falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
80770 Commits
32 Branches
369MB
Tree: 104c357b6a
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '104c357b6a'
${ noResults }
FFmpeg/libavcodec/diracdec.c

2222 lines
76KB
Raw Blame History 

  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 "get_bits.h"

  31. #include "bytestream.h"

  32. #include "internal.h"

  33. #include "golomb.h"

  34. #include "dirac_arith.h"

  35. #include "mpeg12data.h"

  36. #include "libavcodec/mpegvideo.h"

  37. #include "mpegvideoencdsp.h"

  38. #include "dirac_dwt.h"

  39. #include "dirac.h"

  40. #include "diractab.h"

  41. #include "diracdsp.h"

  42. #include "videodsp.h"

  43. 

  44. /**

  45.  * The spec limits this to 3 for frame coding, but in practice can be as high as 6

  46.  */

  47. #define MAX_REFERENCE_FRAMES 8

  48. #define MAX_DELAY 5         /* limit for main profile for frame coding (TODO: field coding) */

  49. #define MAX_FRAMES (MAX_REFERENCE_FRAMES + MAX_DELAY + 1)

  50. #define MAX_QUANT 255        /* max quant for VC-2 */

  51. #define MAX_BLOCKSIZE 32    /* maximum xblen/yblen we support */

  52. 

  53. /**

  54.  * DiracBlock->ref flags, if set then the block does MC from the given ref

  55.  */

  56. #define DIRAC_REF_MASK_REF1   1

  57. #define DIRAC_REF_MASK_REF2   2

  58. #define DIRAC_REF_MASK_GLOBAL 4

  59. 

  60. /**

  61.  * Value of Picture.reference when Picture is not a reference picture, but

  62.  * is held for delayed output.

  63.  */

  64. #define DELAYED_PIC_REF 4

  65. 

  66. #define CALC_PADDING(size, depth)                       \

  67.     (((size + (1 << depth) - 1) >> depth) << depth)

  68. 

  69. #define DIVRNDUP(a, b) (((a) + (b) - 1) / (b))

  70. 

  71. typedef struct {

  72.     AVFrame *avframe;

  73.     int interpolated[3];    /* 1 if hpel[] is valid */

  74.     uint8_t *hpel[3][4];

  75.     uint8_t *hpel_base[3][4];

  76.     int reference;

  77. } DiracFrame;

  78. 

  79. typedef struct {

  80.     union {

  81.         int16_t mv[2][2];

  82.         int16_t dc[3];

  83.     } u; /* anonymous unions aren't in C99 :( */

  84.     uint8_t ref;

  85. } DiracBlock;

  86. 

  87. typedef struct SubBand {

  88.     int level;

  89.     int orientation;

  90.     int stride; /* in bytes */

  91.     int width;

  92.     int height;

  93.     int pshift;

  94.     int quant;

  95.     uint8_t *ibuf;

  96.     struct SubBand *parent;

  97. 

  98.     /* for low delay */

  99.     unsigned length;

  100.     const uint8_t *coeff_data;

  101. } SubBand;

  102. 

  103. typedef struct Plane {

  104.     DWTPlane idwt;

  105. 

  106.     int width;

  107.     int height;

  108.     ptrdiff_t stride;

  109. 

  110.     /* block length */

  111.     uint8_t xblen;

  112.     uint8_t yblen;

  113.     /* block separation (block n+1 starts after this many pixels in block n) */

  114.     uint8_t xbsep;

  115.     uint8_t ybsep;

  116.     /* amount of overspill on each edge (half of the overlap between blocks) */

  117.     uint8_t xoffset;

  118.     uint8_t yoffset;

  119. 

  120.     SubBand band[MAX_DWT_LEVELS][4];

  121. } Plane;

  122. 

  123. typedef struct DiracContext {

  124.     AVCodecContext *avctx;

  125.     MpegvideoEncDSPContext mpvencdsp;

  126.     VideoDSPContext vdsp;

  127.     DiracDSPContext diracdsp;

  128.     DiracVersionInfo version;

  129.     GetBitContext gb;

  130.     AVDiracSeqHeader seq;

  131.     int seen_sequence_header;

  132.     int frame_number;           /* number of the next frame to display       */

  133.     Plane plane[3];

  134.     int chroma_x_shift;

  135.     int chroma_y_shift;

  136. 

  137.     int bit_depth;              /* bit depth                                 */

  138.     int pshift;                 /* pixel shift = bit_depth > 8               */

  139. 

  140.     int zero_res;               /* zero residue flag                         */

  141.     int is_arith;               /* whether coeffs use arith or golomb coding */

  142.     int core_syntax;            /* use core syntax only                      */

  143.     int low_delay;              /* use the low delay syntax                  */

  144.     int hq_picture;             /* high quality picture, enables low_delay   */

  145.     int ld_picture;             /* use low delay picture, turns on low_delay */

  146.     int dc_prediction;          /* has dc prediction                         */

  147.     int globalmc_flag;          /* use global motion compensation            */

  148.     int num_refs;               /* number of reference pictures              */

  149. 

  150.     /* wavelet decoding */

  151.     unsigned wavelet_depth;     /* depth of the IDWT                         */

  152.     unsigned wavelet_idx;

  153. 

  154.     /**

  155.      * schroedinger older than 1.0.8 doesn't store

  156.      * quant delta if only one codebook exists in a band

  157.      */

  158.     unsigned old_delta_quant;

  159.     unsigned codeblock_mode;

  160. 

  161.     unsigned num_x;              /* number of horizontal slices               */

  162.     unsigned num_y;              /* number of vertical slices                 */

  163. 

  164.     struct {

  165.         unsigned width;

  166.         unsigned height;

  167.     } codeblock[MAX_DWT_LEVELS+1];

  168. 

  169.     struct {

  170.         AVRational bytes;       /* average bytes per slice                   */

  171.         uint8_t quant[MAX_DWT_LEVELS][4]; /* [DIRAC_STD] E.1 */

  172.     } lowdelay;

  173. 

  174.     struct {

  175.         unsigned prefix_bytes;

  176.         uint64_t size_scaler;

  177.     } highquality;

  178. 

  179.     struct {

  180.         int pan_tilt[2];        /* pan/tilt vector                           */

  181.         int zrs[2][2];          /* zoom/rotate/shear matrix                  */

  182.         int perspective[2];     /* perspective vector                        */

  183.         unsigned zrs_exp;

  184.         unsigned perspective_exp;

  185.     } globalmc[2];

  186. 

  187.     /* motion compensation */

  188.     uint8_t mv_precision;       /* [DIRAC_STD] REFS_WT_PRECISION             */

  189.     int16_t weight[2];          /* [DIRAC_STD] REF1_WT and REF2_WT           */

  190.     unsigned weight_log2denom;  /* [DIRAC_STD] REFS_WT_PRECISION             */

  191. 

  192.     int blwidth;                /* number of blocks (horizontally)           */

  193.     int blheight;               /* number of blocks (vertically)             */

  194.     int sbwidth;                /* number of superblocks (horizontally)      */

  195.     int sbheight;               /* number of superblocks (vertically)        */

  196. 

  197.     uint8_t *sbsplit;

  198.     DiracBlock *blmotion;

  199. 

  200.     uint8_t *edge_emu_buffer[4];

  201.     uint8_t *edge_emu_buffer_base;

  202. 

  203.     uint16_t *mctmp;            /* buffer holding the MC data multiplied by OBMC weights */

  204.     uint8_t *mcscratch;

  205.     int buffer_stride;

  206. 

  207.     DECLARE_ALIGNED(16, uint8_t, obmc_weight)[3][MAX_BLOCKSIZE*MAX_BLOCKSIZE];

  208. 

  209.     void (*put_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);

  210.     void (*avg_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);

  211.     void (*add_obmc)(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);

  212.     dirac_weight_func weight_func;

  213.     dirac_biweight_func biweight_func;

  214. 

  215.     DiracFrame *current_picture;

  216.     DiracFrame *ref_pics[2];

  217. 

  218.     DiracFrame *ref_frames[MAX_REFERENCE_FRAMES+1];

  219.     DiracFrame *delay_frames[MAX_DELAY+1];

  220.     DiracFrame all_frames[MAX_FRAMES];

  221. } DiracContext;

  222. 

  223. enum dirac_subband {

  224.     subband_ll = 0,

  225.     subband_hl = 1,

  226.     subband_lh = 2,

  227.     subband_hh = 3,

  228.     subband_nb,

  229. };

  230. 

  231. /* magic number division by 3 from schroedinger */

  232. static inline int divide3(int x)

  233. {

  234.     return ((x+1)*21845 + 10922) >> 16;

  235. }

  236. 

  237. static DiracFrame *remove_frame(DiracFrame *framelist[], int picnum)

  238. {

  239.     DiracFrame *remove_pic = NULL;

  240.     int i, remove_idx = -1;

  241. 

  242.     for (i = 0; framelist[i]; i++)

  243.         if (framelist[i]->avframe->display_picture_number == picnum) {

  244.             remove_pic = framelist[i];

  245.             remove_idx = i;

  246.         }

  247. 

  248.     if (remove_pic)

  249.         for (i = remove_idx; framelist[i]; i++)

  250.             framelist[i] = framelist[i+1];

  251. 

  252.     return remove_pic;

  253. }

  254. 

  255. static int add_frame(DiracFrame *framelist[], int maxframes, DiracFrame *frame)

  256. {

  257.     int i;

  258.     for (i = 0; i < maxframes; i++)

  259.         if (!framelist[i]) {

  260.             framelist[i] = frame;

  261.             return 0;

  262.         }

  263.     return -1;

  264. }

  265. 

  266. static int alloc_sequence_buffers(DiracContext *s)

  267. {

  268.     int sbwidth  = DIVRNDUP(s->seq.width,  4);

  269.     int sbheight = DIVRNDUP(s->seq.height, 4);

  270.     int i, w, h, top_padding;

  271. 

  272.     /* todo: think more about this / use or set Plane here */

  273.     for (i = 0; i < 3; i++) {

  274.         int max_xblen = MAX_BLOCKSIZE >> (i ? s->chroma_x_shift : 0);

  275.         int max_yblen = MAX_BLOCKSIZE >> (i ? s->chroma_y_shift : 0);

  276.         w = s->seq.width  >> (i ? s->chroma_x_shift : 0);

  277.         h = s->seq.height >> (i ? s->chroma_y_shift : 0);

  278. 

  279.         /* we allocate the max we support here since num decompositions can

  280.          * change from frame to frame. Stride is aligned to 16 for SIMD, and

  281.          * 1<<MAX_DWT_LEVELS top padding to avoid if(y>0) in arith decoding

  282.          * MAX_BLOCKSIZE padding for MC: blocks can spill up to half of that

  283.          * on each side */

  284.         top_padding = FFMAX(1<<MAX_DWT_LEVELS, max_yblen/2);

  285.         w = FFALIGN(CALC_PADDING(w, MAX_DWT_LEVELS), 8); /* FIXME: Should this be 16 for SSE??? */

  286.         h = top_padding + CALC_PADDING(h, MAX_DWT_LEVELS) + max_yblen/2;

  287. 

  288.         s->plane[i].idwt.buf_base = av_mallocz_array((w+max_xblen), h * (2 << s->pshift));

  289.         s->plane[i].idwt.tmp      = av_malloc_array((w+16), 2 << s->pshift);

  290.         s->plane[i].idwt.buf      = s->plane[i].idwt.buf_base + (top_padding*w)*(2 << s->pshift);

  291.         if (!s->plane[i].idwt.buf_base || !s->plane[i].idwt.tmp)

  292.             return AVERROR(ENOMEM);

  293.     }

  294. 

  295.     /* fixme: allocate using real stride here */

  296.     s->sbsplit  = av_malloc_array(sbwidth, sbheight);

  297.     s->blmotion = av_malloc_array(sbwidth, sbheight * 16 * sizeof(*s->blmotion));

  298. 

  299.     if (!s->sbsplit || !s->blmotion)

  300.         return AVERROR(ENOMEM);

  301.     return 0;

  302. }

  303. 

  304. static int alloc_buffers(DiracContext *s, int stride)

  305. {

  306.     int w = s->seq.width;

  307.     int h = s->seq.height;

  308. 

  309.     av_assert0(stride >= w);

  310.     stride += 64;

  311. 

  312.     if (s->buffer_stride >= stride)

  313.         return 0;

  314.     s->buffer_stride = 0;

  315. 

  316.     av_freep(&s->edge_emu_buffer_base);

  317.     memset(s->edge_emu_buffer, 0, sizeof(s->edge_emu_buffer));

  318.     av_freep(&s->mctmp);

  319.     av_freep(&s->mcscratch);

  320. 

  321.     s->edge_emu_buffer_base = av_malloc_array(stride, MAX_BLOCKSIZE);

  322. 

  323.     s->mctmp     = av_malloc_array((stride+MAX_BLOCKSIZE), (h+MAX_BLOCKSIZE) * sizeof(*s->mctmp));

  324.     s->mcscratch = av_malloc_array(stride, MAX_BLOCKSIZE);

  325. 

