diff options
Diffstat (limited to 'bench')
| -rw-r--r-- | bench/bench.h | 74 | ||||
| -rw-r--r-- | bench/bench_kernels.c | 115 | ||||
| -rw-r--r-- | bench/bench_model.c | 24 | ||||
| -rw-r--r-- | bench/bench_tokenizer.c | 63 |
4 files changed, 138 insertions, 138 deletions
diff --git a/bench/bench.h b/bench/bench.h index 9a46c71..6089174 100644 --- a/bench/bench.h +++ b/bench/bench.h @@ -30,8 +30,8 @@ #include <unistd.h> typedef struct { - long n; /* iterations: the measured loop must run exactly n times */ - double t0_ns; /* set by bench_begin(); end of un-measured setup */ + long n; /* iterations: the measured loop must run exactly n times */ + double t0_ns; /* set by bench_begin(); end of un-measured setup */ } bench_state_t; typedef void (*bench_fn_t)(bench_state_t *); @@ -45,36 +45,29 @@ typedef struct { * compiler cannot delete the measured loop. */ static volatile float bench_sink; -static double bench_now_ns(void) -{ +static double bench_now_ns(void) { struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts); return (double)ts.tv_sec * 1e9 + (double)ts.tv_nsec; } -static void bench_begin(bench_state_t *b) -{ - b->t0_ns = bench_now_ns(); -} +static void bench_begin(bench_state_t *b) { b->t0_ns = bench_now_ns(); } /* One timed run of fn at n iterations; returns ns/op. */ -static double bench_run_once(bench_fn_t fn, long n) -{ - bench_state_t b = { n, 0.0 }; +static double bench_run_once(bench_fn_t fn, long n) { + bench_state_t b = {n, 0.0}; b.t0_ns = bench_now_ns(); fn(&b); double elapsed = bench_now_ns() - b.t0_ns; return elapsed / (double)n; } -static int bench_cmp_double(const void *a, const void *c) -{ +static int bench_cmp_double(const void *a, const void *c) { double x = *(const double *)a, y = *(const double *)c; return (x > y) - (x < y); } -static double bench_median(double *v, int n) -{ +static double bench_median(double *v, int n) { qsort(v, (size_t)n, sizeof(*v), bench_cmp_double); return n % 2 ? v[n / 2] : 0.5 * (v[n / 2 - 1] + v[n / 2]); } @@ -88,13 +81,13 @@ typedef struct { const char *meta; } bench_opts_t; -static void bench_parse_args(bench_opts_t *o, int argc, char **argv) -{ +static void bench_parse_args(bench_opts_t *o, int argc, char **argv) { o->benchtime_ms = 300; o->count = 10; o->filter = NULL; o->json_path = NULL; - if (!o->meta) o->meta = ""; + if (!o->meta) + o->meta = ""; for (int i = 1; i < argc; i++) { if (!strcmp(argv[i], "--benchtime-ms") && i + 1 < argc) o->benchtime_ms = atol(argv[++i]); @@ -105,15 +98,16 @@ static void bench_parse_args(bench_opts_t *o, int argc, char **argv) else if (!strcmp(argv[i], "--json") && i + 1 < argc) o->json_path = argv[++i]; } - if (o->benchtime_ms < 1) o->benchtime_ms = 1; - if (o->count < 1) o->count = 1; - if (o->count > 1000) o->count = 1000; + if (o->benchtime_ms < 1) + o->benchtime_ms = 1; + if (o->count < 1) + o->count = 1; + if (o->count > 1000) + o->count = 1000; } /* Run all cases (subject to the filter) and report. Returns 0 on success. */ -static int bench_main(const bench_opts_t *o, - const bench_case_t *cases, int n_cases) -{ +static int bench_main(const bench_opts_t *o, const bench_case_t *cases, int n_cases) { FILE *jf = NULL; if (o->json_path) { jf = fopen(o->json_path, "w"); @@ -123,14 +117,16 @@ static int bench_main(const bench_opts_t *o, } char host[256] = "unknown"; gethostname(host, sizeof(host) - 1); - fprintf(jf, "{\"host\":\"%s\",\"benchtime_ms\":%ld,\"count\":%d," + fprintf(jf, + "{\"host\":\"%s\",\"benchtime_ms\":%ld,\"count\":%d," "\"meta\":\"%s\",\"benchmarks\":[", host, o->benchtime_ms, o->count, o->meta); } double *samples = (double *)malloc((size_t)o->count * sizeof(double)); if (!samples) { - if (jf) fclose(jf); + if (jf) + fclose(jf); return 1; } @@ -140,7 +136,8 @@ static int bench_main(const bench_opts_t *o, free(all); free(iters); free(samples); - if (jf) fclose(jf); + if (jf) + fclose(jf); return 1; } @@ -156,8 +153,10 @@ static int bench_main(const bench_opts_t *o, while (ns_per_op * (double)n < target_ns) { double want = target_ns / (ns_per_op > 0.0 ? ns_per_op : 1.0); long next = (long)(want * 1.2) + 1; - if (next > n * 100) next = n * 100; /* distrust tiny timings */ - if (next <= n) next = n + 1; + if (next > n * 100) + next = n * 100; /* distrust tiny timings */ + if (next <= n) + next = n + 1; n = next; ns_per_op = bench_run_once(cases[c].fn, n); } @@ -174,8 +173,7 @@ static int bench_main(const bench_opts_t *o, for (int s = 0; s < o->count; s++) { for (int c = 0; c < n_cases; c++) if (iters[c]) - all[(size_t)c * o->count + s] = - bench_run_once(cases[c].fn, iters[c]); + all[(size_t)c * o->count + s] = bench_run_once(cases[c].fn, iters[c]); fprintf(stderr, "pass %d/%d (%.0fs elapsed)\n", s + 1, o->count, (bench_now_ns() - sample_t0) / 1e9); } @@ -184,22 +182,20 @@ static int bench_main(const bench_opts_t *o, for (int c = 0; c < n_cases; c++) { if (!iters[c]) continue; - memcpy(samples, all + (size_t)c * o->count, - (size_t)o->count * sizeof(double)); + memcpy(samples, all + (size_t)c * o->count, (size_t)o->count * sizeof(double)); if (jf) { - fprintf(jf, "%s{\"name\":\"%s\",\"iters\":%ld,\"ns_per_op\":[", - emitted ? "," : "", cases[c].name, iters[c]); + fprintf(jf, "%s{\"name\":\"%s\",\"iters\":%ld,\"ns_per_op\":[", emitted ? "," : "", + cases[c].name, iters[c]); for (int s = 0; s < o->count; s++) fprintf(jf, "%s%.3f", s ? "," : "", samples[s]); fprintf(jf, "]}"); } double med = bench_median(samples, o->count); - double spread = o->count > 1 - ? 100.0 * (samples[o->count - 1] - samples[0]) / med : 0.0; - printf("%-44s %10ld iters %14.1f ns/op (spread %.1f%%)\n", - cases[c].name, iters[c], med, spread); + double spread = o->count > 1 ? 100.0 * (samples[o->count - 1] - samples[0]) / med : 0.0; + printf("%-44s %10ld iters %14.1f ns/op (spread %.1f%%)\n", cases[c].name, iters[c], med, + spread); fflush(stdout); emitted++; } diff --git a/bench/bench_kernels.c b/bench/bench_kernels.c index cc5dc54..4639d18 100644 --- a/bench/bench_kernels.c +++ b/bench/bench_kernels.c @@ -12,12 +12,12 @@ #include <stdint.h> -int qwen_verbose = 0; /* defined by embed.c in full builds */ +int qwen_verbose = 0; /* defined by embed.c in full builds */ enum { HIDDEN = 1024, - Q_DIM = 2048, /* 16 heads * 128 */ - KV_DIM = 1024, /* 8 heads * 128 */ + Q_DIM = 2048, /* 16 heads * 128 */ + KV_DIM = 1024, /* 8 heads * 128 */ HEAD_DIM = 128, INTER = 3072, SEQ = 8, @@ -25,8 +25,7 @@ enum { }; /* Deterministic fill in [-1, 1]; small values keep