image/codecs/
dxt.rs

1//!  Decoding of DXT (S3TC) compression
2//!
3//!  DXT is an image format that supports lossy compression
4//!
5//!  # Related Links
6//!  * <https://www.khronos.org/registry/OpenGL/extensions/EXT/EXT_texture_compression_s3tc.txt> - Description of the DXT compression OpenGL extensions.
7//!
8//!  Note: this module only implements bare DXT encoding/decoding, it does not parse formats that can contain DXT files like .dds
9
10use std::io::{self, Read, Seek, SeekFrom, Write};
11
12use crate::color::ColorType;
13use crate::error::{ImageError, ImageResult, ParameterError, ParameterErrorKind};
14use crate::image::{self, ImageDecoder, ImageDecoderRect, ImageReadBuffer, Progress};
15
16/// What version of DXT compression are we using?
17/// Note that DXT2 and DXT4 are left away as they're
18/// just DXT3 and DXT5 with premultiplied alpha
19#[derive(Clone, Copy, Debug, PartialEq, Eq)]
20pub enum DxtVariant {
21    /// The DXT1 format. 48 bytes of RGB data in a 4x4 pixel square is
22    /// compressed into an 8 byte block of DXT1 data
23    DXT1,
24    /// The DXT3 format. 64 bytes of RGBA data in a 4x4 pixel square is
25    /// compressed into a 16 byte block of DXT3 data
26    DXT3,
27    /// The DXT5 format. 64 bytes of RGBA data in a 4x4 pixel square is
28    /// compressed into a 16 byte block of DXT5 data
29    DXT5,
30}
31
32impl DxtVariant {
33    /// Returns the amount of bytes of raw image data
34    /// that is encoded in a single DXTn block
35    fn decoded_bytes_per_block(self) -> usize {
36        match self {
37            DxtVariant::DXT1 => 48,
38            DxtVariant::DXT3 | DxtVariant::DXT5 => 64,
39        }
40    }
41
42    /// Returns the amount of bytes per block of encoded DXTn data
43    fn encoded_bytes_per_block(self) -> usize {
44        match self {
45            DxtVariant::DXT1 => 8,
46            DxtVariant::DXT3 | DxtVariant::DXT5 => 16,
47        }
48    }
49
50    /// Returns the color type that is stored in this DXT variant
51    pub fn color_type(self) -> ColorType {
52        match self {
53            DxtVariant::DXT1 => ColorType::Rgb8,
54            DxtVariant::DXT3 | DxtVariant::DXT5 => ColorType::Rgba8,
55        }
56    }
57}
58
59/// DXT decoder
60pub struct DxtDecoder<R: Read> {
61    inner: R,
62    width_blocks: u32,
63    height_blocks: u32,
64    variant: DxtVariant,
65    row: u32,
66}
67
68impl<R: Read> DxtDecoder<R> {
69    /// Create a new DXT decoder that decodes from the stream ```r```.
70    /// As DXT is often stored as raw buffers with the width/height
71    /// somewhere else the width and height of the image need
72    /// to be passed in ```width``` and ```height```, as well as the
73    /// DXT variant in ```variant```.
74    /// width and height are required to be powers of 2 and at least 4.
75    /// otherwise an error will be returned
76    pub fn new(
77        r: R,
78        width: u32,
79        height: u32,
80        variant: DxtVariant,
81    ) -> Result<DxtDecoder<R>, ImageError> {
82        if width % 4 != 0 || height % 4 != 0 {
83            // TODO: this is actually a bit of a weird case. We could return `DecodingError` but
84            // it's not really the format that is wrong However, the encoder should surely return
85            // `EncodingError` so it would be the logical choice for symmetry.
86            return Err(ImageError::Parameter(ParameterError::from_kind(
87                ParameterErrorKind::DimensionMismatch,
88            )));
89        }
90        let width_blocks = width / 4;
91        let height_blocks = height / 4;
92        Ok(DxtDecoder {
93            inner: r,
94            width_blocks,
95            height_blocks,
96            variant,
97            row: 0,
98        })
99    }
100
101    fn read_scanline(&mut self, buf: &mut [u8]) -> io::Result<usize> {
102        assert_eq!(
103            u64::try_from(buf.len()),
104            Ok(
105                #[allow(deprecated)]
106                self.scanline_bytes()
107            )
108        );
109
110        let mut src =
111            vec![0u8; self.variant.encoded_bytes_per_block() * self.width_blocks as usize];
112        self.inner.read_exact(&mut src)?;
113        match self.variant {
114            DxtVariant::DXT1 => decode_dxt1_row(&src, buf),
115            DxtVariant::DXT3 => decode_dxt3_row(&src, buf),
116            DxtVariant::DXT5 => decode_dxt5_row(&src, buf),
117        }
118        self.row += 1;
119        Ok(buf.len())
120    }
121}
122
123// Note that, due to the way that DXT compression works, a scanline is considered to consist out of
124// 4 lines of pixels.
