export refs

[libav.git] / libavcodec / h264.c

/*
 * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
 *
 * 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 h264.c
 * H.264 / AVC / MPEG4 part10 codec.
 * @author Michael Niedermayer <michaelni@gmx.at>
 */

#include "dsputil.h"
#include "avcodec.h"
#include "mpegvideo.h"
#include "h264.h"
#include "h264data.h"
#include "h264_parser.h"
#include "golomb.h"
#include "rectangle.h"

#include "cabac.h"
#ifdef ARCH_X86
#include "x86/h264_i386.h"
#endif

//#undef NDEBUG
#include <assert.h>

/**
 * Value of Picture.reference when Picture is not a reference picture, but
 * is held for delayed output.
 */
#define DELAYED_PIC_REF 4

static VLC coeff_token_vlc[4];
static VLC_TYPE coeff_token_vlc_tables[520+332+280+256][2];
static const int coeff_token_vlc_tables_size[4]={520,332,280,256};

static VLC chroma_dc_coeff_token_vlc;
static VLC_TYPE chroma_dc_coeff_token_vlc_table[256][2];
static const int chroma_dc_coeff_token_vlc_table_size = 256;

static VLC total_zeros_vlc[15];
static VLC_TYPE total_zeros_vlc_tables[15][512][2];
static const int total_zeros_vlc_tables_size = 512;

static VLC chroma_dc_total_zeros_vlc[3];
static VLC_TYPE chroma_dc_total_zeros_vlc_tables[3][8][2];
static const int chroma_dc_total_zeros_vlc_tables_size = 8;

static VLC run_vlc[6];
static VLC_TYPE run_vlc_tables[6][8][2];
static const int run_vlc_tables_size = 8;

static VLC run7_vlc;
static VLC_TYPE run7_vlc_table[96][2];
static const int run7_vlc_table_size = 96;

static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
static Picture * remove_long(H264Context *h, int i, int ref_mask);

static av_always_inline uint32_t pack16to32(int a, int b){
#ifdef WORDS_BIGENDIAN
   return (b&0xFFFF) + (a<<16);
#else
   return (a&0xFFFF) + (b<<16);
#endif
}

static const uint8_t rem6[52]={
0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
};

static const uint8_t div6[52]={
0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
};

static const int left_block_options[4][8]={
    {0,1,2,3,7,10,8,11},
    {2,2,3,3,8,11,8,11},
    {0,0,1,1,7,10,7,10},
    {0,2,0,2,7,10,7,10}
};

#define LEVEL_TAB_BITS 8
static int8_t cavlc_level_tab[7][1<<LEVEL_TAB_BITS][2];

static void fill_caches(H264Context *h, int mb_type, int for_deblock){
    MpegEncContext * const s = &h->s;
    const int mb_xy= h->mb_xy;
    int topleft_xy, top_xy, topright_xy, left_xy[2];
    int topleft_type, top_type, topright_type, left_type[2];
    const int * left_block;
    int topleft_partition= -1;
    int i;

    top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);

    //FIXME deblocking could skip the intra and nnz parts.
    if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
        return;

    /* Wow, what a mess, why didn't they simplify the interlacing & intra
     * stuff, I can't imagine that these complex rules are worth it. */

    topleft_xy = top_xy - 1;
    topright_xy= top_xy + 1;
    left_xy[1] = left_xy[0] = mb_xy-1;
    left_block = left_block_options[0];
    if(FRAME_MBAFF){
        const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
        const int top_pair_xy      = pair_xy     - s->mb_stride;
        const int topleft_pair_xy  = top_pair_xy - 1;
        const int topright_pair_xy = top_pair_xy + 1;
        const int topleft_mb_field_flag  = IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
        const int top_mb_field_flag      = IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
        const int topright_mb_field_flag = IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
        const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
        const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
        const int bottom = (s->mb_y & 1);
        tprintf(s->avctx, "fill_caches: curr_mb_field_flag:%d, left_mb_field_flag:%d, topleft_mb_field_flag:%d, top_mb_field_flag:%d, topright_mb_field_flag:%d\n", curr_mb_field_flag, left_mb_field_flag, topleft_mb_field_flag, top_mb_field_flag, topright_mb_field_flag);

        if (curr_mb_field_flag && (bottom || top_mb_field_flag)){
            top_xy -= s->mb_stride;
        }
        if (curr_mb_field_flag && (bottom || topleft_mb_field_flag)){
            topleft_xy -= s->mb_stride;
        } else if(bottom && !curr_mb_field_flag && left_mb_field_flag) {
            topleft_xy += s->mb_stride;
            // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
            topleft_partition = 0;
        }
        if (curr_mb_field_flag && (bottom || topright_mb_field_flag)){
            topright_xy -= s->mb_stride;
        }
        if (left_mb_field_flag != curr_mb_field_flag) {
            left_xy[1] = left_xy[0] = pair_xy - 1;
            if (curr_mb_field_flag) {
                left_xy[1] += s->mb_stride;
                left_block = left_block_options[3];
            } else {
                left_block= left_block_options[2 - bottom];
            }
        }
    }

    h->top_mb_xy = top_xy;
    h->left_mb_xy[0] = left_xy[0];
    h->left_mb_xy[1] = left_xy[1];
    if(for_deblock){
        topleft_type = 0;
        topright_type = 0;
        top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
        left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
        left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;

        if(MB_MBAFF && !IS_INTRA(mb_type)){
            int list;
            for(list=0; list<h->list_count; list++){
                //These values where changed for ease of performing MC, we need to change them back
                //FIXME maybe we can make MC and loop filter use the same values or prevent
                //the MC code from changing ref_cache and rather use a temporary array.
                if(USES_LIST(mb_type,list)){
                    int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
                    *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
                    *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
                    ref += h->b8_stride;
                    *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
                    *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
                }
            }
        }
    }else{
        topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
        top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
        topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
        left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
        left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;

    if(IS_INTRA(mb_type)){
        int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
        h->topleft_samples_available=
        h->top_samples_available=
        h->left_samples_available= 0xFFFF;
        h->topright_samples_available= 0xEEEA;

        if(!(top_type & type_mask)){
            h->topleft_samples_available= 0xB3FF;
            h->top_samples_available= 0x33FF;
            h->topright_samples_available= 0x26EA;
        }
        if(IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[0])){
            if(IS_INTERLACED(mb_type)){
                if(!(left_type[0] & type_mask)){
                    h->topleft_samples_available&= 0xDFFF;
                    h->left_samples_available&= 0x5FFF;
                }
                if(!(left_type[1] & type_mask)){
                    h->topleft_samples_available&= 0xFF5F;
                    h->left_samples_available&= 0xFF5F;
                }
            }else{
                int left_typei = h->slice_table[left_xy[0] + s->mb_stride ] == h->slice_num
                                ? s->current_picture.mb_type[left_xy[0] + s->mb_stride] : 0;
                assert(left_xy[0] == left_xy[1]);
                if(!((left_typei & type_mask) && (left_type[0] & type_mask))){
                    h->topleft_samples_available&= 0xDF5F;
                    h->left_samples_available&= 0x5F5F;
                }
            }
        }else{
            if(!(left_type[0] & type_mask)){
                h->topleft_samples_available&= 0xDF5F;
                h->left_samples_available&= 0x5F5F;
            }
        }

        if(!(topleft_type & type_mask))
            h->topleft_samples_available&= 0x7FFF;

        if(!(topright_type & type_mask))
            h->topright_samples_available&= 0xFBFF;

        if(IS_INTRA4x4(mb_type)){
            if(IS_INTRA4x4(top_type)){
                h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
                h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
                h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
                h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
            }else{
                int pred;
                if(!(top_type & type_mask))
                    pred= -1;
                else{
                    pred= 2;
                }
                h->intra4x4_pred_mode_cache[4+8*0]=
                h->intra4x4_pred_mode_cache[5+8*0]=
                h->intra4x4_pred_mode_cache[6+8*0]=
                h->intra4x4_pred_mode_cache[7+8*0]= pred;
            }
            for(i=0; i<2; i++){
                if(IS_INTRA4x4(left_type[i])){
                    h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
                    h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
                }else{
                    int pred;
                    if(!(left_type[i] & type_mask))
                        pred= -1;
                    else{
                        pred= 2;
                    }
                    h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
                    h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
                }
            }
        }
    }
    }


