annotate vp3.c @ 11409:9b1db41afb30 libavcodec

Remove unused fast_memcpy() function in dct-test

/*
 * Copyright (C) 2003-2004 the ffmpeg project
 *
 * 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 libavcodec/vp3.c
 * On2 VP3 Video Decoder
 *
 * VP3 Video Decoder by Mike Melanson (mike at multimedia.cx)
 * For more information about the VP3 coding process, visit:
 *   http://wiki.multimedia.cx/index.php?title=On2_VP3
 *
 * Theora decoder by Alex Beregszaszi
 */

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include "avcodec.h"
#include "dsputil.h"
#include "get_bits.h"

#include "vp3data.h"
#include "xiph.h"

#define FRAGMENT_PIXELS 8

static av_cold int vp3_decode_end(AVCodecContext *avctx);

//FIXME split things out into their own arrays
typedef struct Vp3Fragment {
    int16_t dc;
    uint8_t coding_method;
    int8_t motion_x;
    int8_t motion_y;
    uint8_t qpi;
} Vp3Fragment;

#define SB_NOT_CODED        0
#define SB_PARTIALLY_CODED  1
#define SB_FULLY_CODED      2

// This is the maximum length of a single long bit run that can be encoded
// for superblock coding or block qps. Theora special-cases this to read a
// bit instead of flipping the current bit to allow for runs longer than 4129.
#define MAXIMUM_LONG_BIT_RUN 4129

#define MODE_INTER_NO_MV      0
#define MODE_INTRA            1
#define MODE_INTER_PLUS_MV    2
#define MODE_INTER_LAST_MV    3
#define MODE_INTER_PRIOR_LAST 4
#define MODE_USING_GOLDEN     5
#define MODE_GOLDEN_MV        6
#define MODE_INTER_FOURMV     7
#define CODING_MODE_COUNT     8

/* special internal mode */
#define MODE_COPY             8

/* There are 6 preset schemes, plus a free-form scheme */
static const int ModeAlphabet[6][CODING_MODE_COUNT] =
{
    /* scheme 1: Last motion vector dominates */
    {    MODE_INTER_LAST_MV,    MODE_INTER_PRIOR_LAST,
         MODE_INTER_PLUS_MV,    MODE_INTER_NO_MV,
         MODE_INTRA,            MODE_USING_GOLDEN,
         MODE_GOLDEN_MV,        MODE_INTER_FOURMV },

    /* scheme 2 */
    {    MODE_INTER_LAST_MV,    MODE_INTER_PRIOR_LAST,
         MODE_INTER_NO_MV,      MODE_INTER_PLUS_MV,
         MODE_INTRA,            MODE_USING_GOLDEN,
         MODE_GOLDEN_MV,        MODE_INTER_FOURMV },

    /* scheme 3 */
    {    MODE_INTER_LAST_MV,    MODE_INTER_PLUS_MV,
         MODE_INTER_PRIOR_LAST, MODE_INTER_NO_MV,
         MODE_INTRA,            MODE_USING_GOLDEN,
         MODE_GOLDEN_MV,        MODE_INTER_FOURMV },

    /* scheme 4 */
    {    MODE_INTER_LAST_MV,    MODE_INTER_PLUS_MV,
         MODE_INTER_NO_MV,      MODE_INTER_PRIOR_LAST,
         MODE_INTRA,            MODE_USING_GOLDEN,
         MODE_GOLDEN_MV,        MODE_INTER_FOURMV },

    /* scheme 5: No motion vector dominates */
    {    MODE_INTER_NO_MV,      MODE_INTER_LAST_MV,
         MODE_INTER_PRIOR_LAST, MODE_INTER_PLUS_MV,
         MODE_INTRA,            MODE_USING_GOLDEN,
         MODE_GOLDEN_MV,        MODE_INTER_FOURMV },

    /* scheme 6 */
    {    MODE_INTER_NO_MV,      MODE_USING_GOLDEN,
         MODE_INTER_LAST_MV,    MODE_INTER_PRIOR_LAST,
         MODE_INTER_PLUS_MV,    MODE_INTRA,
         MODE_GOLDEN_MV,        MODE_INTER_FOURMV },

};

static const uint8_t hilbert_offset[16][2] = {
    {0,0}, {1,0}, {1,1}, {0,1},
    {0,2}, {0,3}, {1,3}, {1,2},
    {2,2}, {2,3}, {3,3}, {3,2},
    {3,1}, {2,1}, {2,0}, {3,0}
};

#define MIN_DEQUANT_VAL 2

typedef struct Vp3DecodeContext {
    AVCodecContext *avctx;
    int theora, theora_tables;
    int version;
    int width, height;
    AVFrame golden_frame;
    AVFrame last_frame;
    AVFrame current_frame;
    int keyframe;
    DSPContext dsp;
    int flipped_image;
    int last_slice_end;

    int qps[3];
    int nqps;
    int last_qps[3];

    int superblock_count;
    int y_superblock_width;
    int y_superblock_height;
    int y_superblock_count;
    int c_superblock_width;
    int c_superblock_height;
    int c_superblock_count;
    int u_superblock_start;
    int v_superblock_start;
    unsigned char *superblock_coding;

    int macroblock_count;
    int macroblock_width;
    int macroblock_height;

    int fragment_count;
    int fragment_width;
    int fragment_height;

    Vp3Fragment *all_fragments;
    int fragment_start[3];
    int data_offset[3];

    ScanTable scantable;

    /* tables */
    uint16_t coded_dc_scale_factor[64];
    uint32_t coded_ac_scale_factor[64];
    uint8_t base_matrix[384][64];
    uint8_t qr_count[2][3];
    uint8_t qr_size [2][3][64];
    uint16_t qr_base[2][3][64];

    /**
     * This is a list of all tokens in bitstream order. Reordering takes place
     * by pulling from each level during IDCT. As a consequence, IDCT must be
     * in Hilbert order, making the minimum slice height 64 for 4:2:0 and 32
     * otherwise. The 32 different tokens with up to 12 bits of extradata are
     * collapsed into 3 types, packed as follows:
     *   (from the low to high bits)
     *
     * 2 bits: type (0,1,2)
     *   0: EOB run, 14 bits for run length (12 needed)
     *   1: zero run, 7 bits for run length
     *                7 bits for the next coefficient (3 needed)
     *   2: coefficient, 14 bits (11 needed)
     *
     * Coefficients are signed, so are packed in the highest bits for automatic
     * sign extension.
     */
    int16_t *dct_tokens[3][64];
    int16_t *dct_tokens_base;
#define TOKEN_EOB(eob_run)              ((eob_run) << 2)
#define TOKEN_ZERO_RUN(coeff, zero_run) (((coeff) << 9) + ((zero_run) << 2) + 1)
#define TOKEN_COEFF(coeff)              (((coeff) << 2) + 2)

    /**
     * number of blocks that contain DCT coefficients at the given level or higher
     */
    int num_coded_frags[3][64];
    int total_num_coded_frags;

    /* this is a list of indexes into the all_fragments array indicating
     * which of the fragments are coded */
    int *coded_fragment_list[3];

    VLC dc_vlc[16];
    VLC ac_vlc_1[16];
    VLC ac_vlc_2[16];
    VLC ac_vlc_3[16];
    VLC ac_vlc_4[16];

    VLC superblock_run_length_vlc;
    VLC fragment_run_length_vlc;
    VLC mode_code_vlc;
    VLC motion_vector_vlc;

    /* these arrays need to be on 16-byte boundaries since SSE2 operations
     * index into them */
    DECLARE_ALIGNED(16, int16_t, qmat)[3][2][3][64];     //<qmat[qpi][is_inter][plane]

    /* This table contains superblock_count * 16 entries. Each set of 16
     * numbers corresponds to the fragment indexes 0..15 of the superblock.
     * An entry will be -1 to indicate that no entry corresponds to that
     * index. */
    int *superblock_fragments;

    /* This is an array that indicates how a particular macroblock
     * is coded. */
    unsigned char *macroblock_coding;

    uint8_t edge_emu_buffer[9*2048]; //FIXME dynamic alloc
    int8_t qscale_table[2048]; //FIXME dynamic alloc (width+15)/16

    /* Huffman decode */
    int hti;
    unsigned int hbits;
    int entries;
    int huff_code_size;
    uint16_t huffman_table[80][32][2];

    uint8_t filter_limit_values[64];
    DECLARE_ALIGNED(8, int, bounding_values_array)[256+2];
} Vp3DecodeContext;

/************************************************************************
 * VP3 specific functions
 ************************************************************************/

/*
 * This function sets up all of the various blocks mappings:
 * superblocks <-> fragments, macroblocks <-> fragments,
 * superblocks <-> macroblocks
 *
 * Returns 0 is successful; returns 1 if *anything* went wrong.
 */
static int init_block_mapping(Vp3DecodeContext *s)
{
    int i, j;
    signed int hilbert_walk_mb[4];

    int current_fragment = 0;
    int current_width = 0;
    int current_height = 0;
    int right_edge = 0;
    int bottom_edge = 0;
    int superblock_row_inc = 0;
    int mapping_index = 0;

    static const signed char travel_width[16] = {
         1,  1,  0, -1,
         0,  0,  1,  0,
         1,  0,  1,  0,
         0, -1,  0,  1
    };

    static const signed char travel_height[16] = {
         0,  0,  1,  0,
         1,  1,  0, -1,
         0,  1,  0, -1,
        -1,  0, -1,  0
    };

    hilbert_walk_mb[0] = 1;
    hilbert_walk_mb[1] = s->macroblock_width;
    hilbert_walk_mb[2] = 1;
    hilbert_walk_mb[3] = -s->macroblock_width;

    /* iterate through each superblock (all planes) and map the fragments */
    for (i = 0; i < s->superblock_count; i++) {
        /* time to re-assign the limits? */
        if (i == 0) {

