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 histo_size = VP8LGetHistogramSize(dst_cache_bits);
55 assert(src->palette_code_bits_ == dst_cache_bits);
56 memcpy(dst, src, histo_size);
57 dst->literal_ = dst_literal;
58 }
59
VP8LGetHistogramSize(int cache_bits)60 int VP8LGetHistogramSize(int cache_bits) {
61 const int literal_size = VP8LHistogramNumCodes(cache_bits);
62 const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size;
63 assert(total_size <= (size_t)0x7fffffff);
64 return (int)total_size;
65 }
66
VP8LFreeHistogram(VP8LHistogram * const histo)67 void VP8LFreeHistogram(VP8LHistogram* const histo) {
68 WebPSafeFree(histo);
69 }
70
VP8LFreeHistogramSet(VP8LHistogramSet * const histo)71 void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
72 WebPSafeFree(histo);
73 }
74
VP8LHistogramStoreRefs(const VP8LBackwardRefs * const refs,VP8LHistogram * const histo)75 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
76 VP8LHistogram* const histo) {
77 VP8LRefsCursor c = VP8LRefsCursorInit(refs);
78 while (VP8LRefsCursorOk(&c)) {
79 VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos, NULL, 0);
80 VP8LRefsCursorNext(&c);
81 }
82 }
83
VP8LHistogramCreate(VP8LHistogram * const p,const VP8LBackwardRefs * const refs,int palette_code_bits)84 void VP8LHistogramCreate(VP8LHistogram* const p,
85 const VP8LBackwardRefs* const refs,
86 int palette_code_bits) {
87 if (palette_code_bits >= 0) {
88 p->palette_code_bits_ = palette_code_bits;
89 }
90 HistogramClear(p);
91 VP8LHistogramStoreRefs(refs, p);
92 }
93
VP8LHistogramInit(VP8LHistogram * const p,int palette_code_bits)94 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) {
95 p->palette_code_bits_ = palette_code_bits;
96 HistogramClear(p);
97 }
98
VP8LAllocateHistogram(int cache_bits)99 VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
100 VP8LHistogram* histo = NULL;
101 const int total_size = VP8LGetHistogramSize(cache_bits);
102 uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
103 if (memory == NULL) return NULL;
104 histo = (VP8LHistogram*)memory;
105 // literal_ won't necessary be aligned.
106 histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
107 VP8LHistogramInit(histo, cache_bits);
108 return histo;
109 }
110
VP8LAllocateHistogramSet(int size,int cache_bits)111 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
112 int i;
113 VP8LHistogramSet* set;
114 const int histo_size = VP8LGetHistogramSize(cache_bits);
115 const size_t total_size =
116 sizeof(*set) + size * (sizeof(*set->histograms) +
117 histo_size + WEBP_ALIGN_CST);
118 uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
119 if (memory == NULL) return NULL;
120
121 set = (VP8LHistogramSet*)memory;
122 memory += sizeof(*set);
123 set->histograms = (VP8LHistogram**)memory;
124 memory += size * sizeof(*set->histograms);
125 set->max_size = size;
126 set->size = size;
127 for (i = 0; i < size; ++i) {
128 memory = (uint8_t*)WEBP_ALIGN(memory);
129 set->histograms[i] = (VP8LHistogram*)memory;
130 // literal_ won't necessary be aligned.
131 set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
132 VP8LHistogramInit(set->histograms[i], cache_bits);
133 memory += histo_size;
134 }
135 return set;
136 }
137
138 // -----------------------------------------------------------------------------
139
VP8LHistogramAddSinglePixOrCopy(VP8LHistogram * const histo,const PixOrCopy * const v,int (* const distance_modifier)(int,int),int distance_modifier_arg0)140 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
141 const PixOrCopy* const v,
142 int (*const distance_modifier)(int, int),
143 int distance_modifier_arg0) {
144 if (PixOrCopyIsLiteral(v)) {
145 ++histo->alpha_[PixOrCopyLiteral(v, 3)];
146 ++histo->red_[PixOrCopyLiteral(v, 2)];
147 ++histo->literal_[PixOrCopyLiteral(v, 1)];
148 ++histo->blue_[PixOrCopyLiteral(v, 0)];
149 } else if (PixOrCopyIsCacheIdx(v)) {
150 const int literal_ix =
151 NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
152 ++histo->literal_[literal_ix];
153 } else {
154 int code, extra_bits;
155 VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
156 ++histo->literal_[NUM_LITERAL_CODES + code];
157 if (distance_modifier == NULL) {
158 VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
159 } else {
160 VP8LPrefixEncodeBits(
161 distance_modifier(distance_modifier_arg0, PixOrCopyDistance(v)),
162 &code, &extra_bits);
163 }
164 ++histo->distance_[code];
165 }
166 }
167
168 // -----------------------------------------------------------------------------
169 // Entropy-related functions.
170
BitsEntropyRefine(const VP8LBitEntropy * entropy)171 static WEBP_INLINE double BitsEntropyRefine(const VP8LBitEntropy* entropy) {
172 double mix;
173 if (entropy->nonzeros < 5) {
174 if (entropy->nonzeros <= 1) {
175 return 0;
176 }
177 // Two symbols, they will be 0 and 1 in a Huffman code.