/*
0 . T T. T T T T
1 L . .L . . . .
2 L . .L . . . .
3 . T TL . . . .
4 L . .L . . . .
5 L . .. . . . .
*/
//FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
    if(top_type){
        h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
        h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
        h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
        h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];

        h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
        h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];

        h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
        h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];

    }else{
        h->non_zero_count_cache[4+8*0]=
        h->non_zero_count_cache[5+8*0]=
        h->non_zero_count_cache[6+8*0]=
        h->non_zero_count_cache[7+8*0]=

        h->non_zero_count_cache[1+8*0]=
        h->non_zero_count_cache[2+8*0]=

        h->non_zero_count_cache[1+8*3]=
        h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;

    }

    for (i=0; i<2; i++) {
        if(left_type[i]){
            h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
            h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
            h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
            h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
        }else{
            h->non_zero_count_cache[3+8*1 + 2*8*i]=
            h->non_zero_count_cache[3+8*2 + 2*8*i]=
            h->non_zero_count_cache[0+8*1 +   8*i]=
            h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
        }
    }

    if( h->pps.cabac ) {
        // top_cbp
        if(top_type) {
            h->top_cbp = h->cbp_table[top_xy];
        } else if(IS_INTRA(mb_type)) {
            h->top_cbp = 0x1C0;
        } else {
            h->top_cbp = 0;
        }
        // left_cbp
        if (left_type[0]) {
            h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
        } else if(IS_INTRA(mb_type)) {
            h->left_cbp = 0x1C0;
        } else {
            h->left_cbp = 0;
        }
        if (left_type[0]) {
            h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
        }
        if (left_type[1]) {
            h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
        }
    }

#if 1
    if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
        int list;
        for(list=0; list<h->list_count; list++){
            if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
                /*if(!h->mv_cache_clean[list]){
                    memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
                    memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
                    h->mv_cache_clean[list]= 1;
                }*/
                continue;
            }
            h->mv_cache_clean[list]= 0;

            if(USES_LIST(top_type, list)){
                const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
                const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
                *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
                *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
                *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
                *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
                h->ref_cache[list][scan8[0] + 0 - 1*8]=
                h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
                h->ref_cache[list][scan8[0] + 2 - 1*8]=
                h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
            }else{
                *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
                *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
                *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
                *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
                *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
            }

            for(i=0; i<2; i++){
                int cache_idx = scan8[0] - 1 + i*2*8;
                if(USES_LIST(left_type[i], list)){
                    const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
                    const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
                    *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
                    *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
                    h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
                    h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
                }else{
                    *(uint32_t*)h->mv_cache [list][cache_idx  ]=
                    *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
                    h->ref_cache[list][cache_idx  ]=
                    h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
                }
            }

            if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
                continue;

            if(USES_LIST(topleft_type, list)){
                const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
                const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
                *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
                h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
            }else{
                *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
                h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
            }

            if(USES_LIST(topright_type, list)){
                const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
                const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
                *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
                h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
            }else{
                *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
                h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
            }

            if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
                continue;

            h->ref_cache[list][scan8[5 ]+1] =
            h->ref_cache[list][scan8[7 ]+1] =
            h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
            h->ref_cache[list][scan8[4 ]] =
            h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
            *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
            *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
            *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
            *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
            *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;

            if( h->pps.cabac ) {
                /* XXX beurk, Load mvd */
                if(USES_LIST(top_type, list)){
                    const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
                    *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
                    *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
                    *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
                    *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
                }else{
                    *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
                    *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
                    *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
                    *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
                }
                if(USES_LIST(left_type[0], list)){
                    const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
                }else{
                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
                }
                if(USES_LIST(left_type[1], list)){
                    const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
                }else{
                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
                }
                *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
                *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
                *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
                *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
                *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;

                if(h->slice_type_nos == FF_B_TYPE){
                    fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);

                    if(IS_DIRECT(top_type)){
                        *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
                    }else if(IS_8X8(top_type)){
                        int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
                        h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
                        h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
                    }else{
                        *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
                    }

                    if(IS_DIRECT(left_type[0]))
                        h->direct_cache[scan8[0] - 1 + 0*8]= 1;
                    else if(IS_8X8(left_type[0]))
                        h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
                    else
                        h->direct_cache[scan8[0] - 1 + 0*8]= 0;

                    if(IS_DIRECT(left_type[1]))
                        h->direct_cache[scan8[0] - 1 + 2*8]= 1;
                    else if(IS_8X8(left_type[1]))
                        h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
                    else
                        h->direct_cache[scan8[0] - 1 + 2*8]= 0;
                }
            }

            if(FRAME_MBAFF){
#define MAP_MVS\
                    MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
                    MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
                    MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
                    MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
                    MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
                    MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
                    MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
                    MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
                    MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
                    MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
                if(MB_FIELD){
#define MAP_F2F(idx, mb_type)\
                    if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
                        h->ref_cache[list][idx] <<= 1;\
                        h->mv_cache[list][idx][1] /= 2;\
                        h->mvd_cache[list][idx][1] /= 2;\
                    }
                    MAP_MVS
#undef MAP_F2F
                }else{
#define MAP_F2F(idx, mb_type)\
                    if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
                        h->ref_cache[list][idx] >>= 1;\
                        h->mv_cache[list][idx][1] <<= 1;\
                        h->mvd_cache[list][idx][1] <<= 1;\
                    }
                    MAP_MVS
#undef MAP_F2F
                }
            }
        }
    }
#endif

    h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
}

static inline void write_back_intra_pred_mode(H264Context *h){
    const int mb_xy= h->mb_xy;

    h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
    h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
    h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
    h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
    h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
    h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
    h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
}

/**
 * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 */
static inline int check_intra4x4_pred_mode(H264Context *h){
    MpegEncContext * const s = &h->s;
    static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
    static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
    int i;

    if(!(h->top_samples_available&0x8000)){
        for(i=0; i<4; i++){
            int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
            if(status<0){
                av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
                return -1;
            } else if(status){
                h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
            }
        }
    }

    if((h->left_samples_available&0x8888)!=0x8888){
        static const int mask[4]={0x8000,0x2000,0x80,0x20};
        for(i=0; i<4; i++){
            if(!(h->left_samples_available&mask[i])){
                int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
                if(status<0){
                    av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
                    return -1;
                } else if(status){
                    h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
                }
            }
        }
    }

    return 0;
} //FIXME cleanup like next

/**
 * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 */
static inline int check_intra_pred_mode(H264Context *h, int mode){
    MpegEncContext * const s = &h->s;
    static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
    static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};

    if(mode > 6U) {
        av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
        return -1;
    }

    if(!(h->top_samples_available&0x8000)){
        mode= top[ mode ];
        if(mode<0){
            av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
            return -1;
        }
    }

    if((h->left_samples_available&0x8080) != 0x8080){
        mode= left[ mode ];
        if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
            mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
        }
        if(mode<0){
            av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
            return -1;
        }
    }

    return mode;
}

/**
 * gets the predicted intra4x4 prediction mode.
 */
static inline int pred_intra_mode(H264Context *h, int n){
    const int index8= scan8[n];
    const int left= h->intra4x4_pred_mode_cache[index8 - 1];
    const int top = h->intra4x4_pred_mode_cache[index8 - 8];
    const int min= FFMIN(left, top);

    tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);

    if(min<0) return DC_PRED;
    else      return min;
}

static inline void write_back_non_zero_count(H264Context *h){
    const int mb_xy= h->mb_xy;

    h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
    h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
    h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
    h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
    h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
    h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
    h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];

    h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
    h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
    h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];

    h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
    h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
    h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
}

/**
 * gets the predicted number of non-zero coefficients.
 * @param n block index
 */
static inline int pred_non_zero_count(H264Context *h, int n){
    const int index8= scan8[n];
    const int left= h->non_zero_count_cache[index8 - 1];
    const int top = h->non_zero_count_cache[index8 - 8];
    int i= left + top;

    if(i<64) i= (i+1)>>1;

    tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);

    return i&31;
}

static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
    const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
    MpegEncContext *s = &h->s;

