mradermacher/gpt-oss-20b-triton-kernel-GGUF
21B • Updated
• 786 • 1
entry_point stringlengths 1 65 | original_triton_code stringlengths 4.5k 619k | python_code stringlengths 208 60.9k | triton_code stringlengths 1.15k 275k | repo_name stringlengths 7 115 | module_name stringlengths 1 65 | synthetic bool 1 class | uuid int64 0 18.5k | licenses listlengths 1 6 | stars int64 0 19.8k | sha stringlengths 40 40 | repo_link stringlengths 72 180 | messages stringlengths 8.63k 1.48M |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
SumAggregator | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/7o/c7otc5ij6whexgxcr56vlxp2l7hzg3oc4onljp557uc6wncu5gvg.py
# Topologically Sorted Source Nodes: [sum_1], Original ATen: [aten.sum]
# Source node to ATen node mapping:
# sum_1 => sum_1
# Graph fragment:
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%arg0_1, [1]), kwargs = {})
triton_poi_fused_sum_0 = async_compile.triton('triton_poi_fused_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[64],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_sum_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = (xindex // 16)
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + (64*x1)), xmask)
tmp1 = tl.load(in_ptr0 + (16 + x0 + (64*x1)), xmask)
tmp3 = tl.load(in_ptr0 + (32 + x0 + (64*x1)), xmask)
tmp5 = tl.load(in_ptr0 + (48 + x0 + (64*x1)), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tl.store(out_ptr0 + (x2), tmp6, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [sum_1], Original ATen: [aten.sum]
stream0 = get_raw_stream(0)
triton_poi_fused_sum_0.run(arg0_1, buf0, 64, grid=grid(64), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class SumAggregator(nn.Module):
def __init__(self):
super(SumAggregator, self).__init__()
def forward(self, neighbor):
return torch.sum(neighbor, dim=1)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_sum_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = xindex // 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)
tmp3 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)
tmp5 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tl.store(out_ptr0 + x2, tmp6, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_sum_0[grid(64)](arg0_1, buf0, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del arg0_1
return buf0,
class SumAggregatorNew(nn.Module):
def __init__(self):
super(SumAggregatorNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| AlexMinhao/NAS_GNN | SumAggregator | false | 0 | [
"Apache-2.0"
] | 0 | 89183988a96e1d6baed910ab3843c13282f8b077 | https://github.com/AlexMinhao/NAS_GNN/tree/89183988a96e1d6baed910ab3843c13282f8b077 | [
{
"content": "entry_point: SumAggregator\n\n# AOT ID: ['0_inference']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\n\n\n# kernel path: runs/run_shard_6/inductor_cache/7o/c7otc5ij6whexgxcr56vlxp2l7hzg3oc4onljp557uc6wncu5gvg.py\n# Topologically Sorted Source Nodes: [sum_1], Original ATen: [aten.sum]\n# Source node to ATen node mapping:\n# sum_1 => sum_1\n# Graph fragment:\n# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%arg0_1, [1]), kwargs = {})\ntriton_poi_fused_sum_0 = async_compile.triton('triton_poi_fused_sum_0', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[64], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_sum_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 64\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 16\n x1 = (xindex // 16)\n x2 = xindex\n tmp0 = tl.load(in_ptr0 + (x0 + (64*x1)), xmask)\n tmp1 = tl.load(in_ptr0 + (16 + x0 + (64*x1)), xmask)\n tmp3 = tl.load(in_ptr0 + (32 + x0 + (64*x1)), xmask)\n tmp5 = tl.load(in_ptr0 + (48 + x0 + (64*x1)), xmask)\n tmp2 = tmp0 + tmp1\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tl.store(out_ptr0 + (x2), tmp6, xmask)\n''', device_str='cuda')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)\n # Topologically Sorted Source Nodes: [sum_1], Original ATen: [aten.sum]\n stream0 = get_raw_stream(0)\n triton_poi_fused_sum_0.run(arg0_1, buf0, 64, grid=grid(64), stream=stream0)\n del arg0_1\n return (buf0, )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)\n fn = lambda: call([arg0_1])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == \"__main__\":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main('None', benchmark_compiled_module)\n",
"role": "system",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_sum_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 64\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 16\n x1 = xindex // 16\n x2 = xindex\n tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)\n tmp1 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)\n tmp3 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)\n tmp5 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)\n tmp2 = tmp0 + tmp1\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tl.store(out_ptr0 + x2, tmp6, xmask)\n\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_sum_0[grid(64)](arg0_1, buf0, 64, XBLOCK=64,\n num_warps=1, num_stages=1)\n del arg0_1\n return buf0,\n\n\nclass SumAggregatorNew(nn.Module):\n\n def __init__(self):\n super(SumAggregatorNew, self).__init__()\n\n def forward(self, input_0):\n arg0_1 = input_0\n output = call([arg0_1])\n return output[0]\n",
"role": "user",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_sum_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 64\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 16\n x1 = xindex // 16\n x2 = xindex\n tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)\n tmp1 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)\n tmp3 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)\n tmp5 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)\n tmp2 = tmp0 + tmp1\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tl.store(out_ptr0 + x2, tmp6, xmask)\n\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_sum_0[grid(64)](arg0_1, buf0, 64, XBLOCK=64,\n num_warps=1, num_stages=1)\n del arg0_1\n return buf0,\n\n\nclass SumAggregatorNew(nn.Module):\n\n def __init__(self):\n super(SumAggregatorNew, self).__init__()\n\n def forward(self, input_0):\n arg0_1 = input_0\n output = call([arg0_1])\n return output[0]\n",
"role": "assistant",
"thinking": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_sum_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 64\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 16\n x1 = xindex // 16\n x2 = xindex\n tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)\n tmp1 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)\n tmp3 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)\n tmp5 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)\n tmp2 = tmp0 + tmp1\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tl.store(out_ptr0 + x2, tmp6, xmask)\n\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_sum_0[grid(64)](arg0_1, buf0, 64, XBLOCK=64,\n num_warps=1, num_stages=1)\n del arg0_1\n return buf0,\n\n\nclass SumAggregatorNew(nn.Module):\n\n def __init__(self):\n super(SumAggregatorNew, self).__init__()\n\n def forward(self, input_0):\n arg0_1 = input_0\n output = call([arg0_1])\n return output[0]\n"
}
] |
LinearEmbedding | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/xz/cxz7i3qbiizfbbzvas22bbwy5nxzvmtfdg5vhhiye56dk4hdonst.py
# Topologically Sorted Source Nodes: [mul], Original ATen: [aten.mul]
# Source node to ATen node mapping:
# mul => mul
# Graph fragment:
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 2.0), kwargs = {})
triton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mul_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 2.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + (x2), tmp4, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
# Topologically Sorted Source Nodes: [], Original ATen: []
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse
# Topologically Sorted Source Nodes: [mul], Original ATen: [aten.mul]
stream0 = get_raw_stream(0)
triton_poi_fused_mul_0.run(buf1, primals_2, 256, grid=grid(256), stream=stream0)
del primals_2
return (buf1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import math
import torch
import torch.utils.data
import torch.nn as nn
class LinearEmbedding(nn.Module):
def __init__(self, inp_size, d_model):
super(LinearEmbedding, self).__init__()
self.lut = nn.Linear(inp_size, d_model)
self.d_model = d_model
def forward(self, x):
return self.lut(x) * math.sqrt(self.d_model)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'inp_size': 4, 'd_model': 4}]
| import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.utils.data
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 2.0
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_mul_0[grid(256)](buf1, primals_2, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_2
return buf1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0)
class LinearEmbeddingNew(nn.Module):
def __init__(self, inp_size, d_model):
super(LinearEmbeddingNew, self).__init__()
self.lut = nn.Linear(inp_size, d_model)
self.d_model = d_model
def forward(self, input_0):
primals_1 = self.lut.weight
primals_2 = self.lut.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Akhil-Raj/Trajectory-Transformer | LinearEmbedding | false | 1 | [
"MIT"
] | 0 | dd09fda99443f6afb59d962026573162219ea6a9 | https://github.com/Akhil-Raj/Trajectory-Transformer/tree/dd09fda99443f6afb59d962026573162219ea6a9 | [
{
