portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-10 19:36:56 +01:00
Code Releases Activity

Major rework of stencil. Half precision and MPI3 now working.

Browse Source
This commit is contained in:
paboyle
2017-04-22 11:33:50 +01:00
parent b9bbe5d188
commit 736bf3c866
9 changed files with 787 additions and 1384 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
lib/qcd/action/fermion/WilsonCompressor.h
View File

@ -241,13 +241,18 @@ public:
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
std::vector<int> same_node;
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)
{ /*Do nothing*/ };
: CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) ,
same_node(npoints)
{
assert(npoints==8);// or 10 if do naive DWF 5d red black ?
};
template < class compressor>
void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress)
@ -257,6 +262,7 @@ public:
this->CommunicateBegin(reqs);
this->CommunicateComplete(reqs);
this->CommsMerge(compress);
this->CommsMergeSHM(compress);
}
template <class compressor>
@ -295,23 +301,23 @@ public:
int face_idx=0;
if ( dag ) {
// std::cout << " Optimised Dagger compress " <<std::endl;
this->HaloGatherDir(source,XpCompress,Xp,face_idx);
this->HaloGatherDir(source,YpCompress,Yp,face_idx);
this->HaloGatherDir(source,ZpCompress,Zp,face_idx);
this->HaloGatherDir(source,TpCompress,Tp,face_idx);
this->HaloGatherDir(source,XmCompress,Xm,face_idx);
this->HaloGatherDir(source,YmCompress,Ym,face_idx);
this->HaloGatherDir(source,ZmCompress,Zm,face_idx);
this->HaloGatherDir(source,TmCompress,Tm,face_idx);
same_node[Xp]=this->HaloGatherDir(source,XpCompress,Xp,face_idx);
same_node[Yp]=this->HaloGatherDir(source,YpCompress,Yp,face_idx);
same_node[Zp]=this->HaloGatherDir(source,ZpCompress,Zp,face_idx);
same_node[Tp]=this->HaloGatherDir(source,TpCompress,Tp,face_idx);
same_node[Xm]=this->HaloGatherDir(source,XmCompress,Xm,face_idx);
same_node[Ym]=this->HaloGatherDir(source,YmCompress,Ym,face_idx);
same_node[Zm]=this->HaloGatherDir(source,ZmCompress,Zm,face_idx);
same_node[Tm]=this->HaloGatherDir(source,TmCompress,Tm,face_idx);
} else {
this->HaloGatherDir(source,XmCompress,Xp,face_idx);
this->HaloGatherDir(source,YmCompress,Yp,face_idx);
this->HaloGatherDir(source,ZmCompress,Zp,face_idx);
this->HaloGatherDir(source,TmCompress,Tp,face_idx);
this->HaloGatherDir(source,XpCompress,Xm,face_idx);
this->HaloGatherDir(source,YpCompress,Ym,face_idx);
this->HaloGatherDir(source,ZpCompress,Zm,face_idx);
this->HaloGatherDir(source,TpCompress,Tm,face_idx);
same_node[Xp]=this->HaloGatherDir(source,XmCompress,Xp,face_idx);
same_node[Yp]=this->HaloGatherDir(source,YmCompress,Yp,face_idx);
same_node[Zp]=this->HaloGatherDir(source,ZmCompress,Zp,face_idx);
same_node[Tp]=this->HaloGatherDir(source,TmCompress,Tp,face_idx);
same_node[Xm]=this->HaloGatherDir(source,XpCompress,Xm,face_idx);
same_node[Ym]=this->HaloGatherDir(source,YpCompress,Ym,face_idx);
same_node[Zm]=this->HaloGatherDir(source,ZpCompress,Zm,face_idx);
same_node[Tm]=this->HaloGatherDir(source,TpCompress,Tm,face_idx);
}
this->face_table_computed=1;
assert(this->u_comm_offset==this->_unified_buffer_size);

52
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -118,48 +118,6 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
// Allocate the required comms buffer
ImportGauge(_Umu);
}
/*
template<class Impl>
WilsonFermion5D<Impl>::WilsonFermion5D(int simd,GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
RealD _M5,const ImplParams &p) :
{
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]);
}
{
}
}
*/
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
@ -415,6 +373,10 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
DhopFaceTime+=usecond();
std::vector<std::vector<CommsRequest_t> > reqs;
// Rely on async comms; start comms before merge of local data
st.CommunicateBegin(reqs);
st.CommsMergeSHM(compressor);
#pragma omp parallel
{
int nthreads = omp_get_num_threads();
@ -426,7 +388,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
if ( me == 0 ) {
DhopCommTime-=usecond();
st.CommunicateBegin(reqs);
st.CommunicateComplete(reqs);
DhopCommTime+=usecond();
} else {
@ -442,10 +403,13 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
// Load imbalance alert. Should use dynamic schedule OMP for loop
// Perhaps create a list of only those sites with face work, and
// load balance process the list.
