id
int64
0
190k
prompt
stringlengths
21
13.4M
docstring
stringlengths
1
12k
23,543
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption, NndctScreenLogger, QError from nndct_shared.quantization import kernel_need_quant from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import maybe_get_quantizer import pytorch_nndct.utils as py_utils import torch.nn.functional as F from pytorch_nndct.utils import Const from .fix_ops import NndctISqrt class deephi_GroupNorm(torch.nn.modules.normalization.GroupNorm): r"""DeePhi group normalization operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_GroupNorm, self).__init__(*args, **kwards) self.params_name = None self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() self.param_saved = False self.param_quantized = False def forward(self, input): if not kernel_need_quant(self.quantizer, self.node): output = super().forward(input) output = quantize_tensors([output], self.node)[0] return output # quantize input tensor qinput = quantize_tensors([input], self.node, tensor_type='input')[0] params = [] if self.weight is not None: params.append(self.weight) if self.bias is not None: params.append(self.bias) param_names = self.params_name[:len(params)] if len(params) != len(param_names): NndctScreenLogger().error2user(QError.PARAM_NUMBER, f"Parameter number error in node {self.node} for InstanceNorm operator!") if (not self.param_quantized) and len(params) > 0: inplace = self.quantizer is not None and self.quantizer.inplace # quantize weights and bias if inplace: _ = quantize_tensors( params, self.node, tensor_names=param_names, tensor_type='param') qparams = [p for p in params] else: qparams = quantize_tensors( params, self.node, tensor_names=param_names, tensor_type='param') if not NndctOption.nndct_quant_off.value: self.param_quantized = True else: qparams = [p for p in params] # quantization configure input_name = self.node.in_nodes[0] input_node = self.quantizer.configer.get_Nndctnode(input_name) if not self.quantizer.configer.node_output_quantizable(input_node): input_name = input_node.in_nodes[0] if NndctOption.nndct_op_groupnorm_mode.value == "ipu_8bw" and (not self.quantizer.exporting): ifp = self.quantizer.get_quant_config(input_name, False)[1] wfp = self.quantizer.get_quant_config(self.params_name[0], False, tensor_type='param')[1] if self.affine else None bfp = self.quantizer.get_quant_config(self.params_name[1], False, tensor_type='param')[1] if self.affine else None output = self.simulateGroupNorm(qinput, self.num_groups, self.affine, qparams[0], qparams[1], ifp, wfp, bfp) output = quantize_tensors([output], self.node, method=4)[0] else: output = torch.nn.functional.group_norm( qinput, self.num_groups, qparams[0] if self.param_quantized else None, qparams[1] if self.param_quantized else None, self.eps ) output = quantize_tensors([output], self.node)[0] return output # e.g. inp (N, C, H, W) def simulateGroupNorm(self, inp, num_groups, affine, weight, bias, ifp, wfp, bfp): # inp fake int8: [-256, 255] inp = torch.floor(inp * (2**ifp)) # float32 inp = inp.to(torch.bfloat16) # bfloat16 # inp reshape: (N, G, C//G, H, W) G = self.num_groups if inp.dim() == 3: # [N, C, L] N, C, L = inp.shape inp_group = inp.reshape(N, G, C//G, L) dim = [2, 3] elif inp.dim() == 4: # [N, C, H, W] N, C, H, W = inp.shape inp_group = inp.reshape(N, G, C//G, H, W) dim = [2, 3, 4] elif inp.dim() == 5: # [N, C, L, H, W] N, C, D, H, W = inp.shape inp_group = inp.reshape(N, G, C//G, D, H, W) dim = [2, 3, 4, 5] else: NndctScreenLogger().error2user(QError.INPUT_DIMENSION, f"Input dimension error in node {self.node}! The dimension of input is {inp.dim()}.") numel = inp_group[0][0].numel() # mean: aie_mean equals to torch.mean(inp, dim, keepdim=True) except adding # aid_add input reshape NG, CG = inp_group.shape[0], inp_group.shape[1] # inp_group: (N, G, C//G, H, W)=(NG, CG, DG, HG, WG) inp_temp1 = inp_group.reshape(NG, CG, -1) # (NG, CG, DG*HG*WG), bfloat16 numel = inp_temp1[0][0].numel() # element number inp_temp1 = inp_group.reshape(-1, inp_temp1.shape[-1]) # (NG*CG, DG*HG*WG), bfloat16 inp_temp2 = inp_temp1.reshape(inp_temp1.shape[0], -1, 8) # (NG*CG, DG*HG*WG/8, 8), bfloat16 inp_v8 = inp_temp2.sum(dim=-2, keepdim=False, dtype=torch.float32) # (NG*CG, 8), float32 # aie add v8 sum = py_utils.aie_add_v8(inp_v8) # aie sum: (NG*CG,), float32 # aie mean sum = sum.reshape(NG, CG) # (NG, CG), float32 mean = sum/numel # (NG, CG), float32 # mean unsqueeze: (NG, CG, 1, 1) for i in range(mean.dim(), inp_group.dim()): mean = torch.unsqueeze(mean, -1) mean = mean.to(torch.bfloat16) # (NG, CG, 1, 1, 1) # x - mu sub = inp_group - mean # (NG, CG, DG, HG, WG), bfloat16 sub = sub.to(torch.float32) # float32 # var square = torch.square(sub) # float32 var = square.mean(dim, keepdim=False, dtype=torch.float32) # (NG, CG), float32: dim=(2, 3, 4) # isqrt: 1/sqrt(var) isqrt = torch.empty_like(var) NndctISqrt(var, isqrt) # CUDA/CPU: float32 isqrt = isqrt.to(torch.bfloat16) # mul: (x-mu)*(1/sigma) for i in range(isqrt.dim(), sub.dim()): # isqrt shape: (NG, CG, 1, 1, 1) isqrt = torch.unsqueeze(isqrt, dim=-1) mul = torch.mul(sub, isqrt) # float32, (NG, CG, DG, HG, WG) mul = mul.to(torch.bfloat16).to(torch.float32) # float32 mul = mul.reshape(N, C, H, W) # (N, C, H, W) # affine: (x - mu)/sigma*gamma + beta if affine: weight = weight.repeat(mul.shape[0], 1) # repeat C into (N, C) bias = bias.repeat(mul.shape[0], 1) # repeat C into (N, C) for i in range(weight.dim(), mul.dim()): # unsqueeze (N, C) into (N, C, 1), (N, C, 1, 1) or (N, C, 1, 1, 1) weight = torch.unsqueeze(weight, -1) # 1d (N, C, 1); 2d (N, C, 1, 1); 3d (N, C, 1, 1, 1) bias = torch.unsqueeze(bias, -1) # 1d (N, C, 1); 2d (N, C, 1, 1); 3d (N, C, 1, 1, 1) out = mul*weight + bias # (x-mu)/sigma*gamma + beta, float32 out = out.to(torch.bfloat16).to(torch.float32) # float32 else: # (x - mu)/sigma out = mul return out def GroupNorm(*args, **kwargs): return deephi_GroupNorm(*args, **kwargs)
null
23,544
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption, NndctScreenLogger, QError, QWarning from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from .quant_noise import eval_qnoise import pytorch_nndct.utils as py_utils import torch.nn.functional as F class deephi_Conv1d(torch.nn.modules.conv.Conv1d): r"""DeePhi Conv1d operation, support float and double""" def __init__(self, *args, **kwargs): super(deephi_Conv1d, self).__init__(*args, **kwargs) self.params_name = None self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() self.param_saved = False self.param_quantized = False # self.weight and self.bias are not quantized float parameters self.weight_bak = None # backup of float bias for bias correction self.bias_bak = None # backup of float bias for bias correction self.stop = False self.rate = NndctOption.nndct_param_corr_rate.value self.efficency = 0.0 self.deviation = 0.0 def forward(self, input): # backup bias for bias correction feature if (not self.param_saved): if NndctOption.nndct_param_corr.value > 0: # backup orignal float parameters if self.quant_mode == 1: self.weight_bak = self.weight.detach().clone() if self.bias is not None: self.bias_bak = self.bias.detach().clone() # adjust bias if self.quant_mode == 2 and self.bias is not None: if not self.quantizer.has_bias_corr(self.node): NndctScreenLogger().error2user(QError.BIAS_CORRECTION, f"Bias correction file in quantization result directory does not match current model.") exit(2) self.bias.data = torch.sub(self.bias.data, torch.tensor( self.quantizer.get_bias_corr(self.node), device=self.bias.data.device, dtype=self.bias.data.dtype)) self.param_saved = True # quantize parameters qweight = None qbias = None inplace = (NndctOption.nndct_quant_off.value or self.quantizer is not None and self.quantizer.inplace) if (not self.param_quantized): if inplace: _ = quantize_tensors( [self.weight], self.node, tensor_names = [self.params_name[0]], tensor_type = 'param')[0] qweight = self.weight if self.bias is not None: _ = quantize_tensors( [self.bias], self.node, tensor_names = [self.params_name[1]], tensor_type = 'param')[0] qbias = self.bias else: qweight = quantize_tensors( [self.weight], self.node, tensor_names = [self.params_name[0]], tensor_type = 'param')[0] if self.bias is not None: qbias = quantize_tensors( [self.bias], self.node, tensor_names = [self.params_name[1]], tensor_type = 'param')[0] if not NndctOption.nndct_quant_off.value: self.param_quantized = True else: qweight = self.weight qbias = self.bias # quantize input tensor qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = torch.nn.functional.conv1d(qinput, weight = qweight, bias = qbias, stride = self.stride, padding = self.padding, dilation = self.dilation, groups = self.groups) output = quantize_tensors([output], self.node)[0] # correct weights and bias in calibation if NndctOption.nndct_param_corr.value > 0: #rate = NndctOption.nndct_param_corr_rate.value # statistic of quantization error if (self.quant_mode == 1 and not self.stop): res_f = torch.nn.functional.conv1d(input, self.weight_bak, bias = self.bias_bak, stride = self.stride, padding = self.padding, dilation = self.dilation, groups = self.groups) error, rate, self.stop, self.efficency, self.deviation = eval_qnoise( output, res_f, self.efficency, self.deviation, self.rate, self.stop) if (not self.stop) and (self.bias is not None): error = error.mean(dim = [0, 1, 2]) self.bias.data = torch.sub(self.bias.data, error, alpha=rate) self.param_quantized = False return output def bias_corr(self): if self.bias is not None and self.bias_bak is not None: bias_err = torch.sub(self.bias_bak, self.bias.data) return bias_err.cpu().numpy().tolist() else: return None def Conv1d(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.Conv1d(*args, **kwargs) return deephi_Conv1d(*args, **kwargs)
null
23,545
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_Sub(torch.nn.Module): def __init__(self): def forward(self, input, other, alpha=1): def Sub(*args, **kwargs): return deephi_Sub(*args, **kwargs)
null
23,546
import torch import numpy as np from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.utils import NndctOption from nndct_shared.base import GLOBAL_MAP, NNDCT_KEYS from .tanh_table import * from .fix_ops import NndctTanhTableLookup, NndctTanhSimulation, NndctTanhTableLookupAIE2 import pytorch_nndct.utils as py_utils class deephi_Tanh(torch.nn.modules.Tanh): r"""DeePhi Tanh operation""" def __init__(self): super(deephi_Tanh, self).__init__() self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): if self.quant_mode == 0 or (not self.node.in_quant_part): return super().forward(input) qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if (NndctOption.nndct_quant_off.value or self.quantizer is None or self.quantizer.exporting or NndctOption.nndct_cv_app.value or NndctOption.nndct_only_int_quant is False): # Method 0: quant input and output (for CV) output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] else: output = torch.empty_like(qinput) input_name = self.node.in_nodes[0] input_node = self.quantizer.configer.get_Nndctnode(input_name) if not self.quantizer.configer.node_output_quantizable(input_node): input_name = input_node.in_nodes[0] fragpos = self.quantizer.get_quant_config(input_name, False)[1] # Method 1: Simulation AIE with 16 bw (for RNNT) if NndctOption.nndct_op_tanh_sigmoid_mode.value == "simulation": NndctTanhSimulation(input, output, fragpos) output = quantize_tensors([output], self.node)[0] # Method 2: Table Look up for AIE2 with 16 bw (based on LUT) elif NndctOption.nndct_op_tanh_sigmoid_mode.value == "aie2_lut_16bw" or NndctOption.nndct_ip_asr.value: NndctTanhTableLookupAIE2(qinput, output, fragpos) output = quantize_tensors([output], self.node)[0] # Method 3: Table Look up for FPGA with 16 bw else: quant_device = qinput.device Ttable = TANH_TABLE.table.to(qinput.dtype).to(quant_device) output = output.to(quant_device) NndctTanhTableLookup(input, Ttable, output, fragpos) bnfp = self.quantizer.get_quant_config(input_name, False) bnfp[1] = 15 self.quantizer.set_quant_config(self.node.name, bnfp) return output import torch def Tanh(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.Tanh(*args, **kwargs) return deephi_Tanh(*args, **kwargs)
null
23,547
import math import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.utils import NndctOption from .fix_ops import NndctExpApprAIE2, NndctLogSoftmaxFastLn, NndctLogSoftmaxSub import pytorch_nndct.utils as py_utils class deephi_LogSoftmax(torch.nn.modules.LogSoftmax): r"""DeePhi LogSoftmax operation""" def __init__(self, dim = None): super(deephi_LogSoftmax, self).__init__() self.dim = dim self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): if self.quant_mode == 0 or (not self.node.in_quant_part): return super().forward(input) qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if (NndctOption.nndct_quant_off.value or self.quantizer is None or self.quantizer.exporting): # Method 0: quant input and output output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] else: # Method: Table Look up for AIE2 (based on LUT) with 16 bw if NndctOption.nndct_op_softmax_mode.value == "aie2_lut_16bw": input_name = self.node.in_nodes[0] input_node = self.quantizer.configer.get_Nndctnode(input_name) if not self.quantizer.configer.node_output_quantizable(input_node): input_name = input_node.in_nodes[0] bw = self.quantizer.get_quant_config(input_name, False)[0] fragpos = self.quantizer.get_quant_config(input_name, False)[1] if bw == 8 and fragpos < 2: if fragpos < 2: qinput_max = torch.max(qinput, dim = self.dim, keepdim = True).values qinput -= qinput_max else: qinput -= 31.75 else: # bw == 16 if fragpos < 10: qinput_max = torch.max(qinput) qinput -= qinput_max else: qinput -= 32 qinput_exp = torch.empty_like(input) NndctExpApprAIE2(qinput, qinput_exp, bw) exp_sum = torch.sum(qinput_exp, dim = self.dim, keepdim = True) ln_sum = torch.empty_like(exp_sum) NndctLogSoftmaxFastLn(exp_sum, ln_sum) output = torch.empty_like(input) NndctLogSoftmaxSub(qinput, output, ln_sum) output = quantize_tensors([output], self.node)[0] else: output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] return output def LogSoftmax(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.LogSoftmax(*args, **kwargs) return deephi_LogSoftmax(*args, **kwargs)
