diff options
Diffstat (limited to 'Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h')
| -rw-r--r-- | Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h | 56 |
1 files changed, 32 insertions, 24 deletions
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 7db3e3d38..e55994900 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -58,27 +58,31 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder, { typedef typename scalar_product_traits<LhsScalar, RhsScalar>::ReturnType ResScalar; static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* _lhs, Index lhsStride, - const RhsScalar* _rhs, Index rhsStride, ResScalar* res, Index resStride, const ResScalar& alpha) + const RhsScalar* _rhs, Index rhsStride, ResScalar* _res, Index resStride, const ResScalar& alpha) { - const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride); - const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride); - typedef gebp_traits<LhsScalar,RhsScalar> Traits; + typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper; + typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper; + typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + Index kc = depth; // cache block size along the K direction Index mc = size; // cache block size along the M direction Index nc = size; // cache block size along the N direction - computeProductBlockingSizes<LhsScalar,RhsScalar>(kc, mc, nc); + computeProductBlockingSizes<LhsScalar,RhsScalar>(kc, mc, nc, 1); // !!! mc must be a multiple of nr: if(mc > Traits::nr) mc = (mc/Traits::nr)*Traits::nr; ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, kc*mc, 0); ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, kc*size, 0); - - gemm_pack_lhs<LhsScalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; - gemm_pack_rhs<RhsScalar, Index, Traits::nr, RhsStorageOrder> pack_rhs; - gebp_kernel <LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp; + + gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; + gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs; + gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp; tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb; for(Index k2=0; k2<depth; k2+=kc) @@ -86,29 +90,30 @@ struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder, const Index actual_kc = (std::min)(k2+kc,depth)-k2; // note that the actual rhs is the transpose/adjoint of mat - pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, size); + pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, size); for(Index i2=0; i2<size; i2+=mc) { const Index actual_mc = (std::min)(i2+mc,size)-i2; - pack_lhs(blockA, &lhs(i2, k2), lhsStride, actual_kc, actual_mc); + pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc); // the selected actual_mc * size panel of res is split into three different part: // 1 - before the diagonal => processed with gebp or skipped // 2 - the actual_mc x actual_mc symmetric block => processed with a special kernel // 3 - after the diagonal => processed with gebp or skipped if (UpLo==Lower) - gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, (std::min)(size,i2), alpha, - -1, -1, 0, 0); + gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, + (std::min)(size,i2), alpha, -1, -1, 0, 0); + - sybb(res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha); + sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha); if (UpLo==Upper) { Index j2 = i2+actual_mc; - gebp(res+resStride*j2+i2, resStride, blockA, blockB+actual_kc*j2, actual_mc, actual_kc, (std::max)(Index(0), size-j2), alpha, - -1, -1, 0, 0); + gebp(res.getSubMapper(i2, j2), blockA, blockB+actual_kc*j2, actual_mc, + actual_kc, (std::max)(Index(0), size-j2), alpha, -1, -1, 0, 0); } } } @@ -129,13 +134,16 @@ struct tribb_kernel { typedef gebp_traits<LhsScalar,RhsScalar,ConjLhs,ConjRhs> Traits; typedef typename Traits::ResScalar ResScalar; - + enum { BlockSize = EIGEN_PLAIN_ENUM_MAX(mr,nr) }; - void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha) + void operator()(ResScalar* _res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha) { - gebp_kernel<LhsScalar, RhsScalar, Index, mr, nr, ConjLhs, ConjRhs> gebp_kernel; + typedef blas_data_mapper<ResScalar, Index, ColMajor> ResMapper; + ResMapper res(_res, resStride); + gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel; + Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer; // let's process the block per panel of actual_mc x BlockSize, @@ -146,7 +154,7 @@ struct tribb_kernel const RhsScalar* actual_b = blockB+j*depth; if(UpLo==Upper) - gebp_kernel(res+j*resStride, resStride, blockA, actual_b, j, depth, actualBlockSize, alpha, + gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha, -1, -1, 0, 0); // selfadjoint micro block @@ -154,12 +162,12 @@ struct tribb_kernel Index i = j; buffer.setZero(); // 1 - apply the kernel on the temporary buffer - gebp_kernel(buffer.data(), BlockSize, blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha, + gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha, -1, -1, 0, 0); // 2 - triangular accumulation for(Index j1=0; j1<actualBlockSize; ++j1) { - ResScalar* r = res + (j+j1)*resStride + i; + ResScalar* r = &res(i, j + j1); for(Index i1=UpLo==Lower ? j1 : 0; UpLo==Lower ? i1<actualBlockSize : i1<=j1; ++i1) r[i1] += buffer(i1,j1); @@ -169,8 +177,8 @@ struct tribb_kernel if(UpLo==Lower) { Index i = j+actualBlockSize; - gebp_kernel(res+j*resStride+i, resStride, blockA+depth*i, actual_b, size-i, depth, actualBlockSize, alpha, - -1, -1, 0, 0); + gebp_kernel(res.getSubMapper(i, j), blockA+depth*i, actual_b, size-i, + depth, actualBlockSize, alpha, -1, -1, 0, 0); } } } |