1 // Copyright 2011 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 (at) google.com) 11 // 12 // Entropy encoding (Huffman) for webp lossless. 13 14 #include <assert.h> 15 #include <stdlib.h> 16 #include <string.h> 17 #include "./huffman_encode.h" 18 #include "../utils/utils.h" 19 #include "../webp/format_constants.h" 20 21 // ----------------------------------------------------------------------------- 22 // Util function to optimize the symbol map for RLE coding 23 24 // Heuristics for selecting the stride ranges to collapse. 25 static int ValuesShouldBeCollapsedToStrideAverage(int a, int b) { 26 return abs(a - b) < 4; 27 } 28 29 // Change the population counts in a way that the consequent 30 // Hufmann tree compression, especially its RLE-part, give smaller output. 31 static int OptimizeHuffmanForRle(int length, int* const counts) { 32 uint8_t* good_for_rle; 33 // 1) Let's make the Huffman code more compatible with rle encoding. 34 int i; 35 for (; length >= 0; --length) { 36 if (length == 0) { 37 return 1; // All zeros. 38 } 39 if (counts[length - 1] != 0) { 40 // Now counts[0..length - 1] does not have trailing zeros. 41 break; 42 } 43 } 44 // 2) Let's mark all population counts that already can be encoded 45 // with an rle code. 46 good_for_rle = (uint8_t*)calloc(length, 1); 47 if (good_for_rle == NULL) { 48 return 0; 49 } 50 { 51 // Let's not spoil any of the existing good rle codes. 52 // Mark any seq of 0's that is longer as 5 as a good_for_rle. 53 // Mark any seq of non-0's that is longer as 7 as a good_for_rle. 54 int symbol = counts[0]; 55 int stride = 0; 56 for (i = 0; i < length + 1; ++i) { 57 if (i == length || counts[i] != symbol) { 58 if ((symbol == 0 && stride >= 5) || 59 (symbol != 0 && stride >= 7)) { 60 int k; 61 for (k = 0; k < stride; ++k) { 62 good_for_rle[i - k - 1] = 1; 63 } 64 } 65 stride = 1; 66 if (i != length) { 67 symbol = counts[i]; 68 } 69 } else { 70 ++stride; 71 } 72 } 73 } 74 // 3) Let's replace those population counts that lead to more rle codes. 75 { 76 int stride = 0; 77 int limit = counts[0]; 78 int sum = 0; 79 for (i = 0; i < length + 1; ++i) { 80 if (i == length || good_for_rle[i] || 81 (i != 0 && good_for_rle[i - 1]) || 82 !ValuesShouldBeCollapsedToStrideAverage(counts[i], limit)) { 83 if (stride >= 4 || (stride >= 3 && sum == 0)) { 84 int k; 85 // The stride must end, collapse what we have, if we have enough (4). 86 int count = (sum + stride / 2) / stride; 87 if (count < 1) { 88 count = 1; 89 } 90 if (sum == 0) { 91 // Don't make an all zeros stride to be upgraded to ones. 92 count = 0; 93 } 94 for (k = 0; k < stride; ++k) { 95 // We don't want to change value at counts[i], 96 // that is already belonging to the next stride. Thus - 1. 97 counts[i - k - 1] = count; 98 } 99 } 100 stride = 0; 101 sum = 0; 102 if (i < length - 3) { 103 // All interesting strides have a count of at least 4, 104 // at least when non-zeros. 105 limit = (counts[i] + counts[i + 1] + 106 counts[i + 2] + counts[i + 3] + 2) / 4; 107 } else if (i < length) { 108 limit = counts[i]; 109 } else { 110 limit = 0; 111 } 112 } 113 ++stride; 114 if (i != length) { 115 sum += counts[i]; 116 if (stride >= 4) { 117 limit = (sum + stride / 2) / stride; 118 } 119 } 120 } 121 } 122 free(good_for_rle); 123 return 1; 124 } 125 126 typedef struct { 127 int total_count_; 128 int value_; 129 int pool_index_left_; 130 int pool_index_right_; 131 } HuffmanTree; 132 133 // A comparer function for two Huffman trees: sorts first by 'total count' 134 // (more comes first), and then by 'value' (more comes first). 