test: tmem_load column 1 only

This commit is contained in:
2026-05-28 09:45:34 +00:00
parent 6c38c6e442
commit ee3fe6d6b2

View File

@@ -90,28 +90,11 @@ test_umma_qk_hd16(
// Use separate tmem_load calls with __syncthreads between them
// to avoid the loop crash
if (wid == 0) {
#define READ_COL(c) do { \
uint32_t u0, u1, u2, u3; \
tmem_load(tmem_base + c, u0, u1, u2, u3); \
if (lane == 0) s_out[c] = u32_to_f32(u0); \
} while(0)
READ_COL(0);
uint32_t u0, u1, u2, u3;
tmem_load(tmem_base + 1, u0, u1, u2, u3);
if (lane == 0) s_out[0] = u32_to_f32(u0);
}
__syncthreads();
if (wid == 0) { READ_COL(1); }
__syncthreads();
if (wid == 0) { READ_COL(2); }
__syncthreads();
if (wid == 0) { READ_COL(3); }
__syncthreads();
if (wid == 0) { READ_COL(4); }
__syncthreads();
if (wid == 0) { READ_COL(5); }
__syncthreads();
if (wid == 0) { READ_COL(6); }
__syncthreads();
if (wid == 0) { READ_COL(7); }
__syncthreads();
// Scalar reference
if (tid == 0) {