#pragma omp parallel
{
int nthreads = omp_get_num_threads();
int me = omp_get_thread_num();
int me = omp_get_thread_num();
int myoff, mywork;
GridThread::GetWork(len,me,mywork,myoff,nthreads);

680
lib/qcd/action/fermion/WilsonKernels.cc
View File

@ -36,62 +36,78 @@ namespace QCD {
int WilsonKernelsStatic::Opt = WilsonKernelsStatic::OptGeneric;
int WilsonKernelsStatic::Comms = WilsonKernelsStatic::CommsAndCompute;
#ifdef QPX
#include <spi/include/kernel/location.h>
#include <spi/include/l1p/types.h>
#include <hwi/include/bqc/l1p_mmio.h>
#include <hwi/include/bqc/A2_inlines.h>
#endif
void bgq_l1p_optimisation(int mode)
{
#ifdef QPX
#undef L1P_CFG_PF_USR
#define L1P_CFG_PF_USR (0x3fde8000108ll) /* (64 bit reg, 23 bits wide, user/unpriv) */
uint64_t cfg_pf_usr;
if ( mode ) {
cfg_pf_usr =
L1P_CFG_PF_USR_ifetch_depth(0)
| L1P_CFG_PF_USR_ifetch_max_footprint(1)
| L1P_CFG_PF_USR_pf_stream_est_on_dcbt
| L1P_CFG_PF_USR_pf_stream_establish_enable
| L1P_CFG_PF_USR_pf_stream_optimistic
| L1P_CFG_PF_USR_pf_adaptive_throttle(0xF) ;
// if ( sizeof(Float) == sizeof(double) ) {
cfg_pf_usr |= L1P_CFG_PF_USR_dfetch_depth(2)| L1P_CFG_PF_USR_dfetch_max_footprint(3) ;
// } else {
// cfg_pf_usr |= L1P_CFG_PF_USR_dfetch_depth(1)| L1P_CFG_PF_USR_dfetch_max_footprint(2) ;
// }
} else {
cfg_pf_usr = L1P_CFG_PF_USR_dfetch_depth(1)
| L1P_CFG_PF_USR_dfetch_max_footprint(2)
| L1P_CFG_PF_USR_ifetch_depth(0)
| L1P_CFG_PF_USR_ifetch_max_footprint(1)
| L1P_CFG_PF_USR_pf_stream_est_on_dcbt
| L1P_CFG_PF_USR_pf_stream_establish_enable
| L1P_CFG_PF_USR_pf_stream_optimistic
| L1P_CFG_PF_USR_pf_stream_prefetch_enable;
}
*((uint64_t *)L1P_CFG_PF_USR) = cfg_pf_usr;
#endif
}
template <class Impl>
WilsonKernels<Impl>::WilsonKernels(const ImplParams &p) : Base(p){};
////////////////////////////////////////////
// Generic implementation; move to different file?
