diskann-wide 0.45.0

DiskANN is a fast approximate nearest neighbor search library for high dimensional data
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

/*
 * Copyright (c) Microsoft Corporation.
 * Licensed under the MIT license.
 */

#[cfg(target_arch = "x86_64")]
use diskann_wide::{SIMDMulAdd, SIMDSumTree, SIMDVector};

#[cfg(target_arch = "x86_64")]
type V3 = diskann_wide::arch::x86_64::V3;

#[cfg(target_arch = "x86_64")]
type V4 = diskann_wide::arch::x86_64::V4;

struct InnerProduct;

impl diskann_wide::arch::Target2<diskann_wide::arch::Scalar, f32, &[f32], &[f32]> for InnerProduct {
    #[inline(always)]
    fn run(self, _: diskann_wide::arch::Scalar, x: &[f32], y: &[f32]) -> f32 {
        std::iter::zip(x.iter(), y.iter()).map(|(a, b)| a * b).sum()
    }
}

#[cfg(target_arch = "x86_64")]
impl diskann_wide::arch::Target2<V3, f32, &[f32], &[f32]> for InnerProduct {
    #[inline(always)]
    fn run(self, arch: V3, a: &[f32], b: &[f32]) -> f32 {
        diskann_wide::alias!(f32s = <V3>::f32x8);

        // The number of lanes to use in a SIMD register.
        const LANES: usize = f32s::LANES;

        let len = a.len();
        let a = a.as_ptr();
        let b = b.as_ptr();

        let mut sum = f32s::default(arch);
        let trips = len / LANES;
        let remainder = len % LANES;
        for i in 0..trips {
            // SAFETY: By loop construction, `a.add(LANES * (i + 1) - 1)` is always in-bounds.
            let wa = unsafe { f32s::load_simd(arch, a.add(LANES * i)) };
            // SAFETY: By loop construction, `b.add(LANES * (i + 1) - 1)` is always in-bounds.
            let wb = unsafe { f32s::load_simd(arch, b.add(LANES * i)) };
            sum = wa.mul_add_simd(wb, sum);
        }

        // Handle and remaining using predicated loads.
        if remainder != 0 {
            // SAFETY: By loop construction, `a.add(LANES * trips)` is always in-bounds.
            let wa = unsafe { f32s::load_simd_first(arch, a.add(trips * LANES), remainder) };
            // SAFETY: By loop construction, `b.add(LANES * trips)` is always in-bounds.
            let wb = unsafe { f32s::load_simd_first(arch, b.add(trips * LANES), remainder) };
            sum = wa.mul_add_simd(wb, sum);
        }

        sum.sum_tree()
    }
}

#[cfg(target_arch = "x86_64")]
impl diskann_wide::arch::Target2<V4, f32, &[f32], &[f32]> for InnerProduct {
    #[inline(always)]
    fn run(self, arch: V4, a: &[f32], b: &[f32]) -> f32 {
        diskann_wide::alias!(f32s = <V4>::f32x16);

        // The number of lanes to use in a SIMD register.
        const LANES: usize = f32s::LANES;

        let len = a.len();
        let a = a.as_ptr();
        let b = b.as_ptr();

        let mut sum = f32s::default(arch);
        let trips = len / LANES;
        let remainder = len % LANES;
        for i in 0..trips {
            // SAFETY: By loop construction, `a.add(LANES * (i + 1) - 1)` is always in-bounds.
            let wa = unsafe { f32s::load_simd(arch, a.add(LANES * i)) };
            // SAFETY: By loop construction, `b.add(LANES * (i + 1) - 1)` is always in-bounds.
            let wb = unsafe { f32s::load_simd(arch, b.add(LANES * i)) };
            sum = wa.mul_add_simd(wb, sum);
        }

        // Handle and remaining using predicated loads.
        if remainder != 0 {
            // SAFETY: By loop construction, `a.add(LANES * trips)` is always in-bounds.
            let wa = unsafe { f32s::load_simd_first(arch, a.add(trips * LANES), remainder) };
            // SAFETY: By loop construction, `b.add(LANES * trips)` is always in-bounds.
            let wb = unsafe { f32s::load_simd_first(arch, b.add(trips * LANES), remainder) };
            sum = wa.mul_add_simd(wb, sum);
        }

        sum.sum_tree()
    }
}

// This example
struct SquaredL2;

impl diskann_wide::arch::Target2<diskann_wide::arch::Scalar, f32, &[f32], &[f32]> for SquaredL2 {
    #[inline(always)]
    fn run(self, _: diskann_wide::arch::Scalar, x: &[f32], y: &[f32]) -> f32 {
        std::iter::zip(x.iter(), y.iter())
            .map(|(a, b)| {
                let d = a - b;
                d * d
            })
            .sum()
    }
}

#[cfg(target_arch = "x86_64")]
impl diskann_wide::arch::Target2<V3, f32, &[f32], &[f32]> for SquaredL2 {
    #[inline(always)]
    fn run(self, arch: V3, x: &[f32], y: &[f32]) -> f32 {
        let arch: diskann_wide::arch::Scalar = arch.into();
        self.run(arch, x, y)
    }
}

#[cfg(target_arch = "x86_64")]
impl diskann_wide::arch::Target2<V4, f32, &[f32], &[f32]> for SquaredL2 {
    #[inline(always)]
    fn run(self, arch: V4, x: &[f32], y: &[f32]) -> f32 {
        let arch: diskann_wide::arch::Scalar = arch.into();
        self.run(arch, x, y)
    }
}

#[inline(never)]
pub fn test_inner_product(a: &[f32], b: &[f32]) -> f32 {
    diskann_wide::arch::dispatch2(InnerProduct, a, b)
}

#[inline(never)]
pub fn test_euclidean(a: &[f32], b: &[f32]) -> f32 {
    diskann_wide::arch::dispatch2(SquaredL2, a, b)
}

#[test]
fn test() {
    assert_eq!(test_inner_product(&[1.0], &[2.0]), 2.0);
    assert_eq!(test_inner_product(&[2.0], &[4.0]), 8.0);

    assert_eq!(test_euclidean(&[1.0], &[2.0]), 1.0);
    assert_eq!(test_euclidean(&[2.0], &[4.0]), 4.0);
}