go-libwebp

Experimental translation from libwebp to Go source.
Log | Files | Refs | README | LICENSE

vp8l_dec.c (62514B)


      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@gmail.com)
     13 //          Jyrki Alakuijala (jyrki@google.com)
     14 
     15 #include <assert.h>
     16 #include <stddef.h>
     17 #include <stdlib.h>
     18 #include <string.h>
     19 
     20 #include "src/dec/alphai_dec.h"
     21 #include "src/dec/vp8_dec.h"
     22 #include "src/dec/vp8li_dec.h"
     23 #include "src/dec/webpi_dec.h"
     24 #include "src/dsp/dsp.h"
     25 #include "src/dsp/lossless.h"
     26 #include "src/dsp/lossless_common.h"
     27 #include "src/utils/bit_reader_utils.h"
     28 #include "src/utils/color_cache_utils.h"
     29 #include "src/utils/huffman_utils.h"
     30 #include "src/utils/rescaler_utils.h"
     31 #include "src/utils/utils.h"
     32 #include "src/webp/decode.h"
     33 #include "src/webp/format_constants.h"
     34 #include "src/webp/types.h"
     35 
     36 #define NUM_ARGB_CACHE_ROWS          16
     37 
     38 static const int kCodeLengthLiterals = 16;
     39 static const int kCodeLengthRepeatCode = 16;
     40 static const uint8_t kCodeLengthExtraBits[3] = { 2, 3, 7 };
     41 static const uint8_t kCodeLengthRepeatOffsets[3] = { 3, 3, 11 };
     42 
     43 // -----------------------------------------------------------------------------
     44 //  Five Huffman codes are used at each meta code:
     45 //  1. green + length prefix codes + color cache codes,
     46 //  2. alpha,
     47 //  3. red,
     48 //  4. blue, and,
     49 //  5. distance prefix codes.
     50 typedef enum {
     51   GREEN = 0,
     52   RED   = 1,
     53   BLUE  = 2,
     54   ALPHA = 3,
     55   DIST  = 4
     56 } HuffIndex;
     57 
     58 static const uint16_t kAlphabetSize[HUFFMAN_CODES_PER_META_CODE] = {
     59   NUM_LITERAL_CODES + NUM_LENGTH_CODES,
     60   NUM_LITERAL_CODES, NUM_LITERAL_CODES, NUM_LITERAL_CODES,
     61   NUM_DISTANCE_CODES
     62 };
     63 
     64 static const uint8_t kLiteralMap[HUFFMAN_CODES_PER_META_CODE] = {
     65   0, 1, 1, 1, 0
     66 };
     67 
     68 #define NUM_CODE_LENGTH_CODES       19
     69 static const uint8_t kCodeLengthCodeOrder[NUM_CODE_LENGTH_CODES] = {
     70   17, 18, 0, 1, 2, 3, 4, 5, 16, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
     71 };
     72 
     73 #define CODE_TO_PLANE_CODES        120
     74 static const uint8_t kCodeToPlane[CODE_TO_PLANE_CODES] = {
     75   0x18, 0x07, 0x17, 0x19, 0x28, 0x06, 0x27, 0x29, 0x16, 0x1a,
     76   0x26, 0x2a, 0x38, 0x05, 0x37, 0x39, 0x15, 0x1b, 0x36, 0x3a,
     77   0x25, 0x2b, 0x48, 0x04, 0x47, 0x49, 0x14, 0x1c, 0x35, 0x3b,
     78   0x46, 0x4a, 0x24, 0x2c, 0x58, 0x45, 0x4b, 0x34, 0x3c, 0x03,
     79   0x57, 0x59, 0x13, 0x1d, 0x56, 0x5a, 0x23, 0x2d, 0x44, 0x4c,
     80   0x55, 0x5b, 0x33, 0x3d, 0x68, 0x02, 0x67, 0x69, 0x12, 0x1e,
     81   0x66, 0x6a, 0x22, 0x2e, 0x54, 0x5c, 0x43, 0x4d, 0x65, 0x6b,
     82   0x32, 0x3e, 0x78, 0x01, 0x77, 0x79, 0x53, 0x5d, 0x11, 0x1f,
     83   0x64, 0x6c, 0x42, 0x4e, 0x76, 0x7a, 0x21, 0x2f, 0x75, 0x7b,
     84   0x31, 0x3f, 0x63, 0x6d, 0x52, 0x5e, 0x00, 0x74, 0x7c, 0x41,
     85   0x4f, 0x10, 0x20, 0x62, 0x6e, 0x30, 0x73, 0x7d, 0x51, 0x5f,
     86   0x40, 0x72, 0x7e, 0x61, 0x6f, 0x50, 0x71, 0x7f, 0x60, 0x70
     87 };
     88 
     89 // Memory needed for lookup tables of one Huffman tree group. Red, blue, alpha
     90 // and distance alphabets are constant (256 for red, blue and alpha, 40 for
     91 // distance) and lookup table sizes for them in worst case are 630 and 410
     92 // respectively. Size of green alphabet depends on color cache size and is equal
     93 // to 256 (green component values) + 24 (length prefix values)
     94 // + color_cache_size (between 0 and 2048).
     95 // All values computed for 8-bit first level lookup with Mark Adler's tool:
     96 // https://github.com/madler/zlib/blob/v1.2.5/examples/enough.c
     97 #define FIXED_TABLE_SIZE (630 * 3 + 410)
     98 static const uint16_t kTableSize[12] = {
     99   FIXED_TABLE_SIZE + 654,
    100   FIXED_TABLE_SIZE + 656,
    101   FIXED_TABLE_SIZE + 658,
    102   FIXED_TABLE_SIZE + 662,
    103   FIXED_TABLE_SIZE + 670,
    104   FIXED_TABLE_SIZE + 686,
    105   FIXED_TABLE_SIZE + 718,
    106   FIXED_TABLE_SIZE + 782,
    107   FIXED_TABLE_SIZE + 912,
    108   FIXED_TABLE_SIZE + 1168,
    109   FIXED_TABLE_SIZE + 1680,
    110   FIXED_TABLE_SIZE + 2704
    111 };
    112 
    113 static int VP8LSetError(VP8LDecoder* const dec, VP8StatusCode error) {
    114   // The oldest error reported takes precedence over the new one.
    115   if (dec->status == VP8_STATUS_OK || dec->status == VP8_STATUS_SUSPENDED) {
    116     dec->status = error;
    117   }
    118   return 0;
    119 }
    120 
    121 static int DecodeImageStream(int xsize, int ysize,
    122                              int is_level0,
    123                              VP8LDecoder* const dec,
    124                              uint32_t** const decoded_data);
    125 
    126 //------------------------------------------------------------------------------
    127 
    128 int VP8LCheckSignature(const uint8_t* const data, size_t size) {
    129   return (size >= VP8L_FRAME_HEADER_SIZE &&
    130           data[0] == VP8L_MAGIC_BYTE &&
    131           (data[4] >> 5) == 0);  // version
    132 }
    133 
    134 static int ReadImageInfo(VP8LBitReader* const br,
    135                          int* const width, int* const height,
    136                          int* const has_alpha) {
    137   if (VP8LReadBits(br, 8) != VP8L_MAGIC_BYTE) return 0;
    138   *width = VP8LReadBits(br, VP8L_IMAGE_SIZE_BITS) + 1;
    139   *height = VP8LReadBits(br, VP8L_IMAGE_SIZE_BITS) + 1;
    140   *has_alpha = VP8LReadBits(br, 1);
    141   if (VP8LReadBits(br, VP8L_VERSION_BITS) != 0) return 0;
    142   return !br->eos;
    143 }
    144 
    145 int VP8LGetInfo(const uint8_t* data, size_t data_size,
    146                 int* const width, int* const height, int* const has_alpha) {
    147   if (data == NULL || data_size < VP8L_FRAME_HEADER_SIZE) {
    148     return 0;         // not enough data
    149   } else if (!VP8LCheckSignature(data, data_size)) {
    150     return 0;         // bad signature
    151   } else {
    152     int w, h, a;
    153     VP8LBitReader br;
    154     VP8LInitBitReader(&br, data, data_size);
    155     if (!ReadImageInfo(&br, &w, &h, &a)) {
    156       return 0;
    157     }
    158     if (width != NULL) *width = w;
    159     if (height != NULL) *height = h;
    160     if (has_alpha != NULL) *has_alpha = a;
    161     return 1;
    162   }
    163 }
    164 
    165 //------------------------------------------------------------------------------
    166 
    167 static WEBP_INLINE int GetCopyDistance(int distance_symbol,
    168                                        VP8LBitReader* const br) {
    169   int extra_bits, offset;
    170   if (distance_symbol < 4) {
    171     return distance_symbol + 1;
    172   }
    173   extra_bits = (distance_symbol - 2) >> 1;
    174   offset = (2 + (distance_symbol & 1)) << extra_bits;
    175   return offset + VP8LReadBits(br, extra_bits) + 1;
    176 }
    177 
    178 static WEBP_INLINE int GetCopyLength(int length_symbol,
    179                                      VP8LBitReader* const br) {
    180   // Length and distance prefixes are encoded the same way.
    181   return GetCopyDistance(length_symbol, br);
    182 }
    183 
    184 static WEBP_INLINE int PlaneCodeToDistance(int xsize, int plane_code) {
    185   if (plane_code > CODE_TO_PLANE_CODES) {
    186     return plane_code - CODE_TO_PLANE_CODES;
    187   } else {
    188     const int dist_code = kCodeToPlane[plane_code - 1];
    189     const int yoffset = dist_code >> 4;
    190     const int xoffset = 8 - (dist_code & 0xf);
    191     const int dist = yoffset * xsize + xoffset;
    192     return (dist >= 1) ? dist : 1;  // dist<1 can happen if xsize is very small
    193   }
    194 }
    195 
    196 //------------------------------------------------------------------------------
    197 // Decodes the next Huffman code from bit-stream.
    198 // VP8LFillBitWindow(br) needs to be called at minimum every second call
    199 // to ReadSymbol, in order to pre-fetch enough bits.
