lattice-inference 0.2.2

Pure Rust transformer inference engine — safetensors loading, SIMD matmul, BGE/Qwen3 embeddings
Documentation 
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

// ---------------------------------------------------------------------------
// AVX2+FMA tiled matmul_bt — cache-blocked with 4×8 AVX2 microkernel
// ---------------------------------------------------------------------------
//
// Same tile structure as NEON (4i × 8j), but uses 256-bit AVX2 lanes (8 f32s).
// Inner k-loop unrolled 2× (processes 16 k-values per iteration).
//

#[cfg(target_arch = "x86_64")]
use super::arch_kernels::hsum_m256;
#[cfg(target_arch = "x86_64")]
use super::tiled::{TILE_I, TILE_J, TILE_K};

#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2", enable = "fma")]
pub(super) unsafe fn matmul_bt_tiled_avx2(
    a: &[f32],
    b: &[f32],
    c: &mut [f32],
    m: usize,
    k: usize,
    n: usize,
) {
    use std::arch::x86_64::*;

    let a_ptr = a.as_ptr();
    let b_ptr = b.as_ptr();
    let c_ptr = c.as_mut_ptr();

    let mut j_start = 0;
    while j_start < n {
        let j_end = (j_start + TILE_J).min(n);
        let j_count = j_end - j_start;

        let mut i_start = 0;
        while i_start < m {
            let i_end = (i_start + TILE_I).min(m);
            let i_count = i_end - i_start;

            let mut k_start = 0;
            while k_start < k {
                let k_end = (k_start + TILE_K).min(k);
                let k_len = k_end - k_start;

                // Fast path: full 4×8 tile with k_len >= 16 (AVX2 2× unrolled)
                if i_count == TILE_I && j_count == TILE_J && k_len >= 16 {
                    // SAFETY: Same bounds reasoning as NEON path.
                    // AVX2 loads read 8 contiguous f32s; all offsets stay within bounds.
                    let mut acc0 = [_mm256_setzero_ps(); TILE_J];
                    let mut acc1 = [_mm256_setzero_ps(); TILE_J];
                    let mut acc2 = [_mm256_setzero_ps(); TILE_J];
                    let mut acc3 = [_mm256_setzero_ps(); TILE_J];

                    let a0_base = a_ptr.add(i_start * k);
                    let a1_base = a_ptr.add((i_start + 1) * k);
                    let a2_base = a_ptr.add((i_start + 2) * k);
                    let a3_base = a_ptr.add((i_start + 3) * k);

                    let b0_base = b_ptr.add(j_start * k);
                    let b1_base = b_ptr.add((j_start + 1) * k);
                    let b2_base = b_ptr.add((j_start + 2) * k);
                    let b3_base = b_ptr.add((j_start + 3) * k);
                    let b4_base = b_ptr.add((j_start + 4) * k);
                    let b5_base = b_ptr.add((j_start + 5) * k);
                    let b6_base = b_ptr.add((j_start + 6) * k);
                    let b7_base = b_ptr.add((j_start + 7) * k);

                    // Main loop: 16 k-values per iteration (2× unrolled 8-wide AVX2)
                    let k_pairs = k_len / 16;
                    for kp in 0..k_pairs {
                        let ko = k_start + kp * 16;

                        // --- First group of 8 k-values ---
                        let bv0a = _mm256_loadu_ps(b0_base.add(ko));
                        let bv1a = _mm256_loadu_ps(b1_base.add(ko));
                        let bv2a = _mm256_loadu_ps(b2_base.add(ko));
                        let bv3a = _mm256_loadu_ps(b3_base.add(ko));
                        let bv4a = _mm256_loadu_ps(b4_base.add(ko));
                        let bv5a = _mm256_loadu_ps(b5_base.add(ko));
                        let bv6a = _mm256_loadu_ps(b6_base.add(ko));
                        let bv7a = _mm256_loadu_ps(b7_base.add(ko));

