2 * Texture block decompression
3 * Copyright (C) 2009 Benjamin Dobell, Glass Echidna
4 * Copyright (C) 2012 Matthäus G. "Anteru" Chajdas (http://anteru.net)
5 * Copyright (C) 2015 Vittorio Giovara <vittorio.giovara@gmail.com>
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included
14 * in all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "libavutil/attributes.h"
29 #include "libavutil/common.h"
30 #include "libavutil/intreadwrite.h"
32 #include "texturedsp.h"
34 #define RGBA(r, g, b, a) ((r) | ((g) << 8) | ((b) << 16) | ((a) << 24))
36 static av_always_inline void extract_color(uint32_t colors[4],
42 uint8_t r0, g0, b0, r1, g1, b1;
43 uint8_t a = dxtn ? 0 : 255;
45 tmp = (color0 >> 11) * 255 + 16;
46 r0 = (uint8_t) ((tmp / 32 + tmp) / 32);
47 tmp = ((color0 & 0x07E0) >> 5) * 255 + 32;
48 g0 = (uint8_t) ((tmp / 64 + tmp) / 64);
49 tmp = (color0 & 0x001F) * 255 + 16;
50 b0 = (uint8_t) ((tmp / 32 + tmp) / 32);
52 tmp = (color1 >> 11) * 255 + 16;
53 r1 = (uint8_t) ((tmp / 32 + tmp) / 32);
54 tmp = ((color1 & 0x07E0) >> 5) * 255 + 32;
55 g1 = (uint8_t) ((tmp / 64 + tmp) / 64);
56 tmp = (color1 & 0x001F) * 255 + 16;
57 b1 = (uint8_t) ((tmp / 32 + tmp) / 32);
59 if (dxtn || color0 > color1) {
60 colors[0] = RGBA(r0, g0, b0, a);
61 colors[1] = RGBA(r1, g1, b1, a);
62 colors[2] = RGBA((2 * r0 + r1) / 3,
66 colors[3] = RGBA((2 * r1 + r0) / 3,
71 colors[0] = RGBA(r0, g0, b0, a);
72 colors[1] = RGBA(r1, g1, b1, a);
73 colors[2] = RGBA((r0 + r1) / 2,
77 colors[3] = RGBA(0, 0, 0, alpha);
81 static inline void dxt1_block_internal(uint8_t *dst, ptrdiff_t stride,
82 const uint8_t *block, uint8_t alpha)
86 uint16_t color0 = AV_RL16(block + 0);
87 uint16_t color1 = AV_RL16(block + 2);
88 uint32_t code = AV_RL32(block + 4);
90 extract_color(colors, color0, color1, 0, alpha);
92 for (y = 0; y < 4; y++) {
93 for (x = 0; x < 4; x++) {
94 uint32_t pixel = colors[code & 3];
96 AV_WL32(dst + x * 4, pixel);
103 * Decompress one block of a DXT1 texture and store the resulting
104 * RGBA pixels in 'dst'. Alpha component is fully opaque.
106 * @param dst output buffer.
107 * @param stride scanline in bytes.
108 * @param block block to decompress.
109 * @return how much texture data has been consumed.
111 static int dxt1_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
113 dxt1_block_internal(dst, stride, block, 255);
119 * Decompress one block of a DXT1 with 1-bit alpha texture and store
120 * the resulting RGBA pixels in 'dst'. Alpha is either fully opaque or
123 * @param dst output buffer.
124 * @param stride scanline in bytes.
125 * @param block block to decompress.
126 * @return how much texture data has been consumed.
