/src/libwebp/src/enc/histogram_enc.c
Line | Count | Source (jump to first uncovered line) |
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 <assert.h> |
17 | | #include <stdlib.h> |
18 | | #include <string.h> |
19 | | |
20 | | #include "src/dsp/lossless.h" |
21 | | #include "src/dsp/lossless_common.h" |
22 | | #include "src/enc/backward_references_enc.h" |
23 | | #include "src/enc/histogram_enc.h" |
24 | | #include "src/enc/vp8i_enc.h" |
25 | | #include "src/utils/utils.h" |
26 | | #include "src/webp/encode.h" |
27 | | #include "src/webp/format_constants.h" |
28 | | #include "src/webp/types.h" |
29 | | |
30 | | // Number of partitions for the three dominant (literal, red and blue) symbol |
31 | | // costs. |
32 | 0 | #define NUM_PARTITIONS 4 |
33 | | // The size of the bin-hash corresponding to the three dominant costs. |
34 | 0 | #define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS) |
35 | | // Maximum number of histograms allowed in greedy combining algorithm. |
36 | 0 | #define MAX_HISTO_GREEDY 100 |
37 | | |
38 | | // Enum to meaningfully access the elements of the Histogram arrays. |
39 | | typedef enum { |
40 | | LITERAL = 0, |
41 | | RED, |
42 | | BLUE, |
43 | | ALPHA, |
44 | | DISTANCE |
45 | | } HistogramIndex; |
46 | | |
47 | | // Return the size of the histogram for a given cache_bits. |
48 | 0 | static int GetHistogramSize(int cache_bits) { |
49 | 0 | const int literal_size = VP8LHistogramNumCodes(cache_bits); |
50 | 0 | const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size; |
51 | 0 | assert(total_size <= (size_t)0x7fffffff); |
52 | 0 | return (int)total_size; |
53 | 0 | } |
54 | | |
55 | 0 | static void HistogramStatsClear(VP8LHistogram* const h) { |
56 | 0 | int i; |
57 | 0 | for (i = 0; i < 5; ++i) { |
58 | 0 | h->trivial_symbol[i] = VP8L_NON_TRIVIAL_SYM; |
59 | | // By default, the histogram is assumed to be used. |
60 | 0 | h->is_used[i] = 1; |
61 | 0 | } |
62 | 0 | h->bit_cost = 0; |
63 | 0 | memset(h->costs, 0, sizeof(h->costs)); |
64 | 0 | } |
65 | | |
66 | 0 | static void HistogramClear(VP8LHistogram* const h) { |
67 | 0 | uint32_t* const literal = h->literal; |
68 | 0 | const int cache_bits = h->palette_code_bits; |
69 | 0 | const int histo_size = GetHistogramSize(cache_bits); |
70 | 0 | memset(h, 0, histo_size); |
71 | 0 | h->palette_code_bits = cache_bits; |
72 | 0 | h->literal = literal; |
73 | 0 | HistogramStatsClear(h); |
74 | 0 | } |
75 | | |
76 | | // Swap two histogram pointers. |
77 | 0 | static void HistogramSwap(VP8LHistogram** const h1, VP8LHistogram** const h2) { |
78 | 0 | VP8LHistogram* const tmp = *h1; |
79 | 0 | *h1 = *h2; |
80 | 0 | *h2 = tmp; |
81 | 0 | } |
82 | | |
83 | | static void HistogramCopy(const VP8LHistogram* const src, |
84 | 0 | VP8LHistogram* const dst) { |
85 | 0 | uint32_t* const dst_literal = dst->literal; |
86 | 0 | const int dst_cache_bits = dst->palette_code_bits; |
87 | 0 | const int literal_size = VP8LHistogramNumCodes(dst_cache_bits); |
88 | 0 | const int histo_size = GetHistogramSize(dst_cache_bits); |
89 | 0 | assert(src->palette_code_bits == dst_cache_bits); |
90 | 0 | memcpy(dst, src, histo_size); |
91 | 0 | dst->literal = dst_literal; |
92 | 0 | memcpy(dst->literal, src->literal, literal_size * sizeof(*dst->literal)); |
93 | 0 | } |
94 | | |
95 | 0 | void VP8LFreeHistogram(VP8LHistogram* const h) { WebPSafeFree(h); } |
96 | | |
97 | 0 | void VP8LFreeHistogramSet(VP8LHistogramSet* const histograms) { |
98 | 0 | WebPSafeFree(histograms); |
99 | 0 | } |
100 | | |
101 | | void VP8LHistogramCreate(VP8LHistogram* const h, |
102 | | const VP8LBackwardRefs* const refs, |
103 | 0 | int palette_code_bits) { |
104 | 0 | if (palette_code_bits >= 0) { |
105 | 0 | h->palette_code_bits = palette_code_bits; |
106 | 0 | } |
107 | 0 | HistogramClear(h); |
108 | 0 | VP8LHistogramStoreRefs(refs, /*distance_modifier=*/NULL, |
109 | 0 | /*distance_modifier_arg0=*/0, h); |
110 | 0 | } |
111 | | |
112 | | void VP8LHistogramInit(VP8LHistogram* const h, int palette_code_bits, |
113 | 0 | int init_arrays) { |
114 | 0 | h->palette_code_bits = palette_code_bits; |
115 | 0 | if (init_arrays) { |
116 | 0 | HistogramClear(h); |
117 | 0 | } else { |
118 | 0 | HistogramStatsClear(h); |
119 | 0 | } |
120 | 0 | } |
121 | | |
122 | 0 | VP8LHistogram* VP8LAllocateHistogram(int cache_bits) { |
123 | 0 | VP8LHistogram* histo = NULL; |
124 | 0 | const int total_size = GetHistogramSize(cache_bits); |
125 | 0 | uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory)); |
126 | 0 | if (memory == NULL) return NULL; |
127 | 0 | histo = (VP8LHistogram*)memory; |
128 | | // 'literal' won't necessary be aligned. |
129 | 0 | histo->literal = (uint32_t*)(memory + sizeof(VP8LHistogram)); |
130 | 0 | VP8LHistogramInit(histo, cache_bits, /*init_arrays=*/ 0); |
131 | 0 | return histo; |
132 | 0 | } |
133 | | |
134 | | // Resets the pointers of the histograms to point to the bit buffer in the set. |
135 | | static void HistogramSetResetPointers(VP8LHistogramSet* const set, |
136 | 0 | int cache_bits) { |
137 | 0 | int i; |
138 | 0 | const int histo_size = GetHistogramSize(cache_bits); |
139 | 0 | uint8_t* memory = (uint8_t*) (set->histograms); |
140 | 0 | memory += set->max_size * sizeof(*set->histograms); |
141 | 0 | for (i = 0; i < set->max_size; ++i) { |
142 | 0 | memory = (uint8_t*) WEBP_ALIGN(memory); |
143 | 0 | set->histograms[i] = (VP8LHistogram*) memory; |
144 | | // 'literal' won't necessary be aligned. |
145 | 0 | set->histograms[i]->literal = (uint32_t*)(memory + sizeof(VP8LHistogram)); |
146 | 0 | memory += histo_size; |
147 | 0 | } |
148 | 0 | } |
149 | | |
150 | | // Returns the total size of the VP8LHistogramSet. |
151 | 0 | static size_t HistogramSetTotalSize(int size, int cache_bits) { |
152 | 0 | const int histo_size = GetHistogramSize(cache_bits); |
153 | 0 | return (sizeof(VP8LHistogramSet) + size * (sizeof(VP8LHistogram*) + |
154 | 0 | histo_size + WEBP_ALIGN_CST)); |
155 | 0 | } |
156 | | |
157 | 0 | VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) { |
158 | 0 | int i; |
159 | 0 | VP8LHistogramSet* set; |
160 | 0 | const size_t total_size = HistogramSetTotalSize(size, cache_bits); |
161 | 0 | uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory)); |
162 | 0 | if (memory == NULL) return NULL; |
163 | | |
164 | 0 | set = (VP8LHistogramSet*)memory; |
165 | 0 | memory += sizeof(*set); |
166 | 0 | set->histograms = (VP8LHistogram**)memory; |
167 | 0 | set->max_size = size; |
168 | 0 | set->size = size; |