"content": "entry_point: LinearEmbedding\n\n# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\n\n\n# kernel path: runs/run_shard_6/inductor_cache/xz/cxz7i3qbiizfbbzvas22bbwy5nxzvmtfdg5vhhiye56dk4hdonst.py\n# Topologically Sorted Source Nodes: [mul], Original ATen: [aten.mul]\n# Source node to ATen node mapping:\n# mul => mul\n# Graph fragment:\n# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%view_1, 2.0), kwargs = {})\ntriton_poi_fused_mul_0 = async_compile.triton('triton_poi_fused_mul_0', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[256], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_mul_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x0 = xindex % 4\n tmp0 = tl.load(in_out_ptr0 + (x2), xmask)\n tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')\n tmp2 = tmp0 + tmp1\n tmp3 = 2.0\n tmp4 = tmp2 * tmp3\n tl.store(in_out_ptr0 + (x2), tmp4, xmask)\n''', device_str='cuda')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4), (4, 1))\n assert_size_stride(primals_2, (4, ), (1, ))\n assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)\n # Topologically Sorted Source Nodes: [], Original ATen: []\n extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)\n del primals_1\n buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0); del buf0 # reuse\n # Topologically Sorted Source Nodes: [mul], Original ATen: [aten.mul]\n stream0 = get_raw_stream(0)\n triton_poi_fused_mul_0.run(buf1, primals_2, 256, grid=grid(256), stream=stream0)\n del primals_2\n return (buf1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n primals_1 = rand_strided((4, 4), (4, 1), device='cuda:0', dtype=torch.float32)\n primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n primals_3 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)\n fn = lambda: call([primals_1, primals_2, primals_3])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == \"__main__\":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main('None', benchmark_compiled_module)\n",
"role": "system",
"thinking": null
},
{
"content": "import torch\nfrom torch._inductor.select_algorithm import extern_kernels\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nimport torch.utils.data\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\n\n\n@triton.jit\ndef triton_poi_fused_mul_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x0 = xindex % 4\n tmp0 = tl.load(in_out_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp0 + tmp1\n tmp3 = 2.0\n tmp4 = tmp2 * tmp3\n tl.store(in_out_ptr0 + x2, tmp4, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4), (4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)\n extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),\n reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)\n del primals_1\n buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)\n del buf0\n get_raw_stream(0)\n triton_poi_fused_mul_0[grid(256)](buf1, primals_2, 256, XBLOCK=256,\n num_warps=4, num_stages=1)\n del primals_2\n return buf1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0)\n\n\nclass LinearEmbeddingNew(nn.Module):\n\n def __init__(self, inp_size, d_model):\n super(LinearEmbeddingNew, self).__init__()\n self.lut = nn.Linear(inp_size, d_model)\n self.d_model = d_model\n\n def forward(self, input_0):\n primals_1 = self.lut.weight\n primals_2 = self.lut.bias\n primals_3 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "user",
"thinking": null
},
{
"content": "import torch\nfrom torch._inductor.select_algorithm import extern_kernels\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nimport torch.utils.data\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\n\n\n@triton.jit\ndef triton_poi_fused_mul_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x0 = xindex % 4\n tmp0 = tl.load(in_out_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp0 + tmp1\n tmp3 = 2.0\n tmp4 = tmp2 * tmp3\n tl.store(in_out_ptr0 + x2, tmp4, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4), (4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)\n extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),\n reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)\n del primals_1\n buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)\n del buf0\n get_raw_stream(0)\n triton_poi_fused_mul_0[grid(256)](buf1, primals_2, 256, XBLOCK=256,\n num_warps=4, num_stages=1)\n del primals_2\n return buf1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0)\n\n\nclass LinearEmbeddingNew(nn.Module):\n\n def __init__(self, inp_size, d_model):\n super(LinearEmbeddingNew, self).__init__()\n self.lut = nn.Linear(inp_size, d_model)\n self.d_model = d_model\n\n def forward(self, input_0):\n primals_1 = self.lut.weight\n primals_2 = self.lut.bias\n primals_3 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "assistant",
"thinking": "import torch\nfrom torch._inductor.select_algorithm import extern_kernels\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nimport torch.utils.data\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\n\n\n@triton.jit\ndef triton_poi_fused_mul_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x0 = xindex % 4\n tmp0 = tl.load(in_out_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp0 + tmp1\n tmp3 = 2.0\n tmp4 = tmp2 * tmp3\n tl.store(in_out_ptr0 + x2, tmp4, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4), (4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)\n extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),\n reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)\n del primals_1\n buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)\n del buf0\n get_raw_stream(0)\n triton_poi_fused_mul_0[grid(256)](buf1, primals_2, 256, XBLOCK=256,\n num_warps=4, num_stages=1)\n del primals_2\n return buf1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0)\n\n\nclass LinearEmbeddingNew(nn.Module):\n\n def __init__(self, inp_size, d_model):\n super(LinearEmbeddingNew, self).__init__()\n self.lut = nn.Linear(inp_size, d_model)\n self.d_model = d_model\n\n def forward(self, input_0):\n primals_1 = self.lut.weight\n primals_2 = self.lut.bias\n primals_3 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n"
}
] |
CustomizeLayer | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/ne/cnem3ewvyvy2ry3vceibv7j67edhj5jen3qlzvzibg5tpkhxkwsq.py
# Topologically Sorted Source Nodes: [pow_1, sum_1, norm, x, mul, add], Original ATen: [aten.pow, aten.sum, aten.sqrt, aten.div, aten.mul, aten.add]
# Source node to ATen node mapping:
# add => add
# mul => mul
# norm => sqrt
# pow_1 => pow_1
# sum_1 => sum_1
# x => div
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_1, 2), kwargs = {})
# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [1], True), kwargs = {})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%sum_1,), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_1, %sqrt), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%div, %primals_2), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %primals_3), kwargs = {})
triton_poi_fused_add_div_mul_pow_sqrt_sum_0 = async_compile.triton('triton_poi_fused_add_div_mul_pow_sqrt_sum_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mul_pow_sqrt_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_mul_pow_sqrt_sum_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x3 = (xindex // 64)
x5 = xindex % 16
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x4), xmask)
tmp1 = tl.load(in_ptr0 + (x5 + (64*x3)), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x5 + (64*x3)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x5 + (64*x3)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x5 + (64*x3)), xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = tmp0 / tmp12
tmp15 = tmp13 * tmp14
tmp17 = tmp15 + tmp16
tl.store(out_ptr0 + (x4), tmp17, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [pow_1, sum_1, norm, x, mul, add], Original ATen: [aten.pow, aten.sum, aten.sqrt, aten.div, aten.mul, aten.add]
stream0 = get_raw_stream(0)
triton_poi_fused_add_div_mul_pow_sqrt_sum_0.run(primals_1, primals_2, primals_3, buf0, 256, grid=grid(256), stream=stream0)
del primals_2
del primals_3
return (buf0, primals_1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class CustomizeLayer(nn.Module):
def __init__(self, in_dim):
super().__init__()
self.in_dim = in_dim
self.scale = nn.Parameter(torch.Tensor(self.in_dim))
self.bias = nn.Parameter(torch.Tensor(self.in_dim))
def forward(self, x):
norm = x.pow(2).sum(dim=1, keepdim=True).sqrt()
x = torch.div(x, norm)
return x * self.scale + self.bias
def __repr__(self):
return 'CustomizedLayer(in_dim=%d)' % self.in_dim
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_dim': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_div_mul_pow_sqrt_sum_0(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x3 = xindex // 64
x5 = xindex % 16
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr0 + (x5 + 64 * x3), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr0 + (16 + x5 + 64 * x3), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (32 + x5 + 64 * x3), xmask, eviction_policy=
'evict_last')
tmp9 = tl.load(in_ptr0 + (48 + x5 + 64 * x3), xmask, eviction_policy=
'evict_last')
tmp14 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp16 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = libdevice.sqrt(tmp11)
tmp13 = tmp0 / tmp12
tmp15 = tmp13 * tmp14
tmp17 = tmp15 + tmp16
tl.store(out_ptr0 + x4, tmp17, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_mul_pow_sqrt_sum_0[grid(256)](primals_1,
primals_2, primals_3, buf0, 256, XBLOCK=128, num_warps=4,
num_stages=1)
del primals_2
del primals_3
return buf0, primals_1
class CustomizeLayerNew(nn.Module):
def __init__(self, in_dim):
super().__init__()
self.in_dim = in_dim
self.scale = nn.Parameter(torch.Tensor(self.in_dim))
self.bias = nn.Parameter(torch.Tensor(self.in_dim))
def __repr__(self):
return 'CustomizedLayer(in_dim=%d)' % self.in_dim
def forward(self, input_0):
primals_2 = self.scale
primals_3 = self.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Abhishekvats1997/Torch-Pruning | CustomizeLayer | false | 2 | [
