diskann-quantization 0.51.0

DiskANN is a fast approximate nearest neighbor search library for high dimensional data
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

// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT license.

use diskann_wide::Architecture;
use diskann_wide::arch::Scalar;
#[cfg(target_arch = "aarch64")]
use diskann_wide::arch::aarch64::Neon;
#[cfg(target_arch = "x86_64")]
use diskann_wide::arch::x86_64::{V3, V4};

use super::{DynQueryComputer, Prepared, QueryComputer, build_prepared};
use crate::multi_vector::distance::kernels::f16::F16Entry;
use crate::multi_vector::{BlockTransposed, BlockTransposedRef, MatRef, Standard};
use diskann_utils::Reborrow;

impl QueryComputer<half::f16> {
    /// Build an f16 query computer, selecting the optimal architecture and
    /// GROUP for the current CPU at runtime.
    pub fn new(query: MatRef<'_, Standard<half::f16>>) -> Self {
        diskann_wide::arch::dispatch1_no_features(BuildComputer, query)
    }
}

impl<A, const GROUP: usize> DynQueryComputer<half::f16>
    for Prepared<A, BlockTransposed<half::f16, GROUP>>
where
    A: Architecture,
    F16Entry<GROUP>: for<'a> diskann_wide::arch::Target3<
            A,
            (),
            BlockTransposedRef<'a, half::f16, GROUP>,
            MatRef<'a, Standard<half::f16>>,
            &'a mut [f32],
        >,
{
    fn compute_max_sim(&self, doc: MatRef<'_, Standard<half::f16>>, scores: &mut [f32]) {
        let mut scratch = vec![f32::MIN; self.prepared.padded_nrows()];
        self.arch.run3(
            F16Entry::<GROUP>,
            self.prepared.reborrow(),
            doc,
            &mut scratch,
        );
        for (dst, &src) in scores.iter_mut().zip(&scratch[..self.prepared.nrows()]) {
            *dst = -src;
        }
    }

    fn nrows(&self) -> usize {
        self.prepared.nrows()
    }
}

#[derive(Debug, Clone, Copy)]
pub(super) struct BuildComputer;

impl diskann_wide::arch::Target1<Scalar, QueryComputer<half::f16>, MatRef<'_, Standard<half::f16>>>
    for BuildComputer
{
    fn run(self, arch: Scalar, query: MatRef<'_, Standard<half::f16>>) -> QueryComputer<half::f16> {
        QueryComputer {
            inner: Box::new(build_prepared::<half::f16, _, 8>(arch, query)),
        }
    }
}

#[cfg(target_arch = "x86_64")]
impl diskann_wide::arch::Target1<V3, QueryComputer<half::f16>, MatRef<'_, Standard<half::f16>>>
    for BuildComputer
{
    fn run(self, arch: V3, query: MatRef<'_, Standard<half::f16>>) -> QueryComputer<half::f16> {
        QueryComputer {
            inner: Box::new(build_prepared::<half::f16, _, 16>(arch, query)),
        }
    }
}

#[cfg(target_arch = "x86_64")]
impl diskann_wide::arch::Target1<V4, QueryComputer<half::f16>, MatRef<'_, Standard<half::f16>>>
    for BuildComputer
{
    fn run(self, arch: V4, query: MatRef<'_, Standard<half::f16>>) -> QueryComputer<half::f16> {
        let arch = arch.retarget();
        QueryComputer {
            inner: Box::new(build_prepared::<half::f16, _, 16>(arch, query)),
        }
    }
}

#[cfg(target_arch = "aarch64")]
impl diskann_wide::arch::Target1<Neon, QueryComputer<half::f16>, MatRef<'_, Standard<half::f16>>>
    for BuildComputer
{
    fn run(self, arch: Neon, query: MatRef<'_, Standard<half::f16>>) -> QueryComputer<half::f16> {
        let arch = arch.retarget();
        QueryComputer {
            inner: Box::new(build_prepared::<half::f16, _, 8>(arch, query)),
        }
    }
}