169 | 0 | HistogramSetResetPointers(set, cache_bits); |
170 | 0 | for (i = 0; i < size; ++i) { |
171 | 0 | VP8LHistogramInit(set->histograms[i], cache_bits, /*init_arrays=*/ 0); |
172 | 0 | } |
173 | 0 | return set; |
174 | 0 | } |
175 | | |
176 | 0 | void VP8LHistogramSetClear(VP8LHistogramSet* const set) { |
177 | 0 | int i; |
178 | 0 | const int cache_bits = set->histograms[0]->palette_code_bits; |
179 | 0 | const int size = set->max_size; |
180 | 0 | const size_t total_size = HistogramSetTotalSize(size, cache_bits); |
181 | 0 | uint8_t* memory = (uint8_t*)set; |
182 | |
|
183 | 0 | memset(memory, 0, total_size); |
184 | 0 | memory += sizeof(*set); |
185 | 0 | set->histograms = (VP8LHistogram**)memory; |
186 | 0 | set->max_size = size; |
187 | 0 | set->size = size; |
188 | 0 | HistogramSetResetPointers(set, cache_bits); |
189 | 0 | for (i = 0; i < size; ++i) { |
190 | 0 | set->histograms[i]->palette_code_bits = cache_bits; |
191 | 0 | } |
192 | 0 | } |
193 | | |
194 | | // Removes the histogram 'i' from 'set'. |
195 | 0 | static void HistogramSetRemoveHistogram(VP8LHistogramSet* const set, int i) { |
196 | 0 | set->histograms[i] = set->histograms[set->size - 1]; |
197 | 0 | --set->size; |
198 | 0 | assert(set->size > 0); |
199 | 0 | } |
200 | | |
201 | | // ----------------------------------------------------------------------------- |
202 | | |
203 | | static void HistogramAddSinglePixOrCopy( |
204 | | VP8LHistogram* const histo, const PixOrCopy* const v, |
205 | 0 | int (*const distance_modifier)(int, int), int distance_modifier_arg0) { |
206 | 0 | if (PixOrCopyIsLiteral(v)) { |
207 | 0 | ++histo->alpha[PixOrCopyLiteral(v, 3)]; |
208 | 0 | ++histo->red[PixOrCopyLiteral(v, 2)]; |
209 | 0 | ++histo->literal[PixOrCopyLiteral(v, 1)]; |
210 | 0 | ++histo->blue[PixOrCopyLiteral(v, 0)]; |
211 | 0 | } else if (PixOrCopyIsCacheIdx(v)) { |
212 | 0 | const int literal_ix = |
213 | 0 | NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v); |
214 | 0 | assert(histo->palette_code_bits != 0); |
215 | 0 | ++histo->literal[literal_ix]; |
216 | 0 | } else { |
217 | 0 | int code, extra_bits; |
218 | 0 | VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits); |
219 | 0 | ++histo->literal[NUM_LITERAL_CODES + code]; |
220 | 0 | if (distance_modifier == NULL) { |
221 | 0 | VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits); |
222 | 0 | } else { |
223 | 0 | VP8LPrefixEncodeBits( |
224 | 0 | distance_modifier(distance_modifier_arg0, PixOrCopyDistance(v)), |
225 | 0 | &code, &extra_bits); |
226 | 0 | } |
227 | 0 | ++histo->distance[code]; |
228 | 0 | } |
229 | 0 | } |
230 | | |
231 | | void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs, |
232 | | int (*const distance_modifier)(int, int), |
233 | | int distance_modifier_arg0, |
234 | 0 | VP8LHistogram* const histo) { |
235 | 0 | VP8LRefsCursor c = VP8LRefsCursorInit(refs); |
236 | 0 | while (VP8LRefsCursorOk(&c)) { |
237 | 0 | HistogramAddSinglePixOrCopy(histo, c.cur_pos, distance_modifier, |
238 | 0 | distance_modifier_arg0); |
239 | 0 | VP8LRefsCursorNext(&c); |
240 | 0 | } |
241 | 0 | } |
242 | | |
243 | | // ----------------------------------------------------------------------------- |
244 | | // Entropy-related functions. |
245 | | |
246 | 0 | static WEBP_INLINE uint64_t BitsEntropyRefine(const VP8LBitEntropy* entropy) { |
247 | 0 | uint64_t mix; |
248 | 0 | if (entropy->nonzeros < 5) { |
249 | 0 | if (entropy->nonzeros <= 1) { |
250 | 0 | return 0; |
251 | 0 | } |
252 | | // Two symbols, they will be 0 and 1 in a Huffman code. |
253 | | // Let's mix in a bit of entropy to favor good clustering when |
254 | | // distributions of these are combined. |
255 | 0 | if (entropy->nonzeros == 2) { |
256 | 0 | return DivRound(99 * ((uint64_t)entropy->sum << LOG_2_PRECISION_BITS) + |
257 | 0 | entropy->entropy, |
258 | 0 | 100); |
259 | 0 | } |
260 | | // No matter what the entropy says, we cannot be better than min_limit |
261 | | // with Huffman coding. I am mixing a bit of entropy into the |
262 | | // min_limit since it produces much better (~0.5 %) compression results |
263 | | // perhaps because of better entropy clustering. |
264 | 0 | if (entropy->nonzeros == 3) { |
265 | 0 | mix = 950; |
266 | 0 | } else { |
267 | 0 | mix = 700; // nonzeros == 4. |
268 | 0 | } |
269 | 0 | } else { |
270 | 0 | mix = 627; |
271 | 0 | } |
272 | | |
273 | 0 | { |
274 | 0 | uint64_t min_limit = (uint64_t)(2 * entropy->sum - entropy->max_val) |
275 | 0 | << LOG_2_PRECISION_BITS; |
276 | 0 | min_limit = |
277 | 0 | DivRound(mix * min_limit + (1000 - mix) * entropy->entropy, 1000); |
278 | 0 | return (entropy->entropy < min_limit) ? min_limit : entropy->entropy; |
279 | 0 | } |
280 | 0 | } |
281 | | |
282 | 0 | uint64_t VP8LBitsEntropy(const uint32_t* const array, int n) { |
283 | 0 | VP8LBitEntropy entropy; |
284 | 0 | VP8LBitsEntropyUnrefined(array, n, &entropy); |
285 | |
|
286 | 0 | return BitsEntropyRefine(&entropy); |
287 | 0 | } |
288 | | |
289 | 0 | static uint64_t InitialHuffmanCost(void) { |
290 | | // Small bias because Huffman code length is typically not stored in |
291 | | // full length. |
292 | 0 | static const uint64_t kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3; |
293 | | // Subtract a bias of 9.1. |
294 | 0 | return (kHuffmanCodeOfHuffmanCodeSize << LOG_2_PRECISION_BITS) - |
295 | 0 | DivRound(91ll << LOG_2_PRECISION_BITS, 10); |
296 | 0 | } |
297 | | |
298 | | // Finalize the Huffman cost based on streak numbers and length type (<3 or >=3) |
299 | 0 | static uint64_t FinalHuffmanCost(const VP8LStreaks* const stats) { |
300 | | // The constants in this function are empirical and got rounded from |
301 | | // their original values in 1/8 when switched to 1/1024. |
302 | 0 | uint64_t retval = InitialHuffmanCost(); |
303 | | // Second coefficient: Many zeros in the histogram are covered efficiently |
304 | | // by a run-length encode. Originally 2/8. |
305 | 0 | uint32_t retval_extra = stats->counts[0] * 1600 + 240 * stats->streaks[0][1]; |
306 | | // Second coefficient: Constant values are encoded less efficiently, but still |
307 | | // RLE'ed. Originally 6/8. |
308 | 0 | retval_extra += stats->counts[1] * 2640 + 720 * stats->streaks[1][1]; |
309 | | // 0s are usually encoded more efficiently than non-0s. |
310 | | // Originally 15/8. |
311 | 0 | retval_extra += 1840 * stats->streaks[0][0]; |
312 | | // Originally 26/8. |
313 | 0 | retval_extra += 3360 * stats->streaks[1][0]; |