"MIT"
] | 0 | b322a42d1c9032cc9644332d33a9662ca6ed44ac | https://github.com/Abhishekvats1997/Torch-Pruning/tree/b322a42d1c9032cc9644332d33a9662ca6ed44ac | [
{
"content": "entry_point: CustomizeLayer\n\n# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\n\n\n# kernel path: runs/run_shard_6/inductor_cache/ne/cnem3ewvyvy2ry3vceibv7j67edhj5jen3qlzvzibg5tpkhxkwsq.py\n# Topologically Sorted Source Nodes: [pow_1, sum_1, norm, x, mul, add], Original ATen: [aten.pow, aten.sum, aten.sqrt, aten.div, aten.mul, aten.add]\n# Source node to ATen node mapping:\n# add => add\n# mul => mul\n# norm => sqrt\n# pow_1 => pow_1\n# sum_1 => sum_1\n# x => div\n# Graph fragment:\n# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%primals_1, 2), kwargs = {})\n# %sum_1 : [num_users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%pow_1, [1], True), kwargs = {})\n# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%sum_1,), kwargs = {})\n# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%primals_1, %sqrt), kwargs = {})\n# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%div, %primals_2), kwargs = {})\n# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %primals_3), kwargs = {})\ntriton_poi_fused_add_div_mul_pow_sqrt_sum_0 = async_compile.triton('triton_poi_fused_add_div_mul_pow_sqrt_sum_0', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[256], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mul_pow_sqrt_sum_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_add_div_mul_pow_sqrt_sum_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x4 = xindex\n x3 = (xindex // 64)\n x5 = xindex % 16\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + (x4), xmask)\n tmp1 = tl.load(in_ptr0 + (x5 + (64*x3)), xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr0 + (16 + x5 + (64*x3)), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (32 + x5 + (64*x3)), xmask, eviction_policy='evict_last')\n tmp9 = tl.load(in_ptr0 + (48 + x5 + (64*x3)), xmask, eviction_policy='evict_last')\n tmp14 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')\n tmp16 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = libdevice.sqrt(tmp11)\n tmp13 = tmp0 / tmp12\n tmp15 = tmp13 * tmp14\n tmp17 = tmp15 + tmp16\n tl.store(out_ptr0 + (x4), tmp17, xmask)\n''', device_str='cuda')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4, ), (1, ))\n assert_size_stride(primals_3, (4, ), (1, ))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n # Topologically Sorted Source Nodes: [pow_1, sum_1, norm, x, mul, add], Original ATen: [aten.pow, aten.sum, aten.sqrt, aten.div, aten.mul, aten.add]\n stream0 = get_raw_stream(0)\n triton_poi_fused_add_div_mul_pow_sqrt_sum_0.run(primals_1, primals_2, primals_3, buf0, 256, grid=grid(256), stream=stream0)\n del primals_2\n del primals_3\n return (buf0, primals_1, )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)\n primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n fn = lambda: call([primals_1, primals_2, primals_3])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == \"__main__\":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main('None', benchmark_compiled_module)\n",
"role": "system",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mul_pow_sqrt_sum_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x4 = xindex\n x3 = xindex // 64\n x5 = xindex % 16\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + x4, xmask)\n tmp1 = tl.load(in_ptr0 + (x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp3 = tl.load(in_ptr0 + (16 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp6 = tl.load(in_ptr0 + (32 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp9 = tl.load(in_ptr0 + (48 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp14 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')\n tmp16 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = libdevice.sqrt(tmp11)\n tmp13 = tmp0 / tmp12\n tmp15 = tmp13 * tmp14\n tmp17 = tmp15 + tmp16\n tl.store(out_ptr0 + x4, tmp17, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mul_pow_sqrt_sum_0[grid(256)](primals_1,\n primals_2, primals_3, buf0, 256, XBLOCK=128, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass CustomizeLayerNew(nn.Module):\n\n def __init__(self, in_dim):\n super().__init__()\n self.in_dim = in_dim\n self.scale = nn.Parameter(torch.Tensor(self.in_dim))\n self.bias = nn.Parameter(torch.Tensor(self.in_dim))\n\n def __repr__(self):\n return 'CustomizedLayer(in_dim=%d)' % self.in_dim\n\n def forward(self, input_0):\n primals_2 = self.scale\n primals_3 = self.bias\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "user",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mul_pow_sqrt_sum_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x4 = xindex\n x3 = xindex // 64\n x5 = xindex % 16\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + x4, xmask)\n tmp1 = tl.load(in_ptr0 + (x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp3 = tl.load(in_ptr0 + (16 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp6 = tl.load(in_ptr0 + (32 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp9 = tl.load(in_ptr0 + (48 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp14 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')\n tmp16 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = libdevice.sqrt(tmp11)\n tmp13 = tmp0 / tmp12\n tmp15 = tmp13 * tmp14\n tmp17 = tmp15 + tmp16\n tl.store(out_ptr0 + x4, tmp17, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mul_pow_sqrt_sum_0[grid(256)](primals_1,\n primals_2, primals_3, buf0, 256, XBLOCK=128, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass CustomizeLayerNew(nn.Module):\n\n def __init__(self, in_dim):\n super().__init__()\n self.in_dim = in_dim\n self.scale = nn.Parameter(torch.Tensor(self.in_dim))\n self.bias = nn.Parameter(torch.Tensor(self.in_dim))\n\n def __repr__(self):\n return 'CustomizedLayer(in_dim=%d)' % self.in_dim\n\n def forward(self, input_0):\n primals_2 = self.scale\n primals_3 = self.bias\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "assistant",
"thinking": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mul_pow_sqrt_sum_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x4 = xindex\n x3 = xindex // 64\n x5 = xindex % 16\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + x4, xmask)\n tmp1 = tl.load(in_ptr0 + (x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp3 = tl.load(in_ptr0 + (16 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp6 = tl.load(in_ptr0 + (32 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp9 = tl.load(in_ptr0 + (48 + x5 + 64 * x3), xmask, eviction_policy=\n 'evict_last')\n tmp14 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')\n tmp16 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = libdevice.sqrt(tmp11)\n tmp13 = tmp0 / tmp12\n tmp15 = tmp13 * tmp14\n tmp17 = tmp15 + tmp16\n tl.store(out_ptr0 + x4, tmp17, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mul_pow_sqrt_sum_0[grid(256)](primals_1,\n primals_2, primals_3, buf0, 256, XBLOCK=128, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass CustomizeLayerNew(nn.Module):\n\n def __init__(self, in_dim):\n super().__init__()\n self.in_dim = in_dim\n self.scale = nn.Parameter(torch.Tensor(self.in_dim))\n self.bias = nn.Parameter(torch.Tensor(self.in_dim))\n\n def __repr__(self):\n return 'CustomizedLayer(in_dim=%d)' % self.in_dim\n\n def forward(self, input_0):\n primals_2 = self.scale\n primals_3 = self.bias\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n"
}
] |
LayerNorm | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/yq/cyqficfzvi7m26ugfmcjnrvh3fwc36rw64jccffjyl3m2zuwe2to.py
# Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# add_1 => add_1
# mean => mean
# mul => mul
# std => sqrt, var
# sub => sub
# truediv => div
# Graph fragment:
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%primals_1, [-1], True), kwargs = {})
# %var : [num_users=1] = call_function[target=torch.ops.aten.var.correction](args = (%primals_1, [-1]), kwargs = {correction: 1.0, keepdim: True})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%var,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %mean), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %sub), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sqrt, 1e-05), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%mul, %add), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %primals_3), kwargs = {})
triton_poi_fused_add_div_mean_mul_std_sub_0 = async_compile.triton('triton_poi_fused_add_div_mean_mul_std_sub_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mean_mul_std_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (x2), xmask)
tmp2 = tl.load(in_ptr1 + (4*x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp30 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp8 = tmp6 + tmp7
tmp9 = 4.0
tmp10 = tmp8 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp0 * tmp11
tmp13 = tmp2 - tmp10
tmp14 = tmp13 * tmp13
tmp15 = tmp3 - tmp10
tmp16 = tmp15 * tmp15
tmp17 = tmp14 + tmp16
tmp18 = tmp5 - tmp10
tmp19 = tmp18 * tmp18
tmp20 = tmp17 + tmp19
tmp21 = tmp7 - tmp10
tmp22 = tmp21 * tmp21
tmp23 = tmp20 + tmp22
tmp24 = 3.0
tmp25 = tmp23 / tmp24
tmp26 = libdevice.sqrt(tmp25)
tmp27 = 1e-05
tmp28 = tmp26 + tmp27
tmp29 = tmp12 / tmp28
tmp31 = tmp29 + tmp30
tl.store(out_ptr0 + (x2), tmp31, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_add_div_mean_mul_std_sub_0.run(primals_2, primals_1, primals_3, buf0, 256, grid=grid(256), stream=stream0)
del primals_2
del primals_3
return (buf0, primals_1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class LayerNorm(nn.Module):
def __init__(self, weights, eps=1e-05):
