go-libwebp

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

vp8_dec.c (22777B)

      1 // Copyright 2010 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 // Author: Skal (pascal.massimino@gmail.com)
     13 
     14 #include <assert.h>
     15 #include <stdlib.h>
     16 #include <string.h>
     17 
     18 #include "src/dec/alphai_dec.h"
     19 #include "src/dec/common_dec.h"
     20 #include "src/dec/vp8_dec.h"
     21 #include "src/dec/vp8i_dec.h"
     22 #include "src/dec/vp8li_dec.h"
     23 #include "src/dec/webpi_dec.h"
     24 #include "src/dsp/cpu.h"
     25 #include "src/dsp/dsp.h"
     26 #include "src/utils/bit_reader_inl_utils.h"
     27 #include "src/utils/bit_reader_utils.h"
     28 #include "src/utils/thread_utils.h"
     29 #include "src/utils/utils.h"
     30 #include "src/webp/decode.h"
     31 #include "src/webp/format_constants.h"
     32 #include "src/webp/types.h"
     33 
     34 //------------------------------------------------------------------------------
     35 
     36 int WebPGetDecoderVersion(void) {
     37   return (DEC_MAJ_VERSION << 16) | (DEC_MIN_VERSION << 8) | DEC_REV_VERSION;
     38 }
     39 
     40 //------------------------------------------------------------------------------
     41 // Signature and pointer-to-function for GetCoeffs() variants below.
     42 
     43 typedef int (*GetCoeffsFunc)(VP8BitReader* const br,
     44                              const VP8BandProbas* const prob[],
     45                              int ctx, const quant_t dq, int n, int16_t* out);
     46 static volatile GetCoeffsFunc GetCoeffs = NULL;
     47 
     48 static void InitGetCoeffs(void);
     49 
     50 //------------------------------------------------------------------------------
     51 // VP8Decoder
     52 
     53 static void SetOk(VP8Decoder* const dec) {
     54   dec->status = VP8_STATUS_OK;
     55   dec->error_msg = "OK";
     56 }
     57 
     58 int VP8InitIoInternal(VP8Io* const io, int version) {
     59   if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DECODER_ABI_VERSION)) {
     60     return 0;  // mismatch error
     61   }
     62   if (io != NULL) {
     63     memset(io, 0, sizeof(*io));
     64   }
     65   return 1;
     66 }
     67 
     68 VP8Decoder* VP8New(void) {
     69   VP8Decoder* const dec = (VP8Decoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
     70   if (dec != NULL) {
     71     SetOk(dec);
     72     WebPGetWorkerInterface()->Init(&dec->worker);
     73     dec->ready = 0;
     74     dec->num_parts_minus_one = 0;
     75     InitGetCoeffs();
     76   }
     77   return dec;
     78 }
     79 
     80 VP8StatusCode VP8Status(VP8Decoder* const dec) {
     81   if (!dec) return VP8_STATUS_INVALID_PARAM;
     82   return dec->status;
     83 }
     84 
     85 const char* VP8StatusMessage(VP8Decoder* const dec) {
     86   if (dec == NULL) return "no object";
     87   if (!dec->error_msg) return "OK";
     88   return dec->error_msg;
     89 }
     90 
     91 void VP8Delete(VP8Decoder* const dec) {
     92   if (dec != NULL) {
     93     VP8Clear(dec);
     94     WebPSafeFree(dec);
     95   }
     96 }
     97 
     98 int VP8SetError(VP8Decoder* const dec,
     99                 VP8StatusCode error, const char* const msg) {
    100   // VP8_STATUS_SUSPENDED is only meaningful in incremental decoding.
    101   assert(dec->incremental || error != VP8_STATUS_SUSPENDED);
    102   // The oldest error reported takes precedence over the new one.
