Subversion Repositories Kolibri OS

/*

 * Copyright (C) 2007 Marco Gerards 

 * Copyright (C) 2009 David Conrad

 * Copyright (C) 2011 Jordi Ortiz

 *

 * This file is part of FFmpeg.

 *

 * FFmpeg is free software; you can redistribute it and/or

 * modify it under the terms of the GNU Lesser General Public

 * License as published by the Free Software Foundation; either

 * version 2.1 of the License, or (at your option) any later version.

 *

 * FFmpeg is distributed in the hope that it will be useful,

 * but WITHOUT ANY WARRANTY; without even the implied warranty of

 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

 * Lesser General Public License for more details.

 *

 * You should have received a copy of the GNU Lesser General Public

 * License along with FFmpeg; if not, write to the Free Software

 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

 */

/**

 * @file

 * Dirac Decoder

 * @author Marco Gerards , David Conrad, Jordi Ortiz 

 */

#include "avcodec.h"

#include "dsputil.h"

#include "get_bits.h"

#include "bytestream.h"

#include "internal.h"

#include "golomb.h"

#include "dirac_arith.h"

#include "mpeg12data.h"

#include "dirac_dwt.h"

#include "dirac.h"

#include "diracdsp.h"

#include "videodsp.h" // for ff_emulated_edge_mc_8

/**

 * The spec limits the number of wavelet decompositions to 4 for both

 * level 1 (VC-2) and 128 (long-gop default).

 * 5 decompositions is the maximum before >16-bit buffers are needed.

 * Schroedinger allows this for DD 9,7 and 13,7 wavelets only, limiting

 * the others to 4 decompositions (or 3 for the fidelity filter).

 *

 * We use this instead of MAX_DECOMPOSITIONS to save some memory.

 */

#define MAX_DWT_LEVELS 5

/**

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

 */

#define MAX_REFERENCE_FRAMES 8

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

#define MAX_FRAMES (MAX_REFERENCE_FRAMES + MAX_DELAY + 1)

#define MAX_QUANT 68        /* max quant for VC-2 */

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

/**

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

 */

#define DIRAC_REF_MASK_REF1   1

#define DIRAC_REF_MASK_REF2   2

#define DIRAC_REF_MASK_GLOBAL 4

/**

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

 * is held for delayed output.

 */

#define DELAYED_PIC_REF 4

#define ff_emulated_edge_mc ff_emulated_edge_mc_8 /* Fix: change the calls to this function regarding bit depth */

#define CALC_PADDING(size, depth)                       \

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

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

typedef struct {

    AVFrame avframe;

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

    uint8_t *hpel[3][4];

    uint8_t *hpel_base[3][4];

} DiracFrame;

typedef struct {

    union {

        int16_t mv[2][2];

        int16_t dc[3];

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

    uint8_t ref;

} DiracBlock;

typedef struct SubBand {

    int level;

    int orientation;

    int stride;

    int width;

    int height;

    int quant;

    IDWTELEM *ibuf;

    struct SubBand *parent;

    /* for low delay */

    unsigned length;

    const uint8_t *coeff_data;

} SubBand;

typedef struct Plane {

    int width;

    int height;

    ptrdiff_t stride;

    int idwt_width;

    int idwt_height;

    int idwt_stride;

    IDWTELEM *idwt_buf;

    IDWTELEM *idwt_buf_base;

    IDWTELEM *idwt_tmp;

    /* block length */

    uint8_t xblen;

    uint8_t yblen;

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

    uint8_t xbsep;

    uint8_t ybsep;

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

    uint8_t xoffset;

    uint8_t yoffset;

    SubBand band[MAX_DWT_LEVELS][4];

} Plane;

typedef struct DiracContext {

    AVCodecContext *avctx;

    DSPContext dsp;

    DiracDSPContext diracdsp;

    GetBitContext gb;

    dirac_source_params source;

    int seen_sequence_header;

