| 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 | // Utilities for building and looking up Huffman trees. |
| 11 | // |
| 12 | // Author: Urvang Joshi (urvang@google.com) |
| 13 | |
| 14 | #include <assert.h> |
| 15 | #include <stdlib.h> |
| 16 | #include <string.h> |
| 17 | #include "src/utils/huffman_utils.h" |
| 18 | #include "src/utils/utils.h" |
| 19 | #include "src/webp/format_constants.h" |
| 20 | |
| 21 | // Huffman data read via DecodeImageStream is represented in two (red and green) |
| 22 | // bytes. |
| 23 | #define MAX_HTREE_GROUPS 0x10000 |
| 24 | |
| 25 | HTreeGroup* VP8LHtreeGroupsNew(int num_htree_groups) { |
| 26 | HTreeGroup* const htree_groups = |
| 27 | (HTreeGroup*)WebPSafeMalloc(num_htree_groups, sizeof(*htree_groups)); |
| 28 | if (htree_groups == NULL) { |
| 29 | return NULL; |
| 30 | } |
| 31 | assert(num_htree_groups <= MAX_HTREE_GROUPS); |
| 32 | return htree_groups; |
| 33 | } |
| 34 | |
| 35 | void VP8LHtreeGroupsFree(HTreeGroup* const htree_groups) { |
| 36 | if (htree_groups != NULL) { |
| 37 | WebPSafeFree(htree_groups); |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | // Returns reverse(reverse(key, len) + 1, len), where reverse(key, len) is the |
| 42 | // bit-wise reversal of the len least significant bits of key. |
| 43 | static WEBP_INLINE uint32_t GetNextKey(uint32_t key, int len) { |
| 44 | uint32_t step = 1 << (len - 1); |
| 45 | while (key & step) { |
| 46 | step >>= 1; |
| 47 | } |
| 48 | return step ? (key & (step - 1)) + step : key; |
| 49 | } |
| 50 | |
| 51 | // Stores code in table[0], table[step], table[2*step], ..., table[end]. |
| 52 | // Assumes that end is an integer multiple of step. |
| 53 | static WEBP_INLINE void ReplicateValue(HuffmanCode* table, |
| 54 | int step, int end, |
| 55 | HuffmanCode code) { |
| 56 | assert(end % step == 0); |
| 57 | do { |
| 58 | end -= step; |
| 59 | table[end] = code; |
| 60 | } while (end > 0); |
| 61 | } |
| 62 | |
| 63 | // Returns the table width of the next 2nd level table. count is the histogram |
| 64 | // of bit lengths for the remaining symbols, len is the code length of the next |
| 65 | // processed symbol |
| 66 | static WEBP_INLINE int NextTableBitSize(const int* const count, |
| 67 | int len, int root_bits) { |
| 68 | int left = 1 << (len - root_bits); |
| 69 | while (len < MAX_ALLOWED_CODE_LENGTH) { |
| 70 | left -= count[len]; |
| 71 | if (left <= 0) break; |
| 72 | ++len; |
| 73 | left <<= 1; |
| 74 | } |
| 75 | return len - root_bits; |
| 76 | } |
| 77 | |
| 78 | // sorted[code_lengths_size] is a pre-allocated array for sorting symbols |
| 79 | // by code length. |
| 80 | static int BuildHuffmanTable(HuffmanCode* const root_table, int root_bits, |
| 81 | const int code_lengths[], int code_lengths_size, |
| 82 | uint16_t sorted[]) { |
| 83 | HuffmanCode* table = root_table; // next available space in table |
| 84 | int total_size = 1 << root_bits; // total size root table + 2nd level table |
| 85 | int len; // current code length |
| 86 | int symbol; // symbol index in original or sorted table |
| 87 | // number of codes of each length: |
| 88 | int count[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 }; |
| 89 | // offsets in sorted table for each length: |
| 90 | int offset[MAX_ALLOWED_CODE_LENGTH + 1]; |
| 91 | |
| 92 | assert(code_lengths_size != 0); |
| 93 | assert(code_lengths != NULL); |
| 94 | assert((root_table != NULL && sorted != NULL) || |
| 95 | (root_table == NULL && sorted == NULL)); |
| 96 | assert(root_bits > 0); |
| 97 | |
| 98 | // Build histogram of code lengths. |
| 99 | for (symbol = 0; symbol < code_lengths_size; ++symbol) { |
| 100 | if (code_lengths[symbol] > MAX_ALLOWED_CODE_LENGTH) { |
| 101 | return 0; |
| 102 | } |
| 103 | ++count[code_lengths[symbol]]; |
| 104 | } |
| 105 | |
| 106 | // Error, all code lengths are zeros. |
| 107 | if (count[0] == code_lengths_size) { |