125impl<'a, R: 'a + Read> ImageDecoder<'a> for DxtDecoder<R> {
126    type Reader = DxtReader<R>;
127
128    fn dimensions(&self) -> (u32, u32) {
129        (self.width_blocks * 4, self.height_blocks * 4)
130    }
131
132    fn color_type(&self) -> ColorType {
133        self.variant.color_type()
134    }
135
136    fn scanline_bytes(&self) -> u64 {
137        self.variant.decoded_bytes_per_block() as u64 * u64::from(self.width_blocks)
138    }
139
140    fn into_reader(self) -> ImageResult<Self::Reader> {
141        Ok(DxtReader {
142            buffer: ImageReadBuffer::new(
143                #[allow(deprecated)]
144                self.scanline_bytes(),
145                self.total_bytes(),
146            ),
147            decoder: self,
148        })
149    }
150
151    fn read_image(mut self, buf: &mut [u8]) -> ImageResult<()> {
152        assert_eq!(u64::try_from(buf.len()), Ok(self.total_bytes()));
153
154        #[allow(deprecated)]
155        for chunk in buf.chunks_mut(self.scanline_bytes().max(1) as usize) {
156            self.read_scanline(chunk)?;
157        }
158        Ok(())
159    }
160}
161
162impl<'a, R: 'a + Read + Seek> ImageDecoderRect<'a> for DxtDecoder<R> {
163    fn read_rect_with_progress<F: Fn(Progress)>(
164        &mut self,
165        x: u32,
166        y: u32,
167        width: u32,
168        height: u32,
169        buf: &mut [u8],
170        progress_callback: F,
171    ) -> ImageResult<()> {
172        let encoded_scanline_bytes =
173            self.variant.encoded_bytes_per_block() as u64 * u64::from(self.width_blocks);
174
175        let start = self.inner.stream_position()?;
176        image::load_rect(
177            x,
178            y,
179            width,
180            height,
181            buf,
182            progress_callback,
183            self,
184            |s, scanline| {
185                s.inner
186                    .seek(SeekFrom::Start(start + scanline * encoded_scanline_bytes))?;
187                Ok(())
188            },
189            |s, buf| s.read_scanline(buf).map(|_| ()),
190        )?;
191        self.inner.seek(SeekFrom::Start(start))?;
192        Ok(())
193    }
194}
195
196/// DXT reader
197pub struct DxtReader<R: Read> {
198    buffer: ImageReadBuffer,
199    decoder: DxtDecoder<R>,
200}
201
202impl<R: Read> Read for DxtReader<R> {
203    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
204        let decoder = &mut self.decoder;
205        self.buffer.read(buf, |buf| decoder.read_scanline(buf))
206    }
207}
208
209/// DXT encoder
210pub struct DxtEncoder<W: Write> {
211    w: W,
212}
213
214impl<W: Write> DxtEncoder<W> {
215    /// Create a new encoder that writes its output to ```w```
216    pub fn new(w: W) -> DxtEncoder<W> {
217        DxtEncoder { w }
218    }
219
220    /// Encodes the image data ```data```
221    /// that has dimensions ```width``` and ```height```
222    /// in ```DxtVariant``` ```variant```
223    /// data is assumed to be in variant.color_type()
224    pub fn encode(
225        mut self,
226        data: &[u8],
227        width: u32,
228        height: u32,
229        variant: DxtVariant,
230    ) -> ImageResult<()> {
231        if width % 4 != 0 || height % 4 != 0 {
232            // TODO: this is not very idiomatic yet. Should return an EncodingError.
