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
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
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
/*
 * // Copyright (c) Radzivon Bartoshyk 7/2026. 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.
 */

use crate::afv;
use crate::quant_weights::DequantMatrices;
use crate::thread_pool::ThreadPool;
use crate::{
    Speed, ac_strategy, adaptive_quant, color_correlation, dct, frame, group, inflated_cost, xyb,
};

#[inline]
fn channel_weights_for_bias(bias: f32, distance: f32) -> [f32; 3] {
    let matrix_mix = ((bias - xyb::B_BIAS) / (0.85 - xyb::B_BIAS)).clamp(0.0, 1.0);
    // The coarse yellow transform protects chroma by itself. Carrying the
    // fine-quality B error boost along with it mostly buys extra coefficients.
    // Mild yellow matrices (bias <= 0.70) retain their existing weights.
    let coarse = crate::yellow_opsin::coarse_weight_strength(distance)
        * ((bias - 0.70) / 0.15).clamp(0.0, 1.0);
    [
        0.30 + (0.10 - 0.30) * matrix_mix,
        1.0,
        0.28 + (0.83 - 0.28) * matrix_mix * (1.0 - coarse),
    ]
}

/// Per-encode dispatch table. The individual modules still own their `OnceLock`
/// selectors, but hot inner loops receive these already-resolved function
/// references instead of touching a static guard for every block / band / token.
/// Place of a VarDCT frame in a layered codestream (see
/// `EncodeConfig::ink_layer`).
#[derive(Clone, Copy, Debug, PartialEq, Eq, Default)]
pub(crate) enum FrameRole {
    /// The only frame.
    #[default]
    Last,
    /// First frame, saved to reference slot 0 after the color transform;
    /// the ink frame is blended over it.
    SavedFirst,
}

