portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-16 23:07:05 +01:00
Code Releases Activity

Enable reordering of the loops in the assembler for cache friendly.

Browse Source 
This gets in the way of L2 prefetching however. Do next next link in stencil
prefetching.
This commit is contained in:
paboyle
2016-06-19 11:45:58 -07:00
parent d6737e4bd8
commit 6d58cb2a68
15 changed files with 670 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
lib/qcd/action/fermion/WilsonKernelsHand.cc
View File

@ -312,7 +312,7 @@ namespace QCD {
template<class Impl>
void WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out)
{
@ -555,7 +555,7 @@ void WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeFiel
}
template<class Impl>
void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out)
{
@ -803,7 +803,7 @@ void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeF
// Specialise Gparity to simple implementation
////////////////////////////////////////////////
template<>
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
@ -811,7 +811,7 @@ void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,Dou
}
template<>
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
@ -819,7 +819,7 @@ void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,
}
template<>
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
@ -827,7 +827,7 @@ void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,Dou
}
template<>
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
@ -839,44 +839,44 @@ void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,
////////////// Wilson ; uses this implementation /////////////////////
// Need Nc=3 though //
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);