add the sync barrier for persistent kernel (#2977)

2026-04-19 14:29:05 +00:00 · 2025-10-07 11:54:04 -07:00
parent 19415d0b6f
commit ae9f29b7d5
1 changed files with 1 additions and 0 deletions
--- a/include/ck_tile/ops/gemm/kernel/universal_gemm_kernel.hpp
+++ b/include/ck_tile/ops/gemm/kernel/universal_gemm_kernel.hpp
@@ -1134,6 +1134,7 @@ struct UniversalGemmKernel

        while(block_id < num_work)
        {
+            s_waitcnt_barrier();
            // Get the tile index for this block
            const auto tile_idx = amd_wave_read_first_lane(block_id % num_tiles);
            const auto [iM, iN] = TilePartitioner{kargs.M, kargs.N}.GetOutputTileIndex(tile_idx);