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 "../webp/config.h" 14 #endif 15 16 #include <math.h> 17 18 #include "./backward_references.h" 19 #include "./histogram.h" 20 #include "../dsp/lossless.h" 21 #include "../utils/utils.h" 22 23 #define MAX_COST 1.e38 24 25 // Number of partitions for the three dominant (literal, red and blue) symbol 26 // costs. 27 #define NUM_PARTITIONS 4 28 // The size of the bin-hash corresponding to the three dominant costs. 29 #define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS) 30 // Maximum number of histograms allowed in greedy combining algorithm. 31 #define MAX_HISTO_GREEDY 100 32 33 static void HistogramClear(VP8LHistogram* const p) { 34 uint32_t* const literal = p->literal_; 35 const int cache_bits = p->palette_code_bits_; 36 const int histo_size = VP8LGetHistogramSize(cache_bits); 37 memset(p, 0, histo_size); 38 p->palette_code_bits_ = cache_bits; 39 p->literal_ = literal; 40 } 41 42 // Swap two histogram pointers. 43 static void HistogramSwap(VP8LHistogram** const A, VP8LHistogram** const B) { 44 VP8LHistogram* const tmp = *A; 45 *A = *B; 46 *B = tmp; 47 } 48 49 static void HistogramCopy(const VP8LHistogram* const src, 50 VP8LHistogram* const dst) { 51 uint32_t* const dst_literal = dst->literal_; 52 const int dst_cache_bits = dst->palette_code_bits_; 53 const int histo_size = VP8LGetHistogramSize(dst_cache_bits); 54 assert(src->palette_code_bits_ == dst_cache_bits); 55 memcpy(dst, src, histo_size); 56 dst->literal_ = dst_literal; 57 } 58 59 int VP8LGetHistogramSize(int cache_bits) { 60 const int literal_size = VP8LHistogramNumCodes(cache_bits); 61 const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size; 62 assert(total_size <= (size_t)0x7fffffff); 63 return (int)total_size; 64 } 65 66 void VP8LFreeHistogram(VP8LHistogram* const histo) { 67 WebPSafeFree(histo); 68 } 69 70 void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) { 71 WebPSafeFree(histo); 72 } 73 74 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs, 75 VP8LHistogram* const histo) { 76 VP8LRefsCursor c = VP8LRefsCursorInit(refs); 77 while (VP8LRefsCursorOk(&c)) { 78 VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos); 79 VP8LRefsCursorNext(&c); 80 } 81 } 82 83 void VP8LHistogramCreate(VP8LHistogram* const p, 84 const VP8LBackwardRefs* const refs, 85 int palette_code_bits) { 86 if (palette_code_bits >= 0) { 87 p->palette_code_bits_ = palette_code_bits; 88 } 89 HistogramClear(p); 90 VP8LHistogramStoreRefs(refs, p); 91 } 92 93 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) { 94 p->palette_code_bits_ = palette_code_bits; 95 HistogramClear(p); 96 } 97 98 VP8LHistogram* VP8LAllocateHistogram(int cache_bits) { 99 VP8LHistogram* histo = NULL; 100 const int total_size = VP8LGetHistogramSize(cache_bits); 101 uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory)); 102 if (memory == NULL) return NULL; 103 histo = (VP8LHistogram*)memory; 104 // literal_ won't necessary be aligned. 105 histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram)); 106 VP8LHistogramInit(histo, cache_bits); 107 return histo; 108 } 109 110 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) { 111 int i; 112 VP8LHistogramSet* set; 113 const int histo_size = VP8LGetHistogramSize(cache_bits); 114 const size_t total_size = 115 sizeof(*set) + size * (sizeof(*set->histograms) + 116 histo_size + WEBP_ALIGN_CST); 117 uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory)); 118 if (memory == NULL) return NULL; 119 120 set = (VP8LHistogramSet*)memory; 121 memory += sizeof(*set); 122 set->histograms = (VP8LHistogram**)memory; 123 memory += size * sizeof(*set->histograms); 124 set->max_size = size; 125 set->size = size; 126 for (i = 0; i < size; ++i) { 127 memory = (uint8_t*)WEBP_ALIGN(memory); 128 set->histograms[i] = (VP8LHistogram*)memory; 129 // literal_ won't necessary be aligned. 