            /* start of Y superblocks */
            right_edge = s->fragment_width;
            bottom_edge = s->fragment_height;
            current_width = -1;
            current_height = 0;
            superblock_row_inc = 3 * s->fragment_width -
                (s->y_superblock_width * 4 - s->fragment_width);

            /* the first operation for this variable is to advance by 1 */
            current_fragment = -1;

        } else if (i == s->u_superblock_start) {

            /* start of U superblocks */
            right_edge = s->fragment_width / 2;
            bottom_edge = s->fragment_height / 2;
            current_width = -1;
            current_height = 0;
            superblock_row_inc = 3 * (s->fragment_width / 2) -
                (s->c_superblock_width * 4 - s->fragment_width / 2);

            /* the first operation for this variable is to advance by 1 */
            current_fragment = s->fragment_start[1] - 1;

        } else if (i == s->v_superblock_start) {

            /* start of V superblocks */
            right_edge = s->fragment_width / 2;
            bottom_edge = s->fragment_height / 2;
            current_width = -1;
            current_height = 0;
            superblock_row_inc = 3 * (s->fragment_width / 2) -
                (s->c_superblock_width * 4 - s->fragment_width / 2);

            /* the first operation for this variable is to advance by 1 */
            current_fragment = s->fragment_start[2] - 1;

        }

        if (current_width >= right_edge - 1) {
            /* reset width and move to next superblock row */
            current_width = -1;
            current_height += 4;

            /* fragment is now at the start of a new superblock row */
            current_fragment += superblock_row_inc;
        }

        /* iterate through all 16 fragments in a superblock */
        for (j = 0; j < 16; j++) {
            current_fragment += travel_width[j] + right_edge * travel_height[j];
            current_width += travel_width[j];
            current_height += travel_height[j];

            /* check if the fragment is in bounds */
            if ((current_width < right_edge) &&
                (current_height < bottom_edge)) {
                s->superblock_fragments[mapping_index] = current_fragment;
            } else {
                s->superblock_fragments[mapping_index] = -1;
            }

            mapping_index++;
        }
    }

    return 0;  /* successful path out */
}

/*
 * This function wipes out all of the fragment data.
 */
static void init_frame(Vp3DecodeContext *s, GetBitContext *gb)
{
    int i;

    /* zero out all of the fragment information */
    for (i = 0; i < s->fragment_count; i++) {
        s->all_fragments[i].motion_x = 127;
        s->all_fragments[i].motion_y = 127;
        s->all_fragments[i].dc = 0;
        s->all_fragments[i].qpi = 0;
    }
}

/*
 * This function sets up the dequantization tables used for a particular
 * frame.
 */
static void init_dequantizer(Vp3DecodeContext *s, int qpi)
{
    int ac_scale_factor = s->coded_ac_scale_factor[s->qps[qpi]];
    int dc_scale_factor = s->coded_dc_scale_factor[s->qps[qpi]];
    int i, plane, inter, qri, bmi, bmj, qistart;

    for(inter=0; inter<2; inter++){
        for(plane=0; plane<3; plane++){
            int sum=0;
            for(qri=0; qri<s->qr_count[inter][plane]; qri++){
                sum+= s->qr_size[inter][plane][qri];
                if(s->qps[qpi] <= sum)
                    break;
            }
            qistart= sum - s->qr_size[inter][plane][qri];
            bmi= s->qr_base[inter][plane][qri  ];
            bmj= s->qr_base[inter][plane][qri+1];
            for(i=0; i<64; i++){
                int coeff= (  2*(sum    -s->qps[qpi])*s->base_matrix[bmi][i]
                            - 2*(qistart-s->qps[qpi])*s->base_matrix[bmj][i]
                            + s->qr_size[inter][plane][qri])
                           / (2*s->qr_size[inter][plane][qri]);

                int qmin= 8<<(inter + !i);
                int qscale= i ? ac_scale_factor : dc_scale_factor;

                s->qmat[qpi][inter][plane][s->dsp.idct_permutation[i]]= av_clip((qscale * coeff)/100 * 4, qmin, 4096);
            }
            // all DC coefficients use the same quant so as not to interfere with DC prediction
            s->qmat[qpi][inter][plane][0] = s->qmat[0][inter][plane][0];
        }
    }

    memset(s->qscale_table, (FFMAX(s->qmat[0][0][0][1], s->qmat[0][0][1][1])+8)/16, 512); //FIXME finetune
}

/*
 * This function initializes the loop filter boundary limits if the frame's
 * quality index is different from the previous frame's.
 *
 * The filter_limit_values may not be larger than 127.
 */
static void init_loop_filter(Vp3DecodeContext *s)
{
    int *bounding_values= s->bounding_values_array+127;
    int filter_limit;
    int x;
    int value;

    filter_limit = s->filter_limit_values[s->qps[0]];

    /* set up the bounding values */
    memset(s->bounding_values_array, 0, 256 * sizeof(int));
    for (x = 0; x < filter_limit; x++) {
        bounding_values[-x] = -x;
        bounding_values[x] = x;
    }
    for (x = value = filter_limit; x < 128 && value; x++, value--) {
        bounding_values[ x] =  value;
        bounding_values[-x] = -value;
    }
    if (value)
        bounding_values[128] = value;
    bounding_values[129] = bounding_values[130] = filter_limit * 0x02020202;
}

/*
 * This function unpacks all of the superblock/macroblock/fragment coding
 * information from the bitstream.
 */
static int unpack_superblocks(Vp3DecodeContext *s, GetBitContext *gb)
{
    int superblock_starts[3] = { 0, s->u_superblock_start, s->v_superblock_start };
    int bit = 0;
    int current_superblock = 0;
    int current_run = 0;
    int num_partial_superblocks = 0;

    int i, j;
    int current_fragment;
    int plane;

    if (s->keyframe) {
        memset(s->superblock_coding, SB_FULLY_CODED, s->superblock_count);

    } else {

        /* unpack the list of partially-coded superblocks */
        bit = get_bits1(gb);
        while (current_superblock < s->superblock_count) {
                current_run = get_vlc2(gb,
                    s->superblock_run_length_vlc.table, 6, 2) + 1;
                if (current_run == 34)
                    current_run += get_bits(gb, 12);

            if (current_superblock + current_run > s->superblock_count) {
                av_log(s->avctx, AV_LOG_ERROR, "Invalid partially coded superblock run length\n");
                return -1;
            }

            memset(s->superblock_coding + current_superblock, bit, current_run);

            current_superblock += current_run;
            if (bit)
                num_partial_superblocks += current_run;

            if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
                bit = get_bits1(gb);
            else
                bit ^= 1;
        }

        /* unpack the list of fully coded superblocks if any of the blocks were
         * not marked as partially coded in the previous step */
        if (num_partial_superblocks < s->superblock_count) {
            int superblocks_decoded = 0;

            current_superblock = 0;
            bit = get_bits1(gb);
            while (superblocks_decoded < s->superblock_count - num_partial_superblocks) {
                        current_run = get_vlc2(gb,
                            s->superblock_run_length_vlc.table, 6, 2) + 1;
                        if (current_run == 34)
                            current_run += get_bits(gb, 12);

                for (j = 0; j < current_run; current_superblock++) {
                    if (current_superblock >= s->superblock_count) {
                        av_log(s->avctx, AV_LOG_ERROR, "Invalid fully coded superblock run length\n");
                        return -1;
                    }

                /* skip any superblocks already marked as partially coded */
                if (s->superblock_coding[current_superblock] == SB_NOT_CODED) {
                    s->superblock_coding[current_superblock] = 2*bit;
                    j++;
                }
                }
                superblocks_decoded += current_run;

                if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
                    bit = get_bits1(gb);
                else
                    bit ^= 1;
            }
        }

        /* if there were partial blocks, initialize bitstream for
         * unpacking fragment codings */
        if (num_partial_superblocks) {

            current_run = 0;
            bit = get_bits1(gb);
            /* toggle the bit because as soon as the first run length is
             * fetched the bit will be toggled again */
            bit ^= 1;
        }
    }

    /* figure out which fragments are coded; iterate through each
     * superblock (all planes) */
    s->total_num_coded_frags = 0;
    memset(s->macroblock_coding, MODE_COPY, s->macroblock_count);

    for (plane = 0; plane < 3; plane++) {
        int sb_start = superblock_starts[plane];
        int sb_end = sb_start + (plane ? s->c_superblock_count : s->y_superblock_count);
        int num_coded_frags = 0;

    for (i = sb_start; i < sb_end; i++) {

        /* iterate through all 16 fragments in a superblock */
        for (j = 0; j < 16; j++) {

            /* if the fragment is in bounds, check its coding status */
            current_fragment = s->superblock_fragments[i * 16 + j];
            if (current_fragment >= s->fragment_count) {
                av_log(s->avctx, AV_LOG_ERROR, "  vp3:unpack_superblocks(): bad fragment number (%d >= %d)\n",
                    current_fragment, s->fragment_count);
                return 1;
            }
            if (current_fragment != -1) {
                int coded = s->superblock_coding[i];

                if (s->superblock_coding[i] == SB_PARTIALLY_CODED) {

                    /* fragment may or may not be coded; this is the case
                     * that cares about the fragment coding runs */
                    if (current_run-- == 0) {
                        bit ^= 1;
                        current_run = get_vlc2(gb,
                            s->fragment_run_length_vlc.table, 5, 2);
                    }
                    coded = bit;
                }

                    if (coded) {
                        /* default mode; actual mode will be decoded in
                         * the next phase */
                        s->all_fragments[current_fragment].coding_method =
                            MODE_INTER_NO_MV;
                        s->coded_fragment_list[plane][num_coded_frags++] =
                            current_fragment;
                    } else {
                        /* not coded; copy this fragment from the prior frame */
                        s->all_fragments[current_fragment].coding_method =
                            MODE_COPY;
                    }
            }
        }
    }
        s->total_num_coded_frags += num_coded_frags;
        for (i = 0; i < 64; i++)
            s->num_coded_frags[plane][i] = num_coded_frags;
        if (plane < 2)
            s->coded_fragment_list[plane+1] = s->coded_fragment_list[plane] + num_coded_frags;
    }
    return 0;
}

