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
use core::convert::TryInto;

use crate::common::BytesPerPixel;

/// The byte level filter applied to scanlines to prepare them for compression.
///
/// Compression in general benefits from repetitive data. The filter is a content-aware method of
/// compressing the range of occurring byte values to help the compression algorithm. Note that
/// this does not operate on pixels but on raw bytes of a scanline.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[repr(u8)]
pub enum FilterType {
    NoFilter = 0,
    Sub = 1,
    Up = 2,
    Avg = 3,
    Paeth = 4,
}

impl Default for FilterType {
    fn default() -> Self {
        FilterType::Sub
    }
}

impl FilterType {
    /// u8 -> Self. Temporary solution until Rust provides a canonical one.
    pub fn from_u8(n: u8) -> Option<FilterType> {
        match n {
            0 => Some(FilterType::NoFilter),
            1 => Some(FilterType::Sub),
            2 => Some(FilterType::Up),
            3 => Some(FilterType::Avg),
            4 => Some(FilterType::Paeth),
            _ => None,
        }
    }
}

/// The filtering method for preprocessing scanline data before compression.
///
/// Adaptive filtering performs additional computation in an attempt to maximize
/// the compression of the data. [`NonAdaptive`] filtering is the default.
///
/// [`NonAdaptive`]: enum.AdaptiveFilterType.html#variant.NonAdaptive
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[repr(u8)]
pub enum AdaptiveFilterType {
    Adaptive,
    NonAdaptive,
}

impl Default for AdaptiveFilterType {
    fn default() -> Self {
        AdaptiveFilterType::NonAdaptive
    }
}

fn filter_paeth_decode(a: u8, b: u8, c: u8) -> u8 {
    // Decoding seems to optimize better with this algorithm
    let pa = (i16::from(b) - i16::from(c)).abs();
    let pb = (i16::from(a) - i16::from(c)).abs();
    let pc = ((i16::from(a) - i16::from(c)) + (i16::from(b) - i16::from(c))).abs();

    let mut out = a;
    let mut min = pa;

    if pb < min {
        min = pb;
        out = b;
    }
    if pc < min {
        out = c;
    }

    out
}

fn filter_paeth(a: u8, b: u8, c: u8) -> u8 {
    // This is an optimized version of the paeth filter from the PNG specification, proposed by
    // Luca Versari for [FPNGE](https://www.lucaversari.it/FJXL_and_FPNGE.pdf). It operates
    // entirely on unsigned 8-bit quantities, making it more conducive to vectorization.
    //
    //     p = a + b - c
    //     pa = |p - a| = |a + b - c - a| = |b - c| = max(b, c) - min(b, c)
    //     pb = |p - b| = |a + b - c - b| = |a - c| = max(a, c) - min(a, c)
    //     pc = |p - c| = |a + b - c - c| = |(b - c) + (a - c)| = ...
    //
    // Further optimizing the calculation of `pc` a bit tricker. However, notice that:
    //
    //        a > c && b > c
    //    ==> (a - c) > 0 && (b - c) > 0
    //    ==> pc > (a - c) && pc > (b - c)
    //    ==> pc > |a - c| && pc > |b - c|
    //    ==> pc > pb && pc > pa
    //
    // Meaning that if `c` is smaller than `a` and `b`, the value of `pc` is irrelevant. Similar
    // reasoning applies if `c` is larger than the other two inputs. Assuming that `c >= b` and
    // `c <= b` or vice versa:
    //
    //     pc = ||b - c| - |a - c|| =  |pa - pb| = max(pa, pb) - min(pa, pb)
    //
    let pa = b.max(c) - c.min(b);
    let pb = a.max(c) - c.min(a);
    let pc = if (a < c) == (c < b) {
        pa.max(pb) - pa.min(pb)
    } else {
        255
    };

    if pa <= pb && pa <= pc {
        a
    } else if pb <= pc {
        b
    } else {
        c
    }
}

pub(crate) fn unfilter(
    filter: FilterType,
    tbpp: BytesPerPixel,
    previous: &[u8],
    current: &mut [u8],
) {
    use self::FilterType::*;