233            return Err(ImageError::Parameter(ParameterError::from_kind(
234                ParameterErrorKind::DimensionMismatch,
235            )));
236        }
237        let width_blocks = width / 4;
238        let height_blocks = height / 4;
239
240        let stride = variant.decoded_bytes_per_block();
241
242        assert!(data.len() >= width_blocks as usize * height_blocks as usize * stride);
243
244        for chunk in data.chunks(width_blocks as usize * stride) {
245            let data = match variant {
246                DxtVariant::DXT1 => encode_dxt1_row(chunk),
247                DxtVariant::DXT3 => encode_dxt3_row(chunk),
248                DxtVariant::DXT5 => encode_dxt5_row(chunk),
249            };
250            self.w.write_all(&data)?;
251        }
252        Ok(())
253    }
254}
255
256/**
257 * Actual encoding/decoding logic below.
258 */
259use std::mem::swap;
260
261type Rgb = [u8; 3];
262
263/// decodes a 5-bit R, 6-bit G, 5-bit B 16-bit packed color value into 8-bit RGB
264/// mapping is done so min/max range values are preserved. So for 5-bit
265/// values 0x00 -> 0x00 and 0x1F -> 0xFF
266fn enc565_decode(value: u16) -> Rgb {
267    let red = (value >> 11) & 0x1F;
268    let green = (value >> 5) & 0x3F;
269    let blue = (value) & 0x1F;
270    [
271        (red * 0xFF / 0x1F) as u8,
272        (green * 0xFF / 0x3F) as u8,
273        (blue * 0xFF / 0x1F) as u8,
274    ]
275}
276
277/// encodes an 8-bit RGB value into a 5-bit R, 6-bit G, 5-bit B 16-bit packed color value
278/// mapping preserves min/max values. It is guaranteed that i == encode(decode(i)) for all i
279fn enc565_encode(rgb: Rgb) -> u16 {
280    let red = (u16::from(rgb[0]) * 0x1F + 0x7E) / 0xFF;
281    let green = (u16::from(rgb[1]) * 0x3F + 0x7E) / 0xFF;
282    let blue = (u16::from(rgb[2]) * 0x1F + 0x7E) / 0xFF;
283    (red << 11) | (green << 5) | blue
284}
285
286/// utility function: squares a value
287fn square(a: i32) -> i32 {
288    a * a
289}
290
291/// returns the squared error between two RGB values
292fn diff(a: Rgb, b: Rgb) -> i32 {
293    square(i32::from(a[0]) - i32::from(b[0]))
294        + square(i32::from(a[1]) - i32::from(b[1]))
295        + square(i32::from(a[2]) - i32::from(b[2]))
296}
297
298/*
299 * Functions for decoding DXT compression
300 */
301
302/// Constructs the DXT5 alpha lookup table from the two alpha entries
303/// if alpha0 > alpha1, constructs a table of [a0, a1, 6 linearly interpolated values from a0 to a1]
304/// if alpha0 <= alpha1, constructs a table of [a0, a1, 4 linearly interpolated values from a0 to a1, 0, 0xFF]
305fn alpha_table_dxt5(alpha0: u8, alpha1: u8) -> [u8; 8] {
306    let mut table = [alpha0, alpha1, 0, 0, 0, 0, 0, 0xFF];
307    if alpha0 > alpha1 {
308        for i in 2..8u16 {
309            table[i as usize] =
310                (((8 - i) * u16::from(alpha0) + (i - 1) * u16::from(alpha1)) / 7) as u8;
311        }
312    } else {
313        for i in 2..6u16 {
314            table[i as usize] =
315                (((6 - i) * u16::from(alpha0) + (i - 1) * u16::from(alpha1)) / 5) as u8;
316        }
317    }
318    table
319}
320
321/// decodes an 8-byte dxt color block into the RGB channels of a 16xRGB or 16xRGBA block.