    200 static WEBP_INLINE int ReadSymbol(const HuffmanCode* table,
    201                                   VP8LBitReader* const br) {
    202   int nbits;
    203   uint32_t val = VP8LPrefetchBits(br);
    204   table += val & HUFFMAN_TABLE_MASK;
    205   nbits = table->bits - HUFFMAN_TABLE_BITS;
    206   if (nbits > 0) {
    207     VP8LSetBitPos(br, br->bit_pos + HUFFMAN_TABLE_BITS);
    208     val = VP8LPrefetchBits(br);
    209     table += table->value;
    210     table += val & ((1 << nbits) - 1);
    211   }
    212   VP8LSetBitPos(br, br->bit_pos + table->bits);
    213   return table->value;
    214 }
    215 
    216 // Reads packed symbol depending on GREEN channel
    217 #define BITS_SPECIAL_MARKER 0x100  // something large enough (and a bit-mask)
    218 #define PACKED_NON_LITERAL_CODE 0  // must be < NUM_LITERAL_CODES
    219 static WEBP_INLINE int ReadPackedSymbols(const HTreeGroup* group,
    220                                          VP8LBitReader* const br,
    221                                          uint32_t* const dst) {
    222   const uint32_t val = VP8LPrefetchBits(br) & (HUFFMAN_PACKED_TABLE_SIZE - 1);
    223   const HuffmanCode32 code = group->packed_table[val];
    224   assert(group->use_packed_table);
    225   if (code.bits < BITS_SPECIAL_MARKER) {
    226     VP8LSetBitPos(br, br->bit_pos + code.bits);
    227     *dst = code.value;
    228     return PACKED_NON_LITERAL_CODE;
    229   } else {
    230     VP8LSetBitPos(br, br->bit_pos + code.bits - BITS_SPECIAL_MARKER);
    231     assert(code.value >= NUM_LITERAL_CODES);
    232     return code.value;
    233   }
    234 }
    235 
    236 static int AccumulateHCode(HuffmanCode hcode, int shift,
    237                            HuffmanCode32* const huff) {
    238   huff->bits += hcode.bits;
    239   huff->value |= (uint32_t)hcode.value << shift;
    240   assert(huff->bits <= HUFFMAN_TABLE_BITS);
    241   return hcode.bits;
    242 }
    243 
    244 static void BuildPackedTable(HTreeGroup* const htree_group) {
    245   uint32_t code;
    246   for (code = 0; code < HUFFMAN_PACKED_TABLE_SIZE; ++code) {
    247     uint32_t bits = code;
    248     HuffmanCode32* const huff = &htree_group->packed_table[bits];
    249     HuffmanCode hcode = htree_group->htrees[GREEN][bits];
    250     if (hcode.value >= NUM_LITERAL_CODES) {
    251       huff->bits = hcode.bits + BITS_SPECIAL_MARKER;
    252       huff->value = hcode.value;
    253     } else {
    254       huff->bits = 0;
    255       huff->value = 0;
    256       bits >>= AccumulateHCode(hcode, 8, huff);
    257       bits >>= AccumulateHCode(htree_group->htrees[RED][bits], 16, huff);
    258       bits >>= AccumulateHCode(htree_group->htrees[BLUE][bits], 0, huff);
    259       bits >>= AccumulateHCode(htree_group->htrees[ALPHA][bits], 24, huff);
    260       (void)bits;
    261     }
    262   }
    263 }
    264 
    265 static int ReadHuffmanCodeLengths(
    266     VP8LDecoder* const dec, const int* const code_length_code_lengths,
    267     int num_symbols, int* const code_lengths) {
    268   int ok = 0;
    269   VP8LBitReader* const br = &dec->br;
    270   int symbol;
    271   int max_symbol;
    272   int prev_code_len = DEFAULT_CODE_LENGTH;
    273   HuffmanTables tables;
    274 
    275   if (!VP8LHuffmanTablesAllocate(1 << LENGTHS_TABLE_BITS, &tables) ||
    276       !VP8LBuildHuffmanTable(&tables, LENGTHS_TABLE_BITS,
    277                              code_length_code_lengths, NUM_CODE_LENGTH_CODES)) {
    278     goto End;
    279   }
    280 
    281   if (VP8LReadBits(br, 1)) {    // use length
    282     const int length_nbits = 2 + 2 * VP8LReadBits(br, 3);
    283     max_symbol = 2 + VP8LReadBits(br, length_nbits);
    284     if (max_symbol > num_symbols) {
    285       goto End;
    286     }
    287   } else {
    288     max_symbol = num_symbols;
    289   }
    290 
    291   symbol = 0;
    292   while (symbol < num_symbols) {
    293     const HuffmanCode* p;
    294     int code_len;
    295     if (max_symbol-- == 0) break;
    296     VP8LFillBitWindow(br);
    297     p = &tables.curr_segment->start[VP8LPrefetchBits(br) & LENGTHS_TABLE_MASK];
    298     VP8LSetBitPos(br, br->bit_pos + p->bits);
    299     code_len = p->value;
    300     if (code_len < kCodeLengthLiterals) {
    301       code_lengths[symbol++] = code_len;
    302       if (code_len != 0) prev_code_len = code_len;
    303     } else {
    304       const int use_prev = (code_len == kCodeLengthRepeatCode);
    305       const int slot = code_len - kCodeLengthLiterals;
    306       const int extra_bits = kCodeLengthExtraBits[slot];
    307       const int repeat_offset = kCodeLengthRepeatOffsets[slot];
    308       int repeat = VP8LReadBits(br, extra_bits) + repeat_offset;
    309       if (symbol + repeat > num_symbols) {
    310         goto End;
    311       } else {
    312         const int length = use_prev ? prev_code_len : 0;
    313         while (repeat-- > 0) code_lengths[symbol++] = length;
    314       }
    315     }
    316   }
    317   ok = 1;
    318 
    319  End:
    320   VP8LHuffmanTablesDeallocate(&tables);
    321   if (!ok) return VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
    322   return ok;
    323 }
    324 
    325 // 'code_lengths' is pre-allocated temporary buffer, used for creating Huffman
    326 // tree.
    327 static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
    328                            int* const code_lengths,
    329                            HuffmanTables* const table) {
    330   int ok = 0;
    331   int size = 0;
    332   VP8LBitReader* const br = &dec->br;
    333   const int simple_code = VP8LReadBits(br, 1);
    334 
    335   memset(code_lengths, 0, alphabet_size * sizeof(*code_lengths));
    336 
    337   if (simple_code) {  // Read symbols, codes & code lengths directly.
    338     const int num_symbols = VP8LReadBits(br, 1) + 1;
    339     const int first_symbol_len_code = VP8LReadBits(br, 1);
    340     // The first code is either 1 bit or 8 bit code.
    341     int symbol = VP8LReadBits(br, (first_symbol_len_code == 0) ? 1 : 8);
    342     code_lengths[symbol] = 1;
    343     // The second code (if present), is always 8 bits long.
    344     if (num_symbols == 2) {
    345       symbol = VP8LReadBits(br, 8);
    346       code_lengths[symbol] = 1;
    347     }
    348     ok = 1;
    349   } else {  // Decode Huffman-coded code lengths.
    350     int i;
    351     int code_length_code_lengths[NUM_CODE_LENGTH_CODES] = { 0 };
    352     const int num_codes = VP8LReadBits(br, 4) + 4;
    353     assert(num_codes <= NUM_CODE_LENGTH_CODES);
    354 
    355     for (i = 0; i < num_codes; ++i) {
    356       code_length_code_lengths[kCodeLengthCodeOrder[i]] = VP8LReadBits(br, 3);
    357     }
    358     ok = ReadHuffmanCodeLengths(dec, code_length_code_lengths, alphabet_size,
    359                                 code_lengths);
    360   }
    361 
    362   ok = ok && !br->eos;
    363   if (ok) {
    364     size = VP8LBuildHuffmanTable(table, HUFFMAN_TABLE_BITS,
    365                                  code_lengths, alphabet_size);
    366   }
    367   if (!ok || size == 0) {
    368     return VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
    369   }
    370   return size;
    371 }
    372 
    373 static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
    374                             int color_cache_bits, int allow_recursion) {
    375   int i;
    376   VP8LBitReader* const br = &dec->br;
    377   VP8LMetadata* const hdr = &dec->hdr;
    378   uint32_t* huffman_image = NULL;
    379   HTreeGroup* htree_groups = NULL;
    380   HuffmanTables* huffman_tables = &hdr->huffman_tables;
    381   int num_htree_groups = 1;
    382   int num_htree_groups_max = 1;
    383   int* mapping = NULL;
    384   int ok = 0;
    385 
    386   // Check the table has been 0 initialized (through InitMetadata).
    387   assert(huffman_tables->root.start == NULL);
    388   assert(huffman_tables->curr_segment == NULL);
    389 
    390   if (allow_recursion && VP8LReadBits(br, 1)) {
    391     // use meta Huffman codes.
    392     const int huffman_precision =
    393         MIN_HUFFMAN_BITS + VP8LReadBits(br, NUM_HUFFMAN_BITS);
    394     const int huffman_xsize = VP8LSubSampleSize(xsize, huffman_precision);
    395     const int huffman_ysize = VP8LSubSampleSize(ysize, huffman_precision);
    396     const int huffman_pixs = huffman_xsize * huffman_ysize;
    397     if (!DecodeImageStream(huffman_xsize, huffman_ysize, /*is_level0=*/0, dec,
    398                            &huffman_image)) {
    399       goto Error;
    400     }
    401     hdr->huffman_subsample_bits = huffman_precision;
    402     for (i = 0; i < huffman_pixs; ++i) {
    403       // The huffman data is stored in red and green bytes.
    404       const int group = (huffman_image[i] >> 8) & 0xffff;
    405       huffman_image[i] = group;
    406       if (group >= num_htree_groups_max) {
    407         num_htree_groups_max = group + 1;
    408       }
    409     }
    410     // Check the validity of num_htree_groups_max. If it seems too big, use a
    411     // smaller value for later. This will prevent big memory allocations to end
    412     // up with a bad bitstream anyway.
    413     // The value of 1000 is totally arbitrary. We know that num_htree_groups_max
    414     // is smaller than (1 << 16) and should be smaller than the number of pixels
    415     // (though the format allows it to be bigger).
