1 | // This file is part of Eigen, a lightweight C++ template library |
2 | // for linear algebra. |
3 | // |
4 | // Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com> |
5 | // Copyright (C) 2009-2014 Gael Guennebaud <gael.guennebaud@inria.fr> |
6 | // |
7 | // This Source Code Form is subject to the terms of the Mozilla |
8 | // Public License v. 2.0. If a copy of the MPL was not distributed |
9 | // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. |
10 | |
11 | #ifndef EIGEN_TRANSPOSE_H |
12 | #define EIGEN_TRANSPOSE_H |
13 | |
14 | namespace Eigen { |
15 | |
16 | namespace internal { |
17 | template<typename MatrixType> |
18 | struct traits<Transpose<MatrixType> > : public traits<MatrixType> |
19 | { |
20 | typedef typename ref_selector<MatrixType>::type MatrixTypeNested; |
21 | typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedPlain; |
22 | enum { |
23 | RowsAtCompileTime = MatrixType::ColsAtCompileTime, |
24 | ColsAtCompileTime = MatrixType::RowsAtCompileTime, |
25 | MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime, |
26 | MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime, |
27 | FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0, |
28 | Flags0 = traits<MatrixTypeNestedPlain>::Flags & ~(LvalueBit | NestByRefBit), |
29 | Flags1 = Flags0 | FlagsLvalueBit, |
30 | Flags = Flags1 ^ RowMajorBit, |
31 | InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret, |
32 | OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret |
33 | }; |
34 | }; |
35 | } |
36 | |
37 | template<typename MatrixType, typename StorageKind> class TransposeImpl; |
38 | |
39 | /** \class Transpose |
40 | * \ingroup Core_Module |
41 | * |
42 | * \brief Expression of the transpose of a matrix |
43 | * |
44 | * \tparam MatrixType the type of the object of which we are taking the transpose |
45 | * |
46 | * This class represents an expression of the transpose of a matrix. |
47 | * It is the return type of MatrixBase::transpose() and MatrixBase::adjoint() |
48 | * and most of the time this is the only way it is used. |
49 | * |
50 | * \sa MatrixBase::transpose(), MatrixBase::adjoint() |
51 | */ |
52 | template<typename MatrixType> class Transpose |
53 | : public TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind> |
54 | { |
55 | public: |
56 | |
57 | typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested; |
58 | |
59 | typedef typename TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base; |
60 | EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose) |
61 | typedef typename internal::remove_all<MatrixType>::type NestedExpression; |
62 | |
63 | EIGEN_DEVICE_FUNC |
64 | explicit inline Transpose(MatrixType& matrix) : m_matrix(matrix) {} |
65 | |
66 | EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose) |
67 | |
68 | EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.cols(); } |
69 | EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.rows(); } |
70 | |
71 | /** \returns the nested expression */ |
72 | EIGEN_DEVICE_FUNC |
73 | const typename internal::remove_all<MatrixTypeNested>::type& |
74 | nestedExpression() const { return m_matrix; } |
75 | |
76 | /** \returns the nested expression */ |
77 | EIGEN_DEVICE_FUNC |
78 | typename internal::remove_reference<MatrixTypeNested>::type& |
79 | nestedExpression() { return m_matrix; } |
80 | |
81 | /** \internal */ |
82 | void resize(Index nrows, Index ncols) { |
83 | m_matrix.resize(ncols,nrows); |
84 | } |
85 | |
86 | protected: |
87 | typename internal::ref_selector<MatrixType>::non_const_type m_matrix; |
88 | }; |
89 | |
90 | namespace internal { |
91 | |
92 | template<typename MatrixType, bool HasDirectAccess = has_direct_access<MatrixType>::ret> |
93 | struct TransposeImpl_base |
94 | { |
95 | typedef typename dense_xpr_base<Transpose<MatrixType> >::type type; |
96 | }; |
97 | |
98 | template<typename MatrixType> |
99 | struct TransposeImpl_base<MatrixType, false> |
100 | { |
101 | typedef typename dense_xpr_base<Transpose<MatrixType> >::type type; |
102 | }; |
103 | |
104 | } // end namespace internal |
105 | |
106 | // Generic API dispatcher |
