from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for quantized unary operators (point-wise and reduction)."""
# Configs for pointwise and reduction unary ops
qunary_ops_configs_short = op_bench.config_list(
attr_names=['M', 'N'],
attrs=[
[512, 512],
],
cross_product_configs={
'dtype': [torch.quint8],
},
tags=['short']
)
qunary_ops_configs_long = op_bench.cross_product_configs(
M=[256, 1024],
N=[256, 1024],
dtype=[torch.quint8, torch.qint8, torch.qint32],
tags=['long']
)
class QUnaryOpBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, dtype, op_func):
f_input = torch.rand(M, N)
import operator_benchmark as op_bench
import torch
'''Microbenchmarks for the quantized interpolate op.
Note: We are not benchmarking `upsample` as it is being depricated, and calls
the `interpolate` anyway.
'''
qinterpolate_long_configs = op_bench.config_list(
attr_names=['M', 'N', 'K'],
attrs=[
[512, 512, 512],
],
cross_product_configs={
'dtype': [torch.quint8, torch.qint8, torch.qint32],
'mode': ['nearest', 'bilinear'],
'scale': [0.5, 1.0, 2.0],
'contig': [True], # TODO: Add `False` after #29435
},
tags=['long']
)
qinterpolate_short_configs = op_bench.config_list(
attr_names=['M', 'N', 'K', 'dtype', 'mode', 'scale', 'contig'],
attrs=[
[32, 32, 32, torch.quint8, 'nearest', 0.5, True], # Downsample
[32, 32, 32, torch.quint8, 'bilinear', 0.5, True], # Downsample
[32, 32, 32, torch.quint8, 'nearest', 2.0, True], # Upsample
[32, 32, 32, torch.quint8, 'bilinear', 2.0, True], # Upsample
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for Chunk operator"""
# Configs for PT Chunk operator
chunk_short_configs = op_bench.config_list(
attr_names=["M", "N", "chunks"],
attrs=[
[256, 512, 2],
[512, 512, 2],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"],
)
chunks_long_configs = op_bench.cross_product_configs(M=[128, 1024],
N=[128, 1024],
chunks=[2, 4],
device=['cpu', 'cuda'],
tags=['long'])
class ChunkBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, chunks, device):
import operator_benchmark as op_bench
import torch
import random
from typing import List
"""Microbenchmarks for Cat operator"""
# Configs for PT Cat operator
cat_configs_short = op_bench.config_list(
attr_names=['sizes', 'N', 'dim'],
attrs=[
[(1, 1, 1), 2, 0], # noqa
[(512, 512, 2), 2, 1], # noqa
[(128, 1024, 2), 2, 1], # noqa
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=['short'],
)
cat_configs_long = op_bench.config_list(
attr_names=['sizes', 'N', 'dim'],
attrs=[
[(2**10, 2**10, 2), 2, 0], # noqa
[(2**10+1, 2**10-1, 2), 2, 1], # noqa
[(2**10, 2**10, 2), 2, 2], # noqa
[[ lambda: random.randint(2**6, 2**7), 2**7-17, 2**6+1], # noqa
5, 0],
import operator_benchmark as op_bench
import torch
import numpy
"""Microbenchmarks for index_select operator."""
# An example input from this configuration is M=4, N=4, dim=0.
index_select_configs_short = op_bench.config_list(
attr_names=["M", "N", "K", "dim"],
attrs=[
[8, 8, 1, 1],
[256, 512, 1, 1],
[512, 512, 1, 1],
[8, 8, 2, 1],
[256, 512, 2, 1],
[512, 512, 2, 1],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"]
)
index_select_configs_long = op_bench.cross_product_configs(
M=[128, 1024],
N=[128, 1024],
K=[1, 2],
dim=[1],
device=['cpu', 'cuda'],
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
import torch.nn as nn
"""Microbenchmarks for batchnorm operator."""
configs = op_bench.config_list(attrs=[
[1, 32, 10],
[4, 256, 100],
[16, 1024, 256],
],
attr_names=["N", "IN", "OUT"],
tags=["short"])
class LinearBenchmark(op_bench.TorchBenchmarkBase):
def init(self, N, IN, OUT):
self.input_one = torch.rand(N, IN)
self.linear = nn.Linear(IN, OUT)
self.set_module_name("linear")
def forward(self):
return self.linear(self.input_one)
op_bench.generate_pt_test(configs, LinearBenchmark)
import operator_benchmark as op_bench
import torch
import torch.nn.functional as F
"""Microbenchmarks for batchnorm operator."""