    416     if (num_htree_groups_max > 1000 || num_htree_groups_max > xsize * ysize) {
    417       // Create a mapping from the used indices to the minimal set of used
    418       // values [0, num_htree_groups)
    419       mapping = (int*)WebPSafeMalloc(num_htree_groups_max, sizeof(*mapping));
    420       if (mapping == NULL) {
    421         VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
    422         goto Error;
    423       }
    424       // -1 means a value is unmapped, and therefore unused in the Huffman
    425       // image.
    426       memset(mapping, 0xff, num_htree_groups_max * sizeof(*mapping));
    427       for (num_htree_groups = 0, i = 0; i < huffman_pixs; ++i) {
    428         // Get the current mapping for the group and remap the Huffman image.
    429         int* const mapped_group = &mapping[huffman_image[i]];
    430         if (*mapped_group == -1) *mapped_group = num_htree_groups++;
    431         huffman_image[i] = *mapped_group;
    432       }
    433     } else {
    434       num_htree_groups = num_htree_groups_max;
    435     }
    436   }
    437 
    438   if (br->eos) goto Error;
    439 
    440   if (!ReadHuffmanCodesHelper(color_cache_bits, num_htree_groups,
    441                               num_htree_groups_max, mapping, dec,
    442                               huffman_tables, &htree_groups)) {
    443     goto Error;
    444   }
    445   ok = 1;
    446 
    447   // All OK. Finalize pointers.
    448   hdr->huffman_image = huffman_image;
    449   hdr->num_htree_groups = num_htree_groups;
    450   hdr->htree_groups = htree_groups;
    451 
    452  Error:
    453   WebPSafeFree(mapping);
    454   if (!ok) {
    455     WebPSafeFree(huffman_image);
    456     VP8LHuffmanTablesDeallocate(huffman_tables);
    457     VP8LHtreeGroupsFree(htree_groups);
    458   }
    459   return ok;
    460 }
    461 
    462 int ReadHuffmanCodesHelper(int color_cache_bits, int num_htree_groups,
    463                            int num_htree_groups_max, const int* const mapping,
    464                            VP8LDecoder* const dec,
    465                            HuffmanTables* const huffman_tables,
    466                            HTreeGroup** const htree_groups) {
    467   int i, j, ok = 0;
    468   const int max_alphabet_size =
    469       kAlphabetSize[0] + ((color_cache_bits > 0) ? 1 << color_cache_bits : 0);
    470   const int table_size = kTableSize[color_cache_bits];
    471   int* code_lengths = NULL;
    472 
    473   if ((mapping == NULL && num_htree_groups != num_htree_groups_max) ||
    474       num_htree_groups > num_htree_groups_max) {
    475     goto Error;
    476   }
    477 
    478   code_lengths =
    479       (int*)WebPSafeCalloc((uint64_t)max_alphabet_size, sizeof(*code_lengths));
    480   *htree_groups = VP8LHtreeGroupsNew(num_htree_groups);
    481 
    482   if (*htree_groups == NULL || code_lengths == NULL ||
    483       !VP8LHuffmanTablesAllocate(num_htree_groups * table_size,
    484                                  huffman_tables)) {
    485     VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
    486     goto Error;
    487   }
    488 
    489   for (i = 0; i < num_htree_groups_max; ++i) {
    490     // If the index "i" is unused in the Huffman image, just make sure the
    491     // coefficients are valid but do not store them.
    492     if (mapping != NULL && mapping[i] == -1) {
    493       for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
    494         int alphabet_size = kAlphabetSize[j];
    495         if (j == 0 && color_cache_bits > 0) {
    496           alphabet_size += (1 << color_cache_bits);
    497         }
    498         // Passing in NULL so that nothing gets filled.
    499         if (!ReadHuffmanCode(alphabet_size, dec, code_lengths, NULL)) {
    500           goto Error;
    501         }
    502       }
    503     } else {
    504       HTreeGroup* const htree_group =
    505           &(*htree_groups)[(mapping == NULL) ? i : mapping[i]];
    506       HuffmanCode** const htrees = htree_group->htrees;
    507       int size;
    508       int total_size = 0;
    509       int is_trivial_literal = 1;
    510       int max_bits = 0;
    511       for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
    512         int alphabet_size = kAlphabetSize[j];
    513         if (j == 0 && color_cache_bits > 0) {
    514           alphabet_size += (1 << color_cache_bits);
    515         }
    516         size =
    517             ReadHuffmanCode(alphabet_size, dec, code_lengths, huffman_tables);
    518         htrees[j] = huffman_tables->curr_segment->curr_table;
    519         if (size == 0) {
    520           goto Error;
    521         }
    522         if (is_trivial_literal && kLiteralMap[j] == 1) {
    523           is_trivial_literal = (htrees[j]->bits == 0);
    524         }
    525         total_size += htrees[j]->bits;
    526         huffman_tables->curr_segment->curr_table += size;
    527         if (j <= ALPHA) {
    528           int local_max_bits = code_lengths[0];
    529           int k;
    530           for (k = 1; k < alphabet_size; ++k) {
    531             if (code_lengths[k] > local_max_bits) {
    532               local_max_bits = code_lengths[k];
    533             }
    534           }
    535           max_bits += local_max_bits;
    536         }
    537       }
    538       htree_group->is_trivial_literal = is_trivial_literal;
    539       htree_group->is_trivial_code = 0;
    540       if (is_trivial_literal) {
    541         const int red = htrees[RED][0].value;
    542         const int blue = htrees[BLUE][0].value;
    543         const int alpha = htrees[ALPHA][0].value;
    544         htree_group->literal_arb = ((uint32_t)alpha << 24) | (red << 16) | blue;
    545         if (total_size == 0 && htrees[GREEN][0].value < NUM_LITERAL_CODES) {
    546           htree_group->is_trivial_code = 1;
    547           htree_group->literal_arb |= htrees[GREEN][0].value << 8;
    548         }
    549       }
    550       htree_group->use_packed_table =
    551           !htree_group->is_trivial_code && (max_bits < HUFFMAN_PACKED_BITS);
    552       if (htree_group->use_packed_table) BuildPackedTable(htree_group);
    553     }
    554   }
    555   ok = 1;
    556 
    557  Error:
    558   WebPSafeFree(code_lengths);
    559   if (!ok) {
    560     VP8LHuffmanTablesDeallocate(huffman_tables);
    561     VP8LHtreeGroupsFree(*htree_groups);
    562     *htree_groups = NULL;
    563   }
    564   return ok;
    565 }
    566 
    567 //------------------------------------------------------------------------------
    568 // Scaling.
    569 
    570 #if !defined(WEBP_REDUCE_SIZE)
    571 static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
    572   const int num_channels = 4;
    573   const int in_width = io->mb_w;
    574   const int out_width = io->scaled_width;
    575   const int in_height = io->mb_h;
    576   const int out_height = io->scaled_height;
    577   const uint64_t work_size = 2 * num_channels * (uint64_t)out_width;
    578   rescaler_t* work;        // Rescaler work area.
    579   const uint64_t scaled_data_size = (uint64_t)out_width;
    580   uint32_t* scaled_data;  // Temporary storage for scaled BGRA data.
    581   const uint64_t memory_size = sizeof(*dec->rescaler) +
    582                                work_size * sizeof(*work) +
    583                                scaled_data_size * sizeof(*scaled_data);
    584   uint8_t* memory = (uint8_t*)WebPSafeMalloc(memory_size, sizeof(*memory));
    585   if (memory == NULL) {
    586     return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
    587   }
    588   assert(dec->rescaler_memory == NULL);
    589   dec->rescaler_memory = memory;
    590 
    591   dec->rescaler = (WebPRescaler*)memory;
    592   memory += sizeof(*dec->rescaler);
    593   work = (rescaler_t*)memory;
    594   memory += work_size * sizeof(*work);
    595   scaled_data = (uint32_t*)memory;
    596 
    597   if (!WebPRescalerInit(dec->rescaler, in_width, in_height,
    598                         (uint8_t*)scaled_data, out_width, out_height,
    599                         0, num_channels, work)) {
    600     return 0;
    601   }
    602   return 1;
    603 }
    604 #endif   // WEBP_REDUCE_SIZE
    605 
    606 //------------------------------------------------------------------------------
    607 // Export to ARGB
    608 
    609 #if !defined(WEBP_REDUCE_SIZE)
    610 
    611 // We have special "export" function since we need to convert from BGRA
    612 static int Export(WebPRescaler* const rescaler, WEBP_CSP_MODE colorspace,
    613                   int rgba_stride, uint8_t* const rgba) {
    614   uint32_t* const src = (uint32_t*)rescaler->dst;
    615   uint8_t* dst = rgba;
    616   const int dst_width = rescaler->dst_width;
    617   int num_lines_out = 0;
    618   while (WebPRescalerHasPendingOutput(rescaler)) {
    619     WebPRescalerExportRow(rescaler);
    620     WebPMultARGBRow(src, dst_width, 1);
    621     VP8LConvertFromBGRA(src, dst_width, colorspace, dst);
    622     dst += rgba_stride;
    623     ++num_lines_out;
    624   }
    625   return num_lines_out;
    626 }
    627 
    628 // Emit scaled rows.
    629 static int EmitRescaledRowsRGBA(const VP8LDecoder* const dec,
    630                                 uint8_t* in, int in_stride, int mb_h,
    631                                 uint8_t* const out, int out_stride) {
    632   const WEBP_CSP_MODE colorspace = dec->output->colorspace;
    633   int num_lines_in = 0;
    634   int num_lines_out = 0;
    635   while (num_lines_in < mb_h) {
    636     uint8_t* const row_in = in + (ptrdiff_t)num_lines_in * in_stride;
    637     uint8_t* const row_out = out + (ptrdiff_t)num_lines_out * out_stride;
    638     const int lines_left = mb_h - num_lines_in;
    639     const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
    640     int lines_imported;
    641     assert(needed_lines > 0 && needed_lines <= lines_left);
    642     WebPMultARGBRows(row_in, in_stride,
    643                      dec->rescaler->src_width, needed_lines, 0);
    644     lines_imported =
    645         WebPRescalerImport(dec->rescaler, lines_left, row_in, in_stride);
    646     assert(lines_imported == needed_lines);
    647     num_lines_in += lines_imported;
    648     num_lines_out += Export(dec->rescaler, colorspace, out_stride, row_out);
    649   }
    650   return num_lines_out;
    651 }
    652 
    653 #endif   // WEBP_REDUCE_SIZE
    654 
    655 // Emit rows without any scaling.