////////////////////////////////////////////
#define GENERIC_STENCIL_LEG(Dir,spProj,Recon) \
SE = st.GetEntry(ptype, Dir, sF); \
if (SE->_is_local) { \
chi_p = &chi; \
if (SE->_permute) { \
spProj(tmp, in._odata[SE->_offset]); \
permute(chi, tmp, ptype); \
} else { \
spProj(chi, in._odata[SE->_offset]); \
} \
} else { \
chi_p = &buf[SE->_offset]; \
} \
Impl::multLink(Uchi, U._odata[sU], *chi_p, Dir, SE, st); \
Recon(result, Uchi);
#define GENERIC_STENCIL_LEG_INT(Dir,spProj,Recon) \
SE = st.GetEntry(ptype, Dir, sF); \
if (SE->_is_local) { \
chi_p = &chi; \
if (SE->_permute) { \
spProj(tmp, in._odata[SE->_offset]); \
permute(chi, tmp, ptype); \
} else { \
spProj(chi, in._odata[SE->_offset]); \
} \
} else if ( st.same_node[Dir] ) { \
chi_p = &buf[SE->_offset]; \
} \
if (SE->_is_local || st.same_node[Dir] ) { \
Impl::multLink(Uchi, U._odata[sU], *chi_p, Dir, SE, st); \
Recon(result, Uchi); \
}
#define GENERIC_STENCIL_LEG_EXT(Dir,spProj,Recon) \
SE = st.GetEntry(ptype, Dir, sF); \
if ((!SE->_is_local) && (!st.same_node[Dir]) ) { \
chi_p = &buf[SE->_offset]; \
Impl::multLink(Uchi, U._odata[sU], *chi_p, Dir, SE, st); \
Recon(result, Uchi); \
nmu++; \
}
#define GENERIC_DHOPDIR_LEG(Dir,spProj,Recon) \
if (gamma == Dir) { \
if (SE->_is_local && SE->_permute) { \
spProj(tmp, in._odata[SE->_offset]); \
permute(chi, tmp, ptype); \
} else if (SE->_is_local) { \
spProj(chi, in._odata[SE->_offset]); \
} else { \
chi = buf[SE->_offset]; \
} \
Impl::multLink(Uchi, U._odata[sU], chi, dir, SE, st); \
Recon(result, Uchi); \
}
////////////////////////////////////////////////////////////////////
// All legs kernels ; comms then compute
////////////////////////////////////////////////////////////////////
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionField &in, FermionField &out,
int interior,int exterior) {
SiteHalfSpinor *buf, int sF,
int sU, const FermionField &in, FermionField &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
@ -100,174 +116,22 @@ void WilsonKernels<Impl>::GenericDhopSiteDag(StencilImpl &st, LebesgueOrder &lo,
StencilEntry *SE;
int ptype;
///////////////////////////
// Xp
///////////////////////////
SE = st.GetEntry(ptype, Xp, sF);
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]);
}
} else {
chi_p = &buf[SE->_offset];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Xp, SE, st);
spReconXp(result, Uchi);
///////////////////////////
// Yp
///////////////////////////
SE = st.GetEntry(ptype, Yp, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Yp, SE, st);
accumReconYp(result, Uchi);
///////////////////////////
// Zp
///////////////////////////
SE = st.GetEntry(ptype, Zp, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Zp, SE, st);
accumReconZp(result, Uchi);
///////////////////////////
// Tp
///////////////////////////
SE = st.GetEntry(ptype, Tp, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Tp, SE, st);
accumReconTp(result, Uchi);
///////////////////////////
// Xm
///////////////////////////
SE = st.GetEntry(ptype, Xm, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Xm, SE, st);
accumReconXm(result, Uchi);
///////////////////////////
// Ym
///////////////////////////
SE = st.GetEntry(ptype, Ym, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Ym, SE, st);
accumReconYm(result, Uchi);
///////////////////////////
// Zm
///////////////////////////
SE = st.GetEntry(ptype, Zm, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Zm, SE, st);
accumReconZm(result, Uchi);
///////////////////////////
// Tm
///////////////////////////
SE = st.GetEntry(ptype, Tm, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Tm, SE, st);
accumReconTm(result, Uchi);
GENERIC_STENCIL_LEG(Xp,spProjXp,spReconXp);
GENERIC_STENCIL_LEG(Yp,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG(Zp,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG(Tp,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG(Xm,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG(Ym,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG(Zm,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG(Tm,spProjTm,accumReconTm);
vstream(out._odata[sF], result);
};
// Need controls to do interior, exterior, or both
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionField &in, FermionField &out,int interior,int exterior) {
SiteHalfSpinor *buf, int sF,
int sU, const FermionField &in, FermionField &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
@ -276,168 +140,123 @@ void WilsonKernels<Impl>::GenericDhopSite(StencilImpl &st, LebesgueOrder &lo, Do
StencilEntry *SE;
int ptype;
///////////////////////////
// Xp
///////////////////////////
SE = st.GetEntry(ptype, Xm, sF);