    // [2023/01 @okaneco] - Notes on optimizing decoding filters
    //
    // Links:
    // [PR]: https://github.com/image-rs/image-png/pull/382
    // [SWAR]: http://aggregate.org/SWAR/over.html
    // [AVG]: http://aggregate.org/MAGIC/#Average%20of%20Integers
    //
    // #382 heavily refactored and optimized the following filters making the
    // implementation nonobvious. These comments function as a summary of that
    // PR with an explanation of the choices made below.
    //
    // #382 originally started with trying to optimize using a technique called
    // SWAR, SIMD Within a Register. SWAR uses regular integer types like `u32`
    // and `u64` as SIMD registers to perform vertical operations in parallel,
    // usually involving bit-twiddling. This allowed each `BytesPerPixel` (bpp)
    // pixel to be decoded in parallel: 3bpp and 4bpp in a `u32`, 6bpp and 8pp
    // in a `u64`. The `Sub` filter looked like the following code block, `Avg`
    // was similar but used a bitwise average method from [AVG]:
    // ```
    // // See "Unpartitioned Operations With Correction Code" from [SWAR]
    // fn swar_add_u32(x: u32, y: u32) -> u32 {
    //     // 7-bit addition so there's no carry over the most significant bit
    //     let n = (x & 0x7f7f7f7f) + (y & 0x7f7f7f7f); // 0x7F = 0b_0111_1111
    //     // 1-bit parity/XOR addition to fill in the missing MSB
    //     n ^ (x ^ y) & 0x80808080                     // 0x80 = 0b_1000_0000
    // }
    //
    // let mut prev =
    //     u32::from_ne_bytes([current[0], current[1], current[2], current[3]]);
    // for chunk in current[4..].chunks_exact_mut(4) {
    //     let cur = u32::from_ne_bytes([chunk[0], chunk[1], chunk[2], chunk[3]]);
    //     let new_chunk = swar_add_u32(cur, prev);
    //     chunk.copy_from_slice(&new_chunk.to_ne_bytes());
    //     prev = new_chunk;
    // }
    // ```
    // While this provided a measurable increase, @fintelia found that this idea
    // could be taken even further by unrolling the chunks component-wise and
    // avoiding unnecessary byte-shuffling by using byte arrays instead of
    // `u32::from|to_ne_bytes`. The bitwise operations were no longer necessary
    // so they were reverted to their obvious arithmetic equivalent. Lastly,
    // `TryInto` was used instead of `copy_from_slice`. The `Sub` code now
    // looked like this (with asserts to remove `0..bpp` bounds checks):
    // ```
    // assert!(len > 3);
    // let mut prev = [current[0], current[1], current[2], current[3]];
    // for chunk in current[4..].chunks_exact_mut(4) {
    //     let new_chunk = [
    //         chunk[0].wrapping_add(prev[0]),
    //         chunk[1].wrapping_add(prev[1]),
    //         chunk[2].wrapping_add(prev[2]),
    //         chunk[3].wrapping_add(prev[3]),
    //     ];
    //     *TryInto::<&mut [u8; 4]>::try_into(chunk).unwrap() = new_chunk;
    //     prev = new_chunk;
    // }
    // ```
    // The compiler was able to optimize the code to be even faster and this
    // method even sped up Paeth filtering! Assertions were experimentally
    // added within loop bodies which produced better instructions but no
    // difference in speed. Finally, the code was refactored to remove manual
    // slicing and start the previous pixel chunks with arrays of `[0; N]`.
    // ```
    // let mut prev = [0; 4];
    // for chunk in current.chunks_exact_mut(4) {
    //     let new_chunk = [
    //         chunk[0].wrapping_add(prev[0]),
    //         chunk[1].wrapping_add(prev[1]),
    //         chunk[2].wrapping_add(prev[2]),
    //         chunk[3].wrapping_add(prev[3]),
    //     ];
    //     *TryInto::<&mut [u8; 4]>::try_into(chunk).unwrap() = new_chunk;
    //     prev = new_chunk;
    // }
    // ```
    // While we're not manually bit-twiddling anymore, a possible takeaway from
    // this is to "think in SWAR" when dealing with small byte arrays. Unrolling
    // array operations and performing them component-wise may unlock previously
    // unavailable optimizations from the compiler, even when using the
    // `chunks_exact` methods for their potential auto-vectorization benefits.
    match filter {
        NoFilter => {}
        Sub => match tbpp {
            BytesPerPixel::One => {
                current.iter_mut().reduce(|&mut prev, curr| {
                    *curr = curr.wrapping_add(prev);
                    curr
                });
            }
            BytesPerPixel::Two => {
                let mut prev = [0; 2];
                for chunk in current.chunks_exact_mut(2) {
                    let new_chunk = [
                        chunk[0].wrapping_add(prev[0]),
                        chunk[1].wrapping_add(prev[1]),
