portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-20 00:36:55 +01:00
Code Releases Activity

up to +36% performance gain for dslash/dwf on QPACE 4 using GCC 10.1.1

Browse Source
This commit is contained in:
Nils Meyer
2020-12-19 00:54:31 +01:00
parent 3362f8dfa0
commit 4dd9e39e0d
8 changed files with 447 additions and 2920 deletions
Download Patch File Download Diff File Expand all files Collapse all files

268
Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h
View File

@ -38,9 +38,6 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
// undefine everything related to kernels
#include <simd/Fujitsu_A64FX_undef.h>
// enable A64FX body
#define WILSONKERNELSASMBODYA64FX
//#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h")
///////////////////////////////////////////////////////////
// If we are A64FX specialise the single precision routine
@ -63,119 +60,89 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplFH>::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
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplFH>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplFH>::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
#define EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplFH>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplFH>::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
/////////////////////////////////////////////////////////////////
@ -185,119 +152,89 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldV
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplFH>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplFH>::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
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplFH>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplFH>::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
#define EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplF>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<WilsonImplFH>::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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
template<> void
WilsonKernels<ZWilsonImplFH>::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
// undefine
@ -330,119 +267,89 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFie
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#define EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
@ -451,124 +358,93 @@ WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldV
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#undef EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#define EXTERIOR
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
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
// undefs
#undef WILSONKERNELSASMBODYA64FX
#include <simd/Fujitsu_A64FX_undef.h>
#endif //A64FXASM

105
Grid/qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h
View File

@ -25,6 +25,11 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
// GCC 10 messes up SVE instruction scheduling using -O3 only,
// using -O3 -fno-schedule-insns -fno-schedule-insns2 does wonders
// performance is better than armclang 20.2
#ifdef KERNEL_DAG
#define DIR0_PROJ XP_PROJ
#define DIR1_PROJ YP_PROJ
@ -97,7 +102,7 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
PROJ; \
MAYBEPERM(PERMUTE_DIR,perm); \
} else { \
LOAD_CHI(base); \
LOAD_CHI(base); \
} \
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
MULT_2SPIN_1(Dir); \
@ -110,6 +115,15 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
} \
RECON; \
/*
NB: picking PREFETCH_GAUGE_L2(Dir+4); here results in performance penalty
though I expected that it would improve on performance
if (s == 0) { \
if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
} \
*/
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
PREFETCH1_CHIMU(base); \
@ -126,73 +140,63 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
basep = st.GetPFInfo(nent,plocal); nent++; \
if ( local ) { \
LOAD_CHIMU(base); \
LOAD_TABLE(PERMUTE_DIR); \
PROJ; \
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_1(Dir); \
PREFETCH_CHIMU(base); \
/* PREFETCH_GAUGE_L1(NxtDir); */ \
MULT_2SPIN_2; \
if (s == 0) { \
if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
} \
RECON; \
PREFETCH_CHIMU_L2(basep); \
} else { PREFETCH_CHIMU(base); } \
if ( local ) { \
LOAD_CHIMU(base); \
LOAD_TABLE(PERMUTE_DIR); \
PROJ; \
MAYBEPERM(PERMUTE_DIR,perm); \
}else if ( st.same_node[Dir] ) {LOAD_CHI(base);} \
if ( local || st.same_node[Dir] ) { \
MULT_2SPIN_1(Dir); \
MULT_2SPIN_2; \
RECON; \
} \
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
PREFETCH_CHIMU(base); \
PREFETCH_CHIMU_L2(basep); \
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
PREFETCH1_CHIMU(base); \
{ ZERO_PSI; } \
ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
#define RESULT(base,basep) SAVE_RESULT(base,basep);
#endif
////////////////////////////////////////////////////////////////////////////////
// Post comms kernel
////////////////////////////////////////////////////////////////////////////////
#ifdef EXTERIOR
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \
MULT_2SPIN_1(Dir); \
PREFETCH_CHIMU(base); \
/* PREFETCH_GAUGE_L1(NxtDir); */ \
MULT_2SPIN_2; \
if (s == 0) { \
if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
} \
RECON; \
nmu++; \
RECON; \
nmu++; \
}
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
nmu=0; \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++;\
if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
nmu=0; \
{ ZERO_PSI;} \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \
MULT_2SPIN_1(Dir); \
PREFETCH_CHIMU(base); \
/* PREFETCH_GAUGE_L1(NxtDir); */ \
MULT_2SPIN_2; \
if (s == 0) { \
if ((Dir == 0) || (Dir == 4)) { PREFETCH_GAUGE_L2(Dir); } \
} \
RECON; \
nmu++; \
RECON; \
nmu++; \
}
#define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
#endif
{
int nmu;
int local,perm, ptype;
@ -209,7 +213,6 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
int ssn=ssU+1; if(ssn>=nmax) ssn=0;
// int sUn=lo.Reorder(ssn);
int sUn=ssn;
LOCK_GAUGE(0);
#else
int sU =ssU;
int ssn=ssU+1; if(ssn>=nmax) ssn=0;
@ -295,6 +298,11 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
std::cout << "----------------------------------------------------" << std::endl;
#endif
// DC ZVA test
// { uint64_t basestore = (uint64_t)&out[ss];
// PREFETCH_RESULT_L2_STORE(basestore); }
ASM_LEG(Ym,Zm,PERMUTE_DIR2,DIR5_PROJ,DIR5_RECON);
#ifdef SHOW
@ -308,6 +316,11 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
std::cout << "----------------------------------------------------" << std::endl;
#endif
// DC ZVA test
//{ uint64_t basestore = (uint64_t)&out[ss];
// PREFETCH_RESULT_L2_STORE(basestore); }
ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJ,DIR6_RECON);
#ifdef SHOW
@ -321,6 +334,11 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
std::cout << "----------------------------------------------------" << std::endl;
#endif
// DC ZVA test
//{ uint64_t basestore = (uint64_t)&out[ss];
// PREFETCH_RESULT_L2_STORE(basestore);
//}
ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJ,DIR7_RECON);
#ifdef SHOW
@ -341,6 +359,7 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
base = (uint64_t) &out[ss];
basep= st.GetPFInfo(nent,plocal); ent++;
basep = (uint64_t) &out[ssn];
//PREFETCH_RESULT_L1_STORE(base);
RESULT(base,basep);
#ifdef SHOW