314 | 0 | return retval + ((uint64_t)retval_extra << (LOG_2_PRECISION_BITS - 10)); |
315 | 0 | } |
316 | | |
317 | | // Get the symbol entropy for the distribution 'population'. |
318 | | // Set 'trivial_sym', if there's only one symbol present in the distribution. |
319 | | static uint64_t PopulationCost(const uint32_t* const population, int length, |
320 | | uint16_t* const trivial_sym, |
321 | 0 | uint8_t* const is_used) { |
322 | 0 | VP8LBitEntropy bit_entropy; |
323 | 0 | VP8LStreaks stats; |
324 | 0 | VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats); |
325 | 0 | if (trivial_sym != NULL) { |
326 | 0 | *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code |
327 | 0 | : VP8L_NON_TRIVIAL_SYM; |
328 | 0 | } |
329 | 0 | if (is_used != NULL) { |
330 | | // The histogram is used if there is at least one non-zero streak. |
331 | 0 | *is_used = (stats.streaks[1][0] != 0 || stats.streaks[1][1] != 0); |
332 | 0 | } |
333 | |
|
334 | 0 | return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); |
335 | 0 | } |
336 | | |
337 | | static WEBP_INLINE void GetPopulationInfo(const VP8LHistogram* const histo, |
338 | | HistogramIndex index, |
339 | | const uint32_t** population, |
340 | 0 | int* length) { |
341 | 0 | switch (index) { |
342 | 0 | case LITERAL: |
343 | 0 | *population = histo->literal; |
344 | 0 | *length = VP8LHistogramNumCodes(histo->palette_code_bits); |
345 | 0 | break; |
346 | 0 | case RED: |
347 | 0 | *population = histo->red; |
348 | 0 | *length = NUM_LITERAL_CODES; |
349 | 0 | break; |
350 | 0 | case BLUE: |
351 | 0 | *population = histo->blue; |
352 | 0 | *length = NUM_LITERAL_CODES; |
353 | 0 | break; |
354 | 0 | case ALPHA: |
355 | 0 | *population = histo->alpha; |
356 | 0 | *length = NUM_LITERAL_CODES; |
357 | 0 | break; |
358 | 0 | case DISTANCE: |
359 | 0 | *population = histo->distance; |
360 | 0 | *length = NUM_DISTANCE_CODES; |
361 | 0 | break; |
362 | 0 | } |
363 | 0 | } |
364 | | |
365 | | // trivial_at_end is 1 if the two histograms only have one element that is |
366 | | // non-zero: both the zero-th one, or both the last one. |
367 | | // 'index' is the index of the symbol in the histogram (literal, red, blue, |
368 | | // alpha, distance). |
369 | | static WEBP_INLINE uint64_t GetCombinedEntropy(const VP8LHistogram* const h1, |
370 | | const VP8LHistogram* const h2, |
371 | 0 | HistogramIndex index) { |
372 | 0 | const uint32_t* X; |
373 | 0 | const uint32_t* Y; |
374 | 0 | int length; |
375 | 0 | VP8LStreaks stats; |
376 | 0 | VP8LBitEntropy bit_entropy; |
377 | 0 | const int is_h1_used = h1->is_used[index]; |
378 | 0 | const int is_h2_used = h2->is_used[index]; |
379 | 0 | const int is_trivial = h1->trivial_symbol[index] != VP8L_NON_TRIVIAL_SYM && |
380 | 0 | h1->trivial_symbol[index] == h2->trivial_symbol[index]; |
381 | |
|
382 | 0 | if (is_trivial || !is_h1_used || !is_h2_used) { |
383 | 0 | if (is_h1_used) return h1->costs[index]; |
384 | 0 | return h2->costs[index]; |
385 | 0 | } |
386 | 0 | assert(is_h1_used && is_h2_used); |
387 | | |
388 | 0 | GetPopulationInfo(h1, index, &X, &length); |
389 | 0 | GetPopulationInfo(h2, index, &Y, &length); |
390 | 0 | VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats); |
391 | 0 | return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats); |
392 | 0 | } |
393 | | |
394 | | // Estimates the Entropy + Huffman + other block overhead size cost. |
395 | 0 | uint64_t VP8LHistogramEstimateBits(const VP8LHistogram* const h) { |
396 | 0 | int i; |
397 | 0 | uint64_t cost = 0; |
398 | 0 | for (i = 0; i < 5; ++i) { |
399 | 0 | int length; |
400 | 0 | const uint32_t* population; |
401 | 0 | GetPopulationInfo(h, (HistogramIndex)i, &population, &length); |
402 | 0 | cost += PopulationCost(population, length, /*trivial_sym=*/NULL, |
403 | | /*is_used=*/NULL); |
404 | 0 | } |
405 | 0 | cost += ((uint64_t)(VP8LExtraCost(h->literal + NUM_LITERAL_CODES, |
406 | 0 | NUM_LENGTH_CODES) + |
407 | 0 | VP8LExtraCost(h->distance, NUM_DISTANCE_CODES)) |
408 | 0 | << LOG_2_PRECISION_BITS); |
409 | 0 | return cost; |
410 | 0 | } |
411 | | |
412 | | // ----------------------------------------------------------------------------- |
413 | | // Various histogram combine/cost-eval functions |
414 | | |
415 | | // Set a + b in b, saturating at WEBP_INT64_MAX. |
416 | 0 | static WEBP_INLINE void SaturateAdd(uint64_t a, int64_t* b) { |
417 | 0 | if (*b < 0 || (int64_t)a <= WEBP_INT64_MAX - *b) { |
418 | 0 | *b += (int64_t)a; |
419 | 0 | } else { |
420 | 0 | *b = WEBP_INT64_MAX; |
421 | 0 | } |
422 | 0 | } |
423 | | |
424 | | // Returns 1 if the cost of the combined histogram is less than the threshold. |
425 | | // Otherwise returns 0 and the cost is invalid due to early bail-out. |
426 | | WEBP_NODISCARD static int GetCombinedHistogramEntropy( |
427 | | const VP8LHistogram* const a, const VP8LHistogram* const b, |
428 | 0 | int64_t cost_threshold_in, uint64_t* cost, uint64_t costs[5]) { |
429 | 0 | int i; |
430 | 0 | const uint64_t cost_threshold = (uint64_t)cost_threshold_in; |
431 | 0 | assert(a->palette_code_bits == b->palette_code_bits); |
432 | 0 | if (cost_threshold_in <= 0) return 0; |
433 | 0 | *cost = 0; |
434 | | |
435 | | // No need to add the extra cost for length and distance as it is a constant |
436 | | // that does not influence the histograms. |
437 | 0 | for (i = 0; i < 5; ++i) { |
438 | 0 | costs[i] = GetCombinedEntropy(a, b, (HistogramIndex)i); |
439 | 0 | *cost += costs[i]; |
440 | 0 | if (*cost >= cost_threshold) return 0; |
441 | 0 | } |
442 | | |
443 | 0 | return 1; |
444 | 0 | } |
445 | | |
446 | | static WEBP_INLINE void HistogramAdd(const VP8LHistogram* const h1, |
447 | | const VP8LHistogram* const h2, |
448 | 0 | VP8LHistogram* const hout) { |
449 | 0 | int i; |
450 | 0 | assert(h1->palette_code_bits == h2->palette_code_bits); |
451 | | |
452 | 0 | for (i = 0; i < 5; ++i) { |
453 | 0 | int length; |
454 | 0 | const uint32_t *p1, *p2, *pout_const; |
455 | 0 | uint32_t* pout; |
456 | 0 | GetPopulationInfo(h1, (HistogramIndex)i, &p1, &length); |
457 | 0 | GetPopulationInfo(h2, (HistogramIndex)i, &p2, &length); |
458 | 0 | GetPopulationInfo(hout, (HistogramIndex)i, &pout_const, &length); |
459 | 0 | pout = (uint32_t*)pout_const; |
460 | 0 | if (h2 == hout) { |
461 | 0 | if (h1->is_used[i]) { |
462 | 0 | if (hout->is_used[i]) { |
463 | 0 | VP8LAddVectorEq(p1, pout, length); |
464 | 0 | } else { |
465 | 0 | memcpy(pout, p1, length * sizeof(pout[0])); |
466 | 0 | } |
467 | 0 | } |
468 | 0 | } else { |
469 | 0 | if (h1->is_used[i]) { |
