OLD | NEW |
| (Empty) |
1 // Copyright 2015 The Chromium Authors. All rights reserved. | |
2 // Use of this source code is governed by a BSD-style license that can be | |
3 // found in the LICENSE file. | |
4 | |
5 // See the following specification for details on the ETC1 format: | |
6 // https://www.khronos.org/registry/gles/extensions/OES/OES_compressed_ETC1_RGB8
_texture.txt | |
7 | |
8 #include "cc/resources/texture_compressor_etc1.h" | |
9 | |
10 #include <string.h> | |
11 #include <limits> | |
12 | |
13 #include "base/logging.h" | |
14 | |
15 // Defining the following macro will cause the error metric function to weigh | |
16 // each color channel differently depending on how the human eye can perceive | |
17 // them. This can give a slight improvement in image quality at the cost of a | |
18 // performance hit. | |
19 // #define USE_PERCEIVED_ERROR_METRIC | |
20 | |
21 namespace { | |
22 | |
23 template <typename T> | |
24 inline T clamp(T val, T min, T max) { | |
25 return val < min ? min : (val > max ? max : val); | |
26 } | |
27 | |
28 inline uint8_t round_to_5_bits(float val) { | |
29 return clamp<uint8_t>(val * 31.0f / 255.0f + 0.5f, 0, 31); | |
30 } | |
31 | |
32 inline uint8_t round_to_4_bits(float val) { | |
33 return clamp<uint8_t>(val * 15.0f / 255.0f + 0.5f, 0, 15); | |
34 } | |
35 | |
36 union Color { | |
37 struct BgraColorType { | |
38 uint8_t b; | |
39 uint8_t g; | |
40 uint8_t r; | |
41 uint8_t a; | |
42 } channels; | |
43 uint8_t components[4]; | |
44 uint32_t bits; | |
45 }; | |
46 | |
47 /* | |
48 * Codeword tables. | |
49 * See: Table 3.17.2 | |
50 */ | |
51 static const int16_t g_codeword_tables[8][4] = {{-8, -2, 2, 8}, | |
52 {-17, -5, 5, 17}, | |
53 {-29, -9, 9, 29}, | |
54 {-42, -13, 13, 42}, | |
55 {-60, -18, 18, 60}, | |
56 {-80, -24, 24, 80}, | |
57 {-106, -33, 33, 106}, | |
58 {-183, -47, 47, 183}}; | |
59 | |
60 /* | |
61 * Maps modifier indices to pixel index values. | |
62 * See: Table 3.17.3 | |
63 */ | |
64 static const uint8_t g_mod_to_pix[4] = {3, 2, 0, 1}; | |
65 | |
66 /* | |
67 * The ETC1 specification index texels as follows: | |
68 * | |
69 * [a][e][i][m] [ 0][ 4][ 8][12] | |
70 * [b][f][j][n] <-> [ 1][ 5][ 9][13] | |
71 * [c][g][k][o] [ 2][ 6][10][14] | |
72 * [d][h][l][p] [ 3][ 7][11][15] | |
73 * | |
74 * However, when extracting sub blocks from BGRA data the natural array | |
75 * indexing order ends up different: | |
76 * | |
77 * vertical0: [a][e][b][f] horizontal0: [a][e][i][m] | |
78 * [c][g][d][h] [b][f][j][n] | |
79 * vertical1: [i][m][j][n] horizontal1: [c][g][k][o] | |
80 * [k][o][l][p] [d][h][l][p] | |
81 * | |
82 * In order to translate from the natural array indices in a sub block to the | |
83 * indices (number) used by specification and hardware we use this table. | |
84 */ | |
85 static const uint8_t g_idx_to_num[4][8] = { | |
86 {0, 4, 1, 5, 2, 6, 3, 7}, // Vertical block 0. | |
87 {8, 12, 9, 13, 10, 14, 11, 15}, // Vertical block 1. | |
88 {0, 4, 8, 12, 1, 5, 9, 13}, // Horizontal block 0. | |
89 {2, 6, 10, 14, 3, 7, 11, 15} // Horizontal block 1. | |