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]);
}
} else {
chi_p = &buf[SE->_offset];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Xm, SE, st);
spReconXp(result, Uchi);
///////////////////////////
// Yp
///////////////////////////
SE = st.GetEntry(ptype, Ym, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Ym, SE, st);
accumReconYp(result, Uchi);
///////////////////////////
// Zp
///////////////////////////
SE = st.GetEntry(ptype, Zm, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Zm, SE, st);
accumReconZp(result, Uchi);
///////////////////////////
// Tp
///////////////////////////
SE = st.GetEntry(ptype, Tm, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Tm, SE, st);
accumReconTp(result, Uchi);
///////////////////////////
// Xm
///////////////////////////
SE = st.GetEntry(ptype, Xp, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Xp, SE, st);
accumReconXm(result, Uchi);
///////////////////////////
// Ym
///////////////////////////
SE = st.GetEntry(ptype, Yp, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Yp, SE, st);
accumReconYm(result, Uchi);
///////////////////////////
// Zm
///////////////////////////
SE = st.GetEntry(ptype, Zp, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Zp, SE, st);
accumReconZm(result, Uchi);
///////////////////////////
// Tm
///////////////////////////
SE = st.GetEntry(ptype, Tp, sF);
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];
}
Impl::multLink(Uchi, U._odata[sU], *chi_p, Tp, SE, st);
accumReconTm(result, Uchi);
GENERIC_STENCIL_LEG(Xm,spProjXp,spReconXp);
GENERIC_STENCIL_LEG(Ym,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG(Zm,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG(Tm,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG(Xp,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG(Yp,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG(Zp,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG(Tp,spProjTm,accumReconTm);
vstream(out._odata[sF], result);
};
////////////////////////////////////////////////////////////////////
// Interior kernels
////////////////////////////////////////////////////////////////////
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
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;
result=zero;
GENERIC_STENCIL_LEG_INT(Xp,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_INT(Yp,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_INT(Zp,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_INT(Tp,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_INT(Xm,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_INT(Ym,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_INT(Zm,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_INT(Tm,spProjTm,accumReconTm);
vstream(out._odata[sF], result);
};
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteInt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
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;
result=zero;
GENERIC_STENCIL_LEG_INT(Xm,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_INT(Ym,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_INT(Zm,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_INT(Tm,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_INT(Xp,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_INT(Yp,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_INT(Zp,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_INT(Tp,spProjTm,accumReconTm);
vstream(out._odata[sF], result);
};
////////////////////////////////////////////////////////////////////
// Exterior kernels
////////////////////////////////////////////////////////////////////
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
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 nmu=0;
result=zero;
GENERIC_STENCIL_LEG_EXT(Xp,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_EXT(Yp,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_EXT(Zp,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_EXT(Tp,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_EXT(Xm,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_EXT(Ym,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_EXT(Zm,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_EXT(Tm,spProjTm,accumReconTm);
if ( nmu ) {
out._odata[sF] = out._odata[sF] + result;
}
};
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteExt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
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 nmu=0;
result=zero;