  326.     if (!s->edge_emu_buffer_base || !s->mctmp || !s->mcscratch)

  327.         return AVERROR(ENOMEM);

  328. 

  329.     s->buffer_stride = stride;

  330.     return 0;

  331. }

  332. 

  333. static void free_sequence_buffers(DiracContext *s)

  334. {

  335.     int i, j, k;

  336. 

  337.     for (i = 0; i < MAX_FRAMES; i++) {

  338.         if (s->all_frames[i].avframe->data[0]) {

  339.             av_frame_unref(s->all_frames[i].avframe);

  340.             memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));

  341.         }

  342. 

  343.         for (j = 0; j < 3; j++)

  344.             for (k = 1; k < 4; k++)

  345.                 av_freep(&s->all_frames[i].hpel_base[j][k]);

  346.     }

  347. 

  348.     memset(s->ref_frames, 0, sizeof(s->ref_frames));

  349.     memset(s->delay_frames, 0, sizeof(s->delay_frames));

  350. 

  351.     for (i = 0; i < 3; i++) {

  352.         av_freep(&s->plane[i].idwt.buf_base);

  353.         av_freep(&s->plane[i].idwt.tmp);

  354.     }

  355. 

  356.     s->buffer_stride = 0;

  357.     av_freep(&s->sbsplit);

  358.     av_freep(&s->blmotion);

  359.     av_freep(&s->edge_emu_buffer_base);

  360. 

  361.     av_freep(&s->mctmp);

  362.     av_freep(&s->mcscratch);

  363. }

  364. 

  365. static av_cold int dirac_decode_init(AVCodecContext *avctx)

  366. {

  367.     DiracContext *s = avctx->priv_data;

  368.     int i;

  369. 

  370.     s->avctx = avctx;

  371.     s->frame_number = -1;

  372. 

  373.     ff_diracdsp_init(&s->diracdsp);

  374.     ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);

  375.     ff_videodsp_init(&s->vdsp, 8);

  376. 

  377.     for (i = 0; i < MAX_FRAMES; i++) {

  378.         s->all_frames[i].avframe = av_frame_alloc();

  379.         if (!s->all_frames[i].avframe) {

  380.             while (i > 0)

  381.                 av_frame_free(&s->all_frames[--i].avframe);

  382.             return AVERROR(ENOMEM);

  383.         }

  384.     }

  385. 

  386.     return 0;

  387. }

  388. 

  389. static void dirac_decode_flush(AVCodecContext *avctx)

  390. {

  391.     DiracContext *s = avctx->priv_data;

  392.     free_sequence_buffers(s);

  393.     s->seen_sequence_header = 0;

  394.     s->frame_number = -1;

  395. }

  396. 

  397. static av_cold int dirac_decode_end(AVCodecContext *avctx)

  398. {

  399.     DiracContext *s = avctx->priv_data;

  400.     int i;

  401. 

  402.     dirac_decode_flush(avctx);

  403.     for (i = 0; i < MAX_FRAMES; i++)

  404.         av_frame_free(&s->all_frames[i].avframe);

  405. 

  406.     return 0;

  407. }

  408. 

  409. #define SIGN_CTX(x) (CTX_SIGN_ZERO + ((x) > 0) - ((x) < 0))

  410. 

  411. static inline int coeff_unpack_golomb(GetBitContext *gb, int qfactor, int qoffset)

  412. {

  413.     int sign, coeff;

  414.     uint32_t buf;

  415. 

  416.     OPEN_READER(re, gb);

  417.     UPDATE_CACHE(re, gb);

  418.     buf = GET_CACHE(re, gb);

  419. 

  420.     if (buf & 0x80000000) {

  421.         LAST_SKIP_BITS(re,gb,1);

  422.         CLOSE_READER(re, gb);

  423.         return 0;

  424.     }

  425. 

  426.     if (buf & 0xAA800000) {

  427.         buf >>= 32 - 8;

  428.         SKIP_BITS(re, gb, ff_interleaved_golomb_vlc_len[buf]);

  429. 

  430.         coeff = ff_interleaved_ue_golomb_vlc_code[buf];

  431.     } else {

  432.         unsigned ret = 1;

  433. 

  434.         do {

  435.             buf >>= 32 - 8;

  436.             SKIP_BITS(re, gb,

  437.                            FFMIN(ff_interleaved_golomb_vlc_len[buf], 8));

  438. 

  439.             if (ff_interleaved_golomb_vlc_len[buf] != 9) {

  440.                 ret <<= (ff_interleaved_golomb_vlc_len[buf] - 1) >> 1;

  441.                 ret  |= ff_interleaved_dirac_golomb_vlc_code[buf];

  442.                 break;

  443.             }

  444.             ret = (ret << 4) | ff_interleaved_dirac_golomb_vlc_code[buf];

  445.             UPDATE_CACHE(re, gb);

  446.             buf = GET_CACHE(re, gb);

  447.         } while (ret<0x8000000U && BITS_AVAILABLE(re, gb));

  448. 

  449.         coeff = ret - 1;

  450.     }

  451. 

  452.     coeff = (coeff * qfactor + qoffset) >> 2;

  453.     sign  = SHOW_SBITS(re, gb, 1);

  454.     LAST_SKIP_BITS(re, gb, 1);

  455.     coeff = (coeff ^ sign) - sign;

  456. 

  457.     CLOSE_READER(re, gb);

  458.     return coeff;

  459. }

  460. 

  461. #define UNPACK_ARITH(n, type) \

  462.     static inline void coeff_unpack_arith_##n(DiracArith *c, int qfactor, int qoffset, \

  463.                                               SubBand *b, type *buf, int x, int y) \

  464.     { \

  465.         int coeff, sign, sign_pred = 0, pred_ctx = CTX_ZPZN_F1; \

  466.         const int mstride = -(b->stride >> (1+b->pshift)); \

  467.         if (b->parent) { \

  468.             const type *pbuf = (type *)b->parent->ibuf; \

  469.             const int stride = b->parent->stride >> (1+b->parent->pshift); \

  470.             pred_ctx += !!pbuf[stride * (y>>1) + (x>>1)] << 1; \

  471.         } \

  472.         if (b->orientation == subband_hl) \

  473.             sign_pred = buf[mstride]; \

  474.         if (x) { \

  475.             pred_ctx += !(buf[-1] | buf[mstride] | buf[-1 + mstride]); \

  476.             if (b->orientation == subband_lh) \

  477.                 sign_pred = buf[-1]; \

  478.         } else { \

  479.             pred_ctx += !buf[mstride]; \

  480.         } \

  481.         coeff = dirac_get_arith_uint(c, pred_ctx, CTX_COEFF_DATA); \

  482.         if (coeff) { \

  483.             coeff = (coeff * qfactor + qoffset) >> 2; \

  484.             sign  = dirac_get_arith_bit(c, SIGN_CTX(sign_pred)); \

  485.             coeff = (coeff ^ -sign) + sign; \

  486.         } \

  487.         *buf = coeff; \

  488.     } \

  489. 

  490. UNPACK_ARITH(8, int16_t)

  491. UNPACK_ARITH(10, int32_t)

  492. 

  493. /**

  494.  * Decode the coeffs in the rectangle defined by left, right, top, bottom

  495.  * [DIRAC_STD] 13.4.3.2 Codeblock unpacking loop. codeblock()

  496.  */

  497. static inline void codeblock(DiracContext *s, SubBand *b,

  498.                              GetBitContext *gb, DiracArith *c,

  499.                              int left, int right, int top, int bottom,

  500.                              int blockcnt_one, int is_arith)

  501. {

  502.     int x, y, zero_block;

  503.     int qoffset, qfactor;

  504.     uint8_t *buf;

  505. 

  506.     /* check for any coded coefficients in this codeblock */

  507.     if (!blockcnt_one) {

  508.         if (is_arith)

  509.             zero_block = dirac_get_arith_bit(c, CTX_ZERO_BLOCK);

  510.         else

  511.             zero_block = get_bits1(gb);

  512. 

  513.         if (zero_block)

  514.             return;

  515.     }

  516. 

  517.     if (s->codeblock_mode && !(s->old_delta_quant && blockcnt_one)) {

  518.         int quant = b->quant;

  519.         if (is_arith)

  520.             quant += dirac_get_arith_int(c, CTX_DELTA_Q_F, CTX_DELTA_Q_DATA);

  521.         else

  522.             quant += dirac_get_se_golomb(gb);

  523.         if (quant < 0) {

  524.             av_log(s->avctx, AV_LOG_ERROR, "Invalid quant\n");

  525.             return;

  526.         }

  527.         b->quant = quant;

  528.     }

  529. 

  530.     if (b->quant > 115) {

  531.         av_log(s->avctx, AV_LOG_ERROR, "Unsupported quant %d\n", b->quant);

  532.         b->quant = 0;

  533.         return;

  534.     }

  535. 

  536.     qfactor = ff_dirac_qscale_tab[b->quant];

  537.     /* TODO: context pointer? */

  538.     if (!s->num_refs)

  539.         qoffset = ff_dirac_qoffset_intra_tab[b->quant] + 2;

  540.     else

  541.         qoffset = ff_dirac_qoffset_inter_tab[b->quant] + 2;

  542. 

  543.     buf = b->ibuf + top * b->stride;

  544.     if (is_arith) {

  545.         for (y = top; y < bottom; y++) {

  546.             for (x = left; x < right; x++) {

  547.                 if (b->pshift) {

  548.                     coeff_unpack_arith_10(c, qfactor, qoffset, b, (int32_t*)(buf)+x, x, y);

  549.                 } else {

  550.                     coeff_unpack_arith_8(c, qfactor, qoffset, b, (int16_t*)(buf)+x, x, y);

  551.                 }

  552.             }

  553.             buf += b->stride;

  554.         }

  555.     } else {

  556.         for (y = top; y < bottom; y++) {

  557.             for (x = left; x < right; x++) {

  558.                 int val = coeff_unpack_golomb(gb, qfactor, qoffset);

  559.                 if (b->pshift) {

  560.                     AV_WN32(&buf[4*x], val);

  561.                 } else {

  562.                     AV_WN16(&buf[2*x], val);

  563.                 }

  564.             }

  565.             buf += b->stride;

  566.          }

  567.      }

  568. }

  569. 

  570. /**

  571.  * Dirac Specification ->

  572.  * 13.3 intra_dc_prediction(band)

  573.  */

  574. #define INTRA_DC_PRED(n, type) \

  575.     static inline void intra_dc_prediction_##n(SubBand *b) \

  576.     { \

  577.         type *buf = (type*)b->ibuf; \

  578.         int x, y; \

  579.         \

  580.         for (x = 1; x < b->width; x++) \

  581.             buf[x] += buf[x-1]; \

  582.         buf += (b->stride >> (1+b->pshift)); \

  583.         \

  584.         for (y = 1; y < b->height; y++) { \

  585.             buf[0] += buf[-(b->stride >> (1+b->pshift))]; \

  586.             \

  587.             for (x = 1; x < b->width; x++) { \

  588.                 int pred = buf[x - 1] + buf[x - (b->stride >> (1+b->pshift))] + buf[x - (b->stride >> (1+b->pshift))-1]; \

  589.                 buf[x]  += divide3(pred); \

  590.             } \

  591.             buf += (b->stride >> (1+b->pshift)); \

  592.         } \

  593.     } \

  594. 

  595. INTRA_DC_PRED(8, int16_t)

  596. INTRA_DC_PRED(10, int32_t)

  597. 

  598. /**

  599.  * Dirac Specification ->

  600.  * 13.4.2 Non-skipped subbands.  subband_coeffs()

  601.  */

  602. static av_always_inline void decode_subband_internal(DiracContext *s, SubBand *b, int is_arith)

  603. {

  604.     int cb_x, cb_y, left, right, top, bottom;

  605.     DiracArith c;

  606.     GetBitContext gb;

  607.     int cb_width  = s->codeblock[b->level + (b->orientation != subband_ll)].width;

  608.     int cb_height = s->codeblock[b->level + (b->orientation != subband_ll)].height;

  609.     int blockcnt_one = (cb_width + cb_height) == 2;

  610. 

  611.     if (!b->length)

  612.         return;

  613. 

  614.     init_get_bits8(&gb, b->coeff_data, b->length);

  615. 

  616.     if (is_arith)

  617.         ff_dirac_init_arith_decoder(&c, &gb, b->length);

  618. 

  619.     top = 0;

  620.     for (cb_y = 0; cb_y < cb_height; cb_y++) {

  621.         bottom = (b->height * (cb_y+1LL)) / cb_height;

  622.         left = 0;

  623.         for (cb_x = 0; cb_x < cb_width; cb_x++) {

  624.             right = (b->width * (cb_x+1LL)) / cb_width;

  625.             codeblock(s, b, &gb, &c, left, right, top, bottom, blockcnt_one, is_arith);

  626.             left = right;

  627.         }

  628.         top = bottom;

  629.     }

  630. 

  631.     if (b->orientation == subband_ll && s->num_refs == 0) {

  632.         if (s->pshift) {

  633.             intra_dc_prediction_10(b);

  634.         } else {

  635.             intra_dc_prediction_8(b);

  636.         }

  637.     }

  638. }

  639. 

