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
//!  Decoding of DXT (S3TC) compression
//!
//!  DXT is an image format that supports lossy compression
//!
//!  # Related Links
//!  * <https://www.khronos.org/registry/OpenGL/extensions/EXT/EXT_texture_compression_s3tc.txt> - Description of the DXT compression OpenGL extensions.
//!
//!  Note: this module only implements bare DXT encoding/decoding, it does not parse formats that can contain DXT files like .dds

use std::io::{self, Read, Seek, SeekFrom, Write};

use color::ColorType;
use image::{self, ImageDecoder, ImageDecoderExt, ImageError, ImageReadBuffer, ImageResult, Progress};

/// What version of DXT compression are we using?
/// Note that DXT2 and DXT4 are left away as they're
/// just DXT3 and DXT5 with premultiplied alpha
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum DXTVariant {
    /// The DXT1 format. 48 bytes of RGB data in a 4x4 pixel square is
    /// compressed into an 8 byte block of DXT1 data
    DXT1,
    /// The DXT3 format. 64 bytes of RGBA data in a 4x4 pixel square is
    /// compressed into a 16 byte block of DXT3 data
    DXT3,
    /// The DXT5 format. 64 bytes of RGBA data in a 4x4 pixel square is
    /// compressed into a 16 byte block of DXT5 data
    DXT5,
}

impl DXTVariant {
    /// Returns the amount of bytes of raw image data
    /// that is encoded in a single DXTn block
    fn decoded_bytes_per_block(self) -> usize {
        match self {
            DXTVariant::DXT1 => 48,
            DXTVariant::DXT3 | DXTVariant::DXT5 => 64,
        }
    }

    /// Returns the amount of bytes per block of encoded DXTn data
    fn encoded_bytes_per_block(self) -> usize {
        match self {
            DXTVariant::DXT1 => 8,
            DXTVariant::DXT3 | DXTVariant::DXT5 => 16,
        }
    }

    /// Returns the colortype that is stored in this DXT variant
    pub fn colortype(self) -> ColorType {
        match self {
            DXTVariant::DXT1 => ColorType::RGB(8),
            DXTVariant::DXT3 | DXTVariant::DXT5 => ColorType::RGBA(8),
        }
    }
}

/// DXT decoder
pub struct DXTDecoder<R: Read> {
    inner: R,
    width_blocks: u32,
    height_blocks: u32,
    variant: DXTVariant,
    row: u32,
}

impl<R: Read> DXTDecoder<R> {
    /// Create a new DXT decoder that decodes from the stream ```r```.
    /// As DXT is often stored as raw buffers with the width/height
    /// somewhere else the width and height of the image need
    /// to be passed in ```width``` and ```height```, as well as the
    /// DXT variant in ```variant```.
    /// width and height are required to be powers of 2 and at least 4.
    /// otherwise an error will be returned
    pub fn new(
        r: R,
        width: u32,
        height: u32,
        variant: DXTVariant,
    ) -> Result<DXTDecoder<R>, ImageError> {
        if width % 4 != 0 || height % 4 != 0 {
            return Err(ImageError::DimensionError);
        }
        let width_blocks = width / 4;
        let height_blocks = height / 4;
        Ok(DXTDecoder {
            inner: r,
            width_blocks,
            height_blocks,
            variant,
            row: 0,
        })
    }

    fn read_scanline(&mut self, buf: &mut [u8]) -> io::Result<usize> {
        assert_eq!(buf.len() as u64, self.scanline_bytes());

        let mut src =
            vec![0u8; self.variant.encoded_bytes_per_block() * self.width_blocks as usize];
        self.inner.read_exact(&mut src)?;
        match self.variant {
            DXTVariant::DXT1 => decode_dxt1_row(&src, buf),
            DXTVariant::DXT3 => decode_dxt3_row(&src, buf),
            DXTVariant::DXT5 => decode_dxt5_row(&src, buf),
        }
        self.row += 1;
        Ok(buf.len())
    }
}