GENERIC_STENCIL_LEG_EXT(Xm,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_EXT(Ym,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_EXT(Zm,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_EXT(Tm,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_EXT(Xp,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_EXT(Yp,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_EXT(Zp,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_EXT(Tp,spProjTm,accumReconTm);
if ( nmu ) {
out._odata[sF] = out._odata[sF] + result;
}
};
template <class Impl>
void WilsonKernels<Impl>::DhopDir( StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor *buf, int sF,
@ -451,119 +270,14 @@ void WilsonKernels<Impl>::DhopDir( StencilImpl &st, DoubledGaugeField &U,SiteHal
int ptype;
SE = st.GetEntry(ptype, dir, sF);
// Xp
if (gamma == Xp) {
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, dir, SE, st);
spReconXp(result, Uchi);
}
// Yp
if (gamma == Yp) {
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, dir, SE, st);
spReconYp(result, Uchi);
}
// Zp
if (gamma == Zp) {
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, dir, SE, st);
spReconZp(result, Uchi);
}
// Tp
if (gamma == Tp) {
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, dir, SE, st);
spReconTp(result, Uchi);
}
// Xm
if (gamma == Xm) {
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, dir, SE, st);
spReconXm(result, Uchi);
}
// Ym
if (gamma == Ym) {
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, dir, SE, st);
spReconYm(result, Uchi);
}
// Zm
if (gamma == Zm) {
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];
}
Impl::multLink(Uchi, U._odata[sU], chi, dir, SE, st);
spReconZm(result, Uchi);
}
// Tm
if (gamma == Tm) {
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 {
chi = buf[SE->_offset];
}
Impl::multLink(Uchi, U._odata[sU], chi, dir, SE, st);
spReconTm(result, Uchi);
}
GENERIC_DHOPDIR_LEG(Xp,spProjXp,spReconXp);
GENERIC_DHOPDIR_LEG(Yp,spProjYp,spReconYp);
GENERIC_DHOPDIR_LEG(Zp,spProjZp,spReconZp);
GENERIC_DHOPDIR_LEG(Tp,spProjTp,spReconTp);
GENERIC_DHOPDIR_LEG(Xm,spProjXm,spReconXm);
GENERIC_DHOPDIR_LEG(Ym,spProjYm,spReconYm);
GENERIC_DHOPDIR_LEG(Zm,spProjZm,spReconZm);
GENERIC_DHOPDIR_LEG(Tm,spProjTm,spReconTm);
vstream(out._odata[sF], result);
}

69
lib/qcd/action/fermion/WilsonKernels.h
View File

@ -34,8 +34,6 @@ directory
namespace Grid {
namespace QCD {
void bgq_l1p_optimisation(int mode);
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Helper routines that implement Wilson stencil for a single site.
// Common to both the WilsonFermion and WilsonFermion5D
@ -44,9 +42,8 @@ class WilsonKernelsStatic {
public:
enum { OptGeneric, OptHandUnroll, OptInlineAsm };
enum { CommsAndCompute, CommsThenCompute };
// S-direction is INNERMOST and takes no part in the parity.
static int Opt; // these are a temporary hack
static int Comms; // these are a temporary hack
static int Opt;
static int Comms;
};
template<class Impl> class WilsonKernels : public FermionOperator<Impl> , public WilsonKernelsStatic {
@ -75,7 +72,7 @@ public:
case OptHandUnroll:
for (int site = 0; site < Ns; site++) {
for (int s = 0; s < Ls; s++) {
if( exterior) WilsonKernels<Impl>::HandDhopSite(st,lo,U,buf,sF,sU,in,out,interior,exterior);
if( exterior) WilsonKernels<Impl>::HandDhopSite(st,lo,U,buf,sF,sU,in,out);
sF++;
}
sU++;
@ -84,7 +81,10 @@ public:
case OptGeneric:
for (int site = 0; site < Ns; site++) {
for (int s = 0; s < Ls; s++) {
if( exterior) WilsonKernels<Impl>::GenericDhopSite(st,lo,U,buf,sF,sU,in,out,interior,exterior);
if(interior&&exterior) WilsonKernels<Impl>::GenericDhopSite(st,lo,U,buf,sF,sU,in,out);
else if (interior) WilsonKernels<Impl>::GenericDhopSiteInt(st,lo,U,buf,sF,sU,in,out);
else if (exterior) WilsonKernels<Impl>::GenericDhopSiteExt(st,lo,U,buf,sF,sU,in,out);
else assert(0);
sF++;
}
sU++;
@ -99,11 +99,14 @@ public:
template <bool EnableBool = true>
typename std::enable_if<(Impl::Dimension != 3 || (Impl::Dimension == 3 && Nc != 3)) && EnableBool, void>::type
DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out,int interior=1,int exterior=1 ) {
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out,int interior=1,int exterior=1 ) {
// no kernel choice
for (int site = 0; site < Ns; site++) {
for (int s = 0; s < Ls; s++) {
if( exterior) WilsonKernels<Impl>::GenericDhopSite(st, lo, U, buf, sF, sU, in, out,interior,exterior);
if(interior&&exterior) WilsonKernels<Impl>::GenericDhopSite(st,lo,U,buf,sF,sU,in,out);
else if (interior) WilsonKernels<Impl>::GenericDhopSiteInt(st,lo,U,buf,sF,sU,in,out);
else if (exterior) WilsonKernels<Impl>::GenericDhopSiteExt(st,lo,U,buf,sF,sU,in,out);
else assert(0);
sF++;
}
sU++;
@ -113,8 +116,8 @@ public:
template <bool EnableBool = true>