/*
 * This function unpacks all the coding mode data for individual macroblocks
 * from the bitstream.
 */
static int unpack_modes(Vp3DecodeContext *s, GetBitContext *gb)
{
    int i, j, k, sb_x, sb_y;
    int scheme;
    int current_macroblock;
    int current_fragment;
    int coding_mode;
    int custom_mode_alphabet[CODING_MODE_COUNT];
    const int *alphabet;

    if (s->keyframe) {
        for (i = 0; i < s->fragment_count; i++)
            s->all_fragments[i].coding_method = MODE_INTRA;

    } else {

        /* fetch the mode coding scheme for this frame */
        scheme = get_bits(gb, 3);

        /* is it a custom coding scheme? */
        if (scheme == 0) {
            for (i = 0; i < 8; i++)
                custom_mode_alphabet[i] = MODE_INTER_NO_MV;
            for (i = 0; i < 8; i++)
                custom_mode_alphabet[get_bits(gb, 3)] = i;
            alphabet = custom_mode_alphabet;
        } else
            alphabet = ModeAlphabet[scheme-1];

        /* iterate through all of the macroblocks that contain 1 or more
         * coded fragments */
        for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
            for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {

            for (j = 0; j < 4; j++) {
                int mb_x = 2*sb_x +   (j>>1);
                int mb_y = 2*sb_y + (((j>>1)+j)&1);
                current_macroblock = mb_y * s->macroblock_width + mb_x;

                if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height)
                    continue;

#define BLOCK_X (2*mb_x + (k&1))
#define BLOCK_Y (2*mb_y + (k>>1))
                /* coding modes are only stored if the macroblock has at least one
                 * luma block coded, otherwise it must be INTER_NO_MV */
                for (k = 0; k < 4; k++) {
                    current_fragment = BLOCK_Y*s->fragment_width + BLOCK_X;
                    if (s->all_fragments[current_fragment].coding_method != MODE_COPY)
                        break;
                }
                if (k == 4) {
                    s->macroblock_coding[current_macroblock] = MODE_INTER_NO_MV;
                    continue;
                }

                /* mode 7 means get 3 bits for each coding mode */
                if (scheme == 7)
                    coding_mode = get_bits(gb, 3);
                else
                    coding_mode = alphabet
                        [get_vlc2(gb, s->mode_code_vlc.table, 3, 3)];

                s->macroblock_coding[current_macroblock] = coding_mode;
                for (k = 0; k < 4; k++) {
                    current_fragment =
                        BLOCK_Y*s->fragment_width + BLOCK_X;
                    if (s->all_fragments[current_fragment].coding_method !=
                        MODE_COPY)
                        s->all_fragments[current_fragment].coding_method =
                            coding_mode;
                }
                for (k = 0; k < 2; k++) {
                    current_fragment = s->fragment_start[k+1] +
                        mb_y*(s->fragment_width>>1) + mb_x;
                    if (s->all_fragments[current_fragment].coding_method !=
                        MODE_COPY)
                        s->all_fragments[current_fragment].coding_method =
                            coding_mode;
                }
            }
            }
        }
    }

    return 0;
}

/*
 * This function unpacks all the motion vectors for the individual
 * macroblocks from the bitstream.
 */
static int unpack_vectors(Vp3DecodeContext *s, GetBitContext *gb)
{
    int j, k, sb_x, sb_y;
    int coding_mode;
    int motion_x[6];
    int motion_y[6];
    int last_motion_x = 0;
    int last_motion_y = 0;
    int prior_last_motion_x = 0;
    int prior_last_motion_y = 0;
    int current_macroblock;
    int current_fragment;

    if (s->keyframe)
        return 0;

    memset(motion_x, 0, 6 * sizeof(int));
    memset(motion_y, 0, 6 * sizeof(int));

    /* coding mode 0 is the VLC scheme; 1 is the fixed code scheme */
    coding_mode = get_bits1(gb);

    /* iterate through all of the macroblocks that contain 1 or more
     * coded fragments */
    for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
        for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {

        for (j = 0; j < 4; j++) {
            int mb_x = 2*sb_x +   (j>>1);
            int mb_y = 2*sb_y + (((j>>1)+j)&1);
            current_macroblock = mb_y * s->macroblock_width + mb_x;

            if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height ||
                (s->macroblock_coding[current_macroblock] == MODE_COPY))
                continue;

            switch (s->macroblock_coding[current_macroblock]) {

            case MODE_INTER_PLUS_MV:
            case MODE_GOLDEN_MV:
                /* all 6 fragments use the same motion vector */
                if (coding_mode == 0) {
                    motion_x[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
                    motion_y[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
                } else {
                    motion_x[0] = fixed_motion_vector_table[get_bits(gb, 6)];
                    motion_y[0] = fixed_motion_vector_table[get_bits(gb, 6)];
                }

                /* vector maintenance, only on MODE_INTER_PLUS_MV */
                if (s->macroblock_coding[current_macroblock] ==
                    MODE_INTER_PLUS_MV) {
                    prior_last_motion_x = last_motion_x;
                    prior_last_motion_y = last_motion_y;
                    last_motion_x = motion_x[0];
                    last_motion_y = motion_y[0];
                }
                break;

            case MODE_INTER_FOURMV:
                /* vector maintenance */
                prior_last_motion_x = last_motion_x;
                prior_last_motion_y = last_motion_y;

                /* fetch 4 vectors from the bitstream, one for each
                 * Y fragment, then average for the C fragment vectors */
                motion_x[4] = motion_y[4] = 0;
                for (k = 0; k < 4; k++) {
                    current_fragment = BLOCK_Y*s->fragment_width + BLOCK_X;
                    if (s->all_fragments[current_fragment].coding_method != MODE_COPY) {
                        if (coding_mode == 0) {
                            motion_x[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
                            motion_y[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
                        } else {
                            motion_x[k] = fixed_motion_vector_table[get_bits(gb, 6)];
                            motion_y[k] = fixed_motion_vector_table[get_bits(gb, 6)];
                        }
                        last_motion_x = motion_x[k];
                        last_motion_y = motion_y[k];
                    } else {
                        motion_x[k] = 0;
                        motion_y[k] = 0;
                    }
                    motion_x[4] += motion_x[k];
                    motion_y[4] += motion_y[k];
                }

                motion_x[5]=
                motion_x[4]= RSHIFT(motion_x[4], 2);
                motion_y[5]=
                motion_y[4]= RSHIFT(motion_y[4], 2);
                break;

            case MODE_INTER_LAST_MV:
                /* all 6 fragments use the last motion vector */
                motion_x[0] = last_motion_x;
                motion_y[0] = last_motion_y;

                /* no vector maintenance (last vector remains the
                 * last vector) */
                break;

            case MODE_INTER_PRIOR_LAST:
                /* all 6 fragments use the motion vector prior to the
                 * last motion vector */
                motion_x[0] = prior_last_motion_x;
                motion_y[0] = prior_last_motion_y;

                /* vector maintenance */
                prior_last_motion_x = last_motion_x;
                prior_last_motion_y = last_motion_y;
                last_motion_x = motion_x[0];
                last_motion_y = motion_y[0];
                break;

            default:
                /* covers intra, inter without MV, golden without MV */
                motion_x[0] = 0;
                motion_y[0] = 0;

                /* no vector maintenance */
                break;
            }

            /* assign the motion vectors to the correct fragments */
            for (k = 0; k < 4; k++) {
                current_fragment =
                    BLOCK_Y*s->fragment_width + BLOCK_X;
                if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
                    s->all_fragments[current_fragment].motion_x = motion_x[k];
                    s->all_fragments[current_fragment].motion_y = motion_y[k];
                } else {
                    s->all_fragments[current_fragment].motion_x = motion_x[0];
                    s->all_fragments[current_fragment].motion_y = motion_y[0];
                }
            }
            for (k = 0; k < 2; k++) {
                current_fragment = s->fragment_start[k+1] +
                    mb_y*(s->fragment_width>>1) + mb_x;
                if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
                    s->all_fragments[current_fragment].motion_x = motion_x[k+4];
                    s->all_fragments[current_fragment].motion_y = motion_y[k+4];
                } else {
                    s->all_fragments[current_fragment].motion_x = motion_x[0];
                    s->all_fragments[current_fragment].motion_y = motion_y[0];
                }
            }
        }
        }
    }

    return 0;
}

static int unpack_block_qpis(Vp3DecodeContext *s, GetBitContext *gb)
{
    int qpi, i, j, bit, run_length, blocks_decoded, num_blocks_at_qpi;
    int num_blocks = s->total_num_coded_frags;

    for (qpi = 0; qpi < s->nqps-1 && num_blocks > 0; qpi++) {
        i = blocks_decoded = num_blocks_at_qpi = 0;

        bit = get_bits1(gb);

        do {
            run_length = get_vlc2(gb, s->superblock_run_length_vlc.table, 6, 2) + 1;
            if (run_length == 34)
                run_length += get_bits(gb, 12);
            blocks_decoded += run_length;

            if (!bit)
                num_blocks_at_qpi += run_length;

            for (j = 0; j < run_length; i++) {
                if (i >= s->total_num_coded_frags)
                    return -1;

                if (s->all_fragments[s->coded_fragment_list[0][i]].qpi == qpi) {
                    s->all_fragments[s->coded_fragment_list[0][i]].qpi += bit;
                    j++;
                }
            }

            if (run_length == MAXIMUM_LONG_BIT_RUN)
                bit = get_bits1(gb);
            else
                bit ^= 1;
        } while (blocks_decoded < num_blocks);

        num_blocks -= num_blocks_at_qpi;
    }

    return 0;
}

/*
 * This function is called by unpack_dct_coeffs() to extract the VLCs from
 * the bitstream. The VLCs encode tokens which are used to unpack DCT
 * data. This function unpacks all the VLCs for either the Y plane or both
 * C planes, and is called for DC coefficients or different AC coefficient
 * levels (since different coefficient types require different VLC tables.
 *
 * This function returns a residual eob run. E.g, if a particular token gave
 * instructions to EOB the next 5 fragments and there were only 2 fragments
 * left in the current fragment range, 3 would be returned so that it could
 * be passed into the next call to this same function.
 */
static int unpack_vlcs(Vp3DecodeContext *s, GetBitContext *gb,
                        VLC *table, int coeff_index,
                        int plane,
                        int eob_run)
{
    int i, j = 0;
    int token;
    int zero_run = 0;
    DCTELEM coeff = 0;
    int bits_to_get;
    int blocks_ended;
    int coeff_i = 0;
    int num_coeffs = s->num_coded_frags[plane][coeff_index];
    int16_t *dct_tokens = s->dct_tokens[plane][coeff_index];

