aboutsummaryrefslogtreecommitdiffhomepage
path: root/Eigen/src/Core/products/TriangularMatrixMatrix.h
diff options
context:
space:
mode:
Diffstat (limited to 'Eigen/src/Core/products/TriangularMatrixMatrix.h')
-rw-r--r--Eigen/src/Core/products/TriangularMatrixMatrix.h67
1 files changed, 39 insertions, 28 deletions
diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index c2d0817ea..60c99dcd2 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -108,7 +108,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
Index _rows, Index _cols, Index _depth,
const Scalar* _lhs, Index lhsStride,
const Scalar* _rhs, Index rhsStride,
- Scalar* res, Index resStride,
+ Scalar* _res, Index resStride,
const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
{
// strip zeros
@@ -117,8 +117,12 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
Index depth = IsLower ? diagSize : _depth;
Index cols = _cols;
- const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
- const_blas_data_mapper<Scalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
+ typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+ typedef const_blas_data_mapper<Scalar, 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 = blocking.kc(); // cache block size along the K direction
Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction
@@ -136,9 +140,9 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
else
triangularBuffer.diagonal().setOnes();
- gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
- gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
- gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
+ gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+ gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+ gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
for(Index k2=IsLower ? depth : 0;
IsLower ? k2>0 : k2<depth;
@@ -154,7 +158,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
k2 = k2+actual_kc-kc;
}
- pack_rhs(blockB, &rhs(actual_k2,0), rhsStride, actual_kc, cols);
+ pack_rhs(blockB, rhs.getSubMapper(actual_k2,0), actual_kc, cols);
// the selected lhs's panel has to be split in three different parts:
// 1 - the part which is zero => skip it
@@ -182,9 +186,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
for (Index i=IsLower ? k+1 : 0; IsLower ? i<actualPanelWidth : i<k; ++i)
triangularBuffer.coeffRef(i,k) = lhs(startBlock+i,startBlock+k);
}
- pack_lhs(blockA, triangularBuffer.data(), triangularBuffer.outerStride(), actualPanelWidth, actualPanelWidth);
+ pack_lhs(blockA, LhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()), actualPanelWidth, actualPanelWidth);
- gebp_kernel(res+startBlock, resStride, blockA, blockB, actualPanelWidth, actualPanelWidth, cols, alpha,
+ gebp_kernel(res.getSubMapper(startBlock, 0), blockA, blockB,
+ actualPanelWidth, actualPanelWidth, cols, alpha,
actualPanelWidth, actual_kc, 0, blockBOffset);
// GEBP with remaining micro panel
@@ -192,9 +197,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
{
Index startTarget = IsLower ? actual_k2+k1+actualPanelWidth : actual_k2;
- pack_lhs(blockA, &lhs(startTarget,startBlock), lhsStride, actualPanelWidth, lengthTarget);
+ pack_lhs(blockA, lhs.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget);
- gebp_kernel(res+startTarget, resStride, blockA, blockB, lengthTarget, actualPanelWidth, cols, alpha,
+ gebp_kernel(res.getSubMapper(startTarget, 0), blockA, blockB,
+ lengthTarget, actualPanelWidth, cols, alpha,
actualPanelWidth, actual_kc, 0, blockBOffset);
}
}
@@ -206,10 +212,11 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
for(Index i2=start; i2<end; i2+=mc)
{
const Index actual_mc = (std::min)(i2+mc,end)-i2;
- gemm_pack_lhs<Scalar, Index, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
- (blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc);
+ gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
+ (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
- gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0);
+ gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
+ actual_kc, cols, alpha, -1, -1, 0, 0);
}
}
}
@@ -247,7 +254,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
Index _rows, Index _cols, Index _depth,
const Scalar* _lhs, Index lhsStride,
const Scalar* _rhs, Index rhsStride,
- Scalar* res, Index resStride,
+ Scalar* _res, Index resStride,
const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
{
// strip zeros
@@ -256,8 +263,12 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
Index depth = IsLower ? _depth : diagSize;
Index cols = IsLower ? diagSize : _cols;
- const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride);
- const_blas_data_mapper<Scalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride);
+ typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+ typedef const_blas_data_mapper<Scalar, 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 = blocking.kc(); // cache block size along the K direction
Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction
@@ -275,10 +286,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
else
triangularBuffer.diagonal().setOnes();
- gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
- gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
- gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
- gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
+ gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+ gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+ gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+ gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
for(Index k2=IsLower ? 0 : depth;
IsLower ? k2<depth : k2>0;
@@ -302,7 +313,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
Scalar* geb = blockB+ts*ts;
geb = geb + internal::first_aligned(geb,EIGEN_ALIGN_BYTES/sizeof(Scalar));
- pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, actual_kc, rs);
+ pack_rhs(geb, rhs.getSubMapper(actual_k2,IsLower ? 0 : k2), actual_kc, rs);
// pack the triangular part of the rhs padding the unrolled blocks with zeros
if(ts>0)
@@ -315,7 +326,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2;
// general part
pack_rhs_panel(blockB+j2*actual_kc,
- &rhs(actual_k2+panelOffset, actual_j2), rhsStride,
+ rhs.getSubMapper(actual_k2+panelOffset, actual_j2),
panelLength, actualPanelWidth,
actual_kc, panelOffset);
@@ -329,7 +340,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
}
pack_rhs_panel(blockB+j2*actual_kc,
- triangularBuffer.data(), triangularBuffer.outerStride(),
+ RhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()),
actualPanelWidth, actualPanelWidth,
actual_kc, j2);
}
@@ -338,7 +349,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
for (Index i2=0; i2<rows; i2+=mc)
{
const Index actual_mc = (std::min)(mc,rows-i2);
- pack_lhs(blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc);
+ pack_lhs(blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
// triangular kernel
if(ts>0)
@@ -349,7 +360,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
Index panelLength = IsLower ? actual_kc-j2 : j2+actualPanelWidth;
Index blockOffset = IsLower ? j2 : 0;
- gebp_kernel(res+i2+(actual_k2+j2)*resStride, resStride,
+ gebp_kernel(res.getSubMapper(i2, actual_k2 + j2),
blockA, blockB+j2*actual_kc,
actual_mc, panelLength, actualPanelWidth,
alpha,
@@ -357,7 +368,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
blockOffset, blockOffset);// offsets
}
}
- gebp_kernel(res+i2+(IsLower ? 0 : k2)*resStride, resStride,
+ gebp_kernel(res.getSubMapper(i2, IsLower ? 0 : k2),
blockA, geb, actual_mc, actual_kc, rs,
alpha,
-1, -1, 0, 0);
@@ -402,7 +413,7 @@ struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows()))
: ((IsLower) ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols()));
- BlockingType blocking(stripedRows, stripedCols, stripedDepth);
+ BlockingType blocking(stripedRows, stripedCols, stripedDepth, 1, false);
internal::product_triangular_matrix_matrix<Scalar, Index,
Mode, LhsIsTriangular,
generated by cgit v1.2.3 (git 2.39.1) at 2024-05-27 23:28:44 +0000