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) 2008-2011 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_GENERAL_PRODUCT_H |
---|
12 | #define EIGEN_GENERAL_PRODUCT_H |
---|
13 | |
---|
14 | namespace Eigen { |
---|
15 | |
---|
16 | /** \class GeneralProduct |
---|
17 | * \ingroup Core_Module |
---|
18 | * |
---|
19 | * \brief Expression of the product of two general matrices or vectors |
---|
20 | * |
---|
21 | * \param LhsNested the type used to store the left-hand side |
---|
22 | * \param RhsNested the type used to store the right-hand side |
---|
23 | * \param ProductMode the type of the product |
---|
24 | * |
---|
25 | * This class represents an expression of the product of two general matrices. |
---|
26 | * We call a general matrix, a dense matrix with full storage. For instance, |
---|
27 | * This excludes triangular, selfadjoint, and sparse matrices. |
---|
28 | * It is the return type of the operator* between general matrices. Its template |
---|
29 | * arguments are determined automatically by ProductReturnType. Therefore, |
---|
30 | * GeneralProduct should never be used direclty. To determine the result type of a |
---|
31 | * function which involves a matrix product, use ProductReturnType::Type. |
---|
32 | * |
---|
33 | * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase<OtherDerived>&) |
---|
34 | */ |
---|
35 | template<typename Lhs, typename Rhs, int ProductType = internal::product_type<Lhs,Rhs>::value> |
---|
36 | class GeneralProduct; |
---|
37 | |
---|
38 | enum { |
---|
39 | Large = 2, |
---|
40 | Small = 3 |
---|
41 | }; |
---|
42 | |
---|
43 | namespace internal { |
---|
44 | |
---|
45 | template<int Rows, int Cols, int Depth> struct product_type_selector; |
---|
46 | |
---|
47 | template<int Size, int MaxSize> struct product_size_category |
---|
48 | { |
---|
49 | enum { is_large = MaxSize == Dynamic || |
---|
50 | Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD, |
---|
51 | value = is_large ? Large |
---|
52 | : Size == 1 ? 1 |
---|
53 | : Small |
---|
54 | }; |
---|
55 | }; |
---|
56 | |
---|
57 | template<typename Lhs, typename Rhs> struct product_type |
---|
58 | { |
---|
59 | typedef typename remove_all<Lhs>::type _Lhs; |
---|
60 | typedef typename remove_all<Rhs>::type _Rhs; |
---|
61 | enum { |
---|
62 | MaxRows = _Lhs::MaxRowsAtCompileTime, |
---|
63 | Rows = _Lhs::RowsAtCompileTime, |
---|
64 | MaxCols = _Rhs::MaxColsAtCompileTime, |
---|
65 | Cols = _Rhs::ColsAtCompileTime, |
---|
66 | MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime, |
---|
67 | _Rhs::MaxRowsAtCompileTime), |
---|
68 | Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, |
---|
69 | _Rhs::RowsAtCompileTime), |
---|
70 | LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD |
---|
71 | }; |
---|
72 | |
---|
73 | // the splitting into different lines of code here, introducing the _select enums and the typedef below, |
---|
74 | // is to work around an internal compiler error with gcc 4.1 and 4.2. |
---|
75 | private: |
---|
76 | enum { |
---|
77 | rows_select = product_size_category<Rows,MaxRows>::value, |
---|
78 | cols_select = product_size_category<Cols,MaxCols>::value, |
---|
79 | depth_select = product_size_category<Depth,MaxDepth>::value |
---|
80 | }; |
---|
81 | typedef product_type_selector<rows_select, cols_select, depth_select> selector; |
---|
82 | |
---|
83 | public: |
---|
84 | enum { |
---|
85 | value = selector::ret |
---|
86 | }; |
---|
87 | #ifdef EIGEN_DEBUG_PRODUCT |
---|
88 | static void debug() |
---|
89 | { |
---|
90 | EIGEN_DEBUG_VAR(Rows); |
---|
91 | EIGEN_DEBUG_VAR(Cols); |
---|
92 | EIGEN_DEBUG_VAR(Depth); |
---|
93 | EIGEN_DEBUG_VAR(rows_select); |
---|
94 | EIGEN_DEBUG_VAR(cols_select); |
---|
95 | EIGEN_DEBUG_VAR(depth_select); |
---|
96 | EIGEN_DEBUG_VAR(value); |
---|
97 | } |
---|
