• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* NOLINT(build/header_guard) */
2 /* Copyright 2013 Google Inc. All Rights Reserved.
3 
4    Distributed under MIT license.
5    See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
6 */
7 
8 /* template parameters: FN, CODE */
9 
10 #define HistogramType FN(Histogram)
11 
12 /* Computes the bit cost reduction by combining out[idx1] and out[idx2] and if
13    it is below a threshold, stores the pair (idx1, idx2) in the *pairs queue. */
14 BROTLI_INTERNAL void FN(BrotliCompareAndPushToQueue)(
15     const HistogramType* out, const uint32_t* cluster_size, uint32_t idx1,
16     uint32_t idx2, size_t max_num_pairs, HistogramPair* pairs,
17     size_t* num_pairs) CODE({
18   BROTLI_BOOL is_good_pair = BROTLI_FALSE;
19   HistogramPair p;
20   p.idx1 = p.idx2 = 0;
21   p.cost_diff = p.cost_combo = 0;
22   if (idx1 == idx2) {
23     return;
24   }
25   if (idx2 < idx1) {
26     uint32_t t = idx2;
27     idx2 = idx1;
28     idx1 = t;
29   }
30   p.idx1 = idx1;
31   p.idx2 = idx2;
32   p.cost_diff = 0.5 * ClusterCostDiff(cluster_size[idx1], cluster_size[idx2]);
33   p.cost_diff -= out[idx1].bit_cost_;
34   p.cost_diff -= out[idx2].bit_cost_;
35 
36   if (out[idx1].total_count_ == 0) {
37     p.cost_combo = out[idx2].bit_cost_;
38     is_good_pair = BROTLI_TRUE;
39   } else if (out[idx2].total_count_ == 0) {
40     p.cost_combo = out[idx1].bit_cost_;
41     is_good_pair = BROTLI_TRUE;
42   } else {
43     double threshold = *num_pairs == 0 ? 1e99 :
44         BROTLI_MAX(double, 0.0, pairs[0].cost_diff);
45     HistogramType combo = out[idx1];
46     double cost_combo;
47     FN(HistogramAddHistogram)(&combo, &out[idx2]);
48     cost_combo = FN(BrotliPopulationCost)(&combo);
49     if (cost_combo < threshold - p.cost_diff) {
50       p.cost_combo = cost_combo;
51       is_good_pair = BROTLI_TRUE;
52     }
53   }
54   if (is_good_pair) {
55     p.cost_diff += p.cost_combo;
56     if (*num_pairs > 0 && HistogramPairIsLess(&pairs[0], &p)) {
57       /* Replace the top of the queue if needed. */
58       if (*num_pairs < max_num_pairs) {
59         pairs[*num_pairs] = pairs[0];
60         ++(*num_pairs);
61       }
62       pairs[0] = p;
63     } else if (*num_pairs < max_num_pairs) {
64       pairs[*num_pairs] = p;
65       ++(*num_pairs);
66     }
67   }
68 })
69 
70 BROTLI_INTERNAL size_t FN(BrotliHistogramCombine)(HistogramType* out,
71                                                   uint32_t* cluster_size,
72                                                   uint32_t* symbols,
73                                                   uint32_t* clusters,
74                                                   HistogramPair* pairs,
75                                                   size_t num_clusters,
76                                                   size_t symbols_size,
77                                                   size_t max_clusters,
78                                                   size_t max_num_pairs) CODE({
79   double cost_diff_threshold = 0.0;
80   size_t min_cluster_size = 1;
81   size_t num_pairs = 0;
82 
83   {
84     /* We maintain a vector of histogram pairs, with the property that the pair
85        with the maximum bit cost reduction is the first. */
86     size_t idx1;
87     for (idx1 = 0; idx1 < num_clusters; ++idx1) {
88       size_t idx2;
89       for (idx2 = idx1 + 1; idx2 < num_clusters; ++idx2) {
90         FN(BrotliCompareAndPushToQueue)(out, cluster_size, clusters[idx1],
91             clusters[idx2], max_num_pairs, &pairs[0], &num_pairs);
92       }
93     }
94   }
95 
96   while (num_clusters > min_cluster_size) {
97     uint32_t best_idx1;
98     uint32_t best_idx2;
99     size_t i;
100     if (pairs[0].cost_diff >= cost_diff_threshold) {
101       cost_diff_threshold = 1e99;
102       min_cluster_size = max_clusters;
103       continue;
104     }
105     /* Take the best pair from the top of heap. */
106     best_idx1 = pairs[0].idx1;
107     best_idx2 = pairs[0].idx2;
108     FN(HistogramAddHistogram)(&out[best_idx1], &out[best_idx2]);
