1 // Copyright 2012 Google Inc. All Rights Reserved. 2 // 3 // Use of this source code is governed by a BSD-style license 4 // that can be found in the COPYING file in the root of the source 5 // tree. An additional intellectual property rights grant can be found 6 // in the file PATENTS. All contributing project authors may 7 // be found in the AUTHORS file in the root of the source tree. 8 // ----------------------------------------------------------------------------- 9 // 10 // Author: Jyrki Alakuijala (jyrki (at) google.com) 11 // 12 #ifdef HAVE_CONFIG_H 13 #include "src/webp/config.h" 14 #endif 15 16 #include <math.h> 17 18 #include "src/enc/backward_references_enc.h" 19 #include "src/enc/histogram_enc.h" 20 #include "src/dsp/lossless.h" 21 #include "src/dsp/lossless_common.h" 22 #include "src/utils/utils.h" 23 24 #define MAX_COST 1.e38 25 26 // Number of partitions for the three dominant (literal, red and blue) symbol 27 // costs. 28 #define NUM_PARTITIONS 4 29 // The size of the bin-hash corresponding to the three dominant costs. 30 #define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS) 31 // Maximum number of histograms allowed in greedy combining algorithm. 32 #define MAX_HISTO_GREEDY 100 33 34 static void HistogramClear(VP8LHistogram* const p) { 35 uint32_t* const literal = p->literal_; 36 const int cache_bits = p->palette_code_bits_; 37 const int histo_size = VP8LGetHistogramSize(cache_bits); 38 memset(p, 0, histo_size); 39 p->palette_code_bits_ = cache_bits; 40 p->literal_ = literal; 41 } 42 43 // Swap two histogram pointers. 44 static void HistogramSwap(VP8LHistogram** const A, VP8LHistogram** const B) { 45 VP8LHistogram* const tmp = *A; 46 *A = *B; 47 *B = tmp; 48 } 49 50 static void HistogramCopy(const VP8LHistogram* const src, 51 VP8LHistogram* const dst) { 52 uint32_t* const dst_literal = dst->literal_; 53 const int dst_cache_bits = dst->palette_code_bits_; 54 const int histo_size = VP8LGetHistogramSize(dst_cache_bits); 55 assert(src->palette_code_bits_ == dst_cache_bits); 56 memcpy(dst, src, histo_size); 57 dst->literal_ = dst_literal; 58 } 59 60 int VP8LGetHistogramSize(int cache_bits) { 61 const int literal_size = VP8LHistogramNumCodes(cache_bits); 62 const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size; 63 assert(total_size <= (size_t)0x7fffffff); 64 return (int)total_size; 65 } 66 67 void VP8LFreeHistogram(VP8LHistogram* const histo) { 68 WebPSafeFree(histo); 69 } 70 71 void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) { 72 WebPSafeFree(histo); 73 } 74 75 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs, 76 VP8LHistogram* const histo) { 77 VP8LRefsCursor c = VP8LRefsCursorInit(refs); 78 while (VP8LRefsCursorOk(&c)) { 79 VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos, NULL, 0); 80 VP8LRefsCursorNext(&c); 81 } 82 } 83 84 void VP8LHistogramCreate(VP8LHistogram* const p, 85 const VP8LBackwardRefs* const refs, 86 int palette_code_bits) { 87 if (palette_code_bits >= 0) { 88 p->palette_code_bits_ = palette_code_bits; 89 } 90 HistogramClear(p); 91 VP8LHistogramStoreRefs(refs, p); 92 } 93 94 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) { 95 p->palette_code_bits_ = palette_code_bits; 96 HistogramClear(p); 97 } 98 99 VP8LHistogram* VP8LAllocateHistogram(int cache_bits) { 100 VP8LHistogram* histo = NULL; 101 const int total_size = VP8LGetHistogramSize(cache_bits); 102 uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory)); 103 if (memory == NULL) return NULL; 104 histo = (VP8LHistogram*)memory; 105 // literal_ won't necessary be aligned. 106 histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram)); 107 VP8LHistogramInit(histo, cache_bits); 108 return histo; 109 } 110 111 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) { 112 int i; 113 VP8LHistogramSet* set; 114 const int histo_size = VP8LGetHistogramSize(cache_bits); 115 const size_t total_size = 116 sizeof(*set) + size * (sizeof(*set->histograms) + 117 histo_size + WEBP_ALIGN_CST); 118 uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory)); 119 if (memory == NULL) return NULL; 120 121 set = (VP8LHistogramSet*)memory; 122 memory += sizeof(*set); 123 set->histograms = (VP8LHistogram**)memory; 124 memory += size * sizeof(*set->histograms); 125 set->max_size = size; 126 set->size = size; 127 for (i = 0; i < size; ++i) { 128 memory = (uint8_t*)WEBP_ALIGN(memory); 129 set->histograms[i] = (VP8LHistogram*)memory; 130 // literal_ won't necessary be aligned. 