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 <math.h>
17
18 #include "src/enc/backward_references_enc.h"
19 #include "src/enc/histogram_enc.h"
20 #include "src/dsp/lossless.h"
21 #include "src/dsp/lossless_common.h"
22 #include "src/utils/utils.h"
23
24 #define MAX_COST 1.e38
25
26 // Number of partitions for the three dominant (literal, red and blue) symbol
27 // costs.
28 #define NUM_PARTITIONS 4
29 // The size of the bin-hash corresponding to the three dominant costs.
30 #define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS)
31 // Maximum number of histograms allowed in greedy combining algorithm.
32 #define MAX_HISTO_GREEDY 100
33
HistogramClear(VP8LHistogram * const p)34 static void HistogramClear(VP8LHistogram* const p) {
35 uint32_t* const literal = p->literal_;
36 const int cache_bits = p->palette_code_bits_;
37 const int histo_size = VP8LGetHistogramSize(cache_bits);
38 memset(p, 0, histo_size);
39 p->palette_code_bits_ = cache_bits;
40 p->literal_ = literal;
41 }
42
43 // Swap two histogram pointers.
HistogramSwap(VP8LHistogram ** const A,VP8LHistogram ** const B)44 static void HistogramSwap(VP8LHistogram** const A, VP8LHistogram** const B) {
45 VP8LHistogram* const tmp = *A;
46 *A = *B;
47 *B = tmp;
48 }
49
HistogramCopy(const VP8LHistogram * const src,VP8LHistogram * const dst)50 static void HistogramCopy(const VP8LHistogram* const src,
51 VP8LHistogram* const dst) {
52 uint32_t* const dst_literal = dst->literal_;
53 const int dst_cache_bits = dst->palette_code_bits_;
54 const int literal_size = VP8LHistogramNumCodes(dst_cache_bits);
55 const int histo_size = VP8LGetHistogramSize(dst_cache_bits);
56 assert(src->palette_code_bits_ == dst_cache_bits);
57 memcpy(dst, src, histo_size);
58 dst->literal_ = dst_literal;
59 memcpy(dst->literal_, src->literal_, literal_size * sizeof(*dst->literal_));
60 }
61
VP8LGetHistogramSize(int cache_bits)62 int VP8LGetHistogramSize(int cache_bits) {
63 const int literal_size = VP8LHistogramNumCodes(cache_bits);
64 const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size;
65 assert(total_size <= (size_t)0x7fffffff);
66 return (int)total_size;
67 }
68
VP8LFreeHistogram(VP8LHistogram * const histo)69 void VP8LFreeHistogram(VP8LHistogram* const histo) {
70 WebPSafeFree(histo);
71 }
72
VP8LFreeHistogramSet(VP8LHistogramSet * const histo)73 void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
74 WebPSafeFree(histo);
75 }
76
VP8LHistogramStoreRefs(const VP8LBackwardRefs * const refs,VP8LHistogram * const histo)77 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
78 VP8LHistogram* const histo) {
79 VP8LRefsCursor c = VP8LRefsCursorInit(refs);
80 while (VP8LRefsCursorOk(&c)) {
81 VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos, NULL, 0);
82 VP8LRefsCursorNext(&c);
83 }
84 }
85
VP8LHistogramCreate(VP8LHistogram * const p,const VP8LBackwardRefs * const refs,int palette_code_bits)86 void VP8LHistogramCreate(VP8LHistogram* const p,
87 const VP8LBackwardRefs* const refs,
88 int palette_code_bits) {
89 if (palette_code_bits >= 0) {
90 p->palette_code_bits_ = palette_code_bits;
91 }
92 HistogramClear(p);
93 VP8LHistogramStoreRefs(refs, p);
94 }
95
VP8LHistogramInit(VP8LHistogram * const p,int palette_code_bits,int init_arrays)96 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits,
97 int init_arrays) {
98 p->palette_code_bits_ = palette_code_bits;
99 if (init_arrays) {
100 HistogramClear(p);
101 } else {
102 p->trivial_symbol_ = 0;
103 p->bit_cost_ = 0.;
104 p->literal_cost_ = 0.;
105 p->red_cost_ = 0.;
106 p->blue_cost_ = 0.;
107 memset(p->is_used_, 0, sizeof(p->is_used_));
108 }
109 }
110
VP8LAllocateHistogram(int cache_bits)111 VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
112 VP8LHistogram* histo = NULL;
113 const int total_size = VP8LGetHistogramSize(cache_bits);
114 uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
115 if (memory == NULL) return NULL;
116 histo = (VP8LHistogram*)memory;
117 // literal_ won't necessary be aligned.
118 histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
119 VP8LHistogramInit(histo, cache_bits, /*init_arrays=*/ 0);
120 return histo;
121 }
122
123 // Resets the pointers of the histograms to point to the bit buffer in the set.
HistogramSetResetPointers(VP8LHistogramSet * const set,int cache_bits)124 static void HistogramSetResetPointers(VP8LHistogramSet* const set,
125 int cache_bits) {
126 int i;
127 const int histo_size = VP8LGetHistogramSize(cache_bits);
128 uint8_t* memory = (uint8_t*) (set->histograms);
129 memory += set->max_size * sizeof(*set->histograms);
130 for (i = 0; i < set->max_size; ++i) {
131 memory = (uint8_t*) WEBP_ALIGN(memory);
132 set->histograms[i] = (VP8LHistogram*) memory;
133 // literal_ won't necessary be aligned.
134 set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
135 memory += histo_size;
136 }
137 }
138
139 // Returns the total size of the VP8LHistogramSet.
