GeneralMatrixMatrix.h
Go to the documentation of this file.
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
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_GENERAL_MATRIX_MATRIX_H
11 #define EIGEN_GENERAL_MATRIX_MATRIX_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 template<typename _LhsScalar, typename _RhsScalar> class level3_blocking;
18 
19 /* Specialization for a row-major destination matrix => simple transposition of the product */
20 template<
21  typename Index,
22  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
23  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
24 struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor>
25 {
27  static EIGEN_STRONG_INLINE void run(
28  Index rows, Index cols, Index depth,
29  const LhsScalar* lhs, Index lhsStride,
30  const RhsScalar* rhs, Index rhsStride,
31  ResScalar* res, Index resStride,
32  ResScalar alpha,
34  GemmParallelInfo<Index>* info = 0)
35  {
36  // transpose the product such that the result is column major
38  RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
39  LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
40  ColMajor>
41  ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking,info);
42  }
43 };
44 
45 /* Specialization for a col-major destination matrix
46  * => Blocking algorithm following Goto's paper */
47 template<
48  typename Index,
49  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
50  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
51 struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor>
52 {
53 
55 static void run(Index rows, Index cols, Index depth,
56  const LhsScalar* _lhs, Index lhsStride,
57  const RhsScalar* _rhs, Index rhsStride,
58  ResScalar* res, Index resStride,
59  ResScalar alpha,
61  GemmParallelInfo<Index>* info = 0)
62 {
65 
66  typedef gebp_traits<LhsScalar,RhsScalar> Traits;
67 
68  Index kc = blocking.kc(); // cache block size along the K direction
69  Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction
70  //Index nc = blocking.nc(); // cache block size along the N direction
71 
75 
76 #ifdef EIGEN_HAS_OPENMP
77  if(info)
78  {
79  // this is the parallel version!
80  Index tid = omp_get_thread_num();
81  Index threads = omp_get_num_threads();
82 
83  std::size_t sizeA = kc*mc;
84  std::size_t sizeW = kc*Traits::WorkSpaceFactor;
85  ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, 0);
86  ei_declare_aligned_stack_constructed_variable(RhsScalar, w, sizeW, 0);
87 
88  RhsScalar* blockB = blocking.blockB();
89  eigen_internal_assert(blockB!=0);
90 
91  // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs...
92  for(Index k=0; k<depth; k+=kc)
93  {
94  const Index actual_kc = (std::min)(k+kc,depth)-k; // => rows of B', and cols of the A'
95 
96  // In order to reduce the chance that a thread has to wait for the other,
97  // let's start by packing A'.
98  pack_lhs(blockA, &lhs(0,k), lhsStride, actual_kc, mc);
99 
100  // Pack B_k to B' in a parallel fashion:
101  // each thread packs the sub block B_k,j to B'_j where j is the thread id.
102 
103  // However, before copying to B'_j, we have to make sure that no other thread is still using it,
104  // i.e., we test that info[tid].users equals 0.
105  // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
106  while(info[tid].users!=0) {}
107  info[tid].users += threads;
108 
109  pack_rhs(blockB+info[tid].rhs_start*actual_kc, &rhs(k,info[tid].rhs_start), rhsStride, actual_kc, info[tid].rhs_length);
110 
111  // Notify the other threads that the part B'_j is ready to go.
112  info[tid].sync = k;
113 
114  // Computes C_i += A' * B' per B'_j
115  for(Index shift=0; shift<threads; ++shift)
116  {
117  Index j = (tid+shift)%threads;
118 
119  // At this point we have to make sure that B'_j has been updated by the thread j,
120  // we use testAndSetOrdered to mimic a volatile access.
121  // However, no need to wait for the B' part which has been updated by the current thread!