super().__init__()
self.gamma = nn.Parameter(torch.ones(weights))
self.beta = nn.Parameter(torch.zeros(weights))
self.eps = eps
def forward(self, x):
mean = x.mean(-1, keepdim=True)
std = x.std(-1, keepdim=True)
return self.gamma * (x - mean) / (std + self.eps) + self.beta
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'weights': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp8 = tmp6 + tmp7
tmp9 = 4.0
tmp10 = tmp8 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp0 * tmp11
tmp13 = tmp2 - tmp10
tmp14 = tmp13 * tmp13
tmp15 = tmp3 - tmp10
tmp16 = tmp15 * tmp15
tmp17 = tmp14 + tmp16
tmp18 = tmp5 - tmp10
tmp19 = tmp18 * tmp18
tmp20 = tmp17 + tmp19
tmp21 = tmp7 - tmp10
tmp22 = tmp21 * tmp21
tmp23 = tmp20 + tmp22
tmp24 = 3.0
tmp25 = tmp23 / tmp24
tmp26 = libdevice.sqrt(tmp25)
tmp27 = 1e-05
tmp28 = tmp26 + tmp27
tmp29 = tmp12 / tmp28
tmp31 = tmp29 + tmp30
tl.store(out_ptr0 + x2, tmp31, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,
primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del primals_2
del primals_3
return buf0, primals_1
class LayerNormNew(nn.Module):
def __init__(self, weights, eps=1e-05):
super().__init__()
self.gamma = nn.Parameter(torch.ones(weights))
self.beta = nn.Parameter(torch.zeros(weights))
self.eps = eps
def forward(self, input_0):
primals_2 = self.gamma
primals_3 = self.beta
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| AWilcke/Dissertation | LayerNorm | false | 3 | [
"MIT"
] | 0 | b85ad38a7f336ee290d5883f5e942f54e140d0d0 | https://github.com/AWilcke/Dissertation/tree/b85ad38a7f336ee290d5883f5e942f54e140d0d0 | [
{
"content": "entry_point: LayerNorm\n\n# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\n\n\n# kernel path: runs/run_shard_6/inductor_cache/yq/cyqficfzvi7m26ugfmcjnrvh3fwc36rw64jccffjyl3m2zuwe2to.py\n# Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]\n# Source node to ATen node mapping:\n# add => add\n# add_1 => add_1\n# mean => mean\n# mul => mul\n# std => sqrt, var\n# sub => sub\n# truediv => div\n# Graph fragment:\n# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%primals_1, [-1], True), kwargs = {})\n# %var : [num_users=1] = call_function[target=torch.ops.aten.var.correction](args = (%primals_1, [-1]), kwargs = {correction: 1.0, keepdim: True})\n# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%var,), kwargs = {})\n# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %mean), kwargs = {})\n# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %sub), kwargs = {})\n# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sqrt, 1e-05), kwargs = {})\n# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%mul, %add), kwargs = {})\n# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %primals_3), kwargs = {})\ntriton_poi_fused_add_div_mean_mul_std_sub_0 = async_compile.triton('triton_poi_fused_add_div_mean_mul_std_sub_0', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[256], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mean_mul_std_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = (xindex // 4)\n tmp0 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + (x2), xmask)\n tmp2 = tl.load(in_ptr1 + (4*x1), xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-05\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + (x2), tmp31, xmask)\n''', device_str='cuda')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4, ), (1, ))\n assert_size_stride(primals_3, (4, ), (1, ))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n # Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]\n stream0 = get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0.run(primals_2, primals_1, primals_3, buf0, 256, grid=grid(256), stream=stream0)\n del primals_2\n del primals_3\n return (buf0, primals_1, )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)\n primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n fn = lambda: call([primals_1, primals_2, primals_3])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == \"__main__\":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main('None', benchmark_compiled_module)\n",
"role": "system",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + x2, xmask)\n tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-05\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + x2, tmp31, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,\n primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass LayerNormNew(nn.Module):\n\n def __init__(self, weights, eps=1e-05):\n super().__init__()\n self.gamma = nn.Parameter(torch.ones(weights))\n self.beta = nn.Parameter(torch.zeros(weights))\n self.eps = eps\n\n def forward(self, input_0):\n primals_2 = self.gamma\n primals_3 = self.beta\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "user",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + x2, xmask)\n tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-05\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + x2, tmp31, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,\n primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass LayerNormNew(nn.Module):\n\n def __init__(self, weights, eps=1e-05):\n super().__init__()\n self.gamma = nn.Parameter(torch.ones(weights))\n self.beta = nn.Parameter(torch.zeros(weights))\n self.eps = eps\n\n def forward(self, input_0):\n primals_2 = self.gamma\n primals_3 = self.beta\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "assistant",
"thinking": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + x2, xmask)\n tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-05\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + x2, tmp31, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,\n primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass LayerNormNew(nn.Module):\n\n def __init__(self, weights, eps=1e-05):\n super().__init__()\n self.gamma = nn.Parameter(torch.ones(weights))\n self.beta = nn.Parameter(torch.zeros(weights))\n self.eps = eps\n\n def forward(self, input_0):\n primals_2 = self.gamma\n primals_3 = self.beta\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n"
}
] |
LayerNorm | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/df/cdfcie57v6pcdd6oeaz4mvlgksxgyuxzmlv5bklwemyulqhtcxta.py
# Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]
# Source node to ATen node mapping:
# add => add
# add_1 => add_1
# mean => mean
# mul => mul
# std => sqrt, var
# sub => sub
# truediv => div
# Graph fragment:
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%primals_1, [-1], True), kwargs = {})
# %var : [num_users=1] = call_function[target=torch.ops.aten.var.correction](args = (%primals_1, [-1]), kwargs = {correction: 1.0, keepdim: True})
# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%var,), kwargs = {})
# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %mean), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %sub), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sqrt, 1e-06), kwargs = {})
# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%mul, %add), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %primals_3), kwargs = {})
triton_poi_fused_add_div_mean_mul_std_sub_0 = async_compile.triton('triton_poi_fused_add_div_mean_mul_std_sub_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mean_mul_std_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (x2), xmask)
tmp2 = tl.load(in_ptr1 + (4*x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp30 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp8 = tmp6 + tmp7
tmp9 = 4.0
tmp10 = tmp8 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp0 * tmp11
tmp13 = tmp2 - tmp10
tmp14 = tmp13 * tmp13
tmp15 = tmp3 - tmp10
tmp16 = tmp15 * tmp15
tmp17 = tmp14 + tmp16
tmp18 = tmp5 - tmp10
tmp19 = tmp18 * tmp18
tmp20 = tmp17 + tmp19
tmp21 = tmp7 - tmp10
tmp22 = tmp21 * tmp21
tmp23 = tmp20 + tmp22
tmp24 = 3.0
tmp25 = tmp23 / tmp24
tmp26 = libdevice.sqrt(tmp25)
tmp27 = 1e-06
tmp28 = tmp26 + tmp27
tmp29 = tmp12 / tmp28
tmp31 = tmp29 + tmp30
tl.store(out_ptr0 + (x2), tmp31, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]
stream0 = get_raw_stream(0)
triton_poi_fused_add_div_mean_mul_std_sub_0.run(primals_2, primals_1, primals_3, buf0, 256, grid=grid(256), stream=stream0)
del primals_2
del primals_3
return (buf0, primals_1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.utils.data
import torch.nn as nn
class LayerNorm(nn.Module):
"""
Construct a layernorm module (See citation for details).
"""
def __init__(self, features, eps=1e-06):
super(LayerNorm, self).__init__()
self.a_2 = nn.Parameter(torch.ones(features))
self.b_2 = nn.Parameter(torch.zeros(features))
self.eps = eps
def forward(self, x):
mean = x.mean(dim=-1, keepdim=True)
std = x.std(dim=-1, keepdim=True)
return self.a_2 * (x - mean) / (std + self.eps) + self.b_2
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'features': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.utils.data
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp8 = tmp6 + tmp7
tmp9 = 4.0
tmp10 = tmp8 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp0 * tmp11
tmp13 = tmp2 - tmp10
tmp14 = tmp13 * tmp13
tmp15 = tmp3 - tmp10
tmp16 = tmp15 * tmp15
tmp17 = tmp14 + tmp16
tmp18 = tmp5 - tmp10
tmp19 = tmp18 * tmp18
tmp20 = tmp17 + tmp19
tmp21 = tmp7 - tmp10
tmp22 = tmp21 * tmp21
tmp23 = tmp20 + tmp22
tmp24 = 3.0
tmp25 = tmp23 / tmp24
tmp26 = libdevice.sqrt(tmp25)
tmp27 = 1e-06
tmp28 = tmp26 + tmp27
tmp29 = tmp12 / tmp28
tmp31 = tmp29 + tmp30
tl.store(out_ptr0 + x2, tmp31, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,
primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del primals_2
del primals_3
return buf0, primals_1
class LayerNormNew(nn.Module):
"""
Construct a layernorm module (See citation for details).
