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
|
/* Copyright 2016 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 <cstddef>
#include <cstdlib>
#include <string>
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/kernel_def_builder.h"
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/tensor_types.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/kernels/bounds_check.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/stringpiece.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/util/bcast.h"
namespace tensorflow {
// Position/length can be 32 or 64-bit integers
template <typename T>
class SubstrOp : public OpKernel {
public:
using OpKernel::OpKernel;
void Compute(OpKernelContext* context) override {
// Get inputs
const Tensor& input_tensor = context->input(0);
const Tensor& pos_tensor = context->input(1);
const Tensor& len_tensor = context->input(2);
const TensorShape& input_shape = input_tensor.shape();
const TensorShape& pos_shape = pos_tensor.shape();
bool is_scalar = TensorShapeUtils::IsScalar(pos_shape);
if (is_scalar || input_shape == pos_shape) {
// pos/len are either scalar or match the shape of input_tensor
// Do not need to do broadcasting
// Reshape input
auto input = input_tensor.flat<string>();
// Allocate output
Tensor* output_tensor = nullptr;
OP_REQUIRES_OK(context,
context->allocate_output("output", input_tensor.shape(),
&output_tensor));
auto output = output_tensor->flat<string>();
if (is_scalar) {
// Perform Op with scalar pos/len
const T pos =
tensorflow::internal::SubtleMustCopy(pos_tensor.scalar<T>()());
const T len =
tensorflow::internal::SubtleMustCopy(len_tensor.scalar<T>()());
for (size_t i = 0; i < input_tensor.NumElements(); ++i) {
StringPiece in(input(i));
OP_REQUIRES(
context, FastBoundsCheck(std::abs(pos), in.size() + 1),
errors::InvalidArgument("pos ", pos, " out of range for string",
"b'", in, "' at index ", i));
StringPiece sub_in = in.substr(AdjustedPosIndex(pos, in), len);
output(i).assign(sub_in.data(), sub_in.size());
}
} else {
// Perform Op element-wise with tensor pos/len
auto pos_flat = pos_tensor.flat<T>();
auto len_flat = len_tensor.flat<T>();
for (size_t i = 0; i < input_tensor.NumElements(); ++i) {
StringPiece in(input(i));
const T pos = tensorflow::internal::SubtleMustCopy(pos_flat(i));
const T len = tensorflow::internal::SubtleMustCopy(len_flat(i));
OP_REQUIRES(
context, FastBoundsCheck(std::abs(pos), in.size() + 1),
errors::InvalidArgument("pos ", pos, " out of range for string",
"b'", in, "' at index ", i));
StringPiece sub_in = in.substr(AdjustedPosIndex(pos, in), len);
output(i).assign(sub_in.data(), sub_in.size());
}
}
} else {
// Perform op with broadcasting
// TODO: Use ternary broadcasting for once available in Eigen. Current
// implementation iterates through broadcasted ops element-wise;
// this should be parallelized.
// Create BCast helper with shape of input and pos/len
BCast bcast(BCast::FromShape(input_shape), BCast::FromShape(pos_shape));
OP_REQUIRES(context, bcast.IsValid(),
errors::InvalidArgument(
"Incompatible shapes: ", input_shape.DebugString(),
" vs. ", pos_shape.DebugString()));
TensorShape output_shape = BCast::ToShape(bcast.result_shape());
int ndims = output_shape.dims();
Tensor* output_tensor = nullptr;
OP_REQUIRES_OK(context, context->allocate_output("output", output_shape,
&output_tensor));
switch (ndims) {
case 1: {
// Reshape tensors according to BCast results
auto input = input_tensor.shaped<string, 1>(bcast.x_reshape());
auto output = output_tensor->shaped<string, 1>(bcast.result_shape());
auto pos_shaped = pos_tensor.shaped<T, 1>(bcast.y_reshape());
auto len_shaped = len_tensor.shaped<T, 1>(bcast.y_reshape());
// Allocate temporary buffer for broadcasted input tensor
Tensor input_buffer;
OP_REQUIRES_OK(context, context->allocate_temp(
DT_STRING, output_shape, &input_buffer));
TTypes<string, 1>::Tensor input_bcast =
input_buffer.shaped<string, 1>(bcast.result_shape());
