• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 "src/webp/config.h"
14 #endif
15 
16 #include <float.h>
17 #include <math.h>
18 
19 #include "src/dsp/lossless.h"
20 #include "src/dsp/lossless_common.h"
21 #include "src/enc/backward_references_enc.h"
22 #include "src/enc/histogram_enc.h"
23 #include "src/enc/vp8i_enc.h"
24 #include "src/utils/utils.h"
25 
26 #define MAX_BIT_COST FLT_MAX
27 
28 // Number of partitions for the three dominant (literal, red and blue) symbol
29 // costs.
30 #define NUM_PARTITIONS 4
31 // The size of the bin-hash corresponding to the three dominant costs.
32 #define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS)
33 // Maximum number of histograms allowed in greedy combining algorithm.
34 #define MAX_HISTO_GREEDY 100
35 
HistogramClear(VP8LHistogram * const p)36 static void HistogramClear(VP8LHistogram* const p) {
37   uint32_t* const literal = p->literal_;
38   const int cache_bits = p->palette_code_bits_;
39   const int histo_size = VP8LGetHistogramSize(cache_bits);
40   memset(p, 0, histo_size);
41   p->palette_code_bits_ = cache_bits;
42   p->literal_ = literal;
43 }
44 
45 // Swap two histogram pointers.
HistogramSwap(VP8LHistogram ** const A,VP8LHistogram ** const B)46 static void HistogramSwap(VP8LHistogram** const A, VP8LHistogram** const B) {
47   VP8LHistogram* const tmp = *A;
48   *A = *B;
49   *B = tmp;
50 }
51 
HistogramCopy(const VP8LHistogram * const src,VP8LHistogram * const dst)52 static void HistogramCopy(const VP8LHistogram* const src,
53                           VP8LHistogram* const dst) {
54   uint32_t* const dst_literal = dst->literal_;
55   const int dst_cache_bits = dst->palette_code_bits_;
56   const int literal_size = VP8LHistogramNumCodes(dst_cache_bits);
57   const int histo_size = VP8LGetHistogramSize(dst_cache_bits);
58   assert(src->palette_code_bits_ == dst_cache_bits);
59   memcpy(dst, src, histo_size);
60   dst->literal_ = dst_literal;
61   memcpy(dst->literal_, src->literal_, literal_size * sizeof(*dst->literal_));
62 }
63 
VP8LGetHistogramSize(int cache_bits)64 int VP8LGetHistogramSize(int cache_bits) {
65   const int literal_size = VP8LHistogramNumCodes(cache_bits);
66   const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size;
67   assert(total_size <= (size_t)0x7fffffff);
68   return (int)total_size;
69 }
70 
VP8LFreeHistogram(VP8LHistogram * const histo)71 void VP8LFreeHistogram(VP8LHistogram* const histo) {
72   WebPSafeFree(histo);
73 }
74 
VP8LFreeHistogramSet(VP8LHistogramSet * const histo)75 void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
76   WebPSafeFree(histo);
77 }
78 
VP8LHistogramStoreRefs(const VP8LBackwardRefs * const refs,VP8LHistogram * const histo)79 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
80                             VP8LHistogram* const histo) {
81   VP8LRefsCursor c = VP8LRefsCursorInit(refs);
82   while (VP8LRefsCursorOk(&c)) {
83     VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos, NULL, 0);
84     VP8LRefsCursorNext(&c);
85   }
86 }
87 
VP8LHistogramCreate(VP8LHistogram * const p,const VP8LBackwardRefs * const refs,int palette_code_bits)88 void VP8LHistogramCreate(VP8LHistogram* const p,
89                          const VP8LBackwardRefs* const refs,
90                          int palette_code_bits) {
91   if (palette_code_bits >= 0) {
92     p->palette_code_bits_ = palette_code_bits;
93   }
94   HistogramClear(p);
95   VP8LHistogramStoreRefs(refs, p);
96 }
97 
VP8LHistogramInit(VP8LHistogram * const p,int palette_code_bits,int init_arrays)98 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits,
99                        int init_arrays) {
100   p->palette_code_bits_ = palette_code_bits;
101   if (init_arrays) {
102     HistogramClear(p);
103   } else {
104     p->trivial_symbol_ = 0;
105     p->bit_cost_ = 0.;
106     p->literal_cost_ = 0.;
107     p->red_cost_ = 0.;
108     p->blue_cost_ = 0.;
109     memset(p->is_used_, 0, sizeof(p->is_used_));
110   }
111 }
112 
VP8LAllocateHistogram(int cache_bits)113 VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
114   VP8LHistogram* histo = NULL;
115   const int total_size = VP8LGetHistogramSize(cache_bits);
116   uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
117   if (memory == NULL) return NULL;
118   histo = (VP8LHistogram*)memory;
119   // literal_ won't necessary be aligned.
120   histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
121   VP8LHistogramInit(histo, cache_bits, /*init_arrays=*/ 0);
122   return histo;
123 }
124 
125 // Resets the pointers of the histograms to point to the bit buffer in the set.
HistogramSetResetPointers(VP8LHistogramSet * const set,int cache_bits)126 static void HistogramSetResetPointers(VP8LHistogramSet* const set,
127                                       int cache_bits) {
128   int i;
129   const int histo_size = VP8LGetHistogramSize(cache_bits);
130   uint8_t* memory = (uint8_t*) (set->histograms);
131   memory += set->max_size * sizeof(*set->histograms);
132   for (i = 0; i < set->max_size; ++i) {
133     memory = (uint8_t*) WEBP_ALIGN(memory);
134     set->histograms[i] = (VP8LHistogram*) memory;
135     // literal_ won't necessary be aligned.
136     set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
137     memory += histo_size;
138   }
139 }
140 
141 // Returns the total size of the VP8LHistogramSet.
HistogramSetTotalSize(int size,int cache_bits)142 static size_t HistogramSetTotalSize(int size, int cache_bits) {
143   const int histo_size = VP8LGetHistogramSize(cache_bits);
144   return (sizeof(VP8LHistogramSet) + size * (sizeof(VP8LHistogram*) +
145           histo_size + WEBP_ALIGN_CST));
146 }
147 
VP8LAllocateHistogramSet(int size,int cache_bits)148 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
149   int i;
150   VP8LHistogramSet* set;
151   const size_t total_size = HistogramSetTotalSize(size, cache_bits);
152   uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
153   if (memory == NULL) return NULL;
154 
155   set = (VP8LHistogramSet*)memory;
156   memory += sizeof(*set);
157   set->histograms = (VP8LHistogram**)memory;
158   set->max_size = size;
159   set->size = size;
160   HistogramSetResetPointers(set, cache_bits);
161   for (i = 0; i < size; ++i) {
162     VP8LHistogramInit(set->histograms[i], cache_bits, /*init_arrays=*/ 0);
163   }
164   return set;
165 }
166 
VP8LHistogramSetClear(VP8LHistogramSet * const set)167 void VP8LHistogramSetClear(VP8LHistogramSet* const set) {
168   int i;
169   const int cache_bits = set->histograms[0]->palette_code_bits_;
170   const int size = set->max_size;
171   const size_t total_size = HistogramSetTotalSize(size, cache_bits);
172   uint8_t* memory = (uint8_t*)set;
173 
174   memset(memory, 0, total_size);
175   memory += sizeof(*set);
176   set->histograms = (VP8LHistogram**)memory;
177   set->max_size = size;
178   set->size = size;
179   HistogramSetResetPointers(set, cache_bits);
180   for (i = 0; i < size; ++i) {
181     set->histograms[i]->palette_code_bits_ = cache_bits;
182   }
183 }
184 
185 // Removes the histogram 'i' from 'set' by setting it to NULL.
