jxl-render 0.12.4

JPEG XL image renderer, part of jxl-oxide
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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197

use std::arch::aarch64::*;
use std::arch::is_aarch64_feature_detected;

use jxl_grid::{MutableSubgrid, SharedSubgrid};
use jxl_modular::ChannelShift;
use jxl_vardct::{BlockInfo, TransformType};

use crate::vardct::{dct_common::DctDirection, transform_common::transform_varblocks_inner};

use super::generic;

#[target_feature(enable = "neon")]
unsafe fn transform_dct4_aarch64_neon(coeff: &mut MutableSubgrid<'_>) {
    generic::aux_idct2_in_place_2(coeff);

    let mut scratch_0 = [vdupq_n_f32(0.0); 4];
    let mut scratch_1 = [vdupq_n_f32(0.0); 4];
    for y2 in 0..4 {
        let row_ptr = coeff.get_row(y2 * 2).as_ptr();
        let float32x4x2_t(a, b) = vld2q_f32(row_ptr);
        scratch_0[y2] = a;
        scratch_1[y2] = b;
    }

    let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_0);
    let mut scratch_0 = super::dct::dct4_inverse([v0, v1, v2, v3]);
    let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_1);
    let mut scratch_1 = super::dct::dct4_inverse([v0, v1, v2, v3]);
    for y2 in 0..4 {
        let row_ptr = coeff.get_row_mut(y2).as_mut_ptr();
        vst1q_f32(row_ptr, super::dct::dct4_vec_inverse(scratch_0[y2]));
        vst1q_f32(row_ptr.add(4), super::dct::dct4_vec_inverse(scratch_1[y2]));

        let row_ptr = coeff.get_row(y2 * 2 + 1).as_ptr();
        let float32x4x2_t(a, b) = vld2q_f32(row_ptr);
        scratch_0[y2] = a;
        scratch_1[y2] = b;
    }

    let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_0);
    let scratch_0 = super::dct::dct4_inverse([v0, v1, v2, v3]);
    let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_1);
    let scratch_1 = super::dct::dct4_inverse([v0, v1, v2, v3]);
    for y in 0..4 {
        let row_ptr = coeff.get_row_mut(y + 4).as_mut_ptr();
        vst1q_f32(row_ptr, super::dct::dct4_vec_inverse(scratch_0[y]));
        vst1q_f32(row_ptr.add(4), super::dct::dct4_vec_inverse(scratch_1[y]));
    }
}

#[target_feature(enable = "neon")]
unsafe fn transform_dct4x8_aarch64_neon<const TR: bool>(coeff: &mut MutableSubgrid<'_>) {
    let coeff0 = coeff.get(0, 0);
    let coeff1 = coeff.get(0, 1);
    *coeff.get_mut(0, 0) = coeff0 + coeff1;
    *coeff.get_mut(0, 1) = coeff0 - coeff1;

    if TR {
        let mut scratch_0 = [vdupq_n_f32(0.0); 4];
        let mut scratch_1 = [vdupq_n_f32(0.0); 4];
        for y2 in 0..4 {
            let row_ptr = coeff.get_row(y2 * 2).as_ptr();
            let a = vld1q_f32(row_ptr);
            let b = vld1q_f32(row_ptr.add(4));
            let (l, r) = super::dct::dct8_vec_inverse(a, b);
            scratch_0[y2] = l;
            scratch_1[y2] = r;
        }

        let scratch_0 = super::dct::dct4_inverse(scratch_0);
        let scratch_1 = super::dct::dct4_inverse(scratch_1);
        let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_0);
        let mut scratch_0 = [v0, v1, v2, v3];
        let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_1);
        let mut scratch_1 = [v0, v1, v2, v3];
        for y2 in 0..4 {
            let y = [1, 5, 3, 7][y2];
            let row_ptr = coeff.get_row(y).as_ptr();
            let a = vld1q_f32(row_ptr);
            let b = vld1q_f32(row_ptr.add(4));
            let (l, r) = super::dct::dct8_vec_inverse(a, b);

            let row_ptr = coeff.get_row_mut(y2).as_mut_ptr();
            vst1q_f32(row_ptr, scratch_0[y2]);
            let row_ptr = coeff.get_row_mut(y2 + 4).as_mut_ptr();
            vst1q_f32(row_ptr, scratch_1[y2]);

            scratch_0[y2] = l;
            scratch_1[y2] = r;
        }
        scratch_0.swap(1, 2);
        scratch_1.swap(1, 2);