null
23,548
import math import torch import pytorch_nndct.utils as py_utils from nndct_shared.quantization import maybe_get_quantizer, quantize_tensors from nndct_shared.utils import NndctOption, NndctScreenLogger class deephi_AdaptiveAvgPool2d(torch.nn.modules.AdaptiveAvgPool2d): r"""DeePhi Conv2d operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_AdaptiveAvgPool2d, self).__init__(*args, **kwards) self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() def forward(self, input): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = super().forward(qinput) input_size = [int(dim) for dim in input.shape[2:]] mod = [input_size[i] % self.output_size[i] for i in range(0, len(input_size))] if mod != [0] * len(mod): if self.node is not None: NndctScreenLogger().warning_once(f"AdaptiveAvgpool2d op({self.node.name}) is not quantized. Because it's output size {self.output_size} are not factor of input size {input_size}.") return output # During slow trace, the dim of shape will convert to tensor value which is not support in nndct. kernel = [int(input_size[i] / self.output_size[i]) for i in range(0, len(input_size))] # scale to DPU accuracy if NndctOption.nndct_avg_pool_approximate.value: scale = 1.0 if kernel == [3, 3]: scale = 9.0 * 7.0 / 64.0 elif kernel == [5, 5]: scale = 25.0 * 10.0 / 256.0 elif kernel in [[6, 6], [3, 6], [6, 3]]: scale = 36.0 * 7.0 / 256.0 elif kernel == [7, 7]: scale = 49.0 * 21.0 / 1024.0 elif kernel == [14, 14]: scale = 196.0 * 21.0 / 4096.0 else: rec = kernel[0] * kernel[1] max_factor = math.ceil(math.log(rec * 128, 2)) diff = 1.0 multi_factor = 0.0 shift_factor = 0.0 for shift_factor_ in range(max_factor): factor = round((2 ** shift_factor_)/rec) diff_ = abs(factor / (2 ** shift_factor_) - 1 / rec) if diff_ < diff: multi_factor = factor diff = diff_ shift_factor = shift_factor_ scale = rec * multi_factor / (2 ** shift_factor) output = output * scale output = quantize_tensors([output], self.node)[0] return output def AdaptiveAvgPool2d(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode is None or NndctOption.nndct_quant_off.value: return torch.nn.AdaptiveAvgPool2d(*args, **kwargs) return deephi_AdaptiveAvgPool2d(*args, **kwargs)
null
23,549
import os import re import torch from torch.autograd import Variable import math import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger, create_work_dir from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from .quant_noise import eval_qnoise import pytorch_nndct.utils as py_utils import torch.nn.functional as F from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version class deephi_GELU(torch.nn.GELU): r"""DeePhi GELU operation, support float and double""" def __init__(self, approximate=None): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.12.0"): super(deephi_GELU, self).__init__(approximate) else: super(deephi_GELU, self).__init__() self.approximate = approximate self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() def gelu(self, x): return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) def forward(self, input): if self.quant_mode <= 0 or (not self.node.in_quant_part) or NndctOption.nndct_gemm88.value: return super().forward(input) qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if NndctOption.nndct_quant_off.value or self.quantizer.exporting: output = super().forward(qinput) else: # Method 1: Dynamic table look up with 8 bw if NndctOption.nndct_op_gelu_mode.value == "dynamic_table" or NndctOption.nndct_ip_v70_bert.value: output = self.gelu(qinput) # Method 0: Quant input and output of Softmax elif compare_torch_version(CmpFlag.GREATER_EQUAL, "1.12.0"): output = F.gelu(qinput, approximate=self.approximate) else: output = F.gelu(qinput) output = quantize_tensors([output], self.node)[0] return output def GELU(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.GELU(*args, **kwargs) return deephi_GELU(*args, **kwargs)
null
23,550
import torch from torch.autograd import Variable import torch.nn.functional as F from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.utils import NndctOption import pytorch_nndct.utils as py_utils from nndct_shared.utils import NNDCT_KEYS, GLOBAL_MAP class deephi_Hardsigmoid(torch.nn.Module): def __init__(self, inplace=False, *args, **kwards): def forward(self, input): def Hardsigmoid(*args, **kwargs): return deephi_Hardsigmoid(*args, **kwargs)
null
23,551
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils from nndct_shared.utils import NndctOption class deephi_LeakyReLU(torch.nn.LeakyReLU): r"""DeePhi LeakyReLU operation""" def __init__(self, *args, **kwargs): # only support the specified slope and inplace operation super().__init__(*args, **kwargs) if NndctOption.nndct_leaky_relu_approximate.value: self.negative_slope = 0.1015625 self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] return output def LeakyReLU(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode is None or NndctOption.nndct_quant_off.value: return torch.nn.LeakyReLU(*args, **kwargs) return deephi_LeakyReLU(*args, **kwargs)
null
23,552
import torch from torch.autograd import Variable from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_MaxPool1d(torch.nn.modules.MaxPool1d): r"""DeePhi Conv1d operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_MaxPool1d, self).__init__(*args, **kwards) self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() def forward(self, input): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] return output def MaxPool1d(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.MaxPool1d(*args, **kwargs) return deephi_MaxPool1d(*args, **kwargs)
null
23,553
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_Interpolate(torch.nn.Module): def __init__(self, *args, **kwards): super(deephi_Interpolate, self).__init__(*args, **kwards) self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() def forward(self, input, size=None, scale_factor=None, mode='nearest', align_corners=None): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = torch.nn.functional.interpolate(qinput, size, scale_factor, mode, align_corners) output = quantize_tensors([output], self.node)[0] return output def Interpolate(*args, **kwargs): return deephi_Interpolate(*args, **kwargs)
null
23,554
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption, NndctScreenLogger, QError, QWarning from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from .quant_noise import eval_qnoise import pytorch_nndct.utils as py_utils import torch.nn.functional as F from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version class deephi_ConvTranspose2d(torch.nn.modules.conv.ConvTranspose2d): r"""DeePhi ConvTranspose2d operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_ConvTranspose2d, self).__init__(*args, **kwards) self.params_name = None self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() self.param_saved = False self.param_quantized = False # self.weight and self.bias are not quantized float parameters self.weight_bak = None # backup of float bias for bias correction self.bias_bak = None # backup of float bias for bias correction self.efficency = 0.0 self.deviation = 0.0 self.stop = False self.rate = NndctOption.nndct_param_corr_rate.value def forward(self, input): # backup bias for bias correction feature if (not self.param_saved): if NndctOption.nndct_param_corr.value > 0: # backup orignal float parameters if self.quant_mode == 1: self.weight_bak = self.weight.detach().clone() if self.bias is not None: self.bias_bak = self.bias.detach().clone() # adjust bias if self.quant_mode == 2 and self.bias is not None: if not self.quantizer.has_bias_corr(self.node): NndctScreenLogger().error2user(QError.BIAS_CORRECTION, f"Bias correction file in quantization result directory does not match current model.") exit(2) self.bias.data = torch.sub(self.bias.data, torch.tensor( self.quantizer.get_bias_corr(self.node), device=self.bias.data.device, dtype=self.bias.data.dtype)) self.param_saved = True # quantize parameters qweight = None qbias = None inplace = (NndctOption.nndct_quant_off.value or self.quantizer is not None and self.quantizer.inplace) if (not self.param_quantized): if inplace: _ = quantize_tensors( [self.weight], self.node, tensor_names = [self.params_name[0]], tensor_type = 'param')[0] qweight = self.weight if self.bias is not None: _ = quantize_tensors( [self.bias], self.node, tensor_names = [self.params_name[1]], tensor_type = 'param')[0] qbias = self.bias else: qweight = quantize_tensors( [self.weight], self.node, tensor_names = [self.params_name[0]], tensor_type = 'param')[0] if self.bias is not None: qbias = quantize_tensors( [self.bias], self.node, tensor_names = [self.params_name[1]], tensor_type = 'param')[0] if not NndctOption.nndct_quant_off.value: self.param_quantized = True else: qweight = self.weight qbias = self.bias # quantize input tensor qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if compare_torch_version(CmpFlag.LESS_EQUAL, "1.11.0"): output_padding = self._output_padding(qinput, None, self.stride, self.padding, self.kernel_size) else: num_spatial_dims = 2 output_padding = self._output_padding( qinput, None, self.stride, self.padding, self.kernel_size, num_spatial_dims, self.dilation) output = torch.nn.functional.conv_transpose2d(qinput, weight = qweight, bias = qbias, stride = self.stride, padding = self.padding, output_padding = output_padding, dilation = self.dilation, groups = self.groups) output = quantize_tensors([output], self.node)[0] # correct weights and bias in calibation if NndctOption.nndct_param_corr.value > 0: #rate = NndctOption.nndct_param_corr_rate.value # statistic of quantization error if (self.quant_mode == 1 and not self.stop): if compare_torch_version(CmpFlag.LESS_EQUAL, "1.11.0"): output_padding = self._output_padding(input, None, self.stride, self.padding, self.kernel_size) else: num_spatial_dims = 2 output_padding = self._output_padding(input, None, self.stride, self.padding, self.kernel_size, num_spatial_dims, self.dilation) res_f = torch.nn.functional.conv_transpose2d(input, self.weight_bak, bias = self.bias_bak, stride = self.stride, padding = self.padding, output_padding = output_padding, dilation = self.dilation, groups = self.groups) error, rate, self.stop, self.efficency, self.deviation = eval_qnoise( output, res_f, self.efficency, self.deviation, self.rate, self.stop) if (not self.stop) and (self.bias is not None): error = error.mean(dim = [0, 2, 3]) self.bias.data = torch.sub(self.bias.data, error, alpha=rate) self.param_quantized = False return output def bias_corr(self): if self.bias is not None and self.bias_bak is not None: bias_err = torch.sub(self.bias_bak, self.bias.data) return bias_err.cpu().numpy().tolist() else: return None def ConvTranspose2d(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.ConvTranspose2d(*args, **kwargs) return deephi_ConvTranspose2d(*args, **kwargs)
null
23,555
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from .quant_noise import eval_qnoise import pytorch_nndct.utils as py_utils import torch.nn.functional as F class deephi_PReLU(torch.nn.PReLU): def __init__(self, *args, **kwards): def forward(self, input): def PReLU(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.PReLU(*args, **kwargs) return deephi_PReLU(*args, **kwargs)
null
23,556
import torch from torch.autograd import Variable from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.utils import NndctOption from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_MaxPool2d(torch.nn.modules.MaxPool2d): r"""DeePhi Conv2d operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_MaxPool2d, self).__init__(*args, **kwards) self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() def forward(self, input): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] return output def MaxPool2d(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.MaxPool2d(*args, **kwargs) return deephi_MaxPool2d(*args, **kwargs)
null
23,557
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Int(_PrimModule): def __init__(self): super().__init__() def forward(self, input): output = int(input) return output def Int(*args, **kwargs): return deephi_Int(*args, **kwargs)
null
23,558
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_QuantInput(_PrimModule): def __init__(self): super().__init__() def forward(self, input): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = qinput if NndctOption.nndct_stat.value > 2: print('Channel number of input data: {}'.format(output.shape[1])) print('Input data histogram: {}'.format( output.histc(bins = 10).cpu().detach().numpy() )) print('Network input channel-wise statistic [Min, Max, Mean, Std]:') t = output.transpose(0, 1) for c in range(t.shape[0]): print('[{}, {}, {}, {}]'.format( t[c].min(), t[c].max(), t[c].float().mean(), t[c].float().std() )) print('histogram: {}'.format( t[c].histc(bins = 10).cpu().detach().numpy() )) output = quantize_tensors([output], self.node)[0] return output def quant_input(*args, **kwargs): return deephi_QuantInput(*args, **kwargs)
null
23,559
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_DequantOutput(_PrimModule): def __init__(self): def forward(self, input): def dequant_output(*args, **kwargs): return deephi_DequantOutput(*args, **kwargs)
null
23,560
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Input(_PrimModule): def __init__(self): def forward(self, input): def Input(*args, **kwargs): return deephi_Input(*args, **kwargs)
null
23,561
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_StridedSlice(_PrimModule): def __init__(self): super().__init__() def forward(self, input, dim, start, end, step): size = input.size() break_symbol = ':' symbols = "" start_symbol = [] end_symbol = [] step_symbol = [] for i in range(dim[0]): start_symbol.append(str(0)) end_symbol.append(str(int(size[i]))) step_symbol.append(str(1)) for i in range(len(start)): start_symbol.append(str(int(start[i]))) end_symbol.append(str(int(end[i]))) step_symbol.append(str(int(step[i]))) for i in range(len(start_symbol)): slice_symbol = break_symbol.join([start_symbol[i], end_symbol[i], step_symbol[i]]) if i > 0: symbols += "," + slice_symbol else: symbols = slice_symbol eval_str = f"input[{symbols}]" output = eval(eval_str) output = quantize_tensors([output], self.node)[0] return output def strided_slice(*args, **kwargs): return deephi_StridedSlice(*args, **kwargs)