130 set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram)); 131 VP8LHistogramInit(set->histograms[i], cache_bits); 132 memory += histo_size; 133 } 134 return set; 135 } 136 137 // ----------------------------------------------------------------------------- 138 139 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo, 140 const PixOrCopy* const v) { 141 if (PixOrCopyIsLiteral(v)) { 142 ++histo->alpha_[PixOrCopyLiteral(v, 3)]; 143 ++histo->red_[PixOrCopyLiteral(v, 2)]; 144 ++histo->literal_[PixOrCopyLiteral(v, 1)]; 145 ++histo->blue_[PixOrCopyLiteral(v, 0)]; 146 } else if (PixOrCopyIsCacheIdx(v)) { 147 const int literal_ix = 148 NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v); 149 ++histo->literal_[literal_ix]; 150 } else { 151 int code, extra_bits; 152 VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits); 153 ++histo->literal_[NUM_LITERAL_CODES + code]; 154 VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits); 155 ++histo->distance_[code]; 156 } 157 } 158 159 // ----------------------------------------------------------------------------- 160 // Entropy-related functions. 161 162 static WEBP_INLINE double BitsEntropyRefine(const VP8LBitEntropy* entropy) { 163 double mix; 164 if (entropy->nonzeros < 5) { 165 if (entropy->nonzeros <= 1) { 166 return 0; 167 } 168 // Two symbols, they will be 0 and 1 in a Huffman code. 169 // Let's mix in a bit of entropy to favor good clustering when 170 // distributions of these are combined. 171 if (entropy->nonzeros == 2) { 172 return 0.99 * entropy->sum + 0.01 * entropy->entropy; 173 } 174 // No matter what the entropy says, we cannot be better than min_limit 175 // with Huffman coding. I am mixing a bit of entropy into the 176 // min_limit since it produces much better (~0.5 %) compression results 177 // perhaps because of better entropy clustering. 178 if (entropy->nonzeros == 3) { 179 mix = 0.95; 180 } else { 181 mix = 0.7; // nonzeros == 4. 182 } 183 } else { 184 mix = 0.627; 185 } 186 187 { 188 double min_limit = 2 * entropy->sum - entropy->max_val; 189 min_limit = mix * min_limit + (1.0 - mix) * entropy->entropy; 190 return (entropy->entropy < min_limit) ? min_limit : entropy->entropy; 191 } 192 } 193 194 double VP8LBitsEntropy(const uint32_t* const array, int n, 195 uint32_t* const trivial_symbol) { 196 VP8LBitEntropy entropy; 197 VP8LBitsEntropyUnrefined(array, n, &entropy); 198 if (trivial_symbol != NULL) { 199 *trivial_symbol = 200 (entropy.nonzeros == 1) ? entropy.nonzero_code : VP8L_NON_TRIVIAL_SYM; 201 } 202 203 return BitsEntropyRefine(&entropy); 204 } 205 206 static double InitialHuffmanCost(void) { 207 // Small bias because Huffman code length is typically not stored in 208 // full length. 209 static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3; 210 static const double kSmallBias = 9.1; 211 return kHuffmanCodeOfHuffmanCodeSize - kSmallBias; 212 } 213 214 // Finalize the Huffman cost based on streak numbers and length type (<3 or >=3) 215 static double FinalHuffmanCost(const VP8LStreaks* const stats) { 216 double retval = InitialHuffmanCost(); 217 retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1]; 218 retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1]; 219 retval += 1.796875 * stats->streaks[0][0]; 220 retval += 3.28125 * stats->streaks[1][0]; 221 return retval; 222 } 223 224 // Get the symbol entropy for the distribution 'population'. 225 // Set 'trivial_sym', if there's only one symbol present in the distribution. 226 static double PopulationCost(const uint32_t* const population, int length, 227 uint32_t* const trivial_sym) { 228 VP8LBitEntropy bit_entropy; 229 VP8LStreaks stats; 230 VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats); 231 if (trivial_sym != NULL) { 232 *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code 233 : VP8L_NON_TRIVIAL_SYM; 234 } 235 236 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); 237 } 238 239 static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X, 240 const uint32_t* const Y, 241 int length) { 242 VP8LBitEntropy bit_entropy; 243 VP8LStreaks stats; 244 VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats); 245 246 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); 247 } 248 249 // Estimates the Entropy + Huffman + other block overhead size cost. 250 double VP8LHistogramEstimateBits(const VP8LHistogram* const p) { 251 return 252 PopulationCost( 253 p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), NULL) 254 + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL) 255 + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL) 