98 | #endif |
---|
99 | }; |
---|
100 | |
---|
101 | |
---|
102 | /* The following allows to select the kind of product at compile time |
---|
103 | * based on the three dimensions of the product. |
---|
104 | * This is a compile time mapping from {1,Small,Large}^3 -> {product types} */ |
---|
105 | // FIXME I'm not sure the current mapping is the ideal one. |
---|
106 | template<int M, int N> struct product_type_selector<M,N,1> { enum { ret = OuterProduct }; }; |
---|
107 | template<int Depth> struct product_type_selector<1, 1, Depth> { enum { ret = InnerProduct }; }; |
---|
108 | template<> struct product_type_selector<1, 1, 1> { enum { ret = InnerProduct }; }; |
---|
109 | template<> struct product_type_selector<Small,1, Small> { enum { ret = CoeffBasedProductMode }; }; |
---|
110 | template<> struct product_type_selector<1, Small,Small> { enum { ret = CoeffBasedProductMode }; }; |
---|
111 | template<> struct product_type_selector<Small,Small,Small> { enum { ret = CoeffBasedProductMode }; }; |
---|
112 | template<> struct product_type_selector<Small, Small, 1> { enum { ret = LazyCoeffBasedProductMode }; }; |
---|
113 | template<> struct product_type_selector<Small, Large, 1> { enum { ret = LazyCoeffBasedProductMode }; }; |
---|
114 | template<> struct product_type_selector<Large, Small, 1> { enum { ret = LazyCoeffBasedProductMode }; }; |
---|
115 | template<> struct product_type_selector<1, Large,Small> { enum { ret = CoeffBasedProductMode }; }; |
---|
116 | template<> struct product_type_selector<1, Large,Large> { enum { ret = GemvProduct }; }; |
---|
117 | template<> struct product_type_selector<1, Small,Large> { enum { ret = CoeffBasedProductMode }; }; |
---|
118 | template<> struct product_type_selector<Large,1, Small> { enum { ret = CoeffBasedProductMode }; }; |
---|
119 | template<> struct product_type_selector<Large,1, Large> { enum { ret = GemvProduct }; }; |
---|
120 | template<> struct product_type_selector<Small,1, Large> { enum { ret = CoeffBasedProductMode }; }; |
---|
121 | template<> struct product_type_selector<Small,Small,Large> { enum { ret = GemmProduct }; }; |
---|
122 | template<> struct product_type_selector<Large,Small,Large> { enum { ret = GemmProduct }; }; |
---|
123 | template<> struct product_type_selector<Small,Large,Large> { enum { ret = GemmProduct }; }; |
---|
124 | template<> struct product_type_selector<Large,Large,Large> { enum { ret = GemmProduct }; }; |
---|
125 | template<> struct product_type_selector<Large,Small,Small> { enum { ret = GemmProduct }; }; |
---|
126 | template<> struct product_type_selector<Small,Large,Small> { enum { ret = GemmProduct }; }; |
---|
127 | template<> struct product_type_selector<Large,Large,Small> { enum { ret = GemmProduct }; }; |
---|
128 | |
---|
129 | } // end namespace internal |
---|
130 | |
---|
131 | /** \class ProductReturnType |
---|
132 | * \ingroup Core_Module |
---|
133 | * |
---|
134 | * \brief Helper class to get the correct and optimized returned type of operator* |
---|
135 | * |
---|
136 | * \param Lhs the type of the left-hand side |
---|
137 | * \param Rhs the type of the right-hand side |
---|
138 | * \param ProductMode the type of the product (determined automatically by internal::product_mode) |
---|
139 | * |
---|
140 | * This class defines the typename Type representing the optimized product expression |
---|
141 | * between two matrix expressions. In practice, using ProductReturnType<Lhs,Rhs>::Type |
---|
142 | * is the recommended way to define the result type of a function returning an expression |
---|
143 | * which involve a matrix product. The class Product should never be |
---|
144 | * used directly. |
---|
145 | * |
---|
146 | * \sa class Product, MatrixBase::operator*(const MatrixBase<OtherDerived>&) |
---|
147 | */ |
---|
148 | template<typename Lhs, typename Rhs, int ProductType> |
