Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright (C) 2007 Marco Gerards <marco@gnu.org>
  3.  * Copyright (C) 2009 David Conrad
  4.  * Copyright (C) 2011 Jordi Ortiz
  5.  *
  6.  * This file is part of FFmpeg.
  7.  *
  8.  * FFmpeg is free software; you can redistribute it and/or
  9.  * modify it under the terms of the GNU Lesser General Public
  10.  * License as published by the Free Software Foundation; either
  11.  * version 2.1 of the License, or (at your option) any later version.
  12.  *
  13.  * FFmpeg is distributed in the hope that it will be useful,
  14.  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16.  * Lesser General Public License for more details.
  17.  *
  18.  * You should have received a copy of the GNU Lesser General Public
  19.  * License along with FFmpeg; if not, write to the Free Software
  20.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21.  */
  22.  
  23. /**
  24.  * @file
  25.  * Dirac Decoder
  26.  * @author Marco Gerards <marco@gnu.org>, David Conrad, Jordi Ortiz <nenjordi@gmail.com>
  27.  */
  28.  
  29. #include "avcodec.h"
  30. #include "dsputil.h"
  31. #include "get_bits.h"
  32. #include "bytestream.h"
  33. #include "internal.h"
  34. #include "golomb.h"
  35. #include "dirac_arith.h"
  36. #include "mpeg12data.h"
  37. #include "dirac_dwt.h"
  38. #include "dirac.h"
  39. #include "diracdsp.h"
  40. #include "videodsp.h" // for ff_emulated_edge_mc_8
  41.  
  42. /**
  43.  * The spec limits the number of wavelet decompositions to 4 for both
  44.  * level 1 (VC-2) and 128 (long-gop default).
  45.  * 5 decompositions is the maximum before >16-bit buffers are needed.
  46.  * Schroedinger allows this for DD 9,7 and 13,7 wavelets only, limiting
  47.  * the others to 4 decompositions (or 3 for the fidelity filter).
  48.  *
  49.  * We use this instead of MAX_DECOMPOSITIONS to save some memory.
  50.  */
  51. #define MAX_DWT_LEVELS 5
  52.  
  53. /**
  54.  * The spec limits this to 3 for frame coding, but in practice can be as high as 6
  55.  */
  56. #define MAX_REFERENCE_FRAMES 8
  57. #define MAX_DELAY 5         /* limit for main profile for frame coding (TODO: field coding) */
  58. #define MAX_FRAMES (MAX_REFERENCE_FRAMES + MAX_DELAY + 1)
  59. #define MAX_QUANT 68        /* max quant for VC-2 */
  60. #define MAX_BLOCKSIZE 32    /* maximum xblen/yblen we support */
  61.  
  62. /**
  63.  * DiracBlock->ref flags, if set then the block does MC from the given ref
  64.  */
  65. #define DIRAC_REF_MASK_REF1   1
  66. #define DIRAC_REF_MASK_REF2   2
  67. #define DIRAC_REF_MASK_GLOBAL 4
  68.  
  69. /**
  70.  * Value of Picture.reference when Picture is not a reference picture, but
  71.  * is held for delayed output.
  72.  */
  73. #define DELAYED_PIC_REF 4
  74.  
  75. #define ff_emulated_edge_mc ff_emulated_edge_mc_8 /* Fix: change the calls to this function regarding bit depth */
  76.  
  77. #define CALC_PADDING(size, depth)                       \
  78.     (((size + (1 << depth) - 1) >> depth) << depth)
  79.  
  80. #define DIVRNDUP(a, b) (((a) + (b) - 1) / (b))
  81.  
  82. typedef struct {
  83.     AVFrame avframe;
  84.     int interpolated[3];    /* 1 if hpel[] is valid */
  85.     uint8_t *hpel[3][4];
  86.     uint8_t *hpel_base[3][4];
  87. } DiracFrame;
  88.  
  89. typedef struct {
  90.     union {
  91.         int16_t mv[2][2];
  92.         int16_t dc[3];
  93.     } u; /* anonymous unions aren't in C99 :( */
  94.     uint8_t ref;
  95. } DiracBlock;
  96.  
  97. typedef struct SubBand {
  98.     int level;
  99.     int orientation;
  100.     int stride;
  101.     int width;
  102.     int height;
  103.     int quant;
  104.     IDWTELEM *ibuf;
  105.     struct SubBand *parent;
  106.  
  107.     /* for low delay */
  108.     unsigned length;
  109.     const uint8_t *coeff_data;
  110. } SubBand;
  111.  
  112. typedef struct Plane {
  113.     int width;
  114.     int height;
  115.     ptrdiff_t stride;
  116.  
  117.     int idwt_width;
  118.     int idwt_height;
  119.     int idwt_stride;
  120.     IDWTELEM *idwt_buf;
  121.     IDWTELEM *idwt_buf_base;
  122.     IDWTELEM *idwt_tmp;
  123.  
  124.     /* block length */
  125.     uint8_t xblen;
  126.     uint8_t yblen;
  127.     /* block separation (block n+1 starts after this many pixels in block n) */
  128.     uint8_t xbsep;
  129.     uint8_t ybsep;
  130.     /* amount of overspill on each edge (half of the overlap between blocks) */
  131.     uint8_t xoffset;
  132.     uint8_t yoffset;
  133.  
  134.     SubBand band[MAX_DWT_LEVELS][4];
  135. } Plane;
  136.  
  137. typedef struct DiracContext {
  138.     AVCodecContext *avctx;
  139.     DSPContext dsp;
  140.     DiracDSPContext diracdsp;
  141.     GetBitContext gb;
  142.     dirac_source_params source;
  143.     int seen_sequence_header;
  144.     int frame_number;           /* number of the next frame to display       */
  145.     Plane plane[3];
  146.     int chroma_x_shift;
  147.     int chroma_y_shift;
  148.  
  149.     int zero_res;               /* zero residue flag                         */
  150.     int is_arith;               /* whether coeffs use arith or golomb coding */
  151.     int low_delay;              /* use the low delay syntax                  */
  152.     int globalmc_flag;          /* use global motion compensation            */
  153.     int num_refs;               /* number of reference pictures              */
  154.  
  155.     /* wavelet decoding */
  156.     unsigned wavelet_depth;     /* depth of the IDWT                         */
  157.     unsigned wavelet_idx;
  158.  
  159.     /**
  160.      * schroedinger older than 1.0.8 doesn't store
  161.      * quant delta if only one codebook exists in a band
  162.      */
  163.     unsigned old_delta_quant;
  164.     unsigned codeblock_mode;
  165.  
  166.     struct {
  167.         unsigned width;
  168.         unsigned height;
  169.     } codeblock[MAX_DWT_LEVELS+1];
  170.  
  171.     struct {
  172.         unsigned num_x;         /* number of horizontal slices               */
  173.         unsigned num_y;         /* number of vertical slices                 */
  174.         AVRational bytes;       /* average bytes per slice                   */
  175.         uint8_t quant[MAX_DWT_LEVELS][4]; /* [DIRAC_STD] E.1 */
  176.     } lowdelay;
  177.  
  178.     struct {
  179.         int pan_tilt[2];        /* pan/tilt vector                           */
  180.         int zrs[2][2];          /* zoom/rotate/shear matrix                  */
  181.         int perspective[2];     /* perspective vector                        */
  182.         unsigned zrs_exp;
  183.         unsigned perspective_exp;
  184.     } globalmc[2];
  185.  
  186.     /* motion compensation */
  187.     uint8_t mv_precision;       /* [DIRAC_STD] REFS_WT_PRECISION             */
  188.     int16_t weight[2];          /* [DIRAC_STD] REF1_WT and REF2_WT           */
  189.     unsigned weight_log2denom;  /* [DIRAC_STD] REFS_WT_PRECISION             */
  190.  
  191.     int blwidth;                /* number of blocks (horizontally)           */
  192.     int blheight;               /* number of blocks (vertically)             */
  193.     int sbwidth;                /* number of superblocks (horizontally)      */
  194.     int sbheight;               /* number of superblocks (vertically)        */
  195.  
  196.     uint8_t *sbsplit;
  197.     DiracBlock *blmotion;
  198.  
  199.     uint8_t *edge_emu_buffer[4];
  200.     uint8_t *edge_emu_buffer_base;
  201.  
  202.     uint16_t *mctmp;            /* buffer holding the MC data multipled by OBMC weights */
  203.     uint8_t *mcscratch;
  204.  
  205.     DECLARE_ALIGNED(16, uint8_t, obmc_weight)[3][MAX_BLOCKSIZE*MAX_BLOCKSIZE];
  206.  
  207.     void (*put_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
  208.     void (*avg_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
  209.     void (*add_obmc)(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
  210.     dirac_weight_func weight_func;
  211.     dirac_biweight_func biweight_func;
  212.  
  213.     DiracFrame *current_picture;
  214.     DiracFrame *ref_pics[2];
  215.  
  216.     DiracFrame *ref_frames[MAX_REFERENCE_FRAMES+1];
  217.     DiracFrame *delay_frames[MAX_DELAY+1];
  218.     DiracFrame all_frames[MAX_FRAMES];
  219. } DiracContext;
  220.  
  221. /**
  222.  * Dirac Specification ->
  223.  * Parse code values. 9.6.1 Table 9.1
  224.  */
  225. enum dirac_parse_code {
  226.     pc_seq_header         = 0x00,
  227.     pc_eos                = 0x10,
  228.     pc_aux_data           = 0x20,
  229.     pc_padding            = 0x30,
  230. };
  231.  
  232. enum dirac_subband {
  233.     subband_ll = 0,
  234.     subband_hl = 1,
  235.     subband_lh = 2,
  236.     subband_hh = 3
  237. };
  238.  
  239. static const uint8_t default_qmat[][4][4] = {
  240.     { { 5,  3,  3,  0}, { 0,  4,  4,  1}, { 0,  5,  5,  2}, { 0,  6,  6,  3} },
  241.     { { 4,  2,  2,  0}, { 0,  4,  4,  2}, { 0,  5,  5,  3}, { 0,  7,  7,  5} },
  242.     { { 5,  3,  3,  0}, { 0,  4,  4,  1}, { 0,  5,  5,  2}, { 0,  6,  6,  3} },
  243.     { { 8,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0} },
  244.     { { 8,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0} },
  245.     { { 0,  4,  4,  8}, { 0,  8,  8, 12}, { 0, 13, 13, 17}, { 0, 17, 17, 21} },
  246.     { { 3,  1,  1,  0}, { 0,  4,  4,  2}, { 0,  6,  6,  5}, { 0,  9,  9,  7} },
  247. };
  248.  
  249. static const int qscale_tab[MAX_QUANT+1] = {
  250.     4,     5,     6,     7,     8,    10,    11,    13,
  251.     16,    19,    23,    27,    32,    38,    45,    54,
  252.     64,    76,    91,   108,   128,   152,   181,   215,
  253.     256,   304,   362,   431,   512,   609,   724,   861,
  254.     1024,  1218,  1448,  1722,  2048,  2435,  2896,  3444,
  255.     4096,  4871,  5793,  6889,  8192,  9742, 11585, 13777,
  256.     16384, 19484, 23170, 27554, 32768, 38968, 46341, 55109,
  257.     65536, 77936
  258. };
  259.  
  260. static const int qoffset_intra_tab[MAX_QUANT+1] = {
  261.     1,     2,     3,     4,     4,     5,     6,     7,
  262.     8,    10,    12,    14,    16,    19,    23,    27,
  263.     32,    38,    46,    54,    64,    76,    91,   108,
  264.     128,   152,   181,   216,   256,   305,   362,   431,
  265.     512,   609,   724,   861,  1024,  1218,  1448,  1722,
  266.     2048,  2436,  2897,  3445,  4096,  4871,  5793,  6889,
  267.     8192,  9742, 11585, 13777, 16384, 19484, 23171, 27555,
  268.     32768, 38968
  269. };
  270.  