256 + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL) 257 + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL) 258 + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES) 259 + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES); 260 } 261 262 // ----------------------------------------------------------------------------- 263 // Various histogram combine/cost-eval functions 264 265 static int GetCombinedHistogramEntropy(const VP8LHistogram* const a, 266 const VP8LHistogram* const b, 267 double cost_threshold, 268 double* cost) { 269 const int palette_code_bits = a->palette_code_bits_; 270 assert(a->palette_code_bits_ == b->palette_code_bits_); 271 *cost += GetCombinedEntropy(a->literal_, b->literal_, 272 VP8LHistogramNumCodes(palette_code_bits)); 273 *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES, 274 b->literal_ + NUM_LITERAL_CODES, 275 NUM_LENGTH_CODES); 276 if (*cost > cost_threshold) return 0; 277 278 *cost += GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES); 279 if (*cost > cost_threshold) return 0; 280 281 *cost += GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES); 282 if (*cost > cost_threshold) return 0; 283 284 *cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES); 285 if (*cost > cost_threshold) return 0; 286 287 *cost += GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES); 288 *cost += 289 VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES); 290 if (*cost > cost_threshold) return 0; 291 292 return 1; 293 } 294 295 // Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing 296 // to the threshold value 'cost_threshold'. The score returned is 297 // Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed. 298 // Since the previous score passed is 'cost_threshold', we only need to compare 299 // the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out 300 // early. 301 static double HistogramAddEval(const VP8LHistogram* const a, 302 const VP8LHistogram* const b, 303 VP8LHistogram* const out, 304 double cost_threshold) { 305 double cost = 0; 306 const double sum_cost = a->bit_cost_ + b->bit_cost_; 307 cost_threshold += sum_cost; 308 309 if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) { 310 VP8LHistogramAdd(a, b, out); 311 out->bit_cost_ = cost; 312 out->palette_code_bits_ = a->palette_code_bits_; 313 out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_) ? 314 a->trivial_symbol_ : VP8L_NON_TRIVIAL_SYM; 315 } 316 317 return cost - sum_cost; 318 } 319 320 // Same as HistogramAddEval(), except that the resulting histogram 321 // is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit 322 // the term C(b) which is constant over all the evaluations. 323 static double HistogramAddThresh(const VP8LHistogram* const a, 324 const VP8LHistogram* const b, 325 double cost_threshold) { 326 double cost = -a->bit_cost_; 327 GetCombinedHistogramEntropy(a, b, cost_threshold, &cost); 328 return cost; 329 } 330 331 // ----------------------------------------------------------------------------- 332 333 // The structure to keep track of cost range for the three dominant entropy 334 // symbols. 335 // TODO(skal): Evaluate if float can be used here instead of double for 336 // representing the entropy costs. 337 typedef struct { 338 double literal_max_; 339 double literal_min_; 340 double red_max_; 341 double red_min_; 342 double blue_max_; 343 double blue_min_; 344 } DominantCostRange; 345 346 static void DominantCostRangeInit(DominantCostRange* const c) { 347 c->literal_max_ = 0.; 348 c->literal_min_ = MAX_COST; 349 c->red_max_ = 0.; 350 c->red_min_ = MAX_COST; 351 c->blue_max_ = 0.; 352 c->blue_min_ = MAX_COST; 353 } 354 355 static void UpdateDominantCostRange( 356 const VP8LHistogram* const h, DominantCostRange* const c) { 357 if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_; 358 if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_; 359 if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_; 360 if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_; 361 if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_; 362 if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_; 363 } 364 365 static void UpdateHistogramCost(VP8LHistogram* const h) { 366 