1 // Copyright 2011 Google Inc. All Rights Reserved.
2 //
3 // This code is licensed under the same terms as WebM:
4 // Software License Agreement: http://www.webmproject.org/license/software/
5 // Additional IP Rights Grant: http://www.webmproject.org/license/additional/
6 // -----------------------------------------------------------------------------
7 //
8 // Author: Jyrki Alakuijala (jyrki@google.com)
9 //
10 // Entropy encoding (Huffman) for webp lossless.
11
12 #include <assert.h>
13 #include <stdlib.h>
14 #include <string.h>
15 #include "./huffman_encode.h"
16 #include "../utils/utils.h"
17 #include "webp/format_constants.h"
18
19 // -----------------------------------------------------------------------------
20 // Util function to optimize the symbol map for RLE coding
21
22 // Heuristics for selecting the stride ranges to collapse.
ValuesShouldBeCollapsedToStrideAverage(int a,int b)23 static int ValuesShouldBeCollapsedToStrideAverage(int a, int b) {
24 return abs(a - b) < 4;
25 }
26
27 // Change the population counts in a way that the consequent
28 // Hufmann tree compression, especially its RLE-part, give smaller output.
OptimizeHuffmanForRle(int length,int * const counts)29 static int OptimizeHuffmanForRle(int length, int* const counts) {
30 uint8_t* good_for_rle;
31 // 1) Let's make the Huffman code more compatible with rle encoding.
32 int i;
33 for (; length >= 0; --length) {
34 if (length == 0) {
35 return 1; // All zeros.
36 }
37 if (counts[length - 1] != 0) {
38 // Now counts[0..length - 1] does not have trailing zeros.
39 break;
40 }
41 }
42 // 2) Let's mark all population counts that already can be encoded
43 // with an rle code.
44 good_for_rle = (uint8_t*)calloc(length, 1);
45 if (good_for_rle == NULL) {
46 return 0;
47 }
48 {
49 // Let's not spoil any of the existing good rle codes.
50 // Mark any seq of 0's that is longer as 5 as a good_for_rle.
51 // Mark any seq of non-0's that is longer as 7 as a good_for_rle.
52 int symbol = counts[0];
53 int stride = 0;
54 for (i = 0; i < length + 1; ++i) {
55 if (i == length || counts[i] != symbol) {
56 if ((symbol == 0 && stride >= 5) ||
57 (symbol != 0 && stride >= 7)) {
58 int k;
59 for (k = 0; k < stride; ++k) {
60 good_for_rle[i - k - 1] = 1;
61 }
62 }
63 stride = 1;
64 if (i != length) {
65 symbol = counts[i];
66 }
67 } else {
68 ++stride;
69 }
70 }
71 }
72 // 3) Let's replace those population counts that lead to more rle codes.
73 {
74 int stride = 0;
75 int limit = counts[0];
76 int sum = 0;
77 for (i = 0; i < length + 1; ++i) {
78 if (i == length || good_for_rle[i] ||
79 (i != 0 && good_for_rle[i - 1]) ||
80 !ValuesShouldBeCollapsedToStrideAverage(counts[i], limit)) {
81 if (stride >= 4 || (stride >= 3 && sum == 0)) {
82 int k;
83 // The stride must end, collapse what we have, if we have enough (4).
84 int count = (sum + stride / 2) / stride;
85 if (count < 1) {
86 count = 1;
87 }
88 if (sum == 0) {
89 // Don't make an all zeros stride to be upgraded to ones.
90 count = 0;
91 }
92 for (k = 0; k < stride; ++k) {
93 // We don't want to change value at counts[i],
94 // that is already belonging to the next stride. Thus - 1.
95 counts[i - k - 1] = count;
96 }
97 }
98 stride = 0;
99 sum = 0;
100 if (i < length - 3) {
101 // All interesting strides have a count of at least 4,
102 // at least when non-zeros.
