lattice-inference 0.3.0

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

// ===================================================================
// Layer normalization (in-place) — with SIMD fast path
// ===================================================================

#[cfg(target_arch = "x86_64")]
use super::arch_kernels::hsum_m256;
use super::simd::simd_config;

/// **Unstable**: layer normalization in-place; dispatch between SIMD and scalar paths may change.
///
/// Layer normalization (in-place).
pub fn layer_norm(x: &mut [f32], gamma: &[f32], beta: &[f32], hidden: usize, eps: f32) {
    debug_assert_eq!(gamma.len(), hidden);
    debug_assert_eq!(beta.len(), hidden);
    debug_assert_eq!(x.len() % hidden, 0);

    let config = simd_config();

    #[cfg(target_arch = "aarch64")]
    {
        if config.neon_enabled {
            // SAFETY: NEON is available on aarch64 and the runtime gate ensures this path.
            unsafe {
                layer_norm_neon(x, gamma, beta, hidden, eps);
                return;
            }
        }
    }

    #[cfg(target_arch = "x86_64")]
    {
        if config.avx2_enabled && config.fma_enabled {
            // SAFETY: The runtime feature checks above guarantee AVX2+FMA support.
            unsafe {
                layer_norm_avx2(x, gamma, beta, hidden, eps);
                return;
            }
        }
    }

    layer_norm_scalar(x, gamma, beta, hidden, eps);
}

#[inline]
pub fn layer_norm_scalar(x: &mut [f32], gamma: &[f32], beta: &[f32], hidden: usize, eps: f32) {
    for row in x.chunks_exact_mut(hidden) {
        let mean = row.iter().copied().sum::<f32>() / hidden as f32;
        let variance = row
            .iter()
            .map(|&v| {
                let d = v - mean;
                d * d
            })
            .sum::<f32>()
            / hidden as f32;
        let inv_std = 1.0 / (variance + eps).sqrt();

        for i in 0..hidden {
            row[i] = (row[i] - mean) * inv_std * gamma[i] + beta[i];
        }
    }
}

#[cfg(target_arch = "aarch64")]
#[target_feature(enable = "neon")]
unsafe fn layer_norm_neon(x: &mut [f32], gamma: &[f32], beta: &[f32], hidden: usize, eps: f32) {
    use std::arch::aarch64::*;

    let inv_n = 1.0 / hidden as f32;

    for row in x.chunks_exact_mut(hidden) {
        let chunks = hidden / 16;
        let ptr = row.as_ptr();

        // --- Pass 1: compute sum → mean with 4x unrolled accumulators ---
        let mut sum0 = vdupq_n_f32(0.0);
        let mut sum1 = vdupq_n_f32(0.0);
        let mut sum2 = vdupq_n_f32(0.0);
        let mut sum3 = vdupq_n_f32(0.0);
        for c in 0..chunks {
            let off = c * 16;
            sum0 = vaddq_f32(sum0, vld1q_f32(ptr.add(off)));
            sum1 = vaddq_f32(sum1, vld1q_f32(ptr.add(off + 4)));
            sum2 = vaddq_f32(sum2, vld1q_f32(ptr.add(off + 8)));
            sum3 = vaddq_f32(sum3, vld1q_f32(ptr.add(off + 12)));
        }
        let combined_sum = vaddq_f32(vaddq_f32(sum0, sum1), vaddq_f32(sum2, sum3));
        let mut total_sum = vaddvq_f32(combined_sum);
        for i in (chunks * 16)..hidden {
            total_sum += *ptr.add(i);
        }
        let mean = total_sum * inv_n;

        // --- Pass 2: compute sum of (x - mean)² → variance with 4x unrolled accumulators ---
        let vmean = vdupq_n_f32(mean);
        let mut sq0 = vdupq_n_f32(0.0);
        let mut sq1 = vdupq_n_f32(0.0);
        let mut sq2 = vdupq_n_f32(0.0);
        let mut sq3 = vdupq_n_f32(0.0);
        for c in 0..chunks {
            let off = c * 16;
            let d0 = vsubq_f32(vld1q_f32(ptr.add(off)), vmean);
            let d1 = vsubq_f32(vld1q_f32(ptr.add(off + 4)), vmean);
            let d2 = vsubq_f32(vld1q_f32(ptr.add(off + 8)), vmean);
            let d3 = vsubq_f32(vld1q_f32(ptr.add(off + 12)), vmean);
            sq0 = vfmaq_f32(sq0, d0, d0);
            sq1 = vfmaq_f32(sq1, d1, d1);
            sq2 = vfmaq_f32(sq2, d2, d2);
            sq3 = vfmaq_f32(sq3, d3, d3);
        }
        let combined_sq = vaddq_f32(vaddq_f32(sq0, sq1), vaddq_f32(sq2, sq3));
        let mut total_sq = vaddvq_f32(combined_sq);
        for i in (chunks * 16)..hidden {
            let d = *ptr.add(i) - mean;
            total_sq += d * d;
        }
        let variance = total_sq * inv_n;
        let inv_std = 1.0 / (variance + eps).sqrt();

