mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-28 18:56:59 +00:00
Add cluster launch in test ck_tile mx gemm tdm wmma ## Motivation Add cluster launch test in test_ck_tile_mx_gemm_pipeline_tdm_wmma on gfx1250, so that we can check the performance on gfx1250 hardware. ## Technical Details Added Out-of-bounds guard in RunGemm of MxGemmKernel to skip blocks padded by cluster alignment. Add ClusterEnable/ClusterDisable aliases and extend the tuple in test_mx_gemm_pipeline_kernel_types.hpp by adding two kernel types with ClusterEnable for F8 CompTDMV1 and CompTDMV2 respectively. The existing F4 non-ClusterLaunch kernel types have issue to be fixed, so this PR does not include F4 cases. Read ClusterLaunch from the tuple in test_mx_gemm_pipeline_util.hpp. Update invoke_mx_gemm to branch on ClusterLaunch, including Add cluster size constants, Switch GemmShape type, TilePartitioner type, and the kernel launch call. ## Test Plan Tested the changes on gfx1250 FFM. ## Test Result The added kernel types (instances) passed the tests on gfx1250 FFM. ## Submission Checklist - [x ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.