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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
/*
* // Copyright (c) Radzivon Bartoshyk 10/2026. All rights reserved.
* //
* // Redistribution and use in source and binary forms, with or without modification,
* // are permitted provided that the following conditions are met:
* //
* // 1. Redistributions of source code must retain the above copyright notice, this
* // list of conditions and the following disclaimer.
* //
* // 2. Redistributions in binary form must reproduce the above copyright notice,
* // this list of conditions and the following disclaimer in the documentation
* // and/or other materials provided with the distribution.
* //
* // 3. Neither the name of the copyright holder nor the names of its
* // contributors may be used to endorse or promote products derived from
* // this software without specific prior written permission.
* //
* // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
* // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
//! Saturated pixels are free beyond the display range. The decoder clips RGB,
//! so a reconstruction that overshoots white at a white pixel, or undershoots
//! black at a black one, shows no error, while the coefficient-domain
//! distortion charges it in full. Before the forward transform, each such
//! pixel takes the value the block quantizer reconstructs there, as long as it
//! stays beyond the limit; a hard edge becomes a smooth wave whose lobes land
//! in the clipped region. B moves with Y along the gray axis, where CfL
//! carries the shift at no chroma cost.
use crate::encoding_context::EncodingContext;
const ITERATIONS: usize = 3;
/// Least share of the footprint at a limit for the pass to run.
const MIN_SATURATED_SHARE: f32 = 0.02;
/// Distance from the white or black Y within which a pixel is at that limit.
const Y_TOLERANCE: f32 = 0.002;
const MAX_COVERED_BLOCKS: usize = 16;
const NORMAL: u8 = 0;
const WHITE: u8 = 1;
const BLACK: u8 = 2;
#[derive(Default)]
pub(crate) struct ClipScratch {
source: Vec<f32>,
target: Vec<f32>,
/// Target minus source per footprint pixel, valid after a `true` return.
shift: Vec<f32>,
coefficients: Vec<f32>,
levels: Vec<i32>,
pixels: Vec<f32>,
class: Vec<u8>,
shifted: Vec<f32>,
}
/// Moves the saturated pixels of one luma footprint beyond their limit.
/// `input` is the footprint's luma, `stride` wide, replicated past the image
/// edge. Returns whether any pixel
/// moved; the shift then applies to the Y and B inputs through
/// [`shifted_input`].
#[allow(clippy::too_many_arguments)]
pub(crate) fn continue_saturated(
ctx: &EncodingContext,
raw_strategy: u8,
cx: usize,
cy: usize,
inv_qm: &[f32],
qm: &[f32],
quant_ac: i32,
scale: f32,
distance: f32,
input: &[f32],
stride: usize,
scratch: &mut ClipScratch,
) -> bool {
let covered = cx * cy;
if covered > MAX_COVERED_BLOCKS {
return false;
}
let n = covered * 64;
let (pw, ph) = crate::inflated_cost::strategy_pixel_dims(raw_strategy);
debug_assert_eq!(pw * ph, n);
let (_, white, _) = crate::xyb::rgb_to_xyb_pixel_f32(&ctx.xyb, 1.0, 1.0, 1.0);
let (_, black, _) = crate::xyb::rgb_to_xyb_pixel_f32(&ctx.xyb, 0.0, 0.0, 0.0);
// Padding past the image edge arrives replicated from the edge pixel and
// is classified like it: it is never shown, so where it goes does not
// matter, and a view and a copy of the same stripe agree.
scratch.class.clear();
scratch.class.reserve(n);
let mut saturated = 0;
let (mut any_white, mut any_black, mut any_normal) = (false, false, false);
for row in input.chunks(stride).take(ph) {
scratch.class.extend(row[..pw].iter().map(|&v| {
let class = if v >= white - Y_TOLERANCE {
any_white = true;
WHITE
} else if v <= black + Y_TOLERANCE {
any_black = true;
BLACK
} else {
any_normal = true;
NORMAL
};
saturated += usize::from(class != NORMAL);
class
}));
}
// A footprint at one limit only has nothing to constrain it; one with
// both limits is pinned by the sign pattern and gains the most.
let single_class = !any_normal && (any_white != any_black);
if single_class || (saturated as f32) < MIN_SATURATED_SHARE * n as f32 {
return false;
}
scratch.source.clear();
scratch.source.reserve(n);
for row in input.chunks(stride).take(ph) {
scratch.source.extend_from_slice(&row[..pw]);
}
scratch.target.clone_from(&scratch.source);
scratch.coefficients.resize(n, 0.0);
scratch.levels.resize(n, 0);
scratch.pixels.resize(n, 0.0);
let bias = ctx.dequant_bias(1);
let inv_qac = 1.0 / (scale * quant_ac as f32);
let wc = cx * 8;
for _ in 0..ITERATIONS {
crate::inflated_cost::forward_for(raw_strategy, &scratch.target, &mut scratch.coefficients);
(ctx.quantize_block_ac)(
&scratch.coefficients,
1,
inv_qm,
quant_ac,
scale,
1.0,
distance,
cx,
cy,
&mut scratch.levels,
);
// The cx*cy lowest frequencies ride on the finer DC path and stay exact.
