentenceEncoder(TransformerSentenceEncoder): + """ + Sparse implementation of the TransformerSentenceEncoder + - see SparseMultiheadAttention + """ + + def __init__( + self, + padding_idx: int, + vocab_size: int, + num_encoder_layers: int = 6, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + max_seq_len: int = 256, + num_segments: int = 2, + use_position_embeddings: bool = True, + offset_positions_by_padding: bool = True, + encoder_normalize_before: bool = False, + apply_bert_init: bool = False, + activation_fn: str = "relu", + learned_pos_embedding: bool = True, + embed_scale: float = None, + freeze_embeddings: bool = False, + n_trans_layers_to_freeze: int = 0, + export: bool = False, + is_bidirectional: bool = True, + stride: int = 32, + expressivity: int = 8, + ) -> None: + + super().__init__( + padding_idx, vocab_size, num_encoder_layers, embedding_dim, + ffn_embedding_dim, num_attention_heads, dropout, attention_dropout, + activation_dropout, max_seq_len, num_segments, use_position_embeddings, + offset_positions_by_padding, encoder_normalize_before, apply_bert_init, + activation_fn, learned_pos_embedding, embed_scale, freeze_embeddings, + n_trans_layers_to_freeze, export + ) + + self.layers = nn.ModuleList( + [ + SparseTransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + is_bidirectional=is_bidirectional, + stride=stride, + expressivity=expressivity, + ) + for _ in range(num_encoder_layers) + ] + ) + + def freeze_module_params(m): + if m is not None: + for p in m.parameters(): + p.requires_grad = False + + for layer in range(n_trans_layers_to_freeze): + freeze_module_params(self.layers[layer]) diff --git a/fairseq/modules/sparse_transformer_sentence_encoder_layer.py b/fairseq/modules/sparse_transformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..21c2fe4d5ae0b4279b13ff355365e698300c7c90 --- /dev/null +++ b/fairseq/modules/sparse_transformer_sentence_encoder_layer.py @@ -0,0 +1,45 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.modules import TransformerSentenceEncoderLayer +from fairseq.modules.sparse_multihead_attention import SparseMultiheadAttention + + +class SparseTransformerSentenceEncoderLayer(TransformerSentenceEncoderLayer): + """ + Implements a Sprase Transformer Encoder Layer (see SparseMultiheadAttention) + """ + + def __init__( + self, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = 'relu', + export: bool = False, + is_bidirectional: bool = True, + stride: int = 32, + expressivity: int = 8, + ) -> None: + + super().__init__( + embedding_dim, ffn_embedding_dim, num_attention_heads, dropout, + attention_dropout, activation_dropout, activation_fn, export + ) + + self.self_attn = SparseMultiheadAttention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + add_bias_kv=False, + add_zero_attn=False, + self_attention=True, + is_bidirectional=is_bidirectional, + stride=stride, + expressivity=expressivity, + ) diff --git a/fairseq/modules/transformer_layer.py b/fairseq/modules/transformer_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..037d8e88aeb0e82e224fe3290ee337bb89e48e48 --- /dev/null +++ b/fairseq/modules/transformer_layer.py @@ -0,0 +1,391 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.modules import LayerNorm, MultiheadAttention +from fairseq.modules.quant_noise import quant_noise +from fairseq.modules.fairseq_dropout import FairseqDropout +from torch import Tensor + + +class TransformerEncoderLayer(nn.Module): + """Encoder layer block. + + In the original paper each operation (multi-head attention or FFN) is + postprocessed with: `dropout -> add residual -> layernorm`. In the + tensor2tensor code they suggest that learning is more robust when + preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *args.encoder_normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + + def __init__(self, args): + super().__init__() + self.embed_dim = args.encoder_embed_dim + self.quant_noise = getattr(args, "quant_noise_pq", 0) + self.quant_noise_block_size = getattr(args, "quant_noise_pq_block_size", 8) + self.self_attn = self.build_self_attention(self.embed_dim, args) + self.self_attn_layer_norm = LayerNorm(self.embed_dim) + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + self.activation_fn = utils.get_activation_fn( + activation=getattr(args, "activation_fn", "relu") + ) + activation_dropout_p = getattr(args, "activation_dropout", 0) + if activation_dropout_p == 0: + # for backwards compatibility with models that use args.relu_dropout + activation_dropout_p = getattr(args, "relu_dropout", 0) + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + self.normalize_before = args.encoder_normalize_before + self.fc1 = self.build_fc1( + self.embed_dim, args.encoder_ffn_embed_dim, self.quant_noise, self.quant_noise_block_size + ) + self.fc2 = self.build_fc2( + args.encoder_ffn_embed_dim, self.embed_dim, self.quant_noise, self.quant_noise_block_size + ) + + self.final_layer_norm = LayerNorm(self.embed_dim) + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size) + + def build_self_attention(self, embed_dim, args): + return MultiheadAttention( + embed_dim, + args.encoder_attention_heads, + dropout=args.attention_dropout, + self_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def upgrade_state_dict_named(self, state_dict, name): + """ + Rename layer norm states from `...layer_norms.0.weight` to + `...self_attn_layer_norm.weight` and `...layer_norms.1.weight` to + `...final_layer_norm.weight` + """ + layer_norm_map = {"0": "self_attn_layer_norm", "1": "final_layer_norm"} + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layer_norms.{}.{}".format(name, old, m) + if k in state_dict: + state_dict["{}.{}.{}".format(name, new, m)] = state_dict[k] + del state_dict[k] + + def forward(self, x, encoder_padding_mask, attn_mask: Optional[Tensor] = None): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, seq_len)` where padding elements are indicated by ``1``. + attn_mask (ByteTensor): binary tensor of shape `(tgt_len, src_len)`, + where `tgt_len` is the length of output and `src_len` is the + length of input, though here both are equal to `seq_len`. + `attn_mask[tgt_i, src_j] = 1` means that when calculating the + embedding for `tgt_i`, we exclude (mask out) `src_j`. This is + useful for strided self-attention. + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + # anything in original attn_mask = 1, becomes -1e8 + # anything in original attn_mask = 0, becomes 0 + # Note that we cannot use -inf here, because at some edge cases, + # the attention weight (before softmax) for some padded element in query + # will become -inf, which results in NaN in model parameters + if attn_mask is not None: + attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8) + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + x, _ = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=encoder_padding_mask, + attn_mask=attn_mask, + ) + x = self.dropout_module(x) + x = residual + x + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + if not self.normalize_before: + x = self.final_layer_norm(x) + return x + + +class TransformerDecoderLayer(nn.Module): + """Decoder layer block. + + In the original paper each operation (multi-head attention, encoder + attention or FFN) is postprocessed with: `dropout -> add residual -> + layernorm`. In the tensor2tensor code they suggest that learning is more + robust when preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *args.decoder_normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__() + self.embed_dim = args.decoder_embed_dim + self.dropout_module = FairseqDropout(args.dropout, module_name=self.__class__.__name__) + self.quant_noise = getattr(args, "quant_noise_pq", 0) + self.quant_noise_block_size = getattr(args, "quant_noise_pq_block_size", 8) + + self.cross_self_attention = getattr(args, "cross_self_attention", False) + + self.self_attn = self.build_self_attention( + self.embed_dim, + args, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + self.activation_fn = utils.get_activation_fn( + activation=getattr(args, "activation_fn", "relu") + ) + activation_dropout_p = getattr(args, "activation_dropout", 0) + if activation_dropout_p == 0: + # for backwards compatibility with models that use args.relu_dropout + activation_dropout_p = getattr(args, "relu_dropout", 0) + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__) + self.normalize_before = args.decoder_normalize_before + + # use layerNorm rather than FusedLayerNorm for exporting. + # char_inputs can be used to determint this. + # TODO remove this once we update apex with the fix + export = getattr(args, "char_inputs", False) + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = self.build_encoder_attention(self.embed_dim, args) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + self.fc1 = self.build_fc1( + self.embed_dim, args.decoder_ffn_embed_dim, self.quant_noise, self.quant_noise_block_size + ) + self.fc2 = self.build_fc2( + args.decoder_ffn_embed_dim, self.embed_dim, self.quant_noise, self.quant_noise_block_size + ) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=export) + self.need_attn = True + + self.onnx_trace = False + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_self_attention(self, embed_dim, args, add_bias_kv=False, add_zero_attn=False): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + dropout=args.attention_dropout, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + self_attention=not getattr(args, "cross_self_attention", False), + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def build_encoder_attention(self, embed_dim, args): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def forward( + self, + x, + encoder_out: Optional[torch.Tensor] = None, + encoder_padding_mask: Optional[torch.Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + prev_self_attn_state: Optional[List[torch.Tensor]] = None, + prev_attn_state: Optional[List[torch.Tensor]] = None, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + need_attn: bool = False, + need_head_weights: bool = False, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor, optional): binary + ByteTensor of shape `(batch, src_len)` where padding + elements are indicated by ``1``. + need_attn (bool, optional): return attention weights + need_head_weights (bool, optional): return attention weights + for each head (default: return average over heads). + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + if need_head_weights: + need_attn = True + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[2] + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state, saved_state) + _self_attn_input_buffer = self.self_attn._get_input_buffer(incremental_state) + if self.cross_self_attention and not ( + incremental_state is not None + and _self_attn_input_buffer is not None + and "prev_key" in _self_attn_input_buffer + ): + if self_attn_mask is not None: + assert encoder_out is not None + self_attn_mask = torch.cat( + (x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), dim=1 + ) + if self_attn_padding_mask is not None: + if encoder_padding_mask is None: + assert encoder_out is not None + encoder_padding_mask = self_attn_padding_mask.new_zeros( + encoder_out.size(1), encoder_out.size(0) + ) + self_attn_padding_mask = torch.cat( + (encoder_padding_mask, self_attn_padding_mask), dim=1 + ) + assert encoder_out is not None + y = torch.cat((encoder_out, x), dim=0) + else: + y = x + + x, attn = self.self_attn( + query=x, + key=y, + value=y, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout_module(x) + x = residual + x + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + if self.encoder_attn is not None: + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = residual + x + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + if not self.normalize_before: + x = self.final_layer_norm(x) + if self.onnx_trace and incremental_state is not None: + saved_state = self.self_attn._get_input_buffer(incremental_state) + assert saved_state is not None + if self_attn_padding_mask is not None: + self_attn_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + saved_state["prev_key_padding_mask"], + ] + else: + self_attn_state = [saved_state["prev_key"], saved_state["prev_value"]] + return x, attn, self_attn_state + return x, attn, None + + def make_generation_fast_(self, need_attn: bool = False, **kwargs): + self.need_attn = need_attn + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m diff --git a/fairseq/modules/transformer_sentence_encoder.py b/fairseq/modules/transformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..8a6994181b84c3f3dac318b2539b282cd0f42590 --- /dev/null +++ b/fairseq/modules/transformer_sentence_encoder.py @@ -0,0 +1,278 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from fairseq.modules import ( + FairseqDropout, + LayerDropModuleList, + LayerNorm, + MultiheadAttention, + PositionalEmbedding, + TransformerSentenceEncoderLayer, +) +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ + + +def init_bert_params(module): + """ + Initialize the weights specific to the BERT Model. + This overrides the default initializations depending on the specified arguments. + 1. If normal_init_linear_weights is set then weights of linear + layer will be initialized using the normal distribution and + bais will be set to the specified value. + 2. If normal_init_embed_weights is set then weights of embedding + layer will be initialized using the normal distribution. + 3. If normal_init_proj_weights is set then weights of + in_project_weight for MultiHeadAttention initialized using + the normal distribution (to be validated). + """ + + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=0.02) + if module.bias is not None: + module.bias.data.zero_() + if isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=0.02) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + if isinstance(module, MultiheadAttention): + module.q_proj.weight.data.normal_(mean=0.0, std=0.02) + module.k_proj.weight.data.normal_(mean=0.0, std=0.02) + module.v_proj.weight.data.normal_(mean=0.0, std=0.02) + + +class TransformerSentenceEncoder(nn.Module): + """ + Implementation for a Bi-directional Transformer based Sentence Encoder used + in BERT/XLM style pre-trained models. + + This first computes the token embedding using the token embedding matrix, + position embeddings (if specified) and segment embeddings + (if specified). After applying the specified number of + TransformerEncoderLayers, it outputs all the internal states of the + encoder as well as the final representation associated with the first + token (usually CLS token). + + Input: + - tokens: B x T matrix representing sentences + - segment_labels: B x T matrix representing segment label for tokens + + Output: + - a tuple of the following: + - a list of internal model states used to compute the + predictions where each tensor has shape T x B x C + - sentence representation associated with first input token + in format B x C. + """ + + def __init__( + self, + padding_idx: int, + vocab_size: int, + num_encoder_layers: int = 6, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + layerdrop: float = 0.0, + max_seq_len: int = 256, + num_segments: int = 2, + use_position_embeddings: bool = True, + offset_positions_by_padding: bool = True, + encoder_normalize_before: bool = False, + apply_bert_init: bool = False, + activation_fn: str = "relu", + learned_pos_embedding: bool = True, + embed_scale: float = None, + freeze_embeddings: bool = False, + n_trans_layers_to_freeze: int = 0, + export: bool = False, + traceable: bool = False, + q_noise: float = 0.0, + qn_block_size: int = 8, + ) -> None: + + super().__init__() + self.padding_idx = padding_idx + self.vocab_size = vocab_size + self.dropout_module = FairseqDropout(dropout, module_name=self.__class__.__name__) + self.layerdrop = layerdrop + self.max_seq_len = max_seq_len + self.embedding_dim = embedding_dim + self.num_segments = num_segments + self.use_position_embeddings = use_position_embeddings + self.apply_bert_init = apply_bert_init + self.learned_pos_embedding = learned_pos_embedding + self.traceable = traceable + self.tpu = False # whether we're on TPU + + self.embed_tokens = self.build_embedding( + self.vocab_size, self.embedding_dim, self.padding_idx + ) + self.embed_scale = embed_scale + + if q_noise > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(self.embedding_dim, self.embedding_dim, bias=False), + q_noise, + qn_block_size, + ) + else: + self.quant_noise = None + + self.segment_embeddings = ( + nn.Embedding(self.num_segments, self.embedding_dim, padding_idx=None) + if self.num_segments > 0 + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + self.max_seq_len, + self.embedding_dim, + padding_idx=(self.padding_idx if offset_positions_by_padding else None), + learned=self.learned_pos_embedding, + ) + if self.use_position_embeddings + else None + ) + + if self.layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend([ + self.build_transformer_sentence_encoder_layer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=self.dropout_module.p, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + for _ in range(num_encoder_layers) + ]) + + if encoder_normalize_before: + self.emb_layer_norm = LayerNorm(self.embedding_dim, export=export) + else: + self.emb_layer_norm = None + + # Apply initialization of model params after building the model + if self.apply_bert_init: + self.apply(init_bert_params) + + def freeze_module_params(m): + if m is not None: + for p in m.parameters(): + p.requires_grad = False + + if freeze_embeddings: + freeze_module_params(self.embed_tokens) + freeze_module_params(self.segment_embeddings) + freeze_module_params(self.embed_positions) + freeze_module_params(self.emb_layer_norm) + + for layer in range(n_trans_layers_to_freeze): + freeze_module_params(self.layers[layer]) + + def build_embedding(self, vocab_size, embedding_dim, padding_idx): + return nn.Embedding(vocab_size, embedding_dim, padding_idx) + + def build_transformer_sentence_encoder_layer( + self, + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + activation_fn, + export, + q_noise, + qn_block_size, + ): + return TransformerSentenceEncoderLayer( + embedding_dim=embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + def prepare_for_tpu_(self, **kwargs): + self.tpu = True + + def forward( + self, + tokens: torch.Tensor, + segment_labels: torch.Tensor = None, + last_state_only: bool = False, + positions: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + + # compute padding mask. This is needed for multi-head attention + padding_mask = tokens.eq(self.padding_idx) + if not self.traceable and not self.tpu and not padding_mask.any(): + padding_mask = None + + x = self.embed_tokens(tokens) + + if self.embed_scale is not None: + x *= self.embed_scale + + if self.embed_positions is not None: + x += self.embed_positions(tokens, positions=positions) + + if self.segment_embeddings is not None and segment_labels is not None: + x += self.segment_embeddings(segment_labels) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.emb_layer_norm is not None: + x = self.emb_layer_norm(x) + + x = self.dropout_module(x) + + # account for padding while computing the representation + if padding_mask is not None: + x *= 1 - padding_mask.unsqueeze(-1).type_as(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + inner_states = [] + if not last_state_only: + inner_states.append(x) + + for layer in self.layers: + x, _ = layer(x, self_attn_padding_mask=padding_mask) + if not last_state_only: + inner_states.append(x) + + sentence_rep = x[0, :, :] + + if last_state_only: + inner_states = [x] + + if self.traceable: + return torch.stack(inner_states), sentence_rep + else: + return inner_states, sentence_rep diff --git a/fairseq/modules/transformer_sentence_encoder_layer.py b/fairseq/modules/transformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..383938f68fae6920bb79f4a48bc3bbcf708cc80d --- /dev/null +++ b/fairseq/modules/transformer_sentence_encoder_layer.py @@ -0,0 +1,139 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Callable, Optional + +import torch +import torch.nn as nn + +from fairseq import utils +from fairseq.modules import ( + LayerNorm, + MultiheadAttention, +) +from fairseq.modules.quant_noise import quant_noise +from fairseq.modules.fairseq_dropout import FairseqDropout + + + +class TransformerSentenceEncoderLayer(nn.Module): + """ + Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__( + self, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = 'relu', + export: bool = False, + q_noise: float = 0.0, + qn_block_size: int = 8, + init_fn: Callable = None, + ) -> None: + super().__init__() + + if init_fn is not None: + init_fn() + + # Initialize parameters + self.embedding_dim = embedding_dim + self.dropout_module = FairseqDropout(dropout, module_name=self.__class__.__name__) + self.activation_dropout_module = FairseqDropout(activation_dropout, module_name=self.__class__.__name__) + + # Initialize blocks + self.activation_fn = utils.get_activation_fn(activation_fn) + self.self_attn = self.build_self_attention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + self_attention=True, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + # layer norm associated with the self attention layer + self.self_attn_layer_norm = LayerNorm(self.embedding_dim, export=export) + + self.fc1 = self.build_fc1( + self.embedding_dim, + ffn_embedding_dim, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + self.fc2 = self.build_fc2( + ffn_embedding_dim, + self.embedding_dim, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + # layer norm associated with the position wise feed-forward NN + self.final_layer_norm = LayerNorm(self.embedding_dim, export=export) + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise( + nn.Linear(input_dim, output_dim), q_noise, qn_block_size + ) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise( + nn.Linear(input_dim, output_dim), q_noise, qn_block_size + ) + + def build_self_attention( + self, + embed_dim, + num_attention_heads, + dropout, + self_attention, + q_noise, + qn_block_size, + ): + return MultiheadAttention( + embed_dim, + num_attention_heads, + dropout=dropout, + self_attention=True, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer implementation. + """ + residual = x + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout_module(x) + x = residual + x + x = self.self_attn_layer_norm(x) + + residual = x + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + x = self.final_layer_norm(x) + return x, attn diff --git a/fairseq/modules/transpose_last.py b/fairseq/modules/transpose_last.py new file mode 100644 index 0000000000000000000000000000000000000000..e578b3ec5097bfac5c976b207ea46bec1d9bd4f5 --- /dev/null +++ b/fairseq/modules/transpose_last.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +transpose last 2 dimensions of the input +""" + +import torch.nn as nn + + +class TransposeLast(nn.Module): + def __init__(self, deconstruct_idx=None): + super().__init__() + self.deconstruct_idx = deconstruct_idx + + def forward(self, x): + if self.deconstruct_idx is not None: + x = x[self.deconstruct_idx] + return x.transpose(-2, -1) diff --git a/fairseq/modules/unfold.py b/fairseq/modules/unfold.py new file mode 100644 index 0000000000000000000000000000000000000000..3a142db69868ff1e36241c7f032d0f886b6b9428 --- /dev/null +++ b/fairseq/modules/unfold.py @@ -0,0 +1,17 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn.functional as F + + +def unfold1d(x, kernel_size, padding_l, pad_value=0): + '''unfold T x B x C to T x B x C x K''' + if kernel_size > 1: + T, B, C = x.size() + x = F.pad(x, (0, 0, 0, 0, padding_l, kernel_size - 1 - padding_l), value=pad_value) + x = x.as_strided((T, B, C, kernel_size), (B*C, C, 1, B*C)) + else: + x = x.unsqueeze(3) + return x diff --git a/fairseq/modules/vggblock.py b/fairseq/modules/vggblock.py new file mode 100644 index 0000000000000000000000000000000000000000..ee5ee19a34816c7350c21fba7c4907fec8ca7a61 --- /dev/null +++ b/fairseq/modules/vggblock.py @@ -0,0 +1,116 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +from collections.abc import Iterable +from itertools import repeat + +import torch +import torch.nn as nn + + +def _pair(v): + if isinstance(v, Iterable): + assert len(v) == 2, "len(v) != 2" + return v + return tuple(repeat(v, 2)) + + +def infer_conv_output_dim(conv_op, input_dim, sample_inchannel): + sample_seq_len = 200 + sample_bsz = 10 + x = torch.randn(sample_bsz, sample_inchannel, sample_seq_len, input_dim) + # N x C x H x W + # N: sample_bsz, C: sample_inchannel, H: sample_seq_len, W: input_dim + x = conv_op(x) + # N x C x H x W + x = x.transpose(1, 2) + # N x H x C x W + bsz, seq = x.size()[:2] + per_channel_dim = x.size()[3] + # bsz: N, seq: H, CxW the rest + return x.contiguous().view(bsz, seq, -1).size(-1), per_channel_dim + + +class VGGBlock(torch.nn.Module): + """ + VGG motibated cnn module https://arxiv.org/pdf/1409.1556.pdf + + Args: + in_channels: (int) number of input channels (typically 1) + out_channels: (int) number of output channels + conv_kernel_size: convolution channels + pooling_kernel_size: the size of the pooling window to take a max over + num_conv_layers: (int) number of convolution layers + input_dim: (int) input dimension + conv_stride: the stride of the convolving kernel. + Can be a single number or a tuple (sH, sW) Default: 1 + padding: implicit paddings on both sides of the input. + Can be a single number or a tuple (padH, padW). Default: None + layer_norm: (bool) if layer norm is going to be applied. Default: False + + Shape: + Input: BxCxTxfeat, i.e. (batch_size, input_size, timesteps, features) + Output: BxCxTxfeat, i.e. (batch_size, input_size, timesteps, features) + """ + + def __init__( + self, + in_channels, + out_channels, + conv_kernel_size, + pooling_kernel_size, + num_conv_layers, + input_dim, + conv_stride=1, + padding=None, + layer_norm=False, + ): + assert ( + input_dim is not None + ), "Need input_dim for LayerNorm and infer_conv_output_dim" + super(VGGBlock, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.conv_kernel_size = _pair(conv_kernel_size) + self.pooling_kernel_size = _pair(pooling_kernel_size) + self.num_conv_layers = num_conv_layers + self.padding = ( + tuple(e // 2 for e in self.conv_kernel_size) + if padding is None + else _pair(padding) + ) + self.conv_stride = _pair(conv_stride) + + self.layers = nn.ModuleList() + for layer in range(num_conv_layers): + conv_op = nn.Conv2d( + in_channels if layer == 0 else out_channels, + out_channels, + self.conv_kernel_size, + stride=self.conv_stride, + padding=self.padding, + ) + self.layers.append(conv_op) + if layer_norm: + conv_output_dim, per_channel_dim = infer_conv_output_dim( + conv_op, input_dim, in_channels if layer == 0 else out_channels + ) + self.layers.append(nn.LayerNorm(per_channel_dim)) + input_dim = per_channel_dim + self.layers.append(nn.ReLU()) + + if self.pooling_kernel_size is not None: + pool_op = nn.MaxPool2d(kernel_size=self.pooling_kernel_size, ceil_mode=True) + self.layers.append(pool_op) + self.total_output_dim, self.output_dim = infer_conv_output_dim( + pool_op, input_dim, out_channels + ) + + def forward(self, x): + for i, _ in enumerate(self.layers): + x = self.layers[i](x) + return x diff --git a/fairseq/nan_detector.py b/fairseq/nan_detector.py new file mode 100644 index 0000000000000000000000000000000000000000..89ea982f69af94875be6855130e3e612f0e5eb21 --- /dev/null +++ b/fairseq/nan_detector.py @@ -0,0 +1,89 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import torch + +logger = logging.getLogger(__name__) + + +class NanDetector: + """ + Detects the first NaN or Inf in forward and/or backward pass and logs, together with the module name + """ + + def __init__(self, model, forward=True, backward=True): + self.bhooks = [] + self.fhooks = [] + self.forward = forward + self.backward = backward + self.reset() + + for name, mod in model.named_modules(): + mod.__module_name = name + self.add_hooks(mod) + + def __enter__(self): + return self + + def __exit__(self, exc_type, exc_value, exc_traceback): + self.close() + + def add_hooks(self, module): + if self.forward: + self.fhooks.append(module.register_forward_hook(self.fhook_fn)) + if self.backward: + self.bhooks.append(module.register_backward_hook(self.bhook_fn)) + + def reset(self): + self.has_printed_f = False + self.has_printed_b = False + + def _detect(self, tensor, name, backward): + err = None + if ( + tensor.numel() >= 2 + ): # single value tensors (like the loss) will not provide much info + with torch.no_grad(): + if torch.isnan(tensor).any(): + err = "NaN" + elif torch.isinf(tensor).any(): + err = "Inf" + if err is not None: + err = f"{err} detected in output of {name}, shape: {tensor.shape}, {'backward' if backward else 'forward'}" + return err + + def _apply(self, module, inp, x, backward): + if torch.is_tensor(x): + if isinstance(inp, tuple) and len(inp) > 0: + inp = inp[0] + err = self._detect(x, module.__module_name, backward) + if err is not None: + if torch.is_tensor(inp) and not backward: + err += ( + f" input max: {inp.max().item()}, input min: {inp.min().item()}" + ) + + has_printed_attr = 'has_printed_b' if backward else 'has_printed_f' + logger.warning(err) + setattr(self, has_printed_attr, True) + elif isinstance(x, dict): + for v in x.values(): + self._apply(module, inp, v, backward) + elif isinstance(x, list) or isinstance(x, tuple): + for v in x: + self._apply(module, inp, v, backward) + + def fhook_fn(self, module, inp, output): + if not self.has_printed_f: + self._apply(module, inp, output, backward=False) + + def bhook_fn(self, module, inp, output): + if not self.has_printed_b: + self._apply(module, inp, output, backward=True) + + def close(self): + for hook in self.fhooks + self.bhooks: + hook.remove() diff --git a/fairseq/optim/__init__.py b/fairseq/optim/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..2b8334d8c25b788aea5780079648d55052f64d5c --- /dev/null +++ b/fairseq/optim/__init__.py @@ -0,0 +1,33 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + +from fairseq import registry +from fairseq.optim.fairseq_optimizer import FairseqOptimizer +from fairseq.optim.fp16_optimizer import FP16Optimizer, MemoryEfficientFP16Optimizer +from fairseq.optim.bmuf import FairseqBMUF # noqa + + +__all__ = [ + 'FairseqOptimizer', + 'FP16Optimizer', + 'MemoryEfficientFP16Optimizer', +] + + +build_optimizer, register_optimizer, OPTIMIZER_REGISTRY = registry.setup_registry( + '--optimizer', + base_class=FairseqOptimizer, + default='nag', +) + + +# automatically import any Python files in the optim/ directory +for file in os.listdir(os.path.dirname(__file__)): + if file.endswith('.py') and not file.startswith('_'): + module = file[:file.find('.py')] + importlib.import_module('fairseq.optim.' + module) diff --git a/fairseq/optim/__pycache__/__init__.cpython-310.pyc b/fairseq/optim/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b9ab0231c84df08851d5c6f4c179d6cf129fb6d6 Binary files /dev/null and b/fairseq/optim/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/adadelta.cpython-310.pyc b/fairseq/optim/__pycache__/adadelta.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a4129ee70ebf74c8e82ba39db34c3f1243797be0 Binary files /dev/null and b/fairseq/optim/__pycache__/adadelta.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/adafactor.cpython-310.pyc b/fairseq/optim/__pycache__/adafactor.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d0d6d4fa3f136d954fca4f9a694c338e12b39d1a Binary files /dev/null and b/fairseq/optim/__pycache__/adafactor.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/adagrad.cpython-310.pyc b/fairseq/optim/__pycache__/adagrad.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..03fa6e7012330d6731f583c64d2c93f0b7e88095 Binary files /dev/null and b/fairseq/optim/__pycache__/adagrad.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/adam.cpython-310.pyc b/fairseq/optim/__pycache__/adam.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0c15fb2ea707c606c24c0833365fdb03541aca0b Binary files /dev/null and b/fairseq/optim/__pycache__/adam.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/adamax.cpython-310.pyc b/fairseq/optim/__pycache__/adamax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0dc9db59a20e3bada2772c95e234b69bba742aac Binary files /dev/null and b/fairseq/optim/__pycache__/adamax.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/bmuf.cpython-310.pyc b/fairseq/optim/__pycache__/bmuf.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..29eed3c91781060bf0b29889d754dec4d61e7fa8 Binary files /dev/null and b/fairseq/optim/__pycache__/bmuf.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/dynamic_loss_scaler.cpython-310.pyc b/fairseq/optim/__pycache__/dynamic_loss_scaler.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e2729d79800d541a9432a707f0e3f681445f46b5 Binary files /dev/null and b/fairseq/optim/__pycache__/dynamic_loss_scaler.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/fairseq_optimizer.cpython-310.pyc b/fairseq/optim/__pycache__/fairseq_optimizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cfb2da269c132a6f486191faecb4b39c1d41c75d Binary files /dev/null and b/fairseq/optim/__pycache__/fairseq_optimizer.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/fp16_optimizer.cpython-310.pyc b/fairseq/optim/__pycache__/fp16_optimizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4421b5b1c55128be5b02e9e994f4801621ea29d1 Binary files /dev/null and b/fairseq/optim/__pycache__/fp16_optimizer.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/fused_adam.cpython-310.pyc b/fairseq/optim/__pycache__/fused_adam.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5e36b1a7e75c3ce5cb9b1b53acba942a147c91cb Binary files /dev/null and b/fairseq/optim/__pycache__/fused_adam.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/fused_lamb.cpython-310.pyc b/fairseq/optim/__pycache__/fused_lamb.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..66a245e719e26594d2ce9265f4d82b5af0447548 Binary files /dev/null and b/fairseq/optim/__pycache__/fused_lamb.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/nag.cpython-310.pyc b/fairseq/optim/__pycache__/nag.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8a49e8e1575bef56f71c3f9b8d337feff603ec32 Binary files /dev/null and b/fairseq/optim/__pycache__/nag.cpython-310.pyc differ diff --git a/fairseq/optim/__pycache__/sgd.cpython-310.pyc b/fairseq/optim/__pycache__/sgd.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bb18d98abae412c9f471859695bebc9059bf9c32 Binary files /dev/null and b/fairseq/optim/__pycache__/sgd.cpython-310.pyc differ diff --git a/fairseq/optim/adadelta.py b/fairseq/optim/adadelta.py new file mode 100644 index 0000000000000000000000000000000000000000..0a76e27fe41dd5966b8bcd61768f511dcf4e5d30 --- /dev/null +++ b/fairseq/optim/adadelta.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.optim + +from . import FairseqOptimizer, register_optimizer + + +@register_optimizer('adadelta') +class Adadelta(FairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = torch.optim.Adadelta(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--adadelta-rho', type=float, default=0.9, metavar='RHO', + help='coefficient used for computing a running average of squared gradients') + parser.add_argument('--adadelta-eps', type=float, default=1e-6, metavar='EPS', + help='term added to the denominator to improve numerical stability') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + parser.add_argument('--anneal-eps', action='store_true', help='flag to anneal eps') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + 'lr': self.args.lr[0], + 'rho': self.args.adadelta_rho, + 'eps': self.args.adadelta_eps, + 'weight_decay': self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return True diff --git a/fairseq/optim/adafactor.py b/fairseq/optim/adafactor.py new file mode 100644 index 0000000000000000000000000000000000000000..f52ec0f139b91ed55272011b4fa459a73af16546 --- /dev/null +++ b/fairseq/optim/adafactor.py @@ -0,0 +1,237 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import torch +import torch.optim + +from . import FairseqOptimizer, register_optimizer + + +@register_optimizer('adafactor') +class FairseqAdafactor(FairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = Adafactor(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--adafactor-eps', default='(1e-30, 1e-3)', metavar="E", + help='epsilons for Adafactor optimizer') + parser.add_argument('--clip-threshold', type=float, default=1.0, metavar="C", + help='threshold for clipping update root mean square') + parser.add_argument('--decay-rate', type=float, default=-0.8, metavar="D", + help='decay rate of the second moment estimator') + parser.add_argument('--beta1', type=float, default=None, metavar="B", + help='beta for first moment estimator. Optional') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + parser.add_argument('--scale-parameter', action='store_true', + help='scale learning rate by root mean square of parameter') + parser.add_argument('--relative-step', action='store_true', + help='set learning rate to inverse square root of timestep,' + 'otherwise use external learning rate') + parser.add_argument('--warmup-init', action='store_true', + help='use relative step for warm-up learning rate schedule') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + Note : Convergence issues empirically observed with fp16 on. + Might require search for appropriate configuration. + """ + return { + 'lr': self.args.lr[0], + 'eps': eval(self.args.adafactor_eps), + 'clip_threshold': self.args.clip_threshold, + 'decay_rate': self.args.decay_rate, + 'beta1': self.args.beta1, + 'weight_decay': self.args.weight_decay, + 'scale_parameter': self.args.scale_parameter, # defaults to False + 'relative_step': self.args.relative_step, # defaults to False + 'warmup_init': self.args.warmup_init, + } + + +class Adafactor(torch.optim.Optimizer): + """Implements Adafactor algorithm. + + This implementation is based on: + `Adafactor: Adaptive Learning Rates with Sublinear Memory Cost` + (see https://arxiv.org/abs/1804.04235) + + Note that this optimizer internally adjusts the learning rate + depending on the *scale_parameter*, *relative_step* and + *warmup_init* options. To use a manual (external) learning rate + schedule you should set `scale_parameter=False` and + `relative_step=False`. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): external learning rate (default: None) + eps (tuple[float, float]): regularization constans for square gradient + and parameter scale respectively (default: (1e-30, 1e-3)) + clip_threshold (float): threshold of root mean square of + final gradient update (default: 1.0) + decay_rate (float): coefficient used to compute running averages of square + gradient (default: -0.8) + beta1 (float): coefficient used for computing running averages of gradient + (default: None) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + scale_parameter (bool): if True, learning rate is scaled by root mean square of + parameter (default: True) + relative_step (bool): if True, time-dependent learning rate is computed + instead of external learning rate (default: True) + warmup_init (bool): time-dependent learning rate computation depends on + whether warm-up initialization is being used (default: False) + """ + + def __init__(self, params, lr=None, eps=(1e-30, 1e-3), clip_threshold=1.0, + decay_rate=-0.8, beta1=None, weight_decay=0.0, scale_parameter=True, + relative_step=True, warmup_init=False): + if lr is not None and relative_step: + raise ValueError('Cannot combine manual lr and relative_step options') + if warmup_init and not relative_step: + raise ValueError('warmup_init requires relative_step=True') + + defaults = dict(lr=lr, eps=eps, clip_threshold=clip_threshold, decay_rate=decay_rate, + beta1=beta1, weight_decay=weight_decay, scale_parameter=scale_parameter, + relative_step=relative_step, warmup_init=warmup_init) + super(Adafactor, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return False + + def _get_lr(self, param_group, param_state): + rel_step_sz = param_group['lr'] + if param_group['relative_step']: + min_step = 1e-6 * param_state['step'] if param_group['warmup_init'] else 1e-2 + rel_step_sz = min(min_step, 1.0/math.sqrt(param_state['step'])) + param_scale = 1.0 + if param_group['scale_parameter']: + param_scale = max(param_group['eps'][1], param_state['RMS']) + return param_scale * rel_step_sz + + def _get_options(self, param_group, param_shape): + factored = len(param_shape) >= 2 + use_first_moment = param_group['beta1'] is not None + return factored, use_first_moment + + def _rms(self, tensor): + return tensor.norm(2) / (tensor.numel() ** 0.5) + + def _approx_sq_grad(self, exp_avg_sq_row, exp_avg_sq_col): + r_factor = ( + exp_avg_sq_row / exp_avg_sq_row.mean(dim=-1, keepdim=True) + ).rsqrt_() + c_factor = exp_avg_sq_col.rsqrt() + return torch.mm(r_factor.unsqueeze(-1), c_factor.unsqueeze(0)) + + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group['params']: + if p.grad is None: + continue + grad = p.grad.data + if grad.dtype in {torch.float16, torch.bfloat16}: + grad = grad.float() + if grad.is_sparse: + raise RuntimeError('Adafactor does not support sparse gradients.') + + state = self.state[p] + grad_shape = grad.shape + + factored, use_first_moment = self._get_options(group, grad_shape) + # State Initialization + if len(state) == 0: + state['step'] = 0 + + if use_first_moment: + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(grad) + if factored: + state['exp_avg_sq_row'] = torch.zeros(grad_shape[:-1]).to(grad) + state['exp_avg_sq_col'] = torch.zeros(grad_shape[:-2] + grad_shape[-1:]).to(grad) + else: + state['exp_avg_sq'] = torch.zeros_like(grad) + + state['RMS'] = 0 + else: + if use_first_moment: + state['exp_avg'] = state['exp_avg'].to(grad) + if factored: + state['exp_avg_sq_row'] = state['exp_avg_sq_row'].to(grad) + state['exp_avg_sq_col'] = state['exp_avg_sq_col'].to(grad) + else: + state['exp_avg_sq'] = state['exp_avg_sq'].to(grad) + + p_data_fp32 = p.data + if p.data.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + state['step'] += 1 + state['RMS'] = self._rms(p_data_fp32) + group['lr'] = self._get_lr(group, state) + + beta2t = 1.0 - math.pow(state['step'], group['decay_rate']) + update = (grad**2) + group['eps'][0] + if factored: + exp_avg_sq_row = state['exp_avg_sq_row'] + exp_avg_sq_col = state['exp_avg_sq_col'] + + exp_avg_sq_row.mul_(beta2t).add_(1.0 - beta2t, update.mean(dim=-1)) + exp_avg_sq_col.mul_(beta2t).add_(1.0 - beta2t, update.mean(dim=-2)) + + # Approximation of exponential moving average of square of gradient + update = self._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col) + update.mul_(grad) + else: + exp_avg_sq = state['exp_avg_sq'] + + exp_avg_sq.mul_(beta2t).add_(1.0 - beta2t, update) + update = exp_avg_sq.rsqrt().mul_(grad) + + update.div_( + (self._rms(update) / group['clip_threshold']).clamp_(min=1.0) + ) + update.mul_(group['lr']) + + if use_first_moment: + exp_avg = state['exp_avg'] + exp_avg.mul_(group['beta1']).add_(1 - group['beta1'], update) + update = exp_avg + + if group['weight_decay'] != 0: + p_data_fp32.add_(-group['weight_decay'] * group['lr'], p_data_fp32) + + p_data_fp32.add_(-update) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + return loss diff --git a/fairseq/optim/adagrad.py b/fairseq/optim/adagrad.py new file mode 100644 index 0000000000000000000000000000000000000000..57f83258cfd177c5381210115f679a501241c6ed --- /dev/null +++ b/fairseq/optim/adagrad.py @@ -0,0 +1,40 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.optim + +from . import FairseqOptimizer, register_optimizer + + +@register_optimizer('adagrad') +class Adagrad(FairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = torch.optim.Adagrad(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + 'lr': self.args.lr[0], + 'weight_decay': self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return True diff --git a/fairseq/optim/adam.py b/fairseq/optim/adam.py new file mode 100644 index 0000000000000000000000000000000000000000..d5783b258cfe3a3d9e16411ad7012f2146c30136 --- /dev/null +++ b/fairseq/optim/adam.py @@ -0,0 +1,209 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +import types + +import torch +import torch.optim +import torch.distributed as dist + +from fairseq.optim import FairseqOptimizer, register_optimizer +from fairseq.optim.fused_adam import get_fused_adam_class + +logger = logging.getLogger(__name__) + + +@register_optimizer('adam') +class FairseqAdam(FairseqOptimizer): + """Adam optimizer for fairseq. + + Important note: this optimizer corresponds to the "AdamW" variant of + Adam in its weight decay behavior. As such, it is most closely + analogous to torch.optim.AdamW from PyTorch. + """ + + def __init__(self, args, params): + super().__init__(args) + fused_adam_cls = get_fused_adam_class() + use_fused_adam = ( + not getattr(args, 'use_old_adam', False) + and fused_adam_cls is not None + and torch.cuda.is_available() + ) + if getattr(args, 'tpu', False): + # on TPUs we use the Adam defined here, since it + # automatically casts gradients to FP32 + self._optimizer = Adam(params, **self.optimizer_config) + elif use_fused_adam: + logger.info('using FusedAdam') + self._optimizer = fused_adam_cls(params, **self.optimizer_config) + else: + self._optimizer = Adam(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--adam-betas', default='(0.9, 0.999)', metavar='B', + help='betas for Adam optimizer') + parser.add_argument('--adam-eps', type=float, default=1e-8, metavar='D', + help='epsilon for Adam optimizer') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # Maintain backward compatibility with old checkpoints that have stored + # optimizer state as fairseq.optim.adam.Adam. + parser.add_argument( + "--use-old-adam", + action='store_true', + default=False, + help="Use fairseq.optim.adam.Adam", + ) + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + 'lr': self.args.lr[0], + 'betas': eval(self.args.adam_betas), + 'eps': self.args.adam_eps, + 'weight_decay': self.args.weight_decay, + } + + def average_params(self): + """Reduce Params is only used during BMUF distributed training.""" + state_dict = self.optimizer.state_dict() + total_gpus = float(dist.get_world_size()) + + for _, value in state_dict["state"].items(): + value["exp_avg"] /= total_gpus + value["exp_avg_sq"] /= total_gpus + dist.all_reduce(value["exp_avg"], op=dist.ReduceOp.SUM) + dist.all_reduce(value["exp_avg_sq"], op=dist.ReduceOp.SUM) + + +class Adam(torch.optim.Optimizer): + """Implements Adam algorithm. + + This implementation is modified from torch.optim.Adam based on: + `Fixed Weight Decay Regularization in Adam` + (see https://arxiv.org/abs/1711.05101) + + It has been proposed in `Adam: A Method for Stochastic Optimization`_. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + amsgrad (boolean, optional): whether to use the AMSGrad variant of this + algorithm from the paper `On the Convergence of Adam and Beyond`_ + + .. _Adam\: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, + weight_decay=0, amsgrad=False): + defaults = dict(lr=lr, betas=betas, eps=eps, + weight_decay=weight_decay, amsgrad=amsgrad) + super(Adam, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group['params']: + if p.grad is None: + continue + grad = p.grad.data + if grad.dtype in {torch.float16, torch.bfloat16}: + grad = grad.float() + if grad.is_sparse: + raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead') + amsgrad = group['amsgrad'] + + p_data_fp32 = p.data + if p.data.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + state = self.state[p] + + # State initialization + if len(state) == 0: + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p_data_fp32) + # Exponential moving average of squared gradient values + state['exp_avg_sq'] = torch.zeros_like(p_data_fp32) + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state['max_exp_avg_sq'] = torch.zeros_like(p_data_fp32) + else: + state['exp_avg'] = state['exp_avg'].to(p_data_fp32) + state['exp_avg_sq'] = state['exp_avg_sq'].to(p_data_fp32) + if amsgrad: + state['max_exp_avg_sq'] = state['max_exp_avg_sq'].to(p_data_fp32) + + exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] + if amsgrad: + max_exp_avg_sq = state['max_exp_avg_sq'] + beta1, beta2 = group['betas'] + + state['step'] += 1 + + # Decay the first and second moment running average coefficient + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + if amsgrad: + # Maintains the maximum of all 2nd moment running avg. till now + torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq) + # Use the max. for normalizing running avg. of gradient + denom = max_exp_avg_sq.sqrt().add_(group['eps']) + else: + denom = exp_avg_sq.sqrt().add_(group['eps']) + + bias_correction1 = 1 - beta1 ** state['step'] + bias_correction2 = 1 - beta2 ** state['step'] + step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1 + + if group['weight_decay'] != 0: + p_data_fp32.add_(p_data_fp32, alpha=-group['weight_decay'] * group['lr']) + + p_data_fp32.addcdiv_(exp_avg, denom, value=-step_size) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + return loss diff --git a/fairseq/optim/adamax.py b/fairseq/optim/adamax.py new file mode 100644 index 0000000000000000000000000000000000000000..856215a3ba97ed03ddf5741fed692d0dc32af947 --- /dev/null +++ b/fairseq/optim/adamax.py @@ -0,0 +1,158 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.optim + +from . import FairseqOptimizer, register_optimizer + + +@register_optimizer('adamax') +class FairseqAdamax(FairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = Adamax(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--adamax-betas', default='(0.9, 0.999)', metavar='B', + help='betas for Adam optimizer') + parser.add_argument('--adamax-eps', type=float, default=1e-8, metavar='D', + help='epsilon for Adam optimizer') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + parser.add_argument('--no-bias-correction', default=False, action='store_true', + help='disable bias correction') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + 'lr': self.args.lr[0], + 'betas': eval(self.args.adamax_betas), + 'eps': self.args.adamax_eps, + 'weight_decay': self.args.weight_decay, + 'bias_correction': not self.args.no_bias_correction, + } + + +class Adamax(torch.optim.Optimizer): + """Implements Adamax algorithm (a variant of Adam based on infinity norm). + + It has been proposed in `Adam: A Method for Stochastic Optimization`__. + + Compared to the version in PyTorch, this version implements a fix for weight decay. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 2e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + bias_correction (bool, optional): enable bias correction (default: True) + + __ https://arxiv.org/abs/1412.6980 + """ + + def __init__(self, params, lr=2e-3, betas=(0.9, 0.999), eps=1e-8, + weight_decay=0, bias_correction=True): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0.0 <= weight_decay: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + + defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, + bias_correction=bias_correction) + super(Adamax, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group['params']: + if p.grad is None: + continue + grad = p.grad.data.float() + if grad.is_sparse: + raise RuntimeError('Adamax does not support sparse gradients') + + p_data_fp32 = p.data + if p.data.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + state = self.state[p] + + # State initialization + if len(state) == 0: + state['step'] = 0 + state['exp_avg'] = torch.zeros_like(p_data_fp32) + state['exp_inf'] = torch.zeros_like(p_data_fp32) + else: + state['exp_avg'] = state['exp_avg'].to(p_data_fp32) + state['exp_inf'] = state['exp_inf'].to(p_data_fp32) + + exp_avg, exp_inf = state['exp_avg'], state['exp_inf'] + beta1, beta2 = group['betas'] + eps = group['eps'] + + state['step'] += 1 + + # Update biased first moment estimate. + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + + # Update the exponentially weighted infinity norm. + torch.max( + exp_inf.mul_(beta2), + grad.abs_(), + out=exp_inf, + ) + + step_size = group['lr'] + if group['bias_correction']: + bias_correction = 1 - beta1 ** state['step'] + step_size /= bias_correction + + if group['weight_decay'] != 0: + p_data_fp32.add_(p_data_fp32, alpha=-group['weight_decay'] * group['lr']) + + p_data_fp32.addcdiv_(exp_avg, exp_inf.add(eps), value=-step_size) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + return loss diff --git a/fairseq/optim/bmuf.py b/fairseq/optim/bmuf.py new file mode 100644 index 0000000000000000000000000000000000000000..be7bdd74a777a626dbb8035b6c21c0ed3af75a67 --- /dev/null +++ b/fairseq/optim/bmuf.py @@ -0,0 +1,230 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.distributed as dist + +from . import FairseqOptimizer + + +class FairseqBMUF(FairseqOptimizer): + """ + Implements incremental block distributed data parallelism similar to + https://ieeexplore.ieee.org/document/7472805 + + Paper title: Scalable training of deep learning machines by incremental + block training with intra-block parallel optimization and blockwise + model-update filtering + """ + + def __init__(self, args, optimizer): + + super().__init__(args) + self._optimizer = optimizer + self._num_updates = 0 + self.sync_iter = self.args.global_sync_iter + self.block_momentum = self.args.block_momentum + self.block_lr = self.args.block_lr + self._reset_local_data() + self.warmup_iteration = self.args.warmup_iterations + self.use_nbm = self.args.use_nbm + self.initial_state = self._optimizer.state_dict() + self.average_sync = self.args.average_sync + self.world_size = self.args.distributed_world_size + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + parser.add_argument( + "--block-lr", default=1, type=float, help="block learning rate for bmuf" + ) + parser.add_argument( + "--block-momentum", + default=0.875, + type=float, + help="block momentum for bmuf", + ) + parser.add_argument( + "--global-sync-iter", + default=50, + type=int, + help="Iteration for syncing global model", + ) + parser.add_argument( + "--warmup-iterations", + default=500, + type=int, + help="warmup iterations for model to broadcast", + ) + parser.add_argument( + "--use-nbm", + default=False, + action="store_true", + help="Specify whether you want to use classical BM / Nesterov BM", + ) + parser.add_argument( + "--average-sync", + default=False, + action="store_true", + help="Specify whether you want to average the local momentum after each sync", + ) + + @property + def optimizer(self): + return self._optimizer.optimizer + + @property + def optimizer_config(self): + return self._optimizer.optimizer_config + + def get_lr(self): + return self._optimizer.get_lr() + + def set_lr(self, lr): + self._optimizer.set_lr(lr) + + def state_dict(self): + return self._optimizer.state_dict() + + def load_state_dict(self, state_dict, optimizer_overrides=None): + self._optimizer.load_state_dict(state_dict, optimizer_overrides) + self.initial_state = self._optimizer.state_dict() + + def multiply_grads(self, c): + """Multiplies grads by a constant *c*.""" + self._optimizer.multiply_grads(c) + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm.""" + return self._optimizer.clip_grad_norm(max_norm, aggregate_norm_fn) + + def average_params(self): + self._optimizer.average_params() + + def _block_sync(self): + if self.world_size <= 1: + return + # Update the global model using local models from all GPUs + # (Step-1) Calculate grad between previously synced model and + # currrent local model + if self.block_momentum != 0: + self._calc_grad() + + # (Step-2) Average gradient from all GPUs + self._avg_grad_from_all_gpus() + + # (Step-3) Calculate global momentum and update the global model + if self.block_momentum != 0: + self._update_global_model() + + # (Step-4) Average local optimizer params + if self.average_sync: + self.average_params() + + def _is_warmup_end(self): + # Check whether train iterations is equal to warmup iter + if self.get_num_updates() == self.warmup_iteration: + return True + return False + + def _is_bmuf_iter(self): + # Check whether train iterations is equal to bmuf sync iter + if (self.get_num_updates() > self.warmup_iteration) and ( + self.get_num_updates() % self.sync_iter == 0 + ): + return True + return False + + def _warmup_sync(self, root_rank=0): + if self.world_size <= 1: + return + # Broadcast the local model to all gpus + for param in self.params: + dist.broadcast(param.data, src=root_rank) + + # Update local optimizer state + if self.average_sync: + self._optimizer.average_params() + else: + self._optimizer.load_state_dict(self.initial_state) + + self._reset_local_data() + + def step(self, closure=None): + """Performs a single optimization step.""" + self._optimizer.step(closure) + self.set_num_updates(self.get_num_updates() + 1) + if self._is_warmup_end(): + self._warmup_sync() + elif self._is_bmuf_iter(): + self._block_sync() + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + self._optimizer.zero_grad() + + def get_num_updates(self): + """Get the number of parameters updates.""" + return self._num_updates + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + self._num_updates = num_updates + + @torch.no_grad() + def _reset_local_data(self): + # (Step-0) Initialize global momentum parameters and store global copy on each gpu + self.global_params = [torch.zeros_like(p.data) for p in self.params] + self.smoothed_grads = [p.data.new_zeros(p.data.size()) for p in self.params] + self.grads = [p.data.new_zeros(p.data.size()) for p in self.params] + + # saving the global model locally for calculating gradient during bmuf sync + for param, global_param in zip(self.params, self.global_params): + global_param.copy_(param.data) + + @torch.no_grad() + def _calc_grad(self): + # global_params is basically the global copy from the previously finished + # synchronisation. param.data is local parameter after block_sync_freq + # for the local gpu. so grad is difference between previously synced + # model and currrent local model. + for index, (param, global_param) in enumerate( + zip(self.params, self.global_params) + ): + self.grads[index] = global_param - param.data + + def _avg_grad_from_all_gpus(self): + for index, param in enumerate(self.params): + sync_para = param.data if self.block_momentum == 0 else self.grads[index] + sync_para /= float(dist.get_world_size()) + dist.all_reduce(sync_para, op=dist.ReduceOp.SUM) + + @torch.no_grad() + def _update_global_model(self): + for index, (param, global_param, smoothed_grad, grad) in enumerate( + zip( + self.params, + self.global_params, + self.smoothed_grads, + # all gpus would share the same value of smoothed_grad, since it is + # always computed on synchronized gradients. + self.grads, + ) + ): + # global_param is basically last syncrhornized parameter. though + # smoothed_grad is local, all processes will have same value of + # smoothed_grad and hence param is globally synchronized copy. + # smoothed_grad(t) = BM * smoothed_grad(t-1) + BM_lr * grad(t) + smoothed_grad = self.block_momentum * smoothed_grad + self.block_lr * grad + param.data.copy_(global_param - smoothed_grad) + + # A Nesterov momentum here is to do a partial weight update before + # calculating the gradient + if self.use_nbm: + param.data.copy_(param.data - self.block_momentum * smoothed_grad) + + # backup for the next synchronization. + self.smoothed_grads[index] = smoothed_grad + global_param.copy_(param.data) diff --git a/fairseq/optim/dynamic_loss_scaler.py b/fairseq/optim/dynamic_loss_scaler.py new file mode 100644 index 0000000000000000000000000000000000000000..9d1f0b2c050f03e8a30abecf03b2cece75e5dc27 --- /dev/null +++ b/fairseq/optim/dynamic_loss_scaler.py @@ -0,0 +1,63 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +class DynamicLossScaler(object): + + def __init__( + self, init_scale=2.**15, scale_factor=2., scale_window=2000, + tolerance=0.05, threshold=None, min_loss_scale=1e-4 + ): + self.loss_scale = init_scale + self.scale_factor = scale_factor + self.scale_window = scale_window + self.tolerance = tolerance + self.threshold = threshold + self._iter = 0 + self._last_overflow_iter = -1 + self._last_rescale_iter = -1 + self._overflows_since_rescale = 0 + self.min_loss_scale = min_loss_scale + + def scale(self, outputs): + return self.loss_scale * outputs + + def update(self): + if (self._iter - self._last_overflow_iter) % self.scale_window == 0: + self.loss_scale *= self.scale_factor + self._last_rescale_iter = self._iter + self._iter += 1 + + def _decrease_loss_scale(self): + self.loss_scale /= self.scale_factor + if self.threshold is not None: + self.loss_scale = max(self.loss_scale, self.threshold) + + def check_overflow(self, grad_norm): + # detect inf and nan + if grad_norm == float('inf') or grad_norm != grad_norm: + # overflow has occured + prev_scale = self.loss_scale + iter_since_rescale = self._iter - self._last_rescale_iter + + self._last_overflow_iter = self._iter + self._overflows_since_rescale += 1 + pct_overflow = self._overflows_since_rescale / float(iter_since_rescale) + if pct_overflow >= self.tolerance: + self._decrease_loss_scale() + self._last_rescale_iter = self._iter + self._overflows_since_rescale = 0 + + if self.loss_scale <= self.min_loss_scale: + # Use FloatingPointError as an uncommon error that parent + # functions can safely catch to stop training. + self.loss_scale = prev_scale + raise FloatingPointError(( + 'Minimum loss scale reached ({}). Your loss is probably exploding. ' + 'Try lowering the learning rate, using gradient clipping or ' + 'increasing the batch size.' + ).format(self.min_loss_scale)) + + self._iter += 1 + raise OverflowError('setting loss scale to: ' + str(self.loss_scale)) diff --git a/fairseq/optim/fairseq_optimizer.py b/fairseq/optim/fairseq_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..b1b9c76edb3ce0ae172205c2c4cd96764b939831 --- /dev/null +++ b/fairseq/optim/fairseq_optimizer.py @@ -0,0 +1,133 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from fairseq import utils + + +class FairseqOptimizer(object): + + def __init__(self, args): + super().__init__() + self.args = args + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + pass + + @property + def optimizer(self): + """Return a torch.optim.optimizer.Optimizer instance.""" + if not hasattr(self, '_optimizer'): + raise NotImplementedError + if not isinstance(self._optimizer, torch.optim.Optimizer): + raise ValueError('_optimizer must be an instance of torch.optim.Optimizer') + return self._optimizer + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + raise NotImplementedError + + @property + def params(self): + """Return an iterable of the parameters held by the optimizer.""" + for param_group in self.param_groups: + for p in param_group['params']: + yield p + + @property + def param_groups(self): + return self.optimizer.param_groups + + def __getstate__(self): + return self._optimizer.__getstate__() + + def get_lr(self): + """Return the current learning rate.""" + return self.param_groups[0]['lr'] + + def set_lr(self, lr): + """Set the learning rate.""" + for param_group in self.param_groups: + param_group['lr'] = lr + + def state_dict(self): + """Return the optimizer's state dict.""" + return self.optimizer.state_dict() + + def load_state_dict(self, state_dict, optimizer_overrides=None): + """Load an optimizer state dict. + + In general we should prefer the configuration of the existing optimizer + instance (e.g., learning rate) over that found in the state_dict. This + allows us to resume training from a checkpoint using a new set of + optimizer args. + """ + self.optimizer.load_state_dict(state_dict) + + if optimizer_overrides is not None and len(optimizer_overrides) > 0: + # override learning rate, momentum, etc. with latest values + for group in self.param_groups: + group.update(optimizer_overrides) + + def backward(self, loss): + """Computes the sum of gradients of the given tensor w.r.t. graph leaves.""" + loss.backward() + + def multiply_grads(self, c): + """Multiplies grads by a constant *c*.""" + for p in self.params: + if p.grad is not None: + p.grad.data.mul_(c) + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm.""" + return utils.clip_grad_norm_(self.params, max_norm, aggregate_norm_fn) + + def step(self, closure=None, scale=1.): + """Performs a single optimization step.""" + if self.supports_step_with_scale: + self.optimizer.step(closure, scale=scale) + else: + self.optimizer.step(closure) + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + for p in self.params: + p.grad = None + self.optimizer.zero_grad() + + @property + def supports_memory_efficient_fp16(self): + if hasattr(self.optimizer, 'supports_memory_efficient_fp16'): + return self.optimizer.supports_memory_efficient_fp16 + return False + + @property + def supports_step_with_scale(self): + if hasattr(self.optimizer, 'supports_step_with_scale'): + return self.optimizer.supports_step_with_scale + return False + + @property + def supports_flat_params(self): + """ + Whether the optimizer supports collapsing of the model + parameters/gradients into a single contiguous Tensor. + """ + if hasattr(self.optimizer, 'supports_flat_params'): + return self.optimizer.supports_flat_params + return False + + def average_params(self): + pass diff --git a/fairseq/optim/fp16_optimizer.py b/fairseq/optim/fp16_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..37e94965bbaa2ae6501af080ade8df5d7787ab4d --- /dev/null +++ b/fairseq/optim/fp16_optimizer.py @@ -0,0 +1,413 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from itertools import chain + +import torch + +from fairseq import optim, utils + +from .dynamic_loss_scaler import DynamicLossScaler + + +class _FP16OptimizerMixin(object): + + def __init__(self, *args, **kwargs): + # forward __init__ call to the next class in mro(method resolution order) + super().__init__(*args, **kwargs) + + @property + def has_flat_params(self): + return torch.is_tensor(self.fp32_params) + + @classmethod + def build_fp32_params(cls, params, flatten=True): + # create FP32 copy of parameters and grads + if flatten: + total_param_size = sum(p.data.numel() for p in params) + fp32_params = torch.zeros(total_param_size, dtype=torch.float, device=params[0].device) + offset = 0 + for p in params: + numel = p.data.numel() + fp32_params[offset:offset+numel].copy_(p.data.view(-1)) + offset += numel + fp32_params = torch.nn.Parameter(fp32_params) + fp32_params.grad = fp32_params.data.new(total_param_size) + return fp32_params + else: + fp32_params = [] + for p in params: + p32 = torch.nn.Parameter(p.data.float()) + p32.grad = torch.zeros_like(p32.data) + fp32_params.append(p32) + return fp32_params + + def state_dict(self): + """Return the optimizer's state dict.""" + state_dict = self.fp32_optimizer.state_dict() + if self.scaler is not None: + state_dict['loss_scale'] = self.scaler.loss_scale + return state_dict + + def load_state_dict(self, state_dict, optimizer_overrides=None): + """Load an optimizer state dict. + + In general we should prefer the configuration of the existing optimizer + instance (e.g., learning rate) over that found in the state_dict. This + allows us to resume training from a checkpoint using a new set of + optimizer args. + """ + if 'loss_scale' in state_dict and self.scaler is not None: + self.scaler.loss_scale = state_dict['loss_scale'] + self.fp32_optimizer.load_state_dict(state_dict, optimizer_overrides) + + def backward(self, loss): + """Computes the sum of gradients of the given tensor w.r.t. graph leaves. + + Compared to :func:`fairseq.optim.FairseqOptimizer.backward`, this + function additionally dynamically scales the loss to avoid gradient + underflow. + """ + if self.scaler is not None: + loss = self.scaler.scale(loss) + loss.backward() + self._needs_sync = True + + def _sync_fp16_grads_to_fp32(self, multiply_grads=1.): + if self._needs_sync: + if self.scaler is not None: + # correct for dynamic loss scaler + multiply_grads /= self.scaler.loss_scale + + # copy FP16 grads to FP32 + if self.has_flat_params: + offset = 0 + for p in self.fp16_params: + if not p.requires_grad: + continue + grad_data = p.grad.data if p.grad is not None else p.data.new_zeros(p.data.shape) + numel = grad_data.numel() + self.fp32_params.grad.data[offset:offset+numel].copy_(grad_data.view(-1)) + offset += numel + self.fp32_params.grad.data.mul_(multiply_grads) + else: + for p, p32 in zip(self.fp16_params, self.fp32_params): + if not p.requires_grad: + continue + if p.grad is not None: + p32.grad.data.copy_(p.grad.data) + p32.grad.data.mul_(multiply_grads) + else: + p32.grad = torch.zeros_like(p.data, dtype=torch.float) + + self._needs_sync = False + + def _sync_fp32_grads_to_fp16(self): + # copy FP32 params back into FP16 model + if self.has_flat_params: + offset = 0 + for p in self.fp16_params: + if not p.requires_grad: + continue + numel = p.data.numel() + p.data.copy_(self.fp32_params.data[offset:offset+numel].view_as(p.data)) + offset += numel + else: + for p, p32 in zip(self.fp16_params, self.fp32_params): + if not p.requires_grad: + continue + p.data.copy_(p32.data) + + def multiply_grads(self, c): + """Multiplies grads by a constant ``c``.""" + if self._needs_sync: + self._sync_fp16_grads_to_fp32(c) + elif self.has_flat_params: + self.fp32_params.grad.data.mul_(c) + else: + for p32 in self.fp32_params: + p32.grad.data.mul_(c) + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm and updates dynamic loss scaler.""" + self._sync_fp16_grads_to_fp32() + grad_norm = utils.clip_grad_norm_(self.fp32_params, max_norm, aggregate_norm_fn) + + # detect overflow and adjust loss scale + if self.scaler is not None: + self.scaler.check_overflow(grad_norm) + + return grad_norm + + def step(self, closure=None): + """Performs a single optimization step.""" + self._sync_fp16_grads_to_fp32() + self.fp32_optimizer.step(closure) + + if self.scaler is not None: + self.scaler.update() + + self._sync_fp32_grads_to_fp16() + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + for p in self.fp16_params: + p.grad = None + if self.has_flat_params: + self.fp32_params.grad.zero_() + else: + for p32 in self.fp32_params: + p32.grad.zero_() + self._needs_sync = False + + +class FP16Optimizer(_FP16OptimizerMixin, optim.FairseqOptimizer): + """ + Wrap an *optimizer* to support FP16 (mixed precision) training. + """ + + def __init__(self, args, params, fp32_optimizer, fp32_params): + super().__init__(args) + self.fp16_params = params + self.fp32_optimizer = fp32_optimizer + self.fp32_params = fp32_params + + if getattr(args, 'fp16_scale_window', None) is None: + if len(args.update_freq) > 1: + raise ValueError( + '--fp16-scale-window must be given explicitly when using a ' + 'custom --update-freq schedule' + ) + data_parallel_size = int(args.distributed_world_size / args.model_parallel_size) + scale_window = int(2**14 / data_parallel_size / args.update_freq[0]) + else: + scale_window = args.fp16_scale_window + + if not getattr(args, 'bf16', False): + self.scaler = DynamicLossScaler( + init_scale=args.fp16_init_scale, + scale_window=scale_window, + tolerance=args.fp16_scale_tolerance, + threshold=args.threshold_loss_scale, + min_loss_scale=args.min_loss_scale + ) + else: + # disable loss scaling for bfloat16 + self.scaler = None + + @classmethod + def build_optimizer(cls, args, params): + """ + Args: + args (argparse.Namespace): fairseq args + params (iterable): iterable of parameters to optimize + """ + flatten = not getattr(args, 'fp16_no_flatten_grads', False) + if getattr(args, 'bf16', False): + flatten = False # mixed precision is faster on TPUs without flat grads + fp32_params = cls.build_fp32_params(params, flatten=flatten) + if flatten: + fp32_optimizer = optim.build_optimizer(args, [fp32_params]) + else: + fp32_optimizer = optim.build_optimizer(args, fp32_params) + if flatten and not fp32_optimizer.supports_flat_params: + raise RuntimeError( + 'chosen optimizer does not support flat params, ' + 'please set --fp16-no-flatten-grads' + ) + return cls(args, params, fp32_optimizer, fp32_params) + + @property + def optimizer(self): + return self.fp32_optimizer.optimizer + + @property + def optimizer_config(self): + return self.fp32_optimizer.optimizer_config + + def get_lr(self): + return self.fp32_optimizer.get_lr() + + def set_lr(self, lr): + self.fp32_optimizer.set_lr(lr) + + +class _MemoryEfficientFP16OptimizerMixin(object): + + def __init__(self, *args, **kwargs): + # forward __init__ call to the next class in MRO (method resolution order) + super().__init__(*args, **kwargs) + + @property + def has_flat_params(self): + return False + + def state_dict(self): + """Return the optimizer's state dict.""" + state_dict = self.wrapped_optimizer.state_dict() + if self.scaler is not None: + state_dict['loss_scale'] = self.scaler.loss_scale + return state_dict + + def load_state_dict(self, state_dict, optimizer_overrides=None): + """Load an optimizer state dict. + + In general we should prefer the configuration of the existing optimizer + instance (e.g., learning rate) over that found in the state_dict. This + allows us to resume training from a checkpoint using a new set of + optimizer args. + """ + if 'loss_scale' in state_dict and self.scaler is not None: + self.scaler.loss_scale = state_dict['loss_scale'] + + self.wrapped_optimizer.load_state_dict(state_dict, optimizer_overrides) + + # Hack: PyTorch automatically casts the optimizer state to match the + # type of the current parameters. But with --memory-efficient-fp16 the + # params are FP16 while the optimizer state is FP32 and we don't want + # to cast. A workaround is to manually copy back the original state + # after the optimizer has been loaded. + groups = self.optimizer.param_groups + saved_groups = state_dict['param_groups'] + id_map = { + old_id: p + for old_id, p in zip( + chain(*(g['params'] for g in saved_groups)), + chain(*(g['params'] for g in groups)) + ) + } + for k, v in state_dict['state'].items(): + if k in id_map: + param = id_map[k] + self.optimizer.state[param] = v + + def backward(self, loss): + """Computes the sum of gradients of the given tensor w.r.t. graph leaves. + + Compared to :func:`fairseq.optim.FairseqOptimizer.backward`, this + function additionally dynamically scales the loss to avoid gradient + underflow. + """ + if self.scaler is not None: + loss = self.scaler.scale(loss) + loss.backward() + + def _unscale_grads(self): + if self._multiply_factor != 1.: + self.wrapped_optimizer.multiply_grads(self._multiply_factor) + self._multiply_factor = 1. + + def multiply_grads(self, c): + """Multiplies grads by a constant *c*.""" + self._multiply_factor *= c + + def clip_grad_norm(self, max_norm, aggregate_norm_fn=None): + """Clips gradient norm and updates dynamic loss scaler.""" + max_norm = float(max_norm) + grad_norm = self._multiply_factor * self.wrapped_optimizer.clip_grad_norm(0, aggregate_norm_fn) + + if self.scaler is not None: + grad_norm_cpu = float(grad_norm) + if grad_norm_cpu > max_norm > 0.: + self._multiply_factor *= max_norm / grad_norm_cpu + + # detect overflow and adjust loss scale + self.scaler.check_overflow(grad_norm_cpu) + else: + clip_coef = (max_norm / (grad_norm + 1e-6)).clamp_(max=1) + self._multiply_factor *= clip_coef + + return grad_norm + + def step(self, closure=None): + """Performs a single optimization step.""" + if self.supports_step_with_scale: + # NOTE(msb) optimizer divides by scale factor + self.wrapped_optimizer.step(closure, scale=(1. / self._multiply_factor)) + else: + self._unscale_grads() + self.wrapped_optimizer.step(closure) + + if self.scaler is not None: + self.scaler.update() + + def zero_grad(self): + """Clears the gradients of all optimized parameters.""" + self.wrapped_optimizer.zero_grad() + if self.scaler is not None: + self._multiply_factor = 1. / float(self.scaler.loss_scale) + + +class MemoryEfficientFP16Optimizer(_MemoryEfficientFP16OptimizerMixin, optim.FairseqOptimizer): + """ + Wrap an *optimizer* to support FP16 (mixed precision) training. + + Compared to :class:`fairseq.optim.FP16Optimizer`, this version does not + maintain an FP32 copy of the model. We instead expect the optimizer to + convert the gradients to FP32 internally and sync the results back to the + FP16 model params. This significantly reduces memory usage but slightly + increases the time spent in the optimizer. + + Since this wrapper depends on specific functionality in the wrapped + optimizer (i.e., on-the-fly conversion of grads to FP32), only certain + optimizers can be wrapped. This is determined by the + *supports_memory_efficient_fp16* property. + """ + + def __init__(self, args, params, optimizer): + if not optimizer.supports_memory_efficient_fp16: + raise ValueError( + 'Unsupported optimizer: {}'.format(optimizer.__class__.__name__) + ) + + super().__init__(args) + self.wrapped_optimizer = optimizer + + if getattr(args, 'fp16_scale_window', None) is None: + if len(args.update_freq) > 1: + raise ValueError( + '--fp16-scale-window must be given explicitly when using a ' + 'custom --update-freq schedule' + ) + data_parallel_size = int(args.distributed_world_size / args.model_parallel_size) + scale_window = 2**14 / data_parallel_size / args.update_freq[0] + else: + scale_window = args.fp16_scale_window + + if not getattr(args, 'bf16', False): + self.scaler = DynamicLossScaler( + init_scale=args.fp16_init_scale, + scale_window=scale_window, + tolerance=args.fp16_scale_tolerance, + threshold=args.threshold_loss_scale, + min_loss_scale=args.min_loss_scale + ) + else: + # disable loss scaling for bfloat16 + self.scaler = None + + @classmethod + def build_optimizer(cls, args, params): + """ + Args: + args (argparse.Namespace): fairseq args + params (iterable): iterable of parameters to optimize + """ + fp16_optimizer = optim.build_optimizer(args, params) + return cls(args, params, fp16_optimizer) + + @property + def optimizer(self): + return self.wrapped_optimizer.optimizer + + @property + def optimizer_config(self): + return self.wrapped_optimizer.optimizer_config + + def get_lr(self): + return self.wrapped_optimizer.get_lr() + + def set_lr(self, lr): + self.wrapped_optimizer.set_lr(lr) diff --git a/fairseq/optim/fused_adam.py b/fairseq/optim/fused_adam.py new file mode 100644 index 0000000000000000000000000000000000000000..9024451aff6df740d699001f5f7c4ca4e2cf3111 --- /dev/null +++ b/fairseq/optim/fused_adam.py @@ -0,0 +1,312 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import types + +import torch + + +def get_fused_adam_class(): + """ + Look for the FusedAdam optimizer from apex. We first try to load the + "contrib" interface, which is a bit faster than the main interface, + but is technically deprecated. + """ + try: + # The "deprecated" interface in recent versions of apex is a bit + # faster than the main interface, since we don't use the apex + # optimizer. This can be installed by passing the + # `--deprecated_fused_adam` option when building apex. + global fused_adam_cuda + import importlib + fused_adam_cuda = importlib.import_module("fused_adam_cuda") + return FusedAdamV1 + except ImportError: + try: + # fallback to the newer interface + from apex.optimizers import FusedAdam as _FusedAdam # noqa + from apex.multi_tensor_apply import multi_tensor_applier + if multi_tensor_applier.available: + return FusedAdamV2 + except ImportError: + pass + return None + + +class FusedAdamV1(torch.optim.Optimizer): + """ + Implements Adam algorithm. Currently GPU-only. Requires Apex to be installed via + ``python setup.py install --cuda_ext --cpp_ext``. + + It has been proposed in `Adam: A Method for Stochastic Optimization`_. + + Compared to the original version in Apex, the fairseq version casts grads + and params to FP32 internally to support ``--memory-efficient-fp16``. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups. + lr (float, optional): learning rate. (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square. (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability. (default: 1e-8) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + amsgrad (boolean, optional): whether to use the AMSGrad variant of this + algorithm from the paper `On the Convergence of Adam and Beyond`_ + (default: False) NOT SUPPORTED in FusedAdam! + eps_inside_sqrt (boolean, optional): in the 'update parameters' step, + adds eps to the bias-corrected second moment estimate before + evaluating square root instead of adding it to the square root of + second moment estimate as in the original paper. (default: False) + .. _Adam: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__(self, params, + lr=1e-3, bias_correction=True, + betas=(0.9, 0.999), eps=1e-8, eps_inside_sqrt=False, + weight_decay=0., max_grad_norm=0., amsgrad=False): + global fused_adam_cuda + import importlib + fused_adam_cuda = importlib.import_module("fused_adam_cuda") + + if amsgrad: + raise RuntimeError('FusedAdam does not support the AMSGrad variant.') + defaults = { + 'lr': lr, + 'bias_correction': bias_correction, + 'betas': betas, + 'eps': eps, + 'weight_decay': weight_decay, + 'max_grad_norm': max_grad_norm, + } + super().__init__(params, defaults) + self.eps_mode = 0 if eps_inside_sqrt else 1 + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + @property + def supports_step_with_scale(self): + return True + + def step(self, closure=None, grads=None, scale=1., grad_norms=None): + """Performs a single optimization step. + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + grads (list of tensors, optional): weight gradient to use for the + optimizer update. If gradients have type torch.half, parameters + are expected to be in type torch.float. (default: None) + output params (list of tensors, optional): A reduced precision copy + of the updated weights written out in addition to the regular + updated weights. Have to be of same type as gradients. (default: None) + scale (float, optional): factor to divide gradient tensor values + by before applying to weights. (default: 1) + """ + loss = None + if closure is not None: + loss = closure() + + if grads is None: + grads_group = [None] * len(self.param_groups) + # backward compatibility + # assuming a list/generator of parameter means single group + elif isinstance(grads, types.GeneratorType): + grads_group = [grads] + elif type(grads[0]) != list: + grads_group = [grads] + else: + grads_group = grads + + if grad_norms is None: + grad_norms = [None]*len(self.param_groups) + + for group, grads_this_group, grad_norm in zip(self.param_groups, grads_group, grad_norms): + if grads_this_group is None: + grads_this_group = [None]*len(group['params']) + + # compute combined scale factor for this group + combined_scale = scale + if group.get('max_grad_norm', 0) > 0: + # norm is in fact norm*scale + clip = ((grad_norm / scale) + 1e-6) / group['max_grad_norm'] + if clip > 1: + combined_scale = clip * scale + + bias_correction = 1 if group.get('bias_correction', 1) else 0 + + for p, grad in zip(group['params'], grads_this_group): + # note: p.grad should not ever be set for correct + # operation of mixed precision optimizer that sometimes + # sends None gradients + if p.grad is None and grad is None: + continue + if grad is None: + grad = p.grad.data + if grad.is_sparse: + raise RuntimeError( + 'FusedAdam does not support sparse gradients, ' + 'please consider SparseAdam instead' + ) + + p_data_fp32 = p.data.float() + + state = self.state[p] + + # State initialization + if len(state) == 0: + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p_data_fp32) + # Exponential moving average of squared gradient values + state['exp_avg_sq'] = torch.zeros_like(p_data_fp32) + else: + state['exp_avg'] = state['exp_avg'].to(p_data_fp32) + state['exp_avg_sq'] = state['exp_avg_sq'].to(p_data_fp32) + + exp_avg = state['exp_avg'] + exp_avg_sq = state['exp_avg_sq'] + beta1, beta2 = group['betas'] + + state['step'] += 1 + + out_p = p.data + with torch.cuda.device(p.device): + fused_adam_cuda.adam(p_data_fp32, + out_p, + exp_avg, + exp_avg_sq, + grad, + group['lr'], + beta1, + beta2, + group['eps'], + combined_scale, + state['step'], + self.eps_mode, + bias_correction, + group['weight_decay']) + + return loss + + +try: + from apex.optimizers import FusedAdam + from apex.multi_tensor_apply import multi_tensor_applier + + class FusedAdamV2(FusedAdam): + """ + Compared to the original version in Apex, the fairseq version casts grads + and params to FP32 internally to support ``--memory-efficient-fp16``. + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + if not hasattr(self, 'multi_tensor_adam'): + raise Exception('Apex installation is outdated. Please install an updated version of apex.') + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step(self, closure=None, grads=None, output_params=None, scale=None, grad_norms=None): + """Performs a single optimization step.""" + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + bias_correction = 1 if group['bias_correction'] else 0 + beta1, beta2 = group['betas'] + + # assume same step across group now to simplify things + # per parameter step can be easily support by making it tensor, or pass list into kernel + if 'step' in group: + group['step'] += 1 + else: + group['step'] = 1 + + # create lists for multi-tensor apply + g_16, p_16, orig_p_16, m_16, v_16 = [], [], [], [], [] + g_32, p_32, m_32, v_32 = [], [], [], [] + + for p in group['params']: + if p.grad is None: + continue + if p.grad.data.is_sparse: + raise RuntimeError( + 'FusedAdam does not support sparse gradients, ' + 'please consider SparseAdam instead' + ) + + state = self.state[p] + # State initialization + if len(state) == 0: + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p.data, dtype=torch.float) + # Exponential moving average of squared gradient values + state['exp_avg_sq'] = torch.zeros_like(p.data, dtype=torch.float) + else: + state['exp_avg'] = state['exp_avg'].to(device=p.data.device, dtype=torch.float) + state['exp_avg_sq'] = state['exp_avg_sq'].to(device=p.data.device, dtype=torch.float) + + if p.dtype == torch.float16: + g_16.append(p.grad.data.float()) + p_16.append(p.data.float()) + orig_p_16.append(p.data) + m_16.append(state['exp_avg']) + v_16.append(state['exp_avg_sq']) + elif p.dtype == torch.float32: + g_32.append(p.grad.data) + p_32.append(p.data) + m_32.append(state['exp_avg']) + v_32.append(state['exp_avg_sq']) + else: + raise RuntimeError('FusedAdam only support fp16 and fp32.') + + with torch.cuda.device(p.device): + if(len(g_16) > 0): + multi_tensor_applier(self.multi_tensor_adam, + self._dummy_overflow_buf, + [g_16, p_16, m_16, v_16], + group['lr'], + beta1, + beta2, + group['eps'], + group['step'], + self.adam_w_mode, + bias_correction, + group['weight_decay']) + for orig_p, p in zip(orig_p_16, p_16): + orig_p.copy_(p.data) + if(len(g_32) > 0): + multi_tensor_applier(self.multi_tensor_adam, + self._dummy_overflow_buf, + [g_32, p_32, m_32, v_32], + group['lr'], + beta1, + beta2, + group['eps'], + group['step'], + self.adam_w_mode, + bias_correction, + group['weight_decay']) + + return loss +except ImportError: + pass diff --git a/fairseq/optim/fused_lamb.py b/fairseq/optim/fused_lamb.py new file mode 100644 index 0000000000000000000000000000000000000000..f9b0409c5333a9ca2bf2fe70b3d222a57dcdd0cc --- /dev/null +++ b/fairseq/optim/fused_lamb.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.optim import FairseqOptimizer, register_optimizer + + +@register_optimizer('lamb') +class FairseqLAMB(FairseqOptimizer): + """LAMB optimizer.""" + + def __init__(self, args, params): + super().__init__(args) + try: + from apex.optimizers import FusedLAMB + self._optimizer = FusedLAMB(params, **self.optimizer_config) + except ImportError: + raise ImportError('Please install apex to use LAMB optimizer') + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--lamb-betas', default='(0.9, 0.999)', metavar='B', + help='betas for LAMB optimizer') + parser.add_argument('--lamb-eps', type=float, default=1e-8, metavar='D', + help='epsilon for LAMB optimizer') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + 'lr': self.args.lr[0], + 'betas': eval(self.args.lamb_betas), + 'eps': self.args.lamb_eps, + 'weight_decay': self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return False diff --git a/fairseq/optim/lr_scheduler/__init__.py b/fairseq/optim/lr_scheduler/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..edd0a6a13e1a10e91a8653371a420b36bcb2cc27 --- /dev/null +++ b/fairseq/optim/lr_scheduler/__init__.py @@ -0,0 +1,23 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + +from fairseq import registry +from fairseq.optim.lr_scheduler.fairseq_lr_scheduler import FairseqLRScheduler + + +build_lr_scheduler, register_lr_scheduler, LR_SCHEDULER_REGISTRY = registry.setup_registry( + '--lr-scheduler', + base_class=FairseqLRScheduler, + default='fixed', +) + +# automatically import any Python files in the optim/lr_scheduler/ directory +for file in os.listdir(os.path.dirname(__file__)): + if file.endswith('.py') and not file.startswith('_'): + module = file[:file.find('.py')] + importlib.import_module('fairseq.optim.lr_scheduler.' + module) diff --git a/fairseq/optim/lr_scheduler/__pycache__/__init__.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5ff0cd34b65ec35fdfdc0923c5d1f7a2ebc660cb Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/cosine_lr_scheduler.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/cosine_lr_scheduler.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dcc046b845c58af19d025431930033c71d18cc71 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/cosine_lr_scheduler.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/fairseq_lr_scheduler.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/fairseq_lr_scheduler.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2eca6a51b616673c8527fce24e9a6ca6a7290949 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/fairseq_lr_scheduler.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/fixed_schedule.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/fixed_schedule.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..19452785fd577d1f2eb88af6996e5f3d50610726 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/fixed_schedule.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/inverse_square_root_schedule.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/inverse_square_root_schedule.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b741628a674fb0b26fc5995c1710d23db3d40777 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/inverse_square_root_schedule.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/polynomial_decay_schedule.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/polynomial_decay_schedule.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f412212e2f58dc6ce71a25075536f00620a56d03 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/polynomial_decay_schedule.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/reduce_lr_on_plateau.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/reduce_lr_on_plateau.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1ec8a3c4feab9a7a753dafa6fba812fa68def743 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/reduce_lr_on_plateau.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/tri_stage_lr_scheduler.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/tri_stage_lr_scheduler.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..de654e1b6720ecfa8dab1cd4e1c7d7dea82e36c7 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/tri_stage_lr_scheduler.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/__pycache__/triangular_lr_scheduler.cpython-310.pyc b/fairseq/optim/lr_scheduler/__pycache__/triangular_lr_scheduler.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a2c1f885567c9d3dacd33756cd919a3e4cfd98d9 Binary files /dev/null and b/fairseq/optim/lr_scheduler/__pycache__/triangular_lr_scheduler.cpython-310.pyc differ diff --git a/fairseq/optim/lr_scheduler/cosine_lr_scheduler.py b/fairseq/optim/lr_scheduler/cosine_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..9137e11b786a860cd6ab9fdcf35523c088989781 --- /dev/null +++ b/fairseq/optim/lr_scheduler/cosine_lr_scheduler.py @@ -0,0 +1,118 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from . import FairseqLRScheduler, register_lr_scheduler + + +@register_lr_scheduler('cosine') +class CosineSchedule(FairseqLRScheduler): + """Assign LR based on a cyclical schedule that follows the cosine function. + + See https://arxiv.org/pdf/1608.03983.pdf for details. + + We also support a warmup phase where we linearly increase the learning rate + from some initial learning rate (``--warmup-init-lr``) until the configured + max learning rate (``--max-lr``). + + During warmup:: + + lrs = torch.linspace(args.warmup_init_lr, args.lr, args.warmup_updates) + lr = lrs[update_num] + + After warmup:: + + lr = lr_min + 0.5*(lr_max - lr_min)*(1 + cos(t_curr / t_i)) + + where ``t_curr`` is current percentage of updates within the current period + range and ``t_i`` is the current period range, which is scaled by ``t_mul`` + after every iteration. + """ + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + if len(args.lr) > 1: + raise ValueError( + 'Cannot use a fixed learning rate schedule with cosine.' + ' Consider --lr-scheduler=fixed instead.' + ) + + warmup_end_lr = args.max_lr + if args.warmup_init_lr < 0: + args.warmup_init_lr = args.lr[0] + + self.min_lr = args.lr[0] + self.max_lr = args.max_lr + + assert self.max_lr > self.min_lr, 'max_lr must be more than lr' + + self.t_mult = args.t_mult + self.period = args.lr_period_updates + + if self.period <= 0: + assert args.max_update >= 0, 'Either --max_update or --lr-period-updates must be set' + self.period = args.max_update - args.warmup_updates + + if args.warmup_updates > 0: + # linearly warmup for the first args.warmup_updates + self.lr_step = (warmup_end_lr - args.warmup_init_lr) / args.warmup_updates + else: + self.lr_step = 1 + + self.warmup_updates = args.warmup_updates + self.lr_shrink = args.lr_shrink + + # initial learning rate + self.lr = args.warmup_init_lr + self.optimizer.set_lr(self.lr) + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + # fmt: off + parser.add_argument('--warmup-updates', default=0, type=int, metavar='N', + help='warmup the learning rate linearly for the first N updates') + parser.add_argument('--warmup-init-lr', default=-1, type=float, metavar='LR', + help='initial learning rate during warmup phase; default is args.lr') + parser.add_argument('--max-lr', type=float, metavar='LR', + help='max learning rate, must be more than args.lr') + parser.add_argument('--t-mult', default=1, type=float, metavar='LR', + help='factor to grow the length of each period') + parser.add_argument('--lr-period-updates', default=-1, type=float, metavar='LR', + help='initial number of updates per period') + parser.add_argument('--lr-shrink', default=0.1, type=float, metavar='LS', + help='shrink factor for annealing') + # fmt: on + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if num_updates < self.args.warmup_updates: + self.lr = self.args.warmup_init_lr + num_updates * self.lr_step + else: + curr_updates = num_updates - self.args.warmup_updates + if self.t_mult != 1: + i = math.floor(math.log(1 - curr_updates / self.period * (1 - self.t_mult), self.t_mult)) + t_i = self.t_mult ** i * self.period + t_curr = curr_updates - (1 - self.t_mult ** i) / (1 - self.t_mult) * self.period + else: + i = math.floor(curr_updates / self.period) + t_i = self.period + t_curr = curr_updates - (self.period * i) + + lr_shrink = self.lr_shrink ** i + min_lr = self.min_lr * lr_shrink + max_lr = self.max_lr * lr_shrink + + self.lr = min_lr + 0.5 * (max_lr - min_lr) * (1 + math.cos(math.pi * t_curr / t_i)) + + self.optimizer.set_lr(self.lr) + return self.lr diff --git a/fairseq/optim/lr_scheduler/fairseq_lr_scheduler.py b/fairseq/optim/lr_scheduler/fairseq_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..8b7884829a6311deea1f1160b452791f1485d4d3 --- /dev/null +++ b/fairseq/optim/lr_scheduler/fairseq_lr_scheduler.py @@ -0,0 +1,42 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .. import FairseqOptimizer + + +class FairseqLRScheduler(object): + + def __init__(self, args, optimizer): + super().