1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-22 01:32:03 +01:00

COmms improvements

This commit is contained in:
Peter Boyle
2023-03-21 08:53:56 -07:00
parent a997d24743
commit e1c326558a
9 changed files with 46 additions and 32 deletions

View File

@ -484,24 +484,26 @@ public:
int dag = compress.dag;
int face_idx=0;
#define vet_same_node(a,b) \
{ auto tmp = b; }
if ( dag ) {
assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XpCompress,Xp,face_idx));
vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YpCompress,Yp,face_idx));
vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TpCompress,Tp,face_idx));
vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XmCompress,Xm,face_idx));
vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YmCompress,Ym,face_idx));
vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TmCompress,Tm,face_idx));
} else {
assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XmCompress,Xp,face_idx));
vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YmCompress,Yp,face_idx));
vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TmCompress,Tp,face_idx));
vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XpCompress,Xm,face_idx));
vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YpCompress,Ym,face_idx));
vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TpCompress,Tm,face_idx));
}
this->face_table_computed=1;
assert(this->u_comm_offset==this->_unified_buffer_size);

View File

@ -439,6 +439,17 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
int sF = ss*Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v); \
});
#define ASM_CALL_SLICE(A) \
auto grid = in.Grid() ; \
int nt = grid->LocalDimensions()[4]; \
int nxyz = Nsite/nt ; \
for(int t=0;t<nt;t++){ \
thread_for( sss, nxyz, { \
int ss = t*nxyz+sss; \
int sU = ss; \
int sF = ss*Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v); \
});}
template <class Impl>
void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,