  640. static int decode_subband_arith(AVCodecContext *avctx, void *b)

  641. {

  642.     DiracContext *s = avctx->priv_data;

  643.     decode_subband_internal(s, b, 1);

  644.     return 0;

  645. }

  646. 

  647. static int decode_subband_golomb(AVCodecContext *avctx, void *arg)

  648. {

  649.     DiracContext *s = avctx->priv_data;

  650.     SubBand **b     = arg;

  651.     decode_subband_internal(s, *b, 0);

  652.     return 0;

  653. }

  654. 

  655. /**

  656.  * Dirac Specification ->

  657.  * [DIRAC_STD] 13.4.1 core_transform_data()

  658.  */

  659. static void decode_component(DiracContext *s, int comp)

  660. {

  661.     AVCodecContext *avctx = s->avctx;

  662.     SubBand *bands[3*MAX_DWT_LEVELS+1];

  663.     enum dirac_subband orientation;

  664.     int level, num_bands = 0;

  665. 

  666.     /* Unpack all subbands at all levels. */

  667.     for (level = 0; level < s->wavelet_depth; level++) {

  668.         for (orientation = !!level; orientation < 4; orientation++) {

  669.             SubBand *b = &s->plane[comp].band[level][orientation];

  670.             bands[num_bands++] = b;

  671. 

  672.             align_get_bits(&s->gb);

  673.             /* [DIRAC_STD] 13.4.2 subband() */

  674.             b->length = get_interleaved_ue_golomb(&s->gb);

  675.             if (b->length) {

  676.                 b->quant = get_interleaved_ue_golomb(&s->gb);

  677.                 align_get_bits(&s->gb);

  678.                 b->coeff_data = s->gb.buffer + get_bits_count(&s->gb)/8;

  679.                 b->length = FFMIN(b->length, FFMAX(get_bits_left(&s->gb)/8, 0));

  680.                 skip_bits_long(&s->gb, b->length*8);

  681.             }

  682.         }

  683.         /* arithmetic coding has inter-level dependencies, so we can only execute one level at a time */

  684.         if (s->is_arith)

  685.             avctx->execute(avctx, decode_subband_arith, &s->plane[comp].band[level][!!level],

  686.                            NULL, 4-!!level, sizeof(SubBand));

  687.     }

  688.     /* golomb coding has no inter-level dependencies, so we can execute all subbands in parallel */

  689.     if (!s->is_arith)

  690.         avctx->execute(avctx, decode_subband_golomb, bands, NULL, num_bands, sizeof(SubBand*));

  691. }

  692. 

  693. #define PARSE_VALUES(type, x, gb, ebits, buf1, buf2) \

  694.     type *buf = (type *)buf1; \

  695.     buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset); \

  696.     if (get_bits_count(gb) >= ebits) \

  697.         return; \

  698.     if (buf2) { \

  699.         buf = (type *)buf2; \

  700.         buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset); \

  701.         if (get_bits_count(gb) >= ebits) \

  702.             return; \

  703.     } \

  704. 

  705. static void decode_subband(DiracContext *s, GetBitContext *gb, int quant,

  706.                            int slice_x, int slice_y, int bits_end,

  707.                            SubBand *b1, SubBand *b2)

  708. {

  709.     int left   = b1->width  * slice_x    / s->num_x;

  710.     int right  = b1->width  *(slice_x+1) / s->num_x;

  711.     int top    = b1->height * slice_y    / s->num_y;

  712.     int bottom = b1->height *(slice_y+1) / s->num_y;

  713. 

  714.     int qfactor, qoffset;

  715. 

  716.     uint8_t *buf1 =      b1->ibuf + top * b1->stride;

  717.     uint8_t *buf2 = b2 ? b2->ibuf + top * b2->stride: NULL;

  718.     int x, y;

  719. 

  720.     if (quant > 115) {

  721.         av_log(s->avctx, AV_LOG_ERROR, "Unsupported quant %d\n", quant);

  722.         return;

  723.     }

  724.     qfactor = ff_dirac_qscale_tab[quant & 0x7f];

  725.     qoffset = ff_dirac_qoffset_intra_tab[quant & 0x7f] + 2;

  726.     /* we have to constantly check for overread since the spec explicitly

  727.        requires this, with the meaning that all remaining coeffs are set to 0 */

  728.     if (get_bits_count(gb) >= bits_end)

  729.         return;

  730. 

  731.     if (s->pshift) {

  732.         for (y = top; y < bottom; y++) {

  733.             for (x = left; x < right; x++) {

  734.                 PARSE_VALUES(int32_t, x, gb, bits_end, buf1, buf2);

  735.             }

  736.             buf1 += b1->stride;

  737.             if (buf2)

  738.                 buf2 += b2->stride;

  739.         }

  740.     }

  741.     else {

  742.         for (y = top; y < bottom; y++) {

  743.             for (x = left; x < right; x++) {

  744.                 PARSE_VALUES(int16_t, x, gb, bits_end, buf1, buf2);

  745.             }

  746.             buf1 += b1->stride;

  747.             if (buf2)

  748.                 buf2 += b2->stride;

  749.         }

  750.     }

  751. }

  752. 

  753. /* Used by Low Delay and High Quality profiles */

  754. typedef struct DiracSlice {

  755.     GetBitContext gb;

  756.     int slice_x;

  757.     int slice_y;

  758.     int bytes;

  759. } DiracSlice;

  760. 

  761. 

  762. /**

  763.  * Dirac Specification ->

  764.  * 13.5.2 Slices. slice(sx,sy)

  765.  */

  766. static int decode_lowdelay_slice(AVCodecContext *avctx, void *arg)

  767. {

  768.     DiracContext *s = avctx->priv_data;

  769.     DiracSlice *slice = arg;

  770.     GetBitContext *gb = &slice->gb;

  771.     enum dirac_subband orientation;

  772.     int level, quant, chroma_bits, chroma_end;

  773. 

  774.     int quant_base  = get_bits(gb, 7); /*[DIRAC_STD] qindex */

  775.     int length_bits = av_log2(8 * slice->bytes)+1;

  776.     int luma_bits   = get_bits_long(gb, length_bits);

  777.     int luma_end    = get_bits_count(gb) + FFMIN(luma_bits, get_bits_left(gb));

  778. 

  779.     /* [DIRAC_STD] 13.5.5.2 luma_slice_band */

  780.     for (level = 0; level < s->wavelet_depth; level++)

  781.         for (orientation = !!level; orientation < 4; orientation++) {

  782.             quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);

  783.             decode_subband(s, gb, quant, slice->slice_x, slice->slice_y, luma_end,

  784.                            &s->plane[0].band[level][orientation], NULL);

  785.         }

  786. 

  787.     /* consume any unused bits from luma */

  788.     skip_bits_long(gb, get_bits_count(gb) - luma_end);

  789. 

  790.     chroma_bits = 8*slice->bytes - 7 - length_bits - luma_bits;

  791.     chroma_end  = get_bits_count(gb) + FFMIN(chroma_bits, get_bits_left(gb));

  792.     /* [DIRAC_STD] 13.5.5.3 chroma_slice_band */

  793.     for (level = 0; level < s->wavelet_depth; level++)

  794.         for (orientation = !!level; orientation < 4; orientation++) {

  795.             quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);

  796.             decode_subband(s, gb, quant, slice->slice_x, slice->slice_y, chroma_end,

  797.                            &s->plane[1].band[level][orientation],

  798.                            &s->plane[2].band[level][orientation]);

  799.         }

  800. 

  801.     return 0;

  802. }

  803. 

  804. /**

  805.  * VC-2 Specification ->

  806.  * 13.5.3 hq_slice(sx,sy)

  807.  */

  808. static int decode_hq_slice(AVCodecContext *avctx, void *arg)

  809. {

  810.     int i, quant, level, orientation, quant_idx;

  811.     uint8_t quants[MAX_DWT_LEVELS][4];

  812.     DiracContext *s = avctx->priv_data;

  813.     DiracSlice *slice = arg;

  814.     GetBitContext *gb = &slice->gb;

  815. 

  816.     skip_bits_long(gb, 8*s->highquality.prefix_bytes);

  817.     quant_idx = get_bits(gb, 8);

  818. 

  819.     /* Slice quantization (slice_quantizers() in the specs) */

  820.     for (level = 0; level < s->wavelet_depth; level++) {

  821.         for (orientation = !!level; orientation < 4; orientation++) {

  822.             quant = FFMAX(quant_idx - s->lowdelay.quant[level][orientation], 0);

  823.             quants[level][orientation] = quant;

  824.         }

  825.     }

  826. 

  827.     /* Luma + 2 Chroma planes */

  828.     for (i = 0; i < 3; i++) {

  829.         int64_t length = s->highquality.size_scaler * get_bits(gb, 8);

  830.         int64_t bits_left = 8 * length;

  831.         int64_t bits_end = get_bits_count(gb) + bits_left;

  832. 

  833.         if (bits_end >= INT_MAX) {

  834.             av_log(s->avctx, AV_LOG_ERROR, "end too far away\n");

  835.             return AVERROR_INVALIDDATA;

  836.         }

  837. 

  838.         for (level = 0; level < s->wavelet_depth; level++) {

  839.             for (orientation = !!level; orientation < 4; orientation++) {

  840.                 decode_subband(s, gb, quants[level][orientation], slice->slice_x, slice->slice_y, bits_end,

  841.                                &s->plane[i].band[level][orientation], NULL);

  842.             }

  843.         }

  844.         skip_bits_long(gb, bits_end - get_bits_count(gb));

  845.     }

  846. 

  847.     return 0;

  848. }

  849. 

  850. /**

  851.  * Dirac Specification ->

  852.  * 13.5.1 low_delay_transform_data()

  853.  */

  854. static int decode_lowdelay(DiracContext *s)

  855. {

  856.     AVCodecContext *avctx = s->avctx;

  857.     int slice_x, slice_y, bufsize;

  858.     int64_t bytes = 0;

  859.     const uint8_t *buf;

  860.     DiracSlice *slices;

  861.     int slice_num = 0;

  862. 

  863.     slices = av_mallocz_array(s->num_x, s->num_y * sizeof(DiracSlice));

  864.     if (!slices)

  865.         return AVERROR(ENOMEM);

  866. 

  867.     align_get_bits(&s->gb);

  868.     /*[DIRAC_STD] 13.5.2 Slices. slice(sx,sy) */

  869.     buf = s->gb.buffer + get_bits_count(&s->gb)/8;

  870.     bufsize = get_bits_left(&s->gb);

  871. 

  872.     if (s->hq_picture) {

  873.         int i;

  874. 

  875.         for (slice_y = 0; bufsize > 0 && slice_y < s->num_y; slice_y++) {

  876.             for (slice_x = 0; bufsize > 0 && slice_x < s->num_x; slice_x++) {

  877.                 bytes = s->highquality.prefix_bytes + 1;

  878.                 for (i = 0; i < 3; i++) {

  879.                     if (bytes <= bufsize/8)

  880.                         bytes += buf[bytes] * s->highquality.size_scaler + 1;

  881.                 }

  882.                 if (bytes >= INT_MAX) {

  883.                     av_log(s->avctx, AV_LOG_ERROR, "too many bytes\n");

  884.                     av_free(slices);

  885.                     return AVERROR_INVALIDDATA;

  886.                 }

  887. 

  888.                 slices[slice_num].bytes   = bytes;

  889.                 slices[slice_num].slice_x = slice_x;

  890.                 slices[slice_num].slice_y = slice_y;

  891.                 init_get_bits(&slices[slice_num].gb, buf, bufsize);

  892.                 slice_num++;

  893. 

  894.                 buf     += bytes;

  895.                 if (bufsize/8 >= bytes)

  896.                     bufsize -= bytes*8;

  897.                 else

  898.                     bufsize = 0;

  899.             }

  900.         }

  901.         avctx->execute(avctx, decode_hq_slice, slices, NULL, slice_num,

  902.                        sizeof(DiracSlice));

  903.     } else {

  904.         for (slice_y = 0; bufsize > 0 && slice_y < s->num_y; slice_y++) {

  905.             for (slice_x = 0; bufsize > 0 && slice_x < s->num_x; slice_x++) {

  906.                 bytes = (slice_num+1) * (int64_t)s->lowdelay.bytes.num / s->lowdelay.bytes.den

  907.                        - slice_num    * (int64_t)s->lowdelay.bytes.num / s->lowdelay.bytes.den;

  908.                 slices[slice_num].bytes   = bytes;

  909.                 slices[slice_num].slice_x = slice_x;

  910.                 slices[slice_num].slice_y = slice_y;

  911.                 init_get_bits(&slices[slice_num].gb, buf, bufsize);

  912.                 slice_num++;

  913. 

  914.                 buf     += bytes;

  915.                 if (bufsize/8 >= bytes)

  916.                     bufsize -= bytes*8;

  917.                 else

  918.                     bufsize = 0;

  919.             }

  920.         }

  921.         avctx->execute(avctx, decode_lowdelay_slice, slices, NULL, slice_num,

  922.                        sizeof(DiracSlice)); /* [DIRAC_STD] 13.5.2 Slices */

  923.     }

  924. 