"""
def __init__(self, features, eps=1e-06):
super(LayerNormNew, self).__init__()
self.a_2 = nn.Parameter(torch.ones(features))
self.b_2 = nn.Parameter(torch.zeros(features))
self.eps = eps
def forward(self, input_0):
primals_2 = self.a_2
primals_3 = self.b_2
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Akhil-Raj/Trajectory-Transformer | LayerNorm | false | 4 | [
"MIT"
] | 0 | dd09fda99443f6afb59d962026573162219ea6a9 | https://github.com/Akhil-Raj/Trajectory-Transformer/tree/dd09fda99443f6afb59d962026573162219ea6a9 | [
{
"content": "entry_point: LayerNorm\n\n# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\n\n\n# kernel path: runs/run_shard_6/inductor_cache/df/cdfcie57v6pcdd6oeaz4mvlgksxgyuxzmlv5bklwemyulqhtcxta.py\n# Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]\n# Source node to ATen node mapping:\n# add => add\n# add_1 => add_1\n# mean => mean\n# mul => mul\n# std => sqrt, var\n# sub => sub\n# truediv => div\n# Graph fragment:\n# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%primals_1, [-1], True), kwargs = {})\n# %var : [num_users=1] = call_function[target=torch.ops.aten.var.correction](args = (%primals_1, [-1]), kwargs = {correction: 1.0, keepdim: True})\n# %sqrt : [num_users=1] = call_function[target=torch.ops.aten.sqrt.default](args = (%var,), kwargs = {})\n# %sub : [num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %mean), kwargs = {})\n# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %sub), kwargs = {})\n# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%sqrt, 1e-06), kwargs = {})\n# %div : [num_users=1] = call_function[target=torch.ops.aten.div.Tensor](args = (%mul, %add), kwargs = {})\n# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %primals_3), kwargs = {})\ntriton_poi_fused_add_div_mean_mul_std_sub_0 = async_compile.triton('triton_poi_fused_add_div_mean_mul_std_sub_0', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[256], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_div_mean_mul_std_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = (xindex // 4)\n tmp0 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + (x2), xmask)\n tmp2 = tl.load(in_ptr1 + (4*x1), xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-06\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + (x2), tmp31, xmask)\n''', device_str='cuda')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4, ), (1, ))\n assert_size_stride(primals_3, (4, ), (1, ))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n # Topologically Sorted Source Nodes: [mean, std, sub, mul, add, truediv, add_1], Original ATen: [aten.mean, aten.std, aten.sub, aten.mul, aten.add, aten.div]\n stream0 = get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0.run(primals_2, primals_1, primals_3, buf0, 256, grid=grid(256), stream=stream0)\n del primals_2\n del primals_3\n return (buf0, primals_1, )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)\n primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n fn = lambda: call([primals_1, primals_2, primals_3])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == \"__main__\":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main('None', benchmark_compiled_module)\n",
"role": "system",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.utils.data\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + x2, xmask)\n tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-06\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + x2, tmp31, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,\n primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass LayerNormNew(nn.Module):\n \"\"\"\n Construct a layernorm module (See citation for details).\n \"\"\"\n\n def __init__(self, features, eps=1e-06):\n super(LayerNormNew, self).__init__()\n self.a_2 = nn.Parameter(torch.ones(features))\n self.b_2 = nn.Parameter(torch.zeros(features))\n self.eps = eps\n\n def forward(self, input_0):\n primals_2 = self.a_2\n primals_3 = self.b_2\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "user",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.utils.data\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + x2, xmask)\n tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-06\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + x2, tmp31, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,\n primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass LayerNormNew(nn.Module):\n \"\"\"\n Construct a layernorm module (See citation for details).\n \"\"\"\n\n def __init__(self, features, eps=1e-06):\n super(LayerNormNew, self).__init__()\n self.a_2 = nn.Parameter(torch.ones(features))\n self.b_2 = nn.Parameter(torch.zeros(features))\n self.eps = eps\n\n def forward(self, input_0):\n primals_2 = self.a_2\n primals_3 = self.b_2\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "assistant",
"thinking": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.utils.data\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 4\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')\n tmp1 = tl.load(in_ptr1 + x2, xmask)\n tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp4 = tmp2 + tmp3\n tmp6 = tmp4 + tmp5\n tmp8 = tmp6 + tmp7\n tmp9 = 4.0\n tmp10 = tmp8 / tmp9\n tmp11 = tmp1 - tmp10\n tmp12 = tmp0 * tmp11\n tmp13 = tmp2 - tmp10\n tmp14 = tmp13 * tmp13\n tmp15 = tmp3 - tmp10\n tmp16 = tmp15 * tmp15\n tmp17 = tmp14 + tmp16\n tmp18 = tmp5 - tmp10\n tmp19 = tmp18 * tmp18\n tmp20 = tmp17 + tmp19\n tmp21 = tmp7 - tmp10\n tmp22 = tmp21 * tmp21\n tmp23 = tmp20 + tmp22\n tmp24 = 3.0\n tmp25 = tmp23 / tmp24\n tmp26 = libdevice.sqrt(tmp25)\n tmp27 = 1e-06\n tmp28 = tmp26 + tmp27\n tmp29 = tmp12 / tmp28\n tmp31 = tmp29 + tmp30\n tl.store(out_ptr0 + x2, tmp31, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_add_div_mean_mul_std_sub_0[grid(256)](primals_2,\n primals_1, primals_3, buf0, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del primals_2\n del primals_3\n return buf0, primals_1\n\n\nclass LayerNormNew(nn.Module):\n \"\"\"\n Construct a layernorm module (See citation for details).\n \"\"\"\n\n def __init__(self, features, eps=1e-06):\n super(LayerNormNew, self).__init__()\n self.a_2 = nn.Parameter(torch.ones(features))\n self.b_2 = nn.Parameter(torch.zeros(features))\n self.eps = eps\n\n def forward(self, input_0):\n primals_2 = self.a_2\n primals_3 = self.b_2\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n"
}
] |
Norm | # AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/dg/cdgw6x7nju4bzp2wyuwgeanbco7zcjis6yiusovvnpz6zw3yjd3l.py
# Topologically Sorted Source Nodes: [u, sub], Original ATen: [aten.mean, aten.sub]
# Source node to ATen node mapping:
# sub => sub
# u => mean
# Graph fragment:
# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%primals_1, [-1], True), kwargs = {})
# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %mean), kwargs = {})
triton_poi_fused_mean_sub_0 = async_compile.triton('triton_poi_fused_mean_sub_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mean_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = 4.0
tmp9 = tmp7 / tmp8
tmp10 = tmp0 - tmp9
tl.store(out_ptr0 + (x2), tmp10, xmask)
''', device_str='cuda')
# kernel path: runs/run_shard_6/inductor_cache/v3/cv3kidzozprmsfiixppams3wqi7ij566hegofbo3ur5gfdrhxiik.py
# Topologically Sorted Source Nodes: [square, s, add, rsqrt, x, mul_1, x_1], Original ATen: [aten.pow, aten.mean, aten.add, aten.rsqrt, aten.mul]
# Source node to ATen node mapping:
# add => add
# mul_1 => mul_1
# rsqrt => rsqrt
# s => mean_1
# square => pow_1
# x => mul
# x_1 => add_1
# Graph fragment:
# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})
# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%pow_1, [-1], True), kwargs = {})
# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean_1, 1e-05), kwargs = {})
# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})
# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub, %rsqrt), kwargs = {})
# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_2), kwargs = {})
# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %primals_3), kwargs = {})
triton_poi_fused_add_mean_mul_pow_rsqrt_1 = async_compile.triton('triton_poi_fused_add_mean_mul_pow_rsqrt_1', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[256],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mean_mul_pow_rsqrt_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_add_mean_mul_pow_rsqrt_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = (xindex // 4)
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + (x2), xmask)
tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
tmp20 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = 4.0
tmp13 = tmp11 / tmp12
tmp14 = 1e-05
tmp15 = tmp13 + tmp14
tmp16 = libdevice.rsqrt(tmp15)
tmp17 = tmp0 * tmp16
tmp19 = tmp17 * tmp18
tmp21 = tmp19 + tmp20
tl.store(out_ptr0 + (x2), tmp21, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, ), (1, ))
assert_size_stride(primals_3, (4, ), (1, ))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [u, sub], Original ATen: [aten.mean, aten.sub]
stream0 = get_raw_stream(0)
triton_poi_fused_mean_sub_0.run(primals_1, buf0, 256, grid=grid(256), stream=stream0)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