  271. static const int qoffset_inter_tab[MAX_QUANT+1] = {
  272.     1,     2,     2,     3,     3,     4,     4,     5,
  273.     6,     7,     9,    10,    12,    14,    17,    20,
  274.     24,    29,    34,    41,    48,    57,    68,    81,
  275.     96,   114,   136,   162,   192,   228,   272,   323,
  276.     384,   457,   543,   646,   768,   913,  1086,  1292,
  277.     1536,  1827,  2172,  2583,  3072,  3653,  4344,  5166,
  278.     6144,  7307,  8689, 10333, 12288, 14613, 17378, 20666,
  279.     24576, 29226
  280. };
  281.  
  282. /* magic number division by 3 from schroedinger */
  283. static inline int divide3(int x)
  284. {
  285.     return ((x+1)*21845 + 10922) >> 16;
  286. }
  287.  
  288. static DiracFrame *remove_frame(DiracFrame *framelist[], int picnum)
  289. {
  290.     DiracFrame *remove_pic = NULL;
  291.     int i, remove_idx = -1;
  292.  
  293.     for (i = 0; framelist[i]; i++)
  294.         if (framelist[i]->avframe.display_picture_number == picnum) {
  295.             remove_pic = framelist[i];
  296.             remove_idx = i;
  297.         }
  298.  
  299.     if (remove_pic)
  300.         for (i = remove_idx; framelist[i]; i++)
  301.             framelist[i] = framelist[i+1];
  302.  
  303.     return remove_pic;
  304. }
  305.  
  306. static int add_frame(DiracFrame *framelist[], int maxframes, DiracFrame *frame)
  307. {
  308.     int i;
  309.     for (i = 0; i < maxframes; i++)
  310.         if (!framelist[i]) {
  311.             framelist[i] = frame;
  312.             return 0;
  313.         }
  314.     return -1;
  315. }
  316.  
  317. static int alloc_sequence_buffers(DiracContext *s)
  318. {
  319.     int sbwidth  = DIVRNDUP(s->source.width,  4);
  320.     int sbheight = DIVRNDUP(s->source.height, 4);
  321.     int i, w, h, top_padding;
  322.  
  323.     /* todo: think more about this / use or set Plane here */
  324.     for (i = 0; i < 3; i++) {
  325.         int max_xblen = MAX_BLOCKSIZE >> (i ? s->chroma_x_shift : 0);
  326.         int max_yblen = MAX_BLOCKSIZE >> (i ? s->chroma_y_shift : 0);
  327.         w = s->source.width  >> (i ? s->chroma_x_shift : 0);
  328.         h = s->source.height >> (i ? s->chroma_y_shift : 0);
  329.  
  330.         /* we allocate the max we support here since num decompositions can
  331.          * change from frame to frame. Stride is aligned to 16 for SIMD, and
  332.          * 1<<MAX_DWT_LEVELS top padding to avoid if(y>0) in arith decoding
  333.          * MAX_BLOCKSIZE padding for MC: blocks can spill up to half of that
  334.          * on each side */
  335.         top_padding = FFMAX(1<<MAX_DWT_LEVELS, max_yblen/2);
  336.         w = FFALIGN(CALC_PADDING(w, MAX_DWT_LEVELS), 8); /* FIXME: Should this be 16 for SSE??? */
  337.         h = top_padding + CALC_PADDING(h, MAX_DWT_LEVELS) + max_yblen/2;
  338.  
  339.         s->plane[i].idwt_buf_base = av_mallocz((w+max_xblen)*h * sizeof(IDWTELEM));
  340.         s->plane[i].idwt_tmp      = av_malloc((w+16) * sizeof(IDWTELEM));
  341.         s->plane[i].idwt_buf      = s->plane[i].idwt_buf_base + top_padding*w;
  342.         if (!s->plane[i].idwt_buf_base || !s->plane[i].idwt_tmp)
  343.             return AVERROR(ENOMEM);
  344.     }
  345.  
  346.     w = s->source.width;
  347.     h = s->source.height;
  348.  
  349.     /* fixme: allocate using real stride here */
  350.     s->sbsplit  = av_malloc(sbwidth * sbheight);
  351.     s->blmotion = av_malloc(sbwidth * sbheight * 16 * sizeof(*s->blmotion));
  352.     s->edge_emu_buffer_base = av_malloc((w+64)*MAX_BLOCKSIZE);
  353.  
  354.     s->mctmp     = av_malloc((w+64+MAX_BLOCKSIZE) * (h+MAX_BLOCKSIZE) * sizeof(*s->mctmp));
  355.     s->mcscratch = av_malloc((w+64)*MAX_BLOCKSIZE);
  356.  
  357.     if (!s->sbsplit || !s->blmotion || !s->mctmp || !s->mcscratch)
  358.         return AVERROR(ENOMEM);
  359.     return 0;
  360. }
  361.  
  362. static void free_sequence_buffers(DiracContext *s)
  363. {
  364.     int i, j, k;
  365.  
  366.     for (i = 0; i < MAX_FRAMES; i++) {
  367.         if (s->all_frames[i].avframe.data[0]) {
  368.             av_frame_unref(&s->all_frames[i].avframe);
  369.             memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
  370.         }
  371.  
  372.         for (j = 0; j < 3; j++)
  373.             for (k = 1; k < 4; k++)
  374.                 av_freep(&s->all_frames[i].hpel_base[j][k]);
  375.     }
  376.  
  377.     memset(s->ref_frames, 0, sizeof(s->ref_frames));
  378.     memset(s->delay_frames, 0, sizeof(s->delay_frames));
  379.  
  380.     for (i = 0; i < 3; i++) {
  381.         av_freep(&s->plane[i].idwt_buf_base);
  382.         av_freep(&s->plane[i].idwt_tmp);
  383.     }
  384.  
  385.     av_freep(&s->sbsplit);
  386.     av_freep(&s->blmotion);
  387.     av_freep(&s->edge_emu_buffer_base);
  388.  
  389.     av_freep(&s->mctmp);
  390.     av_freep(&s->mcscratch);
  391. }
  392.  
  393. static av_cold int dirac_decode_init(AVCodecContext *avctx)
  394. {
  395.     DiracContext *s = avctx->priv_data;
  396.     s->avctx = avctx;
  397.     s->frame_number = -1;
  398.  
  399.     if (avctx->flags&CODEC_FLAG_EMU_EDGE) {
  400.         av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported!\n");
  401.         return AVERROR_PATCHWELCOME;
  402.     }
  403.  
  404.     ff_dsputil_init(&s->dsp, avctx);
  405.     ff_diracdsp_init(&s->diracdsp);
  406.  
  407.     return 0;
  408. }
  409.  
  410. static void dirac_decode_flush(AVCodecContext *avctx)
  411. {
  412.     DiracContext *s = avctx->priv_data;
  413.     free_sequence_buffers(s);
  414.     s->seen_sequence_header = 0;
  415.     s->frame_number = -1;
  416. }
  417.  
  418. static av_cold int dirac_decode_end(AVCodecContext *avctx)
  419. {
  420.     dirac_decode_flush(avctx);
  421.     return 0;
  422. }
  423.  
  424. #define SIGN_CTX(x) (CTX_SIGN_ZERO + ((x) > 0) - ((x) < 0))
  425.  
  426. static inline void coeff_unpack_arith(DiracArith *c, int qfactor, int qoffset,
  427.                                       SubBand *b, IDWTELEM *buf, int x, int y)
  428. {
  429.     int coeff, sign;
  430.     int sign_pred = 0;
  431.     int pred_ctx = CTX_ZPZN_F1;
  432.  
  433.     /* Check if the parent subband has a 0 in the corresponding position */
  434.     if (b->parent)
  435.         pred_ctx += !!b->parent->ibuf[b->parent->stride * (y>>1) + (x>>1)] << 1;
  436.  
  437.     if (b->orientation == subband_hl)
  438.         sign_pred = buf[-b->stride];
  439.  
  440.     /* Determine if the pixel has only zeros in its neighbourhood */
  441.     if (x) {
  442.         pred_ctx += !(buf[-1] | buf[-b->stride] | buf[-1-b->stride]);
  443.         if (b->orientation == subband_lh)
  444.             sign_pred = buf[-1];
  445.     } else {
  446.         pred_ctx += !buf[-b->stride];
  447.     }
  448.  
  449.     coeff = dirac_get_arith_uint(c, pred_ctx, CTX_COEFF_DATA);
  450.     if (coeff) {
  451.         coeff = (coeff * qfactor + qoffset + 2) >> 2;
  452.         sign  = dirac_get_arith_bit(c, SIGN_CTX(sign_pred));
  453.         coeff = (coeff ^ -sign) + sign;
  454.     }
  455.     *buf = coeff;
  456. }
  457.  
  458. static inline int coeff_unpack_golomb(GetBitContext *gb, int qfactor, int qoffset)
  459. {
  460.     int sign, coeff;
  461.  
  462.     coeff = svq3_get_ue_golomb(gb);
  463.     if (coeff) {
  464.         coeff = (coeff * qfactor + qoffset + 2) >> 2;
  465.         sign  = get_bits1(gb);
  466.         coeff = (coeff ^ -sign) + sign;
  467.     }
  468.     return coeff;
  469. }
  470.  
  471. /**
  472.  * Decode the coeffs in the rectangle defined by left, right, top, bottom
  473.  * [DIRAC_STD] 13.4.3.2 Codeblock unpacking loop. codeblock()
  474.  */
  475. static inline void codeblock(DiracContext *s, SubBand *b,
  476.                              GetBitContext *gb, DiracArith *c,
  477.                              int left, int right, int top, int bottom,
  478.                              int blockcnt_one, int is_arith)
  479. {
  480.     int x, y, zero_block;
  481.     int qoffset, qfactor;
  482.     IDWTELEM *buf;
  483.  
  484.     /* check for any coded coefficients in this codeblock */
  485.     if (!blockcnt_one) {
  486.         if (is_arith)
  487.             zero_block = dirac_get_arith_bit(c, CTX_ZERO_BLOCK);
  488.         else
  489.             zero_block = get_bits1(gb);
  490.  
  491.         if (zero_block)
  492.             return;
  493.     }
  494.  
  495.     if (s->codeblock_mode && !(s->old_delta_quant && blockcnt_one)) {
  496.         int quant = b->quant;
  497.         if (is_arith)
  498.             quant += dirac_get_arith_int(c, CTX_DELTA_Q_F, CTX_DELTA_Q_DATA);
  499.         else
  500.             quant += dirac_get_se_golomb(gb);
  501.         if (quant < 0) {
  502.             av_log(s->avctx, AV_LOG_ERROR, "Invalid quant\n");
  503.             return;
  504.         }
  505.         b->quant = quant;
  506.     }
  507.  
  508.     b->quant = FFMIN(b->quant, MAX_QUANT);
  509.  
  510.     qfactor = qscale_tab[b->quant];
  511.     /* TODO: context pointer? */
  512.     if (!s->num_refs)
  513.         qoffset = qoffset_intra_tab[b->quant];
  514.     else
  515.         qoffset = qoffset_inter_tab[b->quant];
  516.  
  517.     buf = b->ibuf + top * b->stride;
  518.     for (y = top; y < bottom; y++) {
  519.         for (x = left; x < right; x++) {
  520.             /* [DIRAC_STD] 13.4.4 Subband coefficients. coeff_unpack() */
  521.             if (is_arith)
  522.                 coeff_unpack_arith(c, qfactor, qoffset, b, buf+x, x, y);
  523.             else
  524.                 buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset);
  525.         }
  526.         buf += b->stride;
  527.     }
  528. }
  529.  
