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
|
/*
* embed_kernels.h - Math kernels for Qwen3-architecture inference
*
* Low-level math operations. All operate on float32 tensors in row-major order.
*/
#ifndef EMBED_KERNELS_H
#define EMBED_KERNELS_H
/* Public-API export annotation. The libraries are built with
* -fvisibility=hidden, so only declarations carrying EMBED_API are exported
* from the shared library; everything else stays internal. */
#ifndef EMBED_API
#if defined(__GNUC__)
#define EMBED_API __attribute__((visibility("default")))
#else
#define EMBED_API
#endif
#endif
#include <stddef.h>
#include <stdint.h>
/* ========================================================================
* Basic Operations
* ======================================================================== */
void embed_add_inplace(float *a, const float *b, int n);
/* ========================================================================
* Matrix Operations
* ======================================================================== */
/* C = A @ B^T: A[M,K], B[N,K], C[M,N] */
void embed_matmul_t(float *C, const float *A, const float *B, int M, int K, int N);
/* y = x @ W^T + b: x[seq,in], W[out,in], b[out], y[seq,out] */
void embed_linear(
float *y, const float *x, const float *W, const float *b, int seq_len, int in_dim, int out_dim);
void embed_linear_nobias(
float *y, const float *x, const float *W, int seq_len, int in_dim, int out_dim);
/* bf16 weight variants */
void embed_bf16_to_f32_buf(float *dst, const uint16_t *src, size_t n);
void embed_linear_nobias_bf16(
float *y, const float *x, const uint16_t *W_bf16, int seq_len, int in_dim, int out_dim);
/* Small-sequence BF16 path: compute two projections with one threaded dispatch. */
void embed_linear_nobias_bf16_pair(float *a,
float *b,
const float *x,
const uint16_t *Wa_bf16,
const uint16_t *Wb_bf16,
int seq_len,
int in_dim,
int a_dim,
int b_dim);
/* Small-sequence BF16 path: compute Q/K/V matvecs with one threaded dispatch. */
void embed_linear_nobias_bf16_qkv(float *q,
float *k,
float *v,
const float *x,
const uint16_t *Wq_bf16,
const uint16_t *Wk_bf16,
const uint16_t *Wv_bf16,
int seq_len,
int in_dim,
int q_dim,
int kv_dim);
/* Dot product using the best available local SIMD implementation. */
float embed_dot_f32(const float *a, const float *b, int n);
/* ========================================================================
* Normalization
* ======================================================================== */
/* RMS Normalization: out = x / rms(x) * weight */
void embed_rms_norm(
float *out, const float *x, const float *weight, int seq_len, int hidden, float eps);
/* Per-head RMS Normalization for Q/K norms in decoder
* x: [seq, n_heads, head_dim], weight: [head_dim]
* Normalizes each head independently */
void embed_rms_norm_per_head(
float *x, const float *weight, int seq_len, int n_heads, int head_dim, float eps);
/* Layer Normalization (BERT family): out = gamma * (x - mean) / sqrt(var + eps)
* + beta, with mean and biased (population) variance over the hidden axis per
* row. Unlike RMSNorm this subtracts the mean and adds a bias. */
void embed_layer_norm(float *out,
const float *x,
const float *gamma,
const float *beta,
int seq_len,
int hidden,
float eps);
/* ========================================================================
* Activation Functions
* ======================================================================== */
void embed_softmax(float *x, int rows, int cols);
/* gate = SiLU(gate) * up */
void embed_silu_mul_inplace(float *gate, const float *up, int n);
/* Exact (erf) GeLU in place: x = 0.5 * x * (1 + erf(x / sqrt(2))). This is the
* "gelu" activation HF BERT/BGE/MiniLM use, not the tanh approximation. */
void embed_gelu_inplace(float *x, int n);
/* Tanh-approximation GeLU in place (HF gelu_new / gelu_pytorch_tanh):
* x = 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3))). */
void embed_gelu_tanh_inplace(float *x, int n);
/* ========================================================================
* Attention Operations
* ======================================================================== */
/*
* Bidirectional GQA attention over a packed/ragged batch.
*
* Q: [total_seq, n_heads * head_dim]
* K,V: [total_seq, n_kv_heads * head_dim]
* offsets: [batch + 1], token spans for each sequence.
*
* Attention is block-diagonal: tokens only attend within their own
* offsets[b]..offsets[b+1] span.
*/
void embed_bidirectional_gqa_attention_packed(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);
/* Optional reusable scratch for the BLAS-tiled packed attention path. */
size_t embed_bidirectional_gqa_attention_packed_scratch_bytes(const int *offsets, int batch);
void embed_bidirectional_gqa_attention_packed_with_scratch(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,
float *scratch,
size_t scratch_bytes);
/*
* Causal GQA attention over a packed/ragged batch. Query position i attends
* only to positions at or before i within its sequence.
*/
void embed_causal_gqa_attention_packed(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);
void embed_causal_gqa_attention_packed_with_scratch(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,
float *scratch,
size_t scratch_bytes);
/* ========================================================================
* Position Embeddings
* ======================================================================== */
/*
* NeoX-style RoPE: compute cos/sin for positions.
* cos_out, sin_out: [seq, head_dim]
* cos[d] and cos[half+d] are the same (duplicated for full head_dim).
*/
void embed_compute_rope_neox(
float *cos_out, float *sin_out, const int *positions, int seq, int head_dim, float theta);
/*
* Apply NeoX-style RoPE to Q or K.
* x: [seq, n_heads * head_dim] (in-place)
* cos_vals, sin_vals: [seq, head_dim]
*/
void embed_apply_rope_neox(
float *x, const float *cos_vals, const float *sin_vals, int seq, int n_heads, int head_dim);
/* ========================================================================
* Threading
* ======================================================================== */
/* The persistent worker pool lives in threadpool.{c,h}; included here so kernel
* callers keep getting embed_set_threads / embed_get_num_cpus from one header. */
#include "threadpool.h"
/* Global verbose flag */
EMBED_API extern int embed_verbose;
#endif /* EMBED_KERNELS_H */
|