| 108 | return 0; |
| 109 | } |
| 110 | |
| 111 | // Generate offsets into sorted symbol table by code length. |
| 112 | offset[1] = 0; |
| 113 | for (len = 1; len < MAX_ALLOWED_CODE_LENGTH; ++len) { |
| 114 | if (count[len] > (1 << len)) { |
| 115 | return 0; |
| 116 | } |
| 117 | offset[len + 1] = offset[len] + count[len]; |
| 118 | } |
| 119 | |
| 120 | // Sort symbols by length, by symbol order within each length. |
| 121 | for (symbol = 0; symbol < code_lengths_size; ++symbol) { |
| 122 | const int symbol_code_length = code_lengths[symbol]; |
| 123 | if (code_lengths[symbol] > 0) { |
| 124 | if (sorted != NULL) { |
| 125 | sorted[offset[symbol_code_length]++] = symbol; |
| 126 | } else { |
| 127 | offset[symbol_code_length]++; |
| 128 | } |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | // Special case code with only one value. |
| 133 | if (offset[MAX_ALLOWED_CODE_LENGTH] == 1) { |
| 134 | if (sorted != NULL) { |
| 135 | HuffmanCode code; |
| 136 | code.bits = 0; |
| 137 | code.value = (uint16_t)sorted[0]; |
| 138 | ReplicateValue(table, 1, total_size, code); |
| 139 | } |
| 140 | return total_size; |
| 141 | } |
| 142 | |
| 143 | { |
| 144 | int step; // step size to replicate values in current table |
| 145 | uint32_t low = 0xffffffffu; // low bits for current root entry |
| 146 | uint32_t mask = total_size - 1; // mask for low bits |
| 147 | uint32_t key = 0; // reversed prefix code |
| 148 | int num_nodes = 1; // number of Huffman tree nodes |
| 149 | int num_open = 1; // number of open branches in current tree level |
| 150 | int table_bits = root_bits; // key length of current table |
| 151 | int table_size = 1 << table_bits; // size of current table |
| 152 | symbol = 0; |
| 153 | // Fill in root table. |
| 154 | for (len = 1, step = 2; len <= root_bits; ++len, step <<= 1) { |
| 155 | num_open <<= 1; |
| 156 | num_nodes += num_open; |
| 157 | num_open -= count[len]; |
| 158 | if (num_open < 0) { |
| 159 | return 0; |
| 160 | } |
| 161 | if (root_table == NULL) continue; |
| 162 | for (; count[len] > 0; --count[len]) { |
| 163 | HuffmanCode code; |
| 164 | code.bits = (uint8_t)len; |
| 165 | code.value = (uint16_t)sorted[symbol++]; |
| 166 | ReplicateValue(&table[key], step, table_size, code); |
| 167 | key = GetNextKey(key, len); |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | // Fill in 2nd level tables and add pointers to root table. |
| 172 | for (len = root_bits + 1, step = 2; len <= MAX_ALLOWED_CODE_LENGTH; |
| 173 | ++len, step <<= 1) { |
| 174 | num_open <<= 1; |
| 175 | num_nodes += num_open; |
| 176 | num_open -= count[len]; |
| 177 | if (num_open < 0) { |
| 178 | return 0; |
| 179 | } |
| 180 | for (; count[len] > 0; --count[len]) { |
| 181 | HuffmanCode code; |
| 182 | if ((key & mask) != low) { |
| 183 | if (root_table != NULL) table += table_size; |
| 184 | table_bits = NextTableBitSize(count, len, root_bits); |
| 185 | table_size = 1 << table_bits; |
| 186 | total_size += table_size; |
| 187 | low = key & mask; |
| 188 | if (root_table != NULL) { |
| 189 | root_table[low].bits = (uint8_t)(table_bits + root_bits); |
| 190 | root_table[low].value = (uint16_t)((table - root_table) - low); |
| 191 | } |
| 192 | } |
| 193 | if (root_table != NULL) { |
| 194 | code.bits = (uint8_t)(len - root_bits); |
| 195 | code.value = (uint16_t)sorted[symbol++]; |
| 196 | ReplicateValue(&table[key >> root_bits], step, table_size, code); |
| 197 | } |
| 198 | key = GetNextKey(key, len); |
| 199 | } |
| 200 | } |
| 201 | |
| 202 | // Check if tree is full. |
| 203 | if (num_nodes != 2 * offset[MAX_ALLOWED_CODE_LENGTH] - 1) { |
| 204 | return 0; |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | return total_size; |
| 209 | } |
| 210 | |
| 211 | // Maximum code_lengths_size is 2328 (reached for 11-bit color_cache_bits). |
| 212 | // More commonly, the value is around ~280. |
