Home | History | Annotate | Download | only in dec
      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 // main entry for the decoder
     11 //
     12 // Authors: Vikas Arora (vikaas.arora (at) gmail.com)
     13 //          Jyrki Alakuijala (jyrki (at) google.com)
     14 
     15 #include <stdlib.h>
     16 
     17 #include "src/dec/alphai_dec.h"
     18 #include "src/dec/vp8li_dec.h"
     19 #include "src/dsp/dsp.h"
     20 #include "src/dsp/lossless.h"
     21 #include "src/dsp/lossless_common.h"
     22 #include "src/dsp/yuv.h"
     23 #include "src/utils/endian_inl_utils.h"
     24 #include "src/utils/huffman_utils.h"
     25 #include "src/utils/utils.h"
     26 
     27 #define NUM_ARGB_CACHE_ROWS          16
     28 
     29 static const int kCodeLengthLiterals = 16;
     30 static const int kCodeLengthRepeatCode = 16;
     31 static const uint8_t kCodeLengthExtraBits[3] = { 2, 3, 7 };
     32 static const uint8_t kCodeLengthRepeatOffsets[3] = { 3, 3, 11 };
     33 
     34 // -----------------------------------------------------------------------------
     35 //  Five Huffman codes are used at each meta code:
     36 //  1. green + length prefix codes + color cache codes,
     37 //  2. alpha,
     38 //  3. red,
     39 //  4. blue, and,
     40 //  5. distance prefix codes.
     41 typedef enum {
     42   GREEN = 0,
     43   RED   = 1,
     44   BLUE  = 2,
     45   ALPHA = 3,
     46   DIST  = 4
     47 } HuffIndex;
     48 
     49 static const uint16_t kAlphabetSize[HUFFMAN_CODES_PER_META_CODE] = {
     50   NUM_LITERAL_CODES + NUM_LENGTH_CODES,
     51   NUM_LITERAL_CODES, NUM_LITERAL_CODES, NUM_LITERAL_CODES,
     52   NUM_DISTANCE_CODES
     53 };
     54 
     55 static const uint8_t kLiteralMap[HUFFMAN_CODES_PER_META_CODE] = {
     56   0, 1, 1, 1, 0
     57 };
     58 
     59 #define NUM_CODE_LENGTH_CODES       19
     60 static const uint8_t kCodeLengthCodeOrder[NUM_CODE_LENGTH_CODES] = {
     61   17, 18, 0, 1, 2, 3, 4, 5, 16, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
     62 };
     63 
     64 #define CODE_TO_PLANE_CODES        120
     65 static const uint8_t kCodeToPlane[CODE_TO_PLANE_CODES] = {
     66   0x18, 0x07, 0x17, 0x19, 0x28, 0x06, 0x27, 0x29, 0x16, 0x1a,
     67   0x26, 0x2a, 0x38, 0x05, 0x37, 0x39, 0x15, 0x1b, 0x36, 0x3a,
     68   0x25, 0x2b, 0x48, 0x04, 0x47, 0x49, 0x14, 0x1c, 0x35, 0x3b,
     69   0x46, 0x4a, 0x24, 0x2c, 0x58, 0x45, 0x4b, 0x34, 0x3c, 0x03,
     70   0x57, 0x59, 0x13, 0x1d, 0x56, 0x5a, 0x23, 0x2d, 0x44, 0x4c,
     71   0x55, 0x5b, 0x33, 0x3d, 0x68, 0x02, 0x67, 0x69, 0x12, 0x1e,
     72   0x66, 0x6a, 0x22, 0x2e, 0x54, 0x5c, 0x43, 0x4d, 0x65, 0x6b,
     73   0x32, 0x3e, 0x78, 0x01, 0x77, 0x79, 0x53, 0x5d, 0x11, 0x1f,
     74   0x64, 0x6c, 0x42, 0x4e, 0x76, 0x7a, 0x21, 0x2f, 0x75, 0x7b,
     75   0x31, 0x3f, 0x63, 0x6d, 0x52, 0x5e, 0x00, 0x74, 0x7c, 0x41,
     76   0x4f, 0x10, 0x20, 0x62, 0x6e, 0x30, 0x73, 0x7d, 0x51, 0x5f,
     77   0x40, 0x72, 0x7e, 0x61, 0x6f, 0x50, 0x71, 0x7f, 0x60, 0x70
     78 };
     79 
     80 // Memory needed for lookup tables of one Huffman tree group. Red, blue, alpha
     81 // and distance alphabets are constant (256 for red, blue and alpha, 40 for
     82 // distance) and lookup table sizes for them in worst case are 630 and 410
     83 // respectively. Size of green alphabet depends on color cache size and is equal
     84 // to 256 (green component values) + 24 (length prefix values)
     85 // + color_cache_size (between 0 and 2048).
     86 // All values computed for 8-bit first level lookup with Mark Adler's tool:
     87 // http://www.hdfgroup.org/ftp/lib-external/zlib/zlib-1.2.5/examples/enough.c
     88 #define FIXED_TABLE_SIZE (630 * 3 + 410)
     89 static const uint16_t kTableSize[12] = {
     90   FIXED_TABLE_SIZE + 654,
     91   FIXED_TABLE_SIZE + 656,
     92   FIXED_TABLE_SIZE + 658,
     93   FIXED_TABLE_SIZE + 662,
     94   FIXED_TABLE_SIZE + 670,
     95   FIXED_TABLE_SIZE + 686,
     96   FIXED_TABLE_SIZE + 718,
     97   FIXED_TABLE_SIZE + 782,
     98   FIXED_TABLE_SIZE + 912,
     99   FIXED_TABLE_SIZE + 1168,
    100   FIXED_TABLE_SIZE + 1680,
    101   FIXED_TABLE_SIZE + 2704
    102 };
    103 
    104 static int DecodeImageStream(int xsize, int ysize,
    105                              int is_level0,
    106                              VP8LDecoder* const dec,
    107                              uint32_t** const decoded_data);
    108 
    109 //------------------------------------------------------------------------------
    110 
    111 int VP8LCheckSignature(const uint8_t* const data, size_t size) {
    112   return (size >= VP8L_FRAME_HEADER_SIZE &&
    113           data[0] == VP8L_MAGIC_BYTE &&
    114           (data[4] >> 5) == 0);  // version
    115 }
    116 
    117 static int ReadImageInfo(VP8LBitReader* const br,
    118                          int* const width, int* const height,
    119                          int* const has_alpha) {
    120   if (VP8LReadBits(br, 8) != VP8L_MAGIC_BYTE) return 0;
    121   *width = VP8LReadBits(br, VP8L_IMAGE_SIZE_BITS) + 1;
    122   *height = VP8LReadBits(br, VP8L_IMAGE_SIZE_BITS) + 1;
    123   *has_alpha = VP8LReadBits(br, 1);
    124   if (VP8LReadBits(br, VP8L_VERSION_BITS) != 0) return 0;
    125   return !br->eos_;
    126 }
    127 
    128 int VP8LGetInfo(const uint8_t* data, size_t data_size,
    129                 int* const width, int* const height, int* const has_alpha) {
    130   if (data == NULL || data_size < VP8L_FRAME_HEADER_SIZE) {
    131     return 0;         // not enough data
    132   } else if (!VP8LCheckSignature(data, data_size)) {
    133     return 0;         // bad signature
    134   } else {
    135     int w, h, a;
    136     VP8LBitReader br;
    137     VP8LInitBitReader(&br, data, data_size);
    138     if (!ReadImageInfo(&br, &w, &h, &a)) {
    139       return 0;
    140     }
    141     if (width != NULL) *width = w;
    142     if (height != NULL) *height = h;
    143     if (has_alpha != NULL) *has_alpha = a;
    144     return 1;
    145   }
    146 }
    147 
    148 //------------------------------------------------------------------------------
    149 
    150 static WEBP_INLINE int GetCopyDistance(int distance_symbol,
    151                                        VP8LBitReader* const br) {
    152   int extra_bits, offset;
    153   if (distance_symbol < 4) {
    154     return distance_symbol + 1;
    155   }
    156   extra_bits = (distance_symbol - 2) >> 1;
    157   offset = (2 + (distance_symbol & 1)) << extra_bits;
    158   return offset + VP8LReadBits(br, extra_bits) + 1;
    159 }
    160 
    161 static WEBP_INLINE int GetCopyLength(int length_symbol,
    162                                      VP8LBitReader* const br) {
    163   // Length and distance prefixes are encoded the same way.
    164   return GetCopyDistance(length_symbol, br);
    165 }
    166 
    167 static WEBP_INLINE int PlaneCodeToDistance(int xsize, int plane_code) {
    168   if (plane_code > CODE_TO_PLANE_CODES) {
    169     return plane_code - CODE_TO_PLANE_CODES;
    170   } else {
    171     const int dist_code = kCodeToPlane[plane_code - 1];
    172     const int yoffset = dist_code >> 4;
    173     const int xoffset = 8 - (dist_code & 0xf);
    174     const int dist = yoffset * xsize + xoffset;
    175     return (dist >= 1) ? dist : 1;  // dist<1 can happen if xsize is very small
    176   }
    177 }
    178 
    179 //------------------------------------------------------------------------------
    180 // Decodes the next Huffman code from bit-stream.
    181 // FillBitWindow(br) needs to be called at minimum every second call
    182 // to ReadSymbol, in order to pre-fetch enough bits.
