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@google.com)
11//
12#ifdef HAVE_CONFIG_H
13#include "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
31static void HistogramClear(VP8LHistogram* const p) {
32  uint32_t* const literal = p->literal_;
33  const int cache_bits = p->palette_code_bits_;
34  const int histo_size = VP8LGetHistogramSize(cache_bits);
35  memset(p, 0, histo_size);
36  p->palette_code_bits_ = cache_bits;
37  p->literal_ = literal;
38}
39
40static void HistogramCopy(const VP8LHistogram* const src,
41                          VP8LHistogram* const dst) {
42  uint32_t* const dst_literal = dst->literal_;
43  const int dst_cache_bits = dst->palette_code_bits_;
44  const int histo_size = VP8LGetHistogramSize(dst_cache_bits);
45  assert(src->palette_code_bits_ == dst_cache_bits);
46  memcpy(dst, src, histo_size);
47  dst->literal_ = dst_literal;
48}
49
50int VP8LGetHistogramSize(int cache_bits) {
51  const int literal_size = VP8LHistogramNumCodes(cache_bits);
52  const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size;
53  assert(total_size <= (size_t)0x7fffffff);
54  return (int)total_size;
55}
56
57void VP8LFreeHistogram(VP8LHistogram* const histo) {
58  WebPSafeFree(histo);
59}
60
61void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
62  WebPSafeFree(histo);
63}
64
65void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
66                            VP8LHistogram* const histo) {
67  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
68  while (VP8LRefsCursorOk(&c)) {
69    VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos);
70    VP8LRefsCursorNext(&c);
71  }
72}
73
74void VP8LHistogramCreate(VP8LHistogram* const p,
75                         const VP8LBackwardRefs* const refs,
76                         int palette_code_bits) {
77  if (palette_code_bits >= 0) {
78    p->palette_code_bits_ = palette_code_bits;
79  }
80  HistogramClear(p);
81  VP8LHistogramStoreRefs(refs, p);
82}
83
84void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) {
85  p->palette_code_bits_ = palette_code_bits;
86  HistogramClear(p);
87}
88
89VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
90  VP8LHistogram* histo = NULL;
91  const int total_size = VP8LGetHistogramSize(cache_bits);
92  uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
93  if (memory == NULL) return NULL;
94  histo = (VP8LHistogram*)memory;
95  // literal_ won't necessary be aligned.
96  histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
97  VP8LHistogramInit(histo, cache_bits);
98  return histo;
99}
100
101VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
102  int i;
103  VP8LHistogramSet* set;
104  const size_t total_size = sizeof(*set)
105                            + sizeof(*set->histograms) * size
106                            + (size_t)VP8LGetHistogramSize(cache_bits) * size;
107  uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
108  if (memory == NULL) return NULL;
109
110  set = (VP8LHistogramSet*)memory;
111  memory += sizeof(*set);
112  set->histograms = (VP8LHistogram**)memory;
113  memory += size * sizeof(*set->histograms);
114  set->max_size = size;
115  set->size = size;
116  for (i = 0; i < size; ++i) {
117    set->histograms[i] = (VP8LHistogram*)memory;
118    // literal_ won't necessary be aligned.
119    set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
120    VP8LHistogramInit(set->histograms[i], cache_bits);
121    // There's no padding/alignment between successive histograms.
122    memory += VP8LGetHistogramSize(cache_bits);
123  }
124  return set;
125}
126
127// -----------------------------------------------------------------------------
128
129void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
130                                     const PixOrCopy* const v) {
131  if (PixOrCopyIsLiteral(v)) {
132    ++histo->alpha_[PixOrCopyLiteral(v, 3)];
133    ++histo->red_[PixOrCopyLiteral(v, 2)];
134    ++histo->literal_[PixOrCopyLiteral(v, 1)];
135    ++histo->blue_[PixOrCopyLiteral(v, 0)];
136  } else if (PixOrCopyIsCacheIdx(v)) {
137    const int literal_ix =
138        NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
139    ++histo->literal_[literal_ix];
140  } else {
141    int code, extra_bits;
142    VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
143    ++histo->literal_[NUM_LITERAL_CODES + code];
144    VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
145    ++histo->distance_[code];
146  }
147}
148
149static WEBP_INLINE double BitsEntropyRefine(int nonzeros, int sum, int max_val,
150                                            double retval) {
151  double mix;
152  if (nonzeros < 5) {
153    if (nonzeros <= 1) {
154      return 0;
155    }
156    // Two symbols, they will be 0 and 1 in a Huffman code.