                        let av = _mm256_loadu_ps(a0_base.add(ko));
                        acc0[0] = _mm256_fmadd_ps(av, bv0a, acc0[0]);
                        acc0[1] = _mm256_fmadd_ps(av, bv1a, acc0[1]);
                        acc0[2] = _mm256_fmadd_ps(av, bv2a, acc0[2]);
                        acc0[3] = _mm256_fmadd_ps(av, bv3a, acc0[3]);
                        acc0[4] = _mm256_fmadd_ps(av, bv4a, acc0[4]);
                        acc0[5] = _mm256_fmadd_ps(av, bv5a, acc0[5]);
                        acc0[6] = _mm256_fmadd_ps(av, bv6a, acc0[6]);
                        acc0[7] = _mm256_fmadd_ps(av, bv7a, acc0[7]);

                        let av = _mm256_loadu_ps(a1_base.add(ko));
                        acc1[0] = _mm256_fmadd_ps(av, bv0a, acc1[0]);
                        acc1[1] = _mm256_fmadd_ps(av, bv1a, acc1[1]);
                        acc1[2] = _mm256_fmadd_ps(av, bv2a, acc1[2]);
                        acc1[3] = _mm256_fmadd_ps(av, bv3a, acc1[3]);
                        acc1[4] = _mm256_fmadd_ps(av, bv4a, acc1[4]);
                        acc1[5] = _mm256_fmadd_ps(av, bv5a, acc1[5]);
                        acc1[6] = _mm256_fmadd_ps(av, bv6a, acc1[6]);
                        acc1[7] = _mm256_fmadd_ps(av, bv7a, acc1[7]);

                        let av = _mm256_loadu_ps(a2_base.add(ko));
                        acc2[0] = _mm256_fmadd_ps(av, bv0a, acc2[0]);
                        acc2[1] = _mm256_fmadd_ps(av, bv1a, acc2[1]);
                        acc2[2] = _mm256_fmadd_ps(av, bv2a, acc2[2]);
                        acc2[3] = _mm256_fmadd_ps(av, bv3a, acc2[3]);
                        acc2[4] = _mm256_fmadd_ps(av, bv4a, acc2[4]);
                        acc2[5] = _mm256_fmadd_ps(av, bv5a, acc2[5]);
                        acc2[6] = _mm256_fmadd_ps(av, bv6a, acc2[6]);
                        acc2[7] = _mm256_fmadd_ps(av, bv7a, acc2[7]);

                        let av = _mm256_loadu_ps(a3_base.add(ko));
                        acc3[0] = _mm256_fmadd_ps(av, bv0a, acc3[0]);
                        acc3[1] = _mm256_fmadd_ps(av, bv1a, acc3[1]);
                        acc3[2] = _mm256_fmadd_ps(av, bv2a, acc3[2]);
                        acc3[3] = _mm256_fmadd_ps(av, bv3a, acc3[3]);
                        acc3[4] = _mm256_fmadd_ps(av, bv4a, acc3[4]);
                        acc3[5] = _mm256_fmadd_ps(av, bv5a, acc3[5]);
                        acc3[6] = _mm256_fmadd_ps(av, bv6a, acc3[6]);
                        acc3[7] = _mm256_fmadd_ps(av, bv7a, acc3[7]);

                        // --- Second group of 8 k-values (unrolled) ---
                        let ko2 = ko + 8;
                        let bv0b = _mm256_loadu_ps(b0_base.add(ko2));
                        let bv1b = _mm256_loadu_ps(b1_base.add(ko2));
                        let bv2b = _mm256_loadu_ps(b2_base.add(ko2));
                        let bv3b = _mm256_loadu_ps(b3_base.add(ko2));
                        let bv4b = _mm256_loadu_ps(b4_base.add(ko2));
                        let bv5b = _mm256_loadu_ps(b5_base.add(ko2));
                        let bv6b = _mm256_loadu_ps(b6_base.add(ko2));
                        let bv7b = _mm256_loadu_ps(b7_base.add(ko2));

                        let av = _mm256_loadu_ps(a0_base.add(ko2));
                        acc0[0] = _mm256_fmadd_ps(av, bv0b, acc0[0]);
                        acc0[1] = _mm256_fmadd_ps(av, bv1b, acc0[1]);
                        acc0[2] = _mm256_fmadd_ps(av, bv2b, acc0[2]);
                        acc0[3] = _mm256_fmadd_ps(av, bv3b, acc0[3]);
                        acc0[4] = _mm256_fmadd_ps(av, bv4b, acc0[4]);
                        acc0[5] = _mm256_fmadd_ps(av, bv5b, acc0[5]);
                        acc0[6] = _mm256_fmadd_ps(av, bv6b, acc0[6]);
                        acc0[7] = _mm256_fmadd_ps(av, bv7b, acc0[7]);

