aboutsummaryrefslogtreecommitdiff
path: root/bench/bench_kernels.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-27 00:58:05 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-27 01:03:41 +0200
commitaf55803d0dff9eb692c760c8d486e736c945061d (patch)
tree0875a1112bf98304e445e0544c4309670d6f4769 /bench/bench_kernels.c
parent0f861b7f3098750ddee4f6570f3be228d1da5320 (diff)
downloadffwd-af55803d0dff9eb692c760c8d486e736c945061d.tar.gz
ffwd-af55803d0dff9eb692c760c8d486e736c945061d.zip
chore: clean up comment formatting and wording
Diffstat (limited to 'bench/bench_kernels.c')
-rw-r--r--bench/bench_kernels.c14
1 files changed, 8 insertions, 6 deletions
diff --git a/bench/bench_kernels.c b/bench/bench_kernels.c
index 55f859b..d302396 100644
--- a/bench/bench_kernels.c
+++ b/bench/bench_kernels.c
@@ -1,10 +1,12 @@
-/* bench/bench_kernels.c - regression microbenchmarks for the hot kernels.
+/* Regression microbenchmarks for the hot kernels.
*
* Shapes are the real pplx-embed-v1-0.6b dimensions (hidden 1024, 16 query /
* 8 KV heads x head_dim 128, intermediate 3072) so a slowdown here predicts
- * a slowdown in inference. Single-threaded by default for stable numbers;
- * --threads N opts into the pool. Build and record via `make bench`,
- * compare records with bench/benchstat.py.
+ * a slowdown in inference.
+ *
+ * Single-threaded by default for stable numbers; --threads N opts into the pool.
+ *
+ * Build and record via make bench, compare records with bench/benchstat.py.
*/
#include "bench.h"
@@ -133,7 +135,7 @@ static void bm_bf16_widen(bench_state_t *b) {
free(src);
}
-/* ---- attention pieces ---- */
+/* attention pieces */
static void bm_attn_scores(bench_state_t *b) {
/* one head: scores[seq,seq] = Q[seq,hd] @ K[seq,hd]^T */
@@ -171,7 +173,7 @@ static void bm_rope(bench_state_t *b) {
free(x);
}
-/* ---- norms, activations, reductions ---- */
+/* norms, activations, reductions */
static void bm_rms_norm(bench_state_t *b) {
float *x = alloc_f32((size_t)SEQ * HIDDEN, 23);