mirror of
https://github.com/paboyle/Grid.git
synced 2025-04-25 13:15:55 +01:00
Added hand-unrolled kernel implementations of all the other dslash precision / comms precision combinations with G-parity
This commit is contained in:
Christopher Kelly
parent
46f88e6d72
commit
a0bb8e5b46
@ -786,120 +786,113 @@ void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,D
|
|||||||
const FermionField &in, \
|
const FermionField &in, \
|
||||||
FermionField &out){ assert(0); } \
|
FermionField &out){ assert(0); } \
|
||||||
|
|
||||||
HAND_SPECIALISE_EMPTY(GparityWilsonImplF);
|
|
||||||
//HAND_SPECIALISE_EMPTY(GparityWilsonImplD);
|
|
||||||
HAND_SPECIALISE_EMPTY(GparityWilsonImplFH);
|
|
||||||
HAND_SPECIALISE_EMPTY(GparityWilsonImplDF);
|
|
||||||
|
|
||||||
|
|
||||||
|
#define HAND_SPECIALISE_GPARITY(IMPL) \
|
||||||
|
template<> void \
|
||||||
|
WilsonKernels<IMPL>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
|
||||||
template<> void
|
int ss,int sU,const FermionField &in, FermionField &out) \
|
||||||
WilsonKernels<GparityWilsonImplD>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
|
{ \
|
||||||
int ss,int sU,const FermionField &in, FermionField &out)
|
typedef IMPL Impl; \
|
||||||
{
|
typedef typename Simd::scalar_type S; \
|
||||||
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
|
typedef typename Simd::vector_type V; \
|
||||||
typedef GparityWilsonImplD Impl;
|
\
|
||||||
typedef typename Simd::scalar_type S;
|
HAND_DECLARATIONS(ignore); \
|
||||||
typedef typename Simd::vector_type V;
|
\
|
||||||
|
int offset,local,perm, ptype; \
|
||||||
HAND_DECLARATIONS(ignore);
|
StencilEntry *SE; \
|
||||||
|
HAND_DOP_SITE(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
int offset,local,perm, ptype;
|
HAND_DOP_SITE(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
StencilEntry *SE;
|
} \
|
||||||
HAND_DOP_SITE(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
\
|
||||||
HAND_DOP_SITE(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
template<> \
|
||||||
}
|
void WilsonKernels<IMPL>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
|
||||||
|
int ss,int sU,const FermionField &in, FermionField &out) \
|
||||||
template<>
|
{ \
|
||||||
void WilsonKernels<GparityWilsonImplD>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
|
typedef IMPL Impl; \
|
||||||
int ss,int sU,const FermionField &in, FermionField &out)
|
typedef typename Simd::scalar_type S; \
|
||||||
{
|
typedef typename Simd::vector_type V; \
|
||||||
typedef GparityWilsonImplD Impl;
|
\
|
||||||
typedef typename Simd::scalar_type S;
|
HAND_DECLARATIONS(ignore); \
|
||||||
typedef typename Simd::vector_type V;
|
\
|
||||||
|
StencilEntry *SE; \
|
||||||
HAND_DECLARATIONS(ignore);
|
int offset,local,perm, ptype; \
|
||||||
|
HAND_DOP_SITE_DAG(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
StencilEntry *SE;
|
HAND_DOP_SITE_DAG(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
int offset,local,perm, ptype;
|
} \
|
||||||
HAND_DOP_SITE_DAG(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
\
|
||||||
HAND_DOP_SITE_DAG(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
template<> void \
|
||||||
}
|
WilsonKernels<IMPL>::HandDhopSiteInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
|
||||||
|
int ss,int sU,const FermionField &in, FermionField &out) \
|
||||||
template<> void
|
{ \
|
||||||
WilsonKernels<GparityWilsonImplD>::HandDhopSiteInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
|
typedef IMPL Impl; \
|
||||||
int ss,int sU,const FermionField &in, FermionField &out)
|
typedef typename Simd::scalar_type S; \
|
||||||
{
|
typedef typename Simd::vector_type V; \
|
||||||
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
|
\
|
||||||
typedef GparityWilsonImplD Impl;
|
HAND_DECLARATIONS(ignore); \
|
||||||
typedef typename Simd::scalar_type S;
|
\
|
||||||
typedef typename Simd::vector_type V;
|
int offset,local,perm, ptype; \
|
||||||
|
StencilEntry *SE; \
|
||||||
HAND_DECLARATIONS(ignore);
|
HAND_DOP_SITE_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
|
HAND_DOP_SITE_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