                    ];
                    *TryInto::<&mut [u8; 2]>::try_into(chunk).unwrap() = new_chunk;
                    prev = new_chunk;
                }
            }
            BytesPerPixel::Three => {
                let mut prev = [0; 3];
                for chunk in current.chunks_exact_mut(3) {
                    let new_chunk = [
                        chunk[0].wrapping_add(prev[0]),
                        chunk[1].wrapping_add(prev[1]),
                        chunk[2].wrapping_add(prev[2]),
                    ];
                    *TryInto::<&mut [u8; 3]>::try_into(chunk).unwrap() = new_chunk;
                    prev = new_chunk;
                }
            }
            BytesPerPixel::Four => {
                let mut prev = [0; 4];
                for chunk in current.chunks_exact_mut(4) {
                    let new_chunk = [
                        chunk[0].wrapping_add(prev[0]),
                        chunk[1].wrapping_add(prev[1]),
                        chunk[2].wrapping_add(prev[2]),
                        chunk[3].wrapping_add(prev[3]),
                    ];
                    *TryInto::<&mut [u8; 4]>::try_into(chunk).unwrap() = new_chunk;
                    prev = new_chunk;
                }
            }
            BytesPerPixel::Six => {
                let mut prev = [0; 6];
                for chunk in current.chunks_exact_mut(6) {
                    let new_chunk = [
                        chunk[0].wrapping_add(prev[0]),
                        chunk[1].wrapping_add(prev[1]),
                        chunk[2].wrapping_add(prev[2]),
                        chunk[3].wrapping_add(prev[3]),
                        chunk[4].wrapping_add(prev[4]),
                        chunk[5].wrapping_add(prev[5]),
                    ];
                    *TryInto::<&mut [u8; 6]>::try_into(chunk).unwrap() = new_chunk;
                    prev = new_chunk;
                }
            }
            BytesPerPixel::Eight => {
                let mut prev = [0; 8];
                for chunk in current.chunks_exact_mut(8) {
                    let new_chunk = [
                        chunk[0].wrapping_add(prev[0]),
                        chunk[1].wrapping_add(prev[1]),
                        chunk[2].wrapping_add(prev[2]),
                        chunk[3].wrapping_add(prev[3]),
                        chunk[4].wrapping_add(prev[4]),
                        chunk[5].wrapping_add(prev[5]),
                        chunk[6].wrapping_add(prev[6]),
                        chunk[7].wrapping_add(prev[7]),
                    ];
                    *TryInto::<&mut [u8; 8]>::try_into(chunk).unwrap() = new_chunk;
                    prev = new_chunk;
                }
            }
        },
        Up => {
            for (curr, &above) in current.iter_mut().zip(previous) {
                *curr = curr.wrapping_add(above);
            }
        }
        Avg => match tbpp {
            BytesPerPixel::One => {
                let mut lprev = [0; 1];
                for (chunk, above) in current.chunks_exact_mut(1).zip(previous.chunks_exact(1)) {
                    let new_chunk =
                        [chunk[0].wrapping_add(((above[0] as u16 + lprev[0] as u16) / 2) as u8)];
                    *TryInto::<&mut [u8; 1]>::try_into(chunk).unwrap() = new_chunk;
                    lprev = new_chunk;
                }
            }
            BytesPerPixel::Two => {
                let mut lprev = [0; 2];
                for (chunk, above) in current.chunks_exact_mut(2).zip(previous.chunks_exact(2)) {
                    let new_chunk = [
                        chunk[0].wrapping_add(((above[0] as u16 + lprev[0] as u16) / 2) as u8),
                        chunk[1].wrapping_add(((above[1] as u16 + lprev[1] as u16) / 2) as u8),
                    ];
                    *TryInto::<&mut [u8; 2]>::try_into(chunk).unwrap() = new_chunk;
                    lprev = new_chunk;
                }
            }
            BytesPerPixel::Three => {
                let mut lprev = [0; 3];
                for (chunk, above) in current.chunks_exact_mut(3).zip(previous.chunks_exact(3)) {
                    let new_chunk = [
                        chunk[0].wrapping_add(((above[0] as u16 + lprev[0] as u16) / 2) as u8),
                        chunk[1].wrapping_add(((above[1] as u16 + lprev[1] as u16) / 2) as u8),
                        chunk[2].wrapping_add(((above[2] as u16 + lprev[2] as u16) / 2) as u8),
                    ];
                    *TryInto::<&mut [u8; 3]>::try_into(chunk).unwrap() = new_chunk;
                    lprev = new_chunk;
                }
            }
            BytesPerPixel::Four => {
                let mut lprev = [0; 4];