# Topologically Sorted Source Nodes: [square, s, add, rsqrt, x, mul_1, x_1], Original ATen: [aten.pow, aten.mean, aten.add, aten.rsqrt, aten.mul]
triton_poi_fused_add_mean_mul_pow_rsqrt_1.run(buf0, primals_2, primals_3, buf1, 256, grid=grid(256), stream=stream0)
del buf0
del primals_2
del primals_3
return (buf1, primals_1, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)
fn = lambda: call([primals_1, primals_2, primals_3])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
import torch.nn as nn
class Norm(nn.Module):
def __init__(self, n_state, axis=-1, epsilon=1e-05):
super().__init__()
self.n_state = n_state
self.g = nn.Parameter(torch.ones([self.n_state]))
self.b = nn.Parameter(torch.zeros([self.n_state]))
self.axis = axis
self.epsilon = epsilon
def forward(self, x):
u = torch.mean(x, dim=self.axis, keepdim=True)
s = torch.mean(torch.square(x - u), dim=self.axis, keepdim=True)
x = (x - u) * torch.rsqrt(s + self.epsilon)
x = x * self.g + self.b
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'n_state': 4}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = 4.0
tmp9 = tmp7 / tmp8
tmp10 = tmp0 - tmp9
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_mean_mul_pow_rsqrt_1(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp20 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp1 * tmp1
tmp4 = tmp3 * tmp3
tmp5 = tmp2 + tmp4
tmp7 = tmp6 * tmp6
tmp8 = tmp5 + tmp7
tmp10 = tmp9 * tmp9
tmp11 = tmp8 + tmp10
tmp12 = 4.0
tmp13 = tmp11 / tmp12
tmp14 = 1e-05
tmp15 = tmp13 + tmp14
tmp16 = libdevice.rsqrt(tmp15)
tmp17 = tmp0 * tmp16
tmp19 = tmp17 * tmp18
tmp21 = tmp19 + tmp20
tl.store(out_ptr0 + x2, tmp21, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mean_sub_0[grid(256)](primals_1, buf0, 256, XBLOCK
=128, num_warps=4, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_mean_mul_pow_rsqrt_1[grid(256)](buf0,
primals_2, primals_3, buf1, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del buf0
del primals_2
del primals_3
return buf1, primals_1
class NormNew(nn.Module):
def __init__(self, n_state, axis=-1, epsilon=1e-05):
super().__init__()
self.n_state = n_state
self.g = nn.Parameter(torch.ones([self.n_state]))
self.b = nn.Parameter(torch.zeros([self.n_state]))
self.axis = axis
self.epsilon = epsilon
def forward(self, input_0):
primals_2 = self.g
primals_3 = self.b
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
| Aalanli/MusicGeneration | Norm | false | 5 | [
"MIT"
] | 0 | 7d268322d692013d8ac6e70be31741cea519fa28 | https://github.com/Aalanli/MusicGeneration/tree/7d268322d692013d8ac6e70be31741cea519fa28 | [
{
"content": "entry_point: Norm\n\n# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\n\n\n# kernel path: runs/run_shard_6/inductor_cache/dg/cdgw6x7nju4bzp2wyuwgeanbco7zcjis6yiusovvnpz6zw3yjd3l.py\n# Topologically Sorted Source Nodes: [u, sub], Original ATen: [aten.mean, aten.sub]\n# Source node to ATen node mapping:\n# sub => sub\n# u => mean\n# Graph fragment:\n# %mean : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%primals_1, [-1], True), kwargs = {})\n# %sub : [num_users=2] = call_function[target=torch.ops.aten.sub.Tensor](args = (%primals_1, %mean), kwargs = {})\ntriton_poi_fused_mean_sub_0 = async_compile.triton('triton_poi_fused_mean_sub_0', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[256], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mean_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = (xindex // 4)\n tmp0 = tl.load(in_ptr0 + (x2), xmask)\n tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')\n tmp2 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp4 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp3 = tmp1 + tmp2\n tmp5 = tmp3 + tmp4\n tmp7 = tmp5 + tmp6\n tmp8 = 4.0\n tmp9 = tmp7 / tmp8\n tmp10 = tmp0 - tmp9\n tl.store(out_ptr0 + (x2), tmp10, xmask)\n''', device_str='cuda')\n\n\n# kernel path: runs/run_shard_6/inductor_cache/v3/cv3kidzozprmsfiixppams3wqi7ij566hegofbo3ur5gfdrhxiik.py\n# Topologically Sorted Source Nodes: [square, s, add, rsqrt, x, mul_1, x_1], Original ATen: [aten.pow, aten.mean, aten.add, aten.rsqrt, aten.mul]\n# Source node to ATen node mapping:\n# add => add\n# mul_1 => mul_1\n# rsqrt => rsqrt\n# s => mean_1\n# square => pow_1\n# x => mul\n# x_1 => add_1\n# Graph fragment:\n# %pow_1 : [num_users=1] = call_function[target=torch.ops.aten.pow.Tensor_Scalar](args = (%sub, 2), kwargs = {})\n# %mean_1 : [num_users=1] = call_function[target=torch.ops.aten.mean.dim](args = (%pow_1, [-1], True), kwargs = {})\n# %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mean_1, 1e-05), kwargs = {})\n# %rsqrt : [num_users=1] = call_function[target=torch.ops.aten.rsqrt.default](args = (%add,), kwargs = {})\n# %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%sub, %rsqrt), kwargs = {})\n# %mul_1 : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul, %primals_2), kwargs = {})\n# %add_1 : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul_1, %primals_3), kwargs = {})\ntriton_poi_fused_add_mean_mul_pow_rsqrt_1 = async_compile.triton('triton_poi_fused_add_mean_mul_pow_rsqrt_1', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[256], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mean_mul_pow_rsqrt_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 7, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_add_mean_mul_pow_rsqrt_1(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = (xindex // 4)\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + (x2), xmask)\n tmp1 = tl.load(in_ptr0 + (4*x1), xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr0 + (1 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (2 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp9 = tl.load(in_ptr0 + (3 + (4*x1)), xmask, eviction_policy='evict_last')\n tmp18 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')\n tmp20 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = 4.0\n tmp13 = tmp11 / tmp12\n tmp14 = 1e-05\n tmp15 = tmp13 + tmp14\n tmp16 = libdevice.rsqrt(tmp15)\n tmp17 = tmp0 * tmp16\n tmp19 = tmp17 * tmp18\n tmp21 = tmp19 + tmp20\n tl.store(out_ptr0 + (x2), tmp21, xmask)\n''', device_str='cuda')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4, ), (1, ))\n assert_size_stride(primals_3, (4, ), (1, ))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n # Topologically Sorted Source Nodes: [u, sub], Original ATen: [aten.mean, aten.sub]\n stream0 = get_raw_stream(0)\n triton_poi_fused_mean_sub_0.run(primals_1, buf0, 256, grid=grid(256), stream=stream0)\n buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n # Topologically Sorted Source Nodes: [square, s, add, rsqrt, x, mul_1, x_1], Original ATen: [aten.pow, aten.mean, aten.add, aten.rsqrt, aten.mul]\n triton_poi_fused_add_mean_mul_pow_rsqrt_1.run(buf0, primals_2, primals_3, buf1, 256, grid=grid(256), stream=stream0)\n del buf0\n del primals_2\n del primals_3\n return (buf1, primals_1, )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n primals_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)\n primals_2 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n primals_3 = rand_strided((4, ), (1, ), device='cuda:0', dtype=torch.float32)\n fn = lambda: call([primals_1, primals_2, primals_3])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == \"__main__\":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main('None', benchmark_compiled_module)\n",
"role": "system",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr\n ):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp3 = tmp1 + tmp2\n tmp5 = tmp3 + tmp4\n tmp7 = tmp5 + tmp6\n tmp8 = 4.0\n tmp9 = tmp7 / tmp8\n tmp10 = tmp0 - tmp9\n tl.store(out_ptr0 + x2, tmp10, xmask)\n\n\n@triton.jit\ndef triton_poi_fused_add_mean_mul_pow_rsqrt_1(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = xindex // 4\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')\n tmp20 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = 4.0\n tmp13 = tmp11 / tmp12\n tmp14 = 1e-05\n tmp15 = tmp13 + tmp14\n tmp16 = libdevice.rsqrt(tmp15)\n tmp17 = tmp0 * tmp16\n tmp19 = tmp17 * tmp18\n tmp21 = tmp19 + tmp20\n tl.store(out_ptr0 + x2, tmp21, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_mean_sub_0[grid(256)](primals_1, buf0, 256, XBLOCK\n =128, num_warps=4, num_stages=1)\n buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n triton_poi_fused_add_mean_mul_pow_rsqrt_1[grid(256)](buf0,\n primals_2, primals_3, buf1, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del buf0\n del primals_2\n del primals_3\n return buf1, primals_1\n\n\nclass NormNew(nn.Module):\n\n def __init__(self, n_state, axis=-1, epsilon=1e-05):\n super().__init__()\n self.n_state = n_state\n self.g = nn.Parameter(torch.ones([self.n_state]))\n self.b = nn.Parameter(torch.zeros([self.n_state]))\n self.axis = axis\n self.epsilon = epsilon\n\n def forward(self, input_0):\n primals_2 = self.g\n primals_3 = self.b\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "user",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr\n ):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp3 = tmp1 + tmp2\n tmp5 = tmp3 + tmp4\n tmp7 = tmp5 + tmp6\n tmp8 = 4.0\n tmp9 = tmp7 / tmp8\n tmp10 = tmp0 - tmp9\n tl.store(out_ptr0 + x2, tmp10, xmask)\n\n\n@triton.jit\ndef triton_poi_fused_add_mean_mul_pow_rsqrt_1(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = xindex // 4\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')\n tmp20 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = 4.0\n tmp13 = tmp11 / tmp12\n tmp14 = 1e-05\n tmp15 = tmp13 + tmp14\n tmp16 = libdevice.rsqrt(tmp15)\n tmp17 = tmp0 * tmp16\n tmp19 = tmp17 * tmp18\n tmp21 = tmp19 + tmp20\n tl.store(out_ptr0 + x2, tmp21, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_mean_sub_0[grid(256)](primals_1, buf0, 256, XBLOCK\n =128, num_warps=4, num_stages=1)\n buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n triton_poi_fused_add_mean_mul_pow_rsqrt_1[grid(256)](buf0,\n primals_2, primals_3, buf1, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del buf0\n del primals_2\n del primals_3\n return buf1, primals_1\n\n\nclass NormNew(nn.Module):\n\n def __init__(self, n_state, axis=-1, epsilon=1e-05):\n super().__init__()\n self.n_state = n_state\n self.g = nn.Parameter(torch.ones([self.n_state]))\n self.b = nn.Parameter(torch.zeros([self.n_state]))\n self.axis = axis\n self.epsilon = epsilon\n\n def forward(self, input_0):\n primals_2 = self.g\n primals_3 = self.b\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n",