    183 static WEBP_INLINE int ReadSymbol(const HuffmanCode* table,
    184                                   VP8LBitReader* const br) {
    185   int nbits;
    186   uint32_t val = VP8LPrefetchBits(br);
    187   table += val & HUFFMAN_TABLE_MASK;
    188   nbits = table->bits - HUFFMAN_TABLE_BITS;
    189   if (nbits > 0) {
    190     VP8LSetBitPos(br, br->bit_pos_ + HUFFMAN_TABLE_BITS);
    191     val = VP8LPrefetchBits(br);
    192     table += table->value;
    193     table += val & ((1 << nbits) - 1);
    194   }
    195   VP8LSetBitPos(br, br->bit_pos_ + table->bits);
    196   return table->value;
    197 }
    198 
    199 // Reads packed symbol depending on GREEN channel
    200 #define BITS_SPECIAL_MARKER 0x100  // something large enough (and a bit-mask)
    201 #define PACKED_NON_LITERAL_CODE 0  // must be < NUM_LITERAL_CODES
    202 static WEBP_INLINE int ReadPackedSymbols(const HTreeGroup* group,
    203                                          VP8LBitReader* const br,
    204                                          uint32_t* const dst) {
    205   const uint32_t val = VP8LPrefetchBits(br) & (HUFFMAN_PACKED_TABLE_SIZE - 1);
    206   const HuffmanCode32 code = group->packed_table[val];
    207   assert(group->use_packed_table);
    208   if (code.bits < BITS_SPECIAL_MARKER) {
    209     VP8LSetBitPos(br, br->bit_pos_ + code.bits);
    210     *dst = code.value;
    211     return PACKED_NON_LITERAL_CODE;
    212   } else {
    213     VP8LSetBitPos(br, br->bit_pos_ + code.bits - BITS_SPECIAL_MARKER);
    214     assert(code.value >= NUM_LITERAL_CODES);
    215     return code.value;
    216   }
    217 }
    218 
    219 static int AccumulateHCode(HuffmanCode hcode, int shift,
    220                            HuffmanCode32* const huff) {
    221   huff->bits += hcode.bits;
    222   huff->value |= (uint32_t)hcode.value << shift;
    223   assert(huff->bits <= HUFFMAN_TABLE_BITS);
    224   return hcode.bits;
    225 }
    226 
    227 static void BuildPackedTable(HTreeGroup* const htree_group) {
    228   uint32_t code;
    229   for (code = 0; code < HUFFMAN_PACKED_TABLE_SIZE; ++code) {
    230     uint32_t bits = code;
    231     HuffmanCode32* const huff = &htree_group->packed_table[bits];
    232     HuffmanCode hcode = htree_group->htrees[GREEN][bits];
    233     if (hcode.value >= NUM_LITERAL_CODES) {
    234       huff->bits = hcode.bits + BITS_SPECIAL_MARKER;
    235       huff->value = hcode.value;
    236     } else {
    237       huff->bits = 0;
    238       huff->value = 0;
    239       bits >>= AccumulateHCode(hcode, 8, huff);
    240       bits >>= AccumulateHCode(htree_group->htrees[RED][bits], 16, huff);
    241       bits >>= AccumulateHCode(htree_group->htrees[BLUE][bits], 0, huff);
    242       bits >>= AccumulateHCode(htree_group->htrees[ALPHA][bits], 24, huff);
    243       (void)bits;
    244     }
    245   }
    246 }
    247 
    248 static int ReadHuffmanCodeLengths(
    249     VP8LDecoder* const dec, const int* const code_length_code_lengths,
    250     int num_symbols, int* const code_lengths) {
    251   int ok = 0;
    252   VP8LBitReader* const br = &dec->br_;
    253   int symbol;
    254   int max_symbol;
    255   int prev_code_len = DEFAULT_CODE_LENGTH;
    256   HuffmanCode table[1 << LENGTHS_TABLE_BITS];
    257 
    258   if (!VP8LBuildHuffmanTable(table, LENGTHS_TABLE_BITS,
    259                              code_length_code_lengths,
    260                              NUM_CODE_LENGTH_CODES)) {
    261     goto End;
    262   }
    263 
    264   if (VP8LReadBits(br, 1)) {    // use length
    265     const int length_nbits = 2 + 2 * VP8LReadBits(br, 3);
    266     max_symbol = 2 + VP8LReadBits(br, length_nbits);
    267     if (max_symbol > num_symbols) {
    268       goto End;
    269     }
    270   } else {
    271     max_symbol = num_symbols;
    272   }
    273 
    274   symbol = 0;
    275   while (symbol < num_symbols) {
    276     const HuffmanCode* p;
    277     int code_len;
    278     if (max_symbol-- == 0) break;
    279     VP8LFillBitWindow(br);
    280     p = &table[VP8LPrefetchBits(br) & LENGTHS_TABLE_MASK];
    281     VP8LSetBitPos(br, br->bit_pos_ + p->bits);
    282     code_len = p->value;
    283     if (code_len < kCodeLengthLiterals) {
    284       code_lengths[symbol++] = code_len;
    285       if (code_len != 0) prev_code_len = code_len;
    286     } else {
    287       const int use_prev = (code_len == kCodeLengthRepeatCode);
    288       const int slot = code_len - kCodeLengthLiterals;
    289       const int extra_bits = kCodeLengthExtraBits[slot];
    290       const int repeat_offset = kCodeLengthRepeatOffsets[slot];
    291       int repeat = VP8LReadBits(br, extra_bits) + repeat_offset;
    292       if (symbol + repeat > num_symbols) {
    293         goto End;
    294       } else {
    295         const int length = use_prev ? prev_code_len : 0;
    296         while (repeat-- > 0) code_lengths[symbol++] = length;
    297       }
    298     }
    299   }
    300   ok = 1;
    301 
    302  End:
    303   if (!ok) dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
    304   return ok;
    305 }
    306 
    307 // 'code_lengths' is pre-allocated temporary buffer, used for creating Huffman
    308 // tree.
    309 static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
    310                            int* const code_lengths, HuffmanCode* const table) {
    311   int ok = 0;
    312   int size = 0;
    313   VP8LBitReader* const br = &dec->br_;
    314   const int simple_code = VP8LReadBits(br, 1);
    315 
    316   memset(code_lengths, 0, alphabet_size * sizeof(*code_lengths));
    317 
    318   if (simple_code) {  // Read symbols, codes & code lengths directly.
    319     const int num_symbols = VP8LReadBits(br, 1) + 1;
    320     const int first_symbol_len_code = VP8LReadBits(br, 1);
    321     // The first code is either 1 bit or 8 bit code.
    322     int symbol = VP8LReadBits(br, (first_symbol_len_code == 0) ? 1 : 8);
    323     code_lengths[symbol] = 1;
    324     // The second code (if present), is always 8 bit long.
    325     if (num_symbols == 2) {
    326       symbol = VP8LReadBits(br, 8);
    327       code_lengths[symbol] = 1;
    328     }
    329     ok = 1;
    330   } else {  // Decode Huffman-coded code lengths.
    331     int i;
    332     int code_length_code_lengths[NUM_CODE_LENGTH_CODES] = { 0 };
    333     const int num_codes = VP8LReadBits(br, 4) + 4;
    334     if (num_codes > NUM_CODE_LENGTH_CODES) {
    335       dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
    336       return 0;
    337     }
    338 
    339     for (i = 0; i < num_codes; ++i) {
    340       code_length_code_lengths[kCodeLengthCodeOrder[i]] = VP8LReadBits(br, 3);
    341     }
    342     ok = ReadHuffmanCodeLengths(dec, code_length_code_lengths, alphabet_size,
    343                                 code_lengths);
    344   }
    345 
    346   ok = ok && !br->eos_;
    347   if (ok) {
    348     size = VP8LBuildHuffmanTable(table, HUFFMAN_TABLE_BITS,
    349                                  code_lengths, alphabet_size);
    350   }
    351   if (!ok || size == 0) {
    352     dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
    353     return 0;
    354   }
    355   return size;
    356 }
    357 
    358 static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
    359                             int color_cache_bits, int allow_recursion) {
    360   int i, j;
    361   VP8LBitReader* const br = &dec->br_;
    362   VP8LMetadata* const hdr = &dec->hdr_;
    363   uint32_t* huffman_image = NULL;
    364   HTreeGroup* htree_groups = NULL;
    365   // When reading htrees, some might be unused, as the format allows it.
    366   // We will still read them but put them in this htree_group_bogus.
    367   HTreeGroup htree_group_bogus;
    368   HuffmanCode* huffman_tables = NULL;
    369   HuffmanCode* huffman_tables_bogus = NULL;
    370   HuffmanCode* next = NULL;
    371   int num_htree_groups = 1;
    372   int num_htree_groups_max = 1;
    373   int max_alphabet_size = 0;
    374   int* code_lengths = NULL;
    375   const int table_size = kTableSize[color_cache_bits];
    376   int* mapping = NULL;
    377   int ok = 0;
    378 
    379   if (allow_recursion && VP8LReadBits(br, 1)) {
    380     // use meta Huffman codes.
    381     const int huffman_precision = VP8LReadBits(br, 3) + 2;
    382     const int huffman_xsize = VP8LSubSampleSize(xsize, huffman_precision);
    383     const int huffman_ysize = VP8LSubSampleSize(ysize, huffman_precision);
    384     const int huffman_pixs = huffman_xsize * huffman_ysize;
    385     if (!DecodeImageStream(huffman_xsize, huffman_ysize, 0, dec,
    386                            &huffman_image)) {
    387       goto Error;
    388     }
    389     hdr->huffman_subsample_bits_ = huffman_precision;
    390     for (i = 0; i < huffman_pixs; ++i) {
    391       // The huffman data is stored in red and green bytes.
    392       const int group = (huffman_image[i] >> 8) & 0xffff;
    393       huffman_image[i] = group;
    394       if (group >= num_htree_groups_max) {
    395         num_htree_groups_max = group + 1;
    396       }
    397     }
    398     // Check the validity of num_htree_groups_max. If it seems too big, use a
    399     // smaller value for later. This will prevent big memory allocations to end
    400     // up with a bad bitstream anyway.
    401     // The value of 1000 is totally arbitrary. We know that num_htree_groups_max
    402     // is smaller than (1 << 16) and should be smaller than the number of pixels
    403     // (though the format allows it to be bigger).