    103   if (dec->status == VP8_STATUS_OK) {
    104     dec->status = error;
    105     dec->error_msg = msg;
    106     dec->ready = 0;
    107   }
    108   return 0;
    109 }
    110 
    111 //------------------------------------------------------------------------------
    112 
    113 int VP8CheckSignature(const uint8_t* const data, size_t data_size) {
    114   return (data_size >= 3 &&
    115           data[0] == 0x9d && data[1] == 0x01 && data[2] == 0x2a);
    116 }
    117 
    118 int VP8GetInfo(const uint8_t* data, size_t data_size, size_t chunk_size,
    119                int* const width, int* const height) {
    120   if (data == NULL || data_size < VP8_FRAME_HEADER_SIZE) {
    121     return 0;         // not enough data
    122   }
    123   // check signature
    124   if (!VP8CheckSignature(data + 3, data_size - 3)) {
    125     return 0;         // Wrong signature.
    126   } else {
    127     const uint32_t bits = data[0] | (data[1] << 8) | (data[2] << 16);
    128     const int key_frame = !(bits & 1);
    129     const int w = ((data[7] << 8) | data[6]) & 0x3fff;
    130     const int h = ((data[9] << 8) | data[8]) & 0x3fff;
    131 
    132     if (!key_frame) {   // Not a keyframe.
    133       return 0;
    134     }
    135 
    136     if (((bits >> 1) & 7) > 3) {
    137       return 0;         // unknown profile
    138     }
    139     if (!((bits >> 4) & 1)) {
    140       return 0;         // first frame is invisible!
    141     }
    142     if (((bits >> 5)) >= chunk_size) {  // partition_length
    143       return 0;         // inconsistent size information.
    144     }
    145     if (w == 0 || h == 0) {
    146       return 0;         // We don't support both width and height to be zero.
    147     }
    148 
    149     if (width) {
    150       *width = w;
    151     }
    152     if (height) {
    153       *height = h;
    154     }
    155 
    156     return 1;
    157   }
    158 }
    159 
    160 //------------------------------------------------------------------------------
    161 // Header parsing
    162 
    163 static void ResetSegmentHeader(VP8SegmentHeader* const hdr) {
    164   assert(hdr != NULL);
    165   hdr->use_segment = 0;
    166   hdr->update_map = 0;
    167   hdr->absolute_delta = 1;
    168   memset(hdr->quantizer, 0, sizeof(hdr->quantizer));
    169   memset(hdr->filter_strength, 0, sizeof(hdr->filter_strength));
    170 }
    171 
    172 // Paragraph 9.3
    173 static int ParseSegmentHeader(VP8BitReader* br,
    174                               VP8SegmentHeader* hdr, VP8Proba* proba) {
    175   assert(br != NULL);
    176   assert(hdr != NULL);
    177   hdr->use_segment = VP8Get(br, "global-header");
    178   if (hdr->use_segment) {
    179     hdr->update_map = VP8Get(br, "global-header");
    180     if (VP8Get(br, "global-header")) {   // update data
    181       int s;
    182       hdr->absolute_delta = VP8Get(br, "global-header");
    183       for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
    184         hdr->quantizer[s] = VP8Get(br, "global-header") ?
    185             VP8GetSignedValue(br, 7, "global-header") : 0;
    186       }
    187       for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
    188         hdr->filter_strength[s] = VP8Get(br, "global-header") ?
    189             VP8GetSignedValue(br, 6, "global-header") : 0;
    190       }
    191     }
    192     if (hdr->update_map) {
    193       int s;
    194       for (s = 0; s < MB_FEATURE_TREE_PROBS; ++s) {
    195         proba->segments[s] = VP8Get(br, "global-header") ?
    196             VP8GetValue(br, 8, "global-header") : 255u;
    197       }
    198     }
    199   } else {
    200     hdr->update_map = 0;
    201   }
    202   return !br->eof;
    203 }
    204 
    205 // Paragraph 9.5
    206 // If we don't have all the necessary data in 'buf', this function returns
    207 // VP8_STATUS_SUSPENDED in incremental decoding, VP8_STATUS_NOT_ENOUGH_DATA
    208 // otherwise.