157    // Let's mix in a bit of entropy to favor good clustering when
158    // distributions of these are combined.
159    if (nonzeros == 2) {
160      return 0.99 * sum + 0.01 * retval;
161    }
162    // No matter what the entropy says, we cannot be better than min_limit
163    // with Huffman coding. I am mixing a bit of entropy into the
164    // min_limit since it produces much better (~0.5 %) compression results
165    // perhaps because of better entropy clustering.
166    if (nonzeros == 3) {
167      mix = 0.95;
168    } else {
169      mix = 0.7;  // nonzeros == 4.
170    }
171  } else {
172    mix = 0.627;
173  }
174
175  {
176    double min_limit = 2 * sum - max_val;
177    min_limit = mix * min_limit + (1.0 - mix) * retval;
178    return (retval < min_limit) ? min_limit : retval;
179  }
180}
181
182static double BitsEntropy(const uint32_t* const array, int n) {
183  double retval = 0.;
184  uint32_t sum = 0;
185  int nonzeros = 0;
186  uint32_t max_val = 0;
187  int i;
188  for (i = 0; i < n; ++i) {
189    if (array[i] != 0) {
190      sum += array[i];
191      ++nonzeros;
192      retval -= VP8LFastSLog2(array[i]);
193      if (max_val < array[i]) {
194        max_val = array[i];
195      }
196    }
197  }
198  retval += VP8LFastSLog2(sum);
199  return BitsEntropyRefine(nonzeros, sum, max_val, retval);
200}
201
202static double BitsEntropyCombined(const uint32_t* const X,
203                                  const uint32_t* const Y, int n) {
204  double retval = 0.;
205  int sum = 0;
206  int nonzeros = 0;
207  int max_val = 0;
208  int i;
209  for (i = 0; i < n; ++i) {
210    const int xy = X[i] + Y[i];
211    if (xy != 0) {
212      sum += xy;
213      ++nonzeros;
214      retval -= VP8LFastSLog2(xy);
215      if (max_val < xy) {
216        max_val = xy;
217      }
218    }
219  }
220  retval += VP8LFastSLog2(sum);
221  return BitsEntropyRefine(nonzeros, sum, max_val, retval);
222}
223
224static double InitialHuffmanCost(void) {
225  // Small bias because Huffman code length is typically not stored in
226  // full length.
227  static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
228  static const double kSmallBias = 9.1;
229  return kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
230}
231
232// Finalize the Huffman cost based on streak numbers and length type (<3 or >=3)
233static double FinalHuffmanCost(const VP8LStreaks* const stats) {
234  double retval = InitialHuffmanCost();
235  retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1];
236  retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1];
237  retval += 1.796875 * stats->streaks[0][0];
238  retval += 3.28125 * stats->streaks[1][0];
239  return retval;
240}
241
242// Trampolines
243static double HuffmanCost(const uint32_t* const population, int length) {
244  const VP8LStreaks stats = VP8LHuffmanCostCount(population, length);
245  return FinalHuffmanCost(&stats);
246}
247
248static double HuffmanCostCombined(const uint32_t* const X,
249                                  const uint32_t* const Y, int length) {
250  const VP8LStreaks stats = VP8LHuffmanCostCombinedCount(X, Y, length);
251  return FinalHuffmanCost(&stats);
252}
253
254// Aggregated costs
255static double PopulationCost(const uint32_t* const population, int length) {
256  return BitsEntropy(population, length) + HuffmanCost(population, length);
257}
258
259static double GetCombinedEntropy(const uint32_t* const X,
260                                 const uint32_t* const Y, int length) {
261  return BitsEntropyCombined(X, Y, length) + HuffmanCostCombined(X, Y, length);
262}
263
264// Estimates the Entropy + Huffman + other block overhead size cost.