null
23,562
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_SliceInplaceCopy(_PrimModule): def __init__(self): super().__init__() def forward(self, input, source, dim, index): index = torch.tensor([index]).to(input.device) if input.dtype != source.dtype: source = source.to(input.dtype) output = input.index_copy_(dim, index, source.unsqueeze(dim)) output = quantize_tensors([output], self.node)[0] return output def slice_tensor_inplace_copy(*args, **kwargs): return deephi_SliceInplaceCopy(*args, **kwargs)
null
23,563
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Index(_PrimModule): def __init__(self): super().__init__() def forward(self, input, index): if isinstance(index, (list, tuple)): break_symbol = ':' symbols = "" for i in range(len(index)): if index[i] == None: slice_symbol = break_symbol else: slice_symbol = "index[" + str(i) + "]" if i > 0: symbols += "," + slice_symbol else: symbols = slice_symbol eval_str = f"input[{symbols}]" output = eval(eval_str) output = quantize_tensors([output], self.node)[0] else: output = input[index] return output def Index(*args, **kwargs): return deephi_Index(*args, **kwargs)
null
23,564
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_IndexInputInplace(_PrimModule): def __init__(self): def forward(self, input, indices, values, accumulate): def index_put_(*args, **kwargs): return deephi_IndexInputInplace(*args, **kwargs)
null
23,565
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_ReLUk(_PrimModule): def __init__(self): def forward(self, input:torch.Tensor, channel_max:Union[torch.Tensor, Sequence[Any], float]): def Reluk(*args, **kwargs): return deephi_ReLUk(*args, **kwargs)
null
23,566
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_ChannelScale(_PrimModule): def __init__(self): super().__init__() def forward(self, input:torch.Tensor, channel_scale:Union[torch.Tensor, Sequence[Any], float]): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if isinstance(channel_scale, (list, tuple)): channel_scale = torch.Tensor(channel_scale).to(input.device) elif isinstance(channel_scale, float): channel_scale = torch.Tensor([channel_scale]).to(input.device) ''' if self.node.in_quant_part: channel_scale = quant_channel_scale_params(self.node, channel_scale) ''' output = qinput * channel_scale if self.node.in_quant_part: output = quantize_tensors([output], self.node)[0] return output def Channel_Scale(*args, **kwargs): return deephi_ChannelScale(*args, **kwargs)
null
23,567
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_ExpandAs(_PrimModule): def __init__(self): super().__init__() def forward(self, input, other): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = qinput.expand_as(other).clone() output = quantize_tensors([output], self.node)[0] return output def expand_as(*args, **kwargs): return deephi_ExpandAs(*args, **kwargs)
null
23,568
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Expand(_PrimModule): def __init__(self): super().__init__() def forward(self, input, size): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = qinput.expand(size).clone() output = quantize_tensors([output], self.node)[0] return output def expand(*args, **kwargs): return deephi_Expand(*args, **kwargs)
null
23,569
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Correlation1D_Elemwise(_PrimModule): def __init__(self): super().__init__() def forward(self, input_1:torch.Tensor, input_2:torch.Tensor, pad_size:Union[torch.Tensor, Sequence[Any], int]): qinput_1 = quantize_tensors([input_1], self.node, tensor_type='input')[0] qinput_2 = quantize_tensors([input_2], self.node, tensor_type='input')[0] if isinstance(pad_size, (list, tuple)): pad_size = torch.Tensor(pad_size).to(qinput_1.device) elif isinstance(pad_size, float): pad_size = torch.Tensor([pad_size]).to(qinput_1.device) output_dim = pad_size + 1 B, C, H, W = qinput_1.size() qinput_2 = F.pad(qinput_2, pad=(pad_size,0,0,0), mode="constant",value=0) cv = [] for i in range(output_dim - 1, -1, -1): cost = qinput_1 * qinput_2[:, :, :, i:(i + W)] cost = cost.unsqueeze(2) cv.append(cost) output = torch.cat(cv, 2) if self.node.in_quant_part: output = quantize_tensors([output], self.node)[0] return output def Correlation1d_Elemwise(*args, **kwargs): return deephi_Correlation1D_Elemwise(*args, **kwargs)
null
23,570
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Correlation2D_Elemwise(_PrimModule): def __init__(self): super().__init__() def forward(self, input_1:torch.Tensor, input_2:torch.Tensor, pad_size:Union[torch.Tensor, Sequence[Any], int]): qinput_1 = quantize_tensors([input_1], self.node, tensor_type='input')[0] qinput_2 = quantize_tensors([input_2], self.node, tensor_type='input')[0] if isinstance(pad_size, (list, tuple)): pad_size = torch.Tensor(pad_size).to(qinput_1.device) elif isinstance(pad_size, float): pad_size = torch.Tensor([pad_size]).to(qinput_1.device) output_dim = 2 * pad_size + 1 B, C, H, W = qinput_1.size() qinput_2 = F.pad(qinput_2, [pad_size] * 4) cv = [] for i in range(output_dim): for j in range(output_dim): cost = qinput_1 * qinput_2[:, :, i:(i + H), j:(j + W)] cost = cost.unsqueeze(2) cv.append(cost) output = torch.cat(cv, 2) if self.node.in_quant_part: output = quantize_tensors([output], self.node)[0] return output def Correlation2d_Elemwise(*args, **kwargs): return deephi_Correlation2D_Elemwise(*args, **kwargs)
null
23,571
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_CostVolume(_PrimModule): def __init__(self): def forward(self, input_1:torch.Tensor, input_2:torch.Tensor, maxdisp:Union[torch.Tensor, Sequence[Any], int]): def CostVolume(*args, **kwargs): return deephi_CostVolume(*args, **kwargs)
null
23,572
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_TupleUnpack(_PrimModule): def __init__(self): def forward(self, input): def TupleUnpack(*args, **kwargs): return deephi_TupleUnpack(*args, **kwargs)
null
23,573
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Sub(_PrimModule): def __init__(self, reverse=False): super().__init__() self.reverse = reverse def forward(self, input, other, alpha=1): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if self.reverse is True: output = -(qinput - alpha * other) else: output = qinput - alpha * other output = quantize_tensors([output], self.node)[0] return output def Sub(*args, **kwargs): return deephi_Sub()
null
23,574
import os import torch import torch.nn.functional as F from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.base import NNDCT_CONSTANT from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import quant_reluk_params from nndct_shared.quantization import quant_channel_scale_params import pytorch_nndct.utils as py_utils from typing import Any, Optional, Sequence, Union from torch.autograd import Variable class deephi_Sub(_PrimModule): def __init__(self, reverse=False): super().__init__() self.reverse = reverse def forward(self, input, other, alpha=1): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if self.reverse is True: output = -(qinput - alpha * other) else: output = qinput - alpha * other output = quantize_tensors([output], self.node)[0] return output def Rsub(*args, **kwargs): return deephi_Sub(reverse=True)
null
23,575
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_Mul(torch.nn.Module): def __init__(self): super(deephi_Mul, self).__init__() self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input, other): [qinput, qother] = quantize_tensors( [input, other], self.node, tensor_type='input') output = torch.mul(qinput, qother) output = quantize_tensors([output], self.node)[0] return output def Mul(*args, **kwargs): return deephi_Mul(*args, **kwargs)
null
23,576
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.utils import NndctOption import pytorch_nndct.utils as py_utils from nndct_shared.utils import calculate_op_scale class deephi_Mean(torch.nn.Module): r"""DeePhi Concat operation""" def __init__(self, *args, **kwargs): super(deephi_Mean, self).__init__() self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input, dim, keepdim): if self.quantizer is None or self.quant_mode is None or NndctOption.nndct_quant_off.value: return self._fp32_forward(input, dim, keepdim) else: return self._forward(input, dim, keepdim) def _fp32_forward(self, input, dim, keepdim): if dim is None: output = torch.mean(input) else: output = torch.mean(input, dim, keepdim) return output def _forward(self, input, dim, keepdim): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if dim is None: output = torch.mean(qinput) else: output = torch.mean(qinput, dim, keepdim) input_shape = self.node.in_tensors[0].shape if self.node.node_attr(self.node.op.AttrName.DIMS) == [None]: dim_list = [i for i in range(len(input_shape))] else: dim_list = self.node.node_attr(self.node.op.AttrName.DIMS) rec = 1 for i in dim_list: input_shape = input.shape[i].item() if isinstance(input.shape[i], torch.Tensor) else input.shape[i] rec = rec * input_shape if (rec & (rec - 1)) != 0: scale = calculate_op_scale(rec, self.node) output = output * scale output = quantize_tensors([output], self.node)[0] return output def Mean(*args, **kwargs): return deephi_Mean(*args, **kwargs)
null
23,577
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_Cat(torch.nn.Module): def __init__(self, *args, **kwargs): def forward(self, tensors, dim): def Cat(*args, **kwargs): return deephi_Cat(*args, **kwargs)
null
23,578
import numpy as np def mysigmoid(x): s = 1 / (1 + np.exp(-x)) return s
null
23,579
import numpy as np def mapping_sigm(x, data, shift): scale = 1.0 / 2 ** 15 inv_step = 2 ** shift def __ele_map(x_ele): scale = 2 ** -15 if x_ele >= 8: return 1.0 - scale elif x_ele < -8: return 0.0 else: x_ele = int(x_ele * inv_step) if x_ele >=0: if shift >= 7: pos = (x_ele >> (shift - 7)) % 1024 else: pos = (x_ele << (7 - shift)) % 1024 return data[pos+1024] * scale else: if shift >= 7: pos = (abs(x_ele) >> (shift - 7)) % 1024 else: pos = (abs(x_ele) << (7 - shift)) % 1024 if x_ele >> shift == -8 and pos == 0: return 0.0 else: return data[1024-pos] * scale return np.array([[__ele_map(c) for c in row] for row in x], dtype=np.float32)
null
23,580
import numpy as np def mapping_tanh(x, data, shift): scale = 1.0 / 2 ** 15 inv_step = 2 ** shift def __ele_map(x_ele): if x_ele >= 4: return 1.0 - scale elif x_ele < -4: return -1.0 else: x_ele = int(x_ele * inv_step) if x_ele >= 0: if shift >= 8: pos = (x_ele >> (shift - 8)) % 1024 else: pos = (x_ele << (8 - shift)) % 1024 return data[pos + 1024] * scale else: if shift >= 8: pos = (abs(x_ele) >> (shift - 8)) % 1024 else: pos = (abs(x_ele) << (8 - shift)) % 1024 if x_ele >> shift == -4 and pos == 0: return data[pos] * scale else: return data[1024-pos] * scale return np.array([[__ele_map(c) for c in row] for row in x], dtype=np.float32)
null
23,581
import numpy as np def absolute_shift(x, pos, to='left', bitwidth=16): res = 0 if to == 'left': if pos >= 0: res = np.left_shift(x, pos) else: res = np.right_shift(x, -pos) elif to == 'right': if pos >= 0: res = np.right_shift(x, pos) else: res = np.left_shift(x, -pos) else: raise TypeError("shift to {} is not expected".format(to)) res = np.where(res > (2**(bitwidth - 1) - 1), 2**(bitwidth - 1) - 1, res) res = np.where(res < -2**(bitwidth - 1), -2**(bitwidth - 1), res) return res
null
23,582
import numpy as np def absolute_shift_round(x, pos, to='left', bitwidth=16): res = 0 if to == 'left': if pos >= 0: #res = np.left_shift(x, pos) res = x * (2**pos) else: #res = np.right_shift(x, -pos) res = x * (2**(-pos)) elif to == 'right': if pos >= 0: #res = np.right_shift(x, pos) res = np.round(x / (2**pos)).astype(np.int32) else: res = np.round(x / (2**(-pos))).astype(np.int32) else: raise TypeError("shift to {} is not expected".format(to)) res = np.where(res > (2**(bitwidth - 1) - 1), 2**(bitwidth - 1) - 1, res) res = np.where(res < -2**(bitwidth - 1), -2**(bitwidth - 1), res) return res
null
23,583
from typing import List, Tuple import torch from torch import Tensor from torch.nn.utils.rnn import (PackedSequence, pack_padded_sequence, pad_packed_sequence) from .rnn_cell import LSTMCell from .rnn_layer import LSTMLayer, QuantGruLayer, QuantLstmLayer def init_stacked_lstm(num_layers, layer, first_layer_args, other_layer_args): layers = [layer(*first_layer_args) ] + [layer(*other_layer_args) for _ in range(num_layers - 1)] return torch.nn.ModuleList(layers)
null
23,584
from typing import List, Tuple import torch from torch import Tensor from torch.nn.utils.rnn import (PackedSequence, pack_padded_sequence, pad_packed_sequence) from .rnn_cell import LSTMCell from .rnn_layer import LSTMLayer, QuantGruLayer, QuantLstmLayer class LSTM(torch.nn.Module): def __init__(self, num_layers, layer, first_layer_args, other_layer_args): def forward( self, input: Tensor, states: Tuple[Tensor, Tensor] ) -> Tuple[Tensor, Tuple[Tensor, Tensor]]: def get_normal_lstm_param_name(name): class LSTMCell(torch.nn.Module): def __init__(self, input_size: int, hidden_size: int, bias: bool = True) -> None: def init_weight(self): def forward(self, input: Tensor, state: Tuple[Tensor, Tensor]) -> Tuple[Tensor, Tuple[Tensor, Tensor]]: class LSTMLayer(torch.nn.Module): def __init__(self, cell, input_size: int, hidden_size: int, bias: bool = True): def forward(self, input: Tensor, state: Tuple[Tensor, Tensor]) -> Tuple[Tensor, Tuple[Tensor, Tensor]]: def stacked_lstm(lstm_module: torch.nn.LSTM) -> LSTM: num_layers = lstm_module.num_layers input_size = lstm_module.input_size hidden_size = lstm_module.hidden_size bias = lstm_module.bias stacked_lstm = LSTM( num_layers, LSTMLayer, first_layer_args=[LSTMCell, input_size, hidden_size, bias], other_layer_args=[LSTMCell, hidden_size, hidden_size, bias]) param_num_per_layer = len(list(lstm_module.parameters())) // num_layers param_layer_list = [] for i, (name, param) in enumerate(lstm_module.named_parameters()): if i % param_num_per_layer == 0: param_layer_list.append({}) param_layer_list[-1][get_normal_lstm_param_name(name)] = param for layer_idx in range(num_layers): for name, param in stacked_lstm.layers[layer_idx].named_parameters(): normalized_name = get_normal_lstm_param_name(name) if normalized_name is not None and normalized_name in param_layer_list[layer_idx]: param.data.copy_(param_layer_list[layer_idx][normalized_name].data) stacked_lstm.layers[layer_idx].cell.init_weight() return stacked_lstm
