1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-08-17 11:41:53 +01:00

Make view specify where and drive data motion - first cut.

This is a compile tiime option --enable-unified=yes/no
This commit is contained in:
Peter Boyle
2020-05-21 16:13:16 -04:00
parent ebb60330c9
commit 7860a50f70
48 changed files with 688 additions and 718 deletions

View File

@@ -35,8 +35,8 @@ NAMESPACE_BEGIN(Grid);
template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
auto rhs = rhs_i.View(AcceleratorRead);
auto ret = ret_i.View(AcceleratorWrite);
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),1,{
ret[ss]=pow(rhs[ss],y);
@@ -45,8 +45,8 @@ template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
}
template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
auto rhs = rhs_i.View(AcceleratorRead);
auto ret = ret_i.View(AcceleratorWrite);
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],mod(rhs(ss),y));
@@ -56,8 +56,8 @@ template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
Lattice<obj> ret_i(rhs_i.Grid());
auto ret = ret_i.View();
auto rhs = rhs_i.View();
auto ret = ret_i.View(AcceleratorWrite);
auto rhs = rhs_i.View(AcceleratorRead);
ret.Checkerboard() = rhs_i.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],div(rhs(ss),y));
@@ -67,8 +67,8 @@ template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs_i, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
auto rhs = rhs_i.View(AcceleratorRead);
auto ret = ret_i.View(AcceleratorWrite);
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));