    209 // In incremental decoding, this case is not necessarily an error. Still, no
    210 // bitreader is ever initialized to make it possible to read unavailable memory.
    211 // If we don't even have the partitions' sizes, then VP8_STATUS_NOT_ENOUGH_DATA
    212 // is returned, and this is an unrecoverable error.
    213 // If the partitions were positioned ok, VP8_STATUS_OK is returned.
    214 static VP8StatusCode ParsePartitions(VP8Decoder* const dec,
    215                                      const uint8_t* buf, size_t size) {
    216   VP8BitReader* const br = &dec->br;
    217   const uint8_t* sz = buf;
    218   const uint8_t* buf_end = buf + size;
    219   const uint8_t* part_start;
    220   size_t size_left = size;
    221   size_t last_part;
    222   size_t p;
    223 
    224   dec->num_parts_minus_one = (1 << VP8GetValue(br, 2, "global-header")) - 1;
    225   last_part = dec->num_parts_minus_one;
    226   if (size < 3 * last_part) {
    227     // we can't even read the sizes with sz[]! That's a failure.
    228     return VP8_STATUS_NOT_ENOUGH_DATA;
    229   }
    230   part_start = buf + last_part * 3;
    231   size_left -= last_part * 3;
    232   for (p = 0; p < last_part; ++p) {
    233     size_t psize = sz[0] | (sz[1] << 8) | (sz[2] << 16);
    234     if (psize > size_left) psize = size_left;
    235     VP8InitBitReader(dec->parts + p, part_start, psize);
    236     part_start += psize;
    237     size_left -= psize;
    238     sz += 3;
    239   }
    240   VP8InitBitReader(dec->parts + last_part, part_start, size_left);
    241   if (part_start < buf_end) return VP8_STATUS_OK;
    242   return dec->incremental
    243              ? VP8_STATUS_SUSPENDED  // Init is ok, but there's not enough data
    244              : VP8_STATUS_NOT_ENOUGH_DATA;
    245 }
    246 
    247 // Paragraph 9.4
    248 static int ParseFilterHeader(VP8BitReader* br, VP8Decoder* const dec) {
    249   VP8FilterHeader* const hdr = &dec->filter_hdr;
    250   hdr->simple    = VP8Get(br, "global-header");
    251   hdr->level     = VP8GetValue(br, 6, "global-header");
    252   hdr->sharpness = VP8GetValue(br, 3, "global-header");
    253   hdr->use_lf_delta = VP8Get(br, "global-header");
    254   if (hdr->use_lf_delta) {
    255     if (VP8Get(br, "global-header")) {   // update lf-delta?