  530. /**
  531.  * Dirac Specification ->
  532.  * 13.3 intra_dc_prediction(band)
  533.  */
  534. static inline void intra_dc_prediction(SubBand *b)
  535. {
  536.     IDWTELEM *buf = b->ibuf;
  537.     int x, y;
  538.  
  539.     for (x = 1; x < b->width; x++)
  540.         buf[x] += buf[x-1];
  541.     buf += b->stride;
  542.  
  543.     for (y = 1; y < b->height; y++) {
  544.         buf[0] += buf[-b->stride];
  545.  
  546.         for (x = 1; x < b->width; x++) {
  547.             int pred = buf[x - 1] + buf[x - b->stride] + buf[x - b->stride-1];
  548.             buf[x]  += divide3(pred);
  549.         }
  550.         buf += b->stride;
  551.     }
  552. }
  553.  
  554. /**
  555.  * Dirac Specification ->
  556.  * 13.4.2 Non-skipped subbands.  subband_coeffs()
  557.  */
  558. static av_always_inline void decode_subband_internal(DiracContext *s, SubBand *b, int is_arith)
  559. {
  560.     int cb_x, cb_y, left, right, top, bottom;
  561.     DiracArith c;
  562.     GetBitContext gb;
  563.     int cb_width  = s->codeblock[b->level + (b->orientation != subband_ll)].width;
  564.     int cb_height = s->codeblock[b->level + (b->orientation != subband_ll)].height;
  565.     int blockcnt_one = (cb_width + cb_height) == 2;
  566.  
  567.     if (!b->length)
  568.         return;
  569.  
  570.     init_get_bits8(&gb, b->coeff_data, b->length);
  571.  
  572.     if (is_arith)
  573.         ff_dirac_init_arith_decoder(&c, &gb, b->length);
  574.  
  575.     top = 0;
  576.     for (cb_y = 0; cb_y < cb_height; cb_y++) {
  577.         bottom = (b->height * (cb_y+1)) / cb_height;
  578.         left = 0;
  579.         for (cb_x = 0; cb_x < cb_width; cb_x++) {
  580.             right = (b->width * (cb_x+1)) / cb_width;
  581.             codeblock(s, b, &gb, &c, left, right, top, bottom, blockcnt_one, is_arith);
  582.             left = right;
  583.         }
  584.         top = bottom;
  585.     }
  586.  
  587.     if (b->orientation == subband_ll && s->num_refs == 0)
  588.         intra_dc_prediction(b);
  589. }
  590.  
  591. static int decode_subband_arith(AVCodecContext *avctx, void *b)
  592. {
  593.     DiracContext *s = avctx->priv_data;
  594.     decode_subband_internal(s, b, 1);
  595.     return 0;
  596. }
  597.  
  598. static int decode_subband_golomb(AVCodecContext *avctx, void *arg)
  599. {
  600.     DiracContext *s = avctx->priv_data;
  601.     SubBand **b     = arg;
  602.     decode_subband_internal(s, *b, 0);
  603.     return 0;
  604. }
  605.  
  606. /**
  607.  * Dirac Specification ->
  608.  * [DIRAC_STD] 13.4.1 core_transform_data()
  609.  */
  610. static void decode_component(DiracContext *s, int comp)
  611. {
  612.     AVCodecContext *avctx = s->avctx;
  613.     SubBand *bands[3*MAX_DWT_LEVELS+1];
  614.     enum dirac_subband orientation;
  615.     int level, num_bands = 0;
  616.  
  617.     /* Unpack all subbands at all levels. */
  618.     for (level = 0; level < s->wavelet_depth; level++) {
  619.         for (orientation = !!level; orientation < 4; orientation++) {
  620.             SubBand *b = &s->plane[comp].band[level][orientation];
  621.             bands[num_bands++] = b;
  622.  
  623.             align_get_bits(&s->gb);
  624.             /* [DIRAC_STD] 13.4.2 subband() */
  625.             b->length = svq3_get_ue_golomb(&s->gb);
  626.             if (b->length) {
  627.                 b->quant = svq3_get_ue_golomb(&s->gb);
  628.                 align_get_bits(&s->gb);
  629.                 b->coeff_data = s->gb.buffer + get_bits_count(&s->gb)/8;
  630.                 b->length = FFMIN(b->length, FFMAX(get_bits_left(&s->gb)/8, 0));
  631.                 skip_bits_long(&s->gb, b->length*8);
  632.             }
  633.         }
  634.         /* arithmetic coding has inter-level dependencies, so we can only execute one level at a time */
  635.         if (s->is_arith)
  636.             avctx->execute(avctx, decode_subband_arith, &s->plane[comp].band[level][!!level],
  637.                            NULL, 4-!!level, sizeof(SubBand));
  638.     }
  639.     /* golomb coding has no inter-level dependencies, so we can execute all subbands in parallel */
  640.     if (!s->is_arith)
  641.         avctx->execute(avctx, decode_subband_golomb, bands, NULL, num_bands, sizeof(SubBand*));
  642. }
  643.  
  644. /* [DIRAC_STD] 13.5.5.2 Luma slice subband data. luma_slice_band(level,orient,sx,sy) --> if b2 == NULL */
  645. /* [DIRAC_STD] 13.5.5.3 Chroma slice subband data. chroma_slice_band(level,orient,sx,sy) --> if b2 != NULL */
  646. static void lowdelay_subband(DiracContext *s, GetBitContext *gb, int quant,
  647.                              int slice_x, int slice_y, int bits_end,
  648.                              SubBand *b1, SubBand *b2)
  649. {
  650.     int left   = b1->width  * slice_x    / s->lowdelay.num_x;
  651.     int right  = b1->width  *(slice_x+1) / s->lowdelay.num_x;
  652.     int top    = b1->height * slice_y    / s->lowdelay.num_y;
  653.     int bottom = b1->height *(slice_y+1) / s->lowdelay.num_y;
  654.  
  655.     int qfactor = qscale_tab[FFMIN(quant, MAX_QUANT)];
  656.     int qoffset = qoffset_intra_tab[FFMIN(quant, MAX_QUANT)];
  657.  
  658.     IDWTELEM *buf1 =      b1->ibuf + top * b1->stride;
  659.     IDWTELEM *buf2 = b2 ? b2->ibuf + top * b2->stride : NULL;
  660.     int x, y;
  661.     /* we have to constantly check for overread since the spec explictly
  662.        requires this, with the meaning that all remaining coeffs are set to 0 */
  663.     if (get_bits_count(gb) >= bits_end)
  664.         return;
  665.  
  666.     for (y = top; y < bottom; y++) {
  667.         for (x = left; x < right; x++) {
  668.             buf1[x] = coeff_unpack_golomb(gb, qfactor, qoffset);
  669.             if (get_bits_count(gb) >= bits_end)
  670.                 return;
  671.             if (buf2) {
  672.                 buf2[x] = coeff_unpack_golomb(gb, qfactor, qoffset);
  673.                 if (get_bits_count(gb) >= bits_end)
  674.                     return;
  675.             }
  676.         }
  677.         buf1 += b1->stride;
  678.         if (buf2)
  679.             buf2 += b2->stride;
  680.     }
  681. }
  682.  
  683. struct lowdelay_slice {
  684.     GetBitContext gb;
  685.     int slice_x;
  686.     int slice_y;
  687.     int bytes;
  688. };
  689.  
  690.  
  691. /**
  692.  * Dirac Specification ->
  693.  * 13.5.2 Slices. slice(sx,sy)
  694.  */
  695. static int decode_lowdelay_slice(AVCodecContext *avctx, void *arg)
  696. {
  697.     DiracContext *s = avctx->priv_data;
  698.     struct lowdelay_slice *slice = arg;
  699.     GetBitContext *gb = &slice->gb;
  700.     enum dirac_subband orientation;
  701.     int level, quant, chroma_bits, chroma_end;
  702.  
  703.     int quant_base  = get_bits(gb, 7); /*[DIRAC_STD] qindex */
  704.     int length_bits = av_log2(8 * slice->bytes)+1;
  705.     int luma_bits   = get_bits_long(gb, length_bits);
  706.     int luma_end    = get_bits_count(gb) + FFMIN(luma_bits, get_bits_left(gb));
  707.  
  708.     /* [DIRAC_STD] 13.5.5.2 luma_slice_band */
  709.     for (level = 0; level < s->wavelet_depth; level++)
  710.         for (orientation = !!level; orientation < 4; orientation++) {
  711.             quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
  712.             lowdelay_subband(s, gb, quant, slice->slice_x, slice->slice_y, luma_end,
  713.                              &s->plane[0].band[level][orientation], NULL);
  714.         }
  715.  
  716.     /* consume any unused bits from luma */
  717.     skip_bits_long(gb, get_bits_count(gb) - luma_end);
  718.  
  719.     chroma_bits = 8*slice->bytes - 7 - length_bits - luma_bits;
  720.     chroma_end  = get_bits_count(gb) + FFMIN(chroma_bits, get_bits_left(gb));
  721.     /* [DIRAC_STD] 13.5.5.3 chroma_slice_band */
  722.     for (level = 0; level < s->wavelet_depth; level++)
  723.         for (orientation = !!level; orientation < 4; orientation++) {
  724.             quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
  725.             lowdelay_subband(s, gb, quant, slice->slice_x, slice->slice_y, chroma_end,
  726.                              &s->plane[1].band[level][orientation],
  727.                              &s->plane[2].band[level][orientation]);
  728.         }
  729.  
  730.     return 0;
  731. }
  732.  
  733. /**
  734.  * Dirac Specification ->
  735.  * 13.5.1 low_delay_transform_data()
  736.  */
  737. static void decode_lowdelay(DiracContext *s)
  738. {
  739.     AVCodecContext *avctx = s->avctx;
  740.     int slice_x, slice_y, bytes, bufsize;
  741.     const uint8_t *buf;
  742.     struct lowdelay_slice *slices;
  743.     int slice_num = 0;
  744.  
  745.     slices = av_mallocz(s->lowdelay.num_x * s->lowdelay.num_y * sizeof(struct lowdelay_slice));
  746.  
  747.     align_get_bits(&s->gb);
  748.     /*[DIRAC_STD] 13.5.2 Slices. slice(sx,sy) */
  749.     buf = s->gb.buffer + get_bits_count(&s->gb)/8;
  750.     bufsize = get_bits_left(&s->gb);
  751.  
  752.     for (slice_y = 0; bufsize > 0 && slice_y < s->lowdelay.num_y; slice_y++)
  753.         for (slice_x = 0; bufsize > 0 && slice_x < s->lowdelay.num_x; slice_x++) {
  754.             bytes = (slice_num+1) * s->lowdelay.bytes.num / s->lowdelay.bytes.den
  755.                 - slice_num    * s->lowdelay.bytes.num / s->lowdelay.bytes.den;
  756.  
  757.             slices[slice_num].bytes   = bytes;
  758.             slices[slice_num].slice_x = slice_x;
  759.             slices[slice_num].slice_y = slice_y;
  760.             init_get_bits(&slices[slice_num].gb, buf, bufsize);
  761.             slice_num++;
  762.  
  763.             buf     += bytes;
  764.             bufsize -= bytes*8;
  765.         }
  766.  
  767.     avctx->execute(avctx, decode_lowdelay_slice, slices, NULL, slice_num,
  768.                    sizeof(struct lowdelay_slice)); /* [DIRAC_STD] 13.5.2 Slices */
  769.     intra_dc_prediction(&s->plane[0].band[0][0]);  /* [DIRAC_STD] 13.3 intra_dc_prediction() */
  770.     intra_dc_prediction(&s->plane[1].band[0][0]);  /* [DIRAC_STD] 13.3 intra_dc_prediction() */
  771.     intra_dc_prediction(&s->plane[2].band[0][0]);  /* [DIRAC_STD] 13.3 intra_dc_prediction() */
  772.     av_free(slices);
  773. }
  774.  
