Updated to compile and run fast on CUDA

This commit is contained in:
2025-08-10 00:00:13 +01:00
parent 82cfff2990
commit 5adf2657dd
8 changed files with 60 additions and 389 deletions
@@ -222,7 +222,7 @@ public:
#if 0
static accelerator_inline typename SiteCloverTriangle::vector_type triangle_elem(const SiteCloverTriangle& triangle, int block, int i, int j) {
GRID_ASSERT(i != j);
assert(i != j);
if(i < j) {
return triangle()(block)(triangle_index(i, j));
} else { // i > j
@@ -232,7 +232,7 @@ public:
#else
template<typename vobj>
static accelerator_inline vobj triangle_elem(const iImplCloverTriangle<vobj>& triangle, int block, int i, int j) {
GRID_ASSERT(i != j);
assert(i != j);
if(i < j) {
return triangle()(block)(triangle_index(i, j));
} else { // i > j