HistogramSetRemoveHistogram(VP8LHistogramSet * const set,int i,int * const num_used)186 static void HistogramSetRemoveHistogram(VP8LHistogramSet* const set, int i,
187                                         int* const num_used) {
188   assert(set->histograms[i] != NULL);
189   set->histograms[i] = NULL;
190   --*num_used;
191   // If we remove the last valid one, shrink until the next valid one.
192   if (i == set->size - 1) {
193     while (set->size >= 1 && set->histograms[set->size - 1] == NULL) {
194       --set->size;
195     }
196   }
197 }
198 
199 // -----------------------------------------------------------------------------
200 
VP8LHistogramAddSinglePixOrCopy(VP8LHistogram * const histo,const PixOrCopy * const v,int (* const distance_modifier)(int,int),int distance_modifier_arg0)201 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
202                                      const PixOrCopy* const v,
203                                      int (*const distance_modifier)(int, int),
204                                      int distance_modifier_arg0) {
205   if (PixOrCopyIsLiteral(v)) {
206     ++histo->alpha_[PixOrCopyLiteral(v, 3)];
207     ++histo->red_[PixOrCopyLiteral(v, 2)];
208     ++histo->literal_[PixOrCopyLiteral(v, 1)];
209     ++histo->blue_[PixOrCopyLiteral(v, 0)];
210   } else if (PixOrCopyIsCacheIdx(v)) {
211     const int literal_ix =
212         NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
213     assert(histo->palette_code_bits_ != 0);
214     ++histo->literal_[literal_ix];
215   } else {
216     int code, extra_bits;
217     VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
218     ++histo->literal_[NUM_LITERAL_CODES + code];
219     if (distance_modifier == NULL) {
220       VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
221     } else {
222       VP8LPrefixEncodeBits(
223           distance_modifier(distance_modifier_arg0, PixOrCopyDistance(v)),
224           &code, &extra_bits);
225     }
226     ++histo->distance_[code];
227   }
228 }
229 
230 // -----------------------------------------------------------------------------
231 // Entropy-related functions.
232 
BitsEntropyRefine(const VP8LBitEntropy * entropy)233 static WEBP_INLINE float BitsEntropyRefine(const VP8LBitEntropy* entropy) {
234   float mix;
235   if (entropy->nonzeros < 5) {
236     if (entropy->nonzeros <= 1) {
237       return 0;
238     }
239     // Two symbols, they will be 0 and 1 in a Huffman code.
240     // Let's mix in a bit of entropy to favor good clustering when
241     // distributions of these are combined.
242     if (entropy->nonzeros == 2) {
243       return 0.99f * entropy->sum + 0.01f * entropy->entropy;
244     }
245     // No matter what the entropy says, we cannot be better than min_limit
246     // with Huffman coding. I am mixing a bit of entropy into the
247     // min_limit since it produces much better (~0.5 %) compression results
248     // perhaps because of better entropy clustering.
249     if (entropy->nonzeros == 3) {
250       mix = 0.95f;
251     } else {
252       mix = 0.7f;  // nonzeros == 4.
253     }
254   } else {
255     mix = 0.627f;
256   }
257 
258   {
259     float min_limit = 2.f * entropy->sum - entropy->max_val;
260     min_limit = mix * min_limit + (1.f - mix) * entropy->entropy;
261     return (entropy->entropy < min_limit) ? min_limit : entropy->entropy;
262   }
263 }
264 
VP8LBitsEntropy(const uint32_t * const array,int n)265 float VP8LBitsEntropy(const uint32_t* const array, int n) {
266   VP8LBitEntropy entropy;
267   VP8LBitsEntropyUnrefined(array, n, &entropy);
268 
269   return BitsEntropyRefine(&entropy);
270 }
271 
InitialHuffmanCost(void)272 static float InitialHuffmanCost(void) {
273   // Small bias because Huffman code length is typically not stored in
274   // full length.
275   static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
276   static const float kSmallBias = 9.1f;
277   return kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
278 }
279 
280 // Finalize the Huffman cost based on streak numbers and length type (<3 or >=3)
FinalHuffmanCost(const VP8LStreaks * const stats)281 static float FinalHuffmanCost(const VP8LStreaks* const stats) {
282   // The constants in this function are experimental and got rounded from
283   // their original values in 1/8 when switched to 1/1024.
284   float retval = InitialHuffmanCost();
285   // Second coefficient: Many zeros in the histogram are covered efficiently
286   // by a run-length encode. Originally 2/8.
287   retval += stats->counts[0] * 1.5625f + 0.234375f * stats->streaks[0][1];
288   // Second coefficient: Constant values are encoded less efficiently, but still
289   // RLE'ed. Originally 6/8.
290   retval += stats->counts[1] * 2.578125f + 0.703125f * stats->streaks[1][1];
291   // 0s are usually encoded more efficiently than non-0s.
292   // Originally 15/8.
293   retval += 1.796875f * stats->streaks[0][0];
294   // Originally 26/8.
295   retval += 3.28125f * stats->streaks[1][0];
296   return retval;
297 }
298 
299 // Get the symbol entropy for the distribution 'population'.
300 // Set 'trivial_sym', if there's only one symbol present in the distribution.
PopulationCost(const uint32_t * const population,int length,uint32_t * const trivial_sym,uint8_t * const is_used)301 static float PopulationCost(const uint32_t* const population, int length,
302                             uint32_t* const trivial_sym,
303                             uint8_t* const is_used) {
304   VP8LBitEntropy bit_entropy;
305   VP8LStreaks stats;
306   VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats);
307   if (trivial_sym != NULL) {
308     *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code
309                                                : VP8L_NON_TRIVIAL_SYM;
310   }
311   // The histogram is used if there is at least one non-zero streak.
312   *is_used = (stats.streaks[1][0] != 0 || stats.streaks[1][1] != 0);
313 
314   return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
315 }
316 
317 // trivial_at_end is 1 if the two histograms only have one element that is
318 // non-zero: both the zero-th one, or both the last one.