__init__() + if not isinstance(optimizer, FairseqOptimizer): + raise ValueError('optimizer must be an instance of FairseqOptimizer') + self.args = args + self.optimizer = optimizer + self.best = None + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + pass + + def state_dict(self): + """Return the LR scheduler state dict.""" + return {'best': self.best} + + def load_state_dict(self, state_dict): + """Load an LR scheduler state dict.""" + self.best = state_dict['best'] + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + if val_loss is not None: + if self.best is None: + self.best = val_loss + else: + self.best = min(self.best, val_loss) + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/fixed_schedule.py b/fairseq/optim/lr_scheduler/fixed_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..cc10db16388e1befc893c1cd7c496c5e7da4892e --- /dev/null +++ b/fairseq/optim/lr_scheduler/fixed_schedule.py @@ -0,0 +1,61 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import FairseqLRScheduler, register_lr_scheduler + + +@register_lr_scheduler('fixed') +class FixedSchedule(FairseqLRScheduler): + """Decay the LR on a fixed schedule.""" + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + + # set defaults + args.warmup_updates = getattr(args, 'warmup_updates', 0) or 0 + + self.lr = args.lr[0] + if args.warmup_updates > 0: + self.warmup_factor = 1. / args.warmup_updates + else: + self.warmup_factor = 1 + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + # fmt: off + parser.add_argument('--force-anneal', '--fa', type=int, metavar='N', + help='force annealing at specified epoch') + parser.add_argument('--lr-shrink', default=0.1, type=float, metavar='LS', + help='shrink factor for annealing, lr_new = (lr * lr_shrink)') + parser.add_argument('--warmup-updates', default=0, type=int, metavar='N', + help='warmup the learning rate linearly for the first N updates') + # fmt: on + + def get_next_lr(self, epoch): + lrs = self.args.lr + if self.args.force_anneal is None or epoch < self.args.force_anneal: + # use fixed LR schedule + next_lr = lrs[min(epoch, len(lrs) - 1)] + else: + # annneal based on lr_shrink + next_lr = lrs[-1] * self.args.lr_shrink ** (epoch + 1 - self.args.force_anneal) + return next_lr + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + self.lr = self.get_next_lr(epoch) + self.optimizer.set_lr(self.warmup_factor * self.lr) + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if self.args.warmup_updates > 0 and num_updates < self.args.warmup_updates: + self.warmup_factor = (num_updates + 1) / float(self.args.warmup_updates) + self.optimizer.set_lr(self.warmup_factor * self.lr) + else: + self.optimizer.set_lr(self.lr) + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/inverse_square_root_schedule.py b/fairseq/optim/lr_scheduler/inverse_square_root_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..f98a7c3b997e2aa2f77911da25686e9bdccbad5b --- /dev/null +++ b/fairseq/optim/lr_scheduler/inverse_square_root_schedule.py @@ -0,0 +1,73 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import FairseqLRScheduler, register_lr_scheduler + + +@register_lr_scheduler('inverse_sqrt') +class InverseSquareRootSchedule(FairseqLRScheduler): + """Decay the LR based on the inverse square root of the update number. + + We also support a warmup phase where we linearly increase the learning rate + from some initial learning rate (``--warmup-init-lr``) until the configured + learning rate (``--lr``). Thereafter we decay proportional to the number of + updates, with a decay factor set to align with the configured learning rate. + + During warmup:: + + lrs = torch.linspace(args.warmup_init_lr, args.lr, args.warmup_updates) + lr = lrs[update_num] + + After warmup:: + + decay_factor = args.lr * sqrt(args.warmup_updates) + lr = decay_factor / sqrt(update_num) + """ + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + if len(args.lr) > 1: + raise ValueError( + 'Cannot use a fixed learning rate schedule with inverse_sqrt.' + ' Consider --lr-scheduler=fixed instead.' + ) + warmup_end_lr = args.lr[0] + if args.warmup_init_lr < 0: + args.warmup_init_lr = 0 if args.warmup_updates > 0 else warmup_end_lr + + # linearly warmup for the first args.warmup_updates + self.lr_step = (warmup_end_lr - args.warmup_init_lr) / args.warmup_updates + + # then, decay prop. to the inverse square root of the update number + self.decay_factor = warmup_end_lr * args.warmup_updates**0.5 + + # initial learning rate + self.lr = args.warmup_init_lr + self.optimizer.set_lr(self.lr) + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + # fmt: off + parser.add_argument('--warmup-updates', default=4000, type=int, metavar='N', + help='warmup the learning rate linearly for the first N updates') + parser.add_argument('--warmup-init-lr', default=-1, type=float, metavar='LR', + help='initial learning rate during warmup phase; default is args.lr') + # fmt: on + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if num_updates < self.args.warmup_updates: + self.lr = self.args.warmup_init_lr + num_updates*self.lr_step + else: + self.lr = self.decay_factor * num_updates**-0.5 + self.optimizer.set_lr(self.lr) + return self.lr diff --git a/fairseq/optim/lr_scheduler/polynomial_decay_schedule.py b/fairseq/optim/lr_scheduler/polynomial_decay_schedule.py new file mode 100644 index 0000000000000000000000000000000000000000..aff57f9b93d0221c3685467be50cde258e965327 --- /dev/null +++ b/fairseq/optim/lr_scheduler/polynomial_decay_schedule.py @@ -0,0 +1,70 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import FairseqLRScheduler, register_lr_scheduler + + +@register_lr_scheduler('polynomial_decay') +class PolynomialDecaySchedule(FairseqLRScheduler): + """Decay the LR on a fixed schedule.""" + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + + # set defaults + args.warmup_updates = getattr(args, 'warmup_updates', 0) or 0 + + self.lr = args.lr[0] + if args.warmup_updates > 0: + self.warmup_factor = 1. / args.warmup_updates + else: + self.warmup_factor = 1 + self.end_learning_rate = args.end_learning_rate + self.total_num_update = args.total_num_update + self.power = args.power + self.optimizer.set_lr(self.warmup_factor * self.lr) + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + parser.add_argument('--force-anneal', '--fa', type=int, metavar='N', + help='force annealing at specified epoch') + parser.add_argument('--warmup-updates', default=0, type=int, metavar='N', + help='warmup the learning rate linearly for the first N updates') + parser.add_argument('--end-learning-rate', default=0.0, type=float) + parser.add_argument('--power', default=1.0, type=float) + parser.add_argument('--total-num-update', default=1000000, type=int) + + def get_next_lr(self, epoch): + lrs = self.args.lr + if self.args.force_anneal is None or epoch < self.args.force_anneal: + # use fixed LR schedule + next_lr = lrs[min(epoch, len(lrs) - 1)] + else: + # annneal based on lr_shrink + next_lr = self.optimizer.get_lr() + return next_lr + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + self.lr = self.get_next_lr(epoch) + self.optimizer.set_lr(self.warmup_factor * self.lr) + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + if self.args.warmup_updates > 0 and num_updates <= self.args.warmup_updates: + self.warmup_factor = num_updates / float(self.args.warmup_updates) + lr = self.warmup_factor * self.lr + elif num_updates >= self.total_num_update: + lr = self.end_learning_rate + else: + warmup = self.args.warmup_updates + lr_range = self.lr - self.end_learning_rate + pct_remaining = 1 - (num_updates - warmup) / (self.total_num_update - warmup) + lr = lr_range * pct_remaining ** (self.power) + self.end_learning_rate + self.optimizer.set_lr(lr) + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/reduce_lr_on_plateau.py b/fairseq/optim/lr_scheduler/reduce_lr_on_plateau.py new file mode 100644 index 0000000000000000000000000000000000000000..8128cf0eb81e24e7ac2c838ddc7bedd9feb5df65 --- /dev/null +++ b/fairseq/optim/lr_scheduler/reduce_lr_on_plateau.py @@ -0,0 +1,110 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.optim.lr_scheduler + +from . import FairseqLRScheduler, register_lr_scheduler + + +@register_lr_scheduler('reduce_lr_on_plateau') +class ReduceLROnPlateau(FairseqLRScheduler): + """ + Decay the LR by a factor every time the validation loss plateaus. + Also comes with optional warmup phase, where we linearly increase + the learning rate from some initial learning rate + (``--warmup-init-lr``) until the configured learning rate + (``--lr``). Thereafter the lr is adjusted according to original + reduce_on_plateau scheme. + + During warmup:: + + lrs = torch.linspace( + args.warmup_init_lr, args.lr, args.warmup_updates + ) + lr = lrs[update_num] + """ + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + if len(args.lr) > 1: + raise ValueError( + 'Cannot use a fixed learning rate schedule with reduce_lr_on_plateau.' + ' Consider --lr-scheduler=fixed instead.' + ) + self.lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( + self.optimizer.optimizer, patience=args.lr_patience, factor=args.lr_shrink, + mode='max' if args.maximize_best_checkpoint_metric else 'min', + threshold=args.lr_threshold) + warmup_end_lr = args.lr[0] + # if no warm up, sets initial lr to be args.lr[0] + if args.warmup_init_lr < 0: + args.warmup_init_lr = 0 if args.warmup_updates > 0 else warmup_end_lr + + # linearly warmup for the first args.warmup_updates + if args.warmup_updates > 0: + self.lr_step = (warmup_end_lr - args.warmup_init_lr) / args.warmup_updates + # this flag is either set from arg when no warm up, or set by + # step_update() when warmup finishes + self.warmup_end = True if args.warmup_updates <= 0 else False + # initial learning rate + # this self.lr is used only during init and/or warm up period + self.lr = args.warmup_init_lr + self.optimizer.set_lr(self.lr) + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + # fmt: off + parser.add_argument('--lr-shrink', default=0.1, type=float, metavar='LS', + help='shrink factor for annealing, lr_new = (lr * lr_shrink)') + parser.add_argument('--lr-threshold', default=1e-4, type=float, metavar='LT', + help='threshold for measuring the new optimum, ' + 'to only focus on significant changes') + parser.add_argument('--lr-patience', default=0, type=int, + help='number of epochs with no improvement after which ' + 'learning rate will be reduced') + parser.add_argument('--warmup-updates', default=0, type=int, metavar='N', + help='warmup the learning rate linearly for the first N updates') + parser.add_argument('--warmup-init-lr', default=-1, type=float, metavar='LR', + help='initial learning rate during warmup phase; default is args.lr') + # fmt: on + + def state_dict(self): + """Return the LR scheduler state dict.""" + return { + 'best': self.lr_scheduler.best, + 'last_epoch': self.lr_scheduler.last_epoch, + } + + def load_state_dict(self, state_dict): + """Load an LR scheduler state dict.""" + self.lr_scheduler.best = state_dict['best'] + if 'last_epoch' in state_dict: + self.lr_scheduler.last_epoch = state_dict['last_epoch'] + + def step(self, epoch, val_loss=None): + """ + Update the learning rate at the end of the given epoch if warmup + finishes otherwise no update of lr on epoch boundaries + """ + if val_loss is not None and self.warmup_end is True: + self.lr_scheduler.step(val_loss) + else: + self.lr_scheduler.last_epoch = epoch + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """ + Update the learning rate after each update.""" + # if there is warmup + if self.args.warmup_updates > 0: + if num_updates <= self.args.warmup_updates: + self.lr = self.args.warmup_init_lr + num_updates*self.lr_step + self.optimizer.set_lr(self.lr) + else: + if self.warmup_end is False: + self.warmup_end = True + # else do nothing + return self.optimizer.get_lr() diff --git a/fairseq/optim/lr_scheduler/tri_stage_lr_scheduler.py b/fairseq/optim/lr_scheduler/tri_stage_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..3460fa1226ed256750409a501a8ebb3a6c0806c2 --- /dev/null +++ b/fairseq/optim/lr_scheduler/tri_stage_lr_scheduler.py @@ -0,0 +1,163 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import FairseqLRScheduler, register_lr_scheduler +import math + + +@register_lr_scheduler('tri_stage') +class TriStageLRSchedule(FairseqLRScheduler): + """Tristage learning rate schedulr + + Implement the learning rate scheduler in https://arxiv.org/pdf/1904.08779.pdf + + Similar to inverse_squre_root scheduler, but tri_stage learning rate employs + three stages LR scheduling: + + - warmup stage, starting from `lr` * `init_lr_scale`, linearly + increased to `lr` in `warmup_steps` iterations + + - hold stage, after `warmup_steps`, keep the LR as `lr` for `hold_steps` + iterations + + - decay stage, after hold stage, decay LR exponetially to + `lr` * `final_lr_scale` in `decay_steps`; + after that LR is keep as `final_lr_scale` * `lr` + + During warmup:: + + init_lr = args.init_lr_scale * args.lr + lrs = torch.linspace(init_lr, args.lr, args.warmup_steps) + lr = lrs[update_num] + + During hold:: + + lr = args.lr + + During decay:: + + decay_factor = - math.log(args.final_lr_scale) / args.decay_steps + lr = args.lr * exp(- (update_num - warmup_steps - decay_steps) * decay_factor) + + After that:: + + lr = args.lr * args.final_lr_scale + """ + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + if len(args.lr) > 1: + raise ValueError( + 'Cannot use a fixed learning rate schedule with tri-stage lr.' + ' Consider --lr-scheduler=fixed instead.' + ) + + # calculate LR at each point + self.peak_lr = args.lr[0] + self.init_lr = args.init_lr_scale * args.lr[0] + self.final_lr = args.final_lr_scale * args.lr[0] + + # remember the steps at each stage + self.warmup_steps = args.warmup_steps + self.hold_steps = args.hold_steps + self.decay_steps = args.decay_steps + + self.warmup_rate = ( + (self.peak_lr - self.init_lr) / self.warmup_steps if self.warmup_steps != 0 + else 0 + ) + self.decay_factor = -math.log(args.final_lr_scale) / args.decay_steps + + # initial learning rate + self.lr = self.init_lr + self.optimizer.set_lr(self.lr) + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + # fmt: off + parser.add_argument( + '--warmup-steps', + default=4000, + type=int, + metavar='N', + help='warmup the learning rate linearly for the first N updates' + ) + parser.add_argument( + '--hold-steps', + default=20000, + type=int, + metavar='N', + help='steps in hold stage.' + ) + parser.add_argument( + '--decay-steps', + default=60000, + type=int, + metavar='N', + help='steps in decay stages' + ) + parser.add_argument( + '--init-lr-scale', + default=0.01, + type=float, + help=""" + initial learning rate scale during warmup phase; default is 0.01""") + parser.add_argument( + '--final-lr-scale', + default=0.01, + type=float, + help="final learning rate scale; default to 0.01" + ) + # fmt: on + + def _decide_stage(self, update_step): + """ + return stage, and the corresponding steps within the current stage + """ + if update_step < self.warmup_steps: + # warmup state + return 0, update_step + + offset = self.warmup_steps + + if update_step < offset + self.hold_steps: + # hold stage + return 1, update_step - offset + + offset += self.hold_steps + + if update_step <= offset + self.decay_steps: + # decay stage + return 2, update_step - offset + + offset += self.decay_steps + + # still here ? constant lr stage + return 3, update_step - offset + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + stage, steps_in_stage = self._decide_stage(num_updates) + if stage == 0: + self.lr = self.init_lr + self.warmup_rate * steps_in_stage + elif stage == 1: + self.lr = self.peak_lr + elif stage == 2: + self.lr = self.peak_lr * math.exp(-self.decay_factor * steps_in_stage) + elif stage == 3: + self.lr = self.final_lr + else: + raise ValueError("Undefined stage") + + self.optimizer.set_lr(self.lr) + + return self.lr diff --git a/fairseq/optim/lr_scheduler/triangular_lr_scheduler.py b/fairseq/optim/lr_scheduler/triangular_lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..fed0cf7ef119852c84825eaae89b35226991eedc --- /dev/null +++ b/fairseq/optim/lr_scheduler/triangular_lr_scheduler.py @@ -0,0 +1,74 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from . import FairseqLRScheduler, register_lr_scheduler + + +@register_lr_scheduler('triangular') +class TriangularSchedule(FairseqLRScheduler): + """Assign LR based on a triangular cyclical schedule. + + See https://arxiv.org/pdf/1506.01186.pdf for details. + """ + + def __init__(self, args, optimizer): + super().__init__(args, optimizer) + if len(args.lr) > 1: + raise ValueError( + 'Cannot use a fixed learning rate schedule with triangular.' + ' Consider --lr-scheduler=fixed instead.' + ) + + lr = args.lr[0] + + assert args.max_lr > lr, 'max_lr must be more than lr' + self.min_lr = lr + self.max_lr = args.max_lr + self.stepsize = args.lr_period_updates // 2 + self.lr_shrink = args.lr_shrink + self.shrink_min = args.shrink_min + + # initial learning rate + self.lr = self.min_lr + self.optimizer.set_lr(self.lr) + + @staticmethod + def add_args(parser): + """Add arguments to the parser for this LR scheduler.""" + # fmt: off + parser.add_argument('--max-lr', required=True, type=float, metavar='LR', + help='max learning rate, must be more than args.lr') + parser.add_argument('--lr-period-updates', default=5000, type=float, metavar='LR', + help='initial number of updates per period (cycle length)') + parser.add_argument('--lr-shrink', default=0.1, type=float, metavar='LS', + help='shrink factor for annealing') + parser.add_argument('--shrink-min', action='store_true', + help='if set, also shrinks min lr') + # fmt: on + + def step(self, epoch, val_loss=None): + """Update the learning rate at the end of the given epoch.""" + super().step(epoch, val_loss) + # we don't change the learning rate at epoch boundaries + return self.optimizer.get_lr() + + def step_update(self, num_updates): + """Update the learning rate after each update.""" + cycle = math.floor(num_updates / (2 * self.stepsize)) + + lr_shrink = self.lr_shrink ** cycle + max_lr = self.max_lr * lr_shrink + if self.shrink_min: + min_lr = self.min_lr * lr_shrink + else: + min_lr = self.min_lr + + x = abs(num_updates / self.stepsize - 2 * (cycle + 1) + 1) + self.lr = min_lr + (max_lr - min_lr) * max(0, (1 - x)) + + self.optimizer.set_lr(self.lr) + return self.lr diff --git a/fairseq/optim/nag.py b/fairseq/optim/nag.py new file mode 100644 index 0000000000000000000000000000000000000000..d9b7fb8019575a5cfa3522c8ebe2f9982a55bdd9 --- /dev/null +++ b/fairseq/optim/nag.py @@ -0,0 +1,103 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch.optim.optimizer import Optimizer, required + +from . import FairseqOptimizer, register_optimizer + + +@register_optimizer('nag') +class FairseqNAG(FairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = NAG(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--momentum', default=0.99, type=float, metavar='M', + help='momentum factor') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + 'lr': self.args.lr[0], + 'momentum': self.args.momentum, + 'weight_decay': self.args.weight_decay, + } + + +class NAG(Optimizer): + def __init__(self, params, lr=required, momentum=0, weight_decay=0): + defaults = dict(lr=lr, lr_old=lr, momentum=momentum, weight_decay=weight_decay) + super(NAG, self).__init__(params, defaults) + + @property + def supports_memory_efficient_fp16(self): + return True + + @property + def supports_flat_params(self): + return True + + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + weight_decay = group['weight_decay'] + momentum = group['momentum'] + lr = group['lr'] + lr_old = group.get('lr_old', lr) + lr_correct = lr / lr_old + + for p in group['params']: + if p.grad is None: + continue + + p_data_fp32 = p.data + if p_data_fp32.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + d_p = p.grad.data.float() + param_state = self.state[p] + if 'momentum_buffer' not in param_state: + param_state['momentum_buffer'] = torch.zeros_like(d_p) + else: + param_state['momentum_buffer'] = param_state['momentum_buffer'].to(d_p) + + buf = param_state['momentum_buffer'] + + if weight_decay != 0: + p_data_fp32.mul_(1 - lr * weight_decay) + p_data_fp32.add_(buf, alpha=momentum * momentum * lr_correct) + p_data_fp32.add_(d_p, alpha=-(1 + momentum) * lr) + + buf.mul_(momentum * lr_correct).add_(d_p, alpha=-lr) + + if p.data.dtype in {torch.float16, torch.bfloat16}: + p.data.copy_(p_data_fp32) + + group['lr_old'] = lr + + return loss diff --git a/fairseq/optim/sgd.py b/fairseq/optim/sgd.py new file mode 100644 index 0000000000000000000000000000000000000000..8c4e3e0a809f308941cbe28504197acb8a72d88d --- /dev/null +++ b/fairseq/optim/sgd.py @@ -0,0 +1,43 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.optim + +from . import FairseqOptimizer, register_optimizer + + +@register_optimizer('sgd') +class SGD(FairseqOptimizer): + def __init__(self, args, params): + super().__init__(args) + self._optimizer = torch.optim.SGD(params, **self.optimizer_config) + + @staticmethod + def add_args(parser): + """Add optimizer-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--momentum', default=0.0, type=float, metavar='M', + help='momentum factor') + parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', + help='weight decay') + # fmt: on + + @property + def optimizer_config(self): + """ + Return a kwarg dictionary that will be used to override optimizer + args stored in checkpoints. This allows us to load a checkpoint and + resume training using a different set of optimizer args, e.g., with a + different learning rate. + """ + return { + 'lr': self.args.lr[0], + 'momentum': self.args.momentum, + 'weight_decay': self.args.weight_decay, + } + + @property + def supports_flat_params(self): + return True diff --git a/fairseq/options.py b/fairseq/options.py new file mode 100644 index 0000000000000000000000000000000000000000..e889821ee6c7d483f517b8806a9e46705e203185 --- /dev/null +++ b/fairseq/options.py @@ -0,0 +1,675 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import sys +from typing import Callable, List, Optional + +import torch + +from fairseq import scoring, utils +from fairseq.data.indexed_dataset import get_available_dataset_impl + + +def get_preprocessing_parser(default_task="translation"): + parser = get_parser("Preprocessing", default_task) + add_preprocess_args(parser) + return parser + + +def get_training_parser(default_task="translation"): + parser = get_parser("Trainer", default_task) + add_dataset_args(parser, train=True) + add_distributed_training_args(parser) + add_model_args(parser) + add_optimization_args(parser) + add_checkpoint_args(parser) + return parser + + +def get_generation_parser(interactive=False, default_task="translation"): + parser = get_parser("Generation", default_task) + add_dataset_args(parser, gen=True) + add_distributed_training_args(parser, default_world_size=1) + add_generation_args(parser) + if interactive: + add_interactive_args(parser) + return parser + + +def get_interactive_generation_parser(default_task="translation"): + return get_generation_parser(interactive=True, default_task=default_task) + + +def get_eval_lm_parser(default_task="language_modeling"): + parser = get_parser("Evaluate Language Model", default_task) + add_dataset_args(parser, gen=True) + add_distributed_training_args(parser, default_world_size=1) + add_eval_lm_args(parser) + return parser + + +def get_validation_parser(default_task=None): + parser = get_parser("Validation", default_task) + add_dataset_args(parser, train=True) + add_distributed_training_args(parser, default_world_size=1) + group = parser.add_argument_group("Evaluation") + add_common_eval_args(group) + return parser + + +def csv_str_list(x): + return x.split(',') + + +def eval_str_list(x, type=float): + if x is None: + return None + if isinstance(x, str): + x = eval(x) + try: + return list(map(type, x)) + except TypeError: + return [type(x)] + + +def eval_str_dict(x, type=dict): + if x is None: + return None + if isinstance(x, str): + x = eval(x) + return x + + +def eval_bool(x, default=False): + if x is None: + return default + try: + return bool(eval(x)) + except TypeError: + return default + + +def parse_args_and_arch( + parser: argparse.ArgumentParser, + input_args: List[str] = None, + parse_known: bool = False, + suppress_defaults: bool = False, + modify_parser: Optional[Callable[[argparse.ArgumentParser], None]] = None, +): + """ + Args: + parser (ArgumentParser): the parser + input_args (List[str]): strings to parse, defaults to sys.argv + parse_known (bool): only parse known arguments, similar to + `ArgumentParser.parse_known_args` + suppress_defaults (bool): parse while ignoring all default values + modify_parser (Optional[Callable[[ArgumentParser], None]]): + function to modify the parser, e.g., to set default values + """ + if suppress_defaults: + # Parse args without any default values. This requires us to parse + # twice, once to identify all the necessary task/model args, and a second + # time with all defaults set to None. + args = parse_args_and_arch( + parser, + input_args=input_args, + parse_known=parse_known, + suppress_defaults=False, + ) + suppressed_parser = argparse.ArgumentParser(add_help=False, parents=[parser]) + suppressed_parser.set_defaults(**{k: None for k, v in vars(args).items()}) + args = suppressed_parser.parse_args(input_args) + return argparse.Namespace( + **{k: v for k, v in vars(args).items() if v is not None} + ) + + from fairseq.models import ARCH_MODEL_REGISTRY, ARCH_CONFIG_REGISTRY + + # Before creating the true parser, we need to import optional user module + # in order to eagerly import custom tasks, optimizers, architectures, etc. + usr_parser = argparse.ArgumentParser(add_help=False, allow_abbrev=False) + usr_parser.add_argument("--user-dir", default=None) + usr_args, _ = usr_parser.parse_known_args(input_args) + utils.import_user_module(usr_args) + + if modify_parser is not None: + modify_parser(parser) + + # The parser doesn't know about model/criterion/optimizer-specific args, so + # we parse twice. First we parse the model/criterion/optimizer, then we + # parse a second time after adding the *-specific arguments. + # If input_args is given, we will parse those args instead of sys.argv. + args, _ = parser.parse_known_args(input_args) + + # Add model-specific args to parser. + if hasattr(args, "arch"): + model_specific_group = parser.add_argument_group( + "Model-specific configuration", + # Only include attributes which are explicitly given as command-line + # arguments or which have default values. + argument_default=argparse.SUPPRESS, + ) + ARCH_MODEL_REGISTRY[args.arch].add_args(model_specific_group) + + # Add *-specific args to parser. + from fairseq.registry import REGISTRIES + + for registry_name, REGISTRY in REGISTRIES.items(): + choice = getattr(args, registry_name, None) + if choice is not None: + cls = REGISTRY["registry"][choice] + if hasattr(cls, "add_args"): + cls.add_args(parser) + if hasattr(args, "task"): + from fairseq.tasks import TASK_REGISTRY + + TASK_REGISTRY[args.task].add_args(parser) + if getattr(args, "use_bmuf", False): + # hack to support extra args for block distributed data parallelism + from fairseq.optim.bmuf import FairseqBMUF + + FairseqBMUF.add_args(parser) + + # Modify the parser a second time, since defaults may have been reset + if modify_parser is not None: + modify_parser(parser) + + # Parse a second time. + if parse_known: + args, extra = parser.parse_known_args(input_args) + else: + args = parser.parse_args(input_args) + extra = None + + # Post-process args. + if hasattr(args, "max_sentences_valid") and args.max_sentences_valid is None: + args.max_sentences_valid = args.max_sentences + if hasattr(args, "max_tokens_valid") and args.max_tokens_valid is None: + args.max_tokens_valid = args.max_tokens + if getattr(args, "memory_efficient_fp16", False): + args.fp16 = True + if getattr(args, "memory_efficient_bf16", False): + args.bf16 = True + args.tpu = getattr(args, "tpu", False) + args.bf16 = getattr(args, "bf16", False) + if args.bf16: + args.tpu = True + if args.tpu and args.fp16: + raise ValueError("Cannot combine --fp16 and --tpu, use --bf16 on TPUs") + + if getattr(args, "seed", None) is None: + args.seed = 1 # default seed for training + args.no_seed_provided = True + else: + args.no_seed_provided = False + + # Apply architecture configuration. + if hasattr(args, "arch"): + ARCH_CONFIG_REGISTRY[args.arch](args) + + if parse_known: + return args, extra + else: + return args + + +def get_parser(desc, default_task="translation"): + # Before creating the true parser, we need to import optional user module + # in order to eagerly import custom tasks, optimizers, architectures, etc. + usr_parser = argparse.ArgumentParser(add_help=False, allow_abbrev=False) + usr_parser.add_argument("--user-dir", default=None) + usr_args, _ = usr_parser.parse_known_args() + utils.import_user_module(usr_args) + + parser = argparse.ArgumentParser(allow_abbrev=False) + # fmt: off + parser.add_argument('--no-progress-bar', action='store_true', help='disable progress bar') + parser.add_argument('--log-interval', type=int, default=100, metavar='N', + help='log progress every N batches (when progress bar is disabled)') + parser.add_argument('--log-format', default=None, help='log format to use', + choices=['json', 'none', 'simple', 'tqdm']) + parser.add_argument('--tensorboard-logdir', metavar='DIR', default='', + help='path to save logs for tensorboard, should match --logdir ' + 'of running tensorboard (default: no tensorboard logging)') + parser.add_argument('--seed', default=None, type=int, metavar='N', + help='pseudo random number generator seed') + parser.add_argument('--cpu', action='store_true', help='use CPU instead of CUDA') + parser.add_argument('--tpu', action='store_true', help='use TPU instead of CUDA') + parser.add_argument('--bf16', action='store_true', help='use bfloat16; implies --tpu') + parser.add_argument('--fp16', action='store_true', help='use FP16') + parser.add_argument('--memory-efficient-bf16', action='store_true', + help='use a memory-efficient version of BF16 training; implies --bf16') + parser.add_argument('--memory-efficient-fp16', action='store_true', + help='use a memory-efficient version of FP16 training; implies --fp16') + parser.add_argument('--fp16-no-flatten-grads', action='store_true', + help='don\'t flatten FP16 grads tensor') + parser.add_argument('--fp16-init-scale', default=2 ** 7, type=int, + help='default FP16 loss scale') + parser.add_argument('--fp16-scale-window', type=int, + help='number of updates before increasing loss scale') + parser.add_argument('--fp16-scale-tolerance', default=0.0, type=float, + help='pct of updates that can overflow before decreasing the loss scale') + parser.add_argument('--min-loss-scale', default=1e-4, type=float, metavar='D', + help='minimum FP16 loss scale, after which training is stopped') + parser.add_argument('--threshold-loss-scale', type=float, + help='threshold FP16 loss scale from below') + parser.add_argument('--user-dir', default=None, + help='path to a python module containing custom extensions (tasks and/or architectures)') + parser.add_argument('--empty-cache-freq', default=0, type=int, + help='how often to clear the PyTorch CUDA cache (0 to disable)') + parser.add_argument('--all-gather-list-size', default=16384, type=int, + help='number of bytes reserved for gathering stats from workers') + parser.add_argument('--model-parallel-size', type=int, metavar='N', + default=1, + help='total number of GPUs to parallelize model over') + parser.add_argument('--checkpoint-suffix', default='', + help='suffix to add to the checkpoint file name') + parser.add_argument('--quantization-config-path', default=None, + help='path to quantization config file') + parser.add_argument('--profile', action='store_true', help='enable autograd profiler emit_nvtx') + + from fairseq.registry import REGISTRIES + for registry_name, REGISTRY in REGISTRIES.items(): + parser.add_argument( + '--' + registry_name.replace('_', '-'), + default=REGISTRY['default'], + choices=REGISTRY['registry'].keys(), + ) + + # Task definitions can be found under fairseq/tasks/ + from fairseq.tasks import TASK_REGISTRY + parser.add_argument('--task', metavar='TASK', default=default_task, + choices=TASK_REGISTRY.keys(), + help='task') + # fmt: on + return parser + + +def add_preprocess_args(parser): + group = parser.add_argument_group("Preprocessing") + # fmt: off + group.add_argument("-s", "--source-lang", default=None, metavar="SRC", + help="source language") + group.add_argument("-t", "--target-lang", default=None, metavar="TARGET", + help="target language") + group.add_argument("--trainpref", metavar="FP", default=None, + help="train file prefix") + group.add_argument("--validpref", metavar="FP", default=None, + help="comma separated, valid file prefixes") + group.add_argument("--testpref", metavar="FP", default=None, + help="comma separated, test file prefixes") + group.add_argument("--align-suffix", metavar="FP", default=None, + help="alignment file suffix") + group.add_argument("--destdir", metavar="DIR", default="data-bin", + help="destination dir") + group.add_argument("--thresholdtgt", metavar="N", default=0, type=int, + help="map words appearing less than threshold times to unknown") + group.add_argument("--thresholdsrc", metavar="N", default=0, type=int, + help="map words appearing less than threshold times to unknown") + group.add_argument("--tgtdict", metavar="FP", + help="reuse given target dictionary") + group.add_argument("--srcdict", metavar="FP", + help="reuse given source dictionary") + group.add_argument("--nwordstgt", metavar="N", default=-1, type=int, + help="number of target words to retain") + group.add_argument("--nwordssrc", metavar="N", default=-1, type=int, + help="number of source words to retain") + group.add_argument("--alignfile", metavar="ALIGN", default=None, + help="an alignment file (optional)") + parser.add_argument('--dataset-impl', metavar='FORMAT', default='mmap', + choices=get_available_dataset_impl(), + help='output dataset implementation') + group.add_argument("--joined-dictionary", action="store_true", + help="Generate joined dictionary") + group.add_argument("--only-source", action="store_true", + help="Only process the source language") + group.add_argument("--padding-factor", metavar="N", default=8, type=int, + help="Pad dictionary size to be multiple of N") + group.add_argument("--workers", metavar="N", default=1, type=int, + help="number of parallel workers") + # fmt: on + return parser + + +def add_dataset_args(parser, train=False, gen=False): + group = parser.add_argument_group("Dataset and data loading") + # fmt: off + group.add_argument('--num-workers', default=1, type=int, metavar='N', + help='how many subprocesses to use for data loading') + group.add_argument('--skip-invalid-size-inputs-valid-test', action='store_true', + help='ignore too long or too short lines in valid and test set') + group.add_argument('--max-tokens', type=int, metavar='N', + help='maximum number of tokens in a batch') + group.add_argument('--max-sentences', '--batch-size', type=int, metavar='N', + help='maximum number of sentences in a batch') + group.add_argument('--required-batch-size-multiple', default=8, type=int, metavar='N', + help='batch size will either be less than this value, ' + 'or a multiple of this value') + parser.add_argument('--dataset-impl', metavar='FORMAT', + choices=get_available_dataset_impl(), + help='output dataset implementation') + group.add_argument('--data-buffer-size', default=10, type=int, metavar='N', + help='number of batches to preload') + if train: + group.add_argument('--train-subset', default='train', metavar='SPLIT', + help='data subset to use for training (e.g. train, valid, test)') + group.add_argument('--valid-subset', default='valid', metavar='SPLIT', + help='comma separated list of data subsets to use for validation' + ' (e.g. train, valid, test)') + group.add_argument('--validate-interval', type=int, default=1, metavar='N', + help='validate every N epochs') + group.add_argument('--validate-interval-updates', type=int, default=0, metavar='N', + help='validate every N updates') + group.add_argument('--validate-after-updates', type=int, default=0, metavar='N', + help='dont validate until reaching this many updates') + group.add_argument('--fixed-validation-seed', default=None, type=int, metavar='N', + help='specified random seed for validation') + group.add_argument('--disable-validation', action='store_true', + help='disable validation') + group.add_argument('--max-tokens-valid', type=int, metavar='N', + help='maximum number of tokens in a validation batch' + ' (defaults to --max-tokens)') + group.add_argument('--max-sentences-valid', type=int, metavar='N', + help='maximum number of sentences in a validation batch' + ' (defaults to --max-sentences)') + group.add_argument('--curriculum', default=0, type=int, metavar='N', + help='don\'t shuffle batches for first N epochs') + if gen: + group.add_argument('--gen-subset', default='test', metavar='SPLIT', + help='data subset to generate (train, valid, test)') + group.add_argument('--num-shards', default=1, type=int, metavar='N', + help='shard generation over N shards') + group.add_argument('--shard-id', default=0, type=int, metavar='ID', + help='id of the shard to generate (id < num_shards)') + # fmt: on + return group + + +def add_distributed_training_args(parser, default_world_size=None): + group = parser.add_argument_group("Distributed training") + # fmt: off + if default_world_size is None: + default_world_size = max(1, torch.cuda.device_count()) + group.add_argument('--distributed-world-size', type=int, metavar='N', + default=default_world_size, + help='total number of GPUs across all nodes (default: all visible GPUs)') + group.add_argument('--distributed-rank', default=0, type=int, + help='rank of the current worker') + group.add_argument('--distributed-backend', default='nccl', type=str, + help='distributed backend') + group.add_argument('--distributed-init-method', default=None, type=str, + help='typically tcp://hostname:port that will be used to ' + 'establish initial connetion') + group.add_argument('--distributed-port', default=-1, type=int, + help='port number (not required if using --distributed-init-method)') + group.add_argument('--device-id', '--local_rank', default=0, type=int, + help='which GPU to use (usually configured automatically)') + group.add_argument('--distributed-no-spawn', action='store_true', + help='do not spawn multiple processes even if multiple GPUs are visible') + # "c10d" is PyTorch's DDP implementation and provides the fastest + # training. "no_c10d" is a more robust, but slightly slower DDP + # implementation. Try this if you get warning messages about + # inconsistent gradients between workers, or if some of your model + # parameters are not always used. + group.add_argument('--ddp-backend', default='c10d', type=str, + choices=['c10d', 'no_c10d'], + help='DistributedDataParallel backend') + group.add_argument('--bucket-cap-mb', default=25, type=int, metavar='MB', + help='bucket size for reduction') + group.add_argument('--fix-batches-to-gpus', action='store_true', + help='don\'t shuffle batches between GPUs; this reduces overall ' + 'randomness and may affect precision but avoids the cost of ' + 're-reading the data') + group.add_argument('--find-unused-parameters', default=False, action='store_true', + help='disable unused parameter detection (not applicable to ' + 'no_c10d ddp-backend') + group.add_argument('--fast-stat-sync', default=False, action='store_true', + help='[deprecated] this is now defined per Criterion') + group.add_argument('--broadcast-buffers', default=False, action='store_true', + help='Copy non-trainable parameters between GPUs, such as ' + 'batchnorm population statistics') + + group.add_argument('--distributed-wrapper', default='DDP', type=str, + choices=['DDP', 'SlowMo'], + help='DistributedDataParallel backend') + # Add arguments for SlowMo - these will be used when SlowMo is enabled via above + group.add_argument('--slowmo-momentum', default=None, type=float, + help='SlowMo momentum term; by default use 0.0 for 16 GPUs, ' + '0.2 for 32 GPUs; 0.5 for 64 GPUs, 0.6 for > 64 GPUs') + group.add_argument('--slowmo-algorithm', default='LocalSGD', choices=['LocalSGD', 'SGP'], + help='whether to use LocalSGD or SGP') + group.add_argument('--localsgd-frequency', default=3, type=int, + help='Local SGD allreduce frequency') + group.add_argument('--nprocs-per-node', type=int, metavar='N', + default=max(1, torch.cuda.device_count()), + help='number of GPUs in each node. An allreduce operation across GPUs in ' + 'a node is very fast. Hence, we do allreduce across GPUs in a node, ' + 'and gossip across different nodes') + # fmt: on + return group + + +def add_optimization_args(parser): + group = parser.add_argument_group("Optimization") + # fmt: off + group.add_argument('--max-epoch', '--me', default=0, type=int, metavar='N', + help='force stop training at specified epoch') + group.add_argument('--max-update', '--mu', default=0, type=int, metavar='N', + help='force stop training at specified update') + group.add_argument('--stop-time-hours', default=0, type=float, metavar='N', + help='force stop training after specified cumulative time (if >0)') + group.add_argument('--clip-norm', default=0.0, type=float, metavar='NORM', + help='clip threshold of gradients') + group.add_argument('--sentence-avg', action='store_true', + help='normalize gradients by the number of sentences in a batch' + ' (default is to normalize by number of tokens)') + group.add_argument('--update-freq', default='1', metavar='N1,N2,...,N_K', + type=lambda uf: eval_str_list(uf, type=int), + help='update parameters every N_i batches, when in epoch i') + group.add_argument('--lr', '--learning-rate', default='0.25', type=eval_str_list, + metavar='LR_1,LR_2,...,LR_N', + help='learning rate for the first N epochs; all epochs >N using LR_N' + ' (note: this may be interpreted differently depending on --lr-scheduler)') + group.add_argument('--min-lr', default=-1, type=float, metavar='LR', + help='stop training when the learning rate reaches this minimum') + group.add_argument('--use-bmuf', default=False, action='store_true', + help='specify global optimizer for syncing models on different GPUs/shards') + # fmt: on + return group + + +def add_checkpoint_args(parser): + group = parser.add_argument_group("Checkpointing") + # fmt: off + group.add_argument('--save-dir', metavar='DIR', default='checkpoints', + help='path to save checkpoints') + group.add_argument('--restore-file', default='checkpoint_last.pt', + help='filename from which to load checkpoint ' + '(default: /checkpoint_last.pt') + group.add_argument('--finetune-from-model', default=None, type=str, + help='finetune from a pretrained model; ' + 'note that meters and lr scheduler will be reset') + group.add_argument('--reset-dataloader', action='store_true', + help='if set, does not reload dataloader state from the checkpoint') + group.add_argument('--reset-lr-scheduler', action='store_true', + help='if set, does not load lr scheduler state from the checkpoint') + group.add_argument('--reset-meters', action='store_true', + help='if set, does not load meters from the checkpoint') + group.add_argument('--reset-optimizer', action='store_true', + help='if set, does not load optimizer state from the checkpoint') + group.add_argument('--optimizer-overrides', default="{}", type=str, metavar='DICT', + help='a dictionary used to override optimizer args when loading a checkpoint') + group.add_argument('--save-interval', type=int, default=1, metavar='N', + help='save a checkpoint every N epochs') + group.add_argument('--save-interval-updates', type=int, default=0, metavar='N', + help='save a checkpoint (and validate) every N updates') + group.add_argument('--keep-interval-updates', type=int, default=-1, metavar='N', + help='keep the last N checkpoints saved with --save-interval-updates') + group.add_argument('--keep-last-epochs', type=int, default=-1, metavar='N', + help='keep last N epoch checkpoints') + group.add_argument('--keep-best-checkpoints', type=int, default=-1, metavar='N', + help='keep best N checkpoints based on scores') + group.add_argument('--no-save', action='store_true', + help='don\'t save models or checkpoints') + group.add_argument('--no-epoch-checkpoints', action='store_true', + help='only store last and best checkpoints') + group.add_argument('--no-last-checkpoints', action='store_true', + help='don\'t store last checkpoints') + group.add_argument('--no-save-optimizer-state', action='store_true', + help='don\'t save optimizer-state as part of checkpoint') + group.add_argument('--best-checkpoint-metric', type=str, default='loss', + help='metric to use for saving "best" checkpoints') + group.add_argument('--maximize-best-checkpoint-metric', action='store_true', + help='select the largest metric value for saving "best" checkpoints') + group.add_argument('--patience', type=int, default=-1, metavar='N', + help=('early stop training if valid performance doesn\'t ' + 'improve for N consecutive validation runs; note ' + 'that this is influenced by --validate-interval')) + # fmt: on + return group + + +def add_common_eval_args(group): + # fmt: off + group.add_argument('--path', metavar='FILE', + help='path(s) to model file(s), colon separated') + group.add_argument('--remove-bpe', '--post-process', nargs='?', const='@@ ', default=None, + help='remove BPE tokens before scoring (can be set to sentencepiece)') + group.add_argument('--quiet', action='store_true', + help='only print final scores') + group.add_argument('--model-overrides', default="{}", type=str, metavar='DICT', + help='a dictionary used to override model args at generation ' + 'that were used during model training') + group.add_argument('--results-path', metavar='RESDIR', type=str, default=None, + help='path to save eval results (optional)"') + # fmt: on + + +def add_eval_lm_args(parser): + group = parser.add_argument_group("LM Evaluation") + add_common_eval_args(group) + # fmt: off + group.add_argument('--output-word-probs', action='store_true', + help='if set, outputs words and their predicted log probabilities to standard output') + group.add_argument('--output-word-stats', action='store_true', + help='if set, outputs word statistics such as word count, average probability, etc') + group.add_argument('--context-window', default=0, type=int, metavar='N', + help='ensures that every evaluated token has access to a context of at least this size,' + ' if possible') + group.add_argument('--softmax-batch', default=sys.maxsize, type=int, metavar='N', + help='if BxT is more than this, will batch the softmax over vocab to this amount of tokens' + ' in order to fit into GPU memory') + # fmt: on + + +def add_generation_args(parser): + group = parser.add_argument_group("Generation") + add_common_eval_args(group) + # fmt: off + group.add_argument('--beam', default=5, type=int, metavar='N', + help='beam size') + group.add_argument('--nbest', default=1, type=int, metavar='N', + help='number of hypotheses to output') + group.add_argument('--max-len-a', default=0, type=float, metavar='N', + help=('generate sequences of maximum length ax + b, ' + 'where x is the source length')) + group.add_argument('--max-len-b', default=200, type=int, metavar='N', + help=('generate sequences of maximum length ax + b, ' + 'where x is the source length')) + group.add_argument('--min-len', default=1, type=float, metavar='N', + help=('minimum generation length')) + group.add_argument('--match-source-len', default=False, action='store_true', + help=('generations should match the source length')) + group.add_argument('--no-early-stop', action='store_true', + help='deprecated') + group.add_argument('--unnormalized', action='store_true', + help='compare unnormalized hypothesis scores') + group.add_argument('--no-beamable-mm', action='store_true', + help='don\'t use BeamableMM in attention layers') + group.add_argument('--lenpen', default=1, type=float, + help='length penalty: <1.0 favors shorter, >1.0 favors longer sentences') + group.add_argument('--unkpen', default=0, type=float, + help='unknown word penalty: <0 produces more unks, >0 produces fewer') + group.add_argument('--replace-unk', nargs='?', const=True, default=None, + help='perform unknown replacement (optionally with alignment dictionary)') + group.add_argument('--sacrebleu', action='store_true', + help='score with sacrebleu') + group.add_argument('--score-reference', action='store_true', + help='just score the reference translation') + group.add_argument('--prefix-size', default=0, type=int, metavar='PS', + help='initialize generation by target prefix of given length') + group.add_argument('--no-repeat-ngram-size', default=0, type=int, metavar='N', + help='ngram blocking such that this size ngram cannot be repeated in the generation') + group.add_argument('--sampling', action='store_true', + help='sample hypotheses instead of using beam search') + group.add_argument('--sampling-topk', default=-1, type=int, metavar='PS', + help='sample from top K likely next words instead of all words') + group.add_argument('--sampling-topp', default=-1.0, type=float, metavar='PS', + help='sample from the smallest set whose cumulative probability mass exceeds p for next words') + group.add_argument('--temperature', default=1., type=float, metavar='N', + help='temperature for generation') + group.add_argument('--diverse-beam-groups', default=-1, type=int, metavar='N', + help='number of groups for Diverse Beam Search') + group.add_argument('--diverse-beam-strength', default=0.5, type=float, metavar='N', + help='strength of diversity penalty for Diverse Beam Search') + group.add_argument('--diversity-rate', default=-1.0, type=float, metavar='N', + help='strength of diversity penalty for Diverse Siblings Search') + group.add_argument('--print-alignment', action='store_true', + help='if set, uses attention feedback to compute and print alignment to source tokens') + group.add_argument('--print-step', action='store_true') + + # arguments for iterative refinement generator + group.add_argument('--iter-decode-eos-penalty', default=0.0, type=float, metavar='N', + help='if > 0.0, it penalized early-stopping in decoding.') + group.add_argument('--iter-decode-max-iter', default=10, type=int, metavar='N', + help='maximum iterations for iterative refinement.') + group.add_argument('--iter-decode-force-max-iter', action='store_true', + help='if set, run exact the maximum number of iterations without early stop') + group.add_argument('--iter-decode-with-beam', default=1, type=int, metavar='N', + help='if > 1, model will generate translations varying by the lengths.') + group.add_argument('--iter-decode-with-external-reranker', action='store_true', + help='if set, the last checkpoint are assumed to be a reranker to rescore the translations'), + group.add_argument('--retain-iter-history', action='store_true', + help='if set, decoding returns the whole history of iterative refinement') + group.add_argument('--retain-dropout', action='store_true', + help='Use dropout at inference time') + group.add_argument('--retain-dropout-modules', default=None, nargs='+', type=str, + help='if set, only retain dropout for the specified modules; ' + 'if not set, then dropout will be retained for all modules') + + # special decoding format for advanced decoding. + group.add_argument('--decoding-format', default=None, type=str, choices=['unigram', 'ensemble', 'vote', 'dp', 'bs']) + # fmt: on + return group + + +def add_interactive_args(parser): + group = parser.add_argument_group("Interactive") + # fmt: off + group.add_argument('--buffer-size', default=0, type=int, metavar='N', + help='read this many sentences into a buffer before processing them') + group.add_argument('--input', default='-', type=str, metavar='FILE', + help='file to read from; use - for stdin') + # fmt: on + + +def add_model_args(parser): + group = parser.add_argument_group("Model configuration") + # fmt: off + + # Model definitions can be found under fairseq/models/ + # + # The model architecture can be specified in several ways. + # In increasing order of priority: + # 1) model defaults (lowest priority) + # 2) --arch argument + # 3) --encoder/decoder-* arguments (highest priority) + from fairseq.models import ARCH_MODEL_REGISTRY + group.add_argument('--arch', '-a', default='fconv', metavar='ARCH', + choices=ARCH_MODEL_REGISTRY.keys(), + help='Model Architecture') + # fmt: on + return group diff --git a/fairseq/pdb.py b/fairseq/pdb.py new file mode 100644 index 0000000000000000000000000000000000000000..f1ce3c46bca47fd8a47272770d6c61fcd9c13f75 --- /dev/null +++ b/fairseq/pdb.