1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-25 19:22:03 +01:00

Change to interface to minise comms in evaluating coarse space operator

This commit is contained in:
Peter Boyle
2020-01-06 11:43:59 -05:00
parent 3c3d6a94f3
commit e583035614
20 changed files with 262 additions and 92 deletions

View File

@ -389,6 +389,14 @@ void CayleyFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,in
Meo5D(psi,tmp);
this->DhopDir(tmp,chi,dir,disp);
}
template<class Impl>
void CayleyFermion5D<Impl>::MdirAll(const FermionField &psi, std::vector<FermionField> &out)
{
FermionField tmp(psi.Grid());
Meo5D(psi,tmp);
this->DhopDirAll(tmp,out);
}
// force terms; five routines; default to Dhop on diagonal
template<class Impl>
void CayleyFermion5D<Impl>::MDeriv (GaugeField &mat,const FermionField &U,const FermionField &V,int dag)

View File

@ -143,6 +143,25 @@ void ContinuedFractionFermion5D<Impl>::Mdir (const FermionField &psi, FermionFi
}
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MdirAll (const FermionField &psi, std::vector<FermionField> &chi)
{
int Ls = this->Ls;
this->DhopDirAll(psi,chi); // Dslash on diagonal. g5 Dslash is hermitian
for(int p=0;p<chi.size();p++){
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==(Ls-1) ){
ag5xpby_ssp(chi[p],Beta[s]*ZoloHiInv,chi[p],0.0,chi[p],s,s);
} else {
ag5xpby_ssp(chi[p],cc[s]*Beta[s]*sign*ZoloHiInv,chi[p],0.0,chi[p],s,s);
}
sign=-sign;
}
}
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::Meooe (const FermionField &psi, FermionField &chi)
{
int Ls = this->Ls;

View File

@ -538,10 +538,16 @@ void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void)
// Implement the general interface. Here we use SAME mass on all slices
/////////////////////////////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) {
void ImprovedStaggeredFermion5D<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp)
{
DhopDir(in, out, dir, disp);
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out)
{
assert(0);
}
template <class Impl>
RealD ImprovedStaggeredFermion5D<Impl>::M(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);

View File

@ -362,12 +362,19 @@ void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) {
void ImprovedStaggeredFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp)
{
DhopDir(in, out, dir, disp);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out)
{
assert(0); // Not implemented yet
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp) {
void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp)
{
Compressor compressor;
Stencil.HaloExchange(in, compressor);
@ -380,6 +387,7 @@ void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionFiel
});
};
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,

View File

@ -31,7 +31,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
NAMESPACE_BEGIN(Grid);
template<class Impl>
template<class Impl>
void PartialFractionFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,int dir,int disp){
// this does both dag and undag but is trivial; make a common helper routing
int Ls = this->Ls;
@ -45,8 +45,25 @@ void PartialFractionFermion5D<Impl>::Mdir (const FermionField &psi, FermionFiel
ag5xpby_ssp(chi, scale,chi,0.0,chi,s+1,s+1);
}
ag5xpby_ssp(chi,p[nblock]*scale/amax,chi,0.0,chi,Ls-1,Ls-1);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::MdirAll (const FermionField &psi, std::vector<FermionField> &chi){
// this does both dag and undag but is trivial; make a common helper routing
int Ls = this->Ls;
this->DhopDirAll(psi,chi);
for(int point=0;point<chi.size();point++){
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
ag5xpby_ssp(chi[point],-scale,chi[point],0.0,chi[point],s,s);
ag5xpby_ssp(chi[point], scale,chi[point],0.0,chi[point],s+1,s+1);
}
ag5xpby_ssp(chi[point],p[nblock]*scale/amax,chi[point],0.0,chi[point],Ls-1,Ls-1);
}
}
template<class Impl>
void PartialFractionFermion5D<Impl>::Meooe_internal(const FermionField &psi, FermionField &chi,int dag)
{

View File

@ -241,6 +241,30 @@ void WilsonFermion5D<Impl>::DhopDir(const FermionField &in, FermionField &out,in
Kernels::DhopDirKernel(Stencil,Umu,Stencil.CommBuf(),Ls,Nsite,in,out,dirdisp,gamma);
};
template<class Impl>
void WilsonFermion5D<Impl>::DhopDirAll(const FermionField &in, std::vector<FermionField> &out)
{
Compressor compressor(DaggerNo);
Stencil.HaloExchange(in,compressor);
assert( (out.size()==8)||(out.size()==9));
for(int dir5=1;dir5<=4;dir5++) {
for(int disp=-1;disp<=1;disp+=2){
int dir = dir5-1; // Maps to the ordering above in "directions" that is passed to stencil
// we drop off the innermost fifth dimension
assert( (disp==1)||(disp==-1) );
assert( (dir>=0)&&(dir<4) ); //must do x,y,z or t;
int skip = (disp==1) ? 0 : 1;
int dirdisp = dir+skip*4;
int gamma = dir+(1-skip)*4;
uint64_t Nsite = Umu.Grid()->oSites();
Kernels::DhopDirKernel(Stencil,Umu,Stencil.CommBuf(),Ls,Nsite,in,out[dirdisp],dirdisp,gamma);
}
}
};
template<class Impl>
void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,

View File

@ -319,28 +319,51 @@ void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int d
}
template <class Impl>
void WilsonFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) {
void WilsonFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp)
{
DhopDir(in, out, dir, disp);
}
template <class Impl>
void WilsonFermion<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out)
{
DhopDirAll(in, out);
}
template <class Impl>
void WilsonFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp)
{
Compressor compressor(DaggerNo);
Stencil.HaloExchange(in, compressor);
int skip = (disp == 1) ? 0 : 1;
int dirdisp = dir + skip * 4;
int gamma = dir + (1 - skip) * 4;
DhopDirDisp(in, out, dirdisp, gamma, DaggerNo);
DhopDirCalc(in, out, dirdisp, gamma, DaggerNo);
};
template <class Impl>
void WilsonFermion<Impl>::DhopDirDisp(const FermionField &in, FermionField &out,int dirdisp, int gamma, int dag)
void WilsonFermion<Impl>::DhopDirAll(const FermionField &in, std::vector<FermionField> &out)
{
Compressor compressor(dag);
Compressor compressor(DaggerNo);
Stencil.HaloExchange(in, compressor);
assert((out.size()==8)||(out.size()==9));
for(int dir=0;dir<Nd;dir++){
for(int disp=-1;disp<=1;disp+=2){
int skip = (disp == 1) ? 0 : 1;
int dirdisp = dir + skip * 4;
int gamma = dir + (1 - skip) * 4;
DhopDirCalc(in, out[dirdisp], dirdisp, gamma, DaggerNo);
}
}
}
template <class Impl>
void WilsonFermion<Impl>::DhopDirCalc(const FermionField &in, FermionField &out,int dirdisp, int gamma, int dag)
{
int Ls=1;
int Nsite=in.oSites();
uint64_t Nsite=in.oSites();
Kernels::DhopDirKernel(Stencil, Umu, Stencil.CommBuf(), Ls, Nsite, in, out, dirdisp, gamma);
};
@ -348,7 +371,8 @@ template <class Impl>
void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag) {
FermionField &out, int dag)
{
#ifdef GRID_OMP
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,in,out,dag);