HistogramSetTotalSize(int size,int cache_bits)140 static size_t HistogramSetTotalSize(int size, int cache_bits) {
141 const int histo_size = VP8LGetHistogramSize(cache_bits);
142 return (sizeof(VP8LHistogramSet) + size * (sizeof(VP8LHistogram*) +
143 histo_size + WEBP_ALIGN_CST));
144 }
145
VP8LAllocateHistogramSet(int size,int cache_bits)146 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
147 int i;
148 VP8LHistogramSet* set;
149 const size_t total_size = HistogramSetTotalSize(size, cache_bits);
150 uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
151 if (memory == NULL) return NULL;
152
153 set = (VP8LHistogramSet*)memory;
154 memory += sizeof(*set);
155 set->histograms = (VP8LHistogram**)memory;
156 set->max_size = size;
157 set->size = size;
158 HistogramSetResetPointers(set, cache_bits);
159 for (i = 0; i < size; ++i) {
160 VP8LHistogramInit(set->histograms[i], cache_bits, /*init_arrays=*/ 0);
161 }
162 return set;
163 }
164
VP8LHistogramSetClear(VP8LHistogramSet * const set)165 void VP8LHistogramSetClear(VP8LHistogramSet* const set) {
166 int i;
167 const int cache_bits = set->histograms[0]->palette_code_bits_;
168 const int size = set->max_size;
169 const size_t total_size = HistogramSetTotalSize(size, cache_bits);
170 uint8_t* memory = (uint8_t*)set;
171
172 memset(memory, 0, total_size);
173 memory += sizeof(*set);
174 set->histograms = (VP8LHistogram**)memory;
175 set->max_size = size;
176 set->size = size;
177 HistogramSetResetPointers(set, cache_bits);
178 for (i = 0; i < size; ++i) {
179 set->histograms[i]->palette_code_bits_ = cache_bits;
180 }
181 }
182
183 // Removes the histogram 'i' from 'set' by setting it to NULL.
HistogramSetRemoveHistogram(VP8LHistogramSet * const set,int i,int * const num_used)184 static void HistogramSetRemoveHistogram(VP8LHistogramSet* const set, int i,
185 int* const num_used) {
186 assert(set->histograms[i] != NULL);
187 set->histograms[i] = NULL;
188 --*num_used;
189 // If we remove the last valid one, shrink until the next valid one.
190 if (i == set->size - 1) {
191 while (set->size >= 1 && set->histograms[set->size - 1] == NULL) {
192 --set->size;
193 }
194 }
195 }
196
197 // -----------------------------------------------------------------------------
198
VP8LHistogramAddSinglePixOrCopy(VP8LHistogram * const histo,const PixOrCopy * const v,int (* const distance_modifier)(int,int),int distance_modifier_arg0)199 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
200 const PixOrCopy* const v,
201 int (*const distance_modifier)(int, int),
202 int distance_modifier_arg0) {
203 if (PixOrCopyIsLiteral(v)) {
204 ++histo->alpha_[PixOrCopyLiteral(v, 3)];
205 ++histo->red_[PixOrCopyLiteral(v, 2)];
206 ++histo->literal_[PixOrCopyLiteral(v, 1)];
207 ++histo->blue_[PixOrCopyLiteral(v, 0)];
208 } else if (PixOrCopyIsCacheIdx(v)) {
209 const int literal_ix =
210 NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
211 assert(histo->palette_code_bits_ != 0);
212 ++histo->literal_[literal_ix];
213 } else {
214 int code, extra_bits;
215 VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
216 ++histo->literal_[NUM_LITERAL_CODES + code];
217 if (distance_modifier == NULL) {
218 VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
219 } else {
220 VP8LPrefixEncodeBits(
221 distance_modifier(distance_modifier_arg0, PixOrCopyDistance(v)),
222 &code, &extra_bits);
223 }
224 ++histo->distance_[code];
225 }
226 }
227
228 // -----------------------------------------------------------------------------
229 // Entropy-related functions.
230
BitsEntropyRefine(const VP8LBitEntropy * entropy)231 static WEBP_INLINE double BitsEntropyRefine(const VP8LBitEntropy* entropy) {
232 double mix;
233 if (entropy->nonzeros < 5) {
234 if (entropy->nonzeros <= 1) {
235 return 0;
236 }
237 // Two symbols, they will be 0 and 1 in a Huffman code.
238 // Let's mix in a bit of entropy to favor good clustering when
239 // distributions of these are combined.
240 if (entropy->nonzeros == 2) {
241 return 0.99 * entropy->sum + 0.01 * entropy->entropy;
242 }
243 // No matter what the entropy says, we cannot be better than min_limit
244 // with Huffman coding. I am mixing a bit of entropy into the
245 // min_limit since it produces much better (~0.5 %) compression results
246 // perhaps because of better entropy clustering.
247 if (entropy->nonzeros == 3) {
248 mix = 0.95;
249 } else {
250 mix = 0.7; // nonzeros == 4.
251 }
252 } else {
253 mix = 0.627;
254 }
255
256 {
257 double min_limit = 2 * entropy->sum - entropy->max_val;
258 min_limit = mix * min_limit + (1.0 - mix) * entropy->entropy;
259 return (entropy->entropy < min_limit) ? min_limit : entropy->entropy;
260 }
261 }
262
VP8LBitsEntropy(const uint32_t * const array,int n)263 double VP8LBitsEntropy(const uint32_t* const array, int n) {
264 VP8LBitEntropy entropy;
265 VP8LBitsEntropyUnrefined(array, n, &entropy);
266
267 return BitsEntropyRefine(&entropy);
268 }
269
InitialHuffmanCost(void)270 static double InitialHuffmanCost(void) {
271 // Small bias because Huffman code length is typically not stored in
272 // full length.
273 static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
274 static const double kSmallBias = 9.1;
275 return kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
276 }
277
278 // Finalize the Huffman cost based on streak numbers and length type (<3 or >=3)
FinalHuffmanCost(const VP8LStreaks * const stats)279 static double FinalHuffmanCost(const VP8LStreaks* const stats) {
280 // The constants in this function are experimental and got rounded from
281 // their original values in 1/8 when switched to 1/1024.
282 double retval = InitialHuffmanCost();
283 // Second coefficient: Many zeros in the histogram are covered efficiently
284 // by a run-length encode. Originally 2/8.
285 retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1];
286 // Second coefficient: Constant values are encoded less efficiently, but still
287 // RLE'ed. Originally 6/8.
288 retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1];
289 // 0s are usually encoded more efficiently than non-0s.
290 // Originally 15/8.
291 retval += 1.796875 * stats->streaks[0][0];
292 // Originally 26/8.
293 retval += 3.28125 * stats->streaks[1][0];
294 return retval;
295 }
296
297 // Get the symbol entropy for the distribution 'population'.
