10 #ifndef EIGEN_GENERAL_MATRIX_MATRIX_H 11 #define EIGEN_GENERAL_MATRIX_MATRIX_H 17 template<
typename _LhsScalar,
typename _RhsScalar>
class level3_blocking;
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>
26 typedef gebp_traits<RhsScalar,LhsScalar> Traits;
28 typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
29 static EIGEN_STRONG_INLINE
void run(
31 const LhsScalar* lhs,
Index lhsStride,
32 const RhsScalar* rhs,
Index rhsStride,
33 ResScalar* res,
Index resStride,
35 level3_blocking<RhsScalar,LhsScalar>& blocking,
36 GemmParallelInfo<Index>* info = 0)
39 general_matrix_matrix_product<
Index,
43 ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking,info);
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>
56 typedef gebp_traits<LhsScalar,RhsScalar> Traits;
58 typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
60 const LhsScalar* _lhs,
Index lhsStride,
61 const RhsScalar* _rhs,
Index rhsStride,
62 ResScalar* _res,
Index resStride,
64 level3_blocking<LhsScalar,RhsScalar>& blocking,
65 GemmParallelInfo<Index>* info = 0)
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);
74 Index kc = blocking.kc();
75 Index mc = (std::min)(rows,blocking.mc());
76 Index nc = (std::min)(cols,blocking.nc());
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;
82 #ifdef EIGEN_HAS_OPENMP 86 Index tid = omp_get_thread_num();
87 Index threads = omp_get_num_threads();
89 LhsScalar* blockA = blocking.blockA();
90 eigen_internal_assert(blockA!=0);
92 std::size_t sizeB = kc*nc;
93 ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, 0);
96 for(
Index k=0; k<depth; k+=kc)
98 const Index actual_kc = (std::min)(k+kc,depth)-k;
102 pack_rhs(blockB, rhs.getSubMapper(k,0), actual_kc, nc);
110 while(info[tid].users!=0) {}
111 info[tid].users += threads;
113 pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
119 for(
Index shift=0; shift<threads; ++shift)
121 Index i = (tid+shift)%threads;
127 while(info[i].sync!=k) {
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);
135 for(
Index j=nc; j<cols; j+=nc)
137 const Index actual_nc = (std::min)(j+nc,cols)-j;
140 pack_rhs(blockB, rhs.getSubMapper(k,j), actual_kc, actual_nc);
143 gebp(res.getSubMapper(0, j), blockA, blockB, rows, actual_kc, actual_nc, alpha);
148 for(
Index i=0; i<threads; ++i)
154 #endif // EIGEN_HAS_OPENMP 156 EIGEN_UNUSED_VARIABLE(info);
159 std::size_t sizeA = kc*mc;
160 std::size_t sizeB = kc*nc;
162 ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
163 ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
165 const bool pack_rhs_once = mc!=rows && kc==depth && nc==cols;
168 for(
Index i2=0; i2<rows; i2+=mc)
170 const Index actual_mc = (std::min)(i2+mc,rows)-i2;
172 for(
Index k2=0; k2<depth; k2+=kc)
174 const Index actual_kc = (std::min)(k2+kc,depth)-k2;
180 pack_lhs(blockA, lhs.getSubMapper(i2,k2), actual_kc, actual_mc);
183 for(
Index j2=0; j2<cols; j2+=nc)
185 const Index actual_nc = (std::min)(j2+nc,cols)-j2;
190 if((!pack_rhs_once) || i2==0)
191 pack_rhs(blockB, rhs.getSubMapper(k2,j2), actual_kc, actual_nc);
194 gebp(res.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, alpha);
208 template<
typename Scalar,
typename Index,
typename Gemm,
typename Lhs,
typename Rhs,
typename Dest,
typename BlockingType>
211 gemm_functor(
const Lhs& lhs,
const Rhs& rhs, Dest& dest,
const Scalar& actualAlpha, BlockingType& blocking)
212 : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking)
215 void initParallelSession(
Index num_threads)
const 217 m_blocking.initParallel(m_lhs.rows(), m_rhs.cols(), m_lhs.cols(), num_threads);
218 m_blocking.allocateA();
221 void operator() (
Index row,
Index rows,
Index col=0,
Index cols=-1, GemmParallelInfo<Index>* info=0)
const 226 Gemm::run(rows, cols, m_lhs.cols(),
227 &m_lhs.coeffRef(row,0), m_lhs.outerStride(),
228 &m_rhs.coeffRef(0,col), m_rhs.outerStride(),
229 (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
230 m_actualAlpha, m_blocking, info);
233 typedef typename Gemm::Traits Traits;
239 Scalar m_actualAlpha;
240 BlockingType& m_blocking;
243 template<
int StorageOrder,
typename LhsScalar,
typename RhsScalar,
int MaxRows,
int MaxCols,
int MaxDepth,
int KcFactor=1,
244 bool FiniteAtCompileTime = MaxRows!=
Dynamic && MaxCols!=
Dynamic && MaxDepth !=
Dynamic>
class gemm_blocking_space;
246 template<
typename _LhsScalar,
typename _RhsScalar>
247 class level3_blocking
249 typedef _LhsScalar LhsScalar;
250 typedef _RhsScalar RhsScalar;
263 : m_blockA(0), m_blockB(0), m_mc(0), m_nc(0), m_kc(0)
266 inline Index mc()
const {
return m_mc; }
267 inline Index nc()
const {
return m_nc; }
268 inline Index kc()
const {
return m_kc; }
270 inline LhsScalar* blockA() {
return m_blockA; }
271 inline RhsScalar* blockB() {
return m_blockB; }
274 template<
int StorageOrder,
typename _LhsScalar,
typename _RhsScalar,
int MaxRows,
int MaxCols,
int MaxDepth,
int KcFactor>
275 class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true >
276 :
public level3_blocking<
277 typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
278 typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
282 ActualRows = Transpose ? MaxCols : MaxRows,
283 ActualCols = Transpose ? MaxRows : MaxCols
285 typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
286 typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
287 typedef gebp_traits<LhsScalar,RhsScalar> Traits;
289 SizeA = ActualRows * MaxDepth,
290 SizeB = ActualCols * MaxDepth
293 #if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES 294 EIGEN_ALIGN_MAX LhsScalar m_staticA[SizeA];