90 }; | |
91 | |
92 inline void WriteColors444(uint8_t* block, | |
93 const Color& color0, | |
94 const Color& color1) { | |
95 block[0] = (color0.channels.r & 0xf0) | (color1.channels.r >> 4); | |
96 block[1] = (color0.channels.g & 0xf0) | (color1.channels.g >> 4); | |
97 block[2] = (color0.channels.b & 0xf0) | (color1.channels.b >> 4); | |
98 } | |
99 | |
100 inline void WriteColors555(uint8_t* block, | |
101 const Color& color0, | |
102 const Color& color1) { | |
103 // Table for conversion to 3-bit two complement format. | |
104 static const uint8_t two_compl_trans_table[8] = { | |
105 4, // -4 (100b) | |
106 5, // -3 (101b) | |
107 6, // -2 (110b) | |
108 7, // -1 (111b) | |
109 0, // 0 (000b) | |
110 1, // 1 (001b) | |
111 2, // 2 (010b) | |
112 3, // 3 (011b) | |
113 }; | |
114 | |
115 int16_t delta_r = | |
116 static_cast<int16_t>(color1.channels.r >> 3) - (color0.channels.r >> 3); | |
117 int16_t delta_g = | |
118 static_cast<int16_t>(color1.channels.g >> 3) - (color0.channels.g >> 3); | |
119 int16_t delta_b = | |
120 static_cast<int16_t>(color1.channels.b >> 3) - (color0.channels.b >> 3); | |
121 DCHECK(delta_r >= -4 && delta_r <= 3); | |
122 DCHECK(delta_g >= -4 && delta_g <= 3); | |
123 DCHECK(delta_b >= -4 && delta_b <= 3); | |
124 | |
125 block[0] = (color0.channels.r & 0xf8) | two_compl_trans_table[delta_r + 4]; | |
126 block[1] = (color0.channels.g & 0xf8) | two_compl_trans_table[delta_g + 4]; | |
127 block[2] = (color0.channels.b & 0xf8) | two_compl_trans_table[delta_b + 4]; | |
128 } | |
129 | |
130 inline void WriteCodewordTable(uint8_t* block, | |
131 uint8_t sub_block_id, | |
132 uint8_t table) { | |
133 DCHECK_LT(sub_block_id, 2); | |
134 DCHECK_LT(table, 8); | |
135 | |
136 uint8_t shift = (2 + (3 - sub_block_id * 3)); | |
137 block[3] &= ~(0x07 << shift); | |
138 block[3] |= table << shift; | |
139 } | |
140 | |
141 inline void WritePixelData(uint8_t* block, uint32_t pixel_data) { | |
142 block[4] |= pixel_data >> 24; | |
143 block[5] |= (pixel_data >> 16) & 0xff; | |
144 block[6] |= (pixel_data >> 8) & 0xff; | |
145 block[7] |= pixel_data & 0xff; | |
146 } | |
147 | |
148 inline void WriteFlip(uint8_t* block, bool flip) { | |
149 block[3] &= ~0x01; | |
150 block[3] |= static_cast<uint8_t>(flip); | |
151 } | |
152 | |
153 inline void WriteDiff(uint8_t* block, bool diff) { | |
154 block[3] &= ~0x02; | |
155 block[3] |= static_cast<uint8_t>(diff) << 1; | |
156 } | |
157 | |
158 /** | |
159 * Compress and rounds BGR888 into BGR444. The resulting BGR444 color is | |
160 * expanded to BGR888 as it would be in hardware after decompression. The | |
161 * actual 444-bit data is available in the four most significant bits of each | |
162 * channel. | |
163 */ | |
164 inline Color MakeColor444(const float* bgr) { | |
165 uint8_t b4 = round_to_4_bits(bgr[0]); | |
166 uint8_t g4 = round_to_4_bits(bgr[1]); | |
167 uint8_t r4 = round_to_4_bits(bgr[2]); | |
168 Color bgr444; | |
169 bgr444.channels.b = (b4 << 4) | b4; | |
170 bgr444.channels.g = (g4 << 4) | g4; | |
171 bgr444.channels.r = (r4 << 4) | r4; | |