softmax/silu sane. */ -static void fill_f32(float *x, size_t n, unsigned seed) -{ +static void fill_f32(float *x, size_t n, unsigned seed) { unsigned s = seed * 2654435761u + 1u; for (size_t i = 0; i < n; i++) { s = s * 1664525u + 1013904223u; @@ -34,15 +33,13 @@ static void fill_f32(float *x, size_t n, unsigned seed) } } -static uint16_t f32_to_bf16(float x) -{ +static uint16_t f32_to_bf16(float x) { uint32_t u; memcpy(&u, &x, sizeof(u)); return (uint16_t)(u >> 16); } -static void fill_bf16(uint16_t *x, size_t n, unsigned seed) -{ +static void fill_bf16(uint16_t *x, size_t n, unsigned seed) { unsigned s = seed * 2654435761u + 1u; for (size_t i = 0; i < n; i++) { s = s * 1664525u + 1013904223u; @@ -50,26 +47,25 @@ static void fill_bf16(uint16_t *x, size_t n, unsigned seed) } } -static float *alloc_f32(size_t n, unsigned seed) -{ +static float *alloc_f32(size_t n, unsigned seed) { float *x = (float *)malloc(n * sizeof(float)); - if (!x) exit(2); + if (!x) + exit(2); fill_f32(x, n, seed); return x; } -static uint16_t *alloc_bf16(size_t n, unsigned seed) -{ +static uint16_t *alloc_bf16(size_t n, unsigned seed) { uint16_t *x = (uint16_t *)malloc(n * sizeof(uint16_t)); - if (!x) exit(2); + if (!x) + exit(2); fill_bf16(x, n, seed); return x; } /* ---- F32 projections (matvec at seq 1, GEMM at seq 8) ---- */ -static void bm_linear_f32(bench_state_t *b, int seq, int in, int out) -{ +static void bm_linear_f32(bench_state_t *b, int seq, int in, int out) { float *x = alloc_f32((size_t)seq * in, 1); float *w = alloc_f32((size_t)out * in, 2); float *y = alloc_f32((size_t)seq * out, 3); @@ -77,18 +73,19 @@ static void bm_linear_f32(bench_state_t *b, int seq, int in, int out) for (long i = 0; i < b->n; i++) qwen_linear_nobias(y, x, w, seq, in, out); bench_sink += y[0]; - free(y); free(w); free(x); + free(y); + free(w); + free(x); } -static void bm_linear_f32_s1_qproj(bench_state_t *b) { bm_linear_f32(b, 1, HIDDEN, Q_DIM); } -static void bm_linear_f32_s1_gate(bench_state_t *b) { bm_linear_f32(b, 1, HIDDEN, INTER); } -static void bm_linear_f32_s1_down(bench_state_t *b) { bm_linear_f32(b, 1, INTER, HIDDEN); } -static void bm_linear_f32_s8_gate(bench_state_t *b) { bm_linear_f32(b, SEQ, HIDDEN, INTER); } +static void bm_linear_f32_s1_qproj(bench_state_t *b) { bm_linear_f32(b, 1, HIDDEN, Q_DIM); } +static void bm_linear_f32_s1_gate(bench_state_t *b) { bm_linear_f32(b, 1, HIDDEN, INTER); } +static void bm_linear_f32_s1_down(bench_state_t *b) { bm_linear_f32(b, 1, INTER, HIDDEN); } +static void bm_linear_f32_s8_gate(bench_state_t *b) { bm_linear_f32(b, SEQ, HIDDEN, INTER); } /* ---- BF16 weights ---- */ -static void bm_linear_bf16(bench_state_t *b, int seq, int in, int out) -{ +static void bm_linear_bf16(bench_state_t *b, int seq, int in, int out) { float *x = alloc_f32((size_t)seq * in, 4); uint16_t *w = alloc_bf16((size_t)out * in, 5); float *y = alloc_f32((size_t)seq * out, 6); @@ -96,14 +93,15 @@ static void bm_linear_bf16(bench_state_t *b, int seq, int in, int out) for (long i = 0; i < b->n; i++) qwen_linear_nobias_bf16(y, x, w, seq, in, out); bench_sink += y[0]; - free(y); free(w); free(x); + free(y); + free(w); + free(x); } static void bm_linear_bf16_s1_gate(bench_state_t *b) { bm_linear_bf16(b, 