test: try M=64 in descriptor + idesc to debug 4x factor
This commit is contained in:
@@ -54,9 +54,9 @@ test_umma_qk_hd16(
|
||||
// Descriptors
|
||||
uint32_t sQ_smem = __cvta_generic_to_shared(sQ);
|
||||
uint32_t sK_smem = __cvta_generic_to_shared(sK);
|
||||
uint64_t desc_q = make_umma_desc_kmajor_none(sQ_smem, 128);
|
||||
uint64_t desc_k = make_umma_desc_kmajor_none(sK_smem, 128);
|
||||
uint32_t idesc = 0; // Try idesc=0 (same as gau-nernst)
|
||||
uint64_t desc_q = make_umma_desc_kmajor_none(sQ_smem, 64); // Try M=64
|
||||
uint64_t desc_k = make_umma_desc_kmajor_none(sK_smem, 64); // Try M=64
|
||||
uint32_t idesc = make_idesc(64, 128); // M=64, N=128
|
||||
|
||||
// Verify SMEM Q and K by reading back row 0
|
||||
if (tid == 0) {
|
||||
|
||||
Reference in New Issue
Block a user