diff --git a/tests/unit/test_tmem_cols.cu b/tests/unit/test_tmem_cols.cu index a63c7f7c..a3dcaba5 100644 --- a/tests/unit/test_tmem_cols.cu +++ b/tests/unit/test_tmem_cols.cu @@ -48,18 +48,7 @@ __global__ void test_tmem_2col(float* out) { memcpy(&u0, &v0, 4); memcpy(&u1, &v1, 4); memcpy(&u2, &v2, 4); memcpy(&u3, &v3, 4); tmem_store(tb + 0, u0, u1, u2, u3); - } - tmem_fence(); - __syncthreads(); - { - float v0 = (float)(lane * 4 + 100); - float v1 = (float)(lane * 4 + 101); - float v2 = (float)(lane * 4 + 102); - float v3 = (float)(lane * 4 + 103); - uint32_t u0, u1, u2, u3; - memcpy(&u0, &v0, 4); memcpy(&u1, &v1, 4); - memcpy(&u2, &v2, 4); memcpy(&u3, &v3, 4); - tmem_store(tb + 1, u0, u1, u2, u3); + tmem_store(tb + 0, u1, u2, u3, u0); // 2nd store to SAME column } tmem_fence(); __syncthreads();