                        let av = _mm256_loadu_ps(a1_base.add(ko2));
                        acc1[0] = _mm256_fmadd_ps(av, bv0b, acc1[0]);
                        acc1[1] = _mm256_fmadd_ps(av, bv1b, acc1[1]);
                        acc1[2] = _mm256_fmadd_ps(av, bv2b, acc1[2]);
                        acc1[3] = _mm256_fmadd_ps(av, bv3b, acc1[3]);
                        acc1[4] = _mm256_fmadd_ps(av, bv4b, acc1[4]);
                        acc1[5] = _mm256_fmadd_ps(av, bv5b, acc1[5]);
                        acc1[6] = _mm256_fmadd_ps(av, bv6b, acc1[6]);
                        acc1[7] = _mm256_fmadd_ps(av, bv7b, acc1[7]);

                        let av = _mm256_loadu_ps(a2_base.add(ko2));
                        acc2[0] = _mm256_fmadd_ps(av, bv0b, acc2[0]);
                        acc2[1] = _mm256_fmadd_ps(av, bv1b, acc2[1]);
                        acc2[2] = _mm256_fmadd_ps(av, bv2b, acc2[2]);
                        acc2[3] = _mm256_fmadd_ps(av, bv3b, acc2[3]);
                        acc2[4] = _mm256_fmadd_ps(av, bv4b, acc2[4]);
                        acc2[5] = _mm256_fmadd_ps(av, bv5b, acc2[5]);
                        acc2[6] = _mm256_fmadd_ps(av, bv6b, acc2[6]);
                        acc2[7] = _mm256_fmadd_ps(av, bv7b, acc2[7]);

                        let av = _mm256_loadu_ps(a3_base.add(ko2));
                        acc3[0] = _mm256_fmadd_ps(av, bv0b, acc3[0]);
                        acc3[1] = _mm256_fmadd_ps(av, bv1b, acc3[1]);
                        acc3[2] = _mm256_fmadd_ps(av, bv2b, acc3[2]);
                        acc3[3] = _mm256_fmadd_ps(av, bv3b, acc3[3]);
                        acc3[4] = _mm256_fmadd_ps(av, bv4b, acc3[4]);
                        acc3[5] = _mm256_fmadd_ps(av, bv5b, acc3[5]);
                        acc3[6] = _mm256_fmadd_ps(av, bv6b, acc3[6]);
                        acc3[7] = _mm256_fmadd_ps(av, bv7b, acc3[7]);
                    }

                    // Handle remaining 8-element chunk if k_len not divisible by 16
                    let k_vec_rem_start = k_start + k_pairs * 16;
                    if k_vec_rem_start + 8 <= k_end {
                        let ko = k_vec_rem_start;
                        let bv0 = _mm256_loadu_ps(b0_base.add(ko));
                        let bv1 = _mm256_loadu_ps(b1_base.add(ko));
                        let bv2 = _mm256_loadu_ps(b2_base.add(ko));
                        let bv3 = _mm256_loadu_ps(b3_base.add(ko));
                        let bv4 = _mm256_loadu_ps(b4_base.add(ko));
                        let bv5 = _mm256_loadu_ps(b5_base.add(ko));
                        let bv6 = _mm256_loadu_ps(b6_base.add(ko));
                        let bv7 = _mm256_loadu_ps(b7_base.add(ko));

                        let av = _mm256_loadu_ps(a0_base.add(ko));
                        acc0[0] = _mm256_fmadd_ps(av, bv0, acc0[0]);
                        acc0[1] = _mm256_fmadd_ps(av, bv1, acc0[1]);
                        acc0[2] = _mm256_fmadd_ps(av, bv2, acc0[2]);
                        acc0[3] = _mm256_fmadd_ps(av, bv3, acc0[3]);
                        acc0[4] = _mm256_fmadd_ps(av, bv4, acc0[4]);
                        acc0[5] = _mm256_fmadd_ps(av, bv5, acc0[5]);
                        acc0[6] = _mm256_fmadd_ps(av, bv6, acc0[6]);
                        acc0[7] = _mm256_fmadd_ps(av, bv7, acc0[7]);