131 set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram)); 132 VP8LHistogramInit(set->histograms[i], cache_bits); 133 memory += histo_size; 134 } 135 return set; 136 } 137 138 // ----------------------------------------------------------------------------- 139 140 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo, 141 const PixOrCopy* const v, 142 int (*const distance_modifier)(int, int), 143 int distance_modifier_arg0) { 144 if (PixOrCopyIsLiteral(v)) { 145 ++histo->alpha_[PixOrCopyLiteral(v, 3)]; 146 ++histo->red_[PixOrCopyLiteral(v, 2)]; 147 ++histo->literal_[PixOrCopyLiteral(v, 1)]; 148 ++histo->blue_[PixOrCopyLiteral(v, 0)]; 149 } else if (PixOrCopyIsCacheIdx(v)) { 150 const int literal_ix = 151 NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v); 152 ++histo->literal_[literal_ix]; 153 } else { 154 int code, extra_bits; 155 VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits); 156 ++histo->literal_[NUM_LITERAL_CODES + code]; 157 if (distance_modifier == NULL) { 158 VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits); 159 } else { 160 VP8LPrefixEncodeBits( 161 distance_modifier(distance_modifier_arg0, PixOrCopyDistance(v)), 162 &code, &extra_bits); 163 } 164 ++histo->distance_[code]; 165 } 166 } 167 168 // ----------------------------------------------------------------------------- 169 // Entropy-related functions. 170 171 static WEBP_INLINE double BitsEntropyRefine(const VP8LBitEntropy* entropy) { 172 double mix; 173 if (entropy->nonzeros < 5) { 174 if (entropy->nonzeros <= 1) { 175 return 0; 176 } 177 // Two symbols, they will be 0 and 1 in a Huffman code. 178 // Let's mix in a bit of entropy to favor good clustering when 179 // distributions of these are combined. 180 if (entropy->nonzeros == 2) { 181 return 0.99 * entropy->sum + 0.01 * entropy->entropy; 182 } 183 // No matter what the entropy says, we cannot be better than min_limit 184 // with Huffman coding. I am mixing a bit of entropy into the 185 // min_limit since it produces much better (~0.5 %) compression results 186 // perhaps because of better entropy clustering. 187 if (entropy->nonzeros == 3) { 188 mix = 0.95; 189 } else { 190 mix = 0.7; // nonzeros == 4. 191 } 192 } else { 193 mix = 0.627; 194 } 195 196 { 197 double min_limit = 2 * entropy->sum - entropy->max_val; 198 min_limit = mix * min_limit + (1.0 - mix) * entropy->entropy; 199 return (entropy->entropy < min_limit) ? min_limit : entropy->entropy; 200 } 201 } 202 203 double VP8LBitsEntropy(const uint32_t* const array, int n, 204 uint32_t* const trivial_symbol) { 205 VP8LBitEntropy entropy; 206 VP8LBitsEntropyUnrefined(array, n, &entropy); 207 if (trivial_symbol != NULL) { 208 *trivial_symbol = 209 (entropy.nonzeros == 1) ? entropy.nonzero_code : VP8L_NON_TRIVIAL_SYM; 210 } 211 212 return BitsEntropyRefine(&entropy); 213 } 214 215 static double InitialHuffmanCost(void) { 216 // Small bias because Huffman code length is typically not stored in 217 // full length. 218 static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3; 219 static const double kSmallBias = 9.1; 220 return kHuffmanCodeOfHuffmanCodeSize - kSmallBias; 221 } 222 223 // Finalize the Huffman cost based on streak numbers and length type (<3 or >=3) 224 static double FinalHuffmanCost(const VP8LStreaks* const stats) { 225 // The constants in this function are experimental and got rounded from 226 // their original values in 1/8 when switched to 1/1024. 227 double retval = InitialHuffmanCost(); 228 // Second coefficient: Many zeros in the histogram are covered efficiently 229 // by a run-length encode. Originally 2/8. 230 retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1]; 231 // Second coefficient: Constant values are encoded less efficiently, but still 232 // RLE'ed. Originally 6/8. 233 retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1]; 234 // 0s are usually encoded more efficiently than non-0s. 235 // Originally 15/8. 236 retval += 1.796875 * stats->streaks[0][0]; 237 // Originally 26/8. 238 retval += 3.28125 * stats->streaks[1][0]; 239 return retval; 240 } 241 242 // Get the symbol entropy for the distribution 'population'. 243 // Set 'trivial_sym', if there's only one symbol present in the distribution. 244 static double PopulationCost(const uint32_t* const population, int length, 245 uint32_t* const trivial_sym) { 246 VP8LBitEntropy bit_entropy; 247 VP8LStreaks stats; 248 VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats); 249 if (trivial_sym != NULL) { 250 *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code 251 : VP8L_NON_TRIVIAL_SYM; 252 } 253 254 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); 255 } 256 257 // trivial_at_end is 1 if the two histograms only have one element that is 258 // non-zero: both the zero-th one, or both the last one. 259 static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X, 260 const uint32_t* const Y, 261 int length, int trivial_at_end) { 262 VP8LStreaks stats; 263 if (trivial_at_end) { 264 // This configuration is due to palettization that transforms an indexed 265 // pixel into 0xff000000 | (pixel << 8) in VP8LBundleColorMap. 