pub(crate) struct EncodingContext {
    pub(crate) thread_pool: ThreadPool,
    pub(crate) speed: Speed,
    pub(crate) frame_role: FrameRole,
    /// Lossy arm selection from the public config
    pub(crate) lossy_modular: crate::LossyModular,
    /// Code VarDCT frames' DC as a separate DC frame (see
    /// `EncodeConfig::progressive_dc`).
    pub(crate) progressive_dc: bool,
    /// Squeeze levels of the DC frame (see
    /// `EncodeConfig::progressive_dc_levels`).
    pub(crate) progressive_dc_levels: Option<u32>,
    /// Codestream order of the VarDCT groups (see `EncodeConfig::group_order`).
    pub(crate) group_order: crate::GroupOrder,
    /// Metric the lossy decisions are tuned for (see [`crate::Arbiter`]).
    pub(crate) arbiter: crate::Arbiter,
    /// Decoder-side grain synthesis (see `EncodeConfig::grain`).
    pub(crate) grain: crate::Grain,
    /// Experimental spline coding (see `EncodeConfig::splines`).
    #[cfg(feature = "splines")]
    pub(crate) splines: bool,
    #[cfg(feature = "splines")]
    pub(crate) edge_splines: bool,
    /// Experimental bright/dark dot coding (see `EncodeConfig::dots`).
    #[cfg(feature = "splines")]
    pub(crate) dots: bool,
    #[cfg(feature = "splines")]
    pub(crate) spline_continuous_idct: crate::splines::ContinuousIdctFn,
    #[cfg(feature = "splines")]
    pub(crate) spline_render_row: crate::splines::RenderRowFn,
    #[cfg(feature = "splines")]
    pub(crate) spline_distance: crate::splines::SegmentDistanceFn,
    #[cfg(feature = "splines")]
    pub(crate) spline_ridge_row: crate::splines::RidgeRowFn,
    pub(crate) xyb: xyb::XybMatrix,
    /// Color transform of lossy VarDCT frames.
    pub(crate) coding: crate::coding::CodingTransform,
    /// Per-channel AC dequantization the header signals for `xyb` and `coding`.
    dequant_bias: [crate::group::DequantBias; 3],
    /// Resolved EPF channel scales; `None` uses the spec XYB weights.
    pub(crate) epf_channel_scale: Option<[f32; 3]>,
    /// Tables of the non-XYB frame, resolved from its coded planes; unset
    /// while coding XYB.
    plain_matrices: std::sync::OnceLock<&'static DequantMatrices>,
    /// Distance the non-XYB tables are built for.
    plain_distance: f32,
    /// The frame's chroma-from-luma `base_correlation_x`, `_b` and
    /// `color_factor` (f32 bits): the coding transform's defaults, or a
    /// per-image fit for non-XYB frames.
    cfl_frame: [std::sync::atomic::AtomicU32; 3],
    /// Lattice step of the extra channel (1 = lossless); the plane must be
    /// snapped with `modular::snap_to_lattice` first.
    pub(crate) extra_step: u32,
    /// Squeeze-quantize the extra channel at this luma-quantizer distance
    /// (0 = off; takes precedence over `extra_step`).
    pub(crate) extra_squeeze_distance: f32,
    /// Learned MA trees for the squeezed extra channel.
    pub(crate) extra_learned_trees: bool,
    /// The learned frame-wide tree of a losslessly coded extra channel, at
    /// this budget; None keeps the fixed per-group coding.
    pub(crate) extra_plane_tree: Option<crate::lossless::PlaneTreeBudget>,
    /// Fixed context trees (neighbor-activity splits, Zero prediction of the
    /// residuals) instead of one Gradient leaf per squeezed channel.
    pub(crate) extra_context_trees: bool,
    /// XYB rendition of a non-XYB frame. Adaptive quantization reads its
    /// luma instead of the coded planes.
    pub(crate) aq_source: std::sync::OnceLock<crate::image::Image3F>,
    /// Per-block quant-field multipliers from the frame's alpha (blocks per
    /// row, row-major factors): color under low alpha is coarser.
    pub(crate) alpha_aq: std::sync::OnceLock<(usize, Vec<f32>)>,
    /// Coverage (0..1) of the layer blended over this frame, per pixel: the
    /// final image shows `1 - coverage` of this frame's colors.
    #[cfg(feature = "splines")]
    pub(crate) overlay_coverage: Option<Vec<f32>>,
    /// Cached with the matrix, including a later adaptive yellow selection.
    channel_weights: [f32; 3],
    /// Transform-merge knobs resolved at this encodes distance.
    pub(crate) merge: ac_strategy::MergeTuning,
    /// Price AC-strategy candidates by the frame's own statistics (see
    /// [`ac_strategy::RatePrices`]) instead of the fixed rate model.
    pub(crate) learned_rate: bool,
    base_matrices: &'static DequantMatrices,
    /// Retain the HQ DCT8 table below its ordinary distance gate for sparse
    /// point content, where dropping it reverses chroma quality.
    point_chroma_hq_matrices: Option<&'static DequantMatrices>,
    sat_matrices: &'static DequantMatrices,
    pair_b_matrices: &'static DequantMatrices,
    sat_pair_b_matrices: &'static DequantMatrices,
    x_heavy_matrices: &'static DequantMatrices,
    /// Set once per frame (post-XYB, before any table use) by the
    /// chroma-saturation gate in `frame::encode_frame`.
    chroma_heavy: std::sync::atomic::AtomicBool,
    /// Luma gradient density of the frame (mean |∇Y| per sampled pixel pair,
    /// `frame::chroma`); gates the ExtraSlow selection-scale race.
    luma_texture: std::sync::atomic::AtomicU32,
    /// Selection-only table sets (16/32 family scaled) in force while a
    /// transform-selection race arm sets up its DC groups; `None` otherwise.
    selection_tables: std::sync::RwLock<Option<&'static [&'static DequantMatrices; 4]>>,
    point_chroma: std::sync::atomic::AtomicBool,
    x_heavy: std::sync::atomic::AtomicBool,
    pair_b_fine: std::sync::atomic::AtomicBool,
    /// Blue-axis twin of `x_heavy`: structure lives in B−Y where luma-driven
    /// masking cannot see it (saturated blue detail on low-luma ground).
    b_heavy: std::sync::atomic::AtomicBool,
    /// Screen/graphics content (many exactly constant 8x8 blocks): hard
    /// edges beside flat fills, where the strongest EPF pays on every metric.
    screen_content: std::sync::atomic::AtomicBool,
    b_qm_scale: std::sync::atomic::AtomicU32,
    /// Content floor for the signaled X quant scale (2 = none).
    x_qm_scale_floor: std::sync::atomic::AtomicU32,
    pub(crate) to_xyb_band: xyb::ToXybBandFn,
    pub(crate) round_lattice_pow2: xyb::RoundLatticePow2Fn,
    pub(crate) fill_quant_field: adaptive_quant::FillQuantFieldFn,
    pub(crate) sse_and_rate: inflated_cost::SseAndRateFn,
    pub(crate) sse_and_quantize: inflated_cost::SseAndQuantizeFn,
    pub(crate) recon_dist_and_rate: inflated_cost::ReconDistAndRateFn,
    pub(crate) recon_luma_phase: inflated_cost::ReconLumaPhaseFn,
    pub(crate) recon_chroma_phase: inflated_cost::ReconChromaPhaseFn,
    pub(crate) recon_error_kernels: inflated_cost::ReconErrorKernels,
    pub(crate) rate_log2_lut: &'static inflated_cost::RateLog2Lut,
    pub(crate) quantize_block_ac: group::QuantizeBlockAcFn,
    pub(crate) quantize_dc: group::QuantizeDcFn,
    pub(crate) quantize_dc_cfl: group::QuantizeDcCflFn,
    pub(crate) quantize_dc_i32: group::QuantizeDcI32Fn,
    pub(crate) quantize_dc_cfl_i32: group::QuantizeDcCflI32Fn,
    pub(crate) dc_rows: crate::dc_ops::DcRowKernels,
    pub(crate) apply_cfl: ac_strategy::ApplyCflFn,
    pub(crate) mosaic_seam_stats: crate::mosaic_seam::MosaicSeamStatsFn,
    pub(crate) gradient_region_stats: ac_strategy::GradientRegionStatsFn,
    pub(crate) gradient_region_stats_with_chroma: ac_strategy::GradientRegionStatsFn,
    pub(crate) cfl_regression: color_correlation::CflRegressionFn,
    pub(crate) cfl_rdo_block: color_correlation::CflRdoBlockFn,
    pub(crate) cfl_rdo_refill: color_correlation::CflRdoRefillFn,
    pub(crate) cfl_rdo_stats: color_correlation::CflRdoStatsFn,
    pub(crate) cfl_rd_costs: color_correlation::CflRdCostsFn,
    pub(crate) noise_kernels: crate::noise::NoiseKernels,
    pub(crate) chroma_gradient_sums: frame::ChromaGradientSumsFn,
    pub(crate) fill_ytob_row: color_correlation::FillYtobRowFn,
    pub(crate) accumulate_ytob_weights: color_correlation::AccumulateYtobWeightsFn,
    pub(crate) fill_ytob_residuals: color_correlation::FillYtobResidualsFn,