265double VP8LHistogramEstimateBits(const VP8LHistogram* const p) {
266  return
267      PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_))
268      + PopulationCost(p->red_, NUM_LITERAL_CODES)
269      + PopulationCost(p->blue_, NUM_LITERAL_CODES)
270      + PopulationCost(p->alpha_, NUM_LITERAL_CODES)
271      + PopulationCost(p->distance_, NUM_DISTANCE_CODES)
272      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
273      + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
274}
275
276double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) {
277  return
278      BitsEntropy(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_))
279      + BitsEntropy(p->red_, NUM_LITERAL_CODES)
280      + BitsEntropy(p->blue_, NUM_LITERAL_CODES)
281      + BitsEntropy(p->alpha_, NUM_LITERAL_CODES)
282      + BitsEntropy(p->distance_, NUM_DISTANCE_CODES)
283      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
284      + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
285}
286
287// -----------------------------------------------------------------------------
288// Various histogram combine/cost-eval functions
289
290static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
291                                       const VP8LHistogram* const b,
292                                       double cost_threshold,
293                                       double* cost) {
294  const int palette_code_bits = a->palette_code_bits_;
295  assert(a->palette_code_bits_ == b->palette_code_bits_);
296  *cost += GetCombinedEntropy(a->literal_, b->literal_,
297                              VP8LHistogramNumCodes(palette_code_bits));
298  *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
299                                 b->literal_ + NUM_LITERAL_CODES,
300                                 NUM_LENGTH_CODES);
301  if (*cost > cost_threshold) return 0;
302
303  *cost += GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES);
304  if (*cost > cost_threshold) return 0;
305
306  *cost += GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES);
307  if (*cost > cost_threshold) return 0;
308
309  *cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES);
310  if (*cost > cost_threshold) return 0;
311
312  *cost += GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES);
313  *cost += VP8LExtraCostCombined(a->distance_, b->distance_,
314                                 NUM_DISTANCE_CODES);
315  if (*cost > cost_threshold) return 0;
316
317  return 1;
318}
319
320// Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing
321// to the threshold value 'cost_threshold'. The score returned is
322//  Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed.
323// Since the previous score passed is 'cost_threshold', we only need to compare
324// the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out
325// early.
326static double HistogramAddEval(const VP8LHistogram* const a,
327                               const VP8LHistogram* const b,
328                               VP8LHistogram* const out,
329                               double cost_threshold) {
330  double cost = 0;
331  const double sum_cost = a->bit_cost_ + b->bit_cost_;
332  cost_threshold += sum_cost;
333
334  if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
335    VP8LHistogramAdd(a, b, out);
336    out->bit_cost_ = cost;
337    out->palette_code_bits_ = a->palette_code_bits_;
338  }
339
340  return cost - sum_cost;
341}
342
343// Same as HistogramAddEval(), except that the resulting histogram
344// is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit
345// the term C(b) which is constant over all the evaluations.
346static double HistogramAddThresh(const VP8LHistogram* const a,
347                                 const VP8LHistogram* const b,
348                                 double cost_threshold) {
349  double cost = -a->bit_cost_;
350  GetCombinedHistogramEntropy(a, b, cost_threshold, &cost);
351  return cost;
352}
353
354// -----------------------------------------------------------------------------
355
356// The structure to keep track of cost range for the three dominant entropy
357// symbols.
358// TODO(skal): Evaluate if float can be used here instead of double for
359// representing the entropy costs.