107 | template<typename XprType, typename StorageKind> |
108 | class TransposeImpl |
109 | : public internal::generic_xpr_base<Transpose<XprType> >::type |
110 | { |
111 | public: |
112 | typedef typename internal::generic_xpr_base<Transpose<XprType> >::type Base; |
113 | }; |
114 | |
115 | template<typename MatrixType> class TransposeImpl<MatrixType,Dense> |
116 | : public internal::TransposeImpl_base<MatrixType>::type |
117 | { |
118 | public: |
119 | |
120 | typedef typename internal::TransposeImpl_base<MatrixType>::type Base; |
121 | using Base::coeffRef; |
122 | EIGEN_DENSE_PUBLIC_INTERFACE(Transpose<MatrixType>) |
123 | EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TransposeImpl) |
124 | |
125 | EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride(); } |
126 | EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().nestedExpression().outerStride(); } |
127 | |
128 | typedef typename internal::conditional< |
129 | internal::is_lvalue<MatrixType>::value, |
130 | Scalar, |
131 | const Scalar |
132 | >::type ScalarWithConstIfNotLvalue; |
133 | |
134 | EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } |
135 | EIGEN_DEVICE_FUNC inline const Scalar* data() const { return derived().nestedExpression().data(); } |
136 | |
137 | // FIXME: shall we keep the const version of coeffRef? |
138 | EIGEN_DEVICE_FUNC |
139 | inline const Scalar& coeffRef(Index rowId, Index colId) const |
140 | { |
141 | return derived().nestedExpression().coeffRef(colId, rowId); |
142 | } |
143 | |
144 | EIGEN_DEVICE_FUNC |
145 | inline const Scalar& coeffRef(Index index) const |
146 | { |
147 | return derived().nestedExpression().coeffRef(index); |
148 | } |
149 | }; |
150 | |
151 | /** \returns an expression of the transpose of *this. |
152 | * |
153 | * Example: \include MatrixBase_transpose.cpp |
154 | * Output: \verbinclude MatrixBase_transpose.out |
155 | * |
156 | * \warning If you want to replace a matrix by its own transpose, do \b NOT do this: |
157 | * \code |
158 | * m = m.transpose(); // bug!!! caused by aliasing effect |
159 | * \endcode |
160 | * Instead, use the transposeInPlace() method: |
161 | * \code |
162 | * m.transposeInPlace(); |
163 | * \endcode |
164 | * which gives Eigen good opportunities for optimization, or alternatively you can also do: |
165 | * \code |
166 | * m = m.transpose().eval(); |
167 | * \endcode |
168 | * |
169 | * \sa transposeInPlace(), adjoint() */ |
170 | template<typename Derived> |
171 | inline Transpose<Derived> |
172 | DenseBase<Derived>::transpose() |
173 | { |
174 | return TransposeReturnType(derived()); |
175 | } |
176 | |
177 | /** This is the const version of transpose(). |
178 | * |
179 | * Make sure you read the warning for transpose() ! |
180 | * |
181 | * \sa transposeInPlace(), adjoint() */ |
182 | template<typename Derived> |
183 | inline typename DenseBase<Derived>::ConstTransposeReturnType |
184 | DenseBase<Derived>::transpose() const |
185 | { |
186 | return ConstTransposeReturnType(derived()); |
187 | } |
188 | |
189 | /** \returns an expression of the adjoint (i.e. conjugate transpose) of *this. |
190 | * |
191 | * Example: \include MatrixBase_adjoint.cpp |
192 | * Output: \verbinclude MatrixBase_adjoint.out |
193 | * |
194 | * \warning If you want to replace a matrix by its own adjoint, do \b NOT do this: |
195 | * \code |
196 | * m = m.adjoint(); // bug!!! caused by aliasing effect |
197 | * \endcode |
198 | * Instead, use the adjointInPlace() method: |
199 | * \code |
200 | * m.adjointInPlace(); |
201 | * \endcode |
202 | * which gives Eigen good opportunities for optimization, or alternatively you can also do: |
203 | * \code |
204 | * m = m.adjoint().eval(); |
205 | * \endcode |
206 | * |
207 | * \sa adjointInPlace(), transpose(), conjugate(), class Transpose, class internal::scalar_conjugate_op */ |
208 | template<typename Derived> |
209 | inline const typename MatrixBase<Derived>::AdjointReturnType |
210 | MatrixBase<Derived>::adjoint() const |
211 | { |
212 | return AdjointReturnType(this->transpose()); |
213 | } |
214 | |
215 | /*************************************************************************** |
