mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-07 18:27:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

improve block-size heuristic

Browse Source
This commit is contained in:
Gael Guennebaud 2010-07-20 13:23:50 +02:00
parent 4824ac1363
commit d7fa09bf05
1 changed files with 4 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -112,13 +112,14 @@ void computeProductBlockingSizes(std::ptrdiff_t& k, std::ptrdiff_t& m, std::ptrd
// mc x kc blocks A' on the lhs. A' has to fit into L2 cache. Moreover, B' is processed // mc x kc blocks A' on the lhs. A' has to fit into L2 cache. Moreover, B' is processed
// per kc x nr vertical small panels where nr is the blocking size along the n dimension // per kc x nr vertical small panels where nr is the blocking size along the n dimension
// at the register level. For vectorization purpose, these small vertical panels are unpacked, // at the register level. For vectorization purpose, these small vertical panels are unpacked,
// i.e., each coefficient is replicated to fit a packet. This small vertical panel has to // e.g., each coefficient is replicated to fit a packet. This small vertical panel has to
// stay in L1 cache. // stay in L1 cache.
std::ptrdiff_t l1, l2; std::ptrdiff_t l1, l2;
typedef ei_gebp_traits<LhsScalar,RhsScalar> Traits;
enum { enum {
kdiv = KcFactor * 2 * ei_gebp_traits<LhsScalar,RhsScalar>::nr kdiv = KcFactor * 2 * Traits::nr
* ei_packet_traits<RhsScalar>::size * sizeof(RhsScalar), * Traits::RhsProgress * sizeof(RhsScalar),
mr = ei_gebp_traits<LhsScalar,RhsScalar>::mr, mr = ei_gebp_traits<LhsScalar,RhsScalar>::mr,
mr_mask = (0xffffffff/mr)*mr mr_mask = (0xffffffff/mr)*mr
}; };