172 return bgr444; | |
173 } | |
174 | |
175 /** | |
176 * Compress and rounds BGR888 into BGR555. The resulting BGR555 color is | |
177 * expanded to BGR888 as it would be in hardware after decompression. The | |
178 * actual 555-bit data is available in the five most significant bits of each | |
179 * channel. | |
180 */ | |
181 inline Color MakeColor555(const float* bgr) { | |
182 uint8_t b5 = round_to_5_bits(bgr[0]); | |
183 uint8_t g5 = round_to_5_bits(bgr[1]); | |
184 uint8_t r5 = round_to_5_bits(bgr[2]); | |
185 Color bgr555; | |
186 bgr555.channels.b = (b5 << 3) | (b5 >> 2); | |
187 bgr555.channels.g = (g5 << 3) | (g5 >> 2); | |
188 bgr555.channels.r = (r5 << 3) | (r5 >> 2); | |
189 return bgr555; | |
190 } | |
191 | |
192 /** | |
193 * Constructs a color from a given base color and luminance value. | |
194 */ | |
195 inline Color MakeColor(const Color& base, int16_t lum) { | |
196 int b = static_cast<int>(base.channels.b) + lum; | |
197 int g = static_cast<int>(base.channels.g) + lum; | |
198 int r = static_cast<int>(base.channels.r) + lum; | |
199 Color color; | |
200 color.channels.b = static_cast<uint8_t>(clamp(b, 0, 255)); | |
201 color.channels.g = static_cast<uint8_t>(clamp(g, 0, 255)); | |
202 color.channels.r = static_cast<uint8_t>(clamp(r, 0, 255)); | |
203 return color; | |
204 } | |
205 | |
206 /** | |
207 * Calculates the error metric for two colors. A small error signals that the | |
208 * colors are similar to each other, a large error the signals the opposite. | |
209 */ | |
210 inline uint32_t GetColorError(const Color& u, const Color& v) { | |
211 #ifdef USE_PERCEIVED_ERROR_METRIC | |
212 float delta_b = static_cast<float>(u.channels.b) - v.channels.b; | |
213 float delta_g = static_cast<float>(u.channels.g) - v.channels.g; | |
214 float delta_r = static_cast<float>(u.channels.r) - v.channels.r; | |
215 return static_cast<uint32_t>(0.299f * delta_b * delta_b + | |
216 0.587f * delta_g * delta_g + | |
217 0.114f * delta_r * delta_r); | |
218 #else | |
219 int delta_b = static_cast<int>(u.channels.b) - v.channels.b; | |
220 int delta_g = static_cast<int>(u.channels.g) - v.channels.g; | |
221 int delta_r = static_cast<int>(u.channels.r) - v.channels.r; | |
222 return delta_b * delta_b + delta_g * delta_g + delta_r * delta_r; | |
223 #endif | |
224 } | |
225 | |
226 void GetAverageColor(const Color* src, float* avg_color) { | |
227 uint32_t sum_b = 0, sum_g = 0, sum_r = 0; | |
228 | |
229 for (unsigned int i = 0; i < 8; ++i) { | |
230 sum_b += src[i].channels.b; | |
231 sum_g += src[i].channels.g; | |
232 sum_r += src[i].channels.r; | |
233 } | |
234 | |
235 const float kInv8 = 1.0f / 8.0f; | |
236 avg_color[0] = static_cast<float>(sum_b) * kInv8; | |
237 avg_color[1] = static_cast<float>(sum_g) * kInv8; | |
238 avg_color[2] = static_cast<float>(sum_r) * kInv8; | |
239 } | |
240 | |
241 void ComputeLuminance(uint8_t* block, | |
242 const Color* src, | |
243 const Color& base, | |
244 int sub_block_id, | |
245 const uint8_t* idx_to_num_tab) { | |