    pub(crate) idct: &'static dct::IdctMethods,
    pub(crate) identity8x8: &'static dct::DctFn<8, 8, 64>,
    pub(crate) dct2x2_8x8: &'static dct::DctFn<8, 8, 64>,
    pub(crate) dct8x8: &'static dct::DctFn<8, 8, 64>,
    pub(crate) dct8x16: &'static dct::DctFn<16, 8, 128>,
    pub(crate) dct16x8: &'static dct::DctFn<8, 16, 128>,
    pub(crate) dct16x16: &'static dct::DctFn<16, 16, 256>,
    pub(crate) dct4x4: &'static dct::DctFn<8, 8, 64>,
    pub(crate) dct4x8: &'static dct::DctFn<8, 8, 64>,
    pub(crate) dct8x4: &'static dct::DctFn<8, 8, 64>,
    pub(crate) dct32x32: &'static dct::DctFn<32, 32, 1024>,
    pub(crate) dct64x64: &'static dct::DctFn<64, 64, 4096>,
    pub(crate) dct64x32: &'static dct::DctFn<32, 64, 2048>,
    pub(crate) dct32x64: &'static dct::DctFn<64, 32, 2048>,
    #[cfg(feature = "extraslow")]
    pub(crate) dct128x128: &'static dct::DctFn<128, 128, 16384>,
    #[cfg(feature = "extraslow")]
    pub(crate) dct128x64: &'static dct::DctFn<64, 128, 8192>,
    #[cfg(feature = "extraslow")]
    pub(crate) dct64x128: &'static dct::DctFn<128, 64, 8192>,
    pub(crate) dct32x16: &'static dct::DctFn<16, 32, 512>,
    pub(crate) dct16x32: &'static dct::DctFn<32, 16, 512>,
    pub(crate) dct32x8: &'static dct::DctFn<8, 32, 256>,
    pub(crate) dct8x32: &'static dct::DctFn<32, 8, 256>,
    pub(crate) dc_from_dct8x32: dct::DcFromDct8x32Fn,
    pub(crate) dc_from_dct32x32: dct::DcFromDct32x32Fn,
    pub(crate) dc_from_dct32x16: dct::DcFromDct32x16Fn,
    pub(crate) dc_from_dct16x32: dct::DcFromDct16x32Fn,
    pub(crate) dc_from_dct64x64: dct::DcFromDct64x64Fn,
    pub(crate) dc_from_dct64x32: dct::DcFromDct64x32Fn,
    pub(crate) dc_from_dct32x64: dct::DcFromDct32x64Fn,
    #[cfg(feature = "extraslow")]
    pub(crate) dc_from_dct128x128: dct::DcFromDct128x128Fn,
    #[cfg(feature = "extraslow")]
    pub(crate) dc_from_dct128x64: dct::DcFromDct128x64Fn,
    #[cfg(feature = "extraslow")]
    pub(crate) dc_from_dct64x128: dct::DcFromDct64x128Fn,
    pub(crate) afv0: afv::AfvFn,
    pub(crate) afv1: afv::AfvFn,
    pub(crate) afv2: afv::AfvFn,
    pub(crate) afv3: afv::AfvFn,
}

impl EncodingContext {
    /// The dequant tables for the current frame: the distance-tier default,
    /// the saturated-content variant when the chroma gate fired, or the
    /// X-gradient variant (it outranks the chroma gates: both fire on saturated
    /// red content, where flattened B bands are a strict RD loss). Sparse
    /// point content retains the HQ DCT8 table below its normal distance gate.
    #[inline]
    pub(crate) fn matrices(&self) -> &'static DequantMatrices {
        use std::sync::atomic::Ordering::Relaxed;
        let plain = self.plain_matrices.get();
        debug_assert!(
            plain.is_some() || self.coding.is_xyb(),
            "non-XYB tables are resolved per frame"
        );
        if let Some(&plain) = plain {
            plain
        } else if let Some(sets) = *self
            .selection_tables
            .read()
            .unwrap_or_else(|e| e.into_inner())
            && !self.x_heavy.load(Relaxed)
            && self.point_chroma_hq_matrices().is_none()
        {
            sets[usize::from(self.chroma_heavy.load(Relaxed))
                + 2 * usize::from(self.pair_b_fine.load(Relaxed))]
        } else if self.x_heavy.load(Relaxed) {
            self.x_heavy_matrices
        } else if let Some(matrices) = self.point_chroma_hq_matrices() {
            matrices
        } else {
            match (
                self.chroma_heavy.load(Relaxed),
                self.pair_b_fine.load(Relaxed),
            ) {
                (true, true) => self.sat_pair_b_matrices,
                (true, false) => self.sat_matrices,
                (false, true) => self.pair_b_matrices,
                (false, false) => self.base_matrices,
            }
        }
    }