batchnorm_configs_short = op_bench.config_list(
attr_names=["M", "N", "K"],
attrs=[
[1, 256, 3136],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"]
)
batchnorm_configs_long = op_bench.cross_product_configs(
M=[1, 128],
N=[8192, 2048],
K=[1],
device=['cpu', 'cuda'],
tags=["long"]
)
class BatchNormBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, K, device):
self.input_one = torch.rand(M, N, K, device=device, requires_grad=self.auto_set())
from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for Split operator"""
# Configs for PT Split operator
split_configs_short = op_bench.config_list(
attr_names=["M", "N", "parts"],
attrs=[
[256, 512, 2],
[512, 512, 2],
],
cross_product_configs={
'device': ['cpu'],
},
tags=["short"],
)
split_configs_long = op_bench.cross_product_configs(
M=[128, 1024],
N=[128, 1024],
parts=[2, 4],
device=['cpu'],
tags=['long']
)
import torch.nn as nn
"""
Microbenchmarks for the softmax operators.
"""
# Configs for softmax ops
softmax_configs_short = op_bench.config_list(
attr_names=[
'N', 'C', 'H', 'W'
],
attrs=[
[4, 3, 256, 256],
[8, 3, 512, 512],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=['short']
)
softmax_configs_long = op_bench.cross_product_configs(
N=[8, 16],
C=[3, 64],
H=[64, 128],
W=[64, 128],
device=['cpu', 'cuda'],
tags=['long']
import torch
"""Microbenchmarks for add_ operator. Supports both Caffe2/PyTorch."""
# Configs for PT add operator
add_long_configs = op_bench.cross_product_configs(M=[8, 128],
N=[32, 64],
K=[256, 512],
device=['cpu', 'cuda'],
tags=["long"])
add_short_configs = op_bench.config_list(
attr_names=["M", "N", "K"],
attrs=[
[1, 1, 1],
[64, 64, 64],
[64, 64, 128],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"],
)
class AddBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, K, device):
self.input_one = torch.rand(M,
N,
K,
device=device,
requires_grad=self.auto_set())
self.input_two = torch.rand(M,
[3, 512, 512, torch.quint8, 'D'],
],
'tags': ['short'],
}
quantize_configs_long_dict = {
'C': [3, 5,
8], # this is reused for per-channel: avoid single channel test
'M': [256, 1024],
'N': [256, 1024],
'dtype': [torch.quint8, torch.qint8, torch.qint32],
'mode': ['D', 'Q'],
'tags': ['long'],
}
quantize_per_tensor_configs_short = op_bench.config_list(
**quantize_configs_short_dict)
quantize_per_tensor_configs_long = op_bench.cross_product_configs(
**quantize_configs_long_dict)
class QuantizePerTensorBenchmark(op_bench.TorchBenchmarkBase):
r"""Benchmarks both quantization and dequantization."""
def init(self, C, M, N, dtype, mode):
assert (mode in ('Q', 'D'))
self.input = torch.rand(C, M, N)
self.dtype = dtype
self.op = nnq.Quantize(scale=1.0, zero_point=0, dtype=dtype)
self.set_module_name('QuantizePerTensor')
if mode == 'D':
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for quantized batchnorm operator."""
batchnorm_configs_short = op_bench.config_list(attr_names=["M", "N", "K"],
attrs=[
[1, 256, 3136],
],
cross_product_configs={
'device': ['cpu'],
'dtype': (torch.qint8, ),
},
tags=["short"])
class QBatchNormBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, K, device, dtype):
self._init(M, N, K, device)
x_scale = 0.1
x_zero_point = 0
self.q_input_one = torch.quantize_per_tensor(self.input_one,
scale=x_scale,
zero_point=x_zero_point,
dtype=dtype)
self.mean = torch.rand(N)
self.var = torch.rand(N)
from caffe2.python import core
"""Microbenchmarks for element-wise ReplaceNaN operator."""
# Configs for C2 ReplaceNaN operator
replace_nan_long_configs = op_bench.cross_product_configs(
M=[32, 64, 128], N=range(32, 128, 32), dtype=["float", "double"], tags=["long"]
)
replace_nan_short_configs = op_bench.config_list(
attrs=[
[16, 16, "float"],
[16, 16, "double"],
[64, 64, "float"],
[64, 64, "double"],
],
attr_names=["M", "N", "dtype"],
tags=["short"],
)
class ReplaceNaNBenchmark(op_bench_c2.Caffe2BenchmarkBase):
def init(self, M, N, dtype):
self.input = self.tensor([M, N], dtype)
self.set_module_name("replace_nan")
def forward(self):
op = core.CreateOperator("ReplaceNaN", self.input, self.input, value=1.0)
return op
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for element-wise Add operator. Supports both Caffe2/PyTorch."""
