jxl 0.1.4

High performance Rust implementation of a JPEG XL decoder
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

// Copyright (c) the JPEG XL Project Authors. All rights reserved.
//
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

#![allow(unsafe_code)]

use std::{
    arch::x86_64::{
        __m256i, __m512i, _mm256_add_epi32, _mm256_blend_epi32, _mm256_loadu_si256,
        _mm256_permutevar8x32_epi32, _mm256_set1_epi32, _mm256_storeu_si256, _mm512_loadu_si512,
        _mm512_mask_permutevar_epi32, _mm512_permutex2var_epi32, _mm512_storeu_si512,
    },
    mem::MaybeUninit,
};

#[target_feature(enable = "avx2")]
fn load_avx2_u32(data: &[u32; 8]) -> __m256i {
    // SAFETY: `data` has the correct size.
    unsafe { _mm256_loadu_si256(data.as_ptr() as *const _) }
}

#[target_feature(enable = "avx2")]
fn load_avx2(data: &[u8; 32]) -> __m256i {
    // SAFETY: `data` has the correct size.
    unsafe { _mm256_loadu_si256(data.as_ptr() as *const _) }
}

#[target_feature(enable = "avx2")]
fn store_avx2(data: __m256i, out: &mut [MaybeUninit<u8>; 32]) {
    // SAFETY: `data` has the correct size.
    unsafe { _mm256_storeu_si256(out.as_mut_ptr() as *mut _, data) }
}

#[target_feature(enable = "avx512f")]
fn load_avx512_u32(data: &[u32; 16]) -> __m512i {
    // SAFETY: `data` has the correct size.
    unsafe { _mm512_loadu_si512(data.as_ptr() as *const _) }
}

#[target_feature(enable = "avx512f")]
fn load_avx512(data: &[u8; 64]) -> __m512i {
    // SAFETY: `data` has the correct size.
    unsafe { _mm512_loadu_si512(data.as_ptr() as *const _) }
}

#[target_feature(enable = "avx512f")]
fn store_avx512(data: __m512i, out: &mut [MaybeUninit<u8>; 64]) {
    // SAFETY: `data` has the correct size.
    unsafe { _mm512_storeu_si512(out.as_mut_ptr() as *mut _, data) }
}

#[target_feature(enable = "avx2")]
fn interleave3_32b_avx2(inp: &[&[u8]; 3], out: &mut [MaybeUninit<u8>]) -> usize {
    let [a, b, c] = inp;

    let idx_a0 = load_avx2_u32(&[0, 0, 0, 1, 0, 0, 2, 0]);
    // c1 = idx_a0 + 2
    // b2 = idx_a0 + 5

    let idx_b0 = load_avx2_u32(&[0, 0, 0, 0, 1, 0, 0, 2]);
    // a1 = idx_b0 + 3
    // c2 = idx_b0 + 5

    let idx_c0 = load_avx2_u32(&[0, 0, 0, 0, 0, 1, 0, 0]);
    // b1 = idx_c0 + 3
    // a2 = idx_c0 + 6

    let two = _mm256_set1_epi32(2);
    let three = _mm256_set1_epi32(3);
    let five = _mm256_set1_epi32(5);
    let six = _mm256_set1_epi32(6);

    const LEN: usize = 32;
    let mut processed = 0;
    for (((a, b), c), out) in a
        .chunks_exact(LEN)
        .zip(b.chunks_exact(LEN))
        .zip(c.chunks_exact(LEN))
        .zip(out.chunks_exact_mut(LEN * 3))
    {
        let a = load_avx2(a.try_into().unwrap());
        let b = load_avx2(b.try_into().unwrap());
        let c = load_avx2(c.try_into().unwrap());

        let a0 = _mm256_permutevar8x32_epi32(a, idx_a0);
        let b0 = _mm256_permutevar8x32_epi32(b, idx_b0);
        let c0 = _mm256_permutevar8x32_epi32(c, idx_c0);
        let out0 = _mm256_blend_epi32::<0b10010010>(a0, b0);
        let out0 = _mm256_blend_epi32::<0b00100100>(out0, c0);