    pub(crate) fn set_point_chroma(&self, point_chroma: bool) {
        self.point_chroma
            .store(point_chroma, std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn point_chroma(&self) -> bool {
        use std::sync::atomic::Ordering::Relaxed;
        self.speed.effort().chroma_policy
            && self.point_chroma.load(Relaxed)
            && !self.x_heavy.load(Relaxed)
    }

    /// The same override is used by the residual encoder and spline RD proxy.
    pub(crate) fn point_chroma_hq_matrices(&self) -> Option<&'static DequantMatrices> {
        if self.point_chroma() {
            self.point_chroma_hq_matrices
        } else {
            None
        }
    }

    pub(crate) fn set_pair_b_fine(&self, fine: bool) {
        self.pair_b_fine
            .store(fine, std::sync::atomic::Ordering::Relaxed);
    }

    /// Scale the 16/32 family tables seen by transform selection by `family`
    /// (`None` restores the coding tables). Coding, header accounting and
    /// signalling always use the coding tables.
    #[cfg(feature = "extraslow")]
    pub(crate) fn set_selection_family(&self, distance: f32, family: Option<f32>) {
        let effort = self.speed.effort();
        let sets = family.filter(|&f| f != 1.0).map(|f| {
            DequantMatrices::selection_sets(
                distance,
                self.arbiter,
                effort.rectangles,
                effort.layout_retile && !effort.layout_loop(),
                f,
            )
        });
        *self
            .selection_tables
            .write()
            .unwrap_or_else(|e| e.into_inner()) = sets;
    }

    pub(crate) fn set_luma_texture(&self, stat: f32) {
        self.luma_texture
            .store(stat.to_bits(), std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn luma_texture(&self) -> f32 {
        f32::from_bits(self.luma_texture.load(std::sync::atomic::Ordering::Relaxed))
    }

    pub(crate) fn chroma_heavy(&self) -> bool {
        self.chroma_heavy.load(std::sync::atomic::Ordering::Relaxed)
    }

    pub(crate) fn set_chroma_heavy(&self, heavy: bool) {
        self.chroma_heavy
            .store(heavy, std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn x_heavy(&self) -> bool {
        self.x_heavy.load(std::sync::atomic::Ordering::Relaxed)
    }

    pub(crate) fn set_x_heavy(&self, heavy: bool) {
        self.x_heavy
            .store(heavy, std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn b_heavy(&self) -> bool {
        self.b_heavy.load(std::sync::atomic::Ordering::Relaxed)
    }

    pub(crate) fn set_b_heavy(&self, heavy: bool) {
        self.b_heavy
            .store(heavy, std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn screen_content(&self) -> bool {
        self.screen_content
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    pub(crate) fn set_screen_content(&self, screen: bool) {
        self.screen_content
            .store(screen, std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn raise_x_qm_scale_floor(&self, scale: u32) {
        self.x_qm_scale_floor
            .fetch_max(scale.clamp(2, 7), std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn x_qm_scale_floor(&self) -> u32 {
        self.x_qm_scale_floor
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    pub(crate) fn raise_b_qm_scale(&self, scale: u32) {
        self.b_qm_scale
            .fetch_max(scale.clamp(2, 7), std::sync::atomic::Ordering::Relaxed);
    }

    #[cfg(feature = "extraslow")]
    /// Frame-level race result: replaces the gate-raised scale.
    pub(crate) fn set_b_qm_scale(&self, scale: u32) {
        self.b_qm_scale
            .store(scale.clamp(2, 7), std::sync::atomic::Ordering::Relaxed);
    }

    pub(crate) fn b_qm_scale(&self) -> u32 {
        self.b_qm_scale.load(std::sync::atomic::Ordering::Relaxed)
    }

    #[inline]
    pub(crate) fn b_qm_mul(&self) -> f32 {
        // The signaled scale is an integer in 2..=7. All powers below are
        // exact binary fractions, so a lookup avoids a libm `powf` in the
        // block-level RDO paths without changing the quantizer value.
        static MULTIPLIERS: [f32; 6] = [1.0, 1.25, 1.5625, 1.953125, 2.441_406_2, 3.051_757_8];
        MULTIPLIERS[(self.b_qm_scale().clamp(2, 7) - 2) as usize]
    }