    404     if (num_htree_groups_max > 1000 || num_htree_groups_max > xsize * ysize) {
    405       // Create a mapping from the used indices to the minimal set of used
    406       // values [0, num_htree_groups)
    407       mapping = (int*)WebPSafeMalloc(num_htree_groups_max, sizeof(*mapping));
    408       if (mapping == NULL) {
    409         dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
    410         goto Error;
    411       }
    412       // -1 means a value is unmapped, and therefore unused in the Huffman
    413       // image.
    414       memset(mapping, 0xff, num_htree_groups_max * sizeof(*mapping));
    415       for (num_htree_groups = 0, i = 0; i < huffman_pixs; ++i) {
    416         // Get the current mapping for the group and remap the Huffman image.
    417         int* const mapped_group = &mapping[huffman_image[i]];
    418         if (*mapped_group == -1) *mapped_group = num_htree_groups++;
    419         huffman_image[i] = *mapped_group;
    420       }
    421       huffman_tables_bogus = (HuffmanCode*)WebPSafeMalloc(
    422           table_size, sizeof(*huffman_tables_bogus));
    423       if (huffman_tables_bogus == NULL) {
    424         dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
    425         goto Error;
    426       }
    427     } else {
    428       num_htree_groups = num_htree_groups_max;
    429     }
    430   }
    431 
    432   if (br->eos_) goto Error;
    433 
    434   // Find maximum alphabet size for the htree group.
    435   for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
    436     int alphabet_size = kAlphabetSize[j];
    437     if (j == 0 && color_cache_bits > 0) {
    438       alphabet_size += 1 << color_cache_bits;
    439     }
    440     if (max_alphabet_size < alphabet_size) {
    441       max_alphabet_size = alphabet_size;
    442     }
    443   }
    444 
    445   code_lengths = (int*)WebPSafeCalloc((uint64_t)max_alphabet_size,
    446                                       sizeof(*code_lengths));
    447   huffman_tables = (HuffmanCode*)WebPSafeMalloc(num_htree_groups * table_size,
    448                                                 sizeof(*huffman_tables));
    449   htree_groups = VP8LHtreeGroupsNew(num_htree_groups);
    450 
    451   if (htree_groups == NULL || code_lengths == NULL || huffman_tables == NULL) {
    452     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
    453     goto Error;
    454   }
    455 
    456   next = huffman_tables;
    457   for (i = 0; i < num_htree_groups_max; ++i) {
    458     // If the index "i" is unused in the Huffman image, read the coefficients
    459     // but store them to a bogus htree_group.
    460     const int is_bogus = (mapping != NULL && mapping[i] == -1);
    461     HTreeGroup* const htree_group =
    462         is_bogus ? &htree_group_bogus :
    463         &htree_groups[(mapping == NULL) ? i : mapping[i]];
    464     HuffmanCode** const htrees = htree_group->htrees;
    465     HuffmanCode* huffman_tables_i = is_bogus ? huffman_tables_bogus : next;
    466     int size;
    467     int total_size = 0;
    468     int is_trivial_literal = 1;
    469     int max_bits = 0;
    470     for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
    471       int alphabet_size = kAlphabetSize[j];
    472       htrees[j] = huffman_tables_i;
    473       if (j == 0 && color_cache_bits > 0) {
    474         alphabet_size += 1 << color_cache_bits;
    475       }
    476       size =
    477           ReadHuffmanCode(alphabet_size, dec, code_lengths, huffman_tables_i);
    478       if (size == 0) {
    479         goto Error;
    480       }
    481       if (is_trivial_literal && kLiteralMap[j] == 1) {
    482         is_trivial_literal = (huffman_tables_i->bits == 0);
    483       }
    484       total_size += huffman_tables_i->bits;
    485       huffman_tables_i += size;
    486       if (j <= ALPHA) {
    487         int local_max_bits = code_lengths[0];
    488         int k;
    489         for (k = 1; k < alphabet_size; ++k) {
    490           if (code_lengths[k] > local_max_bits) {
    491             local_max_bits = code_lengths[k];
    492           }
    493         }
    494         max_bits += local_max_bits;
    495       }
    496     }
    497     if (!is_bogus) next = huffman_tables_i;
    498     htree_group->is_trivial_literal = is_trivial_literal;
    499     htree_group->is_trivial_code = 0;
    500     if (is_trivial_literal) {
    501       const int red = htrees[RED][0].value;
    502       const int blue = htrees[BLUE][0].value;
    503       const int alpha = htrees[ALPHA][0].value;
    504       htree_group->literal_arb = ((uint32_t)alpha << 24) | (red << 16) | blue;
    505       if (total_size == 0 && htrees[GREEN][0].value < NUM_LITERAL_CODES) {
    506         htree_group->is_trivial_code = 1;
    507         htree_group->literal_arb |= htrees[GREEN][0].value << 8;
    508       }
    509     }
    510     htree_group->use_packed_table =
    511         !htree_group->is_trivial_code && (max_bits < HUFFMAN_PACKED_BITS);
    512     if (htree_group->use_packed_table) BuildPackedTable(htree_group);
    513   }
    514   ok = 1;
    515 
    516   // All OK. Finalize pointers.
    517   hdr->huffman_image_ = huffman_image;
    518   hdr->num_htree_groups_ = num_htree_groups;
    519   hdr->htree_groups_ = htree_groups;
    520   hdr->huffman_tables_ = huffman_tables;
    521 
    522  Error:
    523   WebPSafeFree(code_lengths);
    524   WebPSafeFree(huffman_tables_bogus);
    525   WebPSafeFree(mapping);
    526   if (!ok) {
    527     WebPSafeFree(huffman_image);
    528     WebPSafeFree(huffman_tables);
    529     VP8LHtreeGroupsFree(htree_groups);
    530   }
    531   return ok;
    532 }
    533 
    534 //------------------------------------------------------------------------------
    535 // Scaling.
    536 
    537 #if !defined(WEBP_REDUCE_SIZE)
    538 static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
    539   const int num_channels = 4;
    540   const int in_width = io->mb_w;
    541   const int out_width = io->scaled_width;
    542   const int in_height = io->mb_h;
    543   const int out_height = io->scaled_height;
    544   const uint64_t work_size = 2 * num_channels * (uint64_t)out_width;
    545   rescaler_t* work;        // Rescaler work area.
    546   const uint64_t scaled_data_size = (uint64_t)out_width;
    547   uint32_t* scaled_data;  // Temporary storage for scaled BGRA data.
    548   const uint64_t memory_size = sizeof(*dec->rescaler) +
    549                                work_size * sizeof(*work) +
    550                                scaled_data_size * sizeof(*scaled_data);
    551   uint8_t* memory = (uint8_t*)WebPSafeMalloc(memory_size, sizeof(*memory));
    552   if (memory == NULL) {
    553     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
    554     return 0;
    555   }
    556   assert(dec->rescaler_memory == NULL);
    557   dec->rescaler_memory = memory;
    558 
    559   dec->rescaler = (WebPRescaler*)memory;
    560   memory += sizeof(*dec->rescaler);
    561   work = (rescaler_t*)memory;
    562   memory += work_size * sizeof(*work);
    563   scaled_data = (uint32_t*)memory;
    564 
    565   WebPRescalerInit(dec->rescaler, in_width, in_height, (uint8_t*)scaled_data,
    566                    out_width, out_height, 0, num_channels, work);
    567   return 1;
    568 }
    569 #endif   // WEBP_REDUCE_SIZE
    570 
    571 //------------------------------------------------------------------------------
    572 // Export to ARGB
    573 
    574 #if !defined(WEBP_REDUCE_SIZE)
    575 
    576 // We have special "export" function since we need to convert from BGRA
    577 static int Export(WebPRescaler* const rescaler, WEBP_CSP_MODE colorspace,
    578                   int rgba_stride, uint8_t* const rgba) {
    579   uint32_t* const src = (uint32_t*)rescaler->dst;
    580   const int dst_width = rescaler->dst_width;
    581   int num_lines_out = 0;
    582   while (WebPRescalerHasPendingOutput(rescaler)) {
    583     uint8_t* const dst = rgba + num_lines_out * rgba_stride;
    584     WebPRescalerExportRow(rescaler);
    585     WebPMultARGBRow(src, dst_width, 1);
    586     VP8LConvertFromBGRA(src, dst_width, colorspace, dst);
    587     ++num_lines_out;
    588   }
    589   return num_lines_out;
    590 }
    591 
    592 // Emit scaled rows.
    593 static int EmitRescaledRowsRGBA(const VP8LDecoder* const dec,
    594                                 uint8_t* in, int in_stride, int mb_h,
    595                                 uint8_t* const out, int out_stride) {
    596   const WEBP_CSP_MODE colorspace = dec->output_->colorspace;
    597   int num_lines_in = 0;
    598   int num_lines_out = 0;
    599   while (num_lines_in < mb_h) {
    600     uint8_t* const row_in = in + num_lines_in * in_stride;
    601     uint8_t* const row_out = out + num_lines_out * out_stride;
    602     const int lines_left = mb_h - num_lines_in;
    603     const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
    604     int lines_imported;
    605     assert(needed_lines > 0 && needed_lines <= lines_left);
    606     WebPMultARGBRows(row_in, in_stride,
    607                      dec->rescaler->src_width, needed_lines, 0);
    608     lines_imported =
    609         WebPRescalerImport(dec->rescaler, lines_left, row_in, in_stride);
    610     assert(lines_imported == needed_lines);
    611     num_lines_in += lines_imported;
    612     num_lines_out += Export(dec->rescaler, colorspace, out_stride, row_out);
    613   }
    614   return num_lines_out;
    615 }
    616 
    617 #endif   // WEBP_REDUCE_SIZE
    618 
    619 // Emit rows without any scaling.