"role": "assistant",
"thinking": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._inductor.runtime.triton_helpers import libdevice\nimport torch.nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr\n ):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = xindex // 4\n tmp0 = tl.load(in_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp3 = tmp1 + tmp2\n tmp5 = tmp3 + tmp4\n tmp7 = tmp5 + tmp6\n tmp8 = 4.0\n tmp9 = tmp7 / tmp8\n tmp10 = tmp0 - tmp9\n tl.store(out_ptr0 + x2, tmp10, xmask)\n\n\n@triton.jit\ndef triton_poi_fused_add_mean_mul_pow_rsqrt_1(in_ptr0, in_ptr1, in_ptr2,\n out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 256\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x2 = xindex\n x1 = xindex // 4\n x0 = xindex % 4\n tmp0 = tl.load(in_ptr0 + x2, xmask)\n tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')\n tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')\n tmp18 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')\n tmp20 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')\n tmp2 = tmp1 * tmp1\n tmp4 = tmp3 * tmp3\n tmp5 = tmp2 + tmp4\n tmp7 = tmp6 * tmp6\n tmp8 = tmp5 + tmp7\n tmp10 = tmp9 * tmp9\n tmp11 = tmp8 + tmp10\n tmp12 = 4.0\n tmp13 = tmp11 / tmp12\n tmp14 = 1e-05\n tmp15 = tmp13 + tmp14\n tmp16 = libdevice.rsqrt(tmp15)\n tmp17 = tmp0 * tmp16\n tmp19 = tmp17 * tmp18\n tmp21 = tmp19 + tmp20\n tl.store(out_ptr0 + x2, tmp21, xmask)\n\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))\n assert_size_stride(primals_2, (4,), (1,))\n assert_size_stride(primals_3, (4,), (1,))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n get_raw_stream(0)\n triton_poi_fused_mean_sub_0[grid(256)](primals_1, buf0, 256, XBLOCK\n =128, num_warps=4, num_stages=1)\n buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)\n triton_poi_fused_add_mean_mul_pow_rsqrt_1[grid(256)](buf0,\n primals_2, primals_3, buf1, 256, XBLOCK=256, num_warps=4,\n num_stages=1)\n del buf0\n del primals_2\n del primals_3\n return buf1, primals_1\n\n\nclass NormNew(nn.Module):\n\n def __init__(self, n_state, axis=-1, epsilon=1e-05):\n super().__init__()\n self.n_state = n_state\n self.g = nn.Parameter(torch.ones([self.n_state]))\n self.b = nn.Parameter(torch.zeros([self.n_state]))\n self.axis = axis\n self.epsilon = epsilon\n\n def forward(self, input_0):\n primals_2 = self.g\n primals_3 = self.b\n primals_1 = input_0\n output = call([primals_1, primals_2, primals_3])\n return output[0]\n"
}
] |
BehlerAngular | # AOT ID: ['0_inference']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
# kernel path: runs/run_shard_6/inductor_cache/u6/cu6io2odujubs4jc63atgai7gg2fi2mzfw74zg77lxaq5wh3dsrx.py
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
# Source node to ATen node mapping:
# cat => cat
# Graph fragment:
# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%mul, %mul_1], -1), kwargs = {})
triton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor
from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties
@triton_heuristics.pointwise(
size_hints=[512],
filename=__file__,
triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 2
x1 = (xindex // 2)
x2 = xindex
tmp0 = x0
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x1), tmp4 & xmask, eviction_policy='evict_last', other=0.0)
tmp6 = 1.0
tmp7 = tmp6 - tmp5
tmp8 = tmp7 * tmp6
tmp9 = tl.full(tmp8.shape, 0.0, tmp8.dtype)
tmp10 = tl.where(tmp4, tmp8, tmp9)
tmp11 = tmp0 >= tmp3
tmp12 = tl.full([1], 2, tl.int64)
tmp13 = tmp0 < tmp12
tmp14 = tl.load(in_ptr0 + (x1), tmp11 & xmask, eviction_policy='evict_last', other=0.0)
tmp15 = tmp14 + tmp6
tmp16 = tmp15 * tmp6
tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)
tmp18 = tl.where(tmp11, tmp16, tmp17)
tmp19 = tl.where(tmp4, tmp10, tmp18)
tl.store(out_ptr0 + (x2), tmp19, xmask)
''', device_str='cuda')
async_compile.wait(globals())
del async_compile
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4, 2), (128, 32, 8, 2, 1), torch.float32)
# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]
stream0 = get_raw_stream(0)
triton_poi_fused_cat_0.run(arg0_1, buf0, 512, grid=grid(512), stream=stream0)
del arg0_1
return (buf0, )
def benchmark_compiled_module(times=10, repeat=10):
from torch._dynamo.testing import rand_strided
from torch._inductor.utils import print_performance
arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)
fn = lambda: call([arg0_1])
return print_performance(fn, times=times, repeat=repeat)
if __name__ == "__main__":
from torch._inductor.wrapper_benchmark import compiled_module_main
compiled_module_main('None', benchmark_compiled_module)
| import torch
from torch import nn as nn
class BehlerAngular(nn.Module):
"""
Compute Behler type angular contribution of the angle spanned by three atoms:
:math:`2^{(1-\\zeta)} (1 + \\lambda \\cos( {\\theta}_{ijk} ) )^\\zeta`
Sets of zetas with lambdas of -1 and +1 are generated automatically.
Args:
zetas (set of int): Set of exponents used to compute angular Behler term (default={1})
"""
def __init__(self, zetas={1}):
super(BehlerAngular, self).__init__()
self.zetas = zetas
def forward(self, cos_theta):
"""
Args:
cos_theta (torch.Tensor): Cosines between all pairs of neighbors of the central atom.
Returns:
torch.Tensor: Tensor containing values of the angular filters.
"""
angular_pos = [(2 ** (1 - zeta) * ((1.0 - cos_theta) ** zeta).
unsqueeze(-1)) for zeta in self.zetas]
angular_neg = [(2 ** (1 - zeta) * ((1.0 + cos_theta) ** zeta).
unsqueeze(-1)) for zeta in self.zetas]
angular_all = angular_pos + angular_neg
return torch.cat(angular_all, -1)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
| import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch import nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 2
x1 = xindex // 2
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + x1, tmp4 & xmask, eviction_policy='evict_last',
other=0.0)
tmp6 = 1.0
tmp7 = tmp6 - tmp5
tmp8 = tmp7 * tmp6
tmp9 = tl.full(tmp8.shape, 0.0, tmp8.dtype)
tmp10 = tl.where(tmp4, tmp8, tmp9)
tmp11 = tmp0 >= tmp3
tl.full([1], 2, tl.int64)
tmp14 = tl.load(in_ptr0 + x1, tmp11 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp15 = tmp14 + tmp6
tmp16 = tmp15 * tmp6
tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)
tmp18 = tl.where(tmp11, tmp16, tmp17)
tmp19 = tl.where(tmp4, tmp10, tmp18)
tl.store(out_ptr0 + x2, tmp19, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4, 2), (128, 32, 8, 2, 1),
torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(512)](arg0_1, buf0, 512, XBLOCK=128,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class BehlerAngularNew(nn.Module):
"""
Compute Behler type angular contribution of the angle spanned by three atoms:
:math:`2^{(1-\\zeta)} (1 + \\lambda \\cos( {\\theta}_{ijk} ) )^\\zeta`
Sets of zetas with lambdas of -1 and +1 are generated automatically.
Args:
zetas (set of int): Set of exponents used to compute angular Behler term (default={1})
"""
def __init__(self, zetas={1}):
super(BehlerAngularNew, self).__init__()
self.zetas = zetas
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
| AlexanderDKazakov/schnetpack | BehlerAngular | false | 6 | [
"MIT"
] | 0 | 97b82469d977981b500e439a6c93696d8dac8a3f | https://github.com/AlexanderDKazakov/schnetpack/tree/97b82469d977981b500e439a6c93696d8dac8a3f | [
{
