diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-20 20:10:56 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-21 05:57:19 +0200 |
| commit | 6b969d25ce136eaf695222bd1a10bff1ecac0220 (patch) | |
| tree | c23afb50c28b9deaddae5c6526a74ee11cf2d865 /libffwd/generic.c | |
| parent | d4bdb8dae753462f621507d5fc3582e0e53da35c (diff) | |
| download | ffwd-6b969d25ce136eaf695222bd1a10bff1ecac0220.tar.gz ffwd-6b969d25ce136eaf695222bd1a10bff1ecac0220.zip | |
all: split Makefile and reorg library/tools
Diffstat (limited to 'libffwd/generic.c')
| -rw-r--r-- | libffwd/generic.c | 44 |
1 files changed, 44 insertions, 0 deletions
diff --git a/libffwd/generic.c b/libffwd/generic.c new file mode 100644 index 0000000..8798107 --- /dev/null +++ b/libffwd/generic.c @@ -0,0 +1,44 @@ +/* + * generic.c - 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]; +} |
