aboutsummaryrefslogtreecommitdiff
path: root/tests/test_kernels.c
blob: 37847c1142eb46156faf1d0833857d319e49a1ab (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
/* tests/test_kernels.c - golden tests for the low-level math kernels.
 * Runs via `make test`, `make -C tests test_kernels test_kernels_generic`,
 * or `./all.bash check-kernel-golden`. No model files required. */

#include "kernels.h"
#include "threadpool.h"
#include "impl.h"

#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

static int failures = 0;

static void expect_close(const char *name, float got, float want, float tol) {
    float diff = fabsf(got - want);
    float scale = fmaxf(1.0f, fmaxf(fabsf(got), fabsf(want)));
    if (diff > tol * scale) {
        fprintf(stderr, "%s: got %.9g want %.9g diff %.9g\n", name, got, want, diff);
        failures++;
    }
}

static uint16_t f32_to_bf16(float x) {
    uint32_t u;
    memcpy(&u, &x, sizeof(u));
    return (uint16_t)(u >> 16);
}

static float bf16_to_f32(uint16_t x) {
    uint32_t u = ((uint32_t)x) << 16;
    float f;
    memcpy(&f, &u, sizeof(f));
    return f;
}

static void test_rms_norm(void) {
    const int seq = 2, hidden = 4;
    const float x[8] = {1.0f, 2.0f, -3.0f, 4.0f, 0.5f, -1.0f, 2.0f, -0.5f};
    const float w[4] = {1.0f, 0.5f, -1.0f, 2.0f};
    float out[8];

    rms_norm(out, x, w, seq, hidden, 1e-6f);

    for (int s = 0; s < seq; s++) {
        float ss = 0.0f;
        for (int d = 0; d < hidden; d++) {
            float v = x[s * hidden + d];
            ss += v * v;
        }
        float inv = 1.0f / sqrtf(ss / hidden + 1e-6f);
        for (int d = 0; d < hidden; d++) {
            float want = x[s * hidden + d] * inv * w[d];
            expect_close("rms_norm", out[s * hidden + d], want, 1e-6f);
        }
    }
}

static void test_layer_norm(void) {
    const int seq = 2, hidden = 4;
    const float x[8] = {1.0f, 2.0f, -3.0f, 4.0f, 0.5f, -1.0f, 2.0f, -0.5f};
    const float g[4] = {1.0f, 0.5f, -1.0f, 2.0f};
    const float b[4] = {0.1f, -0.2f, 0.3f, -0.4f};
    float out[8];

    layer_norm(out, x, g, b, seq, hidden, 1e-12f);

    for (int s = 0; s < seq; s++) {
        float mean = 0.0f;
        for (int d = 0; d < hidden; d++)
            mean += x[s * hidden + d];
        mean /= hidden;
        float var = 0.0f;
        for (int d = 0; d < hidden; d++) {
            float diff = x[s * hidden + d] - mean;
            var += diff * diff;
        }
        var /= hidden;
        float inv = 1.0f / sqrtf(var + 1e-12f);
        for (int d = 0; d < hidden; d++) {
            float want = g[d] * (x[s * hidden + d] - mean) * inv + b[d];
            expect_close("layer_norm", out[s * hidden + d], want, 1e-5f);
        }
    }

    /* In-place (out aliases x) must match the out-of-place result. */
    float z[8];
    memcpy(z, x, sizeof(z));
    layer_norm(z, z, g, b, seq, hidden, 1e-12f);
    for (int i = 0; i < 8; i++)
        expect_close("layer_norm_inplace", z[i], out[i], 1e-6f);
}

static void test_gelu(void) {
    float x[5] = {0.0f, 1.0f, -1.0f, 2.5f, -3.0f};
    float out[5];
    memcpy(out, x, sizeof(out));
    gelu_inplace(out, 5);
    for (int i = 0; i < 5; i++) {
        float want = 0.5f * x[i] * (1.0f + erff(x[i] * 0.70710678118654752f));
        expect_close("gelu", out[i], want, 1e-6f);
    }
    /* Known reference values for the exact erf GeLU. */
    expect_close("gelu0", out[0], 0.0f, 1e-7f);
    expect_close("gelu1", out[1], 0.841344746f, 1e-5f);
    expect_close("gelu_neg1", out[2], -0.158655254f, 1e-5f);
}

