Fix alignment on StencilEntry struct for HIP

CUDA neatly aligns the StencilEntry struct to 16-bytes, which is not used on HIP.
Mapping the StencilEntry struct to 16-bytes on HIP yields around 5% better performance
on MI355X-class GPUs.
This commit is contained in:
Julio Maia
2026-09-11 10:56:02 -05:00
parent 75bb6bdb11
commit 487ec8d410
+1 -1
View File
@@ -82,7 +82,7 @@ void DslashLogPartial(void);
void DslashLogDirichlet(void);
struct StencilEntry {
#ifdef GRID_CUDA
#if defined(GRID_CUDA) || defined(GRID_HIP)
uint64_t _byte_offset; // 8 bytes
uint32_t _offset; // 4 bytes
#else