GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-13 12:19:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

Make sure that the block size computation is tested by our unit test.

Browse Source
This commit is contained in:
Gael Guennebaud 2015-02-26 17:00:36 +01:00
parent 2e9cb06a87
commit 4ec3f04b3a
1 changed files with 13 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -153,8 +153,13 @@ void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads
} }
else { else {
// In unit tests we do not want to use extra large matrices, // In unit tests we do not want to use extra large matrices,
// so we reduce the block size to check the blocking strategy is not flawed // so we reduce the cache size to check the blocking strategy is not flawed
#ifndef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS #ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS
l1 = 4*1024;
l2 = 32*1024;
l3 = 512*1024;
#endif
// Early return for small problems because the computation below are time consuming for small problems. // Early return for small problems because the computation below are time consuming for small problems.
// Perhaps it would make more sense to consider k*n*m?? // Perhaps it would make more sense to consider k*n*m??
// Note that for very tiny problem, this function should be bypassed anyway // Note that for very tiny problem, this function should be bypassed anyway
@ -195,7 +200,13 @@ void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads
// actual_l2 = max(l2, l3/nb_core_sharing_l3) // actual_l2 = max(l2, l3/nb_core_sharing_l3)
// The number below is quite conservative: it is better to underestimate the cache size rather than overestimating it) // The number below is quite conservative: it is better to underestimate the cache size rather than overestimating it)
// For instance, it corresponds to 6MB of L3 shared among 4 cores. // For instance, it corresponds to 6MB of L3 shared among 4 cores.
#ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS
const Index actual_l2 = l3;
#else
const Index actual_l2 = 1572864; // == 1.5 MB const Index actual_l2 = 1572864; // == 1.5 MB
#endif
// Here, nc is chosen such that a block of kc x nc of the rhs fit within half of L2. // Here, nc is chosen such that a block of kc x nc of the rhs fit within half of L2.
// The second half is implicitly reserved to access the result and lhs coefficients. // The second half is implicitly reserved to access the result and lhs coefficients.
@ -240,11 +251,6 @@ void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads
m = (m%mc)==0 ? mc m = (m%mc)==0 ? mc
: (mc - Traits::mr * ((mc/*-1*/-(m%mc))/(Traits::mr*(m/mc+1)))); : (mc - Traits::mr * ((mc/*-1*/-(m%mc))/(Traits::mr*(m/mc+1))));
} }
#else
k = std::min<Index>(k,24);
n = std::min<Index>(n,384/sizeof(RhsScalar));
m = std::min<Index>(m,384/sizeof(RhsScalar));
#endif
} }
} }