  925.     if (s->dc_prediction) {

  926.         if (s->pshift) {

  927.             intra_dc_prediction_10(&s->plane[0].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */

  928.             intra_dc_prediction_10(&s->plane[1].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */

  929.             intra_dc_prediction_10(&s->plane[2].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */

  930.         } else {

  931.             intra_dc_prediction_8(&s->plane[0].band[0][0]);

  932.             intra_dc_prediction_8(&s->plane[1].band[0][0]);

  933.             intra_dc_prediction_8(&s->plane[2].band[0][0]);

  934.         }

  935.     }

  936.     av_free(slices);

  937.     return 0;

  938. }

  939. 

  940. static void init_planes(DiracContext *s)

  941. {

  942.     int i, w, h, level, orientation;

  943. 

  944.     for (i = 0; i < 3; i++) {

  945.         Plane *p = &s->plane[i];

  946. 

  947.         p->width       = s->seq.width  >> (i ? s->chroma_x_shift : 0);

  948.         p->height      = s->seq.height >> (i ? s->chroma_y_shift : 0);

  949.         p->idwt.width  = w = CALC_PADDING(p->width , s->wavelet_depth);

  950.         p->idwt.height = h = CALC_PADDING(p->height, s->wavelet_depth);

  951.         p->idwt.stride = FFALIGN(p->idwt.width, 8) << (1 + s->pshift);

  952. 

  953.         for (level = s->wavelet_depth-1; level >= 0; level--) {

  954.             w = w>>1;

  955.             h = h>>1;

  956.             for (orientation = !!level; orientation < 4; orientation++) {

  957.                 SubBand *b = &p->band[level][orientation];

  958. 

  959.                 b->pshift = s->pshift;

  960.                 b->ibuf   = p->idwt.buf;

  961.                 b->level  = level;

  962.                 b->stride = p->idwt.stride << (s->wavelet_depth - level);

  963.                 b->width  = w;

  964.                 b->height = h;

  965.                 b->orientation = orientation;

  966. 

  967.                 if (orientation & 1)

  968.                     b->ibuf += w << (1+b->pshift);

  969.                 if (orientation > 1)

  970.                     b->ibuf += (b->stride>>1);

  971. 

  972.                 if (level)

  973.                     b->parent = &p->band[level-1][orientation];

  974.             }

  975.         }

  976. 

  977.         if (i > 0) {

  978.             p->xblen = s->plane[0].xblen >> s->chroma_x_shift;

  979.             p->yblen = s->plane[0].yblen >> s->chroma_y_shift;

  980.             p->xbsep = s->plane[0].xbsep >> s->chroma_x_shift;

  981.             p->ybsep = s->plane[0].ybsep >> s->chroma_y_shift;

  982.         }

  983. 

  984.         p->xoffset = (p->xblen - p->xbsep)/2;

  985.         p->yoffset = (p->yblen - p->ybsep)/2;

  986.     }

  987. }

  988. 

  989. /**

  990.  * Unpack the motion compensation parameters

  991.  * Dirac Specification ->

  992.  * 11.2 Picture prediction data. picture_prediction()

  993.  */

  994. static int dirac_unpack_prediction_parameters(DiracContext *s)

  995. {

  996.     static const uint8_t default_blen[] = { 4, 12, 16, 24 };

  997. 

  998.     GetBitContext *gb = &s->gb;

  999.     unsigned idx, ref;

  1000. 

  1001.     align_get_bits(gb);

  1002.     /* [DIRAC_STD] 11.2.2 Block parameters. block_parameters() */

  1003.     /* Luma and Chroma are equal. 11.2.3 */

  1004.     idx = get_interleaved_ue_golomb(gb); /* [DIRAC_STD] index */

  1005. 

  1006.     if (idx > 4) {

  1007.         av_log(s->avctx, AV_LOG_ERROR, "Block prediction index too high\n");

  1008.         return AVERROR_INVALIDDATA;

  1009.     }

  1010. 

  1011.     if (idx == 0) {

  1012.         s->plane[0].xblen = get_interleaved_ue_golomb(gb);

  1013.         s->plane[0].yblen = get_interleaved_ue_golomb(gb);

  1014.         s->plane[0].xbsep = get_interleaved_ue_golomb(gb);

  1015.         s->plane[0].ybsep = get_interleaved_ue_golomb(gb);

  1016.     } else {

  1017.         /*[DIRAC_STD] preset_block_params(index). Table 11.1 */

  1018.         s->plane[0].xblen = default_blen[idx-1];

  1019.         s->plane[0].yblen = default_blen[idx-1];

  1020.         s->plane[0].xbsep = 4 * idx;

  1021.         s->plane[0].ybsep = 4 * idx;

  1022.     }

  1023.     /*[DIRAC_STD] 11.2.4 motion_data_dimensions()

  1024.       Calculated in function dirac_unpack_block_motion_data */

  1025. 

  1026.     if (s->plane[0].xblen % (1 << s->chroma_x_shift) != 0 ||

  1027.         s->plane[0].yblen % (1 << s->chroma_y_shift) != 0 ||

  1028.         !s->plane[0].xblen || !s->plane[0].yblen) {

  1029.         av_log(s->avctx, AV_LOG_ERROR,

  1030.                "invalid x/y block length (%d/%d) for x/y chroma shift (%d/%d)\n",

  1031.                s->plane[0].xblen, s->plane[0].yblen, s->chroma_x_shift, s->chroma_y_shift);

  1032.         return AVERROR_INVALIDDATA;

  1033.     }

  1034.     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) {

  1035.         av_log(s->avctx, AV_LOG_ERROR, "Block separation too small\n");

  1036.         return AVERROR_INVALIDDATA;

  1037.     }

  1038.     if (s->plane[0].xbsep > s->plane[0].xblen || s->plane[0].ybsep > s->plane[0].yblen) {

  1039.         av_log(s->avctx, AV_LOG_ERROR, "Block separation greater than size\n");

  1040.         return AVERROR_INVALIDDATA;

  1041.     }

  1042.     if (FFMAX(s->plane[0].xblen, s->plane[0].yblen) > MAX_BLOCKSIZE) {

  1043.         av_log(s->avctx, AV_LOG_ERROR, "Unsupported large block size\n");

  1044.         return AVERROR_PATCHWELCOME;

  1045.     }

  1046. 

  1047.     /*[DIRAC_STD] 11.2.5 Motion vector precision. motion_vector_precision()

  1048.       Read motion vector precision */

  1049.     s->mv_precision = get_interleaved_ue_golomb(gb);

  1050.     if (s->mv_precision > 3) {

  1051.         av_log(s->avctx, AV_LOG_ERROR, "MV precision finer than eighth-pel\n");

  1052.         return AVERROR_INVALIDDATA;

  1053.     }

  1054. 

  1055.     /*[DIRAC_STD] 11.2.6 Global motion. global_motion()

  1056.       Read the global motion compensation parameters */

  1057.     s->globalmc_flag = get_bits1(gb);

  1058.     if (s->globalmc_flag) {

  1059.         memset(s->globalmc, 0, sizeof(s->globalmc));

  1060.         /* [DIRAC_STD] pan_tilt(gparams) */

  1061.         for (ref = 0; ref < s->num_refs; ref++) {

  1062.             if (get_bits1(gb)) {

  1063.                 s->globalmc[ref].pan_tilt[0] = dirac_get_se_golomb(gb);

  1064.                 s->globalmc[ref].pan_tilt[1] = dirac_get_se_golomb(gb);

  1065.             }

  1066.             /* [DIRAC_STD] zoom_rotate_shear(gparams)

  1067.                zoom/rotation/shear parameters */

  1068.             if (get_bits1(gb)) {

  1069.                 s->globalmc[ref].zrs_exp   = get_interleaved_ue_golomb(gb);

  1070.                 s->globalmc[ref].zrs[0][0] = dirac_get_se_golomb(gb);

  1071.                 s->globalmc[ref].zrs[0][1] = dirac_get_se_golomb(gb);

  1072.                 s->globalmc[ref].zrs[1][0] = dirac_get_se_golomb(gb);

  1073.                 s->globalmc[ref].zrs[1][1] = dirac_get_se_golomb(gb);

  1074.             } else {

  1075.                 s->globalmc[ref].zrs[0][0] = 1;

  1076.                 s->globalmc[ref].zrs[1][1] = 1;

  1077.             }

  1078.             /* [DIRAC_STD] perspective(gparams) */

  1079.             if (get_bits1(gb)) {

  1080.                 s->globalmc[ref].perspective_exp = get_interleaved_ue_golomb(gb);

  1081.                 s->globalmc[ref].perspective[0]  = dirac_get_se_golomb(gb);

  1082.                 s->globalmc[ref].perspective[1]  = dirac_get_se_golomb(gb);

  1083.             }

  1084.         }

  1085.     }

  1086. 

  1087.     /*[DIRAC_STD] 11.2.7 Picture prediction mode. prediction_mode()

  1088.       Picture prediction mode, not currently used. */

  1089.     if (get_interleaved_ue_golomb(gb)) {

  1090.         av_log(s->avctx, AV_LOG_ERROR, "Unknown picture prediction mode\n");

  1091.         return AVERROR_INVALIDDATA;

  1092.     }

  1093. 

  1094.     /* [DIRAC_STD] 11.2.8 Reference picture weight. reference_picture_weights()

  1095.        just data read, weight calculation will be done later on. */

  1096.     s->weight_log2denom = 1;

  1097.     s->weight[0]        = 1;

  1098.     s->weight[1]        = 1;

  1099. 

  1100.     if (get_bits1(gb)) {

  1101.         s->weight_log2denom = get_interleaved_ue_golomb(gb);

  1102.         s->weight[0] = dirac_get_se_golomb(gb);

  1103.         if (s->num_refs == 2)

  1104.             s->weight[1] = dirac_get_se_golomb(gb);

  1105.     }

  1106.     return 0;

  1107. }

  1108. 

  1109. /**

  1110.  * Dirac Specification ->

  1111.  * 11.3 Wavelet transform data. wavelet_transform()

  1112.  */

  1113. static int dirac_unpack_idwt_params(DiracContext *s)

  1114. {

  1115.     GetBitContext *gb = &s->gb;

  1116.     int i, level;

  1117.     unsigned tmp;

  1118. 

  1119. #define CHECKEDREAD(dst, cond, errmsg) \

  1120.     tmp = get_interleaved_ue_golomb(gb); \

  1121.     if (cond) { \

  1122.         av_log(s->avctx, AV_LOG_ERROR, errmsg); \

  1123.         return AVERROR_INVALIDDATA; \

  1124.     }\

  1125.     dst = tmp;

  1126. 

  1127.     align_get_bits(gb);

  1128. 

  1129.     s->zero_res = s->num_refs ? get_bits1(gb) : 0;

  1130.     if (s->zero_res)

  1131.         return 0;

  1132. 

  1133.     /*[DIRAC_STD] 11.3.1 Transform parameters. transform_parameters() */

  1134.     CHECKEDREAD(s->wavelet_idx, tmp > 6, "wavelet_idx is too big\n")

  1135. 

  1136.     CHECKEDREAD(s->wavelet_depth, tmp > MAX_DWT_LEVELS || tmp < 1, "invalid number of DWT decompositions\n")

  1137. 

  1138.     if (!s->low_delay) {

  1139.         /* Codeblock parameters (core syntax only) */

  1140.         if (get_bits1(gb)) {

  1141.             for (i = 0; i <= s->wavelet_depth; i++) {

  1142.                 CHECKEDREAD(s->codeblock[i].width , tmp < 1 || tmp > (s->avctx->width >>s->wavelet_depth-i), "codeblock width invalid\n")

  1143.                 CHECKEDREAD(s->codeblock[i].height, tmp < 1 || tmp > (s->avctx->height>>s->wavelet_depth-i), "codeblock height invalid\n")

  1144.             }

  1145. 

  1146.             CHECKEDREAD(s->codeblock_mode, tmp > 1, "unknown codeblock mode\n")

  1147.         }

  1148.         else {

  1149.             for (i = 0; i <= s->wavelet_depth; i++)

  1150.                 s->codeblock[i].width = s->codeblock[i].height = 1;

  1151.         }

  1152.     }

  1153.     else {

  1154.         s->num_x        = get_interleaved_ue_golomb(gb);

  1155.         s->num_y        = get_interleaved_ue_golomb(gb);

  1156.         if (s->ld_picture) {

  1157.             s->lowdelay.bytes.num = get_interleaved_ue_golomb(gb);

  1158.             s->lowdelay.bytes.den = get_interleaved_ue_golomb(gb);

  1159.             if (s->lowdelay.bytes.den <= 0) {

  1160.                 av_log(s->avctx,AV_LOG_ERROR,"Invalid lowdelay.bytes.den\n");

  1161.                 return AVERROR_INVALIDDATA;

  1162.             }

  1163.         } else if (s->hq_picture) {

  1164.             s->highquality.prefix_bytes = get_interleaved_ue_golomb(gb);

  1165.             s->highquality.size_scaler  = get_interleaved_ue_golomb(gb);

  1166.             if (s->highquality.prefix_bytes >= INT_MAX / 8) {

  1167.                 av_log(s->avctx,AV_LOG_ERROR,"too many prefix bytes\n");

  1168.                 return AVERROR_INVALIDDATA;

  1169.             }

  1170.         }

  1171. 