    620 static int EmitRows(WEBP_CSP_MODE colorspace,
    621                     const uint8_t* row_in, int in_stride,
    622                     int mb_w, int mb_h,
    623                     uint8_t* const out, int out_stride) {
    624   int lines = mb_h;
    625   uint8_t* row_out = out;
    626   while (lines-- > 0) {
    627     VP8LConvertFromBGRA((const uint32_t*)row_in, mb_w, colorspace, row_out);
    628     row_in += in_stride;
    629     row_out += out_stride;
    630   }
    631   return mb_h;  // Num rows out == num rows in.
    632 }
    633 
    634 //------------------------------------------------------------------------------
    635 // Export to YUVA
    636 
    637 static void ConvertToYUVA(const uint32_t* const src, int width, int y_pos,
    638                           const WebPDecBuffer* const output) {
    639   const WebPYUVABuffer* const buf = &output->u.YUVA;
    640 
    641   // first, the luma plane
    642   WebPConvertARGBToY(src, buf->y + y_pos * buf->y_stride, width);
    643 
    644   // then U/V planes
    645   {
    646     uint8_t* const u = buf->u + (y_pos >> 1) * buf->u_stride;
    647     uint8_t* const v = buf->v + (y_pos >> 1) * buf->v_stride;
    648     // even lines: store values
    649     // odd lines: average with previous values
    650     WebPConvertARGBToUV(src, u, v, width, !(y_pos & 1));
    651   }
    652   // Lastly, store alpha if needed.
    653   if (buf->a != NULL) {
    654     uint8_t* const a = buf->a + y_pos * buf->a_stride;
    655 #if defined(WORDS_BIGENDIAN)
    656     WebPExtractAlpha((uint8_t*)src + 0, 0, width, 1, a, 0);
    657 #else
    658     WebPExtractAlpha((uint8_t*)src + 3, 0, width, 1, a, 0);
    659 #endif
    660   }
    661 }
    662 
    663 static int ExportYUVA(const VP8LDecoder* const dec, int y_pos) {
    664   WebPRescaler* const rescaler = dec->rescaler;
    665   uint32_t* const src = (uint32_t*)rescaler->dst;
    666   const int dst_width = rescaler->dst_width;
    667   int num_lines_out = 0;
    668   while (WebPRescalerHasPendingOutput(rescaler)) {
    669     WebPRescalerExportRow(rescaler);
    670     WebPMultARGBRow(src, dst_width, 1);
    671     ConvertToYUVA(src, dst_width, y_pos, dec->output_);
    672     ++y_pos;
    673     ++num_lines_out;
    674   }
    675   return num_lines_out;
    676 }
    677 
    678 static int EmitRescaledRowsYUVA(const VP8LDecoder* const dec,
    679                                 uint8_t* in, int in_stride, int mb_h) {
    680   int num_lines_in = 0;
    681   int y_pos = dec->last_out_row_;
    682   while (num_lines_in < mb_h) {
    683     const int lines_left = mb_h - num_lines_in;
    684     const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
    685     int lines_imported;
    686     WebPMultARGBRows(in, in_stride, dec->rescaler->src_width, needed_lines, 0);
    687     lines_imported =
    688         WebPRescalerImport(dec->rescaler, lines_left, in, in_stride);
    689     assert(lines_imported == needed_lines);
    690     num_lines_in += lines_imported;
    691     in += needed_lines * in_stride;
    692     y_pos += ExportYUVA(dec, y_pos);
    693   }
    694   return y_pos;
    695 }
    696 
    697 static int EmitRowsYUVA(const VP8LDecoder* const dec,
    698                         const uint8_t* in, int in_stride,
    699                         int mb_w, int num_rows) {
    700   int y_pos = dec->last_out_row_;
    701   while (num_rows-- > 0) {
    702     ConvertToYUVA((const uint32_t*)in, mb_w, y_pos, dec->output_);
    703     in += in_stride;
    704     ++y_pos;
    705   }
    706   return y_pos;
    707 }
    708 
    709 //------------------------------------------------------------------------------
    710 // Cropping.
    711 
    712 // Sets io->mb_y, io->mb_h & io->mb_w according to start row, end row and
    713 // crop options. Also updates the input data pointer, so that it points to the
    714 // start of the cropped window. Note that pixels are in ARGB format even if
    715 // 'in_data' is uint8_t*.
    716 // Returns true if the crop window is not empty.
    717 static int SetCropWindow(VP8Io* const io, int y_start, int y_end,
    718                          uint8_t** const in_data, int pixel_stride) {
    719   assert(y_start < y_end);
    720   assert(io->crop_left < io->crop_right);
    721   if (y_end > io->crop_bottom) {
    722     y_end = io->crop_bottom;  // make sure we don't overflow on last row.
    723   }
    724   if (y_start < io->crop_top) {
    725     const int delta = io->crop_top - y_start;
    726     y_start = io->crop_top;
    727     *in_data += delta * pixel_stride;
    728   }
    729   if (y_start >= y_end) return 0;  // Crop window is empty.
    730 
    731   *in_data += io->crop_left * sizeof(uint32_t);
    732 
    733   io->mb_y = y_start - io->crop_top;
    734   io->mb_w = io->crop_right - io->crop_left;
    735   io->mb_h = y_end - y_start;
    736   return 1;  // Non-empty crop window.
    737 }
    738 
    739 //------------------------------------------------------------------------------
    740 
    741 static WEBP_INLINE int GetMetaIndex(
    742     const uint32_t* const image, int xsize, int bits, int x, int y) {
    743   if (bits == 0) return 0;
    744   return image[xsize * (y >> bits) + (x >> bits)];
    745 }
    746 
    747 static WEBP_INLINE HTreeGroup* GetHtreeGroupForPos(VP8LMetadata* const hdr,
    748                                                    int x, int y) {
    749   const int meta_index = GetMetaIndex(hdr->huffman_image_, hdr->huffman_xsize_,
    750                                       hdr->huffman_subsample_bits_, x, y);
    751   assert(meta_index < hdr->num_htree_groups_);
    752   return hdr->htree_groups_ + meta_index;
    753 }
    754 
    755 //------------------------------------------------------------------------------
    756 // Main loop, with custom row-processing function
    757 
    758 typedef void (*ProcessRowsFunc)(VP8LDecoder* const dec, int row);
    759 
    760 static void ApplyInverseTransforms(VP8LDecoder* const dec, int num_rows,
    761                                    const uint32_t* const rows) {
    762   int n = dec->next_transform_;
    763   const int cache_pixs = dec->width_ * num_rows;
    764   const int start_row = dec->last_row_;
    765   const int end_row = start_row + num_rows;
    766   const uint32_t* rows_in = rows;
    767   uint32_t* const rows_out = dec->argb_cache_;
    768 
    769   // Inverse transforms.
    770   while (n-- > 0) {
    771     VP8LTransform* const transform = &dec->transforms_[n];
    772     VP8LInverseTransform(transform, start_row, end_row, rows_in, rows_out);
    773     rows_in = rows_out;
    774   }
    775   if (rows_in != rows_out) {
    776     // No transform called, hence just copy.
    777     memcpy(rows_out, rows_in, cache_pixs * sizeof(*rows_out));
    778   }
    779 }
    780 
    781 // Processes (transforms, scales & color-converts) the rows decoded after the
    782 // last call.
    783 static void ProcessRows(VP8LDecoder* const dec, int row) {
    784   const uint32_t* const rows = dec->pixels_ + dec->width_ * dec->last_row_;
    785   const int num_rows = row - dec->last_row_;
    786 
    787   assert(row <= dec->io_->crop_bottom);
    788   // We can't process more than NUM_ARGB_CACHE_ROWS at a time (that's the size
    789   // of argb_cache_), but we currently don't need more than that.
    790   assert(num_rows <= NUM_ARGB_CACHE_ROWS);
    791   if (num_rows > 0) {    // Emit output.
    792     VP8Io* const io = dec->io_;
    793     uint8_t* rows_data = (uint8_t*)dec->argb_cache_;
    794     const int in_stride = io->width * sizeof(uint32_t);  // in unit of RGBA
    795 
    796     ApplyInverseTransforms(dec, num_rows, rows);
    797     if (!SetCropWindow(io, dec->last_row_, row, &rows_data, in_stride)) {
    798       // Nothing to output (this time).
    799     } else {
    800       const WebPDecBuffer* const output = dec->output_;
    801       if (WebPIsRGBMode(output->colorspace)) {  // convert to RGBA
    802         const WebPRGBABuffer* const buf = &output->u.RGBA;
    803         uint8_t* const rgba = buf->rgba + dec->last_out_row_ * buf->stride;
    804         const int num_rows_out =
    805 #if !defined(WEBP_REDUCE_SIZE)
    806          io->use_scaling ?
    807             EmitRescaledRowsRGBA(dec, rows_data, in_stride, io->mb_h,
    808                                  rgba, buf->stride) :
    809 #endif  // WEBP_REDUCE_SIZE
    810             EmitRows(output->colorspace, rows_data, in_stride,
    811                      io->mb_w, io->mb_h, rgba, buf->stride);
    812         // Update 'last_out_row_'.
    813         dec->last_out_row_ += num_rows_out;
    814       } else {                              // convert to YUVA
    815         dec->last_out_row_ = io->use_scaling ?
    816             EmitRescaledRowsYUVA(dec, rows_data, in_stride, io->mb_h) :
    817             EmitRowsYUVA(dec, rows_data, in_stride, io->mb_w, io->mb_h);
    818       }
    819       assert(dec->last_out_row_ <= output->height);
    820     }
    821   }
    822 
    823   // Update 'last_row_'.
    824   dec->last_row_ = row;
    825   assert(dec->last_row_ <= dec->height_);
    826 }
    827 
    828 // Row-processing for the special case when alpha data contains only one
    829 // transform (color indexing), and trivial non-green literals.
    830 static int Is8bOptimizable(const VP8LMetadata* const hdr) {
    831   int i;
    832   if (hdr->color_cache_size_ > 0) return 0;
    833   // When the Huffman tree contains only one symbol, we can skip the
    834   // call to ReadSymbol() for red/blue/alpha channels.
