aboutsummaryrefslogtreecommitdiff
path: root/bench
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-16 16:27:47 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-16 16:27:47 +0200
commitd1ba1ebc06ef0d62ca932388fae83e6992594719 (patch)
treef962207b779016593258f990348216999a58766a /bench
parentbe20dfda0f836808dccfc14db244592e41ba28bc (diff)
downloadffwd-d1ba1ebc06ef0d62ca932388fae83e6992594719.tar.gz
ffwd-d1ba1ebc06ef0d62ca932388fae83e6992594719.zip
chore: move sources to src/ with proper names, drop qwen_ prefix
Diffstat (limited to 'bench')
-rw-r--r--bench/bench.h2
-rw-r--r--bench/bench_kernels.c26
-rw-r--r--bench/bench_model.c4
-rw-r--r--bench/bench_tokenizer.c24
4 files changed, 28 insertions, 28 deletions
diff --git a/bench/bench.h b/bench/bench.h
index 6089174..1d2b8f2 100644
--- a/bench/bench.h
+++ b/bench/bench.h
@@ -8,7 +8,7 @@
* float *x = make_input(); // setup, not measured
* bench_begin(b);
* for (long i = 0; i < b->n; i++)
- * bench_sink += qwen_dot_f32(x, x, 1024);
+ * bench_sink += embed_dot_f32(x, x, 1024);
* free(x);
* }
*
diff --git a/bench/bench_kernels.c b/bench/bench_kernels.c
index 4639d18..2b2fe47 100644
--- a/bench/bench_kernels.c
+++ b/bench/bench_kernels.c
@@ -8,11 +8,11 @@
*/
#include "bench.h"
-#include "qwen_kernels.h"
+#include "kernels.h"
#include <stdint.h>
-int qwen_verbose = 0; /* defined by embed.c in full builds */
+int embed_verbose = 0; /* defined by embed.c in full builds */
enum {
HIDDEN = 1024,
@@ -71,7 +71,7 @@ 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++)
- qwen_linear_nobias(y, x, w, seq, in, out);
+ embed_linear_nobias(y, x, w, seq, in, out);
bench_sink += y[0];
free(y);
free(w);
@@ -91,7 +91,7 @@ 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++)
- qwen_linear_nobias_bf16(y, x, w, seq, in, out);
+ embed_linear_nobias_bf16(y, x, w, seq, in, out);
bench_sink += y[0];
free(y);
free(w);
@@ -111,7 +111,7 @@ 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++)
- qwen_linear_nobias_bf16_qkv(q, k, v, x, wq, wk, wv, SEQ, HIDDEN, Q_DIM, KV_DIM);
+ embed_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 +128,7 @@ 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++)
- qwen_bf16_to_f32_buf(dst, src, n);
+ embed_bf16_to_f32_buf(dst, src, n);
bench_sink += dst[0];
free(dst);
free(src);
@@ -143,7 +143,7 @@ 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++)
- qwen_matmul_t(s, qm, km, ATTN_SEQ, HEAD_DIM, ATTN_SEQ);
+ embed_matmul_t(s, qm, km, ATTN_SEQ, HEAD_DIM, ATTN_SEQ);
bench_sink += s[0];
free(s);
free(km);
@@ -154,7 +154,7 @@ 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++)
- qwen_softmax(x, ATTN_SEQ, ATTN_SEQ);
+ embed_softmax(x, ATTN_SEQ, ATTN_SEQ);
bench_sink += x[0];
free(x);
}
@@ -165,7 +165,7 @@ 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++)
- qwen_apply_rope_neox(x, cosv, sinv, ATTN_SEQ, Q_DIM / HEAD_DIM, HEAD_DIM);
+ embed_apply_rope_neox(x, cosv, sinv, ATTN_SEQ, Q_DIM / HEAD_DIM, HEAD_DIM);
bench_sink += x[0];
free(sinv);
free(cosv);
@@ -180,7 +180,7 @@ 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++)
- qwen_rms_norm(y, x, w, SEQ, HIDDEN, 1e-6f);
+ embed_rms_norm(y, x, w, SEQ, HIDDEN, 1e-6f);
bench_sink += y[0];
free(y);
free(w);
@@ -197,7 +197,7 @@ 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));
- qwen_silu_mul_inplace(gate, up, (int)n);
+ embed_silu_mul_inplace(gate, up, (int)n);
}
bench_sink += gate[0];
free(up);
@@ -210,7 +210,7 @@ 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 += qwen_dot_f32(x, y, HIDDEN);
+ bench_sink += embed_dot_f32(x, y, HIDDEN);
free(y);
free(x);
}
@@ -240,7 +240,7 @@ int main(int argc, char **argv) {
}
if (threads < 1)
threads = 1;
- qwen_set_threads(threads);
+ embed_set_threads(threads);
char meta[128];
#ifdef USE_BLAS
diff --git a/bench/bench_model.c b/bench/bench_model.c
index 5cff015..b2d4769 100644
--- a/bench/bench_model.c
+++ b/bench/bench_model.c
@@ -11,7 +11,7 @@
#include "bench.h"
#include "embed.h"
-#include "qwen_kernels.h"
+#include "kernels.h"
static embed_model_t *g_model;
static embed_workspace_t *g_ws;
@@ -67,7 +67,7 @@ int main(int argc, char **argv) {
}
if (threads < 1)
threads = 1;
- qwen_set_threads(threads);
+ embed_set_threads(threads);
g_model = embed_model_load(argv[1]);
if (!g_model) {
diff --git a/bench/bench_tokenizer.c b/bench/bench_tokenizer.c
index fb6d37a..bf4d0b7 100644
--- a/bench/bench_tokenizer.c
+++ b/bench/bench_tokenizer.c
@@ -1,14 +1,14 @@
/* bench/bench_tokenizer.c - tokenizer encode-path benchmark.