470 | 0 | if (h2->is_used[i]) { |
471 | 0 | VP8LAddVector(p1, p2, pout, length); |
472 | 0 | } else { |
473 | 0 | memcpy(pout, p1, length * sizeof(pout[0])); |
474 | 0 | } |
475 | 0 | } else if (h2->is_used[i]) { |
476 | 0 | memcpy(pout, p2, length * sizeof(pout[0])); |
477 | 0 | } else { |
478 | 0 | memset(pout, 0, length * sizeof(pout[0])); |
479 | 0 | } |
480 | 0 | } |
481 | 0 | } |
482 | |
|
483 | 0 | for (i = 0; i < 5; ++i) { |
484 | 0 | hout->trivial_symbol[i] = h1->trivial_symbol[i] == h2->trivial_symbol[i] |
485 | 0 | ? h1->trivial_symbol[i] |
486 | 0 | : VP8L_NON_TRIVIAL_SYM; |
487 | 0 | hout->is_used[i] = h1->is_used[i] || h2->is_used[i]; |
488 | 0 | } |
489 | 0 | } |
490 | | |
491 | | static void UpdateHistogramCost(uint64_t bit_cost, uint64_t costs[5], |
492 | 0 | VP8LHistogram* const h) { |
493 | 0 | int i; |
494 | 0 | h->bit_cost = bit_cost; |
495 | 0 | for (i = 0; i < 5; ++i) { |
496 | 0 | h->costs[i] = costs[i]; |
497 | 0 | } |
498 | 0 | } |
499 | | |
500 | | // Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing |
501 | | // to the threshold value 'cost_threshold'. The score returned is |
502 | | // Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed. |
503 | | // Since the previous score passed is 'cost_threshold', we only need to compare |
504 | | // the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out |
505 | | // early. |
506 | | // Returns 1 if the cost is less than the threshold. |
507 | | // Otherwise returns 0 and the cost is invalid due to early bail-out. |
508 | | WEBP_NODISCARD static int HistogramAddEval(const VP8LHistogram* const a, |
509 | | const VP8LHistogram* const b, |
510 | | VP8LHistogram* const out, |
511 | 0 | int64_t cost_threshold) { |
512 | 0 | const uint64_t sum_cost = a->bit_cost + b->bit_cost; |
513 | 0 | uint64_t bit_cost, costs[5]; |
514 | 0 | SaturateAdd(sum_cost, &cost_threshold); |
515 | 0 | if (!GetCombinedHistogramEntropy(a, b, cost_threshold, &bit_cost, costs)) { |
516 | 0 | return 0; |
517 | 0 | } |
518 | | |
519 | 0 | HistogramAdd(a, b, out); |
520 | 0 | UpdateHistogramCost(bit_cost, costs, out); |
521 | 0 | return 1; |
522 | 0 | } |
523 | | |
524 | | // Same as HistogramAddEval(), except that the resulting histogram |
525 | | // is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit |
526 | | // the term C(b) which is constant over all the evaluations. |
527 | | // Returns 1 if the cost is less than the threshold. |
528 | | // Otherwise returns 0 and the cost is invalid due to early bail-out. |
529 | | WEBP_NODISCARD static int HistogramAddThresh(const VP8LHistogram* const a, |
530 | | const VP8LHistogram* const b, |
531 | | int64_t cost_threshold, |
532 | 0 | int64_t* cost_out) { |
533 | 0 | uint64_t cost, costs[5]; |
534 | 0 | assert(a != NULL && b != NULL); |
535 | 0 | SaturateAdd(a->bit_cost, &cost_threshold); |
536 | 0 | if (!GetCombinedHistogramEntropy(a, b, cost_threshold, &cost, costs)) { |
537 | 0 | return 0; |
538 | 0 | } |
539 | | |
540 | 0 | *cost_out = (int64_t)cost - (int64_t)a->bit_cost; |
541 | 0 | return 1; |
542 | 0 | } |
543 | | |
544 | | // ----------------------------------------------------------------------------- |
545 | | |
546 | | // The structure to keep track of cost range for the three dominant entropy |
547 | | // symbols. |
548 | | typedef struct { |
549 | | uint64_t literal_max; |
550 | | uint64_t literal_min; |
551 | | uint64_t red_max; |
552 | | uint64_t red_min; |
553 | | uint64_t blue_max; |
554 | | uint64_t blue_min; |
555 | | } DominantCostRange; |
556 | | |
557 | 0 | static void DominantCostRangeInit(DominantCostRange* const c) { |
558 | 0 | c->literal_max = 0; |
559 | 0 | c->literal_min = WEBP_UINT64_MAX; |
560 | 0 | c->red_max = 0; |
561 | 0 | c->red_min = WEBP_UINT64_MAX; |
562 | 0 | c->blue_max = 0; |
563 | 0 | c->blue_min = WEBP_UINT64_MAX; |
564 | 0 | } |
565 | | |
566 | | static void UpdateDominantCostRange( |
567 | 0 | const VP8LHistogram* const h, DominantCostRange* const c) { |
568 | 0 | if (c->literal_max < h->costs[LITERAL]) c->literal_max = h->costs[LITERAL]; |
569 | 0 | if (c->literal_min > h->costs[LITERAL]) c->literal_min = h->costs[LITERAL]; |
570 | 0 | if (c->red_max < h->costs[RED]) c->red_max = h->costs[RED]; |
571 | 0 | if (c->red_min > h->costs[RED]) c->red_min = h->costs[RED]; |
572 | 0 | if (c->blue_max < h->costs[BLUE]) c->blue_max = h->costs[BLUE]; |
573 | 0 | if (c->blue_min > h->costs[BLUE]) c->blue_min = h->costs[BLUE]; |
574 | 0 | } |
575 | | |
576 | 0 | static void ComputeHistogramCost(VP8LHistogram* const h) { |
577 | 0 | int i; |
578 | | // No need to add the extra cost for length and distance as it is a constant |
579 | | // that does not influence the histograms. |
580 | 0 | for (i = 0; i < 5; ++i) { |
581 | 0 | const uint32_t* population; |
582 | 0 | int length; |
583 | 0 | GetPopulationInfo(h, i, &population, &length); |
584 | 0 | h->costs[i] = PopulationCost(population, length, &h->trivial_symbol[i], |
585 | 0 | &h->is_used[i]); |
586 | 0 | } |
587 | 0 | h->bit_cost = h->costs[LITERAL] + h->costs[RED] + h->costs[BLUE] + |
588 | 0 | h->costs[ALPHA] + h->costs[DISTANCE]; |
589 | 0 | } |
590 | | |
591 | 0 | static int GetBinIdForEntropy(uint64_t min, uint64_t max, uint64_t val) { |
592 | 0 | const uint64_t range = max - min; |
593 | 0 | if (range > 0) { |
594 | 0 | const uint64_t delta = val - min; |
595 | 0 | return (int)((NUM_PARTITIONS - 1e-6) * delta / range); |
596 | 0 | } else { |
597 | 0 | return 0; |
598 | 0 | } |
599 | 0 | } |
600 | | |
601 | | static int GetHistoBinIndex(const VP8LHistogram* const h, |
602 | 0 | const DominantCostRange* const c, int low_effort) { |
603 | 0 | int bin_id = |
604 | 0 | GetBinIdForEntropy(c->literal_min, c->literal_max, h->costs[LITERAL]); |
605 | 0 | assert(bin_id < NUM_PARTITIONS); |
606 | 0 | if (!low_effort) { |
607 | 0 | bin_id = bin_id * NUM_PARTITIONS + |
608 | 0 | GetBinIdForEntropy(c->red_min, c->red_max, h->costs[RED]); |
609 | 0 | bin_id = bin_id * NUM_PARTITIONS + |
610 | 0 | GetBinIdForEntropy(c->blue_min, c->blue_max, h->costs[BLUE]); |
611 | 0 | assert(bin_id < BIN_SIZE); |
612 | 0 | } |
613 | 0 | return bin_id; |
614 | 0 | } |
615 | | |
616 | | // Construct the histograms from backward references. |
617 | | static void HistogramBuild( |
618 | | int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs, |
619 | 0 | VP8LHistogramSet* const image_histo) { |
620 | 0 | int x = 0, y = 0; |
621 | 0 | const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits); |