// Note that, due to the way that DXT compression works, a scanline is considered to consist out of
// 4 lines of pixels.
impl<R: Read> ImageDecoder for DXTDecoder<R> {
    type Reader = DXTReader<R>;

    fn dimensions(&self) -> (u64, u64) {
        (self.width_blocks as u64 * 4, self.height_blocks as u64 * 4)
    }

    fn colortype(&self) -> ColorType {
        self.variant.colortype()
    }

    fn scanline_bytes(&self) -> u64 {
        self.variant.decoded_bytes_per_block() as u64 * self.width_blocks as u64
    }

    fn into_reader(self) -> ImageResult<Self::Reader> {
        if self.total_bytes() > usize::max_value() as u64 {
            return Err(ImageError::InsufficientMemory);
        }

        Ok(DXTReader {
            buffer: ImageReadBuffer::new(self.scanline_bytes() as usize, self.total_bytes() as usize),
            decoder: self,
        })
    }

    fn read_image(mut self) -> ImageResult<Vec<u8>> {
        if self.total_bytes() > usize::max_value() as u64 {
            return Err(ImageError::InsufficientMemory);
        }

        let mut dest = vec![0u8; self.total_bytes() as usize];
        for chunk in dest.chunks_mut(self.scanline_bytes() as usize) {
            self.read_scanline(chunk)?;
        }
        Ok(dest)
    }
}

impl<R: Read + Seek> ImageDecoderExt for DXTDecoder<R> {
    fn read_rect_with_progress<F: Fn(Progress)>(
        &mut self,
        x: u64,
        y: u64,
        width: u64,
        height: u64,
        buf: &mut [u8],
        progress_callback: F,
    ) -> ImageResult<()> {
        let encoded_scanline_bytes = self.variant.encoded_bytes_per_block() as u64
            * self.width_blocks as u64;

        let start = self.inner.seek(SeekFrom::Current(0))?;
        image::load_rect(x, y, width, height, buf, progress_callback, self,
                         |s, scanline| {
                             s.inner.seek(SeekFrom::Start(start + scanline * encoded_scanline_bytes))?;
                             Ok(())
                         },
                         |s, buf| s.read_scanline(buf))?;
        self.inner.seek(SeekFrom::Start(start))?;
        Ok(())
    }
}

/// DXT reader
pub struct DXTReader<R: Read> {
    buffer: ImageReadBuffer,
    decoder: DXTDecoder<R>,
}
impl<R: Read> Read for DXTReader<R> {
    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
        let ref mut decoder = &mut self.decoder;
        self.buffer.read(buf, |buf| decoder.read_scanline(buf))
    }
}

/// DXT encoder
pub struct DXTEncoder<W: Write> {
    w: W,
}

impl<W: Write> DXTEncoder<W> {
    /// Create a new encoder that writes its output to ```w```
    pub fn new(w: W) -> DXTEncoder<W> {
        DXTEncoder { w }
    }

    /// Encodes the image data ```data```
    /// that has dimensions ```width``` and ```height```
    /// in ```DXTVariant``` ```variant```
    /// data is assumed to be in variant.colortype()
    pub fn encode(
        mut self,
        data: &[u8],
        width: u32,
        height: u32,
        variant: DXTVariant,
    ) -> ImageResult<()> {
        if width % 4 != 0 || height % 4 != 0 {
            return Err(ImageError::DimensionError);
        }
        let width_blocks = width / 4;
        let height_blocks = height / 4;

        let stride = variant.decoded_bytes_per_block();

        assert!(data.len() >= width_blocks as usize * height_blocks as usize * stride);

        for chunk in data.chunks(width_blocks as usize * stride) {
            let data = match variant {
                DXTVariant::DXT1 => encode_dxt1_row(chunk),
                DXTVariant::DXT3 => encode_dxt3_row(chunk),
                DXTVariant::DXT5 => encode_dxt5_row(chunk),
            };
            self.w.write_all(&data)?;
        }
        Ok(())
    }
}

/**
 * Actual encoding/decoding logic below.
 */
use std::mem::swap;

type Rgb = [u8; 3];

/// decodes a 5-bit R, 6-bit G, 5-bit B 16-bit packed color value into 8-bit RGB
/// mapping is done so min/max range values are preserved. So for 5-bit
/// values 0x00 -> 0x00 and 0x1F -> 0xFF
fn enc565_decode(value: u16) -> Rgb {
    let red = (value >> 11) & 0x1F;
    let green = (value >> 5) & 0x3F;
    let blue = (value) & 0x1F;
    [
        (red * 0xFF / 0x1F) as u8,
        (green * 0xFF / 0x3F) as u8,
        (blue * 0xFF / 0x1F) as u8,
    ]
}