122  if(shift>0)
123  while(info[j].sync!=k) {}
124 
125  gebp(res+info[j].rhs_start*resStride, resStride, blockA, blockB+info[j].rhs_start*actual_kc, mc, actual_kc, info[j].rhs_length, alpha, -1,-1,0,0, w);
126  }
127 
128  // Then keep going as usual with the remaining A'
129  for(Index i=mc; i<rows; i+=mc)
130  {
131  const Index actual_mc = (std::min)(i+mc,rows)-i;
132 
133  // pack A_i,k to A'
134  pack_lhs(blockA, &lhs(i,k), lhsStride, actual_kc, actual_mc);
135 
136  // C_i += A' * B'
137  gebp(res+i, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1,-1,0,0, w);
138  }
139 
140  // Release all the sub blocks B'_j of B' for the current thread,
141  // i.e., we simply decrement the number of users by 1
142  for(Index j=0; j<threads; ++j)
143  #pragma omp atomic
144  --(info[j].users);
145  }
146  }
147  else
148 #endif // EIGEN_HAS_OPENMP
149  {
150  EIGEN_UNUSED_VARIABLE(info);
151 
152  // this is the sequential version!
153  std::size_t sizeA = kc*mc;
154  std::size_t sizeB = kc*cols;
155  std::size_t sizeW = kc*Traits::WorkSpaceFactor;
156 
157  ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
158  ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
159  ei_declare_aligned_stack_constructed_variable(RhsScalar, blockW, sizeW, blocking.blockW());
160 
161  // For each horizontal panel of the rhs, and corresponding panel of the lhs...
162  // (==GEMM_VAR1)
163  for(Index k2=0; k2<depth; k2+=kc)
164  {
165  const Index actual_kc = (std::min)(k2+kc,depth)-k2;
166 
167  // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs.
168  // => Pack rhs's panel into a sequential chunk of memory (L2 caching)
169  // Note that this panel will be read as many times as the number of blocks in the lhs's
170  // vertical panel which is, in practice, a very low number.
171  pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, cols);
172 
173  // For each mc x kc block of the lhs's vertical panel...
174  // (==GEPP_VAR1)
175  for(Index i2=0; i2<rows; i2+=mc)
176  {
177  const Index actual_mc = (std::min)(i2+mc,rows)-i2;
178 
179  // We pack the lhs's block into a sequential chunk of memory (L1 caching)
180  // Note that this block will be read a very high number of times, which is equal to the number of
181  // micro vertical panel of the large rhs's panel (e.g., cols/4 times).
182  pack_lhs(blockA, &lhs(i2,k2), lhsStride, actual_kc, actual_mc);
183 
184  // Everything is packed, we can now call the block * panel kernel:
185  gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW);
186  }
187  }
188  }
189 }
190 
191 };
192 
193 /*********************************************************************************
194 * Specialization of GeneralProduct<> for "large" GEMM, i.e.,
195 * implementation of the high level wrapper to general_matrix_matrix_product
196 **********************************************************************************/
197 
198 template<typename Lhs, typename Rhs>
200  : traits<ProductBase<GeneralProduct<Lhs,Rhs,GemmProduct>, Lhs, Rhs> >
201 {};
202 
203 template<typename Scalar, typename Index, typename Gemm, typename Lhs, typename Rhs, typename Dest, typename BlockingType>
205 {
206  gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha,
207  BlockingType& blocking)
208  : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking)
209  {}
210 
211  void initParallelSession() const
212  {
213  m_blocking.allocateB();
214  }
215 
216  void operator() (Index row, Index rows, Index col=0, Index cols=-1, GemmParallelInfo<Index>* info=0) const
217  {
218  if(cols==-1)
219  cols = m_rhs.cols();
220 
221  Gemm::run(rows, cols, m_lhs.cols(),
222  /*(const Scalar*)*/&m_lhs.coeffRef(row,0), m_lhs.outerStride(),
223  /*(const Scalar*)*/&m_rhs.coeffRef(0,col), m_rhs.outerStride(),