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

    Plane plane[3];

    int chroma_x_shift;

    int chroma_y_shift;

    int zero_res;               /* zero residue flag                         */

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

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

    int globalmc_flag;          /* use global motion compensation            */

    int num_refs;               /* number of reference pictures              */

    /* wavelet decoding */

    unsigned wavelet_depth;     /* depth of the IDWT                         */

    unsigned wavelet_idx;

    /**

     * schroedinger older than 1.0.8 doesn't store

     * quant delta if only one codebook exists in a band

     */

    unsigned old_delta_quant;

    unsigned codeblock_mode;

    struct {

        unsigned width;

        unsigned height;

    } codeblock[MAX_DWT_LEVELS+1];

    struct {

        unsigned num_x;         /* number of horizontal slices               */

        unsigned num_y;         /* number of vertical slices                 */

        AVRational bytes;       /* average bytes per slice                   */

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

    } lowdelay;

    struct {

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

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

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

        unsigned zrs_exp;

        unsigned perspective_exp;

    } globalmc[2];

    /* motion compensation */

    uint8_t mv_precision;       /* [DIRAC_STD] REFS_WT_PRECISION             */

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

    unsigned weight_log2denom;  /* [DIRAC_STD] REFS_WT_PRECISION             */

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

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

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

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

    uint8_t *sbsplit;

    DiracBlock *blmotion;

    uint8_t *edge_emu_buffer[4];

    uint8_t *edge_emu_buffer_base;

    uint16_t *mctmp;            /* buffer holding the MC data multipled by OBMC weights */

    uint8_t *mcscratch;

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

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

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

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

    dirac_weight_func weight_func;

    dirac_biweight_func biweight_func;

    DiracFrame *current_picture;

    DiracFrame *ref_pics[2];

    DiracFrame *ref_frames[MAX_REFERENCE_FRAMES+1];

    DiracFrame *delay_frames[MAX_DELAY+1];

    DiracFrame all_frames[MAX_FRAMES];

} DiracContext;

/**

 * Dirac Specification ->

 * Parse code values. 9.6.1 Table 9.1

 */

enum dirac_parse_code {

    pc_seq_header         = 0x00,

    pc_eos                = 0x10,

    pc_aux_data           = 0x20,

    pc_padding            = 0x30,

};

enum dirac_subband {

    subband_ll = 0,

    subband_hl = 1,

    subband_lh = 2,

    subband_hh = 3

};

static const uint8_t default_qmat[][4][4] = {

    { { 5,  3,  3,  0}, { 0,  4,  4,  1}, { 0,  5,  5,  2}, { 0,  6,  6,  3} },

    { { 4,  2,  2,  0}, { 0,  4,  4,  2}, { 0,  5,  5,  3}, { 0,  7,  7,  5} },

    { { 5,  3,  3,  0}, { 0,  4,  4,  1}, { 0,  5,  5,  2}, { 0,  6,  6,  3} },

    { { 8,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0} },

    { { 8,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0}, { 0,  4,  4,  0} },

    { { 0,  4,  4,  8}, { 0,  8,  8, 12}, { 0, 13, 13, 17}, { 0, 17, 17, 21} },

    { { 3,  1,  1,  0}, { 0,  4,  4,  2}, { 0,  6,  6,  5}, { 0,  9,  9,  7} },

};

static const int qscale_tab[MAX_QUANT+1] = {

    4,     5,     6,     7,     8,    10,    11,    13,

    16,    19,    23,    27,    32,    38,    45,    54,

    64,    76,    91,   108,   128,   152,   181,   215,

    256,   304,   362,   431,   512,   609,   724,   861,

    1024,  1218,  1448,  1722,  2048,  2435,  2896,  3444,

    4096,  4871,  5793,  6889,  8192,  9742, 11585, 13777,

    16384, 19484, 23170, 27554, 32768, 38968, 46341, 55109,

    65536, 77936

};

