portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 04:37:05 +01:00
Code Releases Activity

Added overlap comms compute mode

Browse Source
This commit is contained in:
paboyle
2016-01-03 01:38:11 +00:00
parent 4aac345bea
commit 331768dcff
13 changed files with 704 additions and 270 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
lib/qcd/action/fermion/FermionOperatorImpl.h
View File

@ -110,10 +110,10 @@ namespace Grid {
// Single flavour four spinors with colour index
///////
template<class S,int Nrepresentation=Nc>
class WilsonImpl : public PeriodicGaugeImpl<S,Nrepresentation> {
class WilsonImpl : public PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > {
public:
typedef PeriodicGaugeImpl<S,Nrepresentation> Gimpl;
typedef PeriodicGaugeImpl< GaugeImplTypes< S,Nrepresentation> > Gimpl;
INHERIT_GIMPL_TYPES(Gimpl);
@ -182,10 +182,10 @@ PARALLEL_FOR_LOOP
////////////////////////////////////////////////////////////////////////////////////////
template<class S,int Nrepresentation>
class GparityWilsonImpl : public ConjugateGaugeImpl<S,Nrepresentation> {
class GparityWilsonImpl : public ConjugateGaugeImpl< GaugeImplTypes<S,Nrepresentation> >{
public:
typedef ConjugateGaugeImpl<S,Nrepresentation> Gimpl;
typedef ConjugateGaugeImpl< GaugeImplTypes<S,Nrepresentation> > Gimpl;
INHERIT_GIMPL_TYPES(Gimpl);

78
lib/qcd/action/fermion/WilsonFermion.cc
View File

@ -227,7 +227,7 @@ PARALLEL_FOR_LOOP
out.checkerboard = in.checkerboard;
DhopInternal(Stencil,Umu,in,out,dag);
DhopInternalCommsCompute(Stencil,Umu,in,out,dag);
}
template<class Impl>
@ -238,7 +238,7 @@ PARALLEL_FOR_LOOP
assert(in.checkerboard==Even);
out.checkerboard = Odd;
DhopInternal(StencilEven,UmuOdd,in,out,dag);
DhopInternalCommsCompute(StencilEven,UmuOdd,in,out,dag);
}
template<class Impl>
@ -249,7 +249,7 @@ PARALLEL_FOR_LOOP
assert(in.checkerboard==Odd);
out.checkerboard = Even;
DhopInternal(StencilOdd,UmuEven,in,out,dag);
DhopInternalCommsCompute(StencilOdd,UmuEven,in,out,dag);
}
template<class Impl>
@ -319,6 +319,78 @@ PARALLEL_FOR_LOOP
}
}
};
template<class Impl>
void WilsonFermion<Impl>::DhopInternalCommsCompute(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) {
assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor(dag);
std::thread comms_thread = st.HaloExchangeBegin(in,comm_buf,compressor);
comms_thread.join();
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,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,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,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,comm_buf,sss,sss,in,out,local,nonlocal);
}
}
}
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,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,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,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,comm_buf,sss,sss,in,out,local,nonlocal);
}
}
}
};
FermOpTemplateInstantiate(WilsonFermion);

3
lib/qcd/action/fermion/WilsonFermion.h
View File

@ -114,6 +114,9 @@ namespace Grid {
void DhopInternal(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) ;
void DhopInternalCommsCompute(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) ;
// Constructor
WilsonFermion(GaugeField &_Umu,

137
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -396,6 +396,137 @@ PARALLEL_FOR_LOOP
}
dslashtime +=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalCommsCompute(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
// assert((dag==DaggerNo) ||(dag==DaggerYes));
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();
std::thread thr = st.HaloExchangeBegin(in,comm_buf,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,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,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,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,comm_buf,sF,sU,in,out,local,nonlocal);
}
}
}
}
dslashtime +=usecond();
commtime -=usecond();
thr.join();
commtime +=usecond();
local = false;
nonlocal = true;
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,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,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,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,comm_buf,sF,sU,in,out,local,nonlocal);
}
}
}
}
dslashtime +=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
{
@ -405,7 +536,7 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
assert(in.checkerboard==Even);
out.checkerboard = Odd;
DhopInternal(StencilEven,LebesgueEvenOdd,UmuOdd,in,out,dag);
DhopInternalCommsCompute(StencilEven,LebesgueEvenOdd,UmuOdd,in,out,dag);
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
@ -416,7 +547,7 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
assert(in.checkerboard==Odd);
out.checkerboard = Even;
DhopInternal(StencilOdd,LebesgueEvenOdd,UmuEven,in,out,dag);
DhopInternalCommsCompute(StencilOdd,LebesgueEvenOdd,UmuEven,in,out,dag);
}
template<class Impl>
void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
@ -426,7 +557,7 @@ void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int d
out.checkerboard = in.checkerboard;
DhopInternal(Stencil,Lebesgue,Umu,in,out,dag);
DhopInternalCommsCompute(Stencil,Lebesgue,Umu,in,out,dag);
}
template<class Impl>
void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag)