  775. static void init_planes(DiracContext *s)
  776. {
  777.     int i, w, h, level, orientation;
  778.  
  779.     for (i = 0; i < 3; i++) {
  780.         Plane *p = &s->plane[i];
  781.  
  782.         p->width       = s->source.width  >> (i ? s->chroma_x_shift : 0);
  783.         p->height      = s->source.height >> (i ? s->chroma_y_shift : 0);
  784.         p->idwt_width  = w = CALC_PADDING(p->width , s->wavelet_depth);
  785.         p->idwt_height = h = CALC_PADDING(p->height, s->wavelet_depth);
  786.         p->idwt_stride = FFALIGN(p->idwt_width, 8);
  787.  
  788.         for (level = s->wavelet_depth-1; level >= 0; level--) {
  789.             w = w>>1;
  790.             h = h>>1;
  791.             for (orientation = !!level; orientation < 4; orientation++) {
  792.                 SubBand *b = &p->band[level][orientation];
  793.  
  794.                 b->ibuf   = p->idwt_buf;
  795.                 b->level  = level;
  796.                 b->stride = p->idwt_stride << (s->wavelet_depth - level);
  797.                 b->width  = w;
  798.                 b->height = h;
  799.                 b->orientation = orientation;
  800.  
  801.                 if (orientation & 1)
  802.                     b->ibuf += w;
  803.                 if (orientation > 1)
  804.                     b->ibuf += b->stride>>1;
  805.  
  806.                 if (level)
  807.                     b->parent = &p->band[level-1][orientation];
  808.             }
  809.         }
  810.  
  811.         if (i > 0) {
  812.             p->xblen = s->plane[0].xblen >> s->chroma_x_shift;
  813.             p->yblen = s->plane[0].yblen >> s->chroma_y_shift;
  814.             p->xbsep = s->plane[0].xbsep >> s->chroma_x_shift;
  815.             p->ybsep = s->plane[0].ybsep >> s->chroma_y_shift;
  816.         }
  817.  
  818.         p->xoffset = (p->xblen - p->xbsep)/2;
  819.         p->yoffset = (p->yblen - p->ybsep)/2;
  820.     }
  821. }
  822.  
  823. /**
  824.  * Unpack the motion compensation parameters
  825.  * Dirac Specification ->
  826.  * 11.2 Picture prediction data. picture_prediction()
  827.  */
  828. static int dirac_unpack_prediction_parameters(DiracContext *s)
  829. {
  830.     static const uint8_t default_blen[] = { 4, 12, 16, 24 };
  831.     static const uint8_t default_bsep[] = { 4,  8, 12, 16 };
  832.  
  833.     GetBitContext *gb = &s->gb;
  834.     unsigned idx, ref;
  835.  
  836.     align_get_bits(gb);
  837.     /* [DIRAC_STD] 11.2.2 Block parameters. block_parameters() */
  838.     /* Luma and Chroma are equal. 11.2.3 */
  839.     idx = svq3_get_ue_golomb(gb); /* [DIRAC_STD] index */
  840.  
  841.     if (idx > 4) {
  842.         av_log(s->avctx, AV_LOG_ERROR, "Block prediction index too high\n");
  843.         return -1;
  844.     }
  845.  
  846.     if (idx == 0) {
  847.         s->plane[0].xblen = svq3_get_ue_golomb(gb);
  848.         s->plane[0].yblen = svq3_get_ue_golomb(gb);
  849.         s->plane[0].xbsep = svq3_get_ue_golomb(gb);
  850.         s->plane[0].ybsep = svq3_get_ue_golomb(gb);
  851.     } else {
  852.         /*[DIRAC_STD] preset_block_params(index). Table 11.1 */
  853.         s->plane[0].xblen = default_blen[idx-1];
  854.         s->plane[0].yblen = default_blen[idx-1];
  855.         s->plane[0].xbsep = default_bsep[idx-1];
  856.         s->plane[0].ybsep = default_bsep[idx-1];
  857.     }
  858.     /*[DIRAC_STD] 11.2.4 motion_data_dimensions()
  859.       Calculated in function dirac_unpack_block_motion_data */
  860.  
  861.     if (!s->plane[0].xbsep || !s->plane[0].ybsep || s->plane[0].xbsep < s->plane[0].xblen/2 || s->plane[0].ybsep < s->plane[0].yblen/2) {
  862.         av_log(s->avctx, AV_LOG_ERROR, "Block separation too small\n");
  863.         return -1;
  864.     }
  865.     if (s->plane[0].xbsep > s->plane[0].xblen || s->plane[0].ybsep > s->plane[0].yblen) {
  866.         av_log(s->avctx, AV_LOG_ERROR, "Block separation greater than size\n");
  867.         return -1;
  868.     }
  869.     if (FFMAX(s->plane[0].xblen, s->plane[0].yblen) > MAX_BLOCKSIZE) {
  870.         av_log(s->avctx, AV_LOG_ERROR, "Unsupported large block size\n");
  871.         return -1;
  872.     }
  873.  
  874.     /*[DIRAC_STD] 11.2.5 Motion vector precision. motion_vector_precision()
  875.       Read motion vector precision */
  876.     s->mv_precision = svq3_get_ue_golomb(gb);
  877.     if (s->mv_precision > 3) {
  878.         av_log(s->avctx, AV_LOG_ERROR, "MV precision finer than eighth-pel\n");
  879.         return -1;
  880.     }
  881.  
  882.     /*[DIRAC_STD] 11.2.6 Global motion. global_motion()
  883.       Read the global motion compensation parameters */
  884.     s->globalmc_flag = get_bits1(gb);
  885.     if (s->globalmc_flag) {
  886.         memset(s->globalmc, 0, sizeof(s->globalmc));
  887.         /* [DIRAC_STD] pan_tilt(gparams) */
  888.         for (ref = 0; ref < s->num_refs; ref++) {
  889.             if (get_bits1(gb)) {
  890.                 s->globalmc[ref].pan_tilt[0] = dirac_get_se_golomb(gb);
  891.                 s->globalmc[ref].pan_tilt[1] = dirac_get_se_golomb(gb);
  892.             }
  893.             /* [DIRAC_STD] zoom_rotate_shear(gparams)
  894.                zoom/rotation/shear parameters */
  895.             if (get_bits1(gb)) {
  896.                 s->globalmc[ref].zrs_exp   = svq3_get_ue_golomb(gb);
  897.                 s->globalmc[ref].zrs[0][0] = dirac_get_se_golomb(gb);
  898.                 s->globalmc[ref].zrs[0][1] = dirac_get_se_golomb(gb);
  899.                 s->globalmc[ref].zrs[1][0] = dirac_get_se_golomb(gb);
  900.                 s->globalmc[ref].zrs[1][1] = dirac_get_se_golomb(gb);
  901.             } else {
  902.                 s->globalmc[ref].zrs[0][0] = 1;
  903.                 s->globalmc[ref].zrs[1][1] = 1;
  904.             }
  905.             /* [DIRAC_STD] perspective(gparams) */
  906.             if (get_bits1(gb)) {
  907.                 s->globalmc[ref].perspective_exp = svq3_get_ue_golomb(gb);
  908.                 s->globalmc[ref].perspective[0]  = dirac_get_se_golomb(gb);
  909.                 s->globalmc[ref].perspective[1]  = dirac_get_se_golomb(gb);
  910.             }
  911.         }
  912.     }
  913.  
  914.     /*[DIRAC_STD] 11.2.7 Picture prediction mode. prediction_mode()
  915.       Picture prediction mode, not currently used. */
  916.     if (svq3_get_ue_golomb(gb)) {
  917.         av_log(s->avctx, AV_LOG_ERROR, "Unknown picture prediction mode\n");
  918.         return -1;
  919.     }
  920.  
  921.     /* [DIRAC_STD] 11.2.8 Reference picture weight. reference_picture_weights()
  922.        just data read, weight calculation will be done later on. */
  923.     s->weight_log2denom = 1;
  924.     s->weight[0]        = 1;
  925.     s->weight[1]        = 1;
  926.  
  927.     if (get_bits1(gb)) {
  928.         s->weight_log2denom = svq3_get_ue_golomb(gb);
  929.         s->weight[0] = dirac_get_se_golomb(gb);
  930.         if (s->num_refs == 2)
  931.             s->weight[1] = dirac_get_se_golomb(gb);
  932.     }
  933.     return 0;
  934. }
  935.  
  936. /**
  937.  * Dirac Specification ->
  938.  * 11.3 Wavelet transform data. wavelet_transform()
  939.  */
  940. static int dirac_unpack_idwt_params(DiracContext *s)
  941. {
  942.     GetBitContext *gb = &s->gb;
  943.     int i, level;
  944.     unsigned tmp;
  945.  
  946. #define CHECKEDREAD(dst, cond, errmsg) \
  947.     tmp = svq3_get_ue_golomb(gb); \
  948.     if (cond) { \
  949.         av_log(s->avctx, AV_LOG_ERROR, errmsg); \
  950.         return -1; \
  951.     }\
  952.     dst = tmp;
  953.  
  954.     align_get_bits(gb);
  955.  
  956.     s->zero_res = s->num_refs ? get_bits1(gb) : 0;
  957.     if (s->zero_res)
  958.         return 0;
  959.  
  960.     /*[DIRAC_STD] 11.3.1 Transform parameters. transform_parameters() */
  961.     CHECKEDREAD(s->wavelet_idx, tmp > 6, "wavelet_idx is too big\n")
  962.  
  963.     CHECKEDREAD(s->wavelet_depth, tmp > MAX_DWT_LEVELS || tmp < 1, "invalid number of DWT decompositions\n")
  964.  
  965.     if (!s->low_delay) {
  966.         /* Codeblock parameters (core syntax only) */
  967.         if (get_bits1(gb)) {
  968.             for (i = 0; i <= s->wavelet_depth; i++) {
  969.                 CHECKEDREAD(s->codeblock[i].width , tmp < 1, "codeblock width invalid\n")
  970.                 CHECKEDREAD(s->codeblock[i].height, tmp < 1, "codeblock height invalid\n")
  971.             }
  972.  
  973.             CHECKEDREAD(s->codeblock_mode, tmp > 1, "unknown codeblock mode\n")
  974.         } else
  975.             for (i = 0; i <= s->wavelet_depth; i++)
  976.                 s->codeblock[i].width = s->codeblock[i].height = 1;
  977.     } else {
  978.         /* Slice parameters + quantization matrix*/
  979.         /*[DIRAC_STD] 11.3.4 Slice coding Parameters (low delay syntax only). slice_parameters() */
  980.         s->lowdelay.num_x     = svq3_get_ue_golomb(gb);
  981.         s->lowdelay.num_y     = svq3_get_ue_golomb(gb);
  982.         s->lowdelay.bytes.num = svq3_get_ue_golomb(gb);
  983.         s->lowdelay.bytes.den = svq3_get_ue_golomb(gb);
  984.  
  985.         if (s->lowdelay.bytes.den <= 0) {
  986.             av_log(s->avctx,AV_LOG_ERROR,"Invalid lowdelay.bytes.den\n");
  987.             return AVERROR_INVALIDDATA;
  988.         }
  989.  