1, HIDDEN, INTER); } static void bm_linear_bf16_s8_gate(bench_state_t *b) { bm_linear_bf16(b, SEQ, HIDDEN, INTER); } -static void bm_qkv_bf16(bench_state_t *b) -{ +static void bm_qkv_bf16(bench_state_t *b) { float *x = alloc_f32((size_t)SEQ * HIDDEN, 7); uint16_t *wq = alloc_bf16((size_t)Q_DIM * HIDDEN, 8); uint16_t *wk = alloc_bf16((size_t)KV_DIM * HIDDEN, 9); @@ -113,14 +111,18 @@ 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); + qwen_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); free(q); free(wv); free(wk); free(wq); free(x); + free(v); + free(k); + free(q); + free(wv); + free(wk); + free(wq); + free(x); } -static void bm_bf16_widen(bench_state_t *b) -{ +static void bm_bf16_widen(bench_state_t *b) { size_t n = (size_t)INTER * HIDDEN; uint16_t *src = alloc_bf16(n, 14); float *dst = alloc_f32(n, 15); @@ -128,13 +130,13 @@ static void bm_bf16_widen(bench_state_t *b) for (long i = 0; i < b->n; i++) qwen_bf16_to_f32_buf(dst, src, n); bench_sink += dst[0]; - free(dst); free(src); + free(dst); + free(src); } /* ---- attention pieces ---- */ -static void bm_attn_scores(bench_state_t *b) -{ +static void bm_attn_scores(bench_state_t *b) { /* one head: scores[seq,seq] = Q[seq,hd] @ K[seq,hd]^T */ float *qm = alloc_f32((size_t)ATTN_SEQ * HEAD_DIM, 16); float *km = alloc_f32((size_t)ATTN_SEQ * HEAD_DIM, 17); @@ -143,11 +145,12 @@ static void bm_attn_scores(bench_state_t *b) for (long i = 0; i < b->n; i++) qwen_matmul_t(s, qm, km, ATTN_SEQ, HEAD_DIM, ATTN_SEQ); bench_sink += s[0]; - free(s); free(km); free(qm); + free(s); + free(km); + free(qm); } -static void bm_softmax(bench_state_t *b) -{ +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++) @@ -156,23 +159,22 @@ static void bm_softmax(bench_state_t *b) free(x); } -static void bm_rope(bench_state_t *b) -{ +static void bm_rope(bench_state_t *b) { float *x = alloc_f32((size_t)ATTN_SEQ * Q_DIM, 20); float *cosv = alloc_f32((size_t)ATTN_SEQ * HEAD_DIM, 21); 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); + qwen_apply_rope_neox(x, cosv, sinv, ATTN_SEQ, Q_DIM / HEAD_DIM, HEAD_DIM); bench_sink += x[0]; - free(sinv); free(cosv); free(x); + free(sinv); + free(cosv); + free(x); } /* ---- norms, activations, reductions ---- */ -static void bm_rms_norm(bench_state_t *b) -{ +static void bm_rms_norm(bench_state_t *b) { float *x = alloc_f32((size_t)SEQ * HIDDEN, 23); float *w = alloc_f32(HIDDEN, 24); float *y = alloc_f32((size_t)SEQ * HIDDEN, 25); @@ -180,13 +182,14 @@ static void bm_rms_norm(bench_state_t *b) for (long i = 0; i < b->n; i++) qwen_rms_norm(y, x, w, SEQ, HIDDEN, 1e-6f); bench_sink += y[0]; - free(y); free(w); free(x); + free(y); + free(w); + free(x); } /* The engine's SwiGLU op mutates gate in place, so each iteration restores * it first; the memcpy is a constant part of the measured loop. */ -static void bm_silu_mul(bench_state_t *b) -{ +static void bm_silu_mul(bench_state_t *b) { size_t n = (size_t)SEQ * INTER; float *gate0 = alloc_f32(n, 26); float *gate = alloc_f32(n, 26); @@ -197,17 +200,19 @@ static void bm_silu_mul(bench_state_t *b) qwen_silu_mul_inplace(gate, up, (int)n); } bench_sink += gate[0]; - free(up); free(gate); free(gate0); + free(up); + free(gate); + free(gate0); } -static