mirror of
https://github.com/paboyle/Grid.git
synced 2025-06-19 08:17:05 +01:00
980 GiB/s Wilson; 680 GiB/s DW (DP)
This commit is contained in:
@ -82,7 +82,11 @@ WilsonKernels<ZWilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
@ -293,157 +297,6 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFie
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
//#undef MAYBEPERM
|
||||
//#undef MULT_2SPIN
|
||||
//#define MAYBEPERM(A,B)
|
||||
//#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
|
||||
|
||||
/////////////////////////////////////////////////////////////////
|
||||
// Ls vectorised, undag Kernel, single
|
||||
/////////////////////////////////////////////////////////////////
|
||||
|
||||
#ifdef DWFVEC5D
|
||||
|
||||
#undef KERNEL_DAG
|
||||
#define INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#define INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#define EXTERIOR
|
||||
#undef MULT_2SPIN
|
||||
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LSNOPF(ptr,pf)
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
/////////////////////////////////////////////////////////////////
|
||||
// Ls vectorised, dag Kernel, single
|
||||
/////////////////////////////////////////////////////////////////
|
||||
#define KERNEL_DAG
|
||||
#define INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#define INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#define EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#endif // VEC 5D
|
||||
|
||||
//#undef COMPLEX_SIGNS
|
||||
//#undef MAYBEPERM
|
||||
//#undef MULT_2SPIN
|
||||
|
||||
// undefine
|
||||
#include <simd/Fujitsu_A64FX_undef.h>
|
||||
@ -478,20 +331,38 @@ WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, Double
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#define INTERIOR
|
||||
@ -499,20 +370,38 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
@ -520,20 +409,38 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldV
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
/////////////////////////////////////////////////////////////////
|
||||
// XYZT vectorised, dag Kernel, double
|
||||
@ -545,20 +452,38 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldV
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#define INTERIOR
|
||||
@ -566,20 +491,38 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldV
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
@ -587,168 +530,40 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFie
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
template<> void
|
||||
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#if defined (WILSONKERNELSASMBODYA64FX)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
|
||||
#else
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
#endif
|
||||
|
||||
|
||||
// KNL stuff
|
||||
#undef MAYBEPERM
|
||||
//#undef MULT_2SPIN
|
||||
#define MAYBEPERM(A,B)
|
||||
//#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
|
||||
|
||||
/////////////////////////////////////////////////////////////////
|
||||
// Ls vectorised, undag Kernel, double
|
||||
/////////////////////////////////////////////////////////////////
|
||||
#ifdef DWFVEC5D
|
||||
|
||||
#undef KERNEL_DAG
|
||||
#define INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#define INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#define EXTERIOR
|
||||
#undef MULT_2SPIN
|
||||
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LSNOPF(ptr,pf)
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
/////////////////////////////////////////////////////////////////
|
||||
// Ls vectorised, dag Kernel, double
|
||||
/////////////////////////////////////////////////////////////////
|
||||
#define KERNEL_DAG
|
||||
#define INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#define INTERIOR
|
||||
#undef EXTERIOR
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#undef INTERIOR_AND_EXTERIOR
|
||||
#undef INTERIOR
|
||||
#define EXTERIOR
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
template<> void
|
||||
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
template<> void
|
||||
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
|
||||
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
|
||||
#include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
|
||||
|
||||
#endif // VEC 5D
|
||||
|
||||
// undefs
|
||||
#undef WILSONKERNELSASMBODYA64FX
|
||||
|
@ -61,6 +61,29 @@ Author: Nils Meyer <nils.meyer@ur.de>
|
||||
#define DIR7_RECON TP_RECON_ACCUM
|
||||
#endif
|
||||
|
||||
//using namespace std;
|
||||
|
||||
#undef SHOW
|
||||
//#define SHOW
|
||||
|
||||
#undef WHERE
|
||||
|
||||
#ifdef INTERIOR_AND_EXTERIOR
|
||||
#define WHERE "INT_AND_EXT"
|
||||
#endif
|
||||
|
||||
#ifdef INTERIOR
|
||||
#define WHERE "INT"
|
||||
#endif
|
||||
|
||||
#ifdef EXTERIOR
|
||||
#define WHERE "EXT"
|
||||
#endif
|
||||
|
||||
//#pragma message("here")
|
||||
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// Comms then compute kernel
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
@ -69,16 +92,17 @@ Author: Nils Meyer <nils.meyer@ur.de>
|
||||
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
|
||||
basep = st.GetPFInfo(nent,plocal); nent++; \
|
||||
if ( local ) { \
|
||||
LOAD64(%r10,isigns); \
|
||||
/* PREFETCH_GAUGE_L1(Dir); slightly worse performance */ \
