portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

Staggered coalseced read

Browse Source
This commit is contained in:
Peter Boyle 2021-03-29 20:01:15 +02:00
parent 8bdadbadac
commit bb89a82a07
3 changed files with 82 additions and 45 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
Grid/qcd/action/fermion/StaggeredImpl.h
View File

@ -72,19 +72,23 @@ public:
StaggeredImpl(const ImplParams &p = ImplParams()) : Params(p){};
static accelerator_inline void multLink(SiteSpinor &phi,
template<class _Spinor>
static accelerator_inline void multLink(_Spinor &phi,
const SiteDoubledGaugeField &U,
const SiteSpinor &chi,
const _Spinor &chi,
int mu)
{
mult(&phi(), &U(mu), &chi());
auto UU = coalescedRead(U(mu));
mult(&phi(), &UU, &chi());
}
static accelerator_inline void multLinkAdd(SiteSpinor &phi,
template<class _Spinor>
static accelerator_inline void multLinkAdd(_Spinor &phi,
const SiteDoubledGaugeField &U,
const SiteSpinor &chi,
const _Spinor &chi,
int mu)
{
mac(&phi(), &U(mu), &chi());
auto UU = coalescedRead(U(mu));
mac(&phi(), &UU, &chi());
}
template <class ref>

35
Grid/qcd/action/fermion/WilsonImpl.h
View File

@ -184,18 +184,22 @@ public:
mat = TraceIndex<SpinIndex>(P);
}
inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds)
{
for (int mu = 0; mu < Nd; mu++)
mat[mu] = PeekIndex<LorentzIndex>(Uds, mu);
}
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu)
{
#undef USE_OLD_INSERT_FORCE
int Ls=Btilde.Grid()->_fdimensions[0];
autoView( mat_v , mat, AcceleratorWrite);
#ifdef USE_OLD_INSERT_FORCE
GaugeLinkField tmp(mat.Grid());
tmp = Zero();
{
const int Nsimd = SiteSpinor::Nsimd();
autoView( tmp_v , tmp, AcceleratorWrite);
autoView( Btilde_v , Btilde, AcceleratorRead);
autoView( Atilde_v , Atilde, AcceleratorRead);
@ -208,6 +212,29 @@ public:
});
}
PokeIndex<LorentzIndex>(mat,tmp,mu);
#else
{
const int Nsimd = SiteSpinor::Nsimd();
autoView( Btilde_v , Btilde, AcceleratorRead);
autoView( Atilde_v , Atilde, AcceleratorRead);
accelerator_for(sss,mat.Grid()->oSites(),Nsimd,{
int sU=sss;
typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
ColorMatrixType sum;
zeroit(sum);
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
for(int spn=0;spn<Ns;spn++){ //sum over spin
auto bb = coalescedRead(Btilde_v[sF]()(spn) ); //color vector
auto aa = coalescedRead(Atilde_v[sF]()(spn) );
auto op = outerProduct(bb,aa);
sum = sum + op;
}
}
coalescedWrite(mat_v[sU](mu)(), sum);
});
}
#endif
}
};

