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
|
/* tests/test_kernels.c - golden tests for the low-level math kernels.
* Runs via `make test` (generic and BLAS variants) and
* scripts/check_kernel_golden.py. No model files required. */
#include "qwen_kernels.h"
#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
static int failures = 0;
int qwen_verbose = 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];
qwen_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_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));
qwen_compute_rope_neox(cosv, sinv, positions, seq, head_dim, 10000.0f);
qwen_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 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++) {
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 < 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 < 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);
}
qwen_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));
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);
}
qwen_set_threads(4);
qwen_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));
qwen_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_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]);
qwen_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]);
}
qwen_set_threads(n_threads);
qwen_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);
qwen_set_threads(1);
}
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]);
}
qwen_set_threads(n_threads);
qwen_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);
qwen_set_threads(1);
}
int main(void) {
qwen_set_threads(1);
test_rms_norm();
test_rope_neox();
test_packed_gqa_attention();
test_packed_gqa_attention_long();
test_bf16_linear();
test_bf16_qkv();
test_bf16_pair();
if (failures != 0) return 1;
puts("ok: kernel golden tests passed");
return 0;
}
|