portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 07:17:06 +01:00
Code Releases Activity

Merge remote-tracking branch 'origin/develop' into temporary-smearing

Browse Source
This commit is contained in:
Guido Cossu
2016-07-04 17:28:40 +01:00
parent 6ce174cd60 d6737e4bd8
commit 9cb90f714e
107 changed files with 7839 additions and 4572 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
lib/qcd/action/Actions.h
View File

@ -109,10 +109,12 @@ typedef SymanzikGaugeAction<ConjugateGimplD> ConjugateSymanzikGaugeAction
#define FermOpTemplateInstantiate(A) \
template class A<WilsonImplF>; \
template class A<WilsonImplD>; \
template class A<WilsonImplD>; \
template class A<GparityWilsonImplF>; \
template class A<GparityWilsonImplD>;
#define GparityFermOpTemplateInstantiate(A)
////////////////////////////////////////////
// Fermion operators / actions
////////////////////////////////////////////
@ -208,6 +210,14 @@ typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
}}
///////////////////////////////////////////////////////////////////////////////
// G5 herm -- this has to live in QCD since dirac matrix is not in the broader sector of code

1
lib/qcd/action/fermion/CayleyFermion5D.cc
View File

@ -527,6 +527,7 @@ namespace QCD {
}
FermOpTemplateInstantiate(CayleyFermion5D);
GparityFermOpTemplateInstantiate(CayleyFermion5D);
}}

110
lib/qcd/action/fermion/FermionOperatorImpl.h
View File

@ -130,7 +130,7 @@ namespace Grid {
typedef WilsonCompressor<SiteHalfSpinor,SiteSpinor> Compressor;
typedef WilsonImplParams ImplParams;
typedef CartesianStencil<SiteSpinor,SiteHalfSpinor,Compressor> StencilImpl;
typedef WilsonStencil<SiteSpinor,SiteHalfSpinor> StencilImpl;
ImplParams Params;
@ -142,6 +142,10 @@ namespace Grid {
mult(&phi(),&U(mu),&chi());
}
template<class ref>
inline void loadLinkElement(Simd & reg,ref &memory){
reg = memory;
}
inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
{
conformable(Uds._grid,GaugeGrid);
@ -181,6 +185,100 @@ PARALLEL_FOR_LOOP
};
///////
// Single flavour four spinors with colour index, 5d redblack
///////
template<class S,int Nrepresentation=Nc>
class DomainWallRedBlack5dImpl : public PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > {
public:
typedef PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > Gimpl;
INHERIT_GIMPL_TYPES(Gimpl);
template<typename vtype> using iImplSpinor = iScalar<iVector<iVector<vtype, Nrepresentation>, Ns> >;
template<typename vtype> using iImplHalfSpinor = iScalar<iVector<iVector<vtype, Nrepresentation>, Nhs> >;
template<typename vtype> using iImplDoubledGaugeField = iVector<iScalar<iMatrix<vtype, Nrepresentation> >, Nds >;
template<typename vtype> using iImplGaugeField = iVector<iScalar<iMatrix<vtype, Nrepresentation> >, Nd >;
template<typename vtype> using iImplGaugeLink = iScalar<iScalar<iMatrix<vtype, Nrepresentation> > >;
typedef iImplSpinor <Simd> SiteSpinor;
typedef iImplHalfSpinor<Simd> SiteHalfSpinor;
typedef Lattice<SiteSpinor> FermionField;
// Make the doubled gauge field a *scalar*
typedef iImplDoubledGaugeField<typename Simd::scalar_type> SiteDoubledGaugeField; // This is a scalar
typedef iImplGaugeField<typename Simd::scalar_type> SiteScalarGaugeField; // scalar
typedef iImplGaugeLink <typename Simd::scalar_type> SiteScalarGaugeLink; // scalar
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef WilsonCompressor<SiteHalfSpinor,SiteSpinor> Compressor;
typedef WilsonImplParams ImplParams;
typedef WilsonStencil<SiteSpinor,SiteHalfSpinor> StencilImpl;
ImplParams Params;
DomainWallRedBlack5dImpl(const ImplParams &p= ImplParams()) : Params(p) {};
bool overlapCommsCompute(void) { return false; };
template<class ref>
inline void loadLinkElement(Simd & reg,ref &memory){
vsplat(reg,memory);
}
inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,StencilImpl &St)
{
SiteGaugeLink UU;
for(int i=0;i<Nrepresentation;i++){
for(int j=0;j<Nrepresentation;j++){
vsplat(UU()()(i,j),U(mu)()(i,j));
}
}
mult(&phi(),&UU(),&chi());
}
inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
{
SiteScalarGaugeField ScalarUmu;
SiteDoubledGaugeField ScalarUds;
GaugeLinkField U (Umu._grid);
GaugeField Uadj(Umu._grid);
for(int mu=0;mu<Nd;mu++){
U = PeekIndex<LorentzIndex>(Umu,mu);
U = adj(Cshift(U,mu,-1));
PokeIndex<LorentzIndex>(Uadj,U,mu);
}
for(int lidx=0;lidx<GaugeGrid->lSites();lidx++){
std::vector<int> lcoor;
GaugeGrid->LocalIndexToLocalCoor(lidx,lcoor);
peekLocalSite(ScalarUmu,Umu,lcoor);
for(int mu=0;mu<4;mu++) ScalarUds(mu) = ScalarUmu(mu);
peekLocalSite(ScalarUmu,Uadj,lcoor);
for(int mu=0;mu<4;mu++) ScalarUds(mu+4) = ScalarUmu(mu);
pokeLocalSite(ScalarUds,Uds,lcoor);
}
}
inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A,int mu){
assert(0);
}
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
assert(0);
}
};
////////////////////////////////////////////////////////////////////////////////////////
// Flavour doubled spinors; is Gparity the only? what about C*?
////////////////////////////////////////////////////////////////////////////////////////
@ -205,7 +303,7 @@ PARALLEL_FOR_LOOP
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef WilsonCompressor<SiteHalfSpinor,SiteSpinor> Compressor;
typedef CartesianStencil<SiteSpinor,SiteHalfSpinor,Compressor> StencilImpl;
typedef WilsonStencil<SiteSpinor,SiteHalfSpinor> StencilImpl;
typedef GparityWilsonImplParams ImplParams;
@ -290,8 +388,8 @@ PARALLEL_FOR_LOOP
conformable(Uds._grid,GaugeGrid);
conformable(Umu._grid,GaugeGrid);
GaugeLinkField Utmp(GaugeGrid);
GaugeLinkField U(GaugeGrid);
GaugeLinkField Utmp (GaugeGrid);
GaugeLinkField U (GaugeGrid);
GaugeLinkField Uconj(GaugeGrid);
Lattice<iScalar<vInteger> > coor(GaugeGrid);
@ -379,6 +477,10 @@ PARALLEL_FOR_LOOP
typedef WilsonImpl<vComplexF,Nc> WilsonImplF; // Float
typedef WilsonImpl<vComplexD,Nc> WilsonImplD; // Double
typedef DomainWallRedBlack5dImpl<vComplex ,Nc> DomainWallRedBlack5dImplR; // Real.. whichever prec
typedef DomainWallRedBlack5dImpl<vComplexF,Nc> DomainWallRedBlack5dImplF; // Float
typedef DomainWallRedBlack5dImpl<vComplexD,Nc> DomainWallRedBlack5dImplD; // Double
typedef GparityWilsonImpl<vComplex ,Nc> GparityWilsonImplR; // Real.. whichever prec
typedef GparityWilsonImpl<vComplexF,Nc> GparityWilsonImplF; // Float
typedef GparityWilsonImpl<vComplexD,Nc> GparityWilsonImplD; // Double

6
lib/qcd/action/fermion/ScaledShamirFermion.h
View File

@ -48,14 +48,16 @@ namespace Grid {
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mass,RealD _M5,
RealD scale) :
// RealD scale):
RealD scale,const ImplParams &p= ImplParams()) :
// b+c=scale, b-c = 1 <=> 2b = scale+1; 2c = scale-1
MobiusFermion<Impl>(_Umu,
FiveDimGrid,
FiveDimRedBlackGrid,
FourDimGrid,
FourDimRedBlackGrid,_mass,_M5,0.5*(scale+1.0),0.5*(scale-1.0))
FourDimRedBlackGrid,_mass,_M5,0.5*(scale+1.0),0.5*(scale-1.0),p)
// FourDimRedBlackGrid,_mass,_M5,0.5*(scale+1.0),0.5*(scale-1.0))
{
}