298 // 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)299 static double PopulationCost(const uint32_t* const population, int length,
300 uint32_t* const trivial_sym,
301 uint8_t* const is_used) {
302 VP8LBitEntropy bit_entropy;
303 VP8LStreaks stats;
304 VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats);
305 if (trivial_sym != NULL) {
306 *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code
307 : VP8L_NON_TRIVIAL_SYM;
308 }
309 // The histogram is used if there is at least one non-zero streak.
310 *is_used = (stats.streaks[1][0] != 0 || stats.streaks[1][1] != 0);
311
312 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
313 }
314
315 // trivial_at_end is 1 if the two histograms only have one element that is
316 // 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)317 static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X,
318 const uint32_t* const Y,
319 int length, int is_X_used,
320 int is_Y_used,
321 int trivial_at_end) {
322 VP8LStreaks stats;
323 if (trivial_at_end) {
324 // This configuration is due to palettization that transforms an indexed
325 // pixel into 0xff000000 | (pixel << 8) in VP8LBundleColorMap.
326 // BitsEntropyRefine is 0 for histograms with only one non-zero value.
327 // Only FinalHuffmanCost needs to be evaluated.
328 memset(&stats, 0, sizeof(stats));
329 // Deal with the non-zero value at index 0 or length-1.
330 stats.streaks[1][0] = 1;
331 // Deal with the following/previous zero streak.
332 stats.counts[0] = 1;
333 stats.streaks[0][1] = length - 1;
334 return FinalHuffmanCost(&stats);
335 } else {
336 VP8LBitEntropy bit_entropy;
337 if (is_X_used) {
338 if (is_Y_used) {
339 VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats);
340 } else {
341 VP8LGetEntropyUnrefined(X, length, &bit_entropy, &stats);
342 }
343 } else {
344 if (is_Y_used) {
345 VP8LGetEntropyUnrefined(Y, length, &bit_entropy, &stats);
346 } else {
347 memset(&stats, 0, sizeof(stats));
348 stats.counts[0] = 1;
349 stats.streaks[0][length > 3] = length;
350 VP8LBitEntropyInit(&bit_entropy);
351 }
352 }
353
354 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
355 }
356 }
357
358 // Estimates the Entropy + Huffman + other block overhead size cost.
VP8LHistogramEstimateBits(VP8LHistogram * const p)359 double VP8LHistogramEstimateBits(VP8LHistogram* const p) {
360 return
361 PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_),
362 NULL, &p->is_used_[0])
363 + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL, &p->is_used_[1])
364 + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL, &p->is_used_[2])
365 + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL, &p->is_used_[3])
366 + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL, &p->is_used_[4])
367 + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
368 + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
369 }
370
371 // -----------------------------------------------------------------------------
372 // Various histogram combine/cost-eval functions
373
GetCombinedHistogramEntropy(const VP8LHistogram * const a,const VP8LHistogram * const b,double cost_threshold,double * cost)374 static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
375 const VP8LHistogram* const b,
376 double cost_threshold,
377 double* 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,double cost_threshold)442 static double HistogramAddEval(const VP8LHistogram* const a,
443 const VP8LHistogram* const b,
444 VP8LHistogram* const out,
445 double cost_threshold) {
446 double cost = 0;
447 const double sum_cost = a->bit_cost_ + b->bit_cost_;
448 cost_threshold += sum_cost;
449
450 if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
451 HistogramAdd(a, b, out);
452 out->bit_cost_ = cost;
453 out->palette_code_bits_ = a->palette_code_bits_;
454 }
455
456 return cost - sum_cost;
457 }
458
459 // Same as HistogramAddEval(), except that the resulting histogram
460 // is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit
461 // the term C(b) which is constant over all the evaluations.
HistogramAddThresh(const VP8LHistogram * const a,const VP8LHistogram * const b,double cost_threshold)462 static double HistogramAddThresh(const VP8LHistogram* const a,
463 const VP8LHistogram* const b,
464 double cost_threshold) {
465 double cost;
466 assert(a != NULL && b != NULL);
467 cost = -a->bit_cost_;
468 GetCombinedHistogramEntropy(a, b, cost_threshold, &cost);
469 return cost;
470 }
471
472 // -----------------------------------------------------------------------------
473
474 // The structure to keep track of cost range for the three dominant entropy
475 // symbols.
476 // TODO(skal): Evaluate if float can be used here instead of double for
477 // representing the entropy costs.
478 typedef struct {
479 double literal_max_;
480 double literal_min_;
481 double red_max_;
482 double red_min_;
483 double blue_max_;
484 double blue_min_;
485 } DominantCostRange;
486
DominantCostRangeInit(DominantCostRange * const c)487 static void DominantCostRangeInit(DominantCostRange* const c) {
488 c->literal_max_ = 0.;
489 c->literal_min_ = MAX_COST;
490 c->red_max_ = 0.;
491 c->red_min_ = MAX_COST;
492 c->blue_max_ = 0.;
493 c->blue_min_ = MAX_COST;
494 }
495
UpdateDominantCostRange(const VP8LHistogram * const h,DominantCostRange * const c)496 static void UpdateDominantCostRange(
497 const VP8LHistogram* const h, DominantCostRange* const c) {
498 if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_;
499 if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_;
500 if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_;
501 if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_;
502 if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_;
503 if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_;
504 }
505
UpdateHistogramCost(VP8LHistogram * const h)506 static void UpdateHistogramCost(VP8LHistogram* const h) {
507 uint32_t alpha_sym, red_sym, blue_sym;
508 const double alpha_cost =
509 PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym,
510 &h->is_used_[3]);
511 const double distance_cost =
512 PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL, &h->is_used_[4]) +
513 VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
514 const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
515 h->literal_cost_ =
516 PopulationCost(h->literal_, num_codes, NULL, &h->is_used_[0]) +
517 VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES);
518 h->red_cost_ =
519 PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym, &h->is_used_[1]);
520 h->blue_cost_ =
521 PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym, &h->is_used_[2]);
522 h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
523 alpha_cost + distance_cost;
524 if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) {
525 h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM;
526 } else {
527 h->trivial_symbol_ =
528 ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0);
529 }
530 }
531
GetBinIdForEntropy(double min,double max,double val)532 static int GetBinIdForEntropy(double min, double max, double val) {
533 const double range = max - min;
534 if (range > 0.) {
535 const double delta = val - min;
536 return (int)((NUM_PARTITIONS - 1e-6) * delta / range);
537 } else {
538 return 0;
539 }
540 }
541
GetHistoBinIndex(const VP8LHistogram * const h,const DominantCostRange * const c,int low_effort)542 static int GetHistoBinIndex(const VP8LHistogram* const h,
543 const DominantCostRange* const c, int low_effort) {
544 int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_,
545 h->literal_cost_);
546 assert(bin_id < NUM_PARTITIONS);
547 if (!low_effort) {
548 bin_id = bin_id * NUM_PARTITIONS
549 + GetBinIdForEntropy(c->red_min_, c->red_max_, h->red_cost_);
550 bin_id = bin_id * NUM_PARTITIONS
551 + GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_);
552 assert(bin_id < BIN_SIZE);
553 }
554 return bin_id;
555 }
556
557 // Construct the histograms from backward references.