  990.         /* [DIRAC_STD] 11.3.5 Quantisation matrices (low-delay syntax). quant_matrix() */
  991.         if (get_bits1(gb)) {
  992.             av_log(s->avctx,AV_LOG_DEBUG,"Low Delay: Has Custom Quantization Matrix!\n");
  993.             /* custom quantization matrix */
  994.             s->lowdelay.quant[0][0] = svq3_get_ue_golomb(gb);
  995.             for (level = 0; level < s->wavelet_depth; level++) {
  996.                 s->lowdelay.quant[level][1] = svq3_get_ue_golomb(gb);
  997.                 s->lowdelay.quant[level][2] = svq3_get_ue_golomb(gb);
  998.                 s->lowdelay.quant[level][3] = svq3_get_ue_golomb(gb);
  999.             }
  1000.         } else {
  1001.             if (s->wavelet_depth > 4) {
  1002.                 av_log(s->avctx,AV_LOG_ERROR,"Mandatory custom low delay matrix missing for depth %d\n", s->wavelet_depth);
  1003.                 return AVERROR_INVALIDDATA;
  1004.             }
  1005.             /* default quantization matrix */
  1006.             for (level = 0; level < s->wavelet_depth; level++)
  1007.                 for (i = 0; i < 4; i++) {
  1008.                     s->lowdelay.quant[level][i] = default_qmat[s->wavelet_idx][level][i];
  1009.                     /* haar with no shift differs for different depths */
  1010.                     if (s->wavelet_idx == 3)
  1011.                         s->lowdelay.quant[level][i] += 4*(s->wavelet_depth-1 - level);
  1012.                 }
  1013.         }
  1014.     }
  1015.     return 0;
  1016. }
  1017.  
  1018. static inline int pred_sbsplit(uint8_t *sbsplit, int stride, int x, int y)
  1019. {
  1020.     static const uint8_t avgsplit[7] = { 0, 0, 1, 1, 1, 2, 2 };
  1021.  
  1022.     if (!(x|y))
  1023.         return 0;
  1024.     else if (!y)
  1025.         return sbsplit[-1];
  1026.     else if (!x)
  1027.         return sbsplit[-stride];
  1028.  
  1029.     return avgsplit[sbsplit[-1] + sbsplit[-stride] + sbsplit[-stride-1]];
  1030. }
  1031.  
  1032. static inline int pred_block_mode(DiracBlock *block, int stride, int x, int y, int refmask)
  1033. {
  1034.     int pred;
  1035.  
  1036.     if (!(x|y))
  1037.         return 0;
  1038.     else if (!y)
  1039.         return block[-1].ref & refmask;
  1040.     else if (!x)
  1041.         return block[-stride].ref & refmask;
  1042.  
  1043.     /* return the majority */
  1044.     pred = (block[-1].ref & refmask) + (block[-stride].ref & refmask) + (block[-stride-1].ref & refmask);
  1045.     return (pred >> 1) & refmask;
  1046. }
  1047.  
  1048. static inline void pred_block_dc(DiracBlock *block, int stride, int x, int y)
  1049. {
  1050.     int i, n = 0;
  1051.  
  1052.     memset(block->u.dc, 0, sizeof(block->u.dc));
  1053.  
  1054.     if (x && !(block[-1].ref & 3)) {
  1055.         for (i = 0; i < 3; i++)
  1056.             block->u.dc[i] += block[-1].u.dc[i];
  1057.         n++;
  1058.     }
  1059.  
  1060.     if (y && !(block[-stride].ref & 3)) {
  1061.         for (i = 0; i < 3; i++)
  1062.             block->u.dc[i] += block[-stride].u.dc[i];
  1063.         n++;
  1064.     }
  1065.  
  1066.     if (x && y && !(block[-1-stride].ref & 3)) {
  1067.         for (i = 0; i < 3; i++)
  1068.             block->u.dc[i] += block[-1-stride].u.dc[i];
  1069.         n++;
  1070.     }
  1071.  
  1072.     if (n == 2) {
  1073.         for (i = 0; i < 3; i++)
  1074.             block->u.dc[i] = (block->u.dc[i]+1)>>1;
  1075.     } else if (n == 3) {
  1076.         for (i = 0; i < 3; i++)
  1077.             block->u.dc[i] = divide3(block->u.dc[i]);
  1078.     }
  1079. }
  1080.  
  1081. static inline void pred_mv(DiracBlock *block, int stride, int x, int y, int ref)
  1082. {
  1083.     int16_t *pred[3];
  1084.     int refmask = ref+1;
  1085.     int mask = refmask | DIRAC_REF_MASK_GLOBAL; /*  exclude gmc blocks */
  1086.     int n = 0;
  1087.  
  1088.     if (x && (block[-1].ref & mask) == refmask)
  1089.         pred[n++] = block[-1].u.mv[ref];
  1090.  
  1091.     if (y && (block[-stride].ref & mask) == refmask)
  1092.         pred[n++] = block[-stride].u.mv[ref];
  1093.  
  1094.     if (x && y && (block[-stride-1].ref & mask) == refmask)
  1095.         pred[n++] = block[-stride-1].u.mv[ref];
  1096.  
  1097.     switch (n) {
  1098.     case 0:
  1099.         block->u.mv[ref][0] = 0;
  1100.         block->u.mv[ref][1] = 0;
  1101.         break;
  1102.     case 1:
  1103.         block->u.mv[ref][0] = pred[0][0];
  1104.         block->u.mv[ref][1] = pred[0][1];
  1105.         break;
  1106.     case 2:
  1107.         block->u.mv[ref][0] = (pred[0][0] + pred[1][0] + 1) >> 1;
  1108.         block->u.mv[ref][1] = (pred[0][1] + pred[1][1] + 1) >> 1;
  1109.         break;
  1110.     case 3:
  1111.         block->u.mv[ref][0] = mid_pred(pred[0][0], pred[1][0], pred[2][0]);
  1112.         block->u.mv[ref][1] = mid_pred(pred[0][1], pred[1][1], pred[2][1]);
  1113.         break;
  1114.     }
  1115. }
  1116.  
  1117. static void global_mv(DiracContext *s, DiracBlock *block, int x, int y, int ref)
  1118. {
  1119.     int ez      = s->globalmc[ref].zrs_exp;
  1120.     int ep      = s->globalmc[ref].perspective_exp;
  1121.     int (*A)[2] = s->globalmc[ref].zrs;
  1122.     int *b      = s->globalmc[ref].pan_tilt;
  1123.     int *c      = s->globalmc[ref].perspective;
  1124.  
  1125.     int m       = (1<<ep) - (c[0]*x + c[1]*y);
  1126.     int mx      = m * ((A[0][0] * x + A[0][1]*y) + (1<<ez) * b[0]);
  1127.     int my      = m * ((A[1][0] * x + A[1][1]*y) + (1<<ez) * b[1]);
  1128.  
  1129.     block->u.mv[ref][0] = (mx + (1<<(ez+ep))) >> (ez+ep);
  1130.     block->u.mv[ref][1] = (my + (1<<(ez+ep))) >> (ez+ep);
  1131. }
  1132.  
  1133. static void decode_block_params(DiracContext *s, DiracArith arith[8], DiracBlock *block,
  1134.                                 int stride, int x, int y)
  1135. {
  1136.     int i;
  1137.  
  1138.     block->ref  = pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF1);
  1139.     block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF1);
  1140.  
  1141.     if (s->num_refs == 2) {
  1142.         block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF2);
  1143.         block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF2) << 1;
  1144.     }
  1145.  
  1146.     if (!block->ref) {
  1147.         pred_block_dc(block, stride, x, y);
  1148.         for (i = 0; i < 3; i++)
  1149.             block->u.dc[i] += dirac_get_arith_int(arith+1+i, CTX_DC_F1, CTX_DC_DATA);
  1150.         return;
  1151.     }
  1152.  
  1153.     if (s->globalmc_flag) {
  1154.         block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_GLOBAL);
  1155.         block->ref ^= dirac_get_arith_bit(arith, CTX_GLOBAL_BLOCK) << 2;
  1156.     }
  1157.  
  1158.     for (i = 0; i < s->num_refs; i++)
  1159.         if (block->ref & (i+1)) {
  1160.             if (block->ref & DIRAC_REF_MASK_GLOBAL) {
  1161.                 global_mv(s, block, x, y, i);
  1162.             } else {
  1163.                 pred_mv(block, stride, x, y, i);
  1164.                 block->u.mv[i][0] += dirac_get_arith_int(arith + 4 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
  1165.                 block->u.mv[i][1] += dirac_get_arith_int(arith + 5 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
  1166.             }
  1167.         }
  1168. }
  1169.  
  1170. /**
  1171.  * Copies the current block to the other blocks covered by the current superblock split mode
  1172.  */
  1173. static void propagate_block_data(DiracBlock *block, int stride, int size)
  1174. {
  1175.     int x, y;
  1176.     DiracBlock *dst = block;
  1177.  
  1178.     for (x = 1; x < size; x++)
  1179.         dst[x] = *block;
  1180.  
  1181.     for (y = 1; y < size; y++) {
  1182.         dst += stride;
  1183.         for (x = 0; x < size; x++)
  1184.             dst[x] = *block;
  1185.     }
  1186. }
  1187.  
  1188. /**
  1189.  * Dirac Specification ->
  1190.  * 12. Block motion data syntax
  1191.  */
  1192. static int dirac_unpack_block_motion_data(DiracContext *s)
  1193. {
  1194.     GetBitContext *gb = &s->gb;
  1195.     uint8_t *sbsplit = s->sbsplit;
  1196.     int i, x, y, q, p;
  1197.     DiracArith arith[8];
  1198.  
  1199.     align_get_bits(gb);
  1200.  
  1201.     /* [DIRAC_STD] 11.2.4 and 12.2.1 Number of blocks and superblocks */
  1202.     s->sbwidth  = DIVRNDUP(s->source.width,  4*s->plane[0].xbsep);
  1203.     s->sbheight = DIVRNDUP(s->source.height, 4*s->plane[0].ybsep);
  1204.     s->blwidth  = 4 * s->sbwidth;
  1205.     s->blheight = 4 * s->sbheight;
  1206.  
  1207.     /* [DIRAC_STD] 12.3.1 Superblock splitting modes. superblock_split_modes()
  1208.        decode superblock split modes */
  1209.     ff_dirac_init_arith_decoder(arith, gb, svq3_get_ue_golomb(gb));     /* svq3_get_ue_golomb(gb) is the length */
  1210.     for (y = 0; y < s->sbheight; y++) {
  1211.         for (x = 0; x < s->sbwidth; x++) {
  1212.             unsigned int split  = dirac_get_arith_uint(arith, CTX_SB_F1, CTX_SB_DATA);
  1213.             if (split > 2)
  1214.                 return -1;
  1215.             sbsplit[x] = (split + pred_sbsplit(sbsplit+x, s->sbwidth, x, y)) % 3;
  1216.         }
  1217.         sbsplit += s->sbwidth;
  1218.     }
  1219.  
  1220.     /* setup arith decoding */
  1221.     ff_dirac_init_arith_decoder(arith, gb, svq3_get_ue_golomb(gb));
  1222.     for (i = 0; i < s->num_refs; i++) {
  1223.         ff_dirac_init_arith_decoder(arith + 4 + 2 * i, gb, svq3_get_ue_golomb(gb));
  1224.         ff_dirac_init_arith_decoder(arith + 5 + 2 * i, gb, svq3_get_ue_golomb(gb));
  1225.     }
  1226.     for (i = 0; i < 3; i++)
  1227.         ff_dirac_init_arith_decoder(arith+1+i, gb, svq3_get_ue_golomb(gb));
  1228.  
  1229.     for (y = 0; y < s->sbheight; y++)
  1230.         for (x = 0; x < s->sbwidth; x++) {
  1231.             int blkcnt = 1 << s->sbsplit[y * s->sbwidth + x];
  1232.             int step   = 4 >> s->sbsplit[y * s->sbwidth + x];
  1233.  