178 // Let's mix in a bit of entropy to favor good clustering when
179 // distributions of these are combined.
180 if (entropy->nonzeros == 2) {
181 return 0.99 * entropy->sum + 0.01 * entropy->entropy;
182 }
183 // No matter what the entropy says, we cannot be better than min_limit
184 // with Huffman coding. I am mixing a bit of entropy into the
185 // min_limit since it produces much better (~0.5 %) compression results
186 // perhaps because of better entropy clustering.
187 if (entropy->nonzeros == 3) {
188 mix = 0.95;
189 } else {
190 mix = 0.7; // nonzeros == 4.
191 }
192 } else {
193 mix = 0.627;
194 }
195
196 {
197 double min_limit = 2 * entropy->sum - entropy->max_val;
198 min_limit = mix * min_limit + (1.0 - mix) * entropy->entropy;
199 return (entropy->entropy < min_limit) ? min_limit : entropy->entropy;
200 }
201 }
202
VP8LBitsEntropy(const uint32_t * const array,int n)203 double VP8LBitsEntropy(const uint32_t* const array, int n) {
204 VP8LBitEntropy entropy;
205 VP8LBitsEntropyUnrefined(array, n, &entropy);
206
207 return BitsEntropyRefine(&entropy);
208 }
209
InitialHuffmanCost(void)210 static double InitialHuffmanCost(void) {
211 // Small bias because Huffman code length is typically not stored in
212 // full length.
213 static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
214 static const double kSmallBias = 9.1;
215 return kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
216 }
217
218 // Finalize the Huffman cost based on streak numbers and length type (<3 or >=3)
FinalHuffmanCost(const VP8LStreaks * const stats)219 static double FinalHuffmanCost(const VP8LStreaks* const stats) {
220 // The constants in this function are experimental and got rounded from
221 // their original values in 1/8 when switched to 1/1024.
222 double retval = InitialHuffmanCost();
223 // Second coefficient: Many zeros in the histogram are covered efficiently
224 // by a run-length encode. Originally 2/8.
225 retval += stats->counts[0] * 1.5625 + 0.234375 * stats->streaks[0][1];
226 // Second coefficient: Constant values are encoded less efficiently, but still
227 // RLE'ed. Originally 6/8.
228 retval += stats->counts[1] * 2.578125 + 0.703125 * stats->streaks[1][1];
229 // 0s are usually encoded more efficiently than non-0s.
230 // Originally 15/8.
231 retval += 1.796875 * stats->streaks[0][0];
232 // Originally 26/8.
233 retval += 3.28125 * stats->streaks[1][0];
234 return retval;
235 }
236
237 // Get the symbol entropy for the distribution 'population'.
238 // 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)239 static double PopulationCost(const uint32_t* const population, int length,
240 uint32_t* const trivial_sym) {
241 VP8LBitEntropy bit_entropy;
242 VP8LStreaks stats;
243 VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats);
244 if (trivial_sym != NULL) {
245 *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code
246 : VP8L_NON_TRIVIAL_SYM;
247 }
248
249 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
250 }
251
252 // trivial_at_end is 1 if the two histograms only have one element that is
253 // 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 trivial_at_end)254 static WEBP_INLINE double GetCombinedEntropy(const uint32_t* const X,
255 const uint32_t* const Y,
256 int length, int trivial_at_end) {
257 VP8LStreaks stats;
258 if (trivial_at_end) {
259 // This configuration is due to palettization that transforms an indexed
260 // pixel into 0xff000000 | (pixel << 8) in VP8LBundleColorMap.
261 // BitsEntropyRefine is 0 for histograms with only one non-zero value.
262 // Only FinalHuffmanCost needs to be evaluated.
263 memset(&stats, 0, sizeof(stats));
264 // Deal with the non-zero value at index 0 or length-1.
265 stats.streaks[1][0] += 1;
266 // Deal with the following/previous zero streak.
267 stats.counts[0] += 1;
268 stats.streaks[0][1] += length - 1;
269 return FinalHuffmanCost(&stats);
270 } else {
271 VP8LBitEntropy bit_entropy;
272 VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats);
273
274 return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
275 }
276 }
277
278 // Estimates the Entropy + Huffman + other block overhead size cost.
VP8LHistogramEstimateBits(const VP8LHistogram * const p)279 double VP8LHistogramEstimateBits(const VP8LHistogram* const p) {
280 return
281 PopulationCost(
282 p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_), NULL)
283 + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL)
284 + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL)
285 + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL)
286 + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL)
287 + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
288 + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
289 }
290
291 // -----------------------------------------------------------------------------
292 // Various histogram combine/cost-eval functions
293
GetCombinedHistogramEntropy(const VP8LHistogram * const a,const VP8LHistogram * const b,double cost_threshold,double * cost)294 static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
295 const VP8LHistogram* const b,
296 double cost_threshold,
297 double* cost) {
298 const int palette_code_bits = a->palette_code_bits_;
299 int trivial_at_end = 0;
300 assert(a->palette_code_bits_ == b->palette_code_bits_);
301 *cost += GetCombinedEntropy(a->literal_, b->literal_,
302 VP8LHistogramNumCodes(palette_code_bits), 0);
303 *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
304 b->literal_ + NUM_LITERAL_CODES,
305 NUM_LENGTH_CODES);
306 if (*cost > cost_threshold) return 0;
307
308 if (a->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM &&
309 a->trivial_symbol_ == b->trivial_symbol_) {
310 // A, R and B are all 0 or 0xff.