322/// source should have a length of 8, dest a length of 48 (RGB) or 64 (RGBA)
323fn decode_dxt_colors(source: &[u8], dest: &mut [u8], is_dxt1: bool) {
324    // sanity checks, also enable the compiler to elide all following bound checks
325    assert!(source.len() == 8 && (dest.len() == 48 || dest.len() == 64));
326    // calculate pitch to store RGB values in dest (3 for RGB, 4 for RGBA)
327    let pitch = dest.len() / 16;
328
329    // extract color data
330    let color0 = u16::from(source[0]) | (u16::from(source[1]) << 8);
331    let color1 = u16::from(source[2]) | (u16::from(source[3]) << 8);
332    let color_table = u32::from(source[4])
333        | (u32::from(source[5]) << 8)
334        | (u32::from(source[6]) << 16)
335        | (u32::from(source[7]) << 24);
336    // let color_table = source[4..8].iter().rev().fold(0, |t, &b| (t << 8) | b as u32);
337
338    // decode the colors to rgb format
339    let mut colors = [[0; 3]; 4];
340    colors[0] = enc565_decode(color0);
341    colors[1] = enc565_decode(color1);
342
343    // determine color interpolation method
344    if color0 > color1 || !is_dxt1 {
345        // linearly interpolate the other two color table entries
346        for i in 0..3 {
347            colors[2][i] = ((u16::from(colors[0][i]) * 2 + u16::from(colors[1][i]) + 1) / 3) as u8;
348            colors[3][i] = ((u16::from(colors[0][i]) + u16::from(colors[1][i]) * 2 + 1) / 3) as u8;
349        }
350    } else {
351        // linearly interpolate one other entry, keep the other at 0
352        for i in 0..3 {
353            colors[2][i] = ((u16::from(colors[0][i]) + u16::from(colors[1][i]) + 1) / 2) as u8;
354        }
355    }
356
357    // serialize the result. Every color is determined by looking up
358    // two bits in color_table which identify which color to actually pick from the 4 possible colors
359    for i in 0..16 {
360        dest[i * pitch..i * pitch + 3]
361            .copy_from_slice(&colors[(color_table >> (i * 2)) as usize & 3]);
362    }
363}
364
365/// Decodes a 16-byte bock of dxt5 data to a 16xRGBA block
366fn decode_dxt5_block(source: &[u8], dest: &mut [u8]) {
367    assert!(source.len() == 16 && dest.len() == 64);
368
369    // extract alpha index table (stored as little endian 64-bit value)
370    let alpha_table = source[2..8]
371        .iter()
372        .rev()
373        .fold(0, |t, &b| (t << 8) | u64::from(b));
374
375    // alhpa level decode
376    let alphas = alpha_table_dxt5(source[0], source[1]);
377
378    // serialize alpha
379    for i in 0..16 {
380        dest[i * 4 + 3] = alphas[(alpha_table >> (i * 3)) as usize & 7];
381    }
382
383    // handle colors
384    decode_dxt_colors(&source[8..16], dest, false);
385}
386
387/// Decodes a 16-byte bock of dxt3 data to a 16xRGBA block
388fn decode_dxt3_block(source: &[u8], dest: &mut [u8]) {
389    assert!(source.len() == 16 && dest.len() == 64);
390
391    // extract alpha index table (stored as little endian 64-bit value)
392    let alpha_table = source[0..8]
393        .iter()
394        .rev()
395        .fold(0, |t, &b| (t << 8) | u64::from(b));
396
397    // serialize alpha (stored as 4-bit values)
398    for i in 0..16 {
399        dest[i * 4 + 3] = ((alpha_table >> (i * 4)) as u8 & 0xF) * 0x11;
400    }
401
402    // handle colors
403    decode_dxt_colors(&source[8..16], dest, false);
404}
405
406/// Decodes a 8-byte bock of dxt5 data to a 16xRGB block
407fn decode_dxt1_block(source: &[u8], dest: &mut [u8]) {
408    assert!(source.len() == 8 && dest.len() == 48);
409    decode_dxt_colors(source, dest, true);
410}
411
412/// Decode a row of DXT1 data to four rows of RGB data.
413/// source.len() should be a multiple of 8, otherwise this panics.
414fn decode_dxt1_row(source: &[u8], dest: &mut [u8]) {
415    assert!(source.len() % 8 == 0);
416    let block_count = source.len() / 8;
417    assert!(dest.len() >= block_count * 48);
418
419    // contains the 16 decoded pixels per block
420    let mut decoded_block = [0u8; 48];
421
422    for (x, encoded_block) in source.chunks(8).enumerate() {
423        decode_dxt1_block(encoded_block, &mut decoded_block);
424
425        // copy the values from the decoded block to linewise RGB layout
426        for line in 0..4 {
427            let offset = (block_count * line + x) * 12;
428            dest[offset..offset + 12].copy_from_slice(&decoded_block[line * 12..(line + 1) * 12]);
429        }
430    }
431}
432
433/// Decode a row of DXT3 data to four rows of RGBA data.
434/// source.len() should be a multiple of 16, otherwise this panics.
