portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Global changes to parallel_for structure.

Browse Source 
Move the comms flags to more sensible names
This commit is contained in:
paboyle
2017-02-21 05:24:27 -05:00
parent 3906cd2149
commit 3ae92fa2e6
43 changed files with 271 additions and 513 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
lib/qcd/utils/LinalgUtils.h
View File

@ -48,8 +48,7 @@ void axpibg5x(Lattice<vobj> &z,const Lattice<vobj> &x,Coeff a,Coeff b)
GridBase *grid=x._grid;
Gamma G5(Gamma::Gamma5);
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss++){
parallel_for(int ss=0;ss<grid->oSites();ss++){
vobj tmp;
tmp = a*x._odata[ss];
tmp = tmp + G5*(b*timesI(x._odata[ss]));
@ -65,8 +64,7 @@ void axpby_ssp(Lattice<vobj> &z, Coeff a,const Lattice<vobj> &x,Coeff b,const La
conformable(x,z);
GridBase *grid=x._grid;
int Ls = grid->_rdimensions[0];
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
parallel_for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
vobj tmp = a*x._odata[ss+s]+b*y._odata[ss+sp];
vstream(z._odata[ss+s],tmp);
}
@ -81,8 +79,7 @@ void ag5xpby_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const L
GridBase *grid=x._grid;
int Ls = grid->_rdimensions[0];
Gamma G5(Gamma::Gamma5);
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
parallel_for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
vobj tmp;
tmp = G5*x._odata[ss+s]*a;
tmp = tmp + b*y._odata[ss+sp];
@ -99,8 +96,7 @@ void axpbg5y_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const L
GridBase *grid=x._grid;
int Ls = grid->_rdimensions[0];
Gamma G5(Gamma::Gamma5);
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
parallel_for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
vobj tmp;
tmp = G5*y._odata[ss+sp]*b;
tmp = tmp + a*x._odata[ss+s];
@ -117,8 +113,7 @@ void ag5xpbg5y_ssp(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,const
GridBase *grid=x._grid;
int Ls = grid->_rdimensions[0];
Gamma G5(Gamma::Gamma5);
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
parallel_for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
vobj tmp1;
vobj tmp2;
tmp1 = a*x._odata[ss+s]+b*y._odata[ss+sp];
@ -135,8 +130,7 @@ void axpby_ssp_pminus(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,co
conformable(x,z);
GridBase *grid=x._grid;
int Ls = grid->_rdimensions[0];
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
parallel_for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
vobj tmp;
spProj5m(tmp,y._odata[ss+sp]);
tmp = a*x._odata[ss+s]+b*tmp;
@ -152,8 +146,7 @@ void axpby_ssp_pplus(Lattice<vobj> &z,Coeff a,const Lattice<vobj> &x,Coeff b,con
conformable(x,z);
GridBase *grid=x._grid;
int Ls = grid->_rdimensions[0];
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
parallel_for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
vobj tmp;
spProj5p(tmp,y._odata[ss+sp]);
tmp = a*x._odata[ss+s]+b*tmp;
@ -169,8 +162,7 @@ void G5R5(Lattice<vobj> &z,const Lattice<vobj> &x)
conformable(x,z);
int Ls = grid->_rdimensions[0];
Gamma G5(Gamma::Gamma5);
PARALLEL_FOR_LOOP
for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
parallel_for(int ss=0;ss<grid->oSites();ss+=Ls){ // adds Ls
vobj tmp;
for(int s=0;s<Ls;s++){
int sp = Ls-1-s;

6
lib/qcd/utils/SUn.h
View File

@ -221,8 +221,7 @@ class SU {
int i0, i1;
su2SubGroupIndex(i0, i1, su2_index);
PARALLEL_FOR_LOOP
for (int ss = 0; ss < grid->oSites(); ss++) {
parallel_for (int ss = 0; ss < grid->oSites(); ss++) {
subgroup._odata[ss]()()(0, 0) = source._odata[ss]()()(i0, i0);
subgroup._odata[ss]()()(0, 1) = source._odata[ss]()()(i0, i1);
subgroup._odata[ss]()()(1, 0) = source._odata[ss]()()(i1, i0);
@ -252,8 +251,7 @@ class SU {
su2SubGroupIndex(i0, i1, su2_index);
dest = 1.0; // start out with identity
PARALLEL_FOR_LOOP
for (int ss = 0; ss < grid->oSites(); ss++) {
parallel_for (int ss = 0; ss < grid->oSites(); ss++) {
dest._odata[ss]()()(i0, i0) = subgroup._odata[ss]()()(0, 0);
dest._odata[ss]()()(i0, i1) = subgroup._odata[ss]()()(0, 1);
dest._odata[ss]()()(i1, i0) = subgroup._odata[ss]()()(1, 0);