  1234.             for (q = 0; q < blkcnt; q++)
  1235.                 for (p = 0; p < blkcnt; p++) {
  1236.                     int bx = 4 * x + p*step;
  1237.                     int by = 4 * y + q*step;
  1238.                     DiracBlock *block = &s->blmotion[by*s->blwidth + bx];
  1239.                     decode_block_params(s, arith, block, s->blwidth, bx, by);
  1240.                     propagate_block_data(block, s->blwidth, step);
  1241.                 }
  1242.         }
  1243.  
  1244.     return 0;
  1245. }
  1246.  
  1247. static int weight(int i, int blen, int offset)
  1248. {
  1249. #define ROLLOFF(i) offset == 1 ? ((i) ? 5 : 3) :        \
  1250.     (1 + (6*(i) + offset - 1) / (2*offset - 1))
  1251.  
  1252.     if (i < 2*offset)
  1253.         return ROLLOFF(i);
  1254.     else if (i > blen-1 - 2*offset)
  1255.         return ROLLOFF(blen-1 - i);
  1256.     return 8;
  1257. }
  1258.  
  1259. static void init_obmc_weight_row(Plane *p, uint8_t *obmc_weight, int stride,
  1260.                                  int left, int right, int wy)
  1261. {
  1262.     int x;
  1263.     for (x = 0; left && x < p->xblen >> 1; x++)
  1264.         obmc_weight[x] = wy*8;
  1265.     for (; x < p->xblen >> right; x++)
  1266.         obmc_weight[x] = wy*weight(x, p->xblen, p->xoffset);
  1267.     for (; x < p->xblen; x++)
  1268.         obmc_weight[x] = wy*8;
  1269.     for (; x < stride; x++)
  1270.         obmc_weight[x] = 0;
  1271. }
  1272.  
  1273. static void init_obmc_weight(Plane *p, uint8_t *obmc_weight, int stride,
  1274.                              int left, int right, int top, int bottom)
  1275. {
  1276.     int y;
  1277.     for (y = 0; top && y < p->yblen >> 1; y++) {
  1278.         init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
  1279.         obmc_weight += stride;
  1280.     }
  1281.     for (; y < p->yblen >> bottom; y++) {
  1282.         int wy = weight(y, p->yblen, p->yoffset);
  1283.         init_obmc_weight_row(p, obmc_weight, stride, left, right, wy);
  1284.         obmc_weight += stride;
  1285.     }
  1286.     for (; y < p->yblen; y++) {
  1287.         init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
  1288.         obmc_weight += stride;
  1289.     }
  1290. }
  1291.  
  1292. static void init_obmc_weights(DiracContext *s, Plane *p, int by)
  1293. {
  1294.     int top = !by;
  1295.     int bottom = by == s->blheight-1;
  1296.  
  1297.     /* don't bother re-initing for rows 2 to blheight-2, the weights don't change */
  1298.     if (top || bottom || by == 1) {
  1299.         init_obmc_weight(p, s->obmc_weight[0], MAX_BLOCKSIZE, 1, 0, top, bottom);
  1300.         init_obmc_weight(p, s->obmc_weight[1], MAX_BLOCKSIZE, 0, 0, top, bottom);
  1301.         init_obmc_weight(p, s->obmc_weight[2], MAX_BLOCKSIZE, 0, 1, top, bottom);
  1302.     }
  1303. }
  1304.  
  1305. static const uint8_t epel_weights[4][4][4] = {
  1306.     {{ 16,  0,  0,  0 },
  1307.      { 12,  4,  0,  0 },
  1308.      {  8,  8,  0,  0 },
  1309.      {  4, 12,  0,  0 }},
  1310.     {{ 12,  0,  4,  0 },
  1311.      {  9,  3,  3,  1 },
  1312.      {  6,  6,  2,  2 },
  1313.      {  3,  9,  1,  3 }},
  1314.     {{  8,  0,  8,  0 },
  1315.      {  6,  2,  6,  2 },
  1316.      {  4,  4,  4,  4 },
  1317.      {  2,  6,  2,  6 }},
  1318.     {{  4,  0, 12,  0 },
  1319.      {  3,  1,  9,  3 },
  1320.      {  2,  2,  6,  6 },
  1321.      {  1,  3,  3,  9 }}
  1322. };
  1323.  
  1324. /**
  1325.  * For block x,y, determine which of the hpel planes to do bilinear
  1326.  * interpolation from and set src[] to the location in each hpel plane
  1327.  * to MC from.
  1328.  *
  1329.  * @return the index of the put_dirac_pixels_tab function to use
  1330.  *  0 for 1 plane (fpel,hpel), 1 for 2 planes (qpel), 2 for 4 planes (qpel), and 3 for epel
  1331.  */
  1332. static int mc_subpel(DiracContext *s, DiracBlock *block, const uint8_t *src[5],
  1333.                      int x, int y, int ref, int plane)
  1334. {
  1335.     Plane *p = &s->plane[plane];
  1336.     uint8_t **ref_hpel = s->ref_pics[ref]->hpel[plane];
  1337.     int motion_x = block->u.mv[ref][0];
  1338.     int motion_y = block->u.mv[ref][1];
  1339.     int mx, my, i, epel, nplanes = 0;
  1340.  
  1341.     if (plane) {
  1342.         motion_x >>= s->chroma_x_shift;
  1343.         motion_y >>= s->chroma_y_shift;
  1344.     }
  1345.  
  1346.     mx         = motion_x & ~(-1 << s->mv_precision);
  1347.     my         = motion_y & ~(-1 << s->mv_precision);
  1348.     motion_x >>= s->mv_precision;
  1349.     motion_y >>= s->mv_precision;
  1350.     /* normalize subpel coordinates to epel */
  1351.     /* TODO: template this function? */
  1352.     mx      <<= 3 - s->mv_precision;
  1353.     my      <<= 3 - s->mv_precision;
  1354.  
  1355.     x += motion_x;
  1356.     y += motion_y;
  1357.     epel = (mx|my)&1;
  1358.  
  1359.     /* hpel position */
  1360.     if (!((mx|my)&3)) {
  1361.         nplanes = 1;
  1362.         src[0] = ref_hpel[(my>>1)+(mx>>2)] + y*p->stride + x;
  1363.     } else {
  1364.         /* qpel or epel */
  1365.         nplanes = 4;
  1366.         for (i = 0; i < 4; i++)
  1367.             src[i] = ref_hpel[i] + y*p->stride + x;
  1368.  
  1369.         /* if we're interpolating in the right/bottom halves, adjust the planes as needed
  1370.            we increment x/y because the edge changes for half of the pixels */
  1371.         if (mx > 4) {
  1372.             src[0] += 1;
  1373.             src[2] += 1;
  1374.             x++;
  1375.         }
  1376.         if (my > 4) {
  1377.             src[0] += p->stride;
  1378.             src[1] += p->stride;
  1379.             y++;
  1380.         }
  1381.  
  1382.         /* hpel planes are:
  1383.            [0]: F  [1]: H
  1384.            [2]: V  [3]: C */
  1385.         if (!epel) {
  1386.             /* check if we really only need 2 planes since either mx or my is
  1387.                a hpel position. (epel weights of 0 handle this there) */
  1388.             if (!(mx&3)) {
  1389.                 /* mx == 0: average [0] and [2]
  1390.                    mx == 4: average [1] and [3] */
  1391.                 src[!mx] = src[2 + !!mx];
  1392.                 nplanes = 2;
  1393.             } else if (!(my&3)) {
  1394.                 src[0] = src[(my>>1)  ];
  1395.                 src[1] = src[(my>>1)+1];
  1396.                 nplanes = 2;
  1397.             }
  1398.         } else {
  1399.             /* adjust the ordering if needed so the weights work */
  1400.             if (mx > 4) {
  1401.                 FFSWAP(const uint8_t *, src[0], src[1]);
  1402.                 FFSWAP(const uint8_t *, src[2], src[3]);
  1403.             }
  1404.             if (my > 4) {
  1405.                 FFSWAP(const uint8_t *, src[0], src[2]);
  1406.                 FFSWAP(const uint8_t *, src[1], src[3]);
  1407.             }
  1408.             src[4] = epel_weights[my&3][mx&3];
  1409.         }
  1410.     }
  1411.  
  1412.     /* fixme: v/h _edge_pos */
  1413.     if (x + p->xblen > p->width +EDGE_WIDTH/2 ||
  1414.         y + p->yblen > p->height+EDGE_WIDTH/2 ||
  1415.         x < 0 || y < 0) {
  1416.         for (i = 0; i < nplanes; i++) {
  1417.             ff_emulated_edge_mc(s->edge_emu_buffer[i], p->stride,
  1418.                                 src[i], p->stride,
  1419.                                 p->xblen, p->yblen, x, y,
  1420.                                 p->width+EDGE_WIDTH/2, p->height+EDGE_WIDTH/2);
  1421.             src[i] = s->edge_emu_buffer[i];
  1422.         }
  1423.     }
  1424.     return (nplanes>>1) + epel;
  1425. }
  1426.  
  1427. static void add_dc(uint16_t *dst, int dc, int stride,
  1428.                    uint8_t *obmc_weight, int xblen, int yblen)
  1429. {
  1430.     int x, y;
  1431.     dc += 128;
  1432.  
  1433.     for (y = 0; y < yblen; y++) {
  1434.         for (x = 0; x < xblen; x += 2) {
  1435.             dst[x  ] += dc * obmc_weight[x  ];
  1436.             dst[x+1] += dc * obmc_weight[x+1];
  1437.         }
  1438.         dst          += stride;
  1439.         obmc_weight  += MAX_BLOCKSIZE;
  1440.     }
  1441. }
  1442.  
  1443. static void block_mc(DiracContext *s, DiracBlock *block,
  1444.                      uint16_t *mctmp, uint8_t *obmc_weight,
  1445.                      int plane, int dstx, int dsty)
  1446. {
  1447.     Plane *p = &s->plane[plane];
  1448.     const uint8_t *src[5];
  1449.     int idx;
  1450.  
  1451.     switch (block->ref&3) {
  1452.     case 0: /* DC */
  1453.         add_dc(mctmp, block->u.dc[plane], p->stride, obmc_weight, p->xblen, p->yblen);
  1454.         return;
  1455.     case 1:
  1456.     case 2:
  1457.         idx = mc_subpel(s, block, src, dstx, dsty, (block->ref&3)-1, plane);
  1458.         s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
  1459.         if (s->weight_func)
  1460.             s->weight_func(s->mcscratch, p->stride, s->weight_log2denom,
  1461.                            s->weight[0] + s->weight[1], p->yblen);
  1462.         break;
  1463.     case 3:
  1464.         idx = mc_subpel(s, block, src, dstx, dsty, 0, plane);
  1465.         s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
  1466.         idx = mc_subpel(s, block, src, dstx, dsty, 1, plane);
  1467.         if (s->biweight_func) {
  1468.             /* fixme: +32 is a quick hack */
  1469.             s->put_pixels_tab[idx](s->mcscratch + 32, src, p->stride, p->yblen);
  1470.             s->biweight_func(s->mcscratch, s->mcscratch+32, p->stride, s->weight_log2denom,
  1471.                              s->weight[0], s->weight[1], p->yblen);
  1472.         } else
  1473.             s->avg_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
  1474.         break;
  1475.     }
  1476.     s->add_obmc(mctmp, s->mcscratch, p->stride, obmc_weight, p->yblen);
  1477. }
  1478.  
  1479. static void mc_row(DiracContext *s, DiracBlock *block, uint16_t *mctmp, int plane, int dsty)
  1480. {
  1481.     Plane *p = &s->plane[plane];
  1482.     int x, dstx = p->xbsep - p->xoffset;
  1483.  
  1484.     block_mc(s, block, mctmp, s->obmc_weight[0], plane, -p->xoffset, dsty);
  1485.     mctmp += p->xbsep;
  1486.  