/// encodes an 8-bit RGB value into a 5-bit R, 6-bit G, 5-bit B 16-bit packed color value
/// mapping preserves min/max values. It is guaranteed that i == encode(decode(i)) for all i
fn enc565_encode(rgb: Rgb) -> u16 {
    let red = (u16::from(rgb[0]) * 0x1F + 0x7E) / 0xFF;
    let green = (u16::from(rgb[1]) * 0x3F + 0x7E) / 0xFF;
    let blue = (u16::from(rgb[2]) * 0x1F + 0x7E) / 0xFF;
    (red << 11) | (green << 5) | blue
}

/// utility function: squares a value
fn square(a: i32) -> i32 {
    a * a
}

/// returns the squared error between two RGB values
fn diff(a: Rgb, b: Rgb) -> i32 {
    square(i32::from(a[0]) - i32::from(b[0])) + square(i32::from(a[1]) - i32::from(b[1]))
        + square(i32::from(a[2]) - i32::from(b[2]))
}

/*
 * Functions for decoding DXT compression
 */

/// Constructs the DXT5 alpha lookup table from the two alpha entries
/// if alpha0 > alpha1, constructs a table of [a0, a1, 6 linearly interpolated values from a0 to a1]
/// if alpha0 <= alpha1, constructs a table of [a0, a1, 4 linearly interpolated values from a0 to a1, 0, 0xFF]
fn alpha_table_dxt5(alpha0: u8, alpha1: u8) -> [u8; 8] {
    let mut table = [alpha0, alpha1, 0, 0, 0, 0, 0, 0xFF];
    if alpha0 > alpha1 {
        for i in 2..8u16 {
            table[i as usize] =
                (((8 - i) * u16::from(alpha0) + (i - 1) * u16::from(alpha1)) / 7) as u8;
        }
    } else {
        for i in 2..6u16 {
            table[i as usize] =
                (((6 - i) * u16::from(alpha0) + (i - 1) * u16::from(alpha1)) / 5) as u8;
        }
    }
    table
}

/// decodes an 8-byte dxt color block into the RGB channels of a 16xRGB or 16xRGBA block.
/// source should have a length of 8, dest a length of 48 (RGB) or 64 (RGBA)
fn decode_dxt_colors(source: &[u8], dest: &mut [u8]) {
    // sanity checks, also enable the compiler to elide all following bound checks
    assert!(source.len() == 8 && (dest.len() == 48 || dest.len() == 64));
    // calculate pitch to store RGB values in dest (3 for RGB, 4 for RGBA)
    let pitch = dest.len() / 16;

    // extract color data
    let color0 = u16::from(source[0]) | (u16::from(source[1]) << 8);
    let color1 = u16::from(source[2]) | (u16::from(source[3]) << 8);
    let color_table = u32::from(source[4]) | (u32::from(source[5]) << 8)
        | (u32::from(source[6]) << 16) | (u32::from(source[7]) << 24);
    // let color_table = source[4..8].iter().rev().fold(0, |t, &b| (t << 8) | b as u32);

    // decode the colors to rgb format
    let mut colors = [[0; 3]; 4];
    colors[0] = enc565_decode(color0);
    colors[1] = enc565_decode(color1);

    // determine color interpolation method
    if color0 > color1 {
        // linearly interpolate the other two color table entries
        for i in 0..3 {
            colors[2][i] = ((u16::from(colors[0][i]) * 2 + u16::from(colors[1][i]) + 1) / 3) as u8;
            colors[3][i] = ((u16::from(colors[0][i]) + u16::from(colors[1][i]) * 2 + 1) / 3) as u8;
        }
    } else {
        // linearly interpolate one other entry, keep the other at 0
        for i in 0..3 {
            colors[2][i] = ((u16::from(colors[0][i]) + u16::from(colors[1][i]) + 1) / 2) as u8;
        }
    }

    // serialize the result. Every color is determined by looking up
    // two bits in color_table which identify which color to actually pick from the 4 possible colors
    for i in 0..16 {
        dest[i * pitch..i * pitch + 3]
            .copy_from_slice(&colors[(color_table >> (i * 2)) as usize & 3]);
    }
}