---|
149 | struct ProductReturnType |
---|
150 | { |
---|
151 | // TODO use the nested type to reduce instanciations ???? |
---|
152 | // typedef typename internal::nested<Lhs,Rhs::ColsAtCompileTime>::type LhsNested; |
---|
153 | // typedef typename internal::nested<Rhs,Lhs::RowsAtCompileTime>::type RhsNested; |
---|
154 | |
---|
155 | typedef GeneralProduct<Lhs/*Nested*/, Rhs/*Nested*/, ProductType> Type; |
---|
156 | }; |
---|
157 | |
---|
158 | template<typename Lhs, typename Rhs> |
---|
159 | struct ProductReturnType<Lhs,Rhs,CoeffBasedProductMode> |
---|
160 | { |
---|
161 | typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested; |
---|
162 | typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested; |
---|
163 | typedef CoeffBasedProduct<LhsNested, RhsNested, EvalBeforeAssigningBit | EvalBeforeNestingBit> Type; |
---|
164 | }; |
---|
165 | |
---|
166 | template<typename Lhs, typename Rhs> |
---|
167 | struct ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode> |
---|
168 | { |
---|
169 | typedef typename internal::nested<Lhs, Rhs::ColsAtCompileTime, typename internal::plain_matrix_type<Lhs>::type >::type LhsNested; |
---|
170 | typedef typename internal::nested<Rhs, Lhs::RowsAtCompileTime, typename internal::plain_matrix_type<Rhs>::type >::type RhsNested; |
---|
171 | typedef CoeffBasedProduct<LhsNested, RhsNested, NestByRefBit> Type; |
---|
172 | }; |
---|
173 | |
---|
174 | // this is a workaround for sun CC |
---|
175 | template<typename Lhs, typename Rhs> |
---|
176 | struct LazyProductReturnType : public ProductReturnType<Lhs,Rhs,LazyCoeffBasedProductMode> |
---|
177 | {}; |
---|
178 | |
---|
179 | /*********************************************************************** |
---|
180 | * Implementation of Inner Vector Vector Product |
---|
181 | ***********************************************************************/ |
---|
182 | |
---|
183 | // FIXME : maybe the "inner product" could return a Scalar |
---|
184 | // instead of a 1x1 matrix ?? |
---|
185 | // Pro: more natural for the user |
---|
186 | // Cons: this could be a problem if in a meta unrolled algorithm a matrix-matrix |
---|
187 | // product ends up to a row-vector times col-vector product... To tackle this use |
---|
188 | // case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x); |
---|
189 | |
---|
190 | namespace internal { |
---|
191 | |
---|
192 | template<typename Lhs, typename Rhs> |
---|
193 | struct traits<GeneralProduct<Lhs,Rhs,InnerProduct> > |
---|
194 | : traits<Matrix<typename scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> > |
---|
195 | {}; |
---|
196 | |
---|
197 | } |
---|
198 | |
---|
199 | template<typename Lhs, typename Rhs> |
---|
200 | class GeneralProduct<Lhs, Rhs, InnerProduct> |
---|
201 | : internal::no_assignment_operator, |
---|
202 | public Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> |
---|
203 | { |
---|
204 | typedef Matrix<typename internal::scalar_product_traits<typename Lhs::Scalar, typename Rhs::Scalar>::ReturnType,1,1> Base; |
---|
205 | public: |
---|
206 | GeneralProduct(const Lhs& lhs, const Rhs& rhs) |
---|
207 | { |
---|
208 | EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value), |
---|
209 | YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) |
---|
210 | |
---|
211 | Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum(); |
---|
212 | } |
---|
213 | |
---|
214 | /** Convertion to scalar */ |
---|
215 | operator const typename Base::Scalar() const { |
---|
216 | return Base::coeff(0,0); |
---|
217 | } |
---|
218 | }; |
---|
219 | |
---|
220 | /*********************************************************************** |
---|
221 | * Implementation of Outer Vector Vector Product |
---|
222 | ***********************************************************************/ |
---|
223 | |