109     out[best_idx1].bit_cost_ = pairs[0].cost_combo;
110     cluster_size[best_idx1] += cluster_size[best_idx2];
111     for (i = 0; i < symbols_size; ++i) {
112       if (symbols[i] == best_idx2) {
113         symbols[i] = best_idx1;
114       }
115     }
116     for (i = 0; i < num_clusters; ++i) {
117       if (clusters[i] == best_idx2) {
118         memmove(&clusters[i], &clusters[i + 1],
119                 (num_clusters - i - 1) * sizeof(clusters[0]));
120         break;
121       }
122     }
123     --num_clusters;
124     {
125       /* Remove pairs intersecting the just combined best pair. */
126       size_t copy_to_idx = 0;
127       for (i = 0; i < num_pairs; ++i) {
128         HistogramPair* p = &pairs[i];
129         if (p->idx1 == best_idx1 || p->idx2 == best_idx1 ||
130             p->idx1 == best_idx2 || p->idx2 == best_idx2) {
131           /* Remove invalid pair from the queue. */
132           continue;
133         }
134         if (HistogramPairIsLess(&pairs[0], p)) {
135           /* Replace the top of the queue if needed. */
136           HistogramPair front = pairs[0];
137           pairs[0] = *p;
138           pairs[copy_to_idx] = front;
139         } else {
140           pairs[copy_to_idx] = *p;
141         }
142         ++copy_to_idx;
143       }
144       num_pairs = copy_to_idx;
145     }
146 
147     /* Push new pairs formed with the combined histogram to the heap. */
148     for (i = 0; i < num_clusters; ++i) {
149       FN(BrotliCompareAndPushToQueue)(out, cluster_size, best_idx1, clusters[i],
150                                       max_num_pairs, &pairs[0], &num_pairs);
151     }
152   }
153   return num_clusters;
154 })
155 
156 /* What is the bit cost of moving histogram from cur_symbol to candidate. */
157 BROTLI_INTERNAL double FN(BrotliHistogramBitCostDistance)(
158     const HistogramType* histogram, const HistogramType* candidate) CODE({
159   if (histogram->total_count_ == 0) {
160     return 0.0;
161   } else {
162     HistogramType tmp = *histogram;
163     FN(HistogramAddHistogram)(&tmp, candidate);
164     return FN(BrotliPopulationCost)(&tmp) - candidate->bit_cost_;
165   }
166 })
167 
168 /* Find the best 'out' histogram for each of the 'in' histograms.
169    When called, clusters[0..num_clusters) contains the unique values from
170    symbols[0..in_size), but this property is not preserved in this function.
171    Note: we assume that out[]->bit_cost_ is already up-to-date. */
172 BROTLI_INTERNAL void FN(BrotliHistogramRemap)(const HistogramType* in,
173     size_t in_size, const uint32_t* clusters, size_t num_clusters,
174     HistogramType* out, uint32_t* symbols) CODE({
175   size_t i;
176   for (i = 0; i < in_size; ++i) {
177     uint32_t best_out = i == 0 ? symbols[0] : symbols[i - 1];
178     double best_bits =
179         FN(BrotliHistogramBitCostDistance)(&in[i], &out[best_out]);
180     size_t j;
181     for (j = 0; j < num_clusters; ++j) {
182       const double cur_bits =
183           FN(BrotliHistogramBitCostDistance)(&in[i], &out[clusters[j]]);
184       if (cur_bits < best_bits) {
185         best_bits = cur_bits;
186         best_out = clusters[j];
187       }
188     }
189     symbols[i] = best_out;
190   }
191 
192   /* Recompute each out based on raw and symbols. */
193   for (i = 0; i < num_clusters; ++i) {
194     FN(HistogramClear)(&out[clusters[i]]);
195   }
196   for (i = 0; i < in_size; ++i) {
197     FN(HistogramAddHistogram)(&out[symbols[i]], &in[i]);
198   }
199 })
200 
201 /* Reorders elements of the out[0..length) array and changes values in
202    symbols[0..length) array in the following way:
203      * when called, symbols[] contains indexes into out[], and has N unique
204        values (possibly N < length)
205      * on return, symbols'[i] = f(symbols[i]) and
206                   out'[symbols'[i]] = out[symbols[i]], for each 0 <= i < length,
207        where f is a bijection between the range of symbols[] and [0..N), and
208        the first occurrences of values in symbols'[i] come in consecutive
209        increasing order.
