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
|
/*
* Architecture-generic hot kernels / non-SIMD fallback.
*/
#include "impl.h"
#include <string.h>
void bf16_matvec_fused_generic(
float *y, const float *x, const uint16_t *W_bf16, const float *bias, int in_dim, int out_dim) {
for (int o = 0; o < out_dim; o++) {
const uint16_t *w_row = W_bf16 + (size_t)o * in_dim;
float sum = bias ? bias[o] : 0.0f;
for (int k = 0; k < in_dim; k++) {
uint32_t f32_bits = ((uint32_t)w_row[k]) << 16;
float w_val;
memcpy(&w_val, &f32_bits, sizeof(float));
sum += w_val * x[k];
}
y[o] = sum;
}
}
float dot_f32_generic(const float *a, const float *b, int n) {
float sum = 0.0f;
for (int i = 0; i < n; i++)
sum += a[i] * b[i];
return sum;
}
void vec_scale_inplace_generic(float *dst, float scale, int n) {
for (int i = 0; i < n; i++)
dst[i] *= scale;
}
void vec_axpy_inplace_generic(float *dst, const float *src, float alpha, int n) {
for (int i = 0; i < n; i++)
dst[i] += alpha * src[i];
}
void vec_scale_add_generic(float *dst, const float *src, float correction, int n) {
for (int i = 0; i < n; i++)
dst[i] = dst[i] * correction + src[i];
}
|