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import multiprocessing +import os +import pdb +import sys + + +__all__ = ['set_trace'] + + +_stdin = [None] +_stdin_lock = multiprocessing.Lock() +try: + _stdin_fd = sys.stdin.fileno() +except Exception: + _stdin_fd = None + + +class MultiprocessingPdb(pdb.Pdb): + """A Pdb wrapper that works in a multiprocessing environment. + + Usage: `from fairseq import pdb; pdb.set_trace()` + """ + + def __init__(self): + pdb.Pdb.__init__(self, nosigint=True) + + def _cmdloop(self): + stdin_bak = sys.stdin + with _stdin_lock: + try: + if _stdin_fd is not None: + if not _stdin[0]: + _stdin[0] = os.fdopen(_stdin_fd) + sys.stdin = _stdin[0] + self.cmdloop() + finally: + sys.stdin = stdin_bak + + +def set_trace(): + pdb = MultiprocessingPdb() + pdb.set_trace(sys._getframe().f_back) diff --git a/fairseq/quantization_utils.py b/fairseq/quantization_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a7f5ade9b31738393c78ee803fb8d66d825e69e1 --- /dev/null +++ b/fairseq/quantization_utils.py @@ -0,0 +1,142 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +from fairseq.modules.quantization import pq, quantization_options, scalar + + +logger = logging.getLogger(__name__) + + +def quantize_model_scalar(model, args): + quant_noise_scalar = getattr(args, 'quant_noise_scalar', 0) + if quant_noise_scalar > 0: + # quantize_model edits the model in place + scalar.quantize_model_(model, p=quant_noise_scalar, bits=8, update_step=1000) + return model + + +class Quantizer(object): + + def __init__(self, config_path, max_epoch, max_update): + try: + import yaml + except ImportError: + raise ImportError('Please install yaml with: pip install yaml') + + # parse config + if config_path: + with open(config_path) as config_file: + config = quantization_options.parse_config_yaml( + yaml.safe_load(config_file) + ) + else: + config = quantization_options.parse_config_yaml({}) + + self.n_centroids_config = config["n_centroids"] + self.block_sizes_config = config["block_sizes"] + self.layers_to_quantize = config["layers_to_quantize"] + + # We assume that training will run for a fixed number of epochs + # (or updates) and that we should train for equal durations + # between iterations of PQ. + num_iterations = len(self.layers_to_quantize) + if max_epoch > 0: + assert max_epoch % num_iterations == 0, ( + 'for iterative PQ, --max-epoch (={}) must be evenly divisible by ' + 'len(layers_to_quantize) (={})'.format(max_epoch, num_iterations) + ) + self.epoch_schedule = max_epoch // num_iterations + else: + self.epoch_schedule = None + if max_update > 0: + assert max_update % num_iterations == 0, ( + 'for iterative PQ, --max-update (={}) must be evenly divisible by ' + 'len(layers_to_quantize) (={})'.format(max_update, num_iterations) + ) + self.update_schedule = max_update // num_iterations + else: + self.update_schedule = None + assert (self.epoch_schedule is not None) ^ (self.update_schedule is not None), \ + 'for iterative PQ, cannot specify both --max-update and --max-epoch' + + # 0 is a special value for quantization step, which will force + # the first call to begin_epoch() to call step() + self.quantization_step = 0 + + def set_trainer(self, trainer): + self.trainer = trainer + self.size_tracker = pq.SizeTracker(self.trainer.get_model()) + + def step(self): + """Move to the next stage of quantization.""" + if self.quantization_step >= len(self.layers_to_quantize): + # Maybe we just finished the last training step or we loaded + # a checkpoint for an iterative PQ model which previously + # finished training. Either way, don't quantize again. + return + + logger.info( + 'quantizing model (step={}; layers_to_quantize[step]={})'.format( + self.quantization_step, self.layers_to_quantize[self.quantization_step] + ) + ) + quantized_layers = pq.quantize_model_( + self.trainer.get_model(), + self.size_tracker, + self.layers_to_quantize, + self.block_sizes_config, + self.n_centroids_config, + step=self.quantization_step, + ) + logger.info('quantized layers: {}'.format(quantized_layers)) + logger.info(self.size_tracker) + + self.quantization_step += 1 + + # reintialize the Trainer since model parameters have changed + self.trainer.reinitialize() + + def begin_epoch(self, epoch): + """Called at the beginning of each epoch (epochs start at 1).""" + if ( + ( + self.epoch_schedule is not None + and epoch > 0 + and (epoch - 1) % self.epoch_schedule == 0 + ) + # we always step once in the beginning, even if using + # update-based quantization + or self.quantization_step == 0 + ): + self.step() + + def step_update(self, num_updates): + """Called at the end of each step.""" + if ( + self.update_schedule is not None + and num_updates > 0 + and num_updates % self.update_schedule == 0 + ): + self.step() + + def state_dict(self): + return { + 'n_centroids_config': self.n_centroids_config, + 'block_sizes_config': self.block_sizes_config, + 'layers_to_quantize': self.layers_to_quantize, + 'epoch_schedule': self.epoch_schedule, + 'update_schedule': self.update_schedule, + 'quantization_step': self.quantization_step, + } + + def load_state_dict(self, state_dict): + self.n_centroids_config = state_dict['n_centroids_config'] + self.block_sizes_config = state_dict['block_sizes_config'] + self.layers_to_quantize = state_dict['layers_to_quantize'] + self.epoch_schedule = state_dict['epoch_schedule'] + self.update_schedule = state_dict['update_schedule'] + self.quantization_step = state_dict['quantization_step'] diff --git a/fairseq/registry.py b/fairseq/registry.py new file mode 100644 index 0000000000000000000000000000000000000000..ed24258c57713075fd42383a8bfa9461c45cf1b1 --- /dev/null +++ b/fairseq/registry.py @@ -0,0 +1,80 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse + + +REGISTRIES = {} + + +def setup_registry( + registry_name: str, + base_class=None, + default=None, +): + assert registry_name.startswith('--') + registry_name = registry_name[2:].replace('-', '_') + + REGISTRY = {} + REGISTRY_CLASS_NAMES = set() + + # maintain a registry of all registries + if registry_name in REGISTRIES: + return # registry already exists + REGISTRIES[registry_name] = { + 'registry': REGISTRY, + 'default': default, + } + + def build_x(args, *extra_args, **extra_kwargs): + choice = getattr(args, registry_name, None) + if choice is None: + return None + cls = REGISTRY[choice] + if hasattr(cls, 'build_' + registry_name): + builder = getattr(cls, 'build_' + registry_name) + else: + builder = cls + set_defaults(args, cls) + return builder(args, *extra_args, **extra_kwargs) + + def register_x(name): + + def register_x_cls(cls): + if name in REGISTRY: + raise ValueError('Cannot register duplicate {} ({})'.format(registry_name, name)) + if cls.__name__ in REGISTRY_CLASS_NAMES: + raise ValueError( + 'Cannot register {} with duplicate class name ({})'.format( + registry_name, cls.__name__, + ) + ) + if base_class is not None and not issubclass(cls, base_class): + raise ValueError('{} must extend {}'.format(cls.__name__, base_class.__name__)) + REGISTRY[name] = cls + REGISTRY_CLASS_NAMES.add(cls.__name__) + return cls + + return register_x_cls + + return build_x, register_x, REGISTRY + + +def set_defaults(args, cls): + """Helper to set default arguments based on *add_args*.""" + if not hasattr(cls, 'add_args'): + return + parser = argparse.ArgumentParser(argument_default=argparse.SUPPRESS, allow_abbrev=False) + cls.add_args(parser) + # copied from argparse.py: + defaults = argparse.Namespace() + for action in parser._actions: + if action.dest is not argparse.SUPPRESS: + if not hasattr(defaults, action.dest): + if action.default is not argparse.SUPPRESS: + setattr(defaults, action.dest, action.default) + for key, default_value in vars(defaults).items(): + if not hasattr(args, key): + setattr(args, key, default_value) diff --git a/fairseq/scoring/__init__.py b/fairseq/scoring/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6e5cc287ba393e83ed2da5c2a0fb8e156654853f --- /dev/null +++ b/fairseq/scoring/__init__.py @@ -0,0 +1,22 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import importlib +import os + +from fairseq import registry + + +build_scoring, register_scoring, SCORING_REGISTRY = registry.setup_registry( + "--scoring", default="bleu" +) + + +# automatically import any Python files in the current directory +for file in os.listdir(os.path.dirname(__file__)): + if file.endswith(".py") and not file.startswith("_"): + module = file[: file.find(".py")] + importlib.import_module("fairseq.scoring." + module) diff --git a/fairseq/scoring/__pycache__/__init__.cpython-310.pyc b/fairseq/scoring/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eace372b9c5541cd35bef5afad465381cceb604b Binary files /dev/null and b/fairseq/scoring/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/scoring/__pycache__/bleu.cpython-310.pyc b/fairseq/scoring/__pycache__/bleu.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..10670697ca22d2f7d65e2caf1ecc44114cba4fa7 Binary files /dev/null and b/fairseq/scoring/__pycache__/bleu.cpython-310.pyc differ diff --git a/fairseq/scoring/__pycache__/scoring_utils.cpython-310.pyc b/fairseq/scoring/__pycache__/scoring_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..901a4cd98a9f3d97ef365bc09677e24658d57b78 Binary files /dev/null and b/fairseq/scoring/__pycache__/scoring_utils.cpython-310.pyc differ diff --git a/fairseq/scoring/__pycache__/wer.cpython-310.pyc b/fairseq/scoring/__pycache__/wer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6ad6fbccc45824b2b528456b452d50b4818d145b Binary files /dev/null and b/fairseq/scoring/__pycache__/wer.cpython-310.pyc differ diff --git a/fairseq/scoring/bleu.py b/fairseq/scoring/bleu.py new file mode 100644 index 0000000000000000000000000000000000000000..40f3440d82f0638c138f131d3a9bdf0e3d2b2a33 --- /dev/null +++ b/fairseq/scoring/bleu.py @@ -0,0 +1,141 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import ctypes +import math +import torch + +from fairseq.scoring import register_scoring + +try: + from fairseq import libbleu +except ImportError as e: + import sys + + sys.stderr.write("ERROR: missing libbleu.so. run `pip install --editable .`\n") + raise e + + +C = ctypes.cdll.LoadLibrary(libbleu.__file__) + + +class BleuStat(ctypes.Structure): + _fields_ = [ + ("reflen", ctypes.c_size_t), + ("predlen", ctypes.c_size_t), + ("match1", ctypes.c_size_t), + ("count1", ctypes.c_size_t), + ("match2", ctypes.c_size_t), + ("count2", ctypes.c_size_t), + ("match3", ctypes.c_size_t), + ("count3", ctypes.c_size_t), + ("match4", ctypes.c_size_t), + ("count4", ctypes.c_size_t), + ] + + +@register_scoring("sacrebleu") +class SacrebleuScorer(object): + def __init__(self, *unused): + import sacrebleu + + self.sacrebleu = sacrebleu + self.reset() + + def reset(self, one_init=False): + if one_init: + raise NotImplementedError + self.ref = [] + self.sys = [] + + def add_string(self, ref, pred): + self.ref.append(ref) + self.sys.append(pred) + + def score(self, order=4): + return self.result_string(order).score + + def result_string(self, order=4): + if order != 4: + raise NotImplementedError + return self.sacrebleu.corpus_bleu(self.sys, [self.ref]).format() + + +@register_scoring("bleu") +class Scorer(object): + def __init__(self, pad, eos, unk): + self.stat = BleuStat() + self.pad = pad + self.eos = eos + self.unk = unk + self.reset() + + def reset(self, one_init=False): + if one_init: + C.bleu_one_init(ctypes.byref(self.stat)) + else: + C.bleu_zero_init(ctypes.byref(self.stat)) + + def add(self, ref, pred): + if not isinstance(ref, torch.IntTensor): + raise TypeError("ref must be a torch.IntTensor (got {})".format(type(ref))) + if not isinstance(pred, torch.IntTensor): + raise TypeError("pred must be a torch.IntTensor(got {})".format(type(pred))) + + # don't match unknown words + rref = ref.clone() + assert not rref.lt(0).any() + rref[rref.eq(self.unk)] = -999 + + rref = rref.contiguous().view(-1) + pred = pred.contiguous().view(-1) + + C.bleu_add( + ctypes.byref(self.stat), + ctypes.c_size_t(rref.size(0)), + ctypes.c_void_p(rref.data_ptr()), + ctypes.c_size_t(pred.size(0)), + ctypes.c_void_p(pred.data_ptr()), + ctypes.c_int(self.pad), + ctypes.c_int(self.eos), + ) + + def score(self, order=4): + psum = sum( + math.log(p) if p > 0 else float("-Inf") for p in self.precision()[:order] + ) + return self.brevity() * math.exp(psum / order) * 100 + + def precision(self): + def ratio(a, b): + return a / b if b > 0 else 0 + + return [ + ratio(self.stat.match1, self.stat.count1), + ratio(self.stat.match2, self.stat.count2), + ratio(self.stat.match3, self.stat.count3), + ratio(self.stat.match4, self.stat.count4), + ] + + def brevity(self): + r = self.stat.reflen / self.stat.predlen + return min(1, math.exp(1 - r)) + + def result_string(self, order=4): + assert order <= 4, "BLEU scores for order > 4 aren't supported" + fmt = "BLEU{} = {:2.2f}, {:2.1f}" + for _ in range(1, order): + fmt += "/{:2.1f}" + fmt += " (BP={:.3f}, ratio={:.3f}, syslen={}, reflen={})" + bleup = [p * 100 for p in self.precision()[:order]] + return fmt.format( + order, + self.score(order=order), + *bleup, + self.brevity(), + self.stat.predlen / self.stat.reflen, + self.stat.predlen, + self.stat.reflen + ) diff --git a/fairseq/scoring/scoring_utils.py b/fairseq/scoring/scoring_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..0b710d5bb87a9487c3a5554de310f7dda8d1393b --- /dev/null +++ b/fairseq/scoring/scoring_utils.py @@ -0,0 +1,22 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import utils +from . import bleu, build_scoring + + +def build_scorer(args, tgt_dict): + if args.sacrebleu: + utils.deprecation_warning( + "--sacrebleu is deprecated. Please use --scoring sacrebleu instead." + ) + args.scoring = "sacrebleu" + + if args.scoring == "bleu": + scorer = bleu.Scorer(tgt_dict.pad(), tgt_dict.eos(), tgt_dict.unk()) + else: + return build_scoring(args) + + return scorer diff --git a/fairseq/scoring/wer.py b/fairseq/scoring/wer.py new file mode 100644 index 0000000000000000000000000000000000000000..6f4521f6cd02f876ab7c4f3f5ef0f058460f0974 --- /dev/null +++ b/fairseq/scoring/wer.py @@ -0,0 +1,32 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import editdistance + +from fairseq.scoring import register_scoring + + +@register_scoring("wer") +class WerScorer(object): + def __init__(self, *unused): + self.reset() + + def reset(self): + self.distance = 0 + self.target_length = 0 + + def add_string(self, ref, pred): + pred_items = ref.split() + targ_items = pred.split() + self.distance += editdistance.eval(pred_items, targ_items) + self.target_length += len(targ_items) + + def result_string(self): + return f"WER: {self.score()}" + + def score(self): + return ( + 100.0 * self.distance / self.target_length if self.target_length > 0 else 0 + ) diff --git a/fairseq/search.py b/fairseq/search.py new file mode 100644 index 0000000000000000000000000000000000000000..9e18581a978e678b4152427c0b8328c35e0b1005 --- /dev/null +++ b/fairseq/search.py @@ -0,0 +1,341 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Optional, List + +import torch +import torch.nn as nn +from torch import Tensor + + +class Search(nn.Module): + def __init__(self, tgt_dict): + super().__init__() + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() + self.vocab_size = len(tgt_dict) + self.src_lengths = torch.tensor(-1) + + def step(self, step, lprobs, scores): + """Take a single search step. + + Args: + step: the current search step, starting at 0 + lprobs: (bsz x input_beam_size x vocab_size) + the model's log-probabilities over the vocabulary at the current step + scores: (bsz x input_beam_size x step) + the historical model scores of each hypothesis up to this point + + Return: A tuple of (scores, indices, beams) where: + scores: (bsz x output_beam_size) + the scores of the chosen elements; output_beam_size can be + larger than input_beam_size, e.g., we may return + 2*input_beam_size to account for EOS + indices: (bsz x output_beam_size) + the indices of the chosen elements + beams: (bsz x output_beam_size) + the hypothesis ids of the chosen elements, in the range [0, input_beam_size) + """ + raise NotImplementedError + + @torch.jit.export + def set_src_lengths(self, src_lengths): + self.src_lengths = src_lengths + + +class BeamSearch(Search): + def __init__(self, tgt_dict): + super().__init__(tgt_dict) + + @torch.jit.export + def step(self, step: int, lprobs, scores: Optional[Tensor]): + bsz, beam_size, vocab_size = lprobs.size() + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + lprobs = lprobs[:, ::beam_size, :].contiguous() + else: + # make probs contain cumulative scores for each hypothesis + assert scores is not None + lprobs = lprobs + scores[:, :, step - 1].unsqueeze(-1) + + top_prediction = torch.topk( + lprobs.view(bsz, -1), + k=min( + # Take the best 2 x beam_size predictions. We'll choose the first + # beam_size of these which don't predict eos to continue with. + beam_size * 2, + lprobs.view(bsz, -1).size(1) - 1, # -1 so we never select pad + ), + ) + scores_buf = top_prediction[0] + indices_buf = top_prediction[1] + beams_buf = indices_buf // vocab_size + indices_buf = indices_buf.fmod(vocab_size) + return scores_buf, indices_buf, beams_buf + + +class LengthConstrainedBeamSearch(Search): + def __init__(self, tgt_dict, min_len_a, min_len_b, max_len_a, max_len_b): + super().__init__(tgt_dict) + self.min_len_a = min_len_a + self.min_len_b = min_len_b + self.max_len_a = max_len_a + self.max_len_b = max_len_b + self.beam = BeamSearch(tgt_dict) + self.needs_src_lengths = True + + def step(self, step: int, lprobs, scores): + min_lens = self.min_len_a * self.src_lengths + self.min_len_b + max_lens = self.max_len_a * self.src_lengths + self.max_len_b + lprobs[step < min_lens, :, self.eos] = -math.inf + lprobs[step >= max_lens, :, self.eos] = 0 + return self.beam.step(step, lprobs, scores) + + +class DiverseBeamSearch(Search): + """Diverse Beam Search. + + See "Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence + Models" for details. + + We only implement the Hamming Diversity penalty here, which performed best + in the original paper. + """ + + def __init__(self, tgt_dict, num_groups, diversity_strength): + super().__init__(tgt_dict) + self.num_groups = num_groups + self.diversity_strength = -diversity_strength + self.beam = BeamSearch(tgt_dict) + + @torch.jit.export + def step(self, step: int, lprobs, scores): + bsz, beam_size, vocab_size = lprobs.size() + if beam_size % self.num_groups != 0: + raise ValueError( + "DiverseBeamSearch requires --beam to be divisible by the number of groups" + ) + + # initialize diversity penalty + diversity_buf = torch.zeros(lprobs[:, 0, :].size()).to(lprobs) + + scores_G, indices_G, beams_G = [], [], [] + for g in range(self.num_groups): + lprobs_g = lprobs[:, g :: self.num_groups, :] + scores_g = scores[:, g :: self.num_groups, :] if step > 0 else None + + # apply diversity penalty + if g > 0: + lprobs_g = torch.add( + lprobs_g, self.diversity_strength, diversity_buf.unsqueeze(1) + ) + else: + lprobs_g = lprobs_g.contiguous() + + scores_buf, indices_buf, beams_buf = self.beam.step( + step, lprobs_g, scores_g + ) + beams_buf.mul_(self.num_groups).add_(g) + + scores_G.append(scores_buf.clone()) + indices_G.append(indices_buf.clone()) + beams_G.append(beams_buf.clone()) + + # update diversity penalty + diversity_buf.scatter_add_( + 1, indices_buf, torch.ones(indices_buf.size()).to(diversity_buf) + ) + + # interleave results from different groups + scores_buf = torch.stack(scores_G, dim=2).view(bsz, -1) + indices_buf = torch.stack(indices_G, dim=2).view(bsz, -1) + beams_buf = torch.stack(beams_G, dim=2).view(bsz, -1) + return scores_buf, indices_buf, beams_buf + + +class Sampling(Search): + sampling_topk: int + sampling_topp: float + + def __init__(self, tgt_dict, sampling_topk=-1, sampling_topp=-1.0): + super().__init__(tgt_dict) + self.sampling_topk = sampling_topk + self.sampling_topp = sampling_topp + + def _sample_topp(self, lprobs): + """Sample among the smallest set of elements whose cumulative probability mass exceeds p. + + See `"The Curious Case of Neural Text Degeneration" + (Holtzman et al., 2019) `_. + + Args: + lprobs: (bsz x input_beam_size x vocab_size) + the model's log-probabilities over the vocabulary at the current step + + Return: A tuple of (trimed_probs, truncated_indices) where: + trimed_probs: (bsz x input_beam_size x ?) + the model's probabilities over the elements selected to sample from. The + width of the third dimension is determined by top-P. + truncated_indices: (bsz x input_beam_size x ?) + the indices of the chosen elements. + """ + probs = lprobs.exp_() + + # sort the last dimension (vocab dimension) in descending order + sorted_probs, sorted_indices = probs.sort(descending=True) + + # compute a mask to indicate the words to be included in the top-P set. + cumsum_probs = sorted_probs.cumsum(dim=2) + mask = cumsum_probs.lt(self.sampling_topp) + + # note that mask was computed by 'lt'. One more word needs to be included + # so that the cumulative probability mass can exceed p. + cumsum_mask = mask.cumsum(dim=2) + last_included = cumsum_mask[:, :, -1:] + last_included.clamp_(0, mask.size()[2] - 1) + mask = mask.scatter_(2, last_included, 1) + + # truncate unnecessary dims. + max_dim = last_included.max() + truncated_mask = mask[:, :, : max_dim + 1] + truncated_probs = sorted_probs[:, :, : max_dim + 1] + truncated_indices = sorted_indices[:, :, : max_dim + 1] + + # trim the words that are not in top-P by setting their probabilities + # to 0, so that they would not be sampled later. + trim_mask = ~truncated_mask + trimed_probs = truncated_probs.masked_fill_(trim_mask, 0) + return trimed_probs, truncated_indices + + @torch.jit.export + def step(self, step: int, lprobs, scores): + bsz, beam_size, vocab_size = lprobs.size() + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + lprobs = lprobs[:, ::beam_size, :].contiguous() + + if self.sampling_topp > 0: + # only sample from the smallest set of words whose cumulative probability mass exceeds p + probs, top_indices = self._sample_topp(lprobs) + elif self.sampling_topk > 0: + # only sample from top-k candidates + lprobs, top_indices = lprobs.topk(self.sampling_topk) + probs = lprobs.exp_() + else: + probs = lprobs.exp_() + + # dummy data to be consistent with true branch for type check + top_indices = torch.empty(0).to(probs) + # sample + if step == 0: + indices_buf = torch.multinomial( + probs.view(bsz, -1), beam_size, replacement=True, + ).view(bsz, beam_size) + else: + indices_buf = torch.multinomial( + probs.view(bsz * beam_size, -1), + 1, + replacement=True, + ).view(bsz, beam_size) + + if step == 0: + # expand to beam size + probs = probs.expand(bsz, beam_size, -1) + + # gather scores + scores_buf = torch.gather( + probs, dim=2, index=indices_buf.unsqueeze(-1) + ) + scores_buf = scores_buf.log_().view(bsz, -1) + + # remap indices if using top-k or top-P sampling + if self.sampling_topk > 0 or self.sampling_topp > 0: + indices_buf = torch.gather( + top_indices.expand(bsz, beam_size, -1), + dim=2, + index=indices_buf.unsqueeze(-1), + ).squeeze(2) + + if step == 0: + beams_buf = indices_buf.new_zeros(bsz, beam_size) + else: + beams_buf = torch.arange(0, beam_size).to(indices_buf).repeat(bsz, 1) + # make scores cumulative + scores_buf.add_( + torch.gather(scores[:, :, step - 1], dim=1, index=beams_buf) + ) + + return scores_buf, indices_buf, beams_buf + + +class DiverseSiblingsSearch(Search): + """ + Beam search with diverse siblings. + + See "A Simple, Fast Diverse Decoding Algorithm for Neural Generation" for details. + https://arxiv.org/abs/1611.08562 + + 1/ Calculate hypotheses for each beam + 2/ Intra-sibling ordering + 3/ Rewrite scores + 4/ Choose top K hypotheses + + if diversity_rate == 0 is equivalent to BeamSearch + """ + + def __init__(self, tgt_dict, diversity_rate): + super().__init__(tgt_dict) + self.diversity_rate = diversity_rate + self.beam = BeamSearch(tgt_dict) + + def step(self, step: int, lprobs, scores): + bsz, beam_size, vocab_size = lprobs.size() + k = min( + # Take the best 2 x beam_size predictions. We'll choose the first + # beam_size of these which don't predict eos to continue with. + beam_size * 2, + lprobs.view(bsz, -1).size(1) - 1, # -1 so we never select pad + ) + s_list: List[Tensor] + i_list: List[Tensor] + s_list = [torch.empty(0).to(lprobs) for i in range(beam_size)] + i_list = [torch.LongTensor().to(device=lprobs.device) for i in range(beam_size)] + sibling_score = torch.arange(1, k + 1).to(lprobs) * self.diversity_rate + + if step == 0: + return self.beam.step(step, lprobs, scores) + lprobs.add_(scores[:, :, step - 1].unsqueeze(-1)) + + # 1/ Calculate hypotheses for each beam + for i in range(beam_size): + torch.topk(lprobs[:, i, :].view(bsz, -1), k, out=(s_list[i], i_list[i])) + i_list[i].fmod_(vocab_size) + + # 2/ Intra-sibling ordering by default from topk + 3/ Rewrite scores + s_list[i].sub_(sibling_score) + + # 4/ Choose top K hypotheses + indices = torch.stack(i_list, dim=1).view(bsz, -1) + + final_scores = torch.empty(0).to(lprobs) + final_indices = torch.LongTensor().to(device=lprobs.device) + final_beams = torch.LongTensor().to(device=lprobs.device) + (final_scores, final_indices) = torch.topk( + torch.stack(s_list, dim=1).view(bsz, -1), + k, + ) + + final_beams = final_indices // k + + for i in range(bsz): + final_indices[i] = indices[i][final_indices[i]] + + return final_scores, final_indices, final_beams diff --git a/fairseq/sequence_generator.py b/fairseq/sequence_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..42012fbbb129033d848eabe0fe79f83f27e62089 --- /dev/null +++ b/fairseq/sequence_generator.py @@ -0,0 +1,919 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, List, Optional + +import torch +import torch.nn as nn +from fairseq import search, utils +from fairseq.data import data_utils +from fairseq.models import FairseqIncrementalDecoder +from fairseq.models.fairseq_encoder import EncoderOut +from torch import Tensor + + +class SequenceGenerator(nn.Module): + def __init__( + self, + models, + tgt_dict, + beam_size=1, + max_len_a=0, + max_len_b=200, + min_len=1, + normalize_scores=True, + len_penalty=1.0, + unk_penalty=0.0, + temperature=1.0, + match_source_len=False, + no_repeat_ngram_size=0, + search_strategy=None, + eos=None, + symbols_to_strip_from_output=None, + ): + """Generates translations of a given source sentence. + + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models, + currently support fairseq.models.TransformerModel for scripting + beam_size (int, optional): beam width (default: 1) + max_len_a/b (int, optional): generate sequences of maximum length + ax + b, where x is the source length + min_len (int, optional): the minimum length of the generated output + (not including end-of-sentence) + normalize_scores (bool, optional): normalize scores by the length + of the output (default: True) + len_penalty (float, optional): length penalty, where <1.0 favors + shorter, >1.0 favors longer sentences (default: 1.0) + unk_penalty (float, optional): unknown word penalty, where <0 + produces more unks, >0 produces fewer (default: 0.0) + temperature (float, optional): temperature, where values + >1.0 produce more uniform samples and values <1.0 produce + sharper samples (default: 1.0) + match_source_len (bool, optional): outputs should match the source + length (default: False) + """ + super().__init__() + if isinstance(models, EnsembleModel): + self.model = models + else: + self.model = EnsembleModel(models) + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() if eos is None else eos + self.symbols_to_strip_from_output = ( + symbols_to_strip_from_output.union({self.eos}) + if symbols_to_strip_from_output is not None else {self.eos}) + self.vocab_size = len(tgt_dict) + self.beam_size = beam_size + # the max beam size is the dictionary size - 1, since we never select pad + self.beam_size = min(beam_size, self.vocab_size - 1) + self.max_len_a = max_len_a + self.max_len_b = max_len_b + self.min_len = min_len + + self.normalize_scores = normalize_scores + self.len_penalty = len_penalty + self.unk_penalty = unk_penalty + self.temperature = temperature + self.match_source_len = match_source_len + self.no_repeat_ngram_size = no_repeat_ngram_size + assert temperature > 0, "--temperature must be greater than 0" + + self.search = ( + search.BeamSearch(tgt_dict) if search_strategy is None else search_strategy + ) + # We only need to set src_lengths in LengthConstrainedBeamSearch. + # As a module attribute, setting it would break in multithread + # settings when the model is shared. + self.should_set_src_lengths = hasattr(self.search, 'needs_src_lengths') and self.search.needs_src_lengths + + self.model.eval() + + def cuda(self): + self.model.cuda() + return self + + @torch.no_grad() + def forward( + self, + sample: Dict[str, Dict[str, Tensor]], + prefix_tokens: Optional[Tensor] = None, + bos_token: Optional[int] = None, + ): + """Generate a batch of translations. + + Args: + sample (dict): batch + prefix_tokens (torch.LongTensor, optional): force decoder to begin + with these tokens + bos_token (int, optional): beginning of sentence token + (default: self.eos) + """ + return self._generate(sample, prefix_tokens, bos_token) + + # TODO(myleott): unused, deprecate after pytorch-translate migration + def generate_batched_itr(self, data_itr, beam_size=None, cuda=False, timer=None): + """Iterate over a batched dataset and yield individual translations. + Args: + cuda (bool, optional): use GPU for generation + timer (StopwatchMeter, optional): time generations + """ + for sample in data_itr: + s = utils.move_to_cuda(sample) if cuda else sample + if "net_input" not in s: + continue + input = s["net_input"] + # model.forward normally channels prev_output_tokens into the decoder + # separately, but SequenceGenerator directly calls model.encoder + encoder_input = { + k: v for k, v in input.items() if k != "prev_output_tokens" + } + if timer is not None: + timer.start() + with torch.no_grad(): + hypos = self.generate(encoder_input) + if timer is not None: + timer.stop(sum(len(h[0]["tokens"]) for h in hypos)) + for i, id in enumerate(s["id"].data): + # remove padding + src = utils.strip_pad(input["src_tokens"].data[i, :], self.pad) + ref = ( + utils.strip_pad(s["target"].data[i, :], self.pad) + if s["target"] is not None + else None + ) + yield id, src, ref, hypos[i] + + @torch.no_grad() + def generate(self, models, sample: Dict[str, Dict[str, Tensor]], **kwargs): + """Generate translations. Match the api of other fairseq generators. + + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models + sample (dict): batch + prefix_tokens (torch.LongTensor, optional): force decoder to begin + with these tokens + bos_token (int, optional): beginning of sentence token + (default: self.eos) + """ + return self._generate(sample, **kwargs) + + def _generate( + self, + sample: Dict[str, Dict[str, Tensor]], + prefix_tokens: Optional[Tensor] = None, + bos_token: Optional[int] = None, + ): + incremental_states = torch.jit.annotate( + List[Dict[str, Dict[str, Optional[Tensor]]]], + [ + torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {}) + for i in range(self.model.models_size) + ], + ) + net_input = sample["net_input"] + + if 'src_tokens' in net_input: + src_tokens = net_input['src_tokens'] + # length of the source text being the character length except EndOfSentence and pad + src_lengths = (src_tokens.ne(self.eos) & src_tokens.ne(self.pad)).long().sum(dim=1) + elif 'source' in net_input: + src_tokens = net_input['source'] + src_lengths = net_input['padding_mask'].size(-1) - net_input['padding_mask'].sum(-1) if net_input['padding_mask'] is not None else torch.tensor(src_tokens.size(-1)) + else: + raise Exception('expected src_tokens or source in net input') + + # bsz: total number of sentences in beam + input_size = src_tokens.size() + bsz, src_len = input_size[0], input_size[1] + beam_size = self.beam_size + + max_len: int = -1 + if self.match_source_len: + max_len = src_lengths.max().item() + else: + max_len = min( + int(self.max_len_a * src_len + self.max_len_b), + # exclude the EOS marker + self.model.max_decoder_positions() - 1, + ) + assert ( + self.min_len <= max_len + ), "min_len cannot be larger than max_len, please adjust these!" + # compute the encoder output for each beam + encoder_outs = self.model.forward_encoder(net_input) + + # placeholder of indices for bsz * beam_size to hold tokens and accumulative scores + new_order = torch.arange(bsz).view(-1, 1).repeat(1, beam_size).view(-1) + new_order = new_order.to(src_tokens.device).long() + encoder_outs = self.model.reorder_encoder_out(encoder_outs, new_order) + # ensure encoder_outs is a List. + assert encoder_outs is not None + + # initialize buffers + scores = ( + torch.zeros(bsz * beam_size, max_len + 1).to(src_tokens).float() + ) # +1 for eos; pad is never choosed for scoring + tokens = ( + torch.zeros(bsz * beam_size, max_len + 2) + .to(src_tokens) + .long() + .fill_(self.pad) + ) # +2 for eos and pad + tokens[:, 0] = self.eos if bos_token is None else bos_token + attn: Optional[Tensor] = None + + # A list that indicates candidates that should be ignored. + # For example, suppose we're sampling and have already finalized 2/5 + # samples. Then cands_to_ignore would mark 2 positions as being ignored, + # so that we only finalize the remaining 3 samples. + cands_to_ignore = ( + torch.zeros(bsz, beam_size).to(src_tokens).eq(-1) + ) # forward and backward-compatible False mask + + # list of completed sentences + finalized = torch.jit.annotate( + List[List[Dict[str, Tensor]]], + [torch.jit.annotate(List[Dict[str, Tensor]], []) for i in range(bsz)], + ) # contains lists of dictionaries of infomation about the hypothesis being finalized at each step + + finished = [ + False for i in range(bsz) + ] # a boolean array indicating if the sentence at the index is finished or not + num_remaining_sent = bsz # number of sentences remaining + + # number of candidate hypos per step + cand_size = 2 * beam_size # 2 x beam size in case half are EOS + + # offset arrays for converting between different indexing schemes + bbsz_offsets = (torch.arange(0, bsz) * beam_size).unsqueeze(1).type_as(tokens) + cand_offsets = torch.arange(0, cand_size).type_as(tokens) + + reorder_state: Optional[Tensor] = None + batch_idxs: Optional[Tensor] = None + for step in range(max_len + 1): # one extra step for EOS marker + # reorder decoder internal states based on the prev choice of beams + # print(f'step: {step}') + if reorder_state is not None: + if batch_idxs is not None: + # update beam indices to take into account removed sentences + corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as( + batch_idxs + ) + reorder_state.view(-1, beam_size).add_( + corr.unsqueeze(-1) * beam_size + ) + self.model.reorder_incremental_state(incremental_states, reorder_state) + encoder_outs = self.model.reorder_encoder_out( + encoder_outs, reorder_state + ) + + lprobs, avg_attn_scores = self.model.forward_decoder( + tokens[:, : step + 1], + encoder_outs, + incremental_states, + self.temperature, + ) + lprobs[lprobs != lprobs] = torch.tensor(-math.inf).to(lprobs) + + lprobs[:, self.pad] = -math.inf # never select pad + lprobs[:, self.unk] -= self.unk_penalty # apply unk penalty + + # handle max length constraint + if step >= max_len: + lprobs[:, : self.eos] = -math.inf + lprobs[:, self.eos + 1 :] = -math.inf + + # handle prefix tokens (possibly with different lengths) + if ( + prefix_tokens is not None + and step < prefix_tokens.size(1) + and step < max_len + ): + lprobs, tokens, scores = self._prefix_tokens( + step, lprobs, scores, tokens, prefix_tokens, beam_size + ) + elif step < self.min_len: + # minimum length constraint (does not apply if using prefix_tokens) + lprobs[:, self.eos] = -math.inf + + # Record attention scores, only support avg_attn_scores is a Tensor + if avg_attn_scores is not None: + if attn is None: + attn = torch.empty( + bsz * beam_size, avg_attn_scores.size(1), max_len + 2 + ).to(scores) + attn[:, :, step + 1].copy_(avg_attn_scores) + + scores = scores.type_as(lprobs) + eos_bbsz_idx = torch.empty(0).to( + tokens + ) # indices of hypothesis ending with eos (finished sentences) + eos_scores = torch.empty(0).to( + scores + ) # scores of hypothesis ending with eos (finished sentences) + + if self.should_set_src_lengths: + self.search.set_src_lengths(src_lengths) + + if self.no_repeat_ngram_size > 0: + lprobs = self._no_repeat_ngram(tokens, lprobs, bsz, beam_size, step) + + cand_scores, cand_indices, cand_beams = self.search.step( + step, + lprobs.view(bsz, -1, self.vocab_size), + scores.view(bsz, beam_size, -1)[:, :, :step], + ) + + # cand_bbsz_idx contains beam indices for the top candidate + # hypotheses, with a range of values: [0, bsz*beam_size), + # and dimensions: [bsz, cand_size] + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + + # finalize hypotheses that end in eos + eos_mask = cand_indices.eq(self.eos) & cand_scores.ne(-math.inf) + eos_mask[:, :beam_size][cands_to_ignore] = torch.tensor(0).to(eos_mask) + + # only consider eos when it's among the top beam_size indices + eos_bbsz_idx = torch.masked_select( + cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size] + ) + + finalized_sents: List[int] = [] + if eos_bbsz_idx.numel() > 0: + eos_scores = torch.masked_select( + cand_scores[:, :beam_size], mask=eos_mask[:, :beam_size] + ) + finalized_sents = self.finalize_hypos( + step, + eos_bbsz_idx, + eos_scores, + tokens, + scores, + finalized, + finished, + beam_size, + attn, + src_lengths, + max_len, + ) + num_remaining_sent -= len(finalized_sents) + + assert num_remaining_sent >= 0 + if num_remaining_sent == 0: + break + assert step < max_len + + if len(finalized_sents) > 0: + new_bsz = bsz - len(finalized_sents) + + # construct batch_idxs which holds indices of batches to keep for the next pass + batch_mask = torch.ones(bsz).to(cand_indices) + batch_mask[ + torch.tensor(finalized_sents).to(cand_indices) + ] = torch.tensor(0).to(batch_mask) + batch_idxs = batch_mask.nonzero().squeeze(-1) + + eos_mask = eos_mask[batch_idxs] + cand_beams = cand_beams[batch_idxs] + bbsz_offsets.resize_(new_bsz, 1) + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + cand_scores = cand_scores[batch_idxs] + cand_indices = cand_indices[batch_idxs] + + if prefix_tokens is not None: + prefix_tokens = prefix_tokens[batch_idxs] + src_lengths = src_lengths[batch_idxs] + cands_to_ignore = cands_to_ignore[batch_idxs] + + scores = scores.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + tokens = tokens.view(bsz, -1)[batch_idxs].view(new_bsz * beam_size, -1) + if attn is not None: + attn = attn.view(bsz, -1)[batch_idxs].view( + new_bsz * beam_size, attn.size(1), -1 + ) + bsz = new_bsz + else: + batch_idxs = None + # set active_mask so that values > cand_size indicate eos hypos + # and values < cand_size indicate candidate active hypos. + # After, the min values per row are the top candidate active hypos + + # Rewrite the operator since the element wise or is not supported in torchscript. + + eos_mask[:, :beam_size] = ~((~cands_to_ignore) & (~eos_mask[:, :beam_size])) + active_mask = torch.add( + eos_mask.type_as(cand_offsets) * cand_size, + cand_offsets[: eos_mask.size(1)], + ) + + # get the top beam_size active hypotheses, which are just the hypos + # with the smallest values in active_mask + new_cands_to_ignore, active_hypos = torch.topk( + active_mask, k=beam_size, dim=1, largest=False + ) + + # update cands_to_ignore to ignore any finalized hypos + cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size] + assert (~cands_to_ignore).any(dim=1).all() + + active_bbsz_idx = torch.gather(cand_bbsz_idx, dim=1, index=active_hypos) + active_scores = torch.gather(cand_scores, dim=1, index=active_hypos) + + active_bbsz_idx = active_bbsz_idx.view(-1) + active_scores = active_scores.view(-1) + + # copy tokens and scores for active hypotheses + tokens[:, : step + 1] = torch.index_select( + tokens[:, : step + 1], dim=0, index=active_bbsz_idx + ) + tokens.view(bsz, beam_size, -1)[:, :, step + 1] = torch.gather( + cand_indices, dim=1, index=active_hypos + ) + if step > 0: + scores[:, :step] = torch.index_select( + scores[:, :step], dim=0, index=active_bbsz_idx + ) + scores.view(bsz, beam_size, -1)[:, :, step] = torch.gather( + cand_scores, dim=1, index=active_hypos + ) + + # copy attention for active hypotheses + if attn is not None: + attn[:, :, : step + 2] = torch.index_select( + attn[:, :, : step + 2], dim=0, index=active_bbsz_idx + ) + + # reorder incremental state in decoder + reorder_state = active_bbsz_idx + + # sort by score descending + for sent in range(len(finalized)): + # make into beam container + BCList = [ + BeamContainer(elem["score"].item(), elem) for elem in finalized[sent] + ] + BCList.sort() + BCList.reverse() + finalized[sent] = torch.jit.annotate( + List[Dict[str, Tensor]], [x.elem for x in BCList] + ) + + return finalized + + def _prefix_tokens( + self, step: int, lprobs, scores, tokens, prefix_tokens, beam_size: int + ): + """Handle prefix tokens""" + prefix_toks = prefix_tokens[:, step].unsqueeze(-1).repeat(1, beam_size).view(-1) + prefix_lprobs = lprobs.gather(-1, prefix_toks.unsqueeze(-1)) + prefix_mask = prefix_toks.ne(self.pad) + lprobs[prefix_mask] = torch.tensor(-math.inf).to(lprobs) + lprobs[prefix_mask] = lprobs[prefix_mask].scatter( + -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_lprobs[prefix_mask] + ) + # if prefix includes eos, then we should make sure tokens and + # scores are the same across all beams + eos_mask = prefix_toks.eq(self.eos) + if eos_mask.any(): + # validate that the first beam matches the prefix + first_beam = tokens[eos_mask].view(-1, beam_size, tokens.size(-1))[ + :, 0, 1 : step + 1 + ] + eos_mask_batch_dim = eos_mask.view(-1, beam_size)[:, 0] + target_prefix = prefix_tokens[eos_mask_batch_dim][:, :step] + assert (first_beam == target_prefix).all() + + # copy tokens, scores and lprobs from the first beam to all beams + tokens = self.replicate_first_beam(tokens, eos_mask_batch_dim, beam_size) + scores = self.replicate_first_beam(scores, eos_mask_batch_dim, beam_size) + lprobs = self.replicate_first_beam(lprobs, eos_mask_batch_dim, beam_size) + return lprobs, tokens, scores + + def replicate_first_beam(self, tensor, mask, beam_size: int): + tensor = tensor.view(-1, beam_size, tensor.size(-1)) + tensor[mask] = tensor[mask][:, :1, :] + return tensor.view(-1, tensor.size(-1)) + + def finalize_hypos( + self, + step: int, + bbsz_idx, + eos_scores, + tokens, + scores, + finalized: List[List[Dict[str, Tensor]]], + finished: List[bool], + beam_size: int, + attn: Optional[Tensor], + src_lengths, + max_len: int, + ): + """Finalize hypothesis, store finalized information in `finalized`, and change `finished` accordingly. + Returns number of sentences being finalized. + Args: + bbsz_idx (Tensor): + """ + assert bbsz_idx.numel() == eos_scores.numel() + + # clone relevant token and attention tensors + tokens_clone = tokens.index_select(0, bbsz_idx)[ + :, 1 : step + 2 + ] # skip the first index, which is EOS + + tokens_clone[:, step] = self.eos + attn_clone = ( + attn.index_select(0, bbsz_idx)[:, :, 1 : step + 2] + if attn is not None + else None + ) + + # compute scores per token position + pos_scores = scores.index_select(0, bbsz_idx)[:, : step + 1] + pos_scores[:, step] = eos_scores + # convert from cumulative to per-position scores + pos_scores[:, 1:] = pos_scores[:, 1:] - pos_scores[:, :-1] + + # normalize sentence-level scores + if self.normalize_scores: + eos_scores /= (step + 1) ** self.len_penalty + + cum_unfin: List[int] = [] + prev = 0 + for f in finished: + if f: + prev += 1 + else: + cum_unfin.append(prev) + + # set() is not supported in script export + sents_seen: Dict[str, Optional[Tensor]] = {} + for i in range(bbsz_idx.size()[0]): + idx = bbsz_idx[i] + score = eos_scores[i] + unfin_idx = idx // beam_size + sent = unfin_idx + cum_unfin[unfin_idx] + # Cannot create dict for key type '(int, int)' in torchscript. + # The workaround is to cast int to string + seen = str(sent.item()) + "_" + str(unfin_idx.item()) + if seen not in sents_seen: + sents_seen[seen] = None + + if self.match_source_len and step > src_lengths[unfin_idx]: + score = torch.tensor(-math.inf).to(score) + + if len(finalized[sent]) < beam_size: + if attn_clone is not None: + # remove padding tokens from attn scores + hypo_attn = attn_clone[i] + else: + hypo_attn = torch.empty(0) + finalized[sent].append( + { + "tokens": tokens_clone[i], + "score": score, + "attention": hypo_attn, # src_len x tgt_len + "alignment": torch.empty(0), + "positional_scores": pos_scores[i], + } + ) + + newly_finished: List[int] = [] + for seen in sents_seen.keys(): + # check termination conditions for this sentence + sent: int = int(float(seen.split("_")[0])) + unfin_idx: int = int(float(seen.split("_")[1])) + if not finished[sent] and self.is_finished( + step, unfin_idx, max_len, len(finalized[sent]), beam_size + ): + finished[sent] = True + newly_finished.append(unfin_idx) + return newly_finished + + def is_finished( + self, + step: int, + unfin_idx: int, + max_len: int, + finalized_sent_len: int, + beam_size: int, + ): + """ + Check whether we've finished generation for a given sentence, by + comparing the worst score among finalized hypotheses to the best + possible score among unfinalized hypotheses. + """ + assert finalized_sent_len <= beam_size + if finalized_sent_len == beam_size or step == max_len: + return True + return False + + def calculate_banned_tokens( + self, + tokens, + step: int, + gen_ngrams: List[Dict[str, List[int]]], + no_repeat_ngram_size: int, + bbsz_idx: int, + ): + tokens_list: List[int] = tokens[ + bbsz_idx, step + 2 - no_repeat_ngram_size : step + 1 + ].tolist() + # before decoding the next token, prevent decoding of ngrams that have already appeared + ngram_index = ",".join([str(x) for x in tokens_list]) + return gen_ngrams[bbsz_idx].get(ngram_index, torch.jit.annotate(List[int], [])) + + def transpose_list(self, l: List[List[int]]): + # GeneratorExp aren't supported in TS so ignoring the lint + min_len = min([len(x) for x in l]) # noqa + l2 = [[row[i] for row in l] for i in range(min_len)] + return l2 + + def _no_repeat_ngram(self, tokens, lprobs, bsz: int, beam_size: int, step: int): + # for each beam and batch sentence, generate a list of previous ngrams + gen_ngrams: List[Dict[str, List[int]]] = [ + torch.jit.annotate(Dict[str, List[int]], {}) + for bbsz_idx in range(bsz * beam_size) + ] + cpu_tokens = tokens.cpu() + for bbsz_idx in range(bsz * beam_size): + gen_tokens: List[int] = cpu_tokens[bbsz_idx].tolist() + for ngram in self.transpose_list( + [gen_tokens[i:] for i in range(self.no_repeat_ngram_size)] + ): + key = ",".join([str(x) for x in ngram[:-1]]) + gen_ngrams[bbsz_idx][key] = gen_ngrams[bbsz_idx].get( + key, torch.jit.annotate(List[int], []) + ) + [ngram[-1]] + + if step + 2 - self.no_repeat_ngram_size >= 0: + # no banned tokens if we haven't generated no_repeat_ngram_size tokens yet + banned_tokens = [ + self.calculate_banned_tokens( + tokens, step, gen_ngrams, self.no_repeat_ngram_size, bbsz_idx + ) + for bbsz_idx in range(bsz * beam_size) + ] + else: + banned_tokens = [ + torch.jit.annotate(List[int], []) for bbsz_idx in range(bsz * beam_size) + ] + for bbsz_idx in range(bsz * beam_size): + lprobs[bbsz_idx][ + torch.tensor(banned_tokens[bbsz_idx]).long() + ] = torch.tensor(-math.inf, dtype=torch.float) + return lprobs + + +class EnsembleModel(nn.Module): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__() + self.models_size = len(models) + # method '__len__' is not supported in ModuleList for torch script + self.single_model = models[0] + self.models = nn.ModuleList(models) + + self.has_incremental: bool = False + if all( + hasattr(m, "decoder") and isinstance(m.decoder, FairseqIncrementalDecoder) + for m in models + ): + self.has_incremental = True + + def forward(self): + pass + + def has_encoder(self): + return hasattr(self.single_model, "encoder") + + def has_incremental_states(self): + return self.has_incremental + + def max_decoder_positions(self): + return min([m.max_decoder_positions() for m in self.models]) + + @torch.jit.export + def forward_encoder(self, net_input: Dict[str, Tensor]): + if not self.has_encoder(): + return None + return [ + model.encoder.forward_torchscript(net_input) + for model in self.models + ] + + @torch.jit.export + def forward_decoder( + self, + tokens, + encoder_outs: List[EncoderOut], + incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]], + temperature: float = 1.0, + ): + log_probs = [] + avg_attn: Optional[Tensor] = None + encoder_out: Optional[EncoderOut] = None + for i, model in enumerate(self.models): + if self.has_encoder(): + encoder_out = encoder_outs[i] + # decode each model + if self.has_incremental_states(): + decoder_out = model.decoder.forward( + tokens, + encoder_out=encoder_out, + incremental_state=incremental_states[i], + ) + else: + decoder_out = model.decoder.forward(tokens, encoder_out=encoder_out) + + attn: Optional[Tensor] = None + decoder_len = len(decoder_out) + if decoder_len > 1 and decoder_out[1] is not None: + if isinstance(decoder_out[1], Tensor): + attn = decoder_out[1] + else: + attn_holder = decoder_out[1]["attn"] + if isinstance(attn_holder, Tensor): + attn = attn_holder + elif attn_holder is not None: + attn = attn_holder[0] + if attn is not None: + attn = attn[:, -1, :] + + decoder_out_tuple = ( + decoder_out[0][:, -1:, :].div_(temperature), + None if decoder_len <= 1 else decoder_out[1], + ) + + probs = model.get_normalized_probs( + decoder_out_tuple, log_probs=True, sample=None + ) + probs = probs[:, -1, :] + if self.models_size == 1: + return probs, attn + + log_probs.append(probs) + if attn is not None: + if avg_attn is None: + avg_attn = attn + else: + avg_attn.add_(attn) + avg_probs = torch.logsumexp(torch.stack(log_probs, dim=0), dim=0) - math.log( + self.models_size + ) + if avg_attn is not None: + avg_attn.div_(self.models_size) + return avg_probs, avg_attn + + @torch.jit.export + def reorder_encoder_out(self, encoder_outs: Optional[List[EncoderOut]], new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + new_outs: List[EncoderOut] = [] + if not self.has_encoder(): + return new_outs + for i, model in enumerate(self.models): + assert encoder_outs is not None + new_outs.append( + model.encoder.reorder_encoder_out(encoder_outs[i], new_order) + ) + return new_outs + + @torch.jit.export + def reorder_incremental_state( + self, + incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]], + new_order, + ): + if not self.has_incremental_states(): + return + for i, model in enumerate(self.models): + model.decoder.reorder_incremental_state_scripting( + incremental_states[i], new_order + ) + + +class SequenceGeneratorWithAlignment(SequenceGenerator): + def __init__(self, models, tgt_dict, left_pad_target=False, **kwargs): + """Generates translations of a given source sentence. + + Produces alignments following "Jointly Learning to Align and + Translate with Transformer Models" (Garg et al., EMNLP 2019). + + Args: + left_pad_target (bool, optional): Whether or not the + hypothesis should be left padded or not when they are + teacher forced for generating alignments. + """ + super().