HistogramBuild(int xsize,int histo_bits,const VP8LBackwardRefs * const backward_refs,VP8LHistogramSet * const image_histo)558 static void HistogramBuild(
559 int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs,
560 VP8LHistogramSet* const image_histo) {
561 int x = 0, y = 0;
562 const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
563 VP8LHistogram** const histograms = image_histo->histograms;
564 VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs);
565 assert(histo_bits > 0);
566 VP8LHistogramSetClear(image_histo);
567 while (VP8LRefsCursorOk(&c)) {
568 const PixOrCopy* const v = c.cur_pos;
569 const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
570 VP8LHistogramAddSinglePixOrCopy(histograms[ix], v, NULL, 0);
571 x += PixOrCopyLength(v);
572 while (x >= xsize) {
573 x -= xsize;
574 ++y;
575 }
576 VP8LRefsCursorNext(&c);
577 }
578 }
579
580 // Copies the histograms and computes its bit_cost.
581 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)582 static void HistogramCopyAndAnalyze(VP8LHistogramSet* const orig_histo,
583 VP8LHistogramSet* const image_histo,
584 int* const num_used,
585 uint16_t* const histogram_symbols) {
586 int i, cluster_id;
587 int num_used_orig = *num_used;
588 VP8LHistogram** const orig_histograms = orig_histo->histograms;
589 VP8LHistogram** const histograms = image_histo->histograms;
590 assert(image_histo->max_size == orig_histo->max_size);
591 for (cluster_id = 0, i = 0; i < orig_histo->max_size; ++i) {
592 VP8LHistogram* const histo = orig_histograms[i];
593 UpdateHistogramCost(histo);
594
595 // Skip the histogram if it is completely empty, which can happen for tiles
596 // with no information (when they are skipped because of LZ77).
597 if (!histo->is_used_[0] && !histo->is_used_[1] && !histo->is_used_[2]
598 && !histo->is_used_[3] && !histo->is_used_[4]) {
599 // The first histogram is always used. If an histogram is empty, we set
600 // its id to be the same as the previous one: this will improve
601 // compressibility for later LZ77.
602 assert(i > 0);
603 HistogramSetRemoveHistogram(image_histo, i, num_used);
604 HistogramSetRemoveHistogram(orig_histo, i, &num_used_orig);
605 histogram_symbols[i] = kInvalidHistogramSymbol;
606 } else {
607 // Copy histograms from orig_histo[] to image_histo[].
608 HistogramCopy(histo, histograms[i]);
609 histogram_symbols[i] = cluster_id++;
610 assert(cluster_id <= image_histo->max_size);
611 }
612 }
613 }
614
615 // Partition histograms to different entropy bins for three dominant (literal,
616 // 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)617 static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo,
618 uint16_t* const bin_map,
619 int low_effort) {
620 int i;
621 VP8LHistogram** const histograms = image_histo->histograms;
622 const int histo_size = image_histo->size;
623 DominantCostRange cost_range;
624 DominantCostRangeInit(&cost_range);
625
626 // Analyze the dominant (literal, red and blue) entropy costs.
627 for (i = 0; i < histo_size; ++i) {
628 if (histograms[i] == NULL) continue;
629 UpdateDominantCostRange(histograms[i], &cost_range);
630 }
631
632 // bin-hash histograms on three of the dominant (literal, red and blue)
633 // symbol costs and store the resulting bin_id for each histogram.
634 for (i = 0; i < histo_size; ++i) {
635 // bin_map[i] is not set to a special value as its use will later be guarded
636 // by another (histograms[i] == NULL).
637 if (histograms[i] == NULL) continue;
638 bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort);
639 }
640 }
641
642 // Merges some histograms with same bin_id together if it's advantageous.
643 // 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,double combine_cost_factor,int low_effort)644 static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo,
645 int* num_used,
646 const uint16_t* const clusters,
647 uint16_t* const cluster_mappings,
648 VP8LHistogram* cur_combo,
649 const uint16_t* const bin_map,
650 int num_bins,
651 double combine_cost_factor,
652 int low_effort) {
653 VP8LHistogram** const histograms = image_histo->histograms;
654 int idx;
655 struct {
656 int16_t first; // position of the histogram that accumulates all
657 // histograms with the same bin_id
658 uint16_t num_combine_failures; // number of combine failures per bin_id
659 } bin_info[BIN_SIZE];
660
661 assert(num_bins <= BIN_SIZE);
662 for (idx = 0; idx < num_bins; ++idx) {
663 bin_info[idx].first = -1;
664 bin_info[idx].num_combine_failures = 0;
665 }
666
667 // By default, a cluster matches itself.
668 for (idx = 0; idx < *num_used; ++idx) cluster_mappings[idx] = idx;
669 for (idx = 0; idx < image_histo->size; ++idx) {
670 int bin_id, first;
671 if (histograms[idx] == NULL) continue;
672 bin_id = bin_map[idx];
673 first = bin_info[bin_id].first;
674 if (first == -1) {
675 bin_info[bin_id].first = idx;
676 } else if (low_effort) {
677 HistogramAdd(histograms[idx], histograms[first], histograms[first]);
678 HistogramSetRemoveHistogram(image_histo, idx, num_used);
679 cluster_mappings[clusters[idx]] = clusters[first];
680 } else {
681 // try to merge #idx into #first (both share the same bin_id)
682 const double bit_cost = histograms[idx]->bit_cost_;
683 const double bit_cost_thresh = -bit_cost * combine_cost_factor;
684 const double curr_cost_diff =
685 HistogramAddEval(histograms[first], histograms[idx],
686 cur_combo, bit_cost_thresh);
687 if (curr_cost_diff < bit_cost_thresh) {
688 // Try to merge two histograms only if the combo is a trivial one or
689 // the two candidate histograms are already non-trivial.
