portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-09 21:50:45 +01:00
Code Releases Activity

bug fix. works now and great face performance

Browse Source
This commit is contained in:
Peter Boyle 2017-04-26 03:14:02 -04:00
parent fd1eb7de13
commit f8797e1e3e
1 changed files with 12 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
lib/qcd/action/fermion/WilsonKernelsAsmBody.h
View File

@ -87,7 +87,8 @@
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \ #define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \ base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
PF_GAUGE(Xp); \ PF_GAUGE(Xp); \
PREFETCH1_CHIMU(base);{ ZERO_PSI; } \ PREFETCH1_CHIMU(base); \
{ ZERO_PSI; } \
ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
#define RESULT(base,basep) SAVE_RESULT(base,basep); #define RESULT(base,basep) SAVE_RESULT(base,basep);
@ -111,6 +112,7 @@
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \ #define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
nmu=0; \ nmu=0; \
{ ZERO_PSI;} \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \ base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
if((!local)&&(!st.same_node[Dir]) ) { \ if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \ LOAD_CHI(base); \
@ -118,7 +120,7 @@
LOAD64(%r10,isigns); \ LOAD64(%r10,isigns); \
RECON; \ RECON; \
nmu++; \ nmu++; \
} else { ZERO_PSI; }; }
#define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);} #define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
@ -134,11 +136,15 @@
MASK_REGS; MASK_REGS;
int nmax=U._grid->oSites(); int nmax=U._grid->oSites();
for(int site=0;site<Ns;site++) { for(int site=0;site<Ns;site++) {
#ifndef EXTERIOR
int sU =lo.Reorder(ssU); int sU =lo.Reorder(ssU);
int ssn=ssU+1; if(ssn>=nmax) ssn=0; int ssn=ssU+1; if(ssn>=nmax) ssn=0;
int sUn=lo.Reorder(ssn); int sUn=lo.Reorder(ssn);
#ifndef EXTERIOR
LOCK_GAUGE(0); LOCK_GAUGE(0);
#else
int sU =ssU;
int ssn=ssU+1; if(ssn>=nmax) ssn=0;
int sUn=ssn;
#endif #endif
for(int s=0;s<Ls;s++) { for(int s=0;s<Ls;s++) {
ss =sU*Ls+s; ss =sU*Ls+s;
@ -156,8 +162,9 @@
ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJMEM,DIR6_RECON); ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJMEM,DIR6_RECON);
ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJMEM,DIR7_RECON); ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJMEM,DIR7_RECON);
#ifndef EXTERIOR #ifdef EXTERIOR
if (nmu==0) break; //Early out if no work if (nmu==0) break;
// if (nmu!=0) std::cout << "EXT "<<sU<<std::endl;
#endif #endif
base = (uint64_t) &out._odata[ss]; base = (uint64_t) &out._odata[ss];
basep= st.GetPFInfo(nent,plocal); nent++; basep= st.GetPFInfo(nent,plocal); nent++;