224  (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
225  m_actualAlpha, m_blocking, info);
226  }
227 
228  protected:
229  const Lhs& m_lhs;
230  const Rhs& m_rhs;
231  Dest& m_dest;
233  BlockingType& m_blocking;
234 };
235 
236 template<int StorageOrder, typename LhsScalar, typename RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor=1,
237 bool FiniteAtCompileTime = MaxRows!=Dynamic && MaxCols!=Dynamic && MaxDepth != Dynamic> class gemm_blocking_space;
238 
239 template<typename _LhsScalar, typename _RhsScalar>
240 class level3_blocking
241 {
242  typedef _LhsScalar LhsScalar;
243  typedef _RhsScalar RhsScalar;
244 
245  protected:
246  LhsScalar* m_blockA;
247  RhsScalar* m_blockB;
248  RhsScalar* m_blockW;
249 
253 
254  public:
255 
257  : m_blockA(0), m_blockB(0), m_blockW(0), m_mc(0), m_nc(0), m_kc(0)
258  {}
259 
260  inline DenseIndex mc() const { return m_mc; }
261  inline DenseIndex nc() const { return m_nc; }
262  inline DenseIndex kc() const { return m_kc; }
263 
264  inline LhsScalar* blockA() { return m_blockA; }
265  inline RhsScalar* blockB() { return m_blockB; }
266  inline RhsScalar* blockW() { return m_blockW; }
267 };
268 
269 template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
270 class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true>
271  : public level3_blocking<
272  typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
273  typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
274 {
275  enum {
276  Transpose = StorageOrder==RowMajor,
277  ActualRows = Transpose ? MaxCols : MaxRows,
278  ActualCols = Transpose ? MaxRows : MaxCols
279  };
283  enum {
284  SizeA = ActualRows * MaxDepth,
285  SizeB = ActualCols * MaxDepth,
286  SizeW = MaxDepth * Traits::WorkSpaceFactor
287  };
288 
289  EIGEN_ALIGN16 LhsScalar m_staticA[SizeA];
290  EIGEN_ALIGN16 RhsScalar m_staticB[SizeB];
291  EIGEN_ALIGN16 RhsScalar m_staticW[SizeW];
292 
293  public:
294 
295  gemm_blocking_space(DenseIndex /*rows*/, DenseIndex /*cols*/, DenseIndex /*depth*/)
296  {
297  this->m_mc = ActualRows;
298  this->m_nc = ActualCols;
299  this->m_kc = MaxDepth;
300  this->m_blockA = m_staticA;
301  this->m_blockB = m_staticB;
302  this->m_blockW = m_staticW;
303  }
304 
305  inline void allocateA() {}
306  inline void allocateB() {}
307  inline void allocateW() {}
308  inline void allocateAll() {}
309 };
310 
311 template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
312 class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, false>
313  : public level3_blocking<
314  typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
315  typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
316 {
317  enum {
318  Transpose = StorageOrder==RowMajor
319  };
323 
327 
328  public:
329 
331  {
332  this->m_mc = Transpose ? cols : rows;
333  this->m_nc = Transpose ? rows : cols;
334  this->m_kc = depth;
335 
336  computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc);
337  m_sizeA = this->m_mc * this->m_kc;
338  m_sizeB = this->m_kc * this->m_nc;
339  m_sizeW = this->m_kc*Traits::WorkSpaceFactor;
340  }
341 
342  void allocateA()
343  {
344  if(this->m_blockA==0)
345  this->m_blockA = aligned_new<LhsScalar>(m_sizeA);
346  }
347 
348  void allocateB()
349  {
350  if(this->m_blockB==0)
351  this->m_blockB = aligned_new<RhsScalar>(m_sizeB);
352  }
353 
354  void allocateW()
355  {
356  if(this->m_blockW==0)
357  this->m_blockW = aligned_new<RhsScalar>(m_sizeW);
358  }
359 
360  void allocateAll()
361  {
362  allocateA();
363  allocateB();
364  allocateW();
365  }
366 
368  {
369  aligned_delete(this->m_blockA, m_sizeA);
370  aligned_delete(this->m_blockB, m_sizeB);
371  aligned_delete(this->m_blockW, m_sizeW);
372  }