    656 static int EmitRows(WEBP_CSP_MODE colorspace,
    657                     const uint8_t* row_in, int in_stride,
    658                     int mb_w, int mb_h,
    659                     uint8_t* const out, int out_stride) {
    660   int lines = mb_h;
    661   uint8_t* row_out = out;
    662   while (lines-- > 0) {
    663     VP8LConvertFromBGRA((const uint32_t*)row_in, mb_w, colorspace, row_out);
    664     row_in += in_stride;
    665     row_out += out_stride;
    666   }
    667   return mb_h;  // Num rows out == num rows in.
    668 }
    669 
    670 //------------------------------------------------------------------------------
    671 // Export to YUVA
    672 
    673 static void ConvertToYUVA(const uint32_t* const src, int width, int y_pos,
    674                           const WebPDecBuffer* const output) {
    675   const WebPYUVABuffer* const buf = &output->u.YUVA;
    676 
    677   // first, the luma plane
    678   WebPConvertARGBToY(src, buf->y + y_pos * buf->y_stride, width);
    679 
    680   // then U/V planes
    681   {
    682     uint8_t* const u = buf->u + (y_pos >> 1) * buf->u_stride;
    683     uint8_t* const v = buf->v + (y_pos >> 1) * buf->v_stride;
    684     // even lines: store values
    685     // odd lines: average with previous values
    686     WebPConvertARGBToUV(src, u, v, width, !(y_pos & 1));
    687   }
    688   // Lastly, store alpha if needed.
    689   if (buf->a != NULL) {
    690     uint8_t* const a = buf->a + y_pos * buf->a_stride;
    691 #if defined(WORDS_BIGENDIAN)
    692     WebPExtractAlpha((uint8_t*)src + 0, 0, width, 1, a, 0);
    693 #else
    694     WebPExtractAlpha((uint8_t*)src + 3, 0, width, 1, a, 0);
    695 #endif
    696   }
    697 }
    698 
    699 static int ExportYUVA(const VP8LDecoder* const dec, int y_pos) {
    700   WebPRescaler* const rescaler = dec->rescaler;
    701   uint32_t* const src = (uint32_t*)rescaler->dst;
    702   const int dst_width = rescaler->dst_width;
    703   int num_lines_out = 0;
    704   while (WebPRescalerHasPendingOutput(rescaler)) {
    705     WebPRescalerExportRow(rescaler);
    706     WebPMultARGBRow(src, dst_width, 1);
    707     ConvertToYUVA(src, dst_width, y_pos, dec->output);
    708     ++y_pos;
    709     ++num_lines_out;
    710   }
    711   return num_lines_out;
    712 }
    713 
    714 static int EmitRescaledRowsYUVA(const VP8LDecoder* const dec,
    715                                 uint8_t* in, int in_stride, int mb_h) {
    716   int num_lines_in = 0;
    717   int y_pos = dec->last_out_row;
    718   while (num_lines_in < mb_h) {
    719     const int lines_left = mb_h - num_lines_in;
    720     const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
    721     int lines_imported;
    722     WebPMultARGBRows(in, in_stride, dec->rescaler->src_width, needed_lines, 0);
    723     lines_imported =
    724         WebPRescalerImport(dec->rescaler, lines_left, in, in_stride);
    725     assert(lines_imported == needed_lines);
    726     num_lines_in += lines_imported;
    727     in += needed_lines * in_stride;
    728     y_pos += ExportYUVA(dec, y_pos);
    729   }
    730   return y_pos;
    731 }
    732 
    733 static int EmitRowsYUVA(const VP8LDecoder* const dec,
    734                         const uint8_t* in, int in_stride,
    735                         int mb_w, int num_rows) {
    736   int y_pos = dec->last_out_row;
    737   while (num_rows-- > 0) {
    738     ConvertToYUVA((const uint32_t*)in, mb_w, y_pos, dec->output);
    739     in += in_stride;
    740     ++y_pos;
    741   }
    742   return y_pos;
    743 }
    744 
    745 //------------------------------------------------------------------------------
    746 // Cropping.
    747 
    748 // Sets io->mb_y, io->mb_h & io->mb_w according to start row, end row and
    749 // crop options. Also updates the input data pointer, so that it points to the
    750 // start of the cropped window. Note that pixels are in ARGB format even if
    751 // 'in_data' is uint8_t*.
    752 // Returns true if the crop window is not empty.
    753 static int SetCropWindow(VP8Io* const io, int y_start, int y_end,
    754                          uint8_t** const in_data, int pixel_stride) {
    755   assert(y_start < y_end);
    756   assert(io->crop_left < io->crop_right);
    757   if (y_end > io->crop_bottom) {
    758     y_end = io->crop_bottom;  // make sure we don't overflow on last row.
    759   }
    760   if (y_start < io->crop_top) {
    761     const int delta = io->crop_top - y_start;
    762     y_start = io->crop_top;
    763     *in_data += delta * pixel_stride;
    764   }
    765   if (y_start >= y_end) return 0;  // Crop window is empty.
    766 
    767   *in_data += io->crop_left * sizeof(uint32_t);
    768 
    769   io->mb_y = y_start - io->crop_top;
    770   io->mb_w = io->crop_right - io->crop_left;
    771   io->mb_h = y_end - y_start;
    772   return 1;  // Non-empty crop window.
    773 }
    774 
    775 //------------------------------------------------------------------------------
    776 
    777 static WEBP_INLINE int GetMetaIndex(
    778     const uint32_t* const image, int xsize, int bits, int x, int y) {
    779   if (bits == 0) return 0;
    780   return image[xsize * (y >> bits) + (x >> bits)];
    781 }
    782 
    783 static WEBP_INLINE HTreeGroup* GetHtreeGroupForPos(VP8LMetadata* const hdr,
    784                                                    int x, int y) {
    785   const int meta_index = GetMetaIndex(hdr->huffman_image, hdr->huffman_xsize,
    786                                       hdr->huffman_subsample_bits, x, y);
    787   assert(meta_index < hdr->num_htree_groups);
    788   return hdr->htree_groups + meta_index;
    789 }
    790 
    791 //------------------------------------------------------------------------------
    792 // Main loop, with custom row-processing function
    793 
    794 typedef void (*ProcessRowsFunc)(VP8LDecoder* const dec, int row);
    795 
    796 static void ApplyInverseTransforms(VP8LDecoder* const dec,
    797                                    int start_row, int num_rows,
    798                                    const uint32_t* const rows) {
    799   int n = dec->next_transform;
    800   const int cache_pixs = dec->width * num_rows;
    801   const int end_row = start_row + num_rows;
    802   const uint32_t* rows_in = rows;
    803   uint32_t* const rows_out = dec->argb_cache;
    804 
    805   // Inverse transforms.
    806   while (n-- > 0) {
    807     VP8LTransform* const transform = &dec->transforms[n];
    808     VP8LInverseTransform(transform, start_row, end_row, rows_in, rows_out);
    809     rows_in = rows_out;
    810   }
    811   if (rows_in != rows_out) {
    812     // No transform called, hence just copy.
    813     memcpy(rows_out, rows_in, cache_pixs * sizeof(*rows_out));
    814   }
    815 }
    816 
    817 // Processes (transforms, scales & color-converts) the rows decoded after the
    818 // last call.
    819 static void ProcessRows(VP8LDecoder* const dec, int row) {
    820   const uint32_t* const rows = dec->pixels + dec->width * dec->last_row;
    821   const int num_rows = row - dec->last_row;
    822 
    823   assert(row <= dec->io->crop_bottom);
    824   // We can't process more than NUM_ARGB_CACHE_ROWS at a time (that's the size
    825   // of argb_cache), but we currently don't need more than that.
    826   assert(num_rows <= NUM_ARGB_CACHE_ROWS);
    827   if (num_rows > 0) {    // Emit output.
    828     VP8Io* const io = dec->io;
    829     uint8_t* rows_data = (uint8_t*)dec->argb_cache;
    830     const int in_stride = io->width * sizeof(uint32_t);  // in unit of RGBA
    831     ApplyInverseTransforms(dec, dec->last_row, num_rows, rows);
    832     if (!SetCropWindow(io, dec->last_row, row, &rows_data, in_stride)) {
    833       // Nothing to output (this time).
    834     } else {
    835       const WebPDecBuffer* const output = dec->output;
    836       if (WebPIsRGBMode(output->colorspace)) {  // convert to RGBA
    837         const WebPRGBABuffer* const buf = &output->u.RGBA;
    838         uint8_t* const rgba =
    839             buf->rgba + (ptrdiff_t)dec->last_out_row * buf->stride;
    840         const int num_rows_out =
    841 #if !defined(WEBP_REDUCE_SIZE)
    842          io->use_scaling ?
    843             EmitRescaledRowsRGBA(dec, rows_data, in_stride, io->mb_h,
    844                                  rgba, buf->stride) :
    845 #endif  // WEBP_REDUCE_SIZE
    846             EmitRows(output->colorspace, rows_data, in_stride,
    847                      io->mb_w, io->mb_h, rgba, buf->stride);
    848         // Update 'last_out_row'.
    849         dec->last_out_row += num_rows_out;
    850       } else {                              // convert to YUVA
    851         dec->last_out_row = io->use_scaling ?
    852             EmitRescaledRowsYUVA(dec, rows_data, in_stride, io->mb_h) :
    853             EmitRowsYUVA(dec, rows_data, in_stride, io->mb_w, io->mb_h);
    854       }
    855       assert(dec->last_out_row <= output->height);
    856     }
    857   }
    858 
    859   // Update 'last_row'.
    860   dec->last_row = row;
    861   assert(dec->last_row <= dec->height);
    862 }
    863 
    864 // Row-processing for the special case when alpha data contains only one
    865 // transform (color indexing), and trivial non-green literals.
