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 | |
34 | static void HistogramClear(VP8LHistogram* const p) { |
35 | uint32_t* const literal = p->literal_; |
36 | const int cache_bits = p->palette_code_bits_; |
37 | const int histo_size = VP8LGetHistogramSize(cache_bits); |
38 | memset(p, 0, histo_size); |
39 | p->palette_code_bits_ = cache_bits; |
40 | p->literal_ = literal; |
41 | } |
42 | |
43 | // Swap two histogram pointers. |
44 | static void HistogramSwap(VP8LHistogram** const A, VP8LHistogram** const B) { |
45 | VP8LHistogram* const tmp = *A; |
46 | *A = *B; |
47 | *B = tmp; |
48 | } |
49 | |
50 | static void HistogramCopy(const VP8LHistogram* const src, |
51 | VP8LHistogram* const dst) { |
52 | uint32_t* const dst_literal = dst->literal_; |
53 | const int dst_cache_bits = dst->palette_code_bits_; |
54 | const int 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 | |
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 | |
69 | void VP8LFreeHistogram(VP8LHistogram* const histo) { |
70 | WebPSafeFree(histo); |
71 | } |
72 | |
73 | void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) { |
74 | WebPSafeFree(histo); |
75 | } |
76 | |
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 | |
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 | |
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 | |
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. |
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. |
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 | |
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 | |
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. |
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 | |
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 | ++histo->literal_[literal_ix]; |
212 | } else { |
213 | int code, ; |
214 | VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits); |
215 | ++histo->literal_[NUM_LITERAL_CODES + code]; |
216 | if (distance_modifier == NULL) { |
217 | VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits); |
218 | } else { |
219 | VP8LPrefixEncodeBits( |
220 | distance_modifier(distance_modifier_arg0, PixOrCopyDistance(v)), |
221 | &code, &extra_bits); |
222 | } |
223 | ++histo->distance_[code]; |
224 | } |
225 | } |
226 | |
227 | // ----------------------------------------------------------------------------- |
228 | // Entropy-related functions. |
229 | |
230 | static WEBP_INLINE double BitsEntropyRefine(const VP8LBitEntropy* entropy) { |
231 | double mix; |
232 | if (entropy->nonzeros < 5) { |
233 | if (entropy->nonzeros <= 1) { |
234 | return 0; |
235 | } |
236 | // Two symbols, they will be 0 and 1 in a Huffman code. |
237 | // Let's mix in a bit of entropy to favor good clustering when |
238 | // distributions of these are combined. |
239 | if (entropy->nonzeros == 2) { |
240 | return 0.99 * entropy->sum + 0.01 * entropy->entropy; |
241 | } |
242 | // No matter what the entropy says, we cannot be better than min_limit |
243 | // with Huffman coding. I am mixing a bit of entropy into the |
244 | // min_limit since it produces much better (~0.5 %) compression results |
245 | // perhaps because of better entropy clustering. |
246 | if (entropy->nonzeros == 3) { |
247 | mix = 0.95; |
248 | } else { |
249 | mix = 0.7; // nonzeros == 4. |
250 | } |
251 | } else { |
252 | mix = 0.627; |
253 | } |
254 | |
255 | { |
256 | double min_limit = 2 * entropy->sum - entropy->max_val; |
257 | min_limit = mix * min_limit + (1.0 - mix) * entropy->entropy; |
258 | return (entropy->entropy < min_limit) ? min_limit : entropy->entropy; |
259 | } |
260 | } |
261 | |
262 | double VP8LBitsEntropy(const uint32_t* const array, int n) { |
263 | VP8LBitEntropy entropy; |
264 | VP8LBitsEntropyUnrefined(array, n, &entropy); |
265 | |
266 | return BitsEntropyRefine(&entropy); |
267 | } |
268 | |
269 | static double InitialHuffmanCost(void) { |
270 | // Small bias because Huffman code length is typically not stored in |
271 | // full length. |
272 | static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3; |
273 | static const double kSmallBias = 9.1; |
274 | return kHuffmanCodeOfHuffmanCodeSize - kSmallBias; |
275 | } |
276 | |
277 | // Finalize the Huffman cost based on streak numbers and length type (<3 or >=3) |
278 | static double FinalHuffmanCost(const VP8LStreaks* const stats) { |
279 | // The constants in this function are experimental and got rounded from |
280 | // their original values in 1/8 when switched to 1/1024. |
281 | double retval = InitialHuffmanCost(); |
282 | // Second coefficient: Many zeros in the histogram are covered efficiently |
283 | // by a run-length encode. Originally 2/8. |
284 | retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1]; |
285 | // Second coefficient: Constant values are encoded less efficiently, but still |
286 | // RLE'ed. Originally 6/8. |
287 | retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1]; |
288 | // 0s are usually encoded more efficiently than non-0s. |
289 | // Originally 15/8. |
290 | retval += 1.796875 * stats->streaks[0][0]; |
291 | // Originally 26/8. |
292 | retval += 3.28125 * stats->streaks[1][0]; |
293 | return retval; |
294 | } |
295 | |
296 | // Get the symbol entropy for the distribution 'population'. |
297 | // Set 'trivial_sym', if there's only one symbol present in the distribution. |
298 | static double PopulationCost(const uint32_t* const population, int length, |
299 | uint32_t* const trivial_sym, |