  1172.         /* [DIRAC_STD] 11.3.5 Quantisation matrices (low-delay syntax). quant_matrix() */

  1173.         if (get_bits1(gb)) {

  1174.             av_log(s->avctx,AV_LOG_DEBUG,"Low Delay: Has Custom Quantization Matrix!\n");

  1175.             /* custom quantization matrix */

  1176.             s->lowdelay.quant[0][0] = get_interleaved_ue_golomb(gb);

  1177.             for (level = 0; level < s->wavelet_depth; level++) {

  1178.                 s->lowdelay.quant[level][1] = get_interleaved_ue_golomb(gb);

  1179.                 s->lowdelay.quant[level][2] = get_interleaved_ue_golomb(gb);

  1180.                 s->lowdelay.quant[level][3] = get_interleaved_ue_golomb(gb);

  1181.             }

  1182.         } else {

  1183.             if (s->wavelet_depth > 4) {

  1184.                 av_log(s->avctx,AV_LOG_ERROR,"Mandatory custom low delay matrix missing for depth %d\n", s->wavelet_depth);

  1185.                 return AVERROR_INVALIDDATA;

  1186.             }

  1187.             /* default quantization matrix */

  1188.             for (level = 0; level < s->wavelet_depth; level++)

  1189.                 for (i = 0; i < 4; i++) {

  1190.                     s->lowdelay.quant[level][i] = ff_dirac_default_qmat[s->wavelet_idx][level][i];

  1191.                     /* haar with no shift differs for different depths */

  1192.                     if (s->wavelet_idx == 3)

  1193.                         s->lowdelay.quant[level][i] += 4*(s->wavelet_depth-1 - level);

  1194.                 }

  1195.         }

  1196.     }

  1197.     return 0;

  1198. }

  1199. 

  1200. static inline int pred_sbsplit(uint8_t *sbsplit, int stride, int x, int y)

  1201. {

  1202.     static const uint8_t avgsplit[7] = { 0, 0, 1, 1, 1, 2, 2 };

  1203. 

  1204.     if (!(x|y))

  1205.         return 0;

  1206.     else if (!y)

  1207.         return sbsplit[-1];

  1208.     else if (!x)

  1209.         return sbsplit[-stride];

  1210. 

  1211.     return avgsplit[sbsplit[-1] + sbsplit[-stride] + sbsplit[-stride-1]];

  1212. }

  1213. 

  1214. static inline int pred_block_mode(DiracBlock *block, int stride, int x, int y, int refmask)

  1215. {

  1216.     int pred;

  1217. 

  1218.     if (!(x|y))

  1219.         return 0;

  1220.     else if (!y)

  1221.         return block[-1].ref & refmask;

  1222.     else if (!x)

  1223.         return block[-stride].ref & refmask;

  1224. 

  1225.     /* return the majority */

  1226.     pred = (block[-1].ref & refmask) + (block[-stride].ref & refmask) + (block[-stride-1].ref & refmask);

  1227.     return (pred >> 1) & refmask;

  1228. }

  1229. 

  1230. static inline void pred_block_dc(DiracBlock *block, int stride, int x, int y)

  1231. {

  1232.     int i, n = 0;

  1233. 

  1234.     memset(block->u.dc, 0, sizeof(block->u.dc));

  1235. 

  1236.     if (x && !(block[-1].ref & 3)) {

  1237.         for (i = 0; i < 3; i++)

  1238.             block->u.dc[i] += block[-1].u.dc[i];

  1239.         n++;

  1240.     }

  1241. 

  1242.     if (y && !(block[-stride].ref & 3)) {

  1243.         for (i = 0; i < 3; i++)

  1244.             block->u.dc[i] += block[-stride].u.dc[i];

  1245.         n++;

  1246.     }

  1247. 

  1248.     if (x && y && !(block[-1-stride].ref & 3)) {

  1249.         for (i = 0; i < 3; i++)

  1250.             block->u.dc[i] += block[-1-stride].u.dc[i];

  1251.         n++;

  1252.     }

  1253. 

  1254.     if (n == 2) {

  1255.         for (i = 0; i < 3; i++)

  1256.             block->u.dc[i] = (block->u.dc[i]+1)>>1;

  1257.     } else if (n == 3) {

  1258.         for (i = 0; i < 3; i++)

  1259.             block->u.dc[i] = divide3(block->u.dc[i]);

  1260.     }

  1261. }

  1262. 

  1263. static inline void pred_mv(DiracBlock *block, int stride, int x, int y, int ref)

  1264. {

  1265.     int16_t *pred[3];

  1266.     int refmask = ref+1;

  1267.     int mask = refmask | DIRAC_REF_MASK_GLOBAL; /*  exclude gmc blocks */

  1268.     int n = 0;

  1269. 

  1270.     if (x && (block[-1].ref & mask) == refmask)

  1271.         pred[n++] = block[-1].u.mv[ref];

  1272. 

  1273.     if (y && (block[-stride].ref & mask) == refmask)

  1274.         pred[n++] = block[-stride].u.mv[ref];

  1275. 

  1276.     if (x && y && (block[-stride-1].ref & mask) == refmask)

  1277.         pred[n++] = block[-stride-1].u.mv[ref];

  1278. 

  1279.     switch (n) {

  1280.     case 0:

  1281.         block->u.mv[ref][0] = 0;

  1282.         block->u.mv[ref][1] = 0;

  1283.         break;

  1284.     case 1:

  1285.         block->u.mv[ref][0] = pred[0][0];

  1286.         block->u.mv[ref][1] = pred[0][1];

  1287.         break;

  1288.     case 2:

  1289.         block->u.mv[ref][0] = (pred[0][0] + pred[1][0] + 1) >> 1;

  1290.         block->u.mv[ref][1] = (pred[0][1] + pred[1][1] + 1) >> 1;

  1291.         break;

  1292.     case 3:

  1293.         block->u.mv[ref][0] = mid_pred(pred[0][0], pred[1][0], pred[2][0]);

  1294.         block->u.mv[ref][1] = mid_pred(pred[0][1], pred[1][1], pred[2][1]);

  1295.         break;

  1296.     }

  1297. }

  1298. 

  1299. static void global_mv(DiracContext *s, DiracBlock *block, int x, int y, int ref)

  1300. {

  1301.     int ez      = s->globalmc[ref].zrs_exp;

  1302.     int ep      = s->globalmc[ref].perspective_exp;

  1303.     int (*A)[2] = s->globalmc[ref].zrs;

  1304.     int *b      = s->globalmc[ref].pan_tilt;

  1305.     int *c      = s->globalmc[ref].perspective;

  1306. 

  1307.     int m       = (1<<ep) - (c[0]*x + c[1]*y);

  1308.     int mx      = m * ((A[0][0] * x + A[0][1]*y) + (1<<ez) * b[0]);

  1309.     int my      = m * ((A[1][0] * x + A[1][1]*y) + (1<<ez) * b[1]);

  1310. 

  1311.     block->u.mv[ref][0] = (mx + (1<<(ez+ep))) >> (ez+ep);

  1312.     block->u.mv[ref][1] = (my + (1<<(ez+ep))) >> (ez+ep);

  1313. }

  1314. 

  1315. static void decode_block_params(DiracContext *s, DiracArith arith[8], DiracBlock *block,

  1316.                                 int stride, int x, int y)

  1317. {

  1318.     int i;

  1319. 

  1320.     block->ref  = pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF1);

  1321.     block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF1);

  1322. 

  1323.     if (s->num_refs == 2) {

  1324.         block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF2);

  1325.         block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF2) << 1;

  1326.     }

  1327. 

  1328.     if (!block->ref) {

  1329.         pred_block_dc(block, stride, x, y);

  1330.         for (i = 0; i < 3; i++)

  1331.             block->u.dc[i] += dirac_get_arith_int(arith+1+i, CTX_DC_F1, CTX_DC_DATA);

  1332.         return;

  1333.     }

  1334. 

  1335.     if (s->globalmc_flag) {

  1336.         block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_GLOBAL);

  1337.         block->ref ^= dirac_get_arith_bit(arith, CTX_GLOBAL_BLOCK) << 2;

  1338.     }

  1339. 

  1340.     for (i = 0; i < s->num_refs; i++)

  1341.         if (block->ref & (i+1)) {

  1342.             if (block->ref & DIRAC_REF_MASK_GLOBAL) {

  1343.                 global_mv(s, block, x, y, i);

  1344.             } else {

  1345.                 pred_mv(block, stride, x, y, i);

  1346.                 block->u.mv[i][0] += dirac_get_arith_int(arith + 4 + 2 * i, CTX_MV_F1, CTX_MV_DATA);

  1347.                 block->u.mv[i][1] += dirac_get_arith_int(arith + 5 + 2 * i, CTX_MV_F1, CTX_MV_DATA);

  1348.             }

  1349.         }

  1350. }

  1351. 

  1352. /**

  1353.  * Copies the current block to the other blocks covered by the current superblock split mode

  1354.  */

  1355. static void propagate_block_data(DiracBlock *block, int stride, int size)

  1356. {

  1357.     int x, y;

  1358.     DiracBlock *dst = block;

  1359. 

  1360.     for (x = 1; x < size; x++)

  1361.         dst[x] = *block;

  1362. 

  1363.     for (y = 1; y < size; y++) {

  1364.         dst += stride;

  1365.         for (x = 0; x < size; x++)

  1366.             dst[x] = *block;

  1367.     }

  1368. }

  1369. 

  1370. /**

  1371.  * Dirac Specification ->

  1372.  * 12. Block motion data syntax

  1373.  */

  1374. static int dirac_unpack_block_motion_data(DiracContext *s)

  1375. {

  1376.     GetBitContext *gb = &s->gb;

  1377.     uint8_t *sbsplit = s->sbsplit;

  1378.     int i, x, y, q, p;

  1379.     DiracArith arith[8];

  1380. 

  1381.     align_get_bits(gb);

  1382. 

  1383.     /* [DIRAC_STD] 11.2.4 and 12.2.1 Number of blocks and superblocks */

  1384.     s->sbwidth  = DIVRNDUP(s->seq.width,  4*s->plane[0].xbsep);

  1385.     s->sbheight = DIVRNDUP(s->seq.height, 4*s->plane[0].ybsep);

  1386.     s->blwidth  = 4 * s->sbwidth;

  1387.     s->blheight = 4 * s->sbheight;

  1388. 

  1389.     /* [DIRAC_STD] 12.3.1 Superblock splitting modes. superblock_split_modes()

  1390.        decode superblock split modes */

  1391.     ff_dirac_init_arith_decoder(arith, gb, get_interleaved_ue_golomb(gb));     /* get_interleaved_ue_golomb(gb) is the length */

  1392.     for (y = 0; y < s->sbheight; y++) {

  1393.         for (x = 0; x < s->sbwidth; x++) {

  1394.             unsigned int split  = dirac_get_arith_uint(arith, CTX_SB_F1, CTX_SB_DATA);

  1395.             if (split > 2)

  1396.                 return AVERROR_INVALIDDATA;

  1397.             sbsplit[x] = (split + pred_sbsplit(sbsplit+x, s->sbwidth, x, y)) % 3;

  1398.         }

  1399.         sbsplit += s->sbwidth;

  1400.     }

  1401. 

  1402.     /* setup arith decoding */

  1403.     ff_dirac_init_arith_decoder(arith, gb, get_interleaved_ue_golomb(gb));

  1404.     for (i = 0; i < s->num_refs; i++) {

  1405.         ff_dirac_init_arith_decoder(arith + 4 + 2 * i, gb, get_interleaved_ue_golomb(gb));

  1406.         ff_dirac_init_arith_decoder(arith + 5 + 2 * i, gb, get_interleaved_ue_golomb(gb));

  1407.     }

  1408.     for (i = 0; i < 3; i++)

  1409.         ff_dirac_init_arith_decoder(arith+1+i, gb, get_interleaved_ue_golomb(gb));

  1410. 

  1411.     for (y = 0; y < s->sbheight; y++)

  1412.         for (x = 0; x < s->sbwidth; x++) {

  1413.             int blkcnt = 1 << s->sbsplit[y * s->sbwidth + x];

  1414.             int step   = 4 >> s->sbsplit[y * s->sbwidth + x];

  1415. 

  1416.             for (q = 0; q < blkcnt; q++)

  1417.                 for (p = 0; p < blkcnt; p++) {

  1418.                     int bx = 4 * x + p*step;

  1419.                     int by = 4 * y + q*step;

  1420.                     DiracBlock *block = &s->blmotion[by*s->blwidth + bx];

  1421.                     decode_block_params(s, arith, block, s->blwidth, bx, by);

  1422.                     propagate_block_data(block, s->blwidth, step);

  1423.                 }

  1424.         }

  1425. 

  1426.     return 0;

  1427. }

  1428. 

  1429. static int weight(int i, int blen, int offset)

  1430. {

  1431. #define ROLLOFF(i) offset == 1 ? ((i) ? 5 : 3) :        \

  1432.     (1 + (6*(i) + offset - 1) / (2*offset - 1))

  1433. 

