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
|
/* 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_CORE_KERNELS_SCATTER_ND_OP_CPU_IMPL_H_
#define TENSORFLOW_CORE_KERNELS_SCATTER_ND_OP_CPU_IMPL_H_
// Functor definitions for ScatterND ops, must be compilable by nvcc.
#define EIGEN_USE_THREADS
#include <atomic>
#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/tensor.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/kernels/bounds_check.h"
#include "tensorflow/core/kernels/fill_functor.h"
#include "tensorflow/core/kernels/scatter_nd_op.h"
#include "tensorflow/core/platform/mutex.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/util/util.h"
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
#ifdef TENSORFLOW_USE_SYCL
typedef Eigen::SyclDevice SYCLDevice;
#endif // TENSORFLOW_USE_SYCL
class OpKernelContext;
// Specialization of UpdateExecutor to CPU
namespace update_executor {
template <typename Input, typename Update, typename Output,
scatter_nd_op::UpdateOp OP>
class UpdateExecutor {
public:
EIGEN_STRONG_INLINE static void Execute(Input value, Update update,
Output output);
};
template <typename Input, typename Update, typename Output>
class UpdateExecutor<Input, Update, Output, scatter_nd_op::UpdateOp::ASSIGN> {
public:
EIGEN_STRONG_INLINE static void Execute(Input /* input */, Update update,
Output output) {
output = update;
}
};
template <typename Input, typename Update, typename Output>
class UpdateExecutor<Input, Update, Output, scatter_nd_op::UpdateOp::ADD> {
public:
EIGEN_STRONG_INLINE static void Execute(Input /* input */, Update update,
Output output) {
output += update;
}
};
template <typename Input, typename Update, typename Output>
class UpdateExecutor<Input, Update, Output, scatter_nd_op::UpdateOp::SUB> {
public:
EIGEN_STRONG_INLINE static void Execute(Input /* input */, Update update,
Output output) {
output -= update;
}
};
} // namespace update_executor
namespace functor {
// Implementation of update functor for CPU.
template <typename T, typename Index, scatter_nd_op::UpdateOp OP, int IXDIM>
struct ScatterNdFunctor<CPUDevice, T, Index, OP, IXDIM> {
Index operator()(
const CPUDevice& d, const Index slice_size,
const Eigen::array<Eigen::DenseIndex, IXDIM> output_shape_prefix,
typename TTypes<T, 2>::Tensor Tparams,
typename TTypes<Index, 2>::ConstTensor Tindices,
typename TTypes<T, 2>::ConstTensor Tupdates,
typename TTypes<T, 2>::Tensor Toutput) {
// error_loc is -1 if there's no out-of-bounds index,
// otherwise it is the location of an OOB index in Tindices.
Index error_loc = -1;
const Eigen::DenseIndex batch_size = Tindices.dimension(0);
Index batch_strides[IXDIM];
for (int dim = IXDIM - 1; dim >= 0; --dim) {
if (dim == IXDIM - 1) {
batch_strides[dim] = 1;
} else {
batch_strides[dim] =
batch_strides[dim + 1] * output_shape_prefix[dim + 1];
}
}
for (Eigen::DenseIndex loc = 0; loc < batch_size; ++loc) {
Index i = 0;
bool out_of_bounds = false;
for (int dim = 0; dim < IXDIM; ++dim) {
const Index ix_d = internal::SubtleMustCopy(Tindices(loc, dim));
out_of_bounds |= !FastBoundsCheck(ix_d, output_shape_prefix[dim]);
i += ix_d * batch_strides[dim];
}
if (TF_PREDICT_FALSE(out_of_bounds)) {
error_loc = loc;
break;
} else {
auto input_chip = Toutput.template chip<0>(i);
auto output_chip = input_chip.device(d);
auto update_chip = Tupdates.template chip<0>(loc);
update_executor::UpdateExecutor<
decltype(input_chip), decltype(update_chip), decltype(output_chip),
OP>::Execute(input_chip, update_chip, output_chip);
}
}
return error_loc;
}
};
#define REGISTER_SCATTER_ND_FULL(T, Index, op) \
template Index \
ScatterNdFunctor<CPUDevice, T, Index, op, CPU_PROVIDED_IXDIM>::operator()( \
const CPUDevice& d, const Index slice_size, \
const Eigen::array<Eigen::DenseIndex, CPU_PROVIDED_IXDIM> \
output_shape_prefix, \
typename TTypes<T, 2>::Tensor Tparams, \
typename TTypes<Index, 2>::ConstTensor Tindices, \
typename TTypes<T, 2>::ConstTensor Tupdates, \
typename TTypes<T, 2>::Tensor Toutput)
#define REGISTER_SCATTER_ND_INDEX(type, op) \
REGISTER_SCATTER_ND_FULL(type, int32, op); \
REGISTER_SCATTER_ND_FULL(type, int64, op)
#define REGISTER_SCATTER_ND_UPDATE(type) \
REGISTER_SCATTER_ND_INDEX(type, scatter_nd_op::UpdateOp::ASSIGN);
#define REGISTER_SCATTER_ND_MATH(type) \
REGISTER_SCATTER_ND_INDEX(type, scatter_nd_op::UpdateOp::ADD); \
REGISTER_SCATTER_ND_INDEX(type, scatter_nd_op::UpdateOp::SUB);
TF_CALL_ALL_TYPES(REGISTER_SCATTER_ND_UPDATE);
REGISTER_SCATTER_ND_INDEX(string, scatter_nd_op::UpdateOp::ADD);
TF_CALL_NUMBER_TYPES(REGISTER_SCATTER_ND_MATH);
TF_CALL_bool(REGISTER_SCATTER_ND_MATH);
#undef REGISTER_SCATTER_ND_MATH
#undef REGISTER_SCATTER_ND_UPDATE
#undef REGISTER_SCATTER_ND_INDEX
#undef REGISTER_SCATTER_ND_FULL
// Implementation of update functor for SYCL.
