diskann-quantization 0.50.1

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
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

/*
 * Copyright (c) Microsoft Corporation.
 * Licensed under the MIT license.
 */

//! The functions defined in this module are meant for the purposes of automated disassambly
//! to inspect code generation.
//!
//! These functions should not be called as a replacement for their publically exposed
//! alternatives.

use diskann_wide::{
    Architecture,
    arch::x86_64::{V3, V4},
};

use crate::{
    bits::{BitSlice, BitTranspose, Dense, Unsigned},
    distances::{self, MathematicalResult},
};

//////////
// Bits //
//////////

type USlice<'a, const N: usize, Perm = Dense> = BitSlice<'a, N, Unsigned, Perm>;
type MR<T> = MathematicalResult<T>;

//----------------//
// Regular Slices //
//----------------//

// V3 //

#[inline(never)]
pub fn bits_v3_l2_bu8_bu8(arch: V3, x: USlice<'_, 8>, y: USlice<'_, 8>) -> MR<u32> {
    arch.run2_inline(distances::SquaredL2, x, y)
}

#[inline(never)]
pub fn bits_v3_l2_bu4_bu4(arch: V3, x: USlice<'_, 4>, y: USlice<'_, 4>) -> MR<u32> {
    arch.run2_inline(distances::SquaredL2, x, y)
}

#[inline(never)]
pub fn bits_v3_l2_bu2_bu2(arch: V3, x: USlice<'_, 2>, y: USlice<'_, 2>) -> MR<u32> {
    arch.run2_inline(distances::SquaredL2, x, y)
}

#[inline(never)]
pub fn bits_v3_l2_bu1_bu1(arch: V3, x: USlice<'_, 1>, y: USlice<'_, 1>) -> MR<u32> {
    arch.run2_inline(distances::SquaredL2, x, y)
}

#[inline(never)]
pub fn bits_v3_ip_bu8_bu8(arch: V3, x: USlice<'_, 8>, y: USlice<'_, 8>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

#[inline(never)]
pub fn bits_v3_ip_bu4_bu4(arch: V3, x: USlice<'_, 4>, y: USlice<'_, 4>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

#[inline(never)]
pub fn bits_v3_ip_bu2_bu2(arch: V3, x: USlice<'_, 2>, y: USlice<'_, 2>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

#[inline(never)]
pub fn bits_v3_ip_bu1_bu1(arch: V3, x: USlice<'_, 1>, y: USlice<'_, 1>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

// V4 //

#[inline(never)]
pub fn bits_v4_l2_bu1_bu1(arch: V4, x: USlice<'_, 1>, y: USlice<'_, 1>) -> MR<u32> {
    arch.run2_inline(distances::SquaredL2, x, y)
}

#[inline(never)]
pub fn bits_v4_ip_bu1_bu1(arch: V4, x: USlice<'_, 1>, y: USlice<'_, 1>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

#[inline(never)]
pub fn bits_v4_ip_bu2_bu2(arch: V4, x: USlice<'_, 2>, y: USlice<'_, 2>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

//------------//
// Transposed //
//------------//

// V3 //

#[inline(never)]
pub fn bits_v3_ip_tu4_bu1(arch: V3, x: USlice<'_, 4, BitTranspose>, y: USlice<'_, 1>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

// V4 //

#[inline(never)]
pub fn bits_v4_ip_tu4_bu1(arch: V4, x: USlice<'_, 4, BitTranspose>, y: USlice<'_, 1>) -> MR<u32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

//----------------//
// Full Precision //
//----------------//

#[inline(never)]
pub fn bits_v3_ip_f32_bu8(arch: V3, x: &[f32], y: USlice<'_, 8>) -> MR<f32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

#[inline(never)]
pub fn bits_v3_ip_f32_bu4(arch: V3, x: &[f32], y: USlice<'_, 4>) -> MR<f32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

#[inline(never)]
pub fn bits_v3_ip_f32_bu2(arch: V3, x: &[f32], y: USlice<'_, 2>) -> MR<f32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}

#[inline(never)]
pub fn bits_v3_ip_f32_bu1(arch: V3, x: &[f32], y: USlice<'_, 1>) -> MR<f32> {
    arch.run2_inline(distances::InnerProduct, x, y)
}