  1434.     if (i < 2*offset)

  1435.         return ROLLOFF(i);

  1436.     else if (i > blen-1 - 2*offset)

  1437.         return ROLLOFF(blen-1 - i);

  1438.     return 8;

  1439. }

  1440. 

  1441. static void init_obmc_weight_row(Plane *p, uint8_t *obmc_weight, int stride,

  1442.                                  int left, int right, int wy)

  1443. {

  1444.     int x;

  1445.     for (x = 0; left && x < p->xblen >> 1; x++)

  1446.         obmc_weight[x] = wy*8;

  1447.     for (; x < p->xblen >> right; x++)

  1448.         obmc_weight[x] = wy*weight(x, p->xblen, p->xoffset);

  1449.     for (; x < p->xblen; x++)

  1450.         obmc_weight[x] = wy*8;

  1451.     for (; x < stride; x++)

  1452.         obmc_weight[x] = 0;

  1453. }

  1454. 

  1455. static void init_obmc_weight(Plane *p, uint8_t *obmc_weight, int stride,

  1456.                              int left, int right, int top, int bottom)

  1457. {

  1458.     int y;

  1459.     for (y = 0; top && y < p->yblen >> 1; y++) {

  1460.         init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);

  1461.         obmc_weight += stride;

  1462.     }

  1463.     for (; y < p->yblen >> bottom; y++) {

  1464.         int wy = weight(y, p->yblen, p->yoffset);

  1465.         init_obmc_weight_row(p, obmc_weight, stride, left, right, wy);

  1466.         obmc_weight += stride;

  1467.     }

  1468.     for (; y < p->yblen; y++) {

  1469.         init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);

  1470.         obmc_weight += stride;

  1471.     }

  1472. }

  1473. 

  1474. static void init_obmc_weights(DiracContext *s, Plane *p, int by)

  1475. {

  1476.     int top = !by;

  1477.     int bottom = by == s->blheight-1;

  1478. 

  1479.     /* don't bother re-initing for rows 2 to blheight-2, the weights don't change */

  1480.     if (top || bottom || by == 1) {

  1481.         init_obmc_weight(p, s->obmc_weight[0], MAX_BLOCKSIZE, 1, 0, top, bottom);

  1482.         init_obmc_weight(p, s->obmc_weight[1], MAX_BLOCKSIZE, 0, 0, top, bottom);

  1483.         init_obmc_weight(p, s->obmc_weight[2], MAX_BLOCKSIZE, 0, 1, top, bottom);

  1484.     }

  1485. }

  1486. 

  1487. static const uint8_t epel_weights[4][4][4] = {

  1488.     {{ 16,  0,  0,  0 },

  1489.      { 12,  4,  0,  0 },

  1490.      {  8,  8,  0,  0 },

  1491.      {  4, 12,  0,  0 }},

  1492.     {{ 12,  0,  4,  0 },

  1493.      {  9,  3,  3,  1 },

  1494.      {  6,  6,  2,  2 },

  1495.      {  3,  9,  1,  3 }},

  1496.     {{  8,  0,  8,  0 },

  1497.      {  6,  2,  6,  2 },

  1498.      {  4,  4,  4,  4 },

  1499.      {  2,  6,  2,  6 }},

  1500.     {{  4,  0, 12,  0 },

  1501.      {  3,  1,  9,  3 },

  1502.      {  2,  2,  6,  6 },

  1503.      {  1,  3,  3,  9 }}

  1504. };

  1505. 

  1506. /**

  1507.  * For block x,y, determine which of the hpel planes to do bilinear

  1508.  * interpolation from and set src[] to the location in each hpel plane

  1509.  * to MC from.

  1510.  *

  1511.  * @return the index of the put_dirac_pixels_tab function to use

  1512.  *  0 for 1 plane (fpel,hpel), 1 for 2 planes (qpel), 2 for 4 planes (qpel), and 3 for epel

  1513.  */

  1514. static int mc_subpel(DiracContext *s, DiracBlock *block, const uint8_t *src[5],

  1515.                      int x, int y, int ref, int plane)

  1516. {

  1517.     Plane *p = &s->plane[plane];

  1518.     uint8_t **ref_hpel = s->ref_pics[ref]->hpel[plane];

  1519.     int motion_x = block->u.mv[ref][0];

  1520.     int motion_y = block->u.mv[ref][1];

  1521.     int mx, my, i, epel, nplanes = 0;

  1522. 

  1523.     if (plane) {

  1524.         motion_x >>= s->chroma_x_shift;

  1525.         motion_y >>= s->chroma_y_shift;

  1526.     }

  1527. 

  1528.     mx         = motion_x & ~(-1U << s->mv_precision);

  1529.     my         = motion_y & ~(-1U << s->mv_precision);

  1530.     motion_x >>= s->mv_precision;

  1531.     motion_y >>= s->mv_precision;

  1532.     /* normalize subpel coordinates to epel */

  1533.     /* TODO: template this function? */

  1534.     mx      <<= 3 - s->mv_precision;

  1535.     my      <<= 3 - s->mv_precision;

  1536. 

  1537.     x += motion_x;

  1538.     y += motion_y;

  1539.     epel = (mx|my)&1;

  1540. 

  1541.     /* hpel position */

  1542.     if (!((mx|my)&3)) {

  1543.         nplanes = 1;

  1544.         src[0] = ref_hpel[(my>>1)+(mx>>2)] + y*p->stride + x;

  1545.     } else {

  1546.         /* qpel or epel */

  1547.         nplanes = 4;

  1548.         for (i = 0; i < 4; i++)

  1549.             src[i] = ref_hpel[i] + y*p->stride + x;

  1550. 

  1551.         /* if we're interpolating in the right/bottom halves, adjust the planes as needed

  1552.            we increment x/y because the edge changes for half of the pixels */

  1553.         if (mx > 4) {

  1554.             src[0] += 1;

  1555.             src[2] += 1;

  1556.             x++;

  1557.         }

  1558.         if (my > 4) {

  1559.             src[0] += p->stride;

  1560.             src[1] += p->stride;

  1561.             y++;

  1562.         }

  1563. 

  1564.         /* hpel planes are:

  1565.            [0]: F  [1]: H

  1566.            [2]: V  [3]: C */

  1567.         if (!epel) {

  1568.             /* check if we really only need 2 planes since either mx or my is

  1569.                a hpel position. (epel weights of 0 handle this there) */

  1570.             if (!(mx&3)) {

  1571.                 /* mx == 0: average [0] and [2]

  1572.                    mx == 4: average [1] and [3] */

  1573.                 src[!mx] = src[2 + !!mx];

  1574.                 nplanes = 2;

  1575.             } else if (!(my&3)) {

  1576.                 src[0] = src[(my>>1)  ];

  1577.                 src[1] = src[(my>>1)+1];

  1578.                 nplanes = 2;

  1579.             }

  1580.         } else {

  1581.             /* adjust the ordering if needed so the weights work */

  1582.             if (mx > 4) {

  1583.                 FFSWAP(const uint8_t *, src[0], src[1]);

  1584.                 FFSWAP(const uint8_t *, src[2], src[3]);

  1585.             }

  1586.             if (my > 4) {

  1587.                 FFSWAP(const uint8_t *, src[0], src[2]);

  1588.                 FFSWAP(const uint8_t *, src[1], src[3]);

  1589.             }

  1590.             src[4] = epel_weights[my&3][mx&3];

  1591.         }

  1592.     }

  1593. 

  1594.     /* fixme: v/h _edge_pos */

  1595.     if (x + p->xblen > p->width +EDGE_WIDTH/2 ||

  1596.         y + p->yblen > p->height+EDGE_WIDTH/2 ||

  1597.         x < 0 || y < 0) {

  1598.         for (i = 0; i < nplanes; i++) {

  1599.             s->vdsp.emulated_edge_mc(s->edge_emu_buffer[i], src[i],

  1600.                                      p->stride, p->stride,

  1601.                                      p->xblen, p->yblen, x, y,

  1602.                                      p->width+EDGE_WIDTH/2, p->height+EDGE_WIDTH/2);

  1603.             src[i] = s->edge_emu_buffer[i];

  1604.         }

  1605.     }

  1606.     return (nplanes>>1) + epel;

  1607. }

  1608. 

  1609. static void add_dc(uint16_t *dst, int dc, int stride,

  1610.                    uint8_t *obmc_weight, int xblen, int yblen)

  1611. {

  1612.     int x, y;

  1613.     dc += 128;

  1614. 

  1615.     for (y = 0; y < yblen; y++) {

  1616.         for (x = 0; x < xblen; x += 2) {

  1617.             dst[x  ] += dc * obmc_weight[x  ];

  1618.             dst[x+1] += dc * obmc_weight[x+1];

  1619.         }

  1620.         dst          += stride;

  1621.         obmc_weight  += MAX_BLOCKSIZE;

  1622.     }

  1623. }

  1624. 

  1625. static void block_mc(DiracContext *s, DiracBlock *block,

  1626.                      uint16_t *mctmp, uint8_t *obmc_weight,

  1627.                      int plane, int dstx, int dsty)

  1628. {

  1629.     Plane *p = &s->plane[plane];

  1630.     const uint8_t *src[5];

  1631.     int idx;

  1632. 

  1633.     switch (block->ref&3) {

  1634.     case 0: /* DC */

  1635.         add_dc(mctmp, block->u.dc[plane], p->stride, obmc_weight, p->xblen, p->yblen);

  1636.         return;

  1637.     case 1:

  1638.     case 2:

  1639.         idx = mc_subpel(s, block, src, dstx, dsty, (block->ref&3)-1, plane);

  1640.         s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);

  1641.         if (s->weight_func)

  1642.             s->weight_func(s->mcscratch, p->stride, s->weight_log2denom,

  1643.                            s->weight[0] + s->weight[1], p->yblen);

  1644.         break;

  1645.     case 3:

  1646.         idx = mc_subpel(s, block, src, dstx, dsty, 0, plane);

  1647.         s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);

  1648.         idx = mc_subpel(s, block, src, dstx, dsty, 1, plane);

  1649.         if (s->biweight_func) {

  1650.             /* fixme: +32 is a quick hack */

  1651.             s->put_pixels_tab[idx](s->mcscratch + 32, src, p->stride, p->yblen);

  1652.             s->biweight_func(s->mcscratch, s->mcscratch+32, p->stride, s->weight_log2denom,

  1653.                              s->weight[0], s->weight[1], p->yblen);

  1654.         } else

  1655.             s->avg_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);

  1656.         break;

  1657.     }

  1658.     s->add_obmc(mctmp, s->mcscratch, p->stride, obmc_weight, p->yblen);

  1659. }

  1660. 

  1661. static void mc_row(DiracContext *s, DiracBlock *block, uint16_t *mctmp, int plane, int dsty)

  1662. {

  1663.     Plane *p = &s->plane[plane];

  1664.     int x, dstx = p->xbsep - p->xoffset;

  1665. 

  1666.     block_mc(s, block, mctmp, s->obmc_weight[0], plane, -p->xoffset, dsty);

  1667.     mctmp += p->xbsep;

  1668. 

  1669.     for (x = 1; x < s->blwidth-1; x++) {

  1670.         block_mc(s, block+x, mctmp, s->obmc_weight[1], plane, dstx, dsty);

  1671.         dstx  += p->xbsep;

  1672.         mctmp += p->xbsep;

  1673.     }

  1674.     block_mc(s, block+x, mctmp, s->obmc_weight[2], plane, dstx, dsty);

  1675. }

  1676. 

  1677. static void select_dsp_funcs(DiracContext *s, int width, int height, int xblen, int yblen)

  1678. {

  1679.     int idx = 0;

  1680.     if (xblen > 8)

  1681.         idx = 1;

  1682.     if (xblen > 16)

  1683.         idx = 2;

  1684. 

  1685.     memcpy(s->put_pixels_tab, s->diracdsp.put_dirac_pixels_tab[idx], sizeof(s->put_pixels_tab));

  1686.     memcpy(s->avg_pixels_tab, s->diracdsp.avg_dirac_pixels_tab[idx], sizeof(s->avg_pixels_tab));

  1687.     s->add_obmc = s->diracdsp.add_dirac_obmc[idx];

  1688.     if (s->weight_log2denom > 1 || s->weight[0] != 1 || s->weight[1] != 1) {

  1689.         s->weight_func   = s->diracdsp.weight_dirac_pixels_tab[idx];

  1690.         s->biweight_func = s->diracdsp.biweight_dirac_pixels_tab[idx];

  1691.     } else {

  1692.         s->weight_func   = NULL;

  1693.         s->biweight_func = NULL;

  1694.     }

  1695. }

  1696. 

  1697. static int interpolate_refplane(DiracContext *s, DiracFrame *ref, int plane, int width, int height)

  1698. {

  1699.     /* chroma allocates an edge of 8 when subsampled

  1700.        which for 4:2:2 means an h edge of 16 and v edge of 8

  1701.        just use 8 for everything for the moment */

  1702.     int i, edge = EDGE_WIDTH/2;

  1703. 

