GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-29 23:34:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

bug #958 - Allow testing specific blocking sizes

Browse Source 
This is only a debugging/testing patch. It allows testing specific
product blocking sizes, typically to study the impact on performance.

Example usage:

int testk, testm, testn;
#define EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
#define EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_K testk
#define EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_M testm
#define EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_N testn
#include <Eigen/Core>
This commit is contained in:
Benoit Jacob 2015-02-18 09:43:55 -05:00
parent c7bb1e8ea8
commit 4a3e6c8be1
1 changed files with 16 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -84,6 +84,22 @@ inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff
template<typename LhsScalar, typename RhsScalar, int KcFactor, typename SizeType> template<typename LhsScalar, typename RhsScalar, int KcFactor, typename SizeType>
void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n, int num_threads) void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n, int num_threads)
{ {
#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
EIGEN_UNUSED_VARIABLE(num_threads);
typedef gebp_traits<LhsScalar,RhsScalar> Traits;
enum {
kr = 16,
mr = Traits::mr,
nr = Traits::nr
};
k = std::min<SizeType>(k, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_K);
if (k > kr) k -= k % kr;
m = std::min<SizeType>(n, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_M);
if (m > mr) m -= m % mr;
n = std::min<SizeType>(k, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_N);
if (n > nr) n -= n % nr;
return;
#endif
// Explanations: // Explanations:
// Let's recall the product algorithms form kc x nc horizontal panels B' on the rhs and // Let's recall the product algorithms form kc x nc horizontal panels B' on the rhs and
// 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