bbqvec 0.0.10

Scalable, embeddable, vector storage for approximate K-nearest-neighbors (AKNN)
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

use std::arch::x86_64::*;
use std::ptr::read_unaligned;

use crate::unaligned_f32::UnalignedF32Slice;

#[target_feature(enable = "avx")]
#[target_feature(enable = "fma")]
unsafe fn hsum256_ps_avx(x: __m256) -> f32 {
    let x128: __m128 = _mm_add_ps(_mm256_extractf128_ps(x, 1), _mm256_castps256_ps128(x));
    let x64: __m128 = _mm_add_ps(x128, _mm_movehl_ps(x128, x128));
    let x32: __m128 = _mm_add_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
    _mm_cvtss_f32(x32)
}

#[target_feature(enable = "avx")]
#[target_feature(enable = "fma")]
pub(crate) unsafe fn euclid_similarity_avx(v1: &UnalignedF32Slice, v2: &UnalignedF32Slice) -> f32 {
    // It is safe to load unaligned floats from a pointer.
    // <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_loadu_ps&ig_expand=4134>

    let n = v1.len();
    let m = n - (n % 32);
    let mut ptr1 = v1.as_ptr() as *const f32;
    let mut ptr2 = v2.as_ptr() as *const f32;
    let mut sum256_1: __m256 = _mm256_setzero_ps();
    let mut sum256_2: __m256 = _mm256_setzero_ps();
    let mut sum256_3: __m256 = _mm256_setzero_ps();
    let mut sum256_4: __m256 = _mm256_setzero_ps();
    let mut i: usize = 0;
    while i < m {
        let sub256_1: __m256 =
            _mm256_sub_ps(_mm256_loadu_ps(ptr1.add(0)), _mm256_loadu_ps(ptr2.add(0)));
        sum256_1 = _mm256_fmadd_ps(sub256_1, sub256_1, sum256_1);

        let sub256_2: __m256 =
            _mm256_sub_ps(_mm256_loadu_ps(ptr1.add(8)), _mm256_loadu_ps(ptr2.add(8)));
        sum256_2 = _mm256_fmadd_ps(sub256_2, sub256_2, sum256_2);

        let sub256_3: __m256 =
            _mm256_sub_ps(_mm256_loadu_ps(ptr1.add(16)), _mm256_loadu_ps(ptr2.add(16)));
        sum256_3 = _mm256_fmadd_ps(sub256_3, sub256_3, sum256_3);

        let sub256_4: __m256 =
            _mm256_sub_ps(_mm256_loadu_ps(ptr1.add(24)), _mm256_loadu_ps(ptr2.add(24)));
        sum256_4 = _mm256_fmadd_ps(sub256_4, sub256_4, sum256_4);

        ptr1 = ptr1.add(32);
        ptr2 = ptr2.add(32);
        i += 32;
    }

    let mut result = hsum256_ps_avx(sum256_1)
        + hsum256_ps_avx(sum256_2)
        + hsum256_ps_avx(sum256_3)
        + hsum256_ps_avx(sum256_4);
    for i in 0..n - m {
        let a = read_unaligned(ptr1.add(i));
        let b = read_unaligned(ptr2.add(i));
        result += (a - b).powi(2);
    }
    result
}

#[target_feature(enable = "avx")]
#[target_feature(enable = "fma")]
pub(crate) unsafe fn dot_similarity_avx(v1: &UnalignedF32Slice, v2: &UnalignedF32Slice) -> f32 {
    // It is safe to load unaligned floats from a pointer.
    // <https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_loadu_ps&ig_expand=4134>

    let n = v1.len();
    let m = n - (n % 32);
    let mut ptr1 = v1.as_ptr() as *const f32;
    let mut ptr2 = v2.as_ptr() as *const f32;
    let mut sum256_1: __m256 = _mm256_setzero_ps();
    let mut sum256_2: __m256 = _mm256_setzero_ps();
    let mut sum256_3: __m256 = _mm256_setzero_ps();
    let mut sum256_4: __m256 = _mm256_setzero_ps();
    let mut i: usize = 0;
    while i < m {
        sum256_1 = _mm256_fmadd_ps(_mm256_loadu_ps(ptr1), _mm256_loadu_ps(ptr2), sum256_1);
        sum256_2 = _mm256_fmadd_ps(
            _mm256_loadu_ps(ptr1.add(8)),
            _mm256_loadu_ps(ptr2.add(8)),
            sum256_2,
        );
        sum256_3 = _mm256_fmadd_ps(
            _mm256_loadu_ps(ptr1.add(16)),
            _mm256_loadu_ps(ptr2.add(16)),
            sum256_3,
        );
        sum256_4 = _mm256_fmadd_ps(
            _mm256_loadu_ps(ptr1.add(24)),
            _mm256_loadu_ps(ptr2.add(24)),
            sum256_4,
        );

        ptr1 = ptr1.add(32);
        ptr2 = ptr2.add(32);
        i += 32;
    }

    let mut result = hsum256_ps_avx(sum256_1)
        + hsum256_ps_avx(sum256_2)
        + hsum256_ps_avx(sum256_3)
        + hsum256_ps_avx(sum256_4);

    for i in 0..n - m {
        let a = read_unaligned(ptr1.add(i));
        let b = read_unaligned(ptr2.add(i));
        result += a * b;
    }
    result
}

#[cfg(test)]
mod tests {
    #[test]
    fn test_spaces_avx() {
        use super::*;
        use crate::spaces::simple::*;

        if is_x86_feature_detected!("avx") && is_x86_feature_detected!("fma") {
            let v1: Vec<f32> = vec![
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                26., 27., 28., 29., 30., 31.,
            ];
            let v2: Vec<f32> = vec![
                40., 41., 42., 43., 44., 45., 46., 47., 48., 49., 50., 51., 52., 53., 54., 55.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                56., 57., 58., 59., 60., 61.,
            ];

            let v1 = UnalignedF32Slice::from_slice(&v1[..]);
            let v2 = UnalignedF32Slice::from_slice(&v2[..]);

            let euclid_simd = unsafe { euclid_similarity_avx(v1, v2) };
            let euclid = euclidean_distance_non_optimized(v1, v2);
            assert_eq!(euclid_simd, euclid);

            let dot_simd = unsafe { dot_similarity_avx(v1, v2) };
            let dot = dot_product_non_optimized(v1, v2);
            assert_eq!(dot_simd, dot);

            // let cosine_simd = unsafe { cosine_preprocess_avx(v1.clone()) };
            // let cosine = cosine_preprocess(v1);
            // assert_eq!(cosine_simd, cosine);
        } else {
            println!("avx test skipped");
        }
    }
}