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
|
#include <functional>
#include <memory>
#include <vector>
#include "tensorflow/core/common_runtime/kernel_benchmark_testlib.h"
#include "tensorflow/core/framework/allocator.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/graph/testlib.h"
#include "tensorflow/core/graph/node_builder.h"
#include "tensorflow/core/kernels/ops_testutil.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/platform/test_benchmark.h"
#include "tensorflow/core/public/tensor.h"
#include <gtest/gtest.h>
#include "tensorflow/core/lib/core/status_test_util.h"
namespace tensorflow {
namespace {
// For the benchmark, we set up two 2-dimensional tensors, each kDim1 x 'dim'
// in size, and concat them together along "concat_dimension"
template <typename T>
static void ConcatHelper(int iters, int concat_dimension, int dim2) {
testing::StopTiming();
RequireDefaultOps();
Graph* g = new Graph(OpRegistry::Global());
DataType dt = DataTypeToEnum<T>::v();
const int kDim1 = 100;
Tensor concat_dim(DT_INT32, TensorShape({}));
concat_dim.scalar<int32>()() = concat_dimension;
Tensor in0(dt, TensorShape({kDim1, dim2}));
in0.flat<T>().setRandom();
Tensor in1(dt, TensorShape({kDim1, dim2}));
in1.flat<T>().setRandom();
Node* node;
TF_CHECK_OK(
NodeBuilder(g->NewName("n"), "Concat")
.Input(test::graph::Constant(g, concat_dim))
.Input({test::graph::Constant(g, in0), test::graph::Constant(g, in1)})
.Attr("N", 2)
.Attr("T", dt)
.Finalize(g, &node));
testing::BytesProcessed(static_cast<int64>(iters) *
((kDim1 * dim2) + (kDim1 * dim2)) * sizeof(T));
testing::StartTiming();
test::Benchmark("cpu", g).Run(iters);
testing::UseRealTime();
}
static void BM_ConcatDim0Float(int iters, int dim2) {
ConcatHelper<float>(iters, 0, dim2);
}
static void BM_ConcatDim1Float(int iters, int dim2) {
ConcatHelper<float>(iters, 1, dim2);
}
BENCHMARK(BM_ConcatDim0Float)->Arg(1000)->Arg(100000)->Arg(1000000);
BENCHMARK(BM_ConcatDim1Float)->Arg(1000)->Arg(100000)->Arg(1000000);
static void BM_ConcatDim1int16(int iters, int dim2) {
ConcatHelper<int16>(iters, 1, dim2);
}
static void BM_ConcatDim1bfloat16(int iters, int dim2) {
ConcatHelper<bfloat16>(iters, 1, dim2);
}
BENCHMARK(BM_ConcatDim1int16)->Arg(1000)->Arg(100000)->Arg(1000000);
BENCHMARK(BM_ConcatDim1bfloat16)->Arg(1000)->Arg(100000)->Arg(1000000);
template <typename T>
static void ConcatManyHelper(int iters, int concat_dimension, int dim2) {
testing::StopTiming();
RequireDefaultOps();
Graph* g = new Graph(OpRegistry::Global());
DataType dt = DataTypeToEnum<T>::v();
const int kDim1 = 40000;
const int kNumInputs = 64;
Tensor concat_dim(DT_INT32, TensorShape({}));
concat_dim.scalar<int32>()() = concat_dimension;
std::vector<NodeBuilder::NodeOut> inputs;
inputs.reserve(kNumInputs);
for (int i = 0; i < kNumInputs; ++i) {
Tensor in(dt, TensorShape({kDim1, dim2}));
in.flat<T>().setRandom();
inputs.push_back(test::graph::Constant(g, in));
}
Node* node;
TF_CHECK_OK(NodeBuilder(g->NewName("n"), "Concat")
.Input(test::graph::Constant(g, concat_dim))
.Input(inputs)
.Attr("N", 64)
.Attr("T", dt)
.Finalize(g, &node));
testing::BytesProcessed(static_cast<int64>(iters) * kDim1 * dim2 *
kNumInputs * sizeof(T));
testing::StartTiming();
test::Benchmark("cpu", g).Run(iters);
testing::UseRealTime();