622 | 0 | VP8LHistogram** const histograms = image_histo->histograms; |
623 | 0 | VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs); |
624 | 0 | assert(histo_bits > 0); |
625 | 0 | VP8LHistogramSetClear(image_histo); |
626 | 0 | while (VP8LRefsCursorOk(&c)) { |
627 | 0 | const PixOrCopy* const v = c.cur_pos; |
628 | 0 | const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits); |
629 | 0 | HistogramAddSinglePixOrCopy(histograms[ix], v, NULL, 0); |
630 | 0 | x += PixOrCopyLength(v); |
631 | 0 | while (x >= xsize) { |
632 | 0 | x -= xsize; |
633 | 0 | ++y; |
634 | 0 | } |
635 | 0 | VP8LRefsCursorNext(&c); |
636 | 0 | } |
637 | 0 | } |
638 | | |
639 | | // Copies the histograms and computes its bit_cost. |
640 | | static void HistogramCopyAndAnalyze(VP8LHistogramSet* const orig_histo, |
641 | 0 | VP8LHistogramSet* const image_histo) { |
642 | 0 | int i; |
643 | 0 | VP8LHistogram** const orig_histograms = orig_histo->histograms; |
644 | 0 | VP8LHistogram** const histograms = image_histo->histograms; |
645 | 0 | assert(image_histo->max_size == orig_histo->max_size); |
646 | 0 | image_histo->size = 0; |
647 | 0 | for (i = 0; i < orig_histo->max_size; ++i) { |
648 | 0 | VP8LHistogram* const histo = orig_histograms[i]; |
649 | 0 | ComputeHistogramCost(histo); |
650 | | |
651 | | // Skip the histogram if it is completely empty, which can happen for tiles |
652 | | // with no information (when they are skipped because of LZ77). |
653 | 0 | if (!histo->is_used[LITERAL] && !histo->is_used[RED] && |
654 | 0 | !histo->is_used[BLUE] && !histo->is_used[ALPHA] && |
655 | 0 | !histo->is_used[DISTANCE]) { |
656 | | // The first histogram is always used. |
657 | 0 | assert(i > 0); |
658 | 0 | orig_histograms[i] = NULL; |
659 | 0 | } else { |
660 | | // Copy histograms from orig_histo[] to image_histo[]. |
661 | 0 | HistogramCopy(histo, histograms[image_histo->size]); |
662 | 0 | ++image_histo->size; |
663 | 0 | } |
664 | 0 | } |
665 | 0 | } |
666 | | |
667 | | // Partition histograms to different entropy bins for three dominant (literal, |
668 | | // red and blue) symbol costs and compute the histogram aggregate bit_cost. |
669 | | static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo, |
670 | 0 | int low_effort) { |
671 | 0 | int i; |
672 | 0 | VP8LHistogram** const histograms = image_histo->histograms; |
673 | 0 | const int histo_size = image_histo->size; |
674 | 0 | DominantCostRange cost_range; |
675 | 0 | DominantCostRangeInit(&cost_range); |
676 | | |
677 | | // Analyze the dominant (literal, red and blue) entropy costs. |
678 | 0 | for (i = 0; i < histo_size; ++i) { |
679 | 0 | UpdateDominantCostRange(histograms[i], &cost_range); |
680 | 0 | } |
681 | | |
682 | | // bin-hash histograms on three of the dominant (literal, red and blue) |
683 | | // symbol costs and store the resulting bin_id for each histogram. |
684 | 0 | for (i = 0; i < histo_size; ++i) { |
685 | 0 | histograms[i]->bin_id = |
686 | 0 | GetHistoBinIndex(histograms[i], &cost_range, low_effort); |
687 | 0 | } |
688 | 0 | } |
689 | | |
690 | | // Merges some histograms with same bin_id together if it's advantageous. |
691 | | // Sets the remaining histograms to NULL. |
692 | | // 'combine_cost_factor' has to be divided by 100. |
693 | | static void HistogramCombineEntropyBin(VP8LHistogramSet* const image_histo, |
694 | | VP8LHistogram* cur_combo, int num_bins, |
695 | | int32_t combine_cost_factor, |
696 | 0 | int low_effort) { |
697 | 0 | VP8LHistogram** const histograms = image_histo->histograms; |
698 | 0 | int idx; |
699 | 0 | struct { |
700 | 0 | int16_t first; // position of the histogram that accumulates all |
701 | | // histograms with the same bin_id |
702 | 0 | uint16_t num_combine_failures; // number of combine failures per bin_id |
703 | 0 | } bin_info[BIN_SIZE]; |
704 | |
|
705 | 0 | assert(num_bins <= BIN_SIZE); |
706 | 0 | for (idx = 0; idx < num_bins; ++idx) { |
707 | 0 | bin_info[idx].first = -1; |
708 | 0 | bin_info[idx].num_combine_failures = 0; |
709 | 0 | } |
710 | |
|
711 | 0 | for (idx = 0; idx < image_histo->size;) { |
712 | 0 | const int bin_id = histograms[idx]->bin_id; |
713 | 0 | const int first = bin_info[bin_id].first; |
714 | 0 | if (first == -1) { |
715 | 0 | bin_info[bin_id].first = idx; |
716 | 0 | ++idx; |
717 | 0 | } else if (low_effort) { |
718 | 0 | HistogramAdd(histograms[idx], histograms[first], histograms[first]); |
719 | 0 | HistogramSetRemoveHistogram(image_histo, idx); |
720 | 0 | } else { |
721 | | // try to merge #idx into #first (both share the same bin_id) |
722 | 0 | const uint64_t bit_cost = histograms[idx]->bit_cost; |
723 | 0 | const int64_t bit_cost_thresh = |
724 | 0 | -DivRound((int64_t)bit_cost * combine_cost_factor, 100); |
725 | 0 | if (HistogramAddEval(histograms[first], histograms[idx], cur_combo, |
726 | 0 | bit_cost_thresh)) { |
727 | 0 | const int max_combine_failures = 32; |
728 | | // Try to merge two histograms only if the combo is a trivial one or |
729 | | // the two candidate histograms are already non-trivial. |
730 | | // For some images, 'try_combine' turns out to be false for a lot of |
731 | | // histogram pairs. In that case, we fallback to combining |
732 | | // histograms as usual to avoid increasing the header size. |
733 | 0 | int try_combine = |
734 | 0 | cur_combo->trivial_symbol[RED] != VP8L_NON_TRIVIAL_SYM && |
735 | 0 | cur_combo->trivial_symbol[BLUE] != VP8L_NON_TRIVIAL_SYM && |
736 | 0 | cur_combo->trivial_symbol[ALPHA] != VP8L_NON_TRIVIAL_SYM; |
737 | 0 | if (!try_combine) { |
738 | 0 | try_combine = |
739 | 0 | histograms[idx]->trivial_symbol[RED] == VP8L_NON_TRIVIAL_SYM || |
740 | 0 | histograms[idx]->trivial_symbol[BLUE] == VP8L_NON_TRIVIAL_SYM || |
741 | 0 | histograms[idx]->trivial_symbol[ALPHA] == VP8L_NON_TRIVIAL_SYM; |
742 | 0 | try_combine &= |
743 | 0 | histograms[first]->trivial_symbol[RED] == VP8L_NON_TRIVIAL_SYM || |
744 | 0 | histograms[first]->trivial_symbol[BLUE] == VP8L_NON_TRIVIAL_SYM || |
745 | 0 | histograms[first]->trivial_symbol[ALPHA] == VP8L_NON_TRIVIAL_SYM; |
746 | 0 | } |
747 | 0 | if (try_combine || |
748 | 0 | bin_info[bin_id].num_combine_failures >= max_combine_failures) { |
749 | | // move the (better) merged histogram to its final slot |
750 | 0 | HistogramSwap(&cur_combo, &histograms[first]); |
751 | 0 | HistogramSetRemoveHistogram(image_histo, idx); |
752 | 0 | } else { |
753 | 0 | ++bin_info[bin_id].num_combine_failures; |
754 | 0 | ++idx; |
755 | 0 | } |
756 | 0 | } else { |
757 | 0 | ++idx; |
758 | 0 | } |