135 static int CompareHuffmanTrees(const void* ptr1, const void* ptr2) { 136 const HuffmanTree* const t1 = (const HuffmanTree*)ptr1; 137 const HuffmanTree* const t2 = (const HuffmanTree*)ptr2; 138 if (t1->total_count_ > t2->total_count_) { 139 return -1; 140 } else if (t1->total_count_ < t2->total_count_) { 141 return 1; 142 } else { 143 assert(t1->value_ != t2->value_); 144 return (t1->value_ < t2->value_) ? -1 : 1; 145 } 146 } 147 148 static void SetBitDepths(const HuffmanTree* const tree, 149 const HuffmanTree* const pool, 150 uint8_t* const bit_depths, int level) { 151 if (tree->pool_index_left_ >= 0) { 152 SetBitDepths(&pool[tree->pool_index_left_], pool, bit_depths, level + 1); 153 SetBitDepths(&pool[tree->pool_index_right_], pool, bit_depths, level + 1); 154 } else { 155 bit_depths[tree->value_] = level; 156 } 157 } 158 159 // Create an optimal Huffman tree. 160 // 161 // (data,length): population counts. 162 // tree_limit: maximum bit depth (inclusive) of the codes. 163 // bit_depths[]: how many bits are used for the symbol. 164 // 165 // Returns 0 when an error has occurred. 166 // 167 // The catch here is that the tree cannot be arbitrarily deep 168 // 169 // count_limit is the value that is to be faked as the minimum value 170 // and this minimum value is raised until the tree matches the 171 // maximum length requirement. 172 // 173 // This algorithm is not of excellent performance for very long data blocks, 174 // especially when population counts are longer than 2**tree_limit, but 175 // we are not planning to use this with extremely long blocks. 176 // 177 // See http://en.wikipedia.org/wiki/Huffman_coding 178 static int GenerateOptimalTree(const int* const histogram, int histogram_size, 179 int tree_depth_limit, 180 uint8_t* const bit_depths) { 181 int count_min; 182 HuffmanTree* tree_pool; 183 HuffmanTree* tree; 184 int tree_size_orig = 0; 185 int i; 186 187 for (i = 0; i < histogram_size; ++i) { 188 if (histogram[i] != 0) { 189 ++tree_size_orig; 190 } 191 } 192 193 if (tree_size_orig == 0) { // pretty optimal already! 194 return 1; 195 } 196 197 // 3 * tree_size is enough to cover all the nodes representing a 198 // population and all the inserted nodes combining two existing nodes. 199 // The tree pool needs 2 * (tree_size_orig - 1) entities, and the 200 // tree needs exactly tree_size_orig entities. 201 tree = (HuffmanTree*)WebPSafeMalloc(3ULL * tree_size_orig, sizeof(*tree)); 202 if (tree == NULL) return 0; 203 tree_pool = tree + tree_size_orig; 204 205 // For block sizes with less than 64k symbols we never need to do a 206 // second iteration of this loop. 207 // If we actually start running inside this loop a lot, we would perhaps 208 // be better off with the Katajainen algorithm. 209 assert(tree_size_orig <= (1 << (tree_depth_limit - 1))); 210 for (count_min = 1; ; count_min *= 2) { 211 int tree_size = tree_size_orig; 212 // We need to pack the Huffman tree in tree_depth_limit bits. 213 // So, we try by faking histogram entries to be at least 'count_min'. 