    835   for (i = 0; i < hdr->num_htree_groups_; ++i) {
    836     HuffmanCode** const htrees = hdr->htree_groups_[i].htrees;
    837     if (htrees[RED][0].bits > 0) return 0;
    838     if (htrees[BLUE][0].bits > 0) return 0;
    839     if (htrees[ALPHA][0].bits > 0) return 0;
    840   }
    841   return 1;
    842 }
    843 
    844 static void AlphaApplyFilter(ALPHDecoder* const alph_dec,
    845                              int first_row, int last_row,
    846                              uint8_t* out, int stride) {
    847   if (alph_dec->filter_ != WEBP_FILTER_NONE) {
    848     int y;
    849     const uint8_t* prev_line = alph_dec->prev_line_;
    850     assert(WebPUnfilters[alph_dec->filter_] != NULL);
    851     for (y = first_row; y < last_row; ++y) {
    852       WebPUnfilters[alph_dec->filter_](prev_line, out, out, stride);
    853       prev_line = out;
    854       out += stride;
    855     }
    856     alph_dec->prev_line_ = prev_line;
    857   }
    858 }
    859 
    860 static void ExtractPalettedAlphaRows(VP8LDecoder* const dec, int last_row) {
    861   // For vertical and gradient filtering, we need to decode the part above the
    862   // crop_top row, in order to have the correct spatial predictors.
    863   ALPHDecoder* const alph_dec = (ALPHDecoder*)dec->io_->opaque;
    864   const int top_row =
    865       (alph_dec->filter_ == WEBP_FILTER_NONE ||
    866        alph_dec->filter_ == WEBP_FILTER_HORIZONTAL) ? dec->io_->crop_top
    867                                                     : dec->last_row_;
    868   const int first_row = (dec->last_row_ < top_row) ? top_row : dec->last_row_;
    869   assert(last_row <= dec->io_->crop_bottom);
    870   if (last_row > first_row) {
    871     // Special method for paletted alpha data. We only process the cropped area.
    872     const int width = dec->io_->width;
    873     uint8_t* out = alph_dec->output_ + width * first_row;
    874     const uint8_t* const in =
    875       (uint8_t*)dec->pixels_ + dec->width_ * first_row;
    876     VP8LTransform* const transform = &dec->transforms_[0];
    877     assert(dec->next_transform_ == 1);
    878     assert(transform->type_ == COLOR_INDEXING_TRANSFORM);
    879     VP8LColorIndexInverseTransformAlpha(transform, first_row, last_row,
    880                                         in, out);
    881     AlphaApplyFilter(alph_dec, first_row, last_row, out, width);
    882   }
    883   dec->last_row_ = dec->last_out_row_ = last_row;
    884 }
    885 
    886 //------------------------------------------------------------------------------
    887 // Helper functions for fast pattern copy (8b and 32b)
    888 
    889 // cyclic rotation of pattern word
    890 static WEBP_INLINE uint32_t Rotate8b(uint32_t V) {
    891 #if defined(WORDS_BIGENDIAN)
    892   return ((V & 0xff000000u) >> 24) | (V << 8);
    893 #else
    894   return ((V & 0xffu) << 24) | (V >> 8);
    895 #endif
    896 }
    897 
    898 // copy 1, 2 or 4-bytes pattern
    899 static WEBP_INLINE void CopySmallPattern8b(const uint8_t* src, uint8_t* dst,
    900                                            int length, uint32_t pattern) {
    901   int i;
    902   // align 'dst' to 4-bytes boundary. Adjust the pattern along the way.
    903   while ((uintptr_t)dst & 3) {
    904     *dst++ = *src++;
    905     pattern = Rotate8b(pattern);
    906     --length;
    907   }
    908   // Copy the pattern 4 bytes at a time.
    909   for (i = 0; i < (length >> 2); ++i) {
    910     ((uint32_t*)dst)[i] = pattern;
    911   }
    912   // Finish with left-overs. 'pattern' is still correctly positioned,
    913   // so no Rotate8b() call is needed.
    914   for (i <<= 2; i < length; ++i) {
    915     dst[i] = src[i];
    916   }
    917 }
    918 
    919 static WEBP_INLINE void CopyBlock8b(uint8_t* const dst, int dist, int length) {
    920   const uint8_t* src = dst - dist;
    921   if (length >= 8) {
    922     uint32_t pattern = 0;
    923     switch (dist) {
    924       case 1:
    925         pattern = src[0];
    926 #if defined(__arm__) || defined(_M_ARM)   // arm doesn't like multiply that much
    927         pattern |= pattern << 8;
    928         pattern |= pattern << 16;
    929 #elif defined(WEBP_USE_MIPS_DSP_R2)
    930         __asm__ volatile ("replv.qb %0, %0" : "+r"(pattern));
    931 #else
    932         pattern = 0x01010101u * pattern;
    933 #endif
    934         break;
    935       case 2:
    936 #if !defined(WORDS_BIGENDIAN)
    937         memcpy(&pattern, src, sizeof(uint16_t));
    938 #else
    939         pattern = ((uint32_t)src[0] << 8) | src[1];
    940 #endif
    941 #if defined(__arm__) || defined(_M_ARM)
    942         pattern |= pattern << 16;
    943 #elif defined(WEBP_USE_MIPS_DSP_R2)
    944         __asm__ volatile ("replv.ph %0, %0" : "+r"(pattern));
    945 #else
    946         pattern = 0x00010001u * pattern;
    947 #endif
    948         break;
    949       case 4:
    950         memcpy(&pattern, src, sizeof(uint32_t));
    951         break;
    952       default:
    953         goto Copy;
    954         break;
    955     }
    956     CopySmallPattern8b(src, dst, length, pattern);
    957     return;
    958   }
    959  Copy:
    960   if (dist >= length) {  // no overlap -> use memcpy()
    961     memcpy(dst, src, length * sizeof(*dst));
    962   } else {
    963     int i;
    964     for (i = 0; i < length; ++i) dst[i] = src[i];
    965   }
    966 }
    967 
    968 // copy pattern of 1 or 2 uint32_t's
    969 static WEBP_INLINE void CopySmallPattern32b(const uint32_t* src,
    970                                             uint32_t* dst,
    971                                             int length, uint64_t pattern) {
    972   int i;
    973   if ((uintptr_t)dst & 4) {           // Align 'dst' to 8-bytes boundary.
    974     *dst++ = *src++;
    975     pattern = (pattern >> 32) | (pattern << 32);
    976     --length;
    977   }
    978   assert(0 == ((uintptr_t)dst & 7));
    979   for (i = 0; i < (length >> 1); ++i) {
    980     ((uint64_t*)dst)[i] = pattern;    // Copy the pattern 8 bytes at a time.
    981   }
    982   if (length & 1) {                   // Finish with left-over.
    983     dst[i << 1] = src[i << 1];
    984   }
    985 }
    986 
    987 static WEBP_INLINE void CopyBlock32b(uint32_t* const dst,
    988                                      int dist, int length) {
    989   const uint32_t* const src = dst - dist;
    990   if (dist <= 2 && length >= 4 && ((uintptr_t)dst & 3) == 0) {
    991     uint64_t pattern;
    992     if (dist == 1) {
    993       pattern = (uint64_t)src[0];
    994       pattern |= pattern << 32;
    995     } else {
    996       memcpy(&pattern, src, sizeof(pattern));
    997     }
    998     CopySmallPattern32b(src, dst, length, pattern);
    999   } else if (dist >= length) {  // no overlap
   1000     memcpy(dst, src, length * sizeof(*dst));
   1001   } else {
   1002     int i;
   1003     for (i = 0; i < length; ++i) dst[i] = src[i];
   1004   }
   1005 }
   1006 
   1007 //------------------------------------------------------------------------------
   1008 
   1009 static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
   1010                            int width, int height, int last_row) {
   1011   int ok = 1;
   1012   int row = dec->last_pixel_ / width;
   1013   int col = dec->last_pixel_ % width;
   1014   VP8LBitReader* const br = &dec->br_;
   1015   VP8LMetadata* const hdr = &dec->hdr_;
   1016   int pos = dec->last_pixel_;         // current position
   1017   const int end = width * height;     // End of data
   1018   const int last = width * last_row;  // Last pixel to decode
   1019   const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
   1020   const int mask = hdr->huffman_mask_;
   1021   const HTreeGroup* htree_group =
   1022       (pos < last) ? GetHtreeGroupForPos(hdr, col, row) : NULL;
   1023   assert(pos <= end);
   1024   assert(last_row <= height);
   1025   assert(Is8bOptimizable(hdr));
   1026 
   1027   while (!br->eos_ && pos < last) {
   1028     int code;
   1029     // Only update when changing tile.
   1030     if ((col & mask) == 0) {
   1031       htree_group = GetHtreeGroupForPos(hdr, col, row);
   1032     }
   1033     assert(htree_group != NULL);
   1034     VP8LFillBitWindow(br);
   1035     code = ReadSymbol(htree_group->htrees[GREEN], br);
   1036     if (code < NUM_LITERAL_CODES) {  // Literal
   1037       data[pos] = code;
   1038       ++pos;
   1039       ++col;
   1040       if (col >= width) {
   1041         col = 0;
   1042         ++row;
   1043         if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1044           ExtractPalettedAlphaRows(dec, row);
   1045         }
   1046       }
   1047     } else if (code < len_code_limit) {  // Backward reference
   1048       int dist_code, dist;
   1049       const int length_sym = code - NUM_LITERAL_CODES;
   1050       const int length = GetCopyLength(length_sym, br);
   1051       const int dist_symbol = ReadSymbol(htree_group->htrees[DIST], br);
   1052       VP8LFillBitWindow(br);
   1053       dist_code = GetCopyDistance(dist_symbol, br);
   1054       dist = PlaneCodeToDistance(width, dist_code);
   1055       if (pos >= dist && end - pos >= length) {
   1056         CopyBlock8b(data + pos, dist, length);
   1057       } else {
   1058         ok = 0;
   1059         goto End;
   1060       }
   1061       pos += length;
   1062       col += length;
   1063       while (col >= width) {
   1064         col -= width;
   1065         ++row;
   1066         if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1067           ExtractPalettedAlphaRows(dec, row);
   1068         }
   1069       }
   1070       if (pos < last && (col & mask)) {
   1071         htree_group = GetHtreeGroupForPos(hdr, col, row);
   1072       }
   1073     } else {  // Not reached
   1074       ok = 0;
   1075       goto End;
   1076     }
   1077     br->eos_ = VP8LIsEndOfStream(br);
   1078   }
   1079   // Process the remaining rows corresponding to last row-block.