input_bcast =
input.broadcast(BCast::ToIndexArray<1>(bcast.x_bcast()));
// Allocate temporary buffer for broadcasted position tensor
Tensor pos_buffer;
OP_REQUIRES_OK(context,
context->allocate_temp(DataTypeToEnum<T>::v(),
output_shape, &pos_buffer));
typename TTypes<T, 1>::Tensor pos_bcast(
pos_buffer.shaped<T, 1>(bcast.result_shape()));
pos_bcast =
pos_shaped.broadcast(BCast::ToIndexArray<1>(bcast.y_bcast()));
// Allocate temporary buffer for broadcasted length tensor
Tensor len_buffer;
OP_REQUIRES_OK(context,
context->allocate_temp(DataTypeToEnum<T>::v(),
output_shape, &len_buffer));
typename TTypes<T, 1>::Tensor len_bcast(
len_buffer.shaped<T, 1>(bcast.result_shape()));
len_bcast =
len_shaped.broadcast(BCast::ToIndexArray<1>(bcast.y_bcast()));
// Iterate through broadcasted tensors and perform substr
for (int i = 0; i < output_shape.dim_size(0); ++i) {
StringPiece in(input_bcast(i));
const T pos = tensorflow::internal::SubtleMustCopy(pos_bcast(i));
const T len = tensorflow::internal::SubtleMustCopy(len_bcast(i));
OP_REQUIRES(
context,
FastBoundsCheck(std::abs(pos), input_bcast(i).size() + 1),
errors::InvalidArgument("pos ", pos, " out of range for string",
"b'", in, "' at index ", i));
StringPiece sub_in = in.substr(AdjustedPosIndex(pos, in), len);
output(i).assign(sub_in.data(), sub_in.size());
}
break;
}
case 2: {
// Reshape tensors according to BCast results
auto input = input_tensor.shaped<string, 2>(bcast.x_reshape());
auto output = output_tensor->shaped<string, 2>(bcast.result_shape());
auto pos_shaped = pos_tensor.shaped<T, 2>(bcast.y_reshape());
auto len_shaped = len_tensor.shaped<T, 2>(bcast.y_reshape());
// Allocate temporary buffer for broadcasted input tensor
Tensor input_buffer;
OP_REQUIRES_OK(context, context->allocate_temp(
DT_STRING, output_shape, &input_buffer));
TTypes<string, 2>::Tensor input_bcast =
input_buffer.shaped<string, 2>(bcast.result_shape());
input_bcast =
input.broadcast(BCast::ToIndexArray<2>(bcast.x_bcast()));
// Allocate temporary buffer for broadcasted position tensor
Tensor pos_buffer;
OP_REQUIRES_OK(context,
context->allocate_temp(DataTypeToEnum<T>::v(),
output_shape, &pos_buffer));
typename TTypes<T, 2>::Tensor pos_bcast(
pos_buffer.shaped<T, 2>(bcast.result_shape()));
pos_bcast =
pos_shaped.broadcast(BCast::ToIndexArray<2>(bcast.y_bcast()));
// Allocate temporary buffer for broadcasted length tensor
Tensor len_buffer;
OP_REQUIRES_OK(context,
context->allocate_temp(DataTypeToEnum<T>::v(),
output_shape, &len_buffer));
typename TTypes<T, 2>::Tensor len_bcast(
len_buffer.shaped<T, 2>(bcast.result_shape()));
len_bcast =
len_shaped.broadcast(BCast::ToIndexArray<2>(bcast.y_bcast()));
// Iterate through broadcasted tensors and perform substr
for (int i = 0; i < output_shape.dim_size(0); ++i) {
for (int j = 0; j < output_shape.dim_size(1); ++j) {
StringPiece in(input_bcast(i, j));
const T pos =
tensorflow::internal::SubtleMustCopy(pos_bcast(i, j));
const T len =
tensorflow::internal::SubtleMustCopy(len_bcast(i, j));
OP_REQUIRES(
context, FastBoundsCheck(std::abs(pos), in.size() + 1),
errors::InvalidArgument("pos ", pos, " out of range for ",
"string b'", in, "' at index (", i,
", ", j, ")"));
StringPiece sub_in = in.substr(AdjustedPosIndex(pos, in), len);
output(i, j).assign(sub_in.data(), sub_in.size());
}
}
break;
}
default: {
context->SetStatus(errors::Unimplemented(
"Substr broadcast not implemented for ", ndims, " dimensions"));
}
}
}
}
private:
// This adjusts the requested position. Note it does not perform any bound
// checks.
T AdjustedPosIndex(const T pos_requested, const StringPiece s) {
if (pos_requested < 0) {
return s.size() + pos_requested;
}
return pos_requested;
}
};
#define REGISTER_SUBSTR(type) \
REGISTER_KERNEL_BUILDER( \
Name("Substr").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
SubstrOp<type>);
REGISTER_SUBSTR(int32);
REGISTER_SUBSTR(int64);
} // namespace tensorflow
|