    /* local references to structure members to avoid repeated deferences */
    int *coded_fragment_list = s->coded_fragment_list[plane];
    Vp3Fragment *all_fragments = s->all_fragments;
    VLC_TYPE (*vlc_table)[2] = table->table;

    if (num_coeffs < 0)
        av_log(s->avctx, AV_LOG_ERROR, "Invalid number of coefficents at level %d\n", coeff_index);

    if (eob_run > num_coeffs) {
        coeff_i = blocks_ended = num_coeffs;
        eob_run -= num_coeffs;
    } else {
        coeff_i = blocks_ended = eob_run;
        eob_run = 0;
    }

    // insert fake EOB token to cover the split between planes or zzi
    if (blocks_ended)
        dct_tokens[j++] = blocks_ended << 2;

    while (coeff_i < num_coeffs) {
            /* decode a VLC into a token */
            token = get_vlc2(gb, vlc_table, 5, 3);
            /* use the token to get a zero run, a coefficient, and an eob run */
            if (token <= 6) {
                eob_run = eob_run_base[token];
                if (eob_run_get_bits[token])
                    eob_run += get_bits(gb, eob_run_get_bits[token]);

                // record only the number of blocks ended in this plane,
                // any spill will be recorded in the next plane.
                if (eob_run > num_coeffs - coeff_i) {
                    dct_tokens[j++] = TOKEN_EOB(num_coeffs - coeff_i);
                    blocks_ended   += num_coeffs - coeff_i;
                    eob_run        -= num_coeffs - coeff_i;
                    coeff_i         = num_coeffs;
                } else {
                    dct_tokens[j++] = TOKEN_EOB(eob_run);
                    blocks_ended   += eob_run;
                    coeff_i        += eob_run;
                    eob_run = 0;
                }
            } else {
                bits_to_get = coeff_get_bits[token];
                if (bits_to_get)
                    bits_to_get = get_bits(gb, bits_to_get);
                coeff = coeff_tables[token][bits_to_get];

                zero_run = zero_run_base[token];
                if (zero_run_get_bits[token])
                    zero_run += get_bits(gb, zero_run_get_bits[token]);

                if (zero_run) {
                    dct_tokens[j++] = TOKEN_ZERO_RUN(coeff, zero_run);
                } else {
                    // Save DC into the fragment structure. DC prediction is
                    // done in raster order, so the actual DC can't be in with
                    // other tokens. We still need the token in dct_tokens[]
                    // however, or else the structure collapses on itself.
                    if (!coeff_index)
                        all_fragments[coded_fragment_list[coeff_i]].dc = coeff;

                    dct_tokens[j++] = TOKEN_COEFF(coeff);
                }

                if (coeff_index + zero_run > 64) {
                    av_log(s->avctx, AV_LOG_ERROR, "Invalid zero run of %d with"
                           " %d coeffs left\n", zero_run, 64-coeff_index);
                    zero_run = 64 - coeff_index;
                }

                // zero runs code multiple coefficients,
                // so don't try to decode coeffs for those higher levels
                for (i = coeff_index+1; i <= coeff_index+zero_run; i++)
                    s->num_coded_frags[plane][i]--;
                coeff_i++;
            }
    }

    if (blocks_ended > s->num_coded_frags[plane][coeff_index])
        av_log(s->avctx, AV_LOG_ERROR, "More blocks ended than coded!\n");

    // decrement the number of blocks that have higher coeffecients for each
    // EOB run at this level
    if (blocks_ended)
        for (i = coeff_index+1; i < 64; i++)
            s->num_coded_frags[plane][i] -= blocks_ended;

    // setup the next buffer
    if (plane < 2)
        s->dct_tokens[plane+1][coeff_index] = dct_tokens + j;
    else if (coeff_index < 63)
        s->dct_tokens[0][coeff_index+1] = dct_tokens + j;

    return eob_run;
}

static void reverse_dc_prediction(Vp3DecodeContext *s,
                                  int first_fragment,
                                  int fragment_width,
                                  int fragment_height);
/*
 * This function unpacks all of the DCT coefficient data from the
 * bitstream.
 */
static int unpack_dct_coeffs(Vp3DecodeContext *s, GetBitContext *gb)
{
    int i;
    int dc_y_table;
    int dc_c_table;
    int ac_y_table;
    int ac_c_table;
    int residual_eob_run = 0;
    VLC *y_tables[64];
    VLC *c_tables[64];

    s->dct_tokens[0][0] = s->dct_tokens_base;

    /* fetch the DC table indexes */
    dc_y_table = get_bits(gb, 4);
    dc_c_table = get_bits(gb, 4);

    /* unpack the Y plane DC coefficients */
    residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_y_table], 0,
        0, residual_eob_run);

    /* reverse prediction of the Y-plane DC coefficients */
    reverse_dc_prediction(s, 0, s->fragment_width, s->fragment_height);

    /* unpack the C plane DC coefficients */
    residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
        1, residual_eob_run);
    residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
        2, residual_eob_run);

    /* reverse prediction of the C-plane DC coefficients */
    if (!(s->avctx->flags & CODEC_FLAG_GRAY))
    {
        reverse_dc_prediction(s, s->fragment_start[1],
            s->fragment_width / 2, s->fragment_height / 2);
        reverse_dc_prediction(s, s->fragment_start[2],
            s->fragment_width / 2, s->fragment_height / 2);
    }

    /* fetch the AC table indexes */
    ac_y_table = get_bits(gb, 4);
    ac_c_table = get_bits(gb, 4);

    /* build tables of AC VLC tables */
    for (i = 1; i <= 5; i++) {
        y_tables[i] = &s->ac_vlc_1[ac_y_table];
        c_tables[i] = &s->ac_vlc_1[ac_c_table];
    }
    for (i = 6; i <= 14; i++) {
        y_tables[i] = &s->ac_vlc_2[ac_y_table];
        c_tables[i] = &s->ac_vlc_2[ac_c_table];
    }
    for (i = 15; i <= 27; i++) {
        y_tables[i] = &s->ac_vlc_3[ac_y_table];
        c_tables[i] = &s->ac_vlc_3[ac_c_table];
    }
    for (i = 28; i <= 63; i++) {
        y_tables[i] = &s->ac_vlc_4[ac_y_table];
        c_tables[i] = &s->ac_vlc_4[ac_c_table];
    }

    /* decode all AC coefficents */
    for (i = 1; i <= 63; i++) {
            residual_eob_run = unpack_vlcs(s, gb, y_tables[i], i,
                0, residual_eob_run);

            residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
                1, residual_eob_run);
            residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
                2, residual_eob_run);
    }

    return 0;
}

/*
 * This function reverses the DC prediction for each coded fragment in
 * the frame. Much of this function is adapted directly from the original
 * VP3 source code.
 */
#define COMPATIBLE_FRAME(x) \
  (compatible_frame[s->all_fragments[x].coding_method] == current_frame_type)
#define DC_COEFF(u) s->all_fragments[u].dc

static void reverse_dc_prediction(Vp3DecodeContext *s,
                                  int first_fragment,
                                  int fragment_width,
                                  int fragment_height)
{

#define PUL 8
#define PU 4
#define PUR 2
#define PL 1

    int x, y;
    int i = first_fragment;

    int predicted_dc;

    /* DC values for the left, up-left, up, and up-right fragments */
    int vl, vul, vu, vur;

    /* indexes for the left, up-left, up, and up-right fragments */
    int l, ul, u, ur;

    /*
     * The 6 fields mean:
     *   0: up-left multiplier
     *   1: up multiplier
     *   2: up-right multiplier
     *   3: left multiplier
     */
    static const int predictor_transform[16][4] = {
        {  0,  0,  0,  0},
        {  0,  0,  0,128},        // PL
        {  0,  0,128,  0},        // PUR
        {  0,  0, 53, 75},        // PUR|PL
        {  0,128,  0,  0},        // PU
        {  0, 64,  0, 64},        // PU|PL
        {  0,128,  0,  0},        // PU|PUR
        {  0,  0, 53, 75},        // PU|PUR|PL
        {128,  0,  0,  0},        // PUL
        {  0,  0,  0,128},        // PUL|PL
        { 64,  0, 64,  0},        // PUL|PUR
        {  0,  0, 53, 75},        // PUL|PUR|PL
        {  0,128,  0,  0},        // PUL|PU
       {-104,116,  0,116},        // PUL|PU|PL
        { 24, 80, 24,  0},        // PUL|PU|PUR
       {-104,116,  0,116}         // PUL|PU|PUR|PL
    };