266 // BitsEntropyRefine is 0 for histograms with only one non-zero value. 267 // Only FinalHuffmanCost needs to be evaluated. 268 memset(&stats, 0, sizeof(stats)); 269 // Deal with the non-zero value at index 0 or length-1. 270 stats.streaks[1][0] += 1; 271 // Deal with the following/previous zero streak. 272 stats.counts[0] += 1; 273 stats.streaks[0][1] += length - 1; 274 return FinalHuffmanCost(&stats); 275 } else { 276 VP8LBitEntropy bit_entropy; 277 VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats); 278 279 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); 280 } 281 } 282 283 // Estimates the Entropy + Huffman + other block overhead size cost. 284 double VP8LHistogramEstimateBits(const VP8LHistogram* const p) { 285 return 286 PopulationCost( 287 p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), NULL) 288 + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL) 289 + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL) 290 + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL) 291 + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL) 292 + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES) 293 + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES); 294 } 295 296 // ----------------------------------------------------------------------------- 297 // Various histogram combine/cost-eval functions 298 299 static int GetCombinedHistogramEntropy(const VP8LHistogram* const a, 300 const VP8LHistogram* const b, 301 double cost_threshold, 302 double* cost) { 303 const int palette_code_bits = a->palette_code_bits_; 304 int trivial_at_end = 0; 305 assert(a->palette_code_bits_ == b->palette_code_bits_); 306 *cost += GetCombinedEntropy(a->literal_, b->literal_, 307 VP8LHistogramNumCodes(palette_code_bits), 0); 308 *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES, 309 b->literal_ + NUM_LITERAL_CODES, 310 NUM_LENGTH_CODES); 311 if (*cost > cost_threshold) return 0; 312 313 if (a->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM && 314 a->trivial_symbol_ == b->trivial_symbol_) { 315 // A, R and B are all 0 or 0xff. 316 const uint32_t color_a = (a->trivial_symbol_ >> 24) & 0xff; 317 const uint32_t color_r = (a->trivial_symbol_ >> 16) & 0xff; 318 const uint32_t color_b = (a->trivial_symbol_ >> 0) & 0xff; 319 if ((color_a == 0 || color_a == 0xff) && 320 (color_r == 0 || color_r == 0xff) && 321 (color_b == 0 || color_b == 0xff)) { 322 trivial_at_end = 1; 323 } 324 } 325 326 *cost += 327 GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, trivial_at_end); 328 if (*cost > cost_threshold) return 0; 329 330 *cost += 331 GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, trivial_at_end); 332 if (*cost > cost_threshold) return 0; 333 334 *cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES, 335 trivial_at_end); 336 if (*cost > cost_threshold) return 0; 337 338 *cost += 339 GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES, 0); 340 *cost += 341 VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES); 342 if (*cost > cost_threshold) return 0; 343 344 return 1; 345 } 346 347 static WEBP_INLINE void HistogramAdd(const VP8LHistogram* const a, 348 const VP8LHistogram* const b, 349 VP8LHistogram* const out) { 350 VP8LHistogramAdd(a, b, out); 351 out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_) 352 ? a->trivial_symbol_ 353 : VP8L_NON_TRIVIAL_SYM; 354 } 355 356 // Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing 357 // to the threshold value 'cost_threshold'. The score returned is 358 // Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed. 359 // Since the previous score passed is 'cost_threshold', we only need to compare 360 // the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out 361 // early. 362 static double HistogramAddEval(const VP8LHistogram* const a, 363 const VP8LHistogram* const b, 364 VP8LHistogram* const out, 365 double cost_threshold) { 366 double cost = 0; 367 const double sum_cost = a->bit_cost_ + b->bit_cost_; 368 cost_threshold += sum_cost; 369 370 if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) { 371 HistogramAdd(a, b, out); 372 out->bit_cost_ = cost; 373 out->palette_code_bits_ = a->palette_code_bits_; 374 } 375 376 return cost - sum_cost; 377 } 378 379 // Same as HistogramAddEval(), except that the resulting histogram 380 // is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit 381 // the term C(b) which is constant over all the evaluations. 382 static double HistogramAddThresh(const VP8LHistogram* const a, 383 const VP8LHistogram* const b, 384 double cost_threshold) { 385 double cost = -a->bit_cost_; 386 GetCombinedHistogramEntropy(a, b, cost_threshold, &cost); 387 return cost; 388 } 389 390 // ----------------------------------------------------------------------------- 391 392 // The structure to keep track of cost range for the three dominant entropy 393 // symbols. 