128 static int dxt1a_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
130 dxt1_block_internal(dst, stride, block, 0);
135 static inline void dxt3_block_internal(uint8_t *dst, ptrdiff_t stride,
136 const uint8_t *block)
140 uint16_t color0 = AV_RL16(block + 8);
141 uint16_t color1 = AV_RL16(block + 10);
142 uint32_t code = AV_RL32(block + 12);
144 extract_color(colors, color0, color1, 1, 0);
146 for (y = 0; y < 4; y++) {
147 const uint16_t alpha_code = AV_RL16(block + 2 * y);
148 uint8_t alpha_values[4];
150 alpha_values[0] = ((alpha_code >> 0) & 0x0F) * 17;
151 alpha_values[1] = ((alpha_code >> 4) & 0x0F) * 17;
152 alpha_values[2] = ((alpha_code >> 8) & 0x0F) * 17;
153 alpha_values[3] = ((alpha_code >> 12) & 0x0F) * 17;
155 for (x = 0; x < 4; x++) {
156 uint8_t alpha = alpha_values[x];
157 uint32_t pixel = colors[code & 3] | (alpha << 24);
160 AV_WL32(dst + x * 4, pixel);
166 /** Convert a premultiplied alpha pixel to a straigth alpha pixel. */
167 static av_always_inline void premult2straight(uint8_t *src)
172 int a = src[3]; /* unchanged */
174 src[0] = (uint8_t) r * a / 255;
175 src[1] = (uint8_t) g * a / 255;
176 src[2] = (uint8_t) b * a / 255;
180 * Decompress one block of a DXT2 texture and store the resulting
181 * RGBA pixels in 'dst'.
183 * @param dst output buffer.
184 * @param stride scanline in bytes.
185 * @param block block to decompress.
186 * @return how much texture data has been consumed.
188 static int dxt2_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
192 dxt3_block_internal(dst, stride, block);
194 /* This format is DXT3, but returns premultiplied alpha. It needs to be
195 * converted because it's what lavc outputs (and swscale expects). */
196 for (y = 0; y < 4; y++)
197 for (x = 0; x < 4; x++)
198 premult2straight(dst + x * 4 + y * stride);
204 * Decompress one block of a DXT3 texture and store the resulting
205 * RGBA pixels in 'dst'.
207 * @param dst output buffer.
208 * @param stride scanline in bytes.
209 * @param block block to decompress.
210 * @return how much texture data has been consumed.
212 static int dxt3_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
214 dxt3_block_internal(dst, stride, block);
220 * Decompress a BC 16x3 index block stored as
227 * | h | g | f | e | d | c | b | a | // Entry
228 * |765 432 107 654 321 076 543 210| // Bit
229 * |0000000000111111111112222222222| // Byte
231 * into 16 8-bit indices.
233 static void decompress_indices(uint8_t *dst, const uint8_t *src)
237 for (block = 0; block < 2; block++) {
238 int tmp = AV_RL24(src);
240 /* Unpack 8x3 bit from last 3 byte block */
241 for (i = 0; i < 8; i++)
242 dst[i] = (tmp >> (i * 3)) & 0x7;
249 static inline void dxt5_block_internal(uint8_t *dst, ptrdiff_t stride,
250 const uint8_t *block)
254 uint8_t alpha_indices[16];
255 uint16_t color0 = AV_RL16(block + 8);
256 uint16_t color1 = AV_RL16(block + 10);
257 uint32_t code = AV_RL32(block + 12);
258 uint8_t alpha0 = *(block);
259 uint8_t alpha1 = *(block + 1);
261 decompress_indices(alpha_indices, block + 2);
263 extract_color(colors, color0, color1, 1, 0);
265 for (y = 0; y < 4; y++) {
266 for (x = 0; x < 4; x++) {
267 int alpha_code = alpha_indices[x + y * 4];
271 if (alpha_code == 0) {
273 } else if (alpha_code == 1) {
276 if (alpha0 > alpha1) {
277 alpha = (uint8_t) (((8 - alpha_code) * alpha0 +
278 (alpha_code - 1) * alpha1) / 7);
280 if (alpha_code == 6) {
282 } else if (alpha_code == 7) {
285 alpha = (uint8_t) (((6 - alpha_code) * alpha0 +
286 (alpha_code - 1) * alpha1) / 5);
290 pixel = colors[code & 3] | (alpha << 24);
292 AV_WL32(dst + x * 4, pixel);
299 * Decompress one block of a DXT4 texture and store the resulting
300 * RGBA pixels in 'dst'.
302 * @param dst output buffer.
303 * @param stride scanline in bytes.
304 * @param block block to decompress.
305 * @return how much texture data has been consumed.
307 static int dxt4_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
311 dxt5_block_internal(dst, stride, block);
313 /* This format is DXT5, but returns premultiplied alpha. It needs to be
314 * converted because it's what lavc outputs (and swscale expects). */
315 for (y = 0; y < 4; y++)
316 for (x = 0; x < 4; x++)
317 premult2straight(dst + x * 4 + y * stride);
323 * Decompress one block of a DXT5 texture and store the resulting
324 * RGBA pixels in 'dst'.