                for (chunk, above) in current.chunks_exact_mut(4).zip(previous.chunks_exact(4)) {
                    let new_chunk = [
                        chunk[0].wrapping_add(((above[0] as u16 + lprev[0] as u16) / 2) as u8),
                        chunk[1].wrapping_add(((above[1] as u16 + lprev[1] as u16) / 2) as u8),
                        chunk[2].wrapping_add(((above[2] as u16 + lprev[2] as u16) / 2) as u8),
                        chunk[3].wrapping_add(((above[3] as u16 + lprev[3] as u16) / 2) as u8),
                    ];
                    *TryInto::<&mut [u8; 4]>::try_into(chunk).unwrap() = new_chunk;
                    lprev = new_chunk;
                }
            }
            BytesPerPixel::Six => {
                let mut lprev = [0; 6];
                for (chunk, above) in current.chunks_exact_mut(6).zip(previous.chunks_exact(6)) {
                    let new_chunk = [
                        chunk[0].wrapping_add(((above[0] as u16 + lprev[0] as u16) / 2) as u8),
                        chunk[1].wrapping_add(((above[1] as u16 + lprev[1] as u16) / 2) as u8),
                        chunk[2].wrapping_add(((above[2] as u16 + lprev[2] as u16) / 2) as u8),
                        chunk[3].wrapping_add(((above[3] as u16 + lprev[3] as u16) / 2) as u8),
                        chunk[4].wrapping_add(((above[4] as u16 + lprev[4] as u16) / 2) as u8),
                        chunk[5].wrapping_add(((above[5] as u16 + lprev[5] as u16) / 2) as u8),
                    ];
                    *TryInto::<&mut [u8; 6]>::try_into(chunk).unwrap() = new_chunk;
                    lprev = new_chunk;
                }
            }
            BytesPerPixel::Eight => {
                let mut lprev = [0; 8];
                for (chunk, above) in current.chunks_exact_mut(8).zip(previous.chunks_exact(8)) {
                    let new_chunk = [
                        chunk[0].wrapping_add(((above[0] as u16 + lprev[0] as u16) / 2) as u8),
                        chunk[1].wrapping_add(((above[1] as u16 + lprev[1] as u16) / 2) as u8),
                        chunk[2].wrapping_add(((above[2] as u16 + lprev[2] as u16) / 2) as u8),
                        chunk[3].wrapping_add(((above[3] as u16 + lprev[3] as u16) / 2) as u8),
                        chunk[4].wrapping_add(((above[4] as u16 + lprev[4] as u16) / 2) as u8),
                        chunk[5].wrapping_add(((above[5] as u16 + lprev[5] as u16) / 2) as u8),
                        chunk[6].wrapping_add(((above[6] as u16 + lprev[6] as u16) / 2) as u8),
                        chunk[7].wrapping_add(((above[7] as u16 + lprev[7] as u16) / 2) as u8),
                    ];
                    *TryInto::<&mut [u8; 8]>::try_into(chunk).unwrap() = new_chunk;
                    lprev = new_chunk;
                }
            }
        },
        Paeth => {
            // Paeth filter pixels:
            // C B D
            // A X
            match tbpp {
                BytesPerPixel::One => {
                    let mut a_bpp = [0; 1];
                    let mut c_bpp = [0; 1];
                    for (chunk, b_bpp) in current.chunks_exact_mut(1).zip(previous.chunks_exact(1))
                    {
                        let new_chunk = [chunk[0]
                            .wrapping_add(filter_paeth_decode(a_bpp[0], b_bpp[0], c_bpp[0]))];
                        *TryInto::<&mut [u8; 1]>::try_into(chunk).unwrap() = new_chunk;
                        a_bpp = new_chunk;
                        c_bpp = b_bpp.try_into().unwrap();
                    }
                }
                BytesPerPixel::Two => {
                    let mut a_bpp = [0; 2];
                    let mut c_bpp = [0; 2];
                    for (chunk, b_bpp) in current.chunks_exact_mut(2).zip(previous.chunks_exact(2))
                    {
                        let new_chunk = [
                            chunk[0]
                                .wrapping_add(filter_paeth_decode(a_bpp[0], b_bpp[0], c_bpp[0])),
                            chunk[1]
                                .wrapping_add(filter_paeth_decode(a_bpp[1], b_bpp[1], c_bpp[1])),
                        ];
                        *TryInto::<&mut [u8; 2]>::try_into(chunk).unwrap() = new_chunk;
                        a_bpp = new_chunk;
                        c_bpp = b_bpp.try_into().unwrap();
                    }
                }
                BytesPerPixel::Three => {
                    let mut a_bpp = [0; 3];
                    let mut c_bpp = [0; 3];
                    for (chunk, b_bpp) in current.chunks_exact_mut(3).zip(previous.chunks_exact(3))
                    {
                        let new_chunk = [
                            chunk[0]