    /// Code lossy frames in `coding`. Non-XYB planes all span about [0, 1],
    /// so they share one luma-shaped table set and equal distortion weights.
    pub(crate) fn set_coding(&mut self, coding: crate::coding::CodingTransform, distance: f32) {
        self.coding = coding;
        self.refresh_dequant_bias();
        self.set_cfl_frame(crate::color_correlation::CflFrame {
            base_x: coding.base_correlation_x(),
            base_b: coding.base_correlation_b(),
            color_factor: crate::color_correlation::K_COLOR_FACTOR,
        });
        // Reducing only B improves SSIMULACRA2, Butteraugli and CVVDP at Slow.
        // Other efforts retain the transform defaults.
        self.epf_channel_scale = match (coding, self.speed) {
            (crate::coding::CodingTransform::Rgb, Speed::Slow | Speed::ExtraSlow) => {
                Some([5.0, 5.0, 2.5])
            }
            _ => coding.epf_channel_scale(),
        };
        self.plain_matrices = std::sync::OnceLock::new();
        self.plain_distance = distance;
        if !coding.is_xyb() {
            self.channel_weights = coding.channel_weights();
            // Splines and dots render on the XYB planes.
            #[cfg(feature = "splines")]
            {
                self.splines = false;
                self.edge_splines = false;
                self.dots = false;
            }
        }
    }

    /// Resolve the non-XYB frame's tables from its coded planes.
    pub(crate) fn set_plain_matrices(&self, coded: &crate::image::Image3F) {
        self.plain_matrices.get_or_init(|| {
            let effort = self.speed.effort();
            DequantMatrices::new_plain(
                self.plain_distance,
                effort.rectangles,
                self.arbiter,
                self.coding,
                self.coding.plane_scale(coded),
                effort.layout_retile && !effort.layout_loop(),
            )
        });
    }

    /// The frame's chroma-from-luma parameters.
    #[inline]
    pub(crate) fn cfl_frame(&self) -> crate::color_correlation::CflFrame {
        use std::sync::atomic::Ordering::Relaxed;
        let [x, b, f] = self
            .cfl_frame
            .each_ref()
            .map(|v| f32::from_bits(v.load(Relaxed)));
        crate::color_correlation::CflFrame {
            base_x: x,
            base_b: b,
            color_factor: f,
        }
    }

    #[inline]
    pub(crate) fn cfl_base_x(&self) -> f32 {
        self.cfl_frame().base_x
    }

    #[inline]
    pub(crate) fn cfl_base_b(&self) -> f32 {
        self.cfl_frame().base_b
    }

    /// Set the frame's parameters; bases must be F16-exact.
    pub(crate) fn set_cfl_frame(&self, cfl: crate::color_correlation::CflFrame) {
        use std::sync::atomic::Ordering::Relaxed;
        for (slot, v) in self
            .cfl_frame
            .iter()
            .zip([cfl.base_x, cfl.base_b, cfl.color_factor])
        {
            slot.store(v.to_bits(), Relaxed);
        }
    }

    /// Update the adaptive matrix and its cached reconstruction weights together.
    #[inline]
    pub(crate) fn set_xyb_matrix(
        &mut self,
        matrix: xyb::XybMatrix,
        distance: f32,
        default_weights: bool,
    ) {
        self.xyb = matrix;
        self.refresh_dequant_bias();
        let weight_bias = if default_weights {
            xyb::B_BIAS
        } else {
            matrix.fwd[8]
        };
        self.channel_weights = channel_weights_for_bias(weight_bias, distance);
    }

    /// An explicit CustomTransformData bundle is written exactly when an XYB
    /// frame uses a non-default opsin matrix.
    fn refresh_dequant_bias(&mut self) {
        let explicit = self.coding.is_xyb() && !self.xyb.is_decoder_default();
        self.dequant_bias = crate::group::DequantBias::channels(explicit);
    }

    /// The decoder's AC dequantization of `channel`.
    #[inline]
    pub(crate) fn dequant_bias(&self, channel: usize) -> crate::group::DequantBias {
        self.dequant_bias[channel]
    }

    #[inline]
    pub(crate) fn channel_weights(&self) -> [f32; 3] {
        self.channel_weights
    }

    #[inline]
    pub(crate) fn channel_weight(&self, channel: usize) -> f32 {
        self.channel_weights()[channel]
    }