311 const uint32_t color_a = (a->trivial_symbol_ >> 24) & 0xff;
312 const uint32_t color_r = (a->trivial_symbol_ >> 16) & 0xff;
313 const uint32_t color_b = (a->trivial_symbol_ >> 0) & 0xff;
314 if ((color_a == 0 || color_a == 0xff) &&
315 (color_r == 0 || color_r == 0xff) &&
316 (color_b == 0 || color_b == 0xff)) {
317 trivial_at_end = 1;
318 }
319 }
320
321 *cost +=
322 GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, trivial_at_end);
323 if (*cost > cost_threshold) return 0;
324
325 *cost +=
326 GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, trivial_at_end);
327 if (*cost > cost_threshold) return 0;
328
329 *cost += GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES,
330 trivial_at_end);
331 if (*cost > cost_threshold) return 0;
332
333 *cost +=
334 GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES, 0);
335 *cost +=
336 VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES);
337 if (*cost > cost_threshold) return 0;
338
339 return 1;
340 }
341
HistogramAdd(const VP8LHistogram * const a,const VP8LHistogram * const b,VP8LHistogram * const out)342 static WEBP_INLINE void HistogramAdd(const VP8LHistogram* const a,
343 const VP8LHistogram* const b,
344 VP8LHistogram* const out) {
345 VP8LHistogramAdd(a, b, out);
346 out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_)
347 ? a->trivial_symbol_
348 : VP8L_NON_TRIVIAL_SYM;
349 }
350
351 // Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing
352 // to the threshold value 'cost_threshold'. The score returned is
353 // Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed.
354 // Since the previous score passed is 'cost_threshold', we only need to compare
355 // the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out
356 // early.
HistogramAddEval(const VP8LHistogram * const a,const VP8LHistogram * const b,VP8LHistogram * const out,double cost_threshold)357 static double HistogramAddEval(const VP8LHistogram* const a,
358 const VP8LHistogram* const b,
359 VP8LHistogram* const out,
360 double cost_threshold) {
361 double cost = 0;
362 const double sum_cost = a->bit_cost_ + b->bit_cost_;
363 cost_threshold += sum_cost;
364
365 if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
366 HistogramAdd(a, b, out);
367 out->bit_cost_ = cost;
368 out->palette_code_bits_ = a->palette_code_bits_;
369 }
370
371 return cost - sum_cost;
372 }
373
374 // Same as HistogramAddEval(), except that the resulting histogram
375 // is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit
376 // the term C(b) which is constant over all the evaluations.
HistogramAddThresh(const VP8LHistogram * const a,const VP8LHistogram * const b,double cost_threshold)377 static double HistogramAddThresh(const VP8LHistogram* const a,
378 const VP8LHistogram* const b,
379 double cost_threshold) {
380 double cost = -a->bit_cost_;
381 GetCombinedHistogramEntropy(a, b, cost_threshold, &cost);
382 return cost;
383 }
384
385 // -----------------------------------------------------------------------------
386
387 // The structure to keep track of cost range for the three dominant entropy
388 // symbols.
389 // TODO(skal): Evaluate if float can be used here instead of double for
390 // representing the entropy costs.