/// Decodes a 16-byte bock of dxt5 data to a 16xRGBA block
fn decode_dxt5_block(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() == 16 && dest.len() == 64);

    // extract alpha index table (stored as little endian 64-bit value)
    let alpha_table = source[2..8]
        .iter()
        .rev()
        .fold(0, |t, &b| (t << 8) | u64::from(b));

    // alhpa level decode
    let alphas = alpha_table_dxt5(source[0], source[1]);

    // serialize alpha
    for i in 0..16 {
        dest[i * 4 + 3] = alphas[(alpha_table >> (i * 3)) as usize & 7];
    }

    // handle colors
    decode_dxt_colors(&source[8..16], dest);
}

/// Decodes a 16-byte bock of dxt3 data to a 16xRGBA block
fn decode_dxt3_block(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() == 16 && dest.len() == 64);

    // extract alpha index table (stored as little endian 64-bit value)
    let alpha_table = source[0..8]
        .iter()
        .rev()
        .fold(0, |t, &b| (t << 8) | u64::from(b));

    // serialize alpha (stored as 4-bit values)
    for i in 0..16 {
        dest[i * 4 + 3] = ((alpha_table >> (i * 4)) as u8 & 0xF) * 0x11;
    }

    // handle colors
    decode_dxt_colors(&source[8..16], dest);
}

/// Decodes a 8-byte bock of dxt5 data to a 16xRGB block
fn decode_dxt1_block(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() == 8 && dest.len() == 48);
    decode_dxt_colors(&source, dest);
}

/// Decode a row of DXT1 data to four rows of RGBA data.
/// source.len() should be a multiple of 8, otherwise this panics.
fn decode_dxt1_row(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() % 8 == 0);
    let block_count = source.len() / 8;
    assert!(dest.len() >= block_count * 48);

    // contains the 16 decoded pixels per block
    let mut decoded_block = [0u8; 48];

    for (x, encoded_block) in source.chunks(8).enumerate() {
        decode_dxt1_block(encoded_block, &mut decoded_block);

        // copy the values from the decoded block to linewise RGB layout
        for line in 0..4 {
            let offset = (block_count * line + x) * 12;
            dest[offset..offset + 12].copy_from_slice(&decoded_block[line * 12..(line + 1) * 12]);
        }
    }
}

/// Decode a row of DXT3 data to four rows of RGBA data.
/// source.len() should be a multiple of 16, otherwise this panics.
fn decode_dxt3_row(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() % 16 == 0);
    let block_count = source.len() / 16;
    assert!(dest.len() >= block_count * 64);

    // contains the 16 decoded pixels per block
    let mut decoded_block = [0u8; 64];

    for (x, encoded_block) in source.chunks(16).enumerate() {
        decode_dxt3_block(encoded_block, &mut decoded_block);

        // copy the values from the decoded block to linewise RGB layout
        for line in 0..4 {
            let offset = (block_count * line + x) * 16;
            dest[offset..offset + 16].copy_from_slice(&decoded_block[line * 16..(line + 1) * 16]);
        }
    }
}

/// Decode a row of DXT5 data to four rows of RGBA data.
/// source.len() should be a multiple of 16, otherwise this panics.
fn decode_dxt5_row(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() % 16 == 0);
    let block_count = source.len() / 16;
    assert!(dest.len() >= block_count * 64);

    // contains the 16 decoded pixels per block
    let mut decoded_block = [0u8; 64];

    for (x, encoded_block) in source.chunks(16).enumerate() {
        decode_dxt5_block(encoded_block, &mut decoded_block);

        // copy the values from the decoded block to linewise RGB layout
        for line in 0..4 {
            let offset = (block_count * line + x) * 16;
            dest[offset..offset + 16].copy_from_slice(&decoded_block[line * 16..(line + 1) * 16]);
        }
    }
}