103 limit = (counts[i] + counts[i + 1] +
104 counts[i + 2] + counts[i + 3] + 2) / 4;
105 } else if (i < length) {
106 limit = counts[i];
107 } else {
108 limit = 0;
109 }
110 }
111 ++stride;
112 if (i != length) {
113 sum += counts[i];
114 if (stride >= 4) {
115 limit = (sum + stride / 2) / stride;
116 }
117 }
118 }
119 }
120 free(good_for_rle);
121 return 1;
122 }
123
124 typedef struct {
125 int total_count_;
126 int value_;
127 int pool_index_left_;
128 int pool_index_right_;
129 } HuffmanTree;
130
131 // A comparer function for two Huffman trees: sorts first by 'total count'
132 // (more comes first), and then by 'value' (more comes first).
CompareHuffmanTrees(const void * ptr1,const void * ptr2)133 static int CompareHuffmanTrees(const void* ptr1, const void* ptr2) {
134 const HuffmanTree* const t1 = (const HuffmanTree*)ptr1;
135 const HuffmanTree* const t2 = (const HuffmanTree*)ptr2;
136 if (t1->total_count_ > t2->total_count_) {
137 return -1;
138 } else if (t1->total_count_ < t2->total_count_) {
139 return 1;
140 } else {
141 assert(t1->value_ != t2->value_);
142 return (t1->value_ < t2->value_) ? -1 : 1;
143 }
144 }
145
SetBitDepths(const HuffmanTree * const tree,const HuffmanTree * const pool,uint8_t * const bit_depths,int level)146 static void SetBitDepths(const HuffmanTree* const tree,
147 const HuffmanTree* const pool,
148 uint8_t* const bit_depths, int level) {
149 if (tree->pool_index_left_ >= 0) {
150 SetBitDepths(&pool[tree->pool_index_left_], pool, bit_depths, level + 1);
151 SetBitDepths(&pool[tree->pool_index_right_], pool, bit_depths, level + 1);
152 } else {
153 bit_depths[tree->value_] = level;
154 }
155 }
156
157 // Create an optimal Huffman tree.
158 //
159 // (data,length): population counts.
160 // tree_limit: maximum bit depth (inclusive) of the codes.
161 // bit_depths[]: how many bits are used for the symbol.
162 //
163 // Returns 0 when an error has occurred.
164 //
165 // The catch here is that the tree cannot be arbitrarily deep
166 //
167 // count_limit is the value that is to be faked as the minimum value
168 // and this minimum value is raised until the tree matches the
169 // maximum length requirement.
170 //
171 // This algorithm is not of excellent performance for very long data blocks,
172 // especially when population counts are longer than 2**tree_limit, but
173 // we are not planning to use this with extremely long blocks.
174 //
175 // See http://en.wikipedia.org/wiki/Huffman_coding
GenerateOptimalTree(const int * const histogram,int histogram_size,int tree_depth_limit,uint8_t * const bit_depths)176 static int GenerateOptimalTree(const int* const histogram, int histogram_size,
177 int tree_depth_limit,
178 uint8_t* const bit_depths) {
179 int count_min;
180 HuffmanTree* tree_pool;
181 HuffmanTree* tree;
182 int tree_size_orig = 0;
183 int i;
184
185 for (i = 0; i < histogram_size; ++i) {
186 if (histogram[i] != 0) {
187 ++tree_size_orig;
188 }
189 }
190
191 if (tree_size_orig == 0) { // pretty optimal already!
192 return 1;
193 }
194
195 // 3 * tree_size is enough to cover all the nodes representing a
196 // population and all the inserted nodes combining two existing nodes.
197 // The tree pool needs 2 * (tree_size_orig - 1) entities, and the
198 // tree needs exactly tree_size_orig entities.
199 tree = (HuffmanTree*)WebPSafeMalloc(3ULL * tree_size_orig, sizeof(*tree));
200 if (tree == NULL) return 0;
201 tree_pool = tree + tree_size_orig;
202
203 // For block sizes with less than 64k symbols we never need to do a
204 // second iteration of this loop.
205 // If we actually start running inside this loop a lot, we would perhaps
206 // be better off with the Katajainen algorithm.
207 assert(tree_size_orig <= (1 << (tree_depth_limit - 1)));
208 for (count_min = 1; ; count_min *= 2) {
209 int tree_size = tree_size_orig;
210 // We need to pack the Huffman tree in tree_depth_limit bits.
211 // So, we try by faking histogram entries to be at least 'count_min'.
