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
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
// SPDX-License-Identifier: Apache-2.0
// Copyright (c) 2026 Fábio Henrique de Lima Silva (fhl.bsb@gmail.com) All rights reserved.
macro_rules! impl_avx512_gemv {
() => {
#[inline(always)]
// SAFETY: a and b are valid slices; CPU supports AVX-512F+VL+F16C (verified by dispatch).
// Kernel uses min(a.len(), b.len()) as bound with unaligned 512-bit loads.
unsafe fn dot_product(a: &[f32], b: &[f32]) -> f32 {
crate::math::gemm::dot_basic::dot_product_avx512(a, b)
}
#[inline(always)]
// SAFETY: a and b are valid u16 slices; no AVX-512 requirement (fallback path).
unsafe fn dot_product_bf16(a: &[u16], b: &[u16]) -> f32 {
dot_product_bf16_fallback(a, b)
}
#[inline(always)]
// SAFETY: weights is a valid &[[u16; 4]] slice, state is a valid f32 slice;
// CPU supports AVX-512F+VL (verified by dispatch). Kernel uses unaligned loads.
unsafe fn dot_product_4x_interleaved(weights: &[[u16; 4]], state: &[f32]) -> [f32; 4] {
crate::math::gemm::dot_4x::dot_product_4x_interleaved_avx512(weights, state)
}
#[inline(always)]
// SAFETY: weights, state_f0, and state_f1 are valid slices;
// state_f0.len() == state_f1.len() == weights.len();
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_4x_interleaved_dual_frame(
weights: &[[u16; 4]],
state_f0: &[f32],
state_f1: &[f32],
) -> ([f32; 4], [f32; 4]) {
crate::math::gemm::dot_4x::dot_product_4x_interleaved_dual_frame_avx512(
weights, state_f0, state_f1,
)
}
#[inline(always)]
// SAFETY: weights and state are valid slices; weights.len() >= state.len();
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_4x_f32(weights: &[[f32; 4]], state: &[f32]) -> [f32; 4] {
crate::math::gemm::dot_4x::dot_product_4x_f32_avx512(weights, state)
}
#[inline(always)]
// SAFETY: weights, state_f0, and state_f1 are valid slices;
// state_f0.len() == state_f1.len() == weights.len();
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_4x_f32_dual(
weights: &[[f32; 4]],
state_f0: &[f32],
state_f1: &[f32],
) -> ([f32; 4], [f32; 4]) {
crate::math::gemm::dot_4x::dot_product_4x_f32_dual_avx512(weights, state_f0, state_f1)
}
#[inline(always)]
// SAFETY: weights and state are valid slices; weights.len() >= state.len();
// CPU supports AVX-512F (verified by dispatch). AVX-512 implies AVX2/FMA,
// so the 8x AVX2 kernel is valid to call from this context.
unsafe fn dot_product_8x_f32(weights: &[[f32; 8]], state: &[f32]) -> [f32; 8] {
crate::math::gemm::dot_8x::dot_product_8x_f32_avx2(weights, state)
}
#[inline(always)]
// SAFETY: weights, state_f0, state_f1 are valid slices;
// weights.len() >= state_f0.len() and weights.len() >= state_f1.len();
// CPU supports AVX-512F (verified by dispatch). AVX-512 implies AVX2/FMA,
// so the 8x dual-frame AVX2 kernel is valid to call from this context.