typename std::enable_if<Impl::Dimension == 3 && Nc == 3 && EnableBool,void>::type
DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out,int interior=1,int exterior=1) {
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out,int interior=1,int exterior=1)
{
bgq_l1p_optimisation(1);
switch(Opt) {
#if defined(AVX512) || defined (QPX)
@ -128,7 +131,7 @@ public:
case OptHandUnroll:
for (int site = 0; site < Ns; site++) {
for (int s = 0; s < Ls; s++) {
if( exterior) WilsonKernels<Impl>::HandDhopSiteDag(st,lo,U,buf,sF,sU,in,out,interior,exterior);
if( exterior) WilsonKernels<Impl>::HandDhopSiteDag(st,lo,U,buf,sF,sU,in,out);
sF++;
}
sU++;
@ -137,7 +140,10 @@ public:
case OptGeneric:
for (int site = 0; site < Ns; site++) {
for (int s = 0; s < Ls; s++) {
if( exterior) WilsonKernels<Impl>::GenericDhopSiteDag(st,lo,U,buf,sF,sU,in,out,interior,exterior);
if(interior&&exterior) WilsonKernels<Impl>::GenericDhopSiteDag(st,lo,U,buf,sF,sU,in,out);
else if (interior) WilsonKernels<Impl>::GenericDhopSiteDagInt(st,lo,U,buf,sF,sU,in,out);
else if (exterior) WilsonKernels<Impl>::GenericDhopSiteDagExt(st,lo,U,buf,sF,sU,in,out);
else assert(0);
sF++;
}
sU++;
@ -156,7 +162,10 @@ public:
for (int site = 0; site < Ns; site++) {
for (int s = 0; s < Ls; s++) {
if( exterior) WilsonKernels<Impl>::GenericDhopSiteDag(st,lo,U,buf,sF,sU,in,out,interior,exterior);
if(interior&&exterior) WilsonKernels<Impl>::GenericDhopSiteDag(st,lo,U,buf,sF,sU,in,out);
else if (interior) WilsonKernels<Impl>::GenericDhopSiteDagInt(st,lo,U,buf,sF,sU,in,out);
else if (exterior) WilsonKernels<Impl>::GenericDhopSiteDagExt(st,lo,U,buf,sF,sU,in,out);
else assert(0);
sF++;
}
sU++;
@ -169,35 +178,47 @@ public:
private:
// Specialised variants
void GenericDhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out,int interior,int exterior);
int sF, int sU, const FermionField &in, FermionField &out);
void GenericDhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out,int interior,int exterior);
int sF, int sU, const FermionField &in, FermionField &out);
void GenericDhopSiteInt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out);
void GenericDhopSiteDagInt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out);
void GenericDhopSiteExt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out);
void GenericDhopSiteDagExt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out);
void AsmDhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in,FermionField &out);
int sF, int sU, int Ls, int Ns, const FermionField &in,FermionField &out);
void AsmDhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out);
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out);
void AsmDhopSiteInt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in,FermionField &out);
int sF, int sU, int Ls, int Ns, const FermionField &in,FermionField &out);
void AsmDhopSiteDagInt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out);
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out);
void AsmDhopSiteExt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in,FermionField &out);
int sF, int sU, int Ls, int Ns, const FermionField &in,FermionField &out);
void AsmDhopSiteDagExt(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out);
int sF, int sU, int Ls, int Ns, const FermionField &in, FermionField &out);
void HandDhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out,int interior,int exterior);
int sF, int sU, const FermionField &in, FermionField &out);
void HandDhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out,int interior,int exterior);
int sF, int sU, const FermionField &in, FermionField &out);
public:

618
lib/qcd/action/fermion/WilsonKernelsHand.cc
View File

@ -307,55 +307,106 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
result_31-= UChi_11; \
result_32-= UChi_12;
namespace Grid {
namespace QCD {
#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU; \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
LOAD_CHI; \
} \
{ \
MULT_2SPIN(DIR); \
} \
RECON;
#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU; \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
if ( st.same_node[DIR] ) { \
LOAD_CHI; \
} \
} \
if (local || st.same_node[DIR] ) { \
MULT_2SPIN(DIR); \
RECON; \
}
#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,Dir,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if((!SE->_is_local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI; \
MULT_2SPIN(DIR); \
RECON; \
}
#define HAND_RESULT(ss) \
{ \
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); \
}
template<class Impl> void
WilsonKernels<Impl>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionField &in, FermionField &out,int interior,int exterior)
