/
pad_op.cc
437 lines (404 loc) · 20.2 KB
/
pad_op.cc
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
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
/* 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/nn_ops.cc.
#define EIGEN_USE_THREADS
#if !defined(PLUGGABLE_DEVICE_SUPPORTED_MACOS) && defined(__APPLE__) && \
!defined(ANDROID) && !defined(__ANDROID__) && \
(!defined(TARGET_OS_IOS) || !TARGET_OS_IOS)
#define PLUGGABLE_DEVICE_SUPPORTED_MACOS 1
#endif
#include "tensorflow/core/kernels/pad_op.h"
#include <memory>
#include <string>
#include <utility>
#include "unsupported/Eigen/CXX11/Tensor" // from @eigen_archive
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.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/platform/logging.h"
#include "tensorflow/core/platform/types.h"
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
template <typename Device, typename T, typename Tpadding>
class PadOp : public OpKernel {
public:
explicit PadOp(OpKernelConstruction* context) : OpKernel(context) {}
void Compute(OpKernelContext* context) override {
const Tensor& in0 = context->input(0);
const Tensor& in1 = context->input(1);
const int dims = in0.dims();
static const int kMinDims = 0;
static const int kMaxDims = 8;
OP_REQUIRES(context, kMinDims <= dims && dims <= kMaxDims,
errors::Unimplemented("inputs rank not in [", kMinDims, ",",
kMaxDims, "]: ", dims));
OP_REQUIRES(
context,
TensorShapeUtils::IsMatrix(in1.shape()) && in1.dim_size(1) == 2,
errors::InvalidArgument("paddings must be a matrix with 2 columns: ",
in1.shape().DebugString()));
OP_REQUIRES(
context, dims == in1.dim_size(0),
errors::InvalidArgument(
"The first dimension of paddings must be the rank of inputs",
in1.shape().DebugString(), " ", in0.shape().DebugString()));
T pad_value = T();
if (context->num_inputs() == 3) {
const Tensor& constant_values = context->input(2);
OP_REQUIRES(
context, TensorShapeUtils::IsScalar(constant_values.shape()),
errors::InvalidArgument("constant_values must be a scalar. Found: ",
constant_values.shape().DebugString()));
pad_value = context->input(2).scalar<T>()();
}
// Compute the shape of the output tensor, and allocate it.
TensorShape output_shape;
typename TTypes<Tpadding>::ConstMatrix paddings = in1.matrix<Tpadding>();
for (int d = 0; d < dims; ++d) {
const Tpadding before_d =
paddings(d, 0); // Pad before existing elements.
const Tpadding after_d = paddings(d, 1); // Pad after existing elements.
OP_REQUIRES(context, before_d >= 0 && after_d >= 0,
errors::InvalidArgument("Paddings must be non-negative: ",
before_d, " ", after_d));
const int64_t size_d = in0.dim_size(d);
OP_REQUIRES_OK(
context, output_shape.AddDimWithStatus(before_d + size_d + after_d));
}
// If there is no padding to be done, forward the input to output.
if (output_shape.num_elements() == in0.NumElements()) {
// When num_elements == 0, shape may have changed.
Tensor out;
CHECK(out.CopyFrom(in0, output_shape));
context->set_output(0, out);
return;
}
TensorShape collapsed_input_shape;
TensorShape collapsed_output_shape;
Tensor collapsed_paddings;
if (dims > 1 && CollapseAdjacentNonPaddedDimensions(
in0.shape(), in1, output_shape, &collapsed_input_shape,
&collapsed_paddings, &collapsed_output_shape)) {
Tensor collapsed_input;
CHECK(collapsed_input.CopyFrom(in0, collapsed_input_shape));
Tensor collapsed_output;
AllocatorAttributes alloc_attrs;
alloc_attrs.set_on_host(context->input_memory_type(0) == HOST_MEMORY);
OP_REQUIRES_OK(context,
context->allocate_temp(collapsed_input.dtype(),
collapsed_output_shape,
&collapsed_output, alloc_attrs));
const Tensor& collapsed_paddings_ref = collapsed_paddings;
typename TTypes<Tpadding>::ConstMatrix collapsed_paddings_matrix =
collapsed_paddings_ref.matrix<Tpadding>();
OperateWithVariableRank(context, collapsed_input_shape.dims(),
collapsed_input, collapsed_paddings_matrix,
pad_value, &collapsed_output);
Tensor output;
CHECK(output.CopyFrom(collapsed_output, output_shape));
context->set_output(0, output);
} else {
Tensor* output = nullptr;
OP_REQUIRES_OK(context,
context->allocate_output(0, output_shape, &output));
OperateWithVariableRank(context, dims, in0, paddings, pad_value, output);
}
}
private:
// Collapses adjacent dimensions that are not padded to one dimension for
// speed. Returns true if any two dimensions are collapsed. For example,
//
// Pad(input_shape=[8, 28, 28, 3],
// paddings=[[0, 0], [0, 0], [0, 0], [0, 1]]
// is equivalent to
// Pad(input_shape=[6272, 3],
// paddings=[[0, 0], [0, 1]])
//
// input_shape: the original input shape.
