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
|
/* 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.
==============================================================================*/
#ifndef TENSORFLOW_KERNELS_STRIDED_SLICE_OP_IMPL_H_
#define TENSORFLOW_KERNELS_STRIDED_SLICE_OP_IMPL_H_
// Functor definition for StridedSliceOp, must be compilable by nvcc.
#include "tensorflow/core/kernels/slice_op.h"
#include "tensorflow/core/kernels/strided_slice_op.h"
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/register_types_traits.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/kernels/bounds_check.h"
#include "tensorflow/core/kernels/dense_update_functor.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/lib/gtl/array_slice.h"
#include "tensorflow/core/platform/mem.h"
namespace tensorflow {
template <typename Device, typename T, int NDIM>
void HandleStridedSliceCase(OpKernelContext* context,
const gtl::ArraySlice<int64>& begin,
const gtl::ArraySlice<int64>& end,
const gtl::ArraySlice<int64>& strides,
const TensorShape& processing_shape,
bool is_simple_slice, Tensor* result);
template <typename Device, typename T, int NDIM>
void HandleStridedSliceGradCase(OpKernelContext* context,
const gtl::ArraySlice<int64>& begin,
const gtl::ArraySlice<int64>& end,
const gtl::ArraySlice<int64>& strides,
const TensorShape& processing_shape,
bool is_simple_slice, Tensor* result);
template <typename Device, typename T, int NDIM>
class HandleStridedSliceAssignCase {
public:
void operator()(OpKernelContext* context, const gtl::ArraySlice<int64>& begin,
const gtl::ArraySlice<int64>& end,
const gtl::ArraySlice<int64>& strides,
const TensorShape& processing_shape, bool is_simple_slice,
Tensor* result);
};
} // namespace tensorflow
// The actual implementation. This is designed so multiple
// translation units can include this file in the form
//
// #define STRIDED_SLICE_INSTANTIATE_DIM 1
// #include <thisfile>
// #undef STRIDED_SLICE_INSTANTIATE_DIM
//
#ifdef STRIDED_SLICE_INSTANTIATE_DIM
namespace tensorflow {
template <typename Device, typename T, int NDIM>
void HandleStridedSliceCase(OpKernelContext* context,
const gtl::ArraySlice<int64>& begin,
const gtl::ArraySlice<int64>& end,
const gtl::ArraySlice<int64>& strides,
const TensorShape& processing_shape,
bool is_simple_slice, Tensor* result) {
typedef typename proxy_type<Device, T>::type Proxy;
gtl::InlinedVector<int64, 4> processing_dims = processing_shape.dim_sizes();
if (is_simple_slice) {
gtl::InlinedVector<int64, 4> sizes(begin.size());
for (int i = 0; i < NDIM; ++i) {
sizes[i] = end[i] - begin[i];
}
const TensorShape final_shape = result->shape();
CHECK(result->CopyFrom(*result, processing_shape));
const Tensor input = context->input(0);
functor::Slice<Device, T, NDIM>()(
context->eigen_device<Device>(), result, input, begin, sizes);
CHECK(result->CopyFrom(*result, final_shape));
} else {
Eigen::DSizes<Eigen::DenseIndex, NDIM> begin_di;
Eigen::DSizes<Eigen::DenseIndex, NDIM> end_di;
Eigen::DSizes<Eigen::DenseIndex, NDIM> strides_di;
for (int i = 0; i < NDIM; ++i) {
begin_di[i] = begin[i];
end_di[i] = end[i];
strides_di[i] = strides[i];
}
functor::StridedSlice<Device, Proxy, NDIM>()(
context->eigen_device<Device>(),
result->bit_casted_shaped<Proxy, NDIM>(processing_dims),
context->input(0).bit_casted_tensor<Proxy, NDIM>(), begin_di, end_di,
strides_di);
}
}
template <typename Device, typename T, int NDIM>
void HandleStridedSliceGradCase(OpKernelContext* context,
const gtl::ArraySlice<int64>& begin,
const gtl::ArraySlice<int64>& end,
const gtl::ArraySlice<int64>& strides,
const TensorShape& processing_shape,