---|
224 | namespace internal { |
---|
225 | template<int StorageOrder> struct outer_product_selector; |
---|
226 | |
---|
227 | template<typename Lhs, typename Rhs> |
---|
228 | struct traits<GeneralProduct<Lhs,Rhs,OuterProduct> > |
---|
229 | : traits<ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> > |
---|
230 | {}; |
---|
231 | |
---|
232 | } |
---|
233 | |
---|
234 | template<typename Lhs, typename Rhs> |
---|
235 | class GeneralProduct<Lhs, Rhs, OuterProduct> |
---|
236 | : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs> |
---|
237 | { |
---|
238 | public: |
---|
239 | EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) |
---|
240 | |
---|
241 | GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) |
---|
242 | { |
---|
243 | EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value), |
---|
244 | YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) |
---|
245 | } |
---|
246 | |
---|
247 | template<typename Dest> void scaleAndAddTo(Dest& dest, Scalar alpha) const |
---|
248 | { |
---|
249 | internal::outer_product_selector<(int(Dest::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(*this, dest, alpha); |
---|
250 | } |
---|
251 | }; |
---|
252 | |
---|
253 | namespace internal { |
---|
254 | |
---|
255 | template<> struct outer_product_selector<ColMajor> { |
---|
256 | template<typename ProductType, typename Dest> |
---|
257 | static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) { |
---|
258 | typedef typename Dest::Index Index; |
---|
259 | // FIXME make sure lhs is sequentially stored |
---|
260 | // FIXME not very good if rhs is real and lhs complex while alpha is real too |
---|
261 | const Index cols = dest.cols(); |
---|
262 | for (Index j=0; j<cols; ++j) |
---|
263 | dest.col(j) += (alpha * prod.rhs().coeff(j)) * prod.lhs(); |
---|
264 | } |
---|
265 | }; |
---|
266 | |
---|
267 | template<> struct outer_product_selector<RowMajor> { |
---|
268 | template<typename ProductType, typename Dest> |
---|
269 | static EIGEN_DONT_INLINE void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) { |
---|
270 | typedef typename Dest::Index Index; |
---|
271 | // FIXME make sure rhs is sequentially stored |
---|
272 | // FIXME not very good if lhs is real and rhs complex while alpha is real too |
---|
273 | const Index rows = dest.rows(); |
---|
274 | for (Index i=0; i<rows; ++i) |
---|
275 | dest.row(i) += (alpha * prod.lhs().coeff(i)) * prod.rhs(); |
---|
276 | } |
---|
277 | }; |
---|
278 | |
---|
279 | } // end namespace internal |
---|
280 | |
---|
281 | /*********************************************************************** |
---|
282 | * Implementation of General Matrix Vector Product |
---|
283 | ***********************************************************************/ |
---|
284 | |
---|
285 | /* According to the shape/flags of the matrix we have to distinghish 3 different cases: |
---|
286 | * 1 - the matrix is col-major, BLAS compatible and M is large => call fast BLAS-like colmajor routine |
---|
287 | * 2 - the matrix is row-major, BLAS compatible and N is large => call fast BLAS-like rowmajor routine |
---|
288 | * 3 - all other cases are handled using a simple loop along the outer-storage direction. |
---|
289 | * Therefore we need a lower level meta selector. |
---|
290 | * Furthermore, if the matrix is the rhs, then the product has to be transposed. |
---|
291 | */ |
---|
292 | namespace internal { |
---|
293 | |
---|
294 | template<typename Lhs, typename Rhs> |
---|
295 | struct traits<GeneralProduct<Lhs,Rhs,GemvProduct> > |
---|
296 | : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> > |
---|
297 | {}; |
---|
298 | |
---|
299 | template<int Side, int StorageOrder, bool BlasCompatible> |
---|
300 | struct gemv_selector; |
---|
301 | |
---|
302 | } // end namespace internal |
---|
303 | |
---|
