aboutsummaryrefslogtreecommitdiff
path: root/bench/bench_kernels.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-20 20:45:39 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-21 06:00:26 +0200
commitbc8bebe78358e3a37bad0867b2d94955cac01d47 (patch)
treed5a49a34b95b45cec0917a2a07707159825fb912 /bench/bench_kernels.c
parent6652fcbc653d84e75cb5c08c5f5edf52263ed50f (diff)
downloadffwd-bc8bebe78358e3a37bad0867b2d94955cac01d47.tar.gz
ffwd-bc8bebe78358e3a37bad0867b2d94955cac01d47.zip
bench: add e2e benchmarks + new API compat fixes
Diffstat (limited to 'bench/bench_kernels.c')
-rw-r--r--bench/bench_kernels.c27
1 files changed, 13 insertions, 14 deletions
diff --git a/bench/bench_kernels.c b/bench/bench_kernels.c
index 2b2fe47..55f859b 100644
--- a/bench/bench_kernels.c
+++ b/bench/bench_kernels.c
@@ -4,16 +4,15 @@
* 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 tools/benchstat.py.
+ * compare records with bench/benchstat.py.
*/
#include "bench.h"
#include "kernels.h"
+#include "threadpool.h"
#include <stdint.h>
-int embed_verbose = 0; /* defined by embed.c in full builds */
-
enum {
HIDDEN = 1024,
Q_DIM = 2048, /* 16 heads * 128 */
@@ -71,7 +70,7 @@ static void bm_linear_f32(bench_state_t *b, int seq, int in, int out) {
float *y = alloc_f32((size_t)seq * out, 3);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_linear_nobias(y, x, w, seq, in, out);
+ linear_nobias(y, x, w, seq, in, out);
bench_sink += y[0];
free(y);
free(w);
@@ -91,7 +90,7 @@ static void bm_linear_bf16(bench_state_t *b, int seq, int in, int out) {
float *y = alloc_f32((size_t)seq * out, 6);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_linear_nobias_bf16(y, x, w, seq, in, out);
+ linear_nobias_bf16(y, x, w, seq, in, out);
bench_sink += y[0];
free(y);
free(w);
@@ -111,7 +110,7 @@ static void bm_qkv_bf16(bench_state_t *b) {
float *v = alloc_f32((size_t)SEQ * KV_DIM, 13);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_linear_nobias_bf16_qkv(q, k, v, x, wq, wk, wv, SEQ, HIDDEN, Q_DIM, KV_DIM);
+ linear_nobias_bf16_qkv(q, k, v, x, wq, wk, wv, SEQ, HIDDEN, Q_DIM, KV_DIM);
bench_sink += q[0] + k[0] + v[0];
free(v);
free(k);
@@ -128,7 +127,7 @@ static void bm_bf16_widen(bench_state_t *b) {
float *dst = alloc_f32(n, 15);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_bf16_to_f32_buf(dst, src, n);
+ bf16_to_f32_buf(dst, src, n);
bench_sink += dst[0];
free(dst);
free(src);
@@ -143,7 +142,7 @@ static void bm_attn_scores(bench_state_t *b) {
float *s = alloc_f32((size_t)ATTN_SEQ * ATTN_SEQ, 18);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_matmul_t(s, qm, km, ATTN_SEQ, HEAD_DIM, ATTN_SEQ);
+ matmul_t(s, qm, km, ATTN_SEQ, HEAD_DIM, ATTN_SEQ);
bench_sink += s[0];
free(s);
free(km);
@@ -154,7 +153,7 @@ static void bm_softmax(bench_state_t *b) {
float *x = alloc_f32((size_t)ATTN_SEQ * ATTN_SEQ, 19);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_softmax(x, ATTN_SEQ, ATTN_SEQ);
+ softmax(x, ATTN_SEQ, ATTN_SEQ);
bench_sink += x[0];
free(x);
}
@@ -165,7 +164,7 @@ static void bm_rope(bench_state_t *b) {
float *sinv = alloc_f32((size_t)ATTN_SEQ * HEAD_DIM, 22);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_apply_rope_neox(x, cosv, sinv, ATTN_SEQ, Q_DIM / HEAD_DIM, HEAD_DIM);
+ apply_rope_neox(x, cosv, sinv, ATTN_SEQ, Q_DIM / HEAD_DIM, HEAD_DIM);
bench_sink += x[0];
free(sinv);
free(cosv);
@@ -180,7 +179,7 @@ static void bm_rms_norm(bench_state_t *b) {
float *y = alloc_f32((size_t)SEQ * HIDDEN, 25);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- embed_rms_norm(y, x, w, SEQ, HIDDEN, 1e-6f);
+ rms_norm(y, x, w, SEQ, HIDDEN, 1e-6f);
bench_sink += y[0];
free(y);
free(w);
@@ -197,7 +196,7 @@ static void bm_silu_mul(bench_state_t *b) {
bench_begin(b);
for (long i = 0; i < b->n; i++) {
memcpy(gate, gate0, n * sizeof(float));
- embed_silu_mul_inplace(gate, up, (int)n);
+ silu_mul_inplace(gate, up, (int)n);
}
bench_sink += gate[0];
free(up);
@@ -210,7 +209,7 @@ static void bm_dot(bench_state_t *b) {
float *y = alloc_f32(HIDDEN, 29);
bench_begin(b);
for (long i = 0; i < b->n; i++)
- bench_sink += embed_dot_f32(x, y, HIDDEN);
+ bench_sink += dot_f32(x, y, HIDDEN);
free(y);
free(x);
}
@@ -240,7 +239,7 @@ int main(int argc, char **argv) {
}
if (threads < 1)
threads = 1;
- embed_set_threads(threads);
+ tp_set_threads(threads);
char meta[128];
#ifdef USE_BLAS