214 int idx = 0; 215 int j; 216 for (j = 0; j < histogram_size; ++j) { 217 if (histogram[j] != 0) { 218 const int count = 219 (histogram[j] < count_min) ? count_min : histogram[j]; 220 tree[idx].total_count_ = count; 221 tree[idx].value_ = j; 222 tree[idx].pool_index_left_ = -1; 223 tree[idx].pool_index_right_ = -1; 224 ++idx; 225 } 226 } 227 228 // Build the Huffman tree. 229 qsort(tree, tree_size, sizeof(*tree), CompareHuffmanTrees); 230 231 if (tree_size > 1) { // Normal case. 232 int tree_pool_size = 0; 233 while (tree_size > 1) { // Finish when we have only one root. 234 int count; 235 tree_pool[tree_pool_size++] = tree[tree_size - 1]; 236 tree_pool[tree_pool_size++] = tree[tree_size - 2]; 237 count = tree_pool[tree_pool_size - 1].total_count_ + 238 tree_pool[tree_pool_size - 2].total_count_; 239 tree_size -= 2; 240 { 241 // Search for the insertion point. 242 int k; 243 for (k = 0; k < tree_size; ++k) { 244 if (tree[k].total_count_ <= count) { 245 break; 246 } 247 } 248 memmove(tree + (k + 1), tree + k, (tree_size - k) * sizeof(*tree)); 249 tree[k].total_count_ = count; 250 tree[k].value_ = -1; 251 252 tree[k].pool_index_left_ = tree_pool_size - 1; 253 tree[k].pool_index_right_ = tree_pool_size - 2; 254 tree_size = tree_size + 1; 255 } 256 } 257 SetBitDepths(&tree[0], tree_pool, bit_depths, 0); 258 } else if (tree_size == 1) { // Trivial case: only one element. 259 bit_depths[tree[0].value_] = 1; 260 } 261 262 { 263 // Test if this Huffman tree satisfies our 'tree_depth_limit' criteria. 264 int max_depth = bit_depths[0]; 265 for (j = 1; j < histogram_size; ++j) { 266 if (max_depth < bit_depths[j]) { 267 max_depth = bit_depths[j]; 268 } 269 } 270 if (max_depth <= tree_depth_limit) { 271 break; 272 } 273 } 274 } 275 free(tree); 276 return 1; 277 } 278 279 // ----------------------------------------------------------------------------- 280 // Coding of the Huffman tree values 281 282 static HuffmanTreeToken* CodeRepeatedValues(int repetitions, 283 HuffmanTreeToken* tokens, 284 int value, int prev_value) { 285 assert(value <= MAX_ALLOWED_CODE_LENGTH); 286 if (value != prev_value) { 287 tokens->code = value; 288 tokens->extra_bits = 0; 289 ++tokens; 290 --repetitions; 291 } 292 while (repetitions >= 1) { 293 if (repetitions < 3) { 294 int i; 295 for (i = 0; i < repetitions; ++i) { 296 tokens->code = value; 297 tokens->extra_bits = 0; 298 ++tokens; 299 } 300 break; 301 } else if (repetitions < 7) { 302 tokens->code = 16; 303 tokens->extra_bits = repetitions - 3; 304 ++tokens; 305 break; 306 } else { 307 tokens->code = 16; 308 tokens->extra_bits = 3; 309 ++tokens; 310 repetitions -= 6; 311 } 312 } 313 return tokens; 314 } 315 316 static HuffmanTreeToken* CodeRepeatedZeros(int repetitions, 317 HuffmanTreeToken* tokens) { 318 while (repetitions >= 1) { 319 if (repetitions < 3) { 320 int i; 321 for (i = 0; i < repetitions; ++i) { 322 tokens->code = 0; // 0-value 323 tokens->extra_bits = 0; 324 ++tokens; 325 } 326 break; 327 } else if (repetitions < 11) { 328 tokens->code = 17; 329 tokens->extra_bits = repetitions - 3; 330 ++tokens; 331 break; 332 } else if (repetitions < 139) { 333 tokens->code = 18; 334 tokens->extra_bits = repetitions - 11; 