uint32_t alpha_sym, red_sym, blue_sym; 367 const double alpha_cost = 368 PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym); 369 const double distance_cost = 370 PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL) + 371 VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES); 372 const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_); 373 h->literal_cost_ = PopulationCost(h->literal_, num_codes, NULL) + 374 VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, 375 NUM_LENGTH_CODES); 376 h->red_cost_ = PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym); 377 h->blue_cost_ = PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym); 378 h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ + 379 alpha_cost + distance_cost; 380 if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) { 381 h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM; 382 } else { 383 h->trivial_symbol_ = 384 ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0); 385 } 386 } 387 388 static int GetBinIdForEntropy(double min, double max, double val) { 389 const double range = max - min + 1e-6; 390 const double delta = val - min; 391 return (int)(NUM_PARTITIONS * delta / range); 392 } 393 394 static int GetHistoBinIndexLowEffort( 395 const VP8LHistogram* const h, const DominantCostRange* const c) { 396 const int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_, 397 h->literal_cost_); 398 assert(bin_id < NUM_PARTITIONS); 399 return bin_id; 400 } 401 402 static int GetHistoBinIndex( 403 const VP8LHistogram* const h, const DominantCostRange* const c) { 404 const int bin_id = 405 GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_) + 406 NUM_PARTITIONS * GetBinIdForEntropy(c->red_min_, c->red_max_, 407 h->red_cost_) + 408 NUM_PARTITIONS * NUM_PARTITIONS * GetBinIdForEntropy(c->literal_min_, 409 c->literal_max_, 410 h->literal_cost_); 411 assert(bin_id < BIN_SIZE); 412 return bin_id; 413 } 414 415 // Construct the histograms from backward references. 416 static void HistogramBuild( 417 int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs, 418 VP8LHistogramSet* const image_histo) { 419 int x = 0, y = 0; 420 const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits); 421 VP8LHistogram** const histograms = image_histo->histograms; 422 VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs); 423 assert(histo_bits > 0); 424 while (VP8LRefsCursorOk(&c)) { 425 const PixOrCopy* const v = c.cur_pos; 426 const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits); 427 VP8LHistogramAddSinglePixOrCopy(histograms[ix], v); 428 x += PixOrCopyLength(v); 429 while (x >= xsize) { 430 x -= xsize; 431 ++y; 432 } 433 VP8LRefsCursorNext(&c); 434 } 435 } 436 437 // Copies the histograms and computes its bit_cost. 438 static void HistogramCopyAndAnalyze( 439 VP8LHistogramSet* const orig_histo, VP8LHistogramSet* const image_histo) { 440 int i; 441 const int histo_size = orig_histo->size; 442 VP8LHistogram** const orig_histograms = orig_histo->histograms; 443 VP8LHistogram** const histograms = image_histo->histograms; 444 for (i = 0; i < histo_size; ++i) { 445 VP8LHistogram* const histo = orig_histograms[i]; 446 UpdateHistogramCost(histo); 447 // Copy histograms from orig_histo[] to image_histo[]. 448 HistogramCopy(histo, histograms[i]); 449 } 450 } 451 452 // Partition histograms to different entropy bins for three dominant (literal, 453 // red and blue) symbol costs and compute the histogram aggregate bit_cost. 454 static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo, 455 int16_t* const bin_map, int low_effort) { 456 int i; 457 VP8LHistogram** const histograms = image_histo->histograms; 458 const int histo_size = image_histo->size; 459 const int bin_depth = histo_size + 1; 460 DominantCostRange cost_range; 461 DominantCostRangeInit(&cost_range); 462 463 // Analyze the dominant (literal, red and blue) entropy costs. 464 for (i = 0; i < histo_size; ++i) { 465 VP8LHistogram* const histo = histograms[i]; 466 UpdateDominantCostRange(histo, &cost_range); 467 } 468 469 // bin-hash histograms on three of the dominant (literal, red and blue) 470 // symbol costs. 471 for (i = 0; i < histo_size; ++i) { 472 int num_histos; 473 VP8LHistogram* const histo = histograms[i]; 474 const int16_t bin_id = low_effort ? 