391 typedef struct {
392 double literal_max_;
393 double literal_min_;
394 double red_max_;
395 double red_min_;
396 double blue_max_;
397 double blue_min_;
398 } DominantCostRange;
399
DominantCostRangeInit(DominantCostRange * const c)400 static void DominantCostRangeInit(DominantCostRange* const c) {
401 c->literal_max_ = 0.;
402 c->literal_min_ = MAX_COST;
403 c->red_max_ = 0.;
404 c->red_min_ = MAX_COST;
405 c->blue_max_ = 0.;
406 c->blue_min_ = MAX_COST;
407 }
408
UpdateDominantCostRange(const VP8LHistogram * const h,DominantCostRange * const c)409 static void UpdateDominantCostRange(
410 const VP8LHistogram* const h, DominantCostRange* const c) {
411 if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_;
412 if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_;
413 if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_;
414 if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_;
415 if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_;
416 if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_;
417 }
418
UpdateHistogramCost(VP8LHistogram * const h)419 static void UpdateHistogramCost(VP8LHistogram* const h) {
420 uint32_t alpha_sym, red_sym, blue_sym;
421 const double alpha_cost =
422 PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym);
423 const double distance_cost =
424 PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL) +
425 VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
426 const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
427 h->literal_cost_ = PopulationCost(h->literal_, num_codes, NULL) +
428 VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES,
429 NUM_LENGTH_CODES);
430 h->red_cost_ = PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym);
431 h->blue_cost_ = PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym);
432 h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
433 alpha_cost + distance_cost;
434 if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) {
435 h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM;
436 } else {
437 h->trivial_symbol_ =
438 ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0);
439 }
440 }
441
GetBinIdForEntropy(double min,double max,double val)442 static int GetBinIdForEntropy(double min, double max, double val) {
443 const double range = max - min;
444 if (range > 0.) {
445 const double delta = val - min;
446 return (int)((NUM_PARTITIONS - 1e-6) * delta / range);
447 } else {
448 return 0;
449 }
450 }
451
GetHistoBinIndex(const VP8LHistogram * const h,const DominantCostRange * const c,int low_effort)452 static int GetHistoBinIndex(const VP8LHistogram* const h,
453 const DominantCostRange* const c, int low_effort) {
454 int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_,
455 h->literal_cost_);
456 assert(bin_id < NUM_PARTITIONS);
457 if (!low_effort) {
458 bin_id = bin_id * NUM_PARTITIONS
459 + GetBinIdForEntropy(c->red_min_, c->red_max_, h->red_cost_);
460 bin_id = bin_id * NUM_PARTITIONS
461 + GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_);
462 assert(bin_id < BIN_SIZE);
463 }
464 return bin_id;
465 }
466
467 // Construct the histograms from backward references.
HistogramBuild(int xsize,int histo_bits,const VP8LBackwardRefs * const backward_refs,VP8LHistogramSet * const image_histo)468 static void HistogramBuild(
469 int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs,
470 VP8LHistogramSet* const image_histo) {
471 int x = 0, y = 0;
472 const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
473 VP8LHistogram** const histograms = image_histo->histograms;
474 VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs);
475 assert(histo_bits > 0);
476 while (VP8LRefsCursorOk(&c)) {
477 const PixOrCopy* const v = c.cur_pos;
478 const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
479 VP8LHistogramAddSinglePixOrCopy(histograms[ix], v, NULL, 0);
480 x += PixOrCopyLength(v);
481 while (x >= xsize) {
482 x -= xsize;
483 ++y;
484 }
485 VP8LRefsCursorNext(&c);
486 }
487 }
488
489 // Copies the histograms and computes its bit_cost.
HistogramCopyAndAnalyze(VP8LHistogramSet * const orig_histo,VP8LHistogramSet * const image_histo)490 static void HistogramCopyAndAnalyze(
491 VP8LHistogramSet* const orig_histo, VP8LHistogramSet* const image_histo) {
492 int i;
493 const int histo_size = orig_histo->size;
494 VP8LHistogram** const orig_histograms = orig_histo->histograms;
495 VP8LHistogram** const histograms = image_histo->histograms;
496 for (i = 0; i < histo_size; ++i) {
497 VP8LHistogram* const histo = orig_histograms[i];
498 UpdateHistogramCost(histo);
499 // Copy histograms from orig_histo[] to image_histo[].
500 HistogramCopy(histo, histograms[i]);
501 }
502 }
503
504 // Partition histograms to different entropy bins for three dominant (literal,
505 // 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)506 static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo,
507 uint16_t* const bin_map,
508 int low_effort) {
509 int i;
510 VP8LHistogram** const histograms = image_histo->histograms;
511 const int histo_size = image_histo->size;
512 DominantCostRange cost_range;
513 DominantCostRangeInit(&cost_range);
514
515 // Analyze the dominant (literal, red and blue) entropy costs.
516 for (i = 0; i < histo_size; ++i) {
517 UpdateDominantCostRange(histograms[i], &cost_range);
518 }
519
520 // bin-hash histograms on three of the dominant (literal, red and blue)
521 // symbol costs and store the resulting bin_id for each histogram.
522 for (i = 0; i < histo_size; ++i) {
523 bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort);
524 }
525 }
526
527 // Compact image_histo[] by merging some histograms with same bin_id together if
528 // it's advantageous.
HistogramCombineEntropyBin(VP8LHistogramSet * const image_histo,VP8LHistogram * cur_combo,const uint16_t * const bin_map,int bin_map_size,int num_bins,double combine_cost_factor,int low_effort)529 static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo,
530 VP8LHistogram* cur_combo,
531 const uint16_t* const bin_map,
532 int bin_map_size, int num_bins,
533 double combine_cost_factor,
534 int low_effort) {
535 VP8LHistogram** const histograms = image_histo->histograms;
536 int idx;
537 // Work in-place: processed histograms are put at the beginning of
538 // image_histo[]. At the end, we just have to truncate the array.
