diff options
Diffstat (limited to 'blas/level3_impl.h')
-rw-r--r-- | blas/level3_impl.h | 305 |
1 files changed, 169 insertions, 136 deletions
diff --git a/blas/level3_impl.h b/blas/level3_impl.h index b2772b190..267a727ef 100644 --- a/blas/level3_impl.h +++ b/blas/level3_impl.h @@ -13,24 +13,29 @@ int EIGEN_BLAS_FUNC(gemm)(char *opa, char *opb, int *m, int *n, int *k, RealScal { // std::cerr << "in gemm " << *opa << " " << *opb << " " << *m << " " << *n << " " << *k << " " << *lda << " " << *ldb << " " << *ldc << " " << *palpha << " " << *pbeta << "\n"; typedef void (*functype)(DenseIndex, DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, Scalar, internal::level3_blocking<Scalar,Scalar>&, Eigen::internal::GemmParallelInfo<DenseIndex>*); - static functype func[12]; - - static bool init = false; - if(!init) - { - for(int i=0; i<12; ++i) - func[i] = 0; - func[NOTR | (NOTR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,ColMajor,false,ColMajor>::run); - func[TR | (NOTR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,false,ColMajor>::run); - func[ADJ | (NOTR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor>::run); - func[NOTR | (TR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,false,ColMajor>::run); - func[TR | (TR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,false,ColMajor>::run); - func[ADJ | (TR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,false,ColMajor>::run); - func[NOTR | (ADJ << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor>::run); - func[TR | (ADJ << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,Conj, ColMajor>::run); - func[ADJ | (ADJ << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,Conj, ColMajor>::run); - init = true; - } + static const functype func[12] = { + // array index: NOTR | (NOTR << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,ColMajor,false,ColMajor>::run), + // array index: TR | (NOTR << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,false,ColMajor>::run), + // array index: ADJ | (NOTR << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor>::run), + 0, + // array index: NOTR | (TR << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,false,ColMajor>::run), + // array index: TR | (TR << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,false,ColMajor>::run), + // array index: ADJ | (TR << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,false,ColMajor>::run), + 0, + // array index: NOTR | (ADJ << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor>::run), + // array index: TR | (ADJ << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,Conj, ColMajor>::run), + // array index: ADJ | (ADJ << 2) + (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,Conj, ColMajor>::run), + 0 + }; Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* b = reinterpret_cast<Scalar*>(pb); @@ -73,49 +78,64 @@ int EIGEN_BLAS_FUNC(trsm)(char *side, char *uplo, char *opa, char *diag, int *m, { // std::cerr << "in trsm " << *side << " " << *uplo << " " << *opa << " " << *diag << " " << *m << "," << *n << " " << *palpha << " " << *lda << " " << *ldb<< "\n"; typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, internal::level3_blocking<Scalar,Scalar>&); - static functype func[32]; - - static bool init = false; - if(!init) - { - for(int i=0; i<32; ++i) - func[i] = 0; - - func[NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,ColMajor,ColMajor>::run); - func[TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,RowMajor,ColMajor>::run); - func[ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, Conj, RowMajor,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,ColMajor,ColMajor>::run); - func[TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,RowMajor,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, Conj, RowMajor,ColMajor>::run); - - func[NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,ColMajor,ColMajor>::run); - func[TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,RowMajor,ColMajor>::run); - func[ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, Conj, RowMajor,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,ColMajor,ColMajor>::run); - func[TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,RowMajor,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, Conj, RowMajor,ColMajor>::run); - - - func[NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,ColMajor,ColMajor>::run); - func[TR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,RowMajor,ColMajor>::run); - func[ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,Conj, RowMajor,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,ColMajor,ColMajor>::run); - func[TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,RowMajor,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,Conj, RowMajor,ColMajor>::run); - - func[NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,ColMajor,ColMajor>::run); - func[TR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,RowMajor,ColMajor>::run); - func[ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,Conj, RowMajor,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,ColMajor,ColMajor>::run); - func[TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,RowMajor,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,Conj, RowMajor,ColMajor>::run); - - init = true; - } + static const functype func[32] = { + // array index: NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,ColMajor,ColMajor>::run), + // array index: TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,RowMajor,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, Conj, RowMajor,ColMajor>::run),\ + 0, + // array index: NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,ColMajor,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,RowMajor,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, Conj, RowMajor,ColMajor>::run), + 0, + // array index: NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,ColMajor,ColMajor>::run), + // array index: TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,RowMajor,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, Conj, RowMajor,ColMajor>::run), + 0, + // array index: NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,ColMajor,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,RowMajor,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, Conj, RowMajor,ColMajor>::run), + 0, + // array index: NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,ColMajor,ColMajor>::run), + // array index: TR | (LEFT << 2) | (UP << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,RowMajor,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,Conj, RowMajor,ColMajor>::run), + 0, + // array index: NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,ColMajor,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,RowMajor,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,Conj, RowMajor,ColMajor>::run), + 0, + // array index: NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,ColMajor,ColMajor>::run), + // array index: TR | (LEFT << 2) | (LO << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,RowMajor,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,Conj, RowMajor,ColMajor>::run), + 0, + // array index: NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,ColMajor,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,RowMajor,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4) + (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,Conj, RowMajor,ColMajor>::run), + 0 + }; Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* b = reinterpret_cast<Scalar*>(pb); @@ -162,47 +182,64 @@ int EIGEN_BLAS_FUNC(trmm)(char *side, char *uplo, char *opa, char *diag, int *m, { // std::cerr << "in trmm " << *side << " " << *uplo << " " << *opa << " " << *diag << " " << *m << " " << *n << " " << *lda << " " << *ldb << " " << *palpha << "\n"; typedef void (*functype)(DenseIndex, DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&); - static functype func[32]; - static bool init = false; - if(!init) - { - for(int k=0; k<32; ++k) - func[k] = 0; - - func[NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,false,ColMajor,false,ColMajor>::run); - func[ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,false,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run); - - func[NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,false,ColMajor,false,ColMajor>::run); - func[ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,false,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run); - - func[NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run); - func[ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run); - - func[NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run); - func[ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run); - - func[NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run); - func[TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run); - func[ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run); - - init = true; - } + static const functype func[32] = { + // array index: NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,false,ColMajor,false,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run), + 0, + // array index: NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,false,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run), + 0, + // array index: NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,false,ColMajor,false,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run), + 0, + // array index: NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,false,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run), + 0, + // array index: NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (LEFT << 2) | (UP << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run), + 0, + // array index: NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run), + 0, + // array index: NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (LEFT << 2) | (LO << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run), + // array index: ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run), + 0, + // array index: NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run), + // array index: TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run), + // array index: ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4) + (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run), + 0 + }; Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* b = reinterpret_cast<Scalar*>(pb); @@ -318,24 +355,22 @@ int EIGEN_BLAS_FUNC(syrk)(char *uplo, char *op, int *n, int *k, RealScalar *palp // std::cerr << "in syrk " << *uplo << " " << *op << " " << *n << " " << *k << " " << *palpha << " " << *lda << " " << *pbeta << " " << *ldc << "\n"; #if !ISCOMPLEX typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&); - static functype func[8]; - - static bool init = false; - if(!init) - { - for(int i=0; i<8; ++i) - func[i] = 0; - - func[NOTR | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Upper>::run); - func[TR | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Upper>::run); - func[ADJ | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Upper>::run); - - func[NOTR | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Lower>::run); - func[TR | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Lower>::run); - func[ADJ | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Lower>::run); - - init = true; - } + static const functype func[8] = { + // array index: NOTR | (UP << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Upper>::run), + // array index: TR | (UP << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Upper>::run), + // array index: ADJ | (UP << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Upper>::run), + 0, + // array index: NOTR | (LO << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Lower>::run), + // array index: TR | (LO << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Lower>::run), + // array index: ADJ | (LO << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Lower>::run), + 0 + }; #endif Scalar* a = reinterpret_cast<Scalar*>(pa); @@ -524,22 +559,20 @@ int EIGEN_BLAS_FUNC(herk)(char *uplo, char *op, int *n, int *k, RealScalar *palp // std::cerr << "in herk " << *uplo << " " << *op << " " << *n << " " << *k << " " << *palpha << " " << *lda << " " << *pbeta << " " << *ldc << "\n"; typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&); - static functype func[8]; - - static bool init = false; - if(!init) - { - for(int i=0; i<8; ++i) - func[i] = 0; - - func[NOTR | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Upper>::run); - func[ADJ | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Upper>::run); - - func[NOTR | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Lower>::run); - func[ADJ | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Lower>::run); - - init = true; - } + static const functype func[8] = { + // array index: NOTR | (UP << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Upper>::run), + 0, + // array index: ADJ | (UP << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Upper>::run), + 0, + // array index: NOTR | (LO << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Lower>::run), + 0, + // array index: ADJ | (LO << 2) + (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Lower>::run), + 0 + }; Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* c = reinterpret_cast<Scalar*>(pc); |