174
lib/qcd/action/fermion/WilsonCompressor.h
View File

@ -48,12 +48,7 @@ namespace QCD {
mu=p;
};
virtual SiteHalfSpinor operator () (const SiteSpinor &in,int dim,int plane,int osite,GridBase *grid) {
return spinproject(in);
}
SiteHalfSpinor spinproject(const SiteSpinor &in)
{
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
int mudag=mu;
if (!dag) {
@ -92,6 +87,173 @@ namespace QCD {
}
};
/////////////////////////
// optimised versions
/////////////////////////
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonXpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjXp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonYpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjYp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonZpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjZp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonTpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjTp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonXmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjXm(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonYmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjYm(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonZmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjZm(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonTmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjTm(ret,in);
return ret;
}
};
// Fast comms buffer manipulation which should inline right through (avoid direction
// dependent logic that prevents inlining
template<class vobj,class cobj>
class WilsonStencil : public CartesianStencil<vobj,cobj> {
public:
WilsonStencil(GridBase *grid,
int npoints,
int checkerboard,
const std::vector<int> &directions,
const std::vector<int> &distances) : CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances)
{ };
template < class compressor>
std::thread HaloExchangeOptBegin(const Lattice<vobj> &source,compressor &compress) {
this->Mergers.resize(0);
this->Packets.resize(0);
this->HaloGatherOpt(source,compress);
return std::thread([&] { this->Communicate(); });
}
template < class compressor>
void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress)
{
auto thr = this->HaloExchangeOptBegin(source,compress);
this->HaloExchangeOptComplete(thr);
}
void HaloExchangeOptComplete(std::thread &thr)
{
this->CommsMerge(); // spins
this->jointime-=usecond();
thr.join();
this->jointime+=usecond();
}
template < class compressor>
void HaloGatherOpt(const Lattice<vobj> &source,compressor &compress)
{
// conformable(source._grid,_grid);
assert(source._grid==this->_grid);
this->halogtime-=usecond();
assert (this->comm_buf.size() == this->_unified_buffer_size );
this->u_comm_offset=0;
int dag = compress.dag;
static std::vector<int> dirs(Nd*2);
for(int mu=0;mu<Nd;mu++){
if ( dag ) {
dirs[mu] =mu;
dirs[mu+4]=mu+Nd;
} else {
dirs[mu] =mu+Nd;
dirs[mu+Nd]=mu;
}
}
WilsonXpCompressor<cobj,vobj> XpCompress;
this->HaloGatherDir(source,XpCompress,dirs[0]);
WilsonYpCompressor<cobj,vobj> YpCompress;
this->HaloGatherDir(source,YpCompress,dirs[1]);
WilsonZpCompressor<cobj,vobj> ZpCompress;
this->HaloGatherDir(source,ZpCompress,dirs[2]);
WilsonTpCompressor<cobj,vobj> TpCompress;
this->HaloGatherDir(source,TpCompress,dirs[3]);
WilsonXmCompressor<cobj,vobj> XmCompress;
this->HaloGatherDir(source,XmCompress,dirs[4]);
WilsonYmCompressor<cobj,vobj> YmCompress;
this->HaloGatherDir(source,YmCompress,dirs[5]);
WilsonZmCompressor<cobj,vobj> ZmCompress;
this->HaloGatherDir(source,ZmCompress,dirs[6]);
WilsonTmCompressor<cobj,vobj> TmCompress;
this->HaloGatherDir(source,TmCompress,dirs[7]);
assert(this->u_comm_offset==this->_unified_buffer_size);
this->halogtime+=usecond();
}
};
}} // namespace close
#endif

108
lib/qcd/action/fermion/WilsonFermion.cc
View File

@ -64,7 +64,9 @@ namespace QCD {
template<class Impl>
void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
{
Impl::DoubleStore(GaugeGrid(),Umu,_Umu);
GaugeField HUmu(_Umu._grid);
HUmu = _Umu*(-0.5);
Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
pickCheckerboard(Even,UmuEven,Umu);
pickCheckerboard(Odd ,UmuOdd,Umu);
}
@ -286,121 +288,27 @@ PARALLEL_FOR_LOOP
void WilsonFermion<Impl>::DhopInternal(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
if ( Impl::overlapCommsCompute () ) {
DhopInternalCommsOverlapCompute(st,U,in,out,dag);
} else {
DhopInternalCommsThenCompute(st,U,in,out,dag);
}
}
template<class Impl>
void WilsonFermion<Impl>::DhopInternalCommsThenCompute(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) {
assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor(dag);
st.HaloExchange(in,compressor);
if ( dag == DaggerYes ) {
if( HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out);
}
} else {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out);
}
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sss,sss,1,1,in,out);
}
} else {
if( HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sss,sss,in,out);
}
} else {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sss,sss,in,out);
}
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sss,sss,1,1,in,out);
}
}
};
template<class Impl>
void WilsonFermion<Impl>::DhopInternalCommsOverlapCompute(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) {
assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor(dag);
auto handle = st.HaloExchangeBegin(in,compressor);
bool local = true;
bool nonlocal = false;
if ( dag == DaggerYes ) {
if( HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
} else {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
}
} else {
if( HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
} else {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
}
}
st.HaloExchangeComplete(handle);
local = false;
nonlocal = true;
if ( dag == DaggerYes ) {
if( HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
} else {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
}
} else {
if( HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
} else {
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sss,sss,in,out,local,nonlocal);
}
}
}
};
FermOpTemplateInstantiate(WilsonFermion);
GparityFermOpTemplateInstantiate(WilsonFermion);
}}

6
lib/qcd/action/fermion/WilsonFermion.h
View File

@ -114,12 +114,6 @@ namespace Grid {
void DhopInternal(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) ;
void DhopInternalCommsThenCompute(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) ;
void DhopInternalCommsOverlapCompute(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) ;
// Constructor
WilsonFermion(GaugeField &_Umu,
GridCartesian &Fgrid,

