use 64bit types

[libav.git] / libavcodec / h264.c

/*
 * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
 *
 * This library 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 of the License, or (at your option) any later version.
 *
 * This library 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 this library; 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 "common.h"
#include "dsputil.h"
#include "avcodec.h"
#include "mpegvideo.h"
#include "h264data.h"
#include "golomb.h"

#include "cabac.h"

#undef NDEBUG
#include <assert.h>

#define interlaced_dct interlaced_dct_is_a_bad_name
#define mb_intra mb_intra_isnt_initalized_see_mb_type

#define LUMA_DC_BLOCK_INDEX   25
#define CHROMA_DC_BLOCK_INDEX 26

#define CHROMA_DC_COEFF_TOKEN_VLC_BITS 8
#define COEFF_TOKEN_VLC_BITS           8
#define TOTAL_ZEROS_VLC_BITS           9
#define CHROMA_DC_TOTAL_ZEROS_VLC_BITS 3
#define RUN_VLC_BITS                   3
#define RUN7_VLC_BITS                  6

#define MAX_SPS_COUNT 32
#define MAX_PPS_COUNT 256

#define MAX_MMCO_COUNT 66

/**
 * Sequence parameter set
 */
typedef struct SPS{

    int profile_idc;
    int level_idc;
    int transform_bypass;              ///< qpprime_y_zero_transform_bypass_flag
    int log2_max_frame_num;            ///< log2_max_frame_num_minus4 + 4
    int poc_type;                      ///< pic_order_cnt_type
    int log2_max_poc_lsb;              ///< log2_max_pic_order_cnt_lsb_minus4
    int delta_pic_order_always_zero_flag;
    int offset_for_non_ref_pic;
    int offset_for_top_to_bottom_field;
    int poc_cycle_length;              ///< num_ref_frames_in_pic_order_cnt_cycle
    int ref_frame_count;               ///< num_ref_frames
    int gaps_in_frame_num_allowed_flag;
    int mb_width;                      ///< frame_width_in_mbs_minus1 + 1
    int mb_height;                     ///< frame_height_in_mbs_minus1 + 1
    int frame_mbs_only_flag;
    int mb_aff;                        ///<mb_adaptive_frame_field_flag
    int direct_8x8_inference_flag;
    int crop;                   ///< frame_cropping_flag
    int crop_left;              ///< frame_cropping_rect_left_offset
    int crop_right;             ///< frame_cropping_rect_right_offset
    int crop_top;               ///< frame_cropping_rect_top_offset
    int crop_bottom;            ///< frame_cropping_rect_bottom_offset
    int vui_parameters_present_flag;
    AVRational sar;
    int timing_info_present_flag;
    uint32_t num_units_in_tick;
    uint32_t time_scale;
    int fixed_frame_rate_flag;
    short offset_for_ref_frame[256]; //FIXME dyn aloc?
    int bitstream_restriction_flag;
    int num_reorder_frames;
    int scaling_matrix_present;
    uint8_t scaling_matrix4[6][16];
    uint8_t scaling_matrix8[2][64];
}SPS;

/**
 * Picture parameter set
 */
typedef struct PPS{
    int sps_id;
    int cabac;                  ///< entropy_coding_mode_flag
    int pic_order_present;      ///< pic_order_present_flag
    int slice_group_count;      ///< num_slice_groups_minus1 + 1
    int mb_slice_group_map_type;
    int ref_count[2];           ///< num_ref_idx_l0/1_active_minus1 + 1
    int weighted_pred;          ///< weighted_pred_flag
    int weighted_bipred_idc;
    int init_qp;                ///< pic_init_qp_minus26 + 26
    int init_qs;                ///< pic_init_qs_minus26 + 26
    int chroma_qp_index_offset;
    int deblocking_filter_parameters_present; ///< deblocking_filter_parameters_present_flag
    int constrained_intra_pred; ///< constrained_intra_pred_flag
    int redundant_pic_cnt_present; ///< redundant_pic_cnt_present_flag
    int transform_8x8_mode;     ///< transform_8x8_mode_flag
    uint8_t scaling_matrix4[6][16];
    uint8_t scaling_matrix8[2][64];
}PPS;

/**
 * Memory management control operation opcode.
 */
typedef enum MMCOOpcode{
    MMCO_END=0,
    MMCO_SHORT2UNUSED,
    MMCO_LONG2UNUSED,
    MMCO_SHORT2LONG,
    MMCO_SET_MAX_LONG,
    MMCO_RESET,
    MMCO_LONG,
} MMCOOpcode;

/**
 * Memory management control operation.
 */
typedef struct MMCO{
    MMCOOpcode opcode;
    int short_frame_num;
    int long_index;
} MMCO;

/**
 * H264Context
 */
typedef struct H264Context{
    MpegEncContext s;
    int nal_ref_idc;
    int nal_unit_type;
#define NAL_SLICE                1
#define NAL_DPA                  2
#define NAL_DPB                  3
#define NAL_DPC                  4
#define NAL_IDR_SLICE            5
#define NAL_SEI                  6
#define NAL_SPS                  7
#define NAL_PPS                  8
#define NAL_AUD                  9
#define NAL_END_SEQUENCE        10
#define NAL_END_STREAM          11
#define NAL_FILLER_DATA         12
#define NAL_SPS_EXT             13
#define NAL_AUXILIARY_SLICE     19
    uint8_t *rbsp_buffer;
    unsigned int rbsp_buffer_size;

    /**
      * Used to parse AVC variant of h264
      */
    int is_avc; ///< this flag is != 0 if codec is avc1
    int got_avcC; ///< flag used to parse avcC data only once
    int nal_length_size; ///< Number of bytes used for nal length (1, 2 or 4)

    int chroma_qp; //QPc

    int prev_mb_skipped; //FIXME remove (IMHO not used)

    //prediction stuff
    int chroma_pred_mode;
    int intra16x16_pred_mode;

    int top_mb_xy;
    int left_mb_xy[2];

    int8_t intra4x4_pred_mode_cache[5*8];
    int8_t (*intra4x4_pred_mode)[8];
    void (*pred4x4  [9+3])(uint8_t *src, uint8_t *topright, int stride);//FIXME move to dsp?
    void (*pred8x8l [9+3])(uint8_t *src, int topleft, int topright, int stride);
    void (*pred8x8  [4+3])(uint8_t *src, int stride);
    void (*pred16x16[4+3])(uint8_t *src, int stride);
    unsigned int topleft_samples_available;
    unsigned int top_samples_available;
    unsigned int topright_samples_available;
    unsigned int left_samples_available;
    uint8_t (*top_borders[2])[16+2*8];
    uint8_t left_border[2*(17+2*9)];

    /**
     * non zero coeff count cache.
     * is 64 if not available.
     */
    DECLARE_ALIGNED_8(uint8_t, non_zero_count_cache[6*8]);
    uint8_t (*non_zero_count)[16];

    /**
     * Motion vector cache.
     */
    DECLARE_ALIGNED_8(int16_t, mv_cache[2][5*8][2]);
    DECLARE_ALIGNED_8(int8_t, ref_cache[2][5*8]);
#define LIST_NOT_USED -1 //FIXME rename?
#define PART_NOT_AVAILABLE -2

    /**
     * is 1 if the specific list MV&references are set to 0,0,-2.
     */
    int mv_cache_clean[2];

    /**
     * number of neighbors (top and/or left) that used 8x8 dct
     */
    int neighbor_transform_size;

    /**
     * block_offset[ 0..23] for frame macroblocks
     * block_offset[24..47] for field macroblocks
     */
    int block_offset[2*(16+8)];

    uint32_t *mb2b_xy; //FIXME are these 4 a good idea?
    uint32_t *mb2b8_xy;
    int b_stride; //FIXME use s->b4_stride
    int b8_stride;

    int halfpel_flag;
    int thirdpel_flag;

    int unknown_svq3_flag;
    int next_slice_index;

