45#if defined(DSLASHINTRIN)
49#pragma message ("A64FX Dslash: asm")
50#include <simd/Fujitsu_A64FX_asm_single.h>
60#define INTERIOR_AND_EXTERIOR
64#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
67 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
68#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
70#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
73 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
74#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
89#undef INTERIOR_AND_EXTERIOR
93#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
96 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
97#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
99#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
102 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
103#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
118#undef INTERIOR_AND_EXTERIOR
122#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
125 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
126#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
128#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
131 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
132#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
152#define INTERIOR_AND_EXTERIOR
156#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
159 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
160#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
162#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
165 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
166#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
181#undef INTERIOR_AND_EXTERIOR
185#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
188 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
189#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
191#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
194 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
195#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
210#undef INTERIOR_AND_EXTERIOR
214#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
217 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
218#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
220#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
223 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
224#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
247#if defined(DSLASHINTRIN)
250#include <simd/Fujitsu_A64FX_asm_double.h>
259#define INTERIOR_AND_EXTERIOR
267#define INTERIOR_AND_EXTERIOR
271#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
274 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
275#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
277#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
280 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
281#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
296#undef INTERIOR_AND_EXTERIOR
300#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
303 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
304#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
306#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
309 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
310#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
325#undef INTERIOR_AND_EXTERIOR
329#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
332 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
333#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
335#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
338 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
339#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
358#define INTERIOR_AND_EXTERIOR
362#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
365 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
366#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
368#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
371 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
372#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
387#undef INTERIOR_AND_EXTERIOR
391#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
394 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
395#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
397#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
400 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
401#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
416#undef INTERIOR_AND_EXTERIOR
420#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
423 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
424#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
426#pragma GCC optimize ("-O3", "-fno-schedule-insns", "-fno-schedule-insns2")
429 int ss,
int ssU,
int Ls,
int Ns,
const FermionFieldView &in, FermionFieldView &out)
430#include <qcd/action/fermion/implementation/WilsonKernelsAsmBodyA64FX.h>
static INTERNAL_PRECISION U
static void AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf, int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out)
static void AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf, int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out)
static void AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf, int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out)
static void AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf, int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out)
static void AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf, int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out)
static void AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf, int sF, int sU, int Ls, int Nsite, const FermionFieldView &in, FermionFieldView &out)