373
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -38,8 +38,6 @@ namespace QCD {
// S-direction is INNERMOST and takes no part in the parity.
const std::vector<int> WilsonFermion5DStatic::directions ({1,2,3,4, 1, 2, 3, 4});
const std::vector<int> WilsonFermion5DStatic::displacements({1,1,1,1,-1,-1,-1,-1});
int WilsonFermion5DStatic::HandOptDslash;
int WilsonFermion5DStatic::AsmOptDslash;
// 5d lattice for DWF.
template<class Impl>
@ -67,10 +65,8 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
// some assertions
assert(FiveDimGrid._ndimension==5);
assert(FourDimGrid._ndimension==4);
assert(FiveDimRedBlackGrid._ndimension==5);
assert(FourDimRedBlackGrid._ndimension==4);
assert(FiveDimRedBlackGrid._checker_dim==1);
// Dimension zero of the five-d is the Ls direction
@ -99,16 +95,74 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
// Allocate the required comms buffer
ImportGauge(_Umu);
alltime=0;
commtime=0;
jointime=0;
dslashtime=0;
dslash1time=0;
}
template<class Impl>
WilsonFermion5D<Impl>::WilsonFermion5D(int simd,GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
RealD _M5,const ImplParams &p) :
Kernels(p),
_FiveDimGrid (&FiveDimGrid),
_FiveDimRedBlackGrid(&FiveDimRedBlackGrid),
_FourDimGrid (&FourDimGrid),
Stencil (_FiveDimGrid,npoint,Even,directions,displacements),
StencilEven(_FiveDimRedBlackGrid,npoint,Even,directions,displacements), // source is Even
StencilOdd (_FiveDimRedBlackGrid,npoint,Odd ,directions,displacements), // source is Odd
M5(_M5),
Umu(_FourDimGrid),
UmuEven(_FourDimGrid),
UmuOdd (_FourDimGrid),
Lebesgue(_FourDimGrid),
LebesgueEvenOdd(_FourDimGrid)
{
int nsimd = Simd::Nsimd();
// some assertions
assert(FiveDimGrid._ndimension==5);
assert(FiveDimRedBlackGrid._ndimension==5);
assert(FiveDimRedBlackGrid._checker_dim==0); // Checkerboard the s-direction
assert(FourDimGrid._ndimension==4);
// Dimension zero of the five-d is the Ls direction
Ls=FiveDimGrid._fdimensions[0];
assert(FiveDimGrid._processors[0] ==1);
assert(FiveDimGrid._simd_layout[0] ==nsimd);
assert(FiveDimRedBlackGrid._fdimensions[0]==Ls);
assert(FiveDimRedBlackGrid._processors[0] ==1);
assert(FiveDimRedBlackGrid._simd_layout[0]==nsimd);
// Other dimensions must match the decomposition of the four-D fields
for(int d=0;d<4;d++){
assert(FiveDimRedBlackGrid._fdimensions[d+1]==FourDimGrid._fdimensions[d]);
assert(FiveDimRedBlackGrid._processors[d+1] ==FourDimGrid._processors[d]);
assert(FourDimGrid._simd_layout[d]=1);
assert(FiveDimRedBlackGrid._simd_layout[d+1]==1);
assert(FiveDimGrid._fdimensions[d+1] ==FourDimGrid._fdimensions[d]);
assert(FiveDimGrid._processors[d+1] ==FourDimGrid._processors[d]);
assert(FiveDimGrid._simd_layout[d+1] ==FourDimGrid._simd_layout[d]);
}
{
GaugeField HUmu(_Umu._grid);
HUmu = _Umu*(-0.5);
Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
UmuEven=Umu;// Really want a reference.
UmuOdd =Umu;
}
}
template<class Impl>
void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
{
Impl::DoubleStore(GaugeGrid(),Umu,_Umu);
GaugeField HUmu(_Umu._grid);
HUmu = _Umu*(-0.5);
Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
pickCheckerboard(Even,UmuEven,Umu);
pickCheckerboard(Odd ,UmuOdd,Umu);
}
@ -232,30 +286,6 @@ void WilsonFermion5D<Impl>::DhopDerivEO(GaugeField &mat,
}
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
{
std::cout<<GridLogMessage << "******************** WilsonFermion"<<std::endl;
std::cout<<GridLogMessage << "Wilson5d time "<<alltime <<" us"<<std::endl;
std::cout<<GridLogMessage << "HaloBegin time "<<commtime <<" us"<<std::endl;
std::cout<<GridLogMessage << "Dslash time "<<dslashtime<<" us"<<std::endl;
std::cout<<GridLogMessage << "Dslash1 time "<<dslash1time<<" us"<<std::endl;
std::cout<<GridLogMessage << "HaloComplete time "<<jointime<<" us"<<std::endl;
std::cout<<GridLogMessage << "******************** Stencil"<<std::endl;
std::cout<<GridLogMessage << "Stencil all gather time "<<Stencil.halogtime<<" us"<<std::endl;
std::cout<<GridLogMessage << "Stencil nosplice gather time "<<Stencil.nosplicetime<<" us"<<std::endl;
std::cout<<GridLogMessage << "Stencil splice gather time "<<Stencil.splicetime<<" us"<<std::endl;
std::cout<<GridLogMessage << "********************"<<std::endl;
std::cout<<GridLogMessage << "Stencil gather "<<Stencil.gathertime<<" us"<<std::endl;
std::cout<<GridLogMessage << "Stencil gather simd "<<Stencil.gathermtime<<" us"<<std::endl;
std::cout<<GridLogMessage << "Stencil merge simd "<<Stencil.mergetime<<" us"<<std::endl;
std::cout<<GridLogMessage << "Stencil spin simd "<<Stencil.spintime<<" us"<<std::endl;
std::cout<<GridLogMessage << "********************"<<std::endl;
std::cout<<GridLogMessage << "Stencil MB/s "<<(double)Stencil.comms_bytes/Stencil.commtime<<std::endl;
std::cout<<GridLogMessage << "Stencil comm time "<<Stencil.commtime<<" us"<<std::endl;
std::cout<<GridLogMessage << "Stencil join time "<<Stencil.jointime<<" us"<<std::endl;
std::cout<<GridLogMessage << "********************"<<std::endl;
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
const FermionField &A,
@ -277,280 +307,32 @@ template<class Impl>
void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
if ( Impl::overlapCommsCompute () ) {
DhopInternalCommsOverlapCompute(st,lo,U,in,out,dag);
} else {
DhopInternalCommsThenCompute(st,lo,U,in,out,dag);
}
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalCommsThenCompute(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
// assert((dag==DaggerNo) ||(dag==DaggerYes));
alltime-=usecond();
Compressor compressor(dag);
// Assume balanced KMP_AFFINITY; this is forced in GridThread.h
int threads = GridThread::GetThreads();
int HT = GridThread::GetHyperThreads();
int cores = GridThread::GetCores();
int nwork = U._grid->oSites();
int LLs = in._grid->_rdimensions[0];
commtime -=usecond();
auto handle = st.HaloExchangeBegin(in,compressor);
st.HaloExchangeComplete(handle);
commtime +=usecond();
jointime -=usecond();
jointime +=usecond();
st.HaloExchange(in,compressor);
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
// Not loop ordering and data layout.
// Designed to create
// - per thread reuse in L1 cache for U
// - 8 linear access unit stride streams per thread for Fermion for hw prefetchable.
dslashtime -=usecond();
if ( dag == DaggerYes ) {
if( this->HandOptDslash ) {
#pragma omp parallel for schedule(static)
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out);
}
}
} else {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
{
int sd;
for(sd=0;sd<Ls;sd++){
int sU=ss;
int sF = sd+Ls*sU;
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out);
}
}
}
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
int sF=LLs*sU;
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,LLs,1,in,out);
}
} else {
if( this->AsmOptDslash ) {
// for(int i=0;i<1;i++){
// for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
// PerformanceCounter Counter(i);
// Counter.Start();
#pragma omp parallel for
for(int t=0;t<threads;t++){
int hyperthread = t%HT;
int core = t/HT;
int sswork, swork,soff,ssoff, sU,sF;
GridThread::GetWork(nwork,core,sswork,ssoff,cores);
GridThread::GetWork(Ls , hyperthread, swork, soff,HT);
for(int ss=0;ss<sswork;ss++){
for(int s=soff;s<soff+swork;s++){
sU=ss+ ssoff;
if ( LebesgueOrder::UseLebesgueOrder ) {
sU = lo.Reorder(sU);
}
sF = s+Ls*sU;
Kernels::DiracOptAsmDhopSite(st,U,st.comm_buf,sF,sU,in,out,(uint64_t *)0);// &buf[0]
}
}
}
// Counter.Stop();
// Counter.Report();
// }
} else if( this->HandOptDslash ) {
/*
#pragma omp parallel for schedule(static)
for(int t=0;t<threads;t++){
int hyperthread = t%HT;
int core = t/HT;
int sswork, swork,soff,ssoff, sU,sF;
GridThread::GetWork(nwork,core,sswork,ssoff,cores);
GridThread::GetWork(Ls , hyperthread, swork, soff,HT);
for(int ss=0;ss<sswork;ss++){
sU=ss+ ssoff;
for(int s=soff;s<soff+swork;s++){
sF = s+Ls*sU;
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sF,sU,in,out);
}
}
}
*/
#pragma omp parallel for schedule(static)
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sF,sU,in,out);
}
}
} else {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,in,out);
}
}
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
int sF=LLs*sU;
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,LLs,1,in,out);
}
}
dslashtime +=usecond();
alltime+=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalCommsOverlapCompute(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
// assert((dag==DaggerNo) ||(dag==DaggerYes));
alltime-=usecond();
int calls;
int updates;
Compressor compressor(dag);
// Assume balanced KMP_AFFINITY; this is forced in GridThread.h
int threads = GridThread::GetThreads();
int HT = GridThread::GetHyperThreads();
int cores = GridThread::GetCores();
int nwork = U._grid->oSites();
commtime -=usecond();
auto handle = st.HaloExchangeBegin(in,compressor);
commtime +=usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
// Not loop ordering and data layout.
// Designed to create
// - per thread reuse in L1 cache for U
// - 8 linear access unit stride streams per thread for Fermion for hw prefetchable.
bool local = true;
bool nonlocal = false;
dslashtime -=usecond();
if ( dag == DaggerYes ) {
if( this->HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
} else {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
{
int sd;
for(sd=0;sd<Ls;sd++){
int sU=ss;
int sF = sd+Ls*sU;
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
}
}
} else {
if( this->HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
} else {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
}
}
dslashtime +=usecond();
jointime -=usecond();
st.HaloExchangeComplete(handle);
jointime +=usecond();
local = false;
nonlocal = true;
dslash1time -=usecond();
if ( dag == DaggerYes ) {
if( this->HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
} else {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
{
int sd;
for(sd=0;sd<Ls;sd++){
int sU=ss;
int sF = sd+Ls*sU;
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
}
}
} else {
if( this->HandOptDslash ) {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
} else {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,in,out,local,nonlocal);
}
}
}
}
dslash1time +=usecond();
alltime+=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
@ -593,7 +375,10 @@ void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag
}
FermOpTemplateInstantiate(WilsonFermion5D);
GparityFermOpTemplateInstantiate(WilsonFermion5D);
template class WilsonFermion5D<DomainWallRedBlack5dImplF>;
template class WilsonFermion5D<DomainWallRedBlack5dImplD>;
}}