300 | uint8_t* const is_used) { |
301 | VP8LBitEntropy bit_entropy; |
302 | VP8LStreaks stats; |
303 | VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats); |
304 | if (trivial_sym != NULL) { |
305 | *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code |
306 | : VP8L_NON_TRIVIAL_SYM; |
307 | } |
308 | // The histogram is used if there is at least one non-zero streak. |
309 | *is_used = (stats.streaks[1][0] != 0 || stats.streaks[1][1] != 0); |
310 | |
311 | return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); |
312 | } |
313 | |
314 | // trivial_at_end is 1 if the two histograms only have one element that is |
315 | // non-zero: both the zero-th one, or both the last one. |
316 | static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X, |
317 | const uint32_t* const Y, |
318 | int length, int is_X_used, |
319 | int is_Y_used, |
320 | int trivial_at_end) { |
321 | VP8LStreaks stats; |
322 | if (trivial_at_end) { |
323 | // This configuration is due to palettization that transforms an indexed |
324 | // pixel into 0xff000000 | (pixel << 8) in VP8LBundleColorMap. |
325 | // BitsEntropyRefine is 0 for histograms with only one non-zero value. |
326 | // Only FinalHuffmanCost needs to be evaluated. |
327 | memset(&stats, 0, sizeof(stats)); |
328 | // Deal with the non-zero value at index 0 or length-1. |
329 | stats.streaks[1][0] = 1; |
330 | // Deal with the following/previous zero streak. |
331 | stats.counts[0] = 1; |
332 | stats.streaks[0][1] = length - 1; |
333 | return FinalHuffmanCost(&stats); |
334 | } else { |
335 | VP8LBitEntropy bit_entropy; |
336 | if (is_X_used) { |
337 | if (is_Y_used) { |
338 | VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats); |
339 | } else { |
340 | VP8LGetEntropyUnrefined(X, length, &bit_entropy, &stats); |
341 | } |
342 | } else { |
343 | if (is_Y_used) { |
344 | VP8LGetEntropyUnrefined(Y, length, &bit_entropy, &stats); |
345 | } else { |
346 | memset(&stats, 0, sizeof(stats)); |
347 | stats.counts[0] = 1; |
348 | stats.streaks[0][length > 3] = length; |
349 | VP8LBitEntropyInit(&bit_entropy); |
350 | } |
351 | } |
352 | |
353 | return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); |
354 | } |
355 | } |
356 | |
357 | // Estimates the Entropy + Huffman + other block overhead size cost. |
358 | double VP8LHistogramEstimateBits(VP8LHistogram* const p) { |
359 | return |
360 | PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), |
361 | NULL, &p->is_used_[0]) |
362 | + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL, &p->is_used_[1]) |
363 | + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL, &p->is_used_[2]) |
364 | + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL, &p->is_used_[3]) |
365 | + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL, &p->is_used_[4]) |
366 | + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES) |
367 | + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES); |
368 | } |
369 | |
370 | // ----------------------------------------------------------------------------- |
371 | // Various histogram combine/cost-eval functions |
372 | |
373 | static int GetCombinedHistogramEntropy(const VP8LHistogram* const a, |
374 | const VP8LHistogram* const b, |
375 | double cost_threshold, |
376 | double* cost) { |
377 | const int palette_code_bits = a->palette_code_bits_; |
378 | int trivial_at_end = 0; |
379 | assert(a->palette_code_bits_ == b->palette_code_bits_); |
380 | *cost += GetCombinedEntropy(a->literal_, b->literal_, |
381 | VP8LHistogramNumCodes(palette_code_bits), |
382 | a->is_used_[0], b->is_used_[0], 0); |
383 | *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES, |
384 | b->literal_ + NUM_LITERAL_CODES, |
385 | NUM_LENGTH_CODES); |
386 | if (*cost > cost_threshold) return 0; |
387 | |
388 | if (a->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM && |
389 | a->trivial_symbol_ == b->trivial_symbol_) { |
390 | // A, R and B are all 0 or 0xff. |
391 | const uint32_t color_a = (a->trivial_symbol_ >> 24) & 0xff; |
392 | const uint32_t color_r = (a->trivial_symbol_ >> 16) & 0xff; |
393 | const uint32_t color_b = (a->trivial_symbol_ >> 0) & 0xff; |
394 | if ((color_a == 0 || color_a == 0xff) && |
395 | (color_r == 0 || color_r == 0xff) && |
396 | (color_b == 0 || color_b == 0xff)) { |
397 | trivial_at_end = 1; |
398 | } |
399 | } |
400 | |
401 | *cost += |
402 | GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, a->is_used_[1], |
403 | b->is_used_[1], trivial_at_end); |
404 | if (*cost > cost_threshold) return 0; |
405 | |
406 | *cost += |
407 | GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, a->is_used_[2], |
408 | b->is_used_[2], trivial_at_end); |
409 | if (*cost > cost_threshold) return 0; |
410 | |
411 | *cost += |
412 | GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES, |
413 | a->is_used_[3], b->is_used_[3], trivial_at_end); |
414 | if (*cost > cost_threshold) return 0; |
415 | |
416 | *cost += |
417 | GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES, |
418 | a->is_used_[4], b->is_used_[4], 0); |
419 | *cost += |
420 | VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES); |
421 | if (*cost > cost_threshold) return 0; |
422 | |
423 | return 1; |
424 | } |
425 | |
426 | static WEBP_INLINE void HistogramAdd(const VP8LHistogram* const a, |
427 | const VP8LHistogram* const b, |
428 | VP8LHistogram* const out) { |
429 | VP8LHistogramAdd(a, b, out); |
430 | out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_) |