void bm_dot(bench_state_t *b) -{ +static void bm_dot(bench_state_t *b) { float *x = alloc_f32(HIDDEN, 28); 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); - free(y); free(x); + free(y); + free(x); } static const bench_case_t CASES[] = { @@ -227,14 +232,14 @@ static const bench_case_t CASES[] = { {"reduce/dot_1024", bm_dot}, }; -int main(int argc, char **argv) -{ +int main(int argc, char **argv) { int threads = 1; for (int i = 1; i < argc; i++) { if (!strcmp(argv[i], "--threads") && i + 1 < argc) threads = atoi(argv[i + 1]); } - if (threads < 1) threads = 1; + if (threads < 1) + threads = 1; qwen_set_threads(threads); char meta[128]; diff --git a/bench/bench_model.c b/bench/bench_model.c index 9b897e9..5cff015 100644 --- a/bench/bench_model.c +++ b/bench/bench_model.c @@ -19,15 +19,15 @@ static int g_hidden; enum { MAX_BATCH = 8 }; -static void bm_embed(bench_state_t *b, int batch, int len) -{ +static void bm_embed(bench_state_t *b, int batch, int len) { int *ids = (int *)malloc((size_t)batch * len * sizeof(int)); float *out = (float *)malloc((size_t)batch * g_hidden * sizeof(float)); - if (!ids || !out) exit(2); + if (!ids || !out) + exit(2); unsigned s = 99u; for (int i = 0; i < batch * len; i++) { s = s * 1664525u + 1013904223u; - ids[i] = (int)(s % 1000u); /* ordinary vocab rows */ + ids[i] = (int)(s % 1000u); /* ordinary vocab rows */ } embed_input_t inputs[MAX_BATCH]; for (int i = 0; i < batch; i++) { @@ -44,8 +44,8 @@ static void bm_embed(bench_state_t *b, int batch, int len) free(ids); } -static void bm_embed_b1_len32(bench_state_t *b) { bm_embed(b, 1, 32); } -static void bm_embed_b8_len32(bench_state_t *b) { bm_embed(b, 8, 32); } +static void bm_embed_b1_len32(bench_state_t *b) { bm_embed(b, 1, 32); } +static void bm_embed_b8_len32(bench_state_t *b) { bm_embed(b, 8, 32); } static void bm_embed_b1_len256(bench_state_t *b) { bm_embed(b, 1, 256); } static const bench_case_t CASES[] = { @@ -54,11 +54,9 @@ static const bench_case_t CASES[] = { {"embed/b1_len256", bm_embed_b1_len256}, }; -int main(int argc, char **argv) -{ +int main(int argc, char **argv) { if (argc < 2 || argv[1][0] == '-') { - fprintf(stderr, "usage: %s MODEL_DIR [--threads N] [bench flags]\n", - argv[0]); + fprintf(stderr, "usage: %s MODEL_DIR [--threads N] [bench flags]\n", argv[0]); return 2; } @@ -67,7 +65,8 @@ int main(int argc, char **argv) if (!strcmp(argv[i], "--threads") && i + 1 < argc) threads = atoi(argv[i + 1]); } - if (threads < 1) threads = 1; + if (threads < 1) + threads = 1; qwen_set_threads(threads); g_model = embed_model_load(argv[1]); @@ -83,8 +82,7 @@ int main(int argc, char **argv) g_hidden = embed_model_config(g_model)->hidden_size; char meta[160]; - snprintf(meta, sizeof(meta), "suite=model threads=%d hidden=%d", - threads, g_hidden); + snprintf(meta, sizeof(meta), "suite=model threads=%d hidden=%d", threads, g_hidden); bench_opts_t opts = {0}; opts.meta = meta; diff --git a/bench/bench_tokenizer.c b/bench/bench_tokenizer.c index 1033f2e..fb6d37a 100644 --- a/bench/bench_tokenizer.c +++ b/bench/bench_tokenizer.c @@ -10,24 +10,23 @@ int qwen_verbose = 0; -static double now_ms(void) -{ +static double now_ms(void) { struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts); return (double)ts.tv_sec * 1000.0 + (double)ts.tv_nsec / 1000000.0; } -static int same_ids(const int *a, int na, const int *b, int nb) -{ - if (na != nb) return 0; +static int same_ids(const int *a, int na, const int *b, int nb) { + if (na != nb) + return 0; for (int i = 0; i < na; i++) { - if (a[i] != b[i]) return 0; + if (a[i] != b[i]) + return 0; } return 1; } -int main(int argc, char **argv) -{ +int main(int argc, char **argv) { int status = 1; int **gold = NULL; int *gold_n = NULL; @@ -41,7 +40,8 @@ int main(int argc, char **argv) const char *vocab = argv[1]; int runs = atoi(argv[2]); - if (runs <= 0) return 2; + if (runs <= 0) + return 2; int n_texts = argc - 3; char **texts = argv + 3; @@ -56,17 +56,18 @@ int main(int argc, char **argv) gold_n = (int *)calloc((size_t)n_texts, sizeof(*gold_n)); bufs = (int **)calloc((size_t)n_texts, sizeof(*bufs)); caps = (int *)calloc((size_t)n_texts, sizeof(*caps)); - if (!gold || !gold_n || !bufs || !caps) goto cleanup; + if (!gold || !gold_n || !bufs || !caps) + goto cleanup; int tokens_per_pass = 0; for (int i = 0; i < n_texts; i++) { gold[i] = qwen_tokenizer_encode(tok, texts[i], &gold_n[i]); - if (!gold[i] || gold_n[i] <= 0) goto cleanup; + 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 = qwen_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; @@ -74,12 +75,13 @@ int main(int argc, char **argv) free(ids_ws); caps[i] = (int)(strlen(texts[i]) * 2 + 8); - if (caps[i] < gold_n[i]) caps[i] = gold_n[i]; + if (caps[i] < gold_n[i]) + caps[i] = gold_n[i]; bufs[i] = (int *)malloc((size_t)caps[i] * sizeof(int)); - if (!bufs[i]) goto cleanup; + 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 (qwen_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; } @@ -105,8 +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 = qwen_tokenizer_encode_with_workspace(tok, ws, texts[i], &n); if (!ids || n != gold_n[i]) { free(ids); goto cleanup; @@ -122,30 +123,30 @@ 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); - if (rc != 0 || n != gold_n[i]) goto cleanup; + int rc = qwen_tokenizer_encode_into(tok, ws, texts[i], bufs[i], caps[i], &n); + if (rc != 0 || n != gold_n[i]) + goto cleanup; into_tokens += n; } } double into_ms = now_ms() - t0; - printf("texts=%d runs=%d tokens/pass=%d\n", n_texts, runs, - tokens_per_pass); + printf("texts=%d runs=%d tokens/pass=%d\n", n_texts, runs, tokens_per_pass); printf("old_malloc: %.3f ms %.0f tokens/s\n", old_ms, (double)old_tokens / (old_ms / 1000.0)); - printf("workspace_malloc: %.3f ms %.0f tokens/s speedup=%.2fx\n", - ws_ms, (double)ws_tokens / (ws_ms / 1000.0), old_ms / ws_ms); - printf("workspace_into: %.3f ms %.0f tokens/s speedup=%.2fx\n", - into_ms, (double)into_tokens / (into_ms / 1000.0), - old_ms / into_ms); + printf("workspace_malloc: %.3f ms %.0f tokens/s speedup=%.2fx\n", ws_ms, + (double)ws_tokens / (ws_ms / 1000.0), old_ms / ws_ms); + printf("workspace_into: %.3f ms %.0f tokens/s speedup=%.2fx\n", into_ms, + (double)into_tokens / (into_ms / 1000.0), old_ms / into_ms); status = 0; cleanup: for (int i = 0; i < n_texts; i++) { - if (gold) free(gold[i]); - if (bufs) free(bufs[i]); + if (gold) + free(gold[i]); + if (bufs) + free(bufs[i]); } free(gold); free(gold_n); |
