1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-24 18:52:02 +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

@ -43,8 +43,8 @@ template<class vobj>
inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
auto rhs_v = rhs.View();
auto ret_v = ret.View();
auto rhs_v = rhs.View(AcceleratorRead);
auto ret_v = ret.View(AcceleratorWrite);
accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
});
@ -56,9 +56,9 @@ template<class vobj>
inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
auto lhs_v = lhs.View(AcceleratorRead);
auto rhs_v = rhs.View(AcceleratorRead);
auto ret_v = ret.View(AcceleratorWrite);
accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
});
@ -73,9 +73,9 @@ inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Latt
typedef decltype(coalescedRead(ll())) sll;
typedef decltype(coalescedRead(rr())) srr;
Lattice<decltype(outerProduct(ll(),rr()))> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
auto lhs_v = lhs.View(AcceleratorRead);
auto rhs_v = rhs.View(AcceleratorRead);
auto ret_v = ret.View(AcceleratorWrite);
accelerator_for(ss,rhs_v.size(),1,{
// FIXME had issues with scalar version of outer
// Use vector [] operator and don't read coalesce this loop