// paddings_as_tensor: the original paddings.
// output_shape: the original output shape.
// collapsed_input_shape: the input shape after collapsing.
// collapsed_paddings_as_tensor: the paddings after collapsing.
// collapsed_output_shape: the output shape after collapsing.
static bool CollapseAdjacentNonPaddedDimensions(
const TensorShape& input_shape, const Tensor& paddings_as_tensor,
const TensorShape& output_shape, TensorShape* collapsed_input_shape,
Tensor* collapsed_paddings_as_tensor,
TensorShape* collapsed_output_shape) {
bool collapsed = false;
typename TTypes<Tpadding>::ConstMatrix paddings =
paddings_as_tensor.matrix<Tpadding>();
std::vector<std::pair<int, int>> collapsed_paddings;
int i = 0;
while (i < paddings.dimension(0)) {
if (paddings(i, 0) != 0 || paddings(i, 1) != 0) {
// If padded, copy the original dimension over.
collapsed_input_shape->InsertDim(collapsed_input_shape->dims(),
input_shape.dim_size(i));
collapsed_output_shape->InsertDim(collapsed_output_shape->dims(),
output_shape.dim_size(i));
collapsed_paddings.push_back({paddings(i, 0), paddings(i, 1)});
++i;
} else {
// If not padded, find the next dimension that is padded and collapse
// all dimensions in between to one dimension.
int64_t collapsed_input_dim_size = input_shape.dim_size(i);
int64_t collapsed_output_dim_size = output_shape.dim_size(i);
++i;
while (i < paddings.dimension(0) && paddings(i, 0) == 0 &&
paddings(i, 1) == 0) {
collapsed = true;
collapsed_input_dim_size *= input_shape.dim_size(i);
collapsed_output_dim_size *= output_shape.dim_size(i);
++i;
}
collapsed_input_shape->InsertDim(collapsed_input_shape->dims(),
collapsed_input_dim_size);
collapsed_output_shape->InsertDim(collapsed_output_shape->dims(),
collapsed_output_dim_size);
collapsed_paddings.push_back({0, 0});
}
}
// Copy collapsed_paddings to collapsed_paddings_as_tensor.
*collapsed_paddings_as_tensor = Tensor(
paddings_as_tensor.dtype(),
TensorShape({static_cast<int64_t>(collapsed_paddings.size()), 2}));
auto collapsed_paddings_as_matrix =
collapsed_paddings_as_tensor->matrix<Tpadding>();
for (size_t i = 0; i < collapsed_paddings.size(); ++i) {
collapsed_paddings_as_matrix(i, 0) = collapsed_paddings[i].first;
collapsed_paddings_as_matrix(i, 1) = collapsed_paddings[i].second;
}
return collapsed;
}
void OperateWithVariableRank(OpKernelContext* context, int fixed_dims,
const Tensor& input,
typename TTypes<Tpadding>::ConstMatrix paddings,
T pad_value, Tensor* output) {
// Invoke the dims-specific implementation.
switch (fixed_dims) {
case 0:
Operate<0>(context, input.tensor<T, 0>(), paddings, pad_value, output);
break;
case 1:
// TODO(irving): Once Pad doesn't need a scalar special case,
// change flat to tensor. That is, once !allow_legacy_scalars().
