• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1import operator_benchmark as op_bench
2
3import torch
4import torch.nn as nn
5
6
7"""
8Microbenchmarks for the softmax operators.
9"""
10
11
12# Configs for softmax ops
13softmax_configs_short = op_bench.config_list(
14    attr_names=["N", "C", "H", "W"],
15    attrs=[
16        [1, 3, 256, 256],
17        [4, 3, 256, 256],
18    ],
19    cross_product_configs={
20        "device": ["cpu", "cuda"],
21    },
22    tags=["short"],
23)
24
25
26softmax_configs_long = op_bench.cross_product_configs(
27    N=[8, 16],
28    C=[3],
29    H=[256, 512],
30    W=[256, 512],
31    device=["cpu", "cuda"],
32    tags=["long"],
33)
34
35
36softmax_ops_list = op_bench.op_list(
37    attr_names=["op_name", "op_func"],
38    attrs=[
39        ["Softmax", nn.Softmax],
40        ["Softmax2d", nn.Softmax2d],
41        ["LogSoftmax", nn.LogSoftmax],
42    ],
43)
44
45softmax_two_dims_ops_list = op_bench.op_list(
46    attr_names=["op_name", "op_func"],
47    attrs=[
48        ["Softmax", nn.Softmax],
49        ["LogSoftmax", nn.LogSoftmax],
50    ],
51)
52
53
54softmax_two_dims_configs = op_bench.config_list(
55    attr_names=["M", "N", "dim"],
56    attrs=[
57        [700, 23258, 0],
58        [700, 23258, 1],
59        [1024, 23258, 1],
60        [128, 128, 1],
61        [48, 128, 1],
62        [16, 1024, 1],
63        [32, 1024, 1],
64        [48, 1024, 1],
65        [16, 512, 1],
66        [32, 512, 1],
67        [48, 512, 1],
68        [16, 256, 1],
69        [32, 256, 1],
70        [48, 256, 1],
71    ],
72    cross_product_configs={
73        "device": ["cpu", "cuda"],
74    },
75    tags=["long"],
76)
77
78
79class SoftmaxBenchmark(op_bench.TorchBenchmarkBase):
80    def init(self, N, C, H, W, device, op_func):
81        self.inputs = {"input": torch.rand(N, C, H, W, device=device)}
82        self.op_func = op_func()
83
84    def forward(self, input):
85        return self.op_func(input)
86
87
88class Softmax2DimsBenchmark(op_bench.TorchBenchmarkBase):
89    def init(self, M, N, dim, device, op_func):
90        self.inputs = {"input": torch.rand(M, N, device=device)}
91        self.op_func = op_func(dim=dim)
92
93    def forward(self, input):
94        return self.op_func(input)
95
96
97op_bench.generate_pt_tests_from_op_list(
98    softmax_ops_list, softmax_configs_short + softmax_configs_long, SoftmaxBenchmark
99)
100
101
102op_bench.generate_pt_tests_from_op_list(
103    softmax_two_dims_ops_list, softmax_two_dims_configs, Softmax2DimsBenchmark
104)
105
106
107if __name__ == "__main__":
108    op_bench.benchmark_runner.main()
109