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
|
/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/core/common_runtime/kernel_benchmark_testlib.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/platform/test_benchmark.h"
namespace tensorflow {
template <typename T>
static Graph* Matmul(int m, int k, int n, bool transpose_a, bool transpose_b,
DataType type) {
Graph* g = new Graph(OpRegistry::Global());
Tensor in0(type, transpose_a ? TensorShape({k, m}) : TensorShape({m, k}));
in0.flat<T>().setRandom();
Tensor in1(type, transpose_b ? TensorShape({n, k}) : TensorShape({k, n}));
in1.flat<T>().setRandom();
test::graph::Matmul(g, test::graph::Constant(g, in0),
test::graph::Constant(g, in1), transpose_a, transpose_b);
return g;
}
#define BM_MatmulDev(M, K, N, TA, TB, T, TFTYPE, DEVICE) \
static void BM_Matmul##_##M##_##K##_##N##_##TA##_##TB##_##TFTYPE##_##DEVICE( \
int iters) { \
testing::UseRealTime(); \
testing::ItemsProcessed(static_cast<int64>(iters) * M * K * N * 2); \
test::Benchmark(#DEVICE, Matmul<T>(M, K, N, TA, TB, TFTYPE)).Run(iters); \
} \
BENCHMARK(BM_Matmul##_##M##_##K##_##N##_##TA##_##TB##_##TFTYPE##_##DEVICE);
#define BM_Matmul(M, K, N, TA, TB) \
BM_MatmulDev(M, K, N, TA, TB, float, DT_FLOAT, cpu); \
BM_MatmulDev(M, K, N, TA, TB, std::complex<float>, DT_COMPLEX64, cpu); \
BM_MatmulDev(M, K, N, TA, TB, float, DT_FLOAT, gpu); \
BM_MatmulDev(M, K, N, TA, TB, std::complex<float>, DT_COMPLEX64, gpu); \
/* Uncomment to enable benchmarks for double/complex128: */ \
// BM_MatmulDev(M, K, N, TA, TB, double, DT_DOUBLE, cpu); \
// BM_MatmulDev(M, K, N, TA, TB, std::complex<double>, DT_COMPLEX128, cpu); \
// BM_MatmulDev(M, K, N, TA, TB, double, DT_DOUBLE, gpu); \
// BM_MatmulDev(M, K, N, TA, TB, std::complex<double>, DT_COMPLEX128, gpu);
// Batch size of 1 included for inference.
// Typical fully connected layers
BM_Matmul(1, 512, 512, false, false);
BM_Matmul(8, 512, 512, false, false);
BM_Matmul(16, 512, 512, false, false);
BM_Matmul(128, 512, 512, false, false);
BM_Matmul(1, 1024, 1024, false, false);
BM_Matmul(8, 1024, 1024, false, false);
BM_Matmul(16, 1024, 1024, false, false);
BM_Matmul(128, 1024, 1024, false, false);
BM_Matmul(4096, 4096, 4096, false, false);
// Backward for fully connected layers
BM_Matmul(1, 1024, 1024, false, true);
BM_Matmul(8, 1024, 1024, false, true);
BM_Matmul(16, 1024, 1024, false, true);
BM_Matmul(128, 1024, 1024, false, true);
// Forward softmax with large output size
BM_Matmul(1, 200, 10000, false, false);
BM_Matmul(8, 200, 10000, false, false);
BM_Matmul(20, 200, 10000, false, false);
BM_Matmul(20, 200, 20000, false, false);
// Backward softmax with large output size
BM_Matmul(1, 10000, 200, false, true);
BM_Matmul(1, 10000, 200, false, false);
BM_Matmul(8, 10000, 200, false, true);
BM_Matmul(20, 10000, 200, false, true);
BM_Matmul(20, 20000, 200, false, true);
// Test some matrix-vector multiplies.
BM_Matmul(50, 50, 1, false, false);
BM_Matmul(50, 50, 1, true, false);
BM_Matmul(50, 50, 1, false, true);
BM_Matmul(50, 50, 1, true, true);
BM_Matmul(500, 500, 1, false, false);
BM_Matmul(500, 500, 1, true, false);
BM_Matmul(500, 500, 1, false, true);
BM_Matmul(500, 500, 1, true, true);
BM_Matmul(2000, 2000, 1, false, false);
BM_Matmul(2000, 2000, 1, true, false);
BM_Matmul(2000, 2000, 1, false, true);
BM_Matmul(2000, 2000, 1, true, true);
// Test some vector-matrix multiplies.
BM_Matmul(1, 50, 50, false, false);
BM_Matmul(1, 50, 50, true, false);
BM_Matmul(1, 50, 50, false, true);
BM_Matmul(1, 50, 50, true, true);
BM_Matmul(1, 500, 500, false, false);
BM_Matmul(1, 500, 500, true, false);
BM_Matmul(1, 500, 500, false, true);
BM_Matmul(1, 500, 500, true, true);
BM_Matmul(1, 2000, 2000, false, false);
BM_Matmul(1, 2000, 2000, true, false);
BM_Matmul(1, 2000, 2000, false, true);
BM_Matmul(1, 2000, 2000, true, true);
// Test some rank-one products.
BM_Matmul(50, 1, 50, false, false);
BM_Matmul(50, 1, 50, true, false);
BM_Matmul(50, 1, 50, false, true);
BM_Matmul(50, 1, 50, true, true);
BM_Matmul(500, 1, 500, false, false);
BM_Matmul(500, 1, 500, true, false);
BM_Matmul(500, 1, 500, false, true);
BM_Matmul(500, 1, 500, true, true);
BM_Matmul(2000, 1, 2000, false, false);
BM_Matmul(2000, 1, 2000, true, false);
BM_Matmul(2000, 1, 2000, false, true);
BM_Matmul(2000, 1, 2000, true, true);
} // end namespace tensorflow
|
