portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-03 21:44:33 +00:00
Code Releases Activity

Lots of debug on performance Mobius

Browse Source
This commit is contained in:
Peter Boyle
2016-12-08 17:28:28 +00:00
parent ff71a8e847
commit fb8d4b2357
8 changed files with 304 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
lib/qcd/action/fermion/CayleyFermion5D.cc
View File

@@ -54,12 +54,11 @@ template<class Impl>
void CayleyFermion5D<Impl>::Dminus(const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
FermionField tmp(psi._grid);
this->DW(psi,tmp,DaggerNo);
this->DW(psi,this->tmp(),DaggerNo);
for(int s=0;s<Ls;s++){
axpby_ssp(chi,Coeff_t(1.0),psi,-cs[s],tmp,s,s);// chi = (1-c[s] D_W) psi
axpby_ssp(chi,Coeff_t(1.0),psi,-cs[s],this->tmp(),s,s);// chi = (1-c[s] D_W) psi
}
}
@@ -87,8 +86,8 @@ template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls : " << MooeeInvCalls << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
// Flops = 9*12*Ls*vol/2
RealD mflops = 9.0*12*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
// Flops = MADD * Ls *Ls *4dvol * spin/colour/complex
RealD mflops = 2.0*24*this->Ls*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
}
@@ -110,12 +109,11 @@ template<class Impl>
void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
FermionField tmp(psi._grid);
this->DW(psi,tmp,DaggerYes);
this->DW(psi,this->tmp(),DaggerYes);
for(int s=0;s<Ls;s++){
axpby_ssp(chi,Coeff_t(1.0),psi,-cs[s],tmp,s,s);// chi = (1-c[s] D_W) psi
axpby_ssp(chi,Coeff_t(1.0),psi,-cs[s],this->tmp(),s,s);// chi = (1-c[s] D_W) psi
}
}
template<class Impl>
@@ -138,6 +136,7 @@ void CayleyFermion5D<Impl>::Meooe5D (const FermionField &psi, FermionField &D
lower[0] =-mass*lower[0];
M5D(psi,psi,Din,lower,diag,upper);
}
// FIXME Redunant with the above routine; check this and eliminate
template<class Impl> void CayleyFermion5D<Impl>::Meo5D (const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
@@ -259,36 +258,33 @@ template<class Impl>
void CayleyFermion5D<Impl>::Meooe (const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
FermionField tmp(psi._grid);
Meooe5D(psi,tmp);
Meooe5D(psi,this->tmp());
if ( psi.checkerboard == Odd ) {
this->DhopEO(tmp,chi,DaggerNo);
this->DhopEO(this->tmp(),chi,DaggerNo);
} else {
this->DhopOE(tmp,chi,DaggerNo);
this->DhopOE(this->tmp(),chi,DaggerNo);
}
}
template<class Impl>
void CayleyFermion5D<Impl>::MeooeDag (const FermionField &psi, FermionField &chi)
{
FermionField tmp(psi._grid);
// Apply 4d dslash
if ( psi.checkerboard == Odd ) {
this->DhopEO(psi,tmp,DaggerYes);
this->DhopEO(psi,this->tmp(),DaggerYes);
} else {
this->DhopOE(psi,tmp,DaggerYes);
this->DhopOE(psi,this->tmp(),DaggerYes);
}
MeooeDag5D(tmp,chi);
MeooeDag5D(this->tmp(),chi);
}
template<class Impl>
void CayleyFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,int dir,int disp){
FermionField tmp(psi._grid);
Meo5D(psi,tmp);
Meo5D(psi,this->tmp());
// Apply 4d dslash fragment
this->DhopDir(tmp,chi,dir,disp);
this->DhopDir(this->tmp(),chi,dir,disp);
}
// force terms; five routines; default to Dhop on diagonal
template<class Impl>

5
lib/qcd/action/fermion/CayleyFermion5D.h
View File

@@ -76,6 +76,11 @@ namespace Grid {
std::vector<Coeff_t> &diag,
std::vector<Coeff_t> &upper);
void MooeeInternal(const FermionField &in, FermionField &out,int dag,int inv);
void MooeeInternalAsm(const FermionField &in, FermionField &out,
int LLs, int site,
Vector<iSinglet<Simd> > &Matp,
Vector<iSinglet<Simd> > &Matm);
virtual void Instantiatable(void)=0;

