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
|
// 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_PATCH_H
#define EIGEN_CXX11_TENSOR_TENSOR_PATCH_H
namespace Eigen {
/** \class TensorPatch
* \ingroup CXX11_Tensor_Module
*
* \brief Tensor patch class.
*
*
*/
namespace internal {
template<typename PatchDim, typename XprType>
struct traits<TensorPatchOp<PatchDim, XprType> > : public traits<XprType>
{
typedef typename XprType::Scalar Scalar;
typedef traits<XprType> XprTraits;
typedef typename XprTraits::StorageKind StorageKind;
typedef typename XprTraits::Index Index;
typedef typename XprType::Nested Nested;
typedef typename remove_reference<Nested>::type _Nested;
static const int NumDimensions = XprTraits::NumDimensions + 1;
static const int Layout = XprTraits::Layout;
typedef typename XprTraits::PointerType PointerType;
};
template<typename PatchDim, typename XprType>
struct eval<TensorPatchOp<PatchDim, XprType>, Eigen::Dense>
{
typedef const TensorPatchOp<PatchDim, XprType>& type;
};
template<typename PatchDim, typename XprType>
struct nested<TensorPatchOp<PatchDim, XprType>, 1, typename eval<TensorPatchOp<PatchDim, XprType> >::type>
{
typedef TensorPatchOp<PatchDim, XprType> type;
};
} // end namespace internal
template<typename PatchDim, typename XprType>
class TensorPatchOp : public TensorBase<TensorPatchOp<PatchDim, XprType>, ReadOnlyAccessors>
{
public:
typedef typename Eigen::internal::traits<TensorPatchOp>::Scalar Scalar;
typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename Eigen::internal::nested<TensorPatchOp>::type Nested;
typedef typename Eigen::internal::traits<TensorPatchOp>::StorageKind StorageKind;
typedef typename Eigen::internal::traits<TensorPatchOp>::Index Index;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorPatchOp(const XprType& expr, const PatchDim& patch_dims)
: m_xpr(expr), m_patch_dims(patch_dims) {}
EIGEN_DEVICE_FUNC
const PatchDim& patch_dims() const { return m_patch_dims; }
EIGEN_DEVICE_FUNC
const typename internal::remove_all<typename XprType::Nested>::type&
expression() const { return m_xpr; }
protected:
typename XprType::Nested m_xpr;
const PatchDim m_patch_dims;
};
// Eval as rvalue
template<typename PatchDim, typename ArgType, typename Device>
struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
{
typedef TensorPatchOp<PatchDim, ArgType> XprType;
typedef typename XprType::Index Index;
static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value + 1;
typedef DSizes<Index, NumDims> Dimensions;
typedef typename XprType::Scalar Scalar;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
typedef StorageMemory<CoeffReturnType, Device> Storage;
typedef typename Storage::Type EvaluatorPointerType;
enum {
IsAligned = false,
PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
BlockAccess = false,
PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
Layout = TensorEvaluator<ArgType, Device>::Layout,
CoordAccess = false,
RawAccess = false
};
//===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
typedef internal::TensorBlockNotImplemented TensorBlock;
//===--------------------------------------------------------------------===//
EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
: m_impl(op.expression(), device)
{
Index num_patches = 1;
const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
const PatchDim& patch_dims = op.patch_dims();
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
for (int i = 0; i < NumDims-1; ++i) {
m_dimensions[i] = patch_dims[i];
num_patches *= (input_dims[i] - patch_dims[i] + 1);
}
m_dimensions[NumDims-1] = num_patches;
m_inputStrides[0] = 1;
m_patchStrides[0] = 1;
for (int i = 1; i < NumDims-1; ++i) {
m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
m_patchStrides[i] = m_patchStrides[i-1] * (input_dims[i-1] - patch_dims[i-1] + 1);
}
m_outputStrides[0] = 1;
for (int i = 1; i < NumDims; ++i) {
m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
}
} else {
for (int i = 0; i < NumDims-1; ++i) {
m_dimensions[i+1] = patch_dims[i];
num_patches *= (input_dims[i] - patch_dims[i] + 1);
}
m_dimensions[0] = num_patches;
m_inputStrides[NumDims-2] = 1;
m_patchStrides[NumDims-2] = 1;
for (int i = NumDims-3; i >= 0; --i) {
m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
m_patchStrides[i] = m_patchStrides[i+1] * (input_dims[i+1] - patch_dims[i+1] + 1);
}
m_outputStrides[NumDims-1] = 1;
for (int i = NumDims-2; i >= 0; --i) {
m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
}
}
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
m_impl.evalSubExprsIfNeeded(NULL);
return true;
}
EIGEN_STRONG_INLINE void cleanup() {
m_impl.cleanup();
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
{
Index output_stride_index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? NumDims - 1 : 0;
// Find the location of the first element of the patch.