null
23,585
import torch from torch.autograd import Variable import torch.nn.functional as F from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.utils import NndctOption import pytorch_nndct.utils as py_utils from .fix_ops import fake_quantize_per_tensor from nndct_shared.utils import NNDCT_KEYS, GLOBAL_MAP class deephi_Hardswish(torch.nn.Module): def __init__(self, inplace=False, *args, **kwards): def forward(self, input): def Hardswish(*args, **kwargs): return deephi_Hardswish(*args, **kwargs)
null
23,586
import torch from nndct_shared.utils import NndctOption from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_Embedding(torch.nn.modules.sparse.Embedding): r"""DeePhi transpose operation, support float and double""" def __init__(self, *args, **kwargs): super(deephi_Embedding, self).__init__(*args, **kwargs) self.params_name = None self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() self.param_quantized = False def forward(self, input): if self.quant_mode <= 0 or (not self.node.in_quant_part): return torch.nn.functional.embedding(input, self.weight, padding_idx=self.padding_idx) inplace = (NndctOption.nndct_quant_off.value or self.quantizer is not None and self.quantizer.inplace) # params = [] # qparams = [] # param_names = [] # for k in self.node.op.params.keys(): # pname = self.node.op.params[k].name # p = getattr(self.quantizer.quant_model, k.value) # param_names.append(pname) # params.append(p) if not self.param_quantized: if inplace: _ = quantize_tensors([self.weight], self.node, tensor_names= [self.params_name[0]], tensor_type='param') qparams = [self.weight] else: qparams = quantize_tensors([self.weight], self.node, tensor_names= [self.params_name[0]], tensor_type='param') if not NndctOption.nndct_quant_off.value: self.param_quantized = True else: qparams = [self.weight] inputs = quantize_tensors([input], self.node, tensor_type='input') output = torch.nn.functional.embedding(inputs[0], qparams[0], padding_idx=self.padding_idx) output = quantize_tensors([output], self.node)[0] return output def embedding(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.Embedding(*args, **kwargs) return deephi_Embedding(*args, **kwargs)
null
23,587
import math import torch import numpy as np from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import kernel_need_quant from nndct_shared.quantization import quantize_tensors from nndct_shared.utils import NndctOption from .fix_ops import NndctSoftmaxExpApproximate, NndctSoftmaxLOD, NndctSoftmaxSimulationPart1, NndctSoftmaxSimulationPart2, NndctExpApprAIE2, NndctInverseAIE2 import pytorch_nndct.utils as py_utils from pytorch_nndct.nn.nonlinear import approx class deephi_Softmax(torch.nn.modules.Softmax): r"""DeePhi Softmax operation""" def __init__(self, dim = None): super(deephi_Softmax, self).__init__() self.dim = dim self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): if (not kernel_need_quant(self.quantizer, self.node) or self.quantizer.exporting) or NndctOption.nndct_gemm88.value: # Method 0: quant input and output output = super().forward(input) output = quantize_tensors([output], self.node)[0] else: # quantize input tensor qinput = quantize_tensors([input], self.node, tensor_type='input')[0] input_name = self.node.in_nodes[0] input_node = self.quantizer.configer.get_Nndctnode(input_name) if not self.quantizer.configer.node_output_quantizable(input_node): input_name = input_node.in_nodes[0] # Method 1: Hardware PL Softmax with 8 bw if NndctOption.nndct_op_softmax_mode.value == "hardware_pl": x_max = torch.max(qinput, dim = self.dim, keepdim = True).values Exp_sum_appr = 0.0 uvi = 47274 / math.pow(2,15) * (qinput - x_max) exp_appr = torch.empty_like(uvi) NndctSoftmaxExpApproximate(uvi, exp_appr) exp_appr = torch.round(exp_appr*10**5) exp_appr = exp_appr/(10**5) Exp_sum_appr = torch.sum(exp_appr, dim = self.dim, keepdim = True) F = Exp_sum_appr w = torch.empty_like(F) NndctSoftmaxLOD(F, w) m = F/(2**w) lnF = torch.round((22713/(2**15))*(m-1+w)*10**5)/10**5 uvi = 47274 / (2**15) * (qinput - x_max - lnF) exp_appr = torch.empty_like(uvi) NndctSoftmaxExpApproximate(uvi, exp_appr) exp_appr = torch.round(exp_appr*10**5)/10**5 output = exp_appr output = quantize_tensors([output], self.node)[0] # Method 2: Liyi Softmax with any bw elif NndctOption.nndct_op_softmax_mode.value == "liyi": x_max = torch.max(qinput, dim = self.dim, keepdim = True).values qinput = qinput - x_max exp_appr = torch.empty_like(qinput) NndctSoftmaxSimulationPart1(qinput, exp_appr) sum = torch.sum(exp_appr, dim = self.dim, keepdim = True) sum1 = torch.empty_like(sum) NndctSoftmaxSimulationPart2(sum, sum1) output = (exp_appr*sum1).bfloat16().float() output = quantize_tensors([output], self.node)[0] # Method 3: Table Look up for AIE2 Softmax with 8 bw and 16 bw(based on LUT) elif NndctOption.nndct_op_softmax_mode.value == "aie2_lut_16bw": bw = self.quantizer.get_quant_config(self.node.name, False)[0] fragpos = self.quantizer.get_quant_config(input_name, False)[1] if bw == 8 and fragpos < 2: if fragpos < 2: qinput_max = torch.max(qinput, dim = self.dim, keepdim = True).values qinput -= qinput_max else: qinput -= 31.75 else: # bw == 16 if fragpos < 10: qinput_max = torch.max(qinput, dim = self.dim, keepdim = True).values qinput -= qinput_max else: qinput -= 32 exp_appr = torch.empty_like(qinput) NndctExpApprAIE2(qinput, exp_appr, bw) sum = torch.sum(exp_appr, dim = self.dim, keepdim = True) sum_inv = torch.empty_like(sum) NndctInverseAIE2(sum, sum_inv) output = (exp_appr*sum_inv).bfloat16().float() output = quantize_tensors([output], self.node)[0] # Method 4: Bert with 8 bw elif NndctOption.nndct_op_softmax_mode.value == "bert_8bw" or NndctOption.nndct_ip_v70_bert.value: # def generate_exp_table(bw, input_scale, table_name): # from bfloat16 import bfloat16 # inputs_table_positive = np.arange(0, 2**(bw - 1)).astype(np.float32) # inputs_table_negatice = np.arange(-(2**(bw - 1)), 0).astype(np.float32) # inputs_table = np.hstack((inputs_table_positive, inputs_table_negatice)) # outputs_table = np.exp(inputs_table / (2**(input_scale))).astype(bfloat16).astype(np.float32) # outputs_table[128] = 0 # return outputs_table def compute_inv(x): exp_mask = 0x7F800000 mantissa_mask= 0x007FFFFF mantissa_Q = 0x00008000 x_f = x.astype(np.float32) B_x = x_f.view(np.uint32) exponent = (B_x & exp_mask)>>23 mantissa = np.where(((B_x & mantissa_Q)==0), ((B_x & mantissa_mask)>>16), ((B_x & mantissa_mask)>>16)+1) inv_exponent = 253-exponent inv_mantissa = np.round(256*128/(128+mantissa)-128) inv_x_val = (np.int32(inv_exponent)<<23) + (np.int32(inv_mantissa)<<16) inv_x = inv_x_val.view(np.float32) return inv_x bw = self.quantizer.get_quant_config(self.node.name, False)[0] fragpos = self.quantizer.get_quant_config(input_name, False)[1] exp_appr = torch.exp(qinput).bfloat16().float() exp_appr = torch.where(qinput > -128/(2 ** fragpos), exp_appr, 0) # # Code for Golden Verification # exp_table = generate_exp_table(bw, fragpos, self.node.name) # qinput_int = torch.where(qinput >=0, qinput * (2**(fragpos)), qinput * (2**(fragpos)) + 2**bw) # exp_arr = exp_table[qinput_int.cpu().numpy().astype(np.int8)] # exp_appr = torch.from_numpy(exp_arr).to(input.device) if self.dim == -1: sum_v32 = exp_appr.reshape((exp_appr.shape[0],exp_appr.shape[1],exp_appr.shape[2],exp_appr.shape[3]//16,16)) sum_v32 = sum_v32.permute((0,1,2,4,3)) sum_v32 = sum_v32.sum(dim = -1) # sum_v32 = sum_v32.sum(dim = -1, keepdim = True) sum_v32_shape = sum_v32.shape sum_v32 = sum_v32.reshape(-1, 16) sum_v32 = py_utils.aie_add_v16(sum_v32) sum_v32_shape = [int(x) for x in sum_v32_shape][:-1] + [1] sum_v32 = sum_v32.reshape(sum_v32_shape) else: sum_v32 = torch.sum(exp_appr, dim = self.dim, keepdim = True).cpu().numpy() sum_inv = compute_inv(sum_v32.cpu().numpy()) sum_inv = torch.from_numpy(sum_inv).bfloat16().float().to(qinput.device) output = (exp_appr*sum_inv).bfloat16().float() output = quantize_tensors([output], self.node, method=4)[0] # Method 5: ipu 8bw elif NndctOption.nndct_op_softmax_mode.value == "ipu_8bw": output = approx.softmax_approx_poly(qinput) # bfloat16, default settings: axis=-1, exp_table_size=1, degree=3 output = output.float() # float32 output = quantize_tensors([output], self.node, method=4)[0] # round: floor # Method 0: Quant input and output of Softmax else: output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] return output def Softmax(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.Softmax(*args, **kwargs) return deephi_Softmax(*args, **kwargs)
null
23,588
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from .quant_noise import eval_qnoise import pytorch_nndct.utils as py_utils class deephi_clamp(torch.nn.Module): r"""DeePhi clamp operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_clamp, self).__init__(*args, **kwards) self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input, min=None, max=None): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] qmin, qmax = None, None if min is not None: min = torch.tensor(min).to(input.device) qmin = quantize_tensors([min], self.node, tensor_names=[self.params_name[0]], tensor_type='param')[0] if max is not None: max = torch.tensor(max).to(input.device) qmax = quantize_tensors([max], self.node, tensor_names=[self.params_name[1]], tensor_type='param')[0] output = torch.clamp(input=qinput, min=qmin, max=qmax) output = quantize_tensors([output], self.node)[0] return output # # quantize parameters # qweight = None # inplace = (NndctOption.nndct_quant_off.value or # self.quantizer is not None and self.quantizer.inplace) # # quantize input tensor # qinput = quantize_tensors([input], self.node, tensor_type='input')[0] # # output = torch.clamp(input=qinput, min=min, max=max) # qmin, qmax = quantize_tensors(torch.tensor([min, max]), self.node, tensor_type='input') # output = torch.clamp(input=qinput, min=qmin, max=qmax) # output = quantize_tensors([output], self.node)[0] # return output def Clamp(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.clamp(*args, **kwargs) return deephi_clamp(*args, **kwargs)
null
23,589
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.utils import NndctOption, NndctScreenLogger, QWarning from nndct_shared.quantization import kernel_need_quant from nndct_shared.quantization import quantize_tensors import numpy as np import pytorch_nndct.utils as py_utils from .fix_ops import NndctAIESqrt class deephi_sqrt(torch.nn.Module): r"""DeePhi sqrt operation""" def __init__(self, *args, **kwargs): super(deephi_sqrt, self).__init__(*args, **kwargs) self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): # ensure input>=0 if torch.nonzero(input<0, as_tuple=False).numel() > 0: NndctScreenLogger().warning2user(QWarning.TENSOR_NEGATIVE, f"Elements in input tensor of node {self.node.name} are negative, which is not permittable for 'sqrt' operation. The negative numbers have been replaced by zero.") zero = torch.zeros_like(input) input = torch.where(input<0, zero, input) if not kernel_need_quant(self.quantizer, self.node): # not quantizable output = torch.sqrt(input) output = quantize_tensors([output], self.node)[0] return output # quantize input tensor qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if NndctOption.nndct_op_sqrt_mode.value == "ipu_8bw": # sqrt=x*(1/sqrt(x)): cuda/cpu output = torch.empty_like(qinput) NndctAIESqrt(qinput, output) # float32 # quantize output output = quantize_tensors([output], self.node, method=4)[0] else: output = torch.sqrt(qinput) output = quantize_tensors([output], self.node)[0] return output def sqrt(*args, **kwargs): return deephi_sqrt(*args, **kwargs)
null
23,590
import torch from nndct_shared.quantization.utils import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_Matmul(torch.nn.Module): def __init__(self): def forward(self, input, other): def Matmul(*args, **kwargs): return deephi_Matmul(*args, **kwargs)
null
23,591