    256       int i;
    257       for (i = 0; i < NUM_REF_LF_DELTAS; ++i) {
    258         if (VP8Get(br, "global-header")) {
    259           hdr->ref_lf_delta[i] = VP8GetSignedValue(br, 6, "global-header");
    260         }
    261       }
    262       for (i = 0; i < NUM_MODE_LF_DELTAS; ++i) {
    263         if (VP8Get(br, "global-header")) {
    264           hdr->mode_lf_delta[i] = VP8GetSignedValue(br, 6, "global-header");
    265         }
    266       }
    267     }
    268   }
    269   dec->filter_type = (hdr->level == 0) ? 0 : hdr->simple ? 1 : 2;
    270   return !br->eof;
    271 }
    272 
    273 // Topmost call
    274 int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
    275   const uint8_t* buf;
    276   size_t buf_size;
    277   VP8FrameHeader* frm_hdr;
    278   VP8PictureHeader* pic_hdr;
    279   VP8BitReader* br;
    280   VP8StatusCode status;
    281 
    282   if (dec == NULL) {
    283     return 0;
    284   }
    285   SetOk(dec);
    286   if (io == NULL) {
    287     return VP8SetError(dec, VP8_STATUS_INVALID_PARAM,
    288                        "null VP8Io passed to VP8GetHeaders()");
    289   }
    290   buf = io->data;
    291   buf_size = io->data_size;
    292   if (buf_size < 4) {
    293     return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
    294                        "Truncated header.");
    295   }
    296 
    297   // Paragraph 9.1
    298   {
    299     const uint32_t bits = buf[0] | (buf[1] << 8) | (buf[2] << 16);
    300     frm_hdr = &dec->frm_hdr;
    301     frm_hdr->key_frame = !(bits & 1);
    302     frm_hdr->profile = (bits >> 1) & 7;
    303     frm_hdr->show = (bits >> 4) & 1;
    304     frm_hdr->partition_length = (bits >> 5);
    305     if (frm_hdr->profile > 3) {
    306       return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
    307                          "Incorrect keyframe parameters.");
    308     }
    309     if (!frm_hdr->show) {
    310       return VP8SetError(dec, VP8_STATUS_UNSUPPORTED_FEATURE,
    311                          "Frame not displayable.");
    312     }
    313     buf += 3;
    314     buf_size -= 3;
    315   }
    316 
    317   pic_hdr = &dec->pic_hdr;
    318   if (frm_hdr->key_frame) {
    319     // Paragraph 9.2
    320     if (buf_size < 7) {
    321       return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
    322                          "cannot parse picture header");
    323     }
    324     if (!VP8CheckSignature(buf, buf_size)) {
    325       return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
    326                          "Bad code word");
    327     }
    328     pic_hdr->width = ((buf[4] << 8) | buf[3]) & 0x3fff;
    329     pic_hdr->xscale = buf[4] >> 6;   // ratio: 1, 5/4 5/3 or 2
    330     pic_hdr->height = ((buf[6] << 8) | buf[5]) & 0x3fff;
    331     pic_hdr->yscale = buf[6] >> 6;
    332     buf += 7;
    333     buf_size -= 7;
    334 
    335     dec->mb_w = (pic_hdr->width + 15) >> 4;
    336     dec->mb_h = (pic_hdr->height + 15) >> 4;
    337 
    338     // Setup default output area (can be later modified during io->setup())
    339     io->width = pic_hdr->width;
    340     io->height = pic_hdr->height;
    341     // IMPORTANT! use some sane dimensions in crop* and scaled* fields.
    342     // So they can be used interchangeably without always testing for
    343     // 'use_cropping'.
    344     io->use_cropping = 0;
    345     io->crop_top  = 0;
    346     io->crop_left = 0;
    347     io->crop_right  = io->width;
    348     io->crop_bottom = io->height;
    349     io->use_scaling  = 0;
    350     io->scaled_width = io->width;
    351     io->scaled_height = io->height;
    352 
    353     io->mb_w = io->width;   // for soundness
    354     io->mb_h = io->height;  // ditto
    355 
    356     VP8ResetProba(&dec->proba);
    357     ResetSegmentHeader(&dec->segment_hdr);
    358   }
    359 
    360   // Check if we have all the partition #0 available, and initialize dec->br
    361   // to read this partition (and this partition only).
