portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Merge commit '1e554350acae0e67fa7177ed0db9d4f684a54af2'

Browse Source
This commit is contained in:
Antonin Portelli
2016-04-30 00:17:52 -07:00
parent ba09cbae3e 1e554350ac
commit f6c53e5039
33 changed files with 1549 additions and 520 deletions
Download Patch File Download Diff File Expand all files Collapse all files

106
lib/qcd/action/fermion/FermionOperatorImpl.h
View File

@ -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*?
////////////////////////////////////////////////////////////////////////////////////////
@ -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

15
lib/qcd/action/fermion/WilsonFermion.cc
View File

@ -288,11 +288,7 @@ 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);
}
DhopInternalCommsThenCompute(st,U,in,out,dag);
}
template<class Impl>
void WilsonFermion<Impl>::DhopInternalCommsThenCompute(StencilImpl & st,DoubledGaugeField & U,
@ -330,15 +326,6 @@ PARALLEL_FOR_LOOP
}
};
template<class Impl>
void WilsonFermion<Impl>::DhopInternalCommsOverlapCompute(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) {
assert(0);
};
FermOpTemplateInstantiate(WilsonFermion);
GparityFermOpTemplateInstantiate(WilsonFermion);

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

@ -116,9 +116,6 @@ namespace Grid {
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,

415
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -68,10 +68,8 @@ 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
@ -106,11 +104,75 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
dslashtime=0;
dslash1time=0;
}
template<class Impl>
WilsonFermion5D<Impl>::WilsonFermion5D(int simd, GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _M5,const ImplParams &p) :
Kernels(p),
_FiveDimGrid (&FiveDimGrid),
_FiveDimRedBlackGrid(&FiveDimRedBlackGrid),
_FourDimGrid (&FourDimGrid),
_FourDimRedBlackGrid(&FourDimRedBlackGrid),
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(_FourDimRedBlackGrid),
UmuOdd (_FourDimRedBlackGrid),
Lebesgue(_FourDimGrid),
LebesgueEvenOdd(_FourDimRedBlackGrid)
{
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);
assert(FourDimRedBlackGrid._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(FourDimRedBlackGrid._fdimensions[d] ==FourDimGrid._fdimensions[d]);
assert(FiveDimRedBlackGrid._fdimensions[d+1]==FourDimGrid._fdimensions[d]);
assert(FourDimRedBlackGrid._processors[d] ==FourDimGrid._processors[d]);
assert(FiveDimRedBlackGrid._processors[d+1] ==FourDimGrid._processors[d]);
assert(FourDimGrid._simd_layout[d]=1);
assert(FourDimRedBlackGrid._simd_layout[d] ==1);
assert(FourDimRedBlackGrid._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]);
}
// Allocate the required comms buffer
ImportGauge(_Umu);
}
template<class Impl>
void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
{
GaugeField HUmu(_Umu._grid);
HUmu = _Umu*(-0.5);
GaugeField HUmu(_Umu._grid);
HUmu = _Umu*(-0.5);
Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
pickCheckerboard(Even,UmuEven,Umu);
pickCheckerboard(Odd ,UmuOdd,Umu);
@ -294,15 +356,12 @@ void WilsonFermion5D<Impl>::DhopInternalCommsThenCompute(StencilImpl & st, Lebes
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);
// auto handle = st.HaloExchangeBegin(in,compressor);
// st.HaloExchangeComplete(handle);
st.HaloExchange(in,compressor);
commtime +=usecond();
jointime -=usecond();
@ -318,97 +377,48 @@ void WilsonFermion5D<Impl>::DhopInternalCommsThenCompute(StencilImpl & st, Lebes
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;
for(int s=0;s<LLs;s++){
int sU=ss;
int sF = s+LLs*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 s=0;s<LLs;s++){
int sU=ss;
int sF = s+LLs*sU;
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,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);
}
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
for(int s=0;s<LLs;s++){
int sU=ss;
int sF = s+LLs*sU;
Kernels::DiracOptAsmDhopSite(st,U,st.comm_buf,sF,sU,in,out);
}
}
// 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);
}
}
}
*/
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;
for(int s=0;s<LLs;s++){
int sU=ss;
int sF = s+LLs*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;
for(int s=0;s<LLs;s++){
int sU=ss;
int sF = s+LLs*sU;
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,in,out);
}
}
@ -418,251 +428,6 @@ PARALLEL_FOR_LOOP
alltime+=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalOMPbench(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();
commtime -=usecond();
auto handle = st.HaloExchangeBegin(in,compressor);
st.HaloExchangeComplete(handle);
commtime +=usecond();
jointime -=usecond();
jointime +=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.
#pragma omp parallel
{
for(int jjj=0;jjj<100;jjj++){
#pragma omp barrier
dslashtime -=usecond();
if ( dag == DaggerYes ) {
if( this->HandOptDslash ) {
#pragma omp for
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 {
#pragma omp for
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);
}
}
}
}
} 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 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);
}
}
}
// Counter.Stop();
// Counter.Report();
// }
} else if( this->HandOptDslash ) {
#pragma omp for
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 {
#pragma omp for
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);
}
}
}
}
}
}
dslashtime +=usecond();
alltime+=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalL1bench(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();
commtime -=usecond();
auto handle = st.HaloExchangeBegin(in,compressor);
st.HaloExchangeComplete(handle);
commtime +=usecond();
jointime -=usecond();
jointime +=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.
#pragma omp parallel
{
for(int jjj=0;jjj<100;jjj++){
#pragma omp barrier
dslashtime -=usecond();
if ( dag == DaggerYes ) {
if( this->HandOptDslash ) {
#pragma omp for
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=0;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out);
}
}
} else {
#pragma omp for
for(int ss=0;ss<U._grid->oSites();ss++){
{
int sd;
for(sd=0;sd<Ls;sd++){
int sU=0;
int sF = sd+Ls*sU;
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,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 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=0;
sF = s+Ls*sU;
Kernels::DiracOptAsmDhopSite(st,U,st.comm_buf,sF,sU,in,out);
}
}
}
// Counter.Stop();
// Counter.Report();
// }
} else if( this->HandOptDslash ) {
#pragma omp for
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=0;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sF,sU,in,out);
}
}
} else {
#pragma omp for
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=0;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,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(0);
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
@ -706,6 +471,8 @@ void WilsonFermion5D<Impl>::DW(const FermionField &in, FermionField &out,int dag
FermOpTemplateInstantiate(WilsonFermion5D);
GparityFermOpTemplateInstantiate(WilsonFermion5D);
template class WilsonFermion5D<DomainWallRedBlack5dImplF>;
template class WilsonFermion5D<DomainWallRedBlack5dImplD>;
}}