246 uint32_t best_tbl_err = std::numeric_limits<uint32_t>::max(); | |
247 uint8_t best_tbl_idx = 0; | |
248 uint8_t best_mod_idx[8][8]; // [table][texel] | |
249 | |
250 // Try all codeword tables to find the one giving the best results for this | |
251 // block. | |
252 for (unsigned int tbl_idx = 0; tbl_idx < 8; ++tbl_idx) { | |
253 // Pre-compute all the candidate colors; combinations of the base color and | |
254 // all available luminance values. | |
255 Color candidate_color[4]; // [modifier] | |
256 for (unsigned int mod_idx = 0; mod_idx < 4; ++mod_idx) { | |
257 int16_t lum = g_codeword_tables[tbl_idx][mod_idx]; | |
258 candidate_color[mod_idx] = MakeColor(base, lum); | |
259 } | |
260 | |
261 uint32_t tbl_err = 0; | |
262 | |
263 for (unsigned int i = 0; i < 8; ++i) { | |
264 // Try all modifiers in the current table to find which one gives the | |
265 // smallest error. | |
266 uint32_t best_mod_err = std::numeric_limits<uint32_t>::max(); | |
267 for (unsigned int mod_idx = 0; mod_idx < 4; ++mod_idx) { | |
268 const Color& color = candidate_color[mod_idx]; | |
269 | |
270 uint32_t mod_err = GetColorError(src[i], color); | |
271 if (mod_err < best_mod_err) { | |
272 best_mod_idx[tbl_idx][i] = mod_idx; | |
273 best_mod_err = mod_err; | |
274 | |
275 if (mod_err == 0) | |
276 break; // We cannot do any better than this. | |
277 } | |
278 } | |
279 | |
280 tbl_err += best_mod_err; | |
281 if (tbl_err > best_tbl_err) | |
282 break; // We're already doing worse than the best table so skip. | |
283 } | |
284 | |
285 if (tbl_err < best_tbl_err) { | |
286 best_tbl_err = tbl_err; | |
287 best_tbl_idx = tbl_idx; | |
288 | |
289 if (tbl_err == 0) | |
290 break; // We cannot do any better than this. | |
291 } | |
292 } | |
293 | |
294 WriteCodewordTable(block, sub_block_id, best_tbl_idx); | |
295 | |
296 uint32_t pix_data = 0; | |
297 | |
298 for (unsigned int i = 0; i < 8; ++i) { | |
299 uint8_t mod_idx = best_mod_idx[best_tbl_idx][i]; | |
300 uint8_t pix_idx = g_mod_to_pix[mod_idx]; | |
301 | |
302 uint32_t lsb = pix_idx & 0x1; | |
303 uint32_t msb = pix_idx >> 1; | |
304 | |
305 // Obtain the texel number as specified in the standard. | |
306 int texel_num = idx_to_num_tab[i]; | |
307 pix_data |= msb << (texel_num + 16); | |
308 pix_data |= lsb << (texel_num); | |
309 } | |
310 | |
311 WritePixelData(block, pix_data); | |
312 } | |
313 | |
314 /** | |
315 * Tries to compress the block under the assumption that it's a single color | |
316 * block. If it's not the function will bail out without writing anything to | |
317 * the destination buffer. | |
318 */ | |
319 bool TryCompressSolidBlock(uint8_t* dst, const Color* src) { | |
320 for (unsigned int i = 1; i < 16; ++i) { | |
321 if (src[i].bits != src[0].bits) | |
322 return false; | |
323 } | |
324 | |
325 // Clear destination buffer so that we can "or" in the results. | |
326 memset(dst, 0, 8); | |
327 | |
328 float src_color_float[3] = {static_cast<float>(src->channels.b), | |
329 static_cast<float>(src->channels.g), | |
330 static_cast<float>(src->channels.r)}; | |
