portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-09 21:50:45 +01:00
Code Releases Activity

Slightly faster Chebyshev; ifdef'ed out the fastest until tested numerics

Browse Source 
Lifteed from HDCR setup
This commit is contained in:
Peter Boyle 2020-05-08 09:20:54 -04:00
parent 93920c4811
commit 1d65e2f62c
1 changed files with 20 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
Grid/algorithms/approx/Chebyshev.h
View File

@ -234,10 +234,9 @@ public:
GridBase *grid=in.Grid(); GridBase *grid=in.Grid();
// std::cout << "Chevyshef(): in.Grid()="<<in.Grid()<<std::endl;
//std::cout <<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
int vol=grid->gSites(); int vol=grid->gSites();
typedef typename Field::vector_type vector_type;
constexpr int Nsimd = vector_type::Nsimd();
Field T0(grid); T0 = in; Field T0(grid); T0 = in;
Field T1(grid); Field T1(grid);
@ -258,14 +257,27 @@ public:
// out = ()*T0 + Coeffs[1]*T1; // out = ()*T0 + Coeffs[1]*T1;
axpby(out,0.5*Coeffs[0],Coeffs[1],T0,T1); axpby(out,0.5*Coeffs[0],Coeffs[1],T0,T1);
for(int n=2;n<order;n++){ for(int n=2;n<order;n++){
Linop.HermOp(*Tn,y); Linop.HermOp(*Tn,y);
// y=xscale*y+mscale*(*Tn); #if 0
// *Tnp=2.0*y-(*Tnm); auto y_v = y.View();
// out=out+Coeffs[n]* (*Tnp); auto Tn_v = Tn->View();
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#else
axpby(y,xscale,mscale,y,(*Tn)); axpby(y,xscale,mscale,y,(*Tn));
axpby(*Tnp,2.0,-1.0,y,(*Tnm)); axpby(*Tnp,2.0,-1.0,y,(*Tnm));
axpy(out,Coeffs[n],*Tnp,out); if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#endif
// Cycle pointers to avoid copies // Cycle pointers to avoid copies
Field *swizzle = Tnm; Field *swizzle = Tnm;
Tnm =Tn; Tnm =Tn;