394 // TODO(skal): Evaluate if float can be used here instead of double for 395 // representing the entropy costs. 396 typedef struct { 397 double literal_max_; 398 double literal_min_; 399 double red_max_; 400 double red_min_; 401 double blue_max_; 402 double blue_min_; 403 } DominantCostRange; 404 405 static void DominantCostRangeInit(DominantCostRange* const c) { 406 c->literal_max_ = 0.; 407 c->literal_min_ = MAX_COST; 408 c->red_max_ = 0.; 409 c->red_min_ = MAX_COST; 410 c->blue_max_ = 0.; 411 c->blue_min_ = MAX_COST; 412 } 413 414 static void UpdateDominantCostRange( 415 const VP8LHistogram* const h, DominantCostRange* const c) { 416 if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_; 417 if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_; 418 if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_; 419 if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_; 420 if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_; 421 if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_; 422 } 423 424 static void UpdateHistogramCost(VP8LHistogram* const h) { 425 uint32_t alpha_sym, red_sym, blue_sym; 426 const double alpha_cost = 427 PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym); 428 const double distance_cost = 429 PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL) + 430 VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES); 431 const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_); 432 h->literal_cost_ = PopulationCost(h->literal_, num_codes, NULL) + 433 VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, 434 NUM_LENGTH_CODES); 435 h->red_cost_ = PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym); 436 h->blue_cost_ = PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym); 437 h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ + 438 alpha_cost + distance_cost; 439 if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) { 440 h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM; 441 } else { 442 h->trivial_symbol_ = 443 ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0); 444 } 445 } 446 447 static int GetBinIdForEntropy(double min, double max, double val) { 448 const double range = max - min; 449 if (range > 0.) { 450 const double delta = val - min; 451 return (int)((NUM_PARTITIONS - 1e-6) * delta / range); 452 } else { 453 return 0; 454 } 455 } 456 457 static int GetHistoBinIndex(const VP8LHistogram* const h, 458 const DominantCostRange* const c, int low_effort) { 459 int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_, 460 h->literal_cost_); 461 assert(bin_id < NUM_PARTITIONS); 462 if (!low_effort) { 463 bin_id = bin_id * NUM_PARTITIONS 464 + GetBinIdForEntropy(c->red_min_, c->red_max_, h->red_cost_); 465 bin_id = bin_id * NUM_PARTITIONS 466 + GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_); 467 assert(bin_id < BIN_SIZE); 468 } 469 return bin_id; 470 } 471 472 // Construct the histograms from backward references. 473 static void HistogramBuild( 474 int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs, 475 VP8LHistogramSet* const image_histo) { 476 int x = 0, y = 0; 477 const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits); 478 VP8LHistogram** const histograms = image_histo->histograms; 479 VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs); 480 assert(histo_bits > 0); 481 while (VP8LRefsCursorOk(&c)) { 482 const PixOrCopy* const v = c.cur_pos; 483 const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits); 484 VP8LHistogramAddSinglePixOrCopy(histograms[ix], v, NULL, 0); 485 x += PixOrCopyLength(v); 486 while (x >= xsize) { 487 x -= xsize; 488 ++y; 489 } 490 VP8LRefsCursorNext(&c); 491 } 492 } 493 494 // Copies the histograms and computes its bit_cost. 495 static void HistogramCopyAndAnalyze( 496 VP8LHistogramSet* const orig_histo, VP8LHistogramSet* const image_histo) { 497 int i; 498 const int histo_size = orig_histo->size; 499 VP8LHistogram** const orig_histograms = orig_histo->histograms; 500 VP8LHistogram** const histograms = image_histo->histograms; 501 for (i = 0; i < histo_size; ++i) { 502 VP8LHistogram* const histo = orig_histograms[i]; 503 UpdateHistogramCost(histo); 504 // Copy histograms from orig_histo[] to image_histo[]. 505 HistogramCopy(histo, histograms[i]); 506 } 507 } 508 509 // Partition histograms to different entropy bins for three dominant (literal, 510 // red and blue) symbol costs and compute the histogram aggregate bit_cost. 511 static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo, 512 uint16_t* const bin_map, 513 int low_effort) { 514 int i; 515 VP8LHistogram** const histograms = image_histo->histograms; 516 const int histo_size = image_histo->size; 517 DominantCostRange cost_range; 518 DominantCostRangeInit(&cost_range); 519 520 // Analyze the dominant (literal, red and blue) entropy costs. 