360typedef struct {
361  double literal_max_;
362  double literal_min_;
363  double red_max_;
364  double red_min_;
365  double blue_max_;
366  double blue_min_;
367} DominantCostRange;
368
369static void DominantCostRangeInit(DominantCostRange* const c) {
370  c->literal_max_ = 0.;
371  c->literal_min_ = MAX_COST;
372  c->red_max_ = 0.;
373  c->red_min_ = MAX_COST;
374  c->blue_max_ = 0.;
375  c->blue_min_ = MAX_COST;
376}
377
378static void UpdateDominantCostRange(
379    const VP8LHistogram* const h, DominantCostRange* const c) {
380  if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_;
381  if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_;
382  if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_;
383  if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_;
384  if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_;
385  if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_;
386}
387
388static void UpdateHistogramCost(VP8LHistogram* const h) {
389  const double alpha_cost = PopulationCost(h->alpha_, NUM_LITERAL_CODES);
390  const double distance_cost =
391      PopulationCost(h->distance_, NUM_DISTANCE_CODES) +
392      VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
393  const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
394  h->literal_cost_ = PopulationCost(h->literal_, num_codes) +
395                     VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES,
396                                   NUM_LENGTH_CODES);
397  h->red_cost_ = PopulationCost(h->red_, NUM_LITERAL_CODES);
398  h->blue_cost_ = PopulationCost(h->blue_, NUM_LITERAL_CODES);
399  h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
400                 alpha_cost + distance_cost;
401}
402
403static int GetBinIdForEntropy(double min, double max, double val) {
404  const double range = max - min + 1e-6;
405  const double delta = val - min;
406  return (int)(NUM_PARTITIONS * delta / range);
407}
408
409// TODO(vikasa): Evaluate, if there's any correlation between red & blue.
410static int GetHistoBinIndex(
411    const VP8LHistogram* const h, const DominantCostRange* const c) {
412  const int bin_id =
413      GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_) +
414      NUM_PARTITIONS * GetBinIdForEntropy(c->red_min_, c->red_max_,
415                                          h->red_cost_) +
416      NUM_PARTITIONS * NUM_PARTITIONS * GetBinIdForEntropy(c->literal_min_,
417                                                           c->literal_max_,
418                                                           h->literal_cost_);
419  assert(bin_id < BIN_SIZE);
420  return bin_id;
421}
422
423// Construct the histograms from backward references.
424static void HistogramBuild(
425    int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs,
426    VP8LHistogramSet* const image_histo) {
427  int x = 0, y = 0;
428  const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
429  VP8LHistogram** const histograms = image_histo->histograms;
430  VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs);
431  assert(histo_bits > 0);
432  // Construct the Histo from a given backward references.
433  while (VP8LRefsCursorOk(&c)) {
434    const PixOrCopy* const v = c.cur_pos;
435    const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
436    VP8LHistogramAddSinglePixOrCopy(histograms[ix], v);
437    x += PixOrCopyLength(v);
438    while (x >= xsize) {
439      x -= xsize;
440      ++y;
441    }
442    VP8LRefsCursorNext(&c);
443  }
444}
445
446// Copies the histograms and computes its bit_cost.
447static void HistogramCopyAndAnalyze(
448    VP8LHistogramSet* const orig_histo, VP8LHistogramSet* const image_histo) {
449  int i;
450  const int histo_size = orig_histo->size;
451  VP8LHistogram** const orig_histograms = orig_histo->histograms;
452  VP8LHistogram** const histograms = image_histo->histograms;
453  for (i = 0; i < histo_size; ++i) {
454    VP8LHistogram* const histo = orig_histograms[i];
455    UpdateHistogramCost(histo);
456    // Copy histograms from orig_histo[] to image_histo[].
457    HistogramCopy(histo, histograms[i]);
458  }
459}
460
461// Partition histograms to different entropy bins for three dominant (literal,
462// red and blue) symbol costs and compute the histogram aggregate bit_cost.
