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
|
#include <gtest/gtest.h>
#include "tensorflow/core/common_runtime/kernel_benchmark_testlib.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/platform/test_benchmark.h"
#include "tensorflow/core/public/session_options.h"
#include "tensorflow/core/public/tensor.h"
namespace tensorflow {
// We focus on the single thread performance of training ops.
static SessionOptions InitSingleThreadedOptions() {
SessionOptions opts;
opts.config.set_intra_op_parallelism_threads(1);
opts.config.set_inter_op_parallelism_threads(1);
return opts;
}
static SessionOptions* GetOptions() {
static SessionOptions opts = InitSingleThreadedOptions();
return &opts;
}
static Node* Var(Graph* g, int n) {
return test::graph::Var(g, DT_FLOAT, TensorShape({n}));
}
static Node* Zeros(Graph* g, int n) {
Tensor data(DT_FLOAT, TensorShape({n}));
data.flat<float>().setZero();
return test::graph::Constant(g, data);
}
static Node* Random(Graph* g, int n) {
Tensor data(DT_FLOAT, TensorShape({n}));
data.flat<float>().setRandom();
return test::graph::Constant(g, data);
}
static Node* Scalar(Graph* g, float val) {
Tensor data(DT_FLOAT, TensorShape({}));
data.flat<float>()(0) = val;
return test::graph::Constant(g, data);
}
static void SGD(int32 n, Graph** init_g, Graph** train_g) {
RequireDefaultOps();
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
test::graph::Assign(g, var, Zeros(g, n));
*init_g = g;
}
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto lr = Scalar(g, 0.01);
auto grad = Random(g, n);
test::graph::Multi(g, "ApplyGradientDescent", {var, lr, grad});
*train_g = g;
}
}
static void BM_SGD(int iters, int params) {
const int64 tot = static_cast<int64>(iters) * params;
testing::ItemsProcessed(tot);
testing::BytesProcessed(tot * sizeof(float));
Graph* init;
Graph* train;
SGD(params, &init, &train);
test::Benchmark("cpu", train, GetOptions(), init).Run(iters);
}
BENCHMARK(BM_SGD)->Arg(128 << 10)->Arg(256 << 10);
static void Adagrad(int32 n, Graph** init_g, Graph** train_g) {
RequireDefaultOps();
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto accum = Var(g, n);
auto zero = Zeros(g, n);
test::graph::Assign(g, var, zero);
test::graph::Assign(g, accum, zero);
*init_g = g;
}
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto accum = Var(g, n);
auto lr = Scalar(g, 0.01);
auto grad = Random(g, n);
test::graph::Multi(g, "ApplyAdagrad", {var, accum, lr, grad});
*train_g = g;
}
}
static void BM_Adagrad(int iters, int params) {
const int64 tot = static_cast<int64>(iters) * params;
testing::ItemsProcessed(tot);
testing::BytesProcessed(tot * sizeof(float));
Graph* init;
Graph* train;
Adagrad(params, &init, &train);
test::Benchmark("cpu", train, GetOptions(), init).Run(iters);
}
BENCHMARK(BM_Adagrad)->Arg(128 << 10)->Arg(256 << 10);
static void Momentum(int32 n, Graph** init_g, Graph** train_g) {
RequireDefaultOps();
TensorShape shape({n});
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto accum = Var(g, n);
auto zero = Zeros(g, n);
test::graph::Assign(g, var, zero);
test::graph::Assign(g, accum, zero);
*init_g = g;
}
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto accum = Var(g, n);
auto lr = Scalar(g, 0.01);
auto grad = Random(g, n);
auto mom = Scalar(g, 0.01);
test::graph::Multi(g, "ApplyMomentum", {var, accum, lr, grad, mom});
*train_g = g;
}
}
static void BM_Momentum(int iters, int params) {
const int64 tot = static_cast<int64>(iters) * params;
testing::ItemsProcessed(tot);
testing::BytesProcessed(tot * sizeof(float));
Graph* init;
Graph* train;
Momentum(params, &init, &train);
test::Benchmark("cpu", train, GetOptions(), init).Run(iters);
}
BENCHMARK(BM_Momentum)->Arg(128 << 10)->Arg(256 << 10);
static void Adam(int32 n, Graph** init_g, Graph** train_g) {
RequireDefaultOps();
TensorShape shape({n});
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto m = Var(g, n);
auto v = Var(g, n);
auto zero = Zeros(g, n);
test::graph::Assign(g, var, zero);
test::graph::Assign(g, m, zero);
test::graph::Assign(g, v, zero);
*init_g = g;
}
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto m = Var(g, n);
auto v = Var(g, n);
auto beta1_power = Scalar(g, 0.9);
auto beta2_power = Scalar(g, 0.99);
auto lr = Scalar(g, 0.01);
auto beta1 = Scalar(g, 0.9);
auto beta2 = Scalar(g, 0.99);
auto epsilon = Scalar(g, 1e-8);
auto grad = Random(g, n);
test::graph::Multi(g, "ApplyAdam", {var, m, v, beta1_power, beta2_power, lr,
beta1, beta2, epsilon, grad});
*train_g = g;
}
}
static void BM_Adam(int iters, int params) {
const int64 tot = static_cast<int64>(iters) * params;
testing::ItemsProcessed(tot);
testing::BytesProcessed(tot * sizeof(float));
Graph* init;
Graph* train;
Adam(params, &init, &train);
test::Benchmark("cpu", train, GetOptions(), init).Run(iters);
}
BENCHMARK(BM_Adam)->Arg(128 << 10)->Arg(256 << 10);
static void RMSProp(int32 n, Graph** init_g, Graph** train_g) {
RequireDefaultOps();
TensorShape shape({n});
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto ms = Var(g, n);
auto mom = Var(g, n);
auto zero = Zeros(g, n);
test::graph::Assign(g, var, zero);
test::graph::Assign(g, ms, zero);
test::graph::Assign(g, mom, zero);
*init_g = g;
}
{
Graph* g = new Graph(OpRegistry::Global());
auto var = Var(g, n);
auto ms = Var(g, n);
auto mom = Var(g, n);
auto lr = Scalar(g, 0.01);
auto rho = Scalar(g, 0.9);
auto momentum = Scalar(g, 0.9);
auto epsilon = Scalar(g, 1e-8);
auto grad = Random(g, n);
test::graph::Multi(g, "ApplyRMSProp",
{var, ms, mom, lr, rho, momentum, epsilon, grad});
*train_g = g;
}
}
static void BM_RMSProp(int iters, int params) {
const int64 tot = static_cast<int64>(iters) * params;
testing::ItemsProcessed(tot);
testing::BytesProcessed(tot * sizeof(float));
Graph* init;
Graph* train;
RMSProp(params, &init, &train);
test::Benchmark("cpu", train, GetOptions(), init).Run(iters);
}
BENCHMARK(BM_RMSProp)->Arg(128 << 10)->Arg(256 << 10);
} // end namespace tensorflow
|
