mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-10 16:28:38 +00:00
95c916369c
[CK_TILE] Fix Stream-K k_size calculation ## Motivation In a recent benchmarking task for CK Tile Stream-K algorithm, we identified that certain instances segfault. This change works to fix the bug and adds necessary regression tests. ## Technical Details The StreamK kernel constructs tensor views using a `k_size` parameter that determines how much of the K dimension to process in each iteration. Previously, this was calculated as: ```cpp index_t k_size = num_loop_sk * TilePartitioner::KPerBlock; ``` This calculation assumes all macro tiles along K are exactly `KPerBlock` in size. However, when `K % KPerBlock != 0`, the final macro tile along K has a remainder size of `K % KPerBlock`, not a full `KPerBlock` (see the figure below): <img width="961" height="488" alt="image" src="https://github.com/user-attachments/assets/3e1cceed-5dcd-4980-8b02-cee24eecf262" /> With the old code, a workgroup working with the `MPerBlock x (K % KPerBlock)` tile in A and B risk accessing illegal memory. Hence, this change ensures that when `K % KPerBlock != 0`, workgroups processing iterations that include the final macro-tile along K calculate the correct `k_size` based on the remainder rather than assuming a full `KPerBlock`. ## Test Plan I added the following tests: 1. Unit tests added for the Stream-K Tile Partitioner: - `StreamKTilePartitionerBaseGetKSize/NoRemainderTiles` - validates full tiles - `StreamKTilePartitionerBaseGetKSize/RemainderTiles` - validates remainder handling 2. Regression tests that test a case where `K % KPerBlock != 0` ## Test Result Tests passed locally on gfx90a, gfx942, and gfx950. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.