319
lib/qcd/action/fermion/CayleyFermion5Dvec.cc
View File

@@ -34,8 +34,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
namespace Grid {
namespace QCD {
/*
namespace QCD { /*
* Dense matrix versions of routines
*/
template<class Impl>
@@ -126,7 +125,6 @@ PARALLEL_FOR_LOOP
for(int v=0;v<LLs;v++){
vprefetch(psi[ss+v+LLs]);
// vprefetch(phi[ss+v+LLs]);
int vp= (v==LLs-1) ? 0 : v+1;
int vm= (v==0 ) ? LLs-1 : v-1;
@@ -145,9 +143,6 @@ PARALLEL_FOR_LOOP
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
// if ( ss==0) std::cout << " hp_00 " <<hp_00<<std::endl;
// if ( ss==0) std::cout << " hm_00 " <<hm_00<<std::endl;
if ( vp<=v ) {
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
@@ -165,42 +160,20 @@ PARALLEL_FOR_LOOP
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
/*
if ( ss==0) std::cout << " dphi_00 " <<d[v]()()() * phi[ss+v]()(0)(0) <<std::endl;
if ( ss==0) std::cout << " dphi_10 " <<d[v]()()() * phi[ss+v]()(1)(0) <<std::endl;
if ( ss==0) std::cout << " dphi_20 " <<d[v]()()() * phi[ss+v]()(2)(0) <<std::endl;
if ( ss==0) std::cout << " dphi_30 " <<d[v]()()() * phi[ss+v]()(3)(0) <<std::endl;
*/
Simd p_00 = d[v]()()() * phi[ss+v]()(0)(0) + l[v]()()()*hm_00;
Simd p_01 = d[v]()()() * phi[ss+v]()(0)(1) + l[v]()()()*hm_01;
Simd p_02 = d[v]()()() * phi[ss+v]()(0)(2) + l[v]()()()*hm_02;
Simd p_10 = d[v]()()() * phi[ss+v]()(1)(0) + l[v]()()()*hm_10;
Simd p_11 = d[v]()()() * phi[ss+v]()(1)(1) + l[v]()()()*hm_11;
Simd p_12 = d[v]()()() * phi[ss+v]()(1)(2) + l[v]()()()*hm_12;
Simd p_20 = d[v]()()() * phi[ss+v]()(2)(0) + u[v]()()()*hp_00;
Simd p_21 = d[v]()()() * phi[ss+v]()(2)(1) + u[v]()()()*hp_01;
Simd p_22 = d[v]()()() * phi[ss+v]()(2)(2) + u[v]()()()*hp_02;
Simd p_30 = d[v]()()() * phi[ss+v]()(3)(0) + u[v]()()()*hp_10;
Simd p_31 = d[v]()()() * phi[ss+v]()(3)(1) + u[v]()()()*hp_11;
Simd p_32 = d[v]()()() * phi[ss+v]()(3)(2) + u[v]()()()*hp_12;
// Can force these to real arithmetic and save 2x.
Simd p_00 = real_mult(d[v]()()(), phi[ss+v]()(0)(0)) + real_mult(l[v]()()(),hm_00);
Simd p_01 = real_mult(d[v]()()(), phi[ss+v]()(0)(1)) + real_mult(l[v]()()(),hm_01);
Simd p_02 = real_mult(d[v]()()(), phi[ss+v]()(0)(2)) + real_mult(l[v]()()(),hm_02);
Simd p_10 = real_mult(d[v]()()(), phi[ss+v]()(1)(0)) + real_mult(l[v]()()(),hm_10);
Simd p_11 = real_mult(d[v]()()(), phi[ss+v]()(1)(1)) + real_mult(l[v]()()(),hm_11);
Simd p_12 = real_mult(d[v]()()(), phi[ss+v]()(1)(2)) + real_mult(l[v]()()(),hm_12);
Simd p_20 = real_mult(d[v]()()(), phi[ss+v]()(2)(0)) + real_mult(u[v]()()(),hp_00);
Simd p_21 = real_mult(d[v]()()(), phi[ss+v]()(2)(1)) + real_mult(u[v]()()(),hp_01);
Simd p_22 = real_mult(d[v]()()(), phi[ss+v]()(2)(2)) + real_mult(u[v]()()(),hp_02);
Simd p_30 = real_mult(d[v]()()(), phi[ss+v]()(3)(0)) + real_mult(u[v]()()(),hp_10);
Simd p_31 = real_mult(d[v]()()(), phi[ss+v]()(3)(1)) + real_mult(u[v]()()(),hp_11);
Simd p_32 = real_mult(d[v]()()(), phi[ss+v]()(3)(2)) + real_mult(u[v]()()(),hp_12);
// if ( ss==0){
/*
