vq 0.2.0

A vector quantization library for Rust
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

use crate::core::error::{VqError, VqResult};

#[cfg(feature = "simd")]
use crate::core::hsdlib_ffi;

/// Supported distance metrics for vector comparisons.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Distance {
    /// Squared Euclidean distance (L2²). Efficient for comparisons as it avoids square roots.
    SquaredEuclidean,
    /// Euclidean distance (L2).
    Euclidean,
    /// Manhattan distance (L1). Sum of absolute differences.
    Manhattan,
    /// Cosine distance, defined as `1.0 - cosine_similarity`.
    CosineDistance,
}

impl Distance {
    /// Returns the name of this distance metric.
    pub const fn name(&self) -> &'static str {
        match self {
            Distance::SquaredEuclidean => "squared_euclidean",
            Distance::Euclidean => "euclidean",
            Distance::Manhattan => "manhattan",
            Distance::CosineDistance => "cosine",
        }
    }

    /// Computes the distance between two vectors using the specified metric.
    ///
    /// If the `simd` feature is enabled, this method will use SIMD-accelerated
    /// implementations when available (AVX/AVX2/FMA for x86, NEON for ARM).
    ///
    /// # Arguments
    ///
    /// * `a` - First vector
    /// * `b` - Second vector
    ///
    /// # Returns
    ///
    /// The computed distance as an `f32`.
    ///
    /// # Errors
    ///
    /// Returns `VqError::DimensionMismatch` if the vectors have different lengths.
    #[inline]
    pub fn compute(&self, a: &[f32], b: &[f32]) -> VqResult<f32> {
        if a.len() != b.len() {
            return Err(VqError::DimensionMismatch {
                expected: a.len(),
                found: b.len(),
            });
        }

        let result = match self {
            Distance::SquaredEuclidean => compute_squared_euclidean(a, b),
            Distance::Euclidean => compute_squared_euclidean(a, b).sqrt(),
            Distance::Manhattan => compute_manhattan(a, b),
            Distance::CosineDistance => compute_cosine_distance(a, b),
        };

        Ok(result)
    }
}

#[inline]
fn compute_squared_euclidean(a: &[f32], b: &[f32]) -> f32 {
    #[cfg(feature = "simd")]
    {
        if let Some(result) = hsdlib_ffi::sqeuclidean_f32(a, b) {
            return result;
        }
    }
    // Fallback to scalar implementation
    a.iter()
        .zip(b.iter())
        .map(|(&x, &y)| {
            let diff = x - y;
            diff * diff
        })
        .sum()
}

#[inline]
fn compute_manhattan(a: &[f32], b: &[f32]) -> f32 {
    #[cfg(feature = "simd")]
    {
        if let Some(result) = hsdlib_ffi::manhattan_f32(a, b) {
            return result;
        }
    }
    // Fallback to scalar implementation
    a.iter().zip(b.iter()).map(|(&x, &y)| (x - y).abs()).sum()
}

#[inline]
fn compute_cosine_distance(a: &[f32], b: &[f32]) -> f32 {
    #[cfg(feature = "simd")]
    {
        // hsdlib returns cosine similarity, we need cosine distance (1 - similarity)
        if let Some(similarity) = hsdlib_ffi::cosine_f32(a, b) {
            return 1.0 - similarity;
        }
    }
    // Fallback to scalar implementation
    let dot: f32 = a.iter().zip(b.iter()).map(|(&x, &y)| x * y).sum();
    let norm_a = a.iter().map(|&x| x * x).sum::<f32>().sqrt();
    let norm_b = b.iter().map(|&x| x * x).sum::<f32>().sqrt();

    // Use epsilon to handle near-zero norms (avoids division by very small numbers)
    const EPSILON: f32 = 1e-10;
    if norm_a < EPSILON || norm_b < EPSILON {
        // Zero or near-zero vectors are considered maximally distant
        1.0
    } else {
        // Clamp result to [0, 1] to handle floating-point errors
        (1.0 - (dot / (norm_a * norm_b))).clamp(0.0, 1.0)
    }
}