__init__(EnsembleModelWithAlignment(models), tgt_dict, **kwargs) + self.left_pad_target = left_pad_target + + @torch.no_grad() + def generate(self, models, sample, **kwargs): + finalized = super()._generate(sample, **kwargs) + + src_tokens = sample["net_input"]["src_tokens"] + bsz = src_tokens.shape[0] + beam_size = self.beam_size + src_tokens, src_lengths, prev_output_tokens, tgt_tokens = self._prepare_batch_for_alignment( + sample, finalized + ) + if any(getattr(m, "full_context_alignment", False) for m in self.model.models): + attn = self.model.forward_align(src_tokens, src_lengths, prev_output_tokens) + else: + attn = [ + finalized[i // beam_size][i % beam_size]["attention"].transpose(1, 0) + for i in range(bsz * beam_size) + ] + + if src_tokens.device != "cpu": + src_tokens = src_tokens.to('cpu') + tgt_tokens = tgt_tokens.to('cpu') + attn = [i.to('cpu') for i in attn] + + # Process the attn matrix to extract hard alignments. + for i in range(bsz * beam_size): + alignment = utils.extract_hard_alignment( + attn[i], src_tokens[i], tgt_tokens[i], self.pad, self.eos + ) + finalized[i // beam_size][i % beam_size]["alignment"] = alignment + return finalized + + def _prepare_batch_for_alignment(self, sample, hypothesis): + src_tokens = sample["net_input"]["src_tokens"] + bsz = src_tokens.shape[0] + src_tokens = ( + src_tokens[:, None, :] + .expand(-1, self.beam_size, -1) + .contiguous() + .view(bsz * self.beam_size, -1) + ) + src_lengths = sample["net_input"]["src_lengths"] + src_lengths = ( + src_lengths[:, None] + .expand(-1, self.beam_size) + .contiguous() + .view(bsz * self.beam_size) + ) + prev_output_tokens = data_utils.collate_tokens( + [beam["tokens"] for example in hypothesis for beam in example], + self.pad, + self.eos, + self.left_pad_target, + move_eos_to_beginning=True, + ) + tgt_tokens = data_utils.collate_tokens( + [beam["tokens"] for example in hypothesis for beam in example], + self.pad, + self.eos, + self.left_pad_target, + move_eos_to_beginning=False, + ) + return src_tokens, src_lengths, prev_output_tokens, tgt_tokens + + +class EnsembleModelWithAlignment(EnsembleModel): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__(models) + + def forward_align(self, src_tokens, src_lengths, prev_output_tokens): + avg_attn = None + for model in self.models: + decoder_out = model(src_tokens, src_lengths, prev_output_tokens) + attn = decoder_out[1]["attn"] + if avg_attn is None: + avg_attn = attn + else: + avg_attn.add_(attn) + if len(self.models) > 1: + avg_attn.div_(len(self.models)) + return avg_attn + + +@torch.jit.script +class BeamContainer(object): + def __init__(self, score: float, elem: Dict[str, Tensor]): + self.score = score + self.elem = elem + + def __lt__(self, other): + # type: (BeamContainer) -> bool + # Due to https://github.com/pytorch/pytorch/issues/20388, + # this has to use old style type annotations + # Match original behavior of sorted function when two scores are equal. + return self.score <= other.score diff --git a/fairseq/sequence_scorer.py b/fairseq/sequence_scorer.py new file mode 100644 index 0000000000000000000000000000000000000000..343c29acc2292d2d1a86cbed3af035c039b4c36f --- /dev/null +++ b/fairseq/sequence_scorer.py @@ -0,0 +1,133 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import sys + +from fairseq import utils + + +class SequenceScorer(object): + """Scores the target for a given source sentence.""" + + def __init__( + self, tgt_dict, softmax_batch=None, compute_alignment=False, eos=None, + symbols_to_strip_from_output=None, + ): + self.pad = tgt_dict.pad() + self.eos = tgt_dict.eos() if eos is None else eos + self.softmax_batch = softmax_batch or sys.maxsize + assert self.softmax_batch > 0 + self.compute_alignment = compute_alignment + self.symbols_to_strip_from_output = ( + symbols_to_strip_from_output.union({self.eos}) + if symbols_to_strip_from_output is not None else {self.eos}) + + @torch.no_grad() + def generate(self, models, sample, **kwargs): + """Score a batch of translations.""" + net_input = sample['net_input'] + + def batch_for_softmax(dec_out, target): + # assumes decoder_out[0] is the only thing needed (may not be correct for future models!) + first, rest = dec_out[0], dec_out[1:] + bsz, tsz, dim = first.shape + if bsz * tsz < self.softmax_batch: + yield dec_out, target, True + else: + flat = first.contiguous().view(1, -1, dim) + flat_tgt = target.contiguous().view(flat.shape[:-1]) + s = 0 + while s < flat.size(1): + e = s + self.softmax_batch + yield (flat[:, s:e],) + rest, flat_tgt[:, s:e], False + s = e + + def gather_target_probs(probs, target): + probs = probs.gather( + dim=2, + index=target.unsqueeze(-1), + ) + return probs + + orig_target = sample['target'] + + # compute scores for each model in the ensemble + avg_probs = None + avg_attn = None + for model in models: + model.eval() + decoder_out = model(**net_input) + attn = decoder_out[1] if len(decoder_out) > 1 else None + if type(attn) is dict: + attn = attn.get('attn', None) + + batched = batch_for_softmax(decoder_out, orig_target) + probs, idx = None, 0 + for bd, tgt, is_single in batched: + sample['target'] = tgt + curr_prob = model.get_normalized_probs(bd, log_probs=len(models) == 1, sample=sample).data + if is_single: + probs = gather_target_probs(curr_prob, orig_target) + else: + if probs is None: + probs = curr_prob.new(orig_target.numel()) + step = curr_prob.size(0) * curr_prob.size(1) + end = step + idx + tgt_probs = gather_target_probs(curr_prob.view(tgt.shape + (curr_prob.size(-1),)), tgt) + probs[idx:end] = tgt_probs.view(-1) + idx = end + sample['target'] = orig_target + + probs = probs.view(sample['target'].shape) + + if avg_probs is None: + avg_probs = probs + else: + avg_probs.add_(probs) + if attn is not None and torch.is_tensor(attn): + attn = attn.data + if avg_attn is None: + avg_attn = attn + else: + avg_attn.add_(attn) + if len(models) > 1: + avg_probs.div_(len(models)) + avg_probs.log_() + if avg_attn is not None: + avg_attn.div_(len(models)) + + bsz = avg_probs.size(0) + hypos = [] + start_idxs = sample['start_indices'] if 'start_indices' in sample else [0] * bsz + for i in range(bsz): + # remove padding from ref + ref = utils.strip_pad(sample['target'][i, start_idxs[i]:], self.pad) \ + if sample['target'] is not None else None + tgt_len = ref.numel() + avg_probs_i = avg_probs[i][start_idxs[i]:start_idxs[i] + tgt_len] + score_i = avg_probs_i.sum() / tgt_len + if avg_attn is not None: + avg_attn_i = avg_attn[i] + if self.compute_alignment: + alignment = utils.extract_hard_alignment( + avg_attn_i, + sample['net_input']['src_tokens'][i], + sample['target'][i], + self.pad, + self.eos, + ) + else: + alignment = None + else: + avg_attn_i = alignment = None + hypos.append([{ + 'tokens': ref, + 'score': score_i, + 'attention': avg_attn_i, + 'alignment': alignment, + 'positional_scores': avg_probs_i, + }]) + return hypos diff --git a/fairseq/tasks/__init__.py b/fairseq/tasks/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b1bb404f1c31cb8ce02bdf88e5063da209853151 --- /dev/null +++ b/fairseq/tasks/__init__.py @@ -0,0 +1,82 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import importlib +import os + +from .fairseq_task import FairseqTask + +TASK_REGISTRY = {} +TASK_CLASS_NAMES = set() + + +def setup_task(args, **kwargs): + return TASK_REGISTRY[args.task].setup_task(args, **kwargs) + + +def register_task(name): + """ + New tasks can be added to fairseq with the + :func:`~fairseq.tasks.register_task` function decorator. + + For example:: + + @register_task('classification') + class ClassificationTask(FairseqTask): + (...) + + .. note:: + + All Tasks must implement the :class:`~fairseq.tasks.FairseqTask` + interface. + + Please see the + + Args: + name (str): the name of the task + """ + + def register_task_cls(cls): + if name in TASK_REGISTRY: + raise ValueError('Cannot register duplicate task ({})'.format(name)) + if not issubclass(cls, FairseqTask): + raise ValueError('Task ({}: {}) must extend FairseqTask'.format(name, cls.__name__)) + if cls.__name__ in TASK_CLASS_NAMES: + raise ValueError('Cannot register task with duplicate class name ({})'.format(cls.__name__)) + TASK_REGISTRY[name] = cls + TASK_CLASS_NAMES.add(cls.__name__) + return cls + + return register_task_cls + + +def get_task(name): + return TASK_REGISTRY[name] + + +# automatically import any Python files in the tasks/ directory +tasks_dir = os.path.dirname(__file__) +for file in os.listdir(tasks_dir): + path = os.path.join(tasks_dir, file) + if ( + not file.startswith('_') + and not file.startswith('.') + and (file.endswith('.py') or os.path.isdir(path)) + ): + task_name = file[:file.find('.py')] if file.endswith('.py') else file + importlib.import_module('fairseq.tasks.' + task_name) + + # expose `task_parser` for sphinx + if task_name in TASK_REGISTRY: + parser = argparse.ArgumentParser(add_help=False) + group_task = parser.add_argument_group('Task name') + # fmt: off + group_task.add_argument('--task', metavar=task_name, + help='Enable this task with: ``--task=' + task_name + '``') + # fmt: on + group_args = parser.add_argument_group('Additional command-line arguments') + TASK_REGISTRY[task_name].add_args(group_args) + globals()[task_name + '_parser'] = parser diff --git a/fairseq/tasks/__pycache__/__init__.cpython-310.pyc b/fairseq/tasks/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..653ec077abf75177851b98f582cd9ecb36895712 Binary files /dev/null and b/fairseq/tasks/__pycache__/__init__.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/audio_pretraining.cpython-310.pyc b/fairseq/tasks/__pycache__/audio_pretraining.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..960eaca3d163910a3aaedcd8e69bea500ac4bb0c Binary files /dev/null and b/fairseq/tasks/__pycache__/audio_pretraining.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/cross_lingual_lm.cpython-310.pyc b/fairseq/tasks/__pycache__/cross_lingual_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6a9bfc6d9be049c602dd8a74cc37ed5d60de26a2 Binary files /dev/null and b/fairseq/tasks/__pycache__/cross_lingual_lm.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/denoising.cpython-310.pyc b/fairseq/tasks/__pycache__/denoising.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..497d4fe8918655f2bc1f96a5f738bb603d8d5431 Binary files /dev/null and b/fairseq/tasks/__pycache__/denoising.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/fairseq_task.cpython-310.pyc b/fairseq/tasks/__pycache__/fairseq_task.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9ac8525a2ba586c024236f1739c9fd8ece597c17 Binary files /dev/null and b/fairseq/tasks/__pycache__/fairseq_task.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/language_modeling.cpython-310.pyc b/fairseq/tasks/__pycache__/language_modeling.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7e14ac0b417382b48d3297c6a1ee9925babb233a Binary files /dev/null and b/fairseq/tasks/__pycache__/language_modeling.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/legacy_masked_lm.cpython-310.pyc b/fairseq/tasks/__pycache__/legacy_masked_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a09de8871a6922957d2d4eafb59fc4b902a07961 Binary files /dev/null and b/fairseq/tasks/__pycache__/legacy_masked_lm.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/masked_lm.cpython-310.pyc b/fairseq/tasks/__pycache__/masked_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eca76e50e07c2dc2431ec6fdf0a1930615eca4b2 Binary files /dev/null and b/fairseq/tasks/__pycache__/masked_lm.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/multilingual_denoising.cpython-310.pyc b/fairseq/tasks/__pycache__/multilingual_denoising.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6eb542676312ef936419040681b3f4c0f9a10f7a Binary files /dev/null and b/fairseq/tasks/__pycache__/multilingual_denoising.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/multilingual_masked_lm.cpython-310.pyc b/fairseq/tasks/__pycache__/multilingual_masked_lm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4a4bea3529db67f6eb18f2a641ea4a8fcb3a629e Binary files /dev/null and b/fairseq/tasks/__pycache__/multilingual_masked_lm.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/multilingual_translation.cpython-310.pyc b/fairseq/tasks/__pycache__/multilingual_translation.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a1ac49efe92aed8de7eedf2ef20702358d4fc6a8 Binary files /dev/null and b/fairseq/tasks/__pycache__/multilingual_translation.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/semisupervised_translation.cpython-310.pyc b/fairseq/tasks/__pycache__/semisupervised_translation.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2944900a1f8c3318dd84ca4dea1b94d93fecb2ff Binary files /dev/null and b/fairseq/tasks/__pycache__/semisupervised_translation.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/sentence_prediction.cpython-310.pyc b/fairseq/tasks/__pycache__/sentence_prediction.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7ebf0068051f05e0b75b22c906eb653117e824d4 Binary files /dev/null and b/fairseq/tasks/__pycache__/sentence_prediction.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/sentence_ranking.cpython-310.pyc b/fairseq/tasks/__pycache__/sentence_ranking.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..78445f2c6f58807c3a95a0b59594b989feae258b Binary files /dev/null and b/fairseq/tasks/__pycache__/sentence_ranking.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/translation.cpython-310.pyc b/fairseq/tasks/__pycache__/translation.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6f82d9a3a2da52d7f7c942b4b53cdf8a4e8b1362 Binary files /dev/null and b/fairseq/tasks/__pycache__/translation.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/translation_from_pretrained_bart.cpython-310.pyc b/fairseq/tasks/__pycache__/translation_from_pretrained_bart.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..94c53a91abe573cdf6380d8ed26c4116bfb765ec Binary files /dev/null and b/fairseq/tasks/__pycache__/translation_from_pretrained_bart.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/translation_from_pretrained_xlm.cpython-310.pyc b/fairseq/tasks/__pycache__/translation_from_pretrained_xlm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0a3a0b03b42fed1ea56821f82c16fab0897155ac Binary files /dev/null and b/fairseq/tasks/__pycache__/translation_from_pretrained_xlm.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/translation_lev.cpython-310.pyc b/fairseq/tasks/__pycache__/translation_lev.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..677b8b14d980c2671d7a3094500554be4401f66a Binary files /dev/null and b/fairseq/tasks/__pycache__/translation_lev.cpython-310.pyc differ diff --git a/fairseq/tasks/__pycache__/translation_multi_simple_epoch.cpython-310.pyc b/fairseq/tasks/__pycache__/translation_multi_simple_epoch.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..252602de4fea2b6450b71b31f7047fbd90aec229 Binary files /dev/null and b/fairseq/tasks/__pycache__/translation_multi_simple_epoch.cpython-310.pyc differ diff --git a/fairseq/tasks/audio_pretraining.py b/fairseq/tasks/audio_pretraining.py new file mode 100644 index 0000000000000000000000000000000000000000..46d164ba9869f90700076b539fd089144583f691 --- /dev/null +++ b/fairseq/tasks/audio_pretraining.py @@ -0,0 +1,137 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +import os +import sys + +from fairseq.data import FileAudioDataset, Dictionary, AddTargetDataset +from . import FairseqTask, register_task + + +class LabelEncoder(object): + def __init__(self, dictionary): + self.dictionary = dictionary + + def __call__(self, label): + return self.dictionary.encode_line( + label, append_eos=False, add_if_not_exist=False + ) + + +@register_task("audio_pretraining") +class AudioPretrainingTask(FairseqTask): + """ + + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument("data", help="path to data directory") + parser.add_argument( + "--sample-rate", + default=16000, + type=int, + help="target sample rate. audio files will be up/down sampled to this rate", + ) + parser.add_argument( + "--normalize", + action="store_true", + help="if set, normalizes input to have 0 mean and unit variance", + ) + parser.add_argument( + "--max-sample-size", + default=None, + type=int, + help="max sample size to crop to for batching. default = min sample length", + ) + parser.add_argument( + "--min-sample-size", + default=None, + type=int, + help="min sample size to crop to for batching. default = same as --max-sample-size", + ) + + parser.add_argument( + "--enable-padding", + action="store_true", + help="pad shorter samples instead of cropping", + ) + + parser.add_argument( + "--labels", + type=str, + default=None, + help="extension of the label file to load, if any", + ) + + def __init__(self, args, source_dictionary=None): + super().__init__(args) + self._target_dictionary = None + self._source_dictionary = source_dictionary + self.is_ctc = args.criterion == "ctc" + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + return cls(args) + + def load_dataset(self, split, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + manifest = os.path.join(self.args.data, "{}.tsv".format(split)) + self.datasets[split] = FileAudioDataset( + manifest, + sample_rate=self.args.sample_rate, + max_sample_size=self.args.max_sample_size, + min_sample_size=self.args.max_sample_size, + min_length=self.args.min_sample_size, + pad=self.args.labels is not None or self.args.enable_padding, + normalize=self.args.normalize, + ) + + if self.args.labels: + dict_path = os.path.join(self.args.data, f"dict.{self.args.labels}.txt") + self._target_dictionary = Dictionary.load(dict_path) + label_path = os.path.join(self.args.data, f"{split}.{self.args.labels}") + labels = [] + with open(label_path, "r") as f: + for line in f: + labels.append(line) + + process_label = LabelEncoder(self.target_dictionary) + + self.datasets[split] = AddTargetDataset( + self.datasets[split], + labels, + pad=self.target_dictionary.pad(), + eos=self.target_dictionary.eos(), + batch_targets=True, + process_label=process_label, + add_to_input=not self.is_ctc, + ) + + @property + def source_dictionary(self): + return self._source_dictionary + + @property + def target_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self._target_dictionary + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return (sys.maxsize, sys.maxsize) diff --git a/fairseq/tasks/cross_lingual_lm.py b/fairseq/tasks/cross_lingual_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..3589492f1115641712489b7982876fda3cc39317 --- /dev/null +++ b/fairseq/tasks/cross_lingual_lm.py @@ -0,0 +1,170 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict +import itertools +import logging +import os + +import numpy as np + +from fairseq import tokenizer +from fairseq.data.legacy.masked_lm_dictionary import MaskedLMDictionary + +from fairseq.data import ( + Dictionary, + ConcatDataset, + data_utils, + TokenBlockDataset, +) +from fairseq.data.legacy.masked_lm_dataset import MaskedLMDataset +from fairseq.data.multi_corpus_sampled_dataset import MultiCorpusSampledDataset +from fairseq.tasks import FairseqTask, register_task +from fairseq import utils + +logger = logging.getLogger(__name__) + + +@register_task('cross_lingual_lm') +class CrossLingualLMTask(FairseqTask): + """ + Task for training cross-lingual language models. + + For more details look at: https://arxiv.org/pdf/1901.07291.pdf + + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('data', help='colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner') + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments' + ' per sample') + parser.add_argument('--monolingual-langs', default='en', type=str, + help='comma separated list of languages for which we' + ' want to train XLM on') + parser.add_argument('--shuffle', action='store_true', + help='shuffle each monolingual dataset while' + ' training') + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + self.distributed_world_size = args.distributed_world_size + self.langs2id = self._lang_to_id(args.monolingual_langs) + + def _lang_to_id( + self, + languages: str + ): + """ + Build a map from languages to ids. These ids are used as segment labels + for cross-lingual LM training. + """ + lang2id = {} + langs = [l.strip() for l in languages.split(',')] + for id, lang in enumerate(langs): + lang2id[lang] = id + return lang2id + + @classmethod + def load_dictionary(cls, filename): + return MaskedLMDictionary.load(filename) + + @classmethod + def build_dictionary(cls, filenames, workers=1, threshold=-1, nwords=-1, padding_factor=8): + d = MaskedLMDictionary() + for filename in filenames: + Dictionary.add_file_to_dictionary(filename, d, tokenizer.tokenize_line, workers) + d.finalize(threshold=threshold, nwords=nwords, padding_factor=padding_factor) + return d + + @property + def target_dictionary(self): + return self.dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task.""" + dictionary = MaskedLMDictionary.load(os.path.join(args.data, 'dict.txt')) + logger.info('dictionary: {} types'.format(len(dictionary))) + return cls(args, dictionary) + + def _load_single_lang_dataset(self, split, epoch): + loaded_datasets = [] + + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else '') + path = os.path.join(data_path, split_k) + + ds = data_utils.load_indexed_dataset(path, self.dictionary, self.args.dataset_impl) + if ds is None: + if k > 0: + break + else: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, data_path)) + + # Since we append each block with the classification_token, + # we need to effectively create blocks of length + # tokens_per_sample-1 + loaded_datasets.append( + TokenBlockDataset( + ds, ds.sizes, self.args.tokens_per_sample - 1, + pad=self.dictionary.pad(), eos=self.dictionary.eos(), + ) + ) + + logger.info('{} {} {} examples'.format(data_path, split_k, len(loaded_datasets[-1]))) + + if len(loaded_datasets) == 1: + dataset = loaded_datasets[0] + sizes = dataset.sizes + else: + dataset = ConcatDataset(loaded_datasets) + sizes = np.concatenate([ds.sizes for ds in loaded_datasets]) + + return dataset, sizes + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + dataset_map = OrderedDict() + + for lang in self.langs2id.keys(): + # Datasets are expected to be in "split.lang" format (Eg: train.en) + language_split = '{}.{}'.format(split, lang) + + block_dataset, sizes = self._load_single_lang_dataset(split=language_split, epoch=epoch) + + dataset_map[lang] = MaskedLMDataset( + dataset=block_dataset, + sizes=sizes, + vocab=self.dictionary, + pad_idx=self.dictionary.pad(), + mask_idx=self.dictionary.mask(), + classif_token_idx=self.dictionary.eos(), + sep_token_idx=self.dictionary.eos(), + shuffle=getattr(self.args, 'shuffle', False), + has_pairs=False, + segment_id=self.langs2id[lang], + seed=self.seed, + ) + + self.datasets[split] = MultiCorpusSampledDataset(dataset_map) + logger.info('{} {} {} examples'.format( + utils.split_paths(self.args.data)[epoch - 1], split, len(self.datasets[split])) + ) diff --git a/fairseq/tasks/denoising.py b/fairseq/tasks/denoising.py new file mode 100644 index 0000000000000000000000000000000000000000..28beb517f2c6902edf51bd697a9dbe3e8d110b75 --- /dev/null +++ b/fairseq/tasks/denoising.py @@ -0,0 +1,171 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +from fairseq.data import ( + data_utils, + Dictionary, + AppendTokenDataset, + DenoisingDataset, + PrependTokenDataset, + StripTokenDataset, + TokenBlockDataset, +) +from fairseq.data.encoders.utils import get_whole_word_mask +from fairseq.tasks import FairseqTask, register_task +from fairseq import utils + + +logger = logging.getLogger(__name__) + + +@register_task('denoising') +class DenoisingTask(FairseqTask): + """ + Denoising task for applying sequence to sequence denoising. (ie. BART) + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('data', help='path to data directory') + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments' + ' per sample for dataset') + parser.add_argument( + '--sample-break-mode', default="complete_doc", type=str, + help='mode for breaking sentence', + ) + parser.add_argument( + '--mask', default=0.0, type=float, + help='fraction of words/subwords that will be masked', + ) + parser.add_argument( + '--mask-random', default=0.0, type=float, + help='instead of using [MASK], use random token this often' + ) + parser.add_argument( + '--insert', default=0.0, type=float, + help='insert this percentage of additional random tokens', + ) + parser.add_argument( + '--permute', default=0.0, type=float, + help='take this proportion of subwords and permute them', + ) + parser.add_argument( + '--rotate', default=0.5, type=float, + help='rotate this proportion of inputs', + ) + parser.add_argument( + '--poisson-lambda', default=3.0, type=float, + help='randomly shuffle sentences for this proportion of inputs' + ) + parser.add_argument( + '--permute-sentences', default=0.0, type=float, + help='shuffle this proportion of sentences in all inputs' + ) + parser.add_argument( + '--mask-length', default="subword", type=str, + choices=['subword', 'word', 'span-poisson'], + help='mask length to choose' + ) + parser.add_argument( + '--replace-length', default=-1, type=int, + help='when masking N tokens, replace with 0, 1, or N tokens (use -1 for N)' + ) + parser.add_argument( + '--max-source-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the source sequence' + ) + parser.add_argument( + '--max-target-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the target sequence' + ) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = self.dictionary.add_symbol('') + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task. + """ + dictionary = Dictionary.load(os.path.join(args.data, 'dict.txt')) + logger.info('dictionary: {} types'.format(len(dictionary))) + if not hasattr(args, 'shuffle_instance'): + args.shuffle_instance = False + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, split_path)) + + dataset = StripTokenDataset(dataset, self.dictionary.eos()) + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 2, # one less for and one for + pad=self.dictionary.pad(), + eos=self.dictionary.eos(), + break_mode=self.args.sample_break_mode, + document_sep_len=0 + ) + + # prepend beginning-of-sentence token (, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + dataset = AppendTokenDataset(dataset, self.source_dictionary.eos()) + + mask_whole_words = get_whole_word_mask(self.args, self.source_dictionary) \ + if self.args.mask_length != 'subword' else None + + self.datasets[split] = DenoisingDataset( + dataset, dataset.sizes, self.dictionary, self.mask_idx, + mask_whole_words, shuffle=self.args.shuffle_instance, + seed=self.seed, args=self.args + ) + logger.info( + "Split: {0}, Loaded {1} samples of denoising_dataset".format( + split, + len(self.datasets[split]), + ) + ) + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + return (self.args.max_source_positions, self.args.max_target_positions) + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary`.""" + return self.dictionary + + @property + def target_dictionary(self): + """Return the target :class:`~fairseq.data.Dictionary`.""" + return self.dictionary diff --git a/fairseq/tasks/fairseq_task.py b/fairseq/tasks/fairseq_task.py new file mode 100644 index 0000000000000000000000000000000000000000..59663b531d59b9b523836cedb29cbd0e7cf88e0b --- /dev/null +++ b/fairseq/tasks/fairseq_task.py @@ -0,0 +1,465 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import warnings + + +import torch + +from fairseq import metrics, search, tokenizer, utils +from fairseq.data import data_utils, FairseqDataset, iterators, Dictionary + +logger = logging.getLogger(__name__) + + +class FairseqTask(object): + """ + Tasks store dictionaries and provide helpers for loading/iterating over + Datasets, initializing the Model/Criterion and calculating the loss. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + pass + + @staticmethod + def logging_outputs_can_be_summed(criterion) -> bool: + """ + Whether the logging outputs returned by `train_step` and `valid_step` can + be summed across workers prior to calling `aggregate_logging_outputs`. + Setting this to True will improves distributed training speed. + """ + return criterion.logging_outputs_can_be_summed() + + def __init__(self, args): + self.args = args + self.datasets = {} + self.dataset_to_epoch_iter = {} + + @classmethod + def load_dictionary(cls, filename): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + return Dictionary.load(filename) + + @classmethod + def build_dictionary( + cls, filenames, workers=1, threshold=-1, nwords=-1, padding_factor=8 + ): + """Build the dictionary + + Args: + filenames (list): list of filenames + workers (int): number of concurrent workers + threshold (int): defines the minimum word count + nwords (int): defines the total number of words in the final dictionary, + including special symbols + padding_factor (int): can be used to pad the dictionary size to be a + multiple of 8, which is important on some hardware (e.g., Nvidia + Tensor Cores). + """ + d = Dictionary() + for filename in filenames: + Dictionary.add_file_to_dictionary( + filename, d, tokenizer.tokenize_line, workers + ) + d.finalize(threshold=threshold, nwords=nwords, padding_factor=padding_factor) + return d + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + return cls(args, **kwargs) + + def has_sharded_data(self, split): + return (os.pathsep in getattr(self.args, 'data', '')) + + def load_dataset(self, split, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + raise NotImplementedError + + def dataset(self, split): + """ + Return a loaded dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + + Returns: + a :class:`~fairseq.data.FairseqDataset` corresponding to *split* + """ + from fairseq.data import FairseqDataset + + if split not in self.datasets: + raise KeyError("Dataset not loaded: " + split) + if not isinstance(self.datasets[split], FairseqDataset): + raise TypeError("Datasets are expected to be of type FairseqDataset") + return self.datasets[split] + + def filter_indices_by_size(self, + indices, + dataset, + max_positions, + ignore_invalid_inputs): + """ + Filter examples that are too large + + Args: + indices (np.array): original array of sample indices + dataset (~fairseq.data.FairseqDataset): dataset to batch + max_positions (optional): max sentence length supported by the + model (default: None). + ignore_invalid_inputs (bool, optional): don't raise Exception for + sentences that are too long (default: False). + Returns: + np.array: array of filtered sample indices + """ + indices, ignored = dataset.filter_indices_by_size(indices, max_positions) + if len(ignored) > 0: + if not ignore_invalid_inputs: + raise Exception(( + 'Size of sample #{} is invalid (={}) since max_positions={}, ' + 'skip this example with --skip-invalid-size-inputs-valid-test' + ).format(ignored[0], dataset.size(ignored[0]), max_positions)) + logger.warning(( + '{} samples have invalid sizes and will be skipped, ' + 'max_positions={}, first few sample ids={}' + ).format(len(ignored), max_positions, ignored[:10])) + return indices + + def get_batch_iterator( + self, + dataset, + max_tokens=None, + max_sentences=None, + max_positions=None, + ignore_invalid_inputs=False, + required_batch_size_multiple=1, + seed=1, + num_shards=1, + shard_id=0, + num_workers=0, + epoch=1 + ): + """ + Get an iterator that yields batches of data from the given dataset. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset to batch + max_tokens (int, optional): max number of tokens in each batch + (default: None). + max_sentences (int, optional): max number of sentences in each + batch (default: None). + max_positions (optional): max sentence length supported by the + model (default: None). + ignore_invalid_inputs (bool, optional): don't raise Exception for + sentences that are too long (default: False). + required_batch_size_multiple (int, optional): require batch size to + be a multiple of N (default: 1). + seed (int, optional): seed for random number generator for + reproducibility (default: 1). + num_shards (int, optional): shard the data iterator into N + shards (default: 1). + shard_id (int, optional): which shard of the data iterator to + return (default: 0). + num_workers (int, optional): how many subprocesses to use for data + loading. 0 means the data will be loaded in the main process + (default: 0). + epoch (int, optional): the epoch to start the iterator from + (default: 1). + Returns: + ~fairseq.iterators.EpochBatchIterator: a batched iterator over the + given dataset split + """ + # For default fairseq task, return same iterator across epochs + # as datasets are not dynamic, can be overridden in task specific + # setting. + if dataset in self.dataset_to_epoch_iter: + return self.dataset_to_epoch_iter[dataset] + + assert isinstance(dataset, FairseqDataset) + + # initialize the dataset with the correct starting epoch + dataset.set_epoch(epoch) + + # get indices ordered by example size + with data_utils.numpy_seed(seed): + indices = dataset.ordered_indices() + + # filter examples that are too large + if max_positions is not None: + indices = self.filter_indices_by_size(indices, + dataset, + max_positions, + ignore_invalid_inputs) + + # create mini-batches with given size constraints + batch_sampler = dataset.batch_by_size( + indices, + max_tokens=max_tokens, + max_sentences=max_sentences, + required_batch_size_multiple=required_batch_size_multiple, + ) + + # return a reusable, sharded iterator + epoch_iter = iterators.EpochBatchIterator( + dataset=dataset, + collate_fn=dataset.collater, + batch_sampler=batch_sampler, + seed=seed, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + epoch=epoch, + buffer_size=getattr(self.args, 'data_buffer_size', 0) + ) + self.dataset_to_epoch_iter[dataset] = epoch_iter + return epoch_iter + + def build_model(self, args): + """ + Build the :class:`~fairseq.models.BaseFairseqModel` instance for this + task. + + Args: + args (argparse.Namespace): parsed command-line arguments + + Returns: + a :class:`~fairseq.models.BaseFairseqModel` instance + """ + from fairseq import models, quantization_utils + model = models.build_model(args, self) + if getattr(args, 'tpu', False): + model.prepare_for_tpu_() + model = quantization_utils.quantize_model_scalar(model, args) + return model + + def build_criterion(self, args): + """ + Build the :class:`~fairseq.criterions.FairseqCriterion` instance for + this task. + + Args: + args (argparse.Namespace): parsed command-line arguments + + Returns: + a :class:`~fairseq.criterions.FairseqCriterion` instance + """ + from fairseq import criterions + + return criterions.build_criterion(args, self) + + def build_generator( + self, models, args, + seq_gen_cls=None, extra_gen_cls_kwargs=None + ): + if getattr(args, "score_reference", False): + from fairseq.sequence_scorer import SequenceScorer + + return SequenceScorer( + self.target_dictionary, + compute_alignment=getattr(args, "print_alignment", False), + ) + + from fairseq.sequence_generator import ( + SequenceGenerator, + SequenceGeneratorWithAlignment, + ) + + # Choose search strategy. Defaults to Beam Search. + sampling = getattr(args, "sampling", False) + sampling_topk = getattr(args, "sampling_topk", -1) + sampling_topp = getattr(args, "sampling_topp", -1.0) + diverse_beam_groups = getattr(args, "diverse_beam_groups", -1) + diverse_beam_strength = getattr(args, "diverse_beam_strength", 0.5) + match_source_len = getattr(args, "match_source_len", False) + diversity_rate = getattr(args, "diversity_rate", -1) + if ( + sum( + int(cond) + for cond in [ + sampling, + diverse_beam_groups > 0, + match_source_len, + diversity_rate > 0, + ] + ) + > 1 + ): + raise ValueError("Provided Search parameters are mutually exclusive.") + assert sampling_topk < 0 or sampling, "--sampling-topk requires --sampling" + assert sampling_topp < 0 or sampling, "--sampling-topp requires --sampling" + + if sampling: + search_strategy = search.Sampling( + self.target_dictionary, sampling_topk, sampling_topp + ) + elif diverse_beam_groups > 0: + search_strategy = search.DiverseBeamSearch( + self.target_dictionary, diverse_beam_groups, diverse_beam_strength + ) + elif match_source_len: + # this is useful for tagging applications where the output + # length should match the input length, so we hardcode the + # length constraints for simplicity + search_strategy = search.LengthConstrainedBeamSearch( + self.target_dictionary, + min_len_a=1, + min_len_b=0, + max_len_a=1, + max_len_b=0, + ) + elif diversity_rate > -1: + search_strategy = search.DiverseSiblingsSearch( + self.target_dictionary, diversity_rate + ) + else: + search_strategy = search.BeamSearch(self.target_dictionary) + + if seq_gen_cls is None: + if getattr(args, "print_alignment", False): + seq_gen_cls = SequenceGeneratorWithAlignment + else: + seq_gen_cls = SequenceGenerator + extra_gen_cls_kwargs = extra_gen_cls_kwargs or {} + return seq_gen_cls( + models, + self.target_dictionary, + beam_size=getattr(args, "beam", 5), + max_len_a=getattr(args, "max_len_a", 0), + max_len_b=getattr(args, "max_len_b", 200), + min_len=getattr(args, "min_len", 1), + normalize_scores=(not getattr(args, "unnormalized", False)), + len_penalty=getattr(args, "lenpen", 1), + unk_penalty=getattr(args, "unkpen", 0), + temperature=getattr(args, "temperature", 1.0), + match_source_len=getattr(args, "match_source_len", False), + no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0), + search_strategy=search_strategy, + **extra_gen_cls_kwargs, + ) + + def train_step( + self, sample, model, criterion, optimizer, update_num, ignore_grad=False + ): + """ + Do forward and backward, and return the loss as computed by *criterion* + for the given *model* and *sample*. + + Args: + sample (dict): the mini-batch. The format is defined by the + :class:`~fairseq.data.FairseqDataset`. + model (~fairseq.models.BaseFairseqModel): the model + criterion (~fairseq.criterions.FairseqCriterion): the criterion + optimizer (~fairseq.optim.FairseqOptimizer): the optimizer + update_num (int): the current update + ignore_grad (bool): multiply loss by 0 if this is set to True + + Returns: + tuple: + - the loss + - the sample size, which is used as the denominator for the + gradient + - logging outputs to display while training + """ + model.train() + model.set_num_updates(update_num) + with torch.autograd.profiler.record_function("forward"): + loss, sample_size, logging_output = criterion(model, sample) + if ignore_grad: + loss *= 0 + with torch.autograd.profiler.record_function("backward"): + optimizer.backward(loss) + return loss, sample_size, logging_output + + def valid_step(self, sample, model, criterion): + model.eval() + with torch.no_grad(): + loss, sample_size, logging_output = criterion(model, sample) + return loss, sample_size, logging_output + + def inference_step(self, generator, models, sample, prefix_tokens=None): + with torch.no_grad(): + return generator.generate(models, sample, prefix_tokens=prefix_tokens) + + def begin_epoch(self, epoch, model): + """Hook function called before the start of each epoch.""" + pass + + def aggregate_logging_outputs(self, logging_outputs, criterion): + """[deprecated] Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + "The aggregate_logging_outputs API is deprecated. " + "Please use the reduce_metrics API instead." + ) + with metrics.aggregate() as agg: + self.reduce_metrics(logging_outputs, criterion) + return agg.get_smoothed_values() + + def reduce_metrics(self, logging_outputs, criterion): + """Aggregate logging outputs from data parallel training.""" + # backward compatibility for tasks that override aggregate_logging_outputs + base_func = FairseqTask.aggregate_logging_outputs + self_func = getattr(self, "aggregate_logging_outputs").__func__ + if self_func is not base_func: + utils.deprecation_warning( + "Tasks should implement the reduce_metrics API. " + "Falling back to deprecated aggregate_logging_outputs API." + ) + agg_logging_outputs = self.aggregate_logging_outputs( + logging_outputs, criterion + ) + for k, v in agg_logging_outputs.items(): + metrics.log_scalar(k, v) + return + + if not any("ntokens" in log for log in logging_outputs): + warnings.warn( + "ntokens not found in Criterion logging outputs, cannot log wpb or wps" + ) + else: + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + metrics.log_scalar("wpb", ntokens, priority=180, round=1) + metrics.log_speed("wps", ntokens, priority=90, round=1) + + if not any("nsentences" in log for log in logging_outputs): + warnings.warn( + "nsentences not found in Criterion logging outputs, cannot log bsz" + ) + else: + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + metrics.log_scalar("bsz", nsentences, priority=190, round=1) + + criterion.__class__.reduce_metrics(logging_outputs) + + def max_positions(self): + """Return the max input length allowed by the task.""" + return None + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary` (if applicable + for this task).""" + raise NotImplementedError + + @property + def target_dictionary(self): + """Return the target :class:`~fairseq.data.Dictionary` (if applicable + for this task).""" + raise NotImplementedError diff --git a/fairseq/tasks/language_modeling.py b/fairseq/tasks/language_modeling.py new file mode 100644 index 0000000000000000000000000000000000000000..a4a98e07bc5b3e9d21cbacf24815e0771b348456 --- /dev/null +++ b/fairseq/tasks/language_modeling.py @@ -0,0 +1,290 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np +import torch + +from fairseq import utils +from fairseq.data import ( + AppendTokenDataset, + data_utils, + Dictionary, + IdDataset, + MonolingualDataset, + NestedDictionaryDataset, + NumelDataset, + PadDataset, + PrependTokenDataset, + StripTokenDataset, + TokenBlockDataset, + TransformEosDataset, + TruncatedDictionary, +) +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import FairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task("language_modeling") +class LanguageModelingTask(FairseqTask): + """ + Train a language model. + + Args: + dictionary (~fairseq.data.Dictionary): the dictionary for the input of + the language model + output_dictionary (~fairseq.data.Dictionary): the dictionary for the + output of the language model. In most cases it will be the same as + *dictionary*, but could possibly be a more limited version of the + dictionary (if ``--output-dictionary-size`` is used). + targets (List[str]): list of the target types that the language model + should predict. Can be one of "self", "future", and "past". + Defaults to "future". + + .. note:: + + The language modeling task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate`, :mod:`fairseq-interactive` and + :mod:`fairseq-eval-lm`. + + The language modeling task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.language_modeling_parser + :prog: + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + parser.add_argument('data', help='path to data directory') + parser.add_argument('--sample-break-mode', default='none', + choices=['none', 'complete', 'complete_doc', 'eos'], + help='If omitted or "none", fills each sample with tokens-per-sample ' + 'tokens. If set to "complete", splits samples only at the end ' + 'of sentence, but may include multiple sentences per sample. ' + '"complete_doc" is similar but respects doc boundaries. ' + 'If set to "eos", includes only one sentence per sample.') + parser.add_argument('--tokens-per-sample', default=1024, type=int, + help='max number of tokens per sample for LM dataset') + parser.add_argument('--output-dictionary-size', default=-1, type=int, + help='limit the size of output dictionary') + parser.add_argument('--self-target', action='store_true', + help='include self target') + parser.add_argument('--future-target', action='store_true', + help='include future target') + parser.add_argument('--past-target', action='store_true', + help='include past target') + parser.add_argument('--add-bos-token', action='store_true', + help='prepend beginning of sentence token ()') + parser.add_argument('--max-target-positions', type=int, metavar='N', + help='max number of tokens in the target sequence') + parser.add_argument('--shorten-method', default='none', + choices=['none', 'truncate', 'random_crop'], + help='if not none, shorten sequences that exceed --tokens-per-sample') + parser.add_argument('--shorten-data-split-list', default='', + help='comma-separated list of dataset splits to apply shortening to, ' + 'e.g., "train,valid" (default: all dataset splits)') + # fmt: on + + def __init__(self, args, dictionary, output_dictionary=None, targets=None): + super().