212 int idx = 0;
213 int j;
214 for (j = 0; j < histogram_size; ++j) {
215 if (histogram[j] != 0) {
216 const int count =
217 (histogram[j] < count_min) ? count_min : histogram[j];
218 tree[idx].total_count_ = count;
219 tree[idx].value_ = j;
220 tree[idx].pool_index_left_ = -1;
221 tree[idx].pool_index_right_ = -1;
222 ++idx;
223 }
224 }
225
226 // Build the Huffman tree.
227 qsort(tree, tree_size, sizeof(*tree), CompareHuffmanTrees);
228
229 if (tree_size > 1) { // Normal case.
230 int tree_pool_size = 0;
231 while (tree_size > 1) { // Finish when we have only one root.
232 int count;
233 tree_pool[tree_pool_size++] = tree[tree_size - 1];
234 tree_pool[tree_pool_size++] = tree[tree_size - 2];
235 count = tree_pool[tree_pool_size - 1].total_count_ +
236 tree_pool[tree_pool_size - 2].total_count_;
237 tree_size -= 2;
238 {
239 // Search for the insertion point.
240 int k;
241 for (k = 0; k < tree_size; ++k) {
242 if (tree[k].total_count_ <= count) {
243 break;
244 }
245 }
246 memmove(tree + (k + 1), tree + k, (tree_size - k) * sizeof(*tree));
247 tree[k].total_count_ = count;
248 tree[k].value_ = -1;
249
250 tree[k].pool_index_left_ = tree_pool_size - 1;
251 tree[k].pool_index_right_ = tree_pool_size - 2;
252 tree_size = tree_size + 1;
253 }
254 }
255 SetBitDepths(&tree[0], tree_pool, bit_depths, 0);
256 } else if (tree_size == 1) { // Trivial case: only one element.
257 bit_depths[tree[0].value_] = 1;
258 }
259
260 {
261 // Test if this Huffman tree satisfies our 'tree_depth_limit' criteria.
262 int max_depth = bit_depths[0];
263 for (j = 1; j < histogram_size; ++j) {
264 if (max_depth < bit_depths[j]) {
265 max_depth = bit_depths[j];
266 }
267 }
268 if (max_depth <= tree_depth_limit) {
269 break;
270 }
271 }
272 }
273 free(tree);
274 return 1;
275 }
276
277 // -----------------------------------------------------------------------------
278 // Coding of the Huffman tree values
279
CodeRepeatedValues(int repetitions,HuffmanTreeToken * tokens,int value,int prev_value)280 static HuffmanTreeToken* CodeRepeatedValues(int repetitions,
281 HuffmanTreeToken* tokens,
282 int value, int prev_value) {
283 assert(value <= MAX_ALLOWED_CODE_LENGTH);
284 if (value != prev_value) {
285 tokens->code = value;
286 tokens->extra_bits = 0;
287 ++tokens;
288 --repetitions;
289 }
290 while (repetitions >= 1) {
291 if (repetitions < 3) {
292 int i;
293 for (i = 0; i < repetitions; ++i) {
294 tokens->code = value;
295 tokens->extra_bits = 0;
296 ++tokens;
297 }
298 break;
299 } else if (repetitions < 7) {
300 tokens->code = 16;
301 tokens->extra_bits = repetitions - 3;
302 ++tokens;
303 break;
304 } else {
305 tokens->code = 16;
306 tokens->extra_bits = 3;
307 ++tokens;
308 repetitions -= 6;
309 }
310 }
311 return tokens;
312 }
313
CodeRepeatedZeros(int repetitions,HuffmanTreeToken * tokens)314 static HuffmanTreeToken* CodeRepeatedZeros(int repetitions,
315 HuffmanTreeToken* tokens) {
316 while (repetitions >= 1) {
317 if (repetitions < 3) {
318 int i;
319 for (i = 0; i < repetitions; ++i) {
320 tokens->code = 0; // 0-value
321 tokens->extra_bits = 0;
322 ++tokens;
323 }
324 break;
325 } else if (repetitions < 11) {
326 tokens->code = 17;
327 tokens->extra_bits = repetitions - 3;
328 ++tokens;
329 break;
330 } else if (repetitions < 139) {
331 tokens->code = 18;
332 tokens->extra_bits = repetitions - 11;
333 ++tokens;
334 break;
335 } else {
336 tokens->code = 18;
337 tokens->extra_bits = 0x7f; // 138 repeated 0s
338 ++tokens;
339 repetitions -= 138;
340 }
341 }
342 return tokens;
343 }
344
VP8LCreateCompressedHuffmanTree(const HuffmanTreeCode * const tree,HuffmanTreeToken * tokens,int max_tokens)345 int VP8LCreateCompressedHuffmanTree(const HuffmanTreeCode* const tree,
346 HuffmanTreeToken* tokens, int max_tokens) {
347 HuffmanTreeToken* const starting_token = tokens;
348 HuffmanTreeToken* const ending_token = tokens + max_tokens;
349 const int depth_size = tree->num_symbols;
350 int prev_value = 8; // 8 is the initial value for rle.
