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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
|
/* 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/graph/node_builder.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/platform/test_benchmark.h"
#include "tensorflow/core/util/tensor_format.h"
namespace tensorflow {
// Creates a Graph which applies a unary "func" on a 3D tensor of
// type T with "num" elements.
template <typename T>
static Graph* Unary(const string& func, int num, DataType dtype) {
Graph* g = new Graph(OpRegistry::Global());
Tensor data(dtype, TensorShape({64, 64, num / (64 * 64)}));
CHECK_GT(data.NumElements(), 0);
data.flat<T>().setRandom();
test::graph::Unary(g, func, test::graph::Constant(g, data), 0);
return g;
}
static int kRows = 100000;
static int RowsAndColsArg(int r, int c) { return r * kRows + c; }
static int RowsFromArg(int arg) { return (arg / kRows); }
static int ColsFromArg(int arg) { return (arg % kRows); }
#define BM_UNARY(DEVICE, FUNC, T, TYPE) \
static void BM_##DEVICE##_##FUNC##_##TYPE(int iters, int num) { \
const int64 tot = static_cast<int64>(iters) * num; \
testing::ItemsProcessed(tot); \
testing::BytesProcessed(tot * sizeof(T)); \
test::Benchmark(#DEVICE, Unary<T>(#FUNC, num, TYPE)).Run(iters); \
} \
BENCHMARK(BM_##DEVICE##_##FUNC##_##TYPE)->Range(4 << 10, 1 << 20);
BM_UNARY(cpu, Floor, float, DT_FLOAT);
#if GOOGLE_CUDA
BM_UNARY(gpu, Floor, float, DT_FLOAT);
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
BM_UNARY(sycl, Floor, float, DT_FLOAT);
#endif // TENSORFLOW_USE_SYCL
BM_UNARY(cpu, Floor, double, DT_DOUBLE);
#if GOOGLE_CUDA
BM_UNARY(gpu, Floor, double, DT_DOUBLE);
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
BM_UNARY(sycl, Floor, double, DT_DOUBLE);
#endif // TENSORFLOW_USE_SYCL
BM_UNARY(cpu, Conj, std::complex<float>, DT_COMPLEX64);
#if GOOGLE_CUDA
BM_UNARY(gpu, Conj, std::complex<float>, DT_COMPLEX64);
#endif // GOOGLE_CUDA
BM_UNARY(cpu, Conj, std::complex<double>, DT_COMPLEX128);
#if GOOGLE_CUDA
BM_UNARY(gpu, Conj, std::complex<double>, DT_COMPLEX128);
#endif // GOOGLE_CUDA
BM_UNARY(cpu, Rint, double, DT_DOUBLE);
#if GOOGLE_CUDA
BM_UNARY(gpu, Rint, double, DT_DOUBLE);
#endif // GOOGLE_CUDA
BM_UNARY(cpu, Rint, float, DT_FLOAT);
#if GOOGLE_CUDA
BM_UNARY(gpu, Rint, float, DT_FLOAT);
#endif // GOOGLE_CUDA
// data func scalar.
static Graph* BinaryScalar(int num, const string& func) {
Graph* g = new Graph(OpRegistry::Global());
Tensor lhs(DT_FLOAT, TensorShape({64, 64, num / (64 * 64)}));
lhs.flat<float>().setRandom();
Tensor rhs(DT_FLOAT, TensorShape({}));
rhs.flat<float>().setRandom();
test::graph::Binary(g, func, test::graph::Constant(g, lhs),
test::graph::Constant(g, rhs));
return g;
}
#define BM_BINARY_SCALAR(DEVICE, FUNC) \
static void BM_##DEVICE##_##FUNC##_scalar(int iters, int num) { \
const int64 tot = static_cast<int64>(iters) * num; \
testing::ItemsProcessed(tot); \
testing::BytesProcessed(tot * sizeof(float)); \
test::Benchmark(#DEVICE, BinaryScalar(num, #FUNC)).Run(iters); \
} \
BENCHMARK(BM_##DEVICE##_##FUNC##_scalar) \
->Arg(4096) /* must >= 4096 */ \
->Arg(32768) \
->Arg(131072) \
->Arg(1048576);