539 int size = 0;
540 struct {
541 int16_t first; // position of the histogram that accumulates all
542 // histograms with the same bin_id
543 uint16_t num_combine_failures; // number of combine failures per bin_id
544 } bin_info[BIN_SIZE];
545
546 assert(num_bins <= BIN_SIZE);
547 for (idx = 0; idx < num_bins; ++idx) {
548 bin_info[idx].first = -1;
549 bin_info[idx].num_combine_failures = 0;
550 }
551
552 for (idx = 0; idx < bin_map_size; ++idx) {
553 const int bin_id = bin_map[idx];
554 const int first = bin_info[bin_id].first;
555 assert(size <= idx);
556 if (first == -1) {
557 // just move histogram #idx to its final position
558 histograms[size] = histograms[idx];
559 bin_info[bin_id].first = size++;
560 } else if (low_effort) {
561 HistogramAdd(histograms[idx], histograms[first], histograms[first]);
562 } else {
563 // try to merge #idx into #first (both share the same bin_id)
564 const double bit_cost = histograms[idx]->bit_cost_;
565 const double bit_cost_thresh = -bit_cost * combine_cost_factor;
566 const double curr_cost_diff =
567 HistogramAddEval(histograms[first], histograms[idx],
568 cur_combo, bit_cost_thresh);
569 if (curr_cost_diff < bit_cost_thresh) {
570 // Try to merge two histograms only if the combo is a trivial one or
571 // the two candidate histograms are already non-trivial.
572 // For some images, 'try_combine' turns out to be false for a lot of
573 // histogram pairs. In that case, we fallback to combining
574 // histograms as usual to avoid increasing the header size.
575 const int try_combine =
576 (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) ||
577 ((histograms[idx]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) &&
578 (histograms[first]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM));
579 const int max_combine_failures = 32;
580 if (try_combine ||
581 bin_info[bin_id].num_combine_failures >= max_combine_failures) {
582 // move the (better) merged histogram to its final slot
583 HistogramSwap(&cur_combo, &histograms[first]);
584 } else {
585 histograms[size++] = histograms[idx];
586 ++bin_info[bin_id].num_combine_failures;
587 }
588 } else {
589 histograms[size++] = histograms[idx];
590 }
591 }
592 }
593 image_histo->size = size;
594 if (low_effort) {
595 // for low_effort case, update the final cost when everything is merged
596 for (idx = 0; idx < size; ++idx) {
597 UpdateHistogramCost(histograms[idx]);
598 }
599 }
600 }
601
602 // Implement a Lehmer random number generator with a multiplicative constant of
603 // 48271 and a modulo constant of 2^31 - 1.
MyRand(uint32_t * const seed)604 static uint32_t MyRand(uint32_t* const seed) {
605 *seed = (uint32_t)(((uint64_t)(*seed) * 48271u) % 2147483647u);
606 assert(*seed > 0);
607 return *seed;
608 }
609
610 // -----------------------------------------------------------------------------
611 // Histogram pairs priority queue
612
613 // Pair of histograms. Negative idx1 value means that pair is out-of-date.
614 typedef struct {
615 int idx1;
616 int idx2;
617 double cost_diff;
618 double cost_combo;
619 } HistogramPair;
620
621 typedef struct {
622 HistogramPair* queue;
623 int size;
624 int max_size;
625 } HistoQueue;
626
HistoQueueInit(HistoQueue * const histo_queue,const int max_index)627 static int HistoQueueInit(HistoQueue* const histo_queue, const int max_index) {
628 histo_queue->size = 0;
629 // max_index^2 for the queue size is safe. If you look at
630 // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes
631 // data to the queue, you insert at most:
632 // - max_index*(max_index-1)/2 (the first two for loops)
633 // - max_index - 1 in the last for loop at the first iteration of the while
634 // loop, max_index - 2 at the second iteration ... therefore
635 // max_index*(max_index-1)/2 overall too
636 histo_queue->max_size = max_index * max_index;
637 // We allocate max_size + 1 because the last element at index "size" is
638 // used as temporary data (and it could be up to max_size).
639 histo_queue->queue = (HistogramPair*)WebPSafeMalloc(
640 histo_queue->max_size + 1, sizeof(*histo_queue->queue));
641 return histo_queue->queue != NULL;
642 }
643
HistoQueueClear(HistoQueue * const histo_queue)644 static void HistoQueueClear(HistoQueue* const histo_queue) {
645 assert(histo_queue != NULL);
646 WebPSafeFree(histo_queue->queue);
647 histo_queue->size = 0;
648 histo_queue->max_size = 0;
649 }
650
651 // Pop a specific pair in the queue by replacing it with the last one
652 // and shrinking the queue.
HistoQueuePopPair(HistoQueue * const histo_queue,HistogramPair * const pair)653 static void HistoQueuePopPair(HistoQueue* const histo_queue,
654 HistogramPair* const pair) {
655 assert(pair >= histo_queue->queue &&
656 pair < (histo_queue->queue + histo_queue->size));
657 assert(histo_queue->size > 0);
658 *pair = histo_queue->queue[histo_queue->size - 1];
659 --histo_queue->size;
660 }
661
662 // Check whether a pair in the queue should be updated as head or not.