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(0)(0) << " bad "<<p_00<<" diff "<<chi[ss+v]()(0)(0)-p_00<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(0)(1) << " bad "<<p_01<<" diff "<<chi[ss+v]()(0)(1)-p_01<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(0)(2) << " bad "<<p_02<<" diff "<<chi[ss+v]()(0)(2)-p_02<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(1)(0) << " bad "<<p_10<<" diff "<<chi[ss+v]()(1)(0)-p_10<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(1)(1) << " bad "<<p_11<<" diff "<<chi[ss+v]()(1)(1)-p_11<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(1)(2) << " bad "<<p_12<<" diff "<<chi[ss+v]()(1)(2)-p_12<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(2)(0) << " bad "<<p_20<<" diff "<<chi[ss+v]()(2)(0)-p_20<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(2)(1) << " bad "<<p_21<<" diff "<<chi[ss+v]()(2)(1)-p_21<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(2)(2) << " bad "<<p_22<<" diff "<<chi[ss+v]()(2)(2)-p_22<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(3)(0) << " bad "<<p_30<<" diff "<<chi[ss+v]()(3)(0)-p_30<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(3)(1) << " bad "<<p_31<<" diff "<<chi[ss+v]()(3)(1)-p_31<<std::endl;
std::cout << ss<<" "<< v<< " good "<< chi[ss+v]()(3)(2) << " bad "<<p_32<<" diff "<<chi[ss+v]()(3)(2)-p_32<<std::endl;
}
*/
vstream(chi[ss+v]()(0)(0),p_00);
vstream(chi[ss+v]()(0)(1),p_01);
vstream(chi[ss+v]()(0)(2),p_02);
@@ -261,7 +234,7 @@ void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi,
M5Dtime-=usecond();
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=LLs){ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
@@ -287,9 +260,231 @@ PARALLEL_FOR_LOOP
chi[ss+v] = chi[ss+v] +l[v]*fm;
}
#else
for(int v=0;v<LLs;v++){
vprefetch(psi[ss+v+LLs]);
int vp= (v==LLs-1) ? 0 : v+1;
int vm= (v==0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if ( vp<=v ) {
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if ( vm>=v ) {
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
Simd p_00 = real_mult(d[v]()()(), phi[ss+v]()(0)(0)) + real_mult(u[v]()()(),hp_00);
Simd p_01 = real_mult(d[v]()()(), phi[ss+v]()(0)(1)) + real_mult(u[v]()()(),hp_01);
Simd p_02 = real_mult(d[v]()()(), phi[ss+v]()(0)(2)) + real_mult(u[v]()()(),hp_02);
Simd p_10 = real_mult(d[v]()()(), phi[ss+v]()(1)(0)) + real_mult(u[v]()()(),hp_10);
Simd p_11 = real_mult(d[v]()()(), phi[ss+v]()(1)(1)) + real_mult(u[v]()()(),hp_11);
Simd p_12 = real_mult(d[v]()()(), phi[ss+v]()(1)(2)) + real_mult(u[v]()()(),hp_12);
Simd p_20 = real_mult(d[v]()()(), phi[ss+v]()(2)(0)) + real_mult(l[v]()()(),hm_00);
Simd p_21 = real_mult(d[v]()()(), phi[ss+v]()(2)(1)) + real_mult(l[v]()()(),hm_01);
Simd p_22 = real_mult(d[v]()()(), phi[ss+v]()(2)(2)) + real_mult(l[v]()()(),hm_02);
Simd p_30 = real_mult(d[v]()()(), phi[ss+v]()(3)(0)) + real_mult(l[v]()()(),hm_10);
Simd p_31 = real_mult(d[v]()()(), phi[ss+v]()(3)(1)) + real_mult(l[v]()()(),hm_11);
Simd p_32 = real_mult(d[v]()()(), phi[ss+v]()(3)(2)) + real_mult(l[v]()()(),hm_12);
vstream(chi[ss+v]()(0)(0),p_00);
vstream(chi[ss+v]()(0)(1),p_01);
vstream(chi[ss+v]()(0)(2),p_02);
vstream(chi[ss+v]()(1)(0),p_10);
vstream(chi[ss+v]()(1)(1),p_11);