                                .wrapping_add(filter_paeth_decode(a_bpp[0], b_bpp[0], c_bpp[0])),
                            chunk[1]
                                .wrapping_add(filter_paeth_decode(a_bpp[1], b_bpp[1], c_bpp[1])),
                            chunk[2]
                                .wrapping_add(filter_paeth_decode(a_bpp[2], b_bpp[2], c_bpp[2])),
                        ];
                        *TryInto::<&mut [u8; 3]>::try_into(chunk).unwrap() = new_chunk;
                        a_bpp = new_chunk;
                        c_bpp = b_bpp.try_into().unwrap();
                    }
                }
                BytesPerPixel::Four => {
                    let mut a_bpp = [0; 4];
                    let mut c_bpp = [0; 4];
                    for (chunk, b_bpp) in current.chunks_exact_mut(4).zip(previous.chunks_exact(4))
                    {
                        let new_chunk = [
                            chunk[0]
                                .wrapping_add(filter_paeth_decode(a_bpp[0], b_bpp[0], c_bpp[0])),
                            chunk[1]
                                .wrapping_add(filter_paeth_decode(a_bpp[1], b_bpp[1], c_bpp[1])),
                            chunk[2]
                                .wrapping_add(filter_paeth_decode(a_bpp[2], b_bpp[2], c_bpp[2])),
                            chunk[3]
                                .wrapping_add(filter_paeth_decode(a_bpp[3], b_bpp[3], c_bpp[3])),
                        ];
                        *TryInto::<&mut [u8; 4]>::try_into(chunk).unwrap() = new_chunk;
                        a_bpp = new_chunk;
                        c_bpp = b_bpp.try_into().unwrap();
                    }
                }
                BytesPerPixel::Six => {
                    let mut a_bpp = [0; 6];
                    let mut c_bpp = [0; 6];
                    for (chunk, b_bpp) in current.chunks_exact_mut(6).zip(previous.chunks_exact(6))
                    {
                        let new_chunk = [
                            chunk[0]
                                .wrapping_add(filter_paeth_decode(a_bpp[0], b_bpp[0], c_bpp[0])),
                            chunk[1]
                                .wrapping_add(filter_paeth_decode(a_bpp[1], b_bpp[1], c_bpp[1])),
                            chunk[2]
                                .wrapping_add(filter_paeth_decode(a_bpp[2], b_bpp[2], c_bpp[2])),
                            chunk[3]
                                .wrapping_add(filter_paeth_decode(a_bpp[3], b_bpp[3], c_bpp[3])),
                            chunk[4]
                                .wrapping_add(filter_paeth_decode(a_bpp[4], b_bpp[4], c_bpp[4])),
                            chunk[5]
                                .wrapping_add(filter_paeth_decode(a_bpp[5], b_bpp[5], c_bpp[5])),
                        ];
                        *TryInto::<&mut [u8; 6]>::try_into(chunk).unwrap() = new_chunk;
                        a_bpp = new_chunk;
                        c_bpp = b_bpp.try_into().unwrap();
                    }
                }
                BytesPerPixel::Eight => {
                    let mut a_bpp = [0; 8];
                    let mut c_bpp = [0; 8];
                    for (chunk, b_bpp) in current.chunks_exact_mut(8).zip(previous.chunks_exact(8))
                    {
                        let new_chunk = [
                            chunk[0]
                                .wrapping_add(filter_paeth_decode(a_bpp[0], b_bpp[0], c_bpp[0])),
                            chunk[1]
                                .wrapping_add(filter_paeth_decode(a_bpp[1], b_bpp[1], c_bpp[1])),
                            chunk[2]
                                .wrapping_add(filter_paeth_decode(a_bpp[2], b_bpp[2], c_bpp[2])),
                            chunk[3]
                                .wrapping_add(filter_paeth_decode(a_bpp[3], b_bpp[3], c_bpp[3])),
                            chunk[4]
                                .wrapping_add(filter_paeth_decode(a_bpp[4], b_bpp[4], c_bpp[4])),
                            chunk[5]
                                .wrapping_add(filter_paeth_decode(a_bpp[5], b_bpp[5], c_bpp[5])),
                            chunk[6]
                                .wrapping_add(filter_paeth_decode(a_bpp[6], b_bpp[6], c_bpp[6])),
                            chunk[7]
                                .wrapping_add(filter_paeth_decode(a_bpp[7], b_bpp[7], c_bpp[7])),
                        ];
                        *TryInto::<&mut [u8; 8]>::try_into(chunk).unwrap() = new_chunk;
                        a_bpp = new_chunk;
                        c_bpp = b_bpp.try_into().unwrap();
                    }
                }
            }
        }
    }
}