335 ++tokens; 336 break; 337 } else { 338 tokens->code = 18; 339 tokens->extra_bits = 0x7f; // 138 repeated 0s 340 ++tokens; 341 repetitions -= 138; 342 } 343 } 344 return tokens; 345 } 346 347 int VP8LCreateCompressedHuffmanTree(const HuffmanTreeCode* const tree, 348 HuffmanTreeToken* tokens, int max_tokens) { 349 HuffmanTreeToken* const starting_token = tokens; 350 HuffmanTreeToken* const ending_token = tokens + max_tokens; 351 const int depth_size = tree->num_symbols; 352 int prev_value = 8; // 8 is the initial value for rle. 353 int i = 0; 354 assert(tokens != NULL); 355 while (i < depth_size) { 356 const int value = tree->code_lengths[i]; 357 int k = i + 1; 358 int runs; 359 while (k < depth_size && tree->code_lengths[k] == value) ++k; 360 runs = k - i; 361 if (value == 0) { 362 tokens = CodeRepeatedZeros(runs, tokens); 363 } else { 364 tokens = CodeRepeatedValues(runs, tokens, value, prev_value); 365 prev_value = value; 366 } 367 i += runs; 368 assert(tokens <= ending_token); 369 } 370 (void)ending_token; // suppress 'unused variable' warning 371 return (int)(tokens - starting_token); 372 } 373 374 // ----------------------------------------------------------------------------- 375 376 // Pre-reversed 4-bit values. 377 static const uint8_t kReversedBits[16] = { 378 0x0, 0x8, 0x4, 0xc, 0x2, 0xa, 0x6, 0xe, 379 0x1, 0x9, 0x5, 0xd, 0x3, 0xb, 0x7, 0xf 380 }; 381 382 static uint32_t ReverseBits(int num_bits, uint32_t bits) { 383 uint32_t retval = 0; 384 int i = 0; 385 while (i < num_bits) { 386 i += 4; 387 retval |= kReversedBits[bits & 0xf] << (MAX_ALLOWED_CODE_LENGTH + 1 - i); 388 bits >>= 4; 389 } 390 retval >>= (MAX_ALLOWED_CODE_LENGTH + 1 - num_bits); 391 return retval; 392 } 393 394 // Get the actual bit values for a tree of bit depths. 395 static void ConvertBitDepthsToSymbols(HuffmanTreeCode* const tree) { 396 // 0 bit-depth means that the symbol does not exist. 397 int i; 398 int len; 399 uint32_t next_code[MAX_ALLOWED_CODE_LENGTH + 1]; 400 int depth_count[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 }; 401 402 assert(tree != NULL); 403 len = tree->num_symbols; 404 for (i = 0; i < len; ++i) { 405 const int code_length = tree->code_lengths[i]; 406 assert(code_length <= MAX_ALLOWED_CODE_LENGTH); 407 ++depth_count[code_length]; 408 } 409 depth_count[0] = 0; // ignore unused symbol 410 next_code[0] = 0; 411 { 412 uint32_t code = 0; 413 for (i = 1; i <= MAX_ALLOWED_CODE_LENGTH; ++i) { 414 code = (code + depth_count[i - 1]) << 1; 415 next_code[i] = code; 416 } 417 } 418 for (i = 0; i < len; ++i) { 419 const int code_length = tree->code_lengths[i]; 420 tree->codes[i] = ReverseBits(code_length, next_code[code_length]++); 421 } 422 } 423 424 // ----------------------------------------------------------------------------- 425 // Main entry point 426 427 int VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit, 428 HuffmanTreeCode* const tree) { 429 const int num_symbols = tree->num_symbols; 430 if (!OptimizeHuffmanForRle(num_symbols, histogram)) { 431 return 0; 432 } 433 if (!GenerateOptimalTree(histogram, num_symbols, 434 tree_depth_limit, tree->code_lengths)) { 435 return 0; 436 } 437 // Create the actual bit codes for the bit lengths. 438 ConvertBitDepthsToSymbols(tree); 439 return 1; 440 } 441