• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 {
26   typedef gebp_traits<RhsScalar,LhsScalar> Traits;
27 
28   typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
29   static EIGEN_STRONG_INLINE void run(
30     Index rows, Index cols, Index depth,
31     const LhsScalar* lhs, Index lhsStride,
32     const RhsScalar* rhs, Index rhsStride,
33     ResScalar* res, Index resStride,
34     ResScalar alpha,
35     level3_blocking<RhsScalar,LhsScalar>& blocking,
36     GemmParallelInfo<Index>* info = 0)
37   {
38     // transpose the product such that the result is column major
39     general_matrix_matrix_product<Index,
40       RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
41       LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
42       ColMajor>
43     ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking,info);
44   }
45 };
46 
47 /*  Specialization for a col-major destination matrix
48  *    => Blocking algorithm following Goto's paper */
49 template<
50   typename Index,
51   typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
52   typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
53 struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor>
54 {
55 
56 typedef gebp_traits<LhsScalar,RhsScalar> Traits;
57 
58 typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
59 static void run(Index rows, Index cols, Index depth,
60   const LhsScalar* _lhs, Index lhsStride,
61   const RhsScalar* _rhs, Index rhsStride,
62   ResScalar* _res, Index resStride,
63   ResScalar alpha,
64   level3_blocking<LhsScalar,RhsScalar>& blocking,
65   GemmParallelInfo<Index>* info = 0)
66 {
67   typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper;
68   typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper;
69   typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
70   LhsMapper lhs(_lhs,lhsStride);
71   RhsMapper rhs(_rhs,rhsStride);
72   ResMapper res(_res, resStride);
73 
74   Index kc = blocking.kc();                   // cache block size along the K direction
75   Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
76   Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
77 
78   gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
79   gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
80   gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
81 
82 #ifdef EIGEN_HAS_OPENMP
83   if(info)
84   {
85     // this is the parallel version!
86     int tid = omp_get_thread_num();
87     int threads = omp_get_num_threads();
88 
89     LhsScalar* blockA = blocking.blockA();
90     eigen_internal_assert(blockA!=0);
91 
92     std::size_t sizeB = kc*nc;
93     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, 0);
94 
95     // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs...
96     for(Index k=0; k<depth; k+=kc)
97     {
98       const Index actual_kc = (std::min)(k+kc,depth)-k; // => rows of B', and cols of the A'
99 
100       // In order to reduce the chance that a thread has to wait for the other,
101       // let's start by packing B'.
102       pack_rhs(blockB, rhs.getSubMapper(k,0), actual_kc, nc);
103 
104       // Pack A_k to A' in a parallel fashion:
105       // each thread packs the sub block A_k,i to A'_i where i is the thread id.
106 
107       // However, before copying to A'_i, we have to make sure that no other thread is still using it,
108       // i.e., we test that info[tid].users equals 0.
109       // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
110       while(info[tid].users!=0) {}
111       info[tid].users += threads;
112 
113       pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
114 
115       // Notify the other threads that the part A'_i is ready to go.
116       info[tid].sync = k;
117 
118       // Computes C_i += A' * B' per A'_i
119       for(int shift=0; shift<threads; ++shift)
120       {
121         int i = (tid+shift)%threads;
122 
123         // At this point we have to make sure that A'_i has been updated by the thread i,
124         // we use testAndSetOrdered to mimic a volatile access.
125         // However, no need to wait for the B' part which has been updated by the current thread!
126         if (shift>0) {
127           while(info[i].sync!=k) {
128           }
129         }
130 
131         gebp(res.getSubMapper(info[i].lhs_start, 0), blockA+info[i].lhs_start*actual_kc, blockB, info[i].lhs_length, actual_kc, nc, alpha);
132       }
133 
134       // Then keep going as usual with the remaining B'
135       for(Index j=nc; j<cols; j+=nc)
136       {
137         const Index actual_nc = (std::min)(j+nc,cols)-j;
138 
139         // pack B_k,j to B'
140         pack_rhs(blockB, rhs.getSubMapper(k,j), actual_kc, actual_nc);
141 
142         // C_j += A' * B'
143         gebp(res.getSubMapper(0, j), blockA, blockB, rows, actual_kc, actual_nc, alpha);
144       }
145 
146       // Release all the sub blocks A'_i of A' for the current thread,
147       // i.e., we simply decrement the number of users by 1
148       for(Index i=0; i<threads; ++i)
149         info[i].users -= 1;
150     }
151   }
152   else
153 #endif // EIGEN_HAS_OPENMP
154   {
155     EIGEN_UNUSED_VARIABLE(info);
156 
157     // this is the sequential version!
158     std::size_t sizeA = kc*mc;
159     std::size_t sizeB = kc*nc;
160 
161     ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
162     ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
163 
164     const bool pack_rhs_once = mc!=rows && kc==depth && nc==cols;
165 
166     // For each horizontal panel of the rhs, and corresponding panel of the lhs...
