portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
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
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
benchmarks/Benchmark_dwf.cc
View File

@ -58,7 +58,7 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
std::vector<int> latt4 = GridDefaultLatt();
const int Ls=16;
const int Ls=8;
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
@ -122,7 +122,7 @@ int main (int argc, char ** argv)
DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5,params);
std::cout<<GridLogMessage << "Calling Dw"<<std::endl;
int ncall=100000;
int ncall=100;
{
double t0=usecond();
for(int i=0;i<ncall;i++){

4
lib/Make.inc
View File

File diff suppressed because one or more lines are too long

2
lib/qcd/action/fermion/WilsonKernelsAsm.cc
View File

@ -30,7 +30,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#if defined(AVX512) || defined (IMCI)
//#if defined (IMCI)
#include <simd/Avx512Asm.h>
#include <simd/Intel512wilson.h>
#undef VLOAD
#undef VSTORE

1044
lib/simd/Avx512Asm.h
View File

File diff suppressed because it is too large Load Diff

10
lib/simd/Intel512avx.h
View File

@ -85,6 +85,16 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define ZEND2d(Criir,Ciirr, tmp) "vshufpd $0x55," #Ciirr "," #Ciirr "," #tmp ";\n"\
"vsubpd " #tmp "," #Ciirr "," #Criir"{%k7};\n" // ri+ir ; ri+ir,rr-ii
#define VMOVRDUPd(OFF,A,DEST) "vpshufd $0x44," #OFF "*64(" #A ")," #DEST ";\n" // 32 bit level: 1,0,3,2
#define VMOVIDUPd(OFF,A,DEST) "vpshufd $0xee," #OFF "*64(" #A ")," #DEST ";\n" // 32 bit level: 3,2,3,2
#define VMOVRDUPf(OFF,PTR,DEST) "vmovsldup " #OFF "*64(" #PTR "), " #DEST ";\n"
#define VMOVIDUPf(OFF,PTR,DEST) "vmovshdup " #OFF "*64(" #PTR "), " #DEST ";\n"
#define VMADDSUBf(A,B,accum) "vfmaddsub231ps " #A "," #B "," #accum ";\n"
#define VMADDSUBd(A,B,accum) "vfmaddsub231pd " #A "," #B "," #accum ";\n"
#define VTIMESI0f(A,DEST, Z) VSHUFf(A,DEST)
#define VTIMESI1f(A,DEST, Z) "vaddps " #DEST "," #Z "," #DEST"{%k6}" ";\n"
#define VTIMESI2f(A,DEST, Z) "vsubps " #DEST "," #Z "," #DEST"{%k7}" ";\n"

14
lib/simd/Intel512avxAddsub.h
View File

@ -28,18 +28,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_ASM_AV512_ADDSUB_H
#define GRID_ASM_AV512_ADDSUB_H
////////////////////////////////////////////////////////////
// Knights Landing specials
////////////////////////////////////////////////////////////
#define VMOVRDUPd(OFF,A,DEST) "vpshufd $0x44," #OFF"*64("#A ")," #DEST ";\n" // 32 bit level: 1,0,3,2
#define VMOVIDUPd(OFF,A,DEST) "vpshufd $0xee," #OFF"*64("#A ")," #DEST ";\n" // 32 bit level: 3,2,3,2
#define VMOVRDUPf(O,P,DEST) "vmovsldup " #OFF "*64(" #PTR "), " #DEST ";\n"
#define VMOVIDUPf(O,P,DEST) "vmovshdup " #OFF "*64(" #PTR "), " #DEST ";\n"
#define VMADDSUBf(Aii,Bri,accum) "vfmaddsub231ps " #A "," #B "," #accum ";\n"
#define VMADDSUBd(Aii,Bri,accum) "vfmaddsub231pd " #A "," #B "," #accum ";\n"
////////////////////////////////////////////////////////////////
// Building blocks for SU3 x 2spinor
@ -48,7 +36,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
// 6 Chi shuffles ir,ri
// 6muls, 30 fmaddsubs
////////////////////////////////////////////////////////////////
#define MULT_ADDSUB_2SPIN_PF(ptr) \
#define MULT_ADDSUB_2SPIN(ptr) \
LOAD64(%r8,ptr) \
__asm__ ( \
VMOVIDUPf(0,%r8,Z0 ) \