435fn decode_dxt3_row(source: &[u8], dest: &mut [u8]) {
436    assert!(source.len() % 16 == 0);
437    let block_count = source.len() / 16;
438    assert!(dest.len() >= block_count * 64);
439
440    // contains the 16 decoded pixels per block
441    let mut decoded_block = [0u8; 64];
442
443    for (x, encoded_block) in source.chunks(16).enumerate() {
444        decode_dxt3_block(encoded_block, &mut decoded_block);
445
446        // copy the values from the decoded block to linewise RGB layout
447        for line in 0..4 {
448            let offset = (block_count * line + x) * 16;
449            dest[offset..offset + 16].copy_from_slice(&decoded_block[line * 16..(line + 1) * 16]);
450        }
451    }
452}
453
454/// Decode a row of DXT5 data to four rows of RGBA data.
455/// source.len() should be a multiple of 16, otherwise this panics.
456fn decode_dxt5_row(source: &[u8], dest: &mut [u8]) {
457    assert!(source.len() % 16 == 0);
458    let block_count = source.len() / 16;
459    assert!(dest.len() >= block_count * 64);
460
461    // contains the 16 decoded pixels per block
462    let mut decoded_block = [0u8; 64];
463
464    for (x, encoded_block) in source.chunks(16).enumerate() {
465        decode_dxt5_block(encoded_block, &mut decoded_block);
466
467        // copy the values from the decoded block to linewise RGB layout
468        for line in 0..4 {
469            let offset = (block_count * line + x) * 16;
470            dest[offset..offset + 16].copy_from_slice(&decoded_block[line * 16..(line + 1) * 16]);
471        }
472    }
473}
474
475/*
476 * Functions for encoding DXT compression
477 */
478
479/// Tries to perform the color encoding part of dxt compression
480/// the approach taken is simple, it picks unique combinations
481/// of the colors present in the block, and attempts to encode the
482/// block with each, picking the encoding that yields the least
483/// squared error out of all of them.
484///
485/// This could probably be faster but is already reasonably fast
486/// and a good reference impl to optimize others against.
487///
488/// Another way to perform this analysis would be to perform a
489/// singular value decomposition of the different colors, and
490/// then pick 2 points on this line as the base colors. But
491/// this is still rather unwieldy math and has issues
492/// with the 3-linear-colors-and-0 case, it's also worse
493/// at conserving the original colors.
494///
495/// source: should be RGBAx16 or RGBx16 bytes of data,
496/// dest 8 bytes of resulting encoded color data
497fn encode_dxt_colors(source: &[u8], dest: &mut [u8], is_dxt1: bool) {
498    // sanity checks and determine stride when parsing the source data
499    assert!((source.len() == 64 || source.len() == 48) && dest.len() == 8);
500    let stride = source.len() / 16;
501
502    // reference colors array
503    let mut colors = [[0u8; 3]; 4];
504
505    // Put the colors we're going to be processing in an array with pure RGB layout
506    // note: we reverse the pixel order here. The reason for this is found in the inner quantization loop.
507    let mut targets = [[0u8; 3]; 16];
508    for (s, d) in source.chunks(stride).rev().zip(&mut targets) {
509        *d = [s[0], s[1], s[2]];
510    }
511
512    // roundtrip all colors through the r5g6b5 encoding
513    for rgb in &mut targets {
514        *rgb = enc565_decode(enc565_encode(*rgb));
515    }
516
517    // and deduplicate the set of colors to choose from as the algorithm is O(N^2) in this
518    let mut colorspace_ = [[0u8; 3]; 16];
519    let mut colorspace_len = 0;
520    for color in &targets {
521        if !colorspace_[..colorspace_len].contains(color) {
522            colorspace_[colorspace_len] = *color;
523            colorspace_len += 1;
524        }
525    }
526    let mut colorspace = &colorspace_[..colorspace_len];
527
528    // in case of slight gradients it can happen that there's only one entry left in the color table.
