Eigen-unsupported  3.4.90 (git rev a4098ac676528a83cfb73d4d26ce1b42ec05f47c)
TensorAssign.h
1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
5//
6// This Source Code Form is subject to the terms of the Mozilla
7// Public License v. 2.0. If a copy of the MPL was not distributed
8// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9
10#ifndef EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H
11#define EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H
12
13#include "./InternalHeaderCheck.h"
14
15namespace Eigen {
16
25namespace internal {
26template<typename LhsXprType, typename RhsXprType>
27struct traits<TensorAssignOp<LhsXprType, RhsXprType> >
28{
29 typedef typename LhsXprType::Scalar Scalar;
30 typedef typename traits<LhsXprType>::StorageKind StorageKind;
31 typedef typename promote_index_type<typename traits<LhsXprType>::Index,
32 typename traits<RhsXprType>::Index>::type Index;
33 typedef typename LhsXprType::Nested LhsNested;
34 typedef typename RhsXprType::Nested RhsNested;
35 typedef typename remove_reference<LhsNested>::type _LhsNested;
36 typedef typename remove_reference<RhsNested>::type _RhsNested;
37 static const std::size_t NumDimensions = internal::traits<LhsXprType>::NumDimensions;
38 static const int Layout = internal::traits<LhsXprType>::Layout;
39 typedef typename traits<LhsXprType>::PointerType PointerType;
40
41 enum {
42 Flags = 0
43 };
44};
45
46template<typename LhsXprType, typename RhsXprType>
47struct eval<TensorAssignOp<LhsXprType, RhsXprType>, Eigen::Dense>
48{
49 typedef const TensorAssignOp<LhsXprType, RhsXprType>& type;
50};
51
52template<typename LhsXprType, typename RhsXprType>
53struct nested<TensorAssignOp<LhsXprType, RhsXprType>, 1, typename eval<TensorAssignOp<LhsXprType, RhsXprType> >::type>
54{
55 typedef TensorAssignOp<LhsXprType, RhsXprType> type;
56};
57
58} // end namespace internal
59
60
61
62template<typename LhsXprType, typename RhsXprType>
63class TensorAssignOp : public TensorBase<TensorAssignOp<LhsXprType, RhsXprType> >
64{
65 public:
66 typedef typename Eigen::internal::traits<TensorAssignOp>::Scalar Scalar;
67 typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
68 typedef typename LhsXprType::CoeffReturnType CoeffReturnType;
69 typedef typename Eigen::internal::nested<TensorAssignOp>::type Nested;
70 typedef typename Eigen::internal::traits<TensorAssignOp>::StorageKind StorageKind;
71 typedef typename Eigen::internal::traits<TensorAssignOp>::Index Index;
72
73 static const int NumDims = Eigen::internal::traits<TensorAssignOp>::NumDimensions;
74
75 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorAssignOp(LhsXprType& lhs, const RhsXprType& rhs)
76 : m_lhs_xpr(lhs), m_rhs_xpr(rhs) {}
77
79 EIGEN_DEVICE_FUNC
80 typename internal::remove_all<typename LhsXprType::Nested>::type&
81 lhsExpression() const { return *((typename internal::remove_all<typename LhsXprType::Nested>::type*)&m_lhs_xpr); }
82
83 EIGEN_DEVICE_FUNC
84 const typename internal::remove_all<typename RhsXprType::Nested>::type&
85 rhsExpression() const { return m_rhs_xpr; }
86
87 protected:
88 typename internal::remove_all<typename LhsXprType::Nested>::type& m_lhs_xpr;
89 const typename internal::remove_all<typename RhsXprType::Nested>::type& m_rhs_xpr;
90};
91
92
93template<typename LeftArgType, typename RightArgType, typename Device>
94struct TensorEvaluator<const TensorAssignOp<LeftArgType, RightArgType>, Device>
95{
96 typedef TensorAssignOp<LeftArgType, RightArgType> XprType;
97 typedef typename XprType::Index Index;
98 typedef typename XprType::Scalar Scalar;
99 typedef typename XprType::CoeffReturnType CoeffReturnType;
100 typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
101 typedef typename TensorEvaluator<RightArgType, Device>::Dimensions Dimensions;
102 typedef StorageMemory<CoeffReturnType, Device> Storage;
103 typedef typename Storage::Type EvaluatorPointerType;
104
105 static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
106 static const int NumDims = XprType::NumDims;
107
108 enum {
109 IsAligned = int(TensorEvaluator<LeftArgType, Device>::IsAligned) &
110 int(TensorEvaluator<RightArgType, Device>::IsAligned),
111 PacketAccess = int(TensorEvaluator<LeftArgType, Device>::PacketAccess) &
112 int(TensorEvaluator<RightArgType, Device>::PacketAccess),
113 BlockAccess = int(TensorEvaluator<LeftArgType, Device>::BlockAccess) &
114 int(TensorEvaluator<RightArgType, Device>::BlockAccess),
115 PreferBlockAccess = int(TensorEvaluator<LeftArgType, Device>::PreferBlockAccess) |
116 int(TensorEvaluator<RightArgType, Device>::PreferBlockAccess),
117 Layout = TensorEvaluator<LeftArgType, Device>::Layout,
118 RawAccess = TensorEvaluator<LeftArgType, Device>::RawAccess
119 };
120
121 //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
122 typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
123 typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
124
125 typedef typename TensorEvaluator<const RightArgType, Device>::TensorBlock
126 RightTensorBlock;
127 //===--------------------------------------------------------------------===//
128
129 TensorEvaluator(const XprType& op, const Device& device) :
130 m_leftImpl(op.lhsExpression(), device),
131 m_rightImpl(op.rhsExpression(), device)
132 {
133 EIGEN_STATIC_ASSERT(
134 (static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) ==
135 static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout)),
136 YOU_MADE_A_PROGRAMMING_MISTAKE);
137 }
138
139 EIGEN_DEVICE_FUNC const Dimensions& dimensions() const
140 {
141 // The dimensions of the lhs and the rhs tensors should be equal to prevent
142 // overflows and ensure the result is fully initialized.
