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
|
/* 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_SCATTER_FUNCTOR_H_
#define TENSORFLOW_KERNELS_SCATTER_FUNCTOR_H_
#include <type_traits>
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/kernels/bounds_check.h"
#include "tensorflow/core/platform/types.h"
namespace tensorflow {
class OpKernelContext;
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
#ifdef TENSORFLOW_USE_SYCL
typedef Eigen::SyclDevice SYCLDevice;
#endif // TENSORFLOW_USE_SYCL
namespace scatter_op {
enum class UpdateOp { ASSIGN, ADD, SUB, MUL, DIV };
namespace internal {
template <scatter_op::UpdateOp Op>
struct Assign {};
template <>
struct Assign<scatter_op::UpdateOp::ASSIGN> {
template <typename Params, typename Update>
static void Run(Params p, Update u) {
p = u;
}
};
template <>
struct Assign<scatter_op::UpdateOp::ADD> {
template <typename Params, typename Update>
static void Run(Params p, Update u) {
p += u;
}
};
template <>
struct Assign<scatter_op::UpdateOp::SUB> {
template <typename Params, typename Update>
static void Run(Params p, Update u) {
p -= u;
}
};
template <>
struct Assign<scatter_op::UpdateOp::MUL> {
template <typename Params, typename Update>
static void Run(Params p, Update u) {
p *= u;
}
};
template <>
struct Assign<scatter_op::UpdateOp::DIV> {
template <typename Params, typename Update>
static void Run(Params p, Update u) {
p /= u;
}
};
} // namespace internal
} // namespace scatter_op
namespace functor {
template <typename Device, typename T, typename Index, scatter_op::UpdateOp op>
struct ScatterFunctor {
Index operator()(OpKernelContext* c, const Device& d,
typename TTypes<T>::Matrix params,
typename TTypes<T>::ConstMatrix updates,
typename TTypes<Index>::ConstFlat indices);
};
template <typename Device, typename T, typename Index, scatter_op::UpdateOp op>
struct ScatterFunctorBase {
Index operator()(OpKernelContext* c, const Device& d,
typename TTypes<T>::Matrix params,
typename TTypes<T>::ConstMatrix updates,
typename TTypes<Index>::ConstFlat indices) {
// indices and params sizes were validated in DoCompute().
const Index N = static_cast<Index>(indices.size());
const Index limit = static_cast<Index>(params.dimension(0));
for (Index i = 0; i < N; i++) {
// Grab the index and check its validity. An earlier version of the
// code checked it and then grabbed it from memory a second time, which
// was a security risk since it could have changed in between.
const Index index = ::tensorflow::internal::SubtleMustCopy(indices(i));
if (!FastBoundsCheck(index, limit)) return i;
// Copy last Ndim-1 dimensions of updates[i] to params[index]
scatter_op::internal::Assign<op>::Run(params.template chip<0>(index),
updates.template chip<0>(i));
}
return -1;
}
};
template <typename T, typename Index>
struct ScatterFunctorBase<CPUDevice, T, Index, scatter_op::UpdateOp::ASSIGN> {
Index operator()(OpKernelContext* c, const CPUDevice& d,
typename TTypes<T>::Matrix params,
typename TTypes<T>::ConstMatrix updates,
typename TTypes<Index>::ConstFlat indices) {
// indices and params sizes were validated in DoCompute().
const Index N = static_cast<Index>(indices.size());
const Index limit = static_cast<Index>(params.dimension(0));
if (!std::is_same<T, string>::value) {
for (Index i = 0; i < N; i++) {
// Grab the index and check its validity. An earlier version of the
// code checked it and then grabbed it from memory a second time, which
// was a security risk since it could have changed in between.
const Index index = ::tensorflow::internal::SubtleMustCopy(indices(i));
if (!FastBoundsCheck(index, limit)) return i;
memmove(params.data() + index * params.dimension(1),
updates.data() + i * updates.dimension(1),
updates.dimension(1) * sizeof(T));
}
} else {
for (Index i = 0; i < N; i++) {
// Grab the index and check its validity. An earlier version of the
// code checked it and then grabbed it from memory a second time, which
// was a security risk since it could have changed in between.
const Index index = ::tensorflow::internal::SubtleMustCopy(indices(i));
if (!FastBoundsCheck(index, limit)) return i;
// Copy last Ndim-1 dimensions of updates[i] to params[index]
scatter_op::internal::Assign<scatter_op::UpdateOp::ASSIGN>::Run(
params.template chip<0>(index), updates.template chip<0>(i));
}
}
return -1;
}
};
template <typename T, typename Index, scatter_op::UpdateOp op>
struct ScatterFunctor<CPUDevice, T, Index, op>
: ScatterFunctorBase<CPUDevice, T, Index, op>{};
#if TENSORFLOW_USE_SYCL
template<typename T, typename Index, scatter_op::UpdateOp op>
struct ScatterFunctor<SYCLDevice, T, Index, op>
: ScatterFunctorBase<SYCLDevice, T, Index, op>{};
#endif // TENSORFLOW_USE_SYCL
} // namespace functor
} // namespace tensorflow
#endif // TENSORFLOW_KERNELS_SCATTER_FUNCTOR_H_
|