BM_BINARY_SCALAR(cpu, Less);
#if GOOGLE_CUDA
BM_BINARY_SCALAR(gpu, Less);
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
BM_BINARY_SCALAR(sycl, Less);
#endif // TENSORFLOW_USE_SYCL
BM_BINARY_SCALAR(cpu, Add);
#if GOOGLE_CUDA
BM_BINARY_SCALAR(gpu, Add);
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
BM_BINARY_SCALAR(sycl, Add);
#endif // TENSORFLOW_USE_SYCL
#undef BM_BINARY_SCALAR
template <class T>
static Graph* BiasAdd(int rows, int cols, DataType type) {
Graph* g = new Graph(OpRegistry::Global());
Tensor lhs(type, TensorShape({rows, cols}));
lhs.template flat<T>().setRandom();
TensorShape rhs_shape;
rhs_shape = TensorShape({cols});
Tensor rhs(type, rhs_shape);
rhs.template flat<T>().setRandom();
test::graph::Binary(g, "BiasAdd", test::graph::Constant(g, lhs),
test::graph::Constant(g, rhs));
return g;
}
#define BM_BIAS_ADD(DEVICE, C_TYPE, TF_TYPE, R, C) \
static void BM_##DEVICE##_##C_TYPE##_BiasAdd_R##R##_C##C(int iters, \
int arg) { \
const int rows = RowsFromArg(arg); \
const int cols = ColsFromArg(arg); \
const int64 tot = static_cast<int64>(iters) * rows * cols; \
testing::ItemsProcessed(tot); \
testing::BytesProcessed(tot * sizeof(C_TYPE)); \
test::Benchmark(#DEVICE, BiasAdd<C_TYPE>(rows, cols, TF_TYPE)).Run(iters); \
} \
BENCHMARK(BM_##DEVICE##_##C_TYPE##_BiasAdd_R##R##_C##C) \
->Arg(RowsAndColsArg(R, C));
#define BM_BIAS_ADD_ALL(DEVICE, C_TYPE, TF_TYPE) \
BM_BIAS_ADD(DEVICE, C_TYPE, TF_TYPE, 512, 2048); \
BM_BIAS_ADD(DEVICE, C_TYPE, TF_TYPE, 512, 4096); \
BM_BIAS_ADD(DEVICE, C_TYPE, TF_TYPE, 2048, 512); \
BM_BIAS_ADD(DEVICE, C_TYPE, TF_TYPE, 4096, 512);
using Eigen::half;
BM_BIAS_ADD_ALL(cpu, float, DT_FLOAT);
#if GOOGLE_CUDA
BM_BIAS_ADD_ALL(gpu, float, DT_FLOAT);
#endif // GOOGLE_CUDA
BM_BIAS_ADD_ALL(cpu, half, DT_HALF);
#if GOOGLE_CUDA
BM_BIAS_ADD_ALL(gpu, half, DT_HALF);
#endif // GOOGLE_CUDA
#undef BM_BIAS_ADD_ALL
#undef BM_BIAS_ADD
template <class T>
static Graph* BiasAddGrad(int rows, int cols, int channels, DataType type,
TensorFormat format) {
Graph* g = new Graph(OpRegistry::Global());
TensorShape lhs_shape;
if (format == FORMAT_NCHW) {
lhs_shape = TensorShape({channels, rows, cols});
} else {
lhs_shape = TensorShape({rows, cols, channels});
}
Tensor lhs(type, lhs_shape);
lhs.template flat<T>().setRandom();
Node* n;
TF_CHECK_OK(NodeBuilder(g->NewName("n"), "BiasAddGrad")
.Attr("data_format", ToString(format))
.Input(test::graph::Constant(g, lhs), /*index=*/0)
.Finalize(g, &n));
return g;
}
#define BM_BIAS_ADD_GRAD(DEVICE, FMT, C_TYPE, TF_TYPE, R, C, CH) \
static void \
BM_##DEVICE##_##FMT##_##C_TYPE##_BiasAddGrad_R##R##_C##C##_CH##CH( \
int iters, int arg, int channels) { \
const int rows = RowsFromArg(arg); \
const int cols = ColsFromArg(arg); \
const int64 tot = static_cast<int64>(iters) * rows * cols * channels; \
testing::ItemsProcessed(tot); \
testing::BytesProcessed(tot * sizeof(C_TYPE)); \
test::Benchmark(#DEVICE, BiasAddGrad<C_TYPE>(rows, cols, channels, \
TF_TYPE, FORMAT_##FMT)) \
.Run(iters); \
} \
BENCHMARK(BM_##DEVICE##_##FMT##_##C_TYPE##_BiasAddGrad_R##R##_C##C##_CH##CH) \
->ArgPair(RowsAndColsArg(R, C), CH);