static const int qoffset_intra_tab[MAX_QUANT+1] = {

    1,     2,     3,     4,     4,     5,     6,     7,

    8,    10,    12,    14,    16,    19,    23,    27,

    32,    38,    46,    54,    64,    76,    91,   108,

    128,   152,   181,   216,   256,   305,   362,   431,

    512,   609,   724,   861,  1024,  1218,  1448,  1722,

    2048,  2436,  2897,  3445,  4096,  4871,  5793,  6889,

    8192,  9742, 11585, 13777, 16384, 19484, 23171, 27555,

    32768, 38968

};

static const int qoffset_inter_tab[MAX_QUANT+1] = {

    1,     2,     2,     3,     3,     4,     4,     5,

    6,     7,     9,    10,    12,    14,    17,    20,

    24,    29,    34,    41,    48,    57,    68,    81,

    96,   114,   136,   162,   192,   228,   272,   323,

    384,   457,   543,   646,   768,   913,  1086,  1292,

    1536,  1827,  2172,  2583,  3072,  3653,  4344,  5166,

    6144,  7307,  8689, 10333, 12288, 14613, 17378, 20666,

    24576, 29226

};

/* magic number division by 3 from schroedinger */

static inline int divide3(int x)

{

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

}

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

{

    DiracFrame *remove_pic = NULL;

    int i, remove_idx = -1;

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

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

            remove_pic = framelist[i];

            remove_idx = i;

        }

    if (remove_pic)

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

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

    return remove_pic;

}

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

{

    int i;

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

        if (!framelist[i]) {

            framelist[i] = frame;

            return 0;

        }

    return -1;

}

static int alloc_sequence_buffers(DiracContext *s)

{

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

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

    int i, w, h, top_padding;

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

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

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

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

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

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

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

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

         * 1<0) in arith decoding

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

         * on each side */

        top_padding = FFMAX(1<

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

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

        s->plane[i].idwt_buf_base = av_mallocz((w+max_xblen)*h * sizeof(IDWTELEM));

        s->plane[i].idwt_tmp      = av_malloc((w+16) * sizeof(IDWTELEM));

        s->plane[i].idwt_buf      = s->plane[i].idwt_buf_base + top_padding*w;

        if (!s->plane[i].idwt_buf_base || !s->plane[i].idwt_tmp)

            return AVERROR(ENOMEM);

    }

    w = s->source.width;

    h = s->source.height;

    /* fixme: allocate using real stride here */

    s->sbsplit  = av_malloc(sbwidth * sbheight);

    s->blmotion = av_malloc(sbwidth * sbheight * 16 * sizeof(*s->blmotion));

    s->edge_emu_buffer_base = av_malloc((w+64)*MAX_BLOCKSIZE);

    s->mctmp     = av_malloc((w+64+MAX_BLOCKSIZE) * (h+MAX_BLOCKSIZE) * sizeof(*s->mctmp));

    s->mcscratch = av_malloc((w+64)*MAX_BLOCKSIZE);

    if (!s->sbsplit || !s->blmotion || !s->mctmp || !s->mcscratch)

        return AVERROR(ENOMEM);

    return 0;

}

static void free_sequence_buffers(DiracContext *s)

{

    int i, j, k;

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

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

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

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

        }

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

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

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

    }

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

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

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

        av_freep(&s->plane[i].idwt_buf_base);

        av_freep(&s->plane[i].idwt_tmp);

    }

    av_freep(&s->sbsplit);

    av_freep(&s->blmotion);

    av_freep(&s->edge_emu_buffer_base);

    av_freep(&s->mctmp);

    av_freep(&s->mcscratch);

}

static av_cold int dirac_decode_init(AVCodecContext *avctx)

{

    DiracContext *s = avctx->priv_data;

    s->avctx = avctx;

    s->frame_number = -1;

    if (avctx->flags&CODEC_FLAG_EMU_EDGE) {

        av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported!\n");

        return AVERROR_PATCHWELCOME;