167     for(Index i2=0; i2<rows; i2+=mc)
168     {
169       const Index actual_mc = (std::min)(i2+mc,rows)-i2;
170 
171       for(Index k2=0; k2<depth; k2+=kc)
172       {
173         const Index actual_kc = (std::min)(k2+kc,depth)-k2;
174 
175         // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs.
176         // => Pack lhs's panel into a sequential chunk of memory (L2/L3 caching)
177         // Note that this panel will be read as many times as the number of blocks in the rhs's
178         // horizontal panel which is, in practice, a very low number.
179         pack_lhs(blockA, lhs.getSubMapper(i2,k2), actual_kc, actual_mc);
180 
181         // For each kc x nc block of the rhs's horizontal panel...
182         for(Index j2=0; j2<cols; j2+=nc)
183         {
184           const Index actual_nc = (std::min)(j2+nc,cols)-j2;
185 
186           // We pack the rhs's block into a sequential chunk of memory (L2 caching)
187           // Note that this block will be read a very high number of times, which is equal to the number of
188           // micro horizontal panel of the large rhs's panel (e.g., rows/12 times).
189           if((!pack_rhs_once) || i2==0)
190             pack_rhs(blockB, rhs.getSubMapper(k2,j2), actual_kc, actual_nc);
191 
192           // Everything is packed, we can now call the panel * block kernel:
193           gebp(res.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, alpha);
194         }
195       }
196     }
197   }
198 }
199 
200 };
201 
202 /*********************************************************************************
203 *  Specialization of generic_product_impl for "large" GEMM, i.e.,
204 *  implementation of the high level wrapper to general_matrix_matrix_product
205 **********************************************************************************/
206 
207 template<typename Scalar, typename Index, typename Gemm, typename Lhs, typename Rhs, typename Dest, typename BlockingType>
208 struct gemm_functor
209 {
210   gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha, BlockingType& blocking)
211     : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking)
212   {}
213 
214   void initParallelSession(Index num_threads) const
215   {
216     m_blocking.initParallel(m_lhs.rows(), m_rhs.cols(), m_lhs.cols(), num_threads);
217     m_blocking.allocateA();
218   }
219 
220   void operator() (Index row, Index rows, Index col=0, Index cols=-1, GemmParallelInfo<Index>* info=0) const
221   {
222     if(cols==-1)
223       cols = m_rhs.cols();
224 
225     Gemm::run(rows, cols, m_lhs.cols(),
226               &m_lhs.coeffRef(row,0), m_lhs.outerStride(),
227               &m_rhs.coeffRef(0,col), m_rhs.outerStride(),
228               (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
229               m_actualAlpha, m_blocking, info);
230   }
231 
232   typedef typename Gemm::Traits Traits;
233 
234   protected:
235     const Lhs& m_lhs;
236     const Rhs& m_rhs;
237     Dest& m_dest;
238     Scalar m_actualAlpha;
239     BlockingType& m_blocking;
240 };
241 
242 template<int StorageOrder, typename LhsScalar, typename RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor=1,
243 bool FiniteAtCompileTime = MaxRows!=Dynamic && MaxCols!=Dynamic && MaxDepth != Dynamic> class gemm_blocking_space;
244 
245 template<typename _LhsScalar, typename _RhsScalar>
246 class level3_blocking
247 {
248     typedef _LhsScalar LhsScalar;
249     typedef _RhsScalar RhsScalar;
250 
251   protected:
252     LhsScalar* m_blockA;
253     RhsScalar* m_blockB;
254 
255     Index m_mc;
256     Index m_nc;
257     Index m_kc;
258 
259   public:
260 
261     level3_blocking()
262       : m_blockA(0), m_blockB(0), m_mc(0), m_nc(0), m_kc(0)
263     {}
264 
265     inline Index mc() const { return m_mc; }
266     inline Index nc() const { return m_nc; }
267     inline Index kc() const { return m_kc; }
268 
269     inline LhsScalar* blockA() { return m_blockA; }
270     inline RhsScalar* blockB() { return m_blockB; }
271 };
272 
273 template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
274 class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true /* == FiniteAtCompileTime */>
275   : public level3_blocking<
276       typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
277       typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
278 {
279     enum {
280       Transpose = StorageOrder==RowMajor,
281       ActualRows = Transpose ? MaxCols : MaxRows,
282       ActualCols = Transpose ? MaxRows : MaxCols
283     };
284     typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
285     typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
286     typedef gebp_traits<LhsScalar,RhsScalar> Traits;
287     enum {
288       SizeA = ActualRows * MaxDepth,
289       SizeB = ActualCols * MaxDepth
290     };
291 
292 #if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES
293     EIGEN_ALIGN_MAX LhsScalar m_staticA[SizeA];
294     EIGEN_ALIGN_MAX RhsScalar m_staticB[SizeB];
295 #else
296     EIGEN_ALIGN_MAX char m_staticA[SizeA * sizeof(LhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
297     EIGEN_ALIGN_MAX char m_staticB[SizeB * sizeof(RhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
298 #endif
299 
300   public:
301 
302     gemm_blocking_space(Index /*rows*/, Index /*cols*/, Index /*depth*/, Index /*num_threads*/, bool /*full_rows = false*/)
303     {
304       this->m_mc = ActualRows;
305       this->m_nc = ActualCols;
306       this->m_kc = MaxDepth;
307 #if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES
308       this->m_blockA = m_staticA;
309       this->m_blockB = m_staticB;
310 #else
311       this->m_blockA = reinterpret_cast<LhsScalar*>((internal::UIntPtr(m_staticA) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
312       this->m_blockB = reinterpret_cast<RhsScalar*>((internal::UIntPtr(m_staticB) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
313 #endif
314     }
315 
316     void initParallel(Index, Index, Index, Index)