30
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@ -87,6 +87,7 @@ 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);
@ -121,32 +122,12 @@ namespace Grid {
FermionField &out,
int dag);
void DhopInternalOMPbench(StencilImpl & st,
LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out,
int dag);
void DhopInternalL1bench(StencilImpl & st,
LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
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,
@ -156,6 +137,15 @@ namespace Grid {
GridRedBlackCartesian &FourDimRedBlackGrid,
double _M5,const ImplParams &p= ImplParams());
// Constructors
WilsonFermion5D(int simd,
GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
double _M5,const ImplParams &p= ImplParams());
// DoubleStore
void ImportGauge(const GaugeField &_Umu);

2
lib/qcd/action/fermion/WilsonKernels.cc
View File

@ -529,5 +529,7 @@ void WilsonKernels<Impl>::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField
#endif
FermOpTemplateInstantiate(WilsonKernels);
template class WilsonKernels<DomainWallRedBlack5dImplF>;
template class WilsonKernels<DomainWallRedBlack5dImplD>;
}}

4
lib/qcd/action/fermion/WilsonKernelsAsm.cc
View File

@ -27,7 +27,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
*************************************************************************************/
/* END LEGAL */
#include <Grid.h>
#if defined(AVX512) || defined (IMCI)
#if defined(AVX512)
//#if defined (IMCI)
#include <simd/Intel512wilson.h>
@ -256,5 +256,7 @@ void WilsonKernels<Impl >::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField
template class WilsonKernels<WilsonImplD>;
template class WilsonKernels<GparityWilsonImplF>;
template class WilsonKernels<GparityWilsonImplD>;
template class WilsonKernels<DomainWallRedBlack5dImplF>;
template class WilsonKernels<DomainWallRedBlack5dImplD>;
}}
#endif