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

    fn approx_eq(a: f32, b: f32, eps: f32) -> bool {
        (a - b).abs() < eps
    }

    #[test]
    fn test_squared_euclidean() {
        let a = vec![1.0f32, 2.0, 3.0];
        let b = vec![4.0f32, 6.0, 8.0];
        let result = Distance::SquaredEuclidean.compute(&a, &b).unwrap();
        assert!(approx_eq(result, 50.0, 1e-6));
    }

    #[test]
    fn test_euclidean() {
        let a = vec![1.0f32, 2.0, 3.0];
        let b = vec![4.0f32, 6.0, 8.0];
        let expected = 50.0f32.sqrt();
        let result = Distance::Euclidean.compute(&a, &b).unwrap();
        assert!(approx_eq(result, expected, 1e-6));
    }

    #[test]
    fn test_cosine_distance() {
        let a = vec![1.0f32, 0.0];
        let b = vec![0.0f32, 1.0];
        let result = Distance::CosineDistance.compute(&a, &b).unwrap();
        assert!(approx_eq(result, 1.0, 1e-6));

        let a = vec![1.0f32, 1.0];
        let b = vec![1.0f32, 1.0];
        let result = Distance::CosineDistance.compute(&a, &b).unwrap();
        assert!(approx_eq(result, 0.0, 1e-6));
    }

    #[test]
    fn test_manhattan() {
        let a = vec![1.0f32, 2.0, 3.0];
        let b = vec![4.0f32, 6.0, 8.0];
        let result = Distance::Manhattan.compute(&a, &b).unwrap();
        assert!(approx_eq(result, 12.0, 1e-6));
    }

    #[test]
    fn test_dimension_mismatch() {
        let a = vec![1.0f32, 2.0];
        let b = vec![1.0f32];
        let result = Distance::Euclidean.compute(&a, &b);
        assert!(result.is_err());
    }

    #[test]
    #[cfg(feature = "simd")]
    fn test_simd_consistency() {
        use crate::core::hsdlib_ffi;

        let mut rng = rand::rng();
        use rand::Rng;

        let len = 100;
        let a: Vec<f32> = (0..len).map(|_| rng.random::<f32>()).collect();
        let b: Vec<f32> = (0..len).map(|_| rng.random::<f32>()).collect();

        // Check L2 Squared
        let scalar_l2sq: f32 = a.iter().zip(b.iter()).map(|(x, y)| (x - y).powi(2)).sum();
        let simd_l2sq = hsdlib_ffi::sqeuclidean_f32(&a, &b).unwrap();
        assert!(
            (scalar_l2sq - simd_l2sq).abs() < 1e-4,
            "L2 Squared mismatch: scalar={}, simd={}",
            scalar_l2sq,
            simd_l2sq
        );

        // Check Manhattan
        let scalar_l1: f32 = a.iter().zip(b.iter()).map(|(x, y)| (x - y).abs()).sum();
        let simd_l1 = hsdlib_ffi::manhattan_f32(&a, &b).unwrap();
        assert!(
            (scalar_l1 - simd_l1).abs() < 1e-4,
            "Manhattan mismatch: scalar={}, simd={}",
            scalar_l1,
            simd_l1
        );

        // Check Cosine Similarity checking
        // Note: Distance::CosineDistance computes 1.0 - similarity
        // hsdlib returns similarity directly
        let dot: f32 = a.iter().zip(b.iter()).map(|(x, y)| x * y).sum();
        let norm_a: f32 = a.iter().map(|x| x * x).sum::<f32>().sqrt();
        let norm_b: f32 = b.iter().map(|x| x * x).sum::<f32>().sqrt();
        let scalar_cos_sim = dot / (norm_a * norm_b);

        let simd_cos_sim = hsdlib_ffi::cosine_f32(&a, &b).unwrap();
        assert!(
            (scalar_cos_sim - simd_cos_sim).abs() < 1e-4,
            "Cosine mismatch: scalar={}, simd={}",
            scalar_cos_sim,
            simd_cos_sim
        );
    }
}