    SPS sps_buffer[MAX_SPS_COUNT];
    SPS sps; ///< current sps

    PPS pps_buffer[MAX_PPS_COUNT];
    /**
     * current pps
     */
    PPS pps; //FIXME move to Picture perhaps? (->no) do we need that?

    uint32_t dequant4_buffer[6][52][16];
    uint32_t dequant8_buffer[2][52][64];
    uint32_t (*dequant4_coeff[6])[16];
    uint32_t (*dequant8_coeff[2])[64];
    int dequant_coeff_pps;     ///< reinit tables when pps changes

    int slice_num;
    uint8_t *slice_table_base;
    uint8_t *slice_table;      ///< slice_table_base + mb_stride + 1
    int slice_type;
    int slice_type_fixed;

    //interlacing specific flags
    int mb_aff_frame;
    int mb_field_decoding_flag;

    int sub_mb_type[4];

    //POC stuff
    int poc_lsb;
    int poc_msb;
    int delta_poc_bottom;
    int delta_poc[2];
    int frame_num;
    int prev_poc_msb;             ///< poc_msb of the last reference pic for POC type 0
    int prev_poc_lsb;             ///< poc_lsb of the last reference pic for POC type 0
    int frame_num_offset;         ///< for POC type 2
    int prev_frame_num_offset;    ///< for POC type 2
    int prev_frame_num;           ///< frame_num of the last pic for POC type 1/2

    /**
     * frame_num for frames or 2*frame_num for field pics.
     */
    int curr_pic_num;

    /**
     * max_frame_num or 2*max_frame_num for field pics.
     */
    int max_pic_num;

    //Weighted pred stuff
    int use_weight;
    int use_weight_chroma;
    int luma_log2_weight_denom;
    int chroma_log2_weight_denom;
    int luma_weight[2][16];
    int luma_offset[2][16];
    int chroma_weight[2][16][2];
    int chroma_offset[2][16][2];
    int implicit_weight[16][16];

    //deblock
    int deblocking_filter;         ///< disable_deblocking_filter_idc with 1<->0
    int slice_alpha_c0_offset;
    int slice_beta_offset;

    int redundant_pic_count;

    int direct_spatial_mv_pred;
    int dist_scale_factor[16];
    int map_col_to_list0[2][16];

    /**
     * num_ref_idx_l0/1_active_minus1 + 1
     */
    int ref_count[2];// FIXME split for AFF
    Picture *short_ref[32];
    Picture *long_ref[32];
    Picture default_ref_list[2][32];
    Picture ref_list[2][32]; //FIXME size?
    Picture field_ref_list[2][32]; //FIXME size?
    Picture *delayed_pic[16]; //FIXME size?
    Picture *delayed_output_pic;

    /**
     * memory management control operations buffer.
     */
    MMCO mmco[MAX_MMCO_COUNT];
    int mmco_index;

    int long_ref_count;  ///< number of actual long term references
    int short_ref_count; ///< number of actual short term references

    //data partitioning
    GetBitContext intra_gb;
    GetBitContext inter_gb;
    GetBitContext *intra_gb_ptr;
    GetBitContext *inter_gb_ptr;

    DECLARE_ALIGNED_8(DCTELEM, mb[16*24]);

    /**
     * Cabac
     */
    CABACContext cabac;
    uint8_t      cabac_state[460];
    int          cabac_init_idc;

    /* 0x100 -> non null luma_dc, 0x80/0x40 -> non null chroma_dc (cb/cr), 0x?0 -> chroma_cbp(0,1,2), 0x0? luma_cbp */
    uint16_t     *cbp_table;
    int top_cbp;
    int left_cbp;
    /* chroma_pred_mode for i4x4 or i16x16, else 0 */
    uint8_t     *chroma_pred_mode_table;
    int         last_qscale_diff;
    int16_t     (*mvd_table[2])[2];
    DECLARE_ALIGNED_8(int16_t, mvd_cache[2][5*8][2]);
    uint8_t     *direct_table;
    uint8_t     direct_cache[5*8];

    uint8_t zigzag_scan[16];
    uint8_t field_scan[16];
    uint8_t zigzag_scan8x8[64];
    uint8_t zigzag_scan8x8_cavlc[64];
    const uint8_t *zigzag_scan_q0;
    const uint8_t *field_scan_q0;
    const uint8_t *zigzag_scan8x8_q0;
    const uint8_t *zigzag_scan8x8_cavlc_q0;

    int x264_build;
}H264Context;

static VLC coeff_token_vlc[4];
static VLC chroma_dc_coeff_token_vlc;

static VLC total_zeros_vlc[15];
static VLC chroma_dc_total_zeros_vlc[3];

static VLC run_vlc[6];
static VLC run7_vlc;

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

/**
 * fill a rectangle.
 * @param h height of the rectangle, should be a constant
 * @param w width of the rectangle, should be a constant
 * @param size the size of val (1 or 4), should be a constant
 */
static always_inline void fill_rectangle(void *vp, int w, int h, int stride, uint32_t val, int size){
    uint8_t *p= (uint8_t*)vp;
    assert(size==1 || size==4);

    w      *= size;
    stride *= size;

    assert((((long)vp)&(FFMIN(w, STRIDE_ALIGN)-1)) == 0);
    assert((stride&(w-1))==0);
//FIXME check what gcc generates for 64 bit on x86 and possibly write a 32 bit ver of it
    if(w==2 && h==2){
        *(uint16_t*)(p + 0)=
        *(uint16_t*)(p + stride)= size==4 ? val : val*0x0101;
    }else if(w==2 && h==4){
        *(uint16_t*)(p + 0*stride)=
        *(uint16_t*)(p + 1*stride)=
        *(uint16_t*)(p + 2*stride)=
        *(uint16_t*)(p + 3*stride)= size==4 ? val : val*0x0101;
    }else if(w==4 && h==1){
        *(uint32_t*)(p + 0*stride)= size==4 ? val : val*0x01010101;
    }else if(w==4 && h==2){
        *(uint32_t*)(p + 0*stride)=
        *(uint32_t*)(p + 1*stride)= size==4 ? val : val*0x01010101;
    }else if(w==4 && h==4){
        *(uint32_t*)(p + 0*stride)=
        *(uint32_t*)(p + 1*stride)=
        *(uint32_t*)(p + 2*stride)=
        *(uint32_t*)(p + 3*stride)= size==4 ? val : val*0x01010101;
    }else if(w==8 && h==1){
        *(uint32_t*)(p + 0)=
        *(uint32_t*)(p + 4)= size==4 ? val : val*0x01010101;
    }else if(w==8 && h==2){
        *(uint32_t*)(p + 0 + 0*stride)=
        *(uint32_t*)(p + 4 + 0*stride)=
        *(uint32_t*)(p + 0 + 1*stride)=
        *(uint32_t*)(p + 4 + 1*stride)=  size==4 ? val : val*0x01010101;
    }else if(w==8 && h==4){
        *(uint64_t*)(p + 0*stride)=
        *(uint64_t*)(p + 1*stride)=
        *(uint64_t*)(p + 2*stride)=
        *(uint64_t*)(p + 3*stride)= size==4 ? val*0x0100000001ULL : val*0x0101010101010101ULL;
    }else if(w==16 && h==2){
        *(uint64_t*)(p + 0+0*stride)=
        *(uint64_t*)(p + 8+0*stride)=
        *(uint64_t*)(p + 0+1*stride)=
        *(uint64_t*)(p + 8+1*stride)= size==4 ? val*0x0100000001ULL : val*0x0101010101010101ULL;
    }else if(w==16 && h==4){
        *(uint64_t*)(p + 0+0*stride)=
        *(uint64_t*)(p + 8+0*stride)=
        *(uint64_t*)(p + 0+1*stride)=
        *(uint64_t*)(p + 8+1*stride)=
        *(uint64_t*)(p + 0+2*stride)=
        *(uint64_t*)(p + 8+2*stride)=
        *(uint64_t*)(p + 0+3*stride)=
        *(uint64_t*)(p + 8+3*stride)= size==4 ? val*0x0100000001ULL : val*0x0101010101010101ULL;
    }else
        assert(0);
}

static void fill_caches(H264Context *h, int mb_type, int for_deblock){
    MpegEncContext * const s = &h->s;
    const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
    int topleft_xy, top_xy, topright_xy, left_xy[2];
    int topleft_type, top_type, topright_type, left_type[2];
    int left_block[8];
    int i;

