wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

flash: Disable perf loop multiplier

Browse Source
This commit is contained in:
Hansung Kim
2024-11-10 22:44:02 -08:00
parent 4448f31fdc
commit 7d7cb5f60a
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/regression/flash_attention/kernel.cpp
View File

@@ -355,7 +355,7 @@ void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
// "inner loop" along the columns of K^T
const uint32_t k_tiles = (dim_seqlen / B_COL);
for (uint32_t tile_k = 0; tile_k < (4 /* for perf measurement */ * k_tiles);
for (uint32_t tile_k = 0; tile_k < (1 /* for perf measurement */ * k_tiles);
tile_k++) {
// float *smem_P_produce = (tile_k % 2) ? smem_P0 : smem_P1;
// float *smem_P_consume = (tile_k % 2) ? smem_P1 : smem_P0;

2
tests/regression/flash_attention/kernel.gemmini.cpp
View File

@@ -347,7 +347,7 @@ void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
// "inner loop" along the columns of K^T
const uint32_t k_tiles = (dim_seqlen / B_COL);
for (uint32_t tile_k = 0;
tile_k < (4 /*for perf measurement*/ *
tile_k < (1 /*for perf measurement*/ *
// virgo kernel is fully pipelined around (2 GEMMs | softmax);
// requires two loop iterations to finish one tile compute
(2 * k_tiles)) +