    pub(crate) fn new(
        speed: Speed,
        arbiter: crate::Arbiter,
        xyb: xyb::XybMatrix,
        distance: f32,
        num_threads: usize,
    ) -> Self {
        let quantize_dc = group::selected_quantize_dc_methods();
        let dc_from_dct = dct::selected_dc_from_dct_methods();
        let afv = afv::selected_afv_methods();
        let effort = speed.effort();
        let retile = effort.layout_retile && !effort.layout_loop();
        let base_matrices = if effort.rectangles {
            DequantMatrices::new(distance, arbiter, retile)
        } else {
            DequantMatrices::new_fast(distance, arbiter)
        };
        Self {
            thread_pool: ThreadPool::new(num_threads),
            speed,
            frame_role: FrameRole::Last,
            lossy_modular: crate::LossyModular::Off,
            progressive_dc: false,
            progressive_dc_levels: None,
            group_order: crate::GroupOrder::Scanline,
            arbiter,
            grain: crate::Grain::Off,
            #[cfg(feature = "splines")]
            splines: false,
            #[cfg(feature = "splines")]
            edge_splines: false,
            #[cfg(feature = "splines")]
            dots: false,
            xyb,
            coding: crate::coding::CodingTransform::Xyb,
            dequant_bias: crate::group::DequantBias::channels(!xyb.is_decoder_default()),
            epf_channel_scale: None,
            plain_matrices: std::sync::OnceLock::new(),
            plain_distance: 0.0,
            cfl_frame: [
                std::sync::atomic::AtomicU32::new(0.0f32.to_bits()),
                std::sync::atomic::AtomicU32::new(1.0f32.to_bits()),
                std::sync::atomic::AtomicU32::new(84.0f32.to_bits()),
            ],
            aq_source: std::sync::OnceLock::new(),
            alpha_aq: std::sync::OnceLock::new(),
            #[cfg(feature = "splines")]
            overlay_coverage: None,
            extra_step: 1,
            extra_squeeze_distance: 0.0,
            extra_learned_trees: false,
            extra_plane_tree: None,
            extra_context_trees: false,
            channel_weights: channel_weights_for_bias(xyb.fwd[8], distance),
            merge: ac_strategy::MergeTuning::new(distance, arbiter, speed),
            learned_rate: false,
            base_matrices,
            point_chroma_hq_matrices: (effort.chroma_policy
                && distance < crate::quant_weights::QM_FLAT_B8_MIN_DISTANCE)
                .then(|| {
                    DequantMatrices::new(
                        crate::quant_weights::QM_FLAT_B8_MIN_DISTANCE,
                        arbiter,
                        retile,
                    )
                }),
            sat_matrices: DequantMatrices::new_saturated(distance, arbiter, retile),
            pair_b_matrices: if effort.chroma_policy {
                DequantMatrices::new_pair_b(distance, arbiter, retile)
            } else {
                base_matrices
            },
            sat_pair_b_matrices: DequantMatrices::new_saturated_pair_b(distance, arbiter, retile),
            x_heavy_matrices: if effort.chroma_policy {
                DequantMatrices::new_x_heavy(distance, arbiter)
            } else {
                DequantMatrices::new(0.0, arbiter, retile)
            },
            chroma_heavy: std::sync::atomic::AtomicBool::new(false),
            selection_tables: std::sync::RwLock::new(None),
            luma_texture: std::sync::atomic::AtomicU32::new(0),
            point_chroma: std::sync::atomic::AtomicBool::new(false),
            x_heavy: std::sync::atomic::AtomicBool::new(false),
            pair_b_fine: std::sync::atomic::AtomicBool::new(false),
            b_heavy: std::sync::atomic::AtomicBool::new(false),
            screen_content: std::sync::atomic::AtomicBool::new(false),
            b_qm_scale: std::sync::atomic::AtomicU32::new(2),
            x_qm_scale_floor: std::sync::atomic::AtomicU32::new(2),
            #[cfg(feature = "splines")]
            spline_continuous_idct: crate::splines::select_continuous_idct_fn(),
            #[cfg(feature = "splines")]
            spline_render_row: crate::splines::select_render_row_fn(),
            #[cfg(feature = "splines")]
            spline_distance: crate::splines::select_segment_distance_fn(),
            #[cfg(feature = "splines")]
            spline_ridge_row: crate::splines::select_ridge_row_fn(),
            to_xyb_band: xyb::selected_to_xyb_band_fn(),
            round_lattice_pow2: xyb::selected_round_lattice_pow2_fn(),
            fill_quant_field: adaptive_quant::selected_fill_quant_field_fn(),
            sse_and_rate: inflated_cost::selected_sse_and_rate_fn(),
            sse_and_quantize: inflated_cost::selected_sse_and_quantize_fn(),
            recon_dist_and_rate: inflated_cost::select_recon_dist_and_rate_fn(),
            recon_luma_phase: inflated_cost::select_recon_luma_phase_fn(),
            recon_chroma_phase: inflated_cost::select_recon_chroma_phase_fn(),
            recon_error_kernels: inflated_cost::ReconErrorKernels {
                gradient_energy: inflated_cost::select_error_gradient_energy_fn(),
                gradient_peak_energy: inflated_cost::select_error_gradient_peak_energy_fn(),
                combine: inflated_cost::select_combine_error_fn(),
                rgb_hue_chroma_edge_loss: inflated_cost::select_rgb_hue_chroma_edge_loss_fn(),
            },
            rate_log2_lut: inflated_cost::rate_log2_lut(),
            quantize_block_ac: group::selected_quantize_block_ac_fn(),
            quantize_dc: quantize_dc.quantize,
            quantize_dc_cfl: quantize_dc.quantize_cfl,
            quantize_dc_i32: quantize_dc.quantize_i32,
            quantize_dc_cfl_i32: quantize_dc.quantize_cfl_i32,
            dc_rows: crate::dc_ops::selected_dc_row_kernels(),
            apply_cfl: ac_strategy::selected_apply_cfl_fn(),
            mosaic_seam_stats: crate::mosaic_seam::select_mosaic_seam_stats_fn(),
            gradient_region_stats: ac_strategy::select_gradient_region_stats_fn(),
            gradient_region_stats_with_chroma:
                ac_strategy::select_gradient_region_stats_with_chroma_fn(),
            cfl_regression: color_correlation::selected_cfl_regression_fn(),
            cfl_rdo_block: color_correlation::selected_cfl_rdo_block_fn(),
            cfl_rdo_refill: color_correlation::selected_cfl_rdo_refill_fn(),
            cfl_rdo_stats: color_correlation::selected_cfl_rdo_stats_fn(),
            cfl_rd_costs: color_correlation::selected_cfl_rd_costs_fn(),
            noise_kernels: crate::noise::selected_noise_kernels(),
            chroma_gradient_sums: frame::selected_chroma_gradient_sums_fn(),
            fill_ytob_row: color_correlation::selected_fill_ytob_row_fn(),
            accumulate_ytob_weights: color_correlation::selected_accumulate_ytob_weights_fn(),
            fill_ytob_residuals: color_correlation::selected_fill_ytob_residuals_fn(),