    //FIXME deblocking can skip fill_caches much of the time with multiple slices too.
    // the actual condition is whether we're on the edge of a slice,
    // and even then the intra and nnz parts are unnecessary.
    if(for_deblock && h->slice_num == 1)
        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

    top_xy     = mb_xy  - s->mb_stride;
    topleft_xy = top_xy - 1;
    topright_xy= top_xy + 1;
    left_xy[1] = left_xy[0] = mb_xy-1;
    left_block[0]= 0;
    left_block[1]= 1;
    left_block[2]= 2;
    left_block[3]= 3;
    left_block[4]= 7;
    left_block[5]= 10;
    left_block[6]= 8;
    left_block[7]= 11;
    if(h->mb_aff_frame){
        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_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
        const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
        const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
        const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
        const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
        const int bottom = (s->mb_y & 1);
        tprintf("fill_caches: curr_mb_frame_flag:%d, left_mb_frame_flag:%d, topleft_mb_frame_flag:%d, top_mb_frame_flag:%d, topright_mb_frame_flag:%d\n", curr_mb_frame_flag, left_mb_frame_flag, topleft_mb_frame_flag, top_mb_frame_flag, topright_mb_frame_flag);
        if (bottom
                ? !curr_mb_frame_flag // bottom macroblock
                : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
                ) {
            top_xy -= s->mb_stride;
        }
        if (bottom
                ? !curr_mb_frame_flag // bottom macroblock
                : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
                ) {
            topleft_xy -= s->mb_stride;
        }
        if (bottom
                ? !curr_mb_frame_flag // bottom macroblock
                : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
                ) {
            topright_xy -= s->mb_stride;
        }
        if (left_mb_frame_flag != curr_mb_frame_flag) {
            left_xy[1] = left_xy[0] = pair_xy - 1;
            if (curr_mb_frame_flag) {
                if (bottom) {
                    left_block[0]= 2;
                    left_block[1]= 2;
                    left_block[2]= 3;
                    left_block[3]= 3;
                    left_block[4]= 8;
                    left_block[5]= 11;
                    left_block[6]= 8;
                    left_block[7]= 11;
                } else {
                    left_block[0]= 0;
                    left_block[1]= 0;
                    left_block[2]= 1;
                    left_block[3]= 1;
                    left_block[4]= 7;
                    left_block[5]= 10;
                    left_block[6]= 7;
                    left_block[7]= 10;
                }
            } else {
                left_xy[1] += s->mb_stride;
                //left_block[0]= 0;
                left_block[1]= 2;
                left_block[2]= 0;
                left_block[3]= 2;
                //left_block[4]= 7;
                left_block[5]= 10;
                left_block[6]= 7;
                left_block[7]= 10;
            }
        }
    }

    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 = h->slice_table[topleft_xy ] < 255 ? s->current_picture.mb_type[topleft_xy] : 0;
        top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
        topright_type= h->slice_table[topright_xy] < 255 ? s->current_picture.mb_type[topright_xy]: 0;
        left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
        left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
    }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)){
        h->topleft_samples_available=
        h->top_samples_available=
        h->left_samples_available= 0xFFFF;
        h->topright_samples_available= 0xEEEA;

        if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
            h->topleft_samples_available= 0xB3FF;
            h->top_samples_available= 0x33FF;
            h->topright_samples_available= 0x26EA;
        }
        for(i=0; i<2; i++){
            if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
                h->topleft_samples_available&= 0xDF5F;
                h->left_samples_available&= 0x5F5F;
            }
        }

        if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
            h->topleft_samples_available&= 0x7FFF;

        if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
            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 || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
                    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] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
                        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 (lets 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
    //FIXME direct mb can skip much of this
    if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
        int list;
        for(list=0; list<1+(h->slice_type==B_TYPE); 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;
            }

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

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

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

            if(USES_LIST(topleft_type, list)){
                const int b_xy = h->mb2b_xy[topleft_xy] + 3 + 3*h->b_stride;
                const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + 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;
            }


            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(topleft_type, list)){
                    const int b_xy = h->mb2b_xy[topleft_xy] + 3 + 3*h->b_stride;
                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy];
                }else{
                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 - 1*8]= 0;
                }

                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 == 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;
                    }

                    //FIXME interlacing
                    if(IS_DIRECT(left_type[0])){
                        h->direct_cache[scan8[0] - 1 + 0*8]=
                        h->direct_cache[scan8[0] - 1 + 2*8]= 1;
                    }else if(IS_8X8(left_type[0])){
                        int b8_xy = h->mb2b8_xy[left_xy[0]] + 1;
                        h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[b8_xy];
                        h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[b8_xy + h->b8_stride];
                    }else{
                        h->direct_cache[scan8[0] - 1 + 0*8]=
                        h->direct_cache[scan8[0] - 1 + 2*8]= 0;
                    }
                }
            }
        }
    }
#endif

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

static inline void write_back_intra_pred_mode(H264Context *h){
    MpegEncContext * const s = &h->s;
    const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;

    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&0x8000)){
        for(i=0; i<4; 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 < 0 || mode > 6) {
        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&0x8000)){
        mode= left[ mode ];
        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("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){
    MpegEncContext * const s = &h->s;
    const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;

    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("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 ];

    if(topright_ref != PART_NOT_AVAILABLE){
        *C= h->mv_cache[list][ i - 8 + part_width ];
        return topright_ref;
    }else{
        tprintf("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("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("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("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("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("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("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("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 == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
       || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){

        *mx = *my = 0;
        return;
    }

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

    return;
}

static inline void direct_dist_scale_factor(H264Context * const h){
    const int poc = h->s.current_picture_ptr->poc;
    const int poc1 = h->ref_list[1][0].poc;
    int i;
    for(i=0; i<h->ref_count[0]; i++){
        int poc0 = h->ref_list[0][i].poc;
        int td = clip(poc1 - poc0, -128, 127);
        if(td == 0 /* FIXME || pic0 is a long-term ref */){
            h->dist_scale_factor[i] = 256;
        }else{
            int tb = clip(poc - poc0, -128, 127);
            int tx = (16384 + (ABS(td) >> 1)) / td;
            h->dist_scale_factor[i] = clip((tb*tx + 32) >> 6, -1024, 1023);
        }
    }
}
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, i, j;
    if(cur->pict_type == I_TYPE)
        cur->ref_count[0] = 0;
    if(cur->pict_type != B_TYPE)
        cur->ref_count[1] = 0;
    for(list=0; list<2; list++){
        cur->ref_count[list] = h->ref_count[list];
        for(j=0; j<h->ref_count[list]; j++)
            cur->ref_poc[list][j] = h->ref_list[list][j].poc;
    }
    if(cur->pict_type != B_TYPE || h->direct_spatial_mv_pred)
        return;
    for(list=0; list<2; list++){
        for(i=0; i<ref1->ref_count[list]; i++){
            const int poc = ref1->ref_poc[list][i];
            h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
            for(j=0; j<h->ref_count[list]; j++)
                if(h->ref_list[list][j].poc == poc){
                    h->map_col_to_list0[list][i] = j;
                    break;
                }
        }
    }
}

static inline void pred_direct_motion(H264Context * const h, int *mb_type){
    MpegEncContext * const s = &h->s;
    const int mb_xy =   s->mb_x +   s->mb_y*s->mb_stride;
    const int b8_xy = 2*s->mb_x + 2*s->mb_y*h->b8_stride;
    const int b4_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
    const int mb_type_col = h->ref_list[1][0].mb_type[mb_xy];
    const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[0][b4_xy];
    const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[1][b4_xy];
    const int8_t *l1ref0 = &h->ref_list[1][0].ref_index[0][b8_xy];
    const int8_t *l1ref1 = &h->ref_list[1][0].ref_index[1][b8_xy];
    const int is_b8x8 = IS_8X8(*mb_type);
    int sub_mb_type;
    int i8, i4;

    if(IS_8X8(mb_type_col) && !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 && (IS_16X16(mb_type_col) || IS_INTRA(mb_type_col))){
        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;
    }
    if(!is_b8x8)
        *mb_type |= MB_TYPE_DIRECT2;

    tprintf("mb_type = %08x, sub_mb_type = %08x, is_b8x8 = %d, mb_type_col = %08x\n", *mb_type, sub_mb_type, is_b8x8, mb_type_col);