34
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@ -1,3 +1,4 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -48,8 +49,6 @@ namespace Grid {
class WilsonFermion5DStatic {
public:
// S-direction is INNERMOST and takes no part in the parity.
static int AsmOptDslash; // these are a temporary hack
static int HandOptDslash; // these are a temporary hack
static const std::vector<int> directions;
static const std::vector<int> displacements;
const int npoint = 8;
@ -61,11 +60,7 @@ namespace Grid {
public:
INHERIT_IMPL_TYPES(Impl);
typedef WilsonKernels<Impl> Kernels;
double alltime;
double jointime;
double commtime;
double dslashtime;
double dslash1time;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////
@ -86,6 +81,7 @@ namespace Grid {
virtual void MeooeDag (const FermionField &in, FermionField &out){assert(0);};
virtual void MooeeDag (const FermionField &in, FermionField &out){assert(0);};
virtual void MooeeInvDag (const FermionField &in, FermionField &out){assert(0);};
virtual void Mdir (const FermionField &in, FermionField &out,int dir,int disp){assert(0);}; // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
// These can be overridden by fancy 5d chiral action
virtual void DhopDeriv (GaugeField &mat,const FermionField &U,const FermionField &V,int dag);
@ -120,19 +116,6 @@ namespace Grid {
FermionField &out,
int dag);
void DhopInternalCommsThenCompute(StencilImpl & st,
LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out,
int dag);
void DhopInternalCommsOverlapCompute(StencilImpl & st,
LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out,
int dag);
// Constructors
WilsonFermion5D(GaugeField &_Umu,
GridCartesian &FiveDimGrid,
@ -141,14 +124,21 @@ namespace Grid {
GridRedBlackCartesian &FourDimRedBlackGrid,
double _M5,const ImplParams &p= ImplParams());
// Constructors
WilsonFermion5D(int simd,
GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
double _M5,const ImplParams &p= ImplParams());
// DoubleStore
void ImportGauge(const GaugeField &_Umu);
void Report(void);
///////////////////////////////////////////////////////////////
// Data members require to support the functionality
///////////////////////////////////////////////////////////////
protected:
public:
// Add these to the support from Wilson
GridBase *_FourDimGrid;

352
lib/qcd/action/fermion/WilsonKernels.cc
View File

@ -31,440 +31,410 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
namespace Grid {
namespace QCD {
int WilsonKernelsStatic::HandOpt;
int WilsonKernelsStatic::AsmOpt;
template<class Impl>
WilsonKernels<Impl>::WilsonKernels(const ImplParams &p): Base(p) {};
// Need controls to do interior, exterior, or both
template<class Impl>
void WilsonKernels<Impl>::DiracOptDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,int Ls, int Ns, const FermionField &in, FermionField &out)
{
if ( AsmOpt ) {
WilsonKernels<Impl>::DiracOptAsmDhopSite(st,U,buf,sF,sU,Ls,Ns,in,out);
} else {
for(int site=0;site<Ns;site++) {
for(int s=0;s<Ls;s++) {
if (HandOpt) WilsonKernels<Impl>::DiracOptHandDhopSite(st,U,buf,sF,sU,in,out);
else WilsonKernels<Impl>::DiracOptGenericDhopSite(st,U,buf,sF,sU,in,out);
sF++;
}
sU++;
}
}
}
template<class Impl>
void WilsonKernels<Impl>::DiracOptDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local, bool nonlocal)
int sF,int sU,int Ls, int Ns, const FermionField &in, FermionField &out)
{
// No asm implementation yet.
// if ( AsmOpt ) WilsonKernels<Impl>::DiracOptAsmDhopSiteDag(st,U,buf,sF,sU,in,out);
// else
for(int site=0;site<Ns;site++) {
for(int s=0;s<Ls;s++) {
if (HandOpt) WilsonKernels<Impl>::DiracOptHandDhopSiteDag(st,U,buf,sF,sU,in,out);
else WilsonKernels<Impl>::DiracOptGenericDhopSiteDag(st,U,buf,sF,sU,in,out);
sF++;
}
sU++;
}
}
////////////////////////////////////////////
// Generic implementation; move to different file?
////////////////////////////////////////////
template<class Impl>
void WilsonKernels<Impl>::DiracOptGenericDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
int num = 0;
result=zero;
///////////////////////////
// Xp
///////////////////////////
SE=st.GetEntry(ptype,Xp,sF);
if (local && SE->_is_local ) {
if (SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjXp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXp(chi,in._odata[SE->_offset]);
}
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
} else {
chi_p=&buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Xp,SE,st);
accumReconXp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Xp,SE,st);
spReconXp(result,Uchi);
///////////////////////////
// Yp
///////////////////////////
SE=st.GetEntry(ptype,Yp,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjYp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYp(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Yp,SE,st);
accumReconYp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Yp,SE,st);
accumReconYp(result,Uchi);
///////////////////////////
// Zp
///////////////////////////
SE=st.GetEntry(ptype,Zp,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjZp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZp(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Zp,SE,st);
accumReconZp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Zp,SE,st);
accumReconZp(result,Uchi);
///////////////////////////
// Tp
///////////////////////////
SE=st.GetEntry(ptype,Tp,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjTp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTp(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Tp,SE,st);
accumReconTp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Tp,SE,st);
accumReconTp(result,Uchi);
///////////////////////////
// Xm
///////////////////////////
SE=st.GetEntry(ptype,Xm,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjXm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Xm,SE,st);
accumReconXm(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Xm,SE,st);
accumReconXm(result,Uchi);
///////////////////////////
// Ym
///////////////////////////
SE=st.GetEntry(ptype,Ym,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjYm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Ym,SE,st);
accumReconYm(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Ym,SE,st);
accumReconYm(result,Uchi);
///////////////////////////
// Zm
///////////////////////////
SE=st.GetEntry(ptype,Zm,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjZm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Zm,SE,st);
accumReconZm(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Zm,SE,st);
accumReconZm(result,Uchi);
///////////////////////////
// Tm
///////////////////////////
SE=st.GetEntry(ptype,Tm,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjTm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Tm,SE,st);
accumReconTm(result,Uchi);
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Tm,SE,st);
accumReconTm(result,Uchi);
num++;
}
if ( local ) {
vstream(out._odata[sF],result*(-0.5));
} else if ( num ) {
vstream(out._odata[sF],out._odata[sF]+result*(-0.5));
}
vstream(out._odata[sF],result);
};
// Need controls to do interior, exterior, or both
template<class Impl>
void WilsonKernels<Impl>::DiracOptDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local, bool nonlocal)
void WilsonKernels<Impl>::DiracOptGenericDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
int num = 0;
result=zero;
///////////////////////////
// Xp
///////////////////////////
SE=st.GetEntry(ptype,Xm,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjXp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXp(chi,in._odata[SE->_offset]);
}
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
} else {
chi_p=&buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Xm,SE,st);
accumReconXp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Xm,SE,st);
spReconXp(result,Uchi);
///////////////////////////
// Yp
///////////////////////////
SE=st.GetEntry(ptype,Ym,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjYp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYp(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Ym,SE,st);
accumReconYp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Ym,SE,st);
accumReconYp(result,Uchi);
///////////////////////////
// Zp
///////////////////////////
SE=st.GetEntry(ptype,Zm,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjZp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZp(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Zm,SE,st);
accumReconZp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Zm,SE,st);
accumReconZp(result,Uchi);
///////////////////////////
// Tp
///////////////////////////
SE=st.GetEntry(ptype,Tm,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjTp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTp(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Tm,SE,st);
accumReconTp(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Tm,SE,st);
accumReconTp(result,Uchi);
///////////////////////////
// Xm
///////////////////////////
SE=st.GetEntry(ptype,Xp,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjXm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Xp,SE,st);
accumReconXm(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Xp,SE,st);
accumReconXm(result,Uchi);
///////////////////////////
// Ym
///////////////////////////
SE=st.GetEntry(ptype,Yp,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjYm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Yp,SE,st);
accumReconYm(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Yp,SE,st);
accumReconYm(result,Uchi);
///////////////////////////
// Zm
///////////////////////////
SE=st.GetEntry(ptype,Zp,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjZm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Zp,SE,st);
accumReconZm(result,Uchi);
num++;
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Zp,SE,st);
accumReconZm(result,Uchi);
///////////////////////////
// Tm
///////////////////////////
SE=st.GetEntry(ptype,Tp,sF);
if (local && SE->_is_local ) {
if ( SE->_is_local ) {
chi_p = &chi;
if ( SE->_permute ) {
spProjTm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTm(chi,in._odata[SE->_offset]);
}
} else {
chi_p=&buf[SE->_offset];
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],*chi_p,Tp,SE,st);
accumReconTm(result,Uchi);
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Tp,SE,st);
accumReconTm(result,Uchi);
num++;
}
if ( local ) {
vstream(out._odata[sF],result*(-0.5));
} else if ( num ) {
vstream(out._odata[sF],out._odata[sF]+result*(-0.5));
}
vstream(out._odata[sF],result);
};
template<class Impl>
@ -593,19 +563,13 @@ void WilsonKernels<Impl>::DiracOptDhopDir(StencilImpl &st,DoubledGaugeField &U,
spReconTm(result,Uchi);
}
vstream(out._odata[sF],result*(-0.5));
vstream(out._odata[sF],result);
}
#if ( ! defined(AVX512) ) && ( ! defined(IMCI) )
template<class Impl>
void WilsonKernels<Impl>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local, bool nonlocal)
{
DiracOptDhopSite(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
}
#endif
FermOpTemplateInstantiate(WilsonKernels);
template class WilsonKernels<DomainWallRedBlack5dImplF>;
template class WilsonKernels<DomainWallRedBlack5dImplD>;
}}