463static void HistogramAnalyzeEntropyBin(
464    VP8LHistogramSet* const image_histo, int16_t* const bin_map) {
465  int i;
466  VP8LHistogram** const histograms = image_histo->histograms;
467  const int histo_size = image_histo->size;
468  const int bin_depth = histo_size + 1;
469  DominantCostRange cost_range;
470  DominantCostRangeInit(&cost_range);
471
472  // Analyze the dominant (literal, red and blue) entropy costs.
473  for (i = 0; i < histo_size; ++i) {
474    VP8LHistogram* const histo = histograms[i];
475    UpdateDominantCostRange(histo, &cost_range);
476  }
477
478  // bin-hash histograms on three of the dominant (literal, red and blue)
479  // symbol costs.
480  for (i = 0; i < histo_size; ++i) {
481    int num_histos;
482    VP8LHistogram* const histo = histograms[i];
483    const int16_t bin_id = (int16_t)GetHistoBinIndex(histo, &cost_range);
484    const int bin_offset = bin_id * bin_depth;
485    // bin_map[n][0] for every bin 'n' maintains the counter for the number of
486    // histograms in that bin.
487    // Get and increment the num_histos in that bin.
488    num_histos = ++bin_map[bin_offset];
489    assert(bin_offset + num_histos < bin_depth * BIN_SIZE);
490    // Add histogram i'th index at num_histos (last) position in the bin_map.
491    bin_map[bin_offset + num_histos] = i;
492  }
493}
494
495// Compact the histogram set by moving the valid one left in the set to the
496// head and moving the ones that have been merged to other histograms towards
497// the end.
498// TODO(vikasa): Evaluate if this method can be avoided by altering the code
499// logic of HistogramCombineEntropyBin main loop.
500static void HistogramCompactBins(VP8LHistogramSet* const image_histo) {
501  int start = 0;
502  int end = image_histo->size - 1;
503  VP8LHistogram** const histograms = image_histo->histograms;
504  while (start < end) {
505    while (start <= end && histograms[start] != NULL &&
506           histograms[start]->bit_cost_ != 0.) {
507      ++start;
508    }
509    while (start <= end && histograms[end]->bit_cost_ == 0.) {
510      histograms[end] = NULL;
511      --end;
512    }
513    if (start < end) {
514      assert(histograms[start] != NULL);
515      assert(histograms[end] != NULL);
516      HistogramCopy(histograms[end], histograms[start]);
517      histograms[end] = NULL;
518      --end;
519    }
520  }
521  image_histo->size = end + 1;
522}
523
524static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo,
525                                       VP8LHistogram* const histos,
526                                       int16_t* const bin_map, int bin_depth,
527                                       double combine_cost_factor) {
528  int bin_id;
529  VP8LHistogram* cur_combo = histos;
530  VP8LHistogram** const histograms = image_histo->histograms;
531
532  for (bin_id = 0; bin_id < BIN_SIZE; ++bin_id) {
533    const int bin_offset = bin_id * bin_depth;
534    const int num_histos = bin_map[bin_offset];
535    const int idx1 = bin_map[bin_offset + 1];
536    int n;
537    for (n = 2; n <= num_histos; ++n) {
538      const int idx2 = bin_map[bin_offset + n];
539      const double bit_cost_idx2 = histograms[idx2]->bit_cost_;
540      if (bit_cost_idx2 > 0.) {
541        const double bit_cost_thresh = -bit_cost_idx2 * combine_cost_factor;
542        const double curr_cost_diff =
543            HistogramAddEval(histograms[idx1], histograms[idx2],
544                             cur_combo, bit_cost_thresh);
545        if (curr_cost_diff < bit_cost_thresh) {
546          HistogramCopy(cur_combo, histograms[idx1]);
547          histograms[idx2]->bit_cost_ = 0.;
548        }
549      }
550    }
551  }
552  HistogramCompactBins(image_histo);
553}
554
555static uint32_t MyRand(uint32_t *seed) {
556  *seed *= 16807U;
557  if (*seed == 0) {
558    *seed = 1;
559  }
560  return *seed;
561}
562
563static void HistogramCombine(VP8LHistogramSet* const image_histo,
564                             VP8LHistogramSet* const histos, int quality) {
565  int iter;
566  uint32_t seed = 0;
567  int tries_with_no_success = 0;
568  int image_histo_size = image_histo->size;
569  const int iter_mult = (quality < 25) ? 2 : 2 + (quality - 25) / 8;
570  const int outer_iters = image_histo_size * iter_mult;
571  const int num_pairs = image_histo_size / 2;
572  const int num_tries_no_success = outer_iters / 2;
573  const int min_cluster_size = 2;
574  VP8LHistogram** const histograms = image_histo->histograms;
575  VP8LHistogram* cur_combo = histos->histograms[0];   // trial histogram
576  VP8LHistogram* best_combo = histos->histograms[1];  // best histogram so far
577
578  // Collapse similar histograms in 'image_histo'.
