Newer
Older
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
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
#define EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H
namespace Eigen {
/** \class TensorConvolution
* \ingroup CXX11_Tensor_Module
*
* \brief Tensor convolution class.
*
*
*/
namespace internal {
template <typename Index, typename InputDims, int NumKernelDims, int Layout>
class IndexMapper {
public:
IndexMapper(const InputDims& input_dims, const array<Index, NumKernelDims>& kernel_dims,
const array<Index, NumKernelDims>& indices) {
array<Index, NumDims> dimensions = input_dims;
for (int i = 0; i < NumKernelDims; ++i) {
const Index index = indices[i];
const Index input_dim = input_dims[index];
const Index kernel_dim = kernel_dims[i];
const Index result_dim = input_dim - kernel_dim + 1;
dimensions[index] = result_dim;
}
array<Index, NumDims> inputStrides;
array<Index, NumDims> outputStrides;
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
inputStrides[0] = 1;
outputStrides[0] = 1;
for (int i = 1; i < NumDims; ++i) {
inputStrides[i] = inputStrides[i-1] * input_dims[i-1];
outputStrides[i] = outputStrides[i-1] * dimensions[i-1];
}
} else {
inputStrides[NumDims - 1] = 1;
outputStrides[NumDims - 1] = 1;
for (int i = static_cast<int>(NumDims) - 2; i >= 0; --i) {
inputStrides[i] = inputStrides[i + 1] * input_dims[i + 1];
outputStrides[i] = outputStrides[i + 1] * dimensions[i + 1];
}
}
array<Index, NumDims> cudaInputDimensions;
array<Index, NumDims> cudaOutputDimensions;
array<Index, NumDims> tmp = dimensions;
array<Index, NumDims> ordering;
const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
? 0
: NumDims - NumKernelDims;
for (int i = 0; i < NumKernelDims; ++i) {
const Index index = i + offset;
ordering[index] = indices[i];
tmp[indices[i]] = -1;
cudaInputDimensions[index] = input_dims[indices[i]];
cudaOutputDimensions[index] = dimensions[indices[i]];
}
int written = static_cast<int>(Layout) == static_cast<int>(ColMajor)
? NumKernelDims
: 0;
for (int i = 0; i < NumDims; ++i) {
if (tmp[i] >= 0) {
ordering[written] = i;
cudaInputDimensions[written] = input_dims[i];
cudaOutputDimensions[written] = dimensions[i];
++written;
}
}
for (int i = 0; i < NumDims; ++i) {
m_inputStrides[i] = inputStrides[ordering[i]];
m_outputStrides[i] = outputStrides[ordering[i]];
}
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
for (int i = 0; i < NumDims; ++i) {
if (i > NumKernelDims) {
m_cudaInputStrides[i] =
m_cudaInputStrides[i - 1] * cudaInputDimensions[i - 1];
m_cudaOutputStrides[i] =
m_cudaOutputStrides[i - 1] * cudaOutputDimensions[i - 1];
} else {
m_cudaInputStrides[i] = 1;
m_cudaOutputStrides[i] = 1;
}
}
} else {
for (int i = NumDims - 1; i >= 0; --i) {
if (i + 1 < offset) {
m_cudaInputStrides[i] =
m_cudaInputStrides[i + 1] * cudaInputDimensions[i + 1];
m_cudaOutputStrides[i] =
m_cudaOutputStrides[i + 1] * cudaOutputDimensions[i + 1];
} else {
m_cudaInputStrides[i] = 1;
m_cudaOutputStrides[i] = 1;
}
}
}
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputPlaneToTensorInputOffset(Index p) const {
Index inputIndex = 0;
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
for (int d = NumDims - 1; d > NumKernelDims; --d) {
const Index idx = p / m_cudaInputStrides[d];
inputIndex += idx * m_inputStrides[d];
p -= idx * m_cudaInputStrides[d];
}
inputIndex += p * m_inputStrides[NumKernelDims];
} else {
std::ptrdiff_t limit = 0;
if (NumKernelDims < NumDims) {
limit = NumDims - NumKernelDims - 1;
}
for (int d = 0; d < limit; ++d) {
const Index idx = p / m_cudaInputStrides[d];
inputIndex += idx * m_inputStrides[d];
p -= idx * m_cudaInputStrides[d];
}
inputIndex += p * m_inputStrides[limit];
}
return inputIndex;
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputPlaneToTensorOutputOffset(Index p) const {
Index outputIndex = 0;
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
for (int d = NumDims - 1; d > NumKernelDims; --d) {
const Index idx = p / m_cudaOutputStrides[d];
outputIndex += idx * m_outputStrides[d];
p -= idx * m_cudaOutputStrides[d];
}
outputIndex += p * m_outputStrides[NumKernelDims];
} else {
std::ptrdiff_t limit = 0;
if (NumKernelDims < NumDims) {
limit = NumDims - NumKernelDims - 1;
}
for (int d = 0; d < limit; ++d) {
const Index idx = p / m_cudaOutputStrides[d];
outputIndex += idx * m_outputStrides[d];
p -= idx * m_cudaOutputStrides[d];
}
outputIndex += p * m_outputStrides[limit];
}
return outputIndex;
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i) const {
const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
? 0
: NumDims - NumKernelDims;
return i * m_inputStrides[offset];
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i) const {
const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
? 0
: NumDims - NumKernelDims;
return i * m_outputStrides[offset];
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i, Index j) const {
const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
? 0
: NumDims - NumKernelDims;
return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1];
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i, Index j) const {
const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
? 0
: NumDims - NumKernelDims;
return i * m_outputStrides[offset] + j * m_outputStrides[offset + 1];
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i, Index j, Index k) const {
const size_t offset = static_cast<int>(Layout) == static_cast<int>(ColMajor)
? 0
: NumDims - NumKernelDims;
return i * m_inputStrides[offset] + j * m_inputStrides[offset + 1] +
k * m_inputStrides[offset + 2];
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i, Index j, Index k) const {
Loading full blame...