529    // as the resulting banding can be quite bad if we would just left the block at the closest
530    // encodable color, we have a special path here that tries to emulate the wanted color
531    // using the linear interpolation between gradients
532    if colorspace.len() == 1 {
533        // the base color we got from colorspace reduction
534        let ref_rgb = colorspace[0];
535        // the unreduced color in this block that's the furthest away from the actual block
536        let mut rgb = targets
537            .iter()
538            .cloned()
539            .max_by_key(|rgb| diff(*rgb, ref_rgb))
540            .unwrap();
541        // amplify differences by 2.5, which should push them to the next quantized value
542        // if possible without overshoot
543        for i in 0..3 {
544            rgb[i] =
545                ((i16::from(rgb[i]) - i16::from(ref_rgb[i])) * 5 / 2 + i16::from(ref_rgb[i])) as u8;
546        }
547
548        // roundtrip it through quantization
549        let encoded = enc565_encode(rgb);
550        let rgb = enc565_decode(encoded);
551
552        // in case this didn't land us a different color the best way to represent this field is
553        // as a single color block
554        if rgb == ref_rgb {
555            dest[0] = encoded as u8;
556            dest[1] = (encoded >> 8) as u8;
557
558            for d in dest.iter_mut().take(8).skip(2) {
559                *d = 0;
560            }
561            return;
562        }
563
564        // we did find a separate value: add it to the options so after one round of quantization
565        // we're done
566        colorspace_[1] = rgb;
567        colorspace = &colorspace_[..2];
568    }
569
570    // block quantization loop: we basically just try every possible combination, returning
571    // the combination with the least squared error
572    // stores the best candidate colors
573    let mut chosen_colors = [[0; 3]; 4];
574    // did this index table use the [0,0,0] variant
575    let mut chosen_use_0 = false;
576    // error calculated for the last entry
577    let mut chosen_error = 0xFFFF_FFFFu32;
578
579    // loop through unique permutations of the colorspace, where c1 != c2
580    'search: for (i, &c1) in colorspace.iter().enumerate() {
581        colors[0] = c1;
582
583        for &c2 in &colorspace[0..i] {
584            colors[1] = c2;
585
586            if is_dxt1 {
587                // what's inside here is ran at most 120 times.
588                for use_0 in 0..2 {
589                    // and 240 times here.
590
591                    if use_0 != 0 {
592                        // interpolate one color, set the other to 0
593                        for i in 0..3 {
594                            colors[2][i] =
595                                ((u16::from(colors[0][i]) + u16::from(colors[1][i]) + 1) / 2) as u8;
596                        }
597                        colors[3] = [0, 0, 0];
598                    } else {
599                        // interpolate to get 2 more colors
600                        for i in 0..3 {
601                            colors[2][i] =
602                                ((u16::from(colors[0][i]) * 2 + u16::from(colors[1][i]) + 1) / 3)
603                                    as u8;
604                            colors[3][i] =
605                                ((u16::from(colors[0][i]) + u16::from(colors[1][i]) * 2 + 1) / 3)
606                                    as u8;
607                        }
608                    }
609
610                    // calculate the total error if we were to quantize the block with these color combinations
611                    // both these loops have statically known iteration counts and are well vectorizable
612                    // note that the inside of this can be run about 15360 times worst case, i.e. 960 times per
613                    // pixel.
614                    let total_error = targets
615                        .iter()
616                        .map(|t| colors.iter().map(|c| diff(*c, *t) as u32).min().unwrap())
617                        .sum();
618
619                    // update the match if we found a better one
620                    if total_error < chosen_error {
621                        chosen_colors = colors;
622                        chosen_use_0 = use_0 != 0;
623                        chosen_error = total_error;
624
625                        // if we've got a perfect or at most 1 LSB off match, we're done
626                        if total_error < 4 {
627                            break 'search;
628                        }
629                    }
630                }
631            } else {
632                // what's inside here is ran at most 120 times.
633
634                // interpolate to get 2 more colors
635                for i in 0..3 {
636                    colors[2][i] =
637                        ((u16::from(colors[0][i]) * 2 + u16::from(colors[1][i]) + 1) / 3) as u8;
638                    colors[3][i] =
639                        ((u16::from(colors[0][i]) + u16::from(colors[1][i]) * 2 + 1) / 3) as u8;
640                }
641
642                // calculate the total error if we were to quantize the block with these color combinations
643                // both these loops have statically known iteration counts and are well vectorizable
644                // note that the inside of this can be run about 15360 times worst case, i.e. 960 times per
645                // pixel.