  1487.     for (x = 1; x < s->blwidth-1; x++) {
  1488.         block_mc(s, block+x, mctmp, s->obmc_weight[1], plane, dstx, dsty);
  1489.         dstx  += p->xbsep;
  1490.         mctmp += p->xbsep;
  1491.     }
  1492.     block_mc(s, block+x, mctmp, s->obmc_weight[2], plane, dstx, dsty);
  1493. }
  1494.  
  1495. static void select_dsp_funcs(DiracContext *s, int width, int height, int xblen, int yblen)
  1496. {
  1497.     int idx = 0;
  1498.     if (xblen > 8)
  1499.         idx = 1;
  1500.     if (xblen > 16)
  1501.         idx = 2;
  1502.  
  1503.     memcpy(s->put_pixels_tab, s->diracdsp.put_dirac_pixels_tab[idx], sizeof(s->put_pixels_tab));
  1504.     memcpy(s->avg_pixels_tab, s->diracdsp.avg_dirac_pixels_tab[idx], sizeof(s->avg_pixels_tab));
  1505.     s->add_obmc = s->diracdsp.add_dirac_obmc[idx];
  1506.     if (s->weight_log2denom > 1 || s->weight[0] != 1 || s->weight[1] != 1) {
  1507.         s->weight_func   = s->diracdsp.weight_dirac_pixels_tab[idx];
  1508.         s->biweight_func = s->diracdsp.biweight_dirac_pixels_tab[idx];
  1509.     } else {
  1510.         s->weight_func   = NULL;
  1511.         s->biweight_func = NULL;
  1512.     }
  1513. }
  1514.  
  1515. static void interpolate_refplane(DiracContext *s, DiracFrame *ref, int plane, int width, int height)
  1516. {
  1517.     /* chroma allocates an edge of 8 when subsampled
  1518.        which for 4:2:2 means an h edge of 16 and v edge of 8
  1519.        just use 8 for everything for the moment */
  1520.     int i, edge = EDGE_WIDTH/2;
  1521.  
  1522.     ref->hpel[plane][0] = ref->avframe.data[plane];
  1523.     s->dsp.draw_edges(ref->hpel[plane][0], ref->avframe.linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM); /* EDGE_TOP | EDGE_BOTTOM values just copied to make it build, this needs to be ensured */
  1524.  
  1525.     /* no need for hpel if we only have fpel vectors */
  1526.     if (!s->mv_precision)
  1527.         return;
  1528.  
  1529.     for (i = 1; i < 4; i++) {
  1530.         if (!ref->hpel_base[plane][i])
  1531.             ref->hpel_base[plane][i] = av_malloc((height+2*edge) * ref->avframe.linesize[plane] + 32);
  1532.         /* we need to be 16-byte aligned even for chroma */
  1533.         ref->hpel[plane][i] = ref->hpel_base[plane][i] + edge*ref->avframe.linesize[plane] + 16;
  1534.     }
  1535.  
  1536.     if (!ref->interpolated[plane]) {
  1537.         s->diracdsp.dirac_hpel_filter(ref->hpel[plane][1], ref->hpel[plane][2],
  1538.                                       ref->hpel[plane][3], ref->hpel[plane][0],
  1539.                                       ref->avframe.linesize[plane], width, height);
  1540.         s->dsp.draw_edges(ref->hpel[plane][1], ref->avframe.linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
  1541.         s->dsp.draw_edges(ref->hpel[plane][2], ref->avframe.linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
  1542.         s->dsp.draw_edges(ref->hpel[plane][3], ref->avframe.linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
  1543.     }
  1544.     ref->interpolated[plane] = 1;
  1545. }
  1546.  
  1547. /**
  1548.  * Dirac Specification ->
  1549.  * 13.0 Transform data syntax. transform_data()
  1550.  */
  1551. static int dirac_decode_frame_internal(DiracContext *s)
  1552. {
  1553.     DWTContext d;
  1554.     int y, i, comp, dsty;
  1555.  
  1556.     if (s->low_delay) {
  1557.         /* [DIRAC_STD] 13.5.1 low_delay_transform_data() */
  1558.         for (comp = 0; comp < 3; comp++) {
  1559.             Plane *p = &s->plane[comp];
  1560.             memset(p->idwt_buf, 0, p->idwt_stride * p->idwt_height * sizeof(IDWTELEM));
  1561.         }
  1562.         if (!s->zero_res)
  1563.             decode_lowdelay(s);
  1564.     }
  1565.  
  1566.     for (comp = 0; comp < 3; comp++) {
  1567.         Plane *p       = &s->plane[comp];
  1568.         uint8_t *frame = s->current_picture->avframe.data[comp];
  1569.  
  1570.         /* FIXME: small resolutions */
  1571.         for (i = 0; i < 4; i++)
  1572.             s->edge_emu_buffer[i] = s->edge_emu_buffer_base + i*FFALIGN(p->width, 16);
  1573.  
  1574.         if (!s->zero_res && !s->low_delay)
  1575.         {
  1576.             memset(p->idwt_buf, 0, p->idwt_stride * p->idwt_height * sizeof(IDWTELEM));
  1577.             decode_component(s, comp); /* [DIRAC_STD] 13.4.1 core_transform_data() */
  1578.         }
  1579.         if (ff_spatial_idwt_init2(&d, p->idwt_buf, p->idwt_width, p->idwt_height, p->idwt_stride,
  1580.                                   s->wavelet_idx+2, s->wavelet_depth, p->idwt_tmp))
  1581.             return -1;
  1582.  
  1583.         if (!s->num_refs) { /* intra */
  1584.             for (y = 0; y < p->height; y += 16) {
  1585.                 ff_spatial_idwt_slice2(&d, y+16); /* decode */
  1586.                 s->diracdsp.put_signed_rect_clamped(frame + y*p->stride, p->stride,
  1587.                                                     p->idwt_buf + y*p->idwt_stride, p->idwt_stride, p->width, 16);
  1588.             }
  1589.         } else { /* inter */
  1590.             int rowheight = p->ybsep*p->stride;
  1591.  
  1592.             select_dsp_funcs(s, p->width, p->height, p->xblen, p->yblen);
  1593.  
  1594.             for (i = 0; i < s->num_refs; i++)
  1595.                 interpolate_refplane(s, s->ref_pics[i], comp, p->width, p->height);
  1596.  
  1597.             memset(s->mctmp, 0, 4*p->yoffset*p->stride);
  1598.  
  1599.             dsty = -p->yoffset;
  1600.             for (y = 0; y < s->blheight; y++) {
  1601.                 int h     = 0,
  1602.                     start = FFMAX(dsty, 0);
  1603.                 uint16_t *mctmp    = s->mctmp + y*rowheight;
  1604.                 DiracBlock *blocks = s->blmotion + y*s->blwidth;
  1605.  
  1606.                 init_obmc_weights(s, p, y);
  1607.  
  1608.                 if (y == s->blheight-1 || start+p->ybsep > p->height)
  1609.                     h = p->height - start;
  1610.                 else
  1611.                     h = p->ybsep - (start - dsty);
  1612.                 if (h < 0)
  1613.                     break;
  1614.  
  1615.                 memset(mctmp+2*p->yoffset*p->stride, 0, 2*rowheight);
  1616.                 mc_row(s, blocks, mctmp, comp, dsty);
  1617.  
  1618.                 mctmp += (start - dsty)*p->stride + p->xoffset;
  1619.                 ff_spatial_idwt_slice2(&d, start + h); /* decode */
  1620.                 s->diracdsp.add_rect_clamped(frame + start*p->stride, mctmp, p->stride,
  1621.                                              p->idwt_buf + start*p->idwt_stride, p->idwt_stride, p->width, h);
  1622.  
  1623.                 dsty += p->ybsep;
  1624.             }
  1625.         }
  1626.     }
  1627.  
  1628.  
  1629.     return 0;
  1630. }
  1631.  
  1632. /**
  1633.  * Dirac Specification ->
  1634.  * 11.1.1 Picture Header. picture_header()
  1635.  */
  1636. static int dirac_decode_picture_header(DiracContext *s)
  1637. {
  1638.     int retire, picnum;
  1639.     int i, j, refnum, refdist;
  1640.     GetBitContext *gb = &s->gb;
  1641.  
  1642.     /* [DIRAC_STD] 11.1.1 Picture Header. picture_header() PICTURE_NUM */
  1643.     picnum = s->current_picture->avframe.display_picture_number = get_bits_long(gb, 32);
  1644.  
  1645.  
  1646.     av_log(s->avctx,AV_LOG_DEBUG,"PICTURE_NUM: %d\n",picnum);
  1647.  
  1648.     /* if this is the first keyframe after a sequence header, start our
  1649.        reordering from here */
  1650.     if (s->frame_number < 0)
  1651.         s->frame_number = picnum;
  1652.  
  1653.     s->ref_pics[0] = s->ref_pics[1] = NULL;
  1654.     for (i = 0; i < s->num_refs; i++) {
  1655.         refnum = picnum + dirac_get_se_golomb(gb);
  1656.         refdist = INT_MAX;
  1657.  
  1658.         /* find the closest reference to the one we want */
  1659.         /* Jordi: this is needed if the referenced picture hasn't yet arrived */
  1660.         for (j = 0; j < MAX_REFERENCE_FRAMES && refdist; j++)
  1661.             if (s->ref_frames[j]
  1662.                 && FFABS(s->ref_frames[j]->avframe.display_picture_number - refnum) < refdist) {
  1663.                 s->ref_pics[i] = s->ref_frames[j];
  1664.                 refdist = FFABS(s->ref_frames[j]->avframe.display_picture_number - refnum);
  1665.             }
  1666.  
  1667.         if (!s->ref_pics[i] || refdist)
  1668.             av_log(s->avctx, AV_LOG_DEBUG, "Reference not found\n");
  1669.  
  1670.         /* if there were no references at all, allocate one */
  1671.         if (!s->ref_pics[i])
  1672.             for (j = 0; j < MAX_FRAMES; j++)
  1673.                 if (!s->all_frames[j].avframe.data[0]) {
  1674.                     s->ref_pics[i] = &s->all_frames[j];
  1675.                     ff_get_buffer(s->avctx, &s->ref_pics[i]->avframe, AV_GET_BUFFER_FLAG_REF);
  1676.                     break;
  1677.                 }
  1678.     }
  1679.  
  1680.     /* retire the reference frames that are not used anymore */
  1681.     if (s->current_picture->avframe.reference) {
  1682.         retire = picnum + dirac_get_se_golomb(gb);
  1683.         if (retire != picnum) {
  1684.             DiracFrame *retire_pic = remove_frame(s->ref_frames, retire);
  1685.  
  1686.             if (retire_pic)
  1687.                 retire_pic->avframe.reference &= DELAYED_PIC_REF;
  1688.             else
  1689.                 av_log(s->avctx, AV_LOG_DEBUG, "Frame to retire not found\n");
  1690.         }
  1691.  
  1692.         /* if reference array is full, remove the oldest as per the spec */
  1693.         while (add_frame(s->ref_frames, MAX_REFERENCE_FRAMES, s->current_picture)) {
  1694.             av_log(s->avctx, AV_LOG_ERROR, "Reference frame overflow\n");
  1695.             remove_frame(s->ref_frames, s->ref_frames[0]->avframe.display_picture_number)->avframe.reference &= DELAYED_PIC_REF;
  1696.         }
  1697.     }
  1698.  
  1699.     if (s->num_refs) {
  1700.         if (dirac_unpack_prediction_parameters(s))  /* [DIRAC_STD] 11.2 Picture Prediction Data. picture_prediction() */
  1701.             return -1;
  1702.         if (dirac_unpack_block_motion_data(s))      /* [DIRAC_STD] 12. Block motion data syntax                       */
  1703.             return -1;
  1704.     }
  1705.     if (dirac_unpack_idwt_params(s))                /* [DIRAC_STD] 11.3 Wavelet transform data                        */
  1706.         return -1;
  1707.  