HistoQueueUpdateHead(HistoQueue * const histo_queue,HistogramPair * const pair)663 static void HistoQueueUpdateHead(HistoQueue* const histo_queue,
664 HistogramPair* const pair) {
665 assert(pair->cost_diff < 0.);
666 assert(pair >= histo_queue->queue &&
667 pair < (histo_queue->queue + histo_queue->size));
668 assert(histo_queue->size > 0);
669 if (pair->cost_diff < histo_queue->queue[0].cost_diff) {
670 // Replace the best pair.
671 const HistogramPair tmp = histo_queue->queue[0];
672 histo_queue->queue[0] = *pair;
673 *pair = tmp;
674 }
675 }
676
677 // Create a pair from indices "idx1" and "idx2" provided its cost
678 // is inferior to "threshold", a negative entropy.
679 // 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)680 static double HistoQueuePush(HistoQueue* const histo_queue,
681 VP8LHistogram** const histograms, int idx1,
682 int idx2, double threshold) {
683 const VP8LHistogram* h1;
684 const VP8LHistogram* h2;
685 HistogramPair pair;
686 double sum_cost;
687
688 assert(threshold <= 0.);
689 if (idx1 > idx2) {
690 const int tmp = idx2;
691 idx2 = idx1;
692 idx1 = tmp;
693 }
694 pair.idx1 = idx1;
695 pair.idx2 = idx2;
696 h1 = histograms[idx1];
697 h2 = histograms[idx2];
698 sum_cost = h1->bit_cost_ + h2->bit_cost_;
699 pair.cost_combo = 0.;
700 GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair.cost_combo);
701 pair.cost_diff = pair.cost_combo - sum_cost;
702
703 // Do not even consider the pair if it does not improve the entropy.
704 if (pair.cost_diff >= threshold) return 0.;
705
706 // We cannot add more elements than the capacity.
707 assert(histo_queue->size < histo_queue->max_size);
708 histo_queue->queue[histo_queue->size++] = pair;
709 HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]);
710
711 return pair.cost_diff;
712 }
713
714 // -----------------------------------------------------------------------------
715
716 // Combines histograms by continuously choosing the one with the highest cost
717 // reduction.
HistogramCombineGreedy(VP8LHistogramSet * const image_histo)718 static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) {
719 int ok = 0;
720 int image_histo_size = image_histo->size;
721 int i, j;
722 VP8LHistogram** const histograms = image_histo->histograms;
723 // Indexes of remaining histograms.
724 int* const clusters =
725 (int*)WebPSafeMalloc(image_histo_size, sizeof(*clusters));
726 // Priority queue of histogram pairs.
727 HistoQueue histo_queue;
728
729 if (!HistoQueueInit(&histo_queue, image_histo_size) || clusters == NULL) {
730 goto End;
731 }
732
733 for (i = 0; i < image_histo_size; ++i) {
734 // Initialize clusters indexes.
735 clusters[i] = i;
736 for (j = i + 1; j < image_histo_size; ++j) {
737 // Initialize positions array.
738 HistoQueuePush(&histo_queue, histograms, i, j, 0.);
739 }
740 }
741
742 while (image_histo_size > 1 && histo_queue.size > 0) {
743 const int idx1 = histo_queue.queue[0].idx1;
744 const int idx2 = histo_queue.queue[0].idx2;
745 HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]);
746 histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
747 // Remove merged histogram.
748 for (i = 0; i + 1 < image_histo_size; ++i) {
749 if (clusters[i] >= idx2) {
750 clusters[i] = clusters[i + 1];
751 }
752 }
753 --image_histo_size;
754
755 // Remove pairs intersecting the just combined best pair.
756 for (i = 0; i < histo_queue.size;) {
757 HistogramPair* const p = histo_queue.queue + i;
758 if (p->idx1 == idx1 || p->idx2 == idx1 ||
759 p->idx1 == idx2 || p->idx2 == idx2) {
760 HistoQueuePopPair(&histo_queue, p);
761 } else {
762 HistoQueueUpdateHead(&histo_queue, p);
763 ++i;
764 }
765 }
766
767 // Push new pairs formed with combined histogram to the queue.
768 for (i = 0; i < image_histo_size; ++i) {
769 if (clusters[i] != idx1) {
770 HistoQueuePush(&histo_queue, histograms, idx1, clusters[i], 0.);
771 }
772 }
773 }
774 // Move remaining histograms to the beginning of the array.
775 for (i = 0; i < image_histo_size; ++i) {
776 if (i != clusters[i]) { // swap the two histograms
777 HistogramSwap(&histograms[i], &histograms[clusters[i]]);
778 }
779 }
780
781 image_histo->size = image_histo_size;
782 ok = 1;
783
784 End:
785 WebPSafeFree(clusters);
786 HistoQueueClear(&histo_queue);
787 return ok;
788 }
789
790 // Perform histogram aggregation using a stochastic approach.
791 // 'do_greedy' is set to 1 if a greedy approach needs to be performed
792 // afterwards, 0 otherwise.
