reduce mfma to 3

2026-07-17 09:08:35 +00:00 · 2025-02-26 12:08:01 +00:00
parent 2b55942a0a
commit d97249ba22
1 changed files with 1 additions and 1 deletions
--- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp
@@ -193,7 +193,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
        static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
            ignore = i;
            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
-            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 3, 0); // MFMA
        });

        // A global