690 // For some images, 'try_combine' turns out to be false for a lot of
691 // histogram pairs. In that case, we fallback to combining
692 // histograms as usual to avoid increasing the header size.
693 const int try_combine =
694 (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) ||
695 ((histograms[idx]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) &&
696 (histograms[first]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM));
697 const int max_combine_failures = 32;
698 if (try_combine ||
699 bin_info[bin_id].num_combine_failures >= max_combine_failures) {
700 // move the (better) merged histogram to its final slot
701 HistogramSwap(&cur_combo, &histograms[first]);
702 HistogramSetRemoveHistogram(image_histo, idx, num_used);
703 cluster_mappings[clusters[idx]] = clusters[first];
704 } else {
705 ++bin_info[bin_id].num_combine_failures;
706 }
707 }
708 }
709 }
710 if (low_effort) {
711 // for low_effort case, update the final cost when everything is merged
712 for (idx = 0; idx < image_histo->size; ++idx) {
713 if (histograms[idx] == NULL) continue;
714 UpdateHistogramCost(histograms[idx]);
715 }
716 }
717 }
718
719 // Implement a Lehmer random number generator with a multiplicative constant of
720 // 48271 and a modulo constant of 2^31 - 1.
MyRand(uint32_t * const seed)721 static uint32_t MyRand(uint32_t* const seed) {
722 *seed = (uint32_t)(((uint64_t)(*seed) * 48271u) % 2147483647u);
723 assert(*seed > 0);
724 return *seed;
725 }
726
727 // -----------------------------------------------------------------------------
728 // Histogram pairs priority queue
729
730 // Pair of histograms. Negative idx1 value means that pair is out-of-date.
731 typedef struct {
732 int idx1;
733 int idx2;
734 double cost_diff;
735 double cost_combo;
736 } HistogramPair;
737
738 typedef struct {
739 HistogramPair* queue;
740 int size;
741 int max_size;
742 } HistoQueue;
743
HistoQueueInit(HistoQueue * const histo_queue,const int max_size)744 static int HistoQueueInit(HistoQueue* const histo_queue, const int max_size) {
745 histo_queue->size = 0;
746 histo_queue->max_size = max_size;
747 // We allocate max_size + 1 because the last element at index "size" is
748 // used as temporary data (and it could be up to max_size).
749 histo_queue->queue = (HistogramPair*)WebPSafeMalloc(
750 histo_queue->max_size + 1, sizeof(*histo_queue->queue));
751 return histo_queue->queue != NULL;
752 }
753
HistoQueueClear(HistoQueue * const histo_queue)754 static void HistoQueueClear(HistoQueue* const histo_queue) {
755 assert(histo_queue != NULL);
756 WebPSafeFree(histo_queue->queue);
757 histo_queue->size = 0;
758 histo_queue->max_size = 0;
759 }
760
761 // Pop a specific pair in the queue by replacing it with the last one
762 // and shrinking the queue.
HistoQueuePopPair(HistoQueue * const histo_queue,HistogramPair * const pair)763 static void HistoQueuePopPair(HistoQueue* const histo_queue,
764 HistogramPair* const pair) {
765 assert(pair >= histo_queue->queue &&
766 pair < (histo_queue->queue + histo_queue->size));
767 assert(histo_queue->size > 0);
768 *pair = histo_queue->queue[histo_queue->size - 1];
769 --histo_queue->size;
770 }
771
772 // Check whether a pair in the queue should be updated as head or not.
HistoQueueUpdateHead(HistoQueue * const histo_queue,HistogramPair * const pair)773 static void HistoQueueUpdateHead(HistoQueue* const histo_queue,
774 HistogramPair* const pair) {
775 assert(pair->cost_diff < 0.);
776 assert(pair >= histo_queue->queue &&
777 pair < (histo_queue->queue + histo_queue->size));
778 assert(histo_queue->size > 0);
779 if (pair->cost_diff < histo_queue->queue[0].cost_diff) {
780 // Replace the best pair.
781 const HistogramPair tmp = histo_queue->queue[0];
782 histo_queue->queue[0] = *pair;
783 *pair = tmp;
784 }
785 }
786
787 // Update the cost diff and combo of a pair of histograms. This needs to be
788 // called when the the histograms have been merged with a third one.
HistoQueueUpdatePair(const VP8LHistogram * const h1,const VP8LHistogram * const h2,double threshold,HistogramPair * const pair)789 static void HistoQueueUpdatePair(const VP8LHistogram* const h1,
790 const VP8LHistogram* const h2,
791 double threshold,
792 HistogramPair* const pair) {
793 const double sum_cost = h1->bit_cost_ + h2->bit_cost_;
794 pair->cost_combo = 0.;
795 GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair->cost_combo);
796 pair->cost_diff = pair->cost_combo - sum_cost;
797 }
798
799 // Create a pair from indices "idx1" and "idx2" provided its cost
800 // is inferior to "threshold", a negative entropy.
801 // 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,double threshold)802 static double HistoQueuePush(HistoQueue* const histo_queue,
803 VP8LHistogram** const histograms, int idx1,
804 int idx2, double threshold) {
805 const VP8LHistogram* h1;
806 const VP8LHistogram* h2;
807 HistogramPair pair;
808
809 // Stop here if the queue is full.
810 if (histo_queue->size == histo_queue->max_size) return 0.;
811 assert(threshold <= 0.);
812 if (idx1 > idx2) {
813 const int tmp = idx2;
814 idx2 = idx1;
815 idx1 = tmp;
816 }
817 pair.idx1 = idx1;
818 pair.idx2 = idx2;
819 h1 = histograms[idx1];
820 h2 = histograms[idx2];
821
822 HistoQueueUpdatePair(h1, h2, threshold, &pair);
823
824 // Do not even consider the pair if it does not improve the entropy.