646                let total_error = targets
647                    .iter()
648                    .map(|t| colors.iter().map(|c| diff(*c, *t) as u32).min().unwrap())
649                    .sum();
650
651                // update the match if we found a better one
652                if total_error < chosen_error {
653                    chosen_colors = colors;
654                    chosen_error = total_error;
655
656                    // if we've got a perfect or at most 1 LSB off match, we're done
657                    if total_error < 4 {
658                        break 'search;
659                    }
660                }
661            }
662        }
663    }
664
665    // calculate the final indices
666    // note that targets is already in reverse pixel order, to make the index computation easy.
667    let mut chosen_indices = 0u32;
668    for t in &targets {
669        let (idx, _) = chosen_colors
670            .iter()
671            .enumerate()
672            .min_by_key(|&(_, c)| diff(*c, *t))
673            .unwrap();
674        chosen_indices = (chosen_indices << 2) | idx as u32;
675    }
676
677    // encode the colors
678    let mut color0 = enc565_encode(chosen_colors[0]);
679    let mut color1 = enc565_encode(chosen_colors[1]);
680
681    // determine encoding. Note that color0 == color1 is impossible at this point
682    if is_dxt1 {
683        if color0 > color1 {
684            if chosen_use_0 {
685                swap(&mut color0, &mut color1);
686                // Indexes are packed 2 bits wide, swap index 0/1 but preserve 2/3.
687                let filter = (chosen_indices & 0xAAAA_AAAA) >> 1;
688                chosen_indices ^= filter ^ 0x5555_5555;
689            }
690        } else if !chosen_use_0 {
691            swap(&mut color0, &mut color1);
692            // Indexes are packed 2 bits wide, swap index 0/1 and 2/3.
693            chosen_indices ^= 0x5555_5555;
694        }
695    }
696
697    // encode everything.
698    dest[0] = color0 as u8;
699    dest[1] = (color0 >> 8) as u8;
700    dest[2] = color1 as u8;
701    dest[3] = (color1 >> 8) as u8;
702    for i in 0..4 {
703        dest[i + 4] = (chosen_indices >> (i * 8)) as u8;
704    }
705}
706
707/// Encodes a buffer of 16 alpha bytes into a dxt5 alpha index table,
708/// where the alpha table they are indexed against is created by
709/// calling alpha_table_dxt5(alpha0, alpha1)
710/// returns the resulting error and alpha table
711fn encode_dxt5_alpha(alpha0: u8, alpha1: u8, alphas: &[u8; 16]) -> (i32, u64) {
712    // create a table for the given alpha ranges
713    let table = alpha_table_dxt5(alpha0, alpha1);
714    let mut indices = 0u64;
715    let mut total_error = 0i32;
716
717    // least error brute force search
718    for (i, &a) in alphas.iter().enumerate() {
719        let (index, error) = table
720            .iter()
721            .enumerate()
722            .map(|(i, &e)| (i, square(i32::from(e) - i32::from(a))))
723            .min_by_key(|&(_, e)| e)
724            .unwrap();
725        total_error += error;
726        indices |= (index as u64) << (i * 3);
727    }
728
729    (total_error, indices)
730}
731
732/// Encodes a RGBAx16 sequence of bytes to a 16 bytes DXT5 block
733fn encode_dxt5_block(source: &[u8], dest: &mut [u8]) {
734    assert!(source.len() == 64 && dest.len() == 16);
735
736    // perform dxt color encoding
737    encode_dxt_colors(source, &mut dest[8..16], false);
738
739    // copy out the alpha bytes
740    let mut alphas = [0; 16];
741    for i in 0..16 {
742        alphas[i] = source[i * 4 + 3];
743    }
744
745    // try both alpha compression methods, see which has the least error.
746    let alpha07 = alphas.iter().cloned().min().unwrap();
747    let alpha17 = alphas.iter().cloned().max().unwrap();
748    let (error7, indices7) = encode_dxt5_alpha(alpha07, alpha17, &alphas);
749
750    // if all alphas are 0 or 255 it doesn't particularly matter what we do here.