bool is_simple_slice, Tensor* result) {
gtl::InlinedVector<int64, 4> processing_dims = processing_shape.dim_sizes();
Eigen::DSizes<Eigen::DenseIndex, NDIM> begin_di;
Eigen::DSizes<Eigen::DenseIndex, NDIM> end_di;
Eigen::DSizes<Eigen::DenseIndex, NDIM> strides_di;
for (int i = 0; i < NDIM; ++i) {
begin_di[i] = begin[i];
end_di[i] = end[i];
strides_di[i] = strides[i];
}
typedef typename proxy_type<Device, T>::type Proxy;
functor::StridedSliceGrad<Device, Proxy, NDIM>()(
context->eigen_device<Device>(), result->bit_casted_tensor<Proxy, NDIM>(),
context->input(4).bit_casted_shaped<Proxy, NDIM>(processing_dims),
begin_di, end_di, strides_di);
}
template <typename Device, typename T, int NDIM>
void HandleStridedSliceAssignCase<Device, T, NDIM>::operator()(
OpKernelContext* context, const gtl::ArraySlice<int64>& begin,
const gtl::ArraySlice<int64>& end, const gtl::ArraySlice<int64>& strides,
const TensorShape& processing_shape, bool is_simple_slice, Tensor* result) {
gtl::InlinedVector<int64, 4> processing_dims = processing_shape.dim_sizes();
typedef typename proxy_type<Device, T>::type Proxy;
Eigen::DSizes<Eigen::DenseIndex, NDIM> begin_di;
Eigen::DSizes<Eigen::DenseIndex, NDIM> end_di;
Eigen::DSizes<Eigen::DenseIndex, NDIM> strides_di;
for (int i = 0; i < NDIM; ++i) {
begin_di[i] = begin[i];
end_di[i] = end[i];
strides_di[i] = strides[i];
}
functor::StridedSliceAssign<Device, Proxy, NDIM>()(
context->eigen_device<Device>(), result->bit_casted_tensor<Proxy, NDIM>(),
context->input(4).bit_casted_shaped<Proxy, NDIM>(processing_dims),
begin_di, end_di, strides_di);
}
template <typename Device, typename T>
class HandleStridedSliceAssignCase<Device, T, 0> {
public:
enum { NDIM_PROXY = 1 };
void operator()(OpKernelContext* context, const gtl::ArraySlice<int64>& begin,
const gtl::ArraySlice<int64>& end,
const gtl::ArraySlice<int64>& strides,
const TensorShape& processing_shape, bool is_simple_slice,
Tensor* result) {
gtl::InlinedVector<int64, 1> processing_dims(1);
processing_dims[0] = 1;
typedef typename proxy_type<Device, T>::type Proxy;
functor::StridedSliceAssignScalar<Device, Proxy>()(
context->eigen_device<Device>(),
result->bit_casted_shaped<Proxy, 1>(processing_dims),
context->input(4).bit_casted_shaped<Proxy, 1>(processing_dims));
}
};
// NODE(aselle): according to bsteiner, we need this because otherwise
// nvcc instantiates templates that are invalid. strided_slice_op_gpu.cu
// handles instantiates externally. It is important that this is done#
// before the HandleXXCase's are instantiated to avoid duplicate
// specialization errors.
#define PREVENT_INSTANTIATE_DIM1_AND_UP(T, NDIM) \
namespace functor { \
template <> \
void StridedSlice<GPUDevice, T, NDIM>::operator()( \
const GPUDevice& d, typename TTypes<T, NDIM>::Tensor output, \
typename TTypes<T, NDIM>::ConstTensor input, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& start, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& stop, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& strides); \
extern template struct StridedSlice<GPUDevice, T, NDIM>; \
template <> \
void Slice<GPUDevice, T, NDIM>::operator()( \
const GPUDevice& d, Tensor* output, const Tensor& input, \
const gtl::ArraySlice<int64>& slice_indices, \
const gtl::ArraySlice<int64>& slice_sizes); \
extern template struct Slice<GPUDevice, T, NDIM>; \
template <> \
void StridedSliceGrad<GPUDevice, T, NDIM>::operator()( \
const GPUDevice& d, typename TTypes<T, NDIM>::Tensor output, \
typename TTypes<T, NDIM>::ConstTensor input, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& start, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& stop, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& strides); \