  1704.     ref->hpel[plane][0] = ref->avframe->data[plane];

  1705.     s->mpvencdsp.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 */

  1706. 

  1707.     /* no need for hpel if we only have fpel vectors */

  1708.     if (!s->mv_precision)

  1709.         return 0;

  1710. 

  1711.     for (i = 1; i < 4; i++) {

  1712.         if (!ref->hpel_base[plane][i])

  1713.             ref->hpel_base[plane][i] = av_malloc((height+2*edge) * ref->avframe->linesize[plane] + 32);

  1714.         if (!ref->hpel_base[plane][i]) {

  1715.             return AVERROR(ENOMEM);

  1716.         }

  1717.         /* we need to be 16-byte aligned even for chroma */

  1718.         ref->hpel[plane][i] = ref->hpel_base[plane][i] + edge*ref->avframe->linesize[plane] + 16;

  1719.     }

  1720. 

  1721.     if (!ref->interpolated[plane]) {

  1722.         s->diracdsp.dirac_hpel_filter(ref->hpel[plane][1], ref->hpel[plane][2],

  1723.                                       ref->hpel[plane][3], ref->hpel[plane][0],

  1724.                                       ref->avframe->linesize[plane], width, height);

  1725.         s->mpvencdsp.draw_edges(ref->hpel[plane][1], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);

  1726.         s->mpvencdsp.draw_edges(ref->hpel[plane][2], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);

  1727.         s->mpvencdsp.draw_edges(ref->hpel[plane][3], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);

  1728.     }

  1729.     ref->interpolated[plane] = 1;

  1730. 

  1731.     return 0;

  1732. }

  1733. 

  1734. /**

  1735.  * Dirac Specification ->

  1736.  * 13.0 Transform data syntax. transform_data()

  1737.  */

  1738. static int dirac_decode_frame_internal(DiracContext *s)

  1739. {

  1740.     DWTContext d;

  1741.     int y, i, comp, dsty;

  1742.     int ret;

  1743. 

  1744.     if (s->low_delay) {

  1745.         /* [DIRAC_STD] 13.5.1 low_delay_transform_data() */

  1746.         for (comp = 0; comp < 3; comp++) {

  1747.             Plane *p = &s->plane[comp];

  1748.             memset(p->idwt.buf, 0, p->idwt.stride * p->idwt.height);

  1749.         }

  1750.         if (!s->zero_res) {

  1751.             if ((ret = decode_lowdelay(s)) < 0)

  1752.                 return ret;

  1753.         }

  1754.     }

  1755. 

  1756.     for (comp = 0; comp < 3; comp++) {

  1757.         Plane *p       = &s->plane[comp];

  1758.         uint8_t *frame = s->current_picture->avframe->data[comp];

  1759. 

  1760.         /* FIXME: small resolutions */

  1761.         for (i = 0; i < 4; i++)

  1762.             s->edge_emu_buffer[i] = s->edge_emu_buffer_base + i*FFALIGN(p->width, 16);

  1763. 

  1764.         if (!s->zero_res && !s->low_delay)

  1765.         {

  1766.             memset(p->idwt.buf, 0, p->idwt.stride * p->idwt.height);

  1767.             decode_component(s, comp); /* [DIRAC_STD] 13.4.1 core_transform_data() */

  1768.         }

  1769.         ret = ff_spatial_idwt_init(&d, &p->idwt, s->wavelet_idx+2,

  1770.                                    s->wavelet_depth, s->bit_depth);

  1771.         if (ret < 0)

  1772.             return ret;

  1773. 

  1774.         if (!s->num_refs) { /* intra */

  1775.             for (y = 0; y < p->height; y += 16) {

  1776.                 int idx = (s->bit_depth - 8) >> 1;

  1777.                 ff_spatial_idwt_slice2(&d, y+16); /* decode */

  1778.                 s->diracdsp.put_signed_rect_clamped[idx](frame + y*p->stride,

  1779.                                                          p->stride,

  1780.                                                          p->idwt.buf + y*p->idwt.stride,

  1781.                                                          p->idwt.stride, p->width, 16);

  1782.             }

  1783.         } else { /* inter */

  1784.             int rowheight = p->ybsep*p->stride;

  1785. 

  1786.             select_dsp_funcs(s, p->width, p->height, p->xblen, p->yblen);

  1787. 

  1788.             for (i = 0; i < s->num_refs; i++) {

  1789.                 int ret = interpolate_refplane(s, s->ref_pics[i], comp, p->width, p->height);

  1790.                 if (ret < 0)

  1791.                     return ret;

  1792.             }

  1793. 

  1794.             memset(s->mctmp, 0, 4*p->yoffset*p->stride);

  1795. 

  1796.             dsty = -p->yoffset;

  1797.             for (y = 0; y < s->blheight; y++) {

  1798.                 int h     = 0,

  1799.                     start = FFMAX(dsty, 0);

  1800.                 uint16_t *mctmp    = s->mctmp + y*rowheight;

  1801.                 DiracBlock *blocks = s->blmotion + y*s->blwidth;

  1802. 

  1803.                 init_obmc_weights(s, p, y);

  1804. 

  1805.                 if (y == s->blheight-1 || start+p->ybsep > p->height)

  1806.                     h = p->height - start;

  1807.                 else

  1808.                     h = p->ybsep - (start - dsty);

  1809.                 if (h < 0)

  1810.                     break;

  1811. 

  1812.                 memset(mctmp+2*p->yoffset*p->stride, 0, 2*rowheight);

  1813.                 mc_row(s, blocks, mctmp, comp, dsty);

  1814. 

  1815.                 mctmp += (start - dsty)*p->stride + p->xoffset;

  1816.                 ff_spatial_idwt_slice2(&d, start + h); /* decode */

  1817.                 /* NOTE: add_rect_clamped hasn't been templated hence the shifts.

  1818.                  * idwt.stride is passed as pixels, not in bytes as in the rest of the decoder */

  1819.                 s->diracdsp.add_rect_clamped(frame + start*p->stride, mctmp, p->stride,

  1820.                                              (int16_t*)(p->idwt.buf) + start*(p->idwt.stride >> 1), (p->idwt.stride >> 1), p->width, h);

  1821. 

  1822.                 dsty += p->ybsep;

  1823.             }

  1824.         }

  1825.     }

  1826. 

  1827. 

  1828.     return 0;

  1829. }

  1830. 

  1831. static int get_buffer_with_edge(AVCodecContext *avctx, AVFrame *f, int flags)

  1832. {

  1833.     int ret, i;

  1834.     int chroma_x_shift, chroma_y_shift;

  1835.     avcodec_get_chroma_sub_sample(avctx->pix_fmt, &chroma_x_shift, &chroma_y_shift);

  1836. 

  1837.     f->width  = avctx->width  + 2 * EDGE_WIDTH;

  1838.     f->height = avctx->height + 2 * EDGE_WIDTH + 2;

  1839.     ret = ff_get_buffer(avctx, f, flags);

  1840.     if (ret < 0)

  1841.         return ret;

  1842. 

  1843.     for (i = 0; f->data[i]; i++) {

  1844.         int offset = (EDGE_WIDTH >> (i && i<3 ? chroma_y_shift : 0)) *

  1845.                      f->linesize[i] + 32;

  1846.         f->data[i] += offset;

  1847.     }

  1848.     f->width  = avctx->width;

  1849.     f->height = avctx->height;

  1850. 

  1851.     return 0;

  1852. }

  1853. 

  1854. /**

  1855.  * Dirac Specification ->

  1856.  * 11.1.1 Picture Header. picture_header()

  1857.  */

  1858. static int dirac_decode_picture_header(DiracContext *s)

  1859. {

  1860.     unsigned retire, picnum;

  1861.     int i, j, ret;

  1862.     int64_t refdist, refnum;

  1863.     GetBitContext *gb = &s->gb;

  1864. 

  1865.     /* [DIRAC_STD] 11.1.1 Picture Header. picture_header() PICTURE_NUM */

  1866.     picnum = s->current_picture->avframe->display_picture_number = get_bits_long(gb, 32);

  1867. 

  1868. 

  1869.     av_log(s->avctx,AV_LOG_DEBUG,"PICTURE_NUM: %d\n",picnum);

  1870. 

  1871.     /* if this is the first keyframe after a sequence header, start our

  1872.        reordering from here */

  1873.     if (s->frame_number < 0)

  1874.         s->frame_number = picnum;

  1875. 

  1876.     s->ref_pics[0] = s->ref_pics[1] = NULL;

  1877.     for (i = 0; i < s->num_refs; i++) {

  1878.         refnum = (picnum + dirac_get_se_golomb(gb)) & 0xFFFFFFFF;

  1879.         refdist = INT64_MAX;

  1880. 

  1881.         /* find the closest reference to the one we want */

  1882.         /* Jordi: this is needed if the referenced picture hasn't yet arrived */

  1883.         for (j = 0; j < MAX_REFERENCE_FRAMES && refdist; j++)

  1884.             if (s->ref_frames[j]

  1885.                 && FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum) < refdist) {

  1886.                 s->ref_pics[i] = s->ref_frames[j];

  1887.                 refdist = FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum);

  1888.             }

  1889. 

  1890.         if (!s->ref_pics[i] || refdist)

  1891.             av_log(s->avctx, AV_LOG_DEBUG, "Reference not found\n");

  1892. 

  1893.         /* if there were no references at all, allocate one */

  1894.         if (!s->ref_pics[i])

  1895.             for (j = 0; j < MAX_FRAMES; j++)

  1896.                 if (!s->all_frames[j].avframe->data[0]) {

  1897.                     s->ref_pics[i] = &s->all_frames[j];

  1898.                     get_buffer_with_edge(s->avctx, s->ref_pics[i]->avframe, AV_GET_BUFFER_FLAG_REF);

  1899.                     break;

  1900.                 }

  1901. 

  1902.         if (!s->ref_pics[i]) {

  1903.             av_log(s->avctx, AV_LOG_ERROR, "Reference could not be allocated\n");

  1904.             return AVERROR_INVALIDDATA;

  1905.         }

  1906. 

  1907.     }

  1908. 

  1909.     /* retire the reference frames that are not used anymore */

  1910.     if (s->current_picture->reference) {

  1911.         retire = (picnum + dirac_get_se_golomb(gb)) & 0xFFFFFFFF;

  1912.         if (retire != picnum) {

  1913.             DiracFrame *retire_pic = remove_frame(s->ref_frames, retire);

  1914. 

  1915.             if (retire_pic)

  1916.                 retire_pic->reference &= DELAYED_PIC_REF;

  1917.             else

  1918.                 av_log(s->avctx, AV_LOG_DEBUG, "Frame to retire not found\n");

  1919.         }

  1920. 

  1921.         /* if reference array is full, remove the oldest as per the spec */

  1922.         while (add_frame(s->ref_frames, MAX_REFERENCE_FRAMES, s->current_picture)) {

  1923.             av_log(s->avctx, AV_LOG_ERROR, "Reference frame overflow\n");

  1924.             remove_frame(s->ref_frames, s->ref_frames[0]->avframe->display_picture_number)->reference &= DELAYED_PIC_REF;

  1925.         }

  1926.     }

  1927. 

  1928.     if (s->num_refs) {

  1929.         ret = dirac_unpack_prediction_parameters(s);  /* [DIRAC_STD] 11.2 Picture Prediction Data. picture_prediction() */

  1930.         if (ret < 0)

  1931.             return ret;

  1932.         ret = dirac_unpack_block_motion_data(s);      /* [DIRAC_STD] 12. Block motion data syntax                       */

  1933.         if (ret < 0)

  1934.             return ret;

  1935.     }

  1936.     ret = dirac_unpack_idwt_params(s);                /* [DIRAC_STD] 11.3 Wavelet transform data                        */

  1937.     if (ret < 0)

  1938.         return ret;

  1939. 

  1940.     init_planes(s);

  1941.     return 0;

  1942. }

  1943. 

  1944. static int get_delayed_pic(DiracContext *s, AVFrame *picture, int *got_frame)

  1945. {

  1946.     DiracFrame *out = s->delay_frames[0];

  1947.     int i, out_idx  = 0;

  1948.     int ret;

  1949. 

  1950.     /* find frame with lowest picture number */

  1951.     for (i = 1; s->delay_frames[i]; i++)

  1952.         if (s->delay_frames[i]->avframe->display_picture_number < out->avframe->display_picture_number) {

  1953.             out     = s->delay_frames[i];

  1954.             out_idx = i;

  1955.         }

  1956. 

  1957.     for (i = out_idx; s->delay_frames[i]; i++)

  1958.         s->delay_frames[i] = s->delay_frames[i+1];

  1959. 

  1960.     if (out) {

  1961.         out->reference ^= DELAYED_PIC_REF;

  1962.         *got_frame = 1;

  1963.         if((ret = av_frame_ref(picture, out->avframe)) < 0)

  1964.             return ret;

  1965.     }

  1966. 

  1967.     return 0;

  1968. }

  1969. 

  1970. /**

  1971.  * Dirac Specification ->

  1972.  * 9.6 Parse Info Header Syntax. parse_info()

  1973.  * 4 byte start code + byte parse code + 4 byte size + 4 byte previous size

  1974.  */

  1975. #define DATA_UNIT_HEADER_SIZE 13

  1976. 