579  for (iter = 0;
580       iter < outer_iters && image_histo_size >= min_cluster_size;
581       ++iter) {
582    double best_cost_diff = 0.;
583    int best_idx1 = -1, best_idx2 = 1;
584    int j;
585    const int num_tries =
586        (num_pairs < image_histo_size) ? num_pairs : image_histo_size;
587    seed += iter;
588    for (j = 0; j < num_tries; ++j) {
589      double curr_cost_diff;
590      // Choose two histograms at random and try to combine them.
591      const uint32_t idx1 = MyRand(&seed) % image_histo_size;
592      const uint32_t tmp = (j & 7) + 1;
593      const uint32_t diff =
594          (tmp < 3) ? tmp : MyRand(&seed) % (image_histo_size - 1);
595      const uint32_t idx2 = (idx1 + diff + 1) % image_histo_size;
596      if (idx1 == idx2) {
597        continue;
598      }
599
600      // Calculate cost reduction on combining.
601      curr_cost_diff = HistogramAddEval(histograms[idx1], histograms[idx2],
602                                        cur_combo, best_cost_diff);
603      if (curr_cost_diff < best_cost_diff) {    // found a better pair?
604        {     // swap cur/best combo histograms
605          VP8LHistogram* const tmp_histo = cur_combo;
606          cur_combo = best_combo;
607          best_combo = tmp_histo;
608        }
609        best_cost_diff = curr_cost_diff;
610        best_idx1 = idx1;
611        best_idx2 = idx2;
612      }
613    }
614
615    if (best_idx1 >= 0) {
616      HistogramCopy(best_combo, histograms[best_idx1]);
617      // swap best_idx2 slot with last one (which is now unused)
618      --image_histo_size;
619      if (best_idx2 != image_histo_size) {
620        HistogramCopy(histograms[image_histo_size], histograms[best_idx2]);
621        histograms[image_histo_size] = NULL;
622      }
623      tries_with_no_success = 0;
624    }
625    if (++tries_with_no_success >= num_tries_no_success) {
626      break;
627    }
628  }
629  image_histo->size = image_histo_size;
630}
631
632// -----------------------------------------------------------------------------
633// Histogram refinement
634
635// Find the best 'out' histogram for each of the 'in' histograms.
636// Note: we assume that out[]->bit_cost_ is already up-to-date.
637static void HistogramRemap(const VP8LHistogramSet* const orig_histo,
638                           const VP8LHistogramSet* const image_histo,
639                           uint16_t* const symbols) {
640  int i;
641  VP8LHistogram** const orig_histograms = orig_histo->histograms;
642  VP8LHistogram** const histograms = image_histo->histograms;
643  for (i = 0; i < orig_histo->size; ++i) {
644    int best_out = 0;
645    double best_bits =
646        HistogramAddThresh(histograms[0], orig_histograms[i], MAX_COST);
647    int k;
648    for (k = 1; k < image_histo->size; ++k) {
649      const double cur_bits =
650          HistogramAddThresh(histograms[k], orig_histograms[i], best_bits);
651      if (cur_bits < best_bits) {
652        best_bits = cur_bits;
653        best_out = k;
654      }
655    }
656    symbols[i] = best_out;
657  }
658
659  // Recompute each out based on raw and symbols.