vstream(chi[ss+v]()(1)(2),p_12);
vstream(chi[ss+v]()(2)(0),p_20);
vstream(chi[ss+v]()(2)(1),p_21);
vstream(chi[ss+v]()(2)(2),p_22);
vstream(chi[ss+v]()(3)(0),p_30);
vstream(chi[ss+v]()(3)(1),p_31);
vstream(chi[ss+v]()(3)(2),p_32);
}
#endif
}
M5Dtime+=usecond();
}
#include <simd/Intel512common.h>
#include <simd/Intel512avx.h>
#include <simd/Intel512single.h>
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternalAsm(const FermionField &psi, FermionField &chi,
int LLs, int site,
Vector<iSinglet<Simd> > &Matp,
Vector<iSinglet<Simd> > &Matm)
{
#if 0
{
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
SiteHalfSpinor SiteChiP;
SiteHalfSpinor SiteChiM;
// Ls*Ls * 2 * 12 * vol flops
for(int s1=0;s1<LLs;s1++){
for(int s2=0;s2<LLs;s2++){
for(int l=0; l<Simd::Nsimd();l++){ // simd lane
int s=s2+l*LLs;
int lex=s2+LLs*site;
if ( s2==0 && l==0) {
SiteChiP=zero;
SiteChiM=zero;
}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vbroadcast(BcastP()(sp )(co),psi[lex]()(sp)(co),l);
}}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vbroadcast(BcastM()(sp )(co),psi[lex]()(sp+2)(co),l);
}}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
SiteChiP()(sp)(co)=real_madd(Matp[LLs*s+s1]()()(),BcastP()(sp)(co),SiteChiP()(sp)(co)); // 1100 us.
SiteChiM()(sp)(co)=real_madd(Matm[LLs*s+s1]()()(),BcastM()(sp)(co),SiteChiM()(sp)(co)); // each found by commenting out
}}
}}
{
int lex = s1+LLs*site;
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vstream(chi[lex]()(sp)(co), SiteChiP()(sp)(co));
vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
}}
}
}
}
#else
{
// pointers
// MASK_REGS;
#define Chi_00 %%zmm1
#define Chi_01 %%zmm2
#define Chi_02 %%zmm3
#define Chi_10 %%zmm4
#define Chi_11 %%zmm5
#define Chi_12 %%zmm6
#define Chi_20 %%zmm7
#define Chi_21 %%zmm8
#define Chi_22 %%zmm9
#define Chi_30 %%zmm10
#define Chi_31 %%zmm11
#define Chi_32 %%zmm12
#define BCAST0 %%zmm13
#define BCAST1 %%zmm14
#define BCAST2 %%zmm15
#define BCAST3 %%zmm16
#define BCAST4 %%zmm17
#define BCAST5 %%zmm18
#define BCAST6 %%zmm19
#define BCAST7 %%zmm20
#define BCAST8 %%zmm21
#define BCAST9 %%zmm22
#define BCAST10 %%zmm23
#define BCAST11 %%zmm24
int incr=LLs*LLs*sizeof(iSinglet<Simd>);
for(int s1=0;s1<LLs;s1++){
for(int s2=0;s2<LLs;s2++){
int lex=s2+LLs*site;
uint64_t a0 = (uint64_t)&Matp[LLs*s2+s1]; // should be cacheable
uint64_t a1 = (uint64_t)&Matm[LLs*s2+s1];
uint64_t a2 = (uint64_t)&psi[lex];
for(int l=0; l<Simd::Nsimd();l++){ // simd lane
if ( (s2+l)==0 ) {
asm (
VPREFETCH1(0,%2) VPREFETCH1(0,%1)
VPREFETCH1(12,%2) VPREFETCH1(13,%2)
VPREFETCH1(14,%2) VPREFETCH1(15,%2)
VBCASTCDUP(0,%2,BCAST0)
VBCASTCDUP(1,%2,BCAST1)
VBCASTCDUP(2,%2,BCAST2)
VBCASTCDUP(3,%2,BCAST3)
VBCASTCDUP(4,%2,BCAST4) VMULMEM (0,%0,BCAST0,Chi_00)
VBCASTCDUP(5,%2,BCAST5) VMULMEM (0,%0,BCAST1,Chi_01)
VBCASTCDUP(6,%2,BCAST6) VMULMEM (0,%0,BCAST2,Chi_02)
VBCASTCDUP(7,%2,BCAST7) VMULMEM (0,%0,BCAST3,Chi_10)
VBCASTCDUP(8,%2,BCAST8) VMULMEM (0,%0,BCAST4,Chi_11)
VBCASTCDUP(9,%2,BCAST9) VMULMEM (0,%0,BCAST5,Chi_12)
VBCASTCDUP(10,%2,BCAST10) VMULMEM (0,%1,BCAST6,Chi_20)
VBCASTCDUP(11,%2,BCAST11) VMULMEM (0,%1,BCAST7,Chi_21)