            idct: dct::selected_idct_methods(),
            identity8x8: dct::selected_identity8x8(),
            dct2x2_8x8: dct::selected_dct2x2_8x8(),
            dct8x8: dct::selected_dct8x8(),
            dct8x16: dct::selected_dct8x16(),
            dct16x8: dct::selected_dct16x8(),
            dct16x16: dct::selected_dct16x16(),
            dct4x4: dct::selected_dct4x4(),
            dct4x8: dct::selected_dct4x8(),
            dct8x4: dct::selected_dct8x4(),
            dct32x32: dct::selected_dct32x32(),
            dct64x64: dct::selected_dct64x64(),
            dct64x32: dct::selected_dct64x32(),
            dct32x64: dct::selected_dct32x64(),
            #[cfg(feature = "extraslow")]
            dct128x128: crate::dct128::selected_dct128x128(),
            #[cfg(feature = "extraslow")]
            dct128x64: crate::dct128::selected_dct128x64(),
            #[cfg(feature = "extraslow")]
            dct64x128: crate::dct128::selected_dct64x128(),
            dct32x16: dct::selected_dct32x16(),
            dct16x32: dct::selected_dct16x32(),
            dct32x8: dct::selected_dct32x8(),
            dct8x32: dct::selected_dct8x32(),
            dc_from_dct8x32: dc_from_dct.dct8x32,
            dc_from_dct32x32: dc_from_dct.dct32x32,
            dc_from_dct32x16: dc_from_dct.dct32x16,
            dc_from_dct16x32: dc_from_dct.dct16x32,
            dc_from_dct64x64: dc_from_dct.dct64x64,
            dc_from_dct64x32: dc_from_dct.dct64x32,
            dc_from_dct32x64: dc_from_dct.dct32x64,
            #[cfg(feature = "extraslow")]
            dc_from_dct128x128: dc_from_dct.dct128x128,
            #[cfg(feature = "extraslow")]
            dc_from_dct128x64: dc_from_dct.dct128x64,
            #[cfg(feature = "extraslow")]
            dc_from_dct64x128: dc_from_dct.dct64x128,
            afv0: afv.afv0,
            afv1: afv.afv1,
            afv2: afv.afv2,
            afv3: afv.afv3,
        }
    }
}

impl Default for EncodingContext {
    #[inline]
    fn default() -> Self {
        Self::new(
            Speed::Fast,
            crate::Arbiter::Ssimulacra2,
            xyb::XybMatrix::SPEC,
            1.0,
            1,
        )
    }
}

#[cfg(test)]
mod tests {
    use super::channel_weights_for_bias;

    #[test]
    fn point_chroma_tables_preserve_existing_class_precedence() {
        use super::{DequantMatrices, EncodingContext, Speed, xyb};
        for distance in [0.295, 0.299, 0.3, 1.25, 2.25] {
            let ctx = EncodingContext::new(
                Speed::Slow,
                crate::Arbiter::Ssimulacra2,
                xyb::XybMatrix::SPEC,
                distance,
                1,
            );
            ctx.set_chroma_heavy(true);
            ctx.set_pair_b_fine(true);
            let previous = ctx.matrices();
            ctx.set_point_chroma(true);
            let expected = if distance < 0.3 {
                DequantMatrices::new(0.3, crate::Arbiter::Ssimulacra2, false)
            } else {
                previous
            };
            assert!(std::ptr::eq(ctx.matrices(), expected));
            ctx.set_x_heavy(true);
            assert!(std::ptr::eq(
                ctx.matrices(),
                DequantMatrices::new_x_heavy(distance, crate::Arbiter::Ssimulacra2)
            ));
            assert!(ctx.point_chroma_hq_matrices().is_none());
            ctx.set_x_heavy(false);
            ctx.set_point_chroma(false);
            assert!(std::ptr::eq(ctx.matrices(), previous));
        }
        let ctx = EncodingContext::new(
            Speed::Fast,
            crate::Arbiter::Ssimulacra2,
            xyb::XybMatrix::SPEC,
            0.295,
            1,
        );
        let previous = ctx.matrices();
        ctx.set_point_chroma(true);
        assert!(std::ptr::eq(ctx.matrices(), previous));
        assert!(ctx.point_chroma_hq_matrices().is_none());
    }