825 if (pair.cost_diff >= threshold) return 0.;
826
827 histo_queue->queue[histo_queue->size++] = pair;
828 HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]);
829
830 return pair.cost_diff;
831 }
832
833 // -----------------------------------------------------------------------------
834
835 // Combines histograms by continuously choosing the one with the highest cost
836 // reduction.
HistogramCombineGreedy(VP8LHistogramSet * const image_histo,int * const num_used)837 static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo,
838 int* const num_used) {
839 int ok = 0;
840 const int image_histo_size = image_histo->size;
841 int i, j;
842 VP8LHistogram** const histograms = image_histo->histograms;
843 // Priority queue of histogram pairs.
844 HistoQueue histo_queue;
845
846 // image_histo_size^2 for the queue size is safe. If you look at
847 // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes
848 // data to the queue, you insert at most:
849 // - image_histo_size*(image_histo_size-1)/2 (the first two for loops)
850 // - image_histo_size - 1 in the last for loop at the first iteration of
851 // the while loop, image_histo_size - 2 at the second iteration ...
852 // therefore image_histo_size*(image_histo_size-1)/2 overall too
853 if (!HistoQueueInit(&histo_queue, image_histo_size * image_histo_size)) {
854 goto End;
855 }
856
857 for (i = 0; i < image_histo_size; ++i) {
858 if (image_histo->histograms[i] == NULL) continue;
859 for (j = i + 1; j < image_histo_size; ++j) {
860 // Initialize queue.
861 if (image_histo->histograms[j] == NULL) continue;
862 HistoQueuePush(&histo_queue, histograms, i, j, 0.);
863 }
864 }
865
866 while (histo_queue.size > 0) {
867 const int idx1 = histo_queue.queue[0].idx1;
868 const int idx2 = histo_queue.queue[0].idx2;
869 HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]);
870 histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
871
872 // Remove merged histogram.
873 HistogramSetRemoveHistogram(image_histo, idx2, num_used);
874
875 // Remove pairs intersecting the just combined best pair.
876 for (i = 0; i < histo_queue.size;) {
877 HistogramPair* const p = histo_queue.queue + i;
878 if (p->idx1 == idx1 || p->idx2 == idx1 ||
879 p->idx1 == idx2 || p->idx2 == idx2) {
880 HistoQueuePopPair(&histo_queue, p);
881 } else {
882 HistoQueueUpdateHead(&histo_queue, p);
883 ++i;
884 }
885 }
886
887 // Push new pairs formed with combined histogram to the queue.
888 for (i = 0; i < image_histo->size; ++i) {
889 if (i == idx1 || image_histo->histograms[i] == NULL) continue;
890 HistoQueuePush(&histo_queue, image_histo->histograms, idx1, i, 0.);
891 }
892 }
893
894 ok = 1;
895
896 End:
897 HistoQueueClear(&histo_queue);
898 return ok;
899 }
900
901 // Perform histogram aggregation using a stochastic approach.
902 // 'do_greedy' is set to 1 if a greedy approach needs to be performed
903 // afterwards, 0 otherwise.
PairComparison(const void * idx1,const void * idx2)904 static int PairComparison(const void* idx1, const void* idx2) {
905 // To be used with bsearch: <0 when *idx1<*idx2, >0 if >, 0 when ==.
906 return (*(int*) idx1 - *(int*) idx2);
907 }
HistogramCombineStochastic(VP8LHistogramSet * const image_histo,int * const num_used,int min_cluster_size,int * const do_greedy)908 static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo,
909 int* const num_used, int min_cluster_size,
910 int* const do_greedy) {
911 int j, iter;
912 uint32_t seed = 1;
913 int tries_with_no_success = 0;
914 const int outer_iters = *num_used;
915 const int num_tries_no_success = outer_iters / 2;
916 VP8LHistogram** const histograms = image_histo->histograms;
917 // Priority queue of histogram pairs. Its size of 'kHistoQueueSize'
918 // impacts the quality of the compression and the speed: the smaller the
919 // faster but the worse for the compression.
920 HistoQueue histo_queue;
921 const int kHistoQueueSize = 9;
922 int ok = 0;
923 // mapping from an index in image_histo with no NULL histogram to the full
924 // blown image_histo.
925 int* mappings;
926
927 if (*num_used < min_cluster_size) {
928 *do_greedy = 1;
929 return 1;
930 }
931
932 mappings = (int*) WebPSafeMalloc(*num_used, sizeof(*mappings));
933 if (mappings == NULL) return 0;
934 if (!HistoQueueInit(&histo_queue, kHistoQueueSize)) goto End;
935 // Fill the initial mapping.
936 for (j = 0, iter = 0; iter < image_histo->size; ++iter) {
937 if (histograms[iter] == NULL) continue;
938 mappings[j++] = iter;
939 }
940 assert(j == *num_used);
941
942 // Collapse similar histograms in 'image_histo'.
943 for (iter = 0;
944 iter < outer_iters && *num_used >= min_cluster_size &&
945 ++tries_with_no_success < num_tries_no_success;
946 ++iter) {
947 int* mapping_index;
948 double best_cost =
949 (histo_queue.size == 0) ? 0. : histo_queue.queue[0].cost_diff;
950 int best_idx1 = -1, best_idx2 = 1;
951 const uint32_t rand_range = (*num_used - 1) * (*num_used);
952 // (*num_used) / 2 was chosen empirically. Less means faster but worse
953 // compression.
954 const int num_tries = (*num_used) / 2;
955
956 // Pick random samples.
957 for (j = 0; *num_used >= 2 && j < num_tries; ++j) {
958 double curr_cost;
959 // Choose two different histograms at random and try to combine them.
960 const uint32_t tmp = MyRand(&seed) % rand_range;
961 uint32_t idx1 = tmp / (*num_used - 1);
962 uint32_t idx2 = tmp % (*num_used - 1);
963 if (idx2 >= idx1) ++idx2;
964 idx1 = mappings[idx1];
965 idx2 = mappings[idx2];
966
967 // Calculate cost reduction on combination.