295 EIGEN_ALIGN_MAX RhsScalar m_staticB[SizeB];
297 EIGEN_ALIGN_MAX
char m_staticA[SizeA *
sizeof(LhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
298 EIGEN_ALIGN_MAX
char m_staticB[SizeB *
sizeof(RhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
305 this->m_mc = ActualRows;
306 this->m_nc = ActualCols;
307 this->m_kc = MaxDepth;
308 #if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES 309 this->m_blockA = m_staticA;
310 this->m_blockB = m_staticB;
312 this->m_blockA =
reinterpret_cast<LhsScalar*
>((internal::UIntPtr(m_staticA) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~
std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
313 this->m_blockB =
reinterpret_cast<RhsScalar*
>((internal::UIntPtr(m_staticB) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~
std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
320 inline void allocateA() {}
321 inline void allocateB() {}
322 inline void allocateAll() {}
325 template<
int StorageOrder,
typename _LhsScalar,
typename _RhsScalar,
int MaxRows,
int MaxCols,
int MaxDepth,
int KcFactor>
326 class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, false>
327 :
public level3_blocking<
328 typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
329 typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
334 typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
335 typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
336 typedef gebp_traits<LhsScalar,RhsScalar> Traits;
345 this->m_mc = Transpose ? cols : rows;
346 this->m_nc = Transpose ? rows : cols;
351 computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc, num_threads);
355 Index n = this->m_nc;
356 computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, n, num_threads);
359 m_sizeA = this->m_mc * this->m_kc;
360 m_sizeB = this->m_kc * this->m_nc;
365 this->m_mc = Transpose ? cols : rows;
366 this->m_nc = Transpose ? rows : cols;
369 eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0);
370 Index m = this->m_mc;
371 computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, m, this->m_nc, num_threads);
372 m_sizeA = this->m_mc * this->m_kc;
373 m_sizeB = this->m_kc * this->m_nc;
378 if(this->m_blockA==0)
379 this->m_blockA = aligned_new<LhsScalar>(m_sizeA);
384 if(this->m_blockB==0)
385 this->m_blockB = aligned_new<RhsScalar>(m_sizeB);
394 ~gemm_blocking_space()
396 aligned_delete(this->m_blockA, m_sizeA);
397 aligned_delete(this->m_blockB, m_sizeB);
405 template<
typename Lhs,
typename Rhs>
406 struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
407 : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> >
409 typedef typename Product<Lhs,Rhs>::Scalar Scalar;
410 typedef typename Lhs::Scalar LhsScalar;
411 typedef typename Rhs::Scalar RhsScalar;
413 typedef internal::blas_traits<Lhs> LhsBlasTraits;
414 typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
415 typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
417 typedef internal::blas_traits<Rhs> RhsBlasTraits;
418 typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
419 typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
422 MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
425 typedef generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> lazyproduct;
427 template<
typename Dst>
428 static void evalTo(Dst& dst,
const Lhs& lhs,
const Rhs& rhs)
430 if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
431 lazyproduct::evalTo(dst, lhs, rhs);
435 scaleAndAddTo(dst, lhs, rhs, Scalar(1));
439 template<
typename Dst>
440 static void addTo(Dst& dst,
const Lhs& lhs,
const Rhs& rhs)
442 if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
443 lazyproduct::addTo(dst, lhs, rhs);
445 scaleAndAddTo(dst,lhs, rhs, Scalar(1));
448 template<
typename Dst>
449 static void subTo(Dst& dst,
const Lhs& lhs,
const Rhs& rhs)
451 if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
452 lazyproduct::subTo(dst, lhs, rhs);
454 scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
457 template<
typename Dest>
458 static void scaleAndAddTo(Dest& dst,
const Lhs& a_lhs,
const Rhs& a_rhs,
const Scalar& alpha)
460 eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols());
461 if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0)
464 typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
465 typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
467 Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
468 * RhsBlasTraits::extractScalarFactor(a_rhs);
471 Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
473 typedef internal::gemm_functor<
475 internal::general_matrix_matrix_product<
477 LhsScalar, (ActualLhsTypeCleaned::Flags&
RowMajorBit) ?
RowMajor : ColMajor,
bool(LhsBlasTraits::NeedToConjugate),
478 RhsScalar, (ActualRhsTypeCleaned::Flags&
RowMajorBit) ?
RowMajor : ColMajor,
bool(RhsBlasTraits::NeedToConjugate),
480 ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor;
482 BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1,
true);
483 internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==
Dynamic)>
484 (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), Dest::Flags&
RowMajorBit);
492 #endif // EIGEN_GENERAL_MATRIX_MATRIX_H void initParallel()
Definition: Parallelizer.h:48
Definition: Constants.h:320
Namespace containing all symbols from the Eigen library.
Definition: Core:271
const unsigned int RowMajorBit
Definition: Constants.h:61
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition: XprHelper.h:35
Definition: Eigen_Colamd.h:50
Definition: Constants.h:322
const int Dynamic
Definition: Constants.h:21