    #[test]
    fn hue_tables_require_the_frame_content_gate() {
        use super::{DequantMatrices, EncodingContext, Speed, xyb};
        let ctx = EncodingContext::new(
            Speed::Slow,
            crate::Arbiter::Ssimulacra2,
            xyb::XybMatrix::SPEC,
            1.0,
            1,
        );
        assert!(std::ptr::eq(
            ctx.matrices(),
            DequantMatrices::new(1.0, crate::Arbiter::Ssimulacra2, false)
        ));
        ctx.set_chroma_heavy(true);
        assert!(std::ptr::eq(
            ctx.matrices(),
            DequantMatrices::new_saturated(1.0, crate::Arbiter::Ssimulacra2, false)
        ));
        ctx.set_x_heavy(true);
        assert!(std::ptr::eq(
            ctx.matrices(),
            DequantMatrices::new_x_heavy(1.0, crate::Arbiter::Ssimulacra2)
        ));
        ctx.set_x_heavy(false);
        assert!(std::ptr::eq(
            ctx.matrices(),
            DequantMatrices::new_saturated(1.0, crate::Arbiter::Ssimulacra2, false)
        ));
    }

    #[test]
    fn channel_weights_follow_the_opsin_b_row() {
        assert_eq!(
            channel_weights_for_bias(crate::xyb::B_BIAS, 1.0),
            [0.30, 1.0, 0.28]
        );
        let strong = channel_weights_for_bias(0.85, 1.0);
        assert!((strong[0] - 0.10).abs() < 1e-6);
        assert_eq!(strong[1], 1.0);
        assert!((strong[2] - 0.83).abs() < 1e-6);

        let mid = channel_weights_for_bias((crate::xyb::B_BIAS + 0.85) * 0.5, 1.0);
        assert!((mid[0] - 0.20).abs() < 1e-6);
        assert!((mid[2] - 0.555).abs() < 1e-6);
    }

    #[test]
    fn adaptive_matrix_updates_cached_coarse_weights() {
        use super::{EncodingContext, Speed, xyb};
        let mut ctx = EncodingContext::new(
            Speed::Slow,
            crate::Arbiter::Ssimulacra2,
            xyb::XybMatrix::SPEC,
            2.0,
            1,
        );
        assert_eq!(ctx.channel_weights(), [0.30, 1.0, 0.28]);
        let matrix = crate::yellow_opsin::matrix_for_bias(0.90);
        ctx.set_xyb_matrix(matrix, 2.0, false);
        assert_eq!(ctx.xyb, matrix);
        let coarse = ctx.channel_weights();
        assert!((coarse[0] - 0.10).abs() < 1e-6);
        assert_eq!(coarse[1], 1.0);
        assert_eq!(coarse[2], 0.28);
        // The smooth-yellow tier and the successful d1 path keep their boost.
        assert_eq!(
            channel_weights_for_bias(0.70, 2.0),
            channel_weights_for_bias(0.70, 1.0)
        );
        ctx.set_xyb_matrix(crate::yellow_opsin::matrix_for_bias(0.85), 1.0, false);
        assert_eq!(ctx.channel_weights(), channel_weights_for_bias(0.85, 1.0));
        // The sparse-yellow fallback signals its matrix but uses the default
        // error budget, even after a previous matrix had stronger weights.
        let mild = crate::yellow_opsin::matrix_for_bias(0.65);
        ctx.set_xyb_matrix(mild, 2.0, true);
        assert_eq!(ctx.xyb, mild);
        assert_eq!(ctx.channel_weights(), [0.30, 1.0, 0.28]);
    }

    #[test]
    fn dequant_bias_follows_the_signaled_bundle() {
        use super::{EncodingContext, Speed, xyb};
        use crate::group::{DequantBias, EXPLICIT_QUANT_BIASES};
        let mut ctx = EncodingContext::new(
            Speed::Slow,
            crate::Arbiter::Ssimulacra2,
            xyb::XybMatrix::SPEC,
            1.0,
            1,
        );
        let biases = |ctx: &EncodingContext| std::array::from_fn(|c| ctx.dequant_bias(c));
        assert_eq!(biases(&ctx), DequantBias::DEFAULT);
        ctx.set_xyb_matrix(crate::yellow_opsin::matrix_for_bias(0.85), 1.0, false);
        let explicit: [DequantBias; 3] = biases(&ctx);
        assert_eq!(explicit[2].one, 1.0);
        for (c, bias) in explicit.iter().enumerate() {
            let f16 = crate::util::f32_to_f16_bits(EXPLICIT_QUANT_BIASES[c]);
            assert_eq!(bias.one, crate::util::f16_bits_to_f32(f16));
            assert!((bias.numerator - 0.145).abs() < 1e-4);
        }
        // Non-XYB frames never write the explicit bundle.
        ctx.set_coding(crate::coding::CodingTransform::YCbCr, 1.0);
        assert_eq!(biases(&ctx), DequantBias::DEFAULT);
    }
}