    362   if (frm_hdr->partition_length > buf_size) {
    363     return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
    364                        "bad partition length");
    365   }
    366 
    367   br = &dec->br;
    368   VP8InitBitReader(br, buf, frm_hdr->partition_length);
    369   buf += frm_hdr->partition_length;
    370   buf_size -= frm_hdr->partition_length;
    371 
    372   if (frm_hdr->key_frame) {
    373     pic_hdr->colorspace = VP8Get(br, "global-header");
    374     pic_hdr->clamp_type = VP8Get(br, "global-header");
    375   }
    376   if (!ParseSegmentHeader(br, &dec->segment_hdr, &dec->proba)) {
    377     return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
    378                        "cannot parse segment header");
    379   }
    380   // Filter specs
    381   if (!ParseFilterHeader(br, dec)) {
    382     return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
    383                        "cannot parse filter header");
    384   }
    385   status = ParsePartitions(dec, buf, buf_size);
    386   if (status != VP8_STATUS_OK) {
    387     return VP8SetError(dec, status, "cannot parse partitions");
    388   }
    389 
    390   // quantizer change
    391   VP8ParseQuant(dec);
    392 
    393   // Frame buffer marking
    394   if (!frm_hdr->key_frame) {
    395     return VP8SetError(dec, VP8_STATUS_UNSUPPORTED_FEATURE,
    396                        "Not a key frame.");
    397   }
    398 
    399   VP8Get(br, "global-header");   // ignore the value of 'update_proba'
    400 
    401   VP8ParseProba(br, dec);
    402 
    403   // sanitized state
    404   dec->ready = 1;
    405   return 1;
    406 }
    407 
    408 //------------------------------------------------------------------------------
    409 // Residual decoding (Paragraph 13.2 / 13.3)
    410 
    411 static const uint8_t kCat3[] = { 173, 148, 140, 0 };
    412 static const uint8_t kCat4[] = { 176, 155, 140, 135, 0 };
    413 static const uint8_t kCat5[] = { 180, 157, 141, 134, 130, 0 };
    414 static const uint8_t kCat6[] =
    415   { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129, 0 };
    416 static const uint8_t* const kCat3456[] = { kCat3, kCat4, kCat5, kCat6 };
    417 static const uint8_t kZigzag[16] = {
    418   0, 1, 4, 8,  5, 2, 3, 6,  9, 12, 13, 10,  7, 11, 14, 15
    419 };
    420 
    421 // See section 13-2: https://datatracker.ietf.org/doc/html/rfc6386#section-13.2
    422 static int GetLargeValue(VP8BitReader* const br, const uint8_t* const p) {
    423   int v;
    424   if (!VP8GetBit(br, p[3], "coeffs")) {
    425     if (!VP8GetBit(br, p[4], "coeffs")) {
    426       v = 2;
    427     } else {
    428       v = 3 + VP8GetBit(br, p[5], "coeffs");
    429     }
    430   } else {
    431     if (!VP8GetBit(br, p[6], "coeffs")) {
    432       if (!VP8GetBit(br, p[7], "coeffs")) {
    433         v = 5 + VP8GetBit(br, 159, "coeffs");
    434       } else {
    435         v = 7 + 2 * VP8GetBit(br, 165, "coeffs");
    436         v += VP8GetBit(br, 145, "coeffs");
    437       }
    438     } else {
    439       const uint8_t* tab;
    440       const int bit1 = VP8GetBit(br, p[8], "coeffs");
    441       const int bit0 = VP8GetBit(br, p[9 + bit1], "coeffs");
    442       const int cat = 2 * bit1 + bit0;
    443       v = 0;
    444       for (tab = kCat3456[cat]; *tab; ++tab) {
    445         v += v + VP8GetBit(br, *tab, "coeffs");
    446       }
    447       v += 3 + (8 << cat);
    448     }
    449   }
    450   return v;
    451 }
    452 
    453 // Returns the position of the last non-zero coeff plus one
    454 static int GetCoeffsFast(VP8BitReader* const br,
    455                          const VP8BandProbas* const prob[],
    456                          int ctx, const quant_t dq, int n, int16_t* out) {
    457   const uint8_t* p = prob[n]->probas[ctx];
    458   for (; n < 16; ++n) {
    459     if (!VP8GetBit(br, p[0], "coeffs")) {
    460       return n;  // previous coeff was last non-zero coeff
    461     }
    462     while (!VP8GetBit(br, p[1], "coeffs")) {       // sequence of zero coeffs
    463       p = prob[++n]->probas[0];
    464       if (n == 16) return 16;
    465     }
    466     {        // non zero coeff
    467       const VP8ProbaArray* const p_ctx = &prob[n + 1]->probas[0];
    468       int v;
    469       if (!VP8GetBit(br, p[2], "coeffs")) {
    470         v = 1;
    471         p = p_ctx[1];
    472       } else {
    473         v = GetLargeValue(br, p);
    474         p = p_ctx[2];
    475       }
    476       out[kZigzag[n]] = VP8GetSigned(br, v, "coeffs") * dq[n > 0];
    477     }
    478   }
    479   return 16;
    480 }
    481 
    482 // This version of GetCoeffs() uses VP8GetBitAlt() which is an alternate version
    483 // of VP8GetBitAlt() targeting specific platforms.