GetCombinedEntropy(const uint32_t * const X,const uint32_t * const Y,int length,int is_X_used,int is_Y_used,int trivial_at_end)319 static WEBP_INLINE float GetCombinedEntropy(const uint32_t* const X,
320                                             const uint32_t* const Y, int length,
321                                             int is_X_used, int is_Y_used,
322                                             int trivial_at_end) {
323   VP8LStreaks stats;
324   if (trivial_at_end) {
325     // This configuration is due to palettization that transforms an indexed
326     // pixel into 0xff000000 | (pixel << 8) in VP8LBundleColorMap.
327     // BitsEntropyRefine is 0 for histograms with only one non-zero value.
328     // Only FinalHuffmanCost needs to be evaluated.
329     memset(&stats, 0, sizeof(stats));
330     // Deal with the non-zero value at index 0 or length-1.
331     stats.streaks[1][0] = 1;
332     // Deal with the following/previous zero streak.
333     stats.counts[0] = 1;
334     stats.streaks[0][1] = length - 1;
335     return FinalHuffmanCost(&stats);
336   } else {
337     VP8LBitEntropy bit_entropy;
338     if (is_X_used) {
339       if (is_Y_used) {
340         VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats);
341       } else {
342         VP8LGetEntropyUnrefined(X, length, &bit_entropy, &stats);
343       }
344     } else {
345       if (is_Y_used) {
346         VP8LGetEntropyUnrefined(Y, length, &bit_entropy, &stats);
347       } else {
348         memset(&stats, 0, sizeof(stats));
349         stats.counts[0] = 1;
350         stats.streaks[0][length > 3] = length;
351         VP8LBitEntropyInit(&bit_entropy);
352       }
353     }
354 
355     return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
356   }
357 }
358 
359 // Estimates the Entropy + Huffman + other block overhead size cost.
VP8LHistogramEstimateBits(VP8LHistogram * const p)360 float VP8LHistogramEstimateBits(VP8LHistogram* const p) {
361   return
362       PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_),
363                      NULL, &p->is_used_[0])
364       + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL, &p->is_used_[1])
365       + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL, &p->is_used_[2])
366       + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL, &p->is_used_[3])
367       + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL, &p->is_used_[4])
368       + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
369       + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
370 }
371 
372 // -----------------------------------------------------------------------------
373 // Various histogram combine/cost-eval functions
374 
GetCombinedHistogramEntropy(const VP8LHistogram * const a,const VP8LHistogram * const b,float cost_threshold,float * cost)375 static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
376                                        const VP8LHistogram* const b,
377                                        float cost_threshold, float* cost) {
378   const int palette_code_bits = a->palette_code_bits_;
379   int trivial_at_end = 0;
380   assert(a->palette_code_bits_ == b->palette_code_bits_);
381   *cost += GetCombinedEntropy(a->literal_, b->literal_,
382                               VP8LHistogramNumCodes(palette_code_bits),
383                               a->is_used_[0], b->is_used_[0], 0);
384   *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
385                                  b->literal_ + NUM_LITERAL_CODES,
386                                  NUM_LENGTH_CODES);
387   if (*cost > cost_threshold) return 0;
388 
389   if (a->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM &&
390       a->trivial_symbol_ == b->trivial_symbol_) {
391     // A, R and B are all 0 or 0xff.
392     const uint32_t color_a = (a->trivial_symbol_ >> 24) & 0xff;
393     const uint32_t color_r = (a->trivial_symbol_ >> 16) & 0xff;
394     const uint32_t color_b = (a->trivial_symbol_ >> 0) & 0xff;
395     if ((color_a == 0 || color_a == 0xff) &&
396         (color_r == 0 || color_r == 0xff) &&
397         (color_b == 0 || color_b == 0xff)) {
398       trivial_at_end = 1;
399     }
400   }
401 
402   *cost +=
403       GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, a->is_used_[1],
404                          b->is_used_[1], trivial_at_end);
405   if (*cost > cost_threshold) return 0;
406 
407   *cost +=
408       GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, a->is_used_[2],
409                          b->is_used_[2], trivial_at_end);
410   if (*cost > cost_threshold) return 0;
411 
412   *cost +=
413       GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES,
414                          a->is_used_[3], b->is_used_[3], trivial_at_end);
415   if (*cost > cost_threshold) return 0;
416 
417   *cost +=
418       GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES,
419                          a->is_used_[4], b->is_used_[4], 0);
420   *cost +=
421       VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES);
422   if (*cost > cost_threshold) return 0;
423 
424   return 1;
425 }
426 
HistogramAdd(const VP8LHistogram * const a,const VP8LHistogram * const b,VP8LHistogram * const out)427 static WEBP_INLINE void HistogramAdd(const VP8LHistogram* const a,
428                                      const VP8LHistogram* const b,
429                                      VP8LHistogram* const out) {
430   VP8LHistogramAdd(a, b, out);
431   out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_)
432                        ? a->trivial_symbol_
433                        : VP8L_NON_TRIVIAL_SYM;
434 }
435 
436 // Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing
437 // to the threshold value 'cost_threshold'. The score returned is
438 //  Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed.
439 // Since the previous score passed is 'cost_threshold', we only need to compare
440 // the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out
441 // early.
HistogramAddEval(const VP8LHistogram * const a,const VP8LHistogram * const b,VP8LHistogram * const out,float cost_threshold)442 static float HistogramAddEval(const VP8LHistogram* const a,
443                               const VP8LHistogram* const b,
444                               VP8LHistogram* const out, float cost_threshold) {
445   float cost = 0;
446   const float sum_cost = a->bit_cost_ + b->bit_cost_;
447   cost_threshold += sum_cost;
448 
449   if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
450     HistogramAdd(a, b, out);
451     out->bit_cost_ = cost;
452     out->palette_code_bits_ = a->palette_code_bits_;
453   }
454 
455   return cost - sum_cost;
456 }
457 
458 // Same as HistogramAddEval(), except that the resulting histogram
459 // is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit
460 // the term C(b) which is constant over all the evaluations.
HistogramAddThresh(const VP8LHistogram * const a,const VP8LHistogram * const b,float cost_threshold)461 static float HistogramAddThresh(const VP8LHistogram* const a,
462                                 const VP8LHistogram* const b,
463                                 float cost_threshold) {
464   float cost;
465   assert(a != NULL && b != NULL);
466   cost = -a->bit_cost_;
467   GetCombinedHistogramEntropy(a, b, cost_threshold, &cost);
468   return cost;
469 }
470 
471 // -----------------------------------------------------------------------------
472 
473 // The structure to keep track of cost range for the three dominant entropy
474 // symbols.