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False import torch from torch.utils.cpp_extension import load, _import_module_from_library The provided code snippet includes necessary dependencies for implementing the `NndctFixNeuron` function. Write a Python function `def NndctFixNeuron(Tinput, Toutput, maxamp, method=2)` to solve the following problem: if Tinput.device == torch.device("cpu"): output = Tinput.cuda() nndct_kernels.FixNeuronV2(output, output, valmax, valamp, method) Tinput.copy_(output.cpu()) return Tinput # cpu fix neuron """ # output = Tinput.cpu().detach().numpy() # output = output * valamp # if method == 2: # output = np.where(output > valmax - 1, (valmax - 1), output) # output = np.where(output < (-valmax), -valmax, output) # output = np.where(np.logical_and(output > 0, np.logical_and(np.floor(output) % 2 == 0, output - np.floor(output) == 0.5)), np.ceil(output), output) # output = np.where(output >= 0, np.round(output), output) # output = np.where(np.logical_and(output < 0, output - np.floor(output) == 0.5), np.ceil(output), output) # output = np.where(output < 0, np.round(output), output) # elif method == 3: # output = np.where(output > valmax - 1, (valmax - 1), output) # output = np.where(output < (-valmax), -valmax, output) # output = np.where(np.logical_and(output > 0, np.logical_and(np.floor(output) % 2 == 0, output - np.floor(output) == 0.5)), np.ceil(output), output) # output = np.where(output >= 0, np.round(output), output) # output = np.where(np.logical_and(output < 0, np.logical_and(np.ceil(output) % 2 == 0, output - np.floor(output) == 0.5)), np.floor(output), output) # output = np.where(output < 0, np.round(output), output) # Tinput.copy_(torch.from_numpy(output)) # Tinput.div_(valamp) # return Tinput """ else: nndct_kernels.FixNeuronV2(Tinput, Toutput, valmax, valamp, method) return Toutput Here is the function: def NndctFixNeuron(Tinput, Toutput, maxamp, method=2): valmax, valamp = maxamp[0], maxamp[1] valmin = -valmax valmax = valmax - 1 device_id = 1 if Tinput.device == torch.device("cpu") else 0 if support_onnx_export(): Toutput = torch.ops.vai.fix_neuron(Tinput, valmin, valmax, valamp, 0, method, device_id, 1) else: nndct_kernels.FixNeuronV2(Tinput, Toutput, valmin, valmax, valamp, 0, method, device_id) return Toutput ''' if Tinput.device == torch.device("cpu"): output = Tinput.cuda() nndct_kernels.FixNeuronV2(output, output, valmax, valamp, method) Tinput.copy_(output.cpu()) return Tinput # cpu fix neuron """ # output = Tinput.cpu().detach().numpy() # output = output * valamp # if method == 2: # output = np.where(output > valmax - 1, (valmax - 1), output) # output = np.where(output < (-valmax), -valmax, output) # output = np.where(np.logical_and(output > 0, np.logical_and(np.floor(output) % 2 == 0, output - np.floor(output) == 0.5)), np.ceil(output), output) # output = np.where(output >= 0, np.round(output), output) # output = np.where(np.logical_and(output < 0, output - np.floor(output) == 0.5), np.ceil(output), output) # output = np.where(output < 0, np.round(output), output) # elif method == 3: # output = np.where(output > valmax - 1, (valmax - 1), output) # output = np.where(output < (-valmax), -valmax, output) # output = np.where(np.logical_and(output > 0, np.logical_and(np.floor(output) % 2 == 0, output - np.floor(output) == 0.5)), np.ceil(output), output) # output = np.where(output >= 0, np.round(output), output) # output = np.where(np.logical_and(output < 0, np.logical_and(np.ceil(output) % 2 == 0, output - np.floor(output) == 0.5)), np.floor(output), output) # output = np.where(output < 0, np.round(output), output) # Tinput.copy_(torch.from_numpy(output)) # Tinput.div_(valamp) # return Tinput """ else: nndct_kernels.FixNeuronV2(Tinput, Toutput, valmax, valamp, method) return Toutput '''
if Tinput.device == torch.device("cpu"): output = Tinput.cuda() nndct_kernels.FixNeuronV2(output, output, valmax, valamp, method) Tinput.copy_(output.cpu()) return Tinput # cpu fix neuron """ # output = Tinput.cpu().detach().numpy() # output = output * valamp # if method == 2: # output = np.where(output > valmax - 1, (valmax - 1), output) # output = np.where(output < (-valmax), -valmax, output) # output = np.where(np.logical_and(output > 0, np.logical_and(np.floor(output) % 2 == 0, output - np.floor(output) == 0.5)), np.ceil(output), output) # output = np.where(output >= 0, np.round(output), output) # output = np.where(np.logical_and(output < 0, output - np.floor(output) == 0.5), np.ceil(output), output) # output = np.where(output < 0, np.round(output), output) # elif method == 3: # output = np.where(output > valmax - 1, (valmax - 1), output) # output = np.where(output < (-valmax), -valmax, output) # output = np.where(np.logical_and(output > 0, np.logical_and(np.floor(output) % 2 == 0, output - np.floor(output) == 0.5)), np.ceil(output), output) # output = np.where(output >= 0, np.round(output), output) # output = np.where(np.logical_and(output < 0, np.logical_and(np.ceil(output) % 2 == 0, output - np.floor(output) == 0.5)), np.floor(output), output) # output = np.where(output < 0, np.round(output), output) # Tinput.copy_(torch.from_numpy(output)) # Tinput.div_(valamp) # return Tinput """ else: nndct_kernels.FixNeuronV2(Tinput, Toutput, valmax, valamp, method) return Toutput
23,592
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor def diffs_fix_pos(input, bit_width, scope, method): # get max and min element in the tensor abs_max = 1 << (bit_width - 1) fix_lb = -abs_max - 0.5; fix_ub = abs_max - 0.5; x_max = torch.max(input) x_min = torch.min(input) # calculate step and fix pos based on max and min value step = torch.max(x_min / fix_lb, x_max / fix_ub); max_scale = torch.floor(torch.log2(1.0/step)) if step > sys.float_info.min else torch.tensor(18) # calculate step based on diffs final_scale = max_scale fixed_diff_min = sys.float_info.max if scope > 1: # avoid clone multiple times input = clone_view_tensor(input) for i in range(0, scope): scale = max_scale + i qinput = fake_quantize_per_tensor( input, pow(2.0, scale), 0, -abs_max, abs_max-1, method, 0) qinput = torch.sub(input, qinput) qinput = torch.pow(qinput, 2.0) diff = torch.sum(qinput).item() if diff < fixed_diff_min: final_scale = scale fixed_diff_min = diff return final_scale import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctDiffsFixPos(Tinput, Tbuffer, Tfixpos, bit_width=8, range=5, method=2): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if support_onnx_export(): torch.ops.vai.diffs_fix_pos(Tinput, Tbuffer, Tfixpos, bit_width, range, method, device_id) else: nndct_kernels.DiffsFixPos(Tinput, Tbuffer, Tfixpos, bit_width, range, method, device_id)
null
23,593
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def diffs_fix_pos(input, bit_width, scope, method): # get max and min element in the tensor abs_max = 1 << (bit_width - 1) fix_lb = -abs_max - 0.5; fix_ub = abs_max - 0.5; x_max = torch.max(input) x_min = torch.min(input) # calculate step and fix pos based on max and min value step = torch.max(x_min / fix_lb, x_max / fix_ub); max_scale = torch.floor(torch.log2(1.0/step)) if step > sys.float_info.min else torch.tensor(18) # calculate step based on diffs final_scale = max_scale fixed_diff_min = sys.float_info.max if scope > 1: # avoid clone multiple times input = clone_view_tensor(input) for i in range(0, scope): scale = max_scale + i qinput = fake_quantize_per_tensor( input, pow(2.0, scale), 0, -abs_max, abs_max-1, method, 0) qinput = torch.sub(input, qinput) qinput = torch.pow(qinput, 2.0) diff = torch.sum(qinput).item() if diff < fixed_diff_min: final_scale = scale fixed_diff_min = diff return final_scale import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctDiffsFixPosChannel(Tinput, Tbuffer, Tfixpos, axis, bit_width=8, scope=5, method=2): device_id = 1 if Tinput.device == torch.device("cpu") else 0 input_split = torch.split(Tinput, 1, dim=axis) buffer_split = torch.split(Tbuffer, 1, dim=axis) # TODO(@kewang): The split is a tensor view operation. Is it neccessary to clone tensor before calib and test ? if support_onnx_export(): for i in range(len(input_split)): torch.ops.vai.diffs_fix_pos(input_split[i], buffer_split[i], Tfixpos[i], bit_width, scope, method, device_id) else: for i in range(len(input_split)): nndct_kernels.DiffsFixPos(input_split[i], buffer_split[i], Tfixpos[i], bit_width, scope, method, device_id)
null
23,594
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctSigmoidTableLookup(Tinput, Ttable, Toutput, fragpos): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if support_onnx_export(): torch.ops.vai.SigmoidTableLookup(Tinput, Ttable, Toutput, fragpos, device_id) else: nndct_kernels.SigmoidTableLookup(Tinput, Ttable, Toutput, fragpos, device_id)
null
23,595
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctSigmoidSimulation(Tinput, Toutput, fragpos): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if device_id == 1: print("Sigmoid simulation does not support CPU") else: if support_onnx_export(): torch.ops.vai.SigmoidSimulation(Tinput, Toutput, fragpos, device_id) else: nndct_kernels.SigmoidSimulation(Tinput, Toutput, fragpos, device_id)
null
23,596
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctTanhTableLookup(Tinput, Ttable, Toutput, fragpos): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if support_onnx_export(): torch.ops.vai.TanhTableLookup(Tinput, Ttable, Toutput, fragpos, device_id) else: nndct_kernels.TanhTableLookup(Tinput, Ttable, Toutput, fragpos, device_id)
null
23,597
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctTanhSimulation(Tinput, Toutput, fragpos): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if device_id == 1: print("Tanh simulation does not support CPU") else: if support_onnx_export(): torch.ops.vai.TanhSimulation(Tinput, Toutput, fragpos, device_id) else: nndct_kernels.TanhSimulation(Tinput, Toutput, fragpos, device_id)
null
23,598
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctSoftmaxExpApproximate(Tinput, Toutput): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if device_id == 1: print("Softmax Exponent Approximate does not support CPU") else: if support_onnx_export(): torch.ops.vai.SoftmaxExpApproximate(Tinput, Toutput, device_id) else: nndct_kernels.SoftmaxExpApproximate(Tinput, Toutput, device_id)
null
23,599
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctSoftmaxLOD(Tinput, Toutput): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if device_id == 1: print("Softmax LOD does not support CPU") else: if support_onnx_export(): torch.ops.vai.SoftmaxLOD(Tinput, Toutput, device_id) else: nndct_kernels.SoftmaxLOD(Tinput, Toutput, device_id)
null
23,600
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctSoftmaxSimulationPart1(Tinput, Toutput): device_id = 1 if Tinput.device == torch.device("cpu") else 0 Tinput = clone_view_tensor(Tinput) if device_id == 1: print("Softmax Simulation Part 1 does not support CPU") else: if support_onnx_export(): torch.ops.vai.SoftmaxSimulationPart1(Tinput, Toutput, device_id) else: nndct_kernels.SoftmaxSimulationPart1(Tinput, Toutput, device_id)
null
23,601
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctSoftmaxSimulationPart2(sum, Toutput): device_id = 1 if Toutput.device == torch.device("cpu") else 0 sum = clone_view_tensor(sum) if device_id == 1: print("Softmax Simulation Part 2 does not support CPU") else: if support_onnx_export(): torch.ops.vai.SoftmaxSimulationPart2(sum, Toutput, device_id) else: nndct_kernels.SoftmaxSimulationPart2(sum, Toutput, device_id)
null
23,602
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def support_onnx_export(): if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): return True else: return False import sys import torch from torch.utils.cpp_extension import load, _import_module_from_library if compare_torch_version(CmpFlag.GREATER_EQUAL, "1.7.0"): new_kernel = True class CmpFlag(Enum): """ Enum for comparison flags """ EQUAL = 0 LESS = 1 LESS_EQUAL = 2 GREATER = 3 GREATER_EQUAL = 4 NOT_EQUAL = 5 def compare_torch_version(compare_type:CmpFlag, version:str): if compare_type == CmpFlag.EQUAL: return LooseVersion(torch.__version__) == LooseVersion(version) if compare_type == CmpFlag.LESS: return LooseVersion(torch.__version__) < LooseVersion(version) if compare_type == CmpFlag.LESS_EQUAL: return LooseVersion(torch.__version__) <= LooseVersion(version) if compare_type == CmpFlag.GREATER: return LooseVersion(torch.__version__) > LooseVersion(version) if compare_type == CmpFlag.GREATER_EQUAL: return LooseVersion(torch.__version__) >= LooseVersion(version) if compare_type == CmpFlag.NOT_EQUAL: return LooseVersion(torch.__version__) != LooseVersion(version) def fake_quantize_per_channel(input, scale_inv, zero_point, axis, quant_min, quant_max, method, inplace): if method == -1: if compare_torch_version(CmpFlag.GREATER, "0.9.0"): zero_point = zero_point.to(torch.int32) else: zero_point = zero_point.to(torch.long) return torch.fake_quantize_per_channel_affine(input, 1.0 / scale_inv, zero_point, axis, quant_min, quant_max) else: device_id = 1 if input.device == torch.device("cpu") else 0 if support_onnx_export(): scale = torch.where(scale_inv<sys.float_info.min, torch.tensor(sys.float_info.max, dtype=scale_inv.dtype, device=scale_inv.device), 1.0/scale_inv).to(torch.float) # api: (tensor(float), int32, int32, tensor(float), tensor(int8), int32, int32, int32, bool) output = torch.ops.vai.fix_neuron_per_channel(input, quant_min, quant_max, scale, zero_point.to(torch.int8), axis, method, device_id, inplace) return output else: input_split = torch.split(input, 1, dim=axis) output_cat = [] for i in range(len(input_split)): output_split = input_split[i].clone() if inplace == 0 else input_split[i] nndct_kernels.FixNeuronV2(input_split[i], output_split, quant_min, quant_max, scale_inv[i], zero_point[i], method, device_id) output_cat.append(output_split) output = torch.cat(output_cat, axis) return output
null
23,603
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version import torch from torch.utils.cpp_extension import load, _import_module_from_library def fake_quantize_per_channel_tensorrt(inputs, amax, min_bound, max_bound, axis=None): # Computation must be in FP32 to prevent potential over flow. if not isinstance(max_bound, torch.Tensor): max_bound = torch.tensor(float(max_bound)) #max_bound = max_bound.double() input_dtype = inputs.dtype if inputs.dtype == torch.half: inputs = inputs.float() min_amax = amax.min() if min_amax < 0: raise ValueError("Negative values in amax") scale = max_bound / amax epsilon = 1. / (1<<24) if min_amax <= epsilon: # Treat amax smaller than minimum representable of fp16 0 zero_amax_mask = (amax <= epsilon) scale[zero_amax_mask] = 0 # Value quantized with amax=0 should all be 0 if axis != None: for x_dim in range(inputs.ndim): if x_dim != axis: scale = torch.unsqueeze(scale, x_dim) outputs = torch.clamp((inputs * scale).round_(), min_bound, max_bound) if min_amax <= epsilon: scale[zero_amax_mask] = 1. # Return 1 makes more sense for values quantized to 0 with amax=0 if input_dtype == torch.half: outputs = outputs.half() outputs = outputs / scale return outputs
null
23,604