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

        /* 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 == -2)
                refc = h->ref_cache[list][scan8[0] - 8 - 1];
            ref[list] = refa;
            if(ref[list] < 0 || (refb < ref[list] && refb >= 0))
                ref[list] = refb;
            if(ref[list] < 0 || (refc < ref[list] && refc >= 0))
                ref[list] = 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){
            *mb_type &= ~MB_TYPE_P0L1;
            sub_mb_type &= ~MB_TYPE_P0L1;
        }else if(ref[0] < 0){
            *mb_type &= ~MB_TYPE_P0L0;
            sub_mb_type &= ~MB_TYPE_P0L0;
        }

        if(IS_16X16(*mb_type)){
            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)
               && (   (l1ref0[0] == 0 && ABS(l1mv0[0][0]) <= 1 && ABS(l1mv0[0][1]) <= 1)
                   || (l1ref0[0]  < 0 && l1ref1[0] == 0 && ABS(l1mv1[0][0]) <= 1 && ABS(l1mv1[0][1]) <= 1
                       && (h->x264_build>33 || !h->x264_build)))){
                if(ref[0] > 0)
                    fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mv[0][0],mv[0][1]), 4);
                else
                    fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
                if(ref[1] > 0)
                    fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, pack16to32(mv[1][0],mv[1][1]), 4);
                else
                    fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
            }else{
                fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mv[0][0],mv[0][1]), 4);
                fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, pack16to32(mv[1][0],mv[1][1]), 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) && (   l1ref0[x8 + y8*h->b8_stride] == 0
                                              || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
                                                  && (h->x264_build>33 || !h->x264_build)))){
                    const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
                    if(IS_SUB_8X8(sub_mb_type)){
                        const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
                        if(ABS(mv_col[0]) <= 1 && ABS(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))*h->b_stride];
                        if(ABS(mv_col[0]) <= 1 && ABS(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 */
        if(IS_16X16(*mb_type)){
            fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
            if(IS_INTRA(mb_type_col)){
                fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, 0, 1);
                fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
                fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
            }else{
                const int ref0 = l1ref0[0] >= 0 ? h->map_col_to_list0[0][l1ref0[0]]
                                                : h->map_col_to_list0[1][l1ref1[0]];
                const int dist_scale_factor = h->dist_scale_factor[ref0];
                const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
                int mv_l0[2];
                mv_l0[0] = (dist_scale_factor * mv_col[0] + 128) >> 8;
                mv_l0[1] = (dist_scale_factor * mv_col[1] + 128) >> 8;
                fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref0, 1);
                fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mv_l0[0],mv_l0[1]), 4);
                fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]), 4);
            }
        }else{
            for(i8=0; i8<4; i8++){
                const int x8 = i8&1;
                const int y8 = i8>>1;
                int ref0, dist_scale_factor;
                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;
                if(IS_INTRA(mb_type_col)){
                    fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
                    fill_rectangle(&h->ref_cache[1][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*h->b8_stride];
                if(ref0 >= 0)
                    ref0 = h->map_col_to_list0[0][ref0];
                else{
                    ref0 = h->map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
                    l1mv= l1mv1;
                }
                dist_scale_factor = h->dist_scale_factor[ref0];

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

    for(list=0; list<2; 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 ) {
            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];
            }
        }

        {
            uint8_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 == 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 uint8_t *decode_nal(H264Context *h, uint8_t *src, int *dst_length, int *consumed, int length){
    int i, si, di;
    uint8_t *dst;

//    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
    for(i=0; i+1<length; i+=2){
        if(src[i]) continue;
        if(i>0 && src[i-1]==0) i--;
        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;
        }
    }

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

    h->rbsp_buffer= av_fast_realloc(h->rbsp_buffer, &h->rbsp_buffer_size, length);
    dst= h->rbsp_buffer;

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

        dst[di++]= src[si++];
    }

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

#if 0
/**
 * @param src the data which should be escaped
 * @param dst the target buffer, dst+1 == src is allowed as a special case
 * @param length the length of the src data
 * @param dst_length the length of the dst array
 * @returns length of escaped data in bytes or -1 if an error occured
 */
static int encode_nal(H264Context *h, uint8_t *dst, uint8_t *src, int length, int dst_length){
    int i, escape_count, si, di;
    uint8_t *temp;

    assert(length>=0);
    assert(dst_length>0);

    dst[0]= (h->nal_ref_idc<<5) + h->nal_unit_type;

    if(length==0) return 1;

    escape_count= 0;
    for(i=0; i<length; i+=2){
        if(src[i]) continue;
        if(i>0 && src[i-1]==0)
            i--;
        if(i+2<length && src[i+1]==0 && src[i+2]<=3){
            escape_count++;
            i+=2;
        }
    }

    if(escape_count==0){
        if(dst+1 != src)
            memcpy(dst+1, src, length);
        return length + 1;
    }

    if(length + escape_count + 1> dst_length)
        return -1;

    //this should be damn rare (hopefully)

    h->rbsp_buffer= av_fast_realloc(h->rbsp_buffer, &h->rbsp_buffer_size, length + escape_count);
    temp= h->rbsp_buffer;
//printf("encoding esc\n");

    si= 0;
    di= 0;
    while(si < length){
        if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
            temp[di++]= 0; si++;
            temp[di++]= 0; si++;
            temp[di++]= 3;
            temp[di++]= src[si++];
        }
        else
            temp[di++]= src[si++];
    }
    memcpy(dst+1, temp, length+escape_count);

    assert(di == length+escape_count);

    return di + 1;
}

/**
 * write 1,10,100,1000,... for alignment, yes its exactly inverse to mpeg4
 */
static void encode_rbsp_trailing(PutBitContext *pb){
    int length;
    put_bits(pb, 1, 1);
    length= (-put_bits_count(pb))&7;
    if(length) put_bits(pb, length, 0);
}
#endif

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

    tprintf("rbsp trailing %X\n", v);

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

/**
 * idct tranforms the 16 dc values and dequantize 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_resdual
        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 tranforms 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(int chroma_qp_index_offset, int qscale){

    return chroma_qp[clip(qscale + chroma_qp_index_offset, 0, 51)];
}


#if 0
static void h264_diff_dct_c(DCTELEM *block, uint8_t *src1, uint8_t *src2, int stride){
    int i;
    //FIXME try int temp instead of block

    for(i=0; i<4; i++){
        const int d0= src1[0 + i*stride] - src2[0 + i*stride];
        const int d1= src1[1 + i*stride] - src2[1 + i*stride];
        const int d2= src1[2 + i*stride] - src2[2 + i*stride];
        const int d3= src1[3 + i*stride] - src2[3 + i*stride];
        const int z0= d0 + d3;
        const int z3= d0 - d3;
        const int z1= d1 + d2;
        const int z2= d1 - d2;