#define BM_BIAS_ADD_GRAD_ALL(DEVICE, FORMAT, C_TYPE, TF_TYPE) \
BM_BIAS_ADD_GRAD(DEVICE, FORMAT, C_TYPE, TF_TYPE, 64, 64, 64); \
BM_BIAS_ADD_GRAD(DEVICE, FORMAT, C_TYPE, TF_TYPE, 512, 512, 4); \
BM_BIAS_ADD_GRAD(DEVICE, FORMAT, C_TYPE, TF_TYPE, 512, 512, 1); \
BM_BIAS_ADD_GRAD(DEVICE, FORMAT, C_TYPE, TF_TYPE, 4096, 4096, 4); \
BM_BIAS_ADD_GRAD(DEVICE, FORMAT, C_TYPE, TF_TYPE, 4096, 4096, 1);
using Eigen::half;
#if GOOGLE_CUDA
BM_BIAS_ADD_GRAD_ALL(gpu, NCHW, float, DT_FLOAT);
BM_BIAS_ADD_GRAD_ALL(gpu, NCHW, half, DT_HALF);
#endif // GOOGLE_CUDA
BM_BIAS_ADD_GRAD_ALL(cpu, NHWC, float, DT_FLOAT);
#if GOOGLE_CUDA
BM_BIAS_ADD_GRAD_ALL(gpu, NHWC, float, DT_FLOAT);
#endif // GOOGLE_CUDA
BM_BIAS_ADD_GRAD_ALL(cpu, NHWC, half, DT_HALF);
#if GOOGLE_CUDA
BM_BIAS_ADD_GRAD_ALL(gpu, NHWC, half, DT_HALF);
#endif // GOOGLE_CUDA
#undef BM_BIAS_ADD_GRAD_ALL
#undef BM_BIAS_ADD_GRAD
static Graph* BcastAdd(int rows, int cols, int dim) {
Graph* g = new Graph(OpRegistry::Global());
Tensor lhs(DT_FLOAT, TensorShape({rows, cols}));
lhs.flat<float>().setRandom();
TensorShape rhs_shape;
if (dim == 0) {
rhs_shape = TensorShape({rows, 1});
} else {
rhs_shape = TensorShape({cols});
}
Tensor rhs(DT_FLOAT, rhs_shape);
rhs.flat<float>().setRandom();
test::graph::Binary(g, "Add", test::graph::Constant(g, lhs),
test::graph::Constant(g, rhs));
return g;
}
#define BM_BCAST_ADD_ROW(DEVICE, R, C) \
static void BM_##DEVICE##_BcastAddRow_R##R##_C##C(int iters, int arg) { \
const int rows = RowsFromArg(arg); \
const int cols = ColsFromArg(arg); \
const int64 tot = static_cast<int64>(iters) * rows * cols; \
testing::ItemsProcessed(tot); \
testing::BytesProcessed(tot * sizeof(float)); \
test::Benchmark(#DEVICE, BcastAdd(rows, cols, 0)).Run(iters); \
} \
BENCHMARK(BM_##DEVICE##_BcastAddRow_R##R##_C##C)->Arg(RowsAndColsArg(R, C));
#define BM_BCAST_ADD_ROW_ALL(DEVICE) \
BM_BCAST_ADD_ROW(DEVICE, 512, 2048); \
BM_BCAST_ADD_ROW(DEVICE, 512, 4096); \
BM_BCAST_ADD_ROW(DEVICE, 2048, 512); \
BM_BCAST_ADD_ROW(DEVICE, 4096, 512);
BM_BCAST_ADD_ROW_ALL(cpu);
#if GOOGLE_CUDA
BM_BCAST_ADD_ROW_ALL(gpu);
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
BM_BCAST_ADD_ROW_ALL(sycl);
#endif // TENSORFLOW_USE_SYCL
#undef BM_BCAST_ADD_ROW_ALL
#undef BM_BCAST_ADD_ROW
#define BM_BCAST_ADD_COL(DEVICE, R, C) \
static void BM_##DEVICE##_BcastAddCol_R##R##_C##C(int iters, int arg) { \
const int rows = RowsFromArg(arg); \
const int cols = ColsFromArg(arg); \
const int64 tot = static_cast<int64>(iters) * rows * cols; \
testing::ItemsProcessed(tot); \
testing::BytesProcessed(tot * sizeof(float)); \
test::Benchmark(#DEVICE, BcastAdd(rows, cols, 1)).Run(iters); \
} \
BENCHMARK(BM_##DEVICE##_BcastAddCol_R##R##_C##C)->Arg(RowsAndColsArg(R, C));
#define BM_BCAST_ADD_COL_ALL(DEVICE) \
BM_BCAST_ADD_COL(DEVICE, 512, 2048); \
BM_BCAST_ADD_COL(DEVICE, 512, 4096); \
BM_BCAST_ADD_COL(DEVICE, 2048, 512); \
BM_BCAST_ADD_COL(DEVICE, 4096, 512);
BM_BCAST_ADD_COL_ALL(cpu);
#if GOOGLE_CUDA
BM_BCAST_ADD_COL_ALL(gpu);
#endif // GOOGLE_CUDA
#ifdef TENSORFLOW_USE_SYCL
BM_BCAST_ADD_COL_ALL(sycl);
#endif // TENSORFLOW_USE_SYCL
#undef BM_BCAST_ADD_COL_ALL
#undef BM_BCAST_ADD_COL
} // end namespace tensorflow
|
