Test 1 PV K-tile from (128,128) P at offset 0
This commit is contained in:
@@ -72,7 +72,7 @@ test_pv_ss_128()
|
||||
{
|
||||
uint64_t dv = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sV), 16);
|
||||
uint32_t idesc = make_idesc(BLOCK_MN, HD);
|
||||
for (int kt = 0; kt < 8; kt++) {
|
||||
for (int kt = 0; kt < 1; kt++) {
|
||||
bf16_t* sp = sP + kt * 2048;
|
||||
uint64_t dp = make_umma_desc_kmajor_none(__cvta_generic_to_shared(sp), BLOCK_MN);
|
||||
if (tid == 0) umma_ss_f16(tb, dp, dv, idesc, true); // ALL accumulate
|
||||
@@ -97,7 +97,7 @@ test_pv_ss_128()
|
||||
printf("O[0,0..15] (raw): ");
|
||||
for (int d=0;d<HD;d++) printf("%.2f ", o_vals[d]);
|
||||
// Expected: sum(P[0,0..15] * V[j,0..15]) = 16 * 0.5 * 1.0 = 8.0 (raw, no MMA scale)
|
||||
printf("\nExpected ~64.0 (8 K-tiles × 16 values × 0.5 × 1.0, MMA scale 1.0)\n");
|
||||
printf("\nExpected ~8.0 (1 K-tile, MMA scale 1.0)\n");
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user