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

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

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

//FIXME need to check that this doesnt overflow signed 32 bit for low qp, i am not sure, it's very close
//FIXME check that gcc inlines this (and optimizes intra & seperate_dc stuff away)
static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int seperate_dc){
    int i;
    const int * const quant_table= quant_coeff[qscale];
    const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
    const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
    const unsigned int threshold2= (threshold1<<1);
    int last_non_zero;

    if(seperate_dc){
        if(qscale<=18){
            //avoid overflows
            const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
            const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
            const unsigned int dc_threshold2= (dc_threshold1<<1);

            int level= block[0]*quant_coeff[qscale+18][0];
            if(((unsigned)(level+dc_threshold1))>dc_threshold2){
                if(level>0){
                    level= (dc_bias + level)>>(QUANT_SHIFT-2);
                    block[0]= level;
                }else{
                    level= (dc_bias - level)>>(QUANT_SHIFT-2);
                    block[0]= -level;
                }
//                last_non_zero = i;
            }else{
                block[0]=0;
            }
        }else{
            const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
            const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
            const unsigned int dc_threshold2= (dc_threshold1<<1);

            int level= block[0]*quant_table[0];
            if(((unsigned)(level+dc_threshold1))>dc_threshold2){
                if(level>0){
                    level= (dc_bias + level)>>(QUANT_SHIFT+1);
                    block[0]= level;
                }else{
                    level= (dc_bias - level)>>(QUANT_SHIFT+1);
                    block[0]= -level;
                }
//                last_non_zero = i;
            }else{
                block[0]=0;
            }
        }
        last_non_zero= 0;
        i=1;
    }else{
        last_non_zero= -1;
        i=0;
    }

    for(; i<16; i++){
        const int j= scantable[i];
        int level= block[j]*quant_table[j];

//        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
//           || bias-level >= (1<<(QMAT_SHIFT - 3))){
        if(((unsigned)(level+threshold1))>threshold2){
            if(level>0){
                level= (bias + level)>>QUANT_SHIFT;
                block[j]= level;
            }else{
                level= (bias - level)>>QUANT_SHIFT;
                block[j]= -level;
            }
            last_non_zero = i;
        }else{
            block[j]=0;
        }
    }

    return last_non_zero;
}

static void pred4x4_vertical_c(uint8_t *src, uint8_t *topright, int stride){
    const uint32_t a= ((uint32_t*)(src-stride))[0];
    ((uint32_t*)(src+0*stride))[0]= a;
    ((uint32_t*)(src+1*stride))[0]= a;
    ((uint32_t*)(src+2*stride))[0]= a;
    ((uint32_t*)(src+3*stride))[0]= a;
}

static void pred4x4_horizontal_c(uint8_t *src, uint8_t *topright, int stride){
    ((uint32_t*)(src+0*stride))[0]= src[-1+0*stride]*0x01010101;
    ((uint32_t*)(src+1*stride))[0]= src[-1+1*stride]*0x01010101;
    ((uint32_t*)(src+2*stride))[0]= src[-1+2*stride]*0x01010101;
    ((uint32_t*)(src+3*stride))[0]= src[-1+3*stride]*0x01010101;
}

static void pred4x4_dc_c(uint8_t *src, uint8_t *topright, int stride){
    const int dc= (  src[-stride] + src[1-stride] + src[2-stride] + src[3-stride]
                   + src[-1+0*stride] + src[-1+1*stride] + src[-1+2*stride] + src[-1+3*stride] + 4) >>3;

    ((uint32_t*)(src+0*stride))[0]=
    ((uint32_t*)(src+1*stride))[0]=
    ((uint32_t*)(src+2*stride))[0]=
    ((uint32_t*)(src+3*stride))[0]= dc* 0x01010101;
}

static void pred4x4_left_dc_c(uint8_t *src, uint8_t *topright, int stride){
    const int dc= (  src[-1+0*stride] + src[-1+1*stride] + src[-1+2*stride] + src[-1+3*stride] + 2) >>2;

    ((uint32_t*)(src+0*stride))[0]=
    ((uint32_t*)(src+1*stride))[0]=
    ((uint32_t*)(src+2*stride))[0]=
    ((uint32_t*)(src+3*stride))[0]= dc* 0x01010101;
}

static void pred4x4_top_dc_c(uint8_t *src, uint8_t *topright, int stride){
    const int dc= (  src[-stride] + src[1-stride] + src[2-stride] + src[3-stride] + 2) >>2;

    ((uint32_t*)(src+0*stride))[0]=
    ((uint32_t*)(src+1*stride))[0]=
    ((uint32_t*)(src+2*stride))[0]=
    ((uint32_t*)(src+3*stride))[0]= dc* 0x01010101;
}

static void pred4x4_128_dc_c(uint8_t *src, uint8_t *topright, int stride){
    ((uint32_t*)(src+0*stride))[0]=
    ((uint32_t*)(src+1*stride))[0]=
    ((uint32_t*)(src+2*stride))[0]=
    ((uint32_t*)(src+3*stride))[0]= 128U*0x01010101U;
}


#define LOAD_TOP_RIGHT_EDGE\
    const int t4= topright[0];\
    const int t5= topright[1];\
    const int t6= topright[2];\
    const int t7= topright[3];\

#define LOAD_LEFT_EDGE\
    const int l0= src[-1+0*stride];\
    const int l1= src[-1+1*stride];\
    const int l2= src[-1+2*stride];\
    const int l3= src[-1+3*stride];\

#define LOAD_TOP_EDGE\
    const int t0= src[ 0-1*stride];\
    const int t1= src[ 1-1*stride];\
    const int t2= src[ 2-1*stride];\
    const int t3= src[ 3-1*stride];\

static void pred4x4_down_right_c(uint8_t *src, uint8_t *topright, int stride){
    const int lt= src[-1-1*stride];
    LOAD_TOP_EDGE
    LOAD_LEFT_EDGE

    src[0+3*stride]=(l3 + 2*l2 + l1 + 2)>>2;
    src[0+2*stride]=
    src[1+3*stride]=(l2 + 2*l1 + l0 + 2)>>2;
    src[0+1*stride]=
    src[1+2*stride]=
    src[2+3*stride]=(l1 + 2*l0 + lt + 2)>>2;
    src[0+0*stride]=
    src[1+1*stride]=
    src[2+2*stride]=
    src[3+3*stride]=(l0 + 2*lt + t0 + 2)>>2;
    src[1+0*stride]=
    src[2+1*stride]=
    src[3+2*stride]=(lt + 2*t0 + t1 + 2)>>2;
    src[2+0*stride]=
    src[3+1*stride]=(t0 + 2*t1 + t2 + 2)>>2;
    src[3+0*stride]=(t1 + 2*t2 + t3 + 2)>>2;
}

static void pred4x4_down_left_c(uint8_t *src, uint8_t *topright, int stride){
    LOAD_TOP_EDGE
    LOAD_TOP_RIGHT_EDGE
//    LOAD_LEFT_EDGE

    src[0+0*stride]=(t0 + t2 + 2*t1 + 2)>>2;
    src[1+0*stride]=
    src[0+1*stride]=(t1 + t3 + 2*t2 + 2)>>2;
    src[2+0*stride]=
    src[1+1*stride]=
    src[0+2*stride]=(t2 + t4 + 2*t3 + 2)>>2;
    src[3+0*stride]=
    src[2+1*stride]=
    src[1+2*stride]=
    src[0+3*stride]=(t3 + t5 + 2*t4 + 2)>>2;
    src[3+1*stride]=
    src[2+2*stride]=
    src[1+3*stride]=(t4 + t6 + 2*t5 + 2)>>2;
    src[3+2*stride]=
    src[2+3*stride]=(t5 + t7 + 2*t6 + 2)>>2;
    src[3+3*stride]=(t6 + 3*t7 + 2)>>2;
}

static void pred4x4_vertical_right_c(uint8_t *src, uint8_t *topright, int stride){
    const int lt= src[-1-1*stride];
    LOAD_TOP_EDGE
    LOAD_LEFT_EDGE
    const __attribute__((unused)) int unu= l3;

    src[0+0*stride]=
    src[1+2*stride]=(lt + t0 + 1)>>1;
    src[1+0*stride]=
    src[2+2*stride]=(t0 + t1 + 1)>>1;
    src[2+0*stride]=
    src[3+2*stride]=(t1 + t2 + 1)>>1;
    src[3+0*stride]=(t2 + t3 + 1)>>1;
    src[0+1*stride]=
    src[1+3*stride]=(l0 + 2*lt + t0 + 2)>>2;
    src[1+1*stride]=
    src[2+3*stride]=(lt + 2*t0 + t1 + 2)>>2;
    src[2+1*stride]=
    src[3+3*stride]=(t0 + 2*t1 + t2 + 2)>>2;
    src[3+1*stride]=(t1 + 2*t2 + t3 + 2)>>2;
    src[0+2*stride]=(lt + 2*l0 + l1 + 2)>>2;
    src[0+3*stride]=(l0 + 2*l1 + l2 + 2)>>2;
}

