diskann-wide 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
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150

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

// x86 intrinsics
use std::arch::x86_64::*;

use super::{f16x8, f16x16, f32x8, f32x16, i8x16, i16x8, i16x16, i32x8, u8x16, u32x8};
use crate::{LoHi, SIMDCast, SIMDReinterpret, SplitJoin, helpers};

/////////////////
// Conversions //
/////////////////

// f16 to f32
cfg_if::cfg_if! {
    // Miri does not have built-in support for these intrinsics.
    //
    // So when we run under Miri, we need to use emulated conversion.
    if #[cfg(not(miri))] {
        helpers::unsafe_map_conversion!(f16x8, f32x8, _mm256_cvtph_ps, "f16c");
        helpers::unsafe_map_cast!(
            f32x8 => (half::f16, f16x8),
            _mm256_cvtps_ph::<_MM_FROUND_TO_NEAREST_INT>,
            "f16c"
        );
        helpers::unsafe_map_cast!(f16x8 => (f32, f32x8), _mm256_cvtph_ps, "f16c");
    } else {
        use crate::{SIMDVector, reference::ReferenceCast};

        impl From<f16x8> for f32x8 {
            fn from(value: f16x8) -> f32x8 {
                let array = value.to_array();
                let converted = array.map(|i| i.reference_cast());
                f32x8::from_array(value.arch(), converted)
            }
        }

        impl SIMDCast<f32> for f16x8 {
            type Cast = f32x8;

            fn simd_cast(self) -> f32x8 {
                self.into()
            }
        }

        impl SIMDCast<half::f16> for f32x8 {
            type Cast = f16x8;

            fn simd_cast(self) -> f16x8 {
                let array = self.to_array();
                let converted = array.map(|i| i.reference_cast());
                f16x8::from_array(self.arch(), converted)
            }
        }
    }
}

impl From<f16x16> for f32x16 {
    #[inline(always)]
    fn from(value: f16x16) -> f32x16 {
        let LoHi { lo, hi } = value.split();
        f32x16::new(lo.into(), hi.into())
    }
}

impl SIMDCast<f32> for f16x16 {
    type Cast = f32x16;
    #[inline(always)]
    fn simd_cast(self) -> f32x16 {
        self.into()
    }
}

impl SIMDCast<half::f16> for f32x16 {
    type Cast = f16x16;
    #[inline(always)]
    fn simd_cast(self) -> f16x16 {
        self.split().map(|x| -> f16x8 { x.simd_cast() }).join()
    }
}

// i8 to i16
helpers::unsafe_map_conversion!(i8x16, i16x16, _mm256_cvtepi8_epi16, "avx2");

// u8 to i16
helpers::unsafe_map_conversion!(u8x16, i16x16, _mm256_cvtepu8_epi16, "avx2");

// i32 to f32
helpers::unsafe_map_cast!(i32x8 => (f32, f32x8), _mm256_cvtepi32_ps, "avx");

//////////////////
// Reinterprets //
//////////////////

impl SIMDReinterpret<i16x16> for u32x8 {
    fn reinterpret_simd(self) -> i16x16 {
        i16x16(self.0)
    }
}

impl SIMDReinterpret<u8x16> for i16x8 {
    fn reinterpret_simd(self) -> u8x16 {
        u8x16(self.0)
    }
}

impl SIMDReinterpret<i16x8> for u8x16 {
    fn reinterpret_simd(self) -> i16x8 {
        i16x8(self.0)
    }
}

///////////
// Tests //
///////////

#[cfg(test)]
mod test_x86_conversions {
    use super::*;
    use crate::{arch::x86_64::V3, test_utils};

    // Lossless Conversions
    #[cfg(not(miri))]
    test_utils::ops::test_lossless_convert!(
        f16x8 => f32x8, 0xa998182f02ff4d0d, V3::new_checked_uncached()
    );

    #[cfg(not(miri))]
    test_utils::ops::test_lossless_convert!(
        f16x16 => f32x16, 0xf63b356d1dfc1d52, V3::new_checked_uncached()
    );

    test_utils::ops::test_lossless_convert!(
        i8x16 => i16x16, 0x84602159fb122584, V3::new_checked_uncached()
    );
    test_utils::ops::test_lossless_convert!(
        u8x16 => i16x16, 0x5ba4b69df84ca568, V3::new_checked_uncached()
    );

    // Numeric Casts
    test_utils::ops::test_cast!(f16x8 => f32x8, 0x37314659b022466a, V3::new_checked_uncached());
    test_utils::ops::test_cast!(f16x16 => f32x16, 0xba8fe343fc9dbeff, V3::new_checked_uncached());

    test_utils::ops::test_cast!(f32x8 => f16x8, 0x8386cb0a7091cc3b, V3::new_checked_uncached());
    test_utils::ops::test_cast!(f32x16 => f16x16, 0x7b9c9afee7e6ac63, V3::new_checked_uncached());

    test_utils::ops::test_cast!(i32x8 => f32x8, 0xde4fbf25c554b29e, V3::new_checked_uncached());
}