portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-14 01:35:36 +00:00
Code Releases Activity

access M5D coeffs through pointers

Browse Source
This commit is contained in:
gfilaci 2019-09-02 11:33:00 +01:00
parent 3ef519aaa4
commit 0efaf3c4fa
3 changed files with 44 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h
View File

@ -55,6 +55,10 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
auto chi = chi_i.View(); auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
int Ls =this->Ls; int Ls =this->Ls;
// 10 = 3 complex mult + 2 complex add // 10 = 3 complex mult + 2 complex add
@ -72,7 +76,7 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
uint64_t idx_l = ss+((s+Ls-1)%Ls); uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5m(tmp1,psi(idx_u)); spProj5m(tmp1,psi(idx_u));
spProj5p(tmp2,psi(idx_l)); spProj5p(tmp2,psi(idx_l));
coalescedWrite(chi[ss+s],diag[s]*phi(ss+s)+upper[s]*tmp1+lower[s]*tmp2); coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
} }
}); });
M5Dtime+=usecond(); M5Dtime+=usecond();
@ -94,6 +98,10 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
auto chi = chi_i.View(); auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
int Ls=this->Ls; int Ls=this->Ls;
// Flops = 6.0*(Nc*Ns) *Ls*vol // Flops = 6.0*(Nc*Ns) *Ls*vol
@ -110,7 +118,7 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
uint64_t idx_l = ss+((s+Ls-1)%Ls); uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5p(tmp1,psi(idx_u)); spProj5p(tmp1,psi(idx_u));
spProj5m(tmp2,psi(idx_l)); spProj5m(tmp2,psi(idx_l));
coalescedWrite(chi[ss+s],diag[s]*phi(ss+s)+upper[s]*tmp1+lower[s]*tmp2); coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
} }
}); });
M5Dtime+=usecond(); M5Dtime+=usecond();

11
Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h
View File

@ -11,6 +11,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local> Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk> Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu> Author: David Murphy <dmurphy@phys.columbia.edu>
Author: Gianluca Filaci <g.filaci@ed.ac.uk>
This program is free software; you can redistribute it and/or modify This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by it under the terms of the GNU General Public License as published by
@ -49,6 +50,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
auto psi = psi_i.View(); auto psi = psi_i.View();
auto chi = chi_i.View(); auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol // Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++; this->M5Dcalls++;
this->M5Dtime -= usecond(); this->M5Dtime -= usecond();
@ -63,7 +67,7 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
uint64_t idx_l = ss+((s+Ls-1)%Ls); uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5m(tmp1, psi(idx_u)); spProj5m(tmp1, psi(idx_u));
spProj5p(tmp2, psi(idx_l)); spProj5p(tmp2, psi(idx_l));
coalescedWrite(chi[ss+s], diag[s]*phi(ss+s) + upper[s]*tmp1 + lower[s]*tmp2); coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
} }
}); });
@ -82,6 +86,9 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
auto phi = phi_i.View(); auto phi = phi_i.View();
auto chi = chi_i.View(); auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol // Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++; this->M5Dcalls++;
@ -97,7 +104,7 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
uint64_t idx_l = ss+((s+Ls-1)%Ls); uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5p(tmp1, psi(idx_u)); spProj5p(tmp1, psi(idx_u));
spProj5m(tmp2, psi(idx_l)); spProj5m(tmp2, psi(idx_l));
coalescedWrite(chi[ss+s], diag[s]*phi(ss+s) + upper[s]*tmp1 + lower[s]*tmp2); coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
} }
}); });

31
Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h
View File

@ -11,6 +11,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local> Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk> Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu> Author: David Murphy <dmurphy@phys.columbia.edu>
Author: Gianluca Filaci <g.filaci@ed.ac.uk>
This program is free software; you can redistribute it and/or modify This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by it under the terms of the GNU General Public License as published by
@ -49,6 +50,10 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol // Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++; this->M5Dcalls++;
this->M5Dtime -= usecond(); this->M5Dtime -= usecond();
@ -64,7 +69,7 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
uint64_t idx_l = ss+((s+Ls-1)%Ls); uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5m(tmp1, psi(idx_u)); spProj5m(tmp1, psi(idx_u));
spProj5p(tmp2, psi(idx_l)); spProj5p(tmp2, psi(idx_l));
coalescedWrite(chi[ss+s], diag[s]*phi(ss+s) + upper[s]*tmp1 + lower[s]*tmp2); coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
} }
}); });
@ -88,6 +93,11 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
auto pshift_coeffs = &shift_coeffs[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol // Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++; this->M5Dcalls++;
this->M5Dtime -= usecond(); this->M5Dtime -= usecond();
@ -108,7 +118,7 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
if(pm == 1){ spProj5p(tmp, psi(ss+shift_s)); } if(pm == 1){ spProj5p(tmp, psi(ss+shift_s)); }
else { spProj5m(tmp, psi(ss+shift_s)); } else { spProj5m(tmp, psi(ss+shift_s)); }
coalescedWrite(chi[ss+s], diag[s]*phi(ss+s) + upper[s]*tmp1 +lower[s]*tmp2 + shift_coeffs[s]*tmp); coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 +plower[s]*tmp2 + pshift_coeffs[s]*tmp);
} }
}); });
@ -128,6 +138,10 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol // Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++; this->M5Dcalls++;
this->M5Dtime -= usecond(); this->M5Dtime -= usecond();
@ -144,7 +158,7 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
uint64_t idx_l = ss+((s+Ls-1)%Ls); uint64_t idx_l = ss+((s+Ls-1)%Ls);
spProj5p(tmp1, psi(idx_u)); spProj5p(tmp1, psi(idx_u));
spProj5m(tmp2, psi(idx_l)); spProj5m(tmp2, psi(idx_l));
coalescedWrite(chi[ss+s], diag[s]*phi(ss+s) + upper[s]*tmp1 + lower[s]*tmp2); coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
} }
}); });
@ -166,6 +180,11 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
assert(phi.Checkerboard() == psi.Checkerboard()); assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
auto plower = &lower[0];
auto pshift_coeffs = &shift_coeffs[0];
// Flops = 6.0*(Nc*Ns) *Ls*vol // Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++; this->M5Dcalls++;
this->M5Dtime -= usecond(); this->M5Dtime -= usecond();
@ -189,12 +208,12 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
spProj5p(tmp1, psi(idx_u)); spProj5p(tmp1, psi(idx_u));
spProj5m(tmp2, psi(idx_l)); spProj5m(tmp2, psi(idx_l));
if(s==(Ls-1)) coalescedWrite(chi[ss+s], chi(ss+s)+ diag[s]*phi(ss+s) + upper[s]*tmp1 + lower[s]*tmp2); if(s==(Ls-1)) coalescedWrite(chi[ss+s], chi(ss+s)+ pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
else coalescedWrite(chi[ss+s], diag[s]*phi(ss+s) + upper[s]*tmp1 + lower[s]*tmp2); else coalescedWrite(chi[ss+s], pdiag[s]*phi(ss+s) + pupper[s]*tmp1 + plower[s]*tmp2);
if(pm == 1){ spProj5p(tmp, psi(ss+s)); } if(pm == 1){ spProj5p(tmp, psi(ss+s)); }
else { spProj5m(tmp, psi(ss+s)); } else { spProj5m(tmp, psi(ss+s)); }
coalescedWrite(chi[ss+shift_s],chi(ss+shift_s)+shift_coeffs[s]*tmp); coalescedWrite(chi[ss+shift_s],chi(ss+shift_s)+pshift_coeffs[s]*tmp);
} }
}); });