751    let alpha05 = alphas
752        .iter()
753        .cloned()
754        .filter(|&i| i != 255)
755        .max()
756        .unwrap_or(255);
757    let alpha15 = alphas
758        .iter()
759        .cloned()
760        .filter(|&i| i != 0)
761        .min()
762        .unwrap_or(0);
763    let (error5, indices5) = encode_dxt5_alpha(alpha05, alpha15, &alphas);
764
765    // pick the best one, encode the min/max values
766    let mut alpha_table = if error5 < error7 {
767        dest[0] = alpha05;
768        dest[1] = alpha15;
769        indices5
770    } else {
771        dest[0] = alpha07;
772        dest[1] = alpha17;
773        indices7
774    };
775
776    // encode the alphas
777    for byte in dest[2..8].iter_mut() {
778        *byte = alpha_table as u8;
779        alpha_table >>= 8;
780    }
781}
782
783/// Encodes a RGBAx16 sequence of bytes into a 16 bytes DXT3 block
784fn encode_dxt3_block(source: &[u8], dest: &mut [u8]) {
785    assert!(source.len() == 64 && dest.len() == 16);
786
787    // perform dxt color encoding
788    encode_dxt_colors(source, &mut dest[8..16], false);
789
790    // DXT3 alpha compression is very simple, just round towards the nearest value
791
792    // index the alpha values into the 64bit alpha table
793    let mut alpha_table = 0u64;
794    for i in 0..16 {
795        let alpha = u64::from(source[i * 4 + 3]);
796        let alpha = (alpha + 0x8) / 0x11;
797        alpha_table |= alpha << (i * 4);
798    }
799
800    // encode the alpha values
801    for byte in &mut dest[0..8] {
802        *byte = alpha_table as u8;
803        alpha_table >>= 8;
804    }
805}
806
807/// Encodes a RGBx16 sequence of bytes into a 8 bytes DXT1 block
808fn encode_dxt1_block(source: &[u8], dest: &mut [u8]) {
809    assert!(source.len() == 48 && dest.len() == 8);
810
811    // perform dxt color encoding
812    encode_dxt_colors(source, dest, true);
813}
814
815/// Decode a row of DXT1 data to four rows of RGBA data.
816/// source.len() should be a multiple of 8, otherwise this panics.
817fn encode_dxt1_row(source: &[u8]) -> Vec<u8> {
818    assert!(source.len() % 48 == 0);
819    let block_count = source.len() / 48;
820
821    let mut dest = vec![0u8; block_count * 8];
822    // contains the 16 decoded pixels per block
823    let mut decoded_block = [0u8; 48];
824
825    for (x, encoded_block) in dest.chunks_mut(8).enumerate() {
826        // copy the values from the decoded block to linewise RGB layout
827        for line in 0..4 {
828            let offset = (block_count * line + x) * 12;
829            decoded_block[line * 12..(line + 1) * 12].copy_from_slice(&source[offset..offset + 12]);
830        }
831
832        encode_dxt1_block(&decoded_block, encoded_block);
833    }
834    dest
835}
836
837/// Decode a row of DXT3 data to four rows of RGBA data.
838/// source.len() should be a multiple of 16, otherwise this panics.
839fn encode_dxt3_row(source: &[u8]) -> Vec<u8> {
840    assert!(source.len() % 64 == 0);
841    let block_count = source.len() / 64;
842
843    let mut dest = vec![0u8; block_count * 16];
844    // contains the 16 decoded pixels per block
845    let mut decoded_block = [0u8; 64];
846
847    for (x, encoded_block) in dest.chunks_mut(16).enumerate() {
848        // copy the values from the decoded block to linewise RGB layout
849        for line in 0..4 {
850            let offset = (block_count * line + x) * 16;
851            decoded_block[line * 16..(line + 1) * 16].copy_from_slice(&source[offset..offset + 16]);
852        }
853
854        encode_dxt3_block(&decoded_block, encoded_block);
855    }
856    dest
857}
858
859/// Decode a row of DXT5 data to four rows of RGBA data.
860/// source.len() should be a multiple of 16, otherwise this panics.
861fn encode_dxt5_row(source: &[u8]) -> Vec<u8> {
862    assert!(source.len() % 64 == 0);
863    let block_count = source.len() / 64;
864
865    let mut dest = vec![0u8; block_count * 16];
866    // contains the 16 decoded pixels per block
867    let mut decoded_block = [0u8; 64];
868
869    for (x, encoded_block) in dest.chunks_mut(16).enumerate() {
870        // copy the values from the decoded block to linewise RGB layout
871        for line in 0..4 {
872            let offset = (block_count * line + x) * 16;
873            decoded_block[line * 16..(line + 1) * 16].copy_from_slice(&source[offset..offset + 16]);
874        }
875
876        encode_dxt5_block(&decoded_block, encoded_block);
877    }
878    dest
879}
image/codecs/dxt.rs

image/codecs/
dxt.rs