521 for (i = 0; i < histo_size; ++i) { 522 UpdateDominantCostRange(histograms[i], &cost_range); 523 } 524 525 // bin-hash histograms on three of the dominant (literal, red and blue) 526 // symbol costs and store the resulting bin_id for each histogram. 527 for (i = 0; i < histo_size; ++i) { 528 bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort); 529 } 530 } 531 532 // Compact image_histo[] by merging some histograms with same bin_id together if 533 // it's advantageous. 534 static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo, 535 VP8LHistogram* cur_combo, 536 const uint16_t* const bin_map, 537 int bin_map_size, int num_bins, 538 double combine_cost_factor, 539 int low_effort) { 540 VP8LHistogram** const histograms = image_histo->histograms; 541 int idx; 542 // Work in-place: processed histograms are put at the beginning of 543 // image_histo[]. At the end, we just have to truncate the array. 544 int size = 0; 545 struct { 546 int16_t first; // position of the histogram that accumulates all 547 // histograms with the same bin_id 548 uint16_t num_combine_failures; // number of combine failures per bin_id 549 } bin_info[BIN_SIZE]; 550 551 assert(num_bins <= BIN_SIZE); 552 for (idx = 0; idx < num_bins; ++idx) { 553 bin_info[idx].first = -1; 554 bin_info[idx].num_combine_failures = 0; 555 } 556 557 for (idx = 0; idx < bin_map_size; ++idx) { 558 const int bin_id = bin_map[idx]; 559 const int first = bin_info[bin_id].first; 560 assert(size <= idx); 561 if (first == -1) { 562 // just move histogram #idx to its final position 563 histograms[size] = histograms[idx]; 564 bin_info[bin_id].first = size++; 565 } else if (low_effort) { 566 HistogramAdd(histograms[idx], histograms[first], histograms[first]); 567 } else { 568 // try to merge #idx into #first (both share the same bin_id) 569 const double bit_cost = histograms[idx]->bit_cost_; 570 const double bit_cost_thresh = -bit_cost * combine_cost_factor; 571 const double curr_cost_diff = 572 HistogramAddEval(histograms[first], histograms[idx], 573 cur_combo, bit_cost_thresh); 574 if (curr_cost_diff < bit_cost_thresh) { 575 // Try to merge two histograms only if the combo is a trivial one or 576 // the two candidate histograms are already non-trivial. 577 // For some images, 'try_combine' turns out to be false for a lot of 578 // histogram pairs. In that case, we fallback to combining 579 // histograms as usual to avoid increasing the header size. 580 const int try_combine = 581 (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) || 582 ((histograms[idx]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) && 583 (histograms[first]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM)); 584 const int max_combine_failures = 32; 585 if (try_combine || 586 bin_info[bin_id].num_combine_failures >= max_combine_failures) { 587 // move the (better) merged histogram to its final slot 588 HistogramSwap(&cur_combo, &histograms[first]); 589 } else { 590 histograms[size++] = histograms[idx]; 591 ++bin_info[bin_id].num_combine_failures; 592 } 593 } else { 594 histograms[size++] = histograms[idx]; 595 } 596 } 597 } 598 image_histo->size = size; 599 if (low_effort) { 600 // for low_effort case, update the final cost when everything is merged 601 for (idx = 0; idx < size; ++idx) { 602 UpdateHistogramCost(histograms[idx]); 603 } 604 } 605 } 606 607 // Implement a Lehmer random number generator with a multiplicative constant of 608 // 48271 and a modulo constant of 2^31 1. 609 static uint32_t MyRand(uint32_t* const seed) { 610 *seed = (uint32_t)(((uint64_t)(*seed) * 48271u) % 2147483647u); 611 assert(*seed > 0); 612 return *seed; 613 } 614 615 // ----------------------------------------------------------------------------- 616 // Histogram pairs priority queue 617 618 // Pair of histograms. Negative idx1 value means that pair is out-of-date. 619 typedef struct { 620 int idx1; 621 int idx2; 622 double cost_diff; 623 double cost_combo; 624 } HistogramPair; 625 626 typedef struct { 627 HistogramPair* queue; 628 int size; 629 int max_size; 630 } HistoQueue; 631 632 static int HistoQueueInit(HistoQueue* const histo_queue, const int max_index) { 633 histo_queue->size = 0; 634 // max_index^2 for the queue size is safe. If you look at 635 // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes 636 // data to the queue, you insert at most: 637 // - max_index*(max_index-1)/2 (the first two for loops) 638 // - max_index - 1 in the last for loop at the first iteration of the while 639 // loop, max_index - 2 at the second iteration ... therefore 640 // max_index*(max_index-1)/2 overall too 641 histo_queue->max_size = max_index * max_index; 642 // We allocate max_size + 1 because the last element at index "size" is 643 // used as temporary data (and it could be up to max_size). 