portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 16:27:05 +01:00
Code Releases Activity

AddSub based alternate SU3 routine

Browse Source
This commit is contained in:
paboyle
2016-03-28 17:55:22 -06:00
parent b6c3bc574b
commit c77b7ee897
9 changed files with 116 additions and 1086 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
tests/Test_zmm.cc
View File

@ -27,7 +27,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */
#include <Grid.h>
#include <PerfCount.h>
#include <simd/Avx512Asm.h>
#include <simd/Intel512wilson.h>
using namespace Grid;
@ -261,6 +261,10 @@ int main(int argc,char **argv)
#undef ZLOAD
#undef ZMUL
#undef ZMADD
#undef VMOVIDUP
#undef VMOVRDUP
#undef VMADDSUB
#undef VSHUF
#define VZERO(A) VZEROd(A)
#define VTIMESI(A,B,C) VTIMESId(A,B,C)
@ -268,15 +272,19 @@ int main(int argc,char **argv)
#define VLOAD(OFF,PTR,DEST) VLOADd(OFF,PTR,DEST)
#define VSTORE(OFF,PTR,SRC) VSTOREd(OFF,PTR,SRC)
#define VMUL(Uri,Uir,Chi,UChi,Z) VMULd(Uri,Uir,Chi,UChi,Z)
#define VMADD(Uri,Uir,Chi,UChi,Z) VMADDd(Uri,Uir,Chi,UChi,Z)
#define ZEND1(A,B,C) ZEND1d(A,B,C)
#define ZEND2(A,B,C) ZEND2d(A,B,C)
#define VMUL(Uri,Uir,Chi) VMULd(Uri,Uir,Chi)
#define VMADD(Uri,Uir,Chi) VMADDd(Uri,Uir,Chi)
#define ZEND1(A,B,C) ZEND1d(A,B,C)
#define ZEND2(A,B,C) ZEND2d(A,B,C)
#define ZLOAD(A,B,C,D) ZLOADd(A,B,C,D)
#define ZMUL(A,B,C,D,E) ZMULd(A,B,C,D,E)
#define ZMADD(A,B,C,D,E) ZMADDd(A,B,C,D,E)
#define ZMULMEM2SP(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr) ZMULMEM2SPd(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr)
#define ZMADDMEM2SP(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr) ZMADDMEM2SPd(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr)
#define VMOVIDUP(A,B,C) VMOVIDUPd(A,B,C)
#define VMOVRDUP(A,B,C) VMOVRDUPd(A,B,C)
#define VMADDSUB(A,B,accum) VMADDSUBd(A,B,accum)
#define VSHUF(A,B) VSHUFd(A,B)
#define zz Z0
@ -424,17 +432,21 @@ void WilsonDslashAvx512(void *ptr1,void *ptr2,void *ptr3)
#undef VTIMESMINUSI
#undef ZMULMEM2SP
#undef ZMADDMEM2SP
#undef VMOVIDUP
#undef VMOVRDUP
#undef VMADDSUB
#undef VSHUF
#define VZERO(A) VZEROf(A)
#define VMOV(A,B) VMOVf(A,B)
#define VADD(A,B,C) VADDf(A,B,C)
#define VSUB(A,B,C) VSUBf(A,B,C)
#define VTIMESI(A,B,C) VTIMESIf(A,B,C)
#define VTIMESMINUSI(A,B,C) VTIMESMINUSIf(A,B,C)
#define VLOAD(OFF,PTR,DEST) VLOADf(OFF,PTR,DEST)
#define VSTORE(OFF,PTR,SRC) VSTOREf(OFF,PTR,SRC)
#define VMUL(Uri,Uir,Chi,UChi,Z) VMULf(Uri,Uir,Chi,UChi,Z)
#define VMADD(Uri,Uir,Chi,UChi,Z) VMADDf(Uri,Uir,Chi,UChi,Z)
#define VMUL(Uri,Uir,Chi) VMULf(Uri,Uir,Chi)
#define VMADD(Uri,Uir,Chi) VMADDf(Uri,Uir,Chi)
#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)
@ -442,6 +454,10 @@ void WilsonDslashAvx512(void *ptr1,void *ptr2,void *ptr3)
#define ZMADD(A,B,C,D,E) ZMADDf(A,B,C,D,E)
#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)
#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)
void ZmulF(void *ptr1,void *ptr2,void *ptr3)
{
@ -528,7 +544,8 @@ void WilsonDslashAvx512F(void *ptr1,void *ptr2,void *ptr3)
LOAD_CHI(ptr1);
MULT_2SPIN(ptr2);
MULT_ADDSUB_2SPIN(ptr2);
//MULT_2SPIN(ptr2);
SAVE_UCHI(ptr3);