/*
 * Functions for encoding DXT compression
 */

/// Tries to perform the color encoding part of dxt compression
/// the approach taken is simple, it picks unique combinations
/// of the colors present in the block, and attempts to encode the
/// block with each, picking the encoding that yields the least
/// squared error out of all of them.
///
/// This could probably be faster but is already reasonably fast
/// and a good reference impl to optimize others against.
///
/// Another way to perform this analysis would be to perform a
/// singular value decomposition of the different colors, and
/// then pick 2 points on this line as the base colors. But
/// this is still rather unwieldly math and has issues
/// with the 3-linear-colors-and-0 case, it's also worse
/// at conserving the original colors.
///
/// source: should be RGBAx16 or RGBx16 bytes of data,
/// dest 8 bytes of resulting encoded color data
fn encode_dxt_colors(source: &[u8], dest: &mut [u8]) {
    // sanity checks and determine stride when parsing the source data
    assert!((source.len() == 64 || source.len() == 48) && dest.len() == 8);
    let stride = source.len() / 16;

    // reference colors array
    let mut colors = [[0u8; 3]; 4];

    // Put the colors we're going to be processing in an array with pure RGB layout
    // note: we reverse the pixel order here. The reason for this is found in the inner quantization loop.
    let mut targets = [[0u8; 3]; 16];
    for (s, d) in source.chunks(stride).rev().zip(&mut targets) {
        *d = [s[0], s[1], s[2]];
    }

    // and a set of colors to pick from.
    let mut colorspace = targets.to_vec();

    // roundtrip all colors through the r5g6b5 encoding
    for rgb in &mut colorspace {
        *rgb = enc565_decode(enc565_encode(*rgb));
    }

    // and deduplicate the set of colors to choose from as the algorithm is O(N^2) in this
    colorspace.dedup();

    // in case of slight gradients it can happen that there's only one entry left in the color table.
    // as the resulting banding can be quite bad if we would just left the block at the closest
    // encodable color, we have a special path here that tries to emulate the wanted color
    // using the linear interpolation between gradients
    if colorspace.len() == 1 {
        // the base color we got from colorspace reduction
        let ref_rgb = colorspace[0];
        // the unreduced color in this block that's the furthest away from the actual block
        let mut rgb = targets
            .iter()
            .cloned()
            .max_by_key(|rgb| diff(*rgb, ref_rgb))
            .unwrap();
        // amplify differences by 2.5, which should push them to the next quantized value
        // if possible without overshoot
        for i in 0..3 {
            rgb[i] =
                ((i16::from(rgb[i]) - i16::from(ref_rgb[i])) * 5 / 2 + i16::from(ref_rgb[i])) as u8;
        }

        // roundtrip it through quantization
        let encoded = enc565_encode(rgb);
        let rgb = enc565_decode(encoded);

        // in case this didn't land us a different color the best way to represent this field is
        // as a single color block
        if rgb == ref_rgb {
            dest[0] = encoded as u8;
            dest[1] = (encoded >> 8) as u8;

            for d in dest.iter_mut().take(8).skip(2) {
                *d = 0;
            }
            return;
        }

        // we did find a separate value: add it to the options so after one round of quantization
        // we're done
        colorspace.push(rgb);
    }

    // block quantization loop: we basically just try every possible combination, returning
    // the combination with the least squared error
    // stores the best candidate colors
    let mut chosen_colors = [[0; 3]; 4];
    // did this index table use the [0,0,0] variant
    let mut chosen_use_0 = false;
    // error calculated for the last entry
    let mut chosen_error = 0xFFFF_FFFFu32;

    // loop through unique permutations of the colorspace, where c1 != c2
    'search: for (i, &c1) in colorspace.iter().enumerate() {
        colors[0] = c1;

        for &c2 in &colorspace[0..i] {
            colors[1] = c2;

            // what's inside here is ran at most 120 times.
            for use_0 in 0..2 {
                // and 240 times here.

                if use_0 != 0 {
                    // interpolate one color, set the other to 0
                    for i in 0..3 {
                        colors[2][i] =
                            ((u16::from(colors[0][i]) + u16::from(colors[1][i]) + 1) / 2) as u8;
                    }
                    colors[3] = [0, 0, 0];
                } else {
                    // interpolate to get 2 more colors
                    for i in 0..3 {
                        colors[2][i] =
                            ((u16::from(colors[0][i]) * 2 + u16::from(colors[1][i]) + 1) / 3) as u8;
                        colors[3][i] =
                            ((u16::from(colors[0][i]) + u16::from(colors[1][i]) * 2 + 1) / 3) as u8;
                    }
                }