| 213 | #define MAX_CODE_LENGTHS_SIZE \ |
| 214 | ((1 << MAX_CACHE_BITS) + NUM_LITERAL_CODES + NUM_LENGTH_CODES) |
| 215 | // Cut-off value for switching between heap and stack allocation. |
| 216 | #define SORTED_SIZE_CUTOFF 512 |
| 217 | int VP8LBuildHuffmanTable(HuffmanTables* const root_table, int root_bits, |
| 218 | const int code_lengths[], int code_lengths_size) { |
| 219 | const int total_size = |
| 220 | BuildHuffmanTable(NULL, root_bits, code_lengths, code_lengths_size, NULL); |
| 221 | assert(code_lengths_size <= MAX_CODE_LENGTHS_SIZE); |
| 222 | if (total_size == 0 || root_table == NULL) return total_size; |
| 223 | |
| 224 | if (root_table->curr_segment->curr_table + total_size >= |
| 225 | root_table->curr_segment->start + root_table->curr_segment->size) { |
| 226 | // If 'root_table' does not have enough memory, allocate a new segment. |
| 227 | // The available part of root_table->curr_segment is left unused because we |
| 228 | // need a contiguous buffer. |
| 229 | const int segment_size = root_table->curr_segment->size; |
| 230 | struct HuffmanTablesSegment* next = |
| 231 | (HuffmanTablesSegment*)WebPSafeMalloc(1, sizeof(*next)); |
| 232 | if (next == NULL) return 0; |
| 233 | // Fill the new segment. |
| 234 | // We need at least 'total_size' but if that value is small, it is better to |
| 235 | // allocate a big chunk to prevent more allocations later. 'segment_size' is |
| 236 | // therefore chosen (any other arbitrary value could be chosen). |
| 237 | next->size = total_size > segment_size ? total_size : segment_size; |
| 238 | next->start = |
| 239 | (HuffmanCode*)WebPSafeMalloc(next->size, sizeof(*next->start)); |
| 240 | if (next->start == NULL) { |
| 241 | WebPSafeFree(next); |
| 242 | return 0; |
| 243 | } |
| 244 | next->curr_table = next->start; |
| 245 | next->next = NULL; |
| 246 | // Point to the new segment. |
| 247 | root_table->curr_segment->next = next; |
| 248 | root_table->curr_segment = next; |
| 249 | } |
| 250 | if (code_lengths_size <= SORTED_SIZE_CUTOFF) { |
| 251 | // use local stack-allocated array. |
| 252 | uint16_t sorted[SORTED_SIZE_CUTOFF]; |
| 253 | BuildHuffmanTable(root_table->curr_segment->curr_table, root_bits, |
| 254 | code_lengths, code_lengths_size, sorted); |
| 255 | } else { // rare case. Use heap allocation. |
| 256 | uint16_t* const sorted = |
| 257 | (uint16_t*)WebPSafeMalloc(code_lengths_size, sizeof(*sorted)); |
| 258 | if (sorted == NULL) return 0; |
| 259 | BuildHuffmanTable(root_table->curr_segment->curr_table, root_bits, |
| 260 | code_lengths, code_lengths_size, sorted); |
| 261 | WebPSafeFree(sorted); |
| 262 | } |
| 263 | return total_size; |
| 264 | } |
| 265 | |
| 266 | int VP8LHuffmanTablesAllocate(int size, HuffmanTables* huffman_tables) { |
| 267 | // Have 'segment' point to the first segment for now, 'root'. |
| 268 | HuffmanTablesSegment* const root = &huffman_tables->root; |
| 269 | huffman_tables->curr_segment = root; |
| 270 | // Allocate root. |
| 271 | root->start = (HuffmanCode*)WebPSafeMalloc(size, sizeof(*root->start)); |
| 272 | if (root->start == NULL) return 0; |
| 273 | root->curr_table = root->start; |
| 274 | root->next = NULL; |
| 275 | root->size = size; |
| 276 | return 1; |
| 277 | } |
| 278 | |
| 279 | void VP8LHuffmanTablesDeallocate(HuffmanTables* const huffman_tables) { |
| 280 | HuffmanTablesSegment *current, *next; |
| 281 | if (huffman_tables == NULL) return; |
| 282 | // Free the root node. |
| 283 | current = &huffman_tables->root; |
| 284 | next = current->next; |
| 285 | WebPSafeFree(current->start); |
| 286 | current->start = NULL; |
| 287 | current->next = NULL; |
| 288 | current = next; |
| 289 | // Free the following nodes. |
| 290 | while (current != NULL) { |
| 291 | next = current->next; |
| 292 | WebPSafeFree(current->start); |
| 293 | WebPSafeFree(current); |
| 294 | current = next; |
| 295 | } |
| 296 | } |
| 297 | |