   1080   ExtractPalettedAlphaRows(dec, row > last_row ? last_row : row);
   1081 
   1082  End:
   1083   br->eos_ = VP8LIsEndOfStream(br);
   1084   if (!ok || (br->eos_ && pos < end)) {
   1085     ok = 0;
   1086     dec->status_ = br->eos_ ? VP8_STATUS_SUSPENDED
   1087                             : VP8_STATUS_BITSTREAM_ERROR;
   1088   } else {
   1089     dec->last_pixel_ = pos;
   1090   }
   1091   return ok;
   1092 }
   1093 
   1094 static void SaveState(VP8LDecoder* const dec, int last_pixel) {
   1095   assert(dec->incremental_);
   1096   dec->saved_br_ = dec->br_;
   1097   dec->saved_last_pixel_ = last_pixel;
   1098   if (dec->hdr_.color_cache_size_ > 0) {
   1099     VP8LColorCacheCopy(&dec->hdr_.color_cache_, &dec->hdr_.saved_color_cache_);
   1100   }
   1101 }
   1102 
   1103 static void RestoreState(VP8LDecoder* const dec) {
   1104   assert(dec->br_.eos_);
   1105   dec->status_ = VP8_STATUS_SUSPENDED;
   1106   dec->br_ = dec->saved_br_;
   1107   dec->last_pixel_ = dec->saved_last_pixel_;
   1108   if (dec->hdr_.color_cache_size_ > 0) {
   1109     VP8LColorCacheCopy(&dec->hdr_.saved_color_cache_, &dec->hdr_.color_cache_);
   1110   }
   1111 }
   1112 
   1113 #define SYNC_EVERY_N_ROWS 8  // minimum number of rows between check-points
   1114 static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
   1115                            int width, int height, int last_row,
   1116                            ProcessRowsFunc process_func) {
   1117   int row = dec->last_pixel_ / width;
   1118   int col = dec->last_pixel_ % width;
   1119   VP8LBitReader* const br = &dec->br_;
   1120   VP8LMetadata* const hdr = &dec->hdr_;
   1121   uint32_t* src = data + dec->last_pixel_;
   1122   uint32_t* last_cached = src;
   1123   uint32_t* const src_end = data + width * height;     // End of data
   1124   uint32_t* const src_last = data + width * last_row;  // Last pixel to decode
   1125   const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
   1126   const int color_cache_limit = len_code_limit + hdr->color_cache_size_;
   1127   int next_sync_row = dec->incremental_ ? row : 1 << 24;
   1128   VP8LColorCache* const color_cache =
   1129       (hdr->color_cache_size_ > 0) ? &hdr->color_cache_ : NULL;
   1130   const int mask = hdr->huffman_mask_;
   1131   const HTreeGroup* htree_group =
   1132       (src < src_last) ? GetHtreeGroupForPos(hdr, col, row) : NULL;
   1133   assert(dec->last_row_ < last_row);
   1134   assert(src_last <= src_end);
   1135 
   1136   while (src < src_last) {
   1137     int code;
   1138     if (row >= next_sync_row) {
   1139       SaveState(dec, (int)(src - data));
   1140       next_sync_row = row + SYNC_EVERY_N_ROWS;
   1141     }
   1142     // Only update when changing tile. Note we could use this test:
   1143     // if "((((prev_col ^ col) | prev_row ^ row)) > mask)" -> tile changed
   1144     // but that's actually slower and needs storing the previous col/row.
   1145     if ((col & mask) == 0) {
   1146       htree_group = GetHtreeGroupForPos(hdr, col, row);
   1147     }
   1148     assert(htree_group != NULL);
   1149     if (htree_group->is_trivial_code) {
   1150       *src = htree_group->literal_arb;
   1151       goto AdvanceByOne;
   1152     }
   1153     VP8LFillBitWindow(br);
   1154     if (htree_group->use_packed_table) {
   1155       code = ReadPackedSymbols(htree_group, br, src);
   1156       if (VP8LIsEndOfStream(br)) break;
   1157       if (code == PACKED_NON_LITERAL_CODE) goto AdvanceByOne;
   1158     } else {
   1159       code = ReadSymbol(htree_group->htrees[GREEN], br);
   1160     }
   1161     if (VP8LIsEndOfStream(br)) break;
   1162     if (code < NUM_LITERAL_CODES) {  // Literal
   1163       if (htree_group->is_trivial_literal) {
   1164         *src = htree_group->literal_arb | (code << 8);
   1165       } else {
   1166         int red, blue, alpha;
   1167         red = ReadSymbol(htree_group->htrees[RED], br);
   1168         VP8LFillBitWindow(br);
   1169         blue = ReadSymbol(htree_group->htrees[BLUE], br);
   1170         alpha = ReadSymbol(htree_group->htrees[ALPHA], br);
   1171         if (VP8LIsEndOfStream(br)) break;
   1172         *src = ((uint32_t)alpha << 24) | (red << 16) | (code << 8) | blue;
   1173       }
   1174     AdvanceByOne:
   1175       ++src;
   1176       ++col;
   1177       if (col >= width) {
   1178         col = 0;
   1179         ++row;
   1180         if (process_func != NULL) {
   1181           if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1182             process_func(dec, row);
   1183           }
   1184         }
   1185         if (color_cache != NULL) {
   1186           while (last_cached < src) {
   1187             VP8LColorCacheInsert(color_cache, *last_cached++);
   1188           }
   1189         }
   1190       }
   1191     } else if (code < len_code_limit) {  // Backward reference
   1192       int dist_code, dist;
   1193       const int length_sym = code - NUM_LITERAL_CODES;
   1194       const int length = GetCopyLength(length_sym, br);
   1195       const int dist_symbol = ReadSymbol(htree_group->htrees[DIST], br);
   1196       VP8LFillBitWindow(br);
   1197       dist_code = GetCopyDistance(dist_symbol, br);
   1198       dist = PlaneCodeToDistance(width, dist_code);
   1199       if (VP8LIsEndOfStream(br)) break;
   1200       if (src - data < (ptrdiff_t)dist || src_end - src < (ptrdiff_t)length) {
   1201         goto Error;
   1202       } else {
   1203         CopyBlock32b(src, dist, length);
   1204       }
   1205       src += length;
   1206       col += length;
   1207       while (col >= width) {
   1208         col -= width;
   1209         ++row;
   1210         if (process_func != NULL) {
   1211           if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1212             process_func(dec, row);
   1213           }
   1214         }
   1215       }
   1216       // Because of the check done above (before 'src' was incremented by
   1217       // 'length'), the following holds true.
   1218       assert(src <= src_end);
   1219       if (col & mask) htree_group = GetHtreeGroupForPos(hdr, col, row);
   1220       if (color_cache != NULL) {
   1221         while (last_cached < src) {
   1222           VP8LColorCacheInsert(color_cache, *last_cached++);
   1223         }
   1224       }
   1225     } else if (code < color_cache_limit) {  // Color cache
   1226       const int key = code - len_code_limit;
   1227       assert(color_cache != NULL);
   1228       while (last_cached < src) {
   1229         VP8LColorCacheInsert(color_cache, *last_cached++);
   1230       }
   1231       *src = VP8LColorCacheLookup(color_cache, key);
   1232       goto AdvanceByOne;
   1233     } else {  // Not reached
   1234       goto Error;
   1235     }
   1236   }
   1237 
   1238   br->eos_ = VP8LIsEndOfStream(br);
   1239   if (dec->incremental_ && br->eos_ && src < src_end) {
   1240     RestoreState(dec);
   1241   } else if (!br->eos_) {
   1242     // Process the remaining rows corresponding to last row-block.
   1243     if (process_func != NULL) {
   1244       process_func(dec, row > last_row ? last_row : row);
   1245     }
   1246     dec->status_ = VP8_STATUS_OK;
   1247     dec->last_pixel_ = (int)(src - data);  // end-of-scan marker
   1248   } else {
   1249     // if not incremental, and we are past the end of buffer (eos_=1), then this
   1250     // is a real bitstream error.
   1251     goto Error;
   1252   }
   1253   return 1;
   1254 
   1255  Error:
   1256   dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
   1257   return 0;
   1258 }
   1259 
   1260 // -----------------------------------------------------------------------------
   1261 // VP8LTransform
   1262 
   1263 static void ClearTransform(VP8LTransform* const transform) {
   1264   WebPSafeFree(transform->data_);
   1265   transform->data_ = NULL;
   1266 }
   1267 
   1268 // For security reason, we need to remap the color map to span
   1269 // the total possible bundled values, and not just the num_colors.
   1270 static int ExpandColorMap(int num_colors, VP8LTransform* const transform) {
   1271   int i;
   1272   const int final_num_colors = 1 << (8 >> transform->bits_);
   1273   uint32_t* const new_color_map =
   1274       (uint32_t*)WebPSafeMalloc((uint64_t)final_num_colors,
   1275                                 sizeof(*new_color_map));
   1276   if (new_color_map == NULL) {
   1277     return 0;
   1278   } else {
   1279     uint8_t* const data = (uint8_t*)transform->data_;
   1280     uint8_t* const new_data = (uint8_t*)new_color_map;
   1281     new_color_map[0] = transform->data_[0];
   1282     for (i = 4; i < 4 * num_colors; ++i) {
   1283       // Equivalent to AddPixelEq(), on a byte-basis.
