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

MooeeInv offloaded to GPU

Browse Source
This commit is contained in:
gfilaci 2019-05-13 12:37:12 +01:00
parent 22e35c9ddd
commit 1179123ac2
2 changed files with 52 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
Grid/qcd/action/fermion/CayleyFermion5D.cc
View File

@ -180,11 +180,16 @@ template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls : " << MooeeInvCalls << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
#ifdef GRID_NVCC
RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*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;
#else
// 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;
#endif
}
}

75
Grid/qcd/action/fermion/CayleyFermion5Dgpu.cc
View File

@ -143,42 +143,59 @@ void CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi_i, FermionField
auto psi = psi_i.View();
auto chi = chi_i.View();
Coeff_t *lee_v = &lee[0];
Coeff_t *leem_v = &leem[0];
Coeff_t *uee_v = &uee[0];
Coeff_t *ueem_v = &ueem[0];
Coeff_t *dee_v = &dee[0];
int Ls=this->Ls;
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites4d = nsimd * grid->oSites() / Ls;
typedef typename SiteSpinor::scalar_object ScalarSiteSpinor;
MooeeInvCalls++;
MooeeInvTime-=usecond();
thread_loop((int ss=0;ss<grid->oSites();ss+=Ls),{ // adds Ls
auto tmp = psi[0];
// flops = 12*2*Ls + 12*2*Ls + 3*12*Ls + 12*2*Ls = 12*Ls * (9) = 108*Ls flops
// Apply (L^{\prime})^{-1}
chi[ss]=psi[ss]; // chi[0]=psi[0]
for(int s=1;s<Ls;s++){
spProj5p(tmp,chi[ss+s-1]);
chi[ss+s] = psi[ss+s]-lee[s-1]*tmp;
accelerator_loopN( sss, sites4d ,{
uint64_t lane = sss % nsimd;
uint64_t ss = Ls * (sss / nsimd);
ScalarSiteSpinor res, tmp, acc;
// X = Nc*Ns
// flops = 2X + (Ls-2)(4X + 4X) + 6X + 1 + 2X + (Ls-1)(10X + 1) = -16X + Ls(1+18X) = -192 + 217*Ls flops
// Apply (L^{\prime})^{-1} L_m^{-1}
res = extractLane(lane,psi[ss]);
spProj5m(tmp,res);
acc = leem_v[0]*tmp;
spProj5p(tmp,res);
insertLane(lane,chi[ss],res);
for(int s=1;s<Ls-1;s++){
res = extractLane(lane,psi[ss+s]);
res -= lee_v[s-1]*tmp;
spProj5m(tmp,res);
acc += leem_v[s]*tmp;
spProj5p(tmp,res);
insertLane(lane,chi[ss+s],res);
}
// L_m^{-1}
for (int s=0;s<Ls-1;s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
spProj5m(tmp,chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - leem[s]*tmp;
}
// U_m^{-1} D^{-1}
for (int s=0;s<Ls-1;s++){
// Chi[s] + 1/d chi[s]
spProj5p(tmp,chi[ss+Ls-1]);
chi[ss+s] = (1.0/dee[s])*chi[ss+s]-(ueem[s]/dee[Ls-1])*tmp;
}
chi[ss+Ls-1]= (1.0/dee[Ls-1])*chi[ss+Ls-1];
// Apply U^{-1}
res = extractLane(lane,psi[ss+Ls-1]);
res = res - lee_v[Ls-2]*tmp - acc;
// Apply U_m^{-1} D^{-1} U^{-1}
res = (1.0/dee_v[Ls-1])*res;
insertLane(lane,chi[ss+Ls-1],res);
spProj5p(acc,res);
spProj5m(tmp,res);
for (int s=Ls-2;s>=0;s--){
spProj5m(tmp,chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - uee[s]*tmp;
res = extractLane(lane,chi[ss+s]);
res = (1.0/dee_v[s])*res - uee_v[s]*tmp - ueem_v[s]*acc;
spProj5m(tmp,res);
insertLane(lane,chi[ss+s],res);
}
});
MooeeInvTime+=usecond();
}