351 int i = 0;
352 assert(tokens != NULL);
353 while (i < depth_size) {
354 const int value = tree->code_lengths[i];
355 int k = i + 1;
356 int runs;
357 while (k < depth_size && tree->code_lengths[k] == value) ++k;
358 runs = k - i;
359 if (value == 0) {
360 tokens = CodeRepeatedZeros(runs, tokens);
361 } else {
362 tokens = CodeRepeatedValues(runs, tokens, value, prev_value);
363 prev_value = value;
364 }
365 i += runs;
366 assert(tokens <= ending_token);
367 }
368 (void)ending_token; // suppress 'unused variable' warning
369 return (int)(tokens - starting_token);
370 }
371
372 // -----------------------------------------------------------------------------
373
374 // Pre-reversed 4-bit values.
375 static const uint8_t kReversedBits[16] = {
376 0x0, 0x8, 0x4, 0xc, 0x2, 0xa, 0x6, 0xe,
377 0x1, 0x9, 0x5, 0xd, 0x3, 0xb, 0x7, 0xf
378 };
379
ReverseBits(int num_bits,uint32_t bits)380 static uint32_t ReverseBits(int num_bits, uint32_t bits) {
381 uint32_t retval = 0;
382 int i = 0;
383 while (i < num_bits) {
384 i += 4;
385 retval |= kReversedBits[bits & 0xf] << (MAX_ALLOWED_CODE_LENGTH + 1 - i);
386 bits >>= 4;
387 }
388 retval >>= (MAX_ALLOWED_CODE_LENGTH + 1 - num_bits);
389 return retval;
390 }
391
392 // Get the actual bit values for a tree of bit depths.
ConvertBitDepthsToSymbols(HuffmanTreeCode * const tree)393 static void ConvertBitDepthsToSymbols(HuffmanTreeCode* const tree) {
394 // 0 bit-depth means that the symbol does not exist.
395 int i;
396 int len;
397 uint32_t next_code[MAX_ALLOWED_CODE_LENGTH + 1];
398 int depth_count[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 };
399
400 assert(tree != NULL);
401 len = tree->num_symbols;
402 for (i = 0; i < len; ++i) {
403 const int code_length = tree->code_lengths[i];
404 assert(code_length <= MAX_ALLOWED_CODE_LENGTH);
405 ++depth_count[code_length];
406 }
407 depth_count[0] = 0; // ignore unused symbol
408 next_code[0] = 0;
409 {
410 uint32_t code = 0;
411 for (i = 1; i <= MAX_ALLOWED_CODE_LENGTH; ++i) {
412 code = (code + depth_count[i - 1]) << 1;
413 next_code[i] = code;
414 }
415 }
416 for (i = 0; i < len; ++i) {
417 const int code_length = tree->code_lengths[i];
418 tree->codes[i] = ReverseBits(code_length, next_code[code_length]++);
419 }
420 }
421
422 // -----------------------------------------------------------------------------
423 // Main entry point
424
VP8LCreateHuffmanTree(int * const histogram,int tree_depth_limit,HuffmanTreeCode * const tree)425 int VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit,
426 HuffmanTreeCode* const tree) {
427 const int num_symbols = tree->num_symbols;
428 if (!OptimizeHuffmanForRle(num_symbols, histogram)) {
429 return 0;
430 }
431 if (!GenerateOptimalTree(histogram, num_symbols,
432 tree_depth_limit, tree->code_lengths)) {
433 return 0;
434 }
435 // Create the actual bit codes for the bit lengths.
436 ConvertBitDepthsToSymbols(tree);
437 return 1;
438 }
439