143 // TODO: use left impl instead if right impl dimensions are known at compile time.
144 return m_rightImpl.dimensions();
145 }
146
147 EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
148 eigen_assert(dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions()));
149 m_leftImpl.evalSubExprsIfNeeded(NULL);
150 // If the lhs provides raw access to its storage area (i.e. if m_leftImpl.data() returns a non
151 // null value), attempt to evaluate the rhs expression in place. Returns true iff in place
152 // evaluation isn't supported and the caller still needs to manually assign the values generated
153 // by the rhs to the lhs.
154 return m_rightImpl.evalSubExprsIfNeeded(m_leftImpl.data());
155 }
156
157#ifdef EIGEN_USE_THREADS
158 template <typename EvalSubExprsCallback>
159 EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(
160 EvaluatorPointerType, EvalSubExprsCallback done) {
161 m_leftImpl.evalSubExprsIfNeededAsync(nullptr, [this, done](bool) {
162 m_rightImpl.evalSubExprsIfNeededAsync(
163 m_leftImpl.data(), [done](bool need_assign) { done(need_assign); });
164 });
165 }
166#endif // EIGEN_USE_THREADS
167
168 EIGEN_STRONG_INLINE void cleanup() {
169 m_leftImpl.cleanup();
170 m_rightImpl.cleanup();
171 }
172
173 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalScalar(Index i) {
174 m_leftImpl.coeffRef(i) = m_rightImpl.coeff(i);
175 }
176 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalPacket(Index i) {
177
178 const int LhsStoreMode = TensorEvaluator<LeftArgType, Device>::IsAligned ? Aligned : Unaligned;
179 const int RhsLoadMode = TensorEvaluator<RightArgType, Device>::IsAligned ? Aligned : Unaligned;
180 m_leftImpl.template writePacket<LhsStoreMode>(i, m_rightImpl.template packet<RhsLoadMode>(i));
181 }
182 EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const
183 {
184 return m_leftImpl.coeff(index);
185 }
186 template<int LoadMode>
187 EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const
188 {
189 return m_leftImpl.template packet<LoadMode>(index);
190 }
191
192 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
193 costPerCoeff(bool vectorized) const {
194 // We assume that evalPacket or evalScalar is called to perform the
195 // assignment and account for the cost of the write here, but reduce left
196 // cost by one load because we are using m_leftImpl.coeffRef.
197 TensorOpCost left = m_leftImpl.costPerCoeff(vectorized);
198 return m_rightImpl.costPerCoeff(vectorized) +
199 TensorOpCost(
200 numext::maxi(0.0, left.bytes_loaded() - sizeof(CoeffReturnType)),
201 left.bytes_stored(), left.compute_cycles()) +
202 TensorOpCost(0, sizeof(CoeffReturnType), 0, vectorized, PacketSize);
203 }
204
205 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
206 internal::TensorBlockResourceRequirements getResourceRequirements() const {
207 return internal::TensorBlockResourceRequirements::merge(
208 m_leftImpl.getResourceRequirements(),
209 m_rightImpl.getResourceRequirements());
210 }
211
212 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalBlock(
213 TensorBlockDesc& desc, TensorBlockScratch& scratch) {
214 if (TensorEvaluator<LeftArgType, Device>::RawAccess &&
215 m_leftImpl.data() != NULL) {
216 // If destination has raw data access, we pass it as a potential
217 // destination for a block descriptor evaluation.
218 desc.template AddDestinationBuffer<Layout>(
219 /*dst_base=*/m_leftImpl.data() + desc.offset(),
220 /*dst_strides=*/internal::strides<Layout>(m_leftImpl.dimensions()));
221 }
222
223 RightTensorBlock block = m_rightImpl.block(desc, scratch, /*root_of_expr_ast=*/true);
224 // If block was evaluated into a destination, there is no need to do assignment.
225 if (block.kind() != internal::TensorBlockKind::kMaterializedInOutput) {
226 m_leftImpl.writeBlock(desc, block);
227 }
228 block.cleanup();
229 }
230
231#ifdef EIGEN_USE_SYCL
232 // binding placeholder accessors to a command group handler for SYCL
233 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
234 m_leftImpl.bind(cgh);
235 m_rightImpl.bind(cgh);
236 }
237#endif
238
239 EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_leftImpl.data(); }
240
241 private:
242 TensorEvaluator<LeftArgType, Device> m_leftImpl;
243 TensorEvaluator<RightArgType, Device> m_rightImpl;
244};
245
246}
247
248
249#endif // EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H
Namespace containing all symbols from the Eigen library.
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index