7
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@ -117,6 +117,13 @@ namespace Grid {
FermionField &out,
int dag);
void DhopInternalCommsCompute(StencilImpl & st,
LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out,
int dag);
// Constructors
WilsonFermion5D(GaugeField &_Umu,
GridCartesian &FiveDimGrid,

567
lib/qcd/action/fermion/WilsonKernels.cc
View File

@ -34,10 +34,11 @@ namespace QCD {
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>::DiracOptDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
int sF,int sU,const FermionField &in, FermionField &out,bool local, bool nonlocal)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
@ -46,231 +47,425 @@ void WilsonKernels<Impl>::DiracOptDhopSiteDag(StencilImpl &st,DoubledGaugeField
StencilEntry *SE;
int ptype;
int num = 0;
result=zero;
///////////////////////////
// Xp
///////////////////////////
SE=st.GetEntry(ptype,Xp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjXp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjXp(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Xp,SE,st);
spReconXp(result,Uchi);
// Yp
SE=st.GetEntry(ptype,Yp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjYp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjYp(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Yp,SE,st);
accumReconYp(result,Uchi);
// Zp
SE=st.GetEntry(ptype,Zp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjZp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjZp(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjXp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXp(chi,in._odata[SE->_offset]);
}
}
Impl::multLink(Uchi,U._odata[sU],chi,Zp,SE,st);
accumReconZp(result,Uchi);
// Tp
SE=st.GetEntry(ptype,Tp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjTp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjTp(chi,in._odata[SE->_offset]);
} else {
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Tp,SE,st);
accumReconTp(result,Uchi);
// Xm
SE=st.GetEntry(ptype,Xm,sF);
if ( SE->_is_local && SE->_permute ) {
spProjXm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjXm(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Xm,SE,st);
accumReconXm(result,Uchi);
// Ym
SE=st.GetEntry(ptype,Ym,sF);
if ( SE->_is_local && SE->_permute ) {
spProjYm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjYm(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Ym,SE,st);
accumReconYm(result,Uchi);
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Xp,SE,st);
accumReconXp(result,Uchi);
num++;
}
///////////////////////////
// Yp
///////////////////////////
SE=st.GetEntry(ptype,Yp,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjYp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYp(chi,in._odata[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++;
}
///////////////////////////
// Zp
///////////////////////////
SE=st.GetEntry(ptype,Zp,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjZp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZp(chi,in._odata[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++;
}
///////////////////////////
// Tp
///////////////////////////
SE=st.GetEntry(ptype,Tp,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjTp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTp(chi,in._odata[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++;
}
///////////////////////////
// Xm
///////////////////////////
SE=st.GetEntry(ptype,Xm,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjXm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXm(chi,in._odata[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++;
}
///////////////////////////
// Ym
///////////////////////////
SE=st.GetEntry(ptype,Ym,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjYm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYm(chi,in._odata[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++;
}
///////////////////////////
// Zm
///////////////////////////
SE=st.GetEntry(ptype,Zm,sF);
if ( SE->_is_local && SE->_permute ) {
spProjZm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjZm(chi,in._odata[SE->_offset]);
} else {
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjZm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZm(chi,in._odata[SE->_offset]);
}
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Zm,SE,st);
accumReconZm(result,Uchi);
if( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Zm,SE,st);
accumReconZm(result,Uchi);
num++;
}
///////////////////////////
// Tm
///////////////////////////
SE=st.GetEntry(ptype,Tm,sF);
if ( SE->_is_local && SE->_permute ) {
spProjTm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjTm(chi,in._odata[SE->_offset]);
} else {
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjTm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTm(chi,in._odata[SE->_offset]);
}
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Tm,SE,st);
accumReconTm(result,Uchi);
vstream(out._odata[sF],result*(-0.5));
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));
}
};
// 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)
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,bool local, bool nonlocal)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteSpinor result;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
int num = 0;
result=zero;
///////////////////////////
// Xp
///////////////////////////
SE=st.GetEntry(ptype,Xm,sF);
if ( SE->_is_local && SE->_permute ) {
spProjXp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjXp(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Xm,SE,st);
spReconXp(result,Uchi);
// Yp
SE=st.GetEntry(ptype,Ym,sF);
if ( SE->_is_local && SE->_permute ) {
spProjYp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjYp(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjXp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXp(chi,in._odata[SE->_offset]);
}
}
Impl::multLink(Uchi,U._odata[sU],chi,Ym,SE,st);
accumReconYp(result,Uchi);
// Zp
SE=st.GetEntry(ptype,Zm,sF);
if ( SE->_is_local && SE->_permute ) {
spProjZp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjZp(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Zm,SE,st);
accumReconZp(result,Uchi);
// Tp
SE=st.GetEntry(ptype,Tm,sF);
if ( SE->_is_local && SE->_permute ) {
spProjTp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjTp(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Tm,SE,st);
accumReconTp(result,Uchi);
// Xm
SE=st.GetEntry(ptype,Xp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjXm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjXm(chi,in._odata[SE->_offset]);
} else {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Xp,SE,st);
accumReconXm(result,Uchi);
// Ym
SE=st.GetEntry(ptype,Yp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjYm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjYm(chi,in._odata[SE->_offset]);
} else {
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Yp,SE,st);
accumReconYm(result,Uchi);
// Zm
SE=st.GetEntry(ptype,Zp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjZm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjZm(chi,in._odata[SE->_offset]);
} else {
chi=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,Zp,SE,st);
accumReconZm(result,Uchi);
// Tm
SE=st.GetEntry(ptype,Tp,sF);
if ( SE->_is_local && SE->_permute ) {
spProjTm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else if ( SE->_is_local ) {
spProjTm(chi,in._odata[SE->_offset]);
} else {
///////////////////////////
// Yp
///////////////////////////
SE=st.GetEntry(ptype,Ym,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjYp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYp(chi,in._odata[SE->_offset]);
}
}
if ( nonlocal && (!SE->_is_local) ) {
chi=buf[SE->_offset];
}
Impl::multLink(Uchi,U._odata[sU],chi,Tp,SE,st);
accumReconTm(result,Uchi);
if ( (local && SE->_is_local) || ( nonlocal && (!SE->_is_local)) ) {
Impl::multLink(Uchi,U._odata[sU],chi,Ym,SE,st);
accumReconYp(result,Uchi);
num++;
}
///////////////////////////
// Zp
///////////////////////////
SE=st.GetEntry(ptype,Zm,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjZp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZp(chi,in._odata[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++;
}
///////////////////////////
// Tp
///////////////////////////
SE=st.GetEntry(ptype,Tm,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjTp(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTp(chi,in._odata[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++;
}
///////////////////////////
// Xm
///////////////////////////
SE=st.GetEntry(ptype,Xp,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjXm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjXm(chi,in._odata[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++;
}
///////////////////////////
// Ym
///////////////////////////
SE=st.GetEntry(ptype,Yp,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjYm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjYm(chi,in._odata[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++;
}
vstream(out._odata[sF],result*(-0.5));
}
///////////////////////////
// Zm
///////////////////////////
SE=st.GetEntry(ptype,Zp,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjZm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjZm(chi,in._odata[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++;
}
///////////////////////////
// Tm
///////////////////////////
SE=st.GetEntry(ptype,Tp,sF);
if (local && SE->_is_local ) {
if ( SE->_permute ) {
spProjTm(tmp,in._odata[SE->_offset]);
permute(chi,tmp,ptype);
} else {
spProjTm(chi,in._odata[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);
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));
}
};
template<class Impl>
void WilsonKernels<Impl>::DiracOptDhopDir(StencilImpl &st,DoubledGaugeField &U,
@ -405,7 +600,7 @@ void WilsonKernels<Impl>::DiracOptDhopDir(StencilImpl &st,DoubledGaugeField &U,
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,uint64_t *p)
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
}