759 | 0 | } |
760 | 0 | } |
761 | 0 | if (low_effort) { |
762 | | // for low_effort case, update the final cost when everything is merged |
763 | 0 | for (idx = 0; idx < image_histo->size; ++idx) { |
764 | 0 | ComputeHistogramCost(histograms[idx]); |
765 | 0 | } |
766 | 0 | } |
767 | 0 | } |
768 | | |
769 | | // Implement a Lehmer random number generator with a multiplicative constant of |
770 | | // 48271 and a modulo constant of 2^31 - 1. |
771 | 0 | static uint32_t MyRand(uint32_t* const seed) { |
772 | 0 | *seed = (uint32_t)(((uint64_t)(*seed) * 48271u) % 2147483647u); |
773 | 0 | assert(*seed > 0); |
774 | 0 | return *seed; |
775 | 0 | } |
776 | | |
777 | | // ----------------------------------------------------------------------------- |
778 | | // Histogram pairs priority queue |
779 | | |
780 | | // Pair of histograms. Negative idx1 value means that pair is out-of-date. |
781 | | typedef struct { |
782 | | int idx1; |
783 | | int idx2; |
784 | | int64_t cost_diff; |
785 | | uint64_t cost_combo; |
786 | | uint64_t costs[5]; |
787 | | } HistogramPair; |
788 | | |
789 | | typedef struct { |
790 | | HistogramPair* queue; |
791 | | int size; |
792 | | int max_size; |
793 | | } HistoQueue; |
794 | | |
795 | 0 | static int HistoQueueInit(HistoQueue* const histo_queue, const int max_size) { |
796 | 0 | histo_queue->size = 0; |
797 | 0 | histo_queue->max_size = max_size; |
798 | | // We allocate max_size + 1 because the last element at index "size" is |
799 | | // used as temporary data (and it could be up to max_size). |
800 | 0 | histo_queue->queue = (HistogramPair*)WebPSafeMalloc( |
801 | 0 | histo_queue->max_size + 1, sizeof(*histo_queue->queue)); |
802 | 0 | return histo_queue->queue != NULL; |
803 | 0 | } |
804 | | |
805 | 0 | static void HistoQueueClear(HistoQueue* const histo_queue) { |
806 | 0 | assert(histo_queue != NULL); |
807 | 0 | WebPSafeFree(histo_queue->queue); |
808 | 0 | histo_queue->size = 0; |
809 | 0 | histo_queue->max_size = 0; |
810 | 0 | } |
811 | | |
812 | | // Pop a specific pair in the queue by replacing it with the last one |
813 | | // and shrinking the queue. |
814 | | static void HistoQueuePopPair(HistoQueue* const histo_queue, |
815 | 0 | HistogramPair* const pair) { |
816 | 0 | assert(pair >= histo_queue->queue && |
817 | 0 | pair < (histo_queue->queue + histo_queue->size)); |
818 | 0 | assert(histo_queue->size > 0); |
819 | 0 | *pair = histo_queue->queue[histo_queue->size - 1]; |
820 | 0 | --histo_queue->size; |
821 | 0 | } |
822 | | |
823 | | // Check whether a pair in the queue should be updated as head or not. |
824 | | static void HistoQueueUpdateHead(HistoQueue* const histo_queue, |
825 | 0 | HistogramPair* const pair) { |
826 | 0 | assert(pair->cost_diff < 0); |
827 | 0 | assert(pair >= histo_queue->queue && |
828 | 0 | pair < (histo_queue->queue + histo_queue->size)); |
829 | 0 | assert(histo_queue->size > 0); |
830 | 0 | if (pair->cost_diff < histo_queue->queue[0].cost_diff) { |
831 | | // Replace the best pair. |
832 | 0 | const HistogramPair tmp = histo_queue->queue[0]; |
833 | 0 | histo_queue->queue[0] = *pair; |
834 | 0 | *pair = tmp; |
835 | 0 | } |
836 | 0 | } |
837 | | |
838 | | // Replaces the bad_id with good_id in the pair. |
839 | | static void HistoQueueFixPair(int bad_id, int good_id, |
840 | 0 | HistogramPair* const pair) { |
841 | 0 | if (pair->idx1 == bad_id) pair->idx1 = good_id; |
842 | 0 | if (pair->idx2 == bad_id) pair->idx2 = good_id; |
843 | 0 | if (pair->idx1 > pair->idx2) { |
844 | 0 | const int tmp = pair->idx1; |
845 | 0 | pair->idx1 = pair->idx2; |
846 | 0 | pair->idx2 = tmp; |
847 | 0 | } |
848 | 0 | } |
849 | | |
850 | | // Update the cost diff and combo of a pair of histograms. This needs to be |
851 | | // called when the histograms have been merged with a third one. |
852 | | // Returns 1 if the cost diff is less than the threshold. |
853 | | // Otherwise returns 0 and the cost is invalid due to early bail-out. |
854 | | WEBP_NODISCARD static int HistoQueueUpdatePair(const VP8LHistogram* const h1, |
855 | | const VP8LHistogram* const h2, |
856 | | int64_t cost_threshold, |
857 | 0 | HistogramPair* const pair) { |
858 | 0 | const int64_t sum_cost = h1->bit_cost + h2->bit_cost; |
859 | 0 | SaturateAdd(sum_cost, &cost_threshold); |
860 | 0 | if (!GetCombinedHistogramEntropy(h1, h2, cost_threshold, &pair->cost_combo, |
861 | 0 | pair->costs)) { |
862 | 0 | return 0; |
863 | 0 | } |
864 | 0 | pair->cost_diff = (int64_t)pair->cost_combo - sum_cost; |
865 | 0 | return 1; |
866 | 0 | } |
867 | | |
868 | | // Create a pair from indices "idx1" and "idx2" provided its cost |
869 | | // is inferior to "threshold", a negative entropy. |
870 | | // It returns the cost of the pair, or 0 if it superior to threshold. |
871 | | static int64_t HistoQueuePush(HistoQueue* const histo_queue, |
872 | | VP8LHistogram** const histograms, int idx1, |
873 | 0 | int idx2, int64_t threshold) { |
874 | 0 | const VP8LHistogram* h1; |
875 | 0 | const VP8LHistogram* h2; |
876 | 0 | HistogramPair pair; |
877 | | |
878 | | // Stop here if the queue is full. |
879 | 0 | if (histo_queue->size == histo_queue->max_size) return 0; |
880 | 0 | assert(threshold <= 0); |
881 | 0 | if (idx1 > idx2) { |
882 | 0 | const int tmp = idx2; |
883 | 0 | idx2 = idx1; |
884 | 0 | idx1 = tmp; |
885 | 0 | } |
886 | 0 | pair.idx1 = idx1; |
887 | 0 | pair.idx2 = idx2; |
888 | 0 | h1 = histograms[idx1]; |
889 | 0 | h2 = histograms[idx2]; |
890 | | |
891 | | // Do not even consider the pair if it does not improve the entropy. |
892 | 0 | if (!HistoQueueUpdatePair(h1, h2, threshold, &pair)) return 0; |
893 | | |
894 | 0 | histo_queue->queue[histo_queue->size++] = pair; |
895 | 0 | HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]); |
896 | |
|
897 | 0 | return pair.cost_diff; |
898 | 0 | } |
899 | | |
900 | | // ----------------------------------------------------------------------------- |
901 | | |
902 | | // Combines histograms by continuously choosing the one with the highest cost |
903 | | // reduction. |
904 | 0 | static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo) { |
905 | 0 | int ok = 0; |
906 | 0 | const int image_histo_size = image_histo->size; |
907 | 0 | int i, j; |
908 | 0 | VP8LHistogram** const histograms = image_histo->histograms; |
909 | | // Priority queue of histogram pairs. |
910 | 0 | HistoQueue histo_queue; |
911 | | |
912 | | // image_histo_size^2 for the queue size is safe. If you look at |