unsafe fn dot_product_8x_f32_dual(
weights: &[[f32; 8]],
state_f0: &[f32],
state_f1: &[f32],
) -> ([f32; 8], [f32; 8]) {
crate::math::gemm::dot_8x::dot_product_8x_f32_dual_avx2(weights, state_f0, state_f1)
}
#[inline(always)]
// SAFETY: weights and state are valid slices; weights.len() >= state.len();
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_16x_f32(weights: &[[f32; 16]], state: &[f32]) -> [f32; 16] {
crate::math::gemm::dot_16x::dot_product_16x_f32_avx512(weights, state)
}
#[inline(always)]
// SAFETY: weights, state_f0, state_f1 are valid slices;
// weights.len() >= state_f0.len() and weights.len() >= state_f1.len();
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_16x_f32_dual(
weights: &[[f32; 16]],
state_f0: &[f32],
state_f1: &[f32],
) -> ([f32; 16], [f32; 16]) {
crate::math::gemm::dot_16x::dot_product_16x_f32_dual_avx512(weights, state_f0, state_f1)
}
#[inline(always)]
// SAFETY: weights, state, init are valid slices; weights.len() >= state.len();
// CPU supports AVX-512F+VL (verified by dispatch).
// Fused kernel: init is loaded into the first accumulator register.
unsafe fn dot_product_4x_f32_accumulate(
weights: &[[f32; 4]],
state: &[f32],
init: &[f32; 4],
) -> [f32; 4] {
crate::math::gemm::dot_4x::dot_product_4x_f32_accumulate_avx512(weights, state, init)
}
#[inline(always)]
// SAFETY: weights, state_f0, state_f1, init_f0, init_f1 are valid slices;
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_4x_f32_dual_accumulate(
weights: &[[f32; 4]],
state_f0: &[f32],
state_f1: &[f32],
init_f0: &[f32; 4],
init_f1: &[f32; 4],
) -> ([f32; 4], [f32; 4]) {
crate::math::gemm::dot_4x::dot_product_4x_f32_dual_accumulate_avx512(
weights, state_f0, state_f1, init_f0, init_f1,
)
}
#[inline(always)]
// SAFETY: weights, state, init are valid slices; weights.len() >= state.len();
// AVX-512 implies AVX2/FMA — delegates to the fused AVX2 8x accumulate kernel.
unsafe fn dot_product_8x_f32_accumulate(
weights: &[[f32; 8]],
state: &[f32],
init: &[f32; 8],
) -> [f32; 8] {
crate::math::gemm::dot_8x::dot_product_8x_f32_accumulate_avx2(weights, state, init)
}
#[inline(always)]
// SAFETY: weights, state_f0, state_f1, init_f0, init_f1 are valid slices;
// AVX-512 implies AVX2/FMA — delegates to the fused AVX2 8x dual accumulate kernel.
unsafe fn dot_product_8x_f32_dual_accumulate(
weights: &[[f32; 8]],
state_f0: &[f32],
state_f1: &[f32],
init_f0: &[f32; 8],
init_f1: &[f32; 8],
) -> ([f32; 8], [f32; 8]) {
crate::math::gemm::dot_8x::dot_product_8x_f32_dual_accumulate_avx2(
weights, state_f0, state_f1, init_f0, init_f1,
)
}
#[inline(always)]
// SAFETY: weights, state, init are valid slices; weights.len() >= state.len();
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_16x_f32_accumulate(
weights: &[[f32; 16]],
state: &[f32],
init: &[f32; 16],
) -> [f32; 16] {
crate::math::gemm::dot_16x::dot_product_16x_f32_accumulate_avx512(weights, state, init)
}
#[inline(always)]
// SAFETY: weights, state_f0, state_f1, init_f0, init_f1 are valid slices;
// CPU supports AVX-512F+VL (verified by dispatch).
unsafe fn dot_product_16x_f32_dual_accumulate(
weights: &[[f32; 16]],
state_f0: &[f32],
state_f1: &[f32],
init_f0: &[f32; 16],
init_f1: &[f32; 16],
) -> ([f32; 16], [f32; 16]) {
crate::math::gemm::dot_16x::dot_product_16x_f32_dual_accumulate_avx512(
weights, state_f0, state_f1, init_f0, init_f1,
)
}
#[inline(always)]
// SAFETY: four weight slices and in_frame are valid u16 slices of equal length;
// fallback path, no AVX-512 requirement.
