hipvs/ivf_pq/
index.rs

1/*
2 * Copyright (c) 2024, NVIDIA CORPORATION.
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *     http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17use std::io::{stderr, Write};
18
19use crate::dlpack::ManagedTensor;
20use crate::error::{check_cuvs, Result};
21use crate::ivf_pq::{IndexParams, SearchParams};
22use crate::resources::Resources;
23
24/// Ivf-Pq ANN Index
25#[derive(Debug)]
26pub struct Index(ffi::cuvsIvfPqIndex_t);
27
28impl Index {
29    /// Builds a new Index from the dataset for efficient search.
30    ///
31    /// # Arguments
32    ///
33    /// * `res` - Resources to use
34    /// * `params` - Parameters for building the index
35    /// * `dataset` - A row-major matrix on either the host or device to index
36    pub fn build<T: Into<ManagedTensor>>(
37        res: &Resources,
38        params: &IndexParams,
39        dataset: T,
40    ) -> Result<Index> {
41        let dataset: ManagedTensor = dataset.into();
42        let index = Index::new()?;
43        unsafe {
44            check_cuvs(ffi::cuvsIvfPqBuild(
45                res.0,
46                params.0,
47                dataset.as_ptr(),
48                index.0,
49            ))?;
50        }
51        Ok(index)
52    }
53
54    /// Creates a new empty index
55    pub fn new() -> Result<Index> {
56        unsafe {
57            let mut index = std::mem::MaybeUninit::<ffi::cuvsIvfPqIndex_t>::uninit();
58            check_cuvs(ffi::cuvsIvfPqIndexCreate(index.as_mut_ptr()))?;
59            Ok(Index(index.assume_init()))
60        }
61    }
62
63    /// Perform a Approximate Nearest Neighbors search on the Index
64    ///
65    /// # Arguments
66    ///
67    /// * `res` - Resources to use
68    /// * `params` - Parameters to use in searching the index
69    /// * `queries` - A matrix in device memory to query for
70    /// * `neighbors` - Matrix in device memory that receives the indices of the nearest neighbors
71    /// * `distances` - Matrix in device memory that receives the distances of the nearest neighbors
72    pub fn search(
73        self,
74        res: &Resources,
75        params: &SearchParams,
76        queries: &ManagedTensor,
77        neighbors: &ManagedTensor,
78        distances: &ManagedTensor,
79    ) -> Result<()> {
80        unsafe {
81            check_cuvs(ffi::cuvsIvfPqSearch(
82                res.0,
83                params.0,
84                self.0,
85                queries.as_ptr(),
86                neighbors.as_ptr(),
87                distances.as_ptr(),
88            ))
89        }
90    }
91}
92
93impl Drop for Index {
94    fn drop(&mut self) {
95        if let Err(e) = check_cuvs(unsafe { ffi::cuvsIvfPqIndexDestroy(self.0) }) {
96            write!(stderr(), "failed to call cuvsIvfPqIndexDestroy {:?}", e)
97                .expect("failed to write to stderr");
98        }
99    }
100}
101
102#[cfg(test)]
103mod tests {
104    use super::*;
105    use ndarray::s;
106    use ndarray_rand::rand_distr::Uniform;
107    use ndarray_rand::RandomExt;
108
109    #[test]
110    fn test_ivf_pq() {
111        let build_params = IndexParams::new().unwrap().set_n_lists(64);
112
113        let res = Resources::new().unwrap();
114
115        // Create a new random dataset to index
116        let n_datapoints = 1024;
117        let n_features = 16;
118        let dataset =
119            ndarray::Array::<f32, _>::random((n_datapoints, n_features), Uniform::new(0., 1.0));
120
121        let dataset_device = ManagedTensor::from(&dataset).to_device(&res).unwrap();
122
123        // build the ivf-pq index
124        let index = Index::build(&res, &build_params, dataset_device)
125            .expect("failed to create ivf-pq index");
126
127        // use the first 4 points from the dataset as queries : will test that we get them back
128        // as their own nearest neighbor
129        let n_queries = 4;
130        let queries = dataset.slice(s![0..n_queries, ..]);
131
132        let k = 10;
133
134        // Ivf-Pq search API requires queries and outputs to be on device memory
135        // copy query data over, and allocate new device memory for the distances/ neighbors
136        // outputs
137        let queries = ManagedTensor::from(&queries).to_device(&res).unwrap();
138        let mut neighbors_host = ndarray::Array::<i64, _>::zeros((n_queries, k));
139        let neighbors = ManagedTensor::from(&neighbors_host)
140            .to_device(&res)
141            .unwrap();
142
143        let mut distances_host = ndarray::Array::<f32, _>::zeros((n_queries, k));
144        let distances = ManagedTensor::from(&distances_host)
145            .to_device(&res)
146            .unwrap();
147
148        let search_params = SearchParams::new().unwrap();
149
150        index
151            .search(&res, &search_params, &queries, &neighbors, &distances)
152            .unwrap();
153
154        // Copy back to host memory
155        distances.to_host(&res, &mut distances_host).unwrap();
156        neighbors.to_host(&res, &mut neighbors_host).unwrap();
157
158        // nearest neighbors should be themselves, since queries are from the
159        // dataset
160        assert_eq!(neighbors_host[[0, 0]], 0);
161        assert_eq!(neighbors_host[[1, 0]], 1);
162        assert_eq!(neighbors_host[[2, 0]], 2);
163        assert_eq!(neighbors_host[[3, 0]], 3);
164    }
165}
hipvs/ivf_pq/index.rs

hipvs/ivf_pq/
index.rs
