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
|
/* Copyright 2015 Google Inc. 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.
==============================================================================*/
#define EIGEN_USE_THREADS
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/kernels/concat_op.h"
#include "tensorflow/core/util/work_sharder.h"
namespace tensorflow {
template <typename T>
static inline void Copy(T* dst, const T* src, int n) {
if (DataTypeCanUseMemcpy(DataTypeToEnum<T>::v())) {
memcpy(dst, src, n * sizeof(T));
} else {
for (int k = 0; k < n; ++k) {
*dst++ = *src++;
}
}
}
template <typename T>
void ConcatCPU(DeviceBase* d,
const std::vector<
std::unique_ptr<typename TTypes<T, 2>::ConstMatrix>>& inputs,
typename TTypes<T, 2>::Matrix* output) {
int num_inputs = inputs.size();
std::vector<ptrdiff_t> sizes;
sizes.reserve(num_inputs);
int row_size = 0;
for (int j = 0; j < num_inputs; ++j) {
sizes.push_back(inputs[j]->dimension(1));
row_size += sizes.back();
}
auto worker_threads = d->tensorflow_cpu_worker_threads();
int num_threads = std::min<int>(std::min(4, worker_threads->num_threads),
output->size() / 4096);
// Single threaded mode.
if (num_threads == 0) {
T* out = &(*output)(0, 0);
std::vector<const T*> inp;
inp.reserve(num_inputs);
for (int j = 0; j < num_inputs; ++j) {
inp.push_back(&(*inputs[j])(0, 0));
}
const int dim0 = output->dimension(0);
for (int i = 0; i < dim0; ++i) {
for (int j = 0; j < num_inputs; ++j) {
auto size = sizes[j];
Copy(out, inp[j], size);
out += size;
inp[j] += size;
}
}
return;
}
// Sharded mode.
auto work = [&row_size, &sizes, &inputs, &output, &num_inputs](int64 start,
int64 end) {
int64 skipped_rows = start / row_size;
T* out = output->data() + skipped_rows * row_size;
T* out_start = output->data() + start;
T* out_end = output->data() + end;
// Handle partial row at start
if (out < out_start) {
for (int j = 0; j < num_inputs; ++j) {
ptrdiff_t size = sizes[j];
ptrdiff_t offset = out_start - out;
if (size <= offset) {
out += size;
continue;
}
const T* inp = &(*inputs[j])(skipped_rows, 0);
if (offset > 0) {
out += offset;
inp += offset;
size -= offset;
}
size = std::min(size, out_end - out);
if (size <= 0) break;
Copy(out, inp, size);
out += size;
}
++skipped_rows;
}
if (out == out_end) return;
CHECK(out >= out_start);
CHECK(out < out_end);
// Copy remaining data.
std::vector<const T*> inp;
inp.reserve(num_inputs);
for (int j = 0; j < num_inputs; ++j) {
inp.push_back(&(*inputs[j])(skipped_rows, 0));
}
const int dim0 = output->dimension(0);
for (int i = skipped_rows; i < dim0; ++i) {
for (int j = 0; j < num_inputs; ++j) {
ptrdiff_t size = std::min(sizes[j], out_end - out);
Copy(out, inp[j], size);
out += size;
inp[j] += size;
if (out == out_end) return;
}
}
};
Shard(num_threads, worker_threads->workers, output->size(), 100, work);
}
#define REGISTER(T) \
template void ConcatCPU<T>( \
DeviceBase*, \
const std::vector<std::unique_ptr<typename TTypes<T, 2>::ConstMatrix>>&, \
typename TTypes<T, 2>::Matrix* output);
TF_CALL_ALL_TYPES(REGISTER)
REGISTER(quint8)
REGISTER(qint8)
REGISTER(qint32)
REGISTER(bfloat16)
} // namespace tensorflow
|