static void test_gelu_tanh(void) {
    float x[5] = {0.0f, 1.0f, -1.0f, 2.5f, -3.0f};
    float out[5];
    memcpy(out, x, sizeof(out));
    gelu_tanh_inplace(out, 5);
    for (int i = 0; i < 5; i++) {
        float v = x[i];
        float inner = 0.79788456080286536f * (v + 0.044715f * v * v * v);
        float want = 0.5f * v * (1.0f + tanhf(inner));
        expect_close("gelu_tanh", out[i], want, 1e-6f);
    }
    /* Reference values for the tanh approximation; distinct from the erf GeLU
     * (erf gelu(1) = 0.841344746) by ~1.5e-4, well outside this tolerance. */
    expect_close("gelu_tanh0", out[0], 0.0f, 1e-7f);
    expect_close("gelu_tanh1", out[1], 0.841191991f, 1e-5f);
    expect_close("gelu_tanh_neg1", out[2], -0.158808009f, 1e-5f);
}

static void test_rope_neox(void) {
    const int seq = 2, heads = 1, head_dim = 4;
    const int positions[2] = {0, 3};
    float cosv[8], sinv[8];
    float x[8] = {1.0f, 2.0f, 3.0f, 4.0f, -1.0f, 0.5f, 2.0f, -0.25f};
    float orig[8];
    memcpy(orig, x, sizeof(x));

    compute_rope_neox(cosv, sinv, positions, seq, head_dim, 10000.0f);
    apply_rope_neox(x, cosv, sinv, seq, heads, head_dim);

    for (int s = 0; s < seq; s++) {
        for (int d = 0; d < head_dim / 2; d++) {
            float freq = 1.0f / powf(10000.0f, (float)(2 * d) / head_dim);
            float angle = (float)positions[s] * freq;
            float c = cosf(angle);
            float sn = sinf(angle);
            float x1 = orig[s * head_dim + d];
            float x2 = orig[s * head_dim + head_dim / 2 + d];
            float want1 = x1 * c - x2 * sn;
            float want2 = x2 * c + x1 * sn;
            expect_close("rope_1", x[s * head_dim + d], want1, 1e-6f);
            expect_close("rope_2", x[s * head_dim + head_dim / 2 + d], want2, 1e-6f);
        }
    }
}

static void reference_packed_gqa(float *out,
                                 const float *Q,
                                 const float *K,
                                 const float *V,
                                 const int *offsets,
                                 int batch,
                                 int n_heads,
                                 int n_kv_heads,
                                 int head_dim,
                                 float scale,
                                 int causal) {
    int q_hidden = n_heads * head_dim;
    int kv_hidden = n_kv_heads * head_dim;
    int heads_per_kv = n_heads / n_kv_heads;

    for (int h = 0; h < n_heads; h++) {
        int kv_h = h / heads_per_kv;
        for (int b = 0; b < batch; b++) {
            int start = offsets[b], end = offsets[b + 1];
            for (int i = start; i < end; i++) {
                int key_end = causal ? i + 1 : end;
                const float *q = Q + i * q_hidden + h * head_dim;
                float *o = out + i * q_hidden + h * head_dim;
                float max_score = -INFINITY;
                for (int j = start; j < key_end; j++) {
                    const float *k = K + j * kv_hidden + kv_h * head_dim;
                    float score = 0.0f;
                    for (int d = 0; d < head_dim; d++)
                        score += q[d] * k[d];
                    score *= scale;
                    if (score > max_score)
                        max_score = score;
                }
                float denom = 0.0f;
                for (int d = 0; d < head_dim; d++)
                    o[d] = 0.0f;
                for (int j = start; j < key_end; j++) {
                    const float *k = K + j * kv_hidden + kv_h * head_dim;
                    const float *v = V + j * kv_hidden + kv_h * head_dim;
                    float score = 0.0f;
                    for (int d = 0; d < head_dim; d++)
                        score += q[d] * k[d];
                    float wt = expf(score * scale - max_score);
                    denom += wt;
                    for (int d = 0; d < head_dim; d++)
                        o[d] += wt * v[d];
                }
                for (int d = 0; d < head_dim; d++)
                    o[d] /= denom;
            }
        }
    }
}

static void test_packed_gqa_attention(void) {
    enum { total = 5, n_heads = 2, n_kv_heads = 1, head_dim = 2 };
    enum { q_hidden = n_heads * head_dim };
    enum { kv_hidden = n_kv_heads * head_dim };
    const int offsets[3] = {0, 2, 5};
    float Q[total * q_hidden];
    float K[total * kv_hidden];
    float V[total * kv_hidden];
    float got[total * q_hidden];
    float want[total * q_hidden];