"content": "entry_point: BehlerAngular\n\n# AOT ID: ['0_inference']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\n\n\n# kernel path: runs/run_shard_6/inductor_cache/u6/cu6io2odujubs4jc63atgai7gg2fi2mzfw74zg77lxaq5wh3dsrx.py\n# Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]\n# Source node to ATen node mapping:\n# cat => cat\n# Graph fragment:\n# %cat : [num_users=1] = call_function[target=torch.ops.aten.cat.default](args = ([%mul, %mul_1], -1), kwargs = {})\ntriton_poi_fused_cat_0 = async_compile.triton('triton_poi_fused_cat_0', '''\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties\n\n@triton_heuristics.pointwise(\n size_hints=[512], \n filename=__file__,\n triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=80, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, multi_processor_count=108), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},\n inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_cat_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 0, 'backend_hash': 'A9C866B4A14FD3277824029365D703C2427B2E685E54EC9B3EF4ADC8D1EEAC1D', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 512\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 2\n x1 = (xindex // 2)\n x2 = xindex\n tmp0 = x0\n tmp1 = tl.full([1], 0, tl.int64)\n tmp2 = tmp0 >= tmp1\n tmp3 = tl.full([1], 1, tl.int64)\n tmp4 = tmp0 < tmp3\n tmp5 = tl.load(in_ptr0 + (x1), tmp4 & xmask, eviction_policy='evict_last', other=0.0)\n tmp6 = 1.0\n tmp7 = tmp6 - tmp5\n tmp8 = tmp7 * tmp6\n tmp9 = tl.full(tmp8.shape, 0.0, tmp8.dtype)\n tmp10 = tl.where(tmp4, tmp8, tmp9)\n tmp11 = tmp0 >= tmp3\n tmp12 = tl.full([1], 2, tl.int64)\n tmp13 = tmp0 < tmp12\n tmp14 = tl.load(in_ptr0 + (x1), tmp11 & xmask, eviction_policy='evict_last', other=0.0)\n tmp15 = tmp14 + tmp6\n tmp16 = tmp15 * tmp6\n tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)\n tmp18 = tl.where(tmp11, tmp16, tmp17)\n tmp19 = tl.where(tmp4, tmp10, tmp18)\n tl.store(out_ptr0 + (x2), tmp19, xmask)\n''', device_str='cuda')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4, 2), (128, 32, 8, 2, 1), torch.float32)\n # Topologically Sorted Source Nodes: [cat], Original ATen: [aten.cat]\n stream0 = get_raw_stream(0)\n triton_poi_fused_cat_0.run(arg0_1, buf0, 512, grid=grid(512), stream=stream0)\n del arg0_1\n return (buf0, )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n arg0_1 = rand_strided((4, 4, 4, 4), (64, 16, 4, 1), device='cuda:0', dtype=torch.float32)\n fn = lambda: call([arg0_1])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == \"__main__\":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main('None', benchmark_compiled_module)\n",
"role": "system",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch import nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 512\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 2\n x1 = xindex // 2\n x2 = xindex\n tmp0 = x0\n tl.full([1], 0, tl.int64)\n tmp3 = tl.full([1], 1, tl.int64)\n tmp4 = tmp0 < tmp3\n tmp5 = tl.load(in_ptr0 + x1, tmp4 & xmask, eviction_policy='evict_last',\n other=0.0)\n tmp6 = 1.0\n tmp7 = tmp6 - tmp5\n tmp8 = tmp7 * tmp6\n tmp9 = tl.full(tmp8.shape, 0.0, tmp8.dtype)\n tmp10 = tl.where(tmp4, tmp8, tmp9)\n tmp11 = tmp0 >= tmp3\n tl.full([1], 2, tl.int64)\n tmp14 = tl.load(in_ptr0 + x1, tmp11 & xmask, eviction_policy=\n 'evict_last', other=0.0)\n tmp15 = tmp14 + tmp6\n tmp16 = tmp15 * tmp6\n tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)\n tmp18 = tl.where(tmp11, tmp16, tmp17)\n tmp19 = tl.where(tmp4, tmp10, tmp18)\n tl.store(out_ptr0 + x2, tmp19, xmask)\n\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4, 2), (128, 32, 8, 2, 1),\n torch.float32)\n get_raw_stream(0)\n triton_poi_fused_cat_0[grid(512)](arg0_1, buf0, 512, XBLOCK=128,\n num_warps=4, num_stages=1)\n del arg0_1\n return buf0,\n\n\nclass BehlerAngularNew(nn.Module):\n \"\"\"\n Compute Behler type angular contribution of the angle spanned by three atoms:\n\n :math:`2^{(1-\\\\zeta)} (1 + \\\\lambda \\\\cos( {\\\\theta}_{ijk} ) )^\\\\zeta`\n\n Sets of zetas with lambdas of -1 and +1 are generated automatically.\n\n Args:\n zetas (set of int): Set of exponents used to compute angular Behler term (default={1})\n\n \"\"\"\n\n def __init__(self, zetas={1}):\n super(BehlerAngularNew, self).__init__()\n self.zetas = zetas\n\n def forward(self, input_0):\n arg0_1 = input_0\n output = call([arg0_1])\n return output[0]\n",
"role": "user",
"thinking": null
},
{
"content": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch import nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 512\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 2\n x1 = xindex // 2\n x2 = xindex\n tmp0 = x0\n tl.full([1], 0, tl.int64)\n tmp3 = tl.full([1], 1, tl.int64)\n tmp4 = tmp0 < tmp3\n tmp5 = tl.load(in_ptr0 + x1, tmp4 & xmask, eviction_policy='evict_last',\n other=0.0)\n tmp6 = 1.0\n tmp7 = tmp6 - tmp5\n tmp8 = tmp7 * tmp6\n tmp9 = tl.full(tmp8.shape, 0.0, tmp8.dtype)\n tmp10 = tl.where(tmp4, tmp8, tmp9)\n tmp11 = tmp0 >= tmp3\n tl.full([1], 2, tl.int64)\n tmp14 = tl.load(in_ptr0 + x1, tmp11 & xmask, eviction_policy=\n 'evict_last', other=0.0)\n tmp15 = tmp14 + tmp6\n tmp16 = tmp15 * tmp6\n tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)\n tmp18 = tl.where(tmp11, tmp16, tmp17)\n tmp19 = tl.where(tmp4, tmp10, tmp18)\n tl.store(out_ptr0 + x2, tmp19, xmask)\n\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4, 2), (128, 32, 8, 2, 1),\n torch.float32)\n get_raw_stream(0)\n triton_poi_fused_cat_0[grid(512)](arg0_1, buf0, 512, XBLOCK=128,\n num_warps=4, num_stages=1)\n del arg0_1\n return buf0,\n\n\nclass BehlerAngularNew(nn.Module):\n \"\"\"\n Compute Behler type angular contribution of the angle spanned by three atoms:\n\n :math:`2^{(1-\\\\zeta)} (1 + \\\\lambda \\\\cos( {\\\\theta}_{ijk} ) )^\\\\zeta`\n\n Sets of zetas with lambdas of -1 and +1 are generated automatically.\n\n Args:\n zetas (set of int): Set of exponents used to compute angular Behler term (default={1})\n\n \"\"\"\n\n def __init__(self, zetas={1}):\n super(BehlerAngularNew, self).__init__()\n self.zetas = zetas\n\n def forward(self, input_0):\n arg0_1 = input_0\n output = call([arg0_1])\n return output[0]\n",
"role": "assistant",
"thinking": "import torch\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import grid\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch import nn as nn\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\n\n\n@triton.jit\ndef triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):\n xnumel = 512\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex % 2\n x1 = xindex // 2\n x2 = xindex\n tmp0 = x0\n tl.full([1], 0, tl.int64)\n tmp3 = tl.full([1], 1, tl.int64)\n tmp4 = tmp0 < tmp3\n tmp5 = tl.load(in_ptr0 + x1, tmp4 & xmask, eviction_policy='evict_last',\n other=0.0)\n tmp6 = 1.0\n tmp7 = tmp6 - tmp5\n tmp8 = tmp7 * tmp6\n tmp9 = tl.full(tmp8.shape, 0.0, tmp8.dtype)\n tmp10 = tl.where(tmp4, tmp8, tmp9)\n tmp11 = tmp0 >= tmp3\n tl.full([1], 2, tl.int64)\n tmp14 = tl.load(in_ptr0 + x1, tmp11 & xmask, eviction_policy=\n 'evict_last', other=0.0)\n tmp15 = tmp14 + tmp6\n tmp16 = tmp15 * tmp6\n tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)\n tmp18 = tl.where(tmp11, tmp16, tmp17)\n tmp19 = tl.where(tmp4, tmp10, tmp18)\n tl.store(out_ptr0 + x2, tmp19, xmask)\n\n\ndef call(args):\n arg0_1, = args\n args.clear()\n assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n buf0 = empty_strided_cuda((4, 4, 4, 4, 2), (128, 32, 8, 2, 1),\n torch.float32)\n get_raw_stream(0)\n triton_poi_fused_cat_0[grid(512)](arg0_1, buf0, 512, XBLOCK=128,\n num_warps=4, num_stages=1)\n del arg0_1\n return buf0,\n\n\nclass BehlerAngularNew(nn.Module):\n \"\"\"\n Compute Behler type angular contribution of the angle spanned by three atoms:\n\n :math:`2^{(1-\\\\zeta)} (1 + \\\\lambda \\\\cos( {\\\\theta}_{ijk} ) )^\\\\zeta`\n\n Sets of zetas with lambdas of -1 and +1 are generated automatically.\n\n Args:\n zetas (set of int): Set of exponents used to compute angular Behler term (default={1})\n\n \"\"\"\n\n def __init__(self, zetas={1}):\n super(BehlerAngularNew, self).__init__()\n self.zetas = zetas\n\n def forward(self, input_0):\n arg0_1 = input_0\n output = call([arg0_1])\n return output[0]\n"
}
] |
BottleneckBlock | "# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimp(...TRUNCATED) | "import math\nimport torch\nimport torch.nn as nn\nimport torch.utils.data\nimport torch.nn.parallel(...TRUNCATED) | "import torch\nfrom torch._inductor.select_algorithm import extern_kernels\nimport triton\nimport tr(...TRUNCATED) | Aamer98/FeatureNorm | BottleneckBlock | false | 7 | [
"MIT"
] | 0 | fbf3d3b4cef81b3351347d272eb51b6cdd9f0cc5 | https://github.com/Aamer98/FeatureNorm/tree/fbf3d3b4cef81b3351347d272eb51b6cdd9f0cc5 | "[\n {\n \"content\": \"entry_point: BottleneckBlock\\n\\n# AOT ID: ['0_forward']\\nfrom ctypes (...TRUNCATED) |
Mlp | "# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimp(...TRUNCATED) | "import torch\nimport torch.nn as nn\n\n\nclass Conv1d(nn.Module):\n\n def __init__(self, nf, nx,(...TRUNCATED) | "import torch\nfrom torch._inductor.select_algorithm import extern_kernels\nimport triton\nimport tr(...TRUNCATED) | Aalanli/MusicGeneration | Mlp | false | 9 | [
"MIT"
] | 0 | 7d268322d692013d8ac6e70be31741cea519fa28 | https://github.com/Aalanli/MusicGeneration/tree/7d268322d692013d8ac6e70be31741cea519fa28 | "[\n {\n \"content\": \"entry_point: Mlp\\n\\n# AOT ID: ['0_forward']\\nfrom ctypes import c_voi(...TRUNCATED) |
GCN | "# AOT ID: ['0_forward']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimp(...TRUNCATED) | "from torch.nn import Module\nimport math\nimport torch\nimport torch.nn.functional as F\nimport tor(...TRUNCATED) | "import torch\nfrom torch._inductor.select_algorithm import extern_kernels\nimport triton\nimport tr(...TRUNCATED) | AlexMinhao/NAS_GNN | GCN | false | 10 | [
"Apache-2.0"
] | 0 | 89183988a96e1d6baed910ab3843c13282f8b077 | https://github.com/AlexMinhao/NAS_GNN/tree/89183988a96e1d6baed910ab3843c13282f8b077 | "[\n {\n \"content\": \"entry_point: GCN\\n\\n# AOT ID: ['0_forward']\\nfrom ctypes import c_voi(...TRUNCATED) |
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 8
Size of downloaded dataset files:
199 MB
Size of the auto-converted Parquet files:
199 MB
Number of rows:
18,162