67
lib/simd/Intel512wilson.h
View File

@ -66,6 +66,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define Uir %zmm24
//#define ONE %zmm24
#define Uri %zmm25
#define T1 %zmm24
#define T2 %zmm25
#define Z0 %zmm26
#define Z1 %zmm27
@ -288,7 +290,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
ZEND2(UChi_02,Z4,Chi_02) \
ZEND2(UChi_12,Z5,Chi_12) );
#define MULT_2SPIN(ptr) MULT_2SPIN_PF(ptr,ptr,VPREFETCHG);
#define MULT_2SPINa(ptr) MULT_2SPIN_PF(ptr,ptr,VPREFETCHG);
#define MULT_2SPIN(ptr) MULT_ADDSUB_2SPIN(ptr);
#define MULT_2SPIN_PFXM(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA)
#define MULT_2SPIN_PFYM(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA)
#define MULT_2SPIN_PFZM(ptr,pf) MULT_2SPIN_PF(ptr,pf,VPREFETCHNTA)
@ -750,8 +754,63 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
VPERM3(Chi_11,Chi_11) \
VPERM3(Chi_12,Chi_12) );
#ifdef AVX512
#include <simd/Intel512avxAddsub.h>
#endif
#define MULT_ADDSUB_2SPIN1(ptr) \
LOAD64(%r8,ptr)
/*
* __asm__ ( \
);
VMUL(Z0,%zmm2,%zmm3) \
*/
#define MULT_ADDSUB_2SPIN(ptr) \
LOAD64(%r8,ptr) \
__asm__ ( \
VMOVIDUP(0,%r8,Z0 ) \
VMOVIDUP(3,%r8,Z1 )\
VMOVIDUP(6,%r8,Z2 )\
VSHUF(Chi_00,T1) \
VSHUF(Chi_10,T2) \
\
VMUL(Z0,T1,UChi_00) VMOVRDUP(0,%r8,Z3 ) \
VMUL(Z0,T2,UChi_10) VMOVRDUP(3,%r8,Z4 ) \
VMUL(Z1,T1,UChi_01) VMOVRDUP(6,%r8,Z5 ) \
VMUL(Z1,T2,UChi_11) VMOVIDUP(1,%r8,Z0 ) \
VMUL(Z2,T1,UChi_02) VMOVIDUP(4,%r8,Z1 ) \
VMUL(Z2,T2,UChi_12) VMOVIDUP(7,%r8,Z2 ) \
\
VMADDSUB(Z3,Chi_00,UChi_00) VSHUF(Chi_01,T1) \
VMADDSUB(Z3,Chi_10,UChi_10) VSHUF(Chi_11,T2) \
VMADDSUB(Z4,Chi_00,UChi_01) VMOVRDUP(1,%r8,Z3 ) \
VMADDSUB(Z4,Chi_10,UChi_11)\
VMADDSUB(Z5,Chi_00,UChi_02) VMOVRDUP(4,%r8,Z4 ) \
VMADDSUB(Z5,Chi_10,UChi_12)\
\
VMADDSUB(Z0,T1,UChi_00) VMOVRDUP(7,%r8,Z5 ) \
VMADDSUB(Z0,T2,UChi_10)\
VMADDSUB(Z1,T1,UChi_01) VMOVIDUP(2,%r8,Z0 ) \
VMADDSUB(Z1,T2,UChi_11)\
VMADDSUB(Z2,T1,UChi_02) VMOVIDUP(5,%r8,Z1 ) \
VMADDSUB(Z2,T2,UChi_12) VMOVIDUP(8,%r8,Z2 ) \
\
VMADDSUB(Z3,Chi_01,UChi_00) VSHUF(Chi_02,T1) \
VMADDSUB(Z3,Chi_11,UChi_10) VSHUF(Chi_12,T2) \
VMADDSUB(Z4,Chi_01,UChi_01) VMOVRDUP(2,%r8,Z3 ) \
VMADDSUB(Z4,Chi_11,UChi_11)\
VMADDSUB(Z5,Chi_01,UChi_02) VMOVRDUP(5,%r8,Z4 ) \
VMADDSUB(Z5,Chi_11,UChi_12)\
\
VMADDSUB(Z0,T1,UChi_00) VMOVRDUP(8,%r8,Z5 ) \
VMADDSUB(Z0,T2,UChi_10)\
VMADDSUB(Z1,T1,UChi_01)\
VMADDSUB(Z1,T2,UChi_11)\
VMADDSUB(Z2,T1,UChi_02)\
VMADDSUB(Z2,T2,UChi_12)\
\
VMADDSUB(Z3,Chi_02,UChi_00)\
VMADDSUB(Z3,Chi_12,UChi_10)\
VMADDSUB(Z4,Chi_02,UChi_01)\
VMADDSUB(Z4,Chi_12,UChi_11)\
VMADDSUB(Z5,Chi_02,UChi_02)\
VMADDSUB(Z5,Chi_12,UChi_12)\
);
#endif

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);