static void pred4x4_vertical_left_c(uint8_t *src, uint8_t *topright, int stride){
    LOAD_TOP_EDGE
    LOAD_TOP_RIGHT_EDGE
    const __attribute__((unused)) int unu= t7;

    src[0+0*stride]=(t0 + t1 + 1)>>1;
    src[1+0*stride]=
    src[0+2*stride]=(t1 + t2 + 1)>>1;
    src[2+0*stride]=
    src[1+2*stride]=(t2 + t3 + 1)>>1;
    src[3+0*stride]=
    src[2+2*stride]=(t3 + t4+ 1)>>1;
    src[3+2*stride]=(t4 + t5+ 1)>>1;
    src[0+1*stride]=(t0 + 2*t1 + t2 + 2)>>2;
    src[1+1*stride]=
    src[0+3*stride]=(t1 + 2*t2 + t3 + 2)>>2;
    src[2+1*stride]=
    src[1+3*stride]=(t2 + 2*t3 + t4 + 2)>>2;
    src[3+1*stride]=
    src[2+3*stride]=(t3 + 2*t4 + t5 + 2)>>2;
    src[3+3*stride]=(t4 + 2*t5 + t6 + 2)>>2;
}

static void pred4x4_horizontal_up_c(uint8_t *src, uint8_t *topright, int stride){
    LOAD_LEFT_EDGE

    src[0+0*stride]=(l0 + l1 + 1)>>1;
    src[1+0*stride]=(l0 + 2*l1 + l2 + 2)>>2;
    src[2+0*stride]=
    src[0+1*stride]=(l1 + l2 + 1)>>1;
    src[3+0*stride]=
    src[1+1*stride]=(l1 + 2*l2 + l3 + 2)>>2;
    src[2+1*stride]=
    src[0+2*stride]=(l2 + l3 + 1)>>1;
    src[3+1*stride]=
    src[1+2*stride]=(l2 + 2*l3 + l3 + 2)>>2;
    src[3+2*stride]=
    src[1+3*stride]=
    src[0+3*stride]=
    src[2+2*stride]=
    src[2+3*stride]=
    src[3+3*stride]=l3;
}

static void pred4x4_horizontal_down_c(uint8_t *src, uint8_t *topright, int stride){
    const int lt= src[-1-1*stride];
    LOAD_TOP_EDGE
    LOAD_LEFT_EDGE
    const __attribute__((unused)) int unu= t3;

    src[0+0*stride]=
    src[2+1*stride]=(lt + l0 + 1)>>1;
    src[1+0*stride]=
    src[3+1*stride]=(l0 + 2*lt + t0 + 2)>>2;
    src[2+0*stride]=(lt + 2*t0 + t1 + 2)>>2;
    src[3+0*stride]=(t0 + 2*t1 + t2 + 2)>>2;
    src[0+1*stride]=
    src[2+2*stride]=(l0 + l1 + 1)>>1;
    src[1+1*stride]=
    src[3+2*stride]=(lt + 2*l0 + l1 + 2)>>2;
    src[0+2*stride]=
    src[2+3*stride]=(l1 + l2+ 1)>>1;
    src[1+2*stride]=
    src[3+3*stride]=(l0 + 2*l1 + l2 + 2)>>2;
    src[0+3*stride]=(l2 + l3 + 1)>>1;
    src[1+3*stride]=(l1 + 2*l2 + l3 + 2)>>2;
}

static void pred16x16_vertical_c(uint8_t *src, int stride){
    int i;
    const uint32_t a= ((uint32_t*)(src-stride))[0];
    const uint32_t b= ((uint32_t*)(src-stride))[1];
    const uint32_t c= ((uint32_t*)(src-stride))[2];
    const uint32_t d= ((uint32_t*)(src-stride))[3];

    for(i=0; i<16; i++){
        ((uint32_t*)(src+i*stride))[0]= a;
        ((uint32_t*)(src+i*stride))[1]= b;
        ((uint32_t*)(src+i*stride))[2]= c;
        ((uint32_t*)(src+i*stride))[3]= d;
    }
}

static void pred16x16_horizontal_c(uint8_t *src, int stride){
    int i;

    for(i=0; i<16; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]=
        ((uint32_t*)(src+i*stride))[2]=
        ((uint32_t*)(src+i*stride))[3]= src[-1+i*stride]*0x01010101;
    }
}

static void pred16x16_dc_c(uint8_t *src, int stride){
    int i, dc=0;

    for(i=0;i<16; i++){
        dc+= src[-1+i*stride];
    }

    for(i=0;i<16; i++){
        dc+= src[i-stride];
    }

    dc= 0x01010101*((dc + 16)>>5);

    for(i=0; i<16; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]=
        ((uint32_t*)(src+i*stride))[2]=
        ((uint32_t*)(src+i*stride))[3]= dc;
    }
}

static void pred16x16_left_dc_c(uint8_t *src, int stride){
    int i, dc=0;

    for(i=0;i<16; i++){
        dc+= src[-1+i*stride];
    }

    dc= 0x01010101*((dc + 8)>>4);

    for(i=0; i<16; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]=
        ((uint32_t*)(src+i*stride))[2]=
        ((uint32_t*)(src+i*stride))[3]= dc;
    }
}

static void pred16x16_top_dc_c(uint8_t *src, int stride){
    int i, dc=0;

    for(i=0;i<16; i++){
        dc+= src[i-stride];
    }
    dc= 0x01010101*((dc + 8)>>4);

    for(i=0; i<16; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]=
        ((uint32_t*)(src+i*stride))[2]=
        ((uint32_t*)(src+i*stride))[3]= dc;
    }
}

static void pred16x16_128_dc_c(uint8_t *src, int stride){
    int i;

    for(i=0; i<16; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]=
        ((uint32_t*)(src+i*stride))[2]=
        ((uint32_t*)(src+i*stride))[3]= 0x01010101U*128U;
    }
}

static inline void pred16x16_plane_compat_c(uint8_t *src, int stride, const int svq3){
  int i, j, k;
  int a;
  uint8_t *cm = cropTbl + MAX_NEG_CROP;
  const uint8_t * const src0 = src+7-stride;
  const uint8_t *src1 = src+8*stride-1;
  const uint8_t *src2 = src1-2*stride;      // == src+6*stride-1;
  int H = src0[1] - src0[-1];
  int V = src1[0] - src2[ 0];
  for(k=2; k<=8; ++k) {
    src1 += stride; src2 -= stride;
    H += k*(src0[k] - src0[-k]);
    V += k*(src1[0] - src2[ 0]);
  }
  if(svq3){
    H = ( 5*(H/4) ) / 16;
    V = ( 5*(V/4) ) / 16;