    866 static int Is8bOptimizable(const VP8LMetadata* const hdr) {
    867   int i;
    868   if (hdr->color_cache_size > 0) return 0;
    869   // When the Huffman tree contains only one symbol, we can skip the
    870   // call to ReadSymbol() for red/blue/alpha channels.
    871   for (i = 0; i < hdr->num_htree_groups; ++i) {
    872     HuffmanCode** const htrees = hdr->htree_groups[i].htrees;
    873     if (htrees[RED][0].bits > 0) return 0;
    874     if (htrees[BLUE][0].bits > 0) return 0;
    875     if (htrees[ALPHA][0].bits > 0) return 0;
    876   }
    877   return 1;
    878 }
    879 
    880 static void AlphaApplyFilter(ALPHDecoder* const alph_dec,
    881                              int first_row, int last_row,
    882                              uint8_t* out, int stride) {
    883   if (alph_dec->filter != WEBP_FILTER_NONE) {
    884     int y;
    885     const uint8_t* prev_line = alph_dec->prev_line;
    886     assert(WebPUnfilters[alph_dec->filter] != NULL);
    887     for (y = first_row; y < last_row; ++y) {
    888       WebPUnfilters[alph_dec->filter](prev_line, out, out, stride);
    889       prev_line = out;
    890       out += stride;
    891     }
    892     alph_dec->prev_line = prev_line;
    893   }
    894 }
    895 
    896 static void ExtractPalettedAlphaRows(VP8LDecoder* const dec, int last_row) {
    897   // For vertical and gradient filtering, we need to decode the part above the
    898   // crop_top row, in order to have the correct spatial predictors.
    899   ALPHDecoder* const alph_dec = (ALPHDecoder*)dec->io->opaque;
    900   const int top_row =
    901       (alph_dec->filter == WEBP_FILTER_NONE ||
    902        alph_dec->filter == WEBP_FILTER_HORIZONTAL) ? dec->io->crop_top
    903                                                     : dec->last_row;
    904   const int first_row = (dec->last_row < top_row) ? top_row : dec->last_row;
    905   assert(last_row <= dec->io->crop_bottom);
    906   if (last_row > first_row) {
    907     // Special method for paletted alpha data. We only process the cropped area.
    908     const int width = dec->io->width;
    909     uint8_t* out = alph_dec->output + width * first_row;
    910     const uint8_t* const in =
    911       (uint8_t*)dec->pixels + dec->width * first_row;
    912     VP8LTransform* const transform = &dec->transforms[0];
    913     assert(dec->next_transform == 1);
    914     assert(transform->type == COLOR_INDEXING_TRANSFORM);
    915     VP8LColorIndexInverseTransformAlpha(transform, first_row, last_row,
    916                                         in, out);
    917     AlphaApplyFilter(alph_dec, first_row, last_row, out, width);
    918   }
    919   dec->last_row = dec->last_out_row = last_row;
    920 }
    921 
    922 //------------------------------------------------------------------------------
    923 // Helper functions for fast pattern copy (8b and 32b)
    924 
    925 // cyclic rotation of pattern word
    926 static WEBP_INLINE uint32_t Rotate8b(uint32_t V) {
    927 #if defined(WORDS_BIGENDIAN)
    928   return ((V & 0xff000000u) >> 24) | (V << 8);
    929 #else
    930   return ((V & 0xffu) << 24) | (V >> 8);
    931 #endif
    932 }
    933 
    934 // copy 1, 2 or 4-bytes pattern
    935 static WEBP_INLINE void CopySmallPattern8b(const uint8_t* src, uint8_t* dst,
    936                                            int length, uint32_t pattern) {
    937   int i;
    938   // align 'dst' to 4-bytes boundary. Adjust the pattern along the way.
    939   while ((uintptr_t)dst & 3) {
    940     *dst++ = *src++;
    941     pattern = Rotate8b(pattern);
    942     --length;
    943   }
    944   // Copy the pattern 4 bytes at a time.
    945   for (i = 0; i < (length >> 2); ++i) {
    946     ((uint32_t*)dst)[i] = pattern;
    947   }
    948   // Finish with left-overs. 'pattern' is still correctly positioned,
    949   // so no Rotate8b() call is needed.
    950   for (i <<= 2; i < length; ++i) {
    951     dst[i] = src[i];
    952   }
    953 }
    954 
    955 static WEBP_INLINE void CopyBlock8b(uint8_t* const dst, int dist, int length) {
    956   const uint8_t* src = dst - dist;
    957   if (length >= 8) {
    958     uint32_t pattern = 0;
    959     switch (dist) {
    960       case 1:
    961         pattern = src[0];
    962 #if defined(__arm__) || defined(_M_ARM)   // arm doesn't like multiply that much
    963         pattern |= pattern << 8;
    964         pattern |= pattern << 16;
    965 #elif defined(WEBP_USE_MIPS_DSP_R2)
    966         __asm__ volatile ("replv.qb %0, %0" : "+r"(pattern));
    967 #else
    968         pattern = 0x01010101u * pattern;
    969 #endif
    970         break;
    971       case 2:
    972 #if !defined(WORDS_BIGENDIAN)
    973         memcpy(&pattern, src, sizeof(uint16_t));
    974 #else
    975         pattern = ((uint32_t)src[0] << 8) | src[1];
    976 #endif
    977 #if defined(__arm__) || defined(_M_ARM)
    978         pattern |= pattern << 16;
    979 #elif defined(WEBP_USE_MIPS_DSP_R2)
    980         __asm__ volatile ("replv.ph %0, %0" : "+r"(pattern));
    981 #else
    982         pattern = 0x00010001u * pattern;
    983 #endif
    984         break;
    985       case 4:
    986         memcpy(&pattern, src, sizeof(uint32_t));
    987         break;
    988       default:
    989         goto Copy;
    990     }
    991     CopySmallPattern8b(src, dst, length, pattern);
    992     return;
    993   }
    994  Copy:
    995   if (dist >= length) {  // no overlap -> use memcpy()
    996     memcpy(dst, src, length * sizeof(*dst));
    997   } else {
    998     int i;
    999     for (i = 0; i < length; ++i) dst[i] = src[i];
   1000   }
   1001 }
   1002 
   1003 // copy pattern of 1 or 2 uint32_t's
   1004 static WEBP_INLINE void CopySmallPattern32b(const uint32_t* src,
   1005                                             uint32_t* dst,
   1006                                             int length, uint64_t pattern) {
   1007   int i;
   1008   if ((uintptr_t)dst & 4) {           // Align 'dst' to 8-bytes boundary.
   1009     *dst++ = *src++;
   1010     pattern = (pattern >> 32) | (pattern << 32);
   1011     --length;
   1012   }
   1013   assert(0 == ((uintptr_t)dst & 7));
   1014   for (i = 0; i < (length >> 1); ++i) {
   1015     ((uint64_t*)dst)[i] = pattern;    // Copy the pattern 8 bytes at a time.
   1016   }
   1017   if (length & 1) {                   // Finish with left-over.
   1018     dst[i << 1] = src[i << 1];
   1019   }
   1020 }
   1021 
   1022 static WEBP_INLINE void CopyBlock32b(uint32_t* const dst,
   1023                                      int dist, int length) {
   1024   const uint32_t* const src = dst - dist;
   1025   if (dist <= 2 && length >= 4 && ((uintptr_t)dst & 3) == 0) {
   1026     uint64_t pattern;
   1027     if (dist == 1) {
   1028       pattern = (uint64_t)src[0];
   1029       pattern |= pattern << 32;
   1030     } else {
   1031       memcpy(&pattern, src, sizeof(pattern));
   1032     }
   1033     CopySmallPattern32b(src, dst, length, pattern);
   1034   } else if (dist >= length) {  // no overlap
   1035     memcpy(dst, src, length * sizeof(*dst));
   1036   } else {
   1037     int i;
   1038     for (i = 0; i < length; ++i) dst[i] = src[i];
   1039   }
   1040 }
   1041 
   1042 //------------------------------------------------------------------------------
   1043 
   1044 static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
   1045                            int width, int height, int last_row) {
   1046   int ok = 1;
   1047   int row = dec->last_pixel / width;
   1048   int col = dec->last_pixel % width;
   1049   VP8LBitReader* const br = &dec->br;
   1050   VP8LMetadata* const hdr = &dec->hdr;
   1051   int pos = dec->last_pixel;          // current position
   1052   const int end = width * height;     // End of data
   1053   const int last = width * last_row;  // Last pixel to decode
   1054   const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
   1055   const int mask = hdr->huffman_mask;
   1056   const HTreeGroup* htree_group =
   1057       (pos < last) ? GetHtreeGroupForPos(hdr, col, row) : NULL;
   1058   assert(pos <= end);
   1059   assert(last_row <= height);
   1060   assert(Is8bOptimizable(hdr));
   1061 
   1062   while (!br->eos && pos < last) {
   1063     int code;
   1064     // Only update when changing tile.
   1065     if ((col & mask) == 0) {
   1066       htree_group = GetHtreeGroupForPos(hdr, col, row);
   1067     }
   1068     assert(htree_group != NULL);
   1069     VP8LFillBitWindow(br);
   1070     code = ReadSymbol(htree_group->htrees[GREEN], br);
   1071     if (code < NUM_LITERAL_CODES) {  // Literal
   1072       data[pos] = code;
   1073       ++pos;
   1074       ++col;
   1075       if (col >= width) {
   1076         col = 0;
   1077         ++row;
   1078         if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1079           ExtractPalettedAlphaRows(dec, row);
   1080         }
   1081       }
   1082     } else if (code < len_code_limit) {  // Backward reference
   1083       int dist_code, dist;
   1084       const int length_sym = code - NUM_LITERAL_CODES;
   1085       const int length = GetCopyLength(length_sym, br);
   1086       const int dist_symbol = ReadSymbol(htree_group->htrees[DIST], br);
   1087       VP8LFillBitWindow(br);
   1088       dist_code = GetCopyDistance(dist_symbol, br);
   1089       dist = PlaneCodeToDistance(width, dist_code);
   1090       if (pos >= dist && end - pos >= length) {
   1091         CopyBlock8b(data + pos, dist, length);
   1092       } else {
   1093         ok = 0;
   1094         goto End;
   1095       }
   1096       pos += length;
   1097       col += length;
   1098       while (col >= width) {
   1099         col -= width;
   1100         ++row;
   1101         if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1102           ExtractPalettedAlphaRows(dec, row);
   1103         }
   1104       }
   1105       if (pos < last && (col & mask)) {
   1106         htree_group = GetHtreeGroupForPos(hdr, col, row);
   1107       }
   1108     } else {  // Not reached
   1109       ok = 0;
   1110       goto End;
   1111     }
   1112     br->eos = VP8LIsEndOfStream(br);
   1113   }
   1114   // Process the remaining rows corresponding to last row-block.