fn filter_internal(
    method: FilterType,
    bpp: usize,
    len: usize,
    previous: &[u8],
    current: &[u8],
    output: &mut [u8],
) -> FilterType {
    use self::FilterType::*;

    // This value was chosen experimentally based on what acheived the best performance. The
    // Rust compiler does auto-vectorization, and 32-bytes per loop iteration seems to enable
    // the fastest code when doing so.
    const CHUNK_SIZE: usize = 32;

    match method {
        NoFilter => {
            output.copy_from_slice(current);
            NoFilter
        }
        Sub => {
            let mut out_chunks = output[bpp..].chunks_exact_mut(CHUNK_SIZE);
            let mut cur_chunks = current[bpp..].chunks_exact(CHUNK_SIZE);
            let mut prev_chunks = current[..len - bpp].chunks_exact(CHUNK_SIZE);

            for ((out, cur), prev) in (&mut out_chunks).zip(&mut cur_chunks).zip(&mut prev_chunks) {
                for i in 0..CHUNK_SIZE {
                    out[i] = cur[i].wrapping_sub(prev[i]);
                }
            }

            for ((out, cur), &prev) in out_chunks
                .into_remainder()
                .iter_mut()
                .zip(cur_chunks.remainder())
                .zip(prev_chunks.remainder())
            {
                *out = cur.wrapping_sub(prev);
            }

            output[..bpp].copy_from_slice(&current[..bpp]);
            Sub
        }
        Up => {
            let mut out_chunks = output.chunks_exact_mut(CHUNK_SIZE);
            let mut cur_chunks = current.chunks_exact(CHUNK_SIZE);
            let mut prev_chunks = previous.chunks_exact(CHUNK_SIZE);

            for ((out, cur), prev) in (&mut out_chunks).zip(&mut cur_chunks).zip(&mut prev_chunks) {
                for i in 0..CHUNK_SIZE {
                    out[i] = cur[i].wrapping_sub(prev[i]);
                }
            }