326 * @param dst output buffer.
327 * @param stride scanline in bytes.
328 * @param block block to decompress.
329 * @return how much texture data has been consumed.
331 static int dxt5_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
333 dxt5_block_internal(dst, stride, block);
339 * Convert a YCoCg buffer to RGBA.
341 * @param src input buffer.
342 * @param scaled variant with scaled chroma components and opaque alpha.
344 static av_always_inline void ycocg2rgba(uint8_t *src, int scaled)
351 int s = scaled ? (b >> 3) + 1 : 1;
353 int co = (r - 128) / s;
354 int cg = (g - 128) / s;
356 src[0] = av_clip_uint8(y + co - cg);
357 src[1] = av_clip_uint8(y + cg);
358 src[2] = av_clip_uint8(y - co - cg);
359 src[3] = scaled ? 255 : b;
363 * Decompress one block of a DXT5 texture with classic YCoCg and store
364 * the resulting RGBA pixels in 'dst'. Alpha component is fully opaque.
366 * @param dst output buffer.
367 * @param stride scanline in bytes.
368 * @param block block to decompress.
369 * @return how much texture data has been consumed.
371 static int dxt5y_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
375 /* This format is basically DXT5, with luma stored in alpha.
376 * Run a normal decompress and then reorder the components. */
377 dxt5_block_internal(dst, stride, block);
379 for (y = 0; y < 4; y++)
380 for (x = 0; x < 4; x++)
381 ycocg2rgba(dst + x * 4 + y * stride, 0);
387 * Decompress one block of a DXT5 texture with scaled YCoCg and store
388 * the resulting RGBA pixels in 'dst'. Alpha component is fully opaque.
390 * @param dst output buffer.
391 * @param stride scanline in bytes.
392 * @param block block to decompress.
393 * @return how much texture data has been consumed.
395 static int dxt5ys_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
399 /* This format is basically DXT5, with luma stored in alpha.
400 * Run a normal decompress and then reorder the components. */
401 dxt5_block_internal(dst, stride, block);
403 for (y = 0; y < 4; y++)
404 for (x = 0; x < 4; x++)
405 ycocg2rgba(dst + x * 4 + y * stride, 1);
410 static inline void rgtc_block_internal(uint8_t *dst, ptrdiff_t stride,
411 const uint8_t *block,
412 const int *color_tab)
417 decompress_indices(indices, block + 2);
419 /* Only one or two channels are stored at most, since it only used to
420 * compress specular (black and white) or normal (red and green) maps.
421 * Although the standard says to zero out unused components, many
422 * implementations fill all of them with the same value. */
423 for (y = 0; y < 4; y++) {
424 for (x = 0; x < 4; x++) {
425 int i = indices[x + y * 4];
426 /* Interval expansion from [-1 1] or [0 1] to [0 255]. */
427 int c = color_tab[i];
428 uint32_t pixel = RGBA(c, c, c, 255U);
429 AV_WL32(dst + x * 4 + y * stride, pixel);
434 static inline void rgtc1_block_internal(uint8_t *dst, ptrdiff_t stride,
435 const uint8_t *block, int sign)
441 /* signed data is in [-128 127] so just offset it to unsigned
442 * and it can be treated exactly the same */
443 r0 = ((int8_t) block[0]) + 128;
444 r1 = ((int8_t) block[1]) + 128;
454 /* 6 interpolated color values */
455 color_table[2] = (6 * r0 + 1 * r1) / 7; // bit code 010
456 color_table[3] = (5 * r0 + 2 * r1) / 7; // bit code 011
457 color_table[4] = (4 * r0 + 3 * r1) / 7; // bit code 100
458 color_table[5] = (3 * r0 + 4 * r1) / 7; // bit code 101
459 color_table[6] = (2 * r0 + 5 * r1) / 7; // bit code 110
460 color_table[7] = (1 * r0 + 6 * r1) / 7; // bit code 111
462 /* 4 interpolated color values */
463 color_table[2] = (4 * r0 + 1 * r1) / 5; // bit code 010
464 color_table[3] = (3 * r0 + 2 * r1) / 5; // bit code 011
465 color_table[4] = (2 * r0 + 3 * r1) / 5; // bit code 100
466 color_table[5] = (1 * r0 + 4 * r1) / 5; // bit code 101
467 color_table[6] = 0; /* min range */ // bit code 110
468 color_table[7] = 255; /* max range */ // bit code 111
471 rgtc_block_internal(dst, stride, block, color_table);
475 * Decompress one block of a RGRC1 texture with signed components
476 * and store the resulting RGBA pixels in 'dst'.