968 curr_cost =
969 HistoQueuePush(&histo_queue, histograms, idx1, idx2, best_cost);
970 if (curr_cost < 0) { // found a better pair?
971 best_cost = curr_cost;
972 // Empty the queue if we reached full capacity.
973 if (histo_queue.size == histo_queue.max_size) break;
974 }
975 }
976 if (histo_queue.size == 0) continue;
977
978 // Get the best histograms.
979 best_idx1 = histo_queue.queue[0].idx1;
980 best_idx2 = histo_queue.queue[0].idx2;
981 assert(best_idx1 < best_idx2);
982 // Pop best_idx2 from mappings.
983 mapping_index = (int*) bsearch(&best_idx2, mappings, *num_used,
984 sizeof(best_idx2), &PairComparison);
985 assert(mapping_index != NULL);
986 memmove(mapping_index, mapping_index + 1, sizeof(*mapping_index) *
987 ((*num_used) - (mapping_index - mappings) - 1));
988 // Merge the histograms and remove best_idx2 from the queue.
989 HistogramAdd(histograms[best_idx2], histograms[best_idx1],
990 histograms[best_idx1]);
991 histograms[best_idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
992 HistogramSetRemoveHistogram(image_histo, best_idx2, num_used);
993 // Parse the queue and update each pair that deals with best_idx1,
994 // best_idx2 or image_histo_size.
995 for (j = 0; j < histo_queue.size;) {
996 HistogramPair* const p = histo_queue.queue + j;
997 const int is_idx1_best = p->idx1 == best_idx1 || p->idx1 == best_idx2;
998 const int is_idx2_best = p->idx2 == best_idx1 || p->idx2 == best_idx2;
999 int do_eval = 0;
1000 // The front pair could have been duplicated by a random pick so
1001 // check for it all the time nevertheless.
1002 if (is_idx1_best && is_idx2_best) {
1003 HistoQueuePopPair(&histo_queue, p);
1004 continue;
1005 }
1006 // Any pair containing one of the two best indices should only refer to
1007 // best_idx1. Its cost should also be updated.
1008 if (is_idx1_best) {
1009 p->idx1 = best_idx1;
1010 do_eval = 1;
1011 } else if (is_idx2_best) {
1012 p->idx2 = best_idx1;
1013 do_eval = 1;
1014 }
1015 // Make sure the index order is respected.
1016 if (p->idx1 > p->idx2) {
1017 const int tmp = p->idx2;
1018 p->idx2 = p->idx1;
1019 p->idx1 = tmp;
1020 }
1021 if (do_eval) {
1022 // Re-evaluate the cost of an updated pair.
1023 HistoQueueUpdatePair(histograms[p->idx1], histograms[p->idx2], 0., p);
1024 if (p->cost_diff >= 0.) {
1025 HistoQueuePopPair(&histo_queue, p);
1026 continue;
1027 }
1028 }
1029 HistoQueueUpdateHead(&histo_queue, p);
1030 ++j;
1031 }
1032 tries_with_no_success = 0;
1033 }
1034 *do_greedy = (*num_used <= min_cluster_size);
1035 ok = 1;
1036
1037 End:
1038 HistoQueueClear(&histo_queue);
1039 WebPSafeFree(mappings);
1040 return ok;
1041 }
1042
1043 // -----------------------------------------------------------------------------
1044 // Histogram refinement
1045
1046 // Find the best 'out' histogram for each of the 'in' histograms.
1047 // At call-time, 'out' contains the histograms of the clusters.
1048 // Note: we assume that out[]->bit_cost_ is already up-to-date.
HistogramRemap(const VP8LHistogramSet * const in,VP8LHistogramSet * const out,uint16_t * const symbols)1049 static void HistogramRemap(const VP8LHistogramSet* const in,
1050 VP8LHistogramSet* const out,
1051 uint16_t* const symbols) {
1052 int i;
1053 VP8LHistogram** const in_histo = in->histograms;
1054 VP8LHistogram** const out_histo = out->histograms;
1055 const int in_size = out->max_size;
1056 const int out_size = out->size;
1057 if (out_size > 1) {
1058 for (i = 0; i < in_size; ++i) {
1059 int best_out = 0;
1060 double best_bits = MAX_COST;
1061 int k;
1062 if (in_histo[i] == NULL) {
1063 // Arbitrarily set to the previous value if unused to help future LZ77.
1064 symbols[i] = symbols[i - 1];
1065 continue;
1066 }
1067 for (k = 0; k < out_size; ++k) {
1068 double cur_bits;
1069 cur_bits = HistogramAddThresh(out_histo[k], in_histo[i], best_bits);
1070 if (k == 0 || cur_bits < best_bits) {
1071 best_bits = cur_bits;
1072 best_out = k;
1073 }
1074 }
1075 symbols[i] = best_out;
1076 }
1077 } else {
1078 assert(out_size == 1);
1079 for (i = 0; i < in_size; ++i) {
1080 symbols[i] = 0;
1081 }
1082 }
1083
1084 // Recompute each out based on raw and symbols.
1085 VP8LHistogramSetClear(out);
1086 out->size = out_size;
1087
1088 for (i = 0; i < in_size; ++i) {
1089 int idx;
1090 if (in_histo[i] == NULL) continue;
1091 idx = symbols[i];
1092 HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]);
1093 }
1094 }
1095
GetCombineCostFactor(int histo_size,int quality)1096 static double GetCombineCostFactor(int histo_size, int quality) {
1097 double combine_cost_factor = 0.16;
1098 if (quality < 90) {
1099 if (histo_size > 256) combine_cost_factor /= 2.;
1100 if (histo_size > 512) combine_cost_factor /= 2.;
1101 if (histo_size > 1024) combine_cost_factor /= 2.;
1102 if (quality <= 50) combine_cost_factor /= 2.;
1103 }
1104 return combine_cost_factor;
1105 }
1106
1107 // Given a HistogramSet 'set', the mapping of clusters 'cluster_mapping' and the
1108 // current assignment of the cells in 'symbols', merge the clusters and
1109 // 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)1110 static void OptimizeHistogramSymbols(const VP8LHistogramSet* const set,
1111 uint16_t* const cluster_mappings,
1112 int num_clusters,
1113 uint16_t* const cluster_mappings_tmp,
1114 uint16_t* const symbols) {
1115 int i, cluster_max;
1116 int do_continue = 1;
1117 // First, assign the lowest cluster to each pixel.
