mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 02:02:46 +00:00
gfx12 falls back from async global-to-LDS loads to sync VGPR loads plus LDS stores. The async raw API relies on buffer OOB behavior instead of tensor-coordinate validity, so keep the sync fallback aligned with that raw-load contract.