478 * @param dst output buffer.
479 * @param stride scanline in bytes.
480 * @param block block to decompress.
481 * @return how much texture data has been consumed.
483 static int rgtc1s_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
485 rgtc1_block_internal(dst, stride, block, 1);
491 * Decompress one block of a RGRC1 texture with unsigned components
492 * and store the resulting RGBA pixels in 'dst'.
494 * @param dst output buffer.
495 * @param stride scanline in bytes.
496 * @param block block to decompress.
497 * @return how much texture data has been consumed.
499 static int rgtc1u_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
501 rgtc1_block_internal(dst, stride, block, 0);
506 static inline void rgtc2_block_internal(uint8_t *dst, ptrdiff_t stride,
507 const uint8_t *block, int sign)
509 /* 4x4 block containing 4 component pixels. */
510 uint8_t c0[4 * 4 * 4];
511 uint8_t c1[4 * 4 * 4];
514 /* Decompress the two channels separately and interleave them afterwards. */
515 rgtc1_block_internal(c0, 16, block, sign);
516 rgtc1_block_internal(c1, 16, block + 8, sign);
518 /* B is rebuilt exactly like a normal map. */
519 for (y = 0; y < 4; y++) {
520 for (x = 0; x < 4; x++) {
521 uint8_t *p = dst + x * 4 + y * stride;
522 int r = c0[x * 4 + y * 16];
523 int g = c1[x * 4 + y * 16];
526 int d = (255 * 255 - r * r - g * g) / 2;
539 * Decompress one block of a RGRC2 texture with signed components
540 * and store the resulting RGBA pixels in 'dst'. Alpha is fully opaque.
542 * @param dst output buffer.
543 * @param stride scanline in bytes.
544 * @param block block to decompress.
545 * @return how much texture data has been consumed.
547 static int rgtc2s_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
549 rgtc2_block_internal(dst, stride, block, 1);
555 * Decompress one block of a RGRC2 texture with unsigned components
556 * and store the resulting RGBA pixels in 'dst'. Alpha is fully opaque.
558 * @param dst output buffer.
559 * @param stride scanline in bytes.
560 * @param block block to decompress.
561 * @return how much texture data has been consumed.
563 static int rgtc2u_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
565 rgtc2_block_internal(dst, stride, block, 0);
571 * Decompress one block of a 3Dc texture with unsigned components
572 * and store the resulting RGBA pixels in 'dst'. Alpha is fully opaque.
574 * @param dst output buffer.
575 * @param stride scanline in bytes.
576 * @param block block to decompress.
577 * @return how much texture data has been consumed.
579 static int dxn3dc_block(uint8_t *dst, ptrdiff_t stride, const uint8_t *block)
582 rgtc2_block_internal(dst, stride, block, 0);
584 /* This is the 3Dc variant of RGTC2, with swapped R and G. */
585 for (y = 0; y < 4; y++) {
586 for (x = 0; x < 4; x++) {
587 uint8_t *p = dst + x * 4 + y * stride;
588 FFSWAP(uint8_t, p[0], p[1]);
595 av_cold void ff_texturedsp_init(TextureDSPContext *c)
597 c->dxt1_block = dxt1_block;
598 c->dxt1a_block = dxt1a_block;
599 c->dxt2_block = dxt2_block;
600 c->dxt3_block = dxt3_block;
601 c->dxt4_block = dxt4_block;
602 c->dxt5_block = dxt5_block;
603 c->dxt5y_block = dxt5y_block;
604 c->dxt5ys_block = dxt5ys_block;
605 c->rgtc1s_block = rgtc1s_block;
606 c->rgtc1u_block = rgtc1u_block;
607 c->rgtc2s_block = rgtc2s_block;
608 c->rgtc2u_block = rgtc2u_block;
609 c->dxn3dc_block = dxn3dc_block;