    for (int i = 0; i < total * q_hidden; i++)
        Q[i] = sinf((float)i * 0.17f) - 0.25f;
    for (int i = 0; i < total * kv_hidden; i++) {
        K[i] = cosf((float)i * 0.11f) + 0.1f;
        V[i] = sinf((float)i * 0.07f) - cosf((float)i * 0.13f);
    }

    bidirectional_gqa_attention_packed(got, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim,
                                       1.0f / sqrtf((float)head_dim));
    reference_packed_gqa(want, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim,
                         1.0f / sqrtf((float)head_dim), 0);

    for (int i = 0; i < total * q_hidden; i++)
        expect_close("packed_gqa", got[i], want[i], 2e-6f);
}

static void test_packed_gqa_attention_long(void) {
    enum { total = 136, n_heads = 4, n_kv_heads = 2, head_dim = 8 };
    enum { q_hidden = n_heads * head_dim };
    enum { kv_hidden = n_kv_heads * head_dim };
    const int offsets[3] = {0, 129, 136};
    float *Q = malloc((size_t)total * q_hidden * sizeof(*Q));
    float *K = malloc((size_t)total * kv_hidden * sizeof(*K));
    float *V = malloc((size_t)total * kv_hidden * sizeof(*V));
    float *got = malloc((size_t)total * q_hidden * sizeof(*got));
    float *want = malloc((size_t)total * q_hidden * sizeof(*want));
    if (!Q || !K || !V || !got || !want)
        exit(2);

    for (int i = 0; i < total * q_hidden; i++)
        Q[i] = sinf((float)i * 0.017f) - 0.25f;
    for (int i = 0; i < total * kv_hidden; i++) {
        K[i] = cosf((float)i * 0.011f) + 0.1f;
        V[i] = sinf((float)i * 0.007f) - cosf((float)i * 0.013f);
    }

    tp_set_threads(4);
    bidirectional_gqa_attention_packed(got, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim,
                                       1.0f / sqrtf((float)head_dim));
    reference_packed_gqa(want, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim,
                         1.0f / sqrtf((float)head_dim), 0);
    tp_set_threads(1);

    for (int i = 0; i < total * q_hidden; i++)
        expect_close("packed_gqa_long", got[i], want[i], 2e-5f);
    free(want);
    free(got);
    free(V);
    free(K);
    free(Q);
}

static void test_packed_causal_gqa_attention(void) {
    enum { total = 136, n_heads = 4, n_kv_heads = 2, head_dim = 8 };
    enum { q_hidden = n_heads * head_dim };
    enum { kv_hidden = n_kv_heads * head_dim };
    const int offsets[3] = {0, 7, 136};
    float *Q = malloc((size_t)total * q_hidden * sizeof(*Q));
    float *K = malloc((size_t)total * kv_hidden * sizeof(*K));
    float *V = malloc((size_t)total * kv_hidden * sizeof(*V));
    float *got = malloc((size_t)total * q_hidden * sizeof(*got));
    float *want = malloc((size_t)total * q_hidden * sizeof(*want));
    if (!Q || !K || !V || !got || !want)
        exit(2);

    for (int i = 0; i < total * q_hidden; i++)
        Q[i] = sinf((float)i * 0.019f) - 0.2f;
    for (int i = 0; i < total * kv_hidden; i++) {
        K[i] = cosf((float)i * 0.013f) + 0.15f;
        V[i] = sinf((float)i * 0.009f) - cosf((float)i * 0.015f);
    }

    tp_set_threads(4);
    causal_gqa_attention_packed(got, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim,
                                1.0f / sqrtf((float)head_dim));
    reference_packed_gqa(want, Q, K, V, offsets, 2, n_heads, n_kv_heads, head_dim,
                         1.0f / sqrtf((float)head_dim), 1);
    tp_set_threads(1);

    for (int i = 0; i < total * q_hidden; i++)
        expect_close("packed_causal_gqa", got[i], want[i], 2e-5f);
    free(want);
    free(got);
    free(V);
    free(K);
    free(Q);
}

static void test_bf16_linear(void) {
    enum { seq = 2, in_dim = 3, out_dim = 2 };
    const float x[6] = {1.0f, -2.0f, 0.5f, -1.0f, 0.25f, 2.0f};
    const float wf32[6] = {0.5f, -1.0f, 2.0f, -0.25f, 0.75f, 1.5f};
    uint16_t wbf16[6];
    float got[seq * out_dim];