unsafe fn dot_product_bf16_4x(
w0: &[u16],
w1: &[u16],
w2: &[u16],
w3: &[u16],
in_frame: &[u16],
) -> [f32; 4] {
dot_product_bf16_4x_fallback(w0, w1, w2, w3, in_frame)
}
#[inline(always)]
// SAFETY: in_frame, weights, bias, out_frame are valid slices with
// weights.len() >= in_frame.len() * out_frame.len(), bias.len() >= out_frame.len();
// CPU supports AVX-512F+VL+F16C (verified by dispatch).
unsafe fn fused_add_gemv(
in_frame: &[f32],
weights: &[f32],
bias: &[f32],
out_frame: &mut [f32],
do_bias: bool,
) {
if out_frame.len() == 16 {
crate::math::gemm::gemv::fused_add_gemv_avx512_small(
in_frame, weights, bias, out_frame, do_bias,
)
} else {
fused_add_gemv_fallback(in_frame, weights, bias, out_frame, do_bias)
}
}
#[inline(always)]
// SAFETY: in_frames, weights, bias, out_frames are valid slices with lengths matching
// num_frames * dimensions; CPU supports AVX-512F+VL+F16C (verified by dispatch).
unsafe fn fused_add_gemm_batch(
in_frames: &[f32],
weights: &[f32],
bias: &[f32],
out_frames: &mut [f32],
num_frames: usize,
do_bias: bool,
) {
// SAFETY: all slice arguments satisfy the function's documented invariants.
unsafe {
crate::math::gemm::gemm_batch::fused_add_gemm_batch_avx512(
in_frames, weights, bias, out_frames, num_frames, do_bias,
)
}
}
#[inline(always)]
// SAFETY: in_frames, weights, bias, residual, out_frames are valid slices with
// lengths matching num_frames * dimensions; CPU supports AVX-512F+VL+F16C.
unsafe fn fused_gemm_residual_batch(
in_frames: &[f32],
weights: &[f32],
bias: &[f32],
residual: &[f32],
out_frames: &mut [f32],
num_frames: usize,
do_bias: bool,
) {
// SAFETY: all slice arguments satisfy the function's documented invariants.
unsafe {
crate::math::gemm::gemm_batch::fused_gemm_residual_batch_avx512(
in_frames, weights, bias, residual, out_frames, num_frames, do_bias,
)
}
}
#[inline(always)]
// SAFETY: in_frames, weights, bias, residual, out_frames are valid f32 slices with
// lengths matching num_frames * dimensions; CPU supports AVX-512F+VL.
unsafe fn fused_gemm_residual_batch_f32(
in_frames: &[f32],
weights: &[f32],
bias: &[f32],
residual: &[f32],
out_frames: &mut [f32],
num_frames: usize,
do_bias: bool,
) {
// SAFETY: all slice arguments satisfy the function's documented invariants.
unsafe {
crate::math::gemm::gemm_batch::fused_gemm_residual_batch_f32_avx512(
in_frames, weights, bias, residual, out_frames, num_frames, do_bias,
)
}
}
#[inline(always)]
// SAFETY: in_frame, weights, bias, out_frame are valid slices with
// weights.len() >= in_frame.len() * out_frame.len(); CPU supports AVX-512F+VL+F16C.
unsafe fn gemv_overwrite(
in_frame: &[f32],
weights: &[f32],
bias: &[f32],
out_frame: &mut [f32],
do_bias: bool,
) {
if out_frame.len() == 16 {
crate::math::gemm::gemv::gemv_overwrite_avx512_small(
in_frame, weights, bias, out_frame, do_bias,
)
} else {
gemv_overwrite_fallback(in_frame, weights, bias, out_frame, do_bias)
}
}
#[inline(always)]
// SAFETY: in_frame (u16 quantized), weights (u16), bias (f32), out_frame (f32) are
// valid slices with matching dimensions; CPU supports AVX-512F (fallback path).
