portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-18 07:47:06 +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

22
tests/Make.inc
View File

@ -1,13 +1,5 @@
bin_PROGRAMS = Test_GaugeAction Test_RectPlaq Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_contfrac_force Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_gpforce Test_dwf_hdcr Test_dwf_lanczos Test_gamma Test_gp_rect_force Test_gparity Test_gpdwf_force Test_gpwilson_even_odd Test_hmc_EODWFRatio Test_hmc_EODWFRatio_Gparity Test_hmc_EOWilsonFermionGauge Test_hmc_EOWilsonRatio Test_hmc_GparityIwasakiGauge Test_hmc_GparityWilsonGauge Test_hmc_IwasakiGauge Test_hmc_RectGauge Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_hmc_WilsonRatio Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_partfrac_force Test_quenched_update Test_rect_force Test_remez Test_rhmc_EOWilson1p1 Test_rhmc_EOWilsonRatio Test_rhmc_Wilson1p1 Test_rhmc_WilsonRatio Test_rng Test_rng_fixed Test_serialisation Test_simd Test_stencil Test_synthetic_lanczos Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi Test_wilson_tm_even_odd
Test_GaugeAction_SOURCES=Test_GaugeAction.cc
Test_GaugeAction_LDADD=-lGrid
Test_RectPlaq_SOURCES=Test_RectPlaq.cc
Test_RectPlaq_LDADD=-lGrid
bin_PROGRAMS = Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_contfrac_force Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_gpforce Test_dwf_hdcr Test_dwf_lanczos Test_gamma Test_GaugeAction Test_gparity Test_gpdwf_force Test_gp_rect_force Test_gpwilson_even_odd Test_hmc_EODWFRatio Test_hmc_EODWFRatio_Gparity Test_hmc_EOWilsonFermionGauge Test_hmc_EOWilsonRatio Test_hmc_GparityIwasakiGauge Test_hmc_GparityWilsonGauge Test_hmc_IwasakiGauge Test_hmc_RectGauge Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_hmc_WilsonRatio Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_partfrac_force Test_quenched_update Test_rect_force Test_RectPlaq Test_remez Test_rhmc_EOWilson1p1 Test_rhmc_EOWilsonRatio Test_rhmc_Wilson1p1 Test_rhmc_WilsonRatio Test_rng Test_rng_fixed Test_serialisation Test_simd Test_stencil Test_synthetic_lanczos Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi Test_wilson_tm_even_odd
Test_cayley_cg_SOURCES=Test_cayley_cg.cc
@ -102,8 +94,8 @@ Test_gamma_SOURCES=Test_gamma.cc
Test_gamma_LDADD=-lGrid
Test_gp_rect_force_SOURCES=Test_gp_rect_force.cc
Test_gp_rect_force_LDADD=-lGrid
Test_GaugeAction_SOURCES=Test_GaugeAction.cc
Test_GaugeAction_LDADD=-lGrid
Test_gparity_SOURCES=Test_gparity.cc
@ -114,6 +106,10 @@ Test_gpdwf_force_SOURCES=Test_gpdwf_force.cc
Test_gpdwf_force_LDADD=-lGrid
Test_gp_rect_force_SOURCES=Test_gp_rect_force.cc
Test_gp_rect_force_LDADD=-lGrid
Test_gpwilson_even_odd_SOURCES=Test_gpwilson_even_odd.cc
Test_gpwilson_even_odd_LDADD=-lGrid
@ -190,6 +186,10 @@ Test_rect_force_SOURCES=Test_rect_force.cc
Test_rect_force_LDADD=-lGrid
Test_RectPlaq_SOURCES=Test_RectPlaq.cc
Test_RectPlaq_LDADD=-lGrid
Test_remez_SOURCES=Test_remez.cc
Test_remez_LDADD=-lGrid

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);