304 | template<typename Lhs, typename Rhs> |
---|
305 | class GeneralProduct<Lhs, Rhs, GemvProduct> |
---|
306 | : public ProductBase<GeneralProduct<Lhs,Rhs,GemvProduct>, Lhs, Rhs> |
---|
307 | { |
---|
308 | public: |
---|
309 | EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) |
---|
310 | |
---|
311 | typedef typename Lhs::Scalar LhsScalar; |
---|
312 | typedef typename Rhs::Scalar RhsScalar; |
---|
313 | |
---|
314 | GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) |
---|
315 | { |
---|
316 | // EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::Scalar, typename Rhs::Scalar>::value), |
---|
317 | // YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) |
---|
318 | } |
---|
319 | |
---|
320 | enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight }; |
---|
321 | typedef typename internal::conditional<int(Side)==OnTheRight,_LhsNested,_RhsNested>::type MatrixType; |
---|
322 | |
---|
323 | template<typename Dest> void scaleAndAddTo(Dest& dst, Scalar alpha) const |
---|
324 | { |
---|
325 | eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols()); |
---|
326 | internal::gemv_selector<Side,(int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor, |
---|
327 | bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)>::run(*this, dst, alpha); |
---|
328 | } |
---|
329 | }; |
---|
330 | |
---|
331 | namespace internal { |
---|
332 | |
---|
333 | // The vector is on the left => transposition |
---|
334 | template<int StorageOrder, bool BlasCompatible> |
---|
335 | struct gemv_selector<OnTheLeft,StorageOrder,BlasCompatible> |
---|
336 | { |
---|
337 | template<typename ProductType, typename Dest> |
---|
338 | static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) |
---|
339 | { |
---|
340 | Transpose<Dest> destT(dest); |
---|
341 | enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; |
---|
342 | gemv_selector<OnTheRight,OtherStorageOrder,BlasCompatible> |
---|
343 | ::run(GeneralProduct<Transpose<const typename ProductType::_RhsNested>,Transpose<const typename ProductType::_LhsNested>, GemvProduct> |
---|
344 | (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha); |
---|
345 | } |
---|
346 | }; |
---|
347 | |
---|
348 | template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if; |
---|
349 | |
---|
350 | template<typename Scalar,int Size,int MaxSize> |
---|
351 | struct gemv_static_vector_if<Scalar,Size,MaxSize,false> |
---|
352 | { |
---|
353 | EIGEN_STRONG_INLINE Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; } |
---|
354 | }; |
---|
355 | |
---|
356 | template<typename Scalar,int Size> |
---|
357 | struct gemv_static_vector_if<Scalar,Size,Dynamic,true> |
---|
358 | { |
---|
359 | EIGEN_STRONG_INLINE Scalar* data() { return 0; } |
---|
360 | }; |
---|
361 | |
---|
362 | template<typename Scalar,int Size,int MaxSize> |
---|
363 | struct gemv_static_vector_if<Scalar,Size,MaxSize,true> |
---|
364 | { |
---|
365 | #if EIGEN_ALIGN_STATICALLY |
---|
366 | internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0> m_data; |
---|
367 | EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; } |
---|
368 | #else |
---|
369 | // Some architectures cannot align on the stack, |
---|
370 | // => let's manually enforce alignment by allocating more data and return the address of the first aligned element. |
---|
371 | enum { |
---|
372 | ForceAlignment = internal::packet_traits<Scalar>::Vectorizable, |
---|
373 | PacketSize = internal::packet_traits<Scalar>::size |
---|
374 | }; |
---|
375 | internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?PacketSize:0),0> m_data; |
---|
376 | EIGEN_STRONG_INLINE Scalar* data() { |
---|
377 | return ForceAlignment |
---|
378 | ? reinterpret_cast<Scalar*>((reinterpret_cast<size_t>(m_data.array) & ~(size_t(15))) + 16) |
---|
379 | : m_data.array; |
---|
380 | } |
---|
381 | #endif |