431 | ? a->trivial_symbol_ |
432 | : VP8L_NON_TRIVIAL_SYM; |
433 | } |
434 | |
435 | // Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing |
436 | // to the threshold value 'cost_threshold'. The score returned is |
437 | // Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed. |
438 | // Since the previous score passed is 'cost_threshold', we only need to compare |
439 | // the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out |
440 | // early. |
441 | static double HistogramAddEval(const VP8LHistogram* const a, |
442 | const VP8LHistogram* const b, |
443 | VP8LHistogram* const out, |
444 | double cost_threshold) { |
445 | double cost = 0; |
446 | const double 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. |
461 | static double HistogramAddThresh(const VP8LHistogram* const a, |
462 | const VP8LHistogram* const b, |
463 | double cost_threshold) { |
464 | double 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 | // TODO(skal): Evaluate if float can be used here instead of double for |
476 | // representing the entropy costs. |
477 | typedef struct { |
478 | double literal_max_; |
479 | double literal_min_; |
480 | double red_max_; |
481 | double red_min_; |
482 | double blue_max_; |
483 | double blue_min_; |
484 | } DominantCostRange; |
485 | |
486 | static void DominantCostRangeInit(DominantCostRange* const c) { |
487 | c->literal_max_ = 0.; |
488 | c->literal_min_ = MAX_COST; |
489 | c->red_max_ = 0.; |
490 | c->red_min_ = MAX_COST; |
491 | c->blue_max_ = 0.; |
492 | c->blue_min_ = MAX_COST; |
493 | } |
494 | |
495 | static void UpdateDominantCostRange( |
496 | const VP8LHistogram* const h, DominantCostRange* const c) { |
497 | if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_; |
498 | if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_; |
499 | if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_; |
500 | if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_; |
501 | if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_; |
502 | if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_; |
503 | } |
504 | |
505 | static void UpdateHistogramCost(VP8LHistogram* const h) { |
506 | uint32_t alpha_sym, red_sym, blue_sym; |
507 | const double alpha_cost = |
508 | PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym, |
509 | &h->is_used_[3]); |
510 | const double distance_cost = |
511 | PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL, &h->is_used_[4]) + |
512 | VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES); |
513 | const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_); |
514 | h->literal_cost_ = |
515 | PopulationCost(h->literal_, num_codes, NULL, &h->is_used_[0]) + |
516 | VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES); |
517 | h->red_cost_ = |
518 | PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym, &h->is_used_[1]); |
519 | h->blue_cost_ = |
520 | PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym, &h->is_used_[2]); |
521 | h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ + |
522 | alpha_cost + distance_cost; |
523 | if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) { |
524 | h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM; |
525 | } else { |
526 | h->trivial_symbol_ = |
527 | ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0); |
528 | } |
529 | } |
530 | |
531 | static int GetBinIdForEntropy(double min, double max, double val) { |
532 | const double range = max - min; |
533 | if (range > 0.) { |
534 | const double delta = val - min; |
535 | return (int)((NUM_PARTITIONS - 1e-6) * delta / range); |
536 | } else { |
537 | return 0; |
538 | } |
539 | } |
540 | |
541 | static int GetHistoBinIndex(const VP8LHistogram* const h, |
542 | const DominantCostRange* const c, int low_effort) { |
543 | int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_, |
544 | h->literal_cost_); |
545 | assert(bin_id < NUM_PARTITIONS); |
546 | if (!low_effort) { |
547 | bin_id = bin_id * NUM_PARTITIONS |
548 | + GetBinIdForEntropy(c->red_min_, c->red_max_, h->red_cost_); |
549 | bin_id = bin_id * NUM_PARTITIONS |
550 | + GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_); |
551 | assert(bin_id < BIN_SIZE); |
552 | } |
553 | return bin_id; |
554 | } |
555 | |
556 | // Construct the histograms from backward references. |
557 | static void HistogramBuild( |
558 | int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs, |
559 | VP8LHistogramSet* const image_histo) { |
560 | int x = 0, y = 0; |
561 | const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits); |
562 | VP8LHistogram** const histograms = image_histo->histograms; |
563 | VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs); |
564 | assert(histo_bits > 0); |
565 | VP8LHistogramSetClear(image_histo); |
566 | while (VP8LRefsCursorOk(&c)) { |
567 | const PixOrCopy* const v = c.cur_pos; |
568 | const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits); |
569 | VP8LHistogramAddSinglePixOrCopy(histograms[ix], v, NULL, 0); |
570 | x += PixOrCopyLength(v); |
571 | while (x >= xsize) { |
572 | x -= xsize; |
573 | ++y; |
574 | } |
575 | VP8LRefsCursorNext(&c); |
576 | } |
577 | } |
578 | |
579 | // Copies the histograms and computes its bit_cost. |
580 | static const uint16_t kInvalidHistogramSymbol = (uint16_t)(-1); |
581 | static void HistogramCopyAndAnalyze(VP8LHistogramSet* const orig_histo, |
582 | VP8LHistogramSet* const image_histo, |
583 | int* const num_used, |