  1977. /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3

  1978.    inside the function parse_sequence() */

  1979. static int dirac_decode_data_unit(AVCodecContext *avctx, const uint8_t *buf, int size)

  1980. {

  1981.     DiracContext *s   = avctx->priv_data;

  1982.     DiracFrame *pic   = NULL;

  1983.     AVDiracSeqHeader *dsh;

  1984.     int ret, i;

  1985.     uint8_t parse_code;

  1986.     unsigned tmp;

  1987. 

  1988.     if (size < DATA_UNIT_HEADER_SIZE)

  1989.         return AVERROR_INVALIDDATA;

  1990. 

  1991.     parse_code = buf[4];

  1992. 

  1993.     init_get_bits(&s->gb, &buf[13], 8*(size - DATA_UNIT_HEADER_SIZE));

  1994. 

  1995.     if (parse_code == DIRAC_PCODE_SEQ_HEADER) {

  1996.         if (s->seen_sequence_header)

  1997.             return 0;

  1998. 

  1999.         /* [DIRAC_STD] 10. Sequence header */

  2000.         ret = av_dirac_parse_sequence_header(&dsh, buf + DATA_UNIT_HEADER_SIZE, size - DATA_UNIT_HEADER_SIZE, avctx);

  2001.         if (ret < 0) {

  2002.             av_log(avctx, AV_LOG_ERROR, "error parsing sequence header");

  2003.             return ret;

  2004.         }

  2005. 

  2006.         ret = ff_set_dimensions(avctx, dsh->width, dsh->height);

  2007.         if (ret < 0) {

  2008.             av_freep(&dsh);

  2009.             return ret;

  2010.         }

  2011. 

  2012.         ff_set_sar(avctx, dsh->sample_aspect_ratio);

  2013.         avctx->pix_fmt         = dsh->pix_fmt;

  2014.         avctx->color_range     = dsh->color_range;

  2015.         avctx->color_trc       = dsh->color_trc;

  2016.         avctx->color_primaries = dsh->color_primaries;

  2017.         avctx->colorspace      = dsh->colorspace;

  2018.         avctx->profile         = dsh->profile;

  2019.         avctx->level           = dsh->level;

  2020.         avctx->framerate       = dsh->framerate;

  2021.         s->bit_depth           = dsh->bit_depth;

  2022.         s->version.major       = dsh->version.major;

  2023.         s->version.minor       = dsh->version.minor;

  2024.         s->seq                 = *dsh;

  2025.         av_freep(&dsh);

  2026. 

  2027.         s->pshift = s->bit_depth > 8;

  2028. 

  2029.         avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_x_shift, &s->chroma_y_shift);

  2030. 

  2031.         ret = alloc_sequence_buffers(s);

  2032.         if (ret < 0)

  2033.             return ret;

  2034. 

  2035.         s->seen_sequence_header = 1;

  2036.     } else if (parse_code == DIRAC_PCODE_END_SEQ) { /* [DIRAC_STD] End of Sequence */

  2037.         free_sequence_buffers(s);

  2038.         s->seen_sequence_header = 0;

  2039.     } else if (parse_code == DIRAC_PCODE_AUX) {

  2040.         if (buf[13] == 1) {     /* encoder implementation/version */

  2041.             int ver[3];

  2042.             /* versions older than 1.0.8 don't store quant delta for

  2043.                subbands with only one codeblock */

  2044.             if (sscanf(buf+14, "Schroedinger %d.%d.%d", ver, ver+1, ver+2) == 3)

  2045.                 if (ver[0] == 1 && ver[1] == 0 && ver[2] <= 7)

  2046.                     s->old_delta_quant = 1;

  2047.         }

  2048.     } else if (parse_code & 0x8) {  /* picture data unit */

  2049.         if (!s->seen_sequence_header) {

  2050.             av_log(avctx, AV_LOG_DEBUG, "Dropping frame without sequence header\n");

  2051.             return AVERROR_INVALIDDATA;

  2052.         }

  2053. 

  2054.         /* find an unused frame */

  2055.         for (i = 0; i < MAX_FRAMES; i++)

  2056.             if (s->all_frames[i].avframe->data[0] == NULL)

  2057.                 pic = &s->all_frames[i];

  2058.         if (!pic) {

  2059.             av_log(avctx, AV_LOG_ERROR, "framelist full\n");

  2060.             return AVERROR_INVALIDDATA;

  2061.         }

  2062. 

  2063.         av_frame_unref(pic->avframe);

  2064. 

  2065.         /* [DIRAC_STD] Defined in 9.6.1 ... */

  2066.         tmp            =  parse_code & 0x03;                   /* [DIRAC_STD] num_refs()      */

  2067.         if (tmp > 2) {

  2068.             av_log(avctx, AV_LOG_ERROR, "num_refs of 3\n");

  2069.             return AVERROR_INVALIDDATA;

  2070.         }

  2071.         s->num_refs      = tmp;

  2072.         s->is_arith      = (parse_code & 0x48) == 0x08;          /* [DIRAC_STD] using_ac()            */

  2073.         s->low_delay     = (parse_code & 0x88) == 0x88;          /* [DIRAC_STD] is_low_delay()        */

  2074.         s->core_syntax   = (parse_code & 0x88) == 0x08;          /* [DIRAC_STD] is_core_syntax()      */

  2075.         s->ld_picture    = (parse_code & 0xF8) == 0xC8;          /* [DIRAC_STD] is_ld_picture()       */

  2076.         s->hq_picture    = (parse_code & 0xF8) == 0xE8;          /* [DIRAC_STD] is_hq_picture()       */

  2077.         s->dc_prediction = (parse_code & 0x28) == 0x08;          /* [DIRAC_STD] using_dc_prediction() */

  2078.         pic->reference   = (parse_code & 0x0C) == 0x0C;          /* [DIRAC_STD] is_reference()        */

  2079.         pic->avframe->key_frame = s->num_refs == 0;              /* [DIRAC_STD] is_intra()            */

  2080.         pic->avframe->pict_type = s->num_refs + 1;               /* Definition of AVPictureType in avutil.h */

  2081. 

  2082.         /* VC-2 Low Delay has a different parse code than the Dirac Low Delay */

  2083.         if (s->version.minor == 2 && parse_code == 0x88)

  2084.             s->ld_picture = 1;

  2085. 

  2086.         if (s->low_delay && !(s->ld_picture || s->hq_picture) ) {

  2087.             av_log(avctx, AV_LOG_ERROR, "Invalid low delay flag\n");

  2088.             return AVERROR_INVALIDDATA;

  2089.         }

  2090. 

  2091.         if ((ret = get_buffer_with_edge(avctx, pic->avframe, (parse_code & 0x0C) == 0x0C ? AV_GET_BUFFER_FLAG_REF : 0)) < 0)

  2092.             return ret;

  2093.         s->current_picture = pic;

  2094.         s->plane[0].stride = pic->avframe->linesize[0];

  2095.         s->plane[1].stride = pic->avframe->linesize[1];

  2096.         s->plane[2].stride = pic->avframe->linesize[2];

  2097. 

  2098.         if (alloc_buffers(s, FFMAX3(FFABS(s->plane[0].stride), FFABS(s->plane[1].stride), FFABS(s->plane[2].stride))) < 0)

  2099.             return AVERROR(ENOMEM);

  2100. 

  2101.         /* [DIRAC_STD] 11.1 Picture parse. picture_parse() */

  2102.         ret = dirac_decode_picture_header(s);

  2103.         if (ret < 0)

  2104.             return ret;

  2105. 

  2106.         /* [DIRAC_STD] 13.0 Transform data syntax. transform_data() */

  2107.         ret = dirac_decode_frame_internal(s);

  2108.         if (ret < 0)

  2109.             return ret;

  2110.     }

  2111.     return 0;

  2112. }

  2113. 

  2114. static int dirac_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *pkt)

  2115. {

  2116.     DiracContext *s     = avctx->priv_data;

  2117.     AVFrame *picture    = data;

  2118.     uint8_t *buf        = pkt->data;

  2119.     int buf_size        = pkt->size;

  2120.     int i, buf_idx      = 0;

  2121.     int ret;

  2122.     unsigned data_unit_size;

  2123. 

  2124.     /* release unused frames */

  2125.     for (i = 0; i < MAX_FRAMES; i++)

  2126.         if (s->all_frames[i].avframe->data[0] && !s->all_frames[i].reference) {

  2127.             av_frame_unref(s->all_frames[i].avframe);

  2128.             memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));

  2129.         }

  2130. 

  2131.     s->current_picture = NULL;

  2132.     *got_frame = 0;

  2133. 

  2134.     /* end of stream, so flush delayed pics */

  2135.     if (buf_size == 0)

  2136.         return get_delayed_pic(s, (AVFrame *)data, got_frame);

  2137. 

  2138.     for (;;) {

  2139.         /*[DIRAC_STD] Here starts the code from parse_info() defined in 9.6

  2140.           [DIRAC_STD] PARSE_INFO_PREFIX = "BBCD" as defined in ISO/IEC 646

  2141.           BBCD start code search */

  2142.         for (; buf_idx + DATA_UNIT_HEADER_SIZE < buf_size; buf_idx++) {

  2143.             if (buf[buf_idx  ] == 'B' && buf[buf_idx+1] == 'B' &&

  2144.                 buf[buf_idx+2] == 'C' && buf[buf_idx+3] == 'D')

  2145.                 break;

  2146.         }

  2147.         /* BBCD found or end of data */

  2148.         if (buf_idx + DATA_UNIT_HEADER_SIZE >= buf_size)

  2149.             break;

  2150. 

  2151.         data_unit_size = AV_RB32(buf+buf_idx+5);

  2152.         if (data_unit_size > buf_size - buf_idx || !data_unit_size) {

  2153.             if(data_unit_size > buf_size - buf_idx)

  2154.             av_log(s->avctx, AV_LOG_ERROR,

  2155.                    "Data unit with size %d is larger than input buffer, discarding\n",

  2156.                    data_unit_size);

  2157.             buf_idx += 4;

  2158.             continue;

  2159.         }

  2160.         /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3 inside the function parse_sequence() */

  2161.         ret = dirac_decode_data_unit(avctx, buf+buf_idx, data_unit_size);

  2162.         if (ret < 0)

  2163.         {

  2164.             av_log(s->avctx, AV_LOG_ERROR,"Error in dirac_decode_data_unit\n");

  2165.             return ret;

  2166.         }

  2167.         buf_idx += data_unit_size;

  2168.     }

  2169. 

  2170.     if (!s->current_picture)

  2171.         return buf_size;

  2172. 

  2173.     if (s->current_picture->avframe->display_picture_number > s->frame_number) {

  2174.         DiracFrame *delayed_frame = remove_frame(s->delay_frames, s->frame_number);

  2175. 

  2176.         s->current_picture->reference |= DELAYED_PIC_REF;

  2177. 

  2178.         if (add_frame(s->delay_frames, MAX_DELAY, s->current_picture)) {

  2179.             int min_num = s->delay_frames[0]->avframe->display_picture_number;

  2180.             /* Too many delayed frames, so we display the frame with the lowest pts */

  2181.             av_log(avctx, AV_LOG_ERROR, "Delay frame overflow\n");

  2182. 

  2183.             for (i = 1; s->delay_frames[i]; i++)

  2184.                 if (s->delay_frames[i]->avframe->display_picture_number < min_num)

  2185.                     min_num = s->delay_frames[i]->avframe->display_picture_number;

  2186. 

  2187.             delayed_frame = remove_frame(s->delay_frames, min_num);

  2188.             add_frame(s->delay_frames, MAX_DELAY, s->current_picture);

  2189.         }

  2190. 

  2191.         if (delayed_frame) {

  2192.             delayed_frame->reference ^= DELAYED_PIC_REF;

  2193.             if((ret=av_frame_ref(data, delayed_frame->avframe)) < 0)

  2194.                 return ret;

  2195.             *got_frame = 1;

  2196.         }

  2197.     } else if (s->current_picture->avframe->display_picture_number == s->frame_number) {

  2198.         /* The right frame at the right time :-) */

  2199.         if((ret=av_frame_ref(data, s->current_picture->avframe)) < 0)

  2200.             return ret;

  2201.         *got_frame = 1;

  2202.     }

  2203. 

  2204.     if (*got_frame)

  2205.         s->frame_number = picture->display_picture_number + 1;

  2206. 

  2207.     return buf_idx;

  2208. }

  2209. 

  2210. AVCodec ff_dirac_decoder = {

  2211.     .name           = "dirac",

  2212.     .long_name      = NULL_IF_CONFIG_SMALL("BBC Dirac VC-2"),

  2213.     .type           = AVMEDIA_TYPE_VIDEO,

  2214.     .id             = AV_CODEC_ID_DIRAC,

  2215.     .priv_data_size = sizeof(DiracContext),

  2216.     .init           = dirac_decode_init,

  2217.     .close          = dirac_decode_end,

  2218.     .decode         = dirac_decode_frame,

  2219.     .capabilities   = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_DR1,

  2220.     .flush          = dirac_decode_flush,

  2221. };