diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-16 16:27:47 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-16 16:27:47 +0200 |
| commit | d1ba1ebc06ef0d62ca932388fae83e6992594719 (patch) | |
| tree | f962207b779016593258f990348216999a58766a /qwen_kernels_impl.h | |
| parent | be20dfda0f836808dccfc14db244592e41ba28bc (diff) | |
| download | ffwd-d1ba1ebc06ef0d62ca932388fae83e6992594719.tar.gz ffwd-d1ba1ebc06ef0d62ca932388fae83e6992594719.zip | |
chore: move sources to src/ with proper names, drop qwen_ prefix
Diffstat (limited to 'qwen_kernels_impl.h')
| -rw-r--r-- | qwen_kernels_impl.h | 53 |
1 files changed, 0 insertions, 53 deletions
diff --git a/qwen_kernels_impl.h b/qwen_kernels_impl.h deleted file mode 100644 index 217f1e8..0000000 --- a/qwen_kernels_impl.h +++ /dev/null @@ -1,53 +0,0 @@ -/* - * qwen_kernels_impl.h - internal architecture dispatch for hot kernels - */ - -#ifndef QWEN_KERNELS_IMPL_H -#define QWEN_KERNELS_IMPL_H - -#include <stdint.h> - -void qwen_bf16_matvec_fused_generic( - float *y, const float *x, const uint16_t *W_bf16, const float *bias, int in_dim, int out_dim); -float qwen_dot_f32_generic(const float *a, const float *b, int n); -void qwen_vec_scale_inplace_generic(float *dst, float scale, int n); -void qwen_vec_axpy_inplace_generic(float *dst, const float *src, float alpha, int n); -void qwen_vec_scale_add_generic(float *dst, const float *src, float correction, int n); - -#ifdef __ARM_NEON -void qwen_bf16_matvec_fused_neon( - float *y, const float *x, const uint16_t *W_bf16, const float *bias, int in_dim, int out_dim); -float qwen_dot_f32_neon(const float *a, const float *b, int n); -void qwen_vec_scale_inplace_neon(float *dst, float scale, int n); -void qwen_vec_axpy_inplace_neon(float *dst, const float *src, float alpha, int n); -void qwen_vec_scale_add_neon(float *dst, const float *src, float correction, int n); - -#define qwen_bf16_matvec_fused_impl qwen_bf16_matvec_fused_neon -#define qwen_dot_f32_impl qwen_dot_f32_neon -#define qwen_vec_scale_inplace_impl qwen_vec_scale_inplace_neon -#define qwen_vec_axpy_inplace_impl qwen_vec_axpy_inplace_neon -#define qwen_vec_scale_add_impl qwen_vec_scale_add_neon - -#elif defined(__AVX2__) && defined(__FMA__) -void qwen_bf16_matvec_fused_avx( - float *y, const float *x, const uint16_t *W_bf16, const float *bias, int in_dim, int out_dim); -float qwen_dot_f32_avx(const float *a, const float *b, int n); -void qwen_vec_scale_inplace_avx(float *dst, float scale, int n); -void qwen_vec_axpy_inplace_avx(float *dst, const float *src, float alpha, int n); -void qwen_vec_scale_add_avx(float *dst, const float *src, float correction, int n); - -#define qwen_bf16_matvec_fused_impl qwen_bf16_matvec_fused_avx -#define qwen_dot_f32_impl qwen_dot_f32_avx -#define qwen_vec_scale_inplace_impl qwen_vec_scale_inplace_avx -#define qwen_vec_axpy_inplace_impl qwen_vec_axpy_inplace_avx -#define qwen_vec_scale_add_impl qwen_vec_scale_add_avx - -#else -#define qwen_bf16_matvec_fused_impl qwen_bf16_matvec_fused_generic -#define qwen_dot_f32_impl qwen_dot_f32_generic -#define qwen_vec_scale_inplace_impl qwen_vec_scale_inplace_generic -#define qwen_vec_axpy_inplace_impl qwen_vec_axpy_inplace_generic -#define qwen_vec_scale_add_impl qwen_vec_scale_add_generic -#endif - -#endif /* QWEN_KERNELS_IMPL_H */ |