  1708.     init_planes(s);
  1709.     return 0;
  1710. }
  1711.  
  1712. static int get_delayed_pic(DiracContext *s, AVFrame *picture, int *got_frame)
  1713. {
  1714.     DiracFrame *out = s->delay_frames[0];
  1715.     int i, out_idx  = 0;
  1716.     int ret;
  1717.  
  1718.     /* find frame with lowest picture number */
  1719.     for (i = 1; s->delay_frames[i]; i++)
  1720.         if (s->delay_frames[i]->avframe.display_picture_number < out->avframe.display_picture_number) {
  1721.             out     = s->delay_frames[i];
  1722.             out_idx = i;
  1723.         }
  1724.  
  1725.     for (i = out_idx; s->delay_frames[i]; i++)
  1726.         s->delay_frames[i] = s->delay_frames[i+1];
  1727.  
  1728.     if (out) {
  1729.         out->avframe.reference ^= DELAYED_PIC_REF;
  1730.         *got_frame = 1;
  1731.         if((ret = av_frame_ref(picture, &out->avframe)) < 0)
  1732.             return ret;
  1733.     }
  1734.  
  1735.     return 0;
  1736. }
  1737.  
  1738. /**
  1739.  * Dirac Specification ->
  1740.  * 9.6 Parse Info Header Syntax. parse_info()
  1741.  * 4 byte start code + byte parse code + 4 byte size + 4 byte previous size
  1742.  */
  1743. #define DATA_UNIT_HEADER_SIZE 13
  1744.  
  1745. /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3
  1746.    inside the function parse_sequence() */
  1747. static int dirac_decode_data_unit(AVCodecContext *avctx, const uint8_t *buf, int size)
  1748. {
  1749.     DiracContext *s   = avctx->priv_data;
  1750.     DiracFrame *pic   = NULL;
  1751.     int ret, i, parse_code = buf[4];
  1752.     unsigned tmp;
  1753.  
  1754.     if (size < DATA_UNIT_HEADER_SIZE)
  1755.         return -1;
  1756.  
  1757.     init_get_bits(&s->gb, &buf[13], 8*(size - DATA_UNIT_HEADER_SIZE));
  1758.  
  1759.     if (parse_code == pc_seq_header) {
  1760.         if (s->seen_sequence_header)
  1761.             return 0;
  1762.  
  1763.         /* [DIRAC_STD] 10. Sequence header */
  1764.         if (avpriv_dirac_parse_sequence_header(avctx, &s->gb, &s->source))
  1765.             return -1;
  1766.  
  1767.         avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_x_shift, &s->chroma_y_shift);
  1768.  
  1769.         if (alloc_sequence_buffers(s))
  1770.             return -1;
  1771.  
  1772.         s->seen_sequence_header = 1;
  1773.     } else if (parse_code == pc_eos) { /* [DIRAC_STD] End of Sequence */
  1774.         free_sequence_buffers(s);
  1775.         s->seen_sequence_header = 0;
  1776.     } else if (parse_code == pc_aux_data) {
  1777.         if (buf[13] == 1) {     /* encoder implementation/version */
  1778.             int ver[3];
  1779.             /* versions older than 1.0.8 don't store quant delta for
  1780.                subbands with only one codeblock */
  1781.             if (sscanf(buf+14, "Schroedinger %d.%d.%d", ver, ver+1, ver+2) == 3)
  1782.                 if (ver[0] == 1 && ver[1] == 0 && ver[2] <= 7)
  1783.                     s->old_delta_quant = 1;
  1784.         }
  1785.     } else if (parse_code & 0x8) {  /* picture data unit */
  1786.         if (!s->seen_sequence_header) {
  1787.             av_log(avctx, AV_LOG_DEBUG, "Dropping frame without sequence header\n");
  1788.             return -1;
  1789.         }
  1790.  
  1791.         /* find an unused frame */
  1792.         for (i = 0; i < MAX_FRAMES; i++)
  1793.             if (s->all_frames[i].avframe.data[0] == NULL)
  1794.                 pic = &s->all_frames[i];
  1795.         if (!pic) {
  1796.             av_log(avctx, AV_LOG_ERROR, "framelist full\n");
  1797.             return -1;
  1798.         }
  1799.  
  1800.         avcodec_get_frame_defaults(&pic->avframe);
  1801.  
  1802.         /* [DIRAC_STD] Defined in 9.6.1 ... */
  1803.         tmp            =  parse_code & 0x03;                   /* [DIRAC_STD] num_refs()      */
  1804.         if (tmp > 2) {
  1805.             av_log(avctx, AV_LOG_ERROR, "num_refs of 3\n");
  1806.             return -1;
  1807.         }
  1808.         s->num_refs    = tmp;
  1809.         s->is_arith    = (parse_code & 0x48) == 0x08;          /* [DIRAC_STD] using_ac()      */
  1810.         s->low_delay   = (parse_code & 0x88) == 0x88;          /* [DIRAC_STD] is_low_delay()  */
  1811.         pic->avframe.reference = (parse_code & 0x0C) == 0x0C;  /* [DIRAC_STD]  is_reference() */
  1812.         pic->avframe.key_frame = s->num_refs == 0;             /* [DIRAC_STD] is_intra()      */
  1813.         pic->avframe.pict_type = s->num_refs + 1;              /* Definition of AVPictureType in avutil.h */
  1814.  
  1815.         if ((ret = ff_get_buffer(avctx, &pic->avframe, (parse_code & 0x0C) == 0x0C ? AV_GET_BUFFER_FLAG_REF : 0)) < 0)
  1816.             return ret;
  1817.         s->current_picture = pic;
  1818.         s->plane[0].stride = pic->avframe.linesize[0];
  1819.         s->plane[1].stride = pic->avframe.linesize[1];
  1820.         s->plane[2].stride = pic->avframe.linesize[2];
  1821.  
  1822.         /* [DIRAC_STD] 11.1 Picture parse. picture_parse() */
  1823.         if (dirac_decode_picture_header(s))
  1824.             return -1;
  1825.  
  1826.         /* [DIRAC_STD] 13.0 Transform data syntax. transform_data() */
  1827.         if (dirac_decode_frame_internal(s))
  1828.             return -1;
  1829.     }
  1830.     return 0;
  1831. }
  1832.  
  1833. static int dirac_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *pkt)
  1834. {
  1835.     DiracContext *s     = avctx->priv_data;
  1836.     DiracFrame *picture = data;
  1837.     uint8_t *buf        = pkt->data;
  1838.     int buf_size        = pkt->size;
  1839.     int i, data_unit_size, buf_idx = 0;
  1840.     int ret;
  1841.  
  1842.     /* release unused frames */
  1843.     for (i = 0; i < MAX_FRAMES; i++)
  1844.         if (s->all_frames[i].avframe.data[0] && !s->all_frames[i].avframe.reference) {
  1845.             av_frame_unref(&s->all_frames[i].avframe);
  1846.             memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
  1847.         }
  1848.  
  1849.     s->current_picture = NULL;
  1850.     *got_frame = 0;
  1851.  
  1852.     /* end of stream, so flush delayed pics */
  1853.     if (buf_size == 0)
  1854.         return get_delayed_pic(s, (AVFrame *)data, got_frame);
  1855.  
  1856.     for (;;) {
  1857.         /*[DIRAC_STD] Here starts the code from parse_info() defined in 9.6
  1858.           [DIRAC_STD] PARSE_INFO_PREFIX = "BBCD" as defined in ISO/IEC 646
  1859.           BBCD start code search */
  1860.         for (; buf_idx + DATA_UNIT_HEADER_SIZE < buf_size; buf_idx++) {
  1861.             if (buf[buf_idx  ] == 'B' && buf[buf_idx+1] == 'B' &&
  1862.                 buf[buf_idx+2] == 'C' && buf[buf_idx+3] == 'D')
  1863.                 break;
  1864.         }
  1865.         /* BBCD found or end of data */
  1866.         if (buf_idx + DATA_UNIT_HEADER_SIZE >= buf_size)
  1867.             break;
  1868.  
  1869.         data_unit_size = AV_RB32(buf+buf_idx+5);
  1870.         if (buf_idx + data_unit_size > buf_size || !data_unit_size) {
  1871.             if(buf_idx + data_unit_size > buf_size)
  1872.             av_log(s->avctx, AV_LOG_ERROR,
  1873.                    "Data unit with size %d is larger than input buffer, discarding\n",
  1874.                    data_unit_size);
  1875.             buf_idx += 4;
  1876.             continue;
  1877.         }
  1878.         /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3 inside the function parse_sequence() */
  1879.         if (dirac_decode_data_unit(avctx, buf+buf_idx, data_unit_size))
  1880.         {
  1881.             av_log(s->avctx, AV_LOG_ERROR,"Error in dirac_decode_data_unit\n");
  1882.             return -1;
  1883.         }
  1884.         buf_idx += data_unit_size;
  1885.     }
  1886.  
  1887.     if (!s->current_picture)
  1888.         return buf_size;
  1889.  
  1890.     if (s->current_picture->avframe.display_picture_number > s->frame_number) {
  1891.         DiracFrame *delayed_frame = remove_frame(s->delay_frames, s->frame_number);
  1892.  
  1893.         s->current_picture->avframe.reference |= DELAYED_PIC_REF;
  1894.  
  1895.         if (add_frame(s->delay_frames, MAX_DELAY, s->current_picture)) {
  1896.             int min_num = s->delay_frames[0]->avframe.display_picture_number;
  1897.             /* Too many delayed frames, so we display the frame with the lowest pts */
  1898.             av_log(avctx, AV_LOG_ERROR, "Delay frame overflow\n");
  1899.             delayed_frame = s->delay_frames[0];
  1900.  
  1901.             for (i = 1; s->delay_frames[i]; i++)
  1902.                 if (s->delay_frames[i]->avframe.display_picture_number < min_num)
  1903.                     min_num = s->delay_frames[i]->avframe.display_picture_number;
  1904.  
  1905.             delayed_frame = remove_frame(s->delay_frames, min_num);
  1906.             add_frame(s->delay_frames, MAX_DELAY, s->current_picture);
  1907.         }
  1908.  
  1909.         if (delayed_frame) {
  1910.             delayed_frame->avframe.reference ^= DELAYED_PIC_REF;
  1911.             if((ret=av_frame_ref(data, &delayed_frame->avframe)) < 0)
  1912.                 return ret;
  1913.             *got_frame = 1;
  1914.         }
  1915.     } else if (s->current_picture->avframe.display_picture_number == s->frame_number) {
  1916.         /* The right frame at the right time :-) */
  1917.         if((ret=av_frame_ref(data, &s->current_picture->avframe)) < 0)
  1918.             return ret;
  1919.         *got_frame = 1;
  1920.     }
  1921.  
  1922.     if (*got_frame)
  1923.         s->frame_number = picture->avframe.display_picture_number + 1;
  1924.  
  1925.     return buf_idx;
  1926. }
  1927.  
  1928. AVCodec ff_dirac_decoder = {
  1929.     .name           = "dirac",
  1930.     .long_name      = NULL_IF_CONFIG_SMALL("BBC Dirac VC-2"),
  1931.     .type           = AVMEDIA_TYPE_VIDEO,
  1932.     .id             = AV_CODEC_ID_DIRAC,
  1933.     .priv_data_size = sizeof(DiracContext),
  1934.     .init           = dirac_decode_init,
  1935.     .close          = dirac_decode_end,
  1936.     .decode         = dirac_decode_frame,
  1937.     .capabilities   = CODEC_CAP_DELAY,
  1938.     .flush          = dirac_decode_flush,
  1939. };
  1940.