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version import torch from torch.utils.cpp_extension import load, _import_module_from_library def fake_quantize_per_tensor_tensorrt(inputs, amax, min_bound, max_bound): # Computation must be in FP32 to prevent potential over flow. if not isinstance(max_bound, torch.Tensor): max_bound = torch.tensor(float(max_bound)) #max_bound = max_bound.double() if not isinstance(amax, torch.Tensor): amax = torch.tensor(float(amax)) #amax = amax.double() input_dtype = inputs.dtype if inputs.dtype == torch.half: inputs = inputs.float() if amax < 0: raise ValueError("Negative values in amax") scale = max_bound / amax epsilon = 1. / (1<<24) if amax <= epsilon: # Treat amax smaller than minimum representable of fp16 0 scale = 0 # Value quantized with amax=0 should all be 0 outputs = torch.clamp((inputs * scale).round_(), min_bound, max_bound) if amax <= epsilon: scale = 1. # Return 1 makes more sense for values quantized to 0 with amax=0 if input_dtype == torch.half: outputs = outputs.half() outputs = outputs / scale return outputs
null
23,605
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctSigmoidTableLookupAIE2(Tinput, Toutput, fragpos): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 torch.ops.vai.SigmoidTableLookupAIE2(Tinput, Toutput, fragpos, device_id)
null
23,606
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctTanhTableLookupAIE2(Tinput, Toutput, fragpos): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 torch.ops.vai.TanhTableLookupAIE2(Tinput, Toutput, fragpos, device_id)
null
23,607
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctExpApprAIE2(Tinput, Toutput, bit_width): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 if device_id == 1: print("Exp Approximation does not support CPU") else: torch.ops.vai.ExpApprAIE2(Tinput, Toutput, device_id, bit_width)
null
23,608
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctLogSoftmaxFastLn(Tinput, Toutput): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 if device_id == 1: print("LogSoftmax fast ln does not support CPU") else: torch.ops.vai.LogSoftmaxFastLn(Tinput, Toutput, device_id)
null
23,609
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctLogSoftmaxSub(Tinput, Toutput, Tsum): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 if device_id == 1: print("LogSoftmax subtraction does not support CPU") else: torch.ops.vai.LogSoftmaxSub(Tinput, Toutput, Tsum, device_id)
null
23,610
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctAIESqrt(Tinput, Toutput): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 torch.ops.vai.AIESqrt(Tinput, Toutput, device_id)
null
23,611
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctAIEISqrt(Tinput, Toutput): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 torch.ops.vai.AIEISqrt(Tinput, Toutput, device_id)
null
23,612
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctISqrt(Tinput, Toutput): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 torch.ops.vai.LayernormISqrt(Tinput, Toutput, device_id)
null
23,613
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): cloned_tensor = tensor if (isinstance(tensor, torch.Tensor) and hasattr(tensor, "storage") and hasattr(tensor, "numel") and tensor.storage().size() != tensor.numel()): cloned_tensor = tensor.clone() return cloned_tensor import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctLayernormInvSqrt(Tinput, Toutput): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 if device_id == 1: print("Layernorm InvSqrt does not support CPU") else: torch.ops.vai.LayernormInvSqrt(Tinput, Toutput, device_id)
null
23,614
import sys import torch from ..load_kernels import * import copy import numpy as np from nndct_shared.utils import NndctOption, NndctScreenLogger from pytorch_nndct.nn.utils.decorator import pre_and_post_process_f16_tensor from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def clone_view_tensor(tensor): import torch from torch.utils.cpp_extension import load, _import_module_from_library def NndctInverseAIE2(Tinput, Toutput): Tinput = clone_view_tensor(Tinput) device_id = 1 if Tinput.device == torch.device("cpu") else 0 if device_id == 1: print("Inverse AIE2 does not support CPU") else: torch.ops.vai.InverseAIE2(Tinput, Toutput, device_id)
null
23,615
import math import torch from torch.autograd import Variable from nndct_shared.quantization import maybe_get_quantizer, quantize_tensors from nndct_shared.utils import NndctOption import pytorch_nndct.utils as py_utils class deephi_AvgPool2d(torch.nn.modules.AvgPool2d): r"""DeePhi Conv2d operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_AvgPool2d, self).__init__(*args, **kwards) self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = super().forward(qinput) # scale to DPU accuracy if NndctOption.nndct_avg_pool_approximate.value: scale = 1.0 if self.node.node_attr(self.node.op.AttrName.KERNEL) == [3, 3]: scale = 9.0 * 7.0 / 64.0 elif self.node.node_attr(self.node.op.AttrName.KERNEL) == [5, 5]: scale = 25.0 * 10.0 / 256.0 elif self.node.node_attr(self.node.op.AttrName.KERNEL) in [[6, 6], [3, 6], [6, 3]]: scale = 36.0 * 7.0 / 256.0 elif self.node.node_attr(self.node.op.AttrName.KERNEL) == [7, 7]: scale = 49.0 * 21.0 / 1024.0 elif self.node.node_attr(self.node.op.AttrName.KERNEL) == [14, 14]: scale = 196.0 * 21.0 / 4096.0 else: rec = self.node.node_attr(self.node.op.AttrName.KERNEL)[0] * self.node.node_attr(self.node.op.AttrName.KERNEL)[1] max_factor = math.ceil(math.log(rec * 128,2)) diff = 1.0 multi_factor = 0.0 shift_factor = 0.0 for shift_factor_ in range(max_factor): factor = round((2 ** shift_factor_)/rec) diff_ = abs(factor / (2 ** shift_factor_) - 1/rec) if diff_ < diff: multi_factor = factor diff = diff_ shift_factor = shift_factor_ scale = rec * multi_factor / (2 ** shift_factor) output = output * scale output = quantize_tensors([output], self.node)[0] return output def AvgPool2d(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode is None or NndctOption.nndct_quant_off.value: return torch.nn.AvgPool2d(*args, **kwargs) return deephi_AvgPool2d(*args, **kwargs)
null
23,616
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.utils import NndctOption from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils class deephi_ReLU(torch.nn.ReLU): r"""DeePhi ReLU operation""" def __init__(self, *args, **kwargs): super(deephi_ReLU, self).__init__(*args, **kwargs) self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): qinput = quantize_tensors([input], self.node, tensor_type='input')[0] output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] return output def ReLU(*args, **kwargs): #quant_mode,_ = maybe_get_quantizer() #if quant_mode==None: # return return deephi_ReLU(*args, **kwargs)
null
23,617
import torch from nndct_shared.quantization.utils import maybe_get_quantizer, quantize_tensors import pytorch_nndct.utils as py_utils class deephi_Add(torch.nn.Module): def __init__(self): super(deephi_Add, self).__init__() self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input, other, alpha=1): [qinput, qother] = quantize_tensors([input, other], self.node, tensor_type='input') output = torch.add(input=qinput, other=qother, alpha=alpha) output = quantize_tensors([output], self.node)[0] return output def Add(*args, **kwargs): return deephi_Add(*args, **kwargs)
null
23,618
import torch import numpy as np from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from nndct_shared.utils import NndctOption from nndct_shared.base import GLOBAL_MAP, NNDCT_KEYS from .sigmoid_table import * from .fix_ops import NndctSigmoidTableLookup, NndctSigmoidSimulation, NndctSigmoidTableLookupAIE2 import pytorch_nndct.utils as py_utils class deephi_Sigmoid(torch.nn.modules.Sigmoid): r"""DeePhi Sigmoid operation""" def __init__(self): super(deephi_Sigmoid, self).__init__() self.quant_mode, self.quantizer = maybe_get_quantizer() self.node = None def forward(self, input): if self.quant_mode == 0 or (not self.node.in_quant_part): return super().forward(input) qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if (NndctOption.nndct_quant_off.value or self.quantizer is None or self.quantizer.exporting or NndctOption.nndct_cv_app.value or NndctOption.nndct_only_int_quant is False): # Method 0: quant input and output (for CV) output = super().forward(qinput) output = quantize_tensors([output], self.node)[0] else: output = torch.empty_like(qinput) input_name = self.node.in_nodes[0] input_node = self.quantizer.configer.get_Nndctnode(input_name) if not self.quantizer.configer.node_output_quantizable(input_node): input_name = input_node.in_nodes[0] elif self.quantizer.configer.will_merge_with_table(input_node, (not NndctOption.nndct_cv_app.value)): output = super().forward(qinput) bnfp = self.quantizer.get_quant_config(input_name, False) bnfp[1] = 15 self.quantizer.set_quant_config(self.node.name, bnfp) return output bw = self.quantizer.get_quant_config(self.node.name, False)[0] fragpos = self.quantizer.get_quant_config(input_name, False)[1] # Method 1: Simulation AIE with 16 bw (for RNNT) if NndctOption.nndct_op_tanh_sigmoid_mode.value == "simulation": NndctSigmoidSimulation(qinput, output, fragpos) output = quantize_tensors([output], self.node)[0] # Method 2: Table Look up for AIE2 with 16 bw (based on LUT) elif NndctOption.nndct_op_tanh_sigmoid_mode.value == "aie2_lut_16bw" or NndctOption.nndct_ip_asr.value: NndctSigmoidTableLookupAIE2(qinput, output, fragpos) output = quantize_tensors([output], self.node)[0] # Method 3: Table Look up for FPGA with 16 bw else: quant_device = qinput.device Ttable = SIGMOID_TABLE.table.to(qinput.dtype).to(quant_device) output = output.to(quant_device) NndctSigmoidTableLookup(input, Ttable, output, fragpos) bnfp = self.quantizer.get_quant_config(input_name, False) bnfp[1] = 15 self.quantizer.set_quant_config(self.node.name, bnfp) return output import torch def Sigmoid(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode is None: return torch.nn.Sigmoid(*args, **kwargs) return deephi_Sigmoid(*args, **kwargs)
null
23,619
import torch from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.utils import NndctOption from nndct_shared.quantization import kernel_need_quant from nndct_shared.quantization import quantize_tensors import pytorch_nndct.utils as py_utils import numpy as np from pytorch_nndct.utils import Const from .fix_ops import NndctISqrt, NndctAIEISqrt class deephi_LayerNorm(torch.nn.LayerNorm): def __init__(self, *args, **kwargs): def forward(self, input): def _simulatedLayerNorm1(self, inp, normalized_shape, elementwise_affine, weight, bias, ifp, wfp, bfp): def _simulatedLayerNorm2(self, inp, normalized_shape, elementwise_affine, weight, bias, ifp, wfp, bfp): def LayerNorm(*args, **kwargs): quant_mode,_ = maybe_get_quantizer() if quant_mode is None: return torch.nn.LayerNorm(*args, **kwargs) return deephi_LayerNorm(*args, **kwargs)
null
23,620
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption, NndctScreenLogger, QError from nndct_shared.quantization import kernel_need_quant from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import maybe_get_quantizer import pytorch_nndct.utils as py_utils import torch.nn.functional as F from pytorch_nndct.utils import Const from .fix_ops import NndctISqrt class deephi_InstanceNorm(torch.nn.modules.instancenorm._InstanceNorm): def __init__(self, *args, **kwards): def forward(self, input): def simulateInstanceNorm(self, inp, params): def instancenorm_process(inp, ifp, weight, wfp, bias, bfp): def simulateInstanceNorm2(inp, affine, weight, bias, ifp, wfp, bfp): def InstanceNorm(*args, **kwargs): return deephi_InstanceNorm(*args, **kwargs)
null
23,621
import torch.nn.functional as F import torch.nn as nn import torch from torch.nn.parameter import Parameter def get_same_padding(kernel_size): if isinstance(kernel_size, (tuple, list)): assert len(kernel_size) == 2, 'invalid kernel size: %s' % kernel_size p1 = get_same_padding(kernel_size[0]) p2 = get_same_padding(kernel_size[1]) return p1, p2 assert isinstance( kernel_size, int), 'kernel size should be either `int` or `tuple` or `list`' assert kernel_size % 2 > 0, 'kernel size should be odd number' return kernel_size // 2
null
23,622
import torch.nn.functional as F import torch.nn as nn import torch from torch.nn.parameter import Parameter def sub_filter_start_end(kernel_size, sub_kernel_size): center = kernel_size // 2 dev = sub_kernel_size // 2 start, end = center - dev, center + dev + 1 assert end - start == sub_kernel_size return start, end
null
23,623
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption, NndctScreenLogger from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from .quant_noise import eval_qnoise import pytorch_nndct.utils as py_utils import torch.nn.functional as F class deephi_Mish(torch.nn.Mish): def __init__(self, *args, **kwards): def forward(self, input): def Mish(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.Mish(*args, **kwargs) return deephi_Mish(*args, **kwargs)
null
23,624
import torch import math def eval_qnoise(output, res_f, efficency, deviation, rate, stop): error = torch.add(output, res_f, alpha=-1).data noise = error.pow(2).mean() if noise > 0: eff = 1.25 * res_f.pow(2).mean().div(noise).log10().detach().cpu().numpy() dev = math.fabs(eff - efficency) if dev > 0: efficency = (efficency * 4 + eff) * 0.2 deviation = (deviation * 4 + dev) * 0.2 #print(node.name, efficency, deviation) if efficency > 4.0: rate = rate * 0.5 if (efficency > 4.3 or (deviation / efficency) < 0.05 or math.fabs(dev - deviation / dev) < 0.05): stop = True else: stop = True else: stop = True return error, rate, stop, efficency, deviation
null
23,625