* Build via `make bench-tokenizer`; needs a model dir (vocab.json). */
-#include "qwen_tokenizer.h"
+#include "tokenizer_bpe.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
-int qwen_verbose = 0;
+int embed_verbose = 0;
static double now_ms(void) {
struct timespec ts;
@@ -45,8 +45,8 @@ int main(int argc, char **argv) {
int n_texts = argc - 3;
char **texts = argv + 3;
- qwen_tokenizer_t *tok = qwen_tokenizer_load(vocab);
- qwen_tokenizer_workspace_t *ws = qwen_tokenizer_workspace_new();
+ embed_tokenizer_t *tok = embed_tokenizer_load(vocab);
+ embed_tokenizer_workspace_t *ws = embed_tokenizer_workspace_new();
if (!tok || !ws) {
fprintf(stderr, "failed to initialize tokenizer\n");
goto cleanup;
@@ -61,13 +61,13 @@ int main(int argc, char **argv) {
int tokens_per_pass = 0;
for (int i = 0; i < n_texts; i++) {
- gold[i] = qwen_tokenizer_encode(tok, texts[i], &gold_n[i]);
+ gold[i] = embed_tokenizer_encode(tok, texts[i], &gold_n[i]);
if (!gold[i] || gold_n[i] <= 0)
goto cleanup;
tokens_per_pass += gold_n[i];
int n_ws = 0;
- int *ids_ws = qwen_tokenizer_encode_with_workspace(tok, ws, texts[i], &n_ws);
+ int *ids_ws = embed_tokenizer_encode_with_workspace(tok, ws, texts[i], &n_ws);
if (!ids_ws || !same_ids(gold[i], gold_n[i], ids_ws, n_ws)) {
free(ids_ws);
goto cleanup;
@@ -81,7 +81,7 @@ int main(int argc, char **argv) {
if (!bufs[i])
goto cleanup;
int n_into = 0;
- if (qwen_tokenizer_encode_into(tok, ws, texts[i], bufs[i], caps[i], &n_into) != 0 ||
+ if (embed_tokenizer_encode_into(tok, ws, texts[i], bufs[i], caps[i], &n_into) != 0 ||
!same_ids(gold[i], gold_n[i], bufs[i], n_into))
goto cleanup;
}
@@ -91,7 +91,7 @@ int main(int argc, char **argv) {
for (int r = 0; r < runs; r++) {
for (int i = 0; i < n_texts; i++) {
int n = 0;
- int *ids = qwen_tokenizer_encode(tok, texts[i], &n);
+ int *ids = embed_tokenizer_encode(tok, texts[i], &n);
if (!ids || n != gold_n[i]) {
free(ids);
goto cleanup;
@@ -107,7 +107,7 @@ int main(int argc, char **argv) {
for (int r = 0; r < runs; r++) {
for (int i = 0; i < n_texts; i++) {
int n = 0;
- int *ids = qwen_tokenizer_encode_with_workspace(tok, ws, texts[i], &n);
+ int *ids = embed_tokenizer_encode_with_workspace(tok, ws, texts[i], &n);
if (!ids || n != gold_n[i]) {
free(ids);
goto cleanup;
@@ -123,7 +123,7 @@ int main(int argc, char **argv) {
for (int r = 0; r < runs; r++) {
for (int i = 0; i < n_texts; i++) {
int n = 0;
- int rc = qwen_tokenizer_encode_into(tok, ws, texts[i], bufs[i], caps[i], &n);
+ int rc = embed_tokenizer_encode_into(tok, ws, texts[i], bufs[i], caps[i], &n);
if (rc != 0 || n != gold_n[i])
goto cleanup;
into_tokens += n;
@@ -152,7 +152,7 @@ cleanup:
free(gold_n);
free(bufs);
free(caps);
- qwen_tokenizer_workspace_free(ws);
- qwen_tokenizer_free(tok);
+ embed_tokenizer_workspace_free(ws);
+ embed_tokenizer_free(tok);
return status;
}