                // calculate the total error if we were to quantize the block with these color combinations
                // both these loops have statically known iteration counts and are well vectorizable
                // note that the inside of this can be run about 15360 times worst case, i.e. 960 times per
                // pixel.
                let total_error = targets
                    .iter()
                    .map(|t| colors.iter().map(|c| diff(*c, *t) as u32).min().unwrap())
                    .sum();

                // update the match if we found a better one
                if total_error < chosen_error {
                    chosen_colors = colors;
                    chosen_use_0 = use_0 != 0;
                    chosen_error = total_error;

                    // if we've got a perfect or at most 1 LSB off match, we're done
                    if total_error < 4 {
                        break 'search;
                    }
                }
            }
        }
    }

    // calculate the final indices
    // note that targets is already in reverse pixel order, to make the index computation easy.
    let mut chosen_indices = 0u32;
    for t in &targets {
        let (idx, _) = chosen_colors
            .iter()
            .enumerate()
            .min_by_key(|&(_, c)| diff(*c, *t))
            .unwrap();
        chosen_indices = (chosen_indices << 2) | idx as u32;
    }

    // encode the colors
    let mut color0 = enc565_encode(chosen_colors[0]);
    let mut color1 = enc565_encode(chosen_colors[1]);

    // determine encoding. Note that color0 == color1 is impossible at this point
    if color0 > color1 {
        if chosen_use_0 {
            swap(&mut color0, &mut color1);
            // Indexes are packed 2 bits wide, swap index 0/1 but preserve 2/3.
            let filter = (chosen_indices & 0xAAAA_AAAA) >> 1;
            chosen_indices ^= filter ^ 0x5555_5555;
        }
    } else if !chosen_use_0 {
        swap(&mut color0, &mut color1);
        // Indexes are packed 2 bits wide, swap index 0/1 and 2/3.
        chosen_indices ^= 0x5555_5555;
    }

    // encode everything.
    dest[0] = color0 as u8;
    dest[1] = (color0 >> 8) as u8;
    dest[2] = color1 as u8;
    dest[3] = (color1 >> 8) as u8;
    for i in 0..4 {
        dest[i + 4] = (chosen_indices >> (i * 8)) as u8;
    }
}

/// Encodes a buffer of 16 alpha bytes into a dxt5 alpha index table,
/// where the alpha table they are indexed against is created by
/// calling alpha_table_dxt5(alpha0, alpha1)
/// returns the resulting error and alpha table
fn encode_dxt5_alpha(alpha0: u8, alpha1: u8, alphas: &[u8; 16]) -> (i32, u64) {
    // create a table for the given alpha ranges
    let table = alpha_table_dxt5(alpha0, alpha1);
    let mut indices = 0u64;
    let mut total_error = 0i32;

    // least error brute force search
    for (i, &a) in alphas.iter().enumerate() {
        let (index, error) = table
            .iter()
            .enumerate()
            .map(|(i, &e)| (i, square(i32::from(e) - i32::from(a))))
            .min_by_key(|&(_, e)| e)
            .unwrap();
        total_error += error;
        indices |= (index as u64) << (i * 3);
    }

    (total_error, indices)
}

/// Encodes a RGBAx16 sequence of bytes to a 16 bytes DXT5 block
fn encode_dxt5_block(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() == 64 && dest.len() == 16);

    // perform dxt color encoding
    encode_dxt_colors(source, &mut dest[8..16]);

    // copy out the alpha bytes
    let mut alphas = [0; 16];
    for i in 0..16 {
        alphas[i] = source[i * 4 + 3];
    }

    // try both alpha compression methods, see which has the least error.
    let alpha07 = alphas.iter().cloned().min().unwrap();
    let alpha17 = alphas.iter().cloned().max().unwrap();
    let (error7, indices7) = encode_dxt5_alpha(alpha07, alpha17, &alphas);

    // if all alphas are 0 or 255 it doesn't particularly matter what we do here.
    let alpha05 = alphas
        .iter()
        .cloned()
        .filter(|&i| i != 255)
        .max()
        .unwrap_or(255);
    let alpha15 = alphas
        .iter()
        .cloned()
        .filter(|&i| i != 0)
        .min()
        .unwrap_or(0);
    let (error5, indices5) = encode_dxt5_alpha(alpha05, alpha15, &alphas);