#ifdef TENSORFLOW_USE_SYCL
template <typename T, typename Index, scatter_nd_op::UpdateOp OP, int IXDIM>
struct ScatterNdFunctor<SYCLDevice, T, Index, OP, IXDIM> {
Index operator()(
const SYCLDevice& d, const Index slice_size,
const Eigen::array<Eigen::DenseIndex, IXDIM> output_shape_prefix,
typename TTypes<T, 2>::Tensor Tparams,
typename TTypes<Index, 2>::ConstTensor Tindices,
typename TTypes<T, 2>::ConstTensor Tupdates,
typename TTypes<T, 2>::Tensor Toutput) {
// error_loc is -1 if there's no out-of-bounds index,
// otherwise it is the location of an OOB index in Tindices.
Index error_loc = -1;
const Eigen::DenseIndex batch_size = Tindices.dimension(0);
Index batch_strides[IXDIM];
for (int dim = IXDIM - 1; dim >= 0; --dim) {
if (dim == IXDIM - 1) {
batch_strides[dim] = 1;
} else {
batch_strides[dim] =
batch_strides[dim + 1] * output_shape_prefix[dim + 1];
}
}
for (Eigen::DenseIndex loc = 0; loc < batch_size; ++loc) {
Index i = 0;
bool out_of_bounds = false;
for (int dim = 0; dim < IXDIM; ++dim) {
const Index ix_d = internal::SubtleMustCopy(Tindices(loc, dim));
out_of_bounds |= !FastBoundsCheck(ix_d, output_shape_prefix[dim]);
i += ix_d * batch_strides[dim];
}
if (TF_PREDICT_FALSE(out_of_bounds)) {
error_loc = loc;
break;
} else {
auto input_chip = Toutput.template chip<0>(i);
auto output_chip = input_chip.device(d);
auto update_chip = Tupdates.template chip<0>(loc);
update_executor::UpdateExecutor<
decltype(input_chip), decltype(update_chip), decltype(output_chip),
OP>::Execute(input_chip, update_chip, output_chip);
}
}
return error_loc;
}
};
#define REGISTER_SCATTER_ND_FULL_SYCL(T, Index, op) \
template Index \
ScatterNdFunctor<SYCLDevice, T, Index, op, CPU_PROVIDED_IXDIM>::operator()( \
const SYCLDevice& d, const Index slice_size, \
const Eigen::array<Eigen::DenseIndex, CPU_PROVIDED_IXDIM> \
output_shape_prefix, \
typename TTypes<T, 2>::Tensor Tparams, \
typename TTypes<Index, 2>::ConstTensor Tindices, \
typename TTypes<T, 2>::ConstTensor Tupdates, \
typename TTypes<T, 2>::Tensor Toutput)
#define REGISTER_SCATTER_ND_INDEX_SYCL(type, op) \
REGISTER_SCATTER_ND_FULL_SYCL(type, int32, op); \
REGISTER_SCATTER_ND_FULL_SYCL(type, int64, op)
#define REGISTER_SCATTER_ND_UPDATE_SYCL(type) \
REGISTER_SCATTER_ND_INDEX_SYCL(type, scatter_nd_op::UpdateOp::ASSIGN);
#define REGISTER_SCATTER_ND_MATH_SYCL(type) \
REGISTER_SCATTER_ND_INDEX_SYCL(type, scatter_nd_op::UpdateOp::ADD); \
REGISTER_SCATTER_ND_INDEX_SYCL(type, scatter_nd_op::UpdateOp::SUB);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_ND_UPDATE_SYCL)
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_ND_MATH_SYCL)
REGISTER_SCATTER_ND_UPDATE_SYCL(int32);
REGISTER_SCATTER_ND_MATH_SYCL(int32);
#undef REGISTER_SCATTER_ND_MATH_SYCL
#undef REGISTER_SCATTER_ND_UPDATE_SYCL
#undef REGISTER_SCATTER_ND_INDEX_SYCL
#undef REGISTER_SCATTER_ND_FULL_SYCL
#endif // TENSORFLOW_USE_SYCL
} // namespace functor
} // namespace tensorflow
#endif // TENSORFLOW_CORE_KERNELS_SCATTER_ND_OP_CPU_IMPL_H_
|