10
lib/qcd/action/fermion/WilsonKernels.h
View File

@ -48,11 +48,11 @@ namespace Grid {
public:
void DiracOptDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out);
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
void DiracOptDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in,FermionField &out);
int sF,int sU,const FermionField &in,FermionField &out,bool local= true, bool nonlocal=true);
void DiracOptDhopDir(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
@ -60,15 +60,15 @@ namespace Grid {
void DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,uint64_t *);
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
void DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out);
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
void DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out);
int sF,int sU,const FermionField &in, FermionField &out,bool local= true, bool nonlocal=true);
WilsonKernels(const ImplParams &p= ImplParams());

56
lib/qcd/action/fermion/WilsonKernelsHand.cc
View File

@ -311,8 +311,8 @@ namespace QCD {
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::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;
@ -555,8 +555,8 @@ void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeF
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)
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
{
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
@ -801,32 +801,32 @@ void WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeFiel
////////////////////////////////////////////////
template<>
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
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
}
template<>
void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
{
DiracOptDhopSiteDag(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
}
template<>
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
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
}
template<>
void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out, bool Local, bool Nonlocal)
{
DiracOptDhopSiteDag(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
}
@ -834,30 +834,30 @@ void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out,bool l,bool n);
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool n);
template void WilsonKernels<WilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool n);
template void WilsonKernels<WilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool n);
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
template void WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
template void WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out, bool l, bool nl);
}}