test: skip MMA, just test descriptor values
This commit is contained in:
@@ -106,22 +106,22 @@ test_umma_qk_hd16(
|
||||
s_out[133] = (float)sQ_smem;
|
||||
s_out[134] = (float)sK_smem;
|
||||
s_out[135] = (float)__cvta_generic_to_shared(sQ);
|
||||
s_out[136] = (float)(sQ_smem >> 4); // start_address for desc
|
||||
s_out[137] = (float)(sQ_smem & 0xF); // alignment check
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
// ================================================================
|
||||
// Call tcgen05.mma SS
|
||||
// SKIP MMA for now — just test descriptor construction
|
||||
// ================================================================
|
||||
if (tid == 0) {
|
||||
umma_ss_f16(tmem_base, desc_q, desc_k, idesc, /*accumulate=*/false);
|
||||
}
|
||||
__syncwarp();
|
||||
|
||||
// Wait for MMA
|
||||
if (wid == 0 && lane == 0) {
|
||||
tmem_fence_store();
|
||||
}
|
||||
__syncthreads();
|
||||
// if (tid == 0) {
|
||||
// umma_ss_f16(tmem_base, desc_q, desc_k, idesc, /*accumulate=*/false);
|
||||
// }
|
||||
// __syncwarp();
|
||||
// if (wid == 0 && lane == 0) {
|
||||
// tmem_fence_store();
|
||||
// }
|
||||
// __syncthreads();
|
||||
|
||||
// ================================================================
|
||||
// Read S from TMEM
|
||||
|
||||
Reference in New Issue
Block a user