---|
382 | }; |
---|
383 | |
---|
384 | template<> struct gemv_selector<OnTheRight,ColMajor,true> |
---|
385 | { |
---|
386 | template<typename ProductType, typename Dest> |
---|
387 | static inline void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) |
---|
388 | { |
---|
389 | typedef typename ProductType::Index Index; |
---|
390 | typedef typename ProductType::LhsScalar LhsScalar; |
---|
391 | typedef typename ProductType::RhsScalar RhsScalar; |
---|
392 | typedef typename ProductType::Scalar ResScalar; |
---|
393 | typedef typename ProductType::RealScalar RealScalar; |
---|
394 | typedef typename ProductType::ActualLhsType ActualLhsType; |
---|
395 | typedef typename ProductType::ActualRhsType ActualRhsType; |
---|
396 | typedef typename ProductType::LhsBlasTraits LhsBlasTraits; |
---|
397 | typedef typename ProductType::RhsBlasTraits RhsBlasTraits; |
---|
398 | typedef Map<Matrix<ResScalar,Dynamic,1>, Aligned> MappedDest; |
---|
399 | |
---|
400 | ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs()); |
---|
401 | ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs()); |
---|
402 | |
---|
403 | ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) |
---|
404 | * RhsBlasTraits::extractScalarFactor(prod.rhs()); |
---|
405 | |
---|
406 | enum { |
---|
407 | // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 |
---|
408 | // on, the other hand it is good for the cache to pack the vector anyways... |
---|
409 | EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1, |
---|
410 | ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex), |
---|
411 | MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal |
---|
412 | }; |
---|
413 | |
---|
414 | gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest; |
---|
415 | |
---|
416 | bool alphaIsCompatible = (!ComplexByReal) || (imag(actualAlpha)==RealScalar(0)); |
---|
417 | bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible; |
---|
418 | |
---|
419 | RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha); |
---|
420 | |
---|
421 | ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), |
---|
422 | evalToDest ? dest.data() : static_dest.data()); |
---|
423 | |
---|
424 | if(!evalToDest) |
---|
425 | { |
---|
426 | #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN |
---|
427 | int size = dest.size(); |
---|
428 | EIGEN_DENSE_STORAGE_CTOR_PLUGIN |
---|
429 | #endif |
---|
430 | if(!alphaIsCompatible) |
---|
431 | { |
---|
432 | MappedDest(actualDestPtr, dest.size()).setZero(); |
---|
433 | compatibleAlpha = RhsScalar(1); |
---|
434 | } |
---|
435 | else |
---|
436 | MappedDest(actualDestPtr, dest.size()) = dest; |
---|
437 | } |
---|
438 | |
---|
439 | general_matrix_vector_product |
---|
440 | <Index,LhsScalar,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run( |
---|
441 | actualLhs.rows(), actualLhs.cols(), |
---|
442 | actualLhs.data(), actualLhs.outerStride(), |
---|
443 | actualRhs.data(), actualRhs.innerStride(), |
---|
444 | actualDestPtr, 1, |
---|
445 | compatibleAlpha); |
---|
446 | |
---|
447 | if (!evalToDest) |
---|
448 | { |
---|
449 | if(!alphaIsCompatible) |
---|
450 | dest += actualAlpha * MappedDest(actualDestPtr, dest.size()); |
---|
451 | else |
---|
452 | dest = MappedDest(actualDestPtr, dest.size()); |
---|
453 | } |
---|
454 | } |
---|
455 | }; |
---|
456 | |
---|
457 | template<> struct gemv_selector<OnTheRight,RowMajor,true> |
---|
458 | { |
---|
459 | template<typename ProductType, typename Dest> |
---|
460 | static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) |
---|
461 | { |
---|
462 | typedef typename ProductType::LhsScalar LhsScalar; |
---|
463 | typedef typename ProductType::RhsScalar RhsScalar; |
---|
464 | typedef typename ProductType::Scalar ResScalar; |
---|
465 | typedef typename ProductType::Index Index; |
---|
