mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-11 00:39:02 +00:00
919096fde8
[CK] Allow skipping split-K C-buffer zero-init in xdl_cshuffle blockscale GEMM (#7935) Add a `skip_zero_init` flag (default false) to the Problem/Argument of the xdl_cshuffle block-scale GEMM device ops (multiple_d ab_scale and blockscale b-preshuffle). When the flag is set, the device invoker skips the internal hipMemsetAsync that zeroes p_c_grid before the KBatch > 1 split-K atomic-accumulation path. The flag is declared on the gridwise Problem struct (inherited by Argument), so it is visible on both the rotating-cache (arg_) and the normal (arg) launch paths in each device op. Why: callers that already pre-zero the output buffer otherwise pay for a redundant device-wide memset before split-K atomic accumulation. Gating the memset behind an opt-in flag lets such callers avoid the duplicate work. Because the flag defaults to false, every existing call site is unaffected and the observable behavior is unchanged. ## Motivation <!-- Explain the purpose of this PR and the goals it aims to achieve. --> ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. Co-authored-by: Cursor <cursoragent@cursor.com>