portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 23:37:06 +01:00
Code Releases Activity

Tweaks

Browse Source
This commit is contained in:
paboyle
2016-06-25 11:08:05 -07:00
parent 51cb2d4328
commit 2d8bb4c594
5 changed files with 242 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
lib/qcd/action/fermion/WilsonKernelsAsmBody.h
View File

@ -1,8 +1,7 @@
{
int locala,perma, ptypea;
int localb,permb, ptypeb;
int localc,permc, ptypec;
uint64_t basea, baseb, basec;
uint64_t basea, baseb;
uint64_t basex;
const uint64_t plocal =(uint64_t) & in._odata[0];
@ -12,22 +11,15 @@
MASK_REGS;
for(int site=0;site<Ns;site++) {
int sU=lo.Reorder(ssU);
int sU=lo.Reorder(ssU);
for(int s=0;s<Ls;s++) {
ss =sU*Ls+s;
ss=sU*Ls+s;
////////////////////////////////
// Xp
////////////////////////////////
int ent=ss*8;// 2*Ndim
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(basea);
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(baseb);
basec = st.GetInfo(ptypec,localc,permc,Zp,ent,plocal); ent++;
PREFETCH_CHIMU(basec);
basex = basea;
if ( locala ) {
@ -47,7 +39,6 @@
// Yp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(basea);
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_PROJMEM(baseb);
@ -56,7 +47,7 @@
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFYP(Yp,basec);
MULT_2SPIN_DIR_PFYP(Yp,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_RECON_ACCUM;
@ -65,16 +56,15 @@
// Zp
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(baseb);
if ( localc ) {
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_PROJMEM(basec);
MAYBEPERM(PERMUTE_DIR1,permc);
ZM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR1,perma);
} else {
LOAD_CHI(basec);
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFZP(Zp,basea);
MULT_2SPIN_DIR_PFZP(Zp,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_RECON_ACCUM;
@ -82,17 +72,16 @@
////////////////////////////////
// Tp
////////////////////////////////
basec = st.GetInfo(ptypec,localc,permc,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(basec);
if ( locala ) {
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR0,perma);
TM_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR0,permb);
} else {
LOAD_CHI(basea);
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFTP(Tp,baseb);
MULT_2SPIN_DIR_PFTP(Tp,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_RECON_ACCUM;
@ -100,17 +89,16 @@
////////////////////////////////
// Xm
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(basea);
if ( localb ) {
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR3,permb);
XP_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR3,perma);
} else {
LOAD_CHI(baseb);
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFXM(Xm,basec);
MULT_2SPIN_DIR_PFXM(Xm,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_RECON_ACCUM;
@ -118,14 +106,13 @@
////////////////////////////////
// Ym
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(baseb);
if ( localc ) {
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_PROJMEM(basec);
MAYBEPERM(PERMUTE_DIR2,permc);
YP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR2,permb);
} else {
LOAD_CHI(basec);
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFYM(Ym,basea);
@ -136,8 +123,7 @@
////////////////////////////////
// Zm
////////////////////////////////
basec = st.GetInfo(ptypec,localc,permc,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(basec);
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_PROJMEM(basea);
@ -155,7 +141,6 @@
// Tm
////////////////////////////////
basea = (uint64_t)&out._odata[ss];
PREFETCH_CHIMU(basea);
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_PROJMEM(baseb);
@ -163,16 +148,16 @@
} else {
LOAD_CHI(baseb);
}
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal);
{
MULT_2SPIN_DIR_PFTM(Tm,basec);
MULT_2SPIN_DIR_PFTM(Tm,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_RECON_ACCUM;
// PREFETCH_CHIMU(basex);
SAVE_RESULT(&out._odata[ss]);
}
SAVE_RESULT(&out._odata[ss],baseb);
}
ssU++;
}
}