475 typedef struct {
476   float literal_max_;
477   float literal_min_;
478   float red_max_;
479   float red_min_;
480   float blue_max_;
481   float blue_min_;
482 } DominantCostRange;
483 
DominantCostRangeInit(DominantCostRange * const c)484 static void DominantCostRangeInit(DominantCostRange* const c) {
485   c->literal_max_ = 0.;
486   c->literal_min_ = MAX_BIT_COST;
487   c->red_max_ = 0.;
488   c->red_min_ = MAX_BIT_COST;
489   c->blue_max_ = 0.;
490   c->blue_min_ = MAX_BIT_COST;
491 }
492 
UpdateDominantCostRange(const VP8LHistogram * const h,DominantCostRange * const c)493 static void UpdateDominantCostRange(
494     const VP8LHistogram* const h, DominantCostRange* const c) {
495   if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_;
496   if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_;
497   if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_;
498   if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_;
499   if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_;
500   if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_;
501 }
502 
UpdateHistogramCost(VP8LHistogram * const h)503 static void UpdateHistogramCost(VP8LHistogram* const h) {
504   uint32_t alpha_sym, red_sym, blue_sym;
505   const float alpha_cost =
506       PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym, &h->is_used_[3]);
507   const float distance_cost =
508       PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL, &h->is_used_[4]) +
509       VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
510   const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
511   h->literal_cost_ =
512       PopulationCost(h->literal_, num_codes, NULL, &h->is_used_[0]) +
513           VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES);
514   h->red_cost_ =
515       PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym, &h->is_used_[1]);
516   h->blue_cost_ =
517       PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym, &h->is_used_[2]);
518   h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
519                  alpha_cost + distance_cost;
520   if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) {
521     h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM;
522   } else {
523     h->trivial_symbol_ =
524         ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0);
525   }
526 }
527 
GetBinIdForEntropy(float min,float max,float val)528 static int GetBinIdForEntropy(float min, float max, float val) {
529   const float range = max - min;
530   if (range > 0.) {
531     const float delta = val - min;
532     return (int)((NUM_PARTITIONS - 1e-6) * delta / range);
533   } else {
534     return 0;
535   }
536 }
537 
GetHistoBinIndex(const VP8LHistogram * const h,const DominantCostRange * const c,int low_effort)538 static int GetHistoBinIndex(const VP8LHistogram* const h,
539                             const DominantCostRange* const c, int low_effort) {
540   int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_,
541                                   h->literal_cost_);
542   assert(bin_id < NUM_PARTITIONS);
543   if (!low_effort) {
544     bin_id = bin_id * NUM_PARTITIONS
545            + GetBinIdForEntropy(c->red_min_, c->red_max_, h->red_cost_);
546     bin_id = bin_id * NUM_PARTITIONS
547            + GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_);
548     assert(bin_id < BIN_SIZE);
549   }
550   return bin_id;
551 }
552 
553 // Construct the histograms from backward references.
HistogramBuild(int xsize,int histo_bits,const VP8LBackwardRefs * const backward_refs,VP8LHistogramSet * const image_histo)554 static void HistogramBuild(
555     int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs,
556     VP8LHistogramSet* const image_histo) {
557   int x = 0, y = 0;
558   const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
559   VP8LHistogram** const histograms = image_histo->histograms;
560   VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs);
561   assert(histo_bits > 0);
562   VP8LHistogramSetClear(image_histo);
563   while (VP8LRefsCursorOk(&c)) {
564     const PixOrCopy* const v = c.cur_pos;
565     const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
566     VP8LHistogramAddSinglePixOrCopy(histograms[ix], v, NULL, 0);
567     x += PixOrCopyLength(v);
568     while (x >= xsize) {
569       x -= xsize;
570       ++y;
571     }
572     VP8LRefsCursorNext(&c);
573   }
574 }
575 
576 // Copies the histograms and computes its bit_cost.
577 static const uint16_t kInvalidHistogramSymbol = (uint16_t)(-1);
HistogramCopyAndAnalyze(VP8LHistogramSet * const orig_histo,VP8LHistogramSet * const image_histo,int * const num_used,uint16_t * const histogram_symbols)578 static void HistogramCopyAndAnalyze(VP8LHistogramSet* const orig_histo,
579                                     VP8LHistogramSet* const image_histo,
580                                     int* const num_used,
581                                     uint16_t* const histogram_symbols) {
582   int i, cluster_id;
583   int num_used_orig = *num_used;
584   VP8LHistogram** const orig_histograms = orig_histo->histograms;
585   VP8LHistogram** const histograms = image_histo->histograms;
586   assert(image_histo->max_size == orig_histo->max_size);
587   for (cluster_id = 0, i = 0; i < orig_histo->max_size; ++i) {
588     VP8LHistogram* const histo = orig_histograms[i];
589     UpdateHistogramCost(histo);
590 
591     // Skip the histogram if it is completely empty, which can happen for tiles
592     // with no information (when they are skipped because of LZ77).
593     if (!histo->is_used_[0] && !histo->is_used_[1] && !histo->is_used_[2]
594         && !histo->is_used_[3] && !histo->is_used_[4]) {
595       // The first histogram is always used. If an histogram is empty, we set
596       // its id to be the same as the previous one: this will improve
597       // compressibility for later LZ77.
598       assert(i > 0);
599       HistogramSetRemoveHistogram(image_histo, i, num_used);
600       HistogramSetRemoveHistogram(orig_histo, i, &num_used_orig);
601       histogram_symbols[i] = kInvalidHistogramSymbol;
602     } else {
603       // Copy histograms from orig_histo[] to image_histo[].
604       HistogramCopy(histo, histograms[i]);
605       histogram_symbols[i] = cluster_id++;
606       assert(cluster_id <= image_histo->max_size);
607     }
608   }
609 }
610 
611 // Partition histograms to different entropy bins for three dominant (literal,
612 // red and blue) symbol costs and compute the histogram aggregate bit_cost.
HistogramAnalyzeEntropyBin(VP8LHistogramSet * const image_histo,uint16_t * const bin_map,int low_effort)613 static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo,
614                                        uint16_t* const bin_map,
615                                        int low_effort) {
616   int i;
617   VP8LHistogram** const histograms = image_histo->histograms;
618   const int histo_size = image_histo->size;
619   DominantCostRange cost_range;
620   DominantCostRangeInit(&cost_range);
621 
622   // Analyze the dominant (literal, red and blue) entropy costs.
623   for (i = 0; i < histo_size; ++i) {
624     if (histograms[i] == NULL) continue;
625     UpdateDominantCostRange(histograms[i], &cost_range);
626   }
627 
628   // bin-hash histograms on three of the dominant (literal, red and blue)
629   // symbol costs and store the resulting bin_id for each histogram.
630   for (i = 0; i < histo_size; ++i) {
631     // bin_map[i] is not set to a special value as its use will later be guarded
632     // by another (histograms[i] == NULL).
633     if (histograms[i] == NULL) continue;
634     bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort);
635   }
636 }
637 
638 // Merges some histograms with same bin_id together if it's advantageous.
639 // Sets the remaining histograms to NULL.