            for ((out, cur), &prev) in out_chunks
                .into_remainder()
                .iter_mut()
                .zip(cur_chunks.remainder())
                .zip(prev_chunks.remainder())
            {
                *out = cur.wrapping_sub(prev);
            }
            Up
        }
        Avg => {
            let mut out_chunks = output[bpp..].chunks_exact_mut(CHUNK_SIZE);
            let mut cur_chunks = current[bpp..].chunks_exact(CHUNK_SIZE);
            let mut cur_minus_bpp_chunks = current[..len - bpp].chunks_exact(CHUNK_SIZE);
            let mut prev_chunks = previous[bpp..].chunks_exact(CHUNK_SIZE);

            for (((out, cur), cur_minus_bpp), prev) in (&mut out_chunks)
                .zip(&mut cur_chunks)
                .zip(&mut cur_minus_bpp_chunks)
                .zip(&mut prev_chunks)
            {
                for i in 0..CHUNK_SIZE {
                    // Bitwise average of two integers without overflow and
                    // without converting to a wider bit-width. See:
                    // http://aggregate.org/MAGIC/#Average%20of%20Integers
                    // If this is unrolled by component, consider reverting to
                    // `((cur_minus_bpp[i] as u16 + prev[i] as u16) / 2) as u8`
                    out[i] = cur[i].wrapping_sub(
                        (cur_minus_bpp[i] & prev[i]) + ((cur_minus_bpp[i] ^ prev[i]) >> 1),
                    );
                }
            }

            for (((out, cur), &cur_minus_bpp), &prev) in out_chunks
                .into_remainder()
                .iter_mut()
                .zip(cur_chunks.remainder())
                .zip(cur_minus_bpp_chunks.remainder())
                .zip(prev_chunks.remainder())
            {
                *out = cur.wrapping_sub((cur_minus_bpp & prev) + ((cur_minus_bpp ^ prev) >> 1));
            }

            for i in 0..bpp {
                output[i] = current[i].wrapping_sub(previous[i] / 2);
            }
            Avg
        }
        Paeth => {
            let mut out_chunks = output[bpp..].chunks_exact_mut(CHUNK_SIZE);
            let mut cur_chunks = current[bpp..].chunks_exact(CHUNK_SIZE);
            let mut a_chunks = current[..len - bpp].chunks_exact(CHUNK_SIZE);
            let mut b_chunks = previous[bpp..].chunks_exact(CHUNK_SIZE);
            let mut c_chunks = previous[..len - bpp].chunks_exact(CHUNK_SIZE);

            for ((((out, cur), a), b), c) in (&mut out_chunks)
                .zip(&mut cur_chunks)
                .zip(&mut a_chunks)
                .zip(&mut b_chunks)
                .zip(&mut c_chunks)
            {
                for i in 0..CHUNK_SIZE {
                    out[i] = cur[i].wrapping_sub(filter_paeth(a[i], b[i], c[i]));
                }
            }

            for ((((out, cur), &a), &b), &c) in out_chunks
                .into_remainder()
                .iter_mut()
                .zip(cur_chunks.remainder())
                .zip(a_chunks.remainder())
                .zip(b_chunks.remainder())
                .zip(c_chunks.remainder())
            {
                *out = cur.wrapping_sub(filter_paeth(a, b, c));
            }

            for i in 0..bpp {
                output[i] = current[i].wrapping_sub(filter_paeth(0, previous[i], 0));
            }
            Paeth
        }
    }
}

pub(crate) fn filter(
    method: FilterType,
    adaptive: AdaptiveFilterType,
    bpp: BytesPerPixel,
    previous: &[u8],
    current: &[u8],
    output: &mut [u8],
) -> FilterType {
    use FilterType::*;
    let bpp = bpp.into_usize();
    let len = current.len();

    match adaptive {
        AdaptiveFilterType::NonAdaptive => {
            filter_internal(method, bpp, len, previous, current, output)
        }
        AdaptiveFilterType::Adaptive => {
            let mut min_sum: u64 = u64::MAX;
            let mut filter_choice = FilterType::NoFilter;
            for &filter in [Sub, Up, Avg, Paeth].iter() {
                filter_internal(filter, bpp, len, previous, current, output);
                let sum = sum_buffer(output);
                if sum <= min_sum {
                    min_sum = sum;
                    filter_choice = filter;
                }
            }

            if filter_choice != Paeth {
                filter_internal(filter_choice, bpp, len, previous, current, output);
            }
            filter_choice
        }
    }
}