import torch from torch.autograd import Variable import math from nndct_shared.utils import NndctOption from nndct_shared.quantization import quantize_tensors from nndct_shared.quantization import maybe_get_quantizer import pytorch_nndct.utils as py_utils import torch.nn.functional as F class deephi_BatchNorm(torch.nn.modules.batchnorm._BatchNorm): r"""DeePhi batchnorm operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_BatchNorm, self).__init__(*args, **kwards) self.params_name = None self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() self.param_saved = False self.param_quantized = False def forward(self, input): params = [self.weight, self.bias] param_names = self.params_name[:2] qinput = quantize_tensors([input], self.node, tensor_type='input')[0] if (not self.param_quantized): inplace = (NndctOption.nndct_quant_off.value or self.quantizer is not None and self.quantizer.inplace) # quantize weights and bias if inplace: _ = quantize_tensors( params, self.node, tensor_names=param_names, tensor_type='param') qparams = [p for p in params] else: qparams = quantize_tensors( params, self.node, tensor_names=param_names, tensor_type='param') if not NndctOption.nndct_quant_off.value: self.param_quantized = True else: qparams = [p for p in params] if self.momentum is None: exponential_average_factor = 0.0 else: exponential_average_factor = self.momentum if self.training and self.track_running_stats: # TODO: if statement only here to tell the jit to skip emitting this when it is None if self.num_batches_tracked is not None: # type: ignore[has-type] self.num_batches_tracked = self.num_batches_tracked + 1 # type: ignore[has-type] if self.momentum is None: # use cumulative moving average exponential_average_factor = 1.0 / float(self.num_batches_tracked) else: # use exponential moving average exponential_average_factor = self.momentum if self.training: bn_training = True else: bn_training = (self.running_mean is None) and (self.running_var is None) output = torch.nn.functional.batch_norm( qinput, # If buffers are not to be tracked, ensure that they won't be updated self.running_mean if not self.training or self.track_running_stats else None, self.running_var if not self.training or self.track_running_stats else None, qparams[0], qparams[1], bn_training, exponential_average_factor, self.eps, ) # quantize output output = quantize_tensors([output], self.node)[0] return output def _check_input_dim(self, input): pass def BatchNorm(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: def _check_input_dim(self, input): pass import types nn = torch.nn.modules.batchnorm._BatchNorm(*args, **kwargs) nn._check_input_dim = types.MethodType(_check_input_dim, nn) return nn return deephi_BatchNorm(*args, **kwargs)
null
23,626
import torch import math from nndct_shared.utils import NndctOption, NndctScreenLogger, QError, QWarning from nndct_shared.quantization import maybe_get_quantizer from nndct_shared.quantization import quantize_tensors from .quant_noise import eval_qnoise import pytorch_nndct.utils as py_utils from .add import Add from .multiply import Mul class deephi_Linear(torch.nn.modules.linear.Linear): r"""DeePhi Linear operation, support float and double""" def __init__(self, *args, **kwards): super(deephi_Linear, self).__init__(*args, **kwards) self.params_name = None self.node = None self.quant_mode, self.quantizer = maybe_get_quantizer() self.param_saved = False self.param_quantized = False # self.weight and self.bias are not quantized float parameters self.weight_bak = None # backup of float bias for bias correction self.bias_bak = None # backup of float bias for bias correction self.stop = False self.rate = NndctOption.nndct_param_corr_rate.value self.efficency = 0.0 self.deviation = 0.0 def forward(self, input): if self.quantizer is None or NndctOption.nndct_quant_off.value is True: return self.fp32_forward(input) else: return self.fake_quantize_forward(input) def fp32_forward(self, input): return super().forward(input) def fake_quantize_forward(self, input): # backup bias for bias correction feature if (not self.param_saved): if NndctOption.nndct_param_corr.value > 0: # backup orignal float parameters if self.quant_mode == 1: self.weight_bak = self.weight.detach().clone() if self.bias is not None: self.bias_bak = self.bias.detach().clone() # adjust bias if self.quant_mode == 2 and self.bias is not None: if not self.quantizer.has_bias_corr(self.node): NndctScreenLogger().error2user(QError.BIAS_CORRECTION, f"Bias correction file in quantization result directory does not match current model.") exit(2) self.bias.data = torch.sub(self.bias.data, torch.tensor( self.quantizer.get_bias_corr(self.node), device=self.bias.data.device, dtype=self.bias.data.dtype)) self.param_saved = True # quantize parameters qweight = None qbias = None inplace = (NndctOption.nndct_quant_off.value or self.quantizer is not None and self.quantizer.inplace) if (not self.param_quantized): if inplace: _ = quantize_tensors( [self.weight], self.node, tensor_names = [self.params_name[0]], tensor_type = 'param')[0] qweight = self.weight if self.bias is not None: _ = quantize_tensors( [self.bias], self.node, tensor_names = [self.params_name[1]], tensor_type = 'param')[0] qbias = self.bias else: qweight = quantize_tensors( [self.weight], self.node, tensor_names = [self.params_name[0]], tensor_type = 'param')[0] if self.bias is not None: qbias = quantize_tensors( [self.bias], self.node, tensor_names = [self.params_name[1]], tensor_type = 'param')[0] if not NndctOption.nndct_quant_off.value: self.param_quantized = True else: qweight = self.weight qbias = self.bias # quantize input tensor qinput = quantize_tensors([input], self.node, tensor_type='input')[0] # split linear to mul and add operations if (self.quant_mode == 2 and self.quantizer.is_lstm): # i * w output = torch.matmul(qinput, torch.transpose(qweight, 0, 1)) datatype = 'int' if NndctOption.nndct_only_int_quant.value is False: datatype = self.quantizer.get_quant_dtype(self.node.name, tensor_type='output') output = self.quantizer.quantize(output, self.node.name, self.node, tensor_type='output', datatype=datatype) # i*w + bias if self.bias is not None: output = torch.add(output, qbias) else: output = torch.nn.functional.linear(qinput, qweight, qbias) output = quantize_tensors([output], self.node)[0] if NndctOption.nndct_param_corr.value > 0: #rate = NndctOption.nndct_param_corr_rate.value # statistic of quantization error if (self.quant_mode == 1 and not self.stop): res_f = torch.matmul(input, torch.transpose(self.weight_bak, 0, 1)) if self.bias is not None: res_f = torch.add(res_f, self.bias_bak) error, rate, self.stop, self.efficency, self.deviation = eval_qnoise( output, res_f, self.efficency, self.deviation, self.rate, self.stop) if (not self.stop) and (self.bias is not None): error = error.mean(dim = [k for k in range(error.dim()-1)]) self.bias.data = torch.sub(self.bias.data, error, alpha=rate) self.param_quantized = False return output def bias_corr(self): if self.bias is not None and self.bias_bak is not None: bias_err = torch.sub(self.bias_bak, self.bias.data) return bias_err.cpu().numpy().tolist() else: return None def Linear(*args, **kwargs): quant_mode, _ = maybe_get_quantizer() if quant_mode == None: return torch.nn.Linear(*args, **kwargs) return deephi_Linear(*args, **kwargs)
null
23,627
class ApproxModes(object): NO_APPROX = 'no_approx' EXP_POLY = 'exp_poly' EXP_LUT = 'exp_lut' QIO = 'quant_input_output' def is_no_approx(mode): return mode == ApproxModes.NO_APPROX
null
23,628
class ApproxModes(object): NO_APPROX = 'no_approx' EXP_POLY = 'exp_poly' EXP_LUT = 'exp_lut' QIO = 'quant_input_output' def is_exp_poly(mode): return mode == ApproxModes.EXP_POLY
null
23,629
class ApproxModes(object): NO_APPROX = 'no_approx' EXP_POLY = 'exp_poly' EXP_LUT = 'exp_lut' QIO = 'quant_input_output' def is_exp_lut(mode): return mode == ApproxModes.EXP_LUT
null
23,630
class ApproxModes(object): def is_quant_input_output(mode): return mode == ApproxModes.QIO
null
23,631
class ApproxModes(object): NO_APPROX = 'no_approx' EXP_POLY = 'exp_poly' EXP_LUT = 'exp_lut' QIO = 'quant_input_output' def available_modes(): return [ApproxModes.NO_APPROX, ApproxModes.EXP_POLY, ApproxModes.EXP_LUT]
null
23,632
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def mult_add(a_bf16, b_bf16, c_fp32): assert a_bf16.dtype == torch.bfloat16 and b_bf16.dtype == torch.bfloat16 out = a_bf16.to(torch.float32) * b_bf16.to(torch.float32) + c_fp32 return out
null
23,633
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal def ploy_HORNER_SCHEME(r, cs, degree): out = mult_add(r, cs[degree].to(torch.bfloat16), cs[degree - 1]) for i in reversed(range(degree - 1)): out = mult_add(r, out.to(torch.bfloat16), cs[i]) out = out.to(r.dtype) return out from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def ploy_HORNER_SCHEME(r, cs, degree): out = mult_add(r, cs[degree].to(torch.bfloat16), cs[degree - 1]) for i in reversed(range(degree - 1)): out = mult_add(r, out.to(torch.bfloat16), cs[i]) out = out.to(r.dtype) return out def get_sigmoid_positive_ploy_coeffcients(degree=4): if not degree in list(SIGMOID_POSITIVE_PLOY_COEFFCIENTS.keys()): raise NotImplementedError( f'degree must in: {list(SIGMOID_POSITIVE_PLOY_COEFFCIENTS.keys())}') cs = np.array(SIGMOID_POSITIVE_PLOY_COEFFCIENTS[degree], dtype=np.float32) return cs def sigmoid_positive_approx(z, degree=4): rge = [-7.0, 7.0] zero_mask = (z < rge[0]).to(torch.bfloat16) one_mask = (z > rge[1]).to(torch.bfloat16) x_sgn, x = torch.sign(z), torch.clamp(torch.abs(z), 0, rge[1]) m_zero_mask = (x_sgn < 0).to(torch.bfloat16) cs = get_sigmoid_positive_ploy_coeffcients(degree) coeffs = torch.from_numpy(cs).to(z.device) # out = coeffs[0] + z * (coeffs[1] + z * (coeffs[2] + z * coeffs[3])) out = ploy_HORNER_SCHEME(x, coeffs, degree) out = m_zero_mask - out * m_zero_mask + out * (1 - m_zero_mask) out = out * (1.0 - zero_mask - one_mask) + one_mask return out
null
23,634
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal def ploy_HORNER_SCHEME(r, cs, degree): out = mult_add(r, cs[degree].to(torch.bfloat16), cs[degree - 1]) for i in reversed(range(degree - 1)): out = mult_add(r, out.to(torch.bfloat16), cs[i]) out = out.to(r.dtype) return out from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def ploy_HORNER_SCHEME(r, cs, degree): out = mult_add(r, cs[degree].to(torch.bfloat16), cs[degree - 1]) for i in reversed(range(degree - 1)): out = mult_add(r, out.to(torch.bfloat16), cs[i]) out = out.to(r.dtype) return out def get_tanh_positive_poly_coeffcients(degree=4): cs = np.array(TANH_POSITIVE_POLY_COEFFCIENTS[degree], dtype=np.float32) return cs def tanh_positive_approx(z, degree=4): assert z.dtype == torch.bfloat16 rge = [-4.0, 4.0] m_one_mask = (z < rge[0]).to(torch.bfloat16) one_mask = (z > rge[1]).to(torch.bfloat16) x_sgn, x = torch.sign(z), torch.clamp(torch.abs(z), 0, rge[1]) cs = get_tanh_positive_poly_coeffcients(degree=degree) coeffs = torch.from_numpy(cs).to(z.device) out = ploy_HORNER_SCHEME(x, coeffs, degree) out = out * x_sgn out = out * (1.0 - m_one_mask - one_mask) + one_mask + m_one_mask * (-1) return out
null
23,635
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def exp_approx_poly(x, exp_table_size=1, degree=3, method='HORNER'): assert method == 'HORNER' assert x.dtype == torch.float32 EXP_TABLE_SIZE = exp_table_size cs = get_exp_poly_coeffcients(exp_table_size, degree) cs = torch.from_numpy(cs).to(x.device) if EXP_TABLE_SIZE > 1: scaling_factor = EXP_TABLE_SIZE / np.log(2) n = _AMD_ROUND(x.to(torch.float32) * scaling_factor) m = _AMD_FLOOR((n / EXP_TABLE_SIZE).to(torch.float32)) #.to(torch.bfloat16) else: scaling_factor = x.to(torch.bfloat16) * ( torch.scalar_tensor(EXP_TABLE_SIZE / LN2).to(torch.bfloat16)) n = _AMD_ROUND(scaling_factor.to(torch.float32)).to(torch.bfloat16) m = _AMD_FLOOR((n / EXP_TABLE_SIZE).to(torch.float32)).to(torch.bfloat16) j = n - m * EXP_TABLE_SIZE #n % EXP_TABLE_SIZE # exp = 2**( (EXP_TABLE_SIZE * m + j + f) / EXP_TABLE_SIZE) # (j/EXP_TABLE_SIZE) < 1.0: Look up table for 2**(j/EXP_TABLE_SIZE). # r = np.exp(f*(LN2/EXP_TABLE_SIZE)) < 1.0: Polynomial r = x - n.to(torch.float32) * (LN2 / EXP_TABLE_SIZE) # exp = 2**m * (2**(j/EXP_TABLE_SIZE)) * np.exp(r) if method == 'HORNER': # exp_x = ( 2**m * (2**(j/EXP_TABLE_SIZE)) * ploy_HORNER_SCHEME(r.to(torch.bfloat16), cs, degree) ).to(torch.bfloat16 exp_x = (2**(j / EXP_TABLE_SIZE).to(torch.bfloat16)) * ploy_HORNER_SCHEME( r.to(torch.bfloat16), cs, degree) if EXP_TABLE_SIZE > 1: assert m.dtype == torch.float32 and exp_x.dtype == torch.bfloat16 else: assert m.dtype == torch.bfloat16 and exp_x.dtype == torch.bfloat16 exp_x = 2**m.to(torch.float32) * exp_x.to(torch.float32) exp_x = torch.clamp(exp_x, _MIN_BFLOAT16, _MAX_BFLOAT16).to(torch.bfloat16) assert exp_x.dtype == torch.bfloat16 return exp_x def reciprocal_approx_moroz(z, output_subnormals=False): assert z.dtype == torch.float32 # @Ephrem (ephremw@xilinx.com) def okay_to_process(x): r"""Indicates which elements in the tensor are ready for reciprocal approximation. The FP32 exponent ranges from -126 to 127. The largest number whose reciprocal does not become subnormal is :math:`2^{126}` because :math:`2^{-126}` is the smallest normal number. This function returns `True` if `x` is normal and :math:`0 < |x| \le 2^{126}`. """ okay_idx = np.logical_and(is_normal(x), x != 0) if output_subnormals: return okay_idx return np.logical_and( okay_idx, x <= np.exp2( min( # maxexp=128, smallest for overflow # pylint: disable=no-member np.abs(np.finfo(np.float32).minexp), np.abs(np.finfo(np.float32).maxexp - 2)))) def mult_add_func(a, b, c): return a * b + c magic_n = np.int32(0x7eb53567) k1 = np.float32(1.9395974) k2 = np.float32(1.436142) magic = [magic_n, k1, k2] x = z.detach().cpu().numpy() frac, exponent = np.frexp(x) reduced_x = np.ldexp(frac, 1, dtype=np.float32) y_as_int = np.int32(magic[0]) - reduced_x.view(np.int32) y_as_fp32 = y_as_int.view(np.float32) y_as_fp32 = np.where( okay_to_process(x), # 1st (modified) Newton-Raphson mult_add_func( mult_add_func(magic[1], y_as_fp32, 0.0), mult_add_func(-reduced_x, y_as_fp32, np.float32(magic[2])), 0.0), y_as_fp32) # r = 1 - xy, y = ry + y # 2nd (classic) Newton-Raphson y_as_fp32 = np.where( okay_to_process(x), mult_add_func(y_as_fp32, mult_add_func(y_as_fp32, -reduced_x, np.float32(1.0)), y_as_fp32), y_as_fp32) out = np.where( np.logical_or(np.isfinite(x) is False, exponent > 126 + 1), np.copysign(np.float32(0), x), np.where( np.logical_or(is_subnormal(x), x == 0), np.copysign(np.float32(np.inf), x), np.ldexp(y_as_fp32, -exponent + 1).astype(np.float32))) out_tensor = torch.from_numpy(out).to(z.device).to(z.dtype) return out_tensor def tanh_with_exp_approx(z, degree=3, exp_table_size=1, output_subnormals=False, dtype=torch.bfloat16): assert z.dtype == dtype sign = torch.sign(z).to(dtype) x = torch.abs(z) out = -2.0 * x out = 1.0 + exp_approx_poly( out.to(torch.float32), exp_table_size=exp_table_size, degree=degree).to(dtype) out = reciprocal_approx_moroz( out.to(torch.float32), output_subnormals=output_subnormals).to(dtype) out = 2.0 * out - 1.0 out = out * sign assert out.dtype == dtype return out