    }

    ff_dsputil_init(&s->dsp, avctx);

    ff_diracdsp_init(&s->diracdsp);

    return 0;

}

static void dirac_decode_flush(AVCodecContext *avctx)

{

    DiracContext *s = avctx->priv_data;

    free_sequence_buffers(s);

    s->seen_sequence_header = 0;

    s->frame_number = -1;

}

static av_cold int dirac_decode_end(AVCodecContext *avctx)

{

    dirac_decode_flush(avctx);

    return 0;

}

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

static inline void coeff_unpack_arith(DiracArith *c, int qfactor, int qoffset,

                                      SubBand *b, IDWTELEM *buf, int x, int y)

{

    int coeff, sign;

    int sign_pred = 0;

    int pred_ctx = CTX_ZPZN_F1;

    /* Check if the parent subband has a 0 in the corresponding position */

    if (b->parent)

        pred_ctx += !!b->parent->ibuf[b->parent->stride * (y>>1) + (x>>1)] << 1;

    if (b->orientation == subband_hl)

        sign_pred = buf[-b->stride];

    /* Determine if the pixel has only zeros in its neighbourhood */

    if (x) {

        pred_ctx += !(buf[-1] | buf[-b->stride] | buf[-1-b->stride]);

        if (b->orientation == subband_lh)

            sign_pred = buf[-1];

    } else {

        pred_ctx += !buf[-b->stride];

    }

    coeff = dirac_get_arith_uint(c, pred_ctx, CTX_COEFF_DATA);

    if (coeff) {

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

        sign  = dirac_get_arith_bit(c, SIGN_CTX(sign_pred));

        coeff = (coeff ^ -sign) + sign;

    }

    *buf = coeff;

}

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

{

    int sign, coeff;

    coeff = svq3_get_ue_golomb(gb);

    if (coeff) {

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

        sign  = get_bits1(gb);

        coeff = (coeff ^ -sign) + sign;

    }

    return coeff;

}

/**

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

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

 */

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

                             GetBitContext *gb, DiracArith *c,

                             int left, int right, int top, int bottom,

                             int blockcnt_one, int is_arith)

{

    int x, y, zero_block;

    int qoffset, qfactor;

    IDWTELEM *buf;

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

    if (!blockcnt_one) {

        if (is_arith)

            zero_block = dirac_get_arith_bit(c, CTX_ZERO_BLOCK);

        else

            zero_block = get_bits1(gb);

        if (zero_block)

            return;

    }

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

        int quant = b->quant;

        if (is_arith)

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

        else

            quant += dirac_get_se_golomb(gb);

        if (quant < 0) {

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

            return;

        }

        b->quant = quant;

    }

    b->quant = FFMIN(b->quant, MAX_QUANT);

    qfactor = qscale_tab[b->quant];

    /* TODO: context pointer? */

    if (!s->num_refs)

        qoffset = qoffset_intra_tab[b->quant];

    else

        qoffset = qoffset_inter_tab[b->quant];

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

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

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

            /* [DIRAC_STD] 13.4.4 Subband coefficients. coeff_unpack() */

            if (is_arith)

                coeff_unpack_arith(c, qfactor, qoffset, b, buf+x, x, y);

            else

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

        }

        buf += b->stride;

    }

}

/**

 * Dirac Specification ->

 * 13.3 intra_dc_prediction(band)

 */

static inline void intra_dc_prediction(SubBand *b)

{

    IDWTELEM *buf = b->ibuf;

    int x, y;

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

        buf[x] += buf[x-1];

    buf += b->stride;

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

        buf[0] += buf[-b->stride];

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

            int pred = buf[x - 1] + buf[x - b->stride] + buf[x - b->stride-1];

            buf[x]  += divide3(pred);

        }

        buf += b->stride;

    }

}

/**

 * Dirac Specification ->

 * 13.4.2 Non-skipped subbands.  subband_coeffs()

 */

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

{

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

    DiracArith c;