210    Returns N, the number of unique values in symbols[]. */
211 BROTLI_INTERNAL size_t FN(BrotliHistogramReindex)(MemoryManager* m,
212     HistogramType* out, uint32_t* symbols, size_t length) CODE({
213   static const uint32_t kInvalidIndex = BROTLI_UINT32_MAX;
214   uint32_t* new_index = BROTLI_ALLOC(m, uint32_t, length);
215   uint32_t next_index;
216   HistogramType* tmp;
217   size_t i;
218   if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(new_index)) return 0;
219   for (i = 0; i < length; ++i) {
220       new_index[i] = kInvalidIndex;
221   }
222   next_index = 0;
223   for (i = 0; i < length; ++i) {
224     if (new_index[symbols[i]] == kInvalidIndex) {
225       new_index[symbols[i]] = next_index;
226       ++next_index;
227     }
228   }
229   /* TODO: by using idea of "cycle-sort" we can avoid allocation of
230      tmp and reduce the number of copying by the factor of 2. */
231   tmp = BROTLI_ALLOC(m, HistogramType, next_index);
232   if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(tmp)) return 0;
233   next_index = 0;
234   for (i = 0; i < length; ++i) {
235     if (new_index[symbols[i]] == next_index) {
236       tmp[next_index] = out[symbols[i]];
237       ++next_index;
238     }
239     symbols[i] = new_index[symbols[i]];
240   }
241   BROTLI_FREE(m, new_index);
242   for (i = 0; i < next_index; ++i) {
243     out[i] = tmp[i];
244   }
245   BROTLI_FREE(m, tmp);
246   return next_index;
247 })
248 
249 BROTLI_INTERNAL void FN(BrotliClusterHistograms)(
250     MemoryManager* m, const HistogramType* in, const size_t in_size,
251     size_t max_histograms, HistogramType* out, size_t* out_size,
252     uint32_t* histogram_symbols) CODE({
253   uint32_t* cluster_size = BROTLI_ALLOC(m, uint32_t, in_size);
254   uint32_t* clusters = BROTLI_ALLOC(m, uint32_t, in_size);
255   size_t num_clusters = 0;
256   const size_t max_input_histograms = 64;
257   size_t pairs_capacity = max_input_histograms * max_input_histograms / 2;
258   /* For the first pass of clustering, we allow all pairs. */
259   HistogramPair* pairs = BROTLI_ALLOC(m, HistogramPair, pairs_capacity + 1);
260   size_t i;
261 
262   if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(cluster_size) ||
263       BROTLI_IS_NULL(clusters) || BROTLI_IS_NULL(pairs)) {
264     return;
265   }
266 
267   for (i = 0; i < in_size; ++i) {
268     cluster_size[i] = 1;
269   }
270 
271   for (i = 0; i < in_size; ++i) {
272     out[i] = in[i];
273     out[i].bit_cost_ = FN(BrotliPopulationCost)(&in[i]);
274     histogram_symbols[i] = (uint32_t)i;
275   }
276 
277   for (i = 0; i < in_size; i += max_input_histograms) {
278     size_t num_to_combine =
279         BROTLI_MIN(size_t, in_size - i, max_input_histograms);
280     size_t num_new_clusters;
281     size_t j;
282     for (j = 0; j < num_to_combine; ++j) {
283       clusters[num_clusters + j] = (uint32_t)(i + j);
284     }
285     num_new_clusters =
286         FN(BrotliHistogramCombine)(out, cluster_size,
287                                    &histogram_symbols[i],
288                                    &clusters[num_clusters], pairs,
289                                    num_to_combine, num_to_combine,
290                                    max_histograms, pairs_capacity);
291     num_clusters += num_new_clusters;
292   }
293 
294   {
295     /* For the second pass, we limit the total number of histogram pairs.
296        After this limit is reached, we only keep searching for the best pair. */
297     size_t max_num_pairs = BROTLI_MIN(size_t,
298         64 * num_clusters, (num_clusters / 2) * num_clusters);
299     BROTLI_ENSURE_CAPACITY(
300         m, HistogramPair, pairs, pairs_capacity, max_num_pairs + 1);
301     if (BROTLI_IS_OOM(m)) return;
302 
303     /* Collapse similar histograms. */
304     num_clusters = FN(BrotliHistogramCombine)(out, cluster_size,
305                                               histogram_symbols, clusters,
306                                               pairs, num_clusters, in_size,
307                                               max_histograms, max_num_pairs);
308   }
309   BROTLI_FREE(m, pairs);
310   BROTLI_FREE(m, cluster_size);
311   /* Find the optimal map from original histograms to the final ones. */
312   FN(BrotliHistogramRemap)(in, in_size, clusters, num_clusters,
313                            out, histogram_symbols);
314   BROTLI_FREE(m, clusters);
315   /* Convert the context map to a canonical form. */
316   *out_size = FN(BrotliHistogramReindex)(m, out, histogram_symbols, in_size);
317   if (BROTLI_IS_OOM(m)) return;
318 })
319 
320 #undef HistogramType
321