test: single-thread MMA (tid==0) for Layout D

2026-05-28 12:56:39 +00:00
parent ba2e390e1e
commit f244c4fdd2
1 changed files with 2 additions and 2 deletions
--- a/tests/unit/test_umma_qk_hd64.cu
+++ b/tests/unit/test_umma_qk_hd64.cu
@@ -107,8 +107,8 @@ test_umma_hd64(const bf16_t* __restrict__ q, const bf16_t* __restrict__ k,
        uint64_t dq = make_umma_desc_kmajor_none(q_addr, BLOCK_MN);
        uint64_t dk = make_umma_desc_kmajor_none(k_addr, BLOCK_MN);

-        // 4 warp leaders call MMA (same as working HD=16 test)
-        if (lane == 0) {
+        // Single thread calls MMA (gau-nernst's elect_one pattern)
+        if (tid == 0) {
            umma_ss_f16(tb, dq, dk, idesc, kt > 0);
        }
        asm volatile("tcgen05.fence::after_thread_sync;" ::: "memory");