        let scratch_0 = super::dct::dct4_inverse(scratch_0);
        let scratch_1 = super::dct::dct4_inverse(scratch_1);
        let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_0);
        let scratch_0 = [v0, v1, v2, v3];
        let float32x4x4_t(v0, v1, v2, v3) = super::dct::transpose_lane(&scratch_1);
        let scratch_1 = [v0, v1, v2, v3];
        for y in 0..4 {
            let row_ptr = coeff.get_row_mut(y).as_mut_ptr().add(4);
            vst1q_f32(row_ptr, scratch_0[y]);
            let row_ptr = coeff.get_row_mut(y + 4).as_mut_ptr().add(4);
            vst1q_f32(row_ptr, scratch_1[y]);
        }
    } else {
        let mut scratch_0 = [vdupq_n_f32(0.0); 4];
        let mut scratch_1 = [vdupq_n_f32(0.0); 4];
        for y2 in 0..4 {
            let row_ptr = coeff.get_row(y2 * 2).as_ptr();
            let a = vld1q_f32(row_ptr);
            let b = vld1q_f32(row_ptr.add(4));
            let (l, r) = super::dct::dct8_vec_inverse(a, b);
            scratch_0[y2] = l;
            scratch_1[y2] = r;
        }

        let mut scratch_0 = super::dct::dct4_inverse(scratch_0);
        let mut scratch_1 = super::dct::dct4_inverse(scratch_1);
        for y2 in 0..4 {
            let row_ptr = coeff.get_row_mut(y2).as_mut_ptr();
            vst1q_f32(row_ptr, scratch_0[y2]);
            vst1q_f32(row_ptr.add(4), scratch_1[y2]);

            let row_ptr = coeff.get_row(y2 * 2 + 1).as_ptr();
            let a = vld1q_f32(row_ptr);
            let b = vld1q_f32(row_ptr.add(4));
            let (l, r) = super::dct::dct8_vec_inverse(a, b);
            scratch_0[y2] = l;
            scratch_1[y2] = r;
        }

        let scratch_0 = super::dct::dct4_inverse(scratch_0);
        let scratch_1 = super::dct::dct4_inverse(scratch_1);
        for y in 0..4 {
            let row_ptr = coeff.get_row_mut(y + 4).as_mut_ptr();
            vst1q_f32(row_ptr, scratch_0[y]);
            vst1q_f32(row_ptr.add(4), scratch_1[y]);
        }
    }
}

#[target_feature(enable = "neon")]
unsafe fn transform_dct_aarch64_neon(coeff: &mut MutableSubgrid<'_>) {
    super::dct::dct_2d_aarch64_neon(coeff, DctDirection::Inverse);
}

#[target_feature(enable = "neon")]
unsafe fn transform_aarch64_neon(coeff: &mut MutableSubgrid<'_>, dct_select: TransformType) {
    use TransformType::*;

    match dct_select {
        Dct2 => generic::transform_dct2(coeff),
        Dct4 => transform_dct4_aarch64_neon(coeff),
        Hornuss => generic::transform_hornuss(coeff),
        Dct4x8 => transform_dct4x8_aarch64_neon::<false>(coeff),
        Dct8x4 => transform_dct4x8_aarch64_neon::<true>(coeff),
        Afv0 => generic::transform_afv::<0>(coeff),
        Afv1 => generic::transform_afv::<1>(coeff),
        Afv2 => generic::transform_afv::<2>(coeff),
        Afv3 => generic::transform_afv::<3>(coeff),
        _ => transform_dct_aarch64_neon(coeff),
    }
}

#[target_feature(enable = "neon")]
unsafe fn transform_varblocks_aarch64_neon(
    lf: &[SharedSubgrid<f32>; 3],
    coeff_out: &mut [MutableSubgrid<'_, f32>; 3],
    shifts_cbycr: [ChannelShift; 3],
    block_info: &SharedSubgrid<BlockInfo>,
) {
    transform_varblocks_inner(
        lf,
        coeff_out,
        shifts_cbycr,
        block_info,
        super::dct::dct_2d_aarch64_neon,
        transform_aarch64_neon,
    );
}

#[inline]
pub fn transform_varblocks(
    lf: &[SharedSubgrid<f32>; 3],
    coeff_out: &mut [MutableSubgrid<'_, f32>; 3],
    shifts_cbycr: [ChannelShift; 3],
    block_info: &SharedSubgrid<BlockInfo>,
) {
    if is_aarch64_feature_detected!("neon") {
        unsafe {
            return transform_varblocks_aarch64_neon(lf, coeff_out, shifts_cbycr, block_info);
        }
    }

    generic::transform_varblocks(lf, coeff_out, shifts_cbycr, block_info);
}