HistogramCombineEntropyBin(VP8LHistogramSet * const image_histo,int * num_used,const uint16_t * const clusters,uint16_t * const cluster_mappings,VP8LHistogram * cur_combo,const uint16_t * const bin_map,int num_bins,float combine_cost_factor,int low_effort)640 static void HistogramCombineEntropyBin(
641     VP8LHistogramSet* const image_histo, int* num_used,
642     const uint16_t* const clusters, uint16_t* const cluster_mappings,
643     VP8LHistogram* cur_combo, const uint16_t* const bin_map, int num_bins,
644     float combine_cost_factor, int low_effort) {
645   VP8LHistogram** const histograms = image_histo->histograms;
646   int idx;
647   struct {
648     int16_t first;    // position of the histogram that accumulates all
649                       // histograms with the same bin_id
650     uint16_t num_combine_failures;   // number of combine failures per bin_id
651   } bin_info[BIN_SIZE];
652 
653   assert(num_bins <= BIN_SIZE);
654   for (idx = 0; idx < num_bins; ++idx) {
655     bin_info[idx].first = -1;
656     bin_info[idx].num_combine_failures = 0;
657   }
658 
659   // By default, a cluster matches itself.
660   for (idx = 0; idx < *num_used; ++idx) cluster_mappings[idx] = idx;
661   for (idx = 0; idx < image_histo->size; ++idx) {
662     int bin_id, first;
663     if (histograms[idx] == NULL) continue;
664     bin_id = bin_map[idx];
665     first = bin_info[bin_id].first;
666     if (first == -1) {
667       bin_info[bin_id].first = idx;
668     } else if (low_effort) {
669       HistogramAdd(histograms[idx], histograms[first], histograms[first]);
670       HistogramSetRemoveHistogram(image_histo, idx, num_used);
671       cluster_mappings[clusters[idx]] = clusters[first];
672     } else {
673       // try to merge #idx into #first (both share the same bin_id)
674       const float bit_cost = histograms[idx]->bit_cost_;
675       const float bit_cost_thresh = -bit_cost * combine_cost_factor;
676       const float curr_cost_diff = HistogramAddEval(
677           histograms[first], histograms[idx], cur_combo, bit_cost_thresh);
678       if (curr_cost_diff < bit_cost_thresh) {
679         // Try to merge two histograms only if the combo is a trivial one or
680         // the two candidate histograms are already non-trivial.
681         // For some images, 'try_combine' turns out to be false for a lot of
682         // histogram pairs. In that case, we fallback to combining
683         // histograms as usual to avoid increasing the header size.
684         const int try_combine =
685             (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) ||
686             ((histograms[idx]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) &&
687              (histograms[first]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM));
688         const int max_combine_failures = 32;
689         if (try_combine ||
690             bin_info[bin_id].num_combine_failures >= max_combine_failures) {
691           // move the (better) merged histogram to its final slot
692           HistogramSwap(&cur_combo, &histograms[first]);
693           HistogramSetRemoveHistogram(image_histo, idx, num_used);
694           cluster_mappings[clusters[idx]] = clusters[first];
695         } else {
696           ++bin_info[bin_id].num_combine_failures;
697         }
698       }
699     }
700   }
701   if (low_effort) {
702     // for low_effort case, update the final cost when everything is merged
703     for (idx = 0; idx < image_histo->size; ++idx) {
704       if (histograms[idx] == NULL) continue;
705       UpdateHistogramCost(histograms[idx]);
706     }
707   }
708 }
709 
710 // Implement a Lehmer random number generator with a multiplicative constant of
711 // 48271 and a modulo constant of 2^31 - 1.
MyRand(uint32_t * const seed)712 static uint32_t MyRand(uint32_t* const seed) {
713   *seed = (uint32_t)(((uint64_t)(*seed) * 48271u) % 2147483647u);
714   assert(*seed > 0);
715   return *seed;
716 }
717 
718 // -----------------------------------------------------------------------------
719 // Histogram pairs priority queue
720 
721 // Pair of histograms. Negative idx1 value means that pair is out-of-date.
722 typedef struct {
723   int idx1;
724   int idx2;
725   float cost_diff;
726   float cost_combo;
727 } HistogramPair;
728 
729 typedef struct {
730   HistogramPair* queue;
731   int size;
732   int max_size;
733 } HistoQueue;
734 
HistoQueueInit(HistoQueue * const histo_queue,const int max_size)735 static int HistoQueueInit(HistoQueue* const histo_queue, const int max_size) {
736   histo_queue->size = 0;
737   histo_queue->max_size = max_size;
738   // We allocate max_size + 1 because the last element at index "size" is
739   // used as temporary data (and it could be up to max_size).
740   histo_queue->queue = (HistogramPair*)WebPSafeMalloc(
741       histo_queue->max_size + 1, sizeof(*histo_queue->queue));
742   return histo_queue->queue != NULL;
743 }
744 
HistoQueueClear(HistoQueue * const histo_queue)745 static void HistoQueueClear(HistoQueue* const histo_queue) {
746   assert(histo_queue != NULL);
747   WebPSafeFree(histo_queue->queue);
748   histo_queue->size = 0;
749   histo_queue->max_size = 0;
750 }
751 
752 // Pop a specific pair in the queue by replacing it with the last one
753 // and shrinking the queue.
HistoQueuePopPair(HistoQueue * const histo_queue,HistogramPair * const pair)754 static void HistoQueuePopPair(HistoQueue* const histo_queue,
755                               HistogramPair* const pair) {
756   assert(pair >= histo_queue->queue &&
757          pair < (histo_queue->queue + histo_queue->size));
758   assert(histo_queue->size > 0);
759   *pair = histo_queue->queue[histo_queue->size - 1];
760   --histo_queue->size;
761 }
762 
763 // Check whether a pair in the queue should be updated as head or not.
HistoQueueUpdateHead(HistoQueue * const histo_queue,HistogramPair * const pair)764 static void HistoQueueUpdateHead(HistoQueue* const histo_queue,
765                                  HistogramPair* const pair) {
766   assert(pair->cost_diff < 0.);
767   assert(pair >= histo_queue->queue &&
768          pair < (histo_queue->queue + histo_queue->size));
769   assert(histo_queue->size > 0);
770   if (pair->cost_diff < histo_queue->queue[0].cost_diff) {
771     // Replace the best pair.
772     const HistogramPair tmp = histo_queue->queue[0];
773     histo_queue->queue[0] = *pair;
774     *pair = tmp;
775   }
776 }
777 
778 // Update the cost diff and combo of a pair of histograms. This needs to be
779 // called when the the histograms have been merged with a third one.
HistoQueueUpdatePair(const VP8LHistogram * const h1,const VP8LHistogram * const h2,float threshold,HistogramPair * const pair)780 static void HistoQueueUpdatePair(const VP8LHistogram* const h1,
781                                  const VP8LHistogram* const h2, float threshold,
782                                  HistogramPair* const pair) {
783   const float sum_cost = h1->bit_cost_ + h2->bit_cost_;
784   pair->cost_combo = 0.;
785   GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair->cost_combo);
786   pair->cost_diff = pair->cost_combo - sum_cost;
787 }
788 
789 // Create a pair from indices "idx1" and "idx2" provided its cost
790 // is inferior to "threshold", a negative entropy.