    484 static int GetCoeffsAlt(VP8BitReader* const br,
    485                         const VP8BandProbas* const prob[],
    486                         int ctx, const quant_t dq, int n, int16_t* out) {
    487   const uint8_t* p = prob[n]->probas[ctx];
    488   for (; n < 16; ++n) {
    489     if (!VP8GetBitAlt(br, p[0], "coeffs")) {
    490       return n;  // previous coeff was last non-zero coeff
    491     }
    492     while (!VP8GetBitAlt(br, p[1], "coeffs")) {       // sequence of zero coeffs
    493       p = prob[++n]->probas[0];
    494       if (n == 16) return 16;
    495     }
    496     {        // non zero coeff
    497       const VP8ProbaArray* const p_ctx = &prob[n + 1]->probas[0];
    498       int v;
    499       if (!VP8GetBitAlt(br, p[2], "coeffs")) {
    500         v = 1;
    501         p = p_ctx[1];
    502       } else {
    503         v = GetLargeValue(br, p);
    504         p = p_ctx[2];
    505       }
    506       out[kZigzag[n]] = VP8GetSigned(br, v, "coeffs") * dq[n > 0];
    507     }
    508   }
    509   return 16;
    510 }
    511 
    512 extern VP8CPUInfo VP8GetCPUInfo;
    513 
    514 WEBP_DSP_INIT_FUNC(InitGetCoeffs) {
    515   if (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kSlowSSSE3)) {
    516     GetCoeffs = GetCoeffsAlt;
    517   } else {
    518     GetCoeffs = GetCoeffsFast;
    519   }
    520 }
    521 
    522 static WEBP_INLINE uint32_t NzCodeBits(uint32_t nz_coeffs, int nz, int dc_nz) {
    523   nz_coeffs <<= 2;
    524   nz_coeffs |= (nz > 3) ? 3 : (nz > 1) ? 2 : dc_nz;
    525   return nz_coeffs;
    526 }
    527 
    528 static int ParseResiduals(VP8Decoder* const dec,
    529                           VP8MB* const mb, VP8BitReader* const token_br) {
    530   const VP8BandProbas* (* const bands)[16 + 1] = dec->proba.bands_ptr;
    531   const VP8BandProbas* const * ac_proba;
    532   VP8MBData* const block = dec->mb_data + dec->mb_x;
    533   const VP8QuantMatrix* const q = &dec->dqm[block->segment];
    534   int16_t* dst = block->coeffs;
    535   VP8MB* const left_mb = dec->mb_info - 1;
    536   uint8_t tnz, lnz;
    537   uint32_t non_zero_y = 0;
    538   uint32_t non_zero_uv = 0;
    539   int x, y, ch;
    540   uint32_t out_t_nz, out_l_nz;
    541   int first;
    542 
    543   memset(dst, 0, 384 * sizeof(*dst));
    544   if (!block->is_i4x4) {    // parse DC
    545     int16_t dc[16] = { 0 };
    546     const int ctx = mb->nz_dc + left_mb->nz_dc;
    547     const int nz = GetCoeffs(token_br, bands[1], ctx, q->y2_mat, 0, dc);
