seismic 0.2.1

Seismic is designed for effective and efficient KNN retrieval over learned sparse embeddings.
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

use crate::{utils::binary_search_branchless, DataType};

/// Computes the dot product between a dense query and a sparse vector.
/// Before using this function, the query must be made dense. This is much faster
/// than computing the dot product with a "merge" style.
///
/// # Arguments
///
/// * `query` - The dense query vector.
/// * `v_components` - The indices of the non-zero components in the vector.
/// * `v_values` - The values of the non-zero components in the vector.
///
/// # Returns
///
/// The dot product between the query and the vector.
///
/// # Examples
///
/// ```
/// use seismic::distances::dot_product_dense_sparse;
///
/// let query = [1.0, 2.0, 3.0, 0.0];
/// let v_components = [0, 2, 3];
/// let v_values = [1.0, 1.0, 1.5];
///
/// let result = dot_product_dense_sparse(&query, &v_components, &v_values);
/// assert_eq!(result, 4.0);
/// ```
#[inline]
#[must_use]
pub fn dot_product_dense_sparse<Q, V>(query: &[Q], v_components: &[u16], v_values: &[V]) -> f32
where
    Q: DataType,
    V: DataType,
{
    const N_LANES: usize = 4;

    let mut result = [0.0; N_LANES];
    let chunk_iter = v_components.iter().zip(v_values).array_chunks::<N_LANES>();

    for chunk in chunk_iter {
        //for i in 0..N_LANES { // Slightly faster withour this for.
        result[0] += query[*chunk[0].0 as usize].to_f32().unwrap() * (chunk[0].1.to_f32().unwrap());
        result[1] += query[*chunk[1].0 as usize].to_f32().unwrap() * chunk[1].1.to_f32().unwrap();
        result[2] += query[*chunk[2].0 as usize].to_f32().unwrap() * chunk[2].1.to_f32().unwrap();
        result[3] += query[*chunk[3].0 as usize].to_f32().unwrap() * chunk[3].1.to_f32().unwrap();
        //result[3] += unsafe { *query.get_unchecked(*chunk[3].0 as usize) } * *chunk[3].1;
        //}
    }

    let l = v_components.len();
    let rem = l % N_LANES;

    if rem > 0 {
        for (&i, &v) in v_components[l - rem..].iter().zip(&v_values[l - rem..]) {
            result[0] += query[i as usize].to_f32().unwrap() * v.to_f32().unwrap();
        }
    }

    result.iter().sum()

    // This is what we would like to write :-)
    // let mut result = 0.0;
    // for (&i, &v) in self.iter_vector(id) {
    //     result += query[i as usize] * v;
    // }

    // result
}

/// Computes the dot product between a sparse query and a sparse vector using binary search.
/// This function should be used when the query has just a few components.
/// Both the query's and vector's terms must be sorted by id.
///
/// # Arguments
///
/// * `query_term_ids` - The ids of the query terms.
/// * `query_values` - The values of the query terms.
/// * `v_terms_ids` - The ids of the vector terms.
/// * `v_values` - The values of the vector terms.
///
/// # Returns
///
/// The dot product between the query and the vector.
///
/// # Examples
///
/// ```
/// use seismic::distances::dot_product_with_binary_search;
///
/// let query_term_ids = [1, 2, 7];
/// let query_values = [1.0, 1.0, 1.0];
/// let v_term_ids = [0, 1, 2, 3, 4];
/// let v_values = [0.1, 1.0, 1.0, 1.0, 0.5];
///
/// let result = dot_product_with_binary_search(&query_term_ids, &query_values, &v_term_ids, &v_values);
/// assert_eq!(result, 2.0);
/// ```
#[inline]
#[must_use]
pub fn dot_product_with_binary_search<Q, V>(
    query_term_ids: &[u16],
    query_values: &[Q],
    v_terms_ids: &[u16],
    v_values: &[V],
) -> f32
where
    Q: DataType,
    V: DataType,
{
    let mut result = 0.0;

    for (&term_id, &value) in query_term_ids.iter().zip(query_values) {
        // Let's use a branchless binary search
        let i = binary_search_branchless(v_terms_ids, term_id);

        // SAFETY: result of binary search is always smaller than v_term_id.len() and v_values.len()
        let cmp = *unsafe { v_terms_ids.get_unchecked(i) } == term_id;
        result += if cmp {
            value.to_f32().unwrap() * unsafe { v_values.get_unchecked(i).to_f32().unwrap() }
        } else {
            0.0
        };
    }
    result
}

/// Computes the dot product between a query and a vector using merge style.
/// This function should be used when the query has just a few components.
/// Both the query's and vector's terms must be sorted by id.
///
/// # Arguments
///
/// * `query_term_ids` - The ids of the query terms.
/// * `query_values` - The values of the query terms.
/// * `v_term_ids` - The ids of the vector terms.
/// * `v_values` - The values of the vector terms.
///
/// # Returns
///
/// The dot product between the query and the vector.
///
/// # Examples
///
/// ```
/// use seismic::distances::dot_product_with_merge;
///
/// let query_term_ids = [1, 2, 7];
/// let query_values = [1.0, 1.0, 1.0];
/// let v_term_ids = [0, 1, 2, 3, 4];
/// let v_values = [0.1, 1.0, 1.0, 1.0, 0.5];
///
/// let result = dot_product_with_merge(&query_term_ids, &query_values, &v_term_ids, &v_values);
/// assert_eq!(result, 2.0);
/// ```
#[inline]
#[must_use]
pub fn dot_product_with_merge<Q, V>(
    query_term_ids: &[u16],
    query_values: &[Q],
    v_term_ids: &[u16],
    v_values: &[V],
) -> f32
where
    Q: DataType,
    V: DataType,
{
    let mut result = 0.0;
    let mut i = 0;
    for (&q_id, &q_v) in query_term_ids.iter().zip(query_values) {
        unsafe {
            while i < v_term_ids.len() && *v_term_ids.get_unchecked(i) < q_id {
                i += 1;
            }

            if i == v_term_ids.len() {
                break;
            }

            if *v_term_ids.get_unchecked(i) == q_id {
                result += (*v_values.get_unchecked(i)).to_f32().unwrap() * q_v.to_f32().unwrap();
            }
        }
    }
    result
}