    /* This table shows which types of blocks can use other blocks for
     * prediction. For example, INTRA is the only mode in this table to
     * have a frame number of 0. That means INTRA blocks can only predict
     * from other INTRA blocks. There are 2 golden frame coding types;
     * blocks encoding in these modes can only predict from other blocks
     * that were encoded with these 1 of these 2 modes. */
    static const unsigned char compatible_frame[9] = {
        1,    /* MODE_INTER_NO_MV */
        0,    /* MODE_INTRA */
        1,    /* MODE_INTER_PLUS_MV */
        1,    /* MODE_INTER_LAST_MV */
        1,    /* MODE_INTER_PRIOR_MV */
        2,    /* MODE_USING_GOLDEN */
        2,    /* MODE_GOLDEN_MV */
        1,    /* MODE_INTER_FOUR_MV */
        3     /* MODE_COPY */
    };
    int current_frame_type;

    /* there is a last DC predictor for each of the 3 frame types */
    short last_dc[3];

    int transform = 0;

    vul = vu = vur = vl = 0;
    last_dc[0] = last_dc[1] = last_dc[2] = 0;

    /* for each fragment row... */
    for (y = 0; y < fragment_height; y++) {

        /* for each fragment in a row... */
        for (x = 0; x < fragment_width; x++, i++) {

            /* reverse prediction if this block was coded */
            if (s->all_fragments[i].coding_method != MODE_COPY) {

                current_frame_type =
                    compatible_frame[s->all_fragments[i].coding_method];

                transform= 0;
                if(x){
                    l= i-1;
                    vl = DC_COEFF(l);
                    if(COMPATIBLE_FRAME(l))
                        transform |= PL;
                }
                if(y){
                    u= i-fragment_width;
                    vu = DC_COEFF(u);
                    if(COMPATIBLE_FRAME(u))
                        transform |= PU;
                    if(x){
                        ul= i-fragment_width-1;
                        vul = DC_COEFF(ul);
                        if(COMPATIBLE_FRAME(ul))
                            transform |= PUL;
                    }
                    if(x + 1 < fragment_width){
                        ur= i-fragment_width+1;
                        vur = DC_COEFF(ur);
                        if(COMPATIBLE_FRAME(ur))
                            transform |= PUR;
                    }
                }

                if (transform == 0) {

                    /* if there were no fragments to predict from, use last
                     * DC saved */
                    predicted_dc = last_dc[current_frame_type];
                } else {

                    /* apply the appropriate predictor transform */
                    predicted_dc =
                        (predictor_transform[transform][0] * vul) +
                        (predictor_transform[transform][1] * vu) +
                        (predictor_transform[transform][2] * vur) +
                        (predictor_transform[transform][3] * vl);

                    predicted_dc /= 128;

                    /* check for outranging on the [ul u l] and
                     * [ul u ur l] predictors */
                    if ((transform == 15) || (transform == 13)) {
                        if (FFABS(predicted_dc - vu) > 128)
                            predicted_dc = vu;
                        else if (FFABS(predicted_dc - vl) > 128)
                            predicted_dc = vl;
                        else if (FFABS(predicted_dc - vul) > 128)
                            predicted_dc = vul;
                    }
                }

                /* at long last, apply the predictor */
                DC_COEFF(i) += predicted_dc;
                /* save the DC */
                last_dc[current_frame_type] = DC_COEFF(i);
            }
        }
    }
}

static void apply_loop_filter(Vp3DecodeContext *s, int plane, int ystart, int yend)
{
    int x, y;
    int *bounding_values= s->bounding_values_array+127;

    int width           = s->fragment_width  >> !!plane;
    int height          = s->fragment_height >> !!plane;
    int fragment        = s->fragment_start        [plane] + ystart * width;
    int stride          = s->current_frame.linesize[plane];
    uint8_t *plane_data = s->current_frame.data    [plane];
    if (!s->flipped_image) stride = -stride;
    plane_data += s->data_offset[plane] + 8*ystart*stride;

    for (y = ystart; y < yend; y++) {

        for (x = 0; x < width; x++) {
            /* This code basically just deblocks on the edges of coded blocks.
             * However, it has to be much more complicated because of the
             * braindamaged deblock ordering used in VP3/Theora. Order matters
             * because some pixels get filtered twice. */
            if( s->all_fragments[fragment].coding_method != MODE_COPY )
            {
                /* do not perform left edge filter for left columns frags */
                if (x > 0) {
                    s->dsp.vp3_h_loop_filter(
                        plane_data + 8*x,
                        stride, bounding_values);
                }

                /* do not perform top edge filter for top row fragments */
                if (y > 0) {
                    s->dsp.vp3_v_loop_filter(
                        plane_data + 8*x,
                        stride, bounding_values);
                }

                /* do not perform right edge filter for right column
                 * fragments or if right fragment neighbor is also coded
                 * in this frame (it will be filtered in next iteration) */
                if ((x < width - 1) &&
                    (s->all_fragments[fragment + 1].coding_method == MODE_COPY)) {
                    s->dsp.vp3_h_loop_filter(
                        plane_data + 8*x + 8,
                        stride, bounding_values);
                }

                /* do not perform bottom edge filter for bottom row
                 * fragments or if bottom fragment neighbor is also coded
                 * in this frame (it will be filtered in the next row) */
                if ((y < height - 1) &&
                    (s->all_fragments[fragment + width].coding_method == MODE_COPY)) {
                    s->dsp.vp3_v_loop_filter(
                        plane_data + 8*x + 8*stride,
                        stride, bounding_values);
                }
            }

            fragment++;
        }
        plane_data += 8*stride;
    }
}

/**
 * Pulls DCT tokens from the 64 levels to decode and dequant the coefficients
 * for the next block in coding order
 */
static inline int vp3_dequant(Vp3DecodeContext *s, Vp3Fragment *frag,
                              int plane, int inter, DCTELEM block[64])
{
    int16_t *dequantizer = s->qmat[frag->qpi][inter][plane];
    uint8_t *perm = s->scantable.permutated;
    int i = 0;

    do {
        int token = *s->dct_tokens[plane][i];
        switch (token & 3) {
        case 0: // EOB
            if (--token < 4) // 0-3 are token types, so the EOB run must now be 0
                s->dct_tokens[plane][i]++;
            else
                *s->dct_tokens[plane][i] = token & ~3;
            goto end;
        case 1: // zero run
            s->dct_tokens[plane][i]++;
            i += (token >> 2) & 0x7f;
            block[perm[i]] = (token >> 9) * dequantizer[perm[i]];
            i++;
            break;
        case 2: // coeff
            block[perm[i]] = (token >> 2) * dequantizer[perm[i]];
            s->dct_tokens[plane][i++]++;
            break;
        default:
            av_log(s->avctx, AV_LOG_ERROR, "internal: invalid token type\n");
            return i;
        }
    } while (i < 64);
end:
    // the actual DC+prediction is in the fragment structure
    block[0] = frag->dc * s->qmat[0][inter][plane][0];
    return i;
}

/**
 * called when all pixels up to row y are complete
 */
static void vp3_draw_horiz_band(Vp3DecodeContext *s, int y)
{
    int h, cy;
    int offset[4];

    if(s->avctx->draw_horiz_band==NULL)
        return;

    h= y - s->last_slice_end;
    y -= h;

    if (!s->flipped_image) {
        if (y == 0)
            h -= s->height - s->avctx->height;  // account for non-mod16
        y = s->height - y - h;
    }

    cy = y >> 1;
    offset[0] = s->current_frame.linesize[0]*y;
    offset[1] = s->current_frame.linesize[1]*cy;
    offset[2] = s->current_frame.linesize[2]*cy;
    offset[3] = 0;

    emms_c();
    s->avctx->draw_horiz_band(s->avctx, &s->current_frame, offset, y, 3, h);
    s->last_slice_end= y + h;
}

/*
 * Perform the final rendering for a particular slice of data.
 * The slice number ranges from 0..(c_superblock_height - 1).
 */
static void render_slice(Vp3DecodeContext *s, int slice)
{
    int x, y, i, j;
    LOCAL_ALIGNED_16(DCTELEM, block, [64]);
    int motion_x = 0xdeadbeef, motion_y = 0xdeadbeef;
    int motion_halfpel_index;
    uint8_t *motion_source;
    int plane, first_pixel;

    if (slice >= s->c_superblock_height)
        return;

    for (plane = 0; plane < 3; plane++) {
        uint8_t *output_plane = s->current_frame.data    [plane] + s->data_offset[plane];
        uint8_t *  last_plane = s->   last_frame.data    [plane] + s->data_offset[plane];
        uint8_t *golden_plane = s-> golden_frame.data    [plane] + s->data_offset[plane];
        int stride            = s->current_frame.linesize[plane];
        int plane_width       = s->width  >> !!plane;
        int plane_height      = s->height >> !!plane;

        int sb_x, sb_y        = slice << !plane;
        int slice_height      = sb_y + (plane ? 1 : 2);
        int slice_width       = plane ? s->c_superblock_width : s->y_superblock_width;

        int fragment_width    = s->fragment_width  >> !!plane;
        int fragment_height   = s->fragment_height >> !!plane;
        int fragment_start    = s->fragment_start[plane];

        if (!s->flipped_image) stride = -stride;
        if (CONFIG_GRAY && plane && (s->avctx->flags & CODEC_FLAG_GRAY))
            continue;


        if(FFABS(stride) > 2048)
            return; //various tables are fixed size

        /* for each superblock row in the slice (both of them)... */
        for (; sb_y < slice_height; sb_y++) {

            /* for each superblock in a row... */
            for (sb_x = 0; sb_x < slice_width; sb_x++) {

                /* for each block in a superblock... */
                for (j = 0; j < 16; j++) {
                    x = 4*sb_x + hilbert_offset[j][0];
                    y = 4*sb_y + hilbert_offset[j][1];

                    i = fragment_start + y*fragment_width + x;

                    // bounds check
                    if (x >= fragment_width || y >= fragment_height)
                        continue;

                first_pixel = 8*y*stride + 8*x;

