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

Partially fixed. But the comms-overlap does not work yet.

Browse Source
This commit is contained in:
Peter Boyle 2017-04-22 18:50:25 -04:00
parent ac58565d0a
commit 5812eb8a8c
1 changed files with 18 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
lib/qcd/action/fermion/WilsonKernelsAsmBody.h
View File

@ -55,7 +55,7 @@
RECON; \ RECON; \
#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,NxtDir,ent,plocal); ent++; \ base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
PF_GAUGE(Xp); \ PF_GAUGE(Xp); \
PREFETCH1_CHIMU(base); \ PREFETCH1_CHIMU(base); \
ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
@ -75,38 +75,34 @@
LOAD64(%r10,isigns); \ LOAD64(%r10,isigns); \
PROJ(base); \ PROJ(base); \
MAYBEPERM(PERMUTE_DIR,perm); \ MAYBEPERM(PERMUTE_DIR,perm); \
} else if ( st.same_dir[Dir] ) { \ } else if ( st.same_node[Dir] ) { \
LOAD_CHI(base); \ LOAD_CHI(base); \
} \ } \
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \ base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
PREFETCH_CHIMU(base); \
if ( local || st.same_node[Dir] ) { \ if ( local || st.same_node[Dir] ) { \
PREFETCH_CHIMU(base); \
MULT_2SPIN_DIR_PF(Dir,basep); \ MULT_2SPIN_DIR_PF(Dir,basep); \
LOAD64(%r10,isigns); \ LOAD64(%r10,isigns); \
RECON; \ RECON; \
} }
#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,NxtDir,ent,plocal); ent++; \ base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
PF_GAUGE(Xp); \ PF_GAUGE(Xp); \
PREFETCH1_CHIMU(base); \ PREFETCH1_CHIMU(base); \
ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) else { ZERO_PSI; } ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) else { ZERO_PSI; }
#define RESULT(base,basep) SAVE_RESULT(base,basep); #define RESULT(base,basep) SAVE_RESULT(base,basep);
#define ZERO_NMU(A)
#endif #endif
//////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////
// Post comms kernel // Post comms kernel
//////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////
#ifdef EXTERIOR #ifdef EXTERIOR
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \ #define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \ base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
if((!SE->_is_local)&&(!st.same_node[DIR]) ) { \ if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \ LOAD_CHI(base); \
MULT_2SPIN_DIR_PF(Dir,base); \ MULT_2SPIN_DIR_PF(Dir,base); \
LOAD64(%r10,isigns); \ LOAD64(%r10,isigns); \
@ -114,11 +110,11 @@
nmu++; \ nmu++; \
} }
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) #define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
nmu=0; \
ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) else { ZERO_PSI; }
#define ZERO_NMU(A) nmu=0; #define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
#define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
#endif #endif
{ {
@ -144,9 +140,7 @@
int ent=ss*8;// 2*Ndim int ent=ss*8;// 2*Ndim
int nent=ssn*8; int nent=ssn*8;
ZERO_NMU(0); ASM_LEG_XP(Xp,Yp,PERMUTE_DIR3,DIR0_PROJMEM,DIR0_RECON);
ASM_LEG_XP(Xp,Yp,PERMUTE_DIR3,DIR0_PROJMEM,DIR0_RECON);
ASM_LEG(Yp,Zp,PERMUTE_DIR2,DIR1_PROJMEM,DIR1_RECON); ASM_LEG(Yp,Zp,PERMUTE_DIR2,DIR1_PROJMEM,DIR1_RECON);
ASM_LEG(Zp,Tp,PERMUTE_DIR1,DIR2_PROJMEM,DIR2_RECON); ASM_LEG(Zp,Tp,PERMUTE_DIR1,DIR2_PROJMEM,DIR2_RECON);
ASM_LEG(Tp,Xm,PERMUTE_DIR0,DIR3_PROJMEM,DIR3_RECON); ASM_LEG(Tp,Xm,PERMUTE_DIR0,DIR3_PROJMEM,DIR3_RECON);
@ -156,10 +150,10 @@
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);
#ifdef EXTERIOR #ifndef EXTERIOR
if( nmu == 0 ) break; //Early out if no work
#endif //if (nmu==0) break;
#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++;
RESULT(base,basep); RESULT(base,basep);
@ -185,10 +179,6 @@
#undef DIR5_RECON #undef DIR5_RECON
#undef DIR6_RECON #undef DIR6_RECON
#undef DIR7_RECON #undef DIR7_RECON
#undef EXTERIOR_BLOCK #undef ASM_LEG
#undef INTERIOR_BLOCK #undef ASM_LEG_XP
#undef EXTERIOR_BLOCK_XP
#undef INTERIOR_BLOCK_XP
#undef COMMON_BLOCK
#undef ZERO_NMU
#undef RESULT #undef RESULT