kannolo 0.1.6

kANNolo is designed for easy prototyping of ANN Search algorithms while ensuring high effectiveness and efficiency over both dense and sparse vectors.
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

use crate::distances::{self, dense_dot_product_unrolled, dense_euclidean_distance_unrolled};
use crate::quantizers::quantizer::{Quantizer, QueryEvaluator};
use crate::topk_selectors::OnlineTopKSelector;
use crate::{dot_product_batch_4, euclidean_distance_batch_4, DArray1, DenseDArray1};
use crate::{Dataset, DistanceType, Float};

use crate::datasets::dense_dataset::DenseDataset;

use serde::{Deserialize, Serialize};
use std::marker::PhantomData;

#[derive(Default, Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct PlainQuantizer<T> {
    d: usize,
    distance: DistanceType,
    _phantom: PhantomData<T>,
}

impl<T> PlainQuantizer<T> {
    pub fn new(d: usize, distance: DistanceType) -> Self {
        PlainQuantizer {
            d,
            distance,
            _phantom: PhantomData,
        }
    }
}

impl<T: Copy + Default + PartialOrd + Sync + Send> Quantizer for PlainQuantizer<T> {
    type InputItem = T;
    type OutputItem = T;

    type DatasetType<'a>
        = DenseDataset<Self>
    where
        T: 'a;
    type Evaluator<'a>
        = QueryEvaluatorPlain<'a, Self::InputItem>
    where
        Self::InputItem: Float
            + 'a
            + distances::euclidean_distance::EuclideanDistance<T>
            + distances::dot_product::DotProduct<T>;

    #[inline]
    fn encode(&self, input_vectors: &[Self::InputItem], output_vectors: &mut [Self::OutputItem]) {
        output_vectors.copy_from_slice(input_vectors);
    }

    #[inline]
    fn m(&self) -> usize {
        self.d
    }

    #[inline]
    fn distance(&self) -> DistanceType {
        self.distance
    }

    fn get_space_usage_bytes(&self) -> usize {
        std::mem::size_of::<usize>()
    }
}

pub struct QueryEvaluatorPlain<
    'a,
    T: Float
        + 'a
        + distances::euclidean_distance::EuclideanDistance<T>
        + distances::dot_product::DotProduct<T>,
> {
    dataset: &'a <<Self as QueryEvaluator<'a>>::Q as Quantizer>::DatasetType<'a>,
    query: <Self as QueryEvaluator<'a>>::QueryType,
}

impl<'a, T: Float> QueryEvaluator<'a> for QueryEvaluatorPlain<'a, T>
where
    T: Float
        + distances::euclidean_distance::EuclideanDistance<T>
        + distances::dot_product::DotProduct<T>,
{
    type Q = PlainQuantizer<T>;
    type QueryType = DenseDArray1<&'a [T]>;

    #[inline]
    fn new(dataset: &'a <Self::Q as Quantizer>::DatasetType<'a>, query: Self::QueryType) -> Self {
        Self { dataset, query }
    }

    fn compute_distance(&self, index: usize) -> f32 {
        let document = self.dataset.get(index);
        match self.dataset.quantizer().distance() {
            DistanceType::Euclidean => dense_euclidean_distance_unrolled(&self.query, &document),
            DistanceType::DotProduct => -dense_dot_product_unrolled(&self.query, &document),
        }
    }

    #[inline]
    fn compute_four_distances(
        &self,
        indexes: impl IntoIterator<Item = usize>,
    ) -> impl Iterator<Item = f32> {
        let chunk: Vec<_> = indexes.into_iter().map(|id| self.dataset.get(id)).collect();
        let query_slice = self.query.values_as_slice();
        let quantizer = self.dataset.quantizer();

        // Process exactly 4 vectors
        let v0 = chunk[0].values_as_slice();
        let v1 = chunk[1].values_as_slice();
        let v2 = chunk[2].values_as_slice();
        let v3 = chunk[3].values_as_slice();
        let vector_batch = [&v0[..], &v1[..], &v2[..], &v3[..]];

        let dist = match quantizer.distance() {
            DistanceType::Euclidean => euclidean_distance_batch_4(query_slice, vector_batch),
            DistanceType::DotProduct => {
                let dps = dot_product_batch_4(query_slice, vector_batch); // Negate distances
                [-dps[0], -dps[1], -dps[2], -dps[3]]
            }
        };

        dist.into_iter()
    }

    #[inline]
    fn topk_retrieval<I, H>(&self, distances: I, heap: &mut H) -> Vec<(f32, usize)>
    where
        I: Iterator<Item = f32>,
        H: OnlineTopKSelector,
    {
        for distance in distances {
            heap.push(distance);
        }

        heap.topk()
    }
}