644 histo_queue->queue = (HistogramPair*)WebPSafeMalloc( 645 histo_queue->max_size + 1, sizeof(*histo_queue->queue)); 646 return histo_queue->queue != NULL; 647 } 648 649 static void HistoQueueClear(HistoQueue* const histo_queue) { 650 assert(histo_queue != NULL); 651 WebPSafeFree(histo_queue->queue); 652 histo_queue->size = 0; 653 histo_queue->max_size = 0; 654 } 655 656 // Pop a specific pair in the queue by replacing it with the last one 657 // and shrinking the queue. 658 static void HistoQueuePopPair(HistoQueue* const histo_queue, 659 HistogramPair* const pair) { 660 assert(pair >= histo_queue->queue && 661 pair < (histo_queue->queue + histo_queue->size)); 662 assert(histo_queue->size > 0); 663 *pair = histo_queue->queue[histo_queue->size - 1]; 664 --histo_queue->size; 665 } 666 667 // Check whether a pair in the queue should be updated as head or not. 668 static void HistoQueueUpdateHead(HistoQueue* const histo_queue, 669 HistogramPair* const pair) { 670 assert(pair->cost_diff < 0.); 671 assert(pair >= histo_queue->queue && 672 pair < (histo_queue->queue + histo_queue->size)); 673 assert(histo_queue->size > 0); 674 if (pair->cost_diff < histo_queue->queue[0].cost_diff) { 675 // Replace the best pair. 676 const HistogramPair tmp = histo_queue->queue[0]; 677 histo_queue->queue[0] = *pair; 678 *pair = tmp; 679 } 680 } 681 682 // Create a pair from indices "idx1" and "idx2" provided its cost 683 // is inferior to "threshold", a negative entropy. 684 // It returns the cost of the pair, or 0. if it superior to threshold. 685 static double HistoQueuePush(HistoQueue* const histo_queue, 686 VP8LHistogram** const histograms, int idx1, 687 int idx2, double threshold) { 688 const VP8LHistogram* h1; 689 const VP8LHistogram* h2; 690 HistogramPair pair; 691 double sum_cost; 692 693 assert(threshold <= 0.); 694 if (idx1 > idx2) { 695 const int tmp = idx2; 696 idx2 = idx1; 697 idx1 = tmp; 698 } 699 pair.idx1 = idx1; 700 pair.idx2 = idx2; 701 h1 = histograms[idx1]; 702 h2 = histograms[idx2]; 703 sum_cost = h1->bit_cost_ + h2->bit_cost_; 704 pair.cost_combo = 0.; 705 GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair.cost_combo); 706 pair.cost_diff = pair.cost_combo - sum_cost; 707 708 // Do not even consider the pair if it does not improve the entropy. 709 if (pair.cost_diff >= threshold) return 0.; 710 711 // We cannot add more elements than the capacity. 712 assert(histo_queue->size < histo_queue->max_size); 713 histo_queue->queue[histo_queue->size++] = pair; 714 HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]); 715 716 return pair.cost_diff; 717 } 718 719 // ----------------------------------------------------------------------------- 720 721 // Combines histograms by continuously choosing the one with the highest cost 722 // reduction. 723 static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) { 724 int ok = 0; 725 int image_histo_size = image_histo->size; 726 int i, j; 727 VP8LHistogram** const histograms = image_histo->histograms; 728 // Indexes of remaining histograms. 729 int* const clusters = 730 (int*)WebPSafeMalloc(image_histo_size, sizeof(*clusters)); 731 // Priority queue of histogram pairs. 732 HistoQueue histo_queue; 733 734 if (!HistoQueueInit(&histo_queue, image_histo_size) || clusters == NULL) { 735 goto End; 736 } 737 738 for (i = 0; i < image_histo_size; ++i) { 739 // Initialize clusters indexes. 740 clusters[i] = i; 741 for (j = i + 1; j < image_histo_size; ++j) { 742 // Initialize positions array. 743 HistoQueuePush(&histo_queue, histograms, i, j, 0.); 744 } 745 } 746 747 while (image_histo_size > 1 && histo_queue.size > 0) { 748 const int idx1 = histo_queue.queue[0].idx1; 749 const int idx2 = histo_queue.queue[0].idx2; 750 HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]); 751 histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo; 752 // Remove merged histogram. 753 for (i = 0; i + 1 < image_histo_size; ++i) { 754 if (clusters[i] >= idx2) { 755 clusters[i] = clusters[i + 1]; 756 } 757 } 758 --image_histo_size; 759 760 // Remove pairs intersecting the just combined best pair. 761 for (i = 0; i < histo_queue.size;) { 762 HistogramPair* const p = histo_queue.queue + i; 763 if (p->idx1 == idx1 || p->idx2 == idx1 || 764 p->idx1 == idx2 || p->idx2 == idx2) { 765 HistoQueuePopPair(&histo_queue, p); 766 } else { 767 HistoQueueUpdateHead(&histo_queue, p); 768 ++i; 769 } 770 } 771 772 // Push new pairs formed with combined histogram to the queue. 