Operate<1>(context, input.flat<T>(), paddings, pad_value, output);
break;
case 2:
Operate<2>(context, input.tensor<T, 2>(), paddings, pad_value, output);
break;
case 3:
Operate<3>(context, input.tensor<T, 3>(), paddings, pad_value, output);
break;
case 4:
Operate<4>(context, input.tensor<T, 4>(), paddings, pad_value, output);
break;
case 5:
Operate<5>(context, input.tensor<T, 5>(), paddings, pad_value, output);
break;
case 6:
Operate<6>(context, input.tensor<T, 6>(), paddings, pad_value, output);
break;
default:
OP_REQUIRES(context, false,
errors::InvalidArgument("Only ranks up to 6 supported: ",
input.shape().DebugString()));
}
}
template <int Dims>
void Operate(OpKernelContext* context,
typename TTypes<T, Dims>::ConstTensor input,
typename TTypes<Tpadding>::ConstMatrix paddings, T pad_value,
Tensor* output) {
CHECK_EQ(Dims, paddings.dimension(0));
CHECK_EQ(2, paddings.dimension(1));
Eigen::array<Eigen::IndexPair<Tpadding>, Dims> paddings_array;
for (int i = 0; i < Dims; ++i) {
paddings_array[i] = {paddings(i, 0), paddings(i, 1)};
}
functor::Pad<Device, T, Tpadding, Dims> functor;
functor(context->eigen_device<Device>(), output->tensor<T, Dims>(), input,
paddings_array, pad_value);
}
};
#define REGISTER_KERNEL(type) \
REGISTER_KERNEL_BUILDER(Name("Pad") \
.Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \
.TypeConstraint<int32>("Tpaddings") \
.HostMemory("paddings"), \
PadOp<CPUDevice, type, int32>); \
REGISTER_KERNEL_BUILDER(Name("Pad") \
.Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \
.TypeConstraint<int64_t>("Tpaddings") \
.HostMemory("paddings"), \
PadOp<CPUDevice, type, int64>); \
REGISTER_KERNEL_BUILDER(Name("PadV2") \
.Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \
.TypeConstraint<int32>("Tpaddings") \
.HostMemory("paddings") \
.HostMemory("constant_values"), \
PadOp<CPUDevice, type, int32>); \
REGISTER_KERNEL_BUILDER(Name("PadV2") \
.Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \
.TypeConstraint<int64_t>("Tpaddings") \
.HostMemory("paddings") \
.HostMemory("constant_values"), \
PadOp<CPUDevice, type, int64>);
TF_CALL_POD_TYPES(REGISTER_KERNEL);
TF_CALL_QUANTIZED_TYPES(REGISTER_KERNEL);
TF_CALL_tstring(REGISTER_KERNEL);
#undef REGISTER_KERNEL
#if defined(PLUGGABLE_DEVICE_SUPPORTED_MACOS)
REGISTER_KERNEL_BUILDER(Name("Pad")
.Device(DEVICE_DEFAULT)
.TypeConstraint<int32>("T")
.TypeConstraint<int32>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("output"),
PadOp<CPUDevice, int32, int32>);
REGISTER_KERNEL_BUILDER(Name("Pad")
.Device(DEVICE_DEFAULT)
.TypeConstraint<int32>("T")
.TypeConstraint<int64>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("output"),
PadOp<CPUDevice, int32, int64>);
REGISTER_KERNEL_BUILDER(Name("PadV2")
.Device(DEVICE_DEFAULT)
.TypeConstraint<int32>("T")
.TypeConstraint<int32>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("constant_values")
.HostMemory("output"),
PadOp<CPUDevice, int32, int32>);
REGISTER_KERNEL_BUILDER(Name("PadV2")
.Device(DEVICE_DEFAULT)
.TypeConstraint<int32>("T")
.TypeConstraint<int64>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("constant_values")
.HostMemory("output"),
PadOp<CPUDevice, int32, int64>);
#endif
#if (defined(GOOGLE_CUDA) && GOOGLE_CUDA) || \
(defined(TENSORFLOW_USE_ROCM) && TENSORFLOW_USE_ROCM)
// Forward declarations of the functor specializations for GPU.
