devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

simplify and optimize block sizes computation for matrix products. They

Browse Source 
are now automatically computed from the L1 and L2 cache sizes which are
themselves automatically determined at runtime.
This commit is contained in:
Gael Guennebaud
2010-06-21 23:28:50 +02:00
parent 4bac6fbe1e
commit 0212eec23f
3 changed files with 55 additions and 107 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
bench/bench_gemm.cpp
View File

@@ -66,10 +66,12 @@ void gemm(const M& a, const M& b, M& c)
int main(int argc, char ** argv)
{
std::cout << "L1 cache size = " << ei_queryL1CacheSize()/1024 << " KB\n";
std::cout << "L2/L3 cache size = " << ei_queryTopLevelCacheSize()/1024 << " KB\n";
std::cout << "L2/L3 cache size = " << ei_queryTopLevelCacheSize()/1024 << " KB\n";
setCpuCacheSizes(ei_queryL1CacheSize()/1,ei_queryTopLevelCacheSize()/2);
int rep = 1; // number of repetitions per try
int tries = 5; // number of tries, we keep the best
int tries = 2; // number of tries, we keep the best
int s = 2048;
int cache_size = -1;
@@ -102,8 +104,8 @@ int main(int argc, char ** argv)
M c(m,p); c.setOnes();
std::cout << "Matrix sizes = " << m << "x" << p << " * " << p << "x" << n << "\n";
std::ptrdiff_t cm, cn, ck;
getBlockingSizes<Scalar>(ck, cm, cn);
std::ptrdiff_t cm(m), cn(n), ck(p);
computeProductBlockingSizes<Scalar,Scalar>(ck, cm, cn);
std::cout << "blocking size = " << cm << " x " << ck << "\n";
M r = c;