584 | uint16_t* const histogram_symbols) { |
585 | int i, cluster_id; |
586 | int num_used_orig = *num_used; |
587 | VP8LHistogram** const orig_histograms = orig_histo->histograms; |
588 | VP8LHistogram** const histograms = image_histo->histograms; |
589 | assert(image_histo->max_size == orig_histo->max_size); |
590 | for (cluster_id = 0, i = 0; i < orig_histo->max_size; ++i) { |
591 | VP8LHistogram* const histo = orig_histograms[i]; |
592 | UpdateHistogramCost(histo); |
593 | |
594 | // Skip the histogram if it is completely empty, which can happen for tiles |
595 | // with no information (when they are skipped because of LZ77). |
596 | if (!histo->is_used_[0] && !histo->is_used_[1] && !histo->is_used_[2] |
597 | && !histo->is_used_[3] && !histo->is_used_[4]) { |
598 | // The first histogram is always used. If an histogram is empty, we set |
599 | // its id to be the same as the previous one: this will improve |
600 | // compressibility for later LZ77. |
601 | assert(i > 0); |
602 | HistogramSetRemoveHistogram(image_histo, i, num_used); |
603 | HistogramSetRemoveHistogram(orig_histo, i, &num_used_orig); |
604 | histogram_symbols[i] = kInvalidHistogramSymbol; |
605 | } else { |
606 | // Copy histograms from orig_histo[] to image_histo[]. |
607 | HistogramCopy(histo, histograms[i]); |
608 | histogram_symbols[i] = cluster_id++; |
609 | assert(cluster_id <= image_histo->max_size); |
610 | } |
611 | } |
612 | } |
613 | |
614 | // Partition histograms to different entropy bins for three dominant (literal, |
615 | // red and blue) symbol costs and compute the histogram aggregate bit_cost. |
616 | static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo, |
617 | uint16_t* const bin_map, |
618 | int low_effort) { |
619 | int i; |
620 | VP8LHistogram** const histograms = image_histo->histograms; |
621 | const int histo_size = image_histo->size; |
622 | DominantCostRange cost_range; |
623 | DominantCostRangeInit(&cost_range); |
624 | |
625 | // Analyze the dominant (literal, red and blue) entropy costs. |
626 | for (i = 0; i < histo_size; ++i) { |
627 | if (histograms[i] == NULL) continue; |
628 | UpdateDominantCostRange(histograms[i], &cost_range); |
629 | } |
630 | |
631 | // bin-hash histograms on three of the dominant (literal, red and blue) |
632 | // symbol costs and store the resulting bin_id for each histogram. |
633 | for (i = 0; i < histo_size; ++i) { |
634 | // bin_map[i] is not set to a special value as its use will later be guarded |
635 | // by another (histograms[i] == NULL). |
636 | if (histograms[i] == NULL) continue; |
637 | bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort); |
638 | } |
639 | } |
640 | |
641 | // Merges some histograms with same bin_id together if it's advantageous. |
642 | // Sets the remaining histograms to NULL. |
643 | static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo, |
644 | int* num_used, |
645 | const uint16_t* const clusters, |
646 | uint16_t* const cluster_mappings, |
647 | VP8LHistogram* cur_combo, |
648 | const uint16_t* const bin_map, |
649 | int num_bins, |
650 | double combine_cost_factor, |
651 | int low_effort) { |
652 | VP8LHistogram** const histograms = image_histo->histograms; |
653 | int idx; |
654 | struct { |
655 | int16_t first; // position of the histogram that accumulates all |
656 | // histograms with the same bin_id |
657 | uint16_t num_combine_failures; // number of combine failures per bin_id |
658 | } bin_info[BIN_SIZE]; |
659 | |
660 | assert(num_bins <= BIN_SIZE); |
661 | for (idx = 0; idx < num_bins; ++idx) { |
662 | bin_info[idx].first = -1; |
663 | bin_info[idx].num_combine_failures = 0; |
664 | } |
665 | |
666 | // By default, a cluster matches itself. |
667 | for (idx = 0; idx < *num_used; ++idx) cluster_mappings[idx] = idx; |
668 | for (idx = 0; idx < image_histo->size; ++idx) { |
669 | int bin_id, first; |
670 | if (histograms[idx] == NULL) continue; |
671 | bin_id = bin_map[idx]; |
672 | first = bin_info[bin_id].first; |
673 | if (first == -1) { |
674 | bin_info[bin_id].first = idx; |
675 | } else if (low_effort) { |
676 | HistogramAdd(histograms[idx], histograms[first], histograms[first]); |
677 | HistogramSetRemoveHistogram(image_histo, idx, num_used); |
678 | cluster_mappings[clusters[idx]] = clusters[first]; |
679 | } else { |
680 | // try to merge #idx into #first (both share the same bin_id) |
681 | const double bit_cost = histograms[idx]->bit_cost_; |
682 | const double bit_cost_thresh = -bit_cost * combine_cost_factor; |
683 | const double curr_cost_diff = |
684 | HistogramAddEval(histograms[first], histograms[idx], |
685 | cur_combo, bit_cost_thresh); |
686 | if (curr_cost_diff < bit_cost_thresh) { |
687 | // Try to merge two histograms only if the combo is a trivial one or |
688 | // the two candidate histograms are already non-trivial. |
689 | // For some images, 'try_combine' turns out to be false for a lot of |
690 | // histogram pairs. In that case, we fallback to combining |
691 | // histograms as usual to avoid increasing the header size. |
692 | const int try_combine = |
693 | (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) || |
694 | ((histograms[idx]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) && |
695 | (histograms[first]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM)); |
696 | const int max_combine_failures = 32; |
697 | if (try_combine || |
698 | bin_info[bin_id].num_combine_failures >= max_combine_failures) { |