    /* there is no consistent mapping of mvs to neighboring locations that will
     * make mbaff happy, so we can't move all this logic to fill_caches */
    if(FRAME_MBAFF){
        const uint32_t *mb_types = s->current_picture_ptr->mb_type;
        const int16_t *mv;
        *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
        *C = h->mv_cache[list][scan8[0]-2];

        if(!MB_FIELD
           && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
            int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
            if(IS_INTERLACED(mb_types[topright_xy])){
#define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
                const int x4 = X4, y4 = Y4;\
                const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
                if(!USES_LIST(mb_type,list))\
                    return LIST_NOT_USED;\
                mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
                h->mv_cache[list][scan8[0]-2][0] = mv[0];\
                h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
                return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;

                SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
            }
        }
        if(topright_ref == PART_NOT_AVAILABLE
           && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
           && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
            if(!MB_FIELD
               && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
                SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
            }
            if(MB_FIELD
               && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
               && i >= scan8[0]+8){
                // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
                SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
            }
        }
#undef SET_DIAG_MV
    }

    if(topright_ref != PART_NOT_AVAILABLE){
        *C= h->mv_cache[list][ i - 8 + part_width ];
        return topright_ref;
    }else{
        tprintf(s->avctx, "topright MV not available\n");

        *C= h->mv_cache[list][ i - 8 - 1 ];
        return h->ref_cache[list][ i - 8 - 1 ];
    }
}

/**
 * gets the predicted MV.
 * @param n the block index
 * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
 * @param mx the x component of the predicted motion vector
 * @param my the y component of the predicted motion vector
 */
static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
    const int index8= scan8[n];
    const int top_ref=      h->ref_cache[list][ index8 - 8 ];
    const int left_ref=     h->ref_cache[list][ index8 - 1 ];
    const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
    const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
    const int16_t * C;
    int diagonal_ref, match_count;

    assert(part_width==1 || part_width==2 || part_width==4);

/* mv_cache
  B . . A T T T T
  U . . L . . , .
  U . . L . . . .
  U . . L . . , .
  . . . L . . . .
*/

    diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
    match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
    tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
    if(match_count > 1){ //most common
        *mx= mid_pred(A[0], B[0], C[0]);
        *my= mid_pred(A[1], B[1], C[1]);
    }else if(match_count==1){
        if(left_ref==ref){
            *mx= A[0];
            *my= A[1];
        }else if(top_ref==ref){
            *mx= B[0];
            *my= B[1];
        }else{
            *mx= C[0];
            *my= C[1];
        }
    }else{
        if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
            *mx= A[0];
            *my= A[1];
        }else{
            *mx= mid_pred(A[0], B[0], C[0]);
            *my= mid_pred(A[1], B[1], C[1]);
        }
    }

    tprintf(h->s.avctx, "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1],                    diagonal_ref, C[0], C[1], left_ref, A[0], A[1], ref, *mx, *my, h->s.mb_x, h->s.mb_y, n, list);
}

/**
 * gets the directionally predicted 16x8 MV.
 * @param n the block index
 * @param mx the x component of the predicted motion vector
 * @param my the y component of the predicted motion vector
 */
static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
    if(n==0){
        const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
        const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];

        tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1], h->s.mb_x, h->s.mb_y, n, list);

        if(top_ref == ref){
            *mx= B[0];
            *my= B[1];
            return;
        }
    }else{
        const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
        const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];

        tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);

        if(left_ref == ref){
            *mx= A[0];
            *my= A[1];
            return;
        }
    }

    //RARE
    pred_motion(h, n, 4, list, ref, mx, my);
}

/**
 * gets the directionally predicted 8x16 MV.
 * @param n the block index
 * @param mx the x component of the predicted motion vector
 * @param my the y component of the predicted motion vector
 */
static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
    if(n==0){
        const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
        const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];

        tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);

        if(left_ref == ref){
            *mx= A[0];
            *my= A[1];
            return;
        }
    }else{
        const int16_t * C;
        int diagonal_ref;

        diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);

        tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", diagonal_ref, C[0], C[1], h->s.mb_x, h->s.mb_y, n, list);

        if(diagonal_ref == ref){
            *mx= C[0];
            *my= C[1];
            return;
        }
    }

    //RARE
    pred_motion(h, n, 2, list, ref, mx, my);
}

static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
    const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
    const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];

    tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);

    if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
       || !( top_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ])
       || !(left_ref | *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ])){

        *mx = *my = 0;
        return;
    }

    pred_motion(h, 0, 4, 0, 0, mx, my);

    return;
}

static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
    int poc0 = h->ref_list[0][i].poc;
    int td = av_clip(poc1 - poc0, -128, 127);
    if(td == 0 || h->ref_list[0][i].long_ref){
        return 256;
    }else{
        int tb = av_clip(poc - poc0, -128, 127);
        int tx = (16384 + (FFABS(td) >> 1)) / td;
        return av_clip((tb*tx + 32) >> 6, -1024, 1023);
    }
}

static inline void direct_dist_scale_factor(H264Context * const h){
    MpegEncContext * const s = &h->s;
    const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
    const int poc1 = h->ref_list[1][0].poc;
    int i, field;
    for(field=0; field<2; field++){
        const int poc  = h->s.current_picture_ptr->field_poc[field];
        const int poc1 = h->ref_list[1][0].field_poc[field];
        for(i=0; i < 2*h->ref_count[0]; i++)
            h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
    }

    for(i=0; i<h->ref_count[0]; i++){
        h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
    }
}

static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
    MpegEncContext * const s = &h->s;
    Picture * const ref1 = &h->ref_list[1][0];
    int j, old_ref, rfield;
    int start= mbafi ? 16                      : 0;
    int end  = mbafi ? 16+2*h->ref_count[list] : h->ref_count[list];
    int interl= mbafi || s->picture_structure != PICT_FRAME;

    /* bogus; fills in for missing frames */
    memset(map[list], 0, sizeof(map[list]));

    for(rfield=0; rfield<2; rfield++){
        for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
            int poc = ref1->ref_poc[colfield][list][old_ref];

            if     (!interl)
                poc |= 3;
            else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
                poc= (poc&~3) + rfield + 1;

            for(j=start; j<end; j++){
                if(4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3) == poc){
                    int cur_ref= mbafi ? (j-16)^field : j;
                    map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
                    if(rfield == field)
                        map[list][old_ref] = cur_ref;
                    break;
                }
            }
        }
    }
}

static inline void direct_ref_list_init(H264Context * const h){
    MpegEncContext * const s = &h->s;
    Picture * const ref1 = &h->ref_list[1][0];
    Picture * const cur = s->current_picture_ptr;
    int list, j, field;
    int sidx= (s->picture_structure&1)^1;
    int ref1sidx= (ref1->reference&1)^1;

    for(list=0; list<2; list++){
        cur->ref_count[sidx][list] = h->ref_count[list];
        for(j=0; j<h->ref_count[list]; j++)
            cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
    }

    if(s->picture_structure == PICT_FRAME){
        memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
        memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
    }

    cur->mbaff= FRAME_MBAFF;

    if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
        return;

    for(list=0; list<2; list++){
        fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
        for(field=0; field<2; field++)
            fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
    }
}

static inline void pred_direct_motion(H264Context * const h, int *mb_type){
    MpegEncContext * const s = &h->s;
    int b8_stride = h->b8_stride;
    int b4_stride = h->b_stride;
    int mb_xy = h->mb_xy;
    int mb_type_col[2];
    const int16_t (*l1mv0)[2], (*l1mv1)[2];
    const int8_t *l1ref0, *l1ref1;
    const int is_b8x8 = IS_8X8(*mb_type);
    unsigned int sub_mb_type;
    int i8, i4;