VMULMEM (0,%1,BCAST8,Chi_22)
VMULMEM (0,%1,BCAST9,Chi_30)
VMULMEM (0,%1,BCAST10,Chi_31)
VMULMEM (0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
} else {
asm (
VBCASTCDUP(0,%2,BCAST0) VMADDMEM (0,%0,BCAST0,Chi_00)
VBCASTCDUP(1,%2,BCAST1) VMADDMEM (0,%0,BCAST1,Chi_01)
VBCASTCDUP(2,%2,BCAST2) VMADDMEM (0,%0,BCAST2,Chi_02)
VBCASTCDUP(3,%2,BCAST3) VMADDMEM (0,%0,BCAST3,Chi_10)
VBCASTCDUP(4,%2,BCAST4) VMADDMEM (0,%0,BCAST4,Chi_11)
VBCASTCDUP(5,%2,BCAST5) VMADDMEM (0,%0,BCAST5,Chi_12)
VBCASTCDUP(6,%2,BCAST6) VMADDMEM (0,%1,BCAST6,Chi_20)
VBCASTCDUP(7,%2,BCAST7) VMADDMEM (0,%1,BCAST7,Chi_21)
VBCASTCDUP(8,%2,BCAST8) VMADDMEM (0,%1,BCAST8,Chi_22)
VBCASTCDUP(9,%2,BCAST9) VMADDMEM (0,%1,BCAST9,Chi_30)
VBCASTCDUP(10,%2,BCAST10) VMADDMEM (0,%1,BCAST10,Chi_31)
VBCASTCDUP(11,%2,BCAST11) VMADDMEM (0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
}
a0 = a0+incr;
a1 = a1+incr;
a2 = a2+sizeof(Simd::scalar_type);
}}
{
int lexa = s1+LLs*site;
asm (
VSTORE(0,%0,Chi_00) VSTORE(1 ,%0,Chi_01) VSTORE(2 ,%0,Chi_02)
VSTORE(3,%0,Chi_10) VSTORE(4 ,%0,Chi_11) VSTORE(5 ,%0,Chi_12)
VSTORE(6,%0,Chi_20) VSTORE(7 ,%0,Chi_21) VSTORE(8 ,%0,Chi_22)
VSTORE(9,%0,Chi_30) VSTORE(10,%0,Chi_31) VSTORE(11,%0,Chi_32)
: : "r" ((uint64_t)&chi[lexa]) : "memory" );
}
}
}
#endif
};
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv)
{
@@ -347,32 +542,33 @@ void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField
scalar_type *sp = (scalar_type *)&Vp;
scalar_type *sm = (scalar_type *)&Vm;
for(int l=0;l<Nsimd;l++){
sp[l] = PplusMat (l*istride+s1*ostride ,s2);
sp[l] = PplusMat (l*istride+s1*ostride,s2);
sp[l] = scalar_type(sp[l].real(),sp[l].real());
sm[l] = PminusMat(l*istride+s1*ostride,s2);
sm[l] = scalar_type(sm[l].real(),sm[l].real());
}
Matp[LLs*s2+s1] = Vp;
Matm[LLs*s2+s1] = Vm;
}
}
}}
MooeeInvCalls++;
MooeeInvTime-=usecond();
// Dynamic allocate on stack to get per thread without serialised heap acces
#if 0
#pragma omp parallel
{
Vector<SiteHalfSpinor> SitePplus(LLs);
Vector<SiteHalfSpinor> SitePminus(LLs);
Vector<SiteHalfSpinor> SiteChiP(LLs);
Vector<SiteHalfSpinor> SiteChiM(LLs);
Vector<SiteSpinor> SiteChi(LLs);
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
std::vector<SiteHalfSpinor> SitePplus(LLs);
std::vector<SiteHalfSpinor> SitePminus(LLs);
std::vector<SiteHalfSpinor> SiteChiP(LLs);
std::vector<SiteHalfSpinor> SiteChiM(LLs);
std::vector<SiteSpinor> SiteChi(LLs);
#pragma omp for
for(auto site=0;site<vol;site++){
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
for(int s=0;s<LLs;s++){
int lex = s+LLs*site;
spProj5p(SitePplus[s] ,psi[lex]);
@@ -391,7 +587,8 @@ void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField
SiteChiM[s1]=SiteChiM[s1]+Matm[LLs*s+s1]*BcastM;
}
s++;
}}
}
}
for(int s=0;s<LLs;s++){
int lex = s+LLs*site;
@@ -399,8 +596,16 @@ void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField
accumRecon5m(SiteChi[s],SiteChiM[s]);
chi[lex] = SiteChi[s]*0.5;
}
}}
#else
PARALLEL_FOR_LOOP
for(auto site=0;site<vol;site++){
MooeeInternalAsm(psi,chi,
LLs,site,
Matp,Matm);
}
}
#endif
MooeeInvTime+=usecond();
}