699 | // move the (better) merged histogram to its final slot |
700 | HistogramSwap(&cur_combo, &histograms[first]); |
701 | HistogramSetRemoveHistogram(image_histo, idx, num_used); |
702 | cluster_mappings[clusters[idx]] = clusters[first]; |
703 | } else { |
704 | ++bin_info[bin_id].num_combine_failures; |
705 | } |
706 | } |
707 | } |
708 | } |
709 | if (low_effort) { |
710 | // for low_effort case, update the final cost when everything is merged |
711 | for (idx = 0; idx < image_histo->size; ++idx) { |
712 | if (histograms[idx] == NULL) continue; |
713 | UpdateHistogramCost(histograms[idx]); |
714 | } |
715 | } |
716 | } |
717 | |
718 | // Implement a Lehmer random number generator with a multiplicative constant of |
719 | // 48271 and a modulo constant of 2^31 - 1. |
720 | static uint32_t MyRand(uint32_t* const seed) { |
721 | *seed = (uint32_t)(((uint64_t)(*seed) * 48271u) % 2147483647u); |
722 | assert(*seed > 0); |
723 | return *seed; |
724 | } |
725 | |
726 | // ----------------------------------------------------------------------------- |
727 | // Histogram pairs priority queue |
728 | |
729 | // Pair of histograms. Negative idx1 value means that pair is out-of-date. |
730 | typedef struct { |
731 | int idx1; |
732 | int idx2; |
733 | double cost_diff; |
734 | double cost_combo; |
735 | } HistogramPair; |
736 | |
737 | typedef struct { |
738 | HistogramPair* queue; |
739 | int size; |
740 | int max_size; |
741 | } HistoQueue; |
742 | |
743 | static int HistoQueueInit(HistoQueue* const histo_queue, const int max_size) { |
744 | histo_queue->size = 0; |
745 | histo_queue->max_size = max_size; |
746 | // We allocate max_size + 1 because the last element at index "size" is |
747 | // used as temporary data (and it could be up to max_size). |
748 | histo_queue->queue = (HistogramPair*)WebPSafeMalloc( |
749 | histo_queue->max_size + 1, sizeof(*histo_queue->queue)); |
750 | return histo_queue->queue != NULL; |
751 | } |
752 | |
753 | static void HistoQueueClear(HistoQueue* const histo_queue) { |
754 | assert(histo_queue != NULL); |
755 | WebPSafeFree(histo_queue->queue); |
756 | histo_queue->size = 0; |
757 | histo_queue->max_size = 0; |
758 | } |
759 | |
760 | // Pop a specific pair in the queue by replacing it with the last one |
761 | // and shrinking the queue. |
762 | static void HistoQueuePopPair(HistoQueue* const histo_queue, |
763 | HistogramPair* const pair) { |
764 | assert(pair >= histo_queue->queue && |
765 | pair < (histo_queue->queue + histo_queue->size)); |
766 | assert(histo_queue->size > 0); |
767 | *pair = histo_queue->queue[histo_queue->size - 1]; |
768 | --histo_queue->size; |
769 | } |
770 | |
771 | // Check whether a pair in the queue should be updated as head or not. |
772 | static void HistoQueueUpdateHead(HistoQueue* const histo_queue, |
773 | HistogramPair* const pair) { |
774 | assert(pair->cost_diff < 0.); |
775 | assert(pair >= histo_queue->queue && |
776 | pair < (histo_queue->queue + histo_queue->size)); |
777 | assert(histo_queue->size > 0); |
778 | if (pair->cost_diff < histo_queue->queue[0].cost_diff) { |
779 | // Replace the best pair. |
780 | const HistogramPair tmp = histo_queue->queue[0]; |
781 | histo_queue->queue[0] = *pair; |
782 | *pair = tmp; |
783 | } |
784 | } |
785 | |
786 | // Update the cost diff and combo of a pair of histograms. This needs to be |
787 | // called when the the histograms have been merged with a third one. |
788 | static void HistoQueueUpdatePair(const VP8LHistogram* const h1, |
789 | const VP8LHistogram* const h2, |
790 | double threshold, |
791 | HistogramPair* const pair) { |
792 | const double sum_cost = h1->bit_cost_ + h2->bit_cost_; |
793 | pair->cost_combo = 0.; |
794 | GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair->cost_combo); |
795 | pair->cost_diff = pair->cost_combo - sum_cost; |
796 | } |
797 | |
798 | // Create a pair from indices "idx1" and "idx2" provided its cost |
799 | // is inferior to "threshold", a negative entropy. |
800 | // It returns the cost of the pair, or 0. if it superior to threshold. |
801 | static double HistoQueuePush(HistoQueue* const histo_queue, |
802 | VP8LHistogram** const histograms, int idx1, |
803 | int idx2, double threshold) { |
804 | const VP8LHistogram* h1; |
805 | const VP8LHistogram* h2; |
806 | HistogramPair pair; |
807 | |
808 | // Stop here if the queue is full. |
809 | if (histo_queue->size == histo_queue->max_size) return 0.; |
810 | assert(threshold <= 0.); |
811 | if (idx1 > idx2) { |
812 | const int tmp = idx2; |
813 | idx2 = idx1; |
814 | idx1 = tmp; |
815 | } |
816 | pair.idx1 = idx1; |
817 | pair.idx2 = idx2; |
818 | h1 = histograms[idx1]; |
819 | h2 = histograms[idx2]; |
820 | |
821 | HistoQueueUpdatePair(h1, h2, threshold, &pair); |
822 | |
823 | // Do not even consider the pair if it does not improve the entropy. |
824 | if (pair.cost_diff >= threshold) return 0.; |
825 | |
826 | histo_queue->queue[histo_queue->size++] = pair; |
827 | HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]); |
828 | |
829 | return pair.cost_diff; |
830 | } |
831 | |
832 | // ----------------------------------------------------------------------------- |
833 | |
834 | // Combines histograms by continuously choosing the one with the highest cost |
835 | // reduction. |
836 | static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo, |
837 | int* const num_used) { |
838 | int ok = 0; |
839 | const int image_histo_size = image_histo->size; |
840 | int i, j; |