475 (int16_t)GetHistoBinIndexLowEffort(histo, &cost_range) : 476 (int16_t)GetHistoBinIndex(histo, &cost_range); 477 const int bin_offset = bin_id * bin_depth; 478 // bin_map[n][0] for every bin 'n' maintains the counter for the number of 479 // histograms in that bin. 480 // Get and increment the num_histos in that bin. 481 num_histos = ++bin_map[bin_offset]; 482 assert(bin_offset + num_histos < bin_depth * BIN_SIZE); 483 // Add histogram i'th index at num_histos (last) position in the bin_map. 484 bin_map[bin_offset + num_histos] = i; 485 } 486 } 487 488 // Compact the histogram set by removing unused entries. 489 static void HistogramCompactBins(VP8LHistogramSet* const image_histo) { 490 VP8LHistogram** const histograms = image_histo->histograms; 491 int i, j; 492 493 for (i = 0, j = 0; i < image_histo->size; ++i) { 494 if (histograms[i] != NULL && histograms[i]->bit_cost_ != 0.) { 495 if (j < i) { 496 histograms[j] = histograms[i]; 497 histograms[i] = NULL; 498 } 499 ++j; 500 } 501 } 502 image_histo->size = j; 503 } 504 505 static VP8LHistogram* HistogramCombineEntropyBin( 506 VP8LHistogramSet* const image_histo, 507 VP8LHistogram* cur_combo, 508 int16_t* const bin_map, int bin_depth, int num_bins, 509 double combine_cost_factor, int low_effort) { 510 int bin_id; 511 VP8LHistogram** const histograms = image_histo->histograms; 512 513 for (bin_id = 0; bin_id < num_bins; ++bin_id) { 514 const int bin_offset = bin_id * bin_depth; 515 const int num_histos = bin_map[bin_offset]; 516 const int idx1 = bin_map[bin_offset + 1]; 517 int num_combine_failures = 0; 518 int n; 519 for (n = 2; n <= num_histos; ++n) { 520 const int idx2 = bin_map[bin_offset + n]; 521 if (low_effort) { 522 // Merge all histograms with the same bin index, irrespective of cost of 523 // the merged histograms. 524 VP8LHistogramAdd(histograms[idx1], histograms[idx2], histograms[idx1]); 525 histograms[idx2]->bit_cost_ = 0.; 526 } else { 527 const double bit_cost_idx2 = histograms[idx2]->bit_cost_; 528 if (bit_cost_idx2 > 0.) { 529 const double bit_cost_thresh = -bit_cost_idx2 * combine_cost_factor; 530 const double curr_cost_diff = 531 HistogramAddEval(histograms[idx1], histograms[idx2], 532 cur_combo, bit_cost_thresh); 533 if (curr_cost_diff < bit_cost_thresh) { 534 // Try to merge two histograms only if the combo is a trivial one or 535 // the two candidate histograms are already non-trivial. 536 // For some images, 'try_combine' turns out to be false for a lot of 537 // histogram pairs. In that case, we fallback to combining 538 // histograms as usual to avoid increasing the header size. 539 const int try_combine = 540 (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) || 541 ((histograms[idx1]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) && 542 (histograms[idx2]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM)); 543 const int max_combine_failures = 32; 544 if (try_combine || (num_combine_failures >= max_combine_failures)) { 545 HistogramSwap(&cur_combo, &histograms[idx1]); 546 histograms[idx2]->bit_cost_ = 0.; 547 } else { 548 ++num_combine_failures; 549 } 550 } 551 } 552 } 553 } 554 if (low_effort) { 555 // Update the bit_cost for the merged histograms (per bin index). 556 UpdateHistogramCost(histograms[idx1]); 557 } 558 } 559 HistogramCompactBins(image_histo); 560 return cur_combo; 561 } 562 563 static uint32_t MyRand(uint32_t *seed) { 564 *seed *= 16807U; 565 if (*seed == 0) { 566 *seed = 1; 567 } 568 return *seed; 569 } 570 571 // ----------------------------------------------------------------------------- 572 // Histogram pairs priority queue 573 574 // Pair of histograms. Negative idx1 value means that pair is out-of-date. 575 typedef struct { 576 int idx1; 577 int idx2; 578 double cost_diff; 579 double cost_combo; 580 } HistogramPair; 581 582 typedef struct { 583 HistogramPair* queue; 584 int size; 585 int max_size; 586 } HistoQueue; 587 588 static int HistoQueueInit(HistoQueue* const histo_queue, const int max_index) { 589 histo_queue->size = 0; 590 // max_index^2 for the queue size is safe. If you look at 591 // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes 592 // data to the queue, you insert at most: 593 // - max_index*(max_index-1)/2 (the first two for loops) 594 // - max_index - 1 in the last for loop at the first iteration of the while 595 // loop, max_index - 2 at the second iteration ... therefore 596 // max_index*(max_index-1)/2 overall too 597 histo_queue->max_size = max_index * max_index; 598 // We allocate max_size + 1 because the last element at index "size" is 599 // used as temporary data (and it could be up to max_size). 