35
lib/qcd/action/fermion/WilsonKernels.h
View File

@ -38,37 +38,56 @@ namespace Grid {
// Helper routines that implement Wilson stencil for a single site.
// Common to both the WilsonFermion and WilsonFermion5D
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
class WilsonKernelsStatic {
public:
// S-direction is INNERMOST and takes no part in the parity.
static int AsmOpt; // these are a temporary hack
static int HandOpt; // these are a temporary hack
};
template<class Impl> class WilsonKernels : public FermionOperator<Impl> {
template<class Impl> class WilsonKernels : public FermionOperator<Impl> , public WilsonKernelsStatic {
public:
INHERIT_IMPL_TYPES(Impl);
typedef FermionOperator<Impl> Base;
public:
void DiracOptDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
int sF, int sU,int Ls, int Ns, const FermionField &in, FermionField &out);
void DiracOptDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in,FermionField &out,bool local= true, bool nonlocal=true);
int sF,int sU,int Ls, int Ns, const FermionField &in,FermionField &out);
void DiracOptDhopDir(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,int dirdisp,int gamma);
private:
// Specialised variants
void DiracOptGenericDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU, const FermionField &in, FermionField &out);
void DiracOptGenericDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in,FermionField &out);
void DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
int sF,int sU,int Ls, int Ns, const FermionField &in, FermionField &out);
int DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
void DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
int sF,int sU,const FermionField &in, FermionField &out);
int DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
void DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
int sF,int sU,const FermionField &in, FermionField &out);
public:
WilsonKernels(const ImplParams &p= ImplParams());

380
lib/qcd/action/fermion/WilsonKernelsAsm.cc
View File

@ -2,6 +2,8 @@
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernelsAsm.cc
Copyright (C) 2015
@ -26,320 +28,88 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid.h>
#if defined(AVX512) || defined (IMCI)
#include <simd/Avx512Asm.h>
#undef VLOAD
#undef VSTORE
#undef VMUL
#undef VMADD
#undef ZEND
#undef ZLOAD
#undef ZMUL
#undef ZMADD
#undef VZERO
#undef VTIMESI
#undef VTIMESMINUSI
#define VZERO(A) VZEROf(A)
#define VMOV(A,B) VMOVf(A,B)
#define VLOAD(OFF,PTR,DEST) VLOADf(OFF,PTR,DEST)
#define VSTORE(OFF,PTR,SRC) VSTOREf(OFF,PTR,SRC)
#define VADD(A,B,C) VADDf(A,B,C)
#define VSUB(A,B,C) VSUBf(A,B,C)
#define VMUL(Uri,Uir,Chi,UChi,Z) VMULf(Uri,Uir,Chi,UChi,Z)
#define VMADD(Uri,Uir,Chi,UChi,Z) VMADDf(Uri,Uir,Chi,UChi,Z)
#define VTIMESI(A,B,C) VTIMESIf(A,B,C)
#define VTIMESMINUSI(A,B,C) VTIMESMINUSIf(A,B,C)
#define VACCTIMESI(A,B,C) VACCTIMESIf(A,B,C)
#define VACCTIMESMINUSI(A,B,C) VACCTIMESMINUSIf(A,B,C)
#define VTIMESI0(A,B,C) VTIMESI0f(A,B,C)
#define VTIMESMINUSI0(A,B,C) VTIMESMINUSI0f(A,B,C)
#define VACCTIMESI0(A,B,C) VACCTIMESI0f(A,B,C)
#define VACCTIMESMINUSI0(A,B,C) VACCTIMESMINUSI0f(A,B,C)
#define VTIMESI1(A,B,C) VTIMESI1f(A,B,C)
#define VTIMESMINUSI1(A,B,C) VTIMESMINUSI1f(A,B,C)
#define VACCTIMESI1(A,B,C) VACCTIMESI1f(A,B,C)
#define VACCTIMESMINUSI1(A,B,C) VACCTIMESMINUSI1f(A,B,C)
#define VTIMESI2(A,B,C) VTIMESI2f(A,B,C)
#define VTIMESMINUSI2(A,B,C) VTIMESMINUSI2f(A,B,C)
#define VACCTIMESI2(A,B,C) VACCTIMESI2f(A,B,C)
#define VACCTIMESMINUSI2(A,B,C) VACCTIMESMINUSI2f(A,B,C)
#define VACCTIMESI1MEM(A,ACC,O,P) VACCTIMESI1MEMf(A,ACC,O,P)
#define VACCTIMESI2MEM(A,ACC,O,P) VACCTIMESI2MEMf(A,ACC,O,P)
#define VACCTIMESMINUSI1MEM(A,ACC,O,P) VACCTIMESMINUSI1MEMf(A,ACC,O,P)
#define VACCTIMESMINUSI2MEM(A,ACC,O,P) VACCTIMESMINUSI2MEMf(A,ACC,O,P)
#define VPERM0(A,B) VPERM0f(A,B)
#define VPERM1(A,B) VPERM1f(A,B)
#define VPERM2(A,B) VPERM2f(A,B)
#define VPERM3(A,B) VPERM3f(A,B)
#define VSHUFMEM(OFF,A,DEST) VSHUFMEMf(OFF,A,DEST)
#define ZEND1(A,B,C) ZEND1f(A,B,C)
#define ZEND2(A,B,C) ZEND2f(A,B,C)
#define ZLOAD(A,B,C,D) ZLOADf(A,B,C,D)
#define ZMUL(A,B,C,D,E) ZMULf(A,B,C,D,E)
#define ZMADD(A,B,C,D,E) ZMADDf(A,B,C,D,E)
#define ZMUL(A,B,C,D,E) ZMULf(A,B,C,D,E)
#define ZMADD(A,B,C,D,E) ZMADDf(A,B,C,D,E)
#define VADDMEM(O,A,B,C) VADDMEMf(O,A,B,C)
#define VSUBMEM(O,A,B,C) VSUBMEMf(O,A,B,C)
#define ZMULMEM2SP(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr) ZMULMEM2SPf(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr)
#define ZMADDMEM2SP(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr) ZMADDMEM2SPf(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr)
namespace Grid {
namespace QCD {
///////////////////////////////////////////////////////////
// Default to no assembler implementation
///////////////////////////////////////////////////////////
template<class Impl>
void WilsonKernels<Impl >::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out,uint64_t *timers)
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out)
{
uint64_t now;
uint64_t first ;
int offset,local,perm, ptype;
const SiteHalfSpinor *pbuf = & buf[0];
const SiteSpinor *plocal = & in._odata[0];
void *pf;
int osites = in._grid->oSites();
StencilEntry *SE;
//#define STAMP(i) timers[i] = __rdtsc() ;
#define STAMP(i) //timers[i] = __rdtsc() ;
MASK_REGS;
first = __rdtsc();
SE=st.GetEntry(ptype,Xm,ss);
#if 0
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
LOAD64(%r9,pf);
__asm__(
VPREFETCH(0,%r9)
VPREFETCH(1,%r9)
VPREFETCH(2,%r9)
VPREFETCH(3,%r9)
VPREFETCH(4,%r9)
VPREFETCH(5,%r9)
VPREFETCH(6,%r9)
VPREFETCH(7,%r9)
VPREFETCH(8,%r9)
VPREFETCH(9,%r9)
VPREFETCH(10,%r9)
VPREFETCH(11,%r9) );
#endif
// Xm
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
// Prefetch
SE=st.GetEntry(ptype,Ym,ss);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
XM_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR3; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFXM(Xm,pf);
}
XM_RECON;
// Ym
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
// Prefetch
SE=st.GetEntry(ptype,Zm,ss);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
YM_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR2; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFYM(Ym,pf);
}
YM_RECON_ACCUM;
// Zm
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
// Prefetch
SE=st.GetEntry(ptype,Tm,ss);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
ZM_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR1; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFZM(Zm,pf);
}
ZM_RECON_ACCUM;
// Tm
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
SE=st.GetEntry(ptype,Tp,ss);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
TM_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR0; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFTM(Tm,pf);
}
TM_RECON_ACCUM;
// Tp
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
// Prefetch
SE=st.GetEntry(ptype,Zp,ss);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
TP_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR0; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFTP(Tp,pf);
}
TP_RECON_ACCUM;
// Zp
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
// Prefetch
SE=st.GetEntry(ptype,Yp,ss);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
ZP_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR1; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFZP(Zp,pf);
}
ZP_RECON_ACCUM;
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
// Prefetch
SE=st.GetEntry(ptype,Xp,ss);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
YP_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR2; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFYP(Yp,pf);
}
YP_RECON_ACCUM;
// Xp
perm = SE->_permute;
offset = SE->_offset;
local = SE->_is_local;
// PREFETCH_R(A);
// Prefetch
SE=st.GetEntry(ptype,Xm,(ss+1)%osites);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
if ( local ) {
XP_PROJMEM(&plocal[offset]);
if ( perm) {
PERMUTE_DIR3; // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI(&pbuf[offset]);
}
{
MULT_2SPIN_DIR_PFXP(Xp,pf);
}
XP_RECON_ACCUM;
debug:
SAVE_RESULT(&out._odata[ss]);
assert(0);
}
template class WilsonKernels<WilsonImplF>;
template class WilsonKernels<WilsonImplD>;
#if defined(AVX512)
///////////////////////////////////////////////////////////
// If we are AVX512 specialise the single precision routine
///////////////////////////////////////////////////////////
#include <simd/Intel512wilson.h>
#include <simd/Intel512single.h>
static Vector<vComplexF> signs;
int setupSigns(void ){
Vector<vComplexF> bother(2);
signs = bother;
vrsign(signs[0]);
visign(signs[1]);
return 1;
}
static int signInit = setupSigns();
#define MAYBEPERM(A,perm) if (perm) { A ; }
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
template<>
void WilsonKernels<WilsonImplF>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef VMOVIDUP
#undef VMOVRDUP
#undef MAYBEPERM
#undef MULT_2SPIN
#define MAYBEPERM(A,B)
#define VMOVIDUP(A,B,C) VBCASTIDUPf(A,B,C)
#define VMOVRDUP(A,B,C) VBCASTRDUPf(A,B,C)
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
template<>
void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
}}
#endif
template void WilsonKernels<WilsonImplF>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out);
template void WilsonKernels<WilsonImplD>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplF>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplD>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,int Ls,int Ns,const FermionField &in, FermionField &out);
}}