841 | VP8LHistogram** const histograms = image_histo->histograms; |
842 | // Priority queue of histogram pairs. |
843 | HistoQueue histo_queue; |
844 | |
845 | // image_histo_size^2 for the queue size is safe. If you look at |
846 | // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes |
847 | // data to the queue, you insert at most: |
848 | // - image_histo_size*(image_histo_size-1)/2 (the first two for loops) |
849 | // - image_histo_size - 1 in the last for loop at the first iteration of |
850 | // the while loop, image_histo_size - 2 at the second iteration ... |
851 | // therefore image_histo_size*(image_histo_size-1)/2 overall too |
852 | if (!HistoQueueInit(&histo_queue, image_histo_size * image_histo_size)) { |
853 | goto End; |
854 | } |
855 | |
856 | for (i = 0; i < image_histo_size; ++i) { |
857 | if (image_histo->histograms[i] == NULL) continue; |
858 | for (j = i + 1; j < image_histo_size; ++j) { |
859 | // Initialize queue. |
860 | if (image_histo->histograms[j] == NULL) continue; |
861 | HistoQueuePush(&histo_queue, histograms, i, j, 0.); |
862 | } |
863 | } |
864 | |
865 | while (histo_queue.size > 0) { |
866 | const int idx1 = histo_queue.queue[0].idx1; |
867 | const int idx2 = histo_queue.queue[0].idx2; |
868 | HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]); |
869 | histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo; |
870 | |
871 | // Remove merged histogram. |
872 | HistogramSetRemoveHistogram(image_histo, idx2, num_used); |
873 | |
874 | // Remove pairs intersecting the just combined best pair. |
875 | for (i = 0; i < histo_queue.size;) { |
876 | HistogramPair* const p = histo_queue.queue + i; |
877 | if (p->idx1 == idx1 || p->idx2 == idx1 || |
878 | p->idx1 == idx2 || p->idx2 == idx2) { |
879 | HistoQueuePopPair(&histo_queue, p); |
880 | } else { |
881 | HistoQueueUpdateHead(&histo_queue, p); |
882 | ++i; |
883 | } |
884 | } |
885 | |
886 | // Push new pairs formed with combined histogram to the queue. |
887 | for (i = 0; i < image_histo->size; ++i) { |
888 | if (i == idx1 || image_histo->histograms[i] == NULL) continue; |
889 | HistoQueuePush(&histo_queue, image_histo->histograms, idx1, i, 0.); |
890 | } |
891 | } |
892 | |
893 | ok = 1; |
894 | |
895 | End: |
896 | HistoQueueClear(&histo_queue); |
897 | return ok; |
898 | } |
899 | |
900 | // Perform histogram aggregation using a stochastic approach. |
901 | // 'do_greedy' is set to 1 if a greedy approach needs to be performed |
902 | // afterwards, 0 otherwise. |
903 | static int PairComparison(const void* idx1, const void* idx2) { |
904 | // To be used with bsearch: <0 when *idx1<*idx2, >0 if >, 0 when ==. |
905 | return (*(int*) idx1 - *(int*) idx2); |
906 | } |
907 | static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, |
908 | int* const num_used, int min_cluster_size, |
909 | int* const do_greedy) { |
910 | int j, iter; |
911 | uint32_t seed = 1; |
912 | int tries_with_no_success = 0; |
913 | const int outer_iters = *num_used; |
914 | const int num_tries_no_success = outer_iters / 2; |
915 | VP8LHistogram** const histograms = image_histo->histograms; |
916 | // Priority queue of histogram pairs. Its size of 'kHistoQueueSize' |
917 | // impacts the quality of the compression and the speed: the smaller the |
918 | // faster but the worse for the compression. |
919 | HistoQueue histo_queue; |
920 | const int kHistoQueueSize = 9; |
921 | int ok = 0; |
922 | // mapping from an index in image_histo with no NULL histogram to the full |
923 | // blown image_histo. |
924 | int* mappings; |
925 | |
926 | if (*num_used < min_cluster_size) { |
927 | *do_greedy = 1; |
928 | return 1; |
929 | } |
930 | |
931 | mappings = (int*) WebPSafeMalloc(*num_used, sizeof(*mappings)); |
932 | if (mappings == NULL) return 0; |
933 | if (!HistoQueueInit(&histo_queue, kHistoQueueSize)) goto End; |
934 | // Fill the initial mapping. |
935 | for (j = 0, iter = 0; iter < image_histo->size; ++iter) { |
936 | if (histograms[iter] == NULL) continue; |
937 | mappings[j++] = iter; |
938 | } |
939 | assert(j == *num_used); |
940 | |
941 | // Collapse similar histograms in 'image_histo'. |
942 | for (iter = 0; |
943 | iter < outer_iters && *num_used >= min_cluster_size && |
944 | ++tries_with_no_success < num_tries_no_success; |
945 | ++iter) { |
946 | int* mapping_index; |
947 | double best_cost = |
948 | (histo_queue.size == 0) ? 0. : histo_queue.queue[0].cost_diff; |
949 | int best_idx1 = -1, best_idx2 = 1; |
950 | const uint32_t rand_range = (*num_used - 1) * (*num_used); |
951 | // (*num_used) / 2 was chosen empirically. Less means faster but worse |
952 | // compression. |
953 | const int num_tries = (*num_used) / 2; |
954 | |
955 | // Pick random samples. |
956 | for (j = 0; *num_used >= 2 && j < num_tries; ++j) { |
957 | double curr_cost; |
958 | // Choose two different histograms at random and try to combine them. |
959 | const uint32_t tmp = MyRand(&seed) % rand_range; |
960 | uint32_t idx1 = tmp / (*num_used - 1); |
961 | uint32_t idx2 = tmp % (*num_used - 1); |
962 | if (idx2 >= idx1) ++idx2; |
963 | idx1 = mappings[idx1]; |
964 | idx2 = mappings[idx2]; |
965 | |
966 | // Calculate cost reduction on combination. |
967 | curr_cost = |
968 | HistoQueuePush(&histo_queue, histograms, idx1, idx2, best_cost); |
969 | if (curr_cost < 0) { // found a better pair? |
970 | best_cost = curr_cost; |
971 | // Empty the queue if we reached full capacity. |
972 | if (histo_queue.size == histo_queue.max_size) break; |