                /* transform if this block was coded */
                if (s->all_fragments[i].coding_method != MODE_COPY) {
                    int intra = s->all_fragments[i].coding_method == MODE_INTRA;

                    if ((s->all_fragments[i].coding_method == MODE_USING_GOLDEN) ||
                        (s->all_fragments[i].coding_method == MODE_GOLDEN_MV))
                        motion_source= golden_plane;
                    else
                        motion_source= last_plane;

                    motion_source += first_pixel;
                    motion_halfpel_index = 0;

                    /* sort out the motion vector if this fragment is coded
                     * using a motion vector method */
                    if ((s->all_fragments[i].coding_method > MODE_INTRA) &&
                        (s->all_fragments[i].coding_method != MODE_USING_GOLDEN)) {
                        int src_x, src_y;
                        motion_x = s->all_fragments[i].motion_x;
                        motion_y = s->all_fragments[i].motion_y;
                        if(plane){
                            motion_x= (motion_x>>1) | (motion_x&1);
                            motion_y= (motion_y>>1) | (motion_y&1);
                        }

                        src_x= (motion_x>>1) + 8*x;
                        src_y= (motion_y>>1) + 8*y;
                        if ((motion_x == 127) || (motion_y == 127))
                            av_log(s->avctx, AV_LOG_ERROR, " help! got invalid motion vector! (%X, %X)\n", motion_x, motion_y);

                        motion_halfpel_index = motion_x & 0x01;
                        motion_source += (motion_x >> 1);

                        motion_halfpel_index |= (motion_y & 0x01) << 1;
                        motion_source += ((motion_y >> 1) * stride);

                        if(src_x<0 || src_y<0 || src_x + 9 >= plane_width || src_y + 9 >= plane_height){
                            uint8_t *temp= s->edge_emu_buffer;
                            if(stride<0) temp -= 9*stride;
                            else temp += 9*stride;

                            ff_emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height);
                            motion_source= temp;
                        }
                    }


                    /* first, take care of copying a block from either the
                     * previous or the golden frame */
                    if (s->all_fragments[i].coding_method != MODE_INTRA) {
                        /* Note, it is possible to implement all MC cases with
                           put_no_rnd_pixels_l2 which would look more like the
                           VP3 source but this would be slower as
                           put_no_rnd_pixels_tab is better optimzed */
                        if(motion_halfpel_index != 3){
                            s->dsp.put_no_rnd_pixels_tab[1][motion_halfpel_index](
                                output_plane + first_pixel,
                                motion_source, stride, 8);
                        }else{
                            int d= (motion_x ^ motion_y)>>31; // d is 0 if motion_x and _y have the same sign, else -1
                            s->dsp.put_no_rnd_pixels_l2[1](
                                output_plane + first_pixel,
                                motion_source - d,
                                motion_source + stride + 1 + d,
                                stride, 8);
                        }
                    }

                        s->dsp.clear_block(block);
                        vp3_dequant(s, s->all_fragments + i, plane, !intra, block);

                    /* invert DCT and place (or add) in final output */

                    if (s->all_fragments[i].coding_method == MODE_INTRA) {
                        if(s->avctx->idct_algo!=FF_IDCT_VP3)
                            block[0] += 128<<3;
                        s->dsp.idct_put(
                            output_plane + first_pixel,
                            stride,
                            block);
                    } else {
                        s->dsp.idct_add(
                            output_plane + first_pixel,
                            stride,
                            block);
                    }
                } else {

                    /* copy directly from the previous frame */
                    s->dsp.put_pixels_tab[1][0](
                        output_plane + first_pixel,
                        last_plane + first_pixel,
                        stride, 8);

                }
                }
            }

            // Filter up to the last row in the superblock row
            apply_loop_filter(s, plane, 4*sb_y - !!sb_y, FFMIN(4*sb_y+3, fragment_height-1));
        }
    }

     /* this looks like a good place for slice dispatch... */
     /* algorithm:
      *   if (slice == s->macroblock_height - 1)
      *     dispatch (both last slice & 2nd-to-last slice);
      *   else if (slice > 0)
      *     dispatch (slice - 1);
      */

    vp3_draw_horiz_band(s, 64*slice + 64-16);
}

/*
 * This is the ffmpeg/libavcodec API init function.
 */
static av_cold int vp3_decode_init(AVCodecContext *avctx)
{
    Vp3DecodeContext *s = avctx->priv_data;
    int i, inter, plane;
    int c_width;
    int c_height;

    if (avctx->codec_tag == MKTAG('V','P','3','0'))
        s->version = 0;
    else
        s->version = 1;

    s->avctx = avctx;
    s->width = FFALIGN(avctx->width, 16);
    s->height = FFALIGN(avctx->height, 16);
    avctx->pix_fmt = PIX_FMT_YUV420P;
    avctx->chroma_sample_location = AVCHROMA_LOC_CENTER;
    if(avctx->idct_algo==FF_IDCT_AUTO)
        avctx->idct_algo=FF_IDCT_VP3;
    dsputil_init(&s->dsp, avctx);

    ff_init_scantable(s->dsp.idct_permutation, &s->scantable, ff_zigzag_direct);

    /* initialize to an impossible value which will force a recalculation
     * in the first frame decode */
    for (i = 0; i < 3; i++)
        s->qps[i] = -1;

    s->y_superblock_width = (s->width + 31) / 32;
    s->y_superblock_height = (s->height + 31) / 32;
    s->y_superblock_count = s->y_superblock_width * s->y_superblock_height;

    /* work out the dimensions for the C planes */
    c_width = s->width / 2;
    c_height = s->height / 2;
    s->c_superblock_width = (c_width + 31) / 32;
    s->c_superblock_height = (c_height + 31) / 32;
    s->c_superblock_count = s->c_superblock_width * s->c_superblock_height;

    s->superblock_count = s->y_superblock_count + (s->c_superblock_count * 2);
    s->u_superblock_start = s->y_superblock_count;
    s->v_superblock_start = s->u_superblock_start + s->c_superblock_count;
    s->superblock_coding = av_malloc(s->superblock_count);

    s->macroblock_width = (s->width + 15) / 16;
    s->macroblock_height = (s->height + 15) / 16;
    s->macroblock_count = s->macroblock_width * s->macroblock_height;

    s->fragment_width = s->width / FRAGMENT_PIXELS;
    s->fragment_height = s->height / FRAGMENT_PIXELS;

    /* fragment count covers all 8x8 blocks for all 3 planes */
    s->fragment_count = s->fragment_width * s->fragment_height * 3 / 2;
    s->fragment_start[1] = s->fragment_width * s->fragment_height;
    s->fragment_start[2] = s->fragment_width * s->fragment_height * 5 / 4;

    s->all_fragments = av_malloc(s->fragment_count * sizeof(Vp3Fragment));
    s->coded_fragment_list[0] = av_malloc(s->fragment_count * sizeof(int));
    s->dct_tokens_base = av_malloc(64*s->fragment_count * sizeof(*s->dct_tokens_base));
    if (!s->superblock_coding || !s->all_fragments || !s->dct_tokens_base ||
        !s->coded_fragment_list[0]) {
        vp3_decode_end(avctx);
        return -1;
    }

    if (!s->theora_tables)
    {
        for (i = 0; i < 64; i++) {
            s->coded_dc_scale_factor[i] = vp31_dc_scale_factor[i];
            s->coded_ac_scale_factor[i] = vp31_ac_scale_factor[i];
            s->base_matrix[0][i] = vp31_intra_y_dequant[i];
            s->base_matrix[1][i] = vp31_intra_c_dequant[i];
            s->base_matrix[2][i] = vp31_inter_dequant[i];
            s->filter_limit_values[i] = vp31_filter_limit_values[i];
        }

        for(inter=0; inter<2; inter++){
            for(plane=0; plane<3; plane++){
                s->qr_count[inter][plane]= 1;
                s->qr_size [inter][plane][0]= 63;
                s->qr_base [inter][plane][0]=
                s->qr_base [inter][plane][1]= 2*inter + (!!plane)*!inter;
            }
        }

        /* init VLC tables */
        for (i = 0; i < 16; i++) {

            /* DC histograms */
            init_vlc(&s->dc_vlc[i], 5, 32,
                &dc_bias[i][0][1], 4, 2,
                &dc_bias[i][0][0], 4, 2, 0);

            /* group 1 AC histograms */
            init_vlc(&s->ac_vlc_1[i], 5, 32,
                &ac_bias_0[i][0][1], 4, 2,
                &ac_bias_0[i][0][0], 4, 2, 0);

            /* group 2 AC histograms */
            init_vlc(&s->ac_vlc_2[i], 5, 32,
                &ac_bias_1[i][0][1], 4, 2,
                &ac_bias_1[i][0][0], 4, 2, 0);

            /* group 3 AC histograms */
            init_vlc(&s->ac_vlc_3[i], 5, 32,
                &ac_bias_2[i][0][1], 4, 2,
                &ac_bias_2[i][0][0], 4, 2, 0);

            /* group 4 AC histograms */
            init_vlc(&s->ac_vlc_4[i], 5, 32,
                &ac_bias_3[i][0][1], 4, 2,
                &ac_bias_3[i][0][0], 4, 2, 0);
        }
    } else {
        for (i = 0; i < 16; i++) {

            /* DC histograms */
            if (init_vlc(&s->dc_vlc[i], 5, 32,
                &s->huffman_table[i][0][1], 4, 2,
                &s->huffman_table[i][0][0], 4, 2, 0) < 0)
                goto vlc_fail;

            /* group 1 AC histograms */
            if (init_vlc(&s->ac_vlc_1[i], 5, 32,
                &s->huffman_table[i+16][0][1], 4, 2,
                &s->huffman_table[i+16][0][0], 4, 2, 0) < 0)
                goto vlc_fail;

