portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-04 19:25:56 +01:00
Code Releases Activity

Benchmark_zmm added

Browse Source
This commit is contained in:
paboyle 2016-04-06 03:00:07 -07:00
parent e8dddb1596
commit b1192a8908
5 changed files with 29 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
benchmarks/Benchmark_zmm.cc
View File

@ -27,21 +27,11 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */ /* END LEGAL */
#include <Grid.h> #include <Grid.h>
#include <PerfCount.h> #include <PerfCount.h>
#include <simd/Intel512wilson.h>
using namespace Grid; using namespace Grid;
using namespace Grid::QCD; using namespace Grid::QCD;
void ZmulF(void *ptr1,void *ptr2,void *ptr3);
void Zmul(void *ptr1,void *ptr2,void *ptr3);
void WilsonDslashAvx512(void *ptr1,void *ptr2,void *ptr3);
void WilsonDslashAvx512F(void *ptr1,void *ptr2,void *ptr3);
void TimesIAvx512F(void *ptr1,void *ptr3);
void TimesIAvx512(void *ptr1,void *ptr3);
void TimesMinusIAvx512F(void *ptr1,void *ptr3);
void TimesMinusIAvx512(void *ptr1,void *ptr3);
int bench(std::ofstream &os, std::vector<int> &latt4,int Ls); int bench(std::ofstream &os, std::vector<int> &latt4,int Ls);
@ -55,7 +45,7 @@ int main(int argc,char **argv)
for(int m=1;m<=2;m++){ for(int m=1;m<=2;m++){
for(int Ls=8;Ls<=16;Ls+=8){ for(int Ls=8;Ls<=16;Ls+=8){
std::vector<int> grid({L,L,m*L,m*L}); std::vector<int> grid({L,L,m*L,m*L});
bench(os,latt4,Ls); bench(os,grid,Ls);
} }
} }
} }
@ -134,6 +124,7 @@ int bench(std::ofstream &os, std::vector<int> &latt4,int Ls)
mfa = flops*ncall/(t1-t0); mfa = flops*ncall/(t1-t0);
std::cout<<GridLogMessage << "Called ASM Dw"<< " mflop/s = "<< mfa<<std::endl; std::cout<<GridLogMessage << "Called ASM Dw"<< " mflop/s = "<< mfa<<std::endl;
int dag=DaggerNo;
t0=usecond(); t0=usecond();
for(int i=0;i<1;i++){ for(int i=0;i<1;i++){
Dw.DhopInternalOMPbench(Dw.StencilOdd,Dw.LebesgueEvenOdd,Dw.UmuEven,srce,junk,dag); Dw.DhopInternalOMPbench(Dw.StencilOdd,Dw.LebesgueEvenOdd,Dw.UmuEven,srce,junk,dag);

6
benchmarks/Make.inc
View File

@ -1,5 +1,5 @@
bin_PROGRAMS = Benchmark_comms Benchmark_dwf Benchmark_memory_asynch Benchmark_memory_bandwidth Benchmark_su3 Benchmark_wilson bin_PROGRAMS = Benchmark_comms Benchmark_dwf Benchmark_memory_asynch Benchmark_memory_bandwidth Benchmark_su3 Benchmark_wilson Benchmark_zmm
Benchmark_comms_SOURCES=Benchmark_comms.cc Benchmark_comms_SOURCES=Benchmark_comms.cc
@ -25,3 +25,7 @@ Benchmark_su3_LDADD=-lGrid
Benchmark_wilson_SOURCES=Benchmark_wilson.cc Benchmark_wilson_SOURCES=Benchmark_wilson.cc
Benchmark_wilson_LDADD=-lGrid Benchmark_wilson_LDADD=-lGrid
Benchmark_zmm_SOURCES=Benchmark_zmm.cc
Benchmark_zmm_LDADD=-lGrid

4
lib/Make.inc
View File

File diff suppressed because one or more lines are too long

10
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@ -1,3 +1,4 @@
/************************************************************************************* /*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid Grid physics library, www.github.com/paboyle/Grid
@ -127,6 +128,13 @@ namespace Grid {
FermionField &out, FermionField &out,
int dag); int dag);
void DhopInternalL1bench(StencilImpl & st,
LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out,
int dag);
void DhopInternalCommsThenCompute(StencilImpl & st, void DhopInternalCommsThenCompute(StencilImpl & st,
LebesgueOrder &lo, LebesgueOrder &lo,
DoubledGaugeField &U, DoubledGaugeField &U,
@ -155,7 +163,7 @@ namespace Grid {
/////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////
// Data members require to support the functionality // Data members require to support the functionality
/////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////
protected: public:
// Add these to the support from Wilson // Add these to the support from Wilson
GridBase *_FourDimGrid; GridBase *_FourDimGrid;

22
tests/Make.inc
View File

@ -1,5 +1,13 @@
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 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
Test_cayley_cg_SOURCES=Test_cayley_cg.cc Test_cayley_cg_SOURCES=Test_cayley_cg.cc
@ -94,8 +102,8 @@ Test_gamma_SOURCES=Test_gamma.cc
Test_gamma_LDADD=-lGrid Test_gamma_LDADD=-lGrid
Test_GaugeAction_SOURCES=Test_GaugeAction.cc Test_gp_rect_force_SOURCES=Test_gp_rect_force.cc
Test_GaugeAction_LDADD=-lGrid Test_gp_rect_force_LDADD=-lGrid
Test_gparity_SOURCES=Test_gparity.cc Test_gparity_SOURCES=Test_gparity.cc
@ -106,10 +114,6 @@ Test_gpdwf_force_SOURCES=Test_gpdwf_force.cc
Test_gpdwf_force_LDADD=-lGrid 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_SOURCES=Test_gpwilson_even_odd.cc
Test_gpwilson_even_odd_LDADD=-lGrid Test_gpwilson_even_odd_LDADD=-lGrid
@ -186,10 +190,6 @@ Test_rect_force_SOURCES=Test_rect_force.cc
Test_rect_force_LDADD=-lGrid Test_rect_force_LDADD=-lGrid
Test_RectPlaq_SOURCES=Test_RectPlaq.cc
Test_RectPlaq_LDADD=-lGrid
Test_remez_SOURCES=Test_remez.cc Test_remez_SOURCES=Test_remez.cc
Test_remez_LDADD=-lGrid Test_remez_LDADD=-lGrid