79
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);\
@ -809,7 +811,6 @@ int WilsonKernels<GparityWilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,Doub
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;
}
@ -843,6 +844,47 @@ int WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,D
//////////////
/*
template<>
int WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
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
return 0;
}
template<>
int WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
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;
}
template<>
int WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
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;
}
template<>
int WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
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;
}
*/
template int WilsonKernels<WilsonImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
@ -870,4 +912,21 @@ template int WilsonKernels<GparityWilsonImplD>::DiracOptHandDhopSiteDag(StencilI
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template int WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template int WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template int WilsonKernels<DomainWallRedBlack5dImplF>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
template int WilsonKernels<DomainWallRedBlack5dImplD>::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out);
}}

4
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;
//

45
lib/qcd/utils/SpaceTimeGrid.cc
View File

@ -41,7 +41,11 @@ GridRedBlackCartesian *SpaceTimeGrid::makeFourDimRedBlackGrid(const GridCartesia
{
return new GridRedBlackCartesian(FourDimGrid);
}
GridCartesian *SpaceTimeGrid::makeFourDimDWFGrid(const std::vector<int> & latt,const std::vector<int> &mpi)
{
std::vector<int> simd(4,1);
return makeFourDimGrid(latt,simd,mpi);
}
GridCartesian *SpaceTimeGrid::makeFiveDimGrid(int Ls,const GridCartesian *FourDimGrid)
{
int N4=FourDimGrid->_ndimension;
@ -58,6 +62,7 @@ GridCartesian *SpaceTimeGrid::makeFiveDimGrid(int Ls,const GridCartesian
return new GridCartesian(latt5,simd5,mpi5);
}
GridRedBlackCartesian *SpaceTimeGrid::makeFiveDimRedBlackGrid(int Ls,const GridCartesian *FourDimGrid)
{
int N4=FourDimGrid->_ndimension;
@ -76,4 +81,42 @@ GridRedBlackCartesian *SpaceTimeGrid::makeFiveDimRedBlackGrid(int Ls,const GridC
return new GridRedBlackCartesian(latt5,simd5,mpi5,cb5,cbd);
}
GridCartesian *SpaceTimeGrid::makeFiveDimDWFGrid(int Ls,const GridCartesian *FourDimGrid)
{
int N4=FourDimGrid->_ndimension;
int nsimd = FourDimGrid->Nsimd();
std::vector<int> latt5(1,Ls);
std::vector<int> simd5(1,nsimd);
std::vector<int> mpi5(1,1);
for(int d=0;d<N4;d++){
latt5.push_back(FourDimGrid->_fdimensions[d]);
simd5.push_back(1);
mpi5.push_back(FourDimGrid->_processors[d]);
}
return new GridCartesian(latt5,simd5,mpi5);
}
GridRedBlackCartesian *SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(int Ls,const GridCartesian *FourDimGrid)
{
int N4=FourDimGrid->_ndimension;
int nsimd = FourDimGrid->Nsimd();
int cbd=0;
std::vector<int> latt5(1,Ls);
std::vector<int> simd5(1,nsimd);
std::vector<int> mpi5(1,1);
std::vector<int> cb5(1,1);
for(int d=0;d<N4;d++){
latt5.push_back(FourDimGrid->_fdimensions[d]);
simd5.push_back(1);
mpi5.push_back(FourDimGrid->_processors[d]);
cb5.push_back(1);
}
return new GridRedBlackCartesian(latt5,simd5,mpi5,cb5,cbd);
}
}}

5
lib/qcd/utils/SpaceTimeGrid.h
View File

@ -35,9 +35,14 @@ class SpaceTimeGrid {
static GridCartesian *makeFourDimGrid(const std::vector<int> & latt,const std::vector<int> &simd,const std::vector<int> &mpi);
static GridRedBlackCartesian *makeFourDimRedBlackGrid (const GridCartesian *FourDimGrid);
static GridCartesian *makeFiveDimGrid (int Ls,const GridCartesian *FourDimGrid);
static GridRedBlackCartesian *makeFiveDimRedBlackGrid(int Ls,const GridCartesian *FourDimGrid);
static GridCartesian *makeFiveDimDWFGrid (int Ls,const GridCartesian *FourDimGrid);
static GridRedBlackCartesian *makeFiveDimDWFRedBlackGrid(int Ls,const GridCartesian *FourDimGrid);
static GridCartesian *makeFourDimDWFGrid (const std::vector<int> & latt,const std::vector<int> &mpi);
};
}}