mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-03 13:11:25 +00:00
782f0a9eed
[CK_TILE][FMHA] Extend pipelines with pssk for gfx11/12 (#4957) ## Motivation Build pipelines with seqlen padding only to support vectorized loads in the hdim dimension. The existing pipelines have either all dims padded or all dims not padded. These pipelines can be used in ComfyUI for slightly better performance. ## Technical Details Also a fix included for correct FLOPS calculation in `tile_example_fmha_fwd` when `seqlen_q * seqlen_k` overflows index_t capacity (signed int32). ## Test Plan The existing test cases will use the new pipelines when parameters allow (seqlens - padded, hdims - not padded): ``` ninja test_ck_tile_fmha_fwd bin/test_ck_tile_fmha_fwd_fp16 bin/test_ck_tile_fmha_fwd_bf16 bin/test_ck_tile_fmha_fwd_fp8bf16 # for gfx12 ``` ## Test Result All tests must pass. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.