1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
|
import operator_benchmark as op_bench
import torch
import torch.nn.functional as F
"""Microbenchmarks for batchnorm operator."""
# Benchmark cudnn if available
if torch.backends.cudnn.is_available:
def cudnn_benchmark_configs(configs):
result = []
for config in configs:
is_cuda = any("cuda" in attr.values() for attr in config)
if is_cuda:
result.append((*config, dict(cudnn=True)))
result.append((*config, dict(cudnn=False)))
return result
else:
def cudnn_benchmark_configs(configs):
return [(*config, dict(cudnn=False)) for config in configs]
batchnorm_configs_short = cudnn_benchmark_configs(
op_bench.config_list(
attr_names=["M", "N", "K"],
attrs=[
[1, 256, 3136],
],
cross_product_configs={
"device": ["cpu", "cuda"],
"training": [True, False],
},
tags=["short"],
)
)
batchnorm_configs_long = cudnn_benchmark_configs(
op_bench.cross_product_configs(
M=[2, 128],
N=[8192, 2048],
K=[1],
device=["cpu", "cuda"],
training=[True, False],
tags=["long"],
)
)
class BatchNormBenchmark(op_bench.TorchBenchmarkBase):
def init(self, M, N, K, device, training, cudnn):
self.inputs = {
"input_one": torch.rand(
M, N, K, device=device, requires_grad=self.auto_set()
),
"mean": torch.rand(N, device=device),
"var": torch.rand(N, device=device),
"weight": torch.rand(N, device=device),
"bias": torch.rand(N, device=device),
"training": training,
"cudnn": cudnn,
}
self.set_module_name("batchnorm")
def forward(self, input_one, mean, var, weight, bias, training, cudnn):
with torch.backends.cudnn.flags(enabled=cudnn):
return F.batch_norm(input_one, mean, var, weight, bias, training)
op_bench.generate_pt_test(
batchnorm_configs_short + batchnorm_configs_long, BatchNormBenchmark
)
op_bench.generate_pt_gradient_test(
batchnorm_configs_short + batchnorm_configs_long, BatchNormBenchmark
)
batchnorm1d_configs_short = cudnn_benchmark_configs(
op_bench.config_list(
attr_names=["N", "C"],
attrs=[
[3136, 256],
],
cross_product_configs={
"device": ["cpu", "cuda"],
"training": [True, False],
},
tags=["short"],
)
)
batchnorm1d_configs_long = cudnn_benchmark_configs(
op_bench.cross_product_configs(
N=[2, 128],
C=[8192, 2048],
device=["cpu", "cuda"],
training=[True, False],
tags=["long"],
)
)
class BatchNorm1dBenchmark(op_bench.TorchBenchmarkBase):
def init(self, N, C, device, training, cudnn):
self.inputs = {
"input_one": torch.rand(N, C, device=device, requires_grad=self.auto_set()),
"mean": torch.rand(C, device=device),
"var": torch.rand(C, device=device),
"weight": torch.rand(C, device=device),
"bias": torch.rand(C, device=device),
"training": training,
"cudnn": cudnn,
}
self.set_module_name("batchnorm")
def forward(self, input_one, mean, var, weight, bias, training, cudnn):
with torch.backends.cudnn.flags(enabled=cudnn):
return F.batch_norm(input_one, mean, var, weight, bias, training)
op_bench.generate_pt_test(
batchnorm1d_configs_short + batchnorm1d_configs_long, BatchNorm1dBenchmark
)
op_bench.generate_pt_gradient_test(
batchnorm1d_configs_short + batchnorm1d_configs_long, BatchNorm1dBenchmark
)
if __name__ == "__main__":
op_bench.benchmark_runner.main()
|