for (y, ((coefficients, levels), weights)) in scratch
.coefficients
.chunks_exact_mut(wc)
.zip(scratch.levels.chunks_exact(wc))
.zip(qm.chunks_exact(wc))
.enumerate()
{
let skip = if y < cy { cx } else { 0 };
for ((coefficient, &level), &weight) in
coefficients.iter_mut().zip(levels).zip(weights).skip(skip)
{
*coefficient = bias.level(level) * weight * inv_qac;
}
}
crate::inflated_cost::reconstruct_error(
ctx.idct,
raw_strategy,
&scratch.coefficients,
&mut scratch.pixels,
);
// Normal pixels retain the source copied above.
for ((target, &recon), &class) in scratch
.target
.iter_mut()
.zip(&scratch.pixels)
.zip(&scratch.class)
{
match class {
WHITE => *target = recon.max(white),
BLACK => *target = recon.min(black),
_ => {}
}
}
}
scratch.shift.clear();
scratch.shift.extend(
scratch
.target
.iter()
.zip(&scratch.source)
.map(|(t, s)| t - s),
);
scratch.shift.iter().any(|&d| d != 0.0)
}
/// `input` (strided) plus the stored shift, contiguous with width `pw`.
pub(crate) fn shifted_input<'a>(
scratch: &'a mut ClipScratch,
input: &[f32],
stride: usize,
pw: usize,
ph: usize,
) -> &'a [f32] {
scratch.shifted.clear();
for y in 0..ph {
let row = &input[y * stride..y * stride + pw];
let shift = &scratch.shift[y * pw..y * pw + pw];
scratch
.shifted
.extend(row.iter().zip(shift).map(|(a, b)| a + b));
}
&scratch.shifted
}
#[cfg(test)]
mod tests {
use super::*;
use crate::Speed;
use crate::dc_group_data::STRATEGY_DCT;
use crate::image::Image3F;
use crate::xyb::{XybMatrix, rgb_to_xyb_pixel_f32};
fn context() -> EncodingContext {
EncodingContext::new(
Speed::Slow,
crate::Arbiter::Ssimulacra2,
XybMatrix::SPEC,
2.0,
1,
)
}
/// 8x8 luma footprint: `left` columns of one gray level, the rest another.
fn image(left_value: f32, right_value: f32, left: usize) -> Image3F {
let mut xyb = Image3F::new(8, 8);
for y in 0..8 {
let row = xyb.plane_row_mut(1, y);
for (x, v) in row.iter_mut().enumerate() {
*v = if x < left { left_value } else { right_value };
}
}
xyb
}
fn run(ctx: &EncodingContext, xyb: &Image3F, scratch: &mut ClipScratch) -> bool {
let matrices = ctx.matrices();
let plane = xyb.plane(1);
continue_saturated(
ctx,
STRATEGY_DCT,
1,
1,
&matrices.inv_matrix(1)[..],
&matrices.matrix(1)[..],
8,
1.0,
2.0,
plane.as_slice(),
plane.xsize(),
scratch,
)
}
#[test]
fn saturated_pixels_only_move_beyond_their_limit() {
let ctx = context();
let (_, white, _) = rgb_to_xyb_pixel_f32(&ctx.xyb, 1.0, 1.0, 1.0);
let (_, black, _) = rgb_to_xyb_pixel_f32(&ctx.xyb, 0.0, 0.0, 0.0);
let xyb = image(black, white, 3);
let mut scratch = ClipScratch::default();
assert!(run(&ctx, &xyb, &mut scratch));
for y in 0..8 {
for x in 0..8 {
let shift = scratch.shift[y * 8 + x];
if x < 3 {
assert!(shift <= 0.0, "black pixel moved up at {x},{y}");
} else {
assert!(shift >= 0.0, "white pixel moved down at {x},{y}");
}
}
}
let shifted = shifted_input(&mut scratch, xyb.plane(1).as_slice(), 8, 8, 8).to_vec();
assert!(
shifted
.iter()
.zip(xyb.plane(1).as_slice())
.any(|(a, b)| a != b)
);
// A pinned column in the middle stays where the source put it.
let mut xyb = image(black, white, 3);
for y in 0..8 {
xyb.plane_row_mut(1, y)[3] = 0.5 * (black + white);
}
let mut scratch = ClipScratch::default();
assert!(run(&ctx, &xyb, &mut scratch));
for y in 0..8 {
assert_eq!(scratch.shift[y * 8 + 3], 0.0);
}
}
#[test]
fn unsaturated_and_single_limit_footprints_are_untouched() {
let ctx = context();
let (_, white, _) = rgb_to_xyb_pixel_f32(&ctx.xyb, 1.0, 1.0, 1.0);
let mut scratch = ClipScratch::default();
assert!(!run(&ctx, &image(0.2, 0.6, 4), &mut scratch));
assert!(!run(&ctx, &image(white, white, 4), &mut scratch));
}
}