973 | } |
974 | } |
975 | if (histo_queue.size == 0) continue; |
976 | |
977 | // Get the best histograms. |
978 | best_idx1 = histo_queue.queue[0].idx1; |
979 | best_idx2 = histo_queue.queue[0].idx2; |
980 | assert(best_idx1 < best_idx2); |
981 | // Pop best_idx2 from mappings. |
982 | mapping_index = (int*) bsearch(&best_idx2, mappings, *num_used, |
983 | sizeof(best_idx2), &PairComparison); |
984 | assert(mapping_index != NULL); |
985 | memmove(mapping_index, mapping_index + 1, sizeof(*mapping_index) * |
986 | ((*num_used) - (mapping_index - mappings) - 1)); |
987 | // Merge the histograms and remove best_idx2 from the queue. |
988 | HistogramAdd(histograms[best_idx2], histograms[best_idx1], |
989 | histograms[best_idx1]); |
990 | histograms[best_idx1]->bit_cost_ = histo_queue.queue[0].cost_combo; |
991 | HistogramSetRemoveHistogram(image_histo, best_idx2, num_used); |
992 | // Parse the queue and update each pair that deals with best_idx1, |
993 | // best_idx2 or image_histo_size. |
994 | for (j = 0; j < histo_queue.size;) { |
995 | HistogramPair* const p = histo_queue.queue + j; |
996 | const int is_idx1_best = p->idx1 == best_idx1 || p->idx1 == best_idx2; |
997 | const int is_idx2_best = p->idx2 == best_idx1 || p->idx2 == best_idx2; |
998 | int do_eval = 0; |
999 | // The front pair could have been duplicated by a random pick so |
1000 | // check for it all the time nevertheless. |
1001 | if (is_idx1_best && is_idx2_best) { |
1002 | HistoQueuePopPair(&histo_queue, p); |
1003 | continue; |
1004 | } |
1005 | // Any pair containing one of the two best indices should only refer to |
1006 | // best_idx1. Its cost should also be updated. |
1007 | if (is_idx1_best) { |
1008 | p->idx1 = best_idx1; |
1009 | do_eval = 1; |
1010 | } else if (is_idx2_best) { |
1011 | p->idx2 = best_idx1; |
1012 | do_eval = 1; |
1013 | } |
1014 | // Make sure the index order is respected. |
1015 | if (p->idx1 > p->idx2) { |
1016 | const int tmp = p->idx2; |
1017 | p->idx2 = p->idx1; |
1018 | p->idx1 = tmp; |
1019 | } |
1020 | if (do_eval) { |
1021 | // Re-evaluate the cost of an updated pair. |
1022 | HistoQueueUpdatePair(histograms[p->idx1], histograms[p->idx2], 0., p); |
1023 | if (p->cost_diff >= 0.) { |
1024 | HistoQueuePopPair(&histo_queue, p); |
1025 | continue; |
1026 | } |
1027 | } |
1028 | HistoQueueUpdateHead(&histo_queue, p); |
1029 | ++j; |
1030 | } |
1031 | tries_with_no_success = 0; |
1032 | } |
1033 | *do_greedy = (*num_used <= min_cluster_size); |
1034 | ok = 1; |
1035 | |
1036 | End: |
1037 | HistoQueueClear(&histo_queue); |
1038 | WebPSafeFree(mappings); |
1039 | return ok; |
1040 | } |
1041 | |
1042 | // ----------------------------------------------------------------------------- |
1043 | // Histogram refinement |
1044 | |
1045 | // Find the best 'out' histogram for each of the 'in' histograms. |
1046 | // At call-time, 'out' contains the histograms of the clusters. |
1047 | // Note: we assume that out[]->bit_cost_ is already up-to-date. |
1048 | static void HistogramRemap(const VP8LHistogramSet* const in, |
1049 | VP8LHistogramSet* const out, |
1050 | uint16_t* const symbols) { |
1051 | int i; |
1052 | VP8LHistogram** const in_histo = in->histograms; |
1053 | VP8LHistogram** const out_histo = out->histograms; |
1054 | const int in_size = out->max_size; |
1055 | const int out_size = out->size; |
1056 | if (out_size > 1) { |
1057 | for (i = 0; i < in_size; ++i) { |
1058 | int best_out = 0; |
1059 | double best_bits = MAX_COST; |
1060 | int k; |
1061 | if (in_histo[i] == NULL) { |
1062 | // Arbitrarily set to the previous value if unused to help future LZ77. |
1063 | symbols[i] = symbols[i - 1]; |
1064 | continue; |
1065 | } |
1066 | for (k = 0; k < out_size; ++k) { |
1067 | double cur_bits; |
1068 | cur_bits = HistogramAddThresh(out_histo[k], in_histo[i], best_bits); |
1069 | if (k == 0 || cur_bits < best_bits) { |
1070 | best_bits = cur_bits; |
1071 | best_out = k; |
1072 | } |
1073 | } |
1074 | symbols[i] = best_out; |
1075 | } |
1076 | } else { |
1077 | assert(out_size == 1); |
1078 | for (i = 0; i < in_size; ++i) { |
1079 | symbols[i] = 0; |
1080 | } |
1081 | } |
1082 | |
1083 | // Recompute each out based on raw and symbols. |
1084 | VP8LHistogramSetClear(out); |
1085 | out->size = out_size; |
1086 | |
1087 | for (i = 0; i < in_size; ++i) { |
1088 | int idx; |
1089 | if (in_histo[i] == NULL) continue; |
1090 | idx = symbols[i]; |
1091 | HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]); |
1092 | } |
1093 | } |
1094 | |
1095 | static double GetCombineCostFactor(int histo_size, int quality) { |
1096 | double combine_cost_factor = 0.16; |
1097 | if (quality < 90) { |
1098 | if (histo_size > 256) combine_cost_factor /= 2.; |
1099 | if (histo_size > 512) combine_cost_factor /= 2.; |
1100 | if (histo_size > 1024) combine_cost_factor /= 2.; |
1101 | if (quality <= 50) combine_cost_factor /= 2.; |
1102 | } |
1103 | return combine_cost_factor; |
1104 | } |
1105 | |
1106 | // Given a HistogramSet 'set', the mapping of clusters 'cluster_mapping' and the |
1107 | // current assignment of the cells in 'symbols', merge the clusters and |
1108 | // assign the smallest possible clusters values. |
1109 | static void OptimizeHistogramSymbols(const VP8LHistogramSet* const set, |
1110 | uint16_t* const cluster_mappings, |
1111 | int num_clusters, |
1112 | uint16_t* const cluster_mappings_tmp, |
1113 | uint16_t* const symbols) { |
1114 | int i, cluster_max; |