HistogramCombineStochastic(VP8LHistogramSet * const image_histo,int min_cluster_size,int * const do_greedy)793 static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo,
794 int min_cluster_size,
795 int* const do_greedy) {
796 int iter;
797 uint32_t seed = 1;
798 int tries_with_no_success = 0;
799 int image_histo_size = image_histo->size;
800 const int outer_iters = image_histo_size;
801 const int num_tries_no_success = outer_iters / 2;
802 VP8LHistogram** const histograms = image_histo->histograms;
803 // Priority queue of histogram pairs. Its size of "kCostHeapSizeSqrt"^2
804 // impacts the quality of the compression and the speed: the smaller the
805 // faster but the worse for the compression.
806 HistoQueue histo_queue;
807 const int kHistoQueueSizeSqrt = 3;
808 int ok = 0;
809
810 if (!HistoQueueInit(&histo_queue, kHistoQueueSizeSqrt)) {
811 goto End;
812 }
813 // Collapse similar histograms in 'image_histo'.
814 ++min_cluster_size;
815 for (iter = 0; iter < outer_iters && image_histo_size >= min_cluster_size &&
816 ++tries_with_no_success < num_tries_no_success;
817 ++iter) {
818 double best_cost =
819 (histo_queue.size == 0) ? 0. : histo_queue.queue[0].cost_diff;
820 int best_idx1 = -1, best_idx2 = 1;
821 int j;
822 const uint32_t rand_range = (image_histo_size - 1) * image_histo_size;
823 // image_histo_size / 2 was chosen empirically. Less means faster but worse
824 // compression.
825 const int num_tries = image_histo_size / 2;
826
827 for (j = 0; j < num_tries; ++j) {
828 double curr_cost;
829 // Choose two different histograms at random and try to combine them.
830 const uint32_t tmp = MyRand(&seed) % rand_range;
831 const uint32_t idx1 = tmp / (image_histo_size - 1);
832 uint32_t idx2 = tmp % (image_histo_size - 1);
833 if (idx2 >= idx1) ++idx2;
834
835 // Calculate cost reduction on combination.
836 curr_cost =
837 HistoQueuePush(&histo_queue, histograms, idx1, idx2, best_cost);
838 if (curr_cost < 0) { // found a better pair?
839 best_cost = curr_cost;
840 // Empty the queue if we reached full capacity.
841 if (histo_queue.size == histo_queue.max_size) break;
842 }
843 }
844 if (histo_queue.size == 0) continue;
845
846 // Merge the two best histograms.
847 best_idx1 = histo_queue.queue[0].idx1;
848 best_idx2 = histo_queue.queue[0].idx2;
849 assert(best_idx1 < best_idx2);
850 HistogramAddEval(histograms[best_idx1], histograms[best_idx2],
851 histograms[best_idx1], 0);
852 // Swap the best_idx2 histogram with the last one (which is now unused).
853 --image_histo_size;
854 if (best_idx2 != image_histo_size) {
855 HistogramSwap(&histograms[image_histo_size], &histograms[best_idx2]);
856 }
857 histograms[image_histo_size] = NULL;
858 // Parse the queue and update each pair that deals with best_idx1,
859 // best_idx2 or image_histo_size.
860 for (j = 0; j < histo_queue.size;) {
861 HistogramPair* const p = histo_queue.queue + j;
862 const int is_idx1_best = p->idx1 == best_idx1 || p->idx1 == best_idx2;
863 const int is_idx2_best = p->idx2 == best_idx1 || p->idx2 == best_idx2;
864 int do_eval = 0;
865 // The front pair could have been duplicated by a random pick so
866 // check for it all the time nevertheless.
867 if (is_idx1_best && is_idx2_best) {
868 HistoQueuePopPair(&histo_queue, p);
869 continue;
870 }
871 // Any pair containing one of the two best indices should only refer to
872 // best_idx1. Its cost should also be updated.
873 if (is_idx1_best) {
874 p->idx1 = best_idx1;
875 do_eval = 1;
876 } else if (is_idx2_best) {
877 p->idx2 = best_idx1;
878 do_eval = 1;
879 }
880 if (p->idx2 == image_histo_size) {
881 // No need to re-evaluate here as it does not involve a pair
882 // containing best_idx1 or best_idx2.
883 p->idx2 = best_idx2;
884 }
885 assert(p->idx2 < image_histo_size);
886 // Make sure the index order is respected.
887 if (p->idx1 > p->idx2) {
888 const int tmp = p->idx2;
889 p->idx2 = p->idx1;
890 p->idx1 = tmp;
891 }
892 if (do_eval) {
893 // Re-evaluate the cost of an updated pair.
894 GetCombinedHistogramEntropy(histograms[p->idx1], histograms[p->idx2], 0,
895 &p->cost_diff);
896 if (p->cost_diff >= 0.) {
897 HistoQueuePopPair(&histo_queue, p);
898 continue;
899 }
900 }
901 HistoQueueUpdateHead(&histo_queue, p);
902 ++j;
903 }
904
905 tries_with_no_success = 0;
906 }
907 image_histo->size = image_histo_size;
908 *do_greedy = (image_histo->size <= min_cluster_size);
909 ok = 1;
910
911 End:
912 HistoQueueClear(&histo_queue);
913 return ok;
914 }
915
916 // -----------------------------------------------------------------------------
917 // Histogram refinement
918
919 // Find the best 'out' histogram for each of the 'in' histograms.