2
lib/qcd/action/fermion/FermionOperator.h
View File

@@ -48,6 +48,8 @@ namespace Grid {
FermionOperator(const ImplParams &p= ImplParams()) : Impl(p) {};
virtual FermionField &tmp(void) = 0;
GridBase * Grid(void) { return FermionGrid(); }; // this is all the linalg routines need to know
GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); };

4
lib/qcd/action/fermion/WilsonFermion.cc
View File

@@ -61,7 +61,9 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
LebesgueEvenOdd(_cbgrid),
Umu(&Fgrid),
UmuEven(&Hgrid),
UmuOdd(&Hgrid) {
UmuOdd(&Hgrid),
_tmp(&Hgrid)
{
// Allocate the required comms buffer
ImportGauge(_Umu);
}

3
lib/qcd/action/fermion/WilsonFermion.h
View File

@@ -58,6 +58,9 @@ class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic {
GridBase *FermionGrid(void) { return _grid; }
GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
//////////////////////////////////////////////////////////////////
// override multiply; cut number routines if pass dagger argument
// and also make interface more uniformly consistent

3
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@@ -60,7 +60,8 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
UmuEven(_FourDimRedBlackGrid),
UmuOdd (_FourDimRedBlackGrid),
Lebesgue(_FourDimGrid),
LebesgueEvenOdd(_FourDimRedBlackGrid)
LebesgueEvenOdd(_FourDimRedBlackGrid),
_tmp(&FiveDimRedBlackGrid)
{
if (Impl::LsVectorised) {

3
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@@ -74,6 +74,9 @@ namespace QCD {
typedef WilsonKernels<Impl> Kernels;
PmuStat stat;
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
void Report(void);
void ZeroCounters(void);
double DhopCalls;