        let a1 = _mm256_permutevar8x32_epi32(a, _mm256_add_epi32(idx_b0, three));
        let b1 = _mm256_permutevar8x32_epi32(b, _mm256_add_epi32(idx_c0, three));
        let c1 = _mm256_permutevar8x32_epi32(c, _mm256_add_epi32(idx_a0, two));
        let out1 = _mm256_blend_epi32::<0b00100100>(a1, b1);
        let out1 = _mm256_blend_epi32::<0b01001001>(out1, c1);

        let a2 = _mm256_permutevar8x32_epi32(a, _mm256_add_epi32(idx_c0, six));
        let b2 = _mm256_permutevar8x32_epi32(b, _mm256_add_epi32(idx_a0, five));
        let c2 = _mm256_permutevar8x32_epi32(c, _mm256_add_epi32(idx_b0, five));
        let out2 = _mm256_blend_epi32::<0b01001001>(a2, b2);
        let out2 = _mm256_blend_epi32::<0b10010010>(out2, c2);

        store_avx2(out0, (&mut out[0..LEN]).try_into().unwrap());
        store_avx2(out1, (&mut out[LEN..2 * LEN]).try_into().unwrap());
        store_avx2(out2, (&mut out[2 * LEN..3 * LEN]).try_into().unwrap());
        processed += LEN / 4;
    }

    processed
}

#[inline(never)]
#[target_feature(enable = "avx512f")]
fn interleave3_32b_avx512(inp: &[&[u8]; 3], out: &mut [MaybeUninit<u8>]) -> usize {
    let [a, b, c] = inp;

    let idx_ab0 = load_avx512_u32(&[0, 16, 0, 1, 17, 0, 2, 18, 0, 3, 19, 0, 4, 20, 0, 5]);
    let idx_c0 = load_avx512_u32(&[0, 0, 0, 0, 0, 1, 0, 0, 2, 0, 0, 3, 0, 0, 4, 0]);

    let idx_ab1 = load_avx512_u32(&[21, 0, 6, 22, 0, 7, 23, 0, 8, 24, 0, 9, 25, 0, 10, 26]);
    let idx_c1 = load_avx512_u32(&[0, 5, 0, 0, 6, 0, 0, 7, 0, 0, 8, 0, 0, 9, 0, 0]);

    let idx_ab2 = load_avx512_u32(&[0, 11, 27, 0, 12, 28, 0, 13, 29, 0, 14, 30, 0, 15, 31, 0]);
    let idx_c2 = load_avx512_u32(&[10, 0, 0, 11, 0, 0, 12, 0, 0, 13, 0, 0, 14, 0, 0, 15]);

    const LEN: usize = 64;
    let mut processed = 0;
    for (((a, b), c), out) in a
        .chunks_exact(LEN)
        .zip(b.chunks_exact(LEN))
        .zip(c.chunks_exact(LEN))
        .zip(out.chunks_exact_mut(LEN * 3))
    {
        let a = load_avx512(a.try_into().unwrap());
        let b = load_avx512(b.try_into().unwrap());
        let c = load_avx512(c.try_into().unwrap());

        let out0 = _mm512_permutex2var_epi32(a, idx_ab0, b);
        let out0 = _mm512_mask_permutevar_epi32(out0, 0b0100100100100100, idx_c0, c);

        let out1 = _mm512_permutex2var_epi32(a, idx_ab1, b);
        let out1 = _mm512_mask_permutevar_epi32(out1, 0b0010010010010010, idx_c1, c);

        let out2 = _mm512_permutex2var_epi32(a, idx_ab2, b);
        let out2 = _mm512_mask_permutevar_epi32(out2, 0b1001001001001001, idx_c2, c);

        store_avx512(out0, (&mut out[0..LEN]).try_into().unwrap());
        store_avx512(out1, (&mut out[LEN..2 * LEN]).try_into().unwrap());
        store_avx512(out2, (&mut out[2 * LEN..3 * LEN]).try_into().unwrap());
        processed += LEN / 4;
    }

    processed
}

/// Safety note: does not write uninit data in `out`.
pub(super) fn interleave3_32b(inp: &[&[u8]; 3], out: &mut [MaybeUninit<u8>]) -> usize {
    if is_x86_feature_detected!("avx512f") {
        // SAFETY: we just checked for avx512f.
        unsafe { interleave3_32b_avx512(inp, out) }
    } else if is_x86_feature_detected!("avx2") {
        // SAFETY: we just checked for avx2.
        unsafe { interleave3_32b_avx2(inp, out) }
    } else {
        0
    }
}