            /* group 2 AC histograms */
            if (init_vlc(&s->ac_vlc_2[i], 5, 32,
                &s->huffman_table[i+16*2][0][1], 4, 2,
                &s->huffman_table[i+16*2][0][0], 4, 2, 0) < 0)
                goto vlc_fail;

            /* group 3 AC histograms */
            if (init_vlc(&s->ac_vlc_3[i], 5, 32,
                &s->huffman_table[i+16*3][0][1], 4, 2,
                &s->huffman_table[i+16*3][0][0], 4, 2, 0) < 0)
                goto vlc_fail;

            /* group 4 AC histograms */
            if (init_vlc(&s->ac_vlc_4[i], 5, 32,
                &s->huffman_table[i+16*4][0][1], 4, 2,
                &s->huffman_table[i+16*4][0][0], 4, 2, 0) < 0)
                goto vlc_fail;
        }
    }

    init_vlc(&s->superblock_run_length_vlc, 6, 34,
        &superblock_run_length_vlc_table[0][1], 4, 2,
        &superblock_run_length_vlc_table[0][0], 4, 2, 0);

    init_vlc(&s->fragment_run_length_vlc, 5, 30,
        &fragment_run_length_vlc_table[0][1], 4, 2,
        &fragment_run_length_vlc_table[0][0], 4, 2, 0);

    init_vlc(&s->mode_code_vlc, 3, 8,
        &mode_code_vlc_table[0][1], 2, 1,
        &mode_code_vlc_table[0][0], 2, 1, 0);

    init_vlc(&s->motion_vector_vlc, 6, 63,
        &motion_vector_vlc_table[0][1], 2, 1,
        &motion_vector_vlc_table[0][0], 2, 1, 0);

    /* work out the block mapping tables */
    s->superblock_fragments = av_malloc(s->superblock_count * 16 * sizeof(int));
    s->macroblock_coding = av_malloc(s->macroblock_count + 1);
    if (!s->superblock_fragments || !s->macroblock_coding) {
        vp3_decode_end(avctx);
        return -1;
    }
    init_block_mapping(s);

    for (i = 0; i < 3; i++) {
        s->current_frame.data[i] = NULL;
        s->last_frame.data[i] = NULL;
        s->golden_frame.data[i] = NULL;
    }

    return 0;

vlc_fail:
    av_log(avctx, AV_LOG_FATAL, "Invalid huffman table\n");
    return -1;
}

/*
 * This is the ffmpeg/libavcodec API frame decode function.
 */
static int vp3_decode_frame(AVCodecContext *avctx,
                            void *data, int *data_size,
                            AVPacket *avpkt)
{
    const uint8_t *buf = avpkt->data;
    int buf_size = avpkt->size;
    Vp3DecodeContext *s = avctx->priv_data;
    GetBitContext gb;
    static int counter = 0;
    int i;

    init_get_bits(&gb, buf, buf_size * 8);

    if (s->theora && get_bits1(&gb))
    {
        av_log(avctx, AV_LOG_ERROR, "Header packet passed to frame decoder, skipping\n");
        return -1;
    }

    s->keyframe = !get_bits1(&gb);
    if (!s->theora)
        skip_bits(&gb, 1);
    for (i = 0; i < 3; i++)
        s->last_qps[i] = s->qps[i];

    s->nqps=0;
    do{
        s->qps[s->nqps++]= get_bits(&gb, 6);
    } while(s->theora >= 0x030200 && s->nqps<3 && get_bits1(&gb));
    for (i = s->nqps; i < 3; i++)
        s->qps[i] = -1;

    if (s->avctx->debug & FF_DEBUG_PICT_INFO)
        av_log(s->avctx, AV_LOG_INFO, " VP3 %sframe #%d: Q index = %d\n",
            s->keyframe?"key":"", counter, s->qps[0]);
    counter++;

    if (s->qps[0] != s->last_qps[0])
        init_loop_filter(s);

    for (i = 0; i < s->nqps; i++)
        // reinit all dequantizers if the first one changed, because
        // the DC of the first quantizer must be used for all matrices
        if (s->qps[i] != s->last_qps[i] || s->qps[0] != s->last_qps[0])
            init_dequantizer(s, i);

    if (avctx->skip_frame >= AVDISCARD_NONKEY && !s->keyframe)
        return buf_size;

    if (s->keyframe) {
        if (!s->theora)
        {
            skip_bits(&gb, 4); /* width code */
            skip_bits(&gb, 4); /* height code */
            if (s->version)
            {
                s->version = get_bits(&gb, 5);
                if (counter == 1)
                    av_log(s->avctx, AV_LOG_DEBUG, "VP version: %d\n", s->version);
            }
        }
        if (s->version || s->theora)
        {
                if (get_bits1(&gb))
                    av_log(s->avctx, AV_LOG_ERROR, "Warning, unsupported keyframe coding type?!\n");
            skip_bits(&gb, 2); /* reserved? */
        }

        if (s->last_frame.data[0] == s->golden_frame.data[0]) {
            if (s->golden_frame.data[0])
                avctx->release_buffer(avctx, &s->golden_frame);
            s->last_frame= s->golden_frame; /* ensure that we catch any access to this released frame */
        } else {
            if (s->golden_frame.data[0])
                avctx->release_buffer(avctx, &s->golden_frame);
            if (s->last_frame.data[0])
                avctx->release_buffer(avctx, &s->last_frame);
        }

        s->golden_frame.reference = 3;
        if(avctx->get_buffer(avctx, &s->golden_frame) < 0) {
            av_log(s->avctx, AV_LOG_ERROR, "vp3: get_buffer() failed\n");
            return -1;
        }

        /* golden frame is also the current frame */
        s->current_frame= s->golden_frame;
    } else {
        /* allocate a new current frame */
        s->current_frame.reference = 3;
        if (!s->golden_frame.data[0]) {
            av_log(s->avctx, AV_LOG_ERROR, "vp3: first frame not a keyframe\n");
            return -1;
        }
        if(avctx->get_buffer(avctx, &s->current_frame) < 0) {
            av_log(s->avctx, AV_LOG_ERROR, "vp3: get_buffer() failed\n");
            return -1;
        }
    }

    s->current_frame.qscale_table= s->qscale_table; //FIXME allocate individual tables per AVFrame
    s->current_frame.qstride= 0;

    init_frame(s, &gb);

    if (unpack_superblocks(s, &gb)){
        av_log(s->avctx, AV_LOG_ERROR, "error in unpack_superblocks\n");
        return -1;
    }
    if (unpack_modes(s, &gb)){
        av_log(s->avctx, AV_LOG_ERROR, "error in unpack_modes\n");
        return -1;
    }
    if (unpack_vectors(s, &gb)){
        av_log(s->avctx, AV_LOG_ERROR, "error in unpack_vectors\n");
        return -1;
    }
    if (unpack_block_qpis(s, &gb)){
        av_log(s->avctx, AV_LOG_ERROR, "error in unpack_block_qpis\n");
        return -1;
    }
    if (unpack_dct_coeffs(s, &gb)){
        av_log(s->avctx, AV_LOG_ERROR, "error in unpack_dct_coeffs\n");
        return -1;
    }

    for (i = 0; i < 3; i++) {
        if (s->flipped_image)
            s->data_offset[i] = 0;
        else
            s->data_offset[i] = ((s->height>>!!i)-1) * s->current_frame.linesize[i];
    }

    s->last_slice_end = 0;
    for (i = 0; i < s->c_superblock_height; i++)
        render_slice(s, i);

    // filter the last row
    for (i = 0; i < 3; i++) {
        int row = (s->height >> (3+!!i)) - 1;
        apply_loop_filter(s, i, row, row+1);
    }
    vp3_draw_horiz_band(s, s->height);

    *data_size=sizeof(AVFrame);
    *(AVFrame*)data= s->current_frame;

    /* release the last frame, if it is allocated and if it is not the
     * golden frame */
    if ((s->last_frame.data[0]) &&
        (s->last_frame.data[0] != s->golden_frame.data[0]))
        avctx->release_buffer(avctx, &s->last_frame);

    /* shuffle frames (last = current) */
    s->last_frame= s->current_frame;
    s->current_frame.data[0]= NULL; /* ensure that we catch any access to this released frame */

    return buf_size;
}

/*
 * This is the ffmpeg/libavcodec API module cleanup function.
 */
static av_cold int vp3_decode_end(AVCodecContext *avctx)
{
    Vp3DecodeContext *s = avctx->priv_data;
    int i;

    av_free(s->superblock_coding);
    av_free(s->all_fragments);
    av_free(s->coded_fragment_list[0]);
    av_free(s->dct_tokens_base);
    av_free(s->superblock_fragments);
    av_free(s->macroblock_coding);

    for (i = 0; i < 16; i++) {
        free_vlc(&s->dc_vlc[i]);
        free_vlc(&s->ac_vlc_1[i]);
        free_vlc(&s->ac_vlc_2[i]);
        free_vlc(&s->ac_vlc_3[i]);
        free_vlc(&s->ac_vlc_4[i]);
    }

    free_vlc(&s->superblock_run_length_vlc);
    free_vlc(&s->fragment_run_length_vlc);
    free_vlc(&s->mode_code_vlc);
    free_vlc(&s->motion_vector_vlc);