913 | | // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes |
914 | | // data to the queue, you insert at most: |
915 | | // - image_histo_size*(image_histo_size-1)/2 (the first two for loops) |
916 | | // - image_histo_size - 1 in the last for loop at the first iteration of |
917 | | // the while loop, image_histo_size - 2 at the second iteration ... |
918 | | // therefore image_histo_size*(image_histo_size-1)/2 overall too |
919 | 0 | if (!HistoQueueInit(&histo_queue, image_histo_size * image_histo_size)) { |
920 | 0 | goto End; |
921 | 0 | } |
922 | | |
923 | | // Initialize the queue. |
924 | 0 | for (i = 0; i < image_histo_size; ++i) { |
925 | 0 | for (j = i + 1; j < image_histo_size; ++j) { |
926 | 0 | HistoQueuePush(&histo_queue, histograms, i, j, 0); |
927 | 0 | } |
928 | 0 | } |
929 | |
|
930 | 0 | while (histo_queue.size > 0) { |
931 | 0 | const int idx1 = histo_queue.queue[0].idx1; |
932 | 0 | const int idx2 = histo_queue.queue[0].idx2; |
933 | 0 | HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]); |
934 | 0 | UpdateHistogramCost(histo_queue.queue[0].cost_combo, |
935 | 0 | histo_queue.queue[0].costs, histograms[idx1]); |
936 | | |
937 | | // Remove merged histogram. |
938 | 0 | HistogramSetRemoveHistogram(image_histo, idx2); |
939 | | |
940 | | // Remove pairs intersecting the just combined best pair. |
941 | 0 | for (i = 0; i < histo_queue.size;) { |
942 | 0 | HistogramPair* const p = histo_queue.queue + i; |
943 | 0 | if (p->idx1 == idx1 || p->idx2 == idx1 || |
944 | 0 | p->idx1 == idx2 || p->idx2 == idx2) { |
945 | 0 | HistoQueuePopPair(&histo_queue, p); |
946 | 0 | } else { |
947 | 0 | HistoQueueFixPair(image_histo->size, idx2, p); |
948 | 0 | HistoQueueUpdateHead(&histo_queue, p); |
949 | 0 | ++i; |
950 | 0 | } |
951 | 0 | } |
952 | | |
953 | | // Push new pairs formed with combined histogram to the queue. |
954 | 0 | for (i = 0; i < image_histo->size; ++i) { |
955 | 0 | if (i == idx1) continue; |
956 | 0 | HistoQueuePush(&histo_queue, image_histo->histograms, idx1, i, 0); |
957 | 0 | } |
958 | 0 | } |
959 | |
|
960 | 0 | ok = 1; |
961 | |
|
962 | 0 | End: |
963 | 0 | HistoQueueClear(&histo_queue); |
964 | 0 | return ok; |
965 | 0 | } |
966 | | |
967 | | // Perform histogram aggregation using a stochastic approach. |
968 | | // 'do_greedy' is set to 1 if a greedy approach needs to be performed |
969 | | // afterwards, 0 otherwise. |
970 | | static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo, |
971 | | int min_cluster_size, |
972 | 0 | int* const do_greedy) { |
973 | 0 | int j, iter; |
974 | 0 | uint32_t seed = 1; |
975 | 0 | int tries_with_no_success = 0; |
976 | 0 | const int outer_iters = image_histo->size; |
977 | 0 | const int num_tries_no_success = outer_iters / 2; |
978 | 0 | VP8LHistogram** const histograms = image_histo->histograms; |
979 | | // Priority queue of histogram pairs. Its size of 'kHistoQueueSize' |
980 | | // impacts the quality of the compression and the speed: the smaller the |
981 | | // faster but the worse for the compression. |
982 | 0 | HistoQueue histo_queue; |
983 | 0 | const int kHistoQueueSize = 9; |
984 | 0 | int ok = 0; |
985 | |
|
986 | 0 | if (image_histo->size < min_cluster_size) { |
987 | 0 | *do_greedy = 1; |
988 | 0 | return 1; |
989 | 0 | } |
990 | | |
991 | 0 | if (!HistoQueueInit(&histo_queue, kHistoQueueSize)) goto End; |
992 | | |
993 | | // Collapse similar histograms in 'image_histo'. |
994 | 0 | for (iter = 0; iter < outer_iters && image_histo->size >= min_cluster_size && |
995 | 0 | ++tries_with_no_success < num_tries_no_success; |
996 | 0 | ++iter) { |
997 | 0 | int64_t best_cost = |
998 | 0 | (histo_queue.size == 0) ? 0 : histo_queue.queue[0].cost_diff; |
999 | 0 | int best_idx1 = -1, best_idx2 = 1; |
1000 | 0 | const uint32_t rand_range = (image_histo->size - 1) * (image_histo->size); |
1001 | | // (image_histo->size) / 2 was chosen empirically. Less means faster but |
1002 | | // worse compression. |
1003 | 0 | const int num_tries = (image_histo->size) / 2; |
1004 | | |
1005 | | // Pick random samples. |
1006 | 0 | for (j = 0; image_histo->size >= 2 && j < num_tries; ++j) { |
1007 | 0 | int64_t curr_cost; |
1008 | | // Choose two different histograms at random and try to combine them. |
1009 | 0 | const uint32_t tmp = MyRand(&seed) % rand_range; |
1010 | 0 | uint32_t idx1 = tmp / (image_histo->size - 1); |
1011 | 0 | uint32_t idx2 = tmp % (image_histo->size - 1); |
1012 | 0 | if (idx2 >= idx1) ++idx2; |
1013 | | |
1014 | | // Calculate cost reduction on combination. |
1015 | 0 | curr_cost = |
1016 | 0 | HistoQueuePush(&histo_queue, histograms, idx1, idx2, best_cost); |
1017 | 0 | if (curr_cost < 0) { // found a better pair? |
1018 | 0 | best_cost = curr_cost; |
1019 | | // Empty the queue if we reached full capacity. |
1020 | 0 | if (histo_queue.size == histo_queue.max_size) break; |
1021 | 0 | } |
1022 | 0 | } |
1023 | 0 | if (histo_queue.size == 0) continue; |
1024 | | |
1025 | | // Get the best histograms. |
1026 | 0 | best_idx1 = histo_queue.queue[0].idx1; |
1027 | 0 | best_idx2 = histo_queue.queue[0].idx2; |
1028 | 0 | assert(best_idx1 < best_idx2); |
1029 | | // Merge the histograms and remove best_idx2 from the queue. |
1030 | 0 | HistogramAdd(histograms[best_idx2], histograms[best_idx1], |
1031 | 0 | histograms[best_idx1]); |
1032 | 0 | UpdateHistogramCost(histo_queue.queue[0].cost_combo, |
1033 | 0 | histo_queue.queue[0].costs, histograms[best_idx1]); |
1034 | 0 | HistogramSetRemoveHistogram(image_histo, best_idx2); |
1035 | | // Parse the queue and update each pair that deals with best_idx1, |
1036 | | // best_idx2 or image_histo_size. |
1037 | 0 | for (j = 0; j < histo_queue.size;) { |
1038 | 0 | HistogramPair* const p = histo_queue.queue + j; |
1039 | 0 | const int is_idx1_best = p->idx1 == best_idx1 || p->idx1 == best_idx2; |
1040 | 0 | const int is_idx2_best = p->idx2 == best_idx1 || p->idx2 == best_idx2; |
1041 | | // The front pair could have been duplicated by a random pick so |
1042 | | // check for it all the time nevertheless. |
1043 | 0 | if (is_idx1_best && is_idx2_best) { |
1044 | 0 | HistoQueuePopPair(&histo_queue, p); |
1045 | 0 | continue; |
1046 | 0 | } |
1047 | | // Any pair containing one of the two best indices should only refer to |
1048 | | // best_idx1. Its cost should also be updated. |
1049 | 0 | if (is_idx1_best || is_idx2_best) { |
1050 | 0 | HistoQueueFixPair(best_idx2, best_idx1, p); |
1051 | | // Re-evaluate the cost of an updated pair. |