   1115   ExtractPalettedAlphaRows(dec, row > last_row ? last_row : row);
   1116 
   1117  End:
   1118   br->eos = VP8LIsEndOfStream(br);
   1119   if (!ok || (br->eos && pos < end)) {
   1120     return VP8LSetError(
   1121         dec, br->eos ? VP8_STATUS_SUSPENDED : VP8_STATUS_BITSTREAM_ERROR);
   1122   }
   1123   dec->last_pixel = pos;
   1124   return ok;
   1125 }
   1126 
   1127 static void SaveState(VP8LDecoder* const dec, int last_pixel) {
   1128   assert(dec->incremental);
   1129   dec->saved_br = dec->br;
   1130   dec->saved_last_pixel = last_pixel;
   1131   if (dec->hdr.color_cache_size > 0) {
   1132     VP8LColorCacheCopy(&dec->hdr.color_cache, &dec->hdr.saved_color_cache);
   1133   }
   1134 }
   1135 
   1136 static void RestoreState(VP8LDecoder* const dec) {
   1137   assert(dec->br.eos);
   1138   dec->status = VP8_STATUS_SUSPENDED;
   1139   dec->br = dec->saved_br;
   1140   dec->last_pixel = dec->saved_last_pixel;
   1141   if (dec->hdr.color_cache_size > 0) {
   1142     VP8LColorCacheCopy(&dec->hdr.saved_color_cache, &dec->hdr.color_cache);
   1143   }
   1144 }
   1145 
   1146 #define SYNC_EVERY_N_ROWS 8  // minimum number of rows between check-points
   1147 static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
   1148                            int width, int height, int last_row,
   1149                            ProcessRowsFunc process_func) {
   1150   int row = dec->last_pixel / width;
   1151   int col = dec->last_pixel % width;
   1152   VP8LBitReader* const br = &dec->br;
   1153   VP8LMetadata* const hdr = &dec->hdr;
   1154   uint32_t* src = data + dec->last_pixel;
   1155   uint32_t* last_cached = src;
   1156   uint32_t* const src_end = data + width * height;     // End of data
   1157   uint32_t* const src_last = data + width * last_row;  // Last pixel to decode
   1158   const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
   1159   const int color_cache_limit = len_code_limit + hdr->color_cache_size;
   1160   int next_sync_row = dec->incremental ? row : 1 << 24;
   1161   VP8LColorCache* const color_cache =
   1162       (hdr->color_cache_size > 0) ? &hdr->color_cache : NULL;
   1163   const int mask = hdr->huffman_mask;
   1164   const HTreeGroup* htree_group =
   1165       (src < src_last) ? GetHtreeGroupForPos(hdr, col, row) : NULL;
   1166   assert(dec->last_row < last_row);
   1167   assert(src_last <= src_end);
   1168 
   1169   while (src < src_last) {
   1170     int code;
   1171     if (row >= next_sync_row) {
   1172       SaveState(dec, (int)(src - data));
   1173       next_sync_row = row + SYNC_EVERY_N_ROWS;
   1174     }
   1175     // Only update when changing tile. Note we could use this test:
   1176     // if "((((prev_col ^ col) | prev_row ^ row)) > mask)" -> tile changed
   1177     // but that's actually slower and needs storing the previous col/row.
   1178     if ((col & mask) == 0) {
   1179       htree_group = GetHtreeGroupForPos(hdr, col, row);
   1180     }
   1181     assert(htree_group != NULL);
   1182     if (htree_group->is_trivial_code) {
   1183       *src = htree_group->literal_arb;
   1184       goto AdvanceByOne;
   1185     }
   1186     VP8LFillBitWindow(br);
   1187     if (htree_group->use_packed_table) {
   1188       code = ReadPackedSymbols(htree_group, br, src);
   1189       if (VP8LIsEndOfStream(br)) break;
   1190       if (code == PACKED_NON_LITERAL_CODE) goto AdvanceByOne;
   1191     } else {
   1192       code = ReadSymbol(htree_group->htrees[GREEN], br);
   1193     }
   1194     if (VP8LIsEndOfStream(br)) break;
   1195     if (code < NUM_LITERAL_CODES) {  // Literal
   1196       if (htree_group->is_trivial_literal) {
   1197         *src = htree_group->literal_arb | (code << 8);
   1198       } else {
   1199         int red, blue, alpha;
   1200         red = ReadSymbol(htree_group->htrees[RED], br);
   1201         VP8LFillBitWindow(br);
   1202         blue = ReadSymbol(htree_group->htrees[BLUE], br);
   1203         alpha = ReadSymbol(htree_group->htrees[ALPHA], br);
   1204         if (VP8LIsEndOfStream(br)) break;
   1205         *src = ((uint32_t)alpha << 24) | (red << 16) | (code << 8) | blue;
   1206       }
   1207     AdvanceByOne:
   1208       ++src;
   1209       ++col;
   1210       if (col >= width) {
   1211         col = 0;
   1212         ++row;
   1213         if (process_func != NULL) {
   1214           if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1215             process_func(dec, row);
   1216           }
   1217         }
   1218         if (color_cache != NULL) {
   1219           while (last_cached < src) {
   1220             VP8LColorCacheInsert(color_cache, *last_cached++);
   1221           }
   1222         }
   1223       }
   1224     } else if (code < len_code_limit) {  // Backward reference
   1225       int dist_code, dist;
   1226       const int length_sym = code - NUM_LITERAL_CODES;
   1227       const int length = GetCopyLength(length_sym, br);
   1228       const int dist_symbol = ReadSymbol(htree_group->htrees[DIST], br);
   1229       VP8LFillBitWindow(br);
   1230       dist_code = GetCopyDistance(dist_symbol, br);
   1231       dist = PlaneCodeToDistance(width, dist_code);
   1232 
   1233       if (VP8LIsEndOfStream(br)) break;
   1234       if (src - data < (ptrdiff_t)dist || src_end - src < (ptrdiff_t)length) {
   1235         goto Error;
   1236       } else {
   1237         CopyBlock32b(src, dist, length);
   1238       }
   1239       src += length;
   1240       col += length;
   1241       while (col >= width) {
   1242         col -= width;
   1243         ++row;
   1244         if (process_func != NULL) {
   1245           if (row <= last_row && (row % NUM_ARGB_CACHE_ROWS == 0)) {
   1246             process_func(dec, row);
   1247           }
   1248         }
   1249       }
   1250       // Because of the check done above (before 'src' was incremented by
   1251       // 'length'), the following holds true.
   1252       assert(src <= src_end);
   1253       if (col & mask) htree_group = GetHtreeGroupForPos(hdr, col, row);
   1254       if (color_cache != NULL) {
   1255         while (last_cached < src) {
   1256           VP8LColorCacheInsert(color_cache, *last_cached++);
   1257         }
   1258       }
   1259     } else if (code < color_cache_limit) {  // Color cache
   1260       const int key = code - len_code_limit;
   1261       assert(color_cache != NULL);
   1262       while (last_cached < src) {
   1263         VP8LColorCacheInsert(color_cache, *last_cached++);
   1264       }
   1265       *src = VP8LColorCacheLookup(color_cache, key);
   1266       goto AdvanceByOne;
   1267     } else {  // Not reached
   1268       goto Error;
   1269     }
   1270   }
   1271 
   1272   br->eos = VP8LIsEndOfStream(br);
   1273   // In incremental decoding:
   1274   // br->eos && src < src_last: if 'br' reached the end of the buffer and
   1275   // 'src_last' has not been reached yet, there is not enough data. 'dec' has to
   1276   // be reset until there is more data.
   1277   // !br->eos && src < src_last: this cannot happen as either the buffer is
   1278   // fully read, either enough has been read to reach 'src_last'.
   1279   // src >= src_last: 'src_last' is reached, all is fine. 'src' can actually go
   1280   // beyond 'src_last' in case the image is cropped and an LZ77 goes further.
   1281   // The buffer might have been enough or there is some left. 'br->eos' does
   1282   // not matter.
   1283   assert(!dec->incremental || (br->eos && src < src_last) || src >= src_last);
   1284   if (dec->incremental && br->eos && src < src_last) {
   1285     RestoreState(dec);
   1286   } else if ((dec->incremental && src >= src_last) || !br->eos) {
   1287     // Process the remaining rows corresponding to last row-block.
   1288     if (process_func != NULL) {
   1289       process_func(dec, row > last_row ? last_row : row);
   1290     }
   1291     dec->status = VP8_STATUS_OK;
   1292     dec->last_pixel = (int)(src - data);  // end-of-scan marker
   1293   } else {
   1294     // if not incremental, and we are past the end of buffer (eos=1), then this
   1295     // is a real bitstream error.
   1296     goto Error;
   1297   }
   1298   return 1;
   1299 
   1300  Error:
   1301   return VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
   1302 }
   1303 
   1304 // -----------------------------------------------------------------------------
   1305 // VP8LTransform
   1306 
   1307 static void ClearTransform(VP8LTransform* const transform) {
   1308   WebPSafeFree(transform->data);
   1309   transform->data = NULL;
   1310 }
   1311 
   1312 // For security reason, we need to remap the color map to span
   1313 // the total possible bundled values, and not just the num_colors.