    for (int i = 0; i < 6; i++)
        wbf16[i] = f32_to_bf16(wf32[i]);
    linear_nobias_bf16(got, x, wbf16, seq, in_dim, out_dim);

    for (int s = 0; s < seq; s++) {
        for (int o = 0; o < out_dim; o++) {
            float want = 0.0f;
            for (int i = 0; i < in_dim; i++)
                want += x[s * in_dim + i] * bf16_to_f32(wbf16[o * in_dim + i]);
            expect_close("bf16_linear", got[s * out_dim + o], want, 1e-6f);
        }
    }
}

static void check_bf16_qkv(int n_threads) {
    enum { seq = 3, in_dim = 4, q_dim = 5, kv_dim = 2 };
    float x[seq * in_dim];
    float wq_f32[q_dim * in_dim];
    float wk_f32[kv_dim * in_dim];
    float wv_f32[kv_dim * in_dim];
    uint16_t wq[q_dim * in_dim];
    uint16_t wk[kv_dim * in_dim];
    uint16_t wv[kv_dim * in_dim];
    float got_q[seq * q_dim], got_k[seq * kv_dim], got_v[seq * kv_dim];

    for (int i = 0; i < seq * in_dim; i++)
        x[i] = sinf((float)i * 0.23f) - 0.4f;
    for (int i = 0; i < q_dim * in_dim; i++) {
        wq_f32[i] = cosf((float)i * 0.17f) + 0.1f;
        wq[i] = f32_to_bf16(wq_f32[i]);
    }
    for (int i = 0; i < kv_dim * in_dim; i++) {
        wk_f32[i] = sinf((float)i * 0.31f) - 0.2f;
        wv_f32[i] = cosf((float)i * 0.29f) + 0.3f;
        wk[i] = f32_to_bf16(wk_f32[i]);
        wv[i] = f32_to_bf16(wv_f32[i]);
    }

    tp_set_threads(n_threads);
    linear_nobias_bf16_qkv(got_q, got_k, got_v, x, wq, wk, wv, seq, in_dim, q_dim, kv_dim);

    for (int s = 0; s < seq; s++) {
        for (int o = 0; o < q_dim; o++) {
            float want = 0.0f;
            for (int i = 0; i < in_dim; i++)
                want += x[s * in_dim + i] * bf16_to_f32(wq[o * in_dim + i]);
            expect_close("bf16_qkv_q", got_q[s * q_dim + o], want, 1e-6f);
        }
        for (int o = 0; o < kv_dim; o++) {
            float want_k = 0.0f, want_v = 0.0f;
            for (int i = 0; i < in_dim; i++) {
                want_k += x[s * in_dim + i] * bf16_to_f32(wk[o * in_dim + i]);
                want_v += x[s * in_dim + i] * bf16_to_f32(wv[o * in_dim + i]);
            }
            expect_close("bf16_qkv_k", got_k[s * kv_dim + o], want_k, 1e-6f);
            expect_close("bf16_qkv_v", got_v[s * kv_dim + o], want_v, 1e-6f);
        }
    }
}

static void test_bf16_qkv(void) {
    check_bf16_qkv(1);
    check_bf16_qkv(3);
    tp_set_threads(1);
}

/* seq==1 routes through bf16_matvec_threaded; a pool splits the output
 * rows across matvec_worker calls. Both must match the scalar reference. */
static void check_bf16_matvec(int n_threads) {
    enum { in_dim = 19, out_dim = 23 };
    float x[in_dim];
    uint16_t w[out_dim * in_dim];
    float got[out_dim];

    for (int i = 0; i < in_dim; i++)
        x[i] = sinf((float)i * 0.37f) - 0.2f;
    for (int i = 0; i < out_dim * in_dim; i++)
        w[i] = f32_to_bf16(cosf((float)i * 0.21f) + 0.05f);

    tp_set_threads(n_threads);
    linear_nobias_bf16(got, x, w, 1, in_dim, out_dim);

    for (int o = 0; o < out_dim; o++) {
        float want = 0.0f;
        for (int i = 0; i < in_dim; i++)
            want += x[i] * bf16_to_f32(w[o * in_dim + i]);
        expect_close("bf16_matvec", got[o], want, 1e-6f);
    }
}

static void test_bf16_matvec(void) {
    check_bf16_matvec(1);
    check_bf16_matvec(4);
    tp_set_threads(1);
}

