mopenzl 0.2.3

OpenZL compression and decompression in safe Rust; encoding is byte-identical to Meta's reference library
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

//! Byte-column transposition kernels of the `transpose` codec.
//!
//! The encoder splits a stream of `W`-byte elements into `W` byte columns
//! ([`split`]); the decoder interleaves the columns back ([`merge`]). For
//! the widths 2, 4 and 8 both run on 16-byte vectors: `log2(W)` rounds of
//! even/odd byte extraction (`unzip`) deinterleave sixteen elements, and the
//! same number of `zip` rounds put them back. Other widths, and the elements
//! after the last full group of sixteen, take the scalar loops, which are
//! also the reference the vector paths are tested against.

#[cfg(any(feature = "compression", test))]
use alloc::vec::Vec;

use fearless_simd::{Level, Simd, SimdBase, dispatch, u8x16};

/// Elements per vector group.
const GROUP: usize = 16;

/// Splits `src` (elements of `lanes.len()` bytes) into one column per byte:
/// `lanes[j]` receives byte `j` of every element, `n` bytes each (the
/// vectors are resized to `n`).
#[cfg(any(feature = "compression", test))]
pub(crate) fn split(level: Level, src: &[u8], lanes: &mut [Vec<u8>]) {
    let width = lanes.len();
    let n = src.len() / width;
    for lane in lanes.iter_mut() {
        lane.clear();
        lane.resize(n, 0);
    }
    let done = match width {
        2 => dispatch!(level, simd => split_groups::<_, 2>(simd, src, lanes)),
        4 => dispatch!(level, simd => split_groups::<_, 4>(simd, src, lanes)),
        8 => dispatch!(level, simd => split_groups::<_, 8>(simd, src, lanes)),
        _ => 0,
    };
    split_scalar(src, lanes, done);
}

/// The scalar columns of the elements from `from` on.
#[cfg(any(feature = "compression", test))]
fn split_scalar(src: &[u8], lanes: &mut [Vec<u8>], from: usize) {
    let width = lanes.len();
    for (index, element) in src.chunks_exact(width).enumerate().skip(from) {
        for (lane, &byte) in lanes.iter_mut().zip(element) {
            lane[index] = byte;
        }
    }
}

/// [`split`] of the full groups of sixteen elements, returning the elements
/// done.
#[cfg(any(feature = "compression", test))]
#[inline(always)]
fn split_groups<S: Simd, const W: usize>(simd: S, src: &[u8], lanes: &mut [Vec<u8>]) -> usize {
    let groups = lanes[0].len() / GROUP;
    for (group, bytes) in src.chunks_exact(GROUP * W).take(groups).enumerate() {
        let mut v: [u8x16<S>; W] =
            core::array::from_fn(|i| u8x16::from_slice(simd, &bytes[i * GROUP..(i + 1) * GROUP]));
        for _ in 0..W.trailing_zeros() {
            let previous = v;
            for j in 0..W / 2 {
                v[j] = simd.unzip_low_u8x16(previous[2 * j], previous[2 * j + 1]);
                v[W / 2 + j] = simd.unzip_high_u8x16(previous[2 * j], previous[2 * j + 1]);
            }
        }
        for (lane, vector) in lanes.iter_mut().zip(v) {
            vector.store_slice(&mut lane[group * GROUP..(group + 1) * GROUP]);
        }
    }
    groups * GROUP
}

/// Interleaves the byte `columns` (each `n` bytes long) into the `n`
/// elements of `columns.len()` bytes that fill `dst`.
#[cfg(any(feature = "decompression", test))]
pub(crate) fn merge(level: Level, dst: &mut [u8], columns: &[&[u8]], n: usize) {
    let done = match columns.len() {
        2 => dispatch!(level, simd => merge_groups::<_, 2>(simd, dst, columns, n)),
        4 => dispatch!(level, simd => merge_groups::<_, 4>(simd, dst, columns, n)),
        8 => dispatch!(level, simd => merge_groups::<_, 8>(simd, dst, columns, n)),
        _ => 0,
    };
    merge_scalar(dst, columns, n, done);
}

/// The scalar elements from `from` on.
#[cfg(any(feature = "decompression", test))]
fn merge_scalar(dst: &mut [u8], columns: &[&[u8]], n: usize, from: usize) {
    let width = columns.len();
    for (position, column) in columns.iter().enumerate() {
        for (element, &byte) in dst.chunks_exact_mut(width).zip(&column[..n]).skip(from) {
            element[position] = byte;
        }
    }
}

/// [`merge`] of the full groups of sixteen elements, returning the elements
/// done.
#[cfg(any(feature = "decompression", test))]
#[inline(always)]
fn merge_groups<S: Simd, const W: usize>(
    simd: S,
    dst: &mut [u8],
    columns: &[&[u8]],
    n: usize,
) -> usize {
    let groups = n / GROUP;
    for (group, out) in dst.chunks_exact_mut(GROUP * W).take(groups).enumerate() {
        let mut v: [u8x16<S>; W] = core::array::from_fn(|i| {
            u8x16::from_slice(simd, &columns[i][group * GROUP..(group + 1) * GROUP])
        });
        for _ in 0..W.trailing_zeros() {
            let previous = v;
            for j in 0..W / 2 {
                let (low, high) = simd.interleave_u8x16(previous[j], previous[W / 2 + j]);
                v[2 * j] = low;
                v[2 * j + 1] = high;
            }
        }
        for (i, vector) in v.into_iter().enumerate() {
            vector.store_slice(&mut out[i * GROUP..(i + 1) * GROUP]);
        }
    }
    groups * GROUP
}

#[cfg(all(test, feature = "compression", feature = "decompression"))]
mod tests {
    use super::*;
    use crate::Backend;

    fn sample(len: usize) -> Vec<u8> {
        (0..len)
            .map(|i| (i.wrapping_mul(0x9E37_79B1) >> 13) as u8)
            .collect()
    }

    #[test]
    fn split_matches_the_scalar_columns_at_every_level() {
        let counts: &[usize] = if cfg!(miri) {
            &[0, 1, 15, 16, 17, 40]
        } else {
            &[0, 1, 15, 16, 17, 31, 32, 33, 100, 1000]
        };
        for width in [1usize, 2, 3, 4, 5, 8, 12] {
            for &n in counts {
                let src = sample(n * width);
                let mut expected: Vec<Vec<u8>> = (0..width).map(|_| alloc::vec![0; n]).collect();
                split_scalar(&src, &mut expected, 0);
                for backend in Backend::available() {
                    let mut lanes: Vec<Vec<u8>> = (0..width).map(|_| Vec::new()).collect();
                    split(backend.0, &src, &mut lanes);
                    assert_eq!(lanes, expected, "{backend} w{width} n{n}");
                }
            }
        }
    }

    #[test]
    fn merge_inverts_split_at_every_level() {
        let counts: &[usize] = if cfg!(miri) {
            &[0, 1, 16, 17, 40]
        } else {
            &[0, 1, 15, 16, 17, 33, 100, 1000]
        };
        for width in [1usize, 2, 3, 4, 8, 16] {
            for &n in counts {
                let src = sample(n * width);
                for backend in Backend::available() {
                    let mut lanes: Vec<Vec<u8>> = (0..width).map(|_| Vec::new()).collect();
                    split(backend.0, &src, &mut lanes);
                    let columns: Vec<&[u8]> = lanes.iter().map(Vec::as_slice).collect();
                    let mut dst = alloc::vec![0u8; n * width];
                    merge(backend.0, &mut dst, &columns, n);
                    assert_eq!(dst, src, "{backend} w{width} n{n}");
                }
            }
        }
    }
}