    548     mb->nz_dc = left_mb->nz_dc = (nz > 0);
    549     if (nz > 1) {   // more than just the DC -> perform the full transform
    550       VP8TransformWHT(dc, dst);
    551     } else {        // only DC is non-zero -> inlined simplified transform
    552       int i;
    553       const int dc0 = (dc[0] + 3) >> 3;
    554       for (i = 0; i < 16 * 16; i += 16) dst[i] = dc0;
    555     }
    556     first = 1;
    557     ac_proba = bands[0];
    558   } else {
    559     first = 0;
    560     ac_proba = bands[3];
    561   }
    562 
    563   tnz = mb->nz & 0x0f;
    564   lnz = left_mb->nz & 0x0f;
    565   for (y = 0; y < 4; ++y) {
    566     int l = lnz & 1;
    567     uint32_t nz_coeffs = 0;
    568     for (x = 0; x < 4; ++x) {
    569       const int ctx = l + (tnz & 1);
    570       const int nz = GetCoeffs(token_br, ac_proba, ctx, q->y1_mat, first, dst);
    571       l = (nz > first);
    572       tnz = (tnz >> 1) | (l << 7);
    573       nz_coeffs = NzCodeBits(nz_coeffs, nz, dst[0] != 0);
    574       dst += 16;
    575     }
    576     tnz >>= 4;
    577     lnz = (lnz >> 1) | (l << 7);
    578     non_zero_y = (non_zero_y << 8) | nz_coeffs;
    579   }
    580   out_t_nz = tnz;
    581   out_l_nz = lnz >> 4;
    582 
    583   for (ch = 0; ch < 4; ch += 2) {
    584     uint32_t nz_coeffs = 0;
    585     tnz = mb->nz >> (4 + ch);
    586     lnz = left_mb->nz >> (4 + ch);
    587     for (y = 0; y < 2; ++y) {
    588       int l = lnz & 1;
    589       for (x = 0; x < 2; ++x) {
    590         const int ctx = l + (tnz & 1);
    591         const int nz = GetCoeffs(token_br, bands[2], ctx, q->uv_mat, 0, dst);
    592         l = (nz > 0);
    593         tnz = (tnz >> 1) | (l << 3);
    594         nz_coeffs = NzCodeBits(nz_coeffs, nz, dst[0] != 0);
    595         dst += 16;
    596       }
    597       tnz >>= 2;
    598       lnz = (lnz >> 1) | (l << 5);
    599     }
    600     // Note: we don't really need the per-4x4 details for U/V blocks.
    601     non_zero_uv |= nz_coeffs << (4 * ch);
    602     out_t_nz |= (tnz << 4) << ch;
    603     out_l_nz |= (lnz & 0xf0) << ch;
    604   }
    605   mb->nz = out_t_nz;
    606   left_mb->nz = out_l_nz;
    607 
    608   block->non_zero_y = non_zero_y;
    609   block->non_zero_uv = non_zero_uv;
    610 
    611   // We look at the mode-code of each block and check if some blocks have less
    612   // than three non-zero coeffs (code < 2). This is to avoid dithering flat and
    613   // empty blocks.