   1284       new_data[i] = (data[i] + new_data[i - 4]) & 0xff;
   1285     }
   1286     for (; i < 4 * final_num_colors; ++i) {
   1287       new_data[i] = 0;  // black tail.
   1288     }
   1289     WebPSafeFree(transform->data_);
   1290     transform->data_ = new_color_map;
   1291   }
   1292   return 1;
   1293 }
   1294 
   1295 static int ReadTransform(int* const xsize, int const* ysize,
   1296                          VP8LDecoder* const dec) {
   1297   int ok = 1;
   1298   VP8LBitReader* const br = &dec->br_;
   1299   VP8LTransform* transform = &dec->transforms_[dec->next_transform_];
   1300   const VP8LImageTransformType type =
   1301       (VP8LImageTransformType)VP8LReadBits(br, 2);
   1302 
   1303   // Each transform type can only be present once in the stream.
   1304   if (dec->transforms_seen_ & (1U << type)) {
   1305     return 0;  // Already there, let's not accept the second same transform.
   1306   }
   1307   dec->transforms_seen_ |= (1U << type);
   1308 
   1309   transform->type_ = type;
   1310   transform->xsize_ = *xsize;
   1311   transform->ysize_ = *ysize;
   1312   transform->data_ = NULL;
   1313   ++dec->next_transform_;
   1314   assert(dec->next_transform_ <= NUM_TRANSFORMS);
   1315 
   1316   switch (type) {
   1317     case PREDICTOR_TRANSFORM:
   1318     case CROSS_COLOR_TRANSFORM:
   1319       transform->bits_ = VP8LReadBits(br, 3) + 2;
   1320       ok = DecodeImageStream(VP8LSubSampleSize(transform->xsize_,
   1321                                                transform->bits_),
   1322                              VP8LSubSampleSize(transform->ysize_,
   1323                                                transform->bits_),
   1324                              0, dec, &transform->data_);
   1325       break;
   1326     case COLOR_INDEXING_TRANSFORM: {
   1327        const int num_colors = VP8LReadBits(br, 8) + 1;
   1328        const int bits = (num_colors > 16) ? 0
   1329                       : (num_colors > 4) ? 1
   1330                       : (num_colors > 2) ? 2
   1331                       : 3;
   1332        *xsize = VP8LSubSampleSize(transform->xsize_, bits);
   1333        transform->bits_ = bits;
   1334        ok = DecodeImageStream(num_colors, 1, 0, dec, &transform->data_);
   1335        ok = ok && ExpandColorMap(num_colors, transform);
   1336       break;
   1337     }
   1338     case SUBTRACT_GREEN:
   1339       break;
   1340     default:
   1341       assert(0);    // can't happen
   1342       break;
   1343   }
   1344 
   1345   return ok;
   1346 }
   1347 
   1348 // -----------------------------------------------------------------------------
   1349 // VP8LMetadata
   1350 
   1351 static void InitMetadata(VP8LMetadata* const hdr) {
   1352   assert(hdr != NULL);
   1353   memset(hdr, 0, sizeof(*hdr));
   1354 }
   1355 
   1356 static void ClearMetadata(VP8LMetadata* const hdr) {
   1357   assert(hdr != NULL);
   1358 
   1359   WebPSafeFree(hdr->huffman_image_);
   1360   WebPSafeFree(hdr->huffman_tables_);
   1361   VP8LHtreeGroupsFree(hdr->htree_groups_);
   1362   VP8LColorCacheClear(&hdr->color_cache_);
   1363   VP8LColorCacheClear(&hdr->saved_color_cache_);
   1364   InitMetadata(hdr);
   1365 }
   1366 
   1367 // -----------------------------------------------------------------------------
   1368 // VP8LDecoder
   1369 
   1370 VP8LDecoder* VP8LNew(void) {
   1371   VP8LDecoder* const dec = (VP8LDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
   1372   if (dec == NULL) return NULL;
   1373   dec->status_ = VP8_STATUS_OK;
   1374   dec->state_ = READ_DIM;
   1375 
   1376   VP8LDspInit();  // Init critical function pointers.
   1377 
   1378   return dec;
   1379 }
   1380 
   1381 void VP8LClear(VP8LDecoder* const dec) {
   1382   int i;
   1383   if (dec == NULL) return;
   1384   ClearMetadata(&dec->hdr_);
   1385 
   1386   WebPSafeFree(dec->pixels_);
   1387   dec->pixels_ = NULL;
   1388   for (i = 0; i < dec->next_transform_; ++i) {
   1389     ClearTransform(&dec->transforms_[i]);
   1390   }
   1391   dec->next_transform_ = 0;
   1392   dec->transforms_seen_ = 0;
   1393 
   1394   WebPSafeFree(dec->rescaler_memory);
   1395   dec->rescaler_memory = NULL;
   1396 
   1397   dec->output_ = NULL;   // leave no trace behind
   1398 }
   1399 
   1400 void VP8LDelete(VP8LDecoder* const dec) {
   1401   if (dec != NULL) {
   1402     VP8LClear(dec);
   1403     WebPSafeFree(dec);
   1404   }
   1405 }
   1406 
   1407 static void UpdateDecoder(VP8LDecoder* const dec, int width, int height) {
   1408   VP8LMetadata* const hdr = &dec->hdr_;
   1409   const int num_bits = hdr->huffman_subsample_bits_;
   1410   dec->width_ = width;
   1411   dec->height_ = height;
   1412 
   1413   hdr->huffman_xsize_ = VP8LSubSampleSize(width, num_bits);
   1414   hdr->huffman_mask_ = (num_bits == 0) ? ~0 : (1 << num_bits) - 1;
   1415 }
   1416 
   1417 static int DecodeImageStream(int xsize, int ysize,
   1418                              int is_level0,
   1419                              VP8LDecoder* const dec,
   1420                              uint32_t** const decoded_data) {
   1421   int ok = 1;
   1422   int transform_xsize = xsize;
   1423   int transform_ysize = ysize;
   1424   VP8LBitReader* const br = &dec->br_;
   1425   VP8LMetadata* const hdr = &dec->hdr_;
   1426   uint32_t* data = NULL;
   1427   int color_cache_bits = 0;
   1428 
   1429   // Read the transforms (may recurse).
   1430   if (is_level0) {
   1431     while (ok && VP8LReadBits(br, 1)) {
   1432       ok = ReadTransform(&transform_xsize, &transform_ysize, dec);
   1433     }
   1434   }
   1435 
   1436   // Color cache
   1437   if (ok && VP8LReadBits(br, 1)) {
   1438     color_cache_bits = VP8LReadBits(br, 4);
   1439     ok = (color_cache_bits >= 1 && color_cache_bits <= MAX_CACHE_BITS);
   1440     if (!ok) {
   1441       dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
   1442       goto End;
   1443     }
   1444   }
   1445 
   1446   // Read the Huffman codes (may recurse).
   1447   ok = ok && ReadHuffmanCodes(dec, transform_xsize, transform_ysize,
   1448                               color_cache_bits, is_level0);
   1449   if (!ok) {
   1450     dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
   1451     goto End;
   1452   }
   1453 
   1454   // Finish setting up the color-cache
   1455   if (color_cache_bits > 0) {
   1456     hdr->color_cache_size_ = 1 << color_cache_bits;
   1457     if (!VP8LColorCacheInit(&hdr->color_cache_, color_cache_bits)) {
   1458       dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
   1459       ok = 0;
   1460       goto End;
   1461     }
   1462   } else {
   1463     hdr->color_cache_size_ = 0;
   1464   }
   1465   UpdateDecoder(dec, transform_xsize, transform_ysize);
   1466 
   1467   if (is_level0) {   // level 0 complete
   1468     dec->state_ = READ_HDR;
   1469     goto End;
   1470   }
   1471 
   1472   {
   1473     const uint64_t total_size = (uint64_t)transform_xsize * transform_ysize;
   1474     data = (uint32_t*)WebPSafeMalloc(total_size, sizeof(*data));
   1475     if (data == NULL) {
   1476       dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
   1477       ok = 0;
   1478       goto End;
   1479     }
   1480   }
   1481 
   1482   // Use the Huffman trees to decode the LZ77 encoded data.
   1483   ok = DecodeImageData(dec, data, transform_xsize, transform_ysize,
   1484                        transform_ysize, NULL);
   1485   ok = ok && !br->eos_;
   1486 
   1487  End:
   1488   if (!ok) {
   1489     WebPSafeFree(data);
   1490     ClearMetadata(hdr);
   1491   } else {
   1492     if (decoded_data != NULL) {
   1493       *decoded_data = data;
   1494     } else {
   1495       // We allocate image data in this function only for transforms. At level 0
   1496       // (that is: not the transforms), we shouldn't have allocated anything.
   1497       assert(data == NULL);
   1498       assert(is_level0);
   1499     }
   1500     dec->last_pixel_ = 0;  // Reset for future DECODE_DATA_FUNC() calls.
   1501     if (!is_level0) ClearMetadata(hdr);  // Clean up temporary data behind.
   1502   }
   1503   return ok;
   1504 }
   1505 
   1506 //------------------------------------------------------------------------------
   1507 // Allocate internal buffers dec->pixels_ and dec->argb_cache_.
   1508 static int AllocateInternalBuffers32b(VP8LDecoder* const dec, int final_width) {
   1509   const uint64_t num_pixels = (uint64_t)dec->width_ * dec->height_;
   1510   // Scratch buffer corresponding to top-prediction row for transforming the
   1511   // first row in the row-blocks. Not needed for paletted alpha.
   1512   const uint64_t cache_top_pixels = (uint16_t)final_width;
   1513   // Scratch buffer for temporary BGRA storage. Not needed for paletted alpha.