# Configs for PT add operator
add_long_configs = op_bench.cross_product_configs(
M=[8, 64, 128],
N=range(2, 10, 3),
K=[2**x for x in range(0, 3)],
tags=["long"])
add_short_configs = op_bench.config_list(
attrs=[
[64, 64, 64],
[64, 64, 128],
],
attr_names=["M", "N", "K"],
tags=["short"],
)
class AddBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, K):
self.input_one = torch.rand(M, N, K)
self.input_two = torch.rand(M, N, K)
self.set_module_name("add")
def forward(self):
return torch.add(self.input_one, self.input_two)
import operator_benchmark as op_bench
import torch
from typing import List
"""Microbenchmarks for as_strided operator"""
# Configs for PT as_strided operator
as_strided_configs_short = op_bench.config_list(
attr_names=["M", "N", "size", "stride", "storage_offset"],
attrs=[
[8, 8, (2, 2), (1, 1), 0],
[256, 256, (32, 32), (1, 1), 0],
[512, 512, (64, 64), (2, 2), 1],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"],
)
as_strided_configs_long = op_bench.cross_product_configs(
M=[512],
N=[1024],
size=[(16, 16), (128, 128)],
stride=[(1, 1)],
storage_offset=[0, 1],
device=['cpu', 'cuda'],
tags=['long'])
class As_stridedBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, size, stride, storage_offset, device):
import torch.nn as nn
"""
Microbenchmarks for the hardsigmoid operator.
"""
# Configs for hardsigmoid ops
hardsigmoid_configs_short = op_bench.config_list(
attr_names=[
'N', 'C', 'H', 'W'
],
attrs=[
[1, 3, 256, 256],
[4, 3, 256, 256],
],
cross_product_configs={
'device': ['cpu'],
},
tags=['short']
)
hardsigmoid_configs_long = op_bench.cross_product_configs(
N=[8, 16],
C=[3],
H=[256, 512],
W=[256, 512],
device=['cpu'],
tags=['long']
LENGTH=range(1, 100),
M=[1],
N=[2],
MAX_LENGTH=range(1, 100),
device=['cpu', 'cuda'],
dtype=[torch.int32],
tags=["long"],
)
clip_ranges_short_configs = op_bench.config_list(
attrs=[
[6, 1, 2, 1, torch.int32],
[7, 1, 2, 2, torch.int32],
[8, 1, 2, 3, torch.int32],
[9, 1, 2, 4, torch.int32],
[10, 1, 2, 5, torch.int32],
],
attr_names=["LENGTH", "M", "N", "MAX_LENGTH", "dtype"],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"],
)
class ClipRangesBenchmark(op_bench.TorchBenchmarkBase):
def init(self, LENGTH, M, N, MAX_LENGTH, device, dtype):
self.inputs = {
"input": torch.rand(LENGTH, M, N, device=device).type(dtype),
"max_length": MAX_LENGTH
}
self.set_module_name("clip_ranges")
import operator_benchmark as op_bench
import torch
# Configs for pointwise and reduction unary ops
qmethods_configs_short = op_bench.config_list(
attr_names=['M', 'N'],
attrs=[
[32, 32],
],
cross_product_configs={
'dtype': [torch.quint8],
'contig': [False, True],
},
tags=['short']
)
qmethods_configs_long = op_bench.cross_product_configs(
M=[256, 1024],
N=[256, 1024],
dtype=[torch.qint8, torch.qint32],
contig=[False, True],
tags=['long']
)
class _QMethodBenchmarkBase(op_bench.TorchBenchmarkBase):
def init(self, M, N, dtype, contig):
f_input = torch.rand(M, N)
scale = 1.0
zero_point = 0
self.q_input = torch.quantize_per_tensor(f_input, scale=scale,
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
import torch.nn as nn
"""
Microbenchmarks for the softmax operators.