extern template struct StridedSliceGrad<GPUDevice, T, NDIM>; \
template <> \
void StridedSliceAssign<GPUDevice, T, NDIM>::operator()( \
const GPUDevice& d, typename TTypes<T, NDIM>::Tensor output, \
typename TTypes<T, NDIM>::ConstTensor input, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& start, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& stop, \
const Eigen::DSizes<Eigen::DenseIndex, NDIM>& strides); \
extern template struct StridedSliceAssign<GPUDevice, T, NDIM>; \
} // namespace functor
#define PREVENT_INSTANTIATE_DIM0_ONLY(T, NDIM) \
namespace functor { \
template <> \
void StridedSliceAssignScalar<GPUDevice, T>::operator()( \
const GPUDevice& d, typename TTypes<T, 1>::Tensor output, \
typename TTypes<T, 1>::ConstTensor input); \
extern template struct StridedSliceAssignScalar<GPUDevice, T>; \
} // namespace functor
// Dimension 0 only instantiates some functors. So we only need
// to prevent ones defined by PREVENT_INSTANTIATE_DIM0_ONLY
#if GOOGLE_CUDA
#if STRIDED_SLICE_INSTANTIATE_DIM == 0
#define PREVENT_INSTANTIATE(T, NDIM) PREVENT_INSTANTIATE_DIM0_ONLY(T, NDIM)
#else
#define PREVENT_INSTANTIATE(T, NDIM) PREVENT_INSTANTIATE_DIM1_AND_UP(T, NDIM)
#endif
#else
#define PREVENT_INSTANTIATE(T, NDIM)
#endif
#define INSTANTIATE_DIM1_AND_UP_HANDLERS(DEVICE, T, DIM) \
template void HandleStridedSliceCase<DEVICE, T, DIM>( \
OpKernelContext * context, const gtl::ArraySlice<int64>& begin, \
const gtl::ArraySlice<int64>& end, \
const gtl::ArraySlice<int64>& strides, \
const TensorShape& processing_shape, bool is_simple_slice, \
Tensor* result); \
template void HandleStridedSliceGradCase<DEVICE, T, DIM>( \
OpKernelContext * context, const gtl::ArraySlice<int64>& begin, \
const gtl::ArraySlice<int64>& end, \
const gtl::ArraySlice<int64>& strides, \
const TensorShape& processing_shape, bool is_simple_slice, \
Tensor* result);
#define INSTANTIATE_DIM0_AND_UP_HANDLERS(DEVICE, T, DIM) \
template class HandleStridedSliceAssignCase<DEVICE, T, DIM>;
// Only some kernels need to be instantiated on dim 0.
#if STRIDED_SLICE_INSTANTIATE_DIM == 0
#define INSTANTIATE(DEVICE, T, DIM) \
INSTANTIATE_DIM0_AND_UP_HANDLERS(DEVICE, T, DIM)
#else
#define INSTANTIATE(DEVICE, T, DIM) \
INSTANTIATE_DIM0_AND_UP_HANDLERS(DEVICE, T, DIM) \
INSTANTIATE_DIM1_AND_UP_HANDLERS(DEVICE, T, DIM)
#endif
#define DECLARE_FOR_N_CPU(T) \
INSTANTIATE(CPUDevice, T, STRIDED_SLICE_INSTANTIATE_DIM)
#define PREVENT_FOR_N_GPU(T) \
PREVENT_INSTANTIATE(T, STRIDED_SLICE_INSTANTIATE_DIM)
#define DECLARE_FOR_N_GPU(T) \
INSTANTIATE(GPUDevice, T, STRIDED_SLICE_INSTANTIATE_DIM)
#if GOOGLE_CUDA
TF_CALL_GPU_PROXY_TYPES(PREVENT_FOR_N_GPU);
TF_CALL_complex64(PREVENT_FOR_N_GPU);
TF_CALL_complex128(PREVENT_FOR_N_GPU);
TF_CALL_GPU_NUMBER_TYPES(DECLARE_FOR_N_GPU);
TF_CALL_complex64(DECLARE_FOR_N_GPU);
TF_CALL_complex128(DECLARE_FOR_N_GPU);
DECLARE_FOR_N_GPU(int32);
#endif // END GOOGLE_CUDA
TF_CALL_ALL_TYPES(DECLARE_FOR_N_CPU);
DECLARE_FOR_N_CPU(bfloat16);
#ifdef TENSORFLOW_USE_SYCL
#define PREVENT_FOR_N_SYCL(T) \
PREVENT_INSTANTIATE(T, STRIDED_SLICE_INSTANTIATE_DIM)
#define DECLARE_FOR_N_SYCL(T) \
INSTANTIATE(SYCLDevice, T, STRIDED_SLICE_INSTANTIATE_DIM)
TF_CALL_SYCL_PROXY_TYPES(PREVENT_FOR_N_SYCL);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(DECLARE_FOR_N_SYCL);
DECLARE_FOR_N_SYCL(int32);
#undef DECLARE_FOR_N_SYCL
#endif // TENSORFLOW_USE_SYCL
#undef INSTANTIATE
#undef DECLARE_FOR_N_CPU
#undef DECLARE_FOR_N_GPU
} // end namespace tensorflow
#endif // END STRIDED_SLICE_INSTANTIATE_DIM
#endif // TENSORFLOW_KERNELS_SLICE_OP_H_
|