1118 while (do_continue) {
1119 do_continue = 0;
1120 for (i = 0; i < num_clusters; ++i) {
1121 int k;
1122 k = cluster_mappings[i];
1123 while (k != cluster_mappings[k]) {
1124 cluster_mappings[k] = cluster_mappings[cluster_mappings[k]];
1125 k = cluster_mappings[k];
1126 }
1127 if (k != cluster_mappings[i]) {
1128 do_continue = 1;
1129 cluster_mappings[i] = k;
1130 }
1131 }
1132 }
1133 // Create a mapping from a cluster id to its minimal version.
1134 cluster_max = 0;
1135 memset(cluster_mappings_tmp, 0,
1136 set->max_size * sizeof(*cluster_mappings_tmp));
1137 assert(cluster_mappings[0] == 0);
1138 // Re-map the ids.
1139 for (i = 0; i < set->max_size; ++i) {
1140 int cluster;
1141 if (symbols[i] == kInvalidHistogramSymbol) continue;
1142 cluster = cluster_mappings[symbols[i]];
1143 assert(symbols[i] < num_clusters);
1144 if (cluster > 0 && cluster_mappings_tmp[cluster] == 0) {
1145 ++cluster_max;
1146 cluster_mappings_tmp[cluster] = cluster_max;
1147 }
1148 symbols[i] = cluster_mappings_tmp[cluster];
1149 }
1150
1151 // Make sure all cluster values are used.
1152 cluster_max = 0;
1153 for (i = 0; i < set->max_size; ++i) {
1154 if (symbols[i] == kInvalidHistogramSymbol) continue;
1155 if (symbols[i] <= cluster_max) continue;
1156 ++cluster_max;
1157 assert(symbols[i] == cluster_max);
1158 }
1159 }
1160
RemoveEmptyHistograms(VP8LHistogramSet * const image_histo)1161 static void RemoveEmptyHistograms(VP8LHistogramSet* const image_histo) {
1162 uint32_t size;
1163 int i;
1164 for (i = 0, size = 0; i < image_histo->size; ++i) {
1165 if (image_histo->histograms[i] == NULL) continue;
1166 image_histo->histograms[size++] = image_histo->histograms[i];
1167 }
1168 image_histo->size = size;
1169 }
1170
VP8LGetHistoImageSymbols(int xsize,int ysize,const VP8LBackwardRefs * const refs,int quality,int low_effort,int histo_bits,int cache_bits,VP8LHistogramSet * const image_histo,VP8LHistogram * const tmp_histo,uint16_t * const histogram_symbols)1171 int VP8LGetHistoImageSymbols(int xsize, int ysize,
1172 const VP8LBackwardRefs* const refs,
1173 int quality, int low_effort,
1174 int histo_bits, int cache_bits,
1175 VP8LHistogramSet* const image_histo,
1176 VP8LHistogram* const tmp_histo,
1177 uint16_t* const histogram_symbols) {
1178 int ok = 0;
1179 const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
1180 const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
1181 const int image_histo_raw_size = histo_xsize * histo_ysize;
1182 VP8LHistogramSet* const orig_histo =
1183 VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
1184 // Don't attempt linear bin-partition heuristic for
1185 // histograms of small sizes (as bin_map will be very sparse) and
1186 // maximum quality q==100 (to preserve the compression gains at that level).
1187 const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE;
1188 int entropy_combine;
1189 uint16_t* const map_tmp =
1190 WebPSafeMalloc(2 * image_histo_raw_size, sizeof(map_tmp));
1191 uint16_t* const cluster_mappings = map_tmp + image_histo_raw_size;
1192 int num_used = image_histo_raw_size;
1193 if (orig_histo == NULL || map_tmp == NULL) goto Error;
1194
1195 // Construct the histograms from backward references.
1196 HistogramBuild(xsize, histo_bits, refs, orig_histo);
1197 // Copies the histograms and computes its bit_cost.
1198 // histogram_symbols is optimized
1199 HistogramCopyAndAnalyze(orig_histo, image_histo, &num_used,
1200 histogram_symbols);
1201
1202 entropy_combine =
1203 (num_used > entropy_combine_num_bins * 2) && (quality < 100);
1204
1205 if (entropy_combine) {
1206 uint16_t* const bin_map = map_tmp;
1207 const double combine_cost_factor =
1208 GetCombineCostFactor(image_histo_raw_size, quality);
1209 const uint32_t num_clusters = num_used;
1210
1211 HistogramAnalyzeEntropyBin(image_histo, bin_map, low_effort);
1212 // Collapse histograms with similar entropy.
1213 HistogramCombineEntropyBin(image_histo, &num_used, histogram_symbols,
1214 cluster_mappings, tmp_histo, bin_map,
1215 entropy_combine_num_bins, combine_cost_factor,
1216 low_effort);
1217 OptimizeHistogramSymbols(image_histo, cluster_mappings, num_clusters,
1218 map_tmp, histogram_symbols);
1219 }
1220
1221 // Don't combine the histograms using stochastic and greedy heuristics for
1222 // low-effort compression mode.
1223 if (!low_effort || !entropy_combine) {
1224 const float x = quality / 100.f;
1225 // cubic ramp between 1 and MAX_HISTO_GREEDY:
1226 const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1));
1227 int do_greedy;
1228 if (!HistogramCombineStochastic(image_histo, &num_used, threshold_size,
1229 &do_greedy)) {
1230 goto Error;
1231 }
1232 if (do_greedy) {
1233 RemoveEmptyHistograms(image_histo);
1234 if (!HistogramCombineGreedy(image_histo, &num_used)) {
1235 goto Error;
1236 }
1237 }
1238 }
1239
1240 // Find the optimal map from original histograms to the final ones.
1241 RemoveEmptyHistograms(image_histo);
1242 HistogramRemap(orig_histo, image_histo, histogram_symbols);
1243
1244 ok = 1;
1245
1246 Error:
1247 VP8LFreeHistogramSet(orig_histo);
1248 WebPSafeFree(map_tmp);
1249 return ok;
1250 }
1251