164
lib/qcd/action/fermion/WilsonKernelsAsmBody.h Normal file
View File

@ -0,0 +1,164 @@
{
int locala,perma, ptypea;
int localb,permb, ptypeb;
uint64_t basea, baseb;
uint64_t basex;
const uint64_t plocal =(uint64_t) & in._odata[0];
// vComplexF isigns[2] = { signs[0], signs[1] };
vComplexF *isigns = &signs[0];
MASK_REGS;
for(int site=0;site<Ns;site++) {
for(int s=0;s<Ls;s++) {
////////////////////////////////
// Xp
////////////////////////////////
int ent=ss*8;// 2*Ndim
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
basex = basea;
if ( locala ) {
LOAD64(%r10,isigns);
XM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR3,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFXP(Xp,baseb);
}
LOAD64(%r10,isigns);
XM_RECON;
////////////////////////////////
// Yp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR2,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFYP(Yp,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_RECON_ACCUM;
////////////////////////////////
// Zp
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR1,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFZP(Zp,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_RECON_ACCUM;
////////////////////////////////
// Tp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR0,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFTP(Tp,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_RECON_ACCUM;
////////////////////////////////
// Xm
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR3,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFXM(Xm,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_RECON_ACCUM;
////////////////////////////////
// Ym
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR2,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFYM(Ym,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_RECON_ACCUM;
////////////////////////////////
// Zm
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR1,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFZM(Zm,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_RECON_ACCUM;
////////////////////////////////
// Tm
////////////////////////////////
basea = (uint64_t)&out._odata[ss];
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR0,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFTM(Tm,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_RECON_ACCUM;
PREFETCH_CHIMU(basex);
SAVE_RESULT(&out._odata[ss]);
ss++;
}
sU++;
}
}

