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
|
/* 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.
==============================================================================*/
// See docs in ../ops/data_flow_ops.cc.
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/kernels/bounds_check.h"
namespace tensorflow {
template <class T>
class DynamicStitchOp : public OpKernel {
public:
explicit DynamicStitchOp(OpKernelConstruction* c) : OpKernel(c) {
// Compute expected input signature
const DataType dt = DataTypeToEnum<T>::v();
const int n = c->num_inputs() / 2;
DataTypeVector expected;
for (int i = 0; i < n; i++) {
expected.push_back(DT_INT32);
}
for (int i = 0; i < n; i++) {
expected.push_back(dt);
}
OP_REQUIRES_OK(c, c->MatchSignature(expected, {dt}));
OP_REQUIRES(
c, c->num_inputs() > 0,
errors::InvalidArgument("DynamicStitchOp: Must have some inputs"));
OP_REQUIRES(c, c->num_inputs() % 2 == 0,
errors::InvalidArgument(
"DynamicStitchOp: Must have even number of arguments"));
}
void Compute(OpKernelContext* c) override {
// Find maximum index in the indices vectors
OpInputList indices_inputs;
OP_REQUIRES_OK(c, c->input_list("indices", &indices_inputs));
int32 max_index = -1;
for (const Tensor& indices : indices_inputs) {
if (indices.NumElements() > 0) {
Eigen::Tensor<int32, 0, Eigen::RowMajor> m =
indices.flat<int32>().maximum();
max_index = std::max(m(), max_index);
}
}
const int first_dim_size = max_index + 1;
// Validate that data[i].shape = indices[i].shape + constant
OpInputList data_inputs;
OP_REQUIRES_OK(c, c->input_list("data", &data_inputs));
const Tensor& data0 = data_inputs[0];
const Tensor& indices0 = indices_inputs[0];
for (int input_num = 0; input_num < indices_inputs.size(); input_num++) {
const Tensor& indices = indices_inputs[input_num];
const Tensor& data = data_inputs[input_num];
OP_REQUIRES(
c, TensorShapeUtils::StartsWith(data.shape(), indices.shape()),
errors::InvalidArgument("data[", input_num, "].shape = ",
data.shape().DebugString(),
" does not start with indices[", input_num,
"].shape = ", indices.shape().DebugString()));
OP_REQUIRES(
c, input_num == 0 || SameExtraShape(data0, indices0, data, indices),
errors::InvalidArgument(
"Need data[0].shape[", indices0.dims(), ":] = data[", input_num,
"].shape[", indices.dims(), ":], got data[0].shape = ",
data0.shape().DebugString(), ", data[", input_num, "].shape = ",
data.shape().DebugString(), ", indices[0].shape = ",
indices0.shape().DebugString(), ", indices[", input_num,
"].shape = ", indices.shape().DebugString()));
}
// Allocate result tensor of shape
// [first_dim_size] + data.shape[indices.dims:]
TensorShape result_shape;
result_shape.AddDim(first_dim_size);
for (int d = indices0.dims(); d < data0.dims(); d++) {
result_shape.AddDim(data0.dim_size(d));
}
Tensor* merged = nullptr;
OP_REQUIRES_OK(c, c->allocate_output(0, result_shape, &merged));
// TODO(jeff): Currently we leave uninitialized any portions of
// merged that aren't covered by an index in indices. What should we do?
if (first_dim_size > 0) {
auto merged_flat = merged->flat_outer_dims<T>();
const int slice_size = merged_flat.dimension(1);
for (int input_num = 0; input_num < indices_inputs.size(); input_num++) {
const Tensor& indices = indices_inputs[input_num];
auto indices_vec = indices.flat<int32>();
const Tensor& data = data_inputs[input_num];
auto data_flat =
data.shaped<T, 2>({indices_vec.dimension(0), slice_size});
if (DataTypeCanUseMemcpy(DataTypeToEnum<T>::v())) {
T* merged_base = &merged_flat(0, 0);
const T* data_base = &data_flat(0, 0);
const size_t slice_bytes = slice_size * sizeof(T);
for (int i = 0; i < indices_vec.size(); i++) {
int32 index = internal::SubtleMustCopy(indices_vec(i));
OP_REQUIRES(
c, FastBoundsCheck(index, first_dim_size),
errors::InvalidArgument("indices[", i, "] is out of range"));
memcpy(merged_base + index * slice_size, data_base + i * slice_size,
slice_bytes);
}
} else {
Eigen::DSizes<Eigen::DenseIndex, 2> sizes(1, slice_size);
for (int i = 0; i < indices_vec.size(); i++) {
// Copy slice data[i] to merged[indices[i]]
Eigen::DSizes<Eigen::DenseIndex, 2> data_indices(i, 0);
int32 index = internal::SubtleMustCopy(indices_vec(i));
OP_REQUIRES(
c, FastBoundsCheck(index, first_dim_size),
errors::InvalidArgument("indices[", i, "] is out of range"));
Eigen::DSizes<Eigen::DenseIndex, 2> merged_indices(index, 0);
merged_flat.slice(merged_indices, sizes) =
data_flat.slice(data_indices, sizes);
}
}
}
}
}
private:
// Check if data0.shape[indices0.dims():] == data1.shape[indices1.dims():]
static bool SameExtraShape(const Tensor& data0, const Tensor& indices0,
const Tensor& data1, const Tensor& indices1) {
const int extra0 = data0.dims() - indices0.dims();
const int extra1 = data1.dims() - indices1.dims();
if (extra0 != extra1) return false;
for (int i = 0; i < extra0; i++) {
if (data0.dim_size(indices0.dims() + i) !=
data1.dim_size(indices1.dims() + i)) {
return false;
}
}
return true;
}
};
#define REGISTER_DYNAMIC_STITCH(type) \
REGISTER_KERNEL_BUILDER(Name("DynamicStitch") \
.Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \
.HostMemory("indices"), \
DynamicStitchOp<type>)
TF_CALL_POD_STRING_TYPES(REGISTER_DYNAMIC_STITCH);
#undef REGISTER_DYNAMIC_STITCH
#if GOOGLE_CUDA
#define REGISTER_DYNAMIC_STITCH_GPU(type) \
REGISTER_KERNEL_BUILDER(Name("DynamicStitch") \
.Device(DEVICE_GPU) \
.TypeConstraint<type>("T") \
.HostMemory("indices") \
.HostMemory("data") \
.HostMemory("merged"), \
DynamicStitchOp<type>)
TF_CALL_POD_STRING_TYPES(REGISTER_DYNAMIC_STITCH_GPU);
#undef REGISTER_DYNAMIC_STITCH_GPU
#endif // GOOGLE_CUDA
} // namespace tensorflow
|