920 // Note: we assume that out[]->bit_cost_ is already up-to-date.
HistogramRemap(const VP8LHistogramSet * const in,const VP8LHistogramSet * const out,uint16_t * const symbols)921 static void HistogramRemap(const VP8LHistogramSet* const in,
922 const VP8LHistogramSet* const out,
923 uint16_t* const symbols) {
924 int i;
925 VP8LHistogram** const in_histo = in->histograms;
926 VP8LHistogram** const out_histo = out->histograms;
927 const int in_size = in->size;
928 const int out_size = out->size;
929 if (out_size > 1) {
930 for (i = 0; i < in_size; ++i) {
931 int best_out = 0;
932 double best_bits = MAX_COST;
933 int k;
934 for (k = 0; k < out_size; ++k) {
935 const double cur_bits =
936 HistogramAddThresh(out_histo[k], in_histo[i], best_bits);
937 if (k == 0 || cur_bits < best_bits) {
938 best_bits = cur_bits;
939 best_out = k;
940 }
941 }
942 symbols[i] = best_out;
943 }
944 } else {
945 assert(out_size == 1);
946 for (i = 0; i < in_size; ++i) {
947 symbols[i] = 0;
948 }
949 }
950
951 // Recompute each out based on raw and symbols.
952 for (i = 0; i < out_size; ++i) {
953 HistogramClear(out_histo[i]);
954 }
955
956 for (i = 0; i < in_size; ++i) {
957 const int idx = symbols[i];
958 HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]);
959 }
960 }
961
GetCombineCostFactor(int histo_size,int quality)962 static double GetCombineCostFactor(int histo_size, int quality) {
963 double combine_cost_factor = 0.16;
964 if (quality < 90) {
965 if (histo_size > 256) combine_cost_factor /= 2.;
966 if (histo_size > 512) combine_cost_factor /= 2.;
967 if (histo_size > 1024) combine_cost_factor /= 2.;
968 if (quality <= 50) combine_cost_factor /= 2.;
969 }
970 return combine_cost_factor;
971 }
972
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)973 int VP8LGetHistoImageSymbols(int xsize, int ysize,
974 const VP8LBackwardRefs* const refs,
975 int quality, int low_effort,
976 int histo_bits, int cache_bits,
977 VP8LHistogramSet* const image_histo,
978 VP8LHistogram* const tmp_histo,
979 uint16_t* const histogram_symbols) {
980 int ok = 0;
981 const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
982 const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
983 const int image_histo_raw_size = histo_xsize * histo_ysize;
984 VP8LHistogramSet* const orig_histo =
985 VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
986 // Don't attempt linear bin-partition heuristic for
987 // histograms of small sizes (as bin_map will be very sparse) and
988 // maximum quality q==100 (to preserve the compression gains at that level).
989 const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE;
990 const int entropy_combine =
991 (orig_histo->size > entropy_combine_num_bins * 2) && (quality < 100);
992
993 if (orig_histo == NULL) goto Error;
994
995 // Construct the histograms from backward references.
996 HistogramBuild(xsize, histo_bits, refs, orig_histo);
997 // Copies the histograms and computes its bit_cost.
998 HistogramCopyAndAnalyze(orig_histo, image_histo);
999
1000 if (entropy_combine) {
1001 const int bin_map_size = orig_histo->size;
1002 // Reuse histogram_symbols storage. By definition, it's guaranteed to be ok.
1003 uint16_t* const bin_map = histogram_symbols;
1004 const double combine_cost_factor =
1005 GetCombineCostFactor(image_histo_raw_size, quality);
1006
1007 HistogramAnalyzeEntropyBin(orig_histo, bin_map, low_effort);
1008 // Collapse histograms with similar entropy.
1009 HistogramCombineEntropyBin(image_histo, tmp_histo, bin_map, bin_map_size,
1010 entropy_combine_num_bins, combine_cost_factor,
1011 low_effort);
1012 }
1013
1014 // Don't combine the histograms using stochastic and greedy heuristics for
1015 // low-effort compression mode.
1016 if (!low_effort || !entropy_combine) {
1017 const float x = quality / 100.f;
1018 // cubic ramp between 1 and MAX_HISTO_GREEDY:
1019 const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1));
1020 int do_greedy;
1021 if (!HistogramCombineStochastic(image_histo, threshold_size, &do_greedy)) {
1022 goto Error;
1023 }
1024 if (do_greedy && !HistogramCombineGreedy(image_histo)) {
1025 goto Error;
1026 }
1027 }
1028
1029 // TODO(vrabaud): Optimize HistogramRemap for low-effort compression mode.
1030 // Find the optimal map from original histograms to the final ones.
1031 HistogramRemap(orig_histo, image_histo, histogram_symbols);
1032
1033 ok = 1;
1034
1035 Error:
1036 VP8LFreeHistogramSet(orig_histo);
1037 return ok;
1038 }
1039