373 };
374 
375 } // end namespace internal
376 
377 template<typename Lhs, typename Rhs>
378 class GeneralProduct<Lhs, Rhs, GemmProduct>
379  : public ProductBase<GeneralProduct<Lhs,Rhs,GemmProduct>, Lhs, Rhs>
380 {
381  enum {
382  MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
383  };
384  public:
386 
387  typedef typename Lhs::Scalar LhsScalar;
388  typedef typename Rhs::Scalar RhsScalar;
389  typedef Scalar ResScalar;
390 
391  GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
392  {
394  EIGEN_CHECK_BINARY_COMPATIBILIY(BinOp,LhsScalar,RhsScalar);
395  }
396 
397  template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
398  {
399  eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
400 
401  typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
402  typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);
403 
404  Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs)
405  * RhsBlasTraits::extractScalarFactor(m_rhs);
406 
408  Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
409 
410  typedef internal::gemm_functor<
411  Scalar, Index,
413  Index,
414  LhsScalar, (_ActualLhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate),
415  RhsScalar, (_ActualRhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate),
416  (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
417  _ActualLhsType, _ActualRhsType, Dest, BlockingType> GemmFunctor;
418 
419  BlockingType blocking(dst.rows(), dst.cols(), lhs.cols());
420 
421  internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit);
422  }
423 };
424 
425 } // end namespace Eigen
426 
427 #endif // EIGEN_GENERAL_MATRIX_MATRIX_H
Expression of the product of two general matrices or vectors.
internal::remove_all< ActualRhsType >::type _ActualRhsType
Definition: ProductBase.h:80
#define EIGEN_PRODUCT_PUBLIC_INTERFACE(Derived)
Definition: ProductBase.h:46
#define EIGEN_ALIGN16
static void run(Index rows, Index cols, Index depth, const LhsScalar *_lhs, Index lhsStride, const RhsScalar *_rhs, Index rhsStride, ResScalar *res, Index resStride, ResScalar alpha, level3_blocking< LhsScalar, RhsScalar > &blocking, GemmParallelInfo< Index > *info=0)
#define EIGEN_STRONG_INLINE
internal::traits< Derived >::Scalar Scalar
Definition: DenseBase.h:63
static EIGEN_STRONG_INLINE void run(Index rows, Index cols, Index depth, const LhsScalar *lhs, Index lhsStride, const RhsScalar *rhs, Index rhsStride, ResScalar *res, Index resStride, ResScalar alpha, level3_blocking< RhsScalar, LhsScalar > &blocking, GemmParallelInfo< Index > *info=0)
#define ei_declare_aligned_stack_constructed_variable(TYPE, NAME, SIZE, BUFFER)
internal::traits< Derived >::Index Index
The type of indices.
Definition: DenseBase.h:61
Expression of the transpose of a matrix.
Definition: Transpose.h:57
#define EIGEN_UNUSED_VARIABLE(var)
iterative scaling algorithm to equilibrate rows and column norms in matrices
Definition: matrix.hpp:471
#define eigen_internal_assert(x)
#define EIGEN_SIZE_MIN_PREFER_FIXED(a, b)
const unsigned int RowMajorBit
gemm_functor(const Lhs &lhs, const Rhs &rhs, Dest &dest, const Scalar &actualAlpha, BlockingType &blocking)
internal::remove_all< ActualLhsType >::type _ActualLhsType
Definition: ProductBase.h:73
void scaleAndAddTo(Dest &dst, const Scalar &alpha) const
void rhs(const real_t *x, real_t *f)
void aligned_delete(T *ptr, size_t size)
EIGEN_DEFAULT_DENSE_INDEX_TYPE DenseIndex
Definition: XprHelper.h:27
RowXpr row(Index i)
Definition: BlockMethods.h:725
ColXpr col(Index i)
Definition: BlockMethods.h:708
#define eigen_assert(x)
#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP, LHS, RHS)
Definition: CwiseBinaryOp.h:96


acado
Author(s): Milan Vukov, Rien Quirynen
autogenerated on Mon Jun 10 2019 12:34:38