    /* required for 100% accuracy */
    i = H; H = V; V = i;
  }else{
    H = ( 5*H+32 ) >> 6;
    V = ( 5*V+32 ) >> 6;
  }

  a = 16*(src1[0] + src2[16] + 1) - 7*(V+H);
  for(j=16; j>0; --j) {
    int b = a;
    a += V;
    for(i=-16; i<0; i+=4) {
      src[16+i] = cm[ (b    ) >> 5 ];
      src[17+i] = cm[ (b+  H) >> 5 ];
      src[18+i] = cm[ (b+2*H) >> 5 ];
      src[19+i] = cm[ (b+3*H) >> 5 ];
      b += 4*H;
    }
    src += stride;
  }
}

static void pred16x16_plane_c(uint8_t *src, int stride){
    pred16x16_plane_compat_c(src, stride, 0);
}

static void pred8x8_vertical_c(uint8_t *src, int stride){
    int i;
    const uint32_t a= ((uint32_t*)(src-stride))[0];
    const uint32_t b= ((uint32_t*)(src-stride))[1];

    for(i=0; i<8; i++){
        ((uint32_t*)(src+i*stride))[0]= a;
        ((uint32_t*)(src+i*stride))[1]= b;
    }
}

static void pred8x8_horizontal_c(uint8_t *src, int stride){
    int i;

    for(i=0; i<8; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]= src[-1+i*stride]*0x01010101;
    }
}

static void pred8x8_128_dc_c(uint8_t *src, int stride){
    int i;

    for(i=0; i<8; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]= 0x01010101U*128U;
    }
}

static void pred8x8_left_dc_c(uint8_t *src, int stride){
    int i;
    int dc0, dc2;

    dc0=dc2=0;
    for(i=0;i<4; i++){
        dc0+= src[-1+i*stride];
        dc2+= src[-1+(i+4)*stride];
    }
    dc0= 0x01010101*((dc0 + 2)>>2);
    dc2= 0x01010101*((dc2 + 2)>>2);

    for(i=0; i<4; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]= dc0;
    }
    for(i=4; i<8; i++){
        ((uint32_t*)(src+i*stride))[0]=
        ((uint32_t*)(src+i*stride))[1]= dc2;
    }
}

static void pred8x8_top_dc_c(uint8_t *src, int stride){
    int i;
    int dc0, dc1;

    dc0=dc1=0;
    for(i=0;i<4; i++){
        dc0+= src[i-stride];
        dc1+= src[4+i-stride];
    }
    dc0= 0x01010101*((dc0 + 2)>>2);
    dc1= 0x01010101*((dc1 + 2)>>2);

    for(i=0; i<4; i++){
        ((uint32_t*)(src+i*stride))[0]= dc0;
        ((uint32_t*)(src+i*stride))[1]= dc1;
    }
    for(i=4; i<8; i++){
        ((uint32_t*)(src+i*stride))[0]= dc0;
        ((uint32_t*)(src+i*stride))[1]= dc1;
    }
}


static void pred8x8_dc_c(uint8_t *src, int stride){
    int i;
    int dc0, dc1, dc2, dc3;

    dc0=dc1=dc2=0;
    for(i=0;i<4; i++){
        dc0+= src[-1+i*stride] + src[i-stride];
        dc1+= src[4+i-stride];
        dc2+= src[-1+(i+4)*stride];
    }
    dc3= 0x01010101*((dc1 + dc2 + 4)>>3);
    dc0= 0x01010101*((dc0 + 4)>>3);
    dc1= 0x01010101*((dc1 + 2)>>2);
    dc2= 0x01010101*((dc2 + 2)>>2);

    for(i=0; i<4; i++){
        ((uint32_t*)(src+i*stride))[0]= dc0;
        ((uint32_t*)(src+i*stride))[1]= dc1;
    }
    for(i=4; i<8; i++){
        ((uint32_t*)(src+i*stride))[0]= dc2;
        ((uint32_t*)(src+i*stride))[1]= dc3;
    }
}

static void pred8x8_plane_c(uint8_t *src, int stride){
  int j, k;
  int a;
  uint8_t *cm = cropTbl + MAX_NEG_CROP;
  const uint8_t * const src0 = src+3-stride;
  const uint8_t *src1 = src+4*stride-1;
  const uint8_t *src2 = src1-2*stride;      // == src+2*stride-1;
  int H = src0[1] - src0[-1];
  int V = src1[0] - src2[ 0];
  for(k=2; k<=4; ++k) {
    src1 += stride; src2 -= stride;
    H += k*(src0[k] - src0[-k]);
    V += k*(src1[0] - src2[ 0]);
  }
  H = ( 17*H+16 ) >> 5;
  V = ( 17*V+16 ) >> 5;

  a = 16*(src1[0] + src2[8]+1) - 3*(V+H);
  for(j=8; j>0; --j) {
    int b = a;
    a += V;
    src[0] = cm[ (b    ) >> 5 ];
    src[1] = cm[ (b+  H) >> 5 ];
    src[2] = cm[ (b+2*H) >> 5 ];
    src[3] = cm[ (b+3*H) >> 5 ];
    src[4] = cm[ (b+4*H) >> 5 ];
    src[5] = cm[ (b+5*H) >> 5 ];
    src[6] = cm[ (b+6*H) >> 5 ];
    src[7] = cm[ (b+7*H) >> 5 ];
    src += stride;
  }
}

#define SRC(x,y) src[(x)+(y)*stride]
#define PL(y) \
    const int l##y = (SRC(-1,y-1) + 2*SRC(-1,y) + SRC(-1,y+1) + 2) >> 2;
#define PREDICT_8x8_LOAD_LEFT \
    const int l0 = ((has_topleft ? SRC(-1,-1) : SRC(-1,0)) \
                     + 2*SRC(-1,0) + SRC(-1,1) + 2) >> 2; \
    PL(1) PL(2) PL(3) PL(4) PL(5) PL(6) \
    const int l7 attribute_unused = (SRC(-1,6) + 3*SRC(-1,7) + 2) >> 2

#define PT(x) \
    const int t##x = (SRC(x-1,-1) + 2*SRC(x,-1) + SRC(x+1,-1) + 2) >> 2;
#define PREDICT_8x8_LOAD_TOP \
    const int t0 = ((has_topleft ? SRC(-1,-1) : SRC(0,-1)) \
                     + 2*SRC(0,-1) + SRC(1,-1) + 2) >> 2; \
    PT(1) PT(2) PT(3) PT(4) PT(5) PT(6) \
    const int t7 attribute_unused = ((has_topright ? SRC(8,-1) : SRC(7,-1)) \
                     + 2*SRC(7,-1) + SRC(6,-1) + 2) >> 2

#define PTR(x) \
    t##x = (SRC(x-1,-1) + 2*SRC(x,-1) + SRC(x+1,-1) + 2) >> 2;
#define PREDICT_8x8_LOAD_TOPRIGHT \
    int t8, t9, t10, t11, t12, t13, t14, t15; \
    if(has_topright) { \
        PTR(8) PTR(9) PTR(10) PTR(11) PTR(12) PTR(13) PTR(14) \
        t15 = (SRC(14,-1) + 3*SRC(15,-1) + 2) >> 2; \
    } else t8=t9=t10=t11=t12=t13=t14=t15= SRC(7,-1);

#define PREDICT_8x8_LOAD_TOPLEFT \
    const int lt = (SRC(-1,0) + 2*SRC(-1,-1) + SRC(0,-1) + 2) >> 2

#define PREDICT_8x8_DC(v) \
    int y; \
    for( y = 0; y < 8; y++ ) { \
        ((uint32_t*)src)[0] = \
        ((uint32_t*)src)[1] = v; \
        src += stride; \
    }