unsafe fn gemv_overwrite_bf16(
in_frame: &[u16],
weights: &[u16],
bias: &[f32],
out_frame: &mut [f32],
do_bias: bool,
) {
gemv_overwrite_bf16_fallback(in_frame, weights, bias, out_frame, do_bias)
}
#[inline(always)]
// SAFETY: in_frame, weights (4*gate concatenated), bias, out_gates are valid slices
// with weights.len() == 4 * in_frame.len() * hidden_size,
// out_gates.len() == 4 * hidden_size, bias.len() == 4 * hidden_size;
// CPU supports AVX-512F+VL+F16C (verified by dispatch).
unsafe fn gemv_overwrite_4gate(
_in_frame: &[f32],
_weights: &[u16],
_bias: &[f32],
_out_gates: &mut [f32],
_hidden_size: usize,
_do_bias: bool,
) {
unreachable!(
"gemv_overwrite_4gate is unused; 4-gate dispatch uses direct kernel functions"
);
}
#[inline(always)]
// SAFETY: in_frame (u16 BF16), weights (u16 BF16), bias (f32), out_gates (f32) are
// valid slices with weights.len() == 4 * in_frame.len() * hidden_size,
// out_gates.len() == 4 * hidden_size; CPU supports AVX-512 VNNI+BF16 (verified by dispatch).
unsafe fn gemv_overwrite_bf16_4gate(
in_frame: &[u16],
weights: &[u16],
bias: &[f32],
out_gates: &mut [f32],
hidden_size: usize,
do_bias: bool,
) {
let ih = in_frame.len();
let stride = ih * hidden_size;
// SAFETY: weights indices [0..4*stride] are within bounds per function invariants;
// AVX-512 VNNI+BF16 ISA verified by caller via dispatch.
unsafe {
crate::math::gemm::gemv_4gate::gemv_4gate_bf16_avx512(
in_frame,
&weights[0..stride],
&weights[stride..2 * stride],
&weights[2 * stride..3 * stride],
&weights[3 * stride..4 * stride],
bias,
out_gates,
do_bias,
)
}
}
#[inline(always)]
// SAFETY: in_frames, weights, bias, out_frames are valid slices with lengths
// matching num_frames * dimensions; CPU supports AVX-512F+VL+F16C.
unsafe fn gemv_overwrite_batch(
in_frames: &[f32],
weights: &[f32],
bias: &[f32],
out_frames: &mut [f32],
num_frames: usize,
do_bias: bool,
) {
// SAFETY: all slice arguments satisfy the function's documented invariants.
unsafe {
crate::math::gemm::gemv::gemv_overwrite_batch_avx512(
in_frames, weights, bias, out_frames, num_frames, do_bias,
)
}
}
#[inline(always)]
// SAFETY: in_frames, weights (f32), bias, out_frames are valid slices with
// lengths matching num_frames * dimensions; CPU supports AVX-512F+VL.
unsafe fn gemv_with_bias_f32(
in_frames: &[f32],
weights: &[f32],
bias: &[f32],
out_frames: &mut [f32],
num_frames: usize,
) {
// SAFETY: all slice arguments satisfy the function's documented invariants.
unsafe {
crate::math::gemm::gemv::gemv_with_bias_f32_avx512(
in_frames, weights, bias, out_frames, num_frames,
)
}
}
#[inline(always)]
// SAFETY: in_frames, weights (f32), out_frames are valid slices with
// lengths matching num_frames * dimensions; CPU supports AVX-512F+VL.
unsafe fn gemv_no_bias_f32(
in_frames: &[f32],
weights: &[f32],
out_frames: &mut [f32],
num_frames: usize,
) {
// SAFETY: all slice arguments satisfy the function's documented invariants.
unsafe {
crate::math::gemm::gemv::gemv_no_bias_f32_avx512(
in_frames, weights, out_frames, num_frames,
)
}
}
};
}