portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

KNL streaming stores, and KNL performance coutners

Browse Source
This commit is contained in:
azusayamaguchi
2016-10-12 11:45:22 +01:00
parent 2d4a45c758
commit 81f2aeaece
12 changed files with 393 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -416,6 +416,28 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
Kernels::DiracOptDhopSiteDag(st, lo, U, st.comm_buf, sF, sU, LLs, 1, in,
out);
}
#ifdef AVX512
} else if (stat.is_init() ) {
int nthreads;
stat.start();
#pragma omp parallel
{
#pragma omp master
nthreads = omp_get_num_threads();
int mythread = omp_get_thread_num();
stat.enter(mythread);
#pragma omp for nowait
for(int ss=0;ss<U._grid->oSites();ss++)
{
int sU=ss;
int sF=LLs*sU;
Kernels::DiracOptDhopSite(st,lo,U,st.comm_buf,sF,sU,LLs,1,in,out);
}
stat.exit(mythread);
}
stat.accum(nthreads);
#endif
} else {
PARALLEL_FOR_LOOP
for (int ss = 0; ss < U._grid->oSites(); ss++) {

3
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@ -31,6 +31,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_QCD_WILSON_FERMION_5D_H
#define GRID_QCD_WILSON_FERMION_5D_H
#include <Grid/Stat.h>
namespace Grid {
namespace QCD {
@ -60,6 +62,7 @@ namespace Grid {
public:
INHERIT_IMPL_TYPES(Impl);
typedef WilsonKernels<Impl> Kernels;
PmuStat stat;
void Report(void);
void ZeroCounters(void);

4
lib/qcd/action/fermion/WilsonKernelsAsmBody.h
View File

@ -134,7 +134,9 @@
////////////////////////////////
// Xm
////////////////////////////////
#ifndef STREAM_STORE
basep= (uint64_t) &out._odata[ss];
#endif
// basep= st.GetPFInfo(nent,plocal); nent++;
if ( local ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
@ -229,7 +231,9 @@
LOAD_CHI(base);
}
base= (uint64_t) &out._odata[ss];
#ifndef STREAM_STORE
PREFETCH_CHIMU(base);
#endif
{
MULT_2SPIN_DIR_PFTM(Tm,basep);
}