216 | * "in place" transpose implementation |
217 | ***************************************************************************/ |
218 | |
219 | namespace internal { |
220 | |
221 | template<typename MatrixType, |
222 | bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic, |
223 | bool MatchPacketSize = |
224 | (int(MatrixType::RowsAtCompileTime) == int(internal::packet_traits<typename MatrixType::Scalar>::size)) |
225 | && (internal::evaluator<MatrixType>::Flags&PacketAccessBit) > |
226 | struct inplace_transpose_selector; |
227 | |
228 | template<typename MatrixType> |
229 | struct inplace_transpose_selector<MatrixType,true,false> { // square matrix |
230 | static void run(MatrixType& m) { |
231 | m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose()); |
232 | } |
233 | }; |
234 | |
235 | // TODO: vectorized path is currently limited to LargestPacketSize x LargestPacketSize cases only. |
236 | template<typename MatrixType> |
237 | struct inplace_transpose_selector<MatrixType,true,true> { // PacketSize x PacketSize |
238 | static void run(MatrixType& m) { |
239 | typedef typename MatrixType::Scalar Scalar; |
240 | typedef typename internal::packet_traits<typename MatrixType::Scalar>::type Packet; |
241 | const Index PacketSize = internal::packet_traits<Scalar>::size; |
242 | const Index Alignment = internal::evaluator<MatrixType>::Alignment; |
243 | PacketBlock<Packet> A; |
244 | for (Index i=0; i<PacketSize; ++i) |
245 | A.packet[i] = m.template packetByOuterInner<Alignment>(i,0); |
246 | internal::ptranspose(A); |
247 | for (Index i=0; i<PacketSize; ++i) |
248 | m.template writePacket<Alignment>(m.rowIndexByOuterInner(i,0), m.colIndexByOuterInner(i,0), A.packet[i]); |
249 | } |
250 | }; |
251 | |
252 | template<typename MatrixType,bool MatchPacketSize> |
253 | struct inplace_transpose_selector<MatrixType,false,MatchPacketSize> { // non square matrix |
254 | static void run(MatrixType& m) { |
255 | if (m.rows()==m.cols()) |
256 | m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose()); |
257 | else |
258 | m = m.transpose().eval(); |
259 | } |
260 | }; |
261 | |
262 | } // end namespace internal |
263 | |
264 | /** This is the "in place" version of transpose(): it replaces \c *this by its own transpose. |
265 | * Thus, doing |
266 | * \code |
267 | * m.transposeInPlace(); |
268 | * \endcode |
269 | * has the same effect on m as doing |
270 | * \code |
271 | * m = m.transpose().eval(); |
272 | * \endcode |
273 | * and is faster and also safer because in the latter line of code, forgetting the eval() results |
274 | * in a bug caused by \ref TopicAliasing "aliasing". |
275 | * |
276 | * Notice however that this method is only useful if you want to replace a matrix by its own transpose. |
277 | * If you just need the transpose of a matrix, use transpose(). |
278 | * |
279 | * \note if the matrix is not square, then \c *this must be a resizable matrix. |
280 | * This excludes (non-square) fixed-size matrices, block-expressions and maps. |
281 | * |
282 | * \sa transpose(), adjoint(), adjointInPlace() */ |
283 | template<typename Derived> |
284 | inline void DenseBase<Derived>::transposeInPlace() |
285 | { |
286 | eigen_assert((rows() == cols() || (RowsAtCompileTime == Dynamic && ColsAtCompileTime == Dynamic)) |
287 | && "transposeInPlace() called on a non-square non-resizable matrix" ); |
288 | internal::inplace_transpose_selector<Derived>::run(derived()); |
289 | } |
290 | |
291 | /*************************************************************************** |
292 | * "in place" adjoint implementation |
293 | ***************************************************************************/ |
294 | |
295 | /** This is the "in place" version of adjoint(): it replaces \c *this by its own transpose. |
296 | * Thus, doing |
297 | * \code |
298 | * m.adjointInPlace(); |
299 | * \endcode |
300 | * has the same effect on m as doing |
301 | * \code |
302 | * m = m.adjoint().eval(); |
303 | * \endcode |
304 | * and is faster and also safer because in the latter line of code, forgetting the eval() results |
305 | * in a bug caused by aliasing. |
306 | * |