   1314 static int ExpandColorMap(int num_colors, VP8LTransform* const transform) {
   1315   int i;
   1316   const int final_num_colors = 1 << (8 >> transform->bits);
   1317   uint32_t* const new_color_map =
   1318       (uint32_t*)WebPSafeMalloc((uint64_t)final_num_colors,
   1319                                 sizeof(*new_color_map));
   1320   if (new_color_map == NULL) {
   1321     return 0;
   1322   } else {
   1323     uint8_t* const data = (uint8_t*)transform->data;
   1324     uint8_t* const new_data = (uint8_t*)new_color_map;
   1325     new_color_map[0] = transform->data[0];
   1326     for (i = 4; i < 4 * num_colors; ++i) {
   1327       // Equivalent to VP8LAddPixels(), on a byte-basis.
   1328       new_data[i] = (data[i] + new_data[i - 4]) & 0xff;
   1329     }
   1330     for (; i < 4 * final_num_colors; ++i) {
   1331       new_data[i] = 0;  // black tail.
   1332     }
   1333     WebPSafeFree(transform->data);
   1334     transform->data = new_color_map;
   1335   }
   1336   return 1;
   1337 }
   1338 
   1339 static int ReadTransform(int* const xsize, int const* ysize,
   1340                          VP8LDecoder* const dec) {
   1341   int ok = 1;
   1342   VP8LBitReader* const br = &dec->br;
   1343   VP8LTransform* transform = &dec->transforms[dec->next_transform];
   1344   const VP8LImageTransformType type =
   1345       (VP8LImageTransformType)VP8LReadBits(br, 2);
   1346 
   1347   // Each transform type can only be present once in the stream.
   1348   if (dec->transforms_seen & (1U << type)) {
   1349     return 0;  // Already there, let's not accept the second same transform.
   1350   }
   1351   dec->transforms_seen |= (1U << type);
   1352 
   1353   transform->type = type;
   1354   transform->xsize = *xsize;
   1355   transform->ysize = *ysize;
   1356   transform->data = NULL;
   1357   ++dec->next_transform;
   1358   assert(dec->next_transform <= NUM_TRANSFORMS);
   1359 
   1360   switch (type) {
   1361     case PREDICTOR_TRANSFORM:
   1362     case CROSS_COLOR_TRANSFORM:
   1363       transform->bits =
   1364           MIN_TRANSFORM_BITS + VP8LReadBits(br, NUM_TRANSFORM_BITS);
   1365       ok = DecodeImageStream(VP8LSubSampleSize(transform->xsize,
   1366                                                transform->bits),
   1367                              VP8LSubSampleSize(transform->ysize,
   1368                                                transform->bits),
   1369                              /*is_level0=*/0, dec, &transform->data);
   1370       break;
   1371     case COLOR_INDEXING_TRANSFORM: {
   1372        const int num_colors = VP8LReadBits(br, 8) + 1;
   1373        const int bits = (num_colors > 16) ? 0
   1374                       : (num_colors > 4) ? 1
   1375                       : (num_colors > 2) ? 2
   1376                       : 3;
   1377        *xsize = VP8LSubSampleSize(transform->xsize, bits);
   1378        transform->bits = bits;
   1379        ok = DecodeImageStream(num_colors, /*ysize=*/1, /*is_level0=*/0, dec,
   1380                               &transform->data);
   1381        if (ok && !ExpandColorMap(num_colors, transform)) {
   1382          return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
   1383        }
   1384       break;
   1385     }
   1386     case SUBTRACT_GREEN_TRANSFORM:
   1387       break;
   1388     default:
   1389       assert(0);    // can't happen
   1390       break;
   1391   }
   1392 
   1393   return ok;
   1394 }
   1395 
   1396 // -----------------------------------------------------------------------------
   1397 // VP8LMetadata
   1398 
   1399 static void InitMetadata(VP8LMetadata* const hdr) {
   1400   assert(hdr != NULL);
   1401   memset(hdr, 0, sizeof(*hdr));
   1402 }
   1403 
   1404 static void ClearMetadata(VP8LMetadata* const hdr) {
   1405   assert(hdr != NULL);
   1406 
   1407   WebPSafeFree(hdr->huffman_image);
   1408   VP8LHuffmanTablesDeallocate(&hdr->huffman_tables);
   1409   VP8LHtreeGroupsFree(hdr->htree_groups);
   1410   VP8LColorCacheClear(&hdr->color_cache);
   1411   VP8LColorCacheClear(&hdr->saved_color_cache);
   1412   InitMetadata(hdr);
   1413 }
   1414 
   1415 // -----------------------------------------------------------------------------
   1416 // VP8LDecoder
   1417 
   1418 VP8LDecoder* VP8LNew(void) {
   1419   VP8LDecoder* const dec = (VP8LDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
   1420   if (dec == NULL) return NULL;
   1421   dec->status = VP8_STATUS_OK;
   1422   dec->state = READ_DIM;
   1423 
   1424   VP8LDspInit();  // Init critical function pointers.
   1425 
   1426   return dec;
   1427 }
   1428 
   1429 // Resets the decoder in its initial state, reclaiming memory.
   1430 // Preserves the dec->status value.
   1431 static void VP8LClear(VP8LDecoder* const dec) {
   1432   int i;
   1433   if (dec == NULL) return;
   1434   ClearMetadata(&dec->hdr);
   1435 
   1436   WebPSafeFree(dec->pixels);
   1437   dec->pixels = NULL;
   1438   for (i = 0; i < dec->next_transform; ++i) {
   1439     ClearTransform(&dec->transforms[i]);
   1440   }
   1441   dec->next_transform = 0;
   1442   dec->transforms_seen = 0;
   1443 
   1444   WebPSafeFree(dec->rescaler_memory);
   1445   dec->rescaler_memory = NULL;
   1446 
   1447   dec->output = NULL;   // leave no trace behind
   1448 }
   1449 
   1450 void VP8LDelete(VP8LDecoder* const dec) {
   1451   if (dec != NULL) {
   1452     VP8LClear(dec);
   1453     WebPSafeFree(dec);
   1454   }
   1455 }
   1456 
   1457 static void UpdateDecoder(VP8LDecoder* const dec, int width, int height) {
   1458   VP8LMetadata* const hdr = &dec->hdr;
   1459   const int num_bits = hdr->huffman_subsample_bits;
   1460   dec->width = width;
   1461   dec->height = height;
   1462 
   1463   hdr->huffman_xsize = VP8LSubSampleSize(width, num_bits);
   1464   hdr->huffman_mask = (num_bits == 0) ? ~0 : (1 << num_bits) - 1;
   1465 }
   1466 
   1467 static int DecodeImageStream(int xsize, int ysize,
   1468                              int is_level0,
   1469                              VP8LDecoder* const dec,
   1470                              uint32_t** const decoded_data) {
   1471   int ok = 1;
   1472   int transform_xsize = xsize;
   1473   int transform_ysize = ysize;
   1474   VP8LBitReader* const br = &dec->br;
   1475   VP8LMetadata* const hdr = &dec->hdr;
   1476   uint32_t* data = NULL;
   1477   int color_cache_bits = 0;
   1478 
   1479   // Read the transforms (may recurse).
   1480   if (is_level0) {
   1481     while (ok && VP8LReadBits(br, 1)) {
   1482       ok = ReadTransform(&transform_xsize, &transform_ysize, dec);
   1483     }
   1484   }
   1485 
   1486   // Color cache
   1487   if (ok && VP8LReadBits(br, 1)) {
   1488     color_cache_bits = VP8LReadBits(br, 4);
   1489     ok = (color_cache_bits >= 1 && color_cache_bits <= MAX_CACHE_BITS);
   1490     if (!ok) {
   1491       VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
   1492       goto End;
   1493     }
   1494   }
   1495 
   1496   // Read the Huffman codes (may recurse).
   1497   ok = ok && ReadHuffmanCodes(dec, transform_xsize, transform_ysize,
   1498                               color_cache_bits, is_level0);
   1499   if (!ok) {
   1500     VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
   1501     goto End;
   1502   }
   1503 
   1504   // Finish setting up the color-cache
   1505   if (color_cache_bits > 0) {
   1506     hdr->color_cache_size = 1 << color_cache_bits;
   1507     if (!VP8LColorCacheInit(&hdr->color_cache, color_cache_bits)) {
   1508       ok = VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
   1509       goto End;
   1510     }
   1511   } else {
   1512     hdr->color_cache_size = 0;
   1513   }
   1514   UpdateDecoder(dec, transform_xsize, transform_ysize);
   1515 
   1516   if (is_level0) {   // level 0 complete
   1517     dec->state = READ_HDR;
   1518     goto End;
   1519   }
   1520 
   1521   {
   1522     const uint64_t total_size = (uint64_t)transform_xsize * transform_ysize;
   1523     data = (uint32_t*)WebPSafeMalloc(total_size, sizeof(*data));
   1524     if (data == NULL) {
   1525       ok = VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
   1526       goto End;
   1527     }
   1528   }
   1529 
   1530   // Use the Huffman trees to decode the LZ77 encoded data.
   1531   ok = DecodeImageData(dec, data, transform_xsize, transform_ysize,
   1532                        transform_ysize, NULL);
   1533   ok = ok && !br->eos;
   1534 
   1535  End:
   1536   if (!ok) {
   1537     WebPSafeFree(data);
   1538     ClearMetadata(hdr);
   1539   } else {
   1540     if (decoded_data != NULL) {
   1541       *decoded_data = data;
   1542     } else {
   1543       // We allocate image data in this function only for transforms. At level 0
   1544       // (that is: not the transforms), we shouldn't have allocated anything.
   1545       assert(data == NULL);
   1546       assert(is_level0);
   1547     }
   1548     dec->last_pixel = 0;  // Reset for future DECODE_DATA_FUNC() calls.
   1549     if (!is_level0) ClearMetadata(hdr);  // Clean up temporary data behind.
   1550   }
   1551   return ok;
   1552 }
   1553 
   1554 //------------------------------------------------------------------------------
   1555 // Allocate internal buffers dec->pixels and dec->argb_cache.
   1556 static int AllocateInternalBuffers32b(VP8LDecoder* const dec, int final_width) {
   1557   const uint64_t num_pixels = (uint64_t)dec->width * dec->height;
   1558   // Scratch buffer corresponding to top-prediction row for transforming the
   1559   // first row in the row-blocks. Not needed for paletted alpha.