466 | typedef typename ProductType::ActualLhsType ActualLhsType; |
---|
467 | typedef typename ProductType::ActualRhsType ActualRhsType; |
---|
468 | typedef typename ProductType::_ActualRhsType _ActualRhsType; |
---|
469 | typedef typename ProductType::LhsBlasTraits LhsBlasTraits; |
---|
470 | typedef typename ProductType::RhsBlasTraits RhsBlasTraits; |
---|
471 | |
---|
472 | typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(prod.lhs()); |
---|
473 | typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(prod.rhs()); |
---|
474 | |
---|
475 | ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) |
---|
476 | * RhsBlasTraits::extractScalarFactor(prod.rhs()); |
---|
477 | |
---|
478 | enum { |
---|
479 | // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1 |
---|
480 | // on, the other hand it is good for the cache to pack the vector anyways... |
---|
481 | DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1 |
---|
482 | }; |
---|
483 | |
---|
484 | gemv_static_vector_if<RhsScalar,_ActualRhsType::SizeAtCompileTime,_ActualRhsType::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs; |
---|
485 | |
---|
486 | ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), |
---|
487 | DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data()); |
---|
488 | |
---|
489 | if(!DirectlyUseRhs) |
---|
490 | { |
---|
491 | #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN |
---|
492 | int size = actualRhs.size(); |
---|
493 | EIGEN_DENSE_STORAGE_CTOR_PLUGIN |
---|
494 | #endif |
---|
495 | Map<typename _ActualRhsType::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs; |
---|
496 | } |
---|
497 | |
---|
498 | general_matrix_vector_product |
---|
499 | <Index,LhsScalar,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsBlasTraits::NeedToConjugate>::run( |
---|
500 | actualLhs.rows(), actualLhs.cols(), |
---|
501 | actualLhs.data(), actualLhs.outerStride(), |
---|
502 | actualRhsPtr, 1, |
---|
503 | dest.data(), dest.innerStride(), |
---|
504 | actualAlpha); |
---|
505 | } |
---|
506 | }; |
---|
507 | |
---|
508 | template<> struct gemv_selector<OnTheRight,ColMajor,false> |
---|
509 | { |
---|
510 | template<typename ProductType, typename Dest> |
---|
511 | static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) |
---|
512 | { |
---|
513 | typedef typename Dest::Index Index; |
---|
514 | // TODO makes sure dest is sequentially stored in memory, otherwise use a temp |
---|
515 | const Index size = prod.rhs().rows(); |
---|
516 | for(Index k=0; k<size; ++k) |
---|
517 | dest += (alpha*prod.rhs().coeff(k)) * prod.lhs().col(k); |
---|
518 | } |
---|
519 | }; |
---|
520 | |
---|
521 | template<> struct gemv_selector<OnTheRight,RowMajor,false> |
---|
522 | { |
---|
523 | template<typename ProductType, typename Dest> |
---|
524 | static void run(const ProductType& prod, Dest& dest, typename ProductType::Scalar alpha) |
---|
525 | { |
---|
526 | typedef typename Dest::Index Index; |
---|
527 | // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp |
---|
528 | const Index rows = prod.rows(); |
---|
529 | for(Index i=0; i<rows; ++i) |
---|
530 | dest.coeffRef(i) += alpha * (prod.lhs().row(i).cwiseProduct(prod.rhs().transpose())).sum(); |
---|
531 | } |
---|
532 | }; |
---|
533 | |
---|
534 | } // end namespace internal |
---|
535 | |
---|
536 | /*************************************************************************** |
---|
537 | * Implementation of matrix base methods |
---|
538 | ***************************************************************************/ |
---|
539 | |
---|
540 | /** \returns the matrix product of \c *this and \a other. |
---|
541 | * |
---|
542 | * \note If instead of the matrix product you want the coefficient-wise product, see Cwise::operator*(). |
---|
543 | * |
---|
544 | * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*() |