   1514   const uint64_t cache_pixels = (uint64_t)final_width * NUM_ARGB_CACHE_ROWS;
   1515   const uint64_t total_num_pixels =
   1516       num_pixels + cache_top_pixels + cache_pixels;
   1517 
   1518   assert(dec->width_ <= final_width);
   1519   dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint32_t));
   1520   if (dec->pixels_ == NULL) {
   1521     dec->argb_cache_ = NULL;    // for sanity check
   1522     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
   1523     return 0;
   1524   }
   1525   dec->argb_cache_ = dec->pixels_ + num_pixels + cache_top_pixels;
   1526   return 1;
   1527 }
   1528 
   1529 static int AllocateInternalBuffers8b(VP8LDecoder* const dec) {
   1530   const uint64_t total_num_pixels = (uint64_t)dec->width_ * dec->height_;
   1531   dec->argb_cache_ = NULL;    // for sanity check
   1532   dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint8_t));
   1533   if (dec->pixels_ == NULL) {
   1534     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
   1535     return 0;
   1536   }
   1537   return 1;
   1538 }
   1539 
   1540 //------------------------------------------------------------------------------
   1541 
   1542 // Special row-processing that only stores the alpha data.
   1543 static void ExtractAlphaRows(VP8LDecoder* const dec, int last_row) {
   1544   int cur_row = dec->last_row_;
   1545   int num_rows = last_row - cur_row;
   1546   const uint32_t* in = dec->pixels_ + dec->width_ * cur_row;
   1547 
   1548   assert(last_row <= dec->io_->crop_bottom);
   1549   while (num_rows > 0) {
   1550     const int num_rows_to_process =
   1551         (num_rows > NUM_ARGB_CACHE_ROWS) ? NUM_ARGB_CACHE_ROWS : num_rows;
   1552     // Extract alpha (which is stored in the green plane).
   1553     ALPHDecoder* const alph_dec = (ALPHDecoder*)dec->io_->opaque;
   1554     uint8_t* const output = alph_dec->output_;
   1555     const int width = dec->io_->width;      // the final width (!= dec->width_)
   1556     const int cache_pixs = width * num_rows_to_process;
   1557     uint8_t* const dst = output + width * cur_row;
   1558     const uint32_t* const src = dec->argb_cache_;
   1559     ApplyInverseTransforms(dec, num_rows_to_process, in);
   1560     WebPExtractGreen(src, dst, cache_pixs);
   1561     AlphaApplyFilter(alph_dec,
   1562                      cur_row, cur_row + num_rows_to_process, dst, width);
   1563     num_rows -= num_rows_to_process;
   1564     in += num_rows_to_process * dec->width_;
   1565     cur_row += num_rows_to_process;
   1566   }
   1567   assert(cur_row == last_row);
   1568   dec->last_row_ = dec->last_out_row_ = last_row;
   1569 }
   1570 
   1571 int VP8LDecodeAlphaHeader(ALPHDecoder* const alph_dec,
   1572                           const uint8_t* const data, size_t data_size) {
   1573   int ok = 0;
   1574   VP8LDecoder* dec = VP8LNew();
   1575 
   1576   if (dec == NULL) return 0;
   1577 
   1578   assert(alph_dec != NULL);
   1579 
   1580   dec->width_ = alph_dec->width_;
   1581   dec->height_ = alph_dec->height_;
   1582   dec->io_ = &alph_dec->io_;
   1583   dec->io_->opaque = alph_dec;
   1584   dec->io_->width = alph_dec->width_;
   1585   dec->io_->height = alph_dec->height_;
   1586 
   1587   dec->status_ = VP8_STATUS_OK;
   1588   VP8LInitBitReader(&dec->br_, data, data_size);
   1589 
   1590   if (!DecodeImageStream(alph_dec->width_, alph_dec->height_, 1, dec, NULL)) {
   1591     goto Err;
   1592   }
   1593 
   1594   // Special case: if alpha data uses only the color indexing transform and
   1595   // doesn't use color cache (a frequent case), we will use DecodeAlphaData()
   1596   // method that only needs allocation of 1 byte per pixel (alpha channel).
   1597   if (dec->next_transform_ == 1 &&
   1598       dec->transforms_[0].type_ == COLOR_INDEXING_TRANSFORM &&
   1599       Is8bOptimizable(&dec->hdr_)) {
   1600     alph_dec->use_8b_decode_ = 1;
   1601     ok = AllocateInternalBuffers8b(dec);
   1602   } else {
   1603     // Allocate internal buffers (note that dec->width_ may have changed here).
   1604     alph_dec->use_8b_decode_ = 0;
   1605     ok = AllocateInternalBuffers32b(dec, alph_dec->width_);
   1606   }
   1607 
   1608   if (!ok) goto Err;
   1609 
   1610   // Only set here, once we are sure it is valid (to avoid thread races).
   1611   alph_dec->vp8l_dec_ = dec;
   1612   return 1;
   1613 
   1614  Err:
   1615   VP8LDelete(dec);
   1616   return 0;
   1617 }
   1618 
   1619 int VP8LDecodeAlphaImageStream(ALPHDecoder* const alph_dec, int last_row) {
   1620   VP8LDecoder* const dec = alph_dec->vp8l_dec_;
   1621   assert(dec != NULL);
   1622   assert(last_row <= dec->height_);
   1623 
   1624   if (dec->last_row_ >= last_row) {
   1625     return 1;  // done
   1626   }
   1627 
   1628   if (!alph_dec->use_8b_decode_) WebPInitAlphaProcessing();
   1629 
   1630   // Decode (with special row processing).
   1631   return alph_dec->use_8b_decode_ ?
   1632       DecodeAlphaData(dec, (uint8_t*)dec->pixels_, dec->width_, dec->height_,
   1633                       last_row) :
   1634       DecodeImageData(dec, dec->pixels_, dec->width_, dec->height_,
   1635                       last_row, ExtractAlphaRows);
   1636 }
   1637 
   1638 //------------------------------------------------------------------------------
   1639 
   1640 int VP8LDecodeHeader(VP8LDecoder* const dec, VP8Io* const io) {
   1641   int width, height, has_alpha;
   1642 
   1643   if (dec == NULL) return 0;
   1644   if (io == NULL) {
   1645     dec->status_ = VP8_STATUS_INVALID_PARAM;
   1646     return 0;
   1647   }
   1648 
   1649   dec->io_ = io;
   1650   dec->status_ = VP8_STATUS_OK;
   1651   VP8LInitBitReader(&dec->br_, io->data, io->data_size);
   1652   if (!ReadImageInfo(&dec->br_, &width, &height, &has_alpha)) {
   1653     dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
   1654     goto Error;
   1655   }
   1656   dec->state_ = READ_DIM;
   1657   io->width = width;
   1658   io->height = height;
   1659 
   1660   if (!DecodeImageStream(width, height, 1, dec, NULL)) goto Error;
   1661   return 1;
   1662 
   1663  Error:
   1664   VP8LClear(dec);
   1665   assert(dec->status_ != VP8_STATUS_OK);
   1666   return 0;
   1667 }
   1668 
   1669 int VP8LDecodeImage(VP8LDecoder* const dec) {
   1670   VP8Io* io = NULL;
   1671   WebPDecParams* params = NULL;
   1672 
   1673   // Sanity checks.
   1674   if (dec == NULL) return 0;
   1675 
   1676   assert(dec->hdr_.huffman_tables_ != NULL);
   1677   assert(dec->hdr_.htree_groups_ != NULL);
   1678   assert(dec->hdr_.num_htree_groups_ > 0);
   1679 
   1680   io = dec->io_;
   1681   assert(io != NULL);
   1682   params = (WebPDecParams*)io->opaque;
   1683   assert(params != NULL);
   1684 
   1685   // Initialization.
   1686   if (dec->state_ != READ_DATA) {
   1687     dec->output_ = params->output;
   1688     assert(dec->output_ != NULL);
   1689 
   1690     if (!WebPIoInitFromOptions(params->options, io, MODE_BGRA)) {
   1691       dec->status_ = VP8_STATUS_INVALID_PARAM;
   1692       goto Err;
   1693     }
   1694 
   1695     if (!AllocateInternalBuffers32b(dec, io->width)) goto Err;
   1696 
   1697 #if !defined(WEBP_REDUCE_SIZE)
   1698     if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
   1699 #else
   1700     if (io->use_scaling) {
   1701       dec->status_ = VP8_STATUS_INVALID_PARAM;
   1702       goto Err;
   1703     }
   1704 #endif
   1705     if (io->use_scaling || WebPIsPremultipliedMode(dec->output_->colorspace)) {
   1706       // need the alpha-multiply functions for premultiplied output or rescaling
   1707       WebPInitAlphaProcessing();
   1708     }
   1709 
   1710     if (!WebPIsRGBMode(dec->output_->colorspace)) {
   1711       WebPInitConvertARGBToYUV();
   1712       if (dec->output_->u.YUVA.a != NULL) WebPInitAlphaProcessing();
   1713     }
   1714     if (dec->incremental_) {
   1715       if (dec->hdr_.color_cache_size_ > 0 &&
   1716           dec->hdr_.saved_color_cache_.colors_ == NULL) {
   1717         if (!VP8LColorCacheInit(&dec->hdr_.saved_color_cache_,
   1718                                 dec->hdr_.color_cache_.hash_bits_)) {
   1719           dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
   1720           goto Err;
   1721         }
   1722       }
   1723     }
   1724     dec->state_ = READ_DATA;
   1725   }
   1726 
   1727   // Decode.
   1728   if (!DecodeImageData(dec, dec->pixels_, dec->width_, dec->height_,
   1729                        io->crop_bottom, ProcessRows)) {
   1730     goto Err;
   1731   }
   1732 
   1733   params->last_y = dec->last_out_row_;
   1734   return 1;
   1735 
   1736  Err:
   1737   VP8LClear(dec);
   1738   assert(dec->status_ != VP8_STATUS_OK);
   1739   return 0;
   1740 }
   1741 
   1742 //------------------------------------------------------------------------------
   1743