   1560   const uint64_t cache_top_pixels = (uint16_t)final_width;
   1561   // Scratch buffer for temporary BGRA storage. Not needed for paletted alpha.
   1562   const uint64_t cache_pixels = (uint64_t)final_width * NUM_ARGB_CACHE_ROWS;
   1563   const uint64_t total_num_pixels =
   1564       num_pixels + cache_top_pixels + cache_pixels;
   1565 
   1566   assert(dec->width <= final_width);
   1567   dec->pixels = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint32_t));
   1568   if (dec->pixels == NULL) {
   1569     dec->argb_cache = NULL;    // for soundness
   1570     return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
   1571   }
   1572   dec->argb_cache = dec->pixels + num_pixels + cache_top_pixels;
   1573   return 1;
   1574 }
   1575 
   1576 static int AllocateInternalBuffers8b(VP8LDecoder* const dec) {
   1577   const uint64_t total_num_pixels = (uint64_t)dec->width * dec->height;
   1578   dec->argb_cache = NULL;    // for soundness
   1579   dec->pixels = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint8_t));
   1580   if (dec->pixels == NULL) {
   1581     return VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
   1582   }
   1583   return 1;
   1584 }
   1585 
   1586 //------------------------------------------------------------------------------
   1587 
   1588 // Special row-processing that only stores the alpha data.
   1589 static void ExtractAlphaRows(VP8LDecoder* const dec, int last_row) {
   1590   int cur_row = dec->last_row;
   1591   int num_rows = last_row - cur_row;
   1592   const uint32_t* in = dec->pixels + dec->width * cur_row;
   1593 
   1594   assert(last_row <= dec->io->crop_bottom);
   1595   while (num_rows > 0) {
   1596     const int num_rows_to_process =
   1597         (num_rows > NUM_ARGB_CACHE_ROWS) ? NUM_ARGB_CACHE_ROWS : num_rows;
   1598     // Extract alpha (which is stored in the green plane).
   1599     ALPHDecoder* const alph_dec = (ALPHDecoder*)dec->io->opaque;
   1600     uint8_t* const output = alph_dec->output;
   1601     const int width = dec->io->width;      // the final width (!= dec->width)
   1602     const int cache_pixs = width * num_rows_to_process;
   1603     uint8_t* const dst = output + width * cur_row;
   1604     const uint32_t* const src = dec->argb_cache;
   1605     ApplyInverseTransforms(dec, cur_row, num_rows_to_process, in);
   1606     WebPExtractGreen(src, dst, cache_pixs);
   1607     AlphaApplyFilter(alph_dec,
   1608                      cur_row, cur_row + num_rows_to_process, dst, width);
   1609     num_rows -= num_rows_to_process;
   1610     in += num_rows_to_process * dec->width;
   1611     cur_row += num_rows_to_process;
   1612   }
   1613   assert(cur_row == last_row);
   1614   dec->last_row = dec->last_out_row = last_row;
   1615 }
   1616 
   1617 int VP8LDecodeAlphaHeader(ALPHDecoder* const alph_dec,
   1618                           const uint8_t* const data, size_t data_size) {
   1619   int ok = 0;
   1620   VP8LDecoder* dec = VP8LNew();
   1621 
   1622   if (dec == NULL) return 0;
   1623 
   1624   assert(alph_dec != NULL);
   1625 
   1626   dec->width = alph_dec->width;
   1627   dec->height = alph_dec->height;
   1628   dec->io = &alph_dec->io;
   1629   dec->io->opaque = alph_dec;
   1630   dec->io->width = alph_dec->width;
   1631   dec->io->height = alph_dec->height;
   1632 
   1633   dec->status = VP8_STATUS_OK;
   1634   VP8LInitBitReader(&dec->br, data, data_size);
   1635 
   1636   if (!DecodeImageStream(alph_dec->width, alph_dec->height, /*is_level0=*/1,
   1637                          dec, /*decoded_data=*/NULL)) {
   1638     goto Err;
   1639   }
   1640 
   1641   // Special case: if alpha data uses only the color indexing transform and
   1642   // doesn't use color cache (a frequent case), we will use DecodeAlphaData()
   1643   // method that only needs allocation of 1 byte per pixel (alpha channel).
   1644   if (dec->next_transform == 1 &&
   1645       dec->transforms[0].type == COLOR_INDEXING_TRANSFORM &&
   1646       Is8bOptimizable(&dec->hdr)) {
   1647     alph_dec->use_8b_decode = 1;
   1648     ok = AllocateInternalBuffers8b(dec);
   1649   } else {
   1650     // Allocate internal buffers (note that dec->width may have changed here).
   1651     alph_dec->use_8b_decode = 0;
   1652     ok = AllocateInternalBuffers32b(dec, alph_dec->width);
   1653   }
   1654 
   1655   if (!ok) goto Err;
   1656 
   1657   // Only set here, once we are sure it is valid (to avoid thread races).
   1658   alph_dec->vp8l_dec = dec;
   1659   return 1;
   1660 
   1661  Err:
   1662   VP8LDelete(dec);
   1663   return 0;
   1664 }
   1665 
   1666 int VP8LDecodeAlphaImageStream(ALPHDecoder* const alph_dec, int last_row) {
   1667   VP8LDecoder* const dec = alph_dec->vp8l_dec;
   1668   assert(dec != NULL);
   1669   assert(last_row <= dec->height);
   1670 
   1671   if (dec->last_row >= last_row) {
   1672     return 1;  // done
   1673   }
   1674 
   1675   if (!alph_dec->use_8b_decode) WebPInitAlphaProcessing();
   1676 
   1677   // Decode (with special row processing).
   1678   return alph_dec->use_8b_decode ?
   1679       DecodeAlphaData(dec, (uint8_t*)dec->pixels, dec->width, dec->height,
   1680                       last_row) :
   1681       DecodeImageData(dec, dec->pixels, dec->width, dec->height,
   1682                       last_row, ExtractAlphaRows);
   1683 }
   1684 
   1685 //------------------------------------------------------------------------------
   1686 
   1687 int VP8LDecodeHeader(VP8LDecoder* const dec, VP8Io* const io) {
   1688   int width, height, has_alpha;
   1689 
   1690   if (dec == NULL) return 0;
   1691   if (io == NULL) {
   1692     return VP8LSetError(dec, VP8_STATUS_INVALID_PARAM);
   1693   }
   1694 
   1695   dec->io = io;
   1696   dec->status = VP8_STATUS_OK;
   1697   VP8LInitBitReader(&dec->br, io->data, io->data_size);
   1698   if (!ReadImageInfo(&dec->br, &width, &height, &has_alpha)) {
   1699     VP8LSetError(dec, VP8_STATUS_BITSTREAM_ERROR);
   1700     goto Error;
   1701   }
   1702   dec->state = READ_DIM;
   1703   io->width = width;
   1704   io->height = height;
   1705 
   1706   if (!DecodeImageStream(width, height, /*is_level0=*/1, dec,
   1707                          /*decoded_data=*/NULL)) {
   1708     goto Error;
   1709   }
   1710   return 1;
   1711 
   1712  Error:
   1713   VP8LClear(dec);
   1714   assert(dec->status != VP8_STATUS_OK);
   1715   return 0;
   1716 }
   1717 
   1718 int VP8LDecodeImage(VP8LDecoder* const dec) {
   1719   VP8Io* io = NULL;
   1720   WebPDecParams* params = NULL;
   1721 
   1722   if (dec == NULL) return 0;
   1723 
   1724   assert(dec->hdr.huffman_tables.root.start != NULL);
   1725   assert(dec->hdr.htree_groups != NULL);
   1726   assert(dec->hdr.num_htree_groups > 0);
   1727 
   1728   io = dec->io;
   1729   assert(io != NULL);
   1730   params = (WebPDecParams*)io->opaque;
   1731   assert(params != NULL);
   1732 
   1733   // Initialization.
   1734   if (dec->state != READ_DATA) {
   1735     dec->output = params->output;
   1736     assert(dec->output != NULL);
   1737 
   1738     if (!WebPIoInitFromOptions(params->options, io, MODE_BGRA)) {
   1739       VP8LSetError(dec, VP8_STATUS_INVALID_PARAM);
   1740       goto Err;
   1741     }
   1742 
   1743     if (!AllocateInternalBuffers32b(dec, io->width)) goto Err;
   1744 
   1745 #if !defined(WEBP_REDUCE_SIZE)
   1746     if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
   1747 #else
   1748     if (io->use_scaling) {
   1749       VP8LSetError(dec, VP8_STATUS_INVALID_PARAM);
   1750       goto Err;
   1751     }
   1752 #endif
   1753     if (io->use_scaling || WebPIsPremultipliedMode(dec->output->colorspace)) {
   1754       // need the alpha-multiply functions for premultiplied output or rescaling
   1755       WebPInitAlphaProcessing();
   1756     }
   1757 
   1758     if (!WebPIsRGBMode(dec->output->colorspace)) {
   1759       WebPInitConvertARGBToYUV();
   1760       if (dec->output->u.YUVA.a != NULL) WebPInitAlphaProcessing();
   1761     }
   1762     if (dec->incremental) {
   1763       if (dec->hdr.color_cache_size > 0 &&
   1764           dec->hdr.saved_color_cache.colors == NULL) {
   1765         if (!VP8LColorCacheInit(&dec->hdr.saved_color_cache,
   1766                                 dec->hdr.color_cache.hash_bits)) {
   1767           VP8LSetError(dec, VP8_STATUS_OUT_OF_MEMORY);
   1768           goto Err;
   1769         }
   1770       }
   1771     }
   1772     dec->state = READ_DATA;
   1773   }
   1774 
   1775   // Decode.
   1776   if (!DecodeImageData(dec, dec->pixels, dec->width, dec->height,
   1777                        io->crop_bottom, ProcessRows)) {
   1778     goto Err;
   1779   }
   1780 
   1781   params->last_y = dec->last_out_row;
   1782   return 1;
   1783 
   1784  Err:
   1785   VP8LClear(dec);
   1786   assert(dec->status != VP8_STATUS_OK);
   1787   return 0;
   1788 }
   1789 
   1790 //------------------------------------------------------------------------------