331 Color base = MakeColor555(src_color_float); | |
332 | |
333 WriteDiff(dst, true); | |
334 WriteFlip(dst, false); | |
335 WriteColors555(dst, base, base); | |
336 | |
337 uint8_t best_tbl_idx = 0; | |
338 uint8_t best_mod_idx = 0; | |
339 uint32_t best_mod_err = std::numeric_limits<uint32_t>::max(); | |
340 | |
341 // Try all codeword tables to find the one giving the best results for this | |
342 // block. | |
343 for (unsigned int tbl_idx = 0; tbl_idx < 8; ++tbl_idx) { | |
344 // Try all modifiers in the current table to find which one gives the | |
345 // smallest error. | |
346 for (unsigned int mod_idx = 0; mod_idx < 4; ++mod_idx) { | |
347 int16_t lum = g_codeword_tables[tbl_idx][mod_idx]; | |
348 const Color& color = MakeColor(base, lum); | |
349 | |
350 uint32_t mod_err = GetColorError(*src, color); | |
351 if (mod_err < best_mod_err) { | |
352 best_tbl_idx = tbl_idx; | |
353 best_mod_idx = mod_idx; | |
354 best_mod_err = mod_err; | |
355 | |
356 if (mod_err == 0) | |
357 break; // We cannot do any better than this. | |
358 } | |
359 } | |
360 | |
361 if (best_mod_err == 0) | |
362 break; | |
363 } | |
364 | |
365 WriteCodewordTable(dst, 0, best_tbl_idx); | |
366 WriteCodewordTable(dst, 1, best_tbl_idx); | |
367 | |
368 uint8_t pix_idx = g_mod_to_pix[best_mod_idx]; | |
369 uint32_t lsb = pix_idx & 0x1; | |
370 uint32_t msb = pix_idx >> 1; | |
371 | |
372 uint32_t pix_data = 0; | |
373 for (unsigned int i = 0; i < 2; ++i) { | |
374 for (unsigned int j = 0; j < 8; ++j) { | |
375 // Obtain the texel number as specified in the standard. | |
376 int texel_num = g_idx_to_num[i][j]; | |
377 pix_data |= msb << (texel_num + 16); | |
378 pix_data |= lsb << (texel_num); | |
379 } | |
380 } | |
381 | |
382 WritePixelData(dst, pix_data); | |
383 return true; | |
384 } | |
385 | |
386 void CompressBlock(uint8_t* dst, const Color* ver_src, const Color* hor_src) { | |
387 if (TryCompressSolidBlock(dst, ver_src)) | |
388 return; | |
389 | |
390 const Color* sub_block_src[4] = {ver_src, ver_src + 8, hor_src, hor_src + 8}; | |
391 | |
392 Color sub_block_avg[4]; | |
393 bool use_differential[2] = {true, true}; | |
394 | |
395 // Compute the average color for each sub block and determine if differential | |
396 // coding can be used. | |
397 for (unsigned int i = 0, j = 1; i < 4; i += 2, j += 2) { | |
398 float avg_color_0[3]; | |
399 GetAverageColor(sub_block_src[i], avg_color_0); | |
400 Color avg_color_555_0 = MakeColor555(avg_color_0); | |
401 | |
402 float avg_color_1[3]; | |
403 GetAverageColor(sub_block_src[j], avg_color_1); | |
404 Color avg_color_555_1 = MakeColor555(avg_color_1); | |
405 | |
406 for (unsigned int light_idx = 0; light_idx < 3; ++light_idx) { | |
407 int u = avg_color_555_0.components[light_idx] >> 3; | |
408 int v = avg_color_555_1.components[light_idx] >> 3; | |
409 | |
410 int component_diff = v - u; | |
411 if (component_diff < -4 || component_diff > 3) { | |
412 use_differential[i / 2] = false; | |
413 sub_block_avg[i] = MakeColor444(avg_color_0); | |
414 sub_block_avg[j] = MakeColor444(avg_color_1); | |
