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

Improved the prefetching when using cache blocking codes

Browse Source
This commit is contained in:
paboyle
2016-06-26 12:54:14 -07:00
parent 1445189361
commit 8fcefc021a
5 changed files with 208 additions and 162 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
lib/qcd/action/fermion/WilsonKernelsAsmBody.h.ab
View File

@ -2,7 +2,6 @@
int locala,perma, ptypea;
int localb,permb, ptypeb;
uint64_t basea, baseb;
uint64_t basex;
const uint64_t plocal =(uint64_t) & in._odata[0];
// vComplexF isigns[2] = { signs[0], signs[1] };
@ -19,9 +18,7 @@
////////////////////////////////
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++;
basex = basea;
if ( locala ) {
LOAD64(%r10,isigns);
@ -39,7 +36,7 @@
////////////////////////////////
// Yp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
basea = st.GetInfo(ptypea,locala,perma,Zp,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_PROJMEM(baseb);
@ -56,7 +53,7 @@
////////////////////////////////
// Zp
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
baseb = st.GetInfo(ptypeb,localb,permb,Tp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_PROJMEM(basea);
@ -73,7 +70,7 @@
////////////////////////////////
// Tp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
basea = st.GetInfo(ptypea,locala,perma,Xm,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_PROJMEM(baseb);
@ -90,7 +87,7 @@
////////////////////////////////
// Xm
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
baseb = st.GetInfo(ptypeb,localb,permb,Ym,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_PROJMEM(basea);
@ -107,7 +104,7 @@
////////////////////////////////
// Ym
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
basea = st.GetInfo(ptypea,locala,perma,Zm,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_PROJMEM(baseb);
@ -124,7 +121,7 @@
////////////////////////////////
// Zm
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
baseb = st.GetInfo(ptypeb,localb,permb,Tm,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_PROJMEM(basea);
@ -149,13 +146,14 @@
} else {
LOAD_CHI(baseb);
}
baseb = st.GetInfo(ptypeb,localb,permb,Xp,ent,plocal);
{
MULT_2SPIN_DIR_PFTM(Tm,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_RECON_ACCUM;
SAVE_RESULT(&out._odata[ss]);
SAVE_RESULT(&out._odata[ss],baseb);
}
ssU++;