    GetBitContext gb;

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

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

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

    if (!b->length)

        return;

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

    if (is_arith)

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

    top = 0;

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

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

        left = 0;

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

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

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

            left = right;

        }

        top = bottom;

    }

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

        intra_dc_prediction(b);

}

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

{

    DiracContext *s = avctx->priv_data;

    decode_subband_internal(s, b, 1);

    return 0;

}

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

{

    DiracContext *s = avctx->priv_data;

    SubBand **b     = arg;

    decode_subband_internal(s, *b, 0);

    return 0;

}

/**

 * Dirac Specification ->

 * [DIRAC_STD] 13.4.1 core_transform_data()

 */

static void decode_component(DiracContext *s, int comp)

{

    AVCodecContext *avctx = s->avctx;

    SubBand *bands[3*MAX_DWT_LEVELS+1];

    enum dirac_subband orientation;

    int level, num_bands = 0;

    /* Unpack all subbands at all levels. */

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

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

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

            bands[num_bands++] = b;

            align_get_bits(&s->gb);

            /* [DIRAC_STD] 13.4.2 subband() */

            b->length = svq3_get_ue_golomb(&s->gb);

            if (b->length) {

                b->quant = svq3_get_ue_golomb(&s->gb);

                align_get_bits(&s->gb);

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

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

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

            }

        }

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

        if (s->is_arith)

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

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

    }

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

    if (!s->is_arith)

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

}

/* [DIRAC_STD] 13.5.5.2 Luma slice subband data. luma_slice_band(level,orient,sx,sy) --> if b2 == NULL */

/* [DIRAC_STD] 13.5.5.3 Chroma slice subband data. chroma_slice_band(level,orient,sx,sy) --> if b2 != NULL */

static void lowdelay_subband(DiracContext *s, GetBitContext *gb, int quant,

                             int slice_x, int slice_y, int bits_end,

                             SubBand *b1, SubBand *b2)

{

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

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

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

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

    int qfactor = qscale_tab[FFMIN(quant, MAX_QUANT)];

    int qoffset = qoffset_intra_tab[FFMIN(quant, MAX_QUANT)];

    IDWTELEM *buf1 =      b1->ibuf + top * b1->stride;

    IDWTELEM *buf2 = b2 ? b2->ibuf + top * b2->stride : NULL;

    int x, y;

    /* we have to constantly check for overread since the spec explictly

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

    if (get_bits_count(gb) >= bits_end)

        return;

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

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

            buf1[x] = coeff_unpack_golomb(gb, qfactor, qoffset);

            if (get_bits_count(gb) >= bits_end)

                return;

            if (buf2) {

                buf2[x] = coeff_unpack_golomb(gb, qfactor, qoffset);

                if (get_bits_count(gb) >= bits_end)

                    return;

            }

        }

        buf1 += b1->stride;

        if (buf2)

            buf2 += b2->stride;

    }

}

struct lowdelay_slice {

    GetBitContext gb;

    int slice_x;

    int slice_y;

    int bytes;

};

/**

 * Dirac Specification ->

 * 13.5.2 Slices. slice(sx,sy)

 */

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

{

    DiracContext *s = avctx->priv_data;

    struct lowdelay_slice *slice = arg;

    GetBitContext *gb = &slice->gb;

    enum dirac_subband orientation;

    int level, quant, chroma_bits, chroma_end;

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

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

    int luma_bits   = get_bits_long(gb, length_bits);

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

    /* [DIRAC_STD] 13.5.5.2 luma_slice_band */

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

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

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

            lowdelay_subband(s, gb, quant, slice->slice_x, slice->slice_y, luma_end,

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

        }

    /* consume any unused bits from luma */

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

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

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

    /* [DIRAC_STD] 13.5.5.3 chroma_slice_band */

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

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

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

            lowdelay_subband(s, gb, quant, slice->slice_x, slice->slice_y, chroma_end,

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

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

        }

    return 0;