int offset,local,perm, ptype;
|
} \
|
||||||
StencilEntry *SE;
|
\
|
||||||
HAND_DOP_SITE_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
template<> \
|
||||||
HAND_DOP_SITE_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
void WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
|
||||||
}
|
int ss,int sU,const FermionField &in, FermionField &out) \
|
||||||
|
{ \
|
||||||
template<>
|
typedef IMPL Impl; \
|
||||||
void WilsonKernels<GparityWilsonImplD>::HandDhopSiteDagInt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
|
typedef typename Simd::scalar_type S; \
|
||||||
int ss,int sU,const FermionField &in, FermionField &out)
|
typedef typename Simd::vector_type V; \
|
||||||
{
|
\
|
||||||
typedef GparityWilsonImplD Impl;
|
HAND_DECLARATIONS(ignore); \
|
||||||
typedef typename Simd::scalar_type S;
|
\
|
||||||
typedef typename Simd::vector_type V;
|
StencilEntry *SE; \
|
||||||
|
int offset,local,perm, ptype; \
|
||||||
HAND_DECLARATIONS(ignore);
|
HAND_DOP_SITE_DAG_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
|
HAND_DOP_SITE_DAG_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
StencilEntry *SE;
|
} \
|
||||||
int offset,local,perm, ptype;
|
\
|
||||||
HAND_DOP_SITE_DAG_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
template<> void \
|
||||||
HAND_DOP_SITE_DAG_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
WilsonKernels<IMPL>::HandDhopSiteExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
|
||||||
}
|
int ss,int sU,const FermionField &in, FermionField &out) \
|
||||||
|
{ \
|
||||||
template<> void
|
typedef IMPL Impl; \
|
||||||
WilsonKernels<GparityWilsonImplD>::HandDhopSiteExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
|
typedef typename Simd::scalar_type S; \
|
||||||
int ss,int sU,const FermionField &in, FermionField &out)
|
typedef typename Simd::vector_type V; \
|
||||||
{
|
\
|
||||||
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
|
HAND_DECLARATIONS(ignore); \
|
||||||
typedef GparityWilsonImplD Impl;
|
\
|
||||||
typedef typename Simd::scalar_type S;
|
int offset,local,perm, ptype; \
|
||||||
typedef typename Simd::vector_type V;
|
StencilEntry *SE; \
|
||||||
|
int nmu=0; \
|
||||||
HAND_DECLARATIONS(ignore);
|
HAND_DOP_SITE_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
|
nmu = 0; \
|
||||||
int offset,local,perm, ptype;
|
HAND_DOP_SITE_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
StencilEntry *SE;
|
} \
|
||||||
int nmu=0;
|
template<> \
|
||||||
HAND_DOP_SITE_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
void WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
|
||||||
nmu = 0;
|
int ss,int sU,const FermionField &in, FermionField &out) \
|
||||||
HAND_DOP_SITE_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
{ \
|
||||||
}
|
typedef IMPL Impl; \
|
||||||
|
typedef typename Simd::scalar_type S; \
|
||||||
template<>
|
typedef typename Simd::vector_type V; \
|
||||||
void WilsonKernels<GparityWilsonImplD>::HandDhopSiteDagExt(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
|
\
|
||||||
int ss,int sU,const FermionField &in, FermionField &out)
|
HAND_DECLARATIONS(ignore); \
|
||||||
{
|
\
|
||||||
typedef GparityWilsonImplD Impl;
|
StencilEntry *SE; \
|
||||||
typedef typename Simd::scalar_type S;
|
int offset,local,perm, ptype; \
|
||||||
typedef typename Simd::vector_type V;
|
int nmu=0; \
|
||||||
|
HAND_DOP_SITE_DAG_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
HAND_DECLARATIONS(ignore);
|
nmu = 0; \
|
||||||
|
HAND_DOP_SITE_DAG_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
|
||||||
StencilEntry *SE;
|
|
||||||
int offset,local,perm, ptype;
|
|
||||||
int nmu=0;
|
|
||||||
HAND_DOP_SITE_DAG_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
|
||||||
nmu = 0;
|
|
||||||
HAND_DOP_SITE_DAG_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
||||||
|
HAND_SPECIALISE_GPARITY(GparityWilsonImplF);
|
||||||
|
HAND_SPECIALISE_GPARITY(GparityWilsonImplD);
|
||||||
|
HAND_SPECIALISE_GPARITY(GparityWilsonImplFH);
|
||||||
|
HAND_SPECIALISE_GPARITY(GparityWilsonImplDF);
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user