600 histo_queue->queue = WebPSafeMalloc(histo_queue->max_size + 1, 601 sizeof(*histo_queue->queue)); 602 return histo_queue->queue != NULL; 603 } 604 605 static void HistoQueueClear(HistoQueue* const histo_queue) { 606 assert(histo_queue != NULL); 607 WebPSafeFree(histo_queue->queue); 608 } 609 610 static void SwapHistogramPairs(HistogramPair *p1, 611 HistogramPair *p2) { 612 const HistogramPair tmp = *p1; 613 *p1 = *p2; 614 *p2 = tmp; 615 } 616 617 // Given a valid priority queue in range [0, queue_size) this function checks 618 // whether histo_queue[queue_size] should be accepted and swaps it with the 619 // front if it is smaller. Otherwise, it leaves it as is. 620 static void UpdateQueueFront(HistoQueue* const histo_queue) { 621 if (histo_queue->queue[histo_queue->size].cost_diff >= 0) return; 622 623 if (histo_queue->queue[histo_queue->size].cost_diff < 624 histo_queue->queue[0].cost_diff) { 625 SwapHistogramPairs(histo_queue->queue, 626 histo_queue->queue + histo_queue->size); 627 } 628 ++histo_queue->size; 629 630 // We cannot add more elements than the capacity. 631 // The allocation adds an extra element to the official capacity so that 632 // histo_queue->queue[histo_queue->max_size] is read/written within bound. 633 assert(histo_queue->size <= histo_queue->max_size); 634 } 635 636 // ----------------------------------------------------------------------------- 637 638 static void PreparePair(VP8LHistogram** histograms, int idx1, int idx2, 639 HistogramPair* const pair, 640 VP8LHistogram* const histos) { 641 if (idx1 > idx2) { 642 const int tmp = idx2; 643 idx2 = idx1; 644 idx1 = tmp; 645 } 646 pair->idx1 = idx1; 647 pair->idx2 = idx2; 648 pair->cost_diff = 649 HistogramAddEval(histograms[idx1], histograms[idx2], histos, 0); 650 pair->cost_combo = histos->bit_cost_; 651 } 652 653 // Combines histograms by continuously choosing the one with the highest cost 654 // reduction. 655 static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo, 656 VP8LHistogram* const histos) { 657 int ok = 0; 658 int image_histo_size = image_histo->size; 659 int i, j; 660 VP8LHistogram** const histograms = image_histo->histograms; 661 // Indexes of remaining histograms. 662 int* const clusters = WebPSafeMalloc(image_histo_size, sizeof(*clusters)); 663 // Priority queue of histogram pairs. 664 HistoQueue histo_queue; 665 666 if (!HistoQueueInit(&histo_queue, image_histo_size) || clusters == NULL) { 667 goto End; 668 } 669 670 for (i = 0; i < image_histo_size; ++i) { 671 // Initialize clusters indexes. 672 clusters[i] = i; 673 for (j = i + 1; j < image_histo_size; ++j) { 674 // Initialize positions array. 675 PreparePair(histograms, i, j, &histo_queue.queue[histo_queue.size], 676 histos); 677 UpdateQueueFront(&histo_queue); 678 } 679 } 680 681 while (image_histo_size > 1 && histo_queue.size > 0) { 682 HistogramPair* copy_to; 683 const int idx1 = histo_queue.queue[0].idx1; 684 const int idx2 = histo_queue.queue[0].idx2; 685 VP8LHistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]); 686 histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo; 687 // Remove merged histogram. 688 for (i = 0; i + 1 < image_histo_size; ++i) { 689 if (clusters[i] >= idx2) { 690 clusters[i] = clusters[i + 1]; 691 } 692 } 693 --image_histo_size; 694 695 // Remove pairs intersecting the just combined best pair. This will 696 // therefore pop the head of the queue. 697 copy_to = histo_queue.queue; 698 for (i = 0; i < histo_queue.size; ++i) { 699 HistogramPair* const p = histo_queue.queue + i; 700 if (p->idx1 == idx1 || p->idx2 == idx1 || 701 p->idx1 == idx2 || p->idx2 == idx2) { 702 // Do not copy the invalid pair. 703 continue; 704 } 705 if (p->cost_diff < histo_queue.queue[0].cost_diff) { 706 // Replace the top of the queue if we found better. 