76
Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h
View File

@ -35,39 +35,32 @@ NAMESPACE_BEGIN(Grid);
#define GENERIC_STENCIL_LEG(U,Dir,skew,multLink) \
SE = st.GetEntry(ptype, Dir+skew, sF); \
if (SE->_is_local ) { \
if (SE->_permute) { \
chi_p = &chi; \
permute(chi, in[SE->_offset], ptype); \
} else { \
chi_p = &in[SE->_offset]; \
} \
int perm= SE->_permute; \
chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
} else { \
chi_p = &buf[SE->_offset]; \
chi = coalescedRead(buf[SE->_offset],lane); \
} \
multLink(Uchi, U[sU], *chi_p, Dir);
acceleratorSynchronise(); \
multLink(Uchi, U[sU], chi, Dir);
#define GENERIC_STENCIL_LEG_INT(U,Dir,skew,multLink) \
SE = st.GetEntry(ptype, Dir+skew, sF); \
if (SE->_is_local ) { \
if (SE->_permute) { \
chi_p = &chi; \
permute(chi, in[SE->_offset], ptype); \
} else { \
chi_p = &in[SE->_offset]; \
} \
int perm= SE->_permute; \
chi = coalescedReadPermute(in[SE->_offset],ptype,perm,lane);\
} else if ( st.same_node[Dir] ) { \
chi_p = &buf[SE->_offset]; \
chi = coalescedRead(buf[SE->_offset],lane); \
} \
if (SE->_is_local || st.same_node[Dir] ) { \
multLink(Uchi, U[sU], *chi_p, Dir); \
multLink(Uchi, U[sU], chi, Dir); \
}
#define GENERIC_STENCIL_LEG_EXT(U,Dir,skew,multLink) \
SE = st.GetEntry(ptype, Dir+skew, sF); \
if ((!SE->_is_local) && (!st.same_node[Dir]) ) { \
nmu++; \
chi_p = &buf[SE->_offset]; \
multLink(Uchi, U[sU], *chi_p, Dir); \
chi = coalescedRead(buf[SE->_offset],lane); \
multLink(Uchi, U[sU], chi, Dir); \
}
template <class Impl>
@ -84,12 +77,14 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out, int dag)
{
const SiteSpinor *chi_p;
SiteSpinor chi;
SiteSpinor Uchi;
typedef decltype(coalescedRead(in[0])) calcSpinor;
calcSpinor chi;
calcSpinor Uchi;
StencilEntry *SE;
int ptype;
int skew;
const int Nsimd = SiteHalfSpinor::Nsimd();
const int lane=acceleratorSIMTlane(Nsimd);
// for(int s=0;s<LLs;s++){
//
@ -118,7 +113,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
if ( dag ) {
Uchi = - Uchi;
}
vstream(out[sF], Uchi);
coalescedWrite(out[sF], Uchi,lane);
}
};
@ -130,13 +125,16 @@ template <int Naik> accelerator_inline
void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag) {
const SiteSpinor *chi_p;
SiteSpinor chi;
SiteSpinor Uchi;
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef decltype(coalescedRead(in[0])) calcSpinor;
calcSpinor chi;
calcSpinor Uchi;
StencilEntry *SE;
int ptype;
int skew ;
const int Nsimd = SiteHalfSpinor::Nsimd();
const int lane=acceleratorSIMTlane(Nsimd);
// for(int s=0;s<LLs;s++){
// int sF=LLs*sU+s;
@ -165,7 +163,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st,
if ( dag ) {
Uchi = - Uchi;
}
vstream(out[sF], Uchi);
coalescedWrite(out[sF], Uchi,lane);
}
};
@ -178,14 +176,17 @@ template <int Naik> accelerator_inline
void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag) {
const SiteSpinor *chi_p;
// SiteSpinor chi;
SiteSpinor Uchi;
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef decltype(coalescedRead(in[0])) calcSpinor;
calcSpinor chi;
calcSpinor Uchi;
StencilEntry *SE;
int ptype;
int nmu=0;
int skew ;
const int Nsimd = SiteHalfSpinor::Nsimd();
const int lane=acceleratorSIMTlane(Nsimd);
// for(int s=0;s<LLs;s++){
// int sF=LLs*sU+s;
@ -211,11 +212,12 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st,
GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
}
if ( nmu ) {
if ( dag ) {
out[sF] = out[sF] - Uchi;
if ( nmu ) {
auto _out = coalescedRead(out[sF],lane);
if ( dag ) {
coalescedWrite(out[sF], _out-Uchi,lane);
} else {
out[sF] = out[sF] + Uchi;
coalescedWrite(out[sF], _out+Uchi,lane);
}
}
}
@ -261,6 +263,8 @@ void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo,
GridBase *FGrid=in.Grid();
GridBase *UGrid=U.Grid();
typedef StaggeredKernels<Impl> ThisKernel;
const int Nsimd = SiteHalfSpinor::Nsimd();
const int lane=acceleratorSIMTlane(Nsimd);
autoView( UUU_v , UUU, AcceleratorRead);
autoView( U_v , U, AcceleratorRead);
autoView( in_v , in, AcceleratorRead);
@ -301,6 +305,8 @@ void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo,
GridBase *FGrid=in.Grid();
GridBase *UGrid=U.Grid();
typedef StaggeredKernels<Impl> ThisKernel;
const int Nsimd = SiteHalfSpinor::Nsimd();
const int lane=acceleratorSIMTlane(Nsimd);
autoView( UUU_v , U, AcceleratorRead);
autoView( U_v , U, AcceleratorRead);
autoView( in_v , in, AcceleratorRead);