791 // It returns the cost of the pair, or 0. if it superior to threshold.
HistoQueuePush(HistoQueue * const histo_queue,VP8LHistogram ** const histograms,int idx1,int idx2,float threshold)792 static float HistoQueuePush(HistoQueue* const histo_queue,
793                             VP8LHistogram** const histograms, int idx1,
794                             int idx2, float threshold) {
795   const VP8LHistogram* h1;
796   const VP8LHistogram* h2;
797   HistogramPair pair;
798 
799   // Stop here if the queue is full.
800   if (histo_queue->size == histo_queue->max_size) return 0.;
801   assert(threshold <= 0.);
802   if (idx1 > idx2) {
803     const int tmp = idx2;
804     idx2 = idx1;
805     idx1 = tmp;
806   }
807   pair.idx1 = idx1;
808   pair.idx2 = idx2;
809   h1 = histograms[idx1];
810   h2 = histograms[idx2];
811 
812   HistoQueueUpdatePair(h1, h2, threshold, &pair);
813 
814   // Do not even consider the pair if it does not improve the entropy.
815   if (pair.cost_diff >= threshold) return 0.;
816 
817   histo_queue->queue[histo_queue->size++] = pair;
818   HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]);
819 
820   return pair.cost_diff;
821 }
822 
823 // -----------------------------------------------------------------------------
824 
825 // Combines histograms by continuously choosing the one with the highest cost
826 // reduction.
HistogramCombineGreedy(VP8LHistogramSet * const image_histo,int * const num_used)827 static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo,
828                                   int* const num_used) {
829   int ok = 0;
830   const int image_histo_size = image_histo->size;
831   int i, j;
832   VP8LHistogram** const histograms = image_histo->histograms;
833   // Priority queue of histogram pairs.
834   HistoQueue histo_queue;
835 
836   // image_histo_size^2 for the queue size is safe. If you look at
837   // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes
838   // data to the queue, you insert at most:
839   // - image_histo_size*(image_histo_size-1)/2 (the first two for loops)
840   // - image_histo_size - 1 in the last for loop at the first iteration of
841   //   the while loop, image_histo_size - 2 at the second iteration ...
842   //   therefore image_histo_size*(image_histo_size-1)/2 overall too
843   if (!HistoQueueInit(&histo_queue, image_histo_size * image_histo_size)) {
844     goto End;
845   }
846 
847   for (i = 0; i < image_histo_size; ++i) {
848     if (image_histo->histograms[i] == NULL) continue;
849     for (j = i + 1; j < image_histo_size; ++j) {
850       // Initialize queue.
851       if (image_histo->histograms[j] == NULL) continue;
852       HistoQueuePush(&histo_queue, histograms, i, j, 0.);
853     }
854   }
855 
856   while (histo_queue.size > 0) {
857     const int idx1 = histo_queue.queue[0].idx1;
858     const int idx2 = histo_queue.queue[0].idx2;
859     HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]);
860     histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
861 
862     // Remove merged histogram.
863     HistogramSetRemoveHistogram(image_histo, idx2, num_used);
864 
865     // Remove pairs intersecting the just combined best pair.
866     for (i = 0; i < histo_queue.size;) {
867       HistogramPair* const p = histo_queue.queue + i;
868       if (p->idx1 == idx1 || p->idx2 == idx1 ||
869           p->idx1 == idx2 || p->idx2 == idx2) {
870         HistoQueuePopPair(&histo_queue, p);
871       } else {
872         HistoQueueUpdateHead(&histo_queue, p);
873         ++i;
874       }
875     }
876 
877     // Push new pairs formed with combined histogram to the queue.
878     for (i = 0; i < image_histo->size; ++i) {
879       if (i == idx1 || image_histo->histograms[i] == NULL) continue;
880       HistoQueuePush(&histo_queue, image_histo->histograms, idx1, i, 0.);
881     }
882   }
883 
884   ok = 1;
885 
886  End:
887   HistoQueueClear(&histo_queue);
888   return ok;
889 }
890 
891 // Perform histogram aggregation using a stochastic approach.
892 // 'do_greedy' is set to 1 if a greedy approach needs to be performed
893 // afterwards, 0 otherwise.
PairComparison(const void * idx1,const void * idx2)894 static int PairComparison(const void* idx1, const void* idx2) {
895   // To be used with bsearch: <0 when *idx1<*idx2, >0 if >, 0 when ==.
896   return (*(int*) idx1 - *(int*) idx2);
897 }
HistogramCombineStochastic(VP8LHistogramSet * const image_histo,int * const num_used,int min_cluster_size,int * const do_greedy)898 static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo,
899                                       int* const num_used, int min_cluster_size,
900                                       int* const do_greedy) {
901   int j, iter;
902   uint32_t seed = 1;
903   int tries_with_no_success = 0;
904   const int outer_iters = *num_used;
905   const int num_tries_no_success = outer_iters / 2;
906   VP8LHistogram** const histograms = image_histo->histograms;
907   // Priority queue of histogram pairs. Its size of 'kHistoQueueSize'
908   // impacts the quality of the compression and the speed: the smaller the
909   // faster but the worse for the compression.
910   HistoQueue histo_queue;
911   const int kHistoQueueSize = 9;
912   int ok = 0;
913   // mapping from an index in image_histo with no NULL histogram to the full
914   // blown image_histo.
915   int* mappings;
916 
917   if (*num_used < min_cluster_size) {
918     *do_greedy = 1;
919     return 1;
920   }
921 
922   mappings = (int*) WebPSafeMalloc(*num_used, sizeof(*mappings));
923   if (mappings == NULL) return 0;
924   if (!HistoQueueInit(&histo_queue, kHistoQueueSize)) goto End;
925   // Fill the initial mapping.
926   for (j = 0, iter = 0; iter < image_histo->size; ++iter) {
927     if (histograms[iter] == NULL) continue;
928     mappings[j++] = iter;
929   }
930   assert(j == *num_used);
931 
932   // Collapse similar histograms in 'image_histo'.
933   for (iter = 0;
934        iter < outer_iters && *num_used >= min_cluster_size &&
935            ++tries_with_no_success < num_tries_no_success;
936        ++iter) {
937     int* mapping_index;
938     float best_cost =
939         (histo_queue.size == 0) ? 0.f : histo_queue.queue[0].cost_diff;
940     int best_idx1 = -1, best_idx2 = 1;
941     const uint32_t rand_range = (*num_used - 1) * (*num_used);
942     // (*num_used) / 2 was chosen empirically. Less means faster but worse
943     // compression.
944     const int num_tries = (*num_used) / 2;
945 
946     // Pick random samples.