// Helper function for Adaptive filter buffer summation
fn sum_buffer(buf: &[u8]) -> u64 {
    const CHUNK_SIZE: usize = 32;

    let mut buf_chunks = buf.chunks_exact(CHUNK_SIZE);
    let mut sum = 0_u64;

    for chunk in &mut buf_chunks {
        // At most, `acc` can be `32 * (i8::MIN as u8) = 32 * 128 = 4096`.
        let mut acc = 0;
        for &b in chunk {
            acc += u64::from((b as i8).unsigned_abs());
        }
        sum = sum.saturating_add(acc);
    }

    let mut acc = 0;
    for &b in buf_chunks.remainder() {
        acc += u64::from((b as i8).unsigned_abs());
    }

    sum.saturating_add(acc)
}

#[cfg(test)]
mod test {
    use super::{filter, unfilter, AdaptiveFilterType, BytesPerPixel, FilterType};
    use core::iter;

    #[test]
    fn roundtrip() {
        // A multiple of 8, 6, 4, 3, 2, 1
        const LEN: u8 = 240;
        let previous: Vec<_> = iter::repeat(1).take(LEN.into()).collect();
        let current: Vec<_> = (0..LEN).collect();
        let expected = current.clone();
        let adaptive = AdaptiveFilterType::NonAdaptive;

        let roundtrip = |kind, bpp: BytesPerPixel| {
            let mut output = vec![0; LEN.into()];
            filter(kind, adaptive, bpp, &previous, &current, &mut output);
            unfilter(kind, bpp, &previous, &mut output);
            assert_eq!(
                output, expected,
                "Filtering {:?} with {:?} does not roundtrip",
                bpp, kind
            );
        };

        let filters = [
            FilterType::NoFilter,
            FilterType::Sub,
            FilterType::Up,
            FilterType::Avg,
            FilterType::Paeth,
        ];

        let bpps = [
            BytesPerPixel::One,
            BytesPerPixel::Two,
            BytesPerPixel::Three,
            BytesPerPixel::Four,
            BytesPerPixel::Six,
            BytesPerPixel::Eight,
        ];

        for &filter in filters.iter() {
            for &bpp in bpps.iter() {
                roundtrip(filter, bpp);
            }
        }
    }

    #[test]
    fn roundtrip_ascending_previous_line() {
        // A multiple of 8, 6, 4, 3, 2, 1
        const LEN: u8 = 240;
        let previous: Vec<_> = (0..LEN).collect();
        let current: Vec<_> = (0..LEN).collect();
        let expected = current.clone();
        let adaptive = AdaptiveFilterType::NonAdaptive;

        let roundtrip = |kind, bpp: BytesPerPixel| {
            let mut output = vec![0; LEN.into()];
            filter(kind, adaptive, bpp, &previous, &current, &mut output);
            unfilter(kind, bpp, &previous, &mut output);
            assert_eq!(
                output, expected,
                "Filtering {:?} with {:?} does not roundtrip",
                bpp, kind
            );
        };

        let filters = [
            FilterType::NoFilter,
            FilterType::Sub,
            FilterType::Up,
            FilterType::Avg,
            FilterType::Paeth,
        ];

        let bpps = [
            BytesPerPixel::One,
            BytesPerPixel::Two,
            BytesPerPixel::Three,
            BytesPerPixel::Four,
            BytesPerPixel::Six,
            BytesPerPixel::Eight,
        ];

        for &filter in filters.iter() {
            for &bpp in bpps.iter() {
                roundtrip(filter, bpp);
            }
        }
    }

    #[test]
    // This tests that converting u8 to i8 doesn't overflow when taking the
    // absolute value for adaptive filtering: -128_i8.abs() will panic in debug
    // or produce garbage in release mode. The sum of 0..=255u8 should equal the
    // sum of the absolute values of -128_i8..=127, or abs(-128..=0) + 1..=127.
    fn sum_buffer_test() {
        let sum = (0..=128).sum::<u64>() + (1..=127).sum::<u64>();
        let buf: Vec<u8> = (0_u8..=255).collect();

        assert_eq!(sum, crate::filter::sum_buffer(&buf));
    }
}