__init__(args) + self.dictionary = dictionary + self.output_dictionary = output_dictionary or dictionary + + if targets is None: + targets = ["future"] + self.targets = targets + + @classmethod + def setup_dictionary(cls, args, **kwargs): + dictionary = None + output_dictionary = None + if args.data: + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt")) + logger.info("dictionary: {} types".format(len(dictionary))) + output_dictionary = dictionary + if args.output_dictionary_size >= 0: + output_dictionary = TruncatedDictionary( + dictionary, args.output_dictionary_size + ) + return (dictionary, output_dictionary) + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + dictionary, output_dictionary = cls.setup_dictionary(args, **kwargs) + + # upgrade old checkpoints + if hasattr(args, "exclude_self_target"): + args.self_target = not args.exclude_self_target + + targets = [] + if getattr(args, "self_target", False): + targets.append("self") + if getattr(args, "future_target", False): + targets.append("future") + if getattr(args, "past_target", False): + targets.append("past") + if len(targets) == 0: + # standard language modeling + targets = ["future"] + + return cls(args, dictionary, output_dictionary, targets=targets) + + def build_model(self, args): + model = super().build_model(args) + + for target in self.targets: + if target not in model.supported_targets: + raise ValueError( + "Unsupported language modeling target: {}".format(target) + ) + + return model + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + dataset = data_utils.load_indexed_dataset( + split_path, self.dictionary, self.args.dataset_impl, combine=combine + ) + if dataset is None: + raise FileNotFoundError( + "Dataset not found: {} ({})".format(split, split_path) + ) + + dataset = maybe_shorten_dataset( + dataset, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.tokens_per_sample, + self.args.seed, + ) + + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample, + pad=self.dictionary.pad(), + eos=self.dictionary.eos(), + break_mode=self.args.sample_break_mode, + include_targets=True, + ) + + add_eos_for_other_targets = ( + self.args.sample_break_mode is not None + and self.args.sample_break_mode != "none" + ) + + self.datasets[split] = self._initialize_dataset( + dataset=dataset, + sizes=dataset.sizes, + src_vocab=self.dictionary, + tgt_vocab=self.output_dictionary, + add_eos_for_other_targets=add_eos_for_other_targets, + shuffle=True, + targets=self.targets, + add_bos_token=self.args.add_bos_token, + ) + + def _initialize_dataset(self, **kwargs): + return MonolingualDataset(**kwargs) + + def build_dataset_for_inference(self, src_tokens, src_lengths, **kwargs): + """ + Generate batches for inference. We prepend an eos token to src_tokens + (or bos if `--add-bos-token` is set) and we append a to target. + This is convenient both for generation with a prefix and LM scoring. + """ + dataset = StripTokenDataset( + TokenBlockDataset( + src_tokens, + src_lengths, + block_size=None, # ignored for "eos" break mode + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode="eos", + ), + # remove eos from (end of) target sequence + self.source_dictionary.eos(), + ) + src_dataset = PrependTokenDataset( + dataset, + token=( + self.source_dictionary.bos() + if getattr(self.args, "add_bos_token", False) + else self.source_dictionary.eos() + ), + ) + tgt_dataset = AppendTokenDataset( + dataset, + token=self.source_dictionary.pad() + ) + return NestedDictionaryDataset( + { + "id": IdDataset(), + "net_input": { + "src_tokens": PadDataset(src_dataset, pad_idx=self.source_dictionary.pad(), left_pad=False), + "src_lengths": NumelDataset(src_dataset, reduce=False), + }, + "target": PadDataset(tgt_dataset, pad_idx=self.source_dictionary.pad(), left_pad=False), + }, + sizes=[np.array(src_lengths)], + ) + + def inference_step(self, generator, models, sample, prefix_tokens=None): + with torch.no_grad(): + # Generation will always be conditioned on bos_token + if getattr(self.args, "add_bos_token", False): + bos_token = self.source_dictionary.bos() + else: + bos_token = self.source_dictionary.eos() + + # SequenceGenerator doesn't use src_tokens directly, we need to + # pass the `prefix_tokens` argument instead + if prefix_tokens is None and sample["net_input"]["src_tokens"].nelement(): + prefix_tokens = sample["net_input"]["src_tokens"] + if prefix_tokens[:, 0].eq(bos_token).all(): + prefix_tokens = prefix_tokens[:, 1:] + + return generator.generate( + models, sample, prefix_tokens=prefix_tokens, bos_token=bos_token, + ) + + @property + def source_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self.dictionary + + @property + def target_dictionary(self): + """Return the :class:`~fairseq.data.Dictionary` for the language + model.""" + return self.output_dictionary diff --git a/fairseq/tasks/legacy_masked_lm.py b/fairseq/tasks/legacy_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..40e27249539b4ff4b2cc9035cf442234733d9d43 --- /dev/null +++ b/fairseq/tasks/legacy_masked_lm.py @@ -0,0 +1,144 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import itertools +import logging +import os + +import numpy as np + +from fairseq import tokenizer +from fairseq.data import ( + ConcatDataset, + indexed_dataset, + data_utils, +) + +from fairseq.data import Dictionary +from fairseq.data.legacy.block_pair_dataset import BlockPairDataset +from fairseq.data.legacy.masked_lm_dataset import MaskedLMDataset +from fairseq.data.legacy.masked_lm_dictionary import BertDictionary +from fairseq.tasks import FairseqTask, register_task +from fairseq import utils + + +logger = logging.getLogger(__name__) + + +@register_task('legacy_masked_lm') +class LegacyMaskedLMTask(FairseqTask): + """ + Task for training Masked LM (BERT) model. + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('data', help='colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner') + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments' + ' per sample for BERT dataset') + parser.add_argument('--break-mode', default="doc", type=str, help='mode for breaking sentence') + parser.add_argument('--shuffle-dataset', action='store_true', default=False) + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + @classmethod + def load_dictionary(cls, filename): + return BertDictionary.load(filename) + + @classmethod + def build_dictionary(cls, filenames, workers=1, threshold=-1, nwords=-1, padding_factor=8): + d = BertDictionary() + for filename in filenames: + Dictionary.add_file_to_dictionary(filename, d, tokenizer.tokenize_line, workers) + d.finalize(threshold=threshold, nwords=nwords, padding_factor=padding_factor) + return d + + @property + def target_dictionary(self): + return self.dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task. + """ + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = BertDictionary.load(os.path.join(paths[0], 'dict.txt')) + logger.info('dictionary: {} types'.format(len(dictionary))) + + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + loaded_datasets = [] + + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + logger.info("data_path", data_path) + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else '') + path = os.path.join(data_path, split_k) + ds = indexed_dataset.make_dataset( + path, + impl=self.args.dataset_impl, + fix_lua_indexing=True, + dictionary=self.dictionary, + ) + + if ds is None: + if k > 0: + break + else: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, data_path)) + + with data_utils.numpy_seed(self.seed + k): + loaded_datasets.append( + BlockPairDataset( + ds, + self.dictionary, + ds.sizes, + self.args.tokens_per_sample, + break_mode=self.args.break_mode, + doc_break_size=1, + ) + ) + + logger.info('{} {} {} examples'.format(data_path, split_k, len(loaded_datasets[-1]))) + + if not combine: + break + + if len(loaded_datasets) == 1: + dataset = loaded_datasets[0] + sizes = dataset.sizes + else: + dataset = ConcatDataset(loaded_datasets) + sizes = np.concatenate([ds.sizes for ds in loaded_datasets]) + + self.datasets[split] = MaskedLMDataset( + dataset=dataset, + sizes=sizes, + vocab=self.dictionary, + pad_idx=self.dictionary.pad(), + mask_idx=self.dictionary.mask(), + classif_token_idx=self.dictionary.cls(), + sep_token_idx=self.dictionary.sep(), + shuffle=self.args.shuffle_dataset, + seed=self.seed, + ) diff --git a/fairseq/tasks/masked_lm.py b/fairseq/tasks/masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..4d7ea54b644d1a3b18b32eb186b57fafb8e86e06 --- /dev/null +++ b/fairseq/tasks/masked_lm.py @@ -0,0 +1,210 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np + +from fairseq.data import ( + data_utils, + Dictionary, + IdDataset, + MaskTokensDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + PadDataset, + PrependTokenDataset, + SortDataset, + TokenBlockDataset, +) +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import FairseqTask, register_task +from fairseq.data.encoders.utils import get_whole_word_mask +from fairseq import utils + + +logger = logging.getLogger(__name__) + + +@register_task('masked_lm') +class MaskedLMTask(FairseqTask): + """Task for training masked language models (e.g., BERT, RoBERTa).""" + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('data', help='colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner') + parser.add_argument('--sample-break-mode', default='complete', + choices=['none', 'complete', 'complete_doc', 'eos'], + help='If omitted or "none", fills each sample with tokens-per-sample ' + 'tokens. If set to "complete", splits samples only at the end ' + 'of sentence, but may include multiple sentences per sample. ' + '"complete_doc" is similar but respects doc boundaries. ' + 'If set to "eos", includes only one sentence per sample.') + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments ' + 'per sample for BERT dataset') + parser.add_argument('--mask-prob', default=0.15, type=float, + help='probability of replacing a token with mask') + parser.add_argument('--leave-unmasked-prob', default=0.1, type=float, + help='probability that a masked token is unmasked') + parser.add_argument('--random-token-prob', default=0.1, type=float, + help='probability of replacing a token with a random token') + parser.add_argument('--freq-weighted-replacement', default=False, action='store_true', + help='sample random replacement words based on word frequencies') + parser.add_argument('--mask-whole-words', default=False, action='store_true', + help='mask whole words; you may also want to set --bpe') + parser.add_argument('--shorten-method', default='none', + choices=['none', 'truncate', 'random_crop'], + help='if not none, shorten sequences that exceed --tokens-per-sample') + parser.add_argument('--shorten-data-split-list', default='', + help='comma-separated list of dataset splits to apply shortening to, ' + 'e.g., "train,valid" (default: all dataset splits)') + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = dictionary.add_symbol('') + + @classmethod + def setup_task(cls, args, **kwargs): + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], 'dict.txt')) + logger.info('dictionary: {} types'.format(len(dictionary))) + return cls(args, dictionary) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, split_path)) + + dataset = maybe_shorten_dataset( + dataset, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.tokens_per_sample, + self.args.seed, + ) + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 1, # one less for + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode=self.args.sample_break_mode, + ) + logger.info('loaded {} blocks from: {}'.format(len(dataset), split_path)) + + # prepend beginning-of-sentence token (, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + + # create masked input and targets + mask_whole_words = get_whole_word_mask(self.args, self.source_dictionary) \ + if self.args.mask_whole_words else None + + src_dataset, tgt_dataset = MaskTokensDataset.apply_mask( + dataset, + self.source_dictionary, + pad_idx=self.source_dictionary.pad(), + mask_idx=self.mask_idx, + seed=self.args.seed, + mask_prob=self.args.mask_prob, + leave_unmasked_prob=self.args.leave_unmasked_prob, + random_token_prob=self.args.random_token_prob, + freq_weighted_replacement=self.args.freq_weighted_replacement, + mask_whole_words=mask_whole_words, + ) + + with data_utils.numpy_seed(self.args.seed + epoch): + shuffle = np.random.permutation(len(src_dataset)) + + self.datasets[split] = SortDataset( + NestedDictionaryDataset( + { + 'id': IdDataset(), + 'net_input': { + 'src_tokens': PadDataset( + src_dataset, + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ), + 'src_lengths': NumelDataset(src_dataset, reduce=False), + }, + 'target': PadDataset( + tgt_dataset, + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ), + 'nsentences': NumSamplesDataset(), + 'ntokens': NumelDataset(src_dataset, reduce=True), + }, + sizes=[src_dataset.sizes], + ), + sort_order=[ + shuffle, + src_dataset.sizes, + ], + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, sort=True): + src_dataset = PadDataset( + TokenBlockDataset( + src_tokens, + src_lengths, + self.args.tokens_per_sample - 1, # one less for + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode='eos', + ), + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ) + src_dataset = PrependTokenDataset(src_dataset, self.source_dictionary.bos()) + src_dataset = NestedDictionaryDataset( + { + 'id': IdDataset(), + 'net_input': { + 'src_tokens': src_dataset, + 'src_lengths': NumelDataset(src_dataset, reduce=False), + }, + }, + sizes=src_lengths, + ) + if sort: + src_dataset = SortDataset(src_dataset, sort_order=[src_lengths]) + return src_dataset + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/tasks/multilingual_denoising.py b/fairseq/tasks/multilingual_denoising.py new file mode 100644 index 0000000000000000000000000000000000000000..18ee717fff0b84154dff36f44dd8f6eb33f96241 --- /dev/null +++ b/fairseq/tasks/multilingual_denoising.py @@ -0,0 +1,225 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np + +from fairseq.data import ( + data_utils, + Dictionary, + AppendTokenDataset, + ConcatDataset, + DenoisingDataset, + PrependTokenDataset, + ResamplingDataset, + SortDataset, + TokenBlockDataset, +) +from .denoising import DenoisingTask +from fairseq.data.encoders.utils import get_whole_word_mask +from fairseq.tasks import register_task + + +logger = logging.getLogger(__name__) + + +@register_task('multilingual_denoising') +class MultilingualDenoisingTask(DenoisingTask): + + @staticmethod + def add_args(parser): + DenoisingTask.add_args(parser) + parser.add_argument('--multilang-sampling-alpha', type=float, default=1.0, + help='smoothing alpha for sample ratios across multiple datasets') + parser.add_argument('--add-lang-token', default=False, action='store_true') + parser.add_argument('--langs', type=str, help="language ids we are considering", default=None) + parser.add_argument('--no-whole-word-mask-langs', type=str, default='', metavar='N', + help='languages without spacing between words dont support whole word masking') + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task. + """ + paths = args.data.split(':') + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], 'dict.txt')) + + data_path = paths[0] + if args.langs is None: + languages = sorted([ + name for name in os.listdir(data_path) + if os.path.isdir(os.path.join(data_path, name)) + ]) + else: + languages = args.langs.split(',') + + if args.add_lang_token: + for lang in languages: + dictionary.add_symbol('[{}]'.format(lang)) + + logger.info("dictionary: {} types".format(len(dictionary))) + if not hasattr(args, 'shuffle_instance'): + args.shuffle_instance = False + return cls(args, dictionary) + + def __init__(self, args, dictionary): + super().__init__(args, dictionary) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = self.dictionary.add_symbol('') + self.langs = args.langs + self.args = args + + def _get_sample_prob(self, dataset_lens): + """ + Get smoothed sampling porbability by languages. This helps low resource + languages by upsampling them. + """ + prob = dataset_lens / dataset_lens.sum() + smoothed_prob = prob ** self.args.multilang_sampling_alpha + smoothed_prob = smoothed_prob / smoothed_prob.sum() + return smoothed_prob + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = self.args.data.split(':') + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + split_path = os.path.join(data_path, split) + + if self.langs is None: + languages = sorted([ + name for name in os.listdir(data_path) + if os.path.isdir(os.path.join(data_path, name)) + ]) + else: + languages = self.langs.split(',') + for name in languages: + p = os.path.join(data_path, name) + assert os.path.exists(p), "data not found: {}".format(p) + + logger.info("Training on {0} languages: {1}".format(len(languages), languages)) + logger.info("Language to id mapping: ", { + lang: id for id, lang in enumerate(languages) + } + ) + + mask_whole_words = get_whole_word_mask(self.args, self.dictionary) + language_without_segmentations = self.args.no_whole_word_mask_langs.split(',') + lang_datasets = [] + for language in languages: + split_path = os.path.join(data_path, language, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, split_path)) + + end_token = self.source_dictionary.index('[{}]'.format(language)) \ + if self.args.add_lang_token else self.source_dictionary.eos() + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 2, # one less for + pad=self.source_dictionary.pad(), + eos=end_token, + break_mode=self.args.sample_break_mode, + ) + logger.info('loaded {} blocks from: {}'.format(len(dataset), split_path)) + + # prepend beginning-of-sentence token (, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + dataset = AppendTokenDataset(dataset, end_token) + + lang_mask_whole_words = mask_whole_words if language not in language_without_segmentations else None + lang_dataset = DenoisingDataset( + dataset, + dataset.sizes, + self.dictionary, + self.mask_idx, + lang_mask_whole_words, + shuffle=self.args.shuffle_instance, + seed=self.seed, + args=self.args, + eos=None if not self.args.add_lang_token else self.source_dictionary.index('[{}]'.format(language)), + ) + lang_datasets.append(lang_dataset) + + dataset_lengths = np.array( + [len(d) for d in lang_datasets], + dtype=float, + ) + logger.info( + 'loaded total {} blocks for all languages'.format( + int(dataset_lengths.sum()), + ) + ) + if split == self.args.train_subset: + # For train subset, additionally up or down sample languages. + sample_probs = self._get_sample_prob(dataset_lengths) + logger.info( + "Sample probability by language: {}".format({ + lang: "{0:.4f}".format(sample_probs[id]) + for id, lang in enumerate(languages) + }) + ) + size_ratio = (sample_probs * dataset_lengths.sum()) / dataset_lengths + logger.info( + "Up/Down Sampling ratio by language: {}".format({ + lang: "{0:.2f}".format(size_ratio[id]) + for id, lang in enumerate(languages) + }) + ) + + resampled_lang_datasets = [ + ResamplingDataset( + lang_datasets[i], + size_ratio=size_ratio[i], + seed=self.args.seed, + epoch=epoch, + replace=size_ratio[i] >= 1.0, + ) + for i, d in enumerate(lang_datasets) + ] + dataset = ConcatDataset( + resampled_lang_datasets, + ) + else: + dataset = ConcatDataset(lang_datasets) + lang_splits = [split] + for lang_id, lang_dataset in enumerate(lang_datasets): + split_name = split + '_' + languages[lang_id] + lang_splits.append(split_name) + self.datasets[split_name] = lang_dataset + + if split in self.args.valid_subset: + self.args.valid_subset = self.args.valid_subset.replace( + split, ','.join(lang_splits) + ) + + with data_utils.numpy_seed(self.args.seed + epoch): + shuffle = np.random.permutation(len(dataset)) + + self.datasets[split] = SortDataset( + dataset, + sort_order=[ + shuffle, + dataset.sizes, + ], + ) diff --git a/fairseq/tasks/multilingual_masked_lm.py b/fairseq/tasks/multilingual_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..248724bd56a02f7a9d9093c2e6183e9623203565 --- /dev/null +++ b/fairseq/tasks/multilingual_masked_lm.py @@ -0,0 +1,317 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np +import torch + +from fairseq.data import ( + data_utils, + Dictionary, + encoders, + ConcatDataset, + IdDataset, + MaskTokensDataset, + NestedDictionaryDataset, + NumelDataset, + NumSamplesDataset, + PadDataset, + PrependTokenDataset, + RawLabelDataset, + ResamplingDataset, + SortDataset, + TokenBlockDataset, +) +from fairseq.tasks import FairseqTask, register_task +from fairseq import utils + + +logger = logging.getLogger(__name__) + + +@register_task('multilingual_masked_lm') +class MultiLingualMaskedLMTask(FairseqTask): + """Task for training masked language models (e.g., BERT, RoBERTa).""" + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('data', help='colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner') + parser.add_argument('--sample-break-mode', default='complete', + choices=['none', 'complete', 'complete_doc', 'eos'], + help='If omitted or "none", fills each sample with tokens-per-sample ' + 'tokens. If set to "complete", splits samples only at the end ' + 'of sentence, but may include multiple sentences per sample. ' + '"complete_doc" is similar but respects doc boundaries. ' + 'If set to "eos", includes only one sentence per sample.') + parser.add_argument('--tokens-per-sample', default=512, type=int, + help='max number of total tokens over all segments ' + 'per sample for BERT dataset') + parser.add_argument('--mask-prob', default=0.15, type=float, + help='probability of replacing a token with mask') + parser.add_argument('--leave-unmasked-prob', default=0.1, type=float, + help='probability that a masked token is unmasked') + parser.add_argument('--random-token-prob', default=0.1, type=float, + help='probability of replacing a token with a random token') + parser.add_argument('--freq-weighted-replacement', action='store_true', + help='sample random replacement words based on word frequencies') + parser.add_argument('--mask-whole-words', default=False, action='store_true', + help='mask whole words; you may also want to set --bpe') + parser.add_argument('--multilang-sampling-alpha', type=float, default=1.0, + help='smoothing alpha for sample rations across multiple datasets') + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + self.seed = args.seed + + # add mask token + self.mask_idx = dictionary.add_symbol('') + + @classmethod + def setup_task(cls, args, **kwargs): + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dictionary = Dictionary.load(os.path.join(paths[0], 'dict.txt')) + logger.info('dictionary: {} types'.format(len(dictionary))) + return cls(args, dictionary) + + def _get_whole_word_mask(self): + # create masked input and targets + if self.args.mask_whole_words: + bpe = encoders.build_bpe(self.args) + if bpe is not None: + + def is_beginning_of_word(i): + if i < self.source_dictionary.nspecial: + # special elements are always considered beginnings + return True + tok = self.source_dictionary[i] + if tok.startswith('madeupword'): + return True + try: + return bpe.is_beginning_of_word(tok) + except ValueError: + return True + + mask_whole_words = torch.ByteTensor(list( + map(is_beginning_of_word, range(len(self.source_dictionary))) + )) + else: + mask_whole_words = None + return mask_whole_words + + def _get_sample_prob(self, dataset_lens): + """ + Get smoothed sampling porbability by languages. This helps low resource + languages by upsampling them. + """ + prob = dataset_lens / dataset_lens.sum() + smoothed_prob = prob ** self.args.multilang_sampling_alpha + smoothed_prob = smoothed_prob / smoothed_prob.sum() + return smoothed_prob + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + languages = sorted( + name for name in os.listdir(data_path) + if os.path.isdir(os.path.join(data_path, name)) + ) + + logger.info("Training on {0} languages: {1}".format(len(languages), languages)) + logger.info("Language to id mapping: ", { + lang: id for id, lang in enumerate(languages) + } + ) + + mask_whole_words = self._get_whole_word_mask() + lang_datasets = [] + for lang_id, language in enumerate(languages): + split_path = os.path.join(data_path, language, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + if dataset is None: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, split_path)) + + # create continuous blocks of tokens + dataset = TokenBlockDataset( + dataset, + dataset.sizes, + self.args.tokens_per_sample - 1, # one less for + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode=self.args.sample_break_mode, + ) + logger.info('loaded {} blocks from: {}'.format(len(dataset), split_path)) + + # prepend beginning-of-sentence token (, equiv. to [CLS] in BERT) + dataset = PrependTokenDataset(dataset, self.source_dictionary.bos()) + + src_dataset, tgt_dataset = MaskTokensDataset.apply_mask( + dataset, + self.source_dictionary, + pad_idx=self.source_dictionary.pad(), + mask_idx=self.mask_idx, + seed=self.args.seed, + mask_prob=self.args.mask_prob, + leave_unmasked_prob=self.args.leave_unmasked_prob, + random_token_prob=self.args.random_token_prob, + freq_weighted_replacement=self.args.freq_weighted_replacement, + mask_whole_words=mask_whole_words, + ) + + lang_dataset = NestedDictionaryDataset( + { + 'net_input': { + 'src_tokens': PadDataset( + src_dataset, + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ), + 'src_lengths': NumelDataset(src_dataset, reduce=False), + }, + 'target': PadDataset( + tgt_dataset, + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ), + 'nsentences': NumSamplesDataset(), + 'ntokens': NumelDataset(src_dataset, reduce=True), + 'lang_id': RawLabelDataset([lang_id] * src_dataset.sizes.shape[0]), + }, + sizes=[src_dataset.sizes], + ) + lang_datasets.append(lang_dataset) + + + dataset_lengths = np.array( + [len(d) for d in lang_datasets], + dtype=float, + ) + logger.info( + 'loaded total {} blocks for all languages'.format( + dataset_lengths.sum(), + ) + ) + if split == self.args.train_subset: + # For train subset, additionally up or down sample languages. + sample_probs = self._get_sample_prob(dataset_lengths) + logger.info("Sample probability by language: ", { + lang: "{0:.4f}".format(sample_probs[id]) + for id, lang in enumerate(languages) + } + ) + size_ratio = (sample_probs * dataset_lengths.sum()) / dataset_lengths + logger.info("Up/Down Sampling ratio by language: ", { + lang: "{0:.2f}".format(size_ratio[id]) + for id, lang in enumerate(languages) + } + ) + + resampled_lang_datasets = [ + ResamplingDataset( + lang_datasets[i], + size_ratio=size_ratio[i], + seed=self.args.seed, + epoch=epoch, + replace=size_ratio[i] >= 1.0, + ) + for i, d in enumerate(lang_datasets) + ] + dataset = ConcatDataset(resampled_lang_datasets) + else: + dataset = ConcatDataset(lang_datasets) + lang_splits = [split] + for lang_id, lang_dataset in enumerate(lang_datasets): + split_name = split + '_' + languages[lang_id] + lang_splits.append(split_name) + self.datasets[split_name] = lang_dataset + + # [TODO]: This is hacky for now to print validation ppl for each + # language individually. Maybe need task API changes to allow it + # in more generic ways. + if split in self.args.valid_subset: + self.args.valid_subset = self.args.valid_subset.replace( + split, ','.join(lang_splits) + ) + + with data_utils.numpy_seed(self.args.seed + epoch): + shuffle = np.random.permutation(len(dataset)) + + self.datasets[split] = SortDataset( + dataset, + sort_order=[ + shuffle, + dataset.sizes, + ], + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths, sort=True): + src_dataset = PadDataset( + TokenBlockDataset( + src_tokens, + src_lengths, + self.args.tokens_per_sample - 1, # one less for + pad=self.source_dictionary.pad(), + eos=self.source_dictionary.eos(), + break_mode='eos', + ), + pad_idx=self.source_dictionary.pad(), + left_pad=False, + ) + src_dataset = PrependTokenDataset(src_dataset, self.source_dictionary.bos()) + src_dataset = NestedDictionaryDataset( + { + 'id': IdDataset(), + 'net_input': { + 'src_tokens': src_dataset, + 'src_lengths': NumelDataset(src_dataset, reduce=False), + }, + }, + sizes=src_lengths, + ) + if sort: + src_dataset = SortDataset(src_dataset, sort_order=[src_lengths]) + return src_dataset + + def get_batch_iterator( + self, dataset, max_tokens=None, max_sentences=None, max_positions=None, + ignore_invalid_inputs=False, required_batch_size_multiple=1, + seed=1, num_shards=1, shard_id=0, num_workers=0, epoch=1, + ): + # Recreate epoch iterator every epoch cause the underlying + # datasets are dynamic due to sampling. + self.dataset_to_epoch_iter = {} + epoch_iter = super().get_batch_iterator( + dataset, max_tokens, max_sentences, max_positions, + ignore_invalid_inputs, required_batch_size_multiple, + seed, num_shards, shard_id, num_workers, epoch, + ) + self.dataset_to_epoch_iter = {} + return epoch_iter + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/tasks/multilingual_translation.py b/fairseq/tasks/multilingual_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..59634131fcddb233c63c4cafd59b3502bb3c0c8f --- /dev/null +++ b/fairseq/tasks/multilingual_translation.py @@ -0,0 +1,358 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict +import logging +import os + +import contextlib +import torch + +from fairseq import metrics, options +from fairseq.data import ( + Dictionary, + LanguagePairDataset, + RoundRobinZipDatasets, + TransformEosLangPairDataset, +) +from fairseq.models import FairseqMultiModel +from fairseq.tasks.translation import load_langpair_dataset + +from . import FairseqTask, register_task +from fairseq import utils + +logger = logging.getLogger(__name__) + + +def _lang_token(lang: str): + return '__{}__'.format(lang) + + +def _lang_token_index(dic: Dictionary, lang: str): + """Return language token index.""" + idx = dic.index(_lang_token(lang)) + assert idx != dic.unk_index, \ + 'cannot find language token for lang {}'.format(lang) + return idx + + +@register_task('multilingual_translation') +class MultilingualTranslationTask(FairseqTask): + """A task for training multiple translation models simultaneously. + + We iterate round-robin over batches from multiple language pairs, ordered + according to the `--lang-pairs` argument. + + The training loop is roughly: + + for i in range(len(epoch)): + for lang_pair in args.lang_pairs: + batch = next_batch_for_lang_pair(lang_pair) + loss = criterion(model_for_lang_pair(lang_pair), batch) + loss.backward() + optimizer.step() + + In practice, `next_batch_for_lang_pair` is abstracted in a FairseqDataset + (e.g., `RoundRobinZipDatasets`) and `model_for_lang_pair` is a model that + implements the `FairseqMultiModel` interface. + + During inference it is required to specify a single `--source-lang` and + `--target-lang`, which indicates the inference langauge direction. + `--lang-pairs`, `--encoder-langtok`, `--decoder-langtok` have to be set to + the same value as training. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + parser.add_argument('data', metavar='DIR', help='path to data directory') + parser.add_argument('--lang-pairs', default=None, metavar='PAIRS', + help='comma-separated list of language pairs (in training order): en-de,en-fr,de-fr') + parser.add_argument('-s', '--source-lang', default=None, metavar='SRC', + help='source language (only needed for inference)') + parser.add_argument('-t', '--target-lang', default=None, metavar='TARGET', + help='target language (only needed for inference)') + parser.add_argument('--left-pad-source', default='True', type=str, metavar='BOOL', + help='pad the source on the left (default: True)') + parser.add_argument('--left-pad-target', default='False', type=str, metavar='BOOL', + help='pad the target on the left (default: False)') + parser.add_argument('--max-source-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the source sequence') + parser.add_argument('--max-target-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the target sequence') + parser.add_argument('--upsample-primary', default=1, type=int, + help='amount to upsample primary dataset') + parser.add_argument('--encoder-langtok', default=None, type=str, choices=['src', 'tgt'], + metavar='SRCTGT', + help='replace beginning-of-sentence in source sentence with source or target ' + 'language token. (src/tgt)') + parser.add_argument('--decoder-langtok', action='store_true', + help='replace beginning-of-sentence in target sentence with target language token') + # fmt: on + + def __init__(self, args, dicts, training): + super().__init__(args) + self.dicts = dicts + self.training = training + if training: + self.lang_pairs = args.lang_pairs + else: + self.lang_pairs = ['{}-{}'.format(args.source_lang, args.target_lang)] + # eval_lang_pairs for multilingual translation is usually all of the + # lang_pairs. However for other multitask settings or when we want to + # optimize for certain languages we want to use a different subset. Thus + # the eval_lang_pairs class variable is provided for classes that extend + # this class. + self.eval_lang_pairs = self.lang_pairs + # model_lang_pairs will be used to build encoder-decoder model pairs in + # models.build_model(). This allows multitask type of sub-class can + # build models other than the input lang_pairs + self.model_lang_pairs = self.lang_pairs + self.langs = list(dicts.keys()) + + @classmethod + def setup_task(cls, args, **kwargs): + dicts, training = cls.prepare(args, **kwargs) + return cls(args, dicts, training) + + @classmethod + def prepare(cls, args, **kargs): + args.left_pad_source = options.eval_bool(args.left_pad_source) + args.left_pad_target = options.eval_bool(args.left_pad_target) + + if args.lang_pairs is None: + raise ValueError('--lang-pairs is required. List all the language pairs in the training objective.') + if isinstance(args.lang_pairs, str): + args.lang_pairs = args.lang_pairs.split(',') + sorted_langs = sorted(list({x for lang_pair in args.lang_pairs for x in lang_pair.split('-')})) + if args.source_lang is not None or args.target_lang is not None: + training = False + else: + training = True + + # load dictionaries + dicts = OrderedDict() + for lang in sorted_langs: + paths = utils.split_paths(args.data) + assert len(paths) > 0 + dicts[lang] = cls.load_dictionary(os.path.join(paths[0], 'dict.{}.txt'.format(lang))) + if len(dicts) > 0: + assert dicts[lang].pad() == dicts[sorted_langs[0]].pad() + assert dicts[lang].eos() == dicts[sorted_langs[0]].eos() + assert dicts[lang].unk() == dicts[sorted_langs[0]].unk() + if args.encoder_langtok is not None or args.decoder_langtok: + for lang_to_add in sorted_langs: + dicts[lang].add_symbol(_lang_token(lang_to_add)) + logger.info('[{}] dictionary: {} types'.format(lang, len(dicts[lang]))) + return dicts, training + + def get_encoder_langtok(self, src_lang, tgt_lang): + if self.args.encoder_langtok is None: + return self.dicts[src_lang].eos() + if self.args.encoder_langtok == 'src': + return _lang_token_index(self.dicts[src_lang], src_lang) + else: + return _lang_token_index(self.dicts[src_lang], tgt_lang) + + def get_decoder_langtok(self, tgt_lang): + if not self.args.decoder_langtok: + return self.dicts[tgt_lang].eos() + return _lang_token_index(self.dicts[tgt_lang], tgt_lang) + + def alter_dataset_langtok(self, lang_pair_dataset, + src_eos=None, src_lang=None, tgt_eos=None, tgt_lang=None): + if self.args.encoder_langtok is None and not self.args.decoder_langtok: + return lang_pair_dataset + + new_src_eos = None + if self.args.encoder_langtok is not None and src_eos is not None \ + and src_lang is not None and tgt_lang is not None: + new_src_eos = self.get_encoder_langtok(src_lang, tgt_lang) + else: + src_eos = None + + new_tgt_bos = None + if self.args.decoder_langtok and tgt_eos is not None and tgt_lang is not None: + new_tgt_bos = self.get_decoder_langtok(tgt_lang) + else: + tgt_eos = None + + return TransformEosLangPairDataset( + lang_pair_dataset, + src_eos=src_eos, + new_src_eos=new_src_eos, + tgt_bos=tgt_eos, + new_tgt_bos=new_tgt_bos, + ) + + def load_dataset(self, split, epoch=1, **kwargs): + """Load a dataset split.""" + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + def language_pair_dataset(lang_pair): + src, tgt = lang_pair.split('-') + langpair_dataset = load_langpair_dataset( + data_path, split, src, self.dicts[src], tgt, self.dicts[tgt], + combine=True, dataset_impl=self.args.dataset_impl, + upsample_primary=self.args.upsample_primary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + max_source_positions=self.args.max_source_positions, + max_target_positions=self.args.max_target_positions, + ) + return self.alter_dataset_langtok( + langpair_dataset, + src_eos=self.dicts[src].eos(), + src_lang=src, + tgt_eos=self.dicts[tgt].eos(), + tgt_lang=tgt, + ) + + self.datasets[split] = RoundRobinZipDatasets( + OrderedDict([ + (lang_pair, language_pair_dataset(lang_pair)) + for lang_pair in self.lang_pairs + ]), + eval_key=None if self.training else "%s-%s" % (self.args.source_lang, self.args.target_lang), + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths): + lang_pair = "%s-%s" % (self.args.source_lang, self.args.target_lang) + return RoundRobinZipDatasets( + OrderedDict([( + lang_pair, + self.alter_dataset_langtok( + LanguagePairDataset( + src_tokens, src_lengths, + self.source_dictionary + ), + src_eos=self.source_dictionary.eos(), + src_lang=self.args.source_lang, + tgt_eos=self.target_dictionary.eos(), + tgt_lang=self.args.target_lang, + ), + )]), + eval_key=lang_pair, + ) + + def build_model(self, args): + def check_args(): + messages = [] + if len(set(self.args.lang_pairs).symmetric_difference(args.lang_pairs)) != 0: + messages.append('--lang-pairs should include all the language pairs {}.'.format(args.lang_pairs)) + if self.args.encoder_langtok != args.encoder_langtok: + messages.append('--encoder-langtok should be {}.'.format(args.encoder_langtok)) + if self.args.decoder_langtok != args.decoder_langtok: + messages.append('--decoder-langtok should {} be set.'.format("" if args.decoder_langtok else "not")) + + if len(messages) > 0: + raise ValueError(' '.join(messages)) + + # Check if task args are consistant with model args + check_args() + + from fairseq import models + model = models.build_model(args, self) + if not isinstance(model, FairseqMultiModel): + raise ValueError('MultilingualTranslationTask requires a FairseqMultiModel architecture') + return model + + def train_step(self, sample, model, criterion, optimizer, update_num, ignore_grad=False): + model.train() + from collections import defaultdict + agg_loss, agg_sample_size, agg_logging_output = 0., 0., defaultdict(float) + curr_lang_pairs = [ + lang_pair + for lang_pair in self.model_lang_pairs + if sample[lang_pair] is not None and len(sample[lang_pair]) != 0 + ] + + for idx, lang_pair in enumerate(curr_lang_pairs): + def maybe_no_sync(): + if ( + self.args.distributed_world_size > 1 + and hasattr(model, 'no_sync') + and idx < len(curr_lang_pairs) - 1 + ): + return model.no_sync() + else: + return contextlib.ExitStack() # dummy contextmanager + with maybe_no_sync(): + loss, sample_size, logging_output = criterion(model.models[lang_pair], sample[lang_pair]) + if ignore_grad: + loss *= 0 + optimizer.backward(loss) + agg_loss += loss.detach().item() + # TODO make summing of the sample sizes configurable + agg_sample_size += sample_size + for k in logging_output: + agg_logging_output[k] += logging_output[k] + agg_logging_output[f"{lang_pair}:{k}"] += logging_output[k] + return agg_loss, agg_sample_size, agg_logging_output + + def valid_step(self, sample, model, criterion): + model.eval() + with torch.no_grad(): + from collections import defaultdict + agg_loss, agg_sample_size, agg_logging_output = 0., 0., defaultdict(float) + for lang_pair in self.eval_lang_pairs: + if lang_pair not in sample or sample[lang_pair] is None or len(sample[lang_pair]) == 0: + continue + loss, sample_size, logging_output = criterion(model.models[lang_pair], sample[lang_pair]) + agg_loss += loss.data.item() + # TODO make summing of the sample sizes configurable + agg_sample_size += sample_size + for k in logging_output: + agg_logging_output[k] += logging_output[k] + agg_logging_output[f"{lang_pair}:{k}"] += logging_output[k] + return agg_loss, agg_sample_size, agg_logging_output + + def inference_step(self, generator, models, sample, prefix_tokens=None): + with torch.no_grad(): + if self.args.decoder_langtok: + bos_token = _lang_token_index(self.target_dictionary, self.args.target_lang) + else: + bos_token = self.target_dictionary.eos() + return generator.generate( + models, + sample, + prefix_tokens=prefix_tokens, + bos_token=bos_token, + ) + + def reduce_metrics(self, logging_outputs, criterion): + with metrics.aggregate(): + # pass 'sample_size', 'nsentences', 'ntokens' stats to fairseq_task + super().reduce_metrics(logging_outputs, criterion) + for k in ['sample_size', 'nsentences', 'ntokens']: + metrics.log_scalar(k, sum(l[k] for l in logging_outputs)) + + @property + def source_dictionary(self): + if self.training: + return next(iter(self.dicts.values())) + else: + return self.dicts[self.args.source_lang] + + @property + def target_dictionary(self): + if self.training: + return next(iter(self.dicts.values())) + else: + return self.dicts[self.args.target_lang] + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + if len(self.datasets.values()) == 0: + return {'%s-%s' % (self.args.source_lang, self.args.target_lang): + (self.args.max_source_positions, self.args.max_target_positions)} + return OrderedDict([ + (key, (self.args.max_source_positions, self.args.max_target_positions)) + for split in self.datasets.keys() + for key in self.datasets[split].datasets.keys() + ]) diff --git a/fairseq/tasks/semisupervised_translation.py b/fairseq/tasks/semisupervised_translation.py new file mode 100644 index 0000000000000000000000000000000000000000..c81d3628862f9b7ccfaf76a336b878f67a3a8f9e --- /dev/null +++ b/fairseq/tasks/semisupervised_translation.py @@ -0,0 +1,385 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from collections import OrderedDict +import logging +import os + +from fairseq.data import ( + BacktranslationDataset, + data_utils, + indexed_dataset, + IndexedCachedDataset, + IndexedDataset, + IndexedRawTextDataset, + LanguagePairDataset, + NoisingDataset, + RoundRobinZipDatasets, +) +from fairseq.models import FairseqMultiModel +from fairseq.sequence_generator import SequenceGenerator + +from .multilingual_translation import MultilingualTranslationTask +from . import register_task +from fairseq import utils + +logger = logging.getLogger(__name__) + + +def _get_bt_dataset_key(lang_pair): + return "bt:" + lang_pair + + +def _get_denoising_dataset_key(lang_pair): + return "denoising:" + lang_pair + + +# ported from UnsupervisedMT +def parse_lambda_config(x): + """ + Parse the configuration of lambda coefficient (for scheduling). + x = "3" # lambda will be a constant equal to x + x = "0:1,1000:0" # lambda will start from 1 and linearly decrease + # to 0 during the first 1000 iterations + x = "0:0,1000:0,2000:1" # lambda will be equal to 0 for the first 1000 + # iterations, then will linearly increase to 1 until iteration 2000 + """ + split = x.split(',') + if len(split) == 1: + return float(x), None + else: + split = [s.split(os.pathsep) for s in split] + assert all(len(s) == 2 for s in split) + assert all(k.isdigit() for k, _ in split) + assert all(int(split[i][0]) < int(split[i + 1][0]) for i in range(len(split) - 1)) + return float(split[0][1]), [(int(k), float(v)) for k, v in split] + + +@register_task('semisupervised_translation') +class SemisupervisedTranslationTask(MultilingualTranslationTask): + """A task for training multiple translation models simultaneously. + + We iterate round-robin over batches from multiple language pairs, ordered + according to the `--lang-pairs` argument. + + The training loop is roughly: + + for i in range(len(epoch)): + for lang_pair in args.lang_pairs: + batch = next_batch_for_lang_pair(lang_pair) + loss = criterion(model_for_lang_pair(lang_pair), batch) + loss.backward() + optimizer.step() + + In practice, `next_batch_for_lang_pair` is abstracted in a FairseqDataset + (e.g., `RoundRobinZipDatasets`) and `model_for_lang_pair` is a model that + implements the `FairseqMultiModel` interface. + + During inference it is required to specify a single `--source-lang` and + `--target-lang`, instead of `--lang-pairs`. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + MultilingualTranslationTask.add_args(parser) + parser.add_argument('--lambda-parallel-config', default="1.0", type=str, metavar='CONFIG', + help='cross-entropy reconstruction coefficient (parallel data). ' + 'use fixed weight during training if set to floating point number. ' + 'use piecewise linear function over number of updates to schedule the ' + 'weight with the format: w0:step0,w1:step1,...') + parser.add_argument('--lambda-denoising-config', default="0.0", type=str, metavar='CONFIG', + help='Cross-entropy reconstruction coefficient (denoising autoencoding)' + 'use fixed weight during training if set to floating point number. ' + 'use piecewise linear function over number of updates to schedule the ' + 'weight with the format: w0:step0,w1:step1,...') + parser.add_argument('--lambda-otf-bt-config', default="0.0", type=str, metavar='CONFIG', + help='cross-entropy reconstruction coefficient (on-the-fly back-translation parallel data)' + 'use fixed weight during training if set to floating point number. ' + 'use piecewise linear function over number of updates to schedule the ' + 'weight with the format: w0:step0,w1:step1,...') + parser.add_argument('--bt-max-len-a', default=1.1, type=float, metavar='N', + help='generate back-translated sequences of maximum length ax + b, where x is the ' + 'source length') + parser.add_argument('--bt-max-len-b', default=10.0, type=float, metavar='N', + help='generate back-translated sequences of maximum length ax + b, where x is the ' + 'source length') + parser.add_argument('--bt-beam-size', default=1, type=int, metavar='N', + help='beam size used in beam search of online back-translation') + parser.add_argument('--max-word-shuffle-distance', default=3.0, type=float, metavar='N', + help='maximum word shuffle distance for denoising autoencoding data generation') + parser.add_argument('--word-dropout-prob', default=0.1, type=float, metavar='N', + help='word dropout probability for denoising autoencoding data generation') + parser.add_argument('--word-blanking-prob', default=0.2, type=float, metavar='N', + help='word blanking probability for denoising autoencoding data generation') + # fmt: on + + def __init__(self, args, dicts, training): + super().__init__(args, dicts, training) + self.lambda_parallel, self.lambda_parallel_steps = parse_lambda_config(args.lambda_parallel_config) + self.lambda_otf_bt, self.lambda_otf_bt_steps = parse_lambda_config(args.lambda_otf_bt_config) + self.lambda_denoising, self.lambda_denoising_steps = parse_lambda_config(args.lambda_denoising_config) + if (self.lambda_denoising > 0.0 or self.lambda_denoising_steps is not None): + denoising_lang_pairs = [ + "%s-%s" % (tgt, tgt) + for tgt in {lang_pair.split('-')[1] for lang_pair in args.lang_pairs} + ] + self.model_lang_pairs = self.model_lang_pairs + denoising_lang_pairs + self.backtranslate_datasets = {} + self.backtranslators = {} + + @classmethod + def setup_task(cls, args, **kwargs): + dicts, training = MultilingualTranslationTask.prepare(args, **kwargs) + return cls(args, dicts, training) + + def load_dataset(self, split, epoch=1, **kwargs): + """Load a dataset split.""" + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + def split_exists(split, src, tgt, lang): + if src is not None: + filename = os.path.join(data_path, '{}.{}-{}.{}'.format(split, src, tgt, lang)) + else: + filename = os.path.join(data_path, '{}.{}-None.{}'.format(split, src, tgt)) + return indexed_dataset.dataset_exists(filename, impl=self.args.dataset_impl) + + def load_indexed_dataset(path, dictionary): + return data_utils.load_indexed_dataset(path, dictionary, self.args.dataset_impl) + + # load parallel datasets + src_datasets, tgt_datasets = {}, {} + if (self.lambda_parallel > 0.0 or self.lambda_parallel_steps is not None or not split.startswith("train")): + for lang_pair in self.lang_pairs: + src, tgt = lang_pair.split('-') + if split_exists(split, src, tgt, src): + prefix = os.path.join(data_path, '{}.{}-{}.'.format(split, src, tgt)) + elif split_exists(split, tgt, src, src): + prefix = os.path.join(data_path, '{}.{}-{}.'.format(split, tgt, src)) + else: + continue + src_datasets[lang_pair] = load_indexed_dataset(prefix + src, self.dicts[src]) + tgt_datasets[lang_pair] = load_indexed_dataset(prefix + tgt, self.dicts[tgt]) + logger.info('parallel-{} {} {} examples'.format(data_path, split, len(src_datasets[lang_pair]))) + if len(src_datasets) == 0: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, data_path)) + + # back translation datasets + backtranslate_datasets = {} + if (self.lambda_otf_bt > 0.0 or self.lambda_otf_bt_steps is not None) and split.startswith("train"): + for lang_pair in self.lang_pairs: + src, tgt = lang_pair.split('-') + if not split_exists(split, tgt, None, tgt): + raise FileNotFoundError('Dataset not found: backtranslation {} ({})'.format(split, data_path)) + filename = os.path.join(data_path, '{}.{}-None.{}'.format(split, tgt, tgt)) + dataset = load_indexed_dataset(filename, self.dicts[tgt]) + lang_pair_dataset_tgt = LanguagePairDataset( + dataset, + dataset.sizes, + self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ) + lang_pair_dataset = LanguagePairDataset( + dataset, + dataset.sizes, + src_dict=self.dicts[src], + tgt=dataset, + tgt_sizes=dataset.sizes, + tgt_dict=self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ) + backtranslate_datasets[lang_pair] = BacktranslationDataset( + tgt_dataset=self.alter_dataset_langtok( + lang_pair_dataset_tgt, + src_eos=self.dicts[tgt].eos(), + src_lang=tgt, + tgt_lang=src, + ), + backtranslation_fn=self.backtranslators[lang_pair], + src_dict=self.dicts[src], tgt_dict=self.dicts[tgt], + output_collater=self.alter_dataset_langtok( + lang_pair_dataset=lang_pair_dataset, + src_eos=self.dicts[src].eos(), + src_lang=src, + tgt_eos=self.dicts[tgt].eos(), + tgt_lang=tgt, + ).collater, + ) + logger.info('backtranslate-{}: {} {} {} examples'.format( + tgt, data_path, split, len(backtranslate_datasets[lang_pair]), + )) + self.backtranslate_datasets[lang_pair] = backtranslate_datasets[lang_pair] + + # denoising autoencoder + noising_datasets = {} + if (self.lambda_denoising > 0.0 or self.lambda_denoising_steps is not None) and split.startswith("train"): + for lang_pair in self.lang_pairs: + _, tgt = lang_pair.split('-') + if not split_exists(split, tgt, None, tgt): + continue + filename = os.path.join(data_path, '{}.{}-None.