#define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)

    if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
        if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
            int cur_poc = s->current_picture_ptr->poc;
            int *col_poc = h->ref_list[1]->field_poc;
            int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
            mb_xy= s->mb_x + ((s->mb_y&~1) + col_parity)*s->mb_stride;
            b8_stride = 0;
        }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){// FL -> FL & differ parity
            int fieldoff= 2*(h->ref_list[1][0].reference)-3;
            mb_xy += s->mb_stride*fieldoff;
        }
        goto single_col;
    }else{                                               // AFL/AFR/FR/FL -> AFR/FR
        if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
            mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
            mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
            mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
            b8_stride *= 3;
            b4_stride *= 6;
            //FIXME IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag
            if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
                && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
                && !is_b8x8){
                sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
                *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
            }else{
                sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
                *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
            }
        }else{                                           //     AFR/FR    -> AFR/FR
single_col:
            mb_type_col[0] =
            mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
            if(IS_8X8(mb_type_col[0]) && !h->sps.direct_8x8_inference_flag){
                /* FIXME save sub mb types from previous frames (or derive from MVs)
                * so we know exactly what block size to use */
                sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
                *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
            }else if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
                sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
                *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
            }else{
                sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
                *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
            }
        }
    }

    l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
    l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
    l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
    l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
    if(!b8_stride){
        if(s->mb_y&1){
            l1ref0 += h->b8_stride;
            l1ref1 += h->b8_stride;
            l1mv0  +=  2*b4_stride;
            l1mv1  +=  2*b4_stride;
        }
    }

    if(h->direct_spatial_mv_pred){
        int ref[2];
        int mv[2][2];
        int list;

        /* FIXME interlacing + spatial direct uses wrong colocated block positions */

        /* ref = min(neighbors) */
        for(list=0; list<2; list++){
            int refa = h->ref_cache[list][scan8[0] - 1];
            int refb = h->ref_cache[list][scan8[0] - 8];
            int refc = h->ref_cache[list][scan8[0] - 8 + 4];
            if(refc == PART_NOT_AVAILABLE)
                refc = h->ref_cache[list][scan8[0] - 8 - 1];
            ref[list] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
            if(ref[list] < 0)
                ref[list] = -1;
        }

        if(ref[0] < 0 && ref[1] < 0){
            ref[0] = ref[1] = 0;
            mv[0][0] = mv[0][1] =
            mv[1][0] = mv[1][1] = 0;
        }else{
            for(list=0; list<2; list++){
                if(ref[list] >= 0)
                    pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
                else
                    mv[list][0] = mv[list][1] = 0;
            }
        }

        if(ref[1] < 0){
            if(!is_b8x8)
                *mb_type &= ~MB_TYPE_L1;
            sub_mb_type &= ~MB_TYPE_L1;
        }else if(ref[0] < 0){
            if(!is_b8x8)
                *mb_type &= ~MB_TYPE_L0;
            sub_mb_type &= ~MB_TYPE_L0;
        }

        if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
            for(i8=0; i8<4; i8++){
                int x8 = i8&1;
                int y8 = i8>>1;
                int xy8 = x8+y8*b8_stride;
                int xy4 = 3*x8+y8*b4_stride;
                int a=0, b=0;

                if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
                    continue;
                h->sub_mb_type[i8] = sub_mb_type;

                fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
                fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
                if(!IS_INTRA(mb_type_col[y8])
                   && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
                       || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
                    if(ref[0] > 0)
                        a= pack16to32(mv[0][0],mv[0][1]);
                    if(ref[1] > 0)
                        b= pack16to32(mv[1][0],mv[1][1]);
                }else{
                    a= pack16to32(mv[0][0],mv[0][1]);
                    b= pack16to32(mv[1][0],mv[1][1]);
                }
                fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
                fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
            }
        }else if(IS_16X16(*mb_type)){
            int a=0, b=0;

            fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
            fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
            if(!IS_INTRA(mb_type_col[0])
               && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
                   || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
                       && (h->x264_build>33 || !h->x264_build)))){
                if(ref[0] > 0)
                    a= pack16to32(mv[0][0],mv[0][1]);
                if(ref[1] > 0)
                    b= pack16to32(mv[1][0],mv[1][1]);
            }else{
                a= pack16to32(mv[0][0],mv[0][1]);
                b= pack16to32(mv[1][0],mv[1][1]);
            }
            fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
            fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
        }else{
            for(i8=0; i8<4; i8++){
                const int x8 = i8&1;
                const int y8 = i8>>1;

                if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
                    continue;
                h->sub_mb_type[i8] = sub_mb_type;

                fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
                fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
                fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
                fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);

                /* col_zero_flag */
                if(!IS_INTRA(mb_type_col[0]) && (   l1ref0[x8 + y8*b8_stride] == 0
                                              || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
                                                  && (h->x264_build>33 || !h->x264_build)))){
                    const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
                    if(IS_SUB_8X8(sub_mb_type)){
                        const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
                        if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
                            if(ref[0] == 0)
                                fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
                            if(ref[1] == 0)
                                fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
                        }
                    }else
                    for(i4=0; i4<4; i4++){
                        const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
                        if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
                            if(ref[0] == 0)
                                *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
                            if(ref[1] == 0)
                                *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
                        }
                    }
                }
            }
        }
    }else{ /* direct temporal mv pred */
        const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
        const int *dist_scale_factor = h->dist_scale_factor;
        int ref_offset= 0;

        if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
            map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
            map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
            dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
        }
        if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0]))
            ref_offset += 16;

        if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
            /* FIXME assumes direct_8x8_inference == 1 */
            int y_shift  = 2*!IS_INTERLACED(*mb_type);

            for(i8=0; i8<4; i8++){
                const int x8 = i8&1;
                const int y8 = i8>>1;
                int ref0, scale;
                const int16_t (*l1mv)[2]= l1mv0;

                if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
                    continue;
                h->sub_mb_type[i8] = sub_mb_type;

                fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
                if(IS_INTRA(mb_type_col[y8])){
                    fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
                    fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
                    fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
                    continue;
                }

                ref0 = l1ref0[x8 + y8*b8_stride];
                if(ref0 >= 0)
                    ref0 = map_col_to_list0[0][ref0 + ref_offset];
                else{
                    ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
                    l1mv= l1mv1;
                }
                scale = dist_scale_factor[ref0];
                fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);

                {
                    const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
                    int my_col = (mv_col[1]<<y_shift)/2;
                    int mx = (scale * mv_col[0] + 128) >> 8;
                    int my = (scale * my_col + 128) >> 8;
                    fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
                    fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
                }
            }
            return;
        }

        /* one-to-one mv scaling */

        if(IS_16X16(*mb_type)){
            int ref, mv0, mv1;

            fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
            if(IS_INTRA(mb_type_col[0])){
                ref=mv0=mv1=0;
            }else{
                const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
                                                : map_col_to_list0[1][l1ref1[0] + ref_offset];
                const int scale = dist_scale_factor[ref0];
                const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
                int mv_l0[2];
                mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
                mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
                ref= ref0;
                mv0= pack16to32(mv_l0[0],mv_l0[1]);
                mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
            }
            fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
            fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
            fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
        }else{
            for(i8=0; i8<4; i8++){
                const int x8 = i8&1;
                const int y8 = i8>>1;
                int ref0, scale;
                const int16_t (*l1mv)[2]= l1mv0;

                if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
                    continue;
                h->sub_mb_type[i8] = sub_mb_type;
                fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
                if(IS_INTRA(mb_type_col[0])){
                    fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
                    fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
                    fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
                    continue;
                }

                ref0 = l1ref0[x8 + y8*b8_stride] + ref_offset;
                if(ref0 >= 0)
                    ref0 = map_col_to_list0[0][ref0];
                else{
                    ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
                    l1mv= l1mv1;
                }
                scale = dist_scale_factor[ref0];

                fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
                if(IS_SUB_8X8(sub_mb_type)){
                    const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
                    int mx = (scale * mv_col[0] + 128) >> 8;
                    int my = (scale * mv_col[1] + 128) >> 8;
                    fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
                    fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
                }else
                for(i4=0; i4<4; i4++){
                    const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
                    int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
                    mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
                    mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
                    *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
                        pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
                }
            }
        }
    }
}

static inline void write_back_motion(H264Context *h, int mb_type){
    MpegEncContext * const s = &h->s;
    const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
    const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
    int list;

    if(!USES_LIST(mb_type, 0))
        fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);

    for(list=0; list<h->list_count; list++){
        int y;
        if(!USES_LIST(mb_type, list))
            continue;

        for(y=0; y<4; y++){
            *(uint64_t*)s->current_picture.motion_val[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mv_cache[list][scan8[0]+0 + 8*y];
            *(uint64_t*)s->current_picture.motion_val[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mv_cache[list][scan8[0]+2 + 8*y];
        }
        if( h->pps.cabac ) {
            if(IS_SKIP(mb_type))
                fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
            else
            for(y=0; y<4; y++){
                *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
                *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
            }
        }