911
lib/qcd/action/fermion/WilsonKernelsHand.cc
View File

@ -54,14 +54,15 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
Chi_11 = ref()(1)(1);\
Chi_12 = ref()(1)(2);
// To splat or not to splat depends on the implementation
#define MULT_2SPIN(A)\
auto & ref(U._odata[sU](A)); \
U_00 = ref()(0,0);\
U_10 = ref()(1,0);\
U_20 = ref()(2,0);\
U_01 = ref()(0,1);\
U_11 = ref()(1,1); \
U_21 = ref()(2,1);\
Impl::loadLinkElement(U_00,ref()(0,0)); \
Impl::loadLinkElement(U_10,ref()(1,0)); \
Impl::loadLinkElement(U_20,ref()(2,0)); \
Impl::loadLinkElement(U_01,ref()(0,1)); \
Impl::loadLinkElement(U_11,ref()(1,1)); \
Impl::loadLinkElement(U_21,ref()(2,1)); \
UChi_00 = U_00*Chi_00;\
UChi_10 = U_00*Chi_10;\
UChi_01 = U_10*Chi_00;\
@ -74,9 +75,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
UChi_11+= U_11*Chi_11;\
UChi_02+= U_21*Chi_01;\
UChi_12+= U_21*Chi_11;\
U_00 = ref()(0,2);\
U_10 = ref()(1,2);\
U_20 = ref()(2,2);\
Impl::loadLinkElement(U_00,ref()(0,2)); \
Impl::loadLinkElement(U_10,ref()(1,2)); \
Impl::loadLinkElement(U_20,ref()(2,2)); \
UChi_00+= U_00*Chi_02;\
UChi_10+= U_00*Chi_12;\
UChi_01+= U_10*Chi_02;\
@ -84,6 +85,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
UChi_02+= U_20*Chi_02;\
UChi_12+= U_20*Chi_12;
#define PERMUTE_DIR(dir) \
permute##dir(Chi_00,Chi_00);\
permute##dir(Chi_01,Chi_01);\
@ -309,546 +311,10 @@ namespace Grid {
namespace QCD {
template<class Impl>
int WilsonKernels<Impl >::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
{
// std::cout << "Hand op Dhop "<<std::endl;
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
REGISTER Simd result_00 ; zeroit(result_00); // 12 regs on knc
REGISTER Simd result_01 ; zeroit(result_01); // 12 regs on knc
REGISTER Simd result_02 ; zeroit(result_02); // 12 regs on knc
REGISTER Simd result_10 ; zeroit(result_10); // 12 regs on knc
REGISTER Simd result_11 ; zeroit(result_11); // 12 regs on knc
REGISTER Simd result_12 ; zeroit(result_12); // 12 regs on knc
REGISTER Simd result_20 ; zeroit(result_20); // 12 regs on knc
REGISTER Simd result_21 ; zeroit(result_21); // 12 regs on knc
REGISTER Simd result_22 ; zeroit(result_22); // 12 regs on knc
REGISTER Simd result_30 ; zeroit(result_30); // 12 regs on knc
REGISTER Simd result_31 ; zeroit(result_31); // 12 regs on knc
REGISTER Simd result_32 ; zeroit(result_32); // 12 regs on knc
REGISTER Simd Chi_00; // two spinor; 6 regs
REGISTER Simd Chi_01;
REGISTER Simd Chi_02;
REGISTER Simd Chi_10;
REGISTER Simd Chi_11;
REGISTER Simd Chi_12; // 14 left
REGISTER Simd UChi_00; // two spinor; 6 regs
REGISTER Simd UChi_01;
REGISTER Simd UChi_02;
REGISTER Simd UChi_10;
REGISTER Simd UChi_11;
REGISTER Simd UChi_12; // 8 left
REGISTER Simd U_00; // two rows of U matrix
REGISTER Simd U_10;
REGISTER Simd U_20;
REGISTER Simd U_01;
REGISTER Simd U_11;
REGISTER Simd U_21; // 2 reg left.
#define Chimu_00 Chi_00
#define Chimu_01 Chi_01
#define Chimu_02 Chi_02
#define Chimu_10 Chi_10
#define Chimu_11 Chi_11
#define Chimu_12 Chi_12
#define Chimu_20 UChi_00
#define Chimu_21 UChi_01
#define Chimu_22 UChi_02
#define Chimu_30 UChi_10
#define Chimu_31 UChi_11
#define Chimu_32 UChi_12
StencilEntry *SE;
int offset, ptype;
int num = 0;
// Xp
SE=st.GetEntry(ptype,Xp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
XP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Xp);
XP_RECON_ACCUM;
num++;
}
// Yp
SE=st.GetEntry(ptype,Yp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
YP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Yp);
YP_RECON_ACCUM;
num++;
}
// Zp
SE=st.GetEntry(ptype,Zp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
ZP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Zp);
ZP_RECON_ACCUM;
num++;
}
// Tp
SE=st.GetEntry(ptype,Tp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
TP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Tp);
TP_RECON_ACCUM;
num++;
}
// Xm
SE=st.GetEntry(ptype,Xm,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
XM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Xm);
XM_RECON_ACCUM;
num++;
}
// Ym
SE=st.GetEntry(ptype,Ym,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
YM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Ym);
YM_RECON_ACCUM;
num++;
}
// Zm
SE=st.GetEntry(ptype,Zm,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
ZM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Zm);
ZM_RECON_ACCUM;
num++;
}
// Tm
SE=st.GetEntry(ptype,Tm,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
TM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Tm);
TM_RECON_ACCUM;
num++;
}
SiteSpinor & ref (out._odata[ss]);
if ( Local ) {
vstream(ref()(0)(0),result_00*(-0.5));
vstream(ref()(0)(1),result_01*(-0.5));
vstream(ref()(0)(2),result_02*(-0.5));
vstream(ref()(1)(0),result_10*(-0.5));
vstream(ref()(1)(1),result_11*(-0.5));
vstream(ref()(1)(2),result_12*(-0.5));
vstream(ref()(2)(0),result_20*(-0.5));
vstream(ref()(2)(1),result_21*(-0.5));
vstream(ref()(2)(2),result_22*(-0.5));
vstream(ref()(3)(0),result_30*(-0.5));
vstream(ref()(3)(1),result_31*(-0.5));
vstream(ref()(3)(2),result_32*(-0.5));
return 1;
} else if ( num ) {
vstream(ref()(0)(0),ref()(0)(0)+result_00*(-0.5));
vstream(ref()(0)(1),ref()(0)(1)+result_01*(-0.5));
vstream(ref()(0)(2),ref()(0)(2)+result_02*(-0.5));
vstream(ref()(1)(0),ref()(1)(0)+result_10*(-0.5));
vstream(ref()(1)(1),ref()(1)(1)+result_11*(-0.5));
vstream(ref()(1)(2),ref()(1)(2)+result_12*(-0.5));
vstream(ref()(2)(0),ref()(2)(0)+result_20*(-0.5));
vstream(ref()(2)(1),ref()(2)(1)+result_21*(-0.5));
vstream(ref()(2)(2),ref()(2)(2)+result_22*(-0.5));
vstream(ref()(3)(0),ref()(3)(0)+result_30*(-0.5));
vstream(ref()(3)(1),ref()(3)(1)+result_31*(-0.5));
vstream(ref()(3)(2),ref()(3)(2)+result_32*(-0.5));
return 1;
}
return 0;
}
template<class Impl>
int WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
{
// std::cout << "Hand op Dhop "<<std::endl;
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
REGISTER Simd result_00 ; zeroit(result_00); // 12 regs on knc
REGISTER Simd result_01 ; zeroit(result_01); // 12 regs on knc
REGISTER Simd result_02 ; zeroit(result_02); // 12 regs on knc
REGISTER Simd result_10 ; zeroit(result_10); // 12 regs on knc
REGISTER Simd result_11 ; zeroit(result_11); // 12 regs on knc
REGISTER Simd result_12 ; zeroit(result_12); // 12 regs on knc
REGISTER Simd result_20 ; zeroit(result_20); // 12 regs on knc
REGISTER Simd result_21 ; zeroit(result_21); // 12 regs on knc
REGISTER Simd result_22 ; zeroit(result_22); // 12 regs on knc
REGISTER Simd result_30 ; zeroit(result_30); // 12 regs on knc
REGISTER Simd result_31 ; zeroit(result_31); // 12 regs on knc
REGISTER Simd result_32 ; zeroit(result_32); // 12 regs on knc
REGISTER Simd Chi_00; // two spinor; 6 regs
REGISTER Simd Chi_01;
REGISTER Simd Chi_02;
REGISTER Simd Chi_10;
REGISTER Simd Chi_11;
REGISTER Simd Chi_12; // 14 left
REGISTER Simd UChi_00; // two spinor; 6 regs
REGISTER Simd UChi_01;
REGISTER Simd UChi_02;
REGISTER Simd UChi_10;
REGISTER Simd UChi_11;
REGISTER Simd UChi_12; // 8 left
REGISTER Simd U_00; // two rows of U matrix
REGISTER Simd U_10;
REGISTER Simd U_20;
REGISTER Simd U_01;
REGISTER Simd U_11;
REGISTER Simd U_21; // 2 reg left.
#define Chimu_00 Chi_00
#define Chimu_01 Chi_01
#define Chimu_02 Chi_02
#define Chimu_10 Chi_10
#define Chimu_11 Chi_11
#define Chimu_12 Chi_12
#define Chimu_20 UChi_00
#define Chimu_21 UChi_01
#define Chimu_22 UChi_02
#define Chimu_30 UChi_10
#define Chimu_31 UChi_11
#define Chimu_32 UChi_12
StencilEntry *SE;
int offset, ptype;
int num = 0;
// Xp
SE=st.GetEntry(ptype,Xp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
XM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Xp);
XM_RECON_ACCUM;
num++;
}
// Yp
SE=st.GetEntry(ptype,Yp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
YM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Yp);
YM_RECON_ACCUM;
num++;
}
// Zp
SE=st.GetEntry(ptype,Zp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
ZM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Zp);
ZM_RECON_ACCUM;
num++;
}
// Tp
SE=st.GetEntry(ptype,Tp,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
TM_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Tp);
TM_RECON_ACCUM;
num++;
}
// Xm
SE=st.GetEntry(ptype,Xm,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
XP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Xm);
XP_RECON_ACCUM;
num++;
}
// Ym
SE=st.GetEntry(ptype,Ym,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
YP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Ym);
YP_RECON_ACCUM;
num++;
}
// Zm
SE=st.GetEntry(ptype,Zm,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
ZP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Zm);
ZP_RECON_ACCUM;
num++;
}
// Tm
SE=st.GetEntry(ptype,Tm,ss);
offset = SE->_offset;
if (Local && SE->_is_local ) {
LOAD_CHIMU;
TP_PROJ;
if ( SE->_permute ) {
PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
}
if ( Nonlocal && (!SE->_is_local) ) {
LOAD_CHI;
}
if ( (Local && SE->_is_local) || ( Nonlocal && (!SE->_is_local)) ) {