773 for (i = 0; i < image_histo_size; ++i) { 774 if (clusters[i] != idx1) { 775 HistoQueuePush(&histo_queue, histograms, idx1, clusters[i], 0.); 776 } 777 } 778 } 779 // Move remaining histograms to the beginning of the array. 780 for (i = 0; i < image_histo_size; ++i) { 781 if (i != clusters[i]) { // swap the two histograms 782 HistogramSwap(&histograms[i], &histograms[clusters[i]]); 783 } 784 } 785 786 image_histo->size = image_histo_size; 787 ok = 1; 788 789 End: 790 WebPSafeFree(clusters); 791 HistoQueueClear(&histo_queue); 792 return ok; 793 } 794 795 // Perform histogram aggregation using a stochastic approach. 796 // 'do_greedy' is set to 1 if a greedy approach needs to be performed 797 // afterwards, 0 otherwise. 798 static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, 799 int min_cluster_size, 800 int* const do_greedy) { 801 int iter; 802 uint32_t seed = 1; 803 int tries_with_no_success = 0; 804 int image_histo_size = image_histo->size; 805 const int outer_iters = image_histo_size; 806 const int num_tries_no_success = outer_iters / 2; 807 VP8LHistogram** const histograms = image_histo->histograms; 808 // Priority queue of histogram pairs. Its size of "kCostHeapSizeSqrt"^2 809 // impacts the quality of the compression and the speed: the smaller the 810 // faster but the worse for the compression. 811 HistoQueue histo_queue; 812 const int kHistoQueueSizeSqrt = 3; 813 int ok = 0; 814 815 if (!HistoQueueInit(&histo_queue, kHistoQueueSizeSqrt)) { 816 goto End; 817 } 818 // Collapse similar histograms in 'image_histo'. 819 ++min_cluster_size; 820 for (iter = 0; iter < outer_iters && image_histo_size >= min_cluster_size && 821 ++tries_with_no_success < num_tries_no_success; 822 ++iter) { 823 double best_cost = 824 (histo_queue.size == 0) ? 0. : histo_queue.queue[0].cost_diff; 825 int best_idx1 = -1, best_idx2 = 1; 826 int j; 827 const uint32_t rand_range = (image_histo_size - 1) * image_histo_size; 828 // image_histo_size / 2 was chosen empirically. Less means faster but worse 829 // compression. 830 const int num_tries = image_histo_size / 2; 831 832 for (j = 0; j < num_tries; ++j) { 833 double curr_cost; 834 // Choose two different histograms at random and try to combine them. 835 const uint32_t tmp = MyRand(&seed) % rand_range; 836 const uint32_t idx1 = tmp / (image_histo_size - 1); 837 uint32_t idx2 = tmp % (image_histo_size - 1); 838 if (idx2 >= idx1) ++idx2; 839 840 // Calculate cost reduction on combination. 841 curr_cost = 842 HistoQueuePush(&histo_queue, histograms, idx1, idx2, best_cost); 843 if (curr_cost < 0) { // found a better pair? 844 best_cost = curr_cost; 845 // Empty the queue if we reached full capacity. 846 if (histo_queue.size == histo_queue.max_size) break; 847 } 848 } 849 if (histo_queue.size == 0) continue; 850 851 // Merge the two best histograms. 852 best_idx1 = histo_queue.queue[0].idx1; 853 best_idx2 = histo_queue.queue[0].idx2; 854 assert(best_idx1 < best_idx2); 855 HistogramAddEval(histograms[best_idx1], histograms[best_idx2], 856 histograms[best_idx1], 0); 857 // Swap the best_idx2 histogram with the last one (which is now unused). 858 --image_histo_size; 859 if (best_idx2 != image_histo_size) { 860 HistogramSwap(&histograms[image_histo_size], &histograms[best_idx2]); 861 } 862 histograms[image_histo_size] = NULL; 863 // Parse the queue and update each pair that deals with best_idx1, 864 // best_idx2 or image_histo_size. 865 for (j = 0; j < histo_queue.size;) { 866 HistogramPair* const p = histo_queue.queue + j; 867 const int is_idx1_best = p->idx1 == best_idx1 || p->idx1 == best_idx2; 868 const int is_idx2_best = p->idx2 == best_idx1 || p->idx2 == best_idx2; 869 int do_eval = 0; 870 // The front pair could have been duplicated by a random pick so 871 // check for it all the time nevertheless. 872 if (is_idx1_best && is_idx2_best) { 873 HistoQueuePopPair(&histo_queue, p); 874 continue; 875 } 876 // Any pair containing one of the two best indices should only refer to 877 // best_idx1. Its cost should also be updated. 878 if (is_idx1_best) { 879 p->idx1 = best_idx1; 880 do_eval = 1; 881 } else if (is_idx2_best) { 882 p->idx2 = best_idx1; 883 do_eval = 1; 884 } 885 if (p->idx2 == image_histo_size) { 886 // No need to re-evaluate here as it does not involve a pair 887 // containing best_idx1 or best_idx2. 888 p->idx2 = best_idx2; 889 } 890 assert(p->idx2 < image_histo_size); 891 // Make sure the index order is respected. 