/* Above TKERN_RMS_NORM_PARALLEL_ELEMS (256k elements) rms_norm fans out to
 * rms_norm_worker; row-wise math is identical, so results must match the
 * single-thread run exactly. */
static void test_rms_norm_threaded(void) {
    enum { seq = 96, hidden = 3072 }; /* 294912 elements */
    size_t elems = (size_t)seq * hidden;
    float *x = (float *)malloc(elems * sizeof(float));
    float *w = (float *)malloc((size_t)hidden * sizeof(float));
    float *one = (float *)malloc(elems * sizeof(float));
    float *many = (float *)malloc(elems * sizeof(float));
    if (!x || !w || !one || !many) {
        fprintf(stderr, "rms_norm_threaded: allocation failure\n");
        failures++;
        free(many);
        free(one);
        free(w);
        free(x);
        return;
    }
    unsigned s = 7u;
    for (size_t i = 0; i < elems; i++) {
        s = s * 1664525u + 1013904223u;
        x[i] = (float)((s >> 8) & 0xFFFF) / 32768.0f - 1.0f;
    }
    for (int i = 0; i < hidden; i++)
        w[i] = 0.8f + 0.01f * (float)(i % 7);

    tp_set_threads(1);
    rms_norm(one, x, w, seq, hidden, 1e-6f);
    tp_set_threads(4);
    rms_norm(many, x, w, seq, hidden, 1e-6f);
    tp_set_threads(1);

    for (size_t i = 0; i < elems; i++) {
        if (one[i] != many[i]) {
            fprintf(stderr, "rms_norm_threaded: mismatch at %zu: %.9g %.9g\n", i, one[i], many[i]);
            failures++;
            break;
        }
    }
    free(many);
    free(one);
    free(w);
    free(x);
}

/* bf16->f32 widening is an exact bit shift; check the SIMD body and the
 * remainder lanes against the scalar definition. */
static void test_bf16_widen_buf(void) {
    enum { NMAXW = 67 };
    static const int sizes[] = {1, 7, 8, 15, 64, NMAXW};
    uint16_t src[NMAXW];
    float got[NMAXW];

    for (int i = 0; i < NMAXW; i++)
        src[i] = f32_to_bf16(sinf((float)i * 0.13f) * 3.0f);

    for (size_t t = 0; t < sizeof(sizes) / sizeof(sizes[0]); t++) {
        int n = sizes[t];
        memset(got, 0, sizeof(got));
        bf16_to_f32_buf(got, src, (size_t)n);
        for (int i = 0; i < n; i++) {
            float want = bf16_to_f32(src[i]);
            uint32_t got_bits, want_bits;
            memcpy(&got_bits, &got[i], sizeof(got_bits));
            memcpy(&want_bits, &want, sizeof(want_bits));
            if (got_bits != want_bits) {
                fprintf(stderr, "bf16_widen n=%d at %d: %.9g want %.9g\n", n, i, got[i], want);
                failures++;
            }
        }
    }
}

static void check_bf16_pair(int n_threads) {
    enum { seq = 4, in_dim = 5, a_dim = 3, b_dim = 7 };
    float x[seq * in_dim];
    float wa_f32[a_dim * in_dim];
    float wb_f32[b_dim * in_dim];
    uint16_t wa[a_dim * in_dim];
    uint16_t wb[b_dim * in_dim];
    float got_a[seq * a_dim], got_b[seq * b_dim];

    for (int i = 0; i < seq * in_dim; i++)
        x[i] = cosf((float)i * 0.19f) - 0.3f;
    for (int i = 0; i < a_dim * in_dim; i++) {
        wa_f32[i] = sinf((float)i * 0.13f) + 0.2f;
        wa[i] = f32_to_bf16(wa_f32[i]);
    }
    for (int i = 0; i < b_dim * in_dim; i++) {
        wb_f32[i] = cosf((float)i * 0.21f) - 0.1f;
        wb[i] = f32_to_bf16(wb_f32[i]);
    }

    tp_set_threads(n_threads);
    linear_nobias_bf16_pair(got_a, got_b, x, wa, wb, seq, in_dim, a_dim, b_dim);

    for (int s = 0; s < seq; s++) {
        for (int o = 0; o < a_dim; o++) {
            float want = 0.0f;
            for (int i = 0; i < in_dim; i++)
                want += x[s * in_dim + i] * bf16_to_f32(wa[o * in_dim + i]);
            expect_close("bf16_pair_a", got_a[s * a_dim + o], want, 1e-6f);
        }
        for (int o = 0; o < b_dim; o++) {
            float want = 0.0f;
            for (int i = 0; i < in_dim; i++)
                want += x[s * in_dim + i] * bf16_to_f32(wb[o * in_dim + i]);
            expect_close("bf16_pair_b", got_b[s * b_dim + o], want, 1e-6f);
        }
    }
}

static void test_bf16_pair(void) {
    check_bf16_pair(1);
    check_bf16_pair(4);
    tp_set_threads(1);
}