MULT_2SPIN(Tm);
TP_RECON_ACCUM;
num++;
}
SiteSpinor & ref (out._odata[ss]);
if ( Local ) {
vstream(ref()(0)(0),result_00*(-0.5));
vstream(ref()(0)(1),result_01*(-0.5));
vstream(ref()(0)(2),result_02*(-0.5));
vstream(ref()(1)(0),result_10*(-0.5));
vstream(ref()(1)(1),result_11*(-0.5));
vstream(ref()(1)(2),result_12*(-0.5));
vstream(ref()(2)(0),result_20*(-0.5));
vstream(ref()(2)(1),result_21*(-0.5));
vstream(ref()(2)(2),result_22*(-0.5));
vstream(ref()(3)(0),result_30*(-0.5));
vstream(ref()(3)(1),result_31*(-0.5));
vstream(ref()(3)(2),result_32*(-0.5));
return 1;
} else if ( num ) {
vstream(ref()(0)(0),ref()(0)(0)+result_00*(-0.5));
vstream(ref()(0)(1),ref()(0)(1)+result_01*(-0.5));
vstream(ref()(0)(2),ref()(0)(2)+result_02*(-0.5));
vstream(ref()(1)(0),ref()(1)(0)+result_10*(-0.5));
vstream(ref()(1)(1),ref()(1)(1)+result_11*(-0.5));
vstream(ref()(1)(2),ref()(1)(2)+result_12*(-0.5));
vstream(ref()(2)(0),ref()(2)(0)+result_20*(-0.5));
vstream(ref()(2)(1),ref()(2)(1)+result_21*(-0.5));
vstream(ref()(2)(2),ref()(2)(2)+result_22*(-0.5));
vstream(ref()(3)(0),ref()(3)(0)+result_30*(-0.5));
vstream(ref()(3)(1),ref()(3)(1)+result_31*(-0.5));
vstream(ref()(3)(2),ref()(3)(2)+result_32*(-0.5));
return 1;
}
return 0;
}
/*
template<class Impl>
void WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out)
{
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
@ -1073,89 +539,346 @@ void WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeFiel
{
SiteSpinor & ref (out._odata[ss]);
vstream(ref()(0)(0),result_00*(-0.5));
vstream(ref()(0)(1),result_01*(-0.5));
vstream(ref()(0)(2),result_02*(-0.5));
vstream(ref()(1)(0),result_10*(-0.5));
vstream(ref()(1)(1),result_11*(-0.5));
vstream(ref()(1)(2),result_12*(-0.5));
vstream(ref()(2)(0),result_20*(-0.5));
vstream(ref()(2)(1),result_21*(-0.5));
vstream(ref()(2)(2),result_22*(-0.5));
vstream(ref()(3)(0),result_30*(-0.5));
vstream(ref()(3)(1),result_31*(-0.5));
vstream(ref()(3)(2),result_32*(-0.5));
vstream(ref()(0)(0),result_00);
vstream(ref()(0)(1),result_01);
vstream(ref()(0)(2),result_02);
vstream(ref()(1)(0),result_10);
vstream(ref()(1)(1),result_11);
vstream(ref()(1)(2),result_12);
vstream(ref()(2)(0),result_20);
vstream(ref()(2)(1),result_21);
vstream(ref()(2)(2),result_22);
vstream(ref()(3)(0),result_30);
vstream(ref()(3)(1),result_31);
vstream(ref()(3)(2),result_32);
}
}
*/
template<class Impl>
void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out)
{
// std::cout << "Hand op Dhop "<<std::endl;
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
REGISTER Simd result_00; // 12 regs on knc
REGISTER Simd result_01;
REGISTER Simd result_02;
REGISTER Simd result_10;
REGISTER Simd result_11;
REGISTER Simd result_12;
REGISTER Simd result_20;
REGISTER Simd result_21;
REGISTER Simd result_22;
REGISTER Simd result_30;
REGISTER Simd result_31;
REGISTER Simd result_32; // 20 left
REGISTER Simd Chi_00; // two spinor; 6 regs
REGISTER Simd Chi_01;
REGISTER Simd Chi_02;
REGISTER Simd Chi_10;
REGISTER Simd Chi_11;
REGISTER Simd Chi_12; // 14 left
REGISTER Simd UChi_00; // two spinor; 6 regs
REGISTER Simd UChi_01;
REGISTER Simd UChi_02;
REGISTER Simd UChi_10;
REGISTER Simd UChi_11;
REGISTER Simd UChi_12; // 8 left
REGISTER Simd U_00; // two rows of U matrix
REGISTER Simd U_10;
REGISTER Simd U_20;
REGISTER Simd U_01;
REGISTER Simd U_11;
REGISTER Simd U_21; // 2 reg left.
#define Chimu_00 Chi_00
#define Chimu_01 Chi_01
#define Chimu_02 Chi_02
#define Chimu_10 Chi_10
#define Chimu_11 Chi_11
#define Chimu_12 Chi_12
#define Chimu_20 UChi_00
#define Chimu_21 UChi_01
#define Chimu_22 UChi_02
#define Chimu_30 UChi_10
#define Chimu_31 UChi_11
#define Chimu_32 UChi_12
StencilEntry *SE;
int offset,local,perm, ptype;
// Xp
SE=st.GetEntry(ptype,Xp,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
XP_PROJ;
if ( perm) {
PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Xp);
}
XP_RECON;
// Yp
SE=st.GetEntry(ptype,Yp,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
YP_PROJ;
if ( perm) {
PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Yp);
}
YP_RECON_ACCUM;
// Zp
SE=st.GetEntry(ptype,Zp,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
ZP_PROJ;
if ( perm) {
PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Zp);
}
ZP_RECON_ACCUM;
// Tp
SE=st.GetEntry(ptype,Tp,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
TP_PROJ;
if ( perm) {
PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Tp);
}
TP_RECON_ACCUM;
// Xm
SE=st.GetEntry(ptype,Xm,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
XM_PROJ;
if ( perm) {
PERMUTE_DIR(3); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Xm);
}
XM_RECON_ACCUM;
// Ym
SE=st.GetEntry(ptype,Ym,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
YM_PROJ;
if ( perm) {
PERMUTE_DIR(2); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Ym);
}
YM_RECON_ACCUM;
// Zm
SE=st.GetEntry(ptype,Zm,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
ZM_PROJ;
if ( perm) {
PERMUTE_DIR(1); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Zm);
}
ZM_RECON_ACCUM;
// Tm
SE=st.GetEntry(ptype,Tm,ss);
offset = SE->_offset;
local = SE->_is_local;
perm = SE->_permute;
if ( local ) {
LOAD_CHIMU;
TM_PROJ;
if ( perm) {
PERMUTE_DIR(0); // T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
}
} else {
LOAD_CHI;
}
{
MULT_2SPIN(Tm);
}
TM_RECON_ACCUM;
{
SiteSpinor & ref (out._odata[ss]);
vstream(ref()(0)(0),result_00);
vstream(ref()(0)(1),result_01);
vstream(ref()(0)(2),result_02);
vstream(ref()(1)(0),result_10);
vstream(ref()(1)(1),result_11);
vstream(ref()(1)(2),result_12);
vstream(ref()(2)(0),result_20);
vstream(ref()(2)(1),result_21);
vstream(ref()(2)(2),result_22);
vstream(ref()(3)(0),result_30);
vstream(ref()(3)(1),result_31);
vstream(ref()(3)(2),result_32);
}
}
////////////////////////////////////////////////
// Specialise Gparity to simple implementation
////////////////////////////////////////////////
template<>
int WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
int sF,int sU,const FermionField &in, FermionField &out)
{
DiracOptDhopSite(st,U,buf,sF,sU,in,out); // returns void, will template override for Wilson Nc=3
//check consistency of return types between these functions and the ones in WilsonKernels.cc
return 0;
assert(0);
}
template<>
int WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
int sF,int sU,const FermionField &in, FermionField &out)
{
DiracOptDhopSiteDag(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
return 0;
assert(0);
}
template<>
int WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
int sF,int sU,const FermionField &in, FermionField &out)
{
DiracOptDhopSite(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
return 0;
assert(0);
}
template<>
int WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
int sF,int sU,const FermionField &in, FermionField &out)
{
DiracOptDhopSiteDag(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
return 0;
assert(0);
}
template int WilsonKernels<WilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
////////////// Wilson ; uses this implementation /////////////////////
// Need Nc=3 though //
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out,bool l,bool n);
template int WilsonKernels<WilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool n);
template int WilsonKernels<WilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool n);
template int WilsonKernels<WilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool n);
int ss,int sU,const FermionField &in, FermionField &out);
template int WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
template int WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
template int WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
template int WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSite(StencilImpl &st,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,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template void WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,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,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
}}

8
lib/qcd/action/gauge/GaugeImpl.h
View File

@ -42,7 +42,9 @@ template<class Gimpl> class WilsonLoops;
#define INHERIT_GIMPL_TYPES(GImpl) \
typedef typename GImpl::Simd Simd;\
typedef typename GImpl::GaugeLinkField GaugeLinkField;\
typedef typename GImpl::GaugeField GaugeField;
typedef typename GImpl::GaugeField GaugeField;\
typedef typename GImpl::SiteGaugeField SiteGaugeField;\
typedef typename GImpl::SiteGaugeLink SiteGaugeLink;
//
template<class S,int Nrepresentation=Nc>
@ -62,9 +64,9 @@ template<class Gimpl> class WilsonLoops;
// Move this elsewhere?
static inline void AddGaugeLink(GaugeField& U, GaugeLinkField& W, int mu){ // U[mu] += W
PARALLEL_FOR_LOOP
PARALLEL_FOR_LOOP
for(auto ss=0;ss<U._grid->oSites();ss++){
U._odata[ss]._internal[mu] = U._odata[ss]._internal[mu] + W._odata[ss]._internal;
U._odata[ss]._internal[mu] = U._odata[ss]._internal[mu] + W._odata[ss]._internal;
}
}