947     for (j = 0; *num_used >= 2 && j < num_tries; ++j) {
948       float curr_cost;
949       // Choose two different histograms at random and try to combine them.
950       const uint32_t tmp = MyRand(&seed) % rand_range;
951       uint32_t idx1 = tmp / (*num_used - 1);
952       uint32_t idx2 = tmp % (*num_used - 1);
953       if (idx2 >= idx1) ++idx2;
954       idx1 = mappings[idx1];
955       idx2 = mappings[idx2];
956 
957       // Calculate cost reduction on combination.
958       curr_cost =
959           HistoQueuePush(&histo_queue, histograms, idx1, idx2, best_cost);
960       if (curr_cost < 0) {  // found a better pair?
961         best_cost = curr_cost;
962         // Empty the queue if we reached full capacity.
963         if (histo_queue.size == histo_queue.max_size) break;
964       }
965     }
966     if (histo_queue.size == 0) continue;
967 
968     // Get the best histograms.
969     best_idx1 = histo_queue.queue[0].idx1;
970     best_idx2 = histo_queue.queue[0].idx2;
971     assert(best_idx1 < best_idx2);
972     // Pop best_idx2 from mappings.
973     mapping_index = (int*) bsearch(&best_idx2, mappings, *num_used,
974                                    sizeof(best_idx2), &PairComparison);
975     assert(mapping_index != NULL);
976     memmove(mapping_index, mapping_index + 1, sizeof(*mapping_index) *
977         ((*num_used) - (mapping_index - mappings) - 1));
978     // Merge the histograms and remove best_idx2 from the queue.
979     HistogramAdd(histograms[best_idx2], histograms[best_idx1],
980                  histograms[best_idx1]);
981     histograms[best_idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
982     HistogramSetRemoveHistogram(image_histo, best_idx2, num_used);
983     // Parse the queue and update each pair that deals with best_idx1,
984     // best_idx2 or image_histo_size.
985     for (j = 0; j < histo_queue.size;) {
986       HistogramPair* const p = histo_queue.queue + j;
987       const int is_idx1_best = p->idx1 == best_idx1 || p->idx1 == best_idx2;
988       const int is_idx2_best = p->idx2 == best_idx1 || p->idx2 == best_idx2;
989       int do_eval = 0;
990       // The front pair could have been duplicated by a random pick so
991       // check for it all the time nevertheless.
992       if (is_idx1_best && is_idx2_best) {
993         HistoQueuePopPair(&histo_queue, p);
994         continue;
995       }
996       // Any pair containing one of the two best indices should only refer to
997       // best_idx1. Its cost should also be updated.
998       if (is_idx1_best) {
999         p->idx1 = best_idx1;
1000         do_eval = 1;
1001       } else if (is_idx2_best) {
1002         p->idx2 = best_idx1;
1003         do_eval = 1;
1004       }
1005       // Make sure the index order is respected.
1006       if (p->idx1 > p->idx2) {
1007         const int tmp = p->idx2;
1008         p->idx2 = p->idx1;
1009         p->idx1 = tmp;
1010       }
1011       if (do_eval) {
1012         // Re-evaluate the cost of an updated pair.
1013         HistoQueueUpdatePair(histograms[p->idx1], histograms[p->idx2], 0., p);
1014         if (p->cost_diff >= 0.) {
1015           HistoQueuePopPair(&histo_queue, p);
1016           continue;
1017         }
1018       }
1019       HistoQueueUpdateHead(&histo_queue, p);
1020       ++j;
1021     }
1022     tries_with_no_success = 0;
1023   }
1024   *do_greedy = (*num_used <= min_cluster_size);
1025   ok = 1;
1026 
1027  End:
1028   HistoQueueClear(&histo_queue);
1029   WebPSafeFree(mappings);
1030   return ok;
1031 }
1032 
1033 // -----------------------------------------------------------------------------
1034 // Histogram refinement
1035 
1036 // Find the best 'out' histogram for each of the 'in' histograms.
1037 // At call-time, 'out' contains the histograms of the clusters.
1038 // Note: we assume that out[]->bit_cost_ is already up-to-date.
HistogramRemap(const VP8LHistogramSet * const in,VP8LHistogramSet * const out,uint16_t * const symbols)1039 static void HistogramRemap(const VP8LHistogramSet* const in,
1040                            VP8LHistogramSet* const out,
1041                            uint16_t* const symbols) {
1042   int i;
1043   VP8LHistogram** const in_histo = in->histograms;
1044   VP8LHistogram** const out_histo = out->histograms;
1045   const int in_size = out->max_size;
1046   const int out_size = out->size;
1047   if (out_size > 1) {
1048     for (i = 0; i < in_size; ++i) {
1049       int best_out = 0;
1050       float best_bits = MAX_BIT_COST;
1051       int k;
1052       if (in_histo[i] == NULL) {
1053         // Arbitrarily set to the previous value if unused to help future LZ77.
1054         symbols[i] = symbols[i - 1];
1055         continue;
1056       }
1057       for (k = 0; k < out_size; ++k) {
1058         float cur_bits;
1059         cur_bits = HistogramAddThresh(out_histo[k], in_histo[i], best_bits);
1060         if (k == 0 || cur_bits < best_bits) {
1061           best_bits = cur_bits;
1062           best_out = k;
1063         }
1064       }
1065       symbols[i] = best_out;
1066     }
1067   } else {
1068     assert(out_size == 1);
1069     for (i = 0; i < in_size; ++i) {
1070       symbols[i] = 0;
1071     }
1072   }
1073 
1074   // Recompute each out based on raw and symbols.
1075   VP8LHistogramSetClear(out);
1076   out->size = out_size;
1077 
1078   for (i = 0; i < in_size; ++i) {
1079     int idx;
1080     if (in_histo[i] == NULL) continue;
1081     idx = symbols[i];
1082     HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]);
1083   }
1084 }
1085 
GetCombineCostFactor(int histo_size,int quality)1086 static float GetCombineCostFactor(int histo_size, int quality) {
1087   float combine_cost_factor = 0.16f;
1088   if (quality < 90) {
1089     if (histo_size > 256) combine_cost_factor /= 2.f;
1090     if (histo_size > 512) combine_cost_factor /= 2.f;
1091     if (histo_size > 1024) combine_cost_factor /= 2.f;
1092     if (quality <= 50) combine_cost_factor /= 2.f;
1093   }
1094   return combine_cost_factor;
1095 }
1096 
1097 // Given a HistogramSet 'set', the mapping of clusters 'cluster_mapping' and the
1098 // current assignment of the cells in 'symbols', merge the clusters and
1099 // assign the smallest possible clusters values.
OptimizeHistogramSymbols(const VP8LHistogramSet * const set,uint16_t * const cluster_mappings,int num_clusters,uint16_t * const cluster_mappings_tmp,uint16_t * const symbols)1100 static void OptimizeHistogramSymbols(const VP8LHistogramSet* const set,
1101                                      uint16_t* const cluster_mappings,
1102                                      int num_clusters,
1103                                      uint16_t* const cluster_mappings_tmp,
1104                                      uint16_t* const symbols) {
1105   int i, cluster_max;
1106   int do_continue = 1;
1107   // First, assign the lowest cluster to each pixel.