"""
# Configs for softmax ops
softmax_configs_short = op_bench.config_list(attrs=[
[1, 3, 32, 32],
[2, 3, 64, 64],
],
attr_names=['N', 'C', 'H', 'W'],
tags=['short'])
softmax_configs_long = op_bench.config_list(attrs=[
[8, 3, 128, 128],
[16, 512, 14, 14],
[16, 256, 28, 28],
],
attr_names=['N', 'C', 'H', 'W'],
tags=['long'])
softmax_ops_list = op_bench.op_list(
attr_names=['op_name', 'op_func'],
attrs=[
['Softmax', nn.Softmax],
import operator_benchmark as op_bench
import torch
import numpy
"""Microbenchmarks for gather operator."""
# An example input from this configuration is M=4, N=4, dim=0.
gather_configs_short = op_bench.config_list(attr_names=["M", "N", "dim"],
attrs=[
[256, 512, 0],
[512, 512, 1],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"])
gather_configs_long = op_bench.cross_product_configs(M=[128, 1024],
N=[128, 1024],
dim=[0, 1],
device=['cpu', 'cuda'],
tags=["long"])
class GatherBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, dim, device):
self.input_one = torch.rand(M, N, device=device)
self.dim = dim
min_val = M if dim == 0 else N
numpy.random.seed((1 << 32) - 1)
self.index = torch.tensor(numpy.random.randint(0, min_val, (M, N)),
device=device)
import operator_benchmark as op_bench
import torch
"""
Microbenchmarks for batch matrix mult with einsum and torch.bmm.
"""
batch_mm_configs_short = op_bench.config_list(
attr_names=["B", "M", "N", "K"],
attrs=[
[4, 5, 3, 2],
[32, 25, 20, 30],
[128, 100, 120, 110],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"],
)
batch_mm_configs_long = op_bench.config_list(
attr_names=["B", "M", "N", "K"],
attrs=[
[128, 256, 128, 256],
[512, 1024, 1024, 512],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["long"],
)
import torch
import operator_benchmark as op_bench
# 2D pooling will have input matrix of rank 3 or 4
qpool2d_long_configs = op_bench.config_list(
attrs=(
# C H W k s p
(1, 3, 3, (3, 3), (1, 1), (0, 0)), # dummy # noqa
(3, 64, 64, (3, 3), (2, 2), (1, 1)), # dummy # noqa
# VGG16 pools with original input shape: (-1, 3, 224, 224)
(64, 224, 224, (2, 2), (2, 2), (0, 0)), # MaxPool2d-4 # noqa
(256, 56, 56, (2, 2), (2, 2), (0, 0)), # MaxPool2d-16 # noqa
),
attr_names=(
'C',
'H',
'W', # Input layout
'k',
's',
'p'), # Pooling parameters
cross_product_configs={
'N': (1, 4),
'contig': (False, True),
'dtype': (torch.quint8, ),
},
tags=('long', ))
qpool2d_short_configs = op_bench.config_list(
attrs=((1, 3, 3, (3, 3), (1, 1), (0, 0)), ), # dummy # noqa
attr_names=(
'C',
from caffe2.python import core
"""Microbenchmarks for MatMul operator"""
# Configs for C2 Matmul operator
mm_long_configs = op_bench.cross_product_configs(M=[8, 64, 128],
N=range(2, 10, 3),
K=[2**x for x in range(0, 3)],
trans_a=[True, False],
trans_b=[True, False],
tags=["long"])
mm_short_configs = op_bench.config_list(
attrs=[
[128, 128, 128, False, True],
[1024, 1024, 256, True, False],
[8192, 8192, 1024, True, False],
],
attr_names=["M", "N", "K", "trans_a", "trans_b"],
tags=["short"],
)
class MatMulBenchmark(op_bench_c2.Caffe2BenchmarkBase):
def init(self, M, N, K, trans_a, trans_b):
self.input_one = self.tensor([N, M]) if trans_a else self.tensor(
[M, N])
self.input_two = self.tensor([K, N]) if trans_b else self.tensor(
[N, K])
self.args = {'trans_a': trans_a, 'trans_b': trans_b}
self.output = self.tensor([M, K])
self.set_module_name("matmul")
self.set_module_name("interpolate")
def forward(self, input_image, output_size, mode):
return torch.nn.functional.interpolate(input_image,
size=output_size,
mode=mode,
align_corners=False)
config_short = op_bench.config_list(
attr_names=["input_size", "output_size"],
attrs=[
[(1, 3, 60, 40), (24, 24)],
[(1, 3, 600, 400), (240, 240)],
[(1, 3, 320, 320), (256, 256)],
],
cross_product_configs={
'channels_last': [True, False],
},
tags=["short"],
)
config_long = op_bench.config_list(
attr_names=["input_size", "output_size"],
attrs=[
[(1, 3, 320, 320), (512, 512)],
[(1, 3, 500, 500), (256, 256)],
[(1, 3, 500, 500), (800, 800)],
[(2, 128, 64, 46), (128, 128)],
],
cross_product_configs={
from __future__ import absolute_import, division, print_function, unicode_literals
import operator_benchmark as op_bench
import torch
# Configs for pointwise unary ops
unary_ops_configs = op_bench.config_list(
attrs=[
[128, 128],
],
attr_names=["M", "N"],
tags=["short"]
)
unary_ops_list = op_bench.op_list(
attr_names=["op_name", "op_func"],
attrs=[
["abs", torch.abs],
["acos", torch.acos],
],
)
class UnaryOpBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, op_func):
self.input_one = torch.rand(M, N)
self.op_func = op_func
def forward(self):
return self.op_func(self.input_one)
import operator_benchmark as op_bench
import torch
from torch import nn
"""
Microbenchmarks for RNNs.