317     {}
318 
319     inline void allocateA() {}
320     inline void allocateB() {}
321     inline void allocateAll() {}
322 };
323 
324 template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
325 class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, false>
326   : public level3_blocking<
327       typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
328       typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
329 {
330     enum {
331       Transpose = StorageOrder==RowMajor
332     };
333     typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
334     typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
335     typedef gebp_traits<LhsScalar,RhsScalar> Traits;
336 
337     Index m_sizeA;
338     Index m_sizeB;
339 
340   public:
341 
342     gemm_blocking_space(Index rows, Index cols, Index depth, Index num_threads, bool l3_blocking)
343     {
344       this->m_mc = Transpose ? cols : rows;
345       this->m_nc = Transpose ? rows : cols;
346       this->m_kc = depth;
347 
348       if(l3_blocking)
349       {
350         computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc, num_threads);
351       }
352       else  // no l3 blocking
353       {
354         Index n = this->m_nc;
355         computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, n, num_threads);
356       }
357 
358       m_sizeA = this->m_mc * this->m_kc;
359       m_sizeB = this->m_kc * this->m_nc;
360     }
361 
362     void initParallel(Index rows, Index cols, Index depth, Index num_threads)
363     {
364       this->m_mc = Transpose ? cols : rows;
365       this->m_nc = Transpose ? rows : cols;
366       this->m_kc = depth;
367 
368       eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0);
369       Index m = this->m_mc;
370       computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, m, this->m_nc, num_threads);
371       m_sizeA = this->m_mc * this->m_kc;
372       m_sizeB = this->m_kc * this->m_nc;
373     }
374 
375     void allocateA()
376     {
377       if(this->m_blockA==0)
378         this->m_blockA = aligned_new<LhsScalar>(m_sizeA);
379     }
380 
381     void allocateB()
382     {
383       if(this->m_blockB==0)
384         this->m_blockB = aligned_new<RhsScalar>(m_sizeB);
385     }
386 
387     void allocateAll()
388     {
389       allocateA();
390       allocateB();
391     }
392 
393     ~gemm_blocking_space()
394     {
395       aligned_delete(this->m_blockA, m_sizeA);
396       aligned_delete(this->m_blockB, m_sizeB);
397     }
398 };
399 
400 } // end namespace internal
401 
402 namespace internal {
403 
404 template<typename Lhs, typename Rhs>
405 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
406   : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> >
407 {
408   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
409   typedef typename Lhs::Scalar LhsScalar;
410   typedef typename Rhs::Scalar RhsScalar;
411 
412   typedef internal::blas_traits<Lhs> LhsBlasTraits;
413   typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
414   typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
415 
416   typedef internal::blas_traits<Rhs> RhsBlasTraits;
417   typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
418   typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
419 
420   enum {
421     MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
422   };
423 
424   typedef generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> lazyproduct;
425 
426   template<typename Dst>
427   static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
428   {
429     if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
430       lazyproduct::evalTo(dst, lhs, rhs);
431     else
432     {
433       dst.setZero();
434       scaleAndAddTo(dst, lhs, rhs, Scalar(1));
435     }
436   }
437 
438   template<typename Dst>
439   static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
440   {
441     if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
442       lazyproduct::addTo(dst, lhs, rhs);
443     else
444       scaleAndAddTo(dst,lhs, rhs, Scalar(1));
445   }
446 
447   template<typename Dst>
448   static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
449   {
450     if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
451       lazyproduct::subTo(dst, lhs, rhs);
452     else
453       scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
454   }
455 
456   template<typename Dest>
457   static void scaleAndAddTo(Dest& dst, const Lhs& a_lhs, const Rhs& a_rhs, const Scalar& alpha)
458   {
459     eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols());
460     if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0)
461       return;
462 
463     typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
464     typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
465 
466     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
467                                * RhsBlasTraits::extractScalarFactor(a_rhs);
468 
469     typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar,
470             Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
471 
472     typedef internal::gemm_functor<
473       Scalar, Index,
474       internal::general_matrix_matrix_product<
475         Index,
476         LhsScalar, (ActualLhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate),
477         RhsScalar, (ActualRhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate),
478         (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
479       ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor;
480 
481     BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1, true);
482     internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>
483         (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), a_lhs.cols(), Dest::Flags&RowMajorBit);
484   }
485 };
486 
487 } // end namespace internal
488 
489 } // end namespace Eigen
490 
491 #endif // EIGEN_GENERAL_MATRIX_MATRIX_H
492