namespace functor {
#define DECLARE_GPU_SPEC(T, Dims) \
template <> \
void Pad<GPUDevice, T, int32, Dims>::operator()( \
const GPUDevice& d, typename TTypes<T, Dims>::Tensor output, \
typename TTypes<T, Dims>::ConstTensor input, \
Eigen::array<Eigen::IndexPair<int32>, Dims> paddings, T pad_value); \
extern template struct Pad<GPUDevice, T, int32, Dims>; \
template <> \
void Pad<GPUDevice, T, int64_t, Dims>::operator()( \
const GPUDevice& d, typename TTypes<T, Dims>::Tensor output, \
typename TTypes<T, Dims>::ConstTensor input, \
Eigen::array<Eigen::IndexPair<int64_t>, Dims> paddings, T pad_value); \
extern template struct Pad<GPUDevice, T, int64_t, Dims>;
#define DECLARE_GPU_SPECS(T) \
DECLARE_GPU_SPEC(T, 0); \
DECLARE_GPU_SPEC(T, 1); \
DECLARE_GPU_SPEC(T, 2); \
DECLARE_GPU_SPEC(T, 3); \
DECLARE_GPU_SPEC(T, 4); \
DECLARE_GPU_SPEC(T, 5); \
DECLARE_GPU_SPEC(T, 6);
TF_CALL_GPU_ALL_TYPES(DECLARE_GPU_SPECS);
TF_CALL_int8(DECLARE_GPU_SPECS);
TF_CALL_uint8(DECLARE_GPU_SPECS);
} // namespace functor
// Registration of the GPU implementations.
#define REGISTER_GPU_KERNEL(T) \
REGISTER_KERNEL_BUILDER(Name("Pad") \
.Device(DEVICE_GPU) \
.TypeConstraint<T>("T") \
.TypeConstraint<int32>("Tpaddings") \
.HostMemory("paddings"), \
PadOp<GPUDevice, T, int32>); \
REGISTER_KERNEL_BUILDER(Name("Pad") \
.Device(DEVICE_GPU) \
.TypeConstraint<T>("T") \
.TypeConstraint<int64_t>("Tpaddings") \
.HostMemory("paddings"), \
PadOp<GPUDevice, T, int64>); \
REGISTER_KERNEL_BUILDER(Name("PadV2") \
.Device(DEVICE_GPU) \
.TypeConstraint<T>("T") \
.TypeConstraint<int32>("Tpaddings") \
.HostMemory("paddings") \
.HostMemory("constant_values"), \
PadOp<GPUDevice, T, int32>) \
REGISTER_KERNEL_BUILDER(Name("PadV2") \
.Device(DEVICE_GPU) \
.TypeConstraint<T>("T") \
.TypeConstraint<int64_t>("Tpaddings") \
.HostMemory("paddings") \
.HostMemory("constant_values"), \
PadOp<GPUDevice, T, int64>)
TF_CALL_GPU_ALL_TYPES(REGISTER_GPU_KERNEL);
TF_CALL_int8(REGISTER_GPU_KERNEL);
TF_CALL_uint8(REGISTER_GPU_KERNEL);
// A special GPU kernel for int32.
// TODO(b/25387198): Also enable int32 in device memory. This kernel
// registration requires all int32 inputs and outputs to be in host memory.
REGISTER_KERNEL_BUILDER(Name("Pad")
.Device(DEVICE_GPU)
.TypeConstraint<int32>("T")
.TypeConstraint<int32>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("output"),
PadOp<CPUDevice, int32, int32>);
REGISTER_KERNEL_BUILDER(Name("Pad")
.Device(DEVICE_GPU)
.TypeConstraint<int32>("T")
.TypeConstraint<int64_t>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("output"),
PadOp<CPUDevice, int32, int64>);
REGISTER_KERNEL_BUILDER(Name("PadV2")
.Device(DEVICE_GPU)
.TypeConstraint<int32>("T")
.TypeConstraint<int32>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("constant_values")
.HostMemory("output"),
PadOp<CPUDevice, int32, int32>);
REGISTER_KERNEL_BUILDER(Name("PadV2")
.Device(DEVICE_GPU)
.TypeConstraint<int32>("T")
.TypeConstraint<int64_t>("Tpaddings")
.HostMemory("input")
.HostMemory("paddings")
.HostMemory("constant_values")
.HostMemory("output"),
PadOp<CPUDevice, int32, int64>);
#endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
} // end namespace tensorflow