        {
            int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
            ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
            ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
            ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
            ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
        }
    }

    if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
        if(IS_8X8(mb_type)){
            uint8_t *direct_table = &h->direct_table[b8_xy];
            direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
            direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
            direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
        }
    }
}

/**
 * Decodes a network abstraction layer unit.
 * @param consumed is the number of bytes used as input
 * @param length is the length of the array
 * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
 * @returns decoded bytes, might be src+1 if no escapes
 */
static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
    int i, si, di;
    uint8_t *dst;
    int bufidx;

//    src[0]&0x80;                //forbidden bit
    h->nal_ref_idc= src[0]>>5;
    h->nal_unit_type= src[0]&0x1F;

    src++; length--;
#if 0
    for(i=0; i<length; i++)
        printf("%2X ", src[i]);
#endif

#ifdef HAVE_FAST_UNALIGNED
# ifdef HAVE_FAST_64BIT
#   define RS 7
    for(i=0; i+1<length; i+=9){
        if(!((~*(uint64_t*)(src+i) & (*(uint64_t*)(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
# else
#   define RS 3
    for(i=0; i+1<length; i+=5){
        if(!((~*(uint32_t*)(src+i) & (*(uint32_t*)(src+i) - 0x01000101U)) & 0x80008080U))
# endif
            continue;
        if(i>0 && !src[i]) i--;
        while(src[i]) i++;
#else
#   define RS 0
    for(i=0; i+1<length; i+=2){
        if(src[i]) continue;
        if(i>0 && src[i-1]==0) i--;
#endif
        if(i+2<length && src[i+1]==0 && src[i+2]<=3){
            if(src[i+2]!=3){
                /* startcode, so we must be past the end */
                length=i;
            }
            break;
        }
        i-= RS;
    }

    if(i>=length-1){ //no escaped 0
        *dst_length= length;
        *consumed= length+1; //+1 for the header
        return src;
    }

    bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
    h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
    dst= h->rbsp_buffer[bufidx];

    if (dst == NULL){
        return NULL;
    }

//printf("decoding esc\n");
    memcpy(dst, src, i);
    si=di=i;
    while(si+2<length){
        //remove escapes (very rare 1:2^22)
        if(src[si+2]>3){
            dst[di++]= src[si++];
            dst[di++]= src[si++];
        }else if(src[si]==0 && src[si+1]==0){
            if(src[si+2]==3){ //escape
                dst[di++]= 0;
                dst[di++]= 0;
                si+=3;
                continue;
            }else //next start code
                goto nsc;
        }

        dst[di++]= src[si++];
    }
    while(si<length)
        dst[di++]= src[si++];
nsc:

    memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);

    *dst_length= di;
    *consumed= si + 1;//+1 for the header
//FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
    return dst;
}

/**
 * identifies the exact end of the bitstream
 * @return the length of the trailing, or 0 if damaged
 */
static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
    int v= *src;
    int r;

    tprintf(h->s.avctx, "rbsp trailing %X\n", v);

    for(r=1; r<9; r++){
        if(v&1) return r;
        v>>=1;
    }
    return 0;
}

/**
 * IDCT transforms the 16 dc values and dequantizes them.
 * @param qp quantization parameter
 */
static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
#define stride 16
    int i;
    int temp[16]; //FIXME check if this is a good idea
    static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
    static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};

//memset(block, 64, 2*256);
//return;
    for(i=0; i<4; i++){
        const int offset= y_offset[i];
        const int z0= block[offset+stride*0] + block[offset+stride*4];
        const int z1= block[offset+stride*0] - block[offset+stride*4];
        const int z2= block[offset+stride*1] - block[offset+stride*5];
        const int z3= block[offset+stride*1] + block[offset+stride*5];

        temp[4*i+0]= z0+z3;
        temp[4*i+1]= z1+z2;
        temp[4*i+2]= z1-z2;
        temp[4*i+3]= z0-z3;
    }

    for(i=0; i<4; i++){
        const int offset= x_offset[i];
        const int z0= temp[4*0+i] + temp[4*2+i];
        const int z1= temp[4*0+i] - temp[4*2+i];
        const int z2= temp[4*1+i] - temp[4*3+i];
        const int z3= temp[4*1+i] + temp[4*3+i];

        block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
        block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
        block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
        block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
    }
}

#if 0
/**
 * DCT transforms the 16 dc values.
 * @param qp quantization parameter ??? FIXME
 */
static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
//    const int qmul= dequant_coeff[qp][0];
    int i;
    int temp[16]; //FIXME check if this is a good idea
    static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
    static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};

    for(i=0; i<4; i++){
        const int offset= y_offset[i];
        const int z0= block[offset+stride*0] + block[offset+stride*4];
        const int z1= block[offset+stride*0] - block[offset+stride*4];
        const int z2= block[offset+stride*1] - block[offset+stride*5];
        const int z3= block[offset+stride*1] + block[offset+stride*5];

        temp[4*i+0]= z0+z3;
        temp[4*i+1]= z1+z2;
        temp[4*i+2]= z1-z2;
        temp[4*i+3]= z0-z3;
    }

    for(i=0; i<4; i++){
        const int offset= x_offset[i];
        const int z0= temp[4*0+i] + temp[4*2+i];
        const int z1= temp[4*0+i] - temp[4*2+i];
        const int z2= temp[4*1+i] - temp[4*3+i];
        const int z3= temp[4*1+i] + temp[4*3+i];

        block[stride*0 +offset]= (z0 + z3)>>1;
        block[stride*2 +offset]= (z1 + z2)>>1;
        block[stride*8 +offset]= (z1 - z2)>>1;
        block[stride*10+offset]= (z0 - z3)>>1;
    }
}
#endif

#undef xStride
#undef stride

static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
    const int stride= 16*2;
    const int xStride= 16;
    int a,b,c,d,e;

    a= block[stride*0 + xStride*0];
    b= block[stride*0 + xStride*1];
    c= block[stride*1 + xStride*0];
    d= block[stride*1 + xStride*1];

    e= a-b;
    a= a+b;
    b= c-d;
    c= c+d;

    block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
    block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
    block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
    block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
}

#if 0
static void chroma_dc_dct_c(DCTELEM *block){
    const int stride= 16*2;
    const int xStride= 16;
    int a,b,c,d,e;

    a= block[stride*0 + xStride*0];
    b= block[stride*0 + xStride*1];
    c= block[stride*1 + xStride*0];
    d= block[stride*1 + xStride*1];

    e= a-b;
    a= a+b;
    b= c-d;
    c= c+d;

    block[stride*0 + xStride*0]= (a+c);
    block[stride*0 + xStride*1]= (e+b);
    block[stride*1 + xStride*0]= (a-c);
    block[stride*1 + xStride*1]= (e-b);
}
#endif

/**
 * gets the chroma qp.
 */
static inline int get_chroma_qp(H264Context *h, int t, int qscale){
    return h->pps.chroma_qp_table[t][qscale];
}

static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int src_x_offset, int src_y_offset,
                           qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
    MpegEncContext * const s = &h->s;
    const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
    int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
    const int luma_xy= (mx&3) + ((my&3)<<2);
    uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
    uint8_t * src_cb, * src_cr;
    int extra_width= h->emu_edge_width;
    int extra_height= h->emu_edge_height;
    int emu=0;
    const int full_mx= mx>>2;
    const int full_my= my>>2;
    const int pic_width  = 16*s->mb_width;
    const int pic_height = 16*s->mb_height >> MB_FIELD;

    if(mx&7) extra_width -= 3;
    if(my&7) extra_height -= 3;

    if(   full_mx < 0-extra_width
       || full_my < 0-extra_height
       || full_mx + 16/*FIXME*/ > pic_width + extra_width
       || full_my + 16/*FIXME*/ > pic_height + extra_height){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
            src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
        emu=1;
    }

    qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
    if(!square){
        qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
    }

    if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;

    if(MB_FIELD){
        // chroma offset when predicting from a field of opposite parity
        my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
        emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
    }
    src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
    src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;

    if(emu){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
            src_cb= s->edge_emu_buffer;
    }
    chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);

    if(emu){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
            src_cr= s->edge_emu_buffer;
    }
    chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
}

static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int x_offset, int y_offset,
                           qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
                           qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
                           int list0, int list1){
    MpegEncContext * const s = &h->s;
    qpel_mc_func *qpix_op=  qpix_put;
    h264_chroma_mc_func chroma_op= chroma_put;

    dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
    dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
    dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
    x_offset += 8*s->mb_x;
    y_offset += 8*(s->mb_y >> MB_FIELD);

    if(list0){
        Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
        mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
                           dest_y, dest_cb, dest_cr, x_offset, y_offset,
                           qpix_op, chroma_op);

        qpix_op=  qpix_avg;
        chroma_op= chroma_avg;
    }

    if(list1){
        Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
        mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
                           dest_y, dest_cb, dest_cr, x_offset, y_offset,
                           qpix_op, chroma_op);
    }
}

static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int x_offset, int y_offset,
                           qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
                           h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
                           h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
                           int list0, int list1){
    MpegEncContext * const s = &h->s;

    dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
    dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
    dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
    x_offset += 8*s->mb_x;
    y_offset += 8*(s->mb_y >> MB_FIELD);

    if(list0 && list1){
        /* don't optimize for luma-only case, since B-frames usually
         * use implicit weights => chroma too. */
        uint8_t *tmp_cb = s->obmc_scratchpad;
        uint8_t *tmp_cr = s->obmc_scratchpad + 8;
        uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
        int refn0 = h->ref_cache[0][ scan8[n] ];
        int refn1 = h->ref_cache[1][ scan8[n] ];

        mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
                    dest_y, dest_cb, dest_cr,
                    x_offset, y_offset, qpix_put, chroma_put);
        mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
                    tmp_y, tmp_cb, tmp_cr,
                    x_offset, y_offset, qpix_put, chroma_put);

        if(h->use_weight == 2){
            int weight0 = h->implicit_weight[refn0][refn1];
            int weight1 = 64 - weight0;
            luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
            chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
            chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
        }else{
            luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
                            h->luma_weight[0][refn0], h->luma_weight[1][refn1],
                            h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
            chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                            h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
                            h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
            chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                            h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
                            h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
        }
    }else{
        int list = list1 ? 1 : 0;
        int refn = h->ref_cache[list][ scan8[n] ];
        Picture *ref= &h->ref_list[list][refn];
        mc_dir_part(h, ref, n, square, chroma_height, delta, list,
                    dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put, chroma_put);

        luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
                       h->luma_weight[list][refn], h->luma_offset[list][refn]);
        if(h->use_weight_chroma){
            chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                             h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
            chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
                             h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
        }
    }
}

static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
                           uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                           int x_offset, int y_offset,
                           qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
                           qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
                           h264_weight_func *weight_op, h264_biweight_func *weight_avg,
                           int list0, int list1){
    if((h->use_weight==2 && list0 && list1
        && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
       || h->use_weight==1)
        mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
                         x_offset, y_offset, qpix_put, chroma_put,
                         weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
    else
        mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
                    x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
}

static inline void prefetch_motion(H264Context *h, int list){
    /* fetch pixels for estimated mv 4 macroblocks ahead
     * optimized for 64byte cache lines */
    MpegEncContext * const s = &h->s;
    const int refn = h->ref_cache[list][scan8[0]];
    if(refn >= 0){
        const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
        const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
        uint8_t **src= h->ref_list[list][refn].data;
        int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
        s->dsp.prefetch(src[0]+off, s->linesize, 4);
        off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
        s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
    }
}

static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
                      qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
                      qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
                      h264_weight_func *weight_op, h264_biweight_func *weight_avg){
    MpegEncContext * const s = &h->s;
    const int mb_xy= h->mb_xy;
    const int mb_type= s->current_picture.mb_type[mb_xy];

    assert(IS_INTER(mb_type));

    prefetch_motion(h, 0);

    if(IS_16X16(mb_type)){
        mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
                qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
                &weight_op[0], &weight_avg[0],
                IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
    }else if(IS_16X8(mb_type)){
        mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
                qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
                &weight_op[1], &weight_avg[1],
                IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
        mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
                qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
                &weight_op[1], &weight_avg[1],
                IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
    }else if(IS_8X16(mb_type)){
        mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
                qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
                &weight_op[2], &weight_avg[2],
                IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
        mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
                qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
                &weight_op[2], &weight_avg[2],
                IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
    }else{
        int i;

        assert(IS_8X8(mb_type));

        for(i=0; i<4; i++){
            const int sub_mb_type= h->sub_mb_type[i];
            const int n= 4*i;
            int x_offset= (i&1)<<2;
            int y_offset= (i&2)<<1;

            if(IS_SUB_8X8(sub_mb_type)){
                mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
                    &weight_op[3], &weight_avg[3],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
            }else if(IS_SUB_8X4(sub_mb_type)){
                mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
                    &weight_op[4], &weight_avg[4],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
                    qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
                    &weight_op[4], &weight_avg[4],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
            }else if(IS_SUB_4X8(sub_mb_type)){
                mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
                    qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
                    &weight_op[5], &weight_avg[5],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
                    qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
                    &weight_op[5], &weight_avg[5],
                    IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
            }else{
                int j;
                assert(IS_SUB_4X4(sub_mb_type));
                for(j=0; j<4; j++){
                    int sub_x_offset= x_offset + 2*(j&1);
                    int sub_y_offset= y_offset +   (j&2);
                    mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
                        qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
                        &weight_op[6], &weight_avg[6],
                        IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
                }
            }
        }
    }

    prefetch_motion(h, 1);
}

static av_cold void init_cavlc_level_tab(void){
    int suffix_length, mask;
    unsigned int i;

    for(suffix_length=0; suffix_length<7; suffix_length++){
        for(i=0; i<(1<<LEVEL_TAB_BITS); i++){
            int prefix= LEVEL_TAB_BITS - av_log2(2*i);
            int level_code= (prefix<<suffix_length) + (i>>(LEVEL_TAB_BITS-prefix-1-suffix_length)) - (1<<suffix_length);

            mask= -(level_code&1);
            level_code= (((2+level_code)>>1) ^ mask) - mask;
            if(prefix + 1 + suffix_length <= LEVEL_TAB_BITS){
                cavlc_level_tab[suffix_length][i][0]= level_code;
                cavlc_level_tab[suffix_length][i][1]= prefix + 1 + suffix_length;
            }else if(prefix + 1 <= LEVEL_TAB_BITS){
                cavlc_level_tab[suffix_length][i][0]= prefix+100;
                cavlc_level_tab[suffix_length][i][1]= prefix + 1;
            }else{
                cavlc_level_tab[suffix_length][i][0]= LEVEL_TAB_BITS+100;
                cavlc_level_tab[suffix_length][i][1]= LEVEL_TAB_BITS;
            }
        }
    }
}

static av_cold void decode_init_vlc(void){
    static int done = 0;

    if (!done) {
        int i;
        int offset;
        done = 1;

        chroma_dc_coeff_token_vlc.table = chroma_dc_coeff_token_vlc_table;
        chroma_dc_coeff_token_vlc.table_allocated = chroma_dc_coeff_token_vlc_table_size;
        init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
                 &chroma_dc_coeff_token_len [0], 1, 1,
                 &chroma_dc_coeff_token_bits[0], 1, 1,
                 INIT_VLC_USE_NEW_STATIC);

        offset = 0;
        for(i=0; i<4; i++){
            coeff_token_vlc[i].table = coeff_token_vlc_tables+offset;
            coeff_token_vlc[i].table_allocated = coeff_token_vlc_tables_size[i];
            init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
                     &coeff_token_len [i][0], 1, 1,
                     &coeff_token_bits[i][0], 1, 1,
                     INIT_VLC_USE_NEW_STATIC);
            offset += coeff_token_vlc_tables_size[i];
        }
        /*
         * This is a one time safety check to make sure that
         * the packed static coeff_token_vlc table sizes
         * were initialized correctly.
         */
        assert(offset == FF_ARRAY_ELEMS(coeff_token_vlc_tables));