660  for (i = 0; i < image_histo->size; ++i) {
661    HistogramClear(histograms[i]);
662  }
663
664  for (i = 0; i < orig_histo->size; ++i) {
665    const int idx = symbols[i];
666    VP8LHistogramAdd(orig_histograms[i], histograms[idx], histograms[idx]);
667  }
668}
669
670static double GetCombineCostFactor(int histo_size, int quality) {
671  double combine_cost_factor = 0.16;
672  if (histo_size > 256) combine_cost_factor /= 2.;
673  if (histo_size > 512) combine_cost_factor /= 2.;
674  if (histo_size > 1024) combine_cost_factor /= 2.;
675  if (quality <= 50) combine_cost_factor /= 2.;
676  return combine_cost_factor;
677}
678
679int VP8LGetHistoImageSymbols(int xsize, int ysize,
680                             const VP8LBackwardRefs* const refs,
681                             int quality, int histo_bits, int cache_bits,
682                             VP8LHistogramSet* const image_histo,
683                             uint16_t* const histogram_symbols) {
684  int ok = 0;
685  const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
686  const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
687  const int image_histo_raw_size = histo_xsize * histo_ysize;
688
689  // The bin_map for every bin follows following semantics:
690  // bin_map[n][0] = num_histo; // The number of histograms in that bin.
691  // bin_map[n][1] = index of first histogram in that bin;
692  // bin_map[n][num_histo] = index of last histogram in that bin;
693  // bin_map[n][num_histo + 1] ... bin_map[n][bin_depth - 1] = un-used indices.
694  const int bin_depth = image_histo_raw_size + 1;
695  int16_t* bin_map = NULL;
696  VP8LHistogramSet* const histos = VP8LAllocateHistogramSet(2, cache_bits);
697  VP8LHistogramSet* const orig_histo =
698      VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
699
700  if (orig_histo == NULL || histos == NULL) {
701    goto Error;
702  }
703
704  // Don't attempt linear bin-partition heuristic for:
705  // histograms of small sizes, as bin_map will be very sparse and;
706  // Higher qualities (> 90), to preserve the compression gains at those
707  // quality settings.
708  if (orig_histo->size > 2 * BIN_SIZE && quality < 90) {
709    const int bin_map_size = bin_depth * BIN_SIZE;
710    bin_map = (int16_t*)WebPSafeCalloc(bin_map_size, sizeof(*bin_map));
711    if (bin_map == NULL) goto Error;
712  }
713
714  // Construct the histograms from backward references.
715  HistogramBuild(xsize, histo_bits, refs, orig_histo);
716  // Copies the histograms and computes its bit_cost.
717  HistogramCopyAndAnalyze(orig_histo, image_histo);
718
719  if (bin_map != NULL) {
720    const double combine_cost_factor =
721        GetCombineCostFactor(image_histo_raw_size, quality);
722    HistogramAnalyzeEntropyBin(orig_histo, bin_map);
723    // Collapse histograms with similar entropy.
724    HistogramCombineEntropyBin(image_histo, histos->histograms[0],
725                               bin_map, bin_depth, combine_cost_factor);
726  }
727
728  // Collapse similar histograms by random histogram-pair compares.
729  HistogramCombine(image_histo, histos, quality);
730
731  // Find the optimal map from original histograms to the final ones.
732  HistogramRemap(orig_histo, image_histo, histogram_symbols);
733
734  ok = 1;
735
736 Error:
737  WebPSafeFree(bin_map);
738  VP8LFreeHistogramSet(orig_histo);
739  VP8LFreeHistogramSet(histos);
740  return ok;
741}
742