---|
545 | */ |
---|
546 | template<typename Derived> |
---|
547 | template<typename OtherDerived> |
---|
548 | inline const typename ProductReturnType<Derived, OtherDerived>::Type |
---|
549 | MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const |
---|
550 | { |
---|
551 | // A note regarding the function declaration: In MSVC, this function will sometimes |
---|
552 | // not be inlined since DenseStorage is an unwindable object for dynamic |
---|
553 | // matrices and product types are holding a member to store the result. |
---|
554 | // Thus it does not help tagging this function with EIGEN_STRONG_INLINE. |
---|
555 | enum { |
---|
556 | ProductIsValid = Derived::ColsAtCompileTime==Dynamic |
---|
557 | || OtherDerived::RowsAtCompileTime==Dynamic |
---|
558 | || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), |
---|
559 | AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, |
---|
560 | SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) |
---|
561 | }; |
---|
562 | // note to the lost user: |
---|
563 | // * for a dot product use: v1.dot(v2) |
---|
564 | // * for a coeff-wise product use: v1.cwiseProduct(v2) |
---|
565 | EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), |
---|
566 | INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) |
---|
567 | EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), |
---|
568 | INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) |
---|
569 | EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) |
---|
570 | #ifdef EIGEN_DEBUG_PRODUCT |
---|
571 | internal::product_type<Derived,OtherDerived>::debug(); |
---|
572 | #endif |
---|
573 | return typename ProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived()); |
---|
574 | } |
---|
575 | |
---|
576 | /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation. |
---|
577 | * |
---|
578 | * The returned product will behave like any other expressions: the coefficients of the product will be |
---|
579 | * computed once at a time as requested. This might be useful in some extremely rare cases when only |
---|
580 | * a small and no coherent fraction of the result's coefficients have to be computed. |
---|
581 | * |
---|
582 | * \warning This version of the matrix product can be much much slower. So use it only if you know |
---|
583 | * what you are doing and that you measured a true speed improvement. |
---|
584 | * |
---|
585 | * \sa operator*(const MatrixBase&) |
---|
586 | */ |
---|
587 | template<typename Derived> |
---|
588 | template<typename OtherDerived> |
---|
589 | const typename LazyProductReturnType<Derived,OtherDerived>::Type |
---|
590 | MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const |
---|
591 | { |
---|
592 | enum { |
---|
593 | ProductIsValid = Derived::ColsAtCompileTime==Dynamic |
---|
594 | || OtherDerived::RowsAtCompileTime==Dynamic |
---|
595 | || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), |
---|
596 | AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, |
---|
597 | SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) |
---|
598 | }; |
---|
599 | // note to the lost user: |
---|
600 | // * for a dot product use: v1.dot(v2) |
---|
601 | // * for a coeff-wise product use: v1.cwiseProduct(v2) |
---|
602 | EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), |
---|
603 | INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) |
---|
604 | EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), |
---|
605 | INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) |
---|
606 | EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) |
---|
607 | |
---|
608 | return typename LazyProductReturnType<Derived,OtherDerived>::Type(derived(), other.derived()); |
---|
609 | } |
---|
610 | |
---|
611 | } // end namespace Eigen |
---|
612 | |
---|
613 | #endif // EIGEN_PRODUCT_H |
---|