307 | * Notice however that this method is only useful if you want to replace a matrix by its own adjoint. |
308 | * If you just need the adjoint of a matrix, use adjoint(). |
309 | * |
310 | * \note if the matrix is not square, then \c *this must be a resizable matrix. |
311 | * This excludes (non-square) fixed-size matrices, block-expressions and maps. |
312 | * |
313 | * \sa transpose(), adjoint(), transposeInPlace() */ |
314 | template<typename Derived> |
315 | inline void MatrixBase<Derived>::adjointInPlace() |
316 | { |
317 | derived() = adjoint().eval(); |
318 | } |
319 | |
320 | #ifndef EIGEN_NO_DEBUG |
321 | |
322 | // The following is to detect aliasing problems in most common cases. |
323 | |
324 | namespace internal { |
325 | |
326 | template<bool DestIsTransposed, typename OtherDerived> |
327 | struct check_transpose_aliasing_compile_time_selector |
328 | { |
329 | enum { ret = bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed }; |
330 | }; |
331 | |
332 | template<bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB> |
333 | struct check_transpose_aliasing_compile_time_selector<DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> > |
334 | { |
335 | enum { ret = bool(blas_traits<DerivedA>::IsTransposed) != DestIsTransposed |
336 | || bool(blas_traits<DerivedB>::IsTransposed) != DestIsTransposed |
337 | }; |
338 | }; |
339 | |
340 | template<typename Scalar, bool DestIsTransposed, typename OtherDerived> |
341 | struct check_transpose_aliasing_run_time_selector |
342 | { |
343 | static bool run(const Scalar* dest, const OtherDerived& src) |
344 | { |
345 | return (bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src)); |
346 | } |
347 | }; |
348 | |
349 | template<typename Scalar, bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB> |
350 | struct check_transpose_aliasing_run_time_selector<Scalar,DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> > |
351 | { |
352 | static bool run(const Scalar* dest, const CwiseBinaryOp<BinOp,DerivedA,DerivedB>& src) |
353 | { |
354 | return ((blas_traits<DerivedA>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.lhs()))) |
355 | || ((blas_traits<DerivedB>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.rhs()))); |
356 | } |
357 | }; |
358 | |
359 | // the following selector, checkTransposeAliasing_impl, based on MightHaveTransposeAliasing, |
360 | // is because when the condition controlling the assert is known at compile time, ICC emits a warning. |
361 | // This is actually a good warning: in expressions that don't have any transposing, the condition is |
362 | // known at compile time to be false, and using that, we can avoid generating the code of the assert again |
363 | // and again for all these expressions that don't need it. |
364 | |
365 | template<typename Derived, typename OtherDerived, |
366 | bool MightHaveTransposeAliasing |
367 | = check_transpose_aliasing_compile_time_selector |
368 | <blas_traits<Derived>::IsTransposed,OtherDerived>::ret |
369 | > |
370 | struct checkTransposeAliasing_impl |
371 | { |
372 | static void run(const Derived& dst, const OtherDerived& other) |
373 | { |
374 | eigen_assert((!check_transpose_aliasing_run_time_selector |
375 | <typename Derived::Scalar,blas_traits<Derived>::IsTransposed,OtherDerived> |
376 | ::run(extract_data(dst), other)) |
377 | && "aliasing detected during transposition, use transposeInPlace() " |
378 | "or evaluate the rhs into a temporary using .eval()" ); |
379 | |
380 | } |
381 | }; |
382 | |
383 | template<typename Derived, typename OtherDerived> |
384 | struct checkTransposeAliasing_impl<Derived, OtherDerived, false> |
385 | { |
386 | static void run(const Derived&, const OtherDerived&) |
387 | { |
388 | } |
389 | }; |
390 | |
391 | template<typename Dst, typename Src> |
392 | void check_for_aliasing(const Dst &dst, const Src &src) |
393 | { |
394 | internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src); |
395 | } |
396 | |
397 | } // end namespace internal |
398 | |
399 | #endif // EIGEN_NO_DEBUG |
400 | |
401 | } // end namespace Eigen |
402 | |
403 | #endif // EIGEN_TRANSPOSE_H |
404 | |