}

/**

 * Dirac Specification ->

 * 13.5.1 low_delay_transform_data()

 */

static void decode_lowdelay(DiracContext *s)

{

    AVCodecContext *avctx = s->avctx;

    int slice_x, slice_y, bytes, bufsize;

    const uint8_t *buf;

    struct lowdelay_slice *slices;

    int slice_num = 0;

    slices = av_mallocz(s->lowdelay.num_x * s->lowdelay.num_y * sizeof(struct lowdelay_slice));

    align_get_bits(&s->gb);

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

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

    bufsize = get_bits_left(&s->gb);

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

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

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

                - slice_num    * s->lowdelay.bytes.num / s->lowdelay.bytes.den;

            slices[slice_num].bytes   = bytes;

            slices[slice_num].slice_x = slice_x;

            slices[slice_num].slice_y = slice_y;

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

            slice_num++;

            buf     += bytes;

            bufsize -= bytes*8;

        }

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

                   sizeof(struct lowdelay_slice)); /* [DIRAC_STD] 13.5.2 Slices */

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

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

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

    av_free(slices);

}

static void init_planes(DiracContext *s)

{

    int i, w, h, level, orientation;

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

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

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

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

        p->idwt_width  = w = CALC_PADDING(p->width , s->wavelet_depth);

        p->idwt_height = h = CALC_PADDING(p->height, s->wavelet_depth);

        p->idwt_stride = FFALIGN(p->idwt_width, 8);

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

            w = w>>1;

            h = h>>1;

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

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

                b->ibuf   = p->idwt_buf;

                b->level  = level;

                b->stride = p->idwt_stride << (s->wavelet_depth - level);

                b->width  = w;

                b->height = h;

                b->orientation = orientation;

                if (orientation & 1)

                    b->ibuf += w;

                if (orientation > 1)

                    b->ibuf += b->stride>>1;

                if (level)

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

            }

        }

        if (i > 0) {

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

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

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

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

        }

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

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

    }

}

/**

 * Unpack the motion compensation parameters

 * Dirac Specification ->

 * 11.2 Picture prediction data. picture_prediction()

 */

static int dirac_unpack_prediction_parameters(DiracContext *s)

{

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

    static const uint8_t default_bsep[] = { 4,  8, 12, 16 };

    GetBitContext *gb = &s->gb;

    unsigned idx, ref;

    align_get_bits(gb);

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

    /* Luma and Chroma are equal. 11.2.3 */

    idx = svq3_get_ue_golomb(gb); /* [DIRAC_STD] index */

    if (idx > 4) {

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

        return -1;

    }

    if (idx == 0) {

        s->plane[0].xblen = svq3_get_ue_golomb(gb);

        s->plane[0].yblen = svq3_get_ue_golomb(gb);

        s->plane[0].xbsep = svq3_get_ue_golomb(gb);

        s->plane[0].ybsep = svq3_get_ue_golomb(gb);

    } else {

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

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

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

        s->plane[0].xbsep = default_bsep[idx-1];

        s->plane[0].ybsep = default_bsep[idx-1];

    }

    /*[DIRAC_STD] 11.2.4 motion_data_dimensions()

      Calculated in function dirac_unpack_block_motion_data */

    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) {

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

        return -1;

    }

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

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

        return -1;

    }

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

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

        return -1;

    }

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

      Read motion vector precision */

    s->mv_precision = svq3_get_ue_golomb(gb);

    if (s->mv_precision > 3) {

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

        return -1;

    }

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

      Read the global motion compensation parameters */

    s->globalmc_flag = get_bits1(gb);

    if (s->globalmc_flag) {

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

        /* [DIRAC_STD] pan_tilt(gparams) */

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

            if (get_bits1(gb)) {

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

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

            }

            /* [DIRAC_STD] zoom_rotate_shear(gparams)

               zoom/rotation/shear parameters */

            if (get_bits1(gb)) {

                s->globalmc[ref].zrs_exp   = svq3_get_ue_golomb(gb);