    /* release all frames */
    if (s->golden_frame.data[0] && s->golden_frame.data[0] != s->last_frame.data[0])
        avctx->release_buffer(avctx, &s->golden_frame);
    if (s->last_frame.data[0])
        avctx->release_buffer(avctx, &s->last_frame);
    /* no need to release the current_frame since it will always be pointing
     * to the same frame as either the golden or last frame */

    return 0;
}

static int read_huffman_tree(AVCodecContext *avctx, GetBitContext *gb)
{
    Vp3DecodeContext *s = avctx->priv_data;

    if (get_bits1(gb)) {
        int token;
        if (s->entries >= 32) { /* overflow */
            av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
            return -1;
        }
        token = get_bits(gb, 5);
        //av_log(avctx, AV_LOG_DEBUG, "hti %d hbits %x token %d entry : %d size %d\n", s->hti, s->hbits, token, s->entries, s->huff_code_size);
        s->huffman_table[s->hti][token][0] = s->hbits;
        s->huffman_table[s->hti][token][1] = s->huff_code_size;
        s->entries++;
    }
    else {
        if (s->huff_code_size >= 32) {/* overflow */
            av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
            return -1;
        }
        s->huff_code_size++;
        s->hbits <<= 1;
        if (read_huffman_tree(avctx, gb))
            return -1;
        s->hbits |= 1;
        if (read_huffman_tree(avctx, gb))
            return -1;
        s->hbits >>= 1;
        s->huff_code_size--;
    }
    return 0;
}

#if CONFIG_THEORA_DECODER
static int theora_decode_header(AVCodecContext *avctx, GetBitContext *gb)
{
    Vp3DecodeContext *s = avctx->priv_data;
    int visible_width, visible_height, colorspace;

    s->theora = get_bits_long(gb, 24);
    av_log(avctx, AV_LOG_DEBUG, "Theora bitstream version %X\n", s->theora);

    /* 3.2.0 aka alpha3 has the same frame orientation as original vp3 */
    /* but previous versions have the image flipped relative to vp3 */
    if (s->theora < 0x030200)
    {
        s->flipped_image = 1;
        av_log(avctx, AV_LOG_DEBUG, "Old (<alpha3) Theora bitstream, flipped image\n");
    }

    visible_width  = s->width  = get_bits(gb, 16) << 4;
    visible_height = s->height = get_bits(gb, 16) << 4;

    if(avcodec_check_dimensions(avctx, s->width, s->height)){
        av_log(avctx, AV_LOG_ERROR, "Invalid dimensions (%dx%d)\n", s->width, s->height);
        s->width= s->height= 0;
        return -1;
    }

    if (s->theora >= 0x030200) {
        visible_width  = get_bits_long(gb, 24);
        visible_height = get_bits_long(gb, 24);

        skip_bits(gb, 8); /* offset x */
        skip_bits(gb, 8); /* offset y */
    }

    skip_bits(gb, 32); /* fps numerator */
    skip_bits(gb, 32); /* fps denumerator */
    skip_bits(gb, 24); /* aspect numerator */
    skip_bits(gb, 24); /* aspect denumerator */

    if (s->theora < 0x030200)
        skip_bits(gb, 5); /* keyframe frequency force */
    colorspace = get_bits(gb, 8);
    skip_bits(gb, 24); /* bitrate */

    skip_bits(gb, 6); /* quality hint */

    if (s->theora >= 0x030200)
    {
        skip_bits(gb, 5); /* keyframe frequency force */
        skip_bits(gb, 2); /* pixel format: 420,res,422,444 */
        skip_bits(gb, 3); /* reserved */
    }

//    align_get_bits(gb);

    if (   visible_width  <= s->width  && visible_width  > s->width-16
        && visible_height <= s->height && visible_height > s->height-16)
        avcodec_set_dimensions(avctx, visible_width, visible_height);
    else
        avcodec_set_dimensions(avctx, s->width, s->height);

    if (colorspace == 1) {
        avctx->color_primaries = AVCOL_PRI_BT470M;
    } else if (colorspace == 2) {
        avctx->color_primaries = AVCOL_PRI_BT470BG;
    }
    if (colorspace == 1 || colorspace == 2) {
        avctx->colorspace = AVCOL_SPC_BT470BG;
        avctx->color_trc  = AVCOL_TRC_BT709;
    }

    return 0;
}

static int theora_decode_tables(AVCodecContext *avctx, GetBitContext *gb)
{
    Vp3DecodeContext *s = avctx->priv_data;
    int i, n, matrices, inter, plane;

    if (s->theora >= 0x030200) {
        n = get_bits(gb, 3);
        /* loop filter limit values table */
        for (i = 0; i < 64; i++) {
            s->filter_limit_values[i] = get_bits(gb, n);
            if (s->filter_limit_values[i] > 127) {
                av_log(avctx, AV_LOG_ERROR, "filter limit value too large (%i > 127), clamping\n", s->filter_limit_values[i]);
                s->filter_limit_values[i] = 127;
            }
        }
    }

    if (s->theora >= 0x030200)
        n = get_bits(gb, 4) + 1;
    else
        n = 16;
    /* quality threshold table */
    for (i = 0; i < 64; i++)
        s->coded_ac_scale_factor[i] = get_bits(gb, n);

    if (s->theora >= 0x030200)
        n = get_bits(gb, 4) + 1;
    else
        n = 16;
    /* dc scale factor table */
    for (i = 0; i < 64; i++)
        s->coded_dc_scale_factor[i] = get_bits(gb, n);

    if (s->theora >= 0x030200)
        matrices = get_bits(gb, 9) + 1;
    else
        matrices = 3;

    if(matrices > 384){
        av_log(avctx, AV_LOG_ERROR, "invalid number of base matrixes\n");
        return -1;
    }

    for(n=0; n<matrices; n++){
        for (i = 0; i < 64; i++)
            s->base_matrix[n][i]= get_bits(gb, 8);
    }

    for (inter = 0; inter <= 1; inter++) {
        for (plane = 0; plane <= 2; plane++) {
            int newqr= 1;
            if (inter || plane > 0)
                newqr = get_bits1(gb);
            if (!newqr) {
                int qtj, plj;
                if(inter && get_bits1(gb)){
                    qtj = 0;
                    plj = plane;
                }else{
                    qtj= (3*inter + plane - 1) / 3;
                    plj= (plane + 2) % 3;
                }
                s->qr_count[inter][plane]= s->qr_count[qtj][plj];
                memcpy(s->qr_size[inter][plane], s->qr_size[qtj][plj], sizeof(s->qr_size[0][0]));
                memcpy(s->qr_base[inter][plane], s->qr_base[qtj][plj], sizeof(s->qr_base[0][0]));
            } else {
                int qri= 0;
                int qi = 0;

                for(;;){
                    i= get_bits(gb, av_log2(matrices-1)+1);
                    if(i>= matrices){
                        av_log(avctx, AV_LOG_ERROR, "invalid base matrix index\n");
                        return -1;
                    }
                    s->qr_base[inter][plane][qri]= i;
                    if(qi >= 63)
                        break;
                    i = get_bits(gb, av_log2(63-qi)+1) + 1;
                    s->qr_size[inter][plane][qri++]= i;
                    qi += i;
                }

                if (qi > 63) {
                    av_log(avctx, AV_LOG_ERROR, "invalid qi %d > 63\n", qi);
                    return -1;
                }
                s->qr_count[inter][plane]= qri;
            }
        }
    }

    /* Huffman tables */
    for (s->hti = 0; s->hti < 80; s->hti++) {
        s->entries = 0;
        s->huff_code_size = 1;
        if (!get_bits1(gb)) {
            s->hbits = 0;
            if(read_huffman_tree(avctx, gb))
                return -1;
            s->hbits = 1;
            if(read_huffman_tree(avctx, gb))
                return -1;
        }
    }

    s->theora_tables = 1;

    return 0;
}

static av_cold int theora_decode_init(AVCodecContext *avctx)
{
    Vp3DecodeContext *s = avctx->priv_data;
    GetBitContext gb;
    int ptype;
    uint8_t *header_start[3];
    int header_len[3];
    int i;

    s->theora = 1;

    if (!avctx->extradata_size)
    {
        av_log(avctx, AV_LOG_ERROR, "Missing extradata!\n");
        return -1;
    }

    if (ff_split_xiph_headers(avctx->extradata, avctx->extradata_size,
                              42, header_start, header_len) < 0) {
        av_log(avctx, AV_LOG_ERROR, "Corrupt extradata\n");
        return -1;
    }

  for(i=0;i<3;i++) {
    init_get_bits(&gb, header_start[i], header_len[i] * 8);

    ptype = get_bits(&gb, 8);

     if (!(ptype & 0x80))
     {
        av_log(avctx, AV_LOG_ERROR, "Invalid extradata!\n");
//        return -1;
     }

    // FIXME: Check for this as well.
    skip_bits_long(&gb, 6*8); /* "theora" */

    switch(ptype)
    {
        case 0x80:
            theora_decode_header(avctx, &gb);
                break;
        case 0x81:
// FIXME: is this needed? it breaks sometimes
//            theora_decode_comments(avctx, gb);
            break;
        case 0x82:
            if (theora_decode_tables(avctx, &gb))
                return -1;
            break;
        default:
            av_log(avctx, AV_LOG_ERROR, "Unknown Theora config packet: %d\n", ptype&~0x80);
            break;
    }
    if(ptype != 0x81 && 8*header_len[i] != get_bits_count(&gb))
        av_log(avctx, AV_LOG_WARNING, "%d bits left in packet %X\n", 8*header_len[i] - get_bits_count(&gb), ptype);
    if (s->theora < 0x030200)
        break;
  }

    return vp3_decode_init(avctx);
}

AVCodec theora_decoder = {
    "theora",
    CODEC_TYPE_VIDEO,
    CODEC_ID_THEORA,
    sizeof(Vp3DecodeContext),
    theora_decode_init,
    NULL,
    vp3_decode_end,
    vp3_decode_frame,
    CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
    NULL,
    .long_name = NULL_IF_CONFIG_SMALL("Theora"),
};
#endif

AVCodec vp3_decoder = {
    "vp3",
    CODEC_TYPE_VIDEO,
    CODEC_ID_VP3,
    sizeof(Vp3DecodeContext),
    vp3_decode_init,
    NULL,
    vp3_decode_end,
    vp3_decode_frame,
    CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
    NULL,
    .long_name = NULL_IF_CONFIG_SMALL("On2 VP3"),
};