1108   while (do_continue) {
1109     do_continue = 0;
1110     for (i = 0; i < num_clusters; ++i) {
1111       int k;
1112       k = cluster_mappings[i];
1113       while (k != cluster_mappings[k]) {
1114         cluster_mappings[k] = cluster_mappings[cluster_mappings[k]];
1115         k = cluster_mappings[k];
1116       }
1117       if (k != cluster_mappings[i]) {
1118         do_continue = 1;
1119         cluster_mappings[i] = k;
1120       }
1121     }
1122   }
1123   // Create a mapping from a cluster id to its minimal version.
1124   cluster_max = 0;
1125   memset(cluster_mappings_tmp, 0,
1126          set->max_size * sizeof(*cluster_mappings_tmp));
1127   assert(cluster_mappings[0] == 0);
1128   // Re-map the ids.
1129   for (i = 0; i < set->max_size; ++i) {
1130     int cluster;
1131     if (symbols[i] == kInvalidHistogramSymbol) continue;
1132     cluster = cluster_mappings[symbols[i]];
1133     assert(symbols[i] < num_clusters);
1134     if (cluster > 0 && cluster_mappings_tmp[cluster] == 0) {
1135       ++cluster_max;
1136       cluster_mappings_tmp[cluster] = cluster_max;
1137     }
1138     symbols[i] = cluster_mappings_tmp[cluster];
1139   }
1140 
1141   // Make sure all cluster values are used.
1142   cluster_max = 0;
1143   for (i = 0; i < set->max_size; ++i) {
1144     if (symbols[i] == kInvalidHistogramSymbol) continue;
1145     if (symbols[i] <= cluster_max) continue;
1146     ++cluster_max;
1147     assert(symbols[i] == cluster_max);
1148   }
1149 }
1150 
RemoveEmptyHistograms(VP8LHistogramSet * const image_histo)1151 static void RemoveEmptyHistograms(VP8LHistogramSet* const image_histo) {
1152   uint32_t size;
1153   int i;
1154   for (i = 0, size = 0; i < image_histo->size; ++i) {
1155     if (image_histo->histograms[i] == NULL) continue;
1156     image_histo->histograms[size++] = image_histo->histograms[i];
1157   }
1158   image_histo->size = size;
1159 }
1160 
VP8LGetHistoImageSymbols(int xsize,int ysize,const VP8LBackwardRefs * const refs,int quality,int low_effort,int histogram_bits,int cache_bits,VP8LHistogramSet * const image_histo,VP8LHistogram * const tmp_histo,uint16_t * const histogram_symbols,const WebPPicture * const pic,int percent_range,int * const percent)1161 int VP8LGetHistoImageSymbols(int xsize, int ysize,
1162                              const VP8LBackwardRefs* const refs, int quality,
1163                              int low_effort, int histogram_bits, int cache_bits,
1164                              VP8LHistogramSet* const image_histo,
1165                              VP8LHistogram* const tmp_histo,
1166                              uint16_t* const histogram_symbols,
1167                              const WebPPicture* const pic, int percent_range,
1168                              int* const percent) {
1169   const int histo_xsize =
1170       histogram_bits ? VP8LSubSampleSize(xsize, histogram_bits) : 1;
1171   const int histo_ysize =
1172       histogram_bits ? VP8LSubSampleSize(ysize, histogram_bits) : 1;
1173   const int image_histo_raw_size = histo_xsize * histo_ysize;
1174   VP8LHistogramSet* const orig_histo =
1175       VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
1176   // Don't attempt linear bin-partition heuristic for
1177   // histograms of small sizes (as bin_map will be very sparse) and
1178   // maximum quality q==100 (to preserve the compression gains at that level).
1179   const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE;
1180   int entropy_combine;
1181   uint16_t* const map_tmp =
1182       WebPSafeMalloc(2 * image_histo_raw_size, sizeof(map_tmp));
1183   uint16_t* const cluster_mappings = map_tmp + image_histo_raw_size;
1184   int num_used = image_histo_raw_size;
1185   if (orig_histo == NULL || map_tmp == NULL) {
1186     WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
1187     goto Error;
1188   }
1189 
1190   // Construct the histograms from backward references.
1191   HistogramBuild(xsize, histogram_bits, refs, orig_histo);
1192   // Copies the histograms and computes its bit_cost.
1193   // histogram_symbols is optimized
1194   HistogramCopyAndAnalyze(orig_histo, image_histo, &num_used,
1195                           histogram_symbols);
1196 
1197   entropy_combine =
1198       (num_used > entropy_combine_num_bins * 2) && (quality < 100);
1199 
1200   if (entropy_combine) {
1201     uint16_t* const bin_map = map_tmp;
1202     const float combine_cost_factor =
1203         GetCombineCostFactor(image_histo_raw_size, quality);
1204     const uint32_t num_clusters = num_used;
1205 
1206     HistogramAnalyzeEntropyBin(image_histo, bin_map, low_effort);
1207     // Collapse histograms with similar entropy.
1208     HistogramCombineEntropyBin(
1209         image_histo, &num_used, histogram_symbols, cluster_mappings, tmp_histo,
1210         bin_map, entropy_combine_num_bins, combine_cost_factor, low_effort);
1211     OptimizeHistogramSymbols(image_histo, cluster_mappings, num_clusters,
1212                              map_tmp, histogram_symbols);
1213   }
1214 
1215   // Don't combine the histograms using stochastic and greedy heuristics for
1216   // low-effort compression mode.
1217   if (!low_effort || !entropy_combine) {
1218     const float x = quality / 100.f;
1219     // cubic ramp between 1 and MAX_HISTO_GREEDY:
1220     const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1));
1221     int do_greedy;
1222     if (!HistogramCombineStochastic(image_histo, &num_used, threshold_size,
1223                                     &do_greedy)) {
1224       WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
1225       goto Error;
1226     }
1227     if (do_greedy) {
1228       RemoveEmptyHistograms(image_histo);
1229       if (!HistogramCombineGreedy(image_histo, &num_used)) {
1230         WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
1231         goto Error;
1232       }
1233     }
1234   }
1235 
1236   // Find the optimal map from original histograms to the final ones.
1237   RemoveEmptyHistograms(image_histo);
1238   HistogramRemap(orig_histo, image_histo, histogram_symbols);
1239 
1240   if (!WebPReportProgress(pic, *percent + percent_range, percent)) {
1241     goto Error;
1242   }
1243 
1244  Error:
1245   VP8LFreeHistogramSet(orig_histo);
1246   WebPSafeFree(map_tmp);
1247   return (pic->error_code == VP8_ENC_OK);
1248 }
1249