892 if (p->idx1 > p->idx2) { 893 const int tmp = p->idx2; 894 p->idx2 = p->idx1; 895 p->idx1 = tmp; 896 } 897 if (do_eval) { 898 // Re-evaluate the cost of an updated pair. 899 GetCombinedHistogramEntropy(histograms[p->idx1], histograms[p->idx2], 0, 900 &p->cost_diff); 901 if (p->cost_diff >= 0.) { 902 HistoQueuePopPair(&histo_queue, p); 903 continue; 904 } 905 } 906 HistoQueueUpdateHead(&histo_queue, p); 907 ++j; 908 } 909 910 tries_with_no_success = 0; 911 } 912 image_histo->size = image_histo_size; 913 *do_greedy = (image_histo->size <= min_cluster_size); 914 ok = 1; 915 916 End: 917 HistoQueueClear(&histo_queue); 918 return ok; 919 } 920 921 // ----------------------------------------------------------------------------- 922 // Histogram refinement 923 924 // Find the best 'out' histogram for each of the 'in' histograms. 925 // Note: we assume that out[]->bit_cost_ is already up-to-date. 926 static void HistogramRemap(const VP8LHistogramSet* const in, 927 const VP8LHistogramSet* const out, 928 uint16_t* const symbols) { 929 int i; 930 VP8LHistogram** const in_histo = in->histograms; 931 VP8LHistogram** const out_histo = out->histograms; 932 const int in_size = in->size; 933 const int out_size = out->size; 934 if (out_size > 1) { 935 for (i = 0; i < in_size; ++i) { 936 int best_out = 0; 937 double best_bits = MAX_COST; 938 int k; 939 for (k = 0; k < out_size; ++k) { 940 const double cur_bits = 941 HistogramAddThresh(out_histo[k], in_histo[i], best_bits); 942 if (k == 0 || cur_bits < best_bits) { 943 best_bits = cur_bits; 944 best_out = k; 945 } 946 } 947 symbols[i] = best_out; 948 } 949 } else { 950 assert(out_size == 1); 951 for (i = 0; i < in_size; ++i) { 952 symbols[i] = 0; 953 } 954 } 955 956 // Recompute each out based on raw and symbols. 957 for (i = 0; i < out_size; ++i) { 958 HistogramClear(out_histo[i]); 959 } 960 961 for (i = 0; i < in_size; ++i) { 962 const int idx = symbols[i]; 963 HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]); 964 } 965 } 966 967 static double GetCombineCostFactor(int histo_size, int quality) { 968 double combine_cost_factor = 0.16; 969 if (quality < 90) { 970 if (histo_size > 256) combine_cost_factor /= 2.; 971 if (histo_size > 512) combine_cost_factor /= 2.; 972 if (histo_size > 1024) combine_cost_factor /= 2.; 973 if (quality <= 50) combine_cost_factor /= 2.; 974 } 975 return combine_cost_factor; 976 } 977 978 int VP8LGetHistoImageSymbols(int xsize, int ysize, 979 const VP8LBackwardRefs* const refs, 980 int quality, int low_effort, 981 int histo_bits, int cache_bits, 982 VP8LHistogramSet* const image_histo, 983 VP8LHistogram* const tmp_histo, 984 uint16_t* const histogram_symbols) { 985 int ok = 0; 986 const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1; 987 const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1; 988 const int image_histo_raw_size = histo_xsize * histo_ysize; 989 VP8LHistogramSet* const orig_histo = 990 VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits); 991 // Don't attempt linear bin-partition heuristic for 992 // histograms of small sizes (as bin_map will be very sparse) and 993 // maximum quality q==100 (to preserve the compression gains at that level). 994 const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE; 995 const int entropy_combine = 996 (orig_histo->size > entropy_combine_num_bins * 2) && (quality < 100); 997 998 if (orig_histo == NULL) goto Error; 999 1000 // Construct the histograms from backward references. 1001 HistogramBuild(xsize, histo_bits, refs, orig_histo); 1002 // Copies the histograms and computes its bit_cost. 1003 HistogramCopyAndAnalyze(orig_histo, image_histo); 1004 1005 if (entropy_combine) { 1006 const int bin_map_size = orig_histo->size; 1007 // Reuse histogram_symbols storage. By definition, it's guaranteed to be ok. 1008 uint16_t* const bin_map = histogram_symbols; 1009 const double combine_cost_factor = 1010 GetCombineCostFactor(image_histo_raw_size, quality); 1011 1012 HistogramAnalyzeEntropyBin(orig_histo, bin_map, low_effort); 1013 // Collapse histograms with similar entropy. 1014 HistogramCombineEntropyBin(image_histo, tmp_histo, bin_map, bin_map_size, 1015 entropy_combine_num_bins, combine_cost_factor, 1016 low_effort); 1017 } 1018 1019 // Don't combine the histograms using stochastic and greedy heuristics for 1020 // low-effort compression mode. 1021 if (!low_effort || !entropy_combine) { 1022 const float x = quality / 100.f; 1023 // cubic ramp between 1 and MAX_HISTO_GREEDY: 1024 const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1)); 1025 int do_greedy; 1026 if (!HistogramCombineStochastic(image_histo, threshold_size, &do_greedy)) { 1027 goto Error; 1028 } 1029 if (do_greedy && !HistogramCombineGreedy(image_histo)) { 1030 goto Error; 1031 } 1032 } 1033 1034 // TODO(vikasa): Optimize HistogramRemap for low-effort compression mode also. 1035 // Find the optimal map from original histograms to the final ones. 1036 HistogramRemap(orig_histo, image_histo, histogram_symbols); 1037 1038 ok = 1; 1039 1040 Error: 1041 VP8LFreeHistogramSet(orig_histo); 1042 return ok; 1043 } 1044
