portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 20:57:06 +01:00
Code Releases Activity

Cleaning up the single/double kernel implementation switch

Browse Source
This commit is contained in:
paboyle
2016-03-31 14:51:32 +01:00
parent 60d965f79e
commit 8052556275
4 changed files with 281 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files

97
lib/qcd/action/fermion/WilsonKernelsAsm.cc
View File

@ -32,81 +32,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <simd/Intel512wilson.h>
#undef VLOAD
#undef VSTORE
#undef VMUL
#undef VMADD
#undef ZEND
#undef ZLOAD
#undef ZMUL
#undef ZMADD
#undef VZERO
#undef VTIMESI
#undef VTIMESMINUSI
#undef VMOVIDUP
#undef VMOVRDUP
#undef VMADDSUB
#undef VSHUF
#include <simd/Intel512single.h>
#define VZERO(A) VZEROf(A)
#define VMOV(A,B) VMOVf(A,B)
#define VLOAD(OFF,PTR,DEST) VLOADf(OFF,PTR,DEST)
#define VSTORE(OFF,PTR,SRC) VSTOREf(OFF,PTR,SRC)
#define VADD(A,B,C) VADDf(A,B,C)
#define VSUB(A,B,C) VSUBf(A,B,C)
#define VMUL(Uri,Uir,Chi) VMULf(Uri,Uir,Chi)
#define VMADD(Uri,Uir,Chi) VMADDf(Uri,Uir,Chi)
#define VTIMESI(A,B,C) VTIMESIf(A,B,C)
#define VTIMESMINUSI(A,B,C) VTIMESMINUSIf(A,B,C)
#define VACCTIMESI(A,B,C) VACCTIMESIf(A,B,C)
#define VACCTIMESMINUSI(A,B,C) VACCTIMESMINUSIf(A,B,C)
#define VTIMESI0(A,B,C) VTIMESI0f(A,B,C)
#define VTIMESMINUSI0(A,B,C) VTIMESMINUSI0f(A,B,C)
#define VACCTIMESI0(A,B,C) VACCTIMESI0f(A,B,C)
#define VACCTIMESMINUSI0(A,B,C) VACCTIMESMINUSI0f(A,B,C)
#define VTIMESI1(A,B,C) VTIMESI1f(A,B,C)
#define VTIMESMINUSI1(A,B,C) VTIMESMINUSI1f(A,B,C)
#define VACCTIMESI1(A,B,C) VACCTIMESI1f(A,B,C)
#define VACCTIMESMINUSI1(A,B,C) VACCTIMESMINUSI1f(A,B,C)
#define VTIMESI2(A,B,C) VTIMESI2f(A,B,C)
#define VTIMESMINUSI2(A,B,C) VTIMESMINUSI2f(A,B,C)
#define VACCTIMESI2(A,B,C) VACCTIMESI2f(A,B,C)
#define VACCTIMESMINUSI2(A,B,C) VACCTIMESMINUSI2f(A,B,C)
#define VACCTIMESI1MEM(A,ACC,O,P) VACCTIMESI1MEMf(A,ACC,O,P)
#define VACCTIMESI2MEM(A,ACC,O,P) VACCTIMESI2MEMf(A,ACC,O,P)
#define VACCTIMESMINUSI1MEM(A,ACC,O,P) VACCTIMESMINUSI1MEMf(A,ACC,O,P)
#define VACCTIMESMINUSI2MEM(A,ACC,O,P) VACCTIMESMINUSI2MEMf(A,ACC,O,P)
#define VPERM0(A,B) VPERM0f(A,B)
#define VPERM1(A,B) VPERM1f(A,B)
#define VPERM2(A,B) VPERM2f(A,B)
#define VPERM3(A,B) VPERM3f(A,B)
#define VSHUFMEM(OFF,A,DEST) VSHUFMEMf(OFF,A,DEST)
#define ZEND1(A,B,C) ZEND1f(A,B,C)
#define ZEND2(A,B,C) ZEND2f(A,B,C)
#define ZLOAD(A,B,C,D) ZLOADf(A,B,C,D)
#define ZMUL(A,B,C,D,E) ZMULf(A,B,C,D,E)
#define ZMADD(A,B,C,D,E) ZMADDf(A,B,C,D,E)
#define ZMUL(A,B,C,D,E) ZMULf(A,B,C,D,E)
#define ZMADD(A,B,C,D,E) ZMADDf(A,B,C,D,E)
#define VADDMEM(O,A,B,C) VADDMEMf(O,A,B,C)
#define VSUBMEM(O,A,B,C) VSUBMEMf(O,A,B,C)
#define VMOVIDUP(A,B,C) VMOVIDUPf(A,B,C)
#define VMOVRDUP(A,B,C) VMOVRDUPf(A,B,C)
#define VMADDSUB(A,B,accum) VMADDSUBf(A,B,accum)
#define VSHUF(A,B) VSHUFf(A,B)
#define ZMULMEM2SP(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr) ZMULMEM2SPf(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr)
#define ZMADDMEM2SP(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr) ZMADDMEM2SPf(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr)
namespace Grid {
namespace QCD {
@ -136,26 +63,6 @@ void WilsonKernels<Impl >::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField
SE=st.GetEntry(ptype,Xm,ss);
#if 0
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];
else pf=(void *)&pbuf[SE->_offset];
LOAD64(%r9,pf);
__asm__(
VPREFETCH(0,%r9)
VPREFETCH(1,%r9)
VPREFETCH(2,%r9)
VPREFETCH(3,%r9)
VPREFETCH(4,%r9)
VPREFETCH(5,%r9)
VPREFETCH(6,%r9)
VPREFETCH(7,%r9)
VPREFETCH(8,%r9)
VPREFETCH(9,%r9)
VPREFETCH(10,%r9)
VPREFETCH(11,%r9) );
#endif
// Xm
offset = SE->_offset;
local = SE->_is_local;
@ -322,8 +229,6 @@ void WilsonKernels<Impl >::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField
offset = SE->_offset;
local = SE->_is_local;
// PREFETCH_R(A);
// Prefetch
SE=st.GetEntry(ptype,Xm,(ss+1)%osites);
if (SE->_is_local) pf=(void *)&plocal[SE->_offset];