        for(i=0; i<3; i++){
            chroma_dc_total_zeros_vlc[i].table = chroma_dc_total_zeros_vlc_tables[i];
            chroma_dc_total_zeros_vlc[i].table_allocated = chroma_dc_total_zeros_vlc_tables_size;
            init_vlc(&chroma_dc_total_zeros_vlc[i],
                     CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
                     &chroma_dc_total_zeros_len [i][0], 1, 1,
                     &chroma_dc_total_zeros_bits[i][0], 1, 1,
                     INIT_VLC_USE_NEW_STATIC);
        }
        for(i=0; i<15; i++){
            total_zeros_vlc[i].table = total_zeros_vlc_tables[i];
            total_zeros_vlc[i].table_allocated = total_zeros_vlc_tables_size;
            init_vlc(&total_zeros_vlc[i],
                     TOTAL_ZEROS_VLC_BITS, 16,
                     &total_zeros_len [i][0], 1, 1,
                     &total_zeros_bits[i][0], 1, 1,
                     INIT_VLC_USE_NEW_STATIC);
        }

        for(i=0; i<6; i++){
            run_vlc[i].table = run_vlc_tables[i];
            run_vlc[i].table_allocated = run_vlc_tables_size;
            init_vlc(&run_vlc[i],
                     RUN_VLC_BITS, 7,
                     &run_len [i][0], 1, 1,
                     &run_bits[i][0], 1, 1,
                     INIT_VLC_USE_NEW_STATIC);
        }
        run7_vlc.table = run7_vlc_table,
        run7_vlc.table_allocated = run7_vlc_table_size;
        init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
                 &run_len [6][0], 1, 1,
                 &run_bits[6][0], 1, 1,
                 INIT_VLC_USE_NEW_STATIC);

        init_cavlc_level_tab();
    }
}

static void free_tables(H264Context *h){
    int i;
    H264Context *hx;
    av_freep(&h->intra4x4_pred_mode);
    av_freep(&h->chroma_pred_mode_table);
    av_freep(&h->cbp_table);
    av_freep(&h->mvd_table[0]);
    av_freep(&h->mvd_table[1]);
    av_freep(&h->direct_table);
    av_freep(&h->non_zero_count);
    av_freep(&h->slice_table_base);
    h->slice_table= NULL;

    av_freep(&h->mb2b_xy);
    av_freep(&h->mb2b8_xy);

    for(i = 0; i < h->s.avctx->thread_count; i++) {
        hx = h->thread_context[i];
        if(!hx) continue;
        av_freep(&hx->top_borders[1]);
        av_freep(&hx->top_borders[0]);
        av_freep(&hx->s.obmc_scratchpad);
    }
}

static void init_dequant8_coeff_table(H264Context *h){
    int i,q,x;
    const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
    h->dequant8_coeff[0] = h->dequant8_buffer[0];
    h->dequant8_coeff[1] = h->dequant8_buffer[1];

    for(i=0; i<2; i++ ){
        if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
            h->dequant8_coeff[1] = h->dequant8_buffer[0];
            break;
        }

        for(q=0; q<52; q++){
            int shift = div6[q];
            int idx = rem6[q];
            for(x=0; x<64; x++)
                h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
                    ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
                    h->pps.scaling_matrix8[i][x]) << shift;
        }
    }
}

static void init_dequant4_coeff_table(H264Context *h){
    int i,j,q,x;
    const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
    for(i=0; i<6; i++ ){
        h->dequant4_coeff[i] = h->dequant4_buffer[i];
        for(j=0; j<i; j++){
            if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
                h->dequant4_coeff[i] = h->dequant4_buffer[j];
                break;
            }
        }
        if(j<i)
            continue;

        for(q=0; q<52; q++){
            int shift = div6[q] + 2;
            int idx = rem6[q];
            for(x=0; x<16; x++)
                h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
                    ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
                    h->pps.scaling_matrix4[i][x]) << shift;
        }
    }
}

static void init_dequant_tables(H264Context *h){
    int i,x;
    init_dequant4_coeff_table(h);
    if(h->pps.transform_8x8_mode)
        init_dequant8_coeff_table(h);
    if(h->sps.transform_bypass){
        for(i=0; i<6; i++)
            for(x=0; x<16; x++)
                h->dequant4_coeff[i][0][x] = 1<<6;
        if(h->pps.transform_8x8_mode)
            for(i=0; i<2; i++)
                for(x=0; x<64; x++)
                    h->dequant8_coeff[i][0][x] = 1<<6;
    }
}


/**
 * allocates tables.
 * needs width/height
 */
static int alloc_tables(H264Context *h){
    MpegEncContext * const s = &h->s;
    const int big_mb_num= s->mb_stride * (s->mb_height+1);
    int x,y;

    CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))

    CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
    CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base))
    CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))

    CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
    CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
    CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
    CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));

    memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
    h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;

    CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
    CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
    for(y=0; y<s->mb_height; y++){
        for(x=0; x<s->mb_width; x++){
            const int mb_xy= x + y*s->mb_stride;
            const int b_xy = 4*x + 4*y*h->b_stride;
            const int b8_xy= 2*x + 2*y*h->b8_stride;

            h->mb2b_xy [mb_xy]= b_xy;
            h->mb2b8_xy[mb_xy]= b8_xy;
        }
    }

    s->obmc_scratchpad = NULL;

    if(!h->dequant4_coeff[0])
        init_dequant_tables(h);

    return 0;
fail:
    free_tables(h);
    return -1;
}

/**
 * Mimic alloc_tables(), but for every context thread.
 */
static void clone_tables(H264Context *dst, H264Context *src){
    dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
    dst->non_zero_count           = src->non_zero_count;
    dst->slice_table              = src->slice_table;
    dst->cbp_table                = src->cbp_table;
    dst->mb2b_xy                  = src->mb2b_xy;
    dst->mb2b8_xy                 = src->mb2b8_xy;
    dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
    dst->mvd_table[0]             = src->mvd_table[0];
    dst->mvd_table[1]             = src->mvd_table[1];
    dst->direct_table             = src->direct_table;

    dst->s.obmc_scratchpad = NULL;
    ff_h264_pred_init(&dst->hpc, src->s.codec_id);
}

/**
 * Init context
 * Allocate buffers which are not shared amongst multiple threads.
 */
static int context_init(H264Context *h){
    CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
    CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))

    return 0;
fail:
    return -1; // free_tables will clean up for us
}

static av_cold void common_init(H264Context *h){
    MpegEncContext * const s = &h->s;

    s->width = s->avctx->width;
    s->height = s->avctx->height;
    s->codec_id= s->avctx->codec->id;

    ff_h264_pred_init(&h->hpc, s->codec_id);

    h->dequant_coeff_pps= -1;
    s->unrestricted_mv=1;
    s->decode=1; //FIXME

    dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early

    memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
    memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
}

static av_cold int decode_init(AVCodecContext *avctx){
    H264Context *h= avctx->priv_data;
    MpegEncContext * const s = &h->s;

    MPV_decode_defaults(s);

    s->avctx = avctx;
    common_init(h);

    s->out_format = FMT_H264;
    s->workaround_bugs= avctx->workaround_bugs;

    // set defaults
//    s->decode_mb= ff_h263_decode_mb;
    s->quarter_sample = 1;
    s->low_delay= 1;

    if(avctx->codec_id == CODEC_ID_SVQ3)
        avctx->pix_fmt= PIX_FMT_YUVJ420P;
    else
        avctx->pix_fmt= PIX_FMT_YUV420P;

    decode_init_vlc();

    if(avctx->extradata_size > 0 && avctx->extradata &&
       *(char *)avctx->extradata == 1){
        h->is_avc = 1;
        h->got_avcC = 0;
    } else {
        h->is_avc = 0;
    }

    h->thread_context[0] = h;
    h->outputed_poc = INT_MIN;
    h->prev_poc_msb= 1<<16;
    return 0;
}

static int frame_start(H264Context *h){
    MpegEncContext * const s = &h->s;
    int i;

    if(MPV_frame_start(s, s->avctx) < 0)
        return -1;
    ff_er_frame_start(s);
    /*
     * MPV_frame_start uses pict_type to derive key_frame.
     * This is incorrect for H.264; IDR markings must be used.
     * Zero here; IDR markings per slice in frame or fields are ORed in later.
     * See decode_nal_units().
     */
    s->current_picture_ptr->key_frame= 0;

    assert(s->linesize && s->uvlinesize);

    for(i=0; i<16; i++){
        h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
        h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
    }
    for(i=0; i<4; i++){
        h->block_offset[16+i]=
        h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
        h->block_offset[24+16+i]=
        h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
    }