1115 | int do_continue = 1; |
1116 | // First, assign the lowest cluster to each pixel. |
1117 | while (do_continue) { |
1118 | do_continue = 0; |
1119 | for (i = 0; i < num_clusters; ++i) { |
1120 | int k; |
1121 | k = cluster_mappings[i]; |
1122 | while (k != cluster_mappings[k]) { |
1123 | cluster_mappings[k] = cluster_mappings[cluster_mappings[k]]; |
1124 | k = cluster_mappings[k]; |
1125 | } |
1126 | if (k != cluster_mappings[i]) { |
1127 | do_continue = 1; |
1128 | cluster_mappings[i] = k; |
1129 | } |
1130 | } |
1131 | } |
1132 | // Create a mapping from a cluster id to its minimal version. |
1133 | cluster_max = 0; |
1134 | memset(cluster_mappings_tmp, 0, |
1135 | set->max_size * sizeof(*cluster_mappings_tmp)); |
1136 | assert(cluster_mappings[0] == 0); |
1137 | // Re-map the ids. |
1138 | for (i = 0; i < set->max_size; ++i) { |
1139 | int cluster; |
1140 | if (symbols[i] == kInvalidHistogramSymbol) continue; |
1141 | cluster = cluster_mappings[symbols[i]]; |
1142 | assert(symbols[i] < num_clusters); |
1143 | if (cluster > 0 && cluster_mappings_tmp[cluster] == 0) { |
1144 | ++cluster_max; |
1145 | cluster_mappings_tmp[cluster] = cluster_max; |
1146 | } |
1147 | symbols[i] = cluster_mappings_tmp[cluster]; |
1148 | } |
1149 | |
1150 | // Make sure all cluster values are used. |
1151 | cluster_max = 0; |
1152 | for (i = 0; i < set->max_size; ++i) { |
1153 | if (symbols[i] == kInvalidHistogramSymbol) continue; |
1154 | if (symbols[i] <= cluster_max) continue; |
1155 | ++cluster_max; |
1156 | assert(symbols[i] == cluster_max); |
1157 | } |
1158 | } |
1159 | |
1160 | static void RemoveEmptyHistograms(VP8LHistogramSet* const image_histo) { |
1161 | uint32_t size; |
1162 | int i; |
1163 | for (i = 0, size = 0; i < image_histo->size; ++i) { |
1164 | if (image_histo->histograms[i] == NULL) continue; |
1165 | image_histo->histograms[size++] = image_histo->histograms[i]; |
1166 | } |
1167 | image_histo->size = size; |
1168 | } |
1169 | |
1170 | int VP8LGetHistoImageSymbols(int xsize, int ysize, |
1171 | const VP8LBackwardRefs* const refs, |
1172 | int quality, int low_effort, |
1173 | int histo_bits, int cache_bits, |
1174 | VP8LHistogramSet* const image_histo, |
1175 | VP8LHistogram* const tmp_histo, |
1176 | uint16_t* const histogram_symbols) { |
1177 | int ok = 0; |
1178 | const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1; |
1179 | const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1; |
1180 | const int image_histo_raw_size = histo_xsize * histo_ysize; |
1181 | VP8LHistogramSet* const orig_histo = |
1182 | VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits); |
1183 | // Don't attempt linear bin-partition heuristic for |
1184 | // histograms of small sizes (as bin_map will be very sparse) and |
1185 | // maximum quality q==100 (to preserve the compression gains at that level). |
1186 | const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE; |
1187 | int entropy_combine; |
1188 | uint16_t* const map_tmp = |
1189 | WebPSafeMalloc(2 * image_histo_raw_size, sizeof(map_tmp)); |
1190 | uint16_t* const cluster_mappings = map_tmp + image_histo_raw_size; |
1191 | int num_used = image_histo_raw_size; |
1192 | if (orig_histo == NULL || map_tmp == NULL) goto Error; |
1193 | |
1194 | // Construct the histograms from backward references. |
1195 | HistogramBuild(xsize, histo_bits, refs, orig_histo); |
1196 | // Copies the histograms and computes its bit_cost. |
1197 | // histogram_symbols is optimized |
1198 | HistogramCopyAndAnalyze(orig_histo, image_histo, &num_used, |
1199 | histogram_symbols); |
1200 | |
1201 | entropy_combine = |
1202 | (num_used > entropy_combine_num_bins * 2) && (quality < 100); |
1203 | |
1204 | if (entropy_combine) { |
1205 | uint16_t* const bin_map = map_tmp; |
1206 | const double combine_cost_factor = |
1207 | GetCombineCostFactor(image_histo_raw_size, quality); |
1208 | const uint32_t num_clusters = num_used; |
1209 | |
1210 | HistogramAnalyzeEntropyBin(image_histo, bin_map, low_effort); |
1211 | // Collapse histograms with similar entropy. |
1212 | HistogramCombineEntropyBin(image_histo, &num_used, histogram_symbols, |
1213 | cluster_mappings, tmp_histo, bin_map, |
1214 | entropy_combine_num_bins, combine_cost_factor, |
1215 | low_effort); |
1216 | OptimizeHistogramSymbols(image_histo, cluster_mappings, num_clusters, |
1217 | map_tmp, histogram_symbols); |
1218 | } |
1219 | |
1220 | // Don't combine the histograms using stochastic and greedy heuristics for |
1221 | // low-effort compression mode. |
1222 | if (!low_effort || !entropy_combine) { |
1223 | const float x = quality / 100.f; |
1224 | // cubic ramp between 1 and MAX_HISTO_GREEDY: |
1225 | const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1)); |
1226 | int do_greedy; |
1227 | if (!HistogramCombineStochastic(image_histo, &num_used, threshold_size, |
1228 | &do_greedy)) { |
1229 | goto Error; |
1230 | } |
1231 | if (do_greedy) { |
1232 | RemoveEmptyHistograms(image_histo); |
1233 | if (!HistogramCombineGreedy(image_histo, &num_used)) { |
1234 | goto Error; |
1235 | } |
1236 | } |
1237 | } |
1238 | |
1239 | // Find the optimal map from original histograms to the final ones. |
1240 | RemoveEmptyHistograms(image_histo); |
1241 | HistogramRemap(orig_histo, image_histo, histogram_symbols); |
1242 | |
1243 | ok = 1; |
1244 | |
1245 | Error: |
1246 | VP8LFreeHistogramSet(orig_histo); |
1247 | WebPSafeFree(map_tmp); |
1248 | return ok; |
1249 | } |
1250 | |