GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-01 17:50:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix various regressions with MKL support

Browse Source 
(transplanted from 8ab0e16e27ea50912ad181ad492b7bd10ffeba98
)
This commit is contained in:
Gael Guennebaud 2012-07-28 16:32:43 +02:00
parent 9f945b6028
commit a03c970c5c
6 changed files with 18 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h
View File

@ -57,11 +57,11 @@ template <typename Index, int Mode, \
struct product_triangular_matrix_matrix<Scalar,Index, Mode, LhsIsTriangular, \
LhsStorageOrder,ConjugateLhs, RhsStorageOrder,ConjugateRhs,ColMajor,Specialized> { \
static inline void run(Index _rows, Index _cols, Index _depth, const Scalar* _lhs, Index lhsStride,\
const Scalar* _rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha) { \
const Scalar* _rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) { \
product_triangular_matrix_matrix_trmm<Scalar,Index,Mode, \
LhsIsTriangular,LhsStorageOrder,ConjugateLhs, \
RhsStorageOrder, ConjugateRhs, ColMajor>::run( \
_rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha); \
_rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
} \
};
@ -96,7 +96,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
const EIGTYPE* _lhs, Index lhsStride, \
const EIGTYPE* _rhs, Index rhsStride, \
EIGTYPE* res, Index resStride, \
EIGTYPE alpha) \
EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
{ \
Index diagSize = (std::min)(_rows,_depth); \
Index rows = IsLower ? _rows : diagSize; \
@ -115,16 +115,16 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
/* Most likely no benefit to call TRMM or GEMM from MKL*/ \
product_triangular_matrix_matrix<EIGTYPE,Index,Mode,true, \
LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
_rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha); \
_rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
/*std::cout << "TRMM_L: A is not square! Go to Eigen TRMM implementation!\n";*/ \
} else { \
/* Make sense to call GEMM */ \
Map<const MatrixLhs, 0, OuterStride<> > lhsMap(_lhs,rows,depth,OuterStride<>(lhsStride)); \
MatrixLhs aa_tmp=lhsMap.template triangularView<Mode>(); \
MKL_INT aStride = aa_tmp.outerStride(); \
gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> blocking(_rows,_cols,_depth); \
gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth); \
general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, resStride, alpha, blocking, 0); \
rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, resStride, alpha, gemm_blocking, 0); \
\
/*std::cout << "TRMM_L: A is not square! Go to MKL GEMM implementation! " << nthr<<" \n";*/ \
} \
@ -210,7 +210,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
const EIGTYPE* _lhs, Index lhsStride, \
const EIGTYPE* _rhs, Index rhsStride, \
EIGTYPE* res, Index resStride, \
EIGTYPE alpha) \
EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
{ \
Index diagSize = (std::min)(_cols,_depth); \
Index rows = _rows; \
@ -229,16 +229,16 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
/* Most likely no benefit to call TRMM or GEMM from MKL*/ \
product_triangular_matrix_matrix<EIGTYPE,Index,Mode,false, \
LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
_rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha); \
_rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
/*std::cout << "TRMM_R: A is not square! Go to Eigen TRMM implementation!\n";*/ \
} else { \
/* Make sense to call GEMM */ \
Map<const MatrixRhs, 0, OuterStride<> > rhsMap(_rhs,depth,cols, OuterStride<>(rhsStride)); \
MatrixRhs aa_tmp=rhsMap.template triangularView<Mode>(); \
MKL_INT aStride = aa_tmp.outerStride(); \
gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> blocking(_rows,_cols,_depth); \
gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth); \
general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, resStride, alpha, blocking, 0); \
rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, resStride, alpha, gemm_blocking, 0); \
\
/*std::cout << "TRMM_R: A is not square! Go to MKL GEMM implementation! " << nthr<<" \n";*/ \
} \

8
Eigen/src/Core/products/TriangularMatrixVector_MKL.h
View File

@ -82,11 +82,11 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
LowUp = IsLower ? Lower : Upper \
}; \
static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \
const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \
{ \
if (ConjLhs || IsZeroDiag) { \
triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor,BuiltIn>::run( \
_rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha, blocking); \
_rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
return; \
}\
Index size = (std::min)(_rows,_cols); \
@ -167,11 +167,11 @@ struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,
LowUp = IsLower ? Lower : Upper \
}; \
static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \
const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \
{ \
if (IsZeroDiag) { \
triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor,BuiltIn>::run( \
_rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha, blocking); \
_rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
return; \
}\
Index size = (std::min)(_rows,_cols); \