/* The dispatch macros in impl.h pick the SIMD variant at
 * compile time, so on arm64/x86 the generic C kernels are linked but never
 * called. Run them directly against the selected implementation on
 * remainder-lane-hostile sizes; on a plain build impl == generic and the
 * comparison is trivially exact. */
static void test_generic_vs_impl(void) {
    enum { IN = 67, OUT = 9, NMAX = 1027 };
    static float x[IN], bias[OUT], a[NMAX], b[NMAX], d0[NMAX], d1[NMAX];
    static uint16_t W[(size_t)OUT * IN];
    uint32_t s = 0x12345678u;
    for (int i = 0; i < NMAX; i++) {
        s = s * 1664525u + 1013904223u;
        a[i] = (float)(int32_t)s * 1e-9f;
        s = s * 1664525u + 1013904223u;
        b[i] = (float)(int32_t)s * 1e-9f;
    }
    for (int i = 0; i < IN; i++)
        x[i] = a[i];
    for (int o = 0; o < OUT; o++)
        bias[o] = b[o];
    for (size_t i = 0; i < sizeof(W) / sizeof(W[0]); i++) {
        s = s * 1664525u + 1013904223u;
        W[i] = f32_to_bf16((float)(int32_t)s * 1e-9f);
    }

    float yg[OUT], yi[OUT];
    bf16_matvec_fused_generic(yg, x, W, bias, IN, OUT);
    bf16_matvec_fused_impl(yi, x, W, bias, IN, OUT);
    for (int o = 0; o < OUT; o++)
        expect_close("matvec_fused generic-vs-impl", yg[o], yi[o], 1e-5f);
    bf16_matvec_fused_generic(yg, x, W, NULL, IN, OUT);
    bf16_matvec_fused_impl(yi, x, W, NULL, IN, OUT);
    for (int o = 0; o < OUT; o++)
        expect_close("matvec_fused nobias generic-vs-impl", yg[o], yi[o], 1e-5f);

    const int sizes[] = {1, 7, 64, NMAX};
    for (size_t t = 0; t < sizeof(sizes) / sizeof(sizes[0]); t++) {
        int n = sizes[t];
        expect_close("dot generic-vs-impl", dot_f32_generic(a, b, n), dot_f32_impl(a, b, n), 1e-5f);

        memcpy(d0, a, (size_t)n * sizeof(float));
        memcpy(d1, a, (size_t)n * sizeof(float));
        vec_scale_inplace_generic(d0, 1.37f, n);
        vec_scale_inplace_impl(d1, 1.37f, n);
        for (int i = 0; i < n; i++)
            expect_close("scale generic-vs-impl", d0[i], d1[i], 1e-6f);

        memcpy(d0, a, (size_t)n * sizeof(float));
        memcpy(d1, a, (size_t)n * sizeof(float));
        vec_axpy_inplace_generic(d0, b, -0.61f, n);
        vec_axpy_inplace_impl(d1, b, -0.61f, n);
        for (int i = 0; i < n; i++)
            expect_close("axpy generic-vs-impl", d0[i], d1[i], 1e-6f);

        memcpy(d0, a, (size_t)n * sizeof(float));
        memcpy(d1, a, (size_t)n * sizeof(float));
        vec_scale_add_generic(d0, b, 0.83f, n);
        vec_scale_add_impl(d1, b, 0.83f, n);
        for (int i = 0; i < n; i++)
            expect_close("scale_add generic-vs-impl", d0[i], d1[i], 1e-6f);
    }
}

int main(void) {
    tp_set_threads(1);
    test_rms_norm();
    test_layer_norm();
    test_gelu();
    test_gelu_tanh();
    test_rope_neox();
    test_packed_gqa_attention();
    test_packed_gqa_attention_long();
    test_packed_causal_gqa_attention();
    test_bf16_linear();
    test_bf16_qkv();
    test_bf16_matvec();
    test_rms_norm_threaded();
    test_bf16_widen_buf();
    test_bf16_pair();
    test_generic_vs_impl();
    if (failures != 0)
        return 1;
    puts("ok: kernel golden tests passed");
    return 0;
}