    /* can't be in alloc_tables because linesize isn't known there.
     * FIXME: redo bipred weight to not require extra buffer? */
    for(i = 0; i < s->avctx->thread_count; i++)
        if(!h->thread_context[i]->s.obmc_scratchpad)
            h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);

    /* some macroblocks will be accessed before they're available */
    if(FRAME_MBAFF || s->avctx->thread_count > 1)
        memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));

//    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;

    // We mark the current picture as non-reference after allocating it, so
    // that if we break out due to an error it can be released automatically
    // in the next MPV_frame_start().
    // SVQ3 as well as most other codecs have only last/next/current and thus
    // get released even with set reference, besides SVQ3 and others do not
    // mark frames as reference later "naturally".
    if(s->codec_id != CODEC_ID_SVQ3)
        s->current_picture_ptr->reference= 0;

    s->current_picture_ptr->field_poc[0]=
    s->current_picture_ptr->field_poc[1]= INT_MAX;
    assert(s->current_picture_ptr->long_ref==0);

    return 0;
}

static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
    MpegEncContext * const s = &h->s;
    int i;
    int step    = 1;
    int offset  = 1;
    int uvoffset= 1;
    int top_idx = 1;
    int skiplast= 0;

    src_y  -=   linesize;
    src_cb -= uvlinesize;
    src_cr -= uvlinesize;

    if(!simple && FRAME_MBAFF){
        if(s->mb_y&1){
            offset  = MB_MBAFF ? 1 : 17;
            uvoffset= MB_MBAFF ? 1 : 9;
            if(!MB_MBAFF){
                *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
                *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
                if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
                    *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+7*uvlinesize);
                    *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
                }
            }
        }else{
            if(!MB_MBAFF){
                h->left_border[0]= h->top_borders[0][s->mb_x][15];
                if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
                    h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
                    h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
                }
                skiplast= 1;
            }
            offset  =
            uvoffset=
            top_idx = MB_MBAFF ? 0 : 1;
        }
        step= MB_MBAFF ? 2 : 1;
    }

    // There are two lines saved, the line above the the top macroblock of a pair,
    // and the line above the bottom macroblock
    h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
    for(i=1; i<17 - skiplast; i++){
        h->left_border[offset+i*step]= src_y[15+i*  linesize];
    }

    *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
    *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);

    if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
        h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
        h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
        for(i=1; i<9 - skiplast; i++){
            h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
            h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
        }
        *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
        *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
    }
}

static inline void xchg_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg, int simple){
    MpegEncContext * const s = &h->s;
    int temp8, i;
    uint64_t temp64;
    int deblock_left;
    int deblock_top;
    int mb_xy;
    int step    = 1;
    int offset  = 1;
    int uvoffset= 1;
    int top_idx = 1;

    if(!simple && FRAME_MBAFF){
        if(s->mb_y&1){
            offset  = MB_MBAFF ? 1 : 17;
            uvoffset= MB_MBAFF ? 1 : 9;
        }else{
            offset  =
            uvoffset=
            top_idx = MB_MBAFF ? 0 : 1;
        }
        step= MB_MBAFF ? 2 : 1;
    }

    if(h->deblocking_filter == 2) {
        mb_xy = h->mb_xy;
        deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
        deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
    } else {
        deblock_left = (s->mb_x > 0);
        deblock_top =  (s->mb_y > !!MB_FIELD);
    }

    src_y  -=   linesize + 1;
    src_cb -= uvlinesize + 1;
    src_cr -= uvlinesize + 1;

#define XCHG(a,b,t,xchg)\
t= a;\
if(xchg)\
    a= b;\
b= t;

    if(deblock_left){
        for(i = !deblock_top; i<16; i++){
            XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
        }
        XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
    }

    if(deblock_top){
        XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
        XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
        if(s->mb_x+1 < s->mb_width){
            XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
        }
    }

    if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
        if(deblock_left){
            for(i = !deblock_top; i<8; i++){
                XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
                XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
            }
            XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
            XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
        }
        if(deblock_top){
            XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
            XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
        }
    }
}

static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
    MpegEncContext * const s = &h->s;
    const int mb_x= s->mb_x;
    const int mb_y= s->mb_y;
    const int mb_xy= h->mb_xy;
    const int mb_type= s->current_picture.mb_type[mb_xy];
    uint8_t  *dest_y, *dest_cb, *dest_cr;
    int linesize, uvlinesize /*dct_offset*/;
    int i;
    int *block_offset = &h->block_offset[0];
    const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
    const int is_h264 = simple || s->codec_id == CODEC_ID_H264;
    void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
    void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);

    dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
    dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
    dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;

    s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
    s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);

    if (!simple && MB_FIELD) {
        linesize   = h->mb_linesize   = s->linesize * 2;
        uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
        block_offset = &h->block_offset[24];
        if(mb_y&1){ //FIXME move out of this function?
            dest_y -= s->linesize*15;
            dest_cb-= s->uvlinesize*7;
            dest_cr-= s->uvlinesize*7;
        }
        if(FRAME_MBAFF) {
            int list;
            for(list=0; list<h->list_count; list++){
                if(!USES_LIST(mb_type, list))
                    continue;
                if(IS_16X16(mb_type)){
                    int8_t *ref = &h->ref_cache[list][scan8[0]];
                    fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
                }else{
                    for(i=0; i<16; i+=4){
                        int ref = h->ref_cache[list][scan8[i]];
                        if(ref >= 0)
                            fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
                    }
                }
            }
        }
    } else {
        linesize   = h->mb_linesize   = s->linesize;
        uvlinesize = h->mb_uvlinesize = s->uvlinesize;
//        dct_offset = s->linesize * 16;
    }

    if (!simple && IS_INTRA_PCM(mb_type)) {
        for (i=0; i<16; i++) {
            memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
        }
        for (i=0; i<8; i++) {
            memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
            memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
        }
    } else {
        if(IS_INTRA(mb_type)){
            if(h->deblocking_filter)
                xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);

            if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
                h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
                h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
            }

            if(IS_INTRA4x4(mb_type)){
                if(simple || !s->encoding){
                    if(IS_8x8DCT(mb_type)){
                        if(transform_bypass){
                            idct_dc_add =
                            idct_add    = s->dsp.add_pixels8;
                        }else{
                            idct_dc_add = s->dsp.h264_idct8_dc_add;
                            idct_add    = s->dsp.h264_idct8_add;
                        }
                        for(i=0; i<16; i+=4){
                            uint8_t * const ptr= dest_y + block_offset[i];
                            const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
                            if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
                                h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
                            }else{
                                const int nnz = h->non_zero_count_cache[ scan8[i] ];
                                h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
                                                            (h->topright_samples_available<<i)&0x4000, linesize);
                                if(nnz){
                                    if(nnz == 1 && h->mb[i*16])
                                        idct_dc_add(ptr, h->mb + i*16, linesize);
                                    else
                                        idct_add   (ptr, h->mb + i*16, linesize);
                                }
                            }
                        }
                    }else{
                        if(transform_bypass){
                            idct_dc_add =
                            idct_add    = s->dsp.add_pixels4;
                        }else{
                            idct_dc_add = s->dsp.h264_idct_dc_add;
                            idct_add    = s->dsp.h264_idct_add;
                        }
                        for(i=0; i<16; i++){
                            uint8_t * const ptr= dest_y + block_offset[i];
                            const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];

                            if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
                                h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
                            }else{
                                uint8_t *topright;
                                int nnz, tr;
                                if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
                                    const int topright_avail= (h->topright_samples_available<<i)&0x8000;
                                    assert(mb_y || linesize <= block_offset[i]);
                                    if(!topright_avail){
                                        tr= ptr[3 - linesize]*0x01010101;
                                        topright= (uint8_t*) &tr;
                                    }else
                                        topright= ptr + 4 - linesize;
                                }else
                                    topright= NULL;

                                h->hpc.pred4x4[ dir ](ptr, topright, linesize);
                                nnz = h->non_zero_count_cache[ scan8[i] ];
                                if(nnz){
                                    if(is_h264){
                                        if(nnz == 1 && h->mb[i*16])
                                            idct_dc_add(ptr, h->mb + i*16, linesize);
                                        else
                                            idct_add   (ptr, h->mb + i*16, linesize);
                                    }else
                                        svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
                                }
                            }
                        }
                    }
                }
            }else{
                h->hpc.pred16x16[ h->i