415 } else { | |
416 sub_block_avg[i] = avg_color_555_0; | |
417 sub_block_avg[j] = avg_color_555_1; | |
418 } | |
419 } | |
420 } | |
421 | |
422 // Compute the error of each sub block before adjusting for luminance. These | |
423 // error values are later used for determining if we should flip the sub | |
424 // block or not. | |
425 uint32_t sub_block_err[4] = {0}; | |
426 for (unsigned int i = 0; i < 4; ++i) { | |
427 for (unsigned int j = 0; j < 8; ++j) { | |
428 sub_block_err[i] += GetColorError(sub_block_avg[i], sub_block_src[i][j]); | |
429 } | |
430 } | |
431 | |
432 bool flip = | |
433 sub_block_err[2] + sub_block_err[3] < sub_block_err[0] + sub_block_err[1]; | |
434 | |
435 // Clear destination buffer so that we can "or" in the results. | |
436 memset(dst, 0, 8); | |
437 | |
438 WriteDiff(dst, use_differential[!!flip]); | |
439 WriteFlip(dst, flip); | |
440 | |
441 uint8_t sub_block_off_0 = flip ? 2 : 0; | |
442 uint8_t sub_block_off_1 = sub_block_off_0 + 1; | |
443 | |
444 if (use_differential[!!flip]) { | |
445 WriteColors555(dst, sub_block_avg[sub_block_off_0], | |
446 sub_block_avg[sub_block_off_1]); | |
447 } else { | |
448 WriteColors444(dst, sub_block_avg[sub_block_off_0], | |
449 sub_block_avg[sub_block_off_1]); | |
450 } | |
451 | |
452 // Compute luminance for the first sub block. | |
453 ComputeLuminance(dst, sub_block_src[sub_block_off_0], | |
454 sub_block_avg[sub_block_off_0], 0, | |
455 g_idx_to_num[sub_block_off_0]); | |
456 // Compute luminance for the second sub block. | |
457 ComputeLuminance(dst, sub_block_src[sub_block_off_1], | |
458 sub_block_avg[sub_block_off_1], 1, | |
459 g_idx_to_num[sub_block_off_1]); | |
460 } | |
461 | |
462 } // namespace | |
463 | |
464 namespace cc { | |
465 | |
466 void TextureCompressorETC1::Compress(const uint8_t* src, | |
467 uint8_t* dst, | |
468 int width, | |
469 int height, | |
470 Quality quality) { | |
471 DCHECK(width >= 4 && (width & 3) == 0); | |
472 DCHECK(height >= 4 && (height & 3) == 0); | |
473 | |
474 Color ver_blocks[16]; | |
475 Color hor_blocks[16]; | |
476 | |
477 for (int y = 0; y < height; y += 4, src += width * 4 * 4) { | |
478 for (int x = 0; x < width; x += 4, dst += 8) { | |
479 const Color* row0 = reinterpret_cast<const Color*>(src + x * 4); | |
480 const Color* row1 = row0 + width; | |
481 const Color* row2 = row1 + width; | |
482 const Color* row3 = row2 + width; | |
483 | |
484 memcpy(ver_blocks, row0, 8); | |
485 memcpy(ver_blocks + 2, row1, 8); | |
486 memcpy(ver_blocks + 4, row2, 8); | |
487 memcpy(ver_blocks + 6, row3, 8); | |
488 memcpy(ver_blocks + 8, row0 + 2, 8); | |
489 memcpy(ver_blocks + 10, row1 + 2, 8); | |
490 memcpy(ver_blocks + 12, row2 + 2, 8); | |
491 memcpy(ver_blocks + 14, row3 + 2, 8); | |
492 | |
493 memcpy(hor_blocks, row0, 16); | |
494 memcpy(hor_blocks + 4, row1, 16); | |
495 memcpy(hor_blocks + 8, row2, 16); | |
496 memcpy(hor_blocks + 12, row3, 16); | |
497 | |
498 CompressBlock(dst, ver_blocks, hor_blocks); | |
499 } | |
500 } | |
501 } | |
502 | |
503 } // namespace cc | |
OLD | NEW |