707 SwapHistogramPairs(histo_queue.queue, p); 708 } 709 SwapHistogramPairs(copy_to, p); 710 ++copy_to; 711 } 712 histo_queue.size = (int)(copy_to - histo_queue.queue); 713 714 // Push new pairs formed with combined histogram to the queue. 715 for (i = 0; i < image_histo_size; ++i) { 716 if (clusters[i] != idx1) { 717 PreparePair(histograms, idx1, clusters[i], 718 &histo_queue.queue[histo_queue.size], histos); 719 UpdateQueueFront(&histo_queue); 720 } 721 } 722 } 723 // Move remaining histograms to the beginning of the array. 724 for (i = 0; i < image_histo_size; ++i) { 725 if (i != clusters[i]) { // swap the two histograms 726 HistogramSwap(&histograms[i], &histograms[clusters[i]]); 727 } 728 } 729 730 image_histo->size = image_histo_size; 731 ok = 1; 732 733 End: 734 WebPSafeFree(clusters); 735 HistoQueueClear(&histo_queue); 736 return ok; 737 } 738 739 static VP8LHistogram* HistogramCombineStochastic( 740 VP8LHistogramSet* const image_histo, 741 VP8LHistogram* tmp_histo, 742 VP8LHistogram* best_combo, 743 int quality, int min_cluster_size) { 744 int iter; 745 uint32_t seed = 0; 746 int tries_with_no_success = 0; 747 int image_histo_size = image_histo->size; 748 const int iter_mult = (quality < 25) ? 2 : 2 + (quality - 25) / 8; 749 const int outer_iters = image_histo_size * iter_mult; 750 const int num_pairs = image_histo_size / 2; 751 const int num_tries_no_success = outer_iters / 2; 752 VP8LHistogram** const histograms = image_histo->histograms; 753 754 // Collapse similar histograms in 'image_histo'. 755 ++min_cluster_size; 756 for (iter = 0; 757 iter < outer_iters && image_histo_size >= min_cluster_size; 758 ++iter) { 759 double best_cost_diff = 0.; 760 int best_idx1 = -1, best_idx2 = 1; 761 int j; 762 const int num_tries = 763 (num_pairs < image_histo_size) ? num_pairs : image_histo_size; 764 seed += iter; 765 for (j = 0; j < num_tries; ++j) { 766 double curr_cost_diff; 767 // Choose two histograms at random and try to combine them. 768 const uint32_t idx1 = MyRand(&seed) % image_histo_size; 769 const uint32_t tmp = (j & 7) + 1; 770 const uint32_t diff = 771 (tmp < 3) ? tmp : MyRand(&seed) % (image_histo_size - 1); 772 const uint32_t idx2 = (idx1 + diff + 1) % image_histo_size; 773 if (idx1 == idx2) { 774 continue; 775 } 776 777 // Calculate cost reduction on combining. 778 curr_cost_diff = HistogramAddEval(histograms[idx1], histograms[idx2], 779 tmp_histo, best_cost_diff); 780 if (curr_cost_diff < best_cost_diff) { // found a better pair? 781 HistogramSwap(&best_combo, &tmp_histo); 782 best_cost_diff = curr_cost_diff; 783 best_idx1 = idx1; 784 best_idx2 = idx2; 785 } 786 } 787 788 if (best_idx1 >= 0) { 789 HistogramSwap(&best_combo, &histograms[best_idx1]); 790 // swap best_idx2 slot with last one (which is now unused) 791 --image_histo_size; 792 if (best_idx2 != image_histo_size) { 793 HistogramSwap(&histograms[image_histo_size], &histograms[best_idx2]); 794 histograms[image_histo_size] = NULL; 795 } 796 tries_with_no_success = 0; 797 } 798 if (++tries_with_no_success >= num_tries_no_success) { 799 break; 800 } 801 } 802 image_histo->size = image_histo_size; 803 return best_combo; 804 } 805 806 // ----------------------------------------------------------------------------- 807 // Histogram refinement 808 809 // Find the best 'out' histogram for each of the 'in' histograms. 810 // Note: we assume that out[]->bit_cost_ is already up-to-date. 811 static void HistogramRemap(const VP8LHistogramSet* const orig_histo, 812 const VP8LHistogramSet* const image_histo, 813 uint16_t* const symbols) { 814 int i; 815 VP8LHistogram** const orig_histograms = orig_histo->histograms; 816 VP8LHistogram** const histograms = image_histo->histograms; 817 const int orig_histo_size = orig_histo->size; 818 const int image_histo_size = image_histo->size; 819 if (image_histo_size > 1) { 820 for (i = 0; i < orig_histo_size; ++i) { 821 int best_out = 0; 822 double best_bits = 823 HistogramAddThresh(histograms[0], orig_histograms[i], MAX_COST); 824 int k; 825 for (k = 1; k < image_histo_size; ++k) { 826 const double cur_bits = 827 HistogramAddThresh(histograms[k], orig_histograms[i], best_bits); 828 if (cur_bits < best_bits) { 829 best_bits = cur_bits; 830 best_out = k; 831 } 832 } 833 symbols[i] = best_out; 834 } 835 } else { 836 assert(image_histo_size == 1); 837 for (i = 0; i < orig_histo_size; ++i) { 838 symbols[i] = 0; 839 } 840 } 841 842 // Recompute each out based on raw and symbols. 