1052 | 0 | if (!HistoQueueUpdatePair(histograms[p->idx1], histograms[p->idx2], 0, |
1053 | 0 | p)) { |
1054 | 0 | HistoQueuePopPair(&histo_queue, p); |
1055 | 0 | continue; |
1056 | 0 | } |
1057 | 0 | } |
1058 | 0 | HistoQueueFixPair(image_histo->size, best_idx2, p); |
1059 | 0 | HistoQueueUpdateHead(&histo_queue, p); |
1060 | 0 | ++j; |
1061 | 0 | } |
1062 | 0 | tries_with_no_success = 0; |
1063 | 0 | } |
1064 | 0 | *do_greedy = (image_histo->size <= min_cluster_size); |
1065 | 0 | ok = 1; |
1066 | |
|
1067 | 0 | End: |
1068 | 0 | HistoQueueClear(&histo_queue); |
1069 | 0 | return ok; |
1070 | 0 | } |
1071 | | |
1072 | | // ----------------------------------------------------------------------------- |
1073 | | // Histogram refinement |
1074 | | |
1075 | | // Find the best 'out' histogram for each of the 'in' histograms. |
1076 | | // At call-time, 'out' contains the histograms of the clusters. |
1077 | | // Note: we assume that out[]->bit_cost is already up-to-date. |
1078 | | static void HistogramRemap(const VP8LHistogramSet* const in, |
1079 | | VP8LHistogramSet* const out, |
1080 | 0 | uint32_t* const symbols) { |
1081 | 0 | int i; |
1082 | 0 | VP8LHistogram** const in_histo = in->histograms; |
1083 | 0 | VP8LHistogram** const out_histo = out->histograms; |
1084 | 0 | const int in_size = out->max_size; |
1085 | 0 | const int out_size = out->size; |
1086 | 0 | if (out_size > 1) { |
1087 | 0 | for (i = 0; i < in_size; ++i) { |
1088 | 0 | int best_out = 0; |
1089 | 0 | int64_t best_bits = WEBP_INT64_MAX; |
1090 | 0 | int k; |
1091 | 0 | if (in_histo[i] == NULL) { |
1092 | | // Arbitrarily set to the previous value if unused to help future LZ77. |
1093 | 0 | symbols[i] = symbols[i - 1]; |
1094 | 0 | continue; |
1095 | 0 | } |
1096 | 0 | for (k = 0; k < out_size; ++k) { |
1097 | 0 | int64_t cur_bits; |
1098 | 0 | if (HistogramAddThresh(out_histo[k], in_histo[i], best_bits, |
1099 | 0 | &cur_bits)) { |
1100 | 0 | best_bits = cur_bits; |
1101 | 0 | best_out = k; |
1102 | 0 | } |
1103 | 0 | } |
1104 | 0 | symbols[i] = best_out; |
1105 | 0 | } |
1106 | 0 | } else { |
1107 | 0 | assert(out_size == 1); |
1108 | 0 | for (i = 0; i < in_size; ++i) { |
1109 | 0 | symbols[i] = 0; |
1110 | 0 | } |
1111 | 0 | } |
1112 | | |
1113 | | // Recompute each out based on raw and symbols. |
1114 | 0 | VP8LHistogramSetClear(out); |
1115 | 0 | out->size = out_size; |
1116 | |
|
1117 | 0 | for (i = 0; i < in_size; ++i) { |
1118 | 0 | int idx; |
1119 | 0 | if (in_histo[i] == NULL) continue; |
1120 | 0 | idx = symbols[i]; |
1121 | 0 | HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]); |
1122 | 0 | } |
1123 | 0 | } |
1124 | | |
1125 | 0 | static int32_t GetCombineCostFactor(int histo_size, int quality) { |
1126 | 0 | int32_t combine_cost_factor = 16; |
1127 | 0 | if (quality < 90) { |
1128 | 0 | if (histo_size > 256) combine_cost_factor /= 2; |
1129 | 0 | if (histo_size > 512) combine_cost_factor /= 2; |
1130 | 0 | if (histo_size > 1024) combine_cost_factor /= 2; |
1131 | 0 | if (quality <= 50) combine_cost_factor /= 2; |
1132 | 0 | } |
1133 | 0 | return combine_cost_factor; |
1134 | 0 | } |
1135 | | |
1136 | | int VP8LGetHistoImageSymbols(int xsize, int ysize, |
1137 | | const VP8LBackwardRefs* const refs, int quality, |
1138 | | int low_effort, int histogram_bits, int cache_bits, |
1139 | | VP8LHistogramSet* const image_histo, |
1140 | | VP8LHistogram* const tmp_histo, |
1141 | | uint32_t* const histogram_symbols, |
1142 | | const WebPPicture* const pic, int percent_range, |
1143 | 0 | int* const percent) { |
1144 | 0 | const int histo_xsize = |
1145 | 0 | histogram_bits ? VP8LSubSampleSize(xsize, histogram_bits) : 1; |
1146 | 0 | const int histo_ysize = |
1147 | 0 | histogram_bits ? VP8LSubSampleSize(ysize, histogram_bits) : 1; |
1148 | 0 | const int image_histo_raw_size = histo_xsize * histo_ysize; |
1149 | 0 | VP8LHistogramSet* const orig_histo = |
1150 | 0 | VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits); |
1151 | | // Don't attempt linear bin-partition heuristic for |
1152 | | // histograms of small sizes (as bin_map will be very sparse) and |
1153 | | // maximum quality q==100 (to preserve the compression gains at that level). |
1154 | 0 | const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE; |
1155 | 0 | int entropy_combine; |
1156 | 0 | if (orig_histo == NULL) { |
1157 | 0 | WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY); |
1158 | 0 | goto Error; |
1159 | 0 | } |
1160 | | |
1161 | | // Construct the histograms from backward references. |
1162 | 0 | HistogramBuild(xsize, histogram_bits, refs, orig_histo); |
1163 | 0 | HistogramCopyAndAnalyze(orig_histo, image_histo); |
1164 | 0 | entropy_combine = |
1165 | 0 | (image_histo->size > entropy_combine_num_bins * 2) && (quality < 100); |
1166 | |
|
1167 | 0 | if (entropy_combine) { |
1168 | 0 | const int32_t combine_cost_factor = |
1169 | 0 | GetCombineCostFactor(image_histo_raw_size, quality); |
1170 | |
|
1171 | 0 | HistogramAnalyzeEntropyBin(image_histo, low_effort); |
1172 | | // Collapse histograms with similar entropy. |
1173 | 0 | HistogramCombineEntropyBin(image_histo, tmp_histo, entropy_combine_num_bins, |
1174 | 0 | combine_cost_factor, low_effort); |
1175 | 0 | } |
1176 | | |
1177 | | // Don't combine the histograms using stochastic and greedy heuristics for |
1178 | | // low-effort compression mode. |
1179 | 0 | if (!low_effort || !entropy_combine) { |
1180 | | // cubic ramp between 1 and MAX_HISTO_GREEDY: |
1181 | 0 | const int threshold_size = |
1182 | 0 | (int)(1 + DivRound(quality * quality * quality * (MAX_HISTO_GREEDY - 1), |
1183 | 0 | 100 * 100 * 100)); |
1184 | 0 | int do_greedy; |
1185 | 0 | if (!HistogramCombineStochastic(image_histo, threshold_size, &do_greedy)) { |
1186 | 0 | WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY); |
1187 | 0 | goto Error; |
1188 | 0 | } |
1189 | 0 | if (do_greedy) { |
1190 | 0 | if (!HistogramCombineGreedy(image_histo)) { |
1191 | 0 | WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY); |
1192 | 0 | goto Error; |
1193 | 0 | } |
1194 | 0 | } |
1195 | 0 | } |
1196 | | |
1197 | | // Find the optimal map from original histograms to the final ones. |
1198 | 0 | HistogramRemap(orig_histo, image_histo, histogram_symbols); |
1199 | |
|
1200 | 0 | if (!WebPReportProgress(pic, *percent + percent_range, percent)) { |
1201 | 0 | goto Error; |
1202 | 0 | } |
1203 | | |
1204 | 0 | Error: |
1205 | 0 | VP8LFreeHistogramSet(orig_histo); |
1206 | 0 | return (pic->error_code == VP8_ENC_OK); |
1207 | 0 | } |