{}'.format(split, tgt, tgt)) + tgt_dataset1 = load_indexed_dataset(filename, self.dicts[tgt]) + tgt_dataset2 = load_indexed_dataset(filename, self.dicts[tgt]) + noising_dataset = NoisingDataset( + tgt_dataset1, + self.dicts[tgt], + seed=1, + max_word_shuffle_distance=self.args.max_word_shuffle_distance, + word_dropout_prob=self.args.word_dropout_prob, + word_blanking_prob=self.args.word_blanking_prob, + ) + noising_datasets[lang_pair] = self.alter_dataset_langtok( + LanguagePairDataset( + noising_dataset, + tgt_dataset1.sizes, + self.dicts[tgt], + tgt_dataset2, + tgt_dataset2.sizes, + self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ), + src_eos=self.dicts[tgt].eos(), + src_lang=tgt, + tgt_eos=self.dicts[tgt].eos(), + tgt_lang=tgt, + ) + logger.info('denoising-{}: {} {} {} examples'.format( + tgt, data_path, split, len(noising_datasets[lang_pair]), + )) + + def language_pair_dataset(lang_pair): + src, tgt = lang_pair.split('-') + src_dataset, tgt_dataset = src_datasets[lang_pair], tgt_datasets[lang_pair] + return self.alter_dataset_langtok( + LanguagePairDataset( + src_dataset, src_dataset.sizes, self.dicts[src], + tgt_dataset, tgt_dataset.sizes, self.dicts[tgt], + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + ), + self.dicts[src].eos(), + src, + self.dicts[tgt].eos(), + tgt, + ) + + self.datasets[split] = RoundRobinZipDatasets( + OrderedDict([ + (lang_pair, language_pair_dataset(lang_pair)) + for lang_pair in src_datasets.keys() + ] + [ + (_get_bt_dataset_key(lang_pair), dataset) + for lang_pair, dataset in backtranslate_datasets.items() + ] + [ + (_get_denoising_dataset_key(lang_pair), dataset) + for lang_pair, dataset in noising_datasets.items() + ]), + eval_key=None if self.training else "%s-%s" % (self.args.source_lang, self.args.target_lang), + ) + + def build_model(self, args): + from fairseq import models + model = models.build_model(args, self) + if not isinstance(model, FairseqMultiModel): + raise ValueError('SemisupervisedTranslationTask requires a FairseqMultiModel architecture') + + # create SequenceGenerator for each model that has backtranslation dependency on it + self.sequence_generators = {} + if (self.lambda_otf_bt > 0.0 or self.lambda_otf_bt_steps is not None) and self.training: + for lang_pair in self.lang_pairs: + src, tgt = lang_pair.split('-') + key = '{}-{}'.format(tgt, src) + self.sequence_generators[key] = SequenceGenerator( + [model.models[key]], + tgt_dict=self.dicts[src], + beam_size=args.bt_beam_size, + max_len_a=args.bt_max_len_a, + max_len_b=args.bt_max_len_b, + ) + decoder_lang_tok_idx = self.get_decoder_langtok(src) + + def backtranslate_fn( + sample, model=model.models[key], + bos_token=decoder_lang_tok_idx, + sequence_generator=self.sequence_generators[key], + ): + return sequence_generator.generate( + [model], + sample, + bos_token=bos_token, + ) + self.backtranslators[lang_pair] = backtranslate_fn + + return model + + def train_step(self, sample, model, criterion, optimizer, update_num, ignore_grad=False): + model.train() + + if update_num > 0: + self.update_step(update_num) + + agg_loss, agg_sample_size, agg_logging_output = 0., 0., {} + + def forward_backward(model, samples, logging_output_key, weight): + nonlocal agg_loss, agg_sample_size, agg_logging_output + if samples is None or len(samples) == 0: + return + loss, sample_size, logging_output = criterion(model, samples) + if ignore_grad: + loss *= 0 + else: + loss *= weight + optimizer.backward(loss) + agg_loss += loss.detach().item() + # TODO make summing of the sample sizes configurable + agg_sample_size += sample_size + for k in logging_output: + agg_logging_output[k] += logging_output[k] + agg_logging_output[logging_output_key] += logging_output[k] + + if self.lambda_parallel > 0.0: + for lang_pair in self.lang_pairs: + forward_backward(model.models[lang_pair], sample[lang_pair], lang_pair, self.lambda_parallel) + + if self.lambda_otf_bt > 0.0: + for lang_pair in self.lang_pairs: + sample_key = _get_bt_dataset_key(lang_pair) + forward_backward(model.models[lang_pair], sample[sample_key], sample_key, self.lambda_otf_bt) + + if self.lambda_denoising > 0.0: + for lang_pair in self.lang_pairs: + _, tgt = lang_pair.split('-') + sample_key = _get_denoising_dataset_key(lang_pair) + forward_backward(model.models['{0}-{0}'.format(tgt)], sample[sample_key], sample_key, self.lambda_denoising) + + return agg_loss, agg_sample_size, agg_logging_output + + def update_step(self, num_updates): + def lambda_step_func(config, n_iter): + """ + Update a lambda value according to its schedule configuration. + """ + ranges = [i for i in range(len(config) - 1) if config[i][0] <= n_iter < config[i + 1][0]] + if len(ranges) == 0: + assert n_iter >= config[-1][0] + return config[-1][1] + assert len(ranges) == 1 + i = ranges[0] + x_a, y_a = config[i] + x_b, y_b = config[i + 1] + return y_a + (n_iter - x_a) * float(y_b - y_a) / float(x_b - x_a) + + if self.lambda_parallel_steps is not None: + self.lambda_parallel = lambda_step_func(self.lambda_parallel_steps, num_updates) + if self.lambda_denoising_steps is not None: + self.lambda_denoising = lambda_step_func(self.lambda_denoising_steps, num_updates) + if self.lambda_otf_bt_steps is not None: + self.lambda_otf_bt = lambda_step_func(self.lambda_otf_bt_steps, num_updates) diff --git a/fairseq/tasks/sentence_prediction.py b/fairseq/tasks/sentence_prediction.py new file mode 100644 index 0000000000000000000000000000000000000000..b50c9922cc9c6472fbc66bb0d248444cd16f46cd --- /dev/null +++ b/fairseq/tasks/sentence_prediction.py @@ -0,0 +1,249 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np + +from fairseq import utils +from fairseq.data import ( + ConcatSentencesDataset, + data_utils, + Dictionary, + IdDataset, + NestedDictionaryDataset, + NumSamplesDataset, + NumelDataset, + OffsetTokensDataset, + PrependTokenDataset, + RawLabelDataset, + RightPadDataset, + RollDataset, + SortDataset, + StripTokenDataset, +) +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import FairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task('sentence_prediction') +class SentencePredictionTask(FairseqTask): + """ + Sentence (or sentence pair) prediction (classification or regression) task. + + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('data', metavar='FILE', + help='file prefix for data') + parser.add_argument('--num-classes', type=int, default=-1, + help='number of classes or regression targets') + parser.add_argument('--init-token', type=int, default=None, + help='add token at the beginning of each batch item') + parser.add_argument('--separator-token', type=int, default=None, + help='add separator token between inputs') + parser.add_argument('--regression-target', action='store_true', default=False) + parser.add_argument('--no-shuffle', action='store_true', default=False) + parser.add_argument('--shorten-method', default='none', + choices=['none', 'truncate', 'random_crop'], + help='if not none, shorten sequences that exceed --tokens-per-sample') + parser.add_argument('--shorten-data-split-list', default='', + help='comma-separated list of dataset splits to apply shortening to, ' + 'e.g., "train,valid" (default: all dataset splits)') + parser.add_argument('--add-prev-output-tokens', action='store_true', default=False, + help='add prev_output_tokens to sample, used for encoder-decoder arch') + + def __init__(self, args, data_dictionary, label_dictionary): + super().__init__(args) + self.dictionary = data_dictionary + self._label_dictionary = label_dictionary + if not hasattr(args, 'max_positions'): + self._max_positions = ( + args.max_source_positions, + args.max_target_positions, + ) + else: + self._max_positions = args.max_positions + args.tokens_per_sample = self._max_positions + + @classmethod + def load_dictionary(cls, args, filename, source=True): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + dictionary = Dictionary.load(filename) + dictionary.add_symbol('') + return dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + assert args.num_classes > 0, 'Must set --num-classes' + + # load data dictionary + data_dict = cls.load_dictionary( + args, + os.path.join(args.data, 'input0', 'dict.txt'), + source=True, + ) + logger.info('[input] dictionary: {} types'.format(len(data_dict))) + + label_dict = None + if not args.regression_target: + # load label dictionary + label_dict = cls.load_dictionary( + args, + os.path.join(args.data, 'label', 'dict.txt'), + source=False, + ) + logger.info('[label] dictionary: {} types'.format(len(label_dict))) + else: + label_dict = data_dict + return SentencePredictionTask(args, data_dict, label_dict) + + def load_dataset(self, split, combine=False, **kwargs): + """Load a given dataset split (e.g., train, valid, test).""" + def get_path(type, split): + return os.path.join(self.args.data, type, split) + + def make_dataset(type, dictionary): + split_path = get_path(type, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + dictionary, + self.args.dataset_impl, + combine=combine, + ) + return dataset + + input0 = make_dataset('input0', self.source_dictionary) + assert input0 is not None, 'could not find dataset: {}'.format(get_path(type, split)) + input1 = make_dataset('input1', self.source_dictionary) + + if self.args.init_token is not None: + input0 = PrependTokenDataset(input0, self.args.init_token) + + if input1 is None: + src_tokens = input0 + else: + if self.args.separator_token is not None: + input1 = PrependTokenDataset(input1, self.args.separator_token) + + src_tokens = ConcatSentencesDataset(input0, input1) + + with data_utils.numpy_seed(self.args.seed): + shuffle = np.random.permutation(len(src_tokens)) + + src_tokens = maybe_shorten_dataset( + src_tokens, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.max_positions, + self.args.seed, + ) + + dataset = { + 'id': IdDataset(), + 'net_input': { + 'src_tokens': RightPadDataset( + src_tokens, + pad_idx=self.source_dictionary.pad(), + ), + 'src_lengths': NumelDataset(src_tokens, reduce=False), + }, + 'nsentences': NumSamplesDataset(), + 'ntokens': NumelDataset(src_tokens, reduce=True), + } + + if self.args.add_prev_output_tokens: + prev_tokens_dataset = RightPadDataset( + RollDataset(src_tokens, 1), + pad_idx=self.dictionary.pad(), + ) + dataset['net_input'].update( + prev_output_tokens=prev_tokens_dataset, + ) + + if not self.args.regression_target: + label_dataset = make_dataset('label', self.label_dictionary) + if label_dataset is not None: + dataset.update( + target=OffsetTokensDataset( + StripTokenDataset( + label_dataset, + id_to_strip=self.label_dictionary.eos(), + ), + offset=-self.label_dictionary.nspecial, + ) + ) + else: + label_path = "{0}.label".format(get_path('label', split)) + if os.path.exists(label_path): + def parse_regression_target(i, line): + values = line.split() + assert len(values) == self.args.num_classes, \ + f'expected num_classes={self.args.num_classes} regression target values on line {i}, found: "{line}"' + return [float(x) for x in values] + dataset.update( + target=RawLabelDataset([ + parse_regression_target(i, line.strip()) for i, line in enumerate(open(label_path).readlines()) + ]) + ) + + nested_dataset = NestedDictionaryDataset( + dataset, + sizes=[src_tokens.sizes], + ) + + if self.args.no_shuffle: + dataset = nested_dataset + else: + dataset = SortDataset( + nested_dataset, + # shuffle + sort_order=[shuffle], + ) + + logger.info("Loaded {0} with #samples: {1}".format(split, len(dataset))) + + self.datasets[split] = dataset + return self.datasets[split] + + def build_model(self, args): + from fairseq import models + model = models.build_model(args, self) + + model.register_classification_head( + getattr(args, 'classification_head_name', 'sentence_classification_head'), + num_classes=self.args.num_classes, + ) + + return model + + def max_positions(self): + return self._max_positions + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary + + @property + def label_dictionary(self): + return self._label_dictionary diff --git a/fairseq/tasks/sentence_ranking.py b/fairseq/tasks/sentence_ranking.py new file mode 100644 index 0000000000000000000000000000000000000000..ea4b50a294a9cac05b28d960a0925a5dc6f30795 --- /dev/null +++ b/fairseq/tasks/sentence_ranking.py @@ -0,0 +1,208 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os + +import numpy as np + +from fairseq import utils +from fairseq.data import ( + ConcatSentencesDataset, + data_utils, + Dictionary, + IdDataset, + NestedDictionaryDataset, + NumSamplesDataset, + NumelDataset, + PrependTokenDataset, + RawLabelDataset, + RightPadDataset, + SortDataset, + TruncateDataset +) +from fairseq.data.shorten_dataset import maybe_shorten_dataset +from fairseq.tasks import FairseqTask, register_task + + +logger = logging.getLogger(__name__) + + +@register_task('sentence_ranking') +class SentenceRankingTask(FairseqTask): + """ + Ranking task on multiple sentences. + + Args: + dictionary (Dictionary): the dictionary for the input of the task + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + parser.add_argument('data', metavar='FILE', + help='file prefix for data') + parser.add_argument('--num-classes', type=int, + help='number of sentences to be ranked') + parser.add_argument('--init-token', type=int, + help='add token at the beginning of each batch item') + parser.add_argument('--separator-token', type=int, + help='add separator token between inputs') + parser.add_argument('--no-shuffle', action='store_true') + parser.add_argument('--shorten-method', default='none', + choices=['none', 'truncate', 'random_crop'], + help='if not none, shorten sequences that exceed --tokens-per-sample') + parser.add_argument('--shorten-data-split-list', default='', + help='comma-separated list of dataset splits to apply shortening to, ' + 'e.g., "train,valid" (default: all dataset splits)') + parser.add_argument('--max-option-length', type=int, + help='max length for each option') + + def __init__(self, args, dictionary): + super().__init__(args) + self.dictionary = dictionary + + @classmethod + def load_dictionary(cls, args, filename, source=True): + """Load the dictionary from the filename + + Args: + filename (str): the filename + """ + dictionary = Dictionary.load(filename) + dictionary.add_symbol('') + return dictionary + + @classmethod + def setup_task(cls, args, **kwargs): + assert args.criterion == 'sentence_ranking', \ + 'Must set --criterion=sentence_ranking' + + # load data dictionary + data_dict = cls.load_dictionary( + args, + os.path.join(args.data, 'input0', 'dict.txt'), + source=True, + ) + logger.info('[input] dictionary: {} types'.format(len(data_dict))) + return SentenceRankingTask(args, data_dict) + + def load_dataset(self, split, combine=False, **kwargs): + """Load a given dataset split (e.g., train, valid, test).""" + + def get_path(type, split): + return os.path.join(self.args.data, type, split) + + def make_dataset(type, dictionary): + split_path = get_path(type, split) + + dataset = data_utils.load_indexed_dataset( + split_path, + self.source_dictionary, + self.args.dataset_impl, + combine=combine, + ) + return dataset + + input0 = make_dataset('input0', self.source_dictionary) + input_options = [ + make_dataset( + 'input{idx}'.format(idx=idx + 1), + self.source_dictionary + ) + for idx in range(self.args.num_classes) + ] + + if self.args.separator_token is not None: + input0 = PrependTokenDataset(input0, self.args.separator_token) + + src_tokens = [] + for input_option in input_options: + if self.args.init_token is not None: + input_option = PrependTokenDataset(input_option, self.args.init_token) + if self.args.max_option_length is not None: + input_option = TruncateDataset(input_option, self.args.max_option_length) + src_token = ConcatSentencesDataset(input_option, input0) + src_token = maybe_shorten_dataset( + src_token, + split, + self.args.shorten_data_split_list, + self.args.shorten_method, + self.args.max_positions, + self.args.seed, + ) + src_tokens.append(src_token) + + with data_utils.numpy_seed(self.args.seed): + shuffle = np.random.permutation(len(src_tokens[0])) + + dataset = { + 'id': IdDataset(), + 'nsentences': NumSamplesDataset(), + 'ntokens': NumelDataset(src_tokens[0], reduce=True), + } + + for src_token_idx in range(len(src_tokens)): + dataset.update( + { + 'net_input{idx}'.format(idx=src_token_idx+1): { + 'src_tokens': RightPadDataset( + src_tokens[src_token_idx], + pad_idx=self.source_dictionary.pad(), + ), + 'src_lengths': NumelDataset(src_tokens[src_token_idx], reduce=False), + } + } + ) + + label_path = '{}.label'.format(get_path('label', split)) + if os.path.exists(label_path): + with open(label_path) as h: + dataset.update( + target=RawLabelDataset([ + int(x.strip()) for x in h.readlines() + ]) + ) + + nested_dataset = NestedDictionaryDataset( + dataset, + sizes=[np.maximum.reduce([src_token.sizes for src_token in src_tokens])], + ) + + if self.args.no_shuffle: + dataset = nested_dataset + else: + dataset = SortDataset( + nested_dataset, + # shuffle + sort_order=[shuffle], + ) + + logger.info("Loaded {0} with #samples: {1}".format(split, len(dataset))) + + self.datasets[split] = dataset + return self.datasets[split] + + def build_model(self, args): + from fairseq import models + model = models.build_model(args, self) + + model.register_classification_head( + getattr(args, 'ranking_head_name', 'sentence_classification_head'), + num_classes=1, + ) + + return model + + def max_positions(self): + return self.args.max_positions + + @property + def source_dictionary(self): + return self.dictionary + + @property + def target_dictionary(self): + return self.dictionary diff --git a/fairseq/tasks/translation.py b/fairseq/tasks/translation.py new file mode 100644 index 0000000000000000000000000000000000000000..7077943c1ef8d51f2393026332f047af4e235c1f --- /dev/null +++ b/fairseq/tasks/translation.py @@ -0,0 +1,389 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from argparse import Namespace +import json +import itertools +import logging +import os + +import numpy as np + +from fairseq import metrics, options, utils +from fairseq.data import ( + AppendTokenDataset, + ConcatDataset, + data_utils, + encoders, + indexed_dataset, + LanguagePairDataset, + PrependTokenDataset, + StripTokenDataset, + TruncateDataset, +) + +from fairseq.tasks import FairseqTask, register_task + +EVAL_BLEU_ORDER = 4 + + +logger = logging.getLogger(__name__) + + +def load_langpair_dataset( + data_path, split, + src, src_dict, + tgt, tgt_dict, + combine, dataset_impl, upsample_primary, + left_pad_source, left_pad_target, max_source_positions, + max_target_positions, prepend_bos=False, load_alignments=False, + truncate_source=False, append_source_id=False, + num_buckets=0, + shuffle=True, +): + + def split_exists(split, src, tgt, lang, data_path): + filename = os.path.join(data_path, '{}.{}-{}.{}'.format(split, src, tgt, lang)) + return indexed_dataset.dataset_exists(filename, impl=dataset_impl) + + src_datasets = [] + tgt_datasets = [] + + for k in itertools.count(): + split_k = split + (str(k) if k > 0 else '') + + # infer langcode + if split_exists(split_k, src, tgt, src, data_path): + prefix = os.path.join(data_path, '{}.{}-{}.'.format(split_k, src, tgt)) + elif split_exists(split_k, tgt, src, src, data_path): + prefix = os.path.join(data_path, '{}.{}-{}.'.format(split_k, tgt, src)) + else: + if k > 0: + break + else: + raise FileNotFoundError('Dataset not found: {} ({})'.format(split, data_path)) + + src_dataset = data_utils.load_indexed_dataset(prefix + src, src_dict, dataset_impl) + if truncate_source: + src_dataset = AppendTokenDataset( + TruncateDataset( + StripTokenDataset(src_dataset, src_dict.eos()), + max_source_positions - 1, + ), + src_dict.eos(), + ) + src_datasets.append(src_dataset) + + tgt_dataset = data_utils.load_indexed_dataset(prefix + tgt, tgt_dict, dataset_impl) + if tgt_dataset is not None: + tgt_datasets.append(tgt_dataset) + + logger.info('{} {} {}-{} {} examples'.format( + data_path, split_k, src, tgt, len(src_datasets[-1]) + )) + + if not combine: + break + + assert len(src_datasets) == len(tgt_datasets) or len(tgt_datasets) == 0 + + if len(src_datasets) == 1: + src_dataset = src_datasets[0] + tgt_dataset = tgt_datasets[0] if len(tgt_datasets) > 0 else None + else: + sample_ratios = [1] * len(src_datasets) + sample_ratios[0] = upsample_primary + src_dataset = ConcatDataset(src_datasets, sample_ratios) + if len(tgt_datasets) > 0: + tgt_dataset = ConcatDataset(tgt_datasets, sample_ratios) + else: + tgt_dataset = None + + if prepend_bos: + assert hasattr(src_dict, "bos_index") and hasattr(tgt_dict, "bos_index") + src_dataset = PrependTokenDataset(src_dataset, src_dict.bos()) + if tgt_dataset is not None: + tgt_dataset = PrependTokenDataset(tgt_dataset, tgt_dict.bos()) + + eos = None + if append_source_id: + src_dataset = AppendTokenDataset(src_dataset, src_dict.index('[{}]'.format(src))) + if tgt_dataset is not None: + tgt_dataset = AppendTokenDataset(tgt_dataset, tgt_dict.index('[{}]'.format(tgt))) + eos = tgt_dict.index('[{}]'.format(tgt)) + + align_dataset = None + if load_alignments: + align_path = os.path.join(data_path, '{}.align.{}-{}'.format(split, src, tgt)) + if indexed_dataset.dataset_exists(align_path, impl=dataset_impl): + align_dataset = data_utils.load_indexed_dataset(align_path, None, dataset_impl) + + tgt_dataset_sizes = tgt_dataset.sizes if tgt_dataset is not None else None + return LanguagePairDataset( + src_dataset, src_dataset.sizes, src_dict, + tgt_dataset, tgt_dataset_sizes, tgt_dict, + left_pad_source=left_pad_source, + left_pad_target=left_pad_target, + align_dataset=align_dataset, eos=eos, + num_buckets=num_buckets, + shuffle=shuffle, + ) + + +@register_task('translation') +class TranslationTask(FairseqTask): + """ + Translate from one (source) language to another (target) language. + + Args: + src_dict (~fairseq.data.Dictionary): dictionary for the source language + tgt_dict (~fairseq.data.Dictionary): dictionary for the target language + + .. note:: + + The translation task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate` and :mod:`fairseq-interactive`. + + The translation task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.translation_parser + :prog: + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + parser.add_argument('data', help='colon separated path to data directories list, \ + will be iterated upon during epochs in round-robin manner') + parser.add_argument('-s', '--source-lang', default=None, metavar='SRC', + help='source language') + parser.add_argument('-t', '--target-lang', default=None, metavar='TARGET', + help='target language') + parser.add_argument('--load-alignments', action='store_true', + help='load the binarized alignments') + parser.add_argument('--left-pad-source', default='True', type=str, metavar='BOOL', + help='pad the source on the left') + parser.add_argument('--left-pad-target', default='False', type=str, metavar='BOOL', + help='pad the target on the left') + parser.add_argument('--max-source-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the source sequence') + parser.add_argument('--max-target-positions', default=1024, type=int, metavar='N', + help='max number of tokens in the target sequence') + parser.add_argument('--upsample-primary', default=1, type=int, + help='amount to upsample primary dataset') + parser.add_argument('--truncate-source', action='store_true', default=False, + help='truncate source to max-source-positions') + parser.add_argument('--num-batch-buckets', default=0, type=int, metavar='N', + help='if >0, then bucket source and target lengths into N ' + 'buckets and pad accordingly; this is useful on TPUs ' + 'to minimize the number of compilations') + + # options for reporting BLEU during validation + parser.add_argument('--eval-bleu', action='store_true', + help='evaluation with BLEU scores') + parser.add_argument('--eval-bleu-detok', type=str, default="space", + help='detokenize before computing BLEU (e.g., "moses"); ' + 'required if using --eval-bleu; use "space" to ' + 'disable detokenization; see fairseq.data.encoders ' + 'for other options') + parser.add_argument('--eval-bleu-detok-args', type=str, metavar='JSON', + help='args for building the tokenizer, if needed') + parser.add_argument('--eval-tokenized-bleu', action='store_true', default=False, + help='compute tokenized BLEU instead of sacrebleu') + parser.add_argument('--eval-bleu-remove-bpe', nargs='?', const='@@ ', default=None, + help='remove BPE before computing BLEU') + parser.add_argument('--eval-bleu-args', type=str, metavar='JSON', + help='generation args for BLUE scoring, ' + 'e.g., \'{"beam": 4, "lenpen": 0.6}\'') + parser.add_argument('--eval-bleu-print-samples', action='store_true', + help='print sample generations during validation') + # fmt: on + + def __init__(self, args, src_dict, tgt_dict): + super().__init__(args) + self.src_dict = src_dict + self.tgt_dict = tgt_dict + + @classmethod + def setup_task(cls, args, **kwargs): + """Setup the task (e.g., load dictionaries). + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + args.left_pad_source = options.eval_bool(args.left_pad_source) + args.left_pad_target = options.eval_bool(args.left_pad_target) + + paths = utils.split_paths(args.data) + assert len(paths) > 0 + # find language pair automatically + if args.source_lang is None or args.target_lang is None: + args.source_lang, args.target_lang = data_utils.infer_language_pair(paths[0]) + if args.source_lang is None or args.target_lang is None: + raise Exception('Could not infer language pair, please provide it explicitly') + + # load dictionaries + src_dict = cls.load_dictionary(os.path.join(paths[0], 'dict.{}.txt'.format(args.source_lang))) + tgt_dict = cls.load_dictionary(os.path.join(paths[0], 'dict.{}.txt'.format(args.target_lang))) + assert src_dict.pad() == tgt_dict.pad() + assert src_dict.eos() == tgt_dict.eos() + assert src_dict.unk() == tgt_dict.unk() + logger.info('[{}] dictionary: {} types'.format(args.source_lang, len(src_dict))) + logger.info('[{}] dictionary: {} types'.format(args.target_lang, len(tgt_dict))) + + return cls(args, src_dict, tgt_dict) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + # infer langcode + src, tgt = self.args.source_lang, self.args.target_lang + + self.datasets[split] = load_langpair_dataset( + data_path, split, src, self.src_dict, tgt, self.tgt_dict, + combine=combine, dataset_impl=self.args.dataset_impl, + upsample_primary=self.args.upsample_primary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + max_source_positions=self.args.max_source_positions, + max_target_positions=self.args.max_target_positions, + load_alignments=self.args.load_alignments, + truncate_source=self.args.truncate_source, + num_buckets=self.args.num_batch_buckets, + shuffle=(split != 'test'), + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths): + return LanguagePairDataset(src_tokens, src_lengths, self.source_dictionary) + + def build_model(self, args): + model = super().build_model(args) + if getattr(args, 'eval_bleu', False): + assert getattr(args, 'eval_bleu_detok', None) is not None, ( + '--eval-bleu-detok is required if using --eval-bleu; ' + 'try --eval-bleu-detok=moses (or --eval-bleu-detok=space ' + 'to disable detokenization, e.g., when using sentencepiece)' + ) + detok_args = json.loads(getattr(args, 'eval_bleu_detok_args', '{}') or '{}') + self.tokenizer = encoders.build_tokenizer(Namespace( + tokenizer=getattr(args, 'eval_bleu_detok', None), + **detok_args + )) + + gen_args = json.loads(getattr(args, 'eval_bleu_args', '{}') or '{}') + self.sequence_generator = self.build_generator([model], Namespace(**gen_args)) + return model + + def valid_step(self, sample, model, criterion): + loss, sample_size, logging_output = super().valid_step(sample, model, criterion) + if self.args.eval_bleu: + bleu = self._inference_with_bleu(self.sequence_generator, sample, model) + logging_output['_bleu_sys_len'] = bleu.sys_len + logging_output['_bleu_ref_len'] = bleu.ref_len + # we split counts into separate entries so that they can be + # summed efficiently across workers using fast-stat-sync + assert len(bleu.counts) == EVAL_BLEU_ORDER + for i in range(EVAL_BLEU_ORDER): + logging_output['_bleu_counts_' + str(i)] = bleu.counts[i] + logging_output['_bleu_totals_' + str(i)] = bleu.totals[i] + return loss, sample_size, logging_output + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + if self.args.eval_bleu: + + def sum_logs(key): + return sum(log.get(key, 0) for log in logging_outputs) + + counts, totals = [], [] + for i in range(EVAL_BLEU_ORDER): + counts.append(sum_logs('_bleu_counts_' + str(i))) + totals.append(sum_logs('_bleu_totals_' + str(i))) + + if max(totals) > 0: + # log counts as numpy arrays -- log_scalar will sum them correctly + metrics.log_scalar('_bleu_counts', np.array(counts)) + metrics.log_scalar('_bleu_totals', np.array(totals)) + metrics.log_scalar('_bleu_sys_len', sum_logs('_bleu_sys_len')) + metrics.log_scalar('_bleu_ref_len', sum_logs('_bleu_ref_len')) + + def compute_bleu(meters): + import inspect + import sacrebleu + fn_sig = inspect.getfullargspec(sacrebleu.compute_bleu)[0] + if 'smooth_method' in fn_sig: + smooth = {'smooth_method': 'exp'} + else: + smooth = {'smooth': 'exp'} + bleu = sacrebleu.compute_bleu( + correct=meters['_bleu_counts'].sum, + total=meters['_bleu_totals'].sum, + sys_len=meters['_bleu_sys_len'].sum, + ref_len=meters['_bleu_ref_len'].sum, + **smooth + ) + return round(bleu.score, 2) + + metrics.log_derived('bleu', compute_bleu) + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + return (self.args.max_source_positions, self.args.max_target_positions) + + @property + def source_dictionary(self): + """Return the source :class:`~fairseq.data.Dictionary`.""" + return self.src_dict + + @property + def target_dictionary(self): + """Return the target :class:`~fairseq.data.Dictionary`.""" + return self.tgt_dict + + def _inference_with_bleu(self, generator, sample, model): + import sacrebleu + + def decode(toks, escape_unk=False): + s = self.tgt_dict.string( + toks.int().cpu(), + self.args.eval_bleu_remove_bpe, + # The default unknown string in fairseq is ``, but + # this is tokenized by sacrebleu as `< unk >`, inflating + # BLEU scores. Instead, we use a somewhat more verbose + # alternative that is unlikely to appear in the real + # reference, but doesn't get split into multiple tokens. + unk_string=( + "UNKNOWNTOKENINREF" if escape_unk else "UNKNOWNTOKENINHYP" + ), + ) + if self.tokenizer: + s = self.tokenizer.decode(s) + return s + + gen_out = self.inference_step(generator, [model], sample, None) + hyps, refs = [], [] + for i in range(len(gen_out)): + hyps.append(decode(gen_out[i][0]['tokens'])) + refs.append(decode( + utils.strip_pad(sample['target'][i], self.tgt_dict.pad()), + escape_unk=True, # don't count as matches to the hypo + )) + if self.args.eval_bleu_print_samples: + logger.info('example hypothesis: ' + hyps[0]) + logger.info('example reference: ' + refs[0]) + if self.args.eval_tokenized_bleu: + return sacrebleu.corpus_bleu(hyps, [refs], tokenize='none') + else: + return sacrebleu.corpus_bleu(hyps, [refs]) diff --git a/fairseq/tasks/translation_from_pretrained_bart.py b/fairseq/tasks/translation_from_pretrained_bart.py new file mode 100644 index 0000000000000000000000000000000000000000..2b7d589ceed8c4d5f33be97a2e05b73e6da24f5b --- /dev/null +++ b/fairseq/tasks/translation_from_pretrained_bart.py @@ -0,0 +1,119 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from fairseq.data import LanguagePairDataset +from fairseq import utils + +from .translation import load_langpair_dataset, TranslationTask +from . import register_task + + +@register_task('translation_from_pretrained_bart') +class TranslationFromPretrainedBARTTask(TranslationTask): + """ + Translate from source language to target language with a model initialized with a multilingual pretrain. + + Args: + src_dict (~fairseq.data.Dictionary): dictionary for the source language + tgt_dict (~fairseq.data.Dictionary): dictionary for the target language + + .. note:: + + The translation task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate` and :mod:`fairseq-interactive`. + + The translation task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.translation_parser + :prog: + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + TranslationTask.add_args(parser) + parser.add_argument('--langs', required=True, metavar='LANG', + help='comma-separated list of monolingual language, ' + 'for example, "en,de,fr". These should match the ' + 'langs from pretraining (and be in the same order). ' + 'You should always add all pretraining language idx ' + 'during finetuning.') + parser.add_argument('--prepend-bos', action='store_true', + help='prepend bos token to each sentence, which matches ' + 'mBART pretraining') + # fmt: on + + def __init__(self, args, src_dict, tgt_dict): + super().__init__(args, src_dict, tgt_dict) + self.langs = args.langs.split(',') + for d in [src_dict, tgt_dict]: + for l in self.langs: + d.add_symbol('[{}]'.format(l)) + d.add_symbol('') + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + # infer langcode + src, tgt = self.args.source_lang, self.args.target_lang + + self.datasets[split] = load_langpair_dataset( + data_path, split, src, self.src_dict, tgt, self.tgt_dict, + combine=combine, dataset_impl=self.args.dataset_impl, + upsample_primary=self.args.upsample_primary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + max_source_positions=getattr(self.args, 'max_source_positions', 1024), + max_target_positions=getattr(self.args, 'max_target_positions', 1024), + load_alignments=self.args.load_alignments, + prepend_bos=getattr(self.args, 'prepend_bos', False), + append_source_id=True + ) + + def build_generator(self, models, args): + if getattr(args, 'score_reference', False): + from fairseq.sequence_scorer import SequenceScorer + return SequenceScorer( + self.target_dictionary, + eos=self.tgt_dict.index('[{}]'.format(self.args.target_lang)) + ) + else: + from fairseq.sequence_generator import SequenceGenerator + return SequenceGenerator( + models, + self.target_dictionary, + beam_size=getattr(args, 'beam', 5), + max_len_a=getattr(args, 'max_len_a', 0), + max_len_b=getattr(args, 'max_len_b', 200), + min_len=getattr(args, 'min_len', 1), + normalize_scores=(not getattr(args, 'unnormalized', False)), + len_penalty=getattr(args, 'lenpen', 1), + unk_penalty=getattr(args, 'unkpen', 0), + temperature=getattr(args, 'temperature', 1.), + match_source_len=getattr(args, 'match_source_len', False), + no_repeat_ngram_size=getattr(args, 'no_repeat_ngram_size', 0), + eos=self.tgt_dict.index('[{}]'.format(self.args.target_lang)) + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths): + src_lang_id = self.source_dictionary.index('[{}]'.format(self.args.source_lang)) + source_tokens = [] + for s_t in src_tokens: + s_t = torch.cat([s_t, s_t.new(1).fill_(src_lang_id)]) + source_tokens.append(s_t) + dataset = LanguagePairDataset(source_tokens, src_lengths, self.source_dictionary) + return dataset diff --git a/fairseq/tasks/translation_from_pretrained_xlm.py b/fairseq/tasks/translation_from_pretrained_xlm.py new file mode 100644 index 0000000000000000000000000000000000000000..347a6eccb7657e6d20d1f1304b76fe31bc731393 --- /dev/null +++ b/fairseq/tasks/translation_from_pretrained_xlm.py @@ -0,0 +1,31 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data.legacy.masked_lm_dictionary import MaskedLMDictionary +from fairseq.tasks.translation import TranslationTask + +from . import register_task + + +@register_task("translation_from_pretrained_xlm") +class TranslationFromPretrainedXLMTask(TranslationTask): + """ + Same as TranslationTask except use the MaskedLMDictionary class so that + we can load data that was binarized with the MaskedLMDictionary class. + + This task should be used for the entire training pipeline when we want to + train an NMT model from a pretrained XLM checkpoint: binarizing NMT data, + training NMT with the pretrained XLM checkpoint, and subsequent evaluation + of that trained model. + """ + + @classmethod + def load_dictionary(cls, filename): + """Load the masked LM dictionary from the filename + + Args: + filename (str): the filename + """ + return MaskedLMDictionary.load(filename) diff --git a/fairseq/tasks/translation_lev.py b/fairseq/tasks/translation_lev.py new file mode 100644 index 0000000000000000000000000000000000000000..845dd8164456fbd76c8477205cba496dad6f3332 --- /dev/null +++ b/fairseq/tasks/translation_lev.py @@ -0,0 +1,169 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import os + +import torch + +from fairseq.data import LanguagePairDataset + +from fairseq.utils import new_arange +from fairseq.tasks import register_task +from fairseq.tasks.translation import TranslationTask, load_langpair_dataset +from fairseq import utils + +@register_task('translation_lev') +class TranslationLevenshteinTask(TranslationTask): + """ + Translation (Sequence Generation) task for Levenshtein Transformer + See `"Levenshtein Transformer" `_. + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + TranslationTask.add_args(parser) + parser.add_argument( + '--noise', + default='random_delete', + choices=['random_delete', 'random_mask', 'no_noise', 'full_mask']) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + paths = utils.split_paths(self.args.data) + assert len(paths) > 0 + data_path = paths[(epoch - 1) % len(paths)] + + # infer langcode + src, tgt = self.args.source_lang, self.args.target_lang + + self.datasets[split] = load_langpair_dataset( + data_path, split, src, self.src_dict, tgt, self.tgt_dict, + combine=combine, dataset_impl=self.args.dataset_impl, + upsample_primary=self.args.upsample_primary, + left_pad_source=self.args.left_pad_source, + left_pad_target=self.args.left_pad_target, + max_source_positions=self.args.max_source_positions, + max_target_positions=self.args.max_target_positions, + prepend_bos=True, + ) + + def inject_noise(self, target_tokens): + def _random_delete(target_tokens): + pad = self.tgt_dict.pad() + bos = self.tgt_dict.bos() + eos = self.tgt_dict.eos() + + max_len = target_tokens.size(1) + target_mask = target_tokens.eq(pad) + target_score = target_tokens.clone().float().uniform_() + target_score.masked_fill_( + target_tokens.eq(bos) | target_tokens.eq(eos), 0.0) + target_score.masked_fill_(target_mask, 1) + target_score, target_rank = target_score.sort(1) + target_length = target_mask.size(1) - target_mask.float().sum( + 1, keepdim=True) + + # do not delete and (we assign 0 score for them) + target_cutoff = 2 + ((target_length - 2) * target_score.new_zeros( + target_score.size(0), 1).uniform_()).long() + target_cutoff = target_score.sort(1)[1] >= target_cutoff + + prev_target_tokens = target_tokens.gather( + 1, target_rank).masked_fill_(target_cutoff, pad).gather( + 1, + target_rank.masked_fill_(target_cutoff, + max_len).sort(1)[1]) + prev_target_tokens = prev_target_tokens[:, :prev_target_tokens. + ne(pad).sum(1).max()] + + return prev_target_tokens + + def _random_mask(target_tokens): + pad = self.tgt_dict.pad() + bos = self.tgt_dict.bos() + eos = self.tgt_dict.eos() + unk = self.tgt_dict.unk() + + target_masks = target_tokens.ne(pad) & \ + target_tokens.ne(bos) & \ + target_tokens.ne(eos) + target_score = target_tokens.clone().float().uniform_() + target_score.masked_fill_(~target_masks, 2.0) + target_length = target_masks.sum(1).float() + target_length = target_length * target_length.clone().uniform_() + target_length = target_length + 1 # make sure to mask at least one token. + + _, target_rank = target_score.sort(1) + target_cutoff = new_arange(target_rank) < target_length[:, None].long() + prev_target_tokens = target_tokens.masked_fill( + target_cutoff.scatter(1, target_rank, target_cutoff), unk) + return prev_target_tokens + + def _full_mask(target_tokens): + pad = self.tgt_dict.pad() + bos = self.tgt_dict.bos() + eos = self.tgt_dict.eos() + unk = self.tgt_dict.unk() + + target_mask = target_tokens.eq(bos) | target_tokens.eq( + eos) | target_tokens.eq(pad) + return target_tokens.masked_fill(~target_mask, unk) + + if self.args.noise == 'random_delete': + return _random_delete(target_tokens) + elif self.args.noise == 'random_mask': + return _random_mask(target_tokens) + elif self.args.noise == 'full_mask': + return _full_mask(target_tokens) + elif self.args.noise == 'no_noise': + return target_tokens + else: + raise NotImplementedError + + def build_generator(self, models, args): + # add models input to match the API for SequenceGenerator + from fairseq.iterative_refinement_generator import IterativeRefinementGenerator + return IterativeRefinementGenerator( + self.target_dictionary, + eos_penalty=getattr(args, 'iter_decode_eos_penalty', 0.0), + max_iter=getattr(args, 'iter_decode_max_iter', 10), + beam_size=getattr(args, 'iter_decode_with_beam', 1), + reranking=getattr(args, 'iter_decode_with_external_reranker', False), + decoding_format=getattr(args, 'decoding_format', None), + adaptive=not getattr(args, 'iter_decode_force_max_iter', False), + retain_history=getattr(args, 'retain_iter_history', False)) + + def build_dataset_for_inference(self, src_tokens, src_lengths): + return LanguagePairDataset( + src_tokens, src_lengths, self.source_dictionary, append_bos=True + ) + + def train_step(self, + sample, + model, + criterion, + optimizer, + update_num, + ignore_grad=False): + model.train() + sample['prev_target'] = self.inject_noise(sample['target']) + loss, sample_size, logging_output = criterion(model, sample) + if ignore_grad: + loss *= 0 + optimizer.backward(loss) + return loss, sample_size, logging_output + + def valid_step(self, sample, model, criterion): + model.eval() + with torch.no_grad(): + sample['prev_target'] = self.inject_noise(sample['target']) + loss, sample_size, logging_output = criterion(model, sample) + return loss, sample_size, logging_output diff --git a/fairseq/tasks/translation_multi_simple_epoch.py b/fairseq/tasks/translation_multi_simple_epoch.py new file mode 100644 index 0000000000000000000000000000000000000000..eba32f17599f195476857edc93d2dd9557497415 --- /dev/null +++ b/fairseq/tasks/translation_multi_simple_epoch.py @@ -0,0 +1,322 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import datetime +import time + +import torch +from fairseq.data import ( + data_utils, + FairseqDataset, + iterators, + LanguagePairDataset, + ListDataset, +) + +from fairseq.tasks import FairseqTask, register_task +from fairseq.data.multilingual.sampling_method import SamplingMethod +from fairseq.data.multilingual.multilingual_data_manager import MultilingualDatasetManager + + +### +def get_time_gap(s, e): + return (datetime.datetime.fromtimestamp(e) - datetime.datetime.fromtimestamp(s)).__str__() +### + + +logger = logging.getLogger(__name__) + + +@register_task('translation_multi_simple_epoch') +class TranslationMultiSimpleEpochTask(FairseqTask): + """ + Translate from one (source) language to another (target) language. + + Args: + langs (List[str]): a list of languages that are being supported + dicts (Dict[str, fairseq.data.Dictionary]): mapping from supported languages to their dictionaries + training (bool): whether the task should be configured for training or not + + .. note:: + + The translation task is compatible with :mod:`fairseq-train`, + :mod:`fairseq-generate` and :mod:`fairseq-interactive`. + + The translation task provides the following additional command-line + arguments: + + .. argparse:: + :ref: fairseq.tasks.translation_parser + :prog: + """ + + @staticmethod + def add_args(parser): + """Add task-specific arguments to the parser.""" + # fmt: off + parser.add_argument('-s', '--source-lang', default=None, metavar='SRC', + help='inference source language') + parser.add_argument('-t', '--target-lang', default=None, metavar='TARGET', + help='inference target language') + parser.add_argument('--lang-pairs', default=None, metavar='PAIRS', + help='comma-separated list of language pairs (in training order): en-de,en-fr,de-fr') + parser.add_argument('--keep-inference-langtok', action='store_true', + help='keep language tokens in inference output (e.g. for analysis or debugging)') + + SamplingMethod.add_arguments(parser) + MultilingualDatasetManager.add_args(parser) + # fmt: on + + def __init__(self, args, langs, dicts, training): + super().__init__(args) + self.langs = langs + self.dicts = dicts + self.training = training + if training: + self.lang_pairs = args.lang_pairs + else: + self.lang_pairs = ['{}-{}'.format(args.source_lang, args.target_lang)] + # eval_lang_pairs for multilingual translation is usually all of the + # lang_pairs. However for other multitask settings or when we want to + # optimize for certain languages we want to use a different subset. Thus + # the eval_lang_pairs class variable is provided for classes that extend + # this class. + self.eval_lang_pairs = self.lang_pairs + # model_lang_pairs will be used to build encoder-decoder model pairs in + # models.build_model(). This allows multitask type of sub-class can + # build models other than the input lang_pairs + self.model_lang_pairs = self.lang_pairs + self.sampling_method = SamplingMethod.build_sampler(args, self) + self.data_manager = MultilingualDatasetManager.setup_data_manager( + args, self.lang_pairs, langs, dicts, self.sampling_method) + + @classmethod + def setup_task(cls, args, **kwargs): + langs, dicts, training = MultilingualDatasetManager.prepare( + cls.load_dictionary, args, **kwargs + ) + return cls(args, langs, dicts, training) + + def has_sharded_data(self, split): + return self.data_manager.has_sharded_data(split) + + def load_dataset(self, split, epoch=1, combine=False, **kwargs): + """Load a given dataset split. + + Args: + split (str): name of the split (e.g., train, valid, test) + """ + if split in self.datasets: + dataset = self.datasets[split] + if self.has_sharded_data(split) and dataset.load_next_shard: + shard_epoch = dataset.shard_epoch + else: + # no need to load next shard so skip loading + # also this avoid always loading from beginning of the data + return + else: + shard_epoch = None + logger.info(f'loading data for {split} epoch={epoch}/{shard_epoch}') + self.datasets[split] = self.data_manager.load_sampled_multi_epoch_dataset( + split, + self.training, + epoch=epoch, combine=combine, shard_epoch=shard_epoch, **kwargs + ) + + def build_dataset_for_inference(self, src_tokens, src_lengths): + src_data = ListDataset(src_tokens, src_lengths) + dataset = LanguagePairDataset(src_data, src_lengths, self.source_dictionary) + src_langtok_spec, tgt_langtok_spec = self.args.langtoks['main'] + if self.args.lang_tok_replacing_bos_eos: + dataset = self.data_manager.alter_dataset_langtok( + dataset, + src_eos=self.source_dictionary.eos(), + src_lang=self.args.source_lang, + tgt_eos=self.target_dictionary.eos(), + tgt_lang=self.args.target_lang, + src_langtok_spec=src_langtok_spec, + tgt_langtok_spec=tgt_langtok_spec, + ) + else: + dataset.src = self.data_manager.src_dataset_tranform_func( + self.args.source_lang, + self.args.target_lang, + dataset=dataset.src, + spec=src_langtok_spec, + ) + return dataset + + def build_generator( + self, models, args, + seq_gen_cls=None, extra_gen_cls_kwargs=None, + ): + if not getattr(args, 'keep_inference_langtok', False): + _, tgt_langtok_spec = self.args.langtoks['main'] + if tgt_langtok_spec: + tgt_lang_tok = self.data_manager.get_decoder_langtok(self.args.target_lang, tgt_langtok_spec) + extra_gen_cls_kwargs = extra_gen_cls_kwargs or {} + extra_gen_cls_kwargs['symbols_to_strip_from_output'] = {tgt_lang_tok} + + return super().build_generator( + models, args, + seq_gen_cls=None, + extra_gen_cls_kwargs=extra_gen_cls_kwargs + ) + + def build_model(self, args): + return super().build_model(args) + + def valid_step(self, sample, model, criterion): + loss, sample_size, logging_output = super().valid_step(sample, model, criterion) + return loss, sample_size, logging_output + + def inference_step(self, generator, models, sample, prefix_tokens=None): + with torch.no_grad(): + _, tgt_langtok_spec = self.args.langtoks['main'] + if not self.args.lang_tok_replacing_bos_eos: + if prefix_tokens is None and tgt_langtok_spec: + tgt_lang_tok = self.data_manager.get_decoder_langtok(self.args.target_lang, tgt_langtok_spec) + src_tokens = sample['net_input']['src_tokens'] + bsz = src_tokens.size(0) + prefix_tokens = torch.LongTensor( + [[tgt_lang_tok]] + ).expand(bsz, 1).to(src_tokens) + return generator.generate( + models, + sample, + prefix_tokens=prefix_tokens, + ) + else: + return generator.generate( + models, + sample, + prefix_tokens=prefix_tokens, + bos_token=self.data_manager.get_decoder_langtok(self.args.target_lang, tgt_langtok_spec) + if tgt_langtok_spec else self.target_dictionary.eos(), + ) + + def reduce_metrics(self, logging_outputs, criterion): + super().reduce_metrics(logging_outputs, criterion) + + def max_positions(self): + """Return the max sentence length allowed by the task.""" + return (self.args.max_source_positions, self.args.max_target_positions) + + @property + def source_dictionary(self): + if self.training: + return next(iter(self.dicts.values())) + else: + return self.dicts[self.args.source_lang] + + @property + def target_dictionary(self): + if self.training: + return next(iter(self.dicts.values())) + else: + return self.dicts[self.args.target_lang] + + def create_batch_sampler_func( + self, max_positions, ignore_invalid_inputs, + max_tokens, max_sentences + ): + def construct_batch_sampler( + dataset, epoch + ): + splits = [s for s, _ in self.datasets.items() if self.datasets[s] == dataset] + split = splits[0] if len(splits) > 0 else None + + if epoch is not None: + dataset.set_epoch(epoch) + start_time = time.time() + # get indices ordered by example size + indices = dataset.ordered_indices() + logger.debug(f'[{split}] @batch_sampler order indices time: {get_time_gap(start_time, time.time())}') + + # filter examples that are too large + if max_positions is not None: + my_time = time.time() + indices = data_utils.filter_by_size( + indices, dataset, max_positions, raise_exception=(not ignore_invalid_inputs), + ) + logger.debug(f'[{split}] @batch_sampler filter_by_size time: {get_time_gap(my_time, time.time())}') + + # create mini-batches with given size constraints + my_time = time.time() + batch_sampler = data_utils.batch_by_size( + indices, dataset.num_tokens, max_tokens=max_tokens, max_sentences=max_sentences, + ) + logger.debug(f'[{split}] @batch_sampler batch_by_size time: {get_time_gap(my_time, time.time())}') + logger.debug(f'[{split}] per epoch batch_sampler set-up time: {get_time_gap(start_time, time.time())}') + return batch_sampler + return construct_batch_sampler + + # we need to override get_batch_iterator because we want to reset the epoch iterator each time + def get_batch_iterator( + self, dataset, max_tokens=None, max_sentences=None, max_positions=None, + ignore_invalid_inputs=False, required_batch_size_multiple=1, + seed=1, num_shards=1, shard_id=0, num_workers=0, epoch=1, + ): + """ + Get an iterator that yields batches of data from the given dataset. + + Args: + dataset (~fairseq.data.FairseqDataset): dataset to batch + max_tokens (int, optional): max number of tokens in each batch + (default: None). + max_sentences (int, optional): max number of sentences in each + batch (default: None). + max_positions (optional): max sentence length supported by the + model (default: None). + ignore_invalid_inputs (bool, optional): don't raise Exception for + sentences that are too long (default: False). + required_batch_size_multiple (int, optional): require batch size to + be a multiple of N (default: 1). + seed (int, optional): seed for random number generator for + reproducibility (default: 1). + num_shards (int, optional): shard the data iterator into N + shards (default: 1). + shard_id (int, optional): which shard of the data iterator to + return (default: 0). + num_workers (int, optional): how many subprocesses to use for data + loading. 