aboutsummaryrefslogtreecommitdiff
path: root/qwen_kernels_impl.h
blob: 43e4fbec25611c0a0e0d8e1424cc264ff31ac71a (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
/*
 * 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 */