granne 0.5.2

Graph-based Retrieval of Approximate Nearest Neighbors
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

#[cfg(feature = "blas")]
use blas;

#[cfg(not(feature = "blas"))]
pub fn dot_product_f32(x: &[f32], y: &[f32]) -> f32 {
    // optimized code to compute the dot product for systems supporting avx2
    // with fallback for other systems

    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    #[target_feature(enable = "avx2,fma")]
    unsafe fn dot_product_avx2(x: &[f32], y: &[f32]) -> f32 {
        // this function will be inlined and take advantage of avx2 auto-vectorization
        dot_product_fallback(x, y)
    }

    #[inline(always)]
    fn dot_product_fallback(x: &[f32], y: &[f32]) -> f32 {
        const CHUNK_SIZE: usize = 32;
        let mut chunk = [0.0f32; CHUNK_SIZE];

        for (a, b) in x.chunks_exact(CHUNK_SIZE).zip(y.chunks_exact(CHUNK_SIZE)) {
            for i in 0..CHUNK_SIZE {
                chunk[i] = a[i].mul_add(b[i], chunk[i]);
            }
        }

        let mut r = 0.0f32;
        for i in 0..CHUNK_SIZE {
            r += chunk[i];
        }

        for (ai, bi) in x
            .chunks_exact(CHUNK_SIZE)
            .remainder()
            .iter()
            .zip(y.chunks_exact(CHUNK_SIZE).remainder())
        {
            r = ai.mul_add(*bi, r);
        }

        r
    }

    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    {
        if is_x86_feature_detected!("avx2") && is_x86_feature_detected!("fma") {
            return unsafe { dot_product_avx2(x, y) };
        }
    }

    dot_product_fallback(x, y)
}

#[cfg(feature = "blas")]
pub fn dot_product_f32(x: &[f32], y: &[f32]) -> f32 {
    unsafe { blas::sdot(x.len() as i32, x, 1, y, 1) }
}

pub fn dot_product_and_squared_norms_i8(x: &[i8], y: &[i8]) -> (i32, i32, i32) {
    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    #[target_feature(enable = "avx2")]
    unsafe fn compute_r_dx_dy_avx2(x: &[i8], y: &[i8]) -> (i32, i32, i32) {
        compute_r_dx_dy_fallback(x, y)
    }

    #[inline(always)]
    fn compute_r_dx_dy_fallback(x: &[i8], y: &[i8]) -> (i32, i32, i32) {
        let mut r = 0i32;
        let mut dx = 0i32;
        let mut dy = 0i32;

        for (xi, yi) in x.iter().map(|&xi| i32::from(xi)).zip(y.iter().map(|&yi| i32::from(yi))) {
            r += xi * yi;
            dx += xi * xi;
            dy += yi * yi;
        }

        (r, dx, dy)
    }

    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    {
        if is_x86_feature_detected!("avx2") {
            return unsafe { compute_r_dx_dy_avx2(x, y) };
        }
    }

    compute_r_dx_dy_fallback(x, y)
}

#[cfg(not(feature = "blas"))]
pub fn sum_into_f32(x: &mut [f32], y: &[f32]) {
    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    #[target_feature(enable = "avx2")]
    unsafe fn sum_into_avx2(x: &mut [f32], y: &[f32]) {
        sum_into_fallback(x, y)
    }

    #[inline(always)]
    fn sum_into_fallback(x: &mut [f32], y: &[f32]) {
        assert_eq!(x.len(), y.len());

        for i in 0..x.len() {
            x[i] += y[i];
        }
    }

    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    {
        if is_x86_feature_detected!("avx2") && is_x86_feature_detected!("fma") {
            return unsafe { sum_into_avx2(x, y) };
        }
    }

    sum_into_fallback(x, y)
}

#[cfg(feature = "blas")]
pub fn sum_into_f32(x: &mut [f32], y: &[f32]) {
    unsafe { blas::saxpy(x.len() as i32, 1f32, y, 1, x, 1) };
}

#[cfg(not(feature = "blas"))]
pub fn normalize_f32(x: &mut [f32]) {
    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    #[target_feature(enable = "avx2,fma")]
    unsafe fn normalize_avx2(x: &mut [f32]) {
        normalize_fallback(x)
    }

    #[inline(always)]
    fn normalize_fallback(x: &mut [f32]) {
        let norm = dot_product_f32(x, x).sqrt();

        if norm > 0.0 {
            for i in 0..x.len() {
                x[i] /= norm;
            }
        }
    }

    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    {
        if is_x86_feature_detected!("avx2") && is_x86_feature_detected!("fma") {
            return unsafe { normalize_avx2(x) };
        }
    }

    normalize_fallback(x)
}

#[cfg(feature = "blas")]
pub fn normalize_f32(x: &mut [f32]) {
    let n = x.len() as i32;
    let norm = unsafe { blas::snrm2(n, x, 1) };
    if norm > 0.0 {
        unsafe { blas::sscal(n, 1.0 / norm, x, 1) };
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::test_helper;

    #[test]
    fn sum() {
        for i in 1..101 {
            let mut x: Vec<f32> = test_helper::random_floats().take(i).collect();
            let y: Vec<f32> = test_helper::random_floats().take(i).collect();

            let mut expected = x.clone();
            for (i, xi) in expected.iter_mut().enumerate() {
                *xi += y[i];
            }

            sum_into_f32(&mut x, &y);

            assert_eq!(expected, x);
        }
    }

    #[test]
    fn dot_product() {
        for i in 1..101 {
            let x: Vec<f32> = test_helper::random_floats().take(i).collect();
            let y: Vec<f32> = test_helper::random_floats().take(i).collect();

            let mut expected = 0.0f32;
            for i in 0..i {
                expected += x[i] * y[i];
            }

            assert!((expected - dot_product_f32(&x, &y)).abs() < 0.000001f32);
        }
    }
}