|
||||
PROJ(base); \
|
||||
/* PREFETCH_GAUGE_L1(Dir); slightly worse performance */ \
|
||||
MAYBEPERM(PERMUTE_DIR,perm); \
|
||||
} else { \
|
||||
LOAD_CHI(base); \
|
||||
} \
|
||||
MULT_2SPIN_DIR_PF(Dir,basep); \
|
||||
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
|
||||
PREFETCH_CHIMU(base); \
|
||||
LOAD64(%r10,isigns); \
|
||||
MULT_2SPIN_DIR_PF(Dir,basep); \
|
||||
PREFETCH_GAUGE_L1(NxtDir); \
|
||||
RECON; \
|
||||
|
||||
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
|
||||
@ -99,17 +123,15 @@ Author: Nils Meyer <nils.meyer@ur.de>
|
||||
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
|
||||
basep = st.GetPFInfo(nent,plocal); nent++; \
|
||||
if ( local ) { \
|
||||
LOAD64(%r10,isigns); \
|
||||
PROJ(base); \
|
||||
MAYBEPERM(PERMUTE_DIR,perm); \
|
||||
}else if ( st.same_node[Dir] ) {LOAD_CHI(base);} \
|
||||
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
|
||||
if ( local || st.same_node[Dir] ) { \
|
||||
MULT_2SPIN_DIR_PF(Dir,basep); \
|
||||
PREFETCH_CHIMU(base); \
|
||||
LOAD64(%r10,isigns); \
|
||||
RECON; \
|
||||
} else { PREFETCH_CHIMU(base); }
|
||||
} \
|
||||
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
|
||||
PREFETCH_CHIMU(base); \
|
||||
|
||||
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
|
||||
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
|
||||
@ -132,7 +154,6 @@ Author: Nils Meyer <nils.meyer@ur.de>
|
||||
if((!local)&&(!st.same_node[Dir]) ) { \
|
||||
LOAD_CHI(base); \
|
||||
MULT_2SPIN_DIR_PF(Dir,base); \
|
||||
LOAD64(%r10,isigns); \
|
||||
RECON; \
|
||||
nmu++; \
|
||||
}
|
||||
@ -144,7 +165,6 @@ Author: Nils Meyer <nils.meyer@ur.de>
|
||||
if((!local)&&(!st.same_node[Dir]) ) { \
|
||||
LOAD_CHI(base); \
|
||||
MULT_2SPIN_DIR_PF(Dir,base); \
|
||||
LOAD64(%r10,isigns); \
|
||||
RECON; \
|
||||
nmu++; \
|
||||
}
|
||||
@ -159,7 +179,6 @@ Author: Nils Meyer <nils.meyer@ur.de>
|
||||
uint64_t basep;
|
||||
const uint64_t plocal =(uint64_t) & in[0];
|
||||
|
||||
COMPLEX_SIGNS(isigns);
|
||||
MASK_REGS;
|
||||
int nmax=U.oSites();
|
||||
for(int site=0;site<Ns;site++) {
|
||||
@ -181,22 +200,137 @@ Author: Nils Meyer <nils.meyer@ur.de>
|
||||
int ent=ss*8;// 2*Ndim
|
||||
int nent=ssn*8;
|
||||
|
||||
uint64_t delta_base, delta_base_p;
|
||||
|
||||
ASM_LEG_XP(Xp,Yp,PERMUTE_DIR3,DIR0_PROJMEM,DIR0_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
float rescale = 64. * 12.;
|
||||
std::cout << "=================================================================" << std::endl;
|
||||
std::cout << "ss = " << ss << " ssn = " << ssn << std::endl;
|
||||
std::cout << "sU = " << sU << " ssU = " << ssU << std::endl;
|
||||
std::cout << " " << std::endl;
|
||||
|
||||
|
||||
std::cout << "Dir = " << Xp << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Xp] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
ASM_LEG(Yp,Zp,PERMUTE_DIR2,DIR1_PROJMEM,DIR1_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = " << Yp << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Yp] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
ASM_LEG(Zp,Tp,PERMUTE_DIR1,DIR2_PROJMEM,DIR2_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = " << Zp << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Zp] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
ASM_LEG(Tp,Xm,PERMUTE_DIR0,DIR3_PROJMEM,DIR3_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = " << Tp << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Tp] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
ASM_LEG(Xm,Ym,PERMUTE_DIR3,DIR4_PROJMEM,DIR4_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = " << Xm << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Xm] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
ASM_LEG(Ym,Zm,PERMUTE_DIR2,DIR5_PROJMEM,DIR5_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = " << Ym << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Ym] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJMEM,DIR6_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = " << Zm << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Zm] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJMEM,DIR7_RECON);
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = " << Tm << " " << WHERE<< std::endl;
|
||||
|
||||
std::cout << "ent nent local perm = " << ent << " " << nent << " " << local << " " << perm << std::endl;
|
||||
std::cout << "st.same_node[Dir] = " << st.same_node[Tm] << std::endl;
|
||||
std::cout << "base = " << (base - plocal)/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
#ifdef EXTERIOR
|
||||
if (nmu==0) break;
|
||||
// if (nmu!=0) std::cout << "EXT "<<sU<<std::endl;
|
||||
#endif
|
||||
base = (uint64_t) &out[ss];
|
||||
basep= st.GetPFInfo(nent,plocal); nent++;
|
||||
|
||||
#ifdef SHOW
|
||||
std::cout << "Dir = FINAL " << WHERE<< std::endl;;
|
||||
|
||||
base_ss = base;
|
||||
std::cout << "base = " << (base - (uint64_t) &out[0])/rescale << std::endl;
|
||||
std::cout << "Basep = " << (basep - plocal)/rescale << std::endl;
|
||||
//printf("U = %llu\n", (uint64_t)&[sU](Dir));
|
||||
std::cout << "----------------------------------------------------" << std::endl;
|
||||
#endif
|
||||
|
||||
//basep = (uint64_t) &out[ssn];
|
||||
RESULT(base,basep);
|
||||
}
|
||||
ssU++;
|
||||
|
Reference in New Issue
Block a user