}
static void BM_ConcatManyDim1bfloat16(int iters, int dim2) {
ConcatManyHelper<bfloat16>(iters, 1, dim2);
}
BENCHMARK(BM_ConcatManyDim1bfloat16)->Arg(18)->Arg(34)->Arg(60);
static void MemcpyAlternativeHelper(int iters, int concat_dimension, int dim2) {
testing::StopTiming();
const int kDim1 = 100;
std::vector<float> data1(kDim1 * dim2, 1.0f);
std::vector<float> data2(kDim1 * dim2, 2.0f);
testing::BytesProcessed(static_cast<int64>(iters) *
((kDim1 * dim2) + (kDim1 * dim2)) * sizeof(float));
testing::StartTiming();
while (--iters > 0) {
const int n0 = data1.size();
const int n1 = data2.size();
float* result = new float[n0 + n1];
memcpy(&result[0], &data1[0], n0 * sizeof(float));
memcpy(&result[n0], &data2[0], n1 * sizeof(float));
delete[] result;
}
}
static void BM_MemcpyAlternativeDim0(int iters, int dim2) {
MemcpyAlternativeHelper(iters, 0, dim2);
}
static void BM_MemcpyAlternativeDim1(int iters, int dim2) {
MemcpyAlternativeHelper(iters, 1, dim2);
}
BENCHMARK(BM_MemcpyAlternativeDim0)->Arg(1000)->Arg(100000)->Arg(1000000);
BENCHMARK(BM_MemcpyAlternativeDim1)->Arg(1000)->Arg(100000)->Arg(1000000);
typedef Eigen::TensorMap<Eigen::Tensor<bfloat16, 1, Eigen::RowMajor>,
Eigen::Unaligned> EigenMap;
static void MemcpyManyAlternative1(int iters, int dim2) {
testing::StopTiming();
const int kDim1 = 40000;
const int kNumCopies = 64;
const int size = kDim1 * dim2 * kNumCopies;
bfloat16* data = new bfloat16[size];
EigenMap map(data, size);
map.setRandom();
testing::BytesProcessed(static_cast<int64>(iters) * kDim1 * dim2 *
kNumCopies * sizeof(bfloat16));
testing::StartTiming();
while (iters-- > 0) {
std::vector<bfloat16*> inputs(kNumCopies);
for (int i = 0; i < kNumCopies; ++i) {
inputs[i] = &data[i * kDim1 * dim2];
}
bfloat16* result = new bfloat16[size];
for (int j = 0; j < kNumCopies; ++j) {
bfloat16* output = &result[j * dim2];
for (int i = 0; i < kDim1; ++i) {
if (i + 1 < kDim1) {
port::prefetch<port::PREFETCH_HINT_T0>(inputs[j] + dim2);
}
memcpy(output, inputs[j], dim2 * sizeof(bfloat16));
inputs[j] += dim2;
output += dim2 * kNumCopies;
}
}
delete[] result;
}
delete[] data;
}
static void MemcpyManyAlternative2(int iters, int dim2) {
testing::StopTiming();
const int kDim1 = 40000;
const int kNumCopies = 64;
const int size = kDim1 * dim2 * kNumCopies;
bfloat16* data = new bfloat16[size];
EigenMap map(data, size);
map.setRandom();
testing::BytesProcessed(static_cast<int64>(iters) * kDim1 * dim2 *
kNumCopies * sizeof(bfloat16));
testing::StartTiming();
std::vector<bfloat16*> inputs(kNumCopies);
while (--iters > 0) {
bfloat16* result = new bfloat16[size];
for (int i = 0; i < kNumCopies; ++i) {
inputs[i] = &data[i * kDim1 * dim2];
}
bfloat16* output = result;
for (int i = 0; i < kDim1; ++i) {
for (int j = 0; j < kNumCopies; ++j) {
if (j + 1 < kNumCopies) {
port::prefetch<port::PREFETCH_HINT_T0>(inputs[j + 1]);
}
memcpy(output, inputs[j], dim2 * sizeof(bfloat16));
inputs[j] += dim2;
output += dim2;
}
}
delete[] result;
}
delete[] data;
}
BENCHMARK(MemcpyManyAlternative1)
->Arg(16)
->Arg(17)
->Arg(18)
->Arg(32)
->Arg(33)
->Arg(34)
->Arg(60)
->Arg(64)
->Arg(65);
BENCHMARK(MemcpyManyAlternative2)
->Arg(16)
->Arg(17)
->Arg(18)
->Arg(32)
->Arg(33)
->Arg(34)
->Arg(60)
->Arg(64)
->Arg(65);
} // namespace
} // namespace tensorflow
|