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

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

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

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

            } else {

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

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

            }

            /* [DIRAC_STD] perspective(gparams) */

            if (get_bits1(gb)) {

                s->globalmc[ref].perspective_exp = svq3_get_ue_golomb(gb);

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

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

            }

        }

    }

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

      Picture prediction mode, not currently used. */

    if (svq3_get_ue_golomb(gb)) {

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

        return -1;

    }

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

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

    s->weight_log2denom = 1;

    s->weight[0]        = 1;

    s->weight[1]        = 1;

    if (get_bits1(gb)) {

        s->weight_log2denom = svq3_get_ue_golomb(gb);

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

        if (s->num_refs == 2)

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

    }

    return 0;

}

/**

 * Dirac Specification ->

 * 11.3 Wavelet transform data. wavelet_transform()

 */

static int dirac_unpack_idwt_params(DiracContext *s)

{

    GetBitContext *gb = &s->gb;

    int i, level;

    unsigned tmp;

#define CHECKEDREAD(dst, cond, errmsg) \

    tmp = svq3_get_ue_golomb(gb); \

    if (cond) { \

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

        return -1; \

    }\

    dst = tmp;

    align_get_bits(gb);

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

    if (s->zero_res)

        return 0;

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

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

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

    if (!s->low_delay) {

        /* Codeblock parameters (core syntax only) */

        if (get_bits1(gb)) {

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

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

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

            }

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

        } else

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

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

    } else {

        /* Slice parameters + quantization matrix*/

        /*[DIRAC_STD] 11.3.4 Slice coding Parameters (low delay syntax only). slice_parameters() */

        s->lowdelay.num_x     = svq3_get_ue_golomb(gb);

        s->lowdelay.num_y     = svq3_get_ue_golomb(gb);

        s->lowdelay.bytes.num = svq3_get_ue_golomb(gb);

        s->lowdelay.bytes.den = svq3_get_ue_golomb(gb);

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

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

            return AVERROR_INVALIDDATA;

        }

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

        if (get_bits1(gb)) {

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

            /* custom quantization matrix */

            s->lowdelay.quant[0][0] = svq3_get_ue_golomb(gb);

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

                s->lowdelay.quant[level][1] = svq3_get_ue_golomb(gb);

                s->lowdelay.quant[level][2] = svq3_get_ue_golomb(gb);

                s->lowdelay.quant[level][3] = svq3_get_ue_golomb(gb);

            }

        } else {

            if (s->wavelet_depth > 4) {

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

                return AVERROR_INVALIDDATA;

            }

            /* default quantization matrix */

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

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

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

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

                    if (s->wavelet_idx == 3)

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

                }

        }

    }

    return 0;

}

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

{

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

    if (!(x|y))

        return 0;

    else if (!y)

        return sbsplit[-1];

    else if (!x)

        return sbsplit[-stride];

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

}

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

{

    int pred;

    if (!(x|y))

        return 0;

    else if (!y)

        return block[-1].ref & refmask;

    else if (!x)

        return block[-stride].ref & refmask;

    /* return the majority */

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

    return (pred >> 1) & refmask;

}

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

{

    int i, n = 0;

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

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

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

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

        n++;

    }

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

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

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

        n++;

    }

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

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

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

        n++;

    }

    if (n == 2) {

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

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

    } else if (n == 3) {

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

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

    }

}

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

{

    int16_t *pred[3];

    int refmask = ref+1;

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

    int n = 0;

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

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

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

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

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

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

    switch (n) {

    case 0:

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

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

        break;

    case 1:

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

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

        break;

    case 2:

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

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

        break;

    case 3:

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

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

        break;

    }

}

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

{

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

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

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

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

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

    int m       = (1<

    int mx      = m * ((A[0][0] * x + A[0][1]*y) + (1<

    int my      = m * ((A[1][0] * x + A[1][1]*y) + (1<