    614   block->dither = (non_zero_uv & 0xaaaa) ? 0 : q->dither;
    615 
    616   return !(non_zero_y | non_zero_uv);  // will be used for further optimization
    617 }
    618 
    619 //------------------------------------------------------------------------------
    620 // Main loop
    621 
    622 int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
    623   VP8MB* const left = dec->mb_info - 1;
    624   VP8MB* const mb = dec->mb_info + dec->mb_x;
    625   VP8MBData* const block = dec->mb_data + dec->mb_x;
    626   int skip = dec->use_skip_proba ? block->skip : 0;
    627 
    628   if (!skip) {
    629     skip = ParseResiduals(dec, mb, token_br);
    630   } else {
    631     left->nz = mb->nz = 0;
    632     if (!block->is_i4x4) {
    633       left->nz_dc = mb->nz_dc = 0;
    634     }
    635     block->non_zero_y = 0;
    636     block->non_zero_uv = 0;
    637     block->dither = 0;
    638   }
    639 
    640   if (dec->filter_type > 0) {  // store filter info
    641     VP8FInfo* const finfo = dec->f_info + dec->mb_x;
    642     *finfo = dec->fstrengths[block->segment][block->is_i4x4];
    643     finfo->f_inner |= !skip;
    644   }
    645 
    646   return !token_br->eof;
    647 }
    648 
    649 void VP8InitScanline(VP8Decoder* const dec) {
    650   VP8MB* const left = dec->mb_info - 1;
    651   left->nz = 0;
    652   left->nz_dc = 0;
    653   memset(dec->intra_l, B_DC_PRED, sizeof(dec->intra_l));
    654   dec->mb_x = 0;
    655 }
    656 
    657 static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
    658   for (dec->mb_y = 0; dec->mb_y < dec->br_mb_y; ++dec->mb_y) {
    659     // Parse bitstream for this row.
    660     VP8BitReader* const token_br =
    661         &dec->parts[dec->mb_y & dec->num_parts_minus_one];
    662     if (!VP8ParseIntraModeRow(&dec->br, dec)) {
    663       return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
    664                          "Premature end-of-partition0 encountered.");
    665     }
    666     for (; dec->mb_x < dec->mb_w; ++dec->mb_x) {
    667       if (!VP8DecodeMB(dec, token_br)) {
    668         return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
    669                            "Premature end-of-file encountered.");
    670       }
    671     }
    672     VP8InitScanline(dec);   // Prepare for next scanline
    673 
    674     // Reconstruct, filter and emit the row.
    675     if (!VP8ProcessRow(dec, io)) {
    676       return VP8SetError(dec, VP8_STATUS_USER_ABORT, "Output aborted.");
    677     }
    678   }
    679   if (dec->mt_method > 0) {
    680     if (!WebPGetWorkerInterface()->Sync(&dec->worker)) return 0;
    681   }
    682 
    683   return 1;
    684 }
    685 
    686 // Main entry point
    687 int VP8Decode(VP8Decoder* const dec, VP8Io* const io) {
    688   int ok = 0;
    689   if (dec == NULL) {
    690     return 0;
    691   }
    692   if (io == NULL) {
    693     return VP8SetError(dec, VP8_STATUS_INVALID_PARAM,
    694                        "NULL VP8Io parameter in VP8Decode().");
    695   }
    696 
    697   if (!dec->ready) {
    698     if (!VP8GetHeaders(dec, io)) {
    699       return 0;
    700     }
    701   }
    702   assert(dec->ready);
    703 
    704   // Finish setting up the decoding parameter. Will call io->setup().
    705   ok = (VP8EnterCritical(dec, io) == VP8_STATUS_OK);
    706   if (ok) {   // good to go.
    707     // Will allocate memory and prepare everything.
    708     if (ok) ok = VP8InitFrame(dec, io);
    709 
    710     // Main decoding loop
    711     if (ok) ok = ParseFrame(dec, io);
    712 
    713     // Exit.
    714     ok &= VP8ExitCritical(dec, io);
    715   }
    716 
    717   if (!ok) {
    718     VP8Clear(dec);
    719     return 0;
    720   }
    721 
    722   dec->ready = 0;
    723   return ok;
    724 }
    725 
    726 void VP8Clear(VP8Decoder* const dec) {
    727   if (dec == NULL) {
    728     return;
    729   }
    730   WebPGetWorkerInterface()->End(&dec->worker);
    731   WebPDeallocateAlphaMemory(dec);
    732   WebPSafeFree(dec->mem);
    733   dec->mem = NULL;
    734   dec->mem_size = 0;
    735   memset(&dec->br, 0, sizeof(dec->br));
    736   dec->ready = 0;
    737 }
    738 
    739 //------------------------------------------------------------------------------