Debug: single PV K-tile
This commit is contained in:
@@ -134,7 +134,7 @@ test_fmha_ts_full(const bf16_t* __restrict__ q, const bf16_t* __restrict__ k,
|
||||
{
|
||||
uint32_t idesc_pv = make_idesc(BLOCK_MN, HD);
|
||||
|
||||
for (int kt = 0; kt < NKT_PV; kt++) {
|
||||
for (int kt = 0; kt < 1; kt++) { // DEBUG: single PV K-tile
|
||||
uint32_t tmem_a = tb + kt * MMA_K_BF16; // A from P's kt-th 16 columns
|
||||
bf16_t* sv = sV + kt * 256;
|
||||
uint64_t dv = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sv), 16);
|
||||
|
||||
Reference in New Issue
Block a user