    // pick the best one, encode the min/max values
    let mut alpha_table = if error5 < error7 {
        dest[0] = alpha05;
        dest[1] = alpha15;
        indices5
    } else {
        dest[0] = alpha07;
        dest[1] = alpha17;
        indices7
    };

    // encode the alphas
    for byte in dest[2..8].iter_mut() {
        *byte = alpha_table as u8;
        alpha_table >>= 8;
    }
}

/// Encodes a RGBAx16 sequence of bytes into a 16 bytes DXT3 block
fn encode_dxt3_block(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() == 64 && dest.len() == 16);

    // perform dxt color encoding
    encode_dxt_colors(source, &mut dest[8..16]);

    // DXT3 alpha compression is very simple, just round towards the nearest value

    // index the alpha values into the 64bit alpha table
    let mut alpha_table = 0u64;
    for i in 0..16 {
        let alpha = u64::from(source[i * 4 + 3]);
        let alpha = (alpha + 0x8) / 0x11;
        alpha_table |= alpha << (i * 4);
    }

    // encode the alpha values
    for byte in &mut dest[0..8] {
        *byte = alpha_table as u8;
        alpha_table >>= 8;
    }
}

/// Encodes a RGBx16 sequence of bytes into a 8 bytes DXT1 block
fn encode_dxt1_block(source: &[u8], dest: &mut [u8]) {
    assert!(source.len() == 48 && dest.len() == 8);

    // perform dxt color encoding
    encode_dxt_colors(source, dest);
}

/// Decode a row of DXT1 data to four rows of RGBA data.
/// source.len() should be a multiple of 8, otherwise this panics.
fn encode_dxt1_row(source: &[u8]) -> Vec<u8> {
    assert!(source.len() % 48 == 0);
    let block_count = source.len() / 48;

    let mut dest = vec![0u8; block_count * 8];
    // contains the 16 decoded pixels per block
    let mut decoded_block = [0u8; 48];

    for (x, encoded_block) in dest.chunks_mut(8).enumerate() {
        // copy the values from the decoded block to linewise RGB layout
        for line in 0..4 {
            let offset = (block_count * line + x) * 12;
            decoded_block[line * 12..(line + 1) * 12].copy_from_slice(&source[offset..offset + 12]);
        }

        encode_dxt1_block(&decoded_block, encoded_block);
    }
    dest
}

/// Decode a row of DXT3 data to four rows of RGBA data.
/// source.len() should be a multiple of 16, otherwise this panics.
fn encode_dxt3_row(source: &[u8]) -> Vec<u8> {
    assert!(source.len() % 64 == 0);
    let block_count = source.len() / 64;

    let mut dest = vec![0u8; block_count * 16];
    // contains the 16 decoded pixels per block
    let mut decoded_block = [0u8; 64];

    for (x, encoded_block) in dest.chunks_mut(16).enumerate() {
        // copy the values from the decoded block to linewise RGB layout
        for line in 0..4 {
            let offset = (block_count * line + x) * 16;
            decoded_block[line * 16..(line + 1) * 16].copy_from_slice(&source[offset..offset + 16]);
        }

        encode_dxt3_block(&decoded_block, encoded_block);
    }
    dest
}

/// Decode a row of DXT5 data to four rows of RGBA data.
/// source.len() should be a multiple of 16, otherwise this panics.
fn encode_dxt5_row(source: &[u8]) -> Vec<u8> {
    assert!(source.len() % 64 == 0);
    let block_count = source.len() / 64;

    let mut dest = vec![0u8; block_count * 16];
    // contains the 16 decoded pixels per block
    let mut decoded_block = [0u8; 64];

    for (x, encoded_block) in dest.chunks_mut(16).enumerate() {
        // copy the values from the decoded block to linewise RGB layout
        for line in 0..4 {
            let offset = (block_count * line + x) * 16;
            decoded_block[line * 16..(line + 1) * 16].copy_from_slice(&source[offset..offset + 16]);
        }

        encode_dxt5_block(&decoded_block, encoded_block);
    }
    dest
}