        // --- Pass 3: normalize using SIMD: (x - mean) * inv_std * gamma + beta ---
        let vinv_std = vdupq_n_f32(inv_std);
        let out_ptr = row.as_mut_ptr();
        let g_ptr = gamma.as_ptr();
        let b_ptr = beta.as_ptr();

        for c in 0..chunks {
            let off = c * 16;
            let d0 = vsubq_f32(vld1q_f32(out_ptr.add(off) as *const f32), vmean);
            let n0 = vfmaq_f32(
                vld1q_f32(b_ptr.add(off)),
                vmulq_f32(d0, vinv_std),
                vld1q_f32(g_ptr.add(off)),
            );
            vst1q_f32(out_ptr.add(off), n0);

            let d1 = vsubq_f32(vld1q_f32(out_ptr.add(off + 4) as *const f32), vmean);
            let n1 = vfmaq_f32(
                vld1q_f32(b_ptr.add(off + 4)),
                vmulq_f32(d1, vinv_std),
                vld1q_f32(g_ptr.add(off + 4)),
            );
            vst1q_f32(out_ptr.add(off + 4), n1);

            let d2 = vsubq_f32(vld1q_f32(out_ptr.add(off + 8) as *const f32), vmean);
            let n2 = vfmaq_f32(
                vld1q_f32(b_ptr.add(off + 8)),
                vmulq_f32(d2, vinv_std),
                vld1q_f32(g_ptr.add(off + 8)),
            );
            vst1q_f32(out_ptr.add(off + 8), n2);

            let d3 = vsubq_f32(vld1q_f32(out_ptr.add(off + 12) as *const f32), vmean);
            let n3 = vfmaq_f32(
                vld1q_f32(b_ptr.add(off + 12)),
                vmulq_f32(d3, vinv_std),
                vld1q_f32(g_ptr.add(off + 12)),
            );
            vst1q_f32(out_ptr.add(off + 12), n3);
        }
        for i in (chunks * 16)..hidden {
            *out_ptr.add(i) = (*out_ptr.add(i) - mean) * inv_std * *g_ptr.add(i) + *b_ptr.add(i);
        }
    }
}

#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2", enable = "fma")]
unsafe fn layer_norm_avx2(x: &mut [f32], gamma: &[f32], beta: &[f32], hidden: usize, eps: f32) {
    use std::arch::x86_64::*;

    let inv_n = 1.0 / hidden as f32;

    for row in x.chunks_exact_mut(hidden) {
        let chunks = hidden / 32;
        let ptr = row.as_ptr();

        // --- Pass 1: compute sum → mean with 4x unrolled accumulators ---
        let mut sum0 = _mm256_setzero_ps();
        let mut sum1 = _mm256_setzero_ps();
        let mut sum2 = _mm256_setzero_ps();
        let mut sum3 = _mm256_setzero_ps();
        for c in 0..chunks {
            let off = c * 32;
            sum0 = _mm256_add_ps(sum0, _mm256_loadu_ps(ptr.add(off)));
            sum1 = _mm256_add_ps(sum1, _mm256_loadu_ps(ptr.add(off + 8)));
            sum2 = _mm256_add_ps(sum2, _mm256_loadu_ps(ptr.add(off + 16)));
            sum3 = _mm256_add_ps(sum3, _mm256_loadu_ps(ptr.add(off + 24)));
        }
        let combined_sum = _mm256_add_ps(_mm256_add_ps(sum0, sum1), _mm256_add_ps(sum2, sum3));
        let mut total_sum = hsum_m256(combined_sum);
        for i in (chunks * 32)..hidden {
            total_sum += *ptr.add(i);
        }
        let mean = total_sum * inv_n;

        // --- Pass 2: compute sum of (x - mean)² → variance with 4x unrolled accumulators ---
        let vmean = _mm256_set1_ps(mean);
        let mut sq0 = _mm256_setzero_ps();
        let mut sq1 = _mm256_setzero_ps();
        let mut sq2 = _mm256_setzero_ps();
        let mut sq3 = _mm256_setzero_ps();
        for c in 0..chunks {
            let off = c * 32;
            let d0 = _mm256_sub_ps(_mm256_loadu_ps(ptr.add(off)), vmean);
            let d1 = _mm256_sub_ps(_mm256_loadu_ps(ptr.add(off + 8)), vmean);
            let d2 = _mm256_sub_ps(_mm256_loadu_ps(ptr.add(off + 16)), vmean);
            let d3 = _mm256_sub_ps(_mm256_loadu_ps(ptr.add(off + 24)), vmean);
            sq0 = _mm256_fmadd_ps(d0, d0, sq0);
            sq1 = _mm256_fmadd_ps(d1, d1, sq1);
            sq2 = _mm256_fmadd_ps(d2, d2, sq2);
            sq3 = _mm256_fmadd_ps(d3, d3, sq3);
        }
        let combined_sq = _mm256_add_ps(_mm256_add_ps(sq0, sq1), _mm256_add_ps(sq2, sq3));
        let mut total_sq = hsum_m256(combined_sq);
        for i in (chunks * 32)..hidden {
            let d = *ptr.add(i) - mean;
            total_sq += d * d;
        }
        let variance = total_sq * inv_n;
        let inv_std = 1.0 / (variance + eps).sqrt();

        // --- Pass 3: normalize ---
        let vinv_std = _mm256_set1_ps(inv_std);
        let out_ptr = row.as_mut_ptr();
        let g_ptr = gamma.as_ptr();
        let b_ptr = beta.as_ptr();

        for c in 0..chunks {
            let off = c * 32;
            for lane_off in (0..32).step_by(8) {
                let p = off + lane_off;
                let d = _mm256_sub_ps(_mm256_loadu_ps(out_ptr.add(p) as *const f32), vmean);
                let scaled = _mm256_mul_ps(d, vinv_std);
                // result = scaled * gamma + beta
                let result = _mm256_fmadd_ps(
                    scaled,
                    _mm256_loadu_ps(g_ptr.add(p)),
                    _mm256_loadu_ps(b_ptr.add(p)),
                );
                _mm256_storeu_ps(out_ptr.add(p), result);
            }
        }
        for i in (chunks * 32)..hidden {
            *out_ptr.add(i) = (*out_ptr.add(i) - mean) * inv_std * *g_ptr.add(i) + *b_ptr.add(i);
        }
    }
}