null
23,636
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def exp_approx_lut(x, lg_table_depth=8): def reciprocal_approx_moroz(z, output_subnormals=False): def tanh_with_exp_approx_lut(z, output_subnormals=False, dtype=torch.bfloat16): assert z.dtype == dtype sign = torch.sign(z).to(dtype) x = torch.abs(z) out = -2.0 * x out = 1.0 + exp_approx_lut(out.to(torch.bfloat16)).to(dtype) out = reciprocal_approx_moroz( out.to(torch.float32), output_subnormals=output_subnormals).to(dtype) out = 2.0 * out - 1.0 out = out * sign assert out.dtype == dtype return out
null
23,637
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def exp_approx_poly(x, exp_table_size=1, degree=3, method='HORNER'): def reciprocal_approx_moroz(z, output_subnormals=False): def sigmoid_with_exp_approx(z, degree=3, exp_table_size=1, output_subnormals=False, dtype=torch.bfloat16): assert z.dtype == dtype z = z.to(torch.float32) out = 1.0 + exp_approx_poly( -z, exp_table_size=exp_table_size, degree=degree).to(dtype) out = reciprocal_approx_moroz( out.to(torch.float32), output_subnormals=output_subnormals) out = out.to(dtype) assert out.dtype == dtype return out
null
23,638
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def exp_approx_lut(x, lg_table_depth=8): assert x.dtype == torch.bfloat16 beta_t = np.exp2(lg_table_depth).astype(int) scaling_factor = beta_t / np.log(2) scaled_x = _AMD_ROUND(x.to(torch.float32) * scaling_factor) assert scaled_x.dtype == torch.float32 quotient = _AMD_FLOOR(scaled_x / beta_t) remainder = scaled_x - quotient * beta_t #scaled_x % beta_t # fraction, exponent = np.frexp(to_dtype(np.exp2(remainder/beta_t), out_dtype)) # exponent = (exponent + quotient).astype(int) # y = np.ldexp(fraction, exponent) y = (2**quotient) * (2**(remainder / beta_t).to(torch.bfloat16)) y = torch.clamp(y, _MIN_BFLOAT16, _MAX_BFLOAT16).to(torch.bfloat16) assert y.dtype == torch.bfloat16 return y def reciprocal_approx_moroz(z, output_subnormals=False): assert z.dtype == torch.float32 # @Ephrem (ephremw@xilinx.com) def okay_to_process(x): r"""Indicates which elements in the tensor are ready for reciprocal approximation. The FP32 exponent ranges from -126 to 127. The largest number whose reciprocal does not become subnormal is :math:`2^{126}` because :math:`2^{-126}` is the smallest normal number. This function returns `True` if `x` is normal and :math:`0 < |x| \le 2^{126}`. """ okay_idx = np.logical_and(is_normal(x), x != 0) if output_subnormals: return okay_idx return np.logical_and( okay_idx, x <= np.exp2( min( # maxexp=128, smallest for overflow # pylint: disable=no-member np.abs(np.finfo(np.float32).minexp), np.abs(np.finfo(np.float32).maxexp - 2)))) def mult_add_func(a, b, c): return a * b + c magic_n = np.int32(0x7eb53567) k1 = np.float32(1.9395974) k2 = np.float32(1.436142) magic = [magic_n, k1, k2] x = z.detach().cpu().numpy() frac, exponent = np.frexp(x) reduced_x = np.ldexp(frac, 1, dtype=np.float32) y_as_int = np.int32(magic[0]) - reduced_x.view(np.int32) y_as_fp32 = y_as_int.view(np.float32) y_as_fp32 = np.where( okay_to_process(x), # 1st (modified) Newton-Raphson mult_add_func( mult_add_func(magic[1], y_as_fp32, 0.0), mult_add_func(-reduced_x, y_as_fp32, np.float32(magic[2])), 0.0), y_as_fp32) # r = 1 - xy, y = ry + y # 2nd (classic) Newton-Raphson y_as_fp32 = np.where( okay_to_process(x), mult_add_func(y_as_fp32, mult_add_func(y_as_fp32, -reduced_x, np.float32(1.0)), y_as_fp32), y_as_fp32) out = np.where( np.logical_or(np.isfinite(x) is False, exponent > 126 + 1), np.copysign(np.float32(0), x), np.where( np.logical_or(is_subnormal(x), x == 0), np.copysign(np.float32(np.inf), x), np.ldexp(y_as_fp32, -exponent + 1).astype(np.float32))) out_tensor = torch.from_numpy(out).to(z.device).to(z.dtype) return out_tensor def sigmoid_with_exp_approx_lut(z, output_subnormals=False, dtype=torch.bfloat16): assert z.dtype == dtype z = z.to(torch.bfloat16) out = 1.0 + exp_approx_lut(-z).to(dtype) out = reciprocal_approx_moroz( out.to(torch.float32), output_subnormals=output_subnormals) out = out.to(dtype) assert out.dtype == dtype return out
null
23,639
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def exp_approx_poly(x, exp_table_size=1, degree=3, method='HORNER'): assert method == 'HORNER' assert x.dtype == torch.float32 EXP_TABLE_SIZE = exp_table_size cs = get_exp_poly_coeffcients(exp_table_size, degree) cs = torch.from_numpy(cs).to(x.device) if EXP_TABLE_SIZE > 1: scaling_factor = EXP_TABLE_SIZE / np.log(2) n = _AMD_ROUND(x.to(torch.float32) * scaling_factor) m = _AMD_FLOOR((n / EXP_TABLE_SIZE).to(torch.float32)) #.to(torch.bfloat16) else: scaling_factor = x.to(torch.bfloat16) * ( torch.scalar_tensor(EXP_TABLE_SIZE / LN2).to(torch.bfloat16)) n = _AMD_ROUND(scaling_factor.to(torch.float32)).to(torch.bfloat16) m = _AMD_FLOOR((n / EXP_TABLE_SIZE).to(torch.float32)).to(torch.bfloat16) j = n - m * EXP_TABLE_SIZE #n % EXP_TABLE_SIZE # exp = 2**( (EXP_TABLE_SIZE * m + j + f) / EXP_TABLE_SIZE) # (j/EXP_TABLE_SIZE) < 1.0: Look up table for 2**(j/EXP_TABLE_SIZE). # r = np.exp(f*(LN2/EXP_TABLE_SIZE)) < 1.0: Polynomial r = x - n.to(torch.float32) * (LN2 / EXP_TABLE_SIZE) # exp = 2**m * (2**(j/EXP_TABLE_SIZE)) * np.exp(r) if method == 'HORNER': # exp_x = ( 2**m * (2**(j/EXP_TABLE_SIZE)) * ploy_HORNER_SCHEME(r.to(torch.bfloat16), cs, degree) ).to(torch.bfloat16 exp_x = (2**(j / EXP_TABLE_SIZE).to(torch.bfloat16)) * ploy_HORNER_SCHEME( r.to(torch.bfloat16), cs, degree) if EXP_TABLE_SIZE > 1: assert m.dtype == torch.float32 and exp_x.dtype == torch.bfloat16 else: assert m.dtype == torch.bfloat16 and exp_x.dtype == torch.bfloat16 exp_x = 2**m.to(torch.float32) * exp_x.to(torch.float32) exp_x = torch.clamp(exp_x, _MIN_BFLOAT16, _MAX_BFLOAT16).to(torch.bfloat16) assert exp_x.dtype == torch.bfloat16 return exp_x def reciprocal_approx_moroz(z, output_subnormals=False): assert z.dtype == torch.float32 # @Ephrem (ephremw@xilinx.com) def okay_to_process(x): r"""Indicates which elements in the tensor are ready for reciprocal approximation. The FP32 exponent ranges from -126 to 127. The largest number whose reciprocal does not become subnormal is :math:`2^{126}` because :math:`2^{-126}` is the smallest normal number. This function returns `True` if `x` is normal and :math:`0 < |x| \le 2^{126}`. """ okay_idx = np.logical_and(is_normal(x), x != 0) if output_subnormals: return okay_idx return np.logical_and( okay_idx, x <= np.exp2( min( # maxexp=128, smallest for overflow # pylint: disable=no-member np.abs(np.finfo(np.float32).minexp), np.abs(np.finfo(np.float32).maxexp - 2)))) def mult_add_func(a, b, c): return a * b + c magic_n = np.int32(0x7eb53567) k1 = np.float32(1.9395974) k2 = np.float32(1.436142) magic = [magic_n, k1, k2] x = z.detach().cpu().numpy() frac, exponent = np.frexp(x) reduced_x = np.ldexp(frac, 1, dtype=np.float32) y_as_int = np.int32(magic[0]) - reduced_x.view(np.int32) y_as_fp32 = y_as_int.view(np.float32) y_as_fp32 = np.where( okay_to_process(x), # 1st (modified) Newton-Raphson mult_add_func( mult_add_func(magic[1], y_as_fp32, 0.0), mult_add_func(-reduced_x, y_as_fp32, np.float32(magic[2])), 0.0), y_as_fp32) # r = 1 - xy, y = ry + y # 2nd (classic) Newton-Raphson y_as_fp32 = np.where( okay_to_process(x), mult_add_func(y_as_fp32, mult_add_func(y_as_fp32, -reduced_x, np.float32(1.0)), y_as_fp32), y_as_fp32) out = np.where( np.logical_or(np.isfinite(x) is False, exponent > 126 + 1), np.copysign(np.float32(0), x), np.where( np.logical_or(is_subnormal(x), x == 0), np.copysign(np.float32(np.inf), x), np.ldexp(y_as_fp32, -exponent + 1).astype(np.float32))) out_tensor = torch.from_numpy(out).to(z.device).to(z.dtype) return out_tensor def softmax_approx_poly(x, axis=-1, exp_table_size=1, degree=3): x = x - torch.max(x.to(torch.float32), axis, keepdims=True)[0].to(x.dtype) exp_x = exp_approx_poly(x, exp_table_size=exp_table_size, degree=degree) # NOTE: approximate: r = exp_x / torch.sum(exp_x, axis, keepdim=True) exp_x_sum = torch.sum(exp_x.to(torch.float32), axis, keepdim=True) exp_x_sum_reciprocal = reciprocal_approx_moroz(exp_x_sum) r = exp_x.to(torch.float32) * exp_x_sum_reciprocal r = r.to(torch.bfloat16) return r
null
23,640
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def exp_approx_lut(x, lg_table_depth=8): assert x.dtype == torch.bfloat16 beta_t = np.exp2(lg_table_depth).astype(int) scaling_factor = beta_t / np.log(2) scaled_x = _AMD_ROUND(x.to(torch.float32) * scaling_factor) assert scaled_x.dtype == torch.float32 quotient = _AMD_FLOOR(scaled_x / beta_t) remainder = scaled_x - quotient * beta_t #scaled_x % beta_t # fraction, exponent = np.frexp(to_dtype(np.exp2(remainder/beta_t), out_dtype)) # exponent = (exponent + quotient).astype(int) # y = np.ldexp(fraction, exponent) y = (2**quotient) * (2**(remainder / beta_t).to(torch.bfloat16)) y = torch.clamp(y, _MIN_BFLOAT16, _MAX_BFLOAT16).to(torch.bfloat16) assert y.dtype == torch.bfloat16 return y def reciprocal_approx_moroz(z, output_subnormals=False): assert z.dtype == torch.float32 # @Ephrem (ephremw@xilinx.com) def okay_to_process(x): r"""Indicates which elements in the tensor are ready for reciprocal approximation. The FP32 exponent ranges from -126 to 127. The largest number whose reciprocal does not become subnormal is :math:`2^{126}` because :math:`2^{-126}` is the smallest normal number. This function returns `True` if `x` is normal and :math:`0 < |x| \le 2^{126}`. """ okay_idx = np.logical_and(is_normal(x), x != 0) if output_subnormals: return okay_idx return np.logical_and( okay_idx, x <= np.exp2( min( # maxexp=128, smallest for overflow # pylint: disable=no-member np.abs(np.finfo(np.float32).minexp), np.abs(np.finfo(np.float32).maxexp - 2)))) def mult_add_func(a, b, c): return a * b + c magic_n = np.int32(0x7eb53567) k1 = np.float32(1.9395974) k2 = np.float32(1.436142) magic = [magic_n, k1, k2] x = z.detach().cpu().numpy() frac, exponent = np.frexp(x) reduced_x = np.ldexp(frac, 1, dtype=np.float32) y_as_int = np.int32(magic[0]) - reduced_x.view(np.int32) y_as_fp32 = y_as_int.view(np.float32) y_as_fp32 = np.where( okay_to_process(x), # 1st (modified) Newton-Raphson mult_add_func( mult_add_func(magic[1], y_as_fp32, 0.0), mult_add_func(-reduced_x, y_as_fp32, np.float32(magic[2])), 0.0), y_as_fp32) # r = 1 - xy, y = ry + y # 2nd (classic) Newton-Raphson y_as_fp32 = np.where( okay_to_process(x), mult_add_func(y_as_fp32, mult_add_func(y_as_fp32, -reduced_x, np.float32(1.0)), y_as_fp32), y_as_fp32) out = np.where( np.logical_or(np.isfinite(x) is False, exponent > 126 + 1), np.copysign(np.float32(0), x), np.where( np.logical_or(is_subnormal(x), x == 0), np.copysign(np.float32(np.inf), x), np.ldexp(y_as_fp32, -exponent + 1).astype(np.float32))) out_tensor = torch.from_numpy(out).to(z.device).to(z.dtype) return out_tensor def softmax_approx_lut(x, axis=-1, lg_table_depth=8): x = x - torch.max(x.to(torch.float32), axis, keepdims=True)[0].to(x.dtype) exp_x = exp_approx_lut(x, lg_table_depth=lg_table_depth) # NOTE: approximate: r = exp_x / torch.sum(exp_x, axis, keepdim=True) exp_x_sum = torch.sum(exp_x.to(torch.float32), axis, keepdim=True) exp_x_sum_reciprocal = reciprocal_approx_moroz(exp_x_sum) r = exp_x.to(torch.float32) * exp_x_sum_reciprocal r = r.to(torch.bfloat16) return r
null
23,641
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal def ploy_HORNER_SCHEME(r, cs, degree): out = mult_add(r, cs[degree].to(torch.bfloat16), cs[degree - 1]) for i in reversed(range(degree - 1)): out = mult_add(r, out.to(torch.bfloat16), cs[i]) out = out.to(r.dtype) return out from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def ploy_HORNER_SCHEME(r, cs, degree): out = mult_add(r, cs[degree].to(torch.bfloat16), cs[degree - 1]) for i in reversed(range(degree - 1)): out = mult_add(r, out.to(torch.bfloat16), cs[i]) out = out.to(r.dtype) return out def get_gelu_tanh_poly_coeffcients(degree=2): cs = np.array(GLUE_TANH_POSITIVE_POLY_COEFFCIENTS[degree], dtype=np.float32) return cs def gelu_approx(x, degree=2): assert x.dtype == torch.bfloat16 sqrt_2 = 1.4142 rge = [-2.5, 2.5] zero_mask = (x < rge[0]).to(x.device).to(torch.bfloat16) y_e_x_mask = (x > rge[1]).to(x.device).to(torch.bfloat16) q = x q_sgn, q = torch.sign(q), torch.clamp(torch.abs(q), 0, rge[1]) cs = get_gelu_tanh_poly_coeffcients(degree=degree) coeffs = torch.from_numpy(cs).to(x.device).to(x.device) out = ploy_HORNER_SCHEME(q, coeffs, degree) out = x * 0.5 * (1 + out * q_sgn) out = out * (1.0 - zero_mask - y_e_x_mask) + x * y_e_x_mask assert out.dtype == torch.bfloat16 return out
null
23,642
import torch import numpy as np from .coefficient import get_sigmoid_positive_ploy_coeffcients, get_exp_poly_coeffcients, get_gelu_tanh_poly_coeffcients, get_tanh_positive_poly_coeffcients from pytorch_nndct.utils.hw_dtype import is_subnormal, is_normal from pytorch_nndct.utils.torch_utils import CmpFlag, compare_torch_version def isqrt_approx_quake(x, mantissa_bit=23, exponent_bit=8): # From fast inverse squre root. # gamma = 0.0450466 first order Taylor error. # magic_number = 3/2 * 2^mantissa_bit * (2^(exponent_bit-1) - 1 - gamma) gamma = 0.0450466 magic_n = np.array( np.round(3.0 / 2.0 * 2.0**mantissa_bit * (2.0**(exponent_bit - 1) - 1.0 - gamma)), dtype=np.int32) # From https://www.mdpi.com/1099-4300/23/1/86/pdf authored by # Walczy, C.J.; Moroz, L.V..; Cie´sli´nski, J.L. # This work is licensed under the Creative Commons Attribution 4.0 # International License. To view a copy of this license, visit # http://creativecommons.org/licenses/by/4.0/ or send a letter to # Creative Commons, PO Box 1866, Mountain View, CA 94042, USA. # magic_n = np.round(2**mantissa_bit *(3* (2.0**(exponent_bit-1) - 1.0) - 1) /2 + np.round(2**mantissa_bit * (3.7315712401613957182292407381942955 - 2)/4 - 0.5)) # magic_n = np.array(magic_n, dtype=np.int32) number = x.cpu().numpy() threehalfs = 1.5 x2 = number * 0.5 y = np.float32(number) i = y.view(np.int32) i = magic_n - np.int32(i >> 1) y = i.view(np.float32) y = y * (threehalfs - (x2 * y * y)) y = y * (threehalfs - (x2 * y * y)) return torch.from_numpy(y).to(x.device).to(x.dtype)
null