"""
qrnn_configs = op_bench.config_list(
attrs=[
[1, 3, 1],
[5, 7, 4],
],
# names: input_size, hidden_size, num_layers
attr_names=["I", "H", "NL"],
cross_product_configs={
"B": (True, ), # Bias always True for quantized
"D": (False, True), # Bidirectional
"dtype": (torch.qint8, ) # Only qint8 dtype works for now
},
tags=["short"])
class LSTMBenchmark(op_bench.TorchBenchmarkBase):
def init(self, I, H, NL, B, D, dtype):
sequence_len = 128
batch_size = 16
# The quantized.dynamic.LSTM has a bug. That's why we create a regular
# LSTM, and quantize it later. See issue #31192.
scale = 1.0 / 256
zero_point = 0
import torch.nn.quantized as nnq
"""
Microbenchmarks for Quantized Linear operators.
"""
# Configs for qlinear
qlinear_configs = op_bench.config_list(
attrs=[
[1024, 1024, 1024],
[64, 800, 320],
[64, 768, 512],
[16, 256, 512],
[128, 128, 128],
[256, 512, 256],
[6400, 15, 141],
[6400, 8, 141],
[16, 211, 2504],
[16, 369, 1434],
[1, 1024, 3496],
[16, 256, 512],
[1, 1600, 3456],
],
attr_names=["N", "OUT", "IN"], # M, N, K
tags=["short"],
)
class QLinearBenchmark(op_bench.TorchBenchmarkBase):
def init(self, N, IN, OUT):
scale = 1.0 / 255
zero_point = 0
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for MatMul operator"""
# Configs for PT Matmul operator
mm_short_configs = op_bench.config_list(
attr_names=["M", "N", "K", "trans_a", "trans_b"],
attrs=[
[1, 1, 1, True, False],
[128, 128, 128, True, False],
[256, 256, 256, False, True],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"],
)
mm_long_configs = op_bench.cross_product_configs(M=[32],
N=[512, 128],
K=[64],
trans_a=[False, True],
trans_b=[True, False],
device=['cpu', 'cuda'],
tags=["long"])
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import operator_benchmark as op_bench
import torch
import torch.nn as nn
"""Microbenchmarks for Linear operator."""
linear_configs_short = op_bench.config_list(attr_names=["N", "IN", "OUT"],
attrs=[
[4, 256, 128],
[16, 512, 256],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
},
tags=["short"])
linear_configs_long = op_bench.cross_product_configs(N=[32, 64],
IN=[128, 512],
OUT=[64, 128],
device=['cpu', 'cuda'],
tags=["long"])
class LinearBenchmark(op_bench.TorchBenchmarkBase):
def init(self, N, IN, OUT, device):
self.input_one = torch.rand(N, IN, device=device)
self.linear = nn.Linear(IN, OUT).to(device=device)
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for element-wise Add operator. Supports both Caffe2/PyTorch."""
add_short_configs = op_bench.config_list(
attr_names=['M', 'N', 'K'],
attrs=[
[8, 16, 32],
[16, 16, 64],
[64, 64, 128],
],
cross_product_configs={
'device': ['cpu', 'cuda'],
'dtype': [torch.float, torch.float64],
},
tags=['short'],
)
class AddBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, K, device, dtype):
self.input_one = torch.rand(M, N, K, device=device, dtype=dtype, requires_grad=True)
self.input_two = torch.rand(M, N, K, device=device, dtype=dtype)
self.set_module_name('add')
def forward(self):
return torch.add(self.input_one, self.input_two)
op_bench.generate_pt_test(add_short_configs, AddBenchmark)