843 for (i = 0; i < image_histo_size; ++i) { 844 HistogramClear(histograms[i]); 845 } 846 847 for (i = 0; i < orig_histo_size; ++i) { 848 const int idx = symbols[i]; 849 VP8LHistogramAdd(orig_histograms[i], histograms[idx], histograms[idx]); 850 } 851 } 852 853 static double GetCombineCostFactor(int histo_size, int quality) { 854 double combine_cost_factor = 0.16; 855 if (quality < 90) { 856 if (histo_size > 256) combine_cost_factor /= 2.; 857 if (histo_size > 512) combine_cost_factor /= 2.; 858 if (histo_size > 1024) combine_cost_factor /= 2.; 859 if (quality <= 50) combine_cost_factor /= 2.; 860 } 861 return combine_cost_factor; 862 } 863 864 int VP8LGetHistoImageSymbols(int xsize, int ysize, 865 const VP8LBackwardRefs* const refs, 866 int quality, int low_effort, 867 int histo_bits, int cache_bits, 868 VP8LHistogramSet* const image_histo, 869 VP8LHistogramSet* const tmp_histos, 870 uint16_t* const histogram_symbols) { 871 int ok = 0; 872 const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1; 873 const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1; 874 const int image_histo_raw_size = histo_xsize * histo_ysize; 875 const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE; 876 877 // The bin_map for every bin follows following semantics: 878 // bin_map[n][0] = num_histo; // The number of histograms in that bin. 879 // bin_map[n][1] = index of first histogram in that bin; 880 // bin_map[n][num_histo] = index of last histogram in that bin; 881 // bin_map[n][num_histo + 1] ... bin_map[n][bin_depth - 1] = unused indices. 882 const int bin_depth = image_histo_raw_size + 1; 883 int16_t* bin_map = NULL; 884 VP8LHistogramSet* const orig_histo = 885 VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits); 886 VP8LHistogram* cur_combo; 887 const int entropy_combine = 888 (orig_histo->size > entropy_combine_num_bins * 2) && (quality < 100); 889 890 if (orig_histo == NULL) goto Error; 891 892 // Don't attempt linear bin-partition heuristic for: 893 // histograms of small sizes, as bin_map will be very sparse and; 894 // Maximum quality (q==100), to preserve the compression gains at that level. 895 if (entropy_combine) { 896 const int bin_map_size = bin_depth * entropy_combine_num_bins; 897 bin_map = (int16_t*)WebPSafeCalloc(bin_map_size, sizeof(*bin_map)); 898 if (bin_map == NULL) goto Error; 899 } 900 901 // Construct the histograms from backward references. 902 HistogramBuild(xsize, histo_bits, refs, orig_histo); 903 // Copies the histograms and computes its bit_cost. 904 HistogramCopyAndAnalyze(orig_histo, image_histo); 905 906 cur_combo = tmp_histos->histograms[1]; // pick up working slot 907 if (entropy_combine) { 908 const double combine_cost_factor = 909 GetCombineCostFactor(image_histo_raw_size, quality); 910 HistogramAnalyzeEntropyBin(orig_histo, bin_map, low_effort); 911 // Collapse histograms with similar entropy. 912 cur_combo = HistogramCombineEntropyBin(image_histo, cur_combo, bin_map, 913 bin_depth, entropy_combine_num_bins, 914 combine_cost_factor, low_effort); 915 } 916 917 // Don't combine the histograms using stochastic and greedy heuristics for 918 // low-effort compression mode. 919 if (!low_effort || !entropy_combine) { 920 const float x = quality / 100.f; 921 // cubic ramp between 1 and MAX_HISTO_GREEDY: 922 const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1)); 923 cur_combo = HistogramCombineStochastic(image_histo, 924 tmp_histos->histograms[0], 925 cur_combo, quality, threshold_size); 926 if ((image_histo->size <= threshold_size) && 927 !HistogramCombineGreedy(image_histo, cur_combo)) { 928 goto Error; 929 } 930 } 931 932 // TODO(vikasa): Optimize HistogramRemap for low-effort compression mode also. 933 // Find the optimal map from original histograms to the final ones. 934 HistogramRemap(orig_histo, image_histo, histogram_symbols); 935 936 ok = 1; 937 938 Error: 939 WebPSafeFree(bin_map); 940 VP8LFreeHistogramSet(orig_histo); 941 return ok; 942 } 943