0 means the data will be loaded in the main process + (default: 0). + epoch (int, optional): the epoch to start the iterator from + (default: 0). + Returns: + ~fairseq.iterators.EpochBatchIterator: a batched iterator over the + given dataset split + """ + # initialize the dataset with the correct starting epoch + assert isinstance(dataset, FairseqDataset) + if dataset in self.dataset_to_epoch_iter: + return self.dataset_to_epoch_iter[dataset] + if ( + self.args.sampling_method == 'RoundRobin' + ): + batch_iter = super().get_batch_iterator( + dataset, max_tokens=max_tokens, max_sentences=max_sentences, max_positions=max_positions, + ignore_invalid_inputs=ignore_invalid_inputs, required_batch_size_multiple=required_batch_size_multiple, + seed=seed, num_shards=num_shards, shard_id=shard_id, num_workers=num_workers, epoch=epoch, + ) + self.dataset_to_epoch_iter[dataset] = batch_iter + return batch_iter + + construct_batch_sampler = self.create_batch_sampler_func( + max_positions, ignore_invalid_inputs, + max_tokens, max_sentences) + + epoch_iter = iterators.EpochBatchIterator( + dataset=dataset, + collate_fn=dataset.collater, + batch_sampler=construct_batch_sampler, + seed=seed, + num_shards=num_shards, + shard_id=shard_id, + num_workers=num_workers, + epoch=epoch, + ) + return epoch_iter diff --git a/fairseq/tokenizer.py b/fairseq/tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..8c4d694aa08be0e73179290785c565419247618b --- /dev/null +++ b/fairseq/tokenizer.py @@ -0,0 +1,14 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + +SPACE_NORMALIZER = re.compile(r"\s+") + + +def tokenize_line(line): + line = SPACE_NORMALIZER.sub(" ", line) + line = line.strip() + return line.split() diff --git a/fairseq/trainer.py b/fairseq/trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..898edb6d6cfd8a8f4bd2e612b96009a1bf232e7d --- /dev/null +++ b/fairseq/trainer.py @@ -0,0 +1,1006 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Train a network across multiple GPUs. +""" + +import contextlib +from itertools import chain +import logging +import sys +import time +from typing import Any, Dict, List + +import torch + +from fairseq import checkpoint_utils, distributed_utils, models, optim, utils +from fairseq.file_io import PathManager +from fairseq.logging import meters, metrics +from fairseq.nan_detector import NanDetector +from fairseq.optim import lr_scheduler + + +logger = logging.getLogger(__name__) + + +class Trainer(object): + """Main class for data parallel training. + + This class supports synchronous distributed data parallel training, + where multiple workers each have a full model replica and gradients + are accumulated across workers before each update. We use + :class:`~torch.nn.parallel.DistributedDataParallel` to handle + communication of the gradients across workers. + """ + + def __init__(self, args, task, model, criterion, quantizer=None): + self.args = args + self.task = task + + # catalog shared parameters + shared_params = _catalog_shared_params(model) + + self.tpu = getattr(args, 'tpu', False) + self.cuda = torch.cuda.is_available() and not args.cpu and not self.tpu + if self.cuda: + self.device = torch.device('cuda') + elif self.tpu: + self.device = utils.get_tpu_device(args) + else: + self.device = torch.device('cpu') + + # copy model and criterion to current device/dtype + self._criterion = criterion + self._model = model + if self.tpu: + import torch_xla.core.xla_model as xm + self._model = xm.send_cpu_data_to_device(self._model, self.device) + if args.fp16: + self._criterion = self._criterion.half() + self._model = self._model.half() + elif args.bf16: + self._criterion = self._criterion.to(dtype=torch.bfloat16) + self._model = self._model.to(dtype=torch.bfloat16) + self._criterion = self._criterion.to(device=self.device) + self._model = self._model.to(device=self.device) + + # check that shared parameters are preserved after device transfer + for shared_param in shared_params: + ref = _get_module_by_path(self._model, shared_param[0]) + for path in shared_param[1:]: + logger.info( + 'detected shared parameter: {} <- {}'.format(shared_param[0], path) + ) + _set_module_by_path(self._model, path, ref) + + self._dummy_batch = "DUMMY" # indicates we don't have a dummy batch at first + self._lr_scheduler = None + self._num_updates = 0 + self._num_xla_compiles = 0 # for TPUs + self._optim_history = None + self._optimizer = None + self._warn_once = set() + self._wrapped_criterion = None + self._wrapped_model = None + + # TODO(myleott): support tpu + if self.cuda and self.data_parallel_world_size > 1: + self._grad_norm_buf = torch.cuda.DoubleTensor(self.data_parallel_world_size) + else: + self._grad_norm_buf = None + + self.quantizer = quantizer + if self.quantizer is not None: + self.quantizer.set_trainer(self) + + # get detailed cuda environment + if self.cuda: + self.cuda_env = utils.CudaEnvironment() + if self.data_parallel_world_size > 1: + self.cuda_env_arr = distributed_utils.all_gather_list(self.cuda_env) + else: + self.cuda_env_arr = [self.cuda_env] + if self.data_parallel_rank == 0: + utils.CudaEnvironment.pretty_print_cuda_env_list(self.cuda_env_arr) + else: + self.cuda_env = None + self.cuda_env_arr = None + + metrics.log_start_time("wall", priority=790, round=0) + + self._start_time = time.time() + self._previous_training_time = 0 + self._cumulative_training_time = None + + def reinitialize(self): + """Reinitialize the Trainer, typically after model params change.""" + self._lr_scheduler = None + self._optimizer = None + self._wrapped_criterion = None + self._wrapped_model = None + + @property + def data_parallel_world_size(self): + return self.args.distributed_world_size + + @property + def data_parallel_process_group(self): + if self.tpu: + return ('tpu', None) + else: + return None + + @property + def data_parallel_rank(self): + return self.args.distributed_rank + + @property + def is_data_parallel_master(self): + return distributed_utils.is_master(self.args) + + @property + def criterion(self): + if self._wrapped_criterion is None: + if ( + utils.has_parameters(self._criterion) + and self.data_parallel_world_size > 1 + and not self.args.use_bmuf + and not self.tpu + ): + self._wrapped_criterion = models.DistributedFairseqModel( + self.args, self._criterion, + process_group=self.data_parallel_process_group + ) + else: + self._wrapped_criterion = self._criterion + return self._wrapped_criterion + + @property + def model(self): + if self._wrapped_model is None: + if ( + self.data_parallel_world_size > 1 + and not self.args.use_bmuf + and not self.tpu + ): + self._wrapped_model = models.DistributedFairseqModel( + self.args, self._model, + process_group=self.data_parallel_process_group + ) + else: + self._wrapped_model = self._model + return self._wrapped_model + + @property + def optimizer(self): + if self._optimizer is None: + self._build_optimizer() + return self._optimizer + + @property + def lr_scheduler(self): + if self._lr_scheduler is None: + self._build_optimizer() # this will initialize self._lr_scheduler + return self._lr_scheduler + + def _build_optimizer(self): + params = list( + filter( + lambda p: p.requires_grad, + chain(self.model.parameters(), self.criterion.parameters()), + ) + ) + + if self.args.fp16 or self.args.bf16: + if self.cuda and torch.cuda.get_device_capability(0)[0] < 7: + logger.info( + "NOTE: your device does NOT support faster training with --fp16, " + "please switch to FP32 which is likely to be faster" + ) + if self.args.memory_efficient_fp16 or self.args.memory_efficient_bf16: + self._optimizer = optim.MemoryEfficientFP16Optimizer.build_optimizer( + self.args, params + ) + else: + self._optimizer = optim.FP16Optimizer.build_optimizer(self.args, params) + else: + if self.cuda and torch.cuda.get_device_capability(0)[0] >= 7: + logger.info("NOTE: your device may support faster training with --fp16") + self._optimizer = optim.build_optimizer(self.args, params) + + if self.args.use_bmuf: + self._optimizer = optim.FairseqBMUF(self.args, self._optimizer) + + # We should initialize the learning rate scheduler immediately after + # building the optimizer, so that the initial learning rate is set. + self._lr_scheduler = lr_scheduler.build_lr_scheduler(self.args, self.optimizer) + self._lr_scheduler.step_update(0) + + def save_checkpoint(self, filename, extra_state): + """Save all training state in a checkpoint file.""" + if self.is_data_parallel_master: # only save one checkpoint + extra_state["metrics"] = metrics.state_dict() + extra_state["previous_training_time"] = self.cumulative_training_time() + checkpoint_utils.save_state( + filename, + self.args, + self.get_model().state_dict(), + self.get_criterion(), + self.optimizer, + self.lr_scheduler, + self.get_num_updates(), + self._optim_history, + extra_state, + ) + + def load_checkpoint( + self, + filename, + reset_optimizer=False, + reset_lr_scheduler=False, + optimizer_overrides=None, + reset_meters=False, + ): + """Load all training state from a checkpoint file.""" + extra_state, self._optim_history, last_optim_state = None, [], None + + bexists = PathManager.isfile(filename) + if bexists: + state = checkpoint_utils.load_checkpoint_to_cpu(filename) + + # load model parameters + try: + self.get_model().load_state_dict( + state["model"], strict=True, args=self.args + ) + if utils.has_parameters(self.get_criterion()): + self.get_criterion().load_state_dict( + state["criterion"], strict=True + ) + except Exception: + raise Exception( + "Cannot load model parameters from checkpoint {}; " + "please ensure that the architectures match.".format(filename) + ) + + extra_state = state["extra_state"] + self._optim_history = state["optimizer_history"] + last_optim_state = state.get("last_optimizer_state", None) + + if last_optim_state is not None and not reset_optimizer: + # rebuild optimizer after loading model, since params may have changed + self._build_optimizer() + + # only reload optimizer and lr_scheduler if they match + last_optim = self._optim_history[-1] + assert ( + last_optim["criterion_name"] == self.get_criterion().__class__.__name__ + ), "Criterion does not match; please reset the optimizer (--reset-optimizer)." + assert ( + last_optim["optimizer_name"] == self.optimizer.__class__.__name__ + ), "Optimizer does not match; please reset the optimizer (--reset-optimizer)." + + if not reset_lr_scheduler: + self.lr_scheduler.load_state_dict(last_optim["lr_scheduler_state"]) + self.optimizer.load_state_dict(last_optim_state, optimizer_overrides) + + self.set_num_updates(last_optim["num_updates"]) + + if extra_state is not None: + epoch = extra_state["train_iterator"]["epoch"] + logger.info( + "loaded checkpoint {} (epoch {} @ {} updates)".format( + filename, epoch, self.get_num_updates() + ) + ) + + if "previous_training_time" in extra_state: + self._previous_training_time = extra_state["previous_training_time"] + self._start_time = time.time() + + self.lr_step(epoch) + + if "metrics" in extra_state and not reset_meters: + metrics.load_state_dict(extra_state["metrics"]) + + # reset TimeMeters, since their start times don't make sense anymore + for meter in metrics.get_meters("default"): + if isinstance(meter, meters.TimeMeter): + meter.reset() + else: + logger.info("no existing checkpoint found {}".format(filename)) + + return extra_state + + def get_train_iterator( + self, + epoch, + combine=True, + load_dataset=True, + data_selector=None, + shard_batch_itr=True, + ): + """Return an EpochBatchIterator over the training set for a given epoch.""" + if load_dataset: + logger.info("loading train data for epoch {}".format(epoch)) + self.task.load_dataset( + self.args.train_subset, + epoch=epoch, + combine=combine, + data_selector=data_selector, + ) + return self.task.get_batch_iterator( + dataset=self.task.dataset(self.args.train_subset), + max_tokens=self.args.max_tokens, + max_sentences=self.args.max_sentences, + max_positions=utils.resolve_max_positions( + self.task.max_positions(), + self.model.max_positions(), + self.args.max_tokens, + ), + ignore_invalid_inputs=True, + required_batch_size_multiple=self.args.required_batch_size_multiple, + seed=self.args.seed, + num_shards=self.data_parallel_world_size if shard_batch_itr else 1, + shard_id=self.data_parallel_rank if shard_batch_itr else 0, + num_workers=self.args.num_workers, + epoch=epoch + ) + + def get_valid_iterator( + self, + subset, + ): + """Return an EpochBatchIterator over given validation subset for a given epoch.""" + return self.task.get_batch_iterator( + dataset=self.task.dataset(subset), + max_tokens=self.args.max_tokens_valid, + max_sentences=self.args.max_sentences_valid, + max_positions=utils.resolve_max_positions( + self.task.max_positions(), + self.model.max_positions(), + ), + ignore_invalid_inputs=self.args.skip_invalid_size_inputs_valid_test, + required_batch_size_multiple=self.args.required_batch_size_multiple, + seed=self.args.seed, + num_shards=self.data_parallel_world_size, + shard_id=self.data_parallel_rank, + num_workers=self.args.num_workers + ) + + def begin_epoch(self, epoch): + """Called at the beginning of each epoch.""" + if self.quantizer is not None: + self.quantizer.begin_epoch(epoch) + + # task specific setup per epoch + self.task.begin_epoch(epoch, self.get_model()) + + @metrics.aggregate("train") + def train_step(self, samples, raise_oom=False): + """Do forward, backward and parameter update.""" + if self._dummy_batch == "DUMMY": + self._dummy_batch = samples[0] + + self._set_seed() + self.model.train() + self.criterion.train() + self.zero_grad() + + metrics.log_start_time("train_wall", priority=800, round=0) + + # forward and backward pass + logging_outputs, sample_size, ooms = [], 0, 0 + for i, sample in enumerate(samples): + sample = self._prepare_sample(sample) + if sample is None: + # when sample is None, run forward/backward on a dummy batch + # and ignore the resulting gradients + sample = self._prepare_sample(self._dummy_batch) + is_dummy_batch = True + else: + is_dummy_batch = False + + def maybe_no_sync(): + """ + Whenever *samples* contains more than one mini-batch, we + want to accumulate gradients locally and only call + all-reduce in the last backwards pass. + """ + if ( + self.data_parallel_world_size > 1 + and hasattr(self.model, "no_sync") + and i < len(samples) - 1 + ): + return self.model.no_sync() + else: + return contextlib.ExitStack() # dummy contextmanager + + try: + with maybe_no_sync(): + # forward and backward + loss, sample_size_i, logging_output = self.task.train_step( + sample=sample, + model=self.model, + criterion=self.criterion, + optimizer=self.optimizer, + update_num=self.get_num_updates(), + ignore_grad=is_dummy_batch, + ) + del loss + + logging_outputs.append(logging_output) + sample_size += sample_size_i + + # emptying the CUDA cache after the first step can + # reduce the chance of OOM + if self.cuda and self.get_num_updates() == 0: + torch.cuda.empty_cache() + except RuntimeError as e: + if "out of memory" in str(e): + self._log_oom(e) + if raise_oom: + raise e + logger.warning( + "attempting to recover from OOM in forward/backward pass" + ) + ooms += 1 + self.zero_grad() + if self.cuda: + torch.cuda.empty_cache() + if self.args.distributed_world_size == 1: + return None + else: + raise e + + if self.tpu and i < len(samples) - 1: + # tpu-comment: every XLA operation before marking step is + # appended to the IR graph, and processing too many batches + # before marking step can lead to OOM errors. + # To handle gradient accumulation use case, we explicitly + # mark step here for every forward pass without a backward pass + import torch_xla.core.xla_model as xm + xm.mark_step() + + if is_dummy_batch: + if torch.is_tensor(sample_size): + sample_size.zero_() + else: + sample_size *= 0. + + if torch.is_tensor(sample_size): + sample_size = sample_size.float() + else: + sample_size = float(sample_size) + + # gather logging outputs from all replicas + if self._sync_stats(): + train_time = self._local_cumulative_training_time() + logging_outputs, (sample_size, ooms, total_train_time) = self._aggregate_logging_outputs( + logging_outputs, sample_size, ooms, train_time, ignore=is_dummy_batch, + ) + self._cumulative_training_time = total_train_time / self.data_parallel_world_size + + if hasattr(self.model, 'all_reduce'): + self.model.all_reduce() + + overflow = False + try: + if self.tpu and self.data_parallel_world_size > 1: + import torch_xla.core.xla_model as xm + gradients = xm._fetch_gradients(self.optimizer.optimizer) + xm.all_reduce('sum', gradients, scale=1.0 / self.data_parallel_world_size) + + with torch.autograd.profiler.record_function("multiply-grads"): + # multiply gradients by (# GPUs / sample_size) since DDP + # already normalizes by the number of GPUs. Thus we get + # (sum_of_gradients / sample_size). + if not self.args.use_bmuf: + self.optimizer.multiply_grads(self.data_parallel_world_size / sample_size) + elif sample_size > 0: # BMUF needs to check sample size + num = self.data_parallel_world_size if self._sync_stats() else 1 + self.optimizer.multiply_grads(num / sample_size) + + with torch.autograd.profiler.record_function("clip-grads"): + # clip grads + grad_norm = self.clip_grad_norm(self.args.clip_norm) + + # check that grad norms are consistent across workers + if ( + not self.args.use_bmuf + and self.args.distributed_wrapper != 'SlowMo' + and not self.tpu + ): + self._check_grad_norms(grad_norm) + + with torch.autograd.profiler.record_function("optimizer"): + # take an optimization step + self.optimizer.step() + except FloatingPointError: + # re-run the forward and backward pass with hooks attached to print + # out where it fails + with NanDetector(self.model): + self.task.train_step( + sample, self.model, self.criterion, self.optimizer, self.get_num_updates(), + ignore_grad=False + ) + raise + except OverflowError as e: + overflow = True + logger.info("NOTE: overflow detected, " + str(e)) + grad_norm = torch.tensor(0.).cuda() + self.zero_grad() + except RuntimeError as e: + if "out of memory" in str(e): + self._log_oom(e) + logger.error("OOM during optimization, irrecoverable") + raise e + + # Some distributed wrappers (e.g., SlowMo) need access to the optimizer after the step + if hasattr(self.model, 'perform_additional_optimizer_actions'): + if hasattr(self.optimizer, 'fp32_params'): + self.model.perform_additional_optimizer_actions(self.optimizer.optimizer, self.optimizer.fp32_params) + else: + self.model.perform_additional_optimizer_actions(self.optimizer.optimizer) + + if not overflow or self.args.distributed_wrapper == 'SlowMo': + self.set_num_updates(self.get_num_updates() + 1) + + if self.tpu: + # mark step on TPUs + import torch_xla.core.xla_model as xm + xm.mark_step() + + # only log stats every log_interval steps + # this causes wps to be misreported when log_interval > 1 + logging_output = {} + if self.get_num_updates() % self.args.log_interval == 0: + logging_output = self._reduce_and_log_stats( + logging_outputs, sample_size, grad_norm, + ) + + # log whenever there's an XLA compilation, since these + # slow down training and may indicate opportunities for + # optimization + self._check_xla_compilation() + else: + # log stats + logging_output = self._reduce_and_log_stats( + logging_outputs, sample_size, grad_norm, + ) + + # clear CUDA cache to reduce memory fragmentation + if ( + self.cuda + and self.args.empty_cache_freq > 0 + and ( + (self.get_num_updates() + self.args.empty_cache_freq - 1) + % self.args.empty_cache_freq + ) == 0 + ): + torch.cuda.empty_cache() + + if self.args.fp16: + metrics.log_scalar("loss_scale", self.optimizer.scaler.loss_scale, priority=700, round=0) + + metrics.log_stop_time("train_wall") + + return logging_output + + @metrics.aggregate("valid") + def valid_step(self, sample, raise_oom=False): + """Do forward pass in evaluation mode.""" + if self._dummy_batch == "DUMMY": + self._dummy_batch = sample + if self.tpu: + import torch_xla.core.xla_model as xm + xm.rendezvous('valid_step') # wait for all workers + xm.mark_step() + + with torch.no_grad(): + self.model.eval() + self.criterion.eval() + + sample = self._prepare_sample(sample) + if sample is None: + sample = self._prepare_sample(self._dummy_batch) + is_dummy_batch = True + else: + is_dummy_batch = False + + try: + _loss, sample_size, logging_output = self.task.valid_step( + sample, self.model, self.criterion + ) + except RuntimeError as e: + if "out of memory" in str(e): + self._log_oom(e) + if not raise_oom: + logger.warning( + "ran out of memory in validation step, retrying batch" + ) + for p in self.model.parameters(): + if p.grad is not None: + p.grad = None # free some memory + if self.cuda: + torch.cuda.empty_cache() + return self.valid_step(sample, raise_oom=True) + raise e + + logging_outputs = [logging_output] + if is_dummy_batch: + if torch.is_tensor(sample_size): + sample_size.zero_() + else: + sample_size *= 0. + + # gather logging outputs from all replicas + if self.data_parallel_world_size > 1: + logging_outputs, (sample_size, ) = self._aggregate_logging_outputs( + logging_outputs, sample_size, ignore=is_dummy_batch, + ) + + # log validation stats + logging_output = self._reduce_and_log_stats(logging_outputs, sample_size) + + return logging_output + + def zero_grad(self): + self.optimizer.zero_grad() + + def lr_step(self, epoch, val_loss=None): + """Adjust the learning rate at the end of the epoch.""" + self.lr_scheduler.step(epoch, val_loss) + # prefer updating the LR based on the number of steps + return self.lr_step_update() + + def lr_step_update(self): + """Update the learning rate after each update.""" + new_lr = self.lr_scheduler.step_update(self.get_num_updates()) + metrics.log_scalar("lr", new_lr, weight=0, priority=300) + return new_lr + + def get_lr(self): + """Get the current learning rate.""" + return self.optimizer.get_lr() + + def get_model(self): + """Get the (non-wrapped) model instance.""" + return self._model + + def get_criterion(self): + """Get the (non-wrapped) criterion instance.""" + return self._criterion + + def get_meter(self, name): + """[deprecated] Get a specific meter by name.""" + from fairseq import meters + + if 'get_meter' not in self._warn_once: + self._warn_once.add('get_meter') + utils.deprecation_warning( + 'Trainer.get_meter is deprecated. Please use fairseq.metrics instead.' + ) + + train_meters = metrics.get_meters("train") + if train_meters is None: + train_meters = {} + + if name == "train_loss" and "loss" in train_meters: + return train_meters["loss"] + elif name == "train_nll_loss": + # support for legacy train.py, which assumed this meter is + # always initialized + m = train_meters.get("nll_loss", None) + return m or meters.AverageMeter() + elif name == "wall": + # support for legacy train.py, which assumed this meter is + # always initialized + m = metrics.get_meter("default", "wall") + return m or meters.TimeMeter() + elif name == "wps": + m = metrics.get_meter("train", "wps") + return m or meters.TimeMeter() + elif name in {"valid_loss", "valid_nll_loss"}: + # support for legacy train.py, which assumed these meters + # are always initialized + k = name[len("valid_"):] + m = metrics.get_meter("valid", k) + return m or meters.AverageMeter() + elif name == "oom": + return meters.AverageMeter() + elif name in train_meters: + return train_meters[name] + return None + + def get_num_updates(self): + """Get the number of parameters updates.""" + return self._num_updates + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + self._num_updates = num_updates + self.lr_step_update() + if self.quantizer: + self.quantizer.step_update(self._num_updates) + metrics.log_scalar("num_updates", self._num_updates, weight=0, priority=200) + + def clip_grad_norm(self, clip_norm): + return self.optimizer.clip_grad_norm(clip_norm, aggregate_norm_fn=None) + + def cumulative_training_time(self): + if self._cumulative_training_time is None: + # single GPU + return self._local_cumulative_training_time() + else: + return self._cumulative_training_time + + def _local_cumulative_training_time(self): + """Aggregate training time in seconds.""" + return time.time() - self._start_time + self._previous_training_time + + def _prepare_sample(self, sample): + if sample == "DUMMY": + raise Exception( + "Trying to use an uninitialized 'dummy' batch. This usually indicates " + "that the total number of batches is smaller than the number of " + "participating GPUs. Try reducing the batch size or using fewer GPUs." + ) + + if sample is None or len(sample) == 0: + return None + + if self.cuda: + sample = utils.move_to_cuda(sample) + + def apply_half(t): + if t.dtype is torch.float32: + return t.half() + return t + + def apply_bfloat16(t): + if t.dtype is torch.float32: + return t.to(dtype=torch.bfloat16) + return t + + if self.args.fp16: + sample = utils.apply_to_sample(apply_half, sample) + + if self.args.bf16: + sample = utils.apply_to_sample(apply_bfloat16, sample) + + return sample + + def _set_seed(self): + # Set seed based on args.seed and the update number so that we get + # reproducible results when resuming from checkpoints + seed = self.args.seed + self.get_num_updates() + utils.set_torch_seed(seed) + + def _sync_stats(self): + # Return True if it's using multiple GPUs and DDP or multiple GPUs with + # BMUF and it's a bmuf sync with warmup iterations completed before. + if self.data_parallel_world_size == 1: + return False + elif self.args.use_bmuf: + return ( + (self.get_num_updates() + 1) % self.args.global_sync_iter == 0 + and (self.get_num_updates() + 1) > self.args.warmup_iterations + ) + else: + return True + + def _log_oom(self, exc): + msg = "OOM: Ran out of memory with exception: {}".format(exc) + logger.warning(msg) + if torch.cuda.is_available() and hasattr(torch.cuda, "memory_summary"): + for device_idx in range(torch.cuda.device_count()): + logger.warning(torch.cuda.memory_summary(device=device_idx)) + sys.stderr.flush() + + def _aggregate_logging_outputs( + self, + logging_outputs: List[Dict[str, Any]], + *extra_stats_to_sum, + ignore=False, + ): + if self.task.__class__.logging_outputs_can_be_summed(self.get_criterion()): + return self._fast_stat_sync_sum( + logging_outputs, *extra_stats_to_sum, ignore=ignore + ) + else: + return self._all_gather_list_sync( + logging_outputs, *extra_stats_to_sum, ignore=ignore + ) + + def _all_gather_list_sync( + self, + logging_outputs: List[Dict[str, Any]], + *extra_stats_to_sum, + ignore=False, + ): + """ + Sync logging outputs across workers. all_gather_list_sync is + suitable when logging outputs are complex types. + """ + if self.tpu: + raise NotImplementedError + if ignore: + logging_outputs = [] + results = list(zip( + *distributed_utils.all_gather_list( + [logging_outputs] + list(extra_stats_to_sum), + max_size=getattr(self.args, 'all_gather_list_size', 16384), + group=self.data_parallel_process_group, + ) + )) + logging_outputs, extra_stats_to_sum = results[0], results[1:] + logging_outputs = list(chain.from_iterable(logging_outputs)) + extra_stats_to_sum = [sum(s) for s in extra_stats_to_sum] + return logging_outputs, extra_stats_to_sum + + def _fast_stat_sync_sum( + self, + logging_outputs: List[Dict[str, Any]], + *extra_stats_to_sum, + ignore=False, + ): + """ + Sync logging outputs across workers. fast_stat_sync_sum is + faster than all_gather_list_sync, but is only suitable when + logging outputs are scalars and can be summed. Note that + *logging_outputs* cannot contain any nested dicts/lists. + """ + data = {} + for i, stat in enumerate(extra_stats_to_sum): + data['extra_stats_' + str(i)] = stat + if len(logging_outputs) > 0: + log_keys = list(logging_outputs[0].keys()) + for k in log_keys: + if not ignore: + v = sum(log[k] for log in logging_outputs if k in log) + else: + v = logging_outputs[0][k] + v = torch.zeros_like(v) if torch.is_tensor(v) else 0 + data['logging_outputs_' + k] = v + else: + log_keys = None + + data = distributed_utils.all_reduce_dict( + data, + device=self.device, + group=self.data_parallel_process_group + ) + + extra_stats_to_sum = [ + data['extra_stats_' + str(i)] for i in range(len(extra_stats_to_sum)) + ] + if log_keys is not None: + logging_outputs = [{k: data['logging_outputs_' + k] for k in log_keys}] + else: + logging_outputs = [] + return logging_outputs, extra_stats_to_sum + + def _check_grad_norms(self, grad_norm): + """Check that grad norms are consistent across workers.""" + if self._grad_norm_buf is not None: + self._grad_norm_buf.zero_() + self._grad_norm_buf[self.data_parallel_rank] = grad_norm + distributed_utils.all_reduce( + self._grad_norm_buf, + group=self.data_parallel_process_group + ) + + def is_consistent(tensor): + max_abs_diff = torch.max(torch.abs(tensor - tensor[0])) + return (max_abs_diff / (tensor[0] + 1e-6) < 1e-6).all() + + if not is_consistent(self._grad_norm_buf): + pretty_detail = "\n".join( + "rank {:3d} = {:.8f}".format(r, n) + for r, n in enumerate(self._grad_norm_buf.tolist()) + ) + error_detail = "grad_norm across the workers:\n{}\n".format(pretty_detail) + raise RuntimeError( + "Fatal error: gradients are inconsistent between workers. " + "Try --ddp-backend=no_c10d. " + "Or are you mixing up different generation of GPUs in training?" + + "\n" + + "-" * 80 + + "\n{}\n".format(error_detail) + + "-" * 80 + ) + + def _reduce_and_log_stats(self, logging_outputs, sample_size, grad_norm=None): + if grad_norm is not None: + metrics.log_speed("ups", 1., priority=100, round=2) + metrics.log_scalar("gnorm", grad_norm, priority=400, round=3) + if self.args.clip_norm > 0: + metrics.log_scalar( + "clip", + torch.where( + grad_norm > self.args.clip_norm, + grad_norm.new_tensor(100), + grad_norm.new_tensor(0), + ), + priority=500, + round=1, + ) + + with metrics.aggregate() as agg: + if logging_outputs is not None: + self.task.reduce_metrics(logging_outputs, self.get_criterion()) + del logging_outputs + + # extra warning for criterions that don't properly log a loss value + if "loss" not in agg: + if "loss" not in self._warn_once: + self._warn_once.add("loss") + logger.warning( + "Criterion.reduce_metrics did not log a 'loss' value, " + "which may break some functionality" + ) + metrics.log_scalar("loss", -1) + + # support legacy interface + if self.tpu: + logging_output = {} + else: + logging_output = agg.get_smoothed_values() + logging_output["sample_size"] = sample_size + for key_to_delete in ["ppl", "wps", "wpb", "bsz"]: + if key_to_delete in logging_output: + del logging_output[key_to_delete] + return logging_output + + def _check_xla_compilation(self, message=None): + import torch_xla.debug.metrics as met + compile_stats = met.metric_data("CompileTime") + if compile_stats is None: + return + num_xla_compiles = compile_stats[0] + if num_xla_compiles > self._num_xla_compiles: + if message is None: + message = ( + "too many of these can lead to slow training, " + "but we expect a few in the beginning" + ) + logging.info("NOTE: XLA compilation detected; {}".format(message)) + self._num_xla_compiles = num_xla_compiles + + +def _catalog_shared_params(module, memo=None, prefix=''): + if memo is None: + first_call = True + memo = {} + else: + first_call = False + for name, param in module._parameters.items(): + param_prefix = prefix + ('.' if prefix else '') + name + if param not in memo: + memo[param] = [] + memo[param].append(param_prefix) + for name, m in module._modules.items(): + if m is None: + continue + submodule_prefix = prefix + ('.' if prefix else '') + name + _catalog_shared_params(m, memo, submodule_prefix) + if first_call: + return [x for x in memo.values() if len(x) > 1] + + +def _get_module_by_path(module, path): + path = path.split('.') + for name in path: + module = getattr(module, name) + return module + + +def _set_module_by_path(module, path, value): + path = path.split('.') + for name in path[:-1]: + module = getattr(module, name) + setattr(module, path[-1], value) diff --git a/fairseq/utils.py b/fairseq/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f68860330c71ba9e32735afb18f8888c66d41004 --- /dev/null +++ b/fairseq/utils.py @@ -0,0 +1,589 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import copy +import importlib.util +import logging +import math +import os +import sys +import warnings +from collections import defaultdict +from itertools import accumulate +from typing import Callable, Dict, List, Optional + +import numpy as np +import torch +import torch.nn.functional as F +from fairseq.logging.meters import safe_round +from fairseq.modules import gelu, gelu_accurate +from fairseq.modules.multihead_attention import MultiheadAttention +from torch import Tensor + +try: + from amp_C import multi_tensor_l2norm + multi_tensor_l2norm_available = True +except ImportError: + multi_tensor_l2norm_available = False + + +logger = logging.getLogger(__name__) + + +MANIFOLD_PATH_SEP = "|" + + +def split_paths(paths: str) -> List[str]: + return paths.split(os.pathsep) if "://" not in paths else paths.split(MANIFOLD_PATH_SEP) + + +def load_ensemble_for_inference(filenames, task, model_arg_overrides=None): + from fairseq import checkpoint_utils + + deprecation_warning( + "utils.load_ensemble_for_inference is deprecated. " + "Please use checkpoint_utils.load_model_ensemble instead." + ) + return checkpoint_utils.load_model_ensemble( + filenames, arg_overrides=model_arg_overrides, task=task + ) + + +def apply_to_sample(f, sample): + if hasattr(sample, '__len__') and len(sample) == 0: + return {} + + def _apply(x): + if torch.is_tensor(x): + return f(x) + elif isinstance(x, dict): + return {key: _apply(value) for key, value in x.items()} + elif isinstance(x, list): + return [_apply(x) for x in x] + elif isinstance(x, tuple): + return tuple(_apply(x) for x in x) + elif isinstance(x, set): + return {_apply(x) for x in x} + else: + return x + + return _apply(sample) + + +def move_to_cuda(sample): + def _move_to_cuda(tensor): + return tensor.cuda() + + return apply_to_sample(_move_to_cuda, sample) + + +def move_to_cpu(sample): + def _move_to_cpu(tensor): + # PyTorch has poor support for half tensors (float16) on CPU. + # Move any such tensors to float32. + if tensor.dtype in {torch.bfloat16, torch.float16}: + tensor = tensor.to(dtype=torch.float32) + return tensor.cpu() + + return apply_to_sample(_move_to_cpu, sample) + + +def get_incremental_state( + module: MultiheadAttention, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, +) -> Optional[Dict[str, Optional[Tensor]]]: + """Helper for getting incremental state for an nn.Module.""" + return module.get_incremental_state(incremental_state, key) + + +def set_incremental_state( + module: MultiheadAttention, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + key: str, + value: Dict[str, Optional[Tensor]], +) -> Optional[Dict[str, Dict[str, Optional[Tensor]]]]: + """Helper for setting incremental state for an nn.Module.""" + if incremental_state is not None: + result = module.set_incremental_state(incremental_state, key, value) + if result is not None: + incremental_state = result + return incremental_state + + +def load_align_dict(replace_unk): + if replace_unk is None: + align_dict = None + elif isinstance(replace_unk, str) and len(replace_unk) > 0: + # Load alignment dictionary for unknown word replacement if it was passed as an argument. + align_dict = {} + with open(replace_unk, "r") as f: + for line in f: + cols = line.split() + align_dict[cols[0]] = cols[1] + else: + # No alignment dictionary provided but we still want to perform unknown word replacement by copying the + # original source word. + align_dict = {} + return align_dict + + +def print_embed_overlap(embed_dict, vocab_dict): + embed_keys = set(embed_dict.keys()) + vocab_keys = set(vocab_dict.symbols) + overlap = len(embed_keys & vocab_keys) + logger.info("found {}/{} types in embedding file".format(overlap, len(vocab_dict))) + + +def parse_embedding(embed_path): + """Parse embedding text file into a dictionary of word and embedding tensors. + + The first line can have vocabulary size and dimension. The following lines + should contain word and embedding separated by spaces. + + Example: + 2 5 + the -0.0230 -0.0264 0.0287 0.0171 0.1403 + at -0.0395 -0.1286 0.0275 0.0254 -0.0932 + """ + embed_dict = {} + with open(embed_path) as f_embed: + next(f_embed) # skip header + for line in f_embed: + pieces = line.rstrip().split(" ") + embed_dict[pieces[0]] = torch.Tensor( + [float(weight) for weight in pieces[1:]] + ) + return embed_dict + + +def load_embedding(embed_dict, vocab, embedding): + for idx in range(len(vocab)): + token = vocab[idx] + if token in embed_dict: + embedding.weight.data[idx] = embed_dict[token] + return embedding + + +def replace_unk(hypo_str, src_str, alignment, align_dict, unk): + from fairseq import tokenizer + + # Tokens are strings here + hypo_tokens = tokenizer.tokenize_line(hypo_str) + # TODO: Very rare cases where the replacement is '' should be handled gracefully + src_tokens = tokenizer.tokenize_line(src_str) + [""] + for i, ht in enumerate(hypo_tokens): + if ht == unk: + src_token = src_tokens[alignment[i]] + # Either take the corresponding value in the aligned dictionary or just copy the original value. + hypo_tokens[i] = align_dict.get(src_token, src_token) + return " ".join(hypo_tokens) + + +def post_process_prediction( + hypo_tokens, src_str, alignment, align_dict, tgt_dict, remove_bpe=None, extra_symbols_to_ignore=None +): + hypo_str = tgt_dict.string(hypo_tokens, remove_bpe, extra_symbols_to_ignore=extra_symbols_to_ignore) + if align_dict is not None: + hypo_str = replace_unk( + hypo_str, src_str, alignment, align_dict, tgt_dict.unk_string() + ) + if align_dict is not None or remove_bpe is not None: + # Convert back to tokens for evaluating with unk replacement or without BPE + # Note that the dictionary can be modified inside the method. + hypo_tokens = tgt_dict.encode_line(hypo_str, add_if_not_exist=True) + return hypo_tokens, hypo_str, alignment + + +def make_positions(tensor, padding_idx: int, onnx_trace: bool = False): + """Replace non-padding symbols with their position numbers. + + Position numbers begin at padding_idx+1. Padding symbols are ignored. + """ + # The series of casts and type-conversions here are carefully + # balanced to both work with ONNX export and XLA. In particular XLA + # prefers ints, cumsum defaults to output longs, and ONNX doesn't know + # how to handle the dtype kwarg in cumsum. + mask = tensor.ne(padding_idx).int() + return (torch.cumsum(mask, dim=1).type_as(mask) * mask).long() + padding_idx + + +def strip_pad(tensor, pad): + return tensor[tensor.ne(pad)] + + +def buffered_arange(max): + if not hasattr(buffered_arange, "buf"): + buffered_arange.buf = torch.LongTensor() + if max > buffered_arange.buf.numel(): + buffered_arange.buf.resize_(max) + torch.arange(max, out=buffered_arange.buf) + return buffered_arange.buf[:max] + + +def convert_padding_direction( + src_tokens, padding_idx, right_to_left: bool = False, left_to_right: bool = False +): + assert right_to_left ^ left_to_right + pad_mask = src_tokens.eq(padding_idx) + if not pad_mask.any(): + # no padding, return early + return src_tokens + if left_to_right and not pad_mask[:, 0].any(): + # already right padded + return src_tokens + if right_to_left and not pad_mask[:, -1].any(): + # already left padded + return src_tokens + max_len = src_tokens.size(1) + buffered = torch.empty(0).long() + if max_len > 0: + torch.arange(max_len, out=buffered) + range = buffered.type_as(src_tokens).expand_as(src_tokens) + num_pads = pad_mask.long().sum(dim=1, keepdim=True) + if right_to_left: + index = torch.remainder(range - num_pads, max_len) + else: + index = torch.remainder(range + num_pads, max_len) + return src_tokens.gather(1, index) + + +def item(tensor): + if hasattr(tensor, "item"): + return tensor.item() + if hasattr(tensor, "__getitem__"): + return tensor[0] + return tensor + + +def multi_tensor_total_norm(grads, chunk_size=2048*32) -> torch.Tensor: + per_device_grads = {} + norms = [] + for grad in grads: + device = grad.device + cur_device_grads = per_device_grads.get(device) + if cur_device_grads is None: + cur_device_grads = [] + per_device_grads[device] = cur_device_grads + cur_device_grads.append(grad) + for device in per_device_grads.keys(): + cur_device_grads = per_device_grads[device] + if device.type == "cuda": + # TODO(msb) return has_inf + has_inf = torch.zeros((1, 1), dtype=torch.int, device=device) + with torch.cuda.device(device): + norm = multi_tensor_l2norm(chunk_size, has_inf, [cur_device_grads], False) + norms.append(norm[0]) + else: + norms += [torch.norm(g, p=2, dtype=torch.float32) for g in cur_device_grads] + total_norm = torch.norm(torch.stack(norms)) + return total_norm + + +def clip_grad_norm_(params, max_norm, aggregate_norm_fn=None) -> torch.Tensor: + if isinstance(params, torch.Tensor): + params = [params] + params = list(params) + grads = [p.grad.detach() for p in filter(lambda p: p.grad is not None, params)] + if len(grads) == 0: + if len(params) > 0: + return params[0].new_tensor(0.) + else: + return torch.tensor(0.) + + if len(grads) == 1: + total_norm = torch.norm(grads[0], p=2, dtype=torch.float32) + else: + if multi_tensor_l2norm_available: + total_norm = multi_tensor_total_norm(grads) + else: + if torch.cuda.is_available(): + warnings.warn( + "amp_C fused kernels unavailable, disabling multi_tensor_l2norm; " + "you may get better performance by installing NVIDIA's apex library" + ) + total_norm = torch.norm( + torch.stack([torch.norm(g, p=2, dtype=torch.float32) for g in grads]) + ) + + if aggregate_norm_fn is not None: + total_norm = aggregate_norm_fn(total_norm) + + if max_norm > 0: + max_norm = float(max_norm) + clip_coef = (max_norm / (total_norm + 1e-6)).clamp_(max=1) + for g in grads: + g.mul_(clip_coef) + return total_norm + + +def fill_with_neg_inf(t): + """FP16-compatible function that fills a tensor with -inf.""" + return t.float().fill_(float("-inf")).type_as(t) + + +def _match_types(arg1, arg2): + """Convert the numerical argument to the same type as the other argument""" + + def upgrade(arg_number, arg_structure): + if isinstance(arg_structure, tuple): + return tuple([arg_number] * len(arg_structure)) + elif isinstance(arg_structure, dict): + arg = copy.deepcopy(arg_structure) + for k in arg: + arg[k] = upgrade(arg_number, arg_structure[k]) + return arg + else: + return arg_number + + if isinstance(arg1, float) or isinstance(arg1, int): + return upgrade(arg1, arg2), arg2 + elif isinstance(arg2, float) or isinstance(arg2, int): + return arg1, upgrade(arg2, arg1) + + return arg1, arg2 + + +def resolve_max_positions(*args): + """Resolve max position constraints from multiple sources.""" + + def map_value_update(d1, d2): + updated_value = copy.deepcopy(d1) + for key in d2: + if key not in updated_value: + updated_value[key] = d2[key] + else: + updated_value[key] = min(d1[key], d2[key]) + return updated_value + + def nullsafe_min(l): + minim = None + for item in l: + if minim is None: + minim = item + elif item is not None and item < minim: + minim = item + return minim + + max_positions = None + for arg in args: + if max_positions is None: + max_positions = arg + elif arg is not None: + max_positions, arg = _match_types(max_positions, arg) + if isinstance(arg, float) or isinstance(arg, int): + max_positions = min(max_positions, arg) + elif isinstance(arg, dict): + max_positions = map_value_update(max_positions, arg) + else: + max_positions = tuple(map(nullsafe_min, zip(max_positions, arg))) + + return max_positions + + +def import_user_module(args): + module_path = getattr(args, "user_dir", None) + if module_path is not None: + module_path = os.path.abspath(args.user_dir) + if not os.path.exists(module_path): + fairseq_rel_path = os.path.join( + os.path.dirname(__file__), "..", args.user_dir + ) + if os.path.exists(fairseq_rel_path): + module_path = fairseq_rel_path + module_parent, module_name = os.path.split(module_path) + + if module_name not in sys.modules: + sys.path.insert(0, module_parent) + importlib.import_module(module_name) + + +def softmax(x, dim: int, onnx_trace: bool = False): + if onnx_trace: + return F.softmax(x.float(), dim=dim) + else: + return F.softmax(x, dim=dim, dtype=torch.float32) + + +def log_softmax(x, dim: int, onnx_trace: bool = False): + if onnx_trace: + return F.log_softmax(x.float(), dim=dim) + else: + return F.log_softmax(x, dim=dim, dtype=torch.float32) + + +def get_perplexity(loss, round=2, base=2): + if loss is None: + return 0. + try: + return safe_round(base ** loss, round) + except OverflowError: + return float('inf') + + +def deprecation_warning(message, stacklevel=3): + # don't use DeprecationWarning, since it's ignored by default + warnings.warn(message, stacklevel=stacklevel) + + +def get_activation_fn(activation: str) -> Callable: + """ Returns the activation function corresponding to `activation` """ + if activation == "relu": + return F.relu + elif activation == "gelu": + return gelu + elif activation == "gelu_fast": + deprecation_warning( + "--activation-fn=gelu_fast has been renamed to gelu_accurate" + ) + return gelu_accurate + elif activation == "gelu_accurate": + return gelu_accurate + elif activation == "tanh": + return torch.tanh + elif activation == "linear": + return lambda x: x + else: + raise RuntimeError("--activation-fn {} not supported".format(activation)) + + +def get_available_activation_fns() -> List: + return [ + "relu", + "gelu", + "gelu_fast", # deprecated + "gelu_accurate", + "tanh", + "linear", + ] + + +@contextlib.contextmanager +def eval(model): + is_training = model.training + model.eval() + yield + model.train(is_training) + + +def has_parameters(module): + try: + next(module.parameters()) + return True + except StopIteration: + return False + + +def set_torch_seed(seed): + # Set seed based on args.seed and the update number so that we get + # reproducible results when resuming from checkpoints + assert isinstance(seed, int) + torch.manual_seed(seed) + torch.cuda.manual_seed(seed) + + +@contextlib.contextmanager +def with_torch_seed(seed): + assert isinstance(seed, int) + rng_state = torch.get_rng_state() + cuda_rng_state = torch.cuda.get_rng_state() + set_torch_seed(seed) + yield + torch.set_rng_state(rng_state) + torch.cuda.set_rng_state(cuda_rng_state) + + +def parse_alignment(line): + """ + Parses a single line from the alingment file. + + Args: + line (str): String containing the alignment of the format: + - - .. + -. All indices are 0 indexed. + + Returns: + torch.IntTensor: packed alignments of shape (2 * m). + """ + alignments = line.strip().split() + parsed_alignment = torch.IntTensor(2 * len(alignments)) + for idx, alignment in enumerate(alignments): + src_idx, tgt_idx = alignment.split("-") + parsed_alignment[2 * idx] = int(src_idx) + parsed_alignment[2 * idx + 1] = int(tgt_idx) + return parsed_alignment + + +def get_token_to_word_mapping(tokens, exclude_list): + n = len(tokens) + word_start = [int(token not in exclude_list) for token in tokens] + word_idx = list(accumulate(word_start)) + token_to_word = {i: word_idx[i] for i in range(n)} + return token_to_word + + +def extract_hard_alignment(attn, src_sent, tgt_sent, pad, eos): + tgt_valid = ((tgt_sent != pad) & (tgt_sent != eos)).nonzero().squeeze(dim=-1) + src_invalid = ((src_sent == pad) | (src_sent == eos)).nonzero().squeeze(dim=-1) + src_token_to_word = get_token_to_word_mapping(src_sent, [eos, pad]) + tgt_token_to_word = get_token_to_word_mapping(tgt_sent, [eos, pad]) + alignment = [] + if len(tgt_valid) != 0 and len(src_invalid) < len(src_sent): + attn_valid = attn[tgt_valid] + attn_valid[:, src_invalid] = float("-inf") + _, src_indices = attn_valid.max(dim=1) + for tgt_idx, src_idx in zip(tgt_valid, src_indices): + alignment.append( + ( + src_token_to_word[src_idx.item()] - 1, + tgt_token_to_word[tgt_idx.item()] - 1, + ) + ) + return alignment + + +def new_arange(x, *size): + """ + Return a Tensor of `size` filled with a range function on the device of x. + If size is empty, using the size of the variable x. + """ + if len(size) == 0: + size = x.size() + return torch.arange(size[-1], device=x.device).expand(*size).contiguous() + + +def get_tpu_device(args): + import torch_xla.core.xla_model as xm + return xm.xla_device() + + +class CudaEnvironment(object): + def __init__(self): + cur_device = torch.cuda.current_device() + prop = torch.cuda.get_device_properties("cuda:{}".format(cur_device)) + self.name = prop.name + self.major = prop.major + self.minor = prop.minor + self.total_memory_in_GB = prop.total_memory / 1024 / 1024 / 1024 + + @staticmethod + def pretty_print_cuda_env_list(cuda_env_list): + """ + Given a list of CudaEnviorments, pretty print them + """ + num_workers = len(cuda_env_list) + center = "CUDA enviroments for all {} workers".format(num_workers) + banner_len = 40 - len(center) // 2 + first_line = "*" * banner_len + center + "*" * banner_len + logger.info(first_line) + for r, env in enumerate(cuda_env_list): + logger.info( + "rank {:3d}: ".format(r) + + "capabilities = {:2d}.{:<2d} ; ".format(env.major, env.minor) + + "total memory = {:.3f} GB ; ".format(env.total_memory_in_GB) + + "name = {:40s}".format(env.name) + ) + logger.info(first_line)