Index patchIndex = index / m_outputStrides[output_stride_index];
// Find the offset of the element wrt the location of the first element.
Index patchOffset = index - patchIndex * m_outputStrides[output_stride_index];
Index inputIndex = 0;
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
EIGEN_UNROLL_LOOP
for (int i = NumDims - 2; i > 0; --i) {
const Index patchIdx = patchIndex / m_patchStrides[i];
patchIndex -= patchIdx * m_patchStrides[i];
const Index offsetIdx = patchOffset / m_outputStrides[i];
patchOffset -= offsetIdx * m_outputStrides[i];
inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i];
}
} else {
EIGEN_UNROLL_LOOP
for (int i = 0; i < NumDims - 2; ++i) {
const Index patchIdx = patchIndex / m_patchStrides[i];
patchIndex -= patchIdx * m_patchStrides[i];
const Index offsetIdx = patchOffset / m_outputStrides[i+1];
patchOffset -= offsetIdx * m_outputStrides[i+1];
inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i];
}
}
inputIndex += (patchIndex + patchOffset);
return m_impl.coeff(inputIndex);
}
template<int LoadMode>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
{
EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
Index output_stride_index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? NumDims - 1 : 0;
Index indices[2] = {index, index + PacketSize - 1};
Index patchIndices[2] = {indices[0] / m_outputStrides[output_stride_index],
indices[1] / m_outputStrides[output_stride_index]};
Index patchOffsets[2] = {indices[0] - patchIndices[0] * m_outputStrides[output_stride_index],
indices[1] - patchIndices[1] * m_outputStrides[output_stride_index]};
Index inputIndices[2] = {0, 0};
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
EIGEN_UNROLL_LOOP
for (int i = NumDims - 2; i > 0; --i) {
const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
patchIndices[1] / m_patchStrides[i]};
patchIndices[0] -= patchIdx[0] * m_patchStrides[i];
patchIndices[1] -= patchIdx[1] * m_patchStrides[i];
const Index offsetIdx[2] = {patchOffsets[0] / m_outputStrides[i],
patchOffsets[1] / m_outputStrides[i]};
patchOffsets[0] -= offsetIdx[0] * m_outputStrides[i];
patchOffsets[1] -= offsetIdx[1] * m_outputStrides[i];
inputIndices[0] += (patchIdx[0] + offsetIdx[0]) * m_inputStrides[i];
inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i];
}
} else {
EIGEN_UNROLL_LOOP
for (int i = 0; i < NumDims - 2; ++i) {
const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
patchIndices[1] / m_patchStrides[i]};
patchIndices[0] -= patchIdx[0] * m_patchStrides[i];
patchIndices[1] -= patchIdx[1] * m_patchStrides[i];
const Index offsetIdx[2] = {patchOffsets[0] / m_outputStrides[i+1],
patchOffsets[1] / m_outputStrides[i+1]};
patchOffsets[0] -= offsetIdx[0] * m_outputStrides[i+1];
patchOffsets[1] -= offsetIdx[1] * m_outputStrides[i+1];
inputIndices[0] += (patchIdx[0] + offsetIdx[0]) * m_inputStrides[i];
inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i];
}
}
inputIndices[0] += (patchIndices[0] + patchOffsets[0]);
inputIndices[1] += (patchIndices[1] + patchOffsets[1]);
if (inputIndices[1] - inputIndices[0] == PacketSize - 1) {
PacketReturnType rslt = m_impl.template packet<Unaligned>(inputIndices[0]);
return rslt;
}
else {
EIGEN_ALIGN_MAX CoeffReturnType values[PacketSize];
values[0] = m_impl.coeff(inputIndices[0]);
values[PacketSize-1] = m_impl.coeff(inputIndices[1]);
EIGEN_UNROLL_LOOP
for (int i = 1; i < PacketSize-1; ++i) {
values[i] = coeff(index+i);
}
PacketReturnType rslt = internal::pload<PacketReturnType>(values);
return rslt;
}
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
const double compute_cost = NumDims * (TensorOpCost::DivCost<Index>() +
TensorOpCost::MulCost<Index>() +
2 * TensorOpCost::AddCost<Index>());
return m_impl.costPerCoeff(vectorized) +
TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
}
EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
#ifdef EIGEN_USE_SYCL
// binding placeholder accessors to a command group handler for SYCL
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
m_impl.bind(cgh);
}
#endif
protected:
Dimensions m_dimensions;
array<Index, NumDims> m_outputStrides;
array<Index, NumDims-1> m_inputStrides;
array<Index, NumDims-1> m_patchStrides;
TensorEvaluator<ArgType, Device> m_impl;
};
} // end namespace Eigen
#endif // EIGEN_CXX11_TENSOR_TENSOR_PATCH_H
|