                        let av = _mm256_loadu_ps(a1_base.add(ko));
                        acc1[0] = _mm256_fmadd_ps(av, bv0, acc1[0]);
                        acc1[1] = _mm256_fmadd_ps(av, bv1, acc1[1]);
                        acc1[2] = _mm256_fmadd_ps(av, bv2, acc1[2]);
                        acc1[3] = _mm256_fmadd_ps(av, bv3, acc1[3]);
                        acc1[4] = _mm256_fmadd_ps(av, bv4, acc1[4]);
                        acc1[5] = _mm256_fmadd_ps(av, bv5, acc1[5]);
                        acc1[6] = _mm256_fmadd_ps(av, bv6, acc1[6]);
                        acc1[7] = _mm256_fmadd_ps(av, bv7, acc1[7]);

                        let av = _mm256_loadu_ps(a2_base.add(ko));
                        acc2[0] = _mm256_fmadd_ps(av, bv0, acc2[0]);
                        acc2[1] = _mm256_fmadd_ps(av, bv1, acc2[1]);
                        acc2[2] = _mm256_fmadd_ps(av, bv2, acc2[2]);
                        acc2[3] = _mm256_fmadd_ps(av, bv3, acc2[3]);
                        acc2[4] = _mm256_fmadd_ps(av, bv4, acc2[4]);
                        acc2[5] = _mm256_fmadd_ps(av, bv5, acc2[5]);
                        acc2[6] = _mm256_fmadd_ps(av, bv6, acc2[6]);
                        acc2[7] = _mm256_fmadd_ps(av, bv7, acc2[7]);

                        let av = _mm256_loadu_ps(a3_base.add(ko));
                        acc3[0] = _mm256_fmadd_ps(av, bv0, acc3[0]);
                        acc3[1] = _mm256_fmadd_ps(av, bv1, acc3[1]);
                        acc3[2] = _mm256_fmadd_ps(av, bv2, acc3[2]);
                        acc3[3] = _mm256_fmadd_ps(av, bv3, acc3[3]);
                        acc3[4] = _mm256_fmadd_ps(av, bv4, acc3[4]);
                        acc3[5] = _mm256_fmadd_ps(av, bv5, acc3[5]);
                        acc3[6] = _mm256_fmadd_ps(av, bv6, acc3[6]);
                        acc3[7] = _mm256_fmadd_ps(av, bv7, acc3[7]);
                    }

                    // Horizontal reduction for AVX2: sum 8 lanes to scalar
                    let c0 = c_ptr.add(i_start * n + j_start);
                    let c1 = c_ptr.add((i_start + 1) * n + j_start);
                    let c2 = c_ptr.add((i_start + 2) * n + j_start);
                    let c3 = c_ptr.add((i_start + 3) * n + j_start);

                    for jj in 0..TILE_J {
                        *c0.add(jj) += hsum_m256(acc0[jj]);
                        *c1.add(jj) += hsum_m256(acc1[jj]);
                        *c2.add(jj) += hsum_m256(acc2[jj]);
                        *c3.add(jj) += hsum_m256(acc3[jj]);
                    }

                    // Scalar remainder for k not covered by AVX2 loads
                    let k_scalar_start = k_start + (k_len / 8) * 8;
                    if k_scalar_start < k_end {
                        for ii in 0..TILE_I {
                            let i = i_start + ii;
                            for jj in 0..TILE_J {
                                let j = j_start + jj;
                                let mut sum = 0.0f32;
                                for p in k_scalar_start..k_end {
                                    sum += *a_ptr.add(i * k + p) * *b_ptr.add(j * k + p);
                                }
                                *c_ptr.add(i * n + j) += sum;
                            }
                        }
                    }
                } else {
                    // Edge tiles: scalar fallback
                    for ii in 0..i_count {
                        let i = i_start + ii;
                        for jj in 0..j_count {
                            let j = j_start + jj;
                            let mut sum = 0.0f32;
                            for p in k_start..k_end {
                                sum += *a_ptr.add(i * k + p) * *b_ptr.add(j * k + p);
                            }
                            *c_ptr.add(i * n + j) += sum;
                        }
                    }
                }

                k_start += TILE_K;
            }
            i_start += TILE_I;
        }
        j_start += TILE_J;
    }
}