fasttext 0.8.0

fastText pure Rust implementation
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

// AVX-512F SIMD kernels (x86_64, 512-bit registers with FMA).
//
// Processes 16 f32 values per instruction. Uses `_mm512_fmadd_ps` for fused
// multiply-add and `_mm512_reduce_add_ps` for horizontal sum.

use std::arch::x86_64::*;

use crate::matrix::{DenseMatrix, Matrix};
use crate::vector::Vector;

/// # Safety
/// Caller must ensure AVX-512F is available.
#[target_feature(enable = "avx512f")]
pub(crate) unsafe fn dot(a: &[f32], b: &[f32]) -> f32 {
    debug_assert_eq!(a.len(), b.len());
    let n = a.len();
    let chunks4 = n / 64; // 4 × 16 f32 = 64 elements per unrolled iteration
    let chunks1 = (n % 64) / 16;
    let remainder = n % 16;

    let mut acc0 = _mm512_setzero_ps();
    let mut acc1 = _mm512_setzero_ps();
    let mut acc2 = _mm512_setzero_ps();
    let mut acc3 = _mm512_setzero_ps();

    let a_ptr = a.as_ptr();
    let b_ptr = b.as_ptr();

    for i in 0..chunks4 {
        let off = i * 64;
        let a0 = _mm512_loadu_ps(a_ptr.add(off));
        let a1 = _mm512_loadu_ps(a_ptr.add(off + 16));
        let a2 = _mm512_loadu_ps(a_ptr.add(off + 32));
        let a3 = _mm512_loadu_ps(a_ptr.add(off + 48));
        let b0 = _mm512_loadu_ps(b_ptr.add(off));
        let b1 = _mm512_loadu_ps(b_ptr.add(off + 16));
        let b2 = _mm512_loadu_ps(b_ptr.add(off + 32));
        let b3 = _mm512_loadu_ps(b_ptr.add(off + 48));
        acc0 = _mm512_fmadd_ps(a0, b0, acc0);
        acc1 = _mm512_fmadd_ps(a1, b1, acc1);
        acc2 = _mm512_fmadd_ps(a2, b2, acc2);
        acc3 = _mm512_fmadd_ps(a3, b3, acc3);
    }

    acc0 = _mm512_add_ps(acc0, acc1);
    acc2 = _mm512_add_ps(acc2, acc3);
    acc0 = _mm512_add_ps(acc0, acc2);

    let rem_start4 = chunks4 * 64;
    for i in 0..chunks1 {
        let off = rem_start4 + i * 16;
        let a0 = _mm512_loadu_ps(a_ptr.add(off));
        let b0 = _mm512_loadu_ps(b_ptr.add(off));
        acc0 = _mm512_fmadd_ps(a0, b0, acc0);
    }

    let mut result = _mm512_reduce_add_ps(acc0);

    let rem_start = rem_start4 + chunks1 * 16;
    for i in 0..remainder {
        result += a[rem_start + i] * b[rem_start + i];
    }

    result
}

/// # Safety
/// Caller must ensure AVX-512F is available.
#[target_feature(enable = "avx512f")]
pub(crate) unsafe fn add_vector(dest: &mut [f32], src: &[f32], scale: f32) {
    debug_assert_eq!(dest.len(), src.len());
    let n = dest.len();
    let chunks4 = n / 64;
    let chunks1 = (n % 64) / 16;
    let remainder = n % 16;

    let scale_v = _mm512_set1_ps(scale);
    let d_ptr = dest.as_mut_ptr();
    let s_ptr = src.as_ptr();

    for i in 0..chunks4 {
        let off = i * 64;
        let d0 = _mm512_loadu_ps(d_ptr.add(off));
        let d1 = _mm512_loadu_ps(d_ptr.add(off + 16));
        let d2 = _mm512_loadu_ps(d_ptr.add(off + 32));
        let d3 = _mm512_loadu_ps(d_ptr.add(off + 48));
        let s0 = _mm512_loadu_ps(s_ptr.add(off));
        let s1 = _mm512_loadu_ps(s_ptr.add(off + 16));
        let s2 = _mm512_loadu_ps(s_ptr.add(off + 32));
        let s3 = _mm512_loadu_ps(s_ptr.add(off + 48));
        _mm512_storeu_ps(d_ptr.add(off), _mm512_fmadd_ps(scale_v, s0, d0));
        _mm512_storeu_ps(d_ptr.add(off + 16), _mm512_fmadd_ps(scale_v, s1, d1));
        _mm512_storeu_ps(d_ptr.add(off + 32), _mm512_fmadd_ps(scale_v, s2, d2));
        _mm512_storeu_ps(d_ptr.add(off + 48), _mm512_fmadd_ps(scale_v, s3, d3));
    }

    let rem_start4 = chunks4 * 64;
    for i in 0..chunks1 {
        let off = rem_start4 + i * 16;
        let d0 = _mm512_loadu_ps(d_ptr.add(off));
        let s0 = _mm512_loadu_ps(s_ptr.add(off));
        _mm512_storeu_ps(d_ptr.add(off), _mm512_fmadd_ps(scale_v, s0, d0));
    }

    let rem_start = rem_start4 + chunks1 * 16;
    for i in 0..remainder {
        dest[rem_start + i] += scale * src[rem_start + i];
    }
}

/// # Safety
/// Caller must ensure AVX-512F is available.
#[target_feature(enable = "avx512f")]
pub(crate) unsafe fn average_rows(x: &mut Vector, rows: &[i32], mat: &DenseMatrix) {
    if rows.is_empty() {
        x.zero();
        return;
    }

    let n = mat.cols() as usize;

    let first_row = mat.row(rows[0] as i64);
    x.data_mut().copy_from_slice(first_row);

    for &row_idx in &rows[1..] {
        let row = mat.row(row_idx as i64);
        add_vector(x.data_mut(), row, 1.0);
    }

    let scale = 1.0 / rows.len() as f32;
    let scale_v = _mm512_set1_ps(scale);
    let n_chunks = n / 16;
    let remainder = n % 16;
    let d_ptr = x.data_mut().as_mut_ptr();

    for i in 0..n_chunks {
        let off = i * 16;
        let v = _mm512_loadu_ps(d_ptr.add(off));
        _mm512_storeu_ps(d_ptr.add(off), _mm512_mul_ps(v, scale_v));
    }

    let rem_start = n_chunks * 16;
    for i in 0..remainder {
        x.data_mut()[rem_start + i] *= scale;
    }
}