{
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 HAND_DECLARATIONS(a) \
Simd result_00; \
Simd result_01; \
Simd result_02; \
Simd result_10; \
Simd result_11; \
Simd result_12; \
Simd result_20; \
Simd result_21; \
Simd result_22; \
Simd result_30; \
Simd result_31; \
Simd result_32; \
Simd Chi_00; \
Simd Chi_01; \
Simd Chi_02; \
Simd Chi_10; \
Simd Chi_11; \
Simd Chi_12; \
Simd UChi_00; \
Simd UChi_01; \
Simd UChi_02; \
Simd UChi_10; \
Simd UChi_11; \
Simd UChi_12; \
Simd U_00; \
Simd U_10; \
Simd U_20; \
Simd U_01; \
Simd U_11; \
Simd U_21;
#define Chimu_00 Chi_00
#define Chimu_01 Chi_01
@ -370,430 +421,54 @@ WilsonKernels<Impl>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGauge
#define Chimu_31 UChi_11
#define Chimu_32 UChi_12
namespace Grid {
namespace QCD {
template<class Impl> void
WilsonKernels<Impl>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionField &in, FermionField &out)
{
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
int offset,local,perm, ptype;
StencilEntry *SE;
// Xp
SE=st.GetEntry(ptype,Xp,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(Xp);
}
XM_RECON;
// Yp
SE=st.GetEntry(ptype,Yp,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(Yp);
}
YM_RECON_ACCUM;
// Zp
SE=st.GetEntry(ptype,Zp,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(Zp);
}
ZM_RECON_ACCUM;
// Tp
SE=st.GetEntry(ptype,Tp,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(Tp);
}
TM_RECON_ACCUM;
// Xm
SE=st.GetEntry(ptype,Xm,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(Xm);
}
XP_RECON_ACCUM;
// Ym
SE=st.GetEntry(ptype,Ym,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(Ym);
}
YP_RECON_ACCUM;
// Zm
SE=st.GetEntry(ptype,Zm,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(Zm);
}
ZP_RECON_ACCUM;
// Tm
SE=st.GetEntry(ptype,Tm,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(Tm);
}
TP_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);
}
HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM);
HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM);
HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM);
HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM);
HAND_RESULT(ss);
}
template<class Impl>
void WilsonKernels<Impl>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionField &in, FermionField &out,int interior,int exterior)
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
HAND_DECLARATIONS(ignore);
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);
}
HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON);
HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM);
HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM);
HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM);
HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM);
HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM);
HAND_RESULT(ss);
}
////////////////////////////////////////////////
@ -801,74 +476,71 @@ void WilsonKernels<Impl>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,Doub
////////////////////////////////////////////////
template<> void
WilsonKernels<GparityWilsonImplF>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
template<> void
WilsonKernels<GparityWilsonImplF>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
template<> void
WilsonKernels<GparityWilsonImplD>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
template<> void
WilsonKernels<GparityWilsonImplD>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
template<> void
WilsonKernels<GparityWilsonImplFH>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
template<> void
WilsonKernels<GparityWilsonImplFH>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
template<> void
WilsonKernels<GparityWilsonImplDF>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
template<> void
WilsonKernels<GparityWilsonImplDF>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
int sF,int sU,const FermionField &in, FermionField &out)
{
assert(0);
}
////////////// Wilson ; uses this implementation /////////////////////
// Need Nc=3 though //
#define INSTANTIATE_THEM(A) \
template void WilsonKernels<A>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionField &in, FermionField &out,int interior,int exterior); \
template void WilsonKernels<A>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionField &in, FermionField &out,int interior,int exterior);
int ss,int sU,const FermionField &in, FermionField &out); \
template void WilsonKernels<A>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionField &in, FermionField &out);
INSTANTIATE_THEM(WilsonImplF);
INSTANTIATE_THEM(WilsonImplD);