aboutsummaryrefslogtreecommitdiffhomepage
path: root/Eigen/src/Core/products/TriangularMatrixMatrix.h
diff options
context:
space:
mode:
authorGravatar Gael Guennebaud <g.gael@free.fr>2010-06-22 16:08:35 +0200
committerGravatar Gael Guennebaud <g.gael@free.fr>2010-06-22 16:08:35 +0200
commitb4fe53f5610f20e9c4f15c676b3b26d252fd2fba (patch)
tree9a29de25cae2ee7c0fd177287d95ffc252b44a68 /Eigen/src/Core/products/TriangularMatrixMatrix.h
parentfd9a9fa0aedea13866481309e8a5408d1d7c652b (diff)
* makes all product use the new API to set the blocking sizes
* fix an issue preventing multithreading (now Dynamic = -1 ...)
Diffstat (limited to 'Eigen/src/Core/products/TriangularMatrixMatrix.h')
-rw-r--r--Eigen/src/Core/products/TriangularMatrixMatrix.h18
1 files changed, 13 insertions, 5 deletions
diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h
index a099160c2..decf515b0 100644
--- a/Eigen/src/Core/products/TriangularMatrixMatrix.h
+++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -114,8 +114,12 @@ struct ei_product_triangular_matrix_matrix<Scalar,Index,Mode,true,
IsLower = (Mode&Lower) == Lower
};
- Index kc = std::min<Index>(Blocking::Max_kc/4,depth); // cache block size along the K direction
- Index mc = std::min<Index>(Blocking::Max_mc,rows); // cache block size along the M direction
+ Index kc = depth; // cache block size along the K direction
+ Index mc = rows; // cache block size along the M direction
+ Index nc = cols; // cache block size along the N direction
+ computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
+ // it is better to use smaller blocks along the diagonal
+ kc /= 4;
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols;
@@ -238,8 +242,12 @@ struct ei_product_triangular_matrix_matrix<Scalar,Index,Mode,false,
IsLower = (Mode&Lower) == Lower
};
- Index kc = std::min<Index>(Blocking::Max_kc/4,depth); // cache block size along the K direction
- Index mc = std::min<Index>(Blocking::Max_mc,rows); // cache block size along the M direction
+ Index kc = depth; // cache block size along the K direction
+ Index mc = rows; // cache block size along the M direction
+ Index nc = cols; // cache block size along the N direction
+ computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
+ // it is better to use smaller blocks along the diagonal
+ kc /= 4;
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols;
@@ -273,7 +281,7 @@ struct ei_product_triangular_matrix_matrix<Scalar,Index,Mode,false,
Index rs = IsLower ? std::min(cols,actual_k2) : cols - k2;
// size of the triangular part
Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc;
-
+
Scalar* geb = blockB+ts*ts;
pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, alpha, actual_kc, rs);