static void pred8x8l_128_dc_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_DC(0x80808080);
}
static void pred8x8l_left_dc_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_LEFT;
    const uint32_t dc = ((l0+l1+l2+l3+l4+l5+l6+l7+4) >> 3) * 0x01010101;
    PREDICT_8x8_DC(dc);
}
static void pred8x8l_top_dc_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_TOP;
    const uint32_t dc = ((t0+t1+t2+t3+t4+t5+t6+t7+4) >> 3) * 0x01010101;
    PREDICT_8x8_DC(dc);
}
static void pred8x8l_dc_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_LEFT;
    PREDICT_8x8_LOAD_TOP;
    const uint32_t dc = ((l0+l1+l2+l3+l4+l5+l6+l7
                         +t0+t1+t2+t3+t4+t5+t6+t7+8) >> 4) * 0x01010101;
    PREDICT_8x8_DC(dc);
}
static void pred8x8l_horizontal_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_LEFT;
#define ROW(y) ((uint32_t*)(src+y*stride))[0] =\
               ((uint32_t*)(src+y*stride))[1] = 0x01010101 * l##y
    ROW(0); ROW(1); ROW(2); ROW(3); ROW(4); ROW(5); ROW(6); ROW(7);
#undef ROW
}
static void pred8x8l_vertical_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    int y;
    PREDICT_8x8_LOAD_TOP;
    src[0] = t0;
    src[1] = t1;
    src[2] = t2;
    src[3] = t3;
    src[4] = t4;
    src[5] = t5;
    src[6] = t6;
    src[7] = t7;
    for( y = 1; y < 8; y++ )
        *(uint64_t*)(src+y*stride) = *(uint64_t*)src;
}
static void pred8x8l_down_left_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_TOP;
    PREDICT_8x8_LOAD_TOPRIGHT;
    SRC(0,0)= (t0 + 2*t1 + t2 + 2) >> 2;
    SRC(0,1)=SRC(1,0)= (t1 + 2*t2 + t3 + 2) >> 2;
    SRC(0,2)=SRC(1,1)=SRC(2,0)= (t2 + 2*t3 + t4 + 2) >> 2;
    SRC(0,3)=SRC(1,2)=SRC(2,1)=SRC(3,0)= (t3 + 2*t4 + t5 + 2) >> 2;
    SRC(0,4)=SRC(1,3)=SRC(2,2)=SRC(3,1)=SRC(4,0)= (t4 + 2*t5 + t6 + 2) >> 2;
    SRC(0,5)=SRC(1,4)=SRC(2,3)=SRC(3,2)=SRC(4,1)=SRC(5,0)= (t5 + 2*t6 + t7 + 2) >> 2;
    SRC(0,6)=SRC(1,5)=SRC(2,4)=SRC(3,3)=SRC(4,2)=SRC(5,1)=SRC(6,0)= (t6 + 2*t7 + t8 + 2) >> 2;
    SRC(0,7)=SRC(1,6)=SRC(2,5)=SRC(3,4)=SRC(4,3)=SRC(5,2)=SRC(6,1)=SRC(7,0)= (t7 + 2*t8 + t9 + 2) >> 2;
    SRC(1,7)=SRC(2,6)=SRC(3,5)=SRC(4,4)=SRC(5,3)=SRC(6,2)=SRC(7,1)= (t8 + 2*t9 + t10 + 2) >> 2;
    SRC(2,7)=SRC(3,6)=SRC(4,5)=SRC(5,4)=SRC(6,3)=SRC(7,2)= (t9 + 2*t10 + t11 + 2) >> 2;
    SRC(3,7)=SRC(4,6)=SRC(5,5)=SRC(6,4)=SRC(7,3)= (t10 + 2*t11 + t12 + 2) >> 2;
    SRC(4,7)=SRC(5,6)=SRC(6,5)=SRC(7,4)= (t11 + 2*t12 + t13 + 2) >> 2;
    SRC(5,7)=SRC(6,6)=SRC(7,5)= (t12 + 2*t13 + t14 + 2) >> 2;
    SRC(6,7)=SRC(7,6)= (t13 + 2*t14 + t15 + 2) >> 2;
    SRC(7,7)= (t14 + 3*t15 + 2) >> 2;
}
static void pred8x8l_down_right_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_TOP;
    PREDICT_8x8_LOAD_LEFT;
    PREDICT_8x8_LOAD_TOPLEFT;
    SRC(0,7)= (l7 + 2*l6 + l5 + 2) >> 2;
    SRC(0,6)=SRC(1,7)= (l6 + 2*l5 + l4 + 2) >> 2;
    SRC(0,5)=SRC(1,6)=SRC(2,7)= (l5 + 2*l4 + l3 + 2) >> 2;
    SRC(0,4)=SRC(1,5)=SRC(2,6)=SRC(3,7)= (l4 + 2*l3 + l2 + 2) >> 2;
    SRC(0,3)=SRC(1,4)=SRC(2,5)=SRC(3,6)=SRC(4,7)= (l3 + 2*l2 + l1 + 2) >> 2;
    SRC(0,2)=SRC(1,3)=SRC(2,4)=SRC(3,5)=SRC(4,6)=SRC(5,7)= (l2 + 2*l1 + l0 + 2) >> 2;
    SRC(0,1)=SRC(1,2)=SRC(2,3)=SRC(3,4)=SRC(4,5)=SRC(5,6)=SRC(6,7)= (l1 + 2*l0 + lt + 2) >> 2;
    SRC(0,0)=SRC(1,1)=SRC(2,2)=SRC(3,3)=SRC(4,4)=SRC(5,5)=SRC(6,6)=SRC(7,7)= (l0 + 2*lt + t0 + 2) >> 2;
    SRC(1,0)=SRC(2,1)=SRC(3,2)=SRC(4,3)=SRC(5,4)=SRC(6,5)=SRC(7,6)= (lt + 2*t0 + t1 + 2) >> 2;
    SRC(2,0)=SRC(3,1)=SRC(4,2)=SRC(5,3)=SRC(6,4)=SRC(7,5)= (t0 + 2*t1 + t2 + 2) >> 2;
    SRC(3,0)=SRC(4,1)=SRC(5,2)=SRC(6,3)=SRC(7,4)= (t1 + 2*t2 + t3 + 2) >> 2;
    SRC(4,0)=SRC(5,1)=SRC(6,2)=SRC(7,3)= (t2 + 2*t3 + t4 + 2) >> 2;
    SRC(5,0)=SRC(6,1)=SRC(7,2)= (t3 + 2*t4 + t5 + 2) >> 2;
    SRC(6,0)=SRC(7,1)= (t4 + 2*t5 + t6 + 2) >> 2;
    SRC(7,0)= (t5 + 2*t6 + t7 + 2) >> 2;

}
static void pred8x8l_vertical_right_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_TOP;
    PREDICT_8x8_LOAD_LEFT;
    PREDICT_8x8_LOAD_TOPLEFT;
    SRC(0,6)= (l5 + 2*l4 + l3 + 2) >> 2;
    SRC(0,7)= (l6 + 2*l5 + l4 + 2) >> 2;
    SRC(0,4)=SRC(1,6)= (l3 + 2*l2 + l1 + 2) >> 2;
    SRC(0,5)=SRC(1,7)= (l4 + 2*l3 + l2 + 2) >> 2;
    SRC(0,2)=SRC(1,4)=SRC(2,6)= (l1 + 2*l0 + lt + 2) >> 2;
    SRC(0,3)=SRC(1,5)=SRC(2,7)= (l2 + 2*l1 + l0 + 2) >> 2;
    SRC(0,1)=SRC(1,3)=SRC(2,5)=SRC(3,7)= (l0 + 2*lt + t0 + 2) >> 2;
    SRC(0,0)=SRC(1,2)=SRC(2,4)=SRC(3,6)= (lt + t0 + 1) >> 1;
    SRC(1,1)=SRC(2,3)=SRC(3,5)=SRC(4,7)= (lt + 2*t0 + t1 + 2) >> 2;
    SRC(1,0)=SRC(2,2)=SRC(3,4)=SRC(4,6)= (t0 + t1 + 1) >> 1;
    SRC(2,1)=SRC(3,3)=SRC(4,5)=SRC(5,7)= (t0 + 2*t1 + t2 + 2) >> 2;
    SRC(2,0)=SRC(3,2)=SRC(4,4)=SRC(5,6)= (t1 + t2 + 1) >> 1;
    SRC(3,1)=SRC(4,3)=SRC(5,5)=SRC(6,7)= (t1 + 2*t2 + t3 + 2) >> 2;
    SRC(3,0)=SRC(4,2)=SRC(5,4)=SRC(6,6)= (t2 + t3 + 1) >> 1;
    SRC(4,1)=SRC(5,3)=SRC(6,5)=SRC(7,7)= (t2 + 2*t3 + t4 + 2) >> 2;
    SRC(4,0)=SRC(5,2)=SRC(6,4)=SRC(7,6)= (t3 + t4 + 1) >> 1;
    SRC(5,1)=SRC(6,3)=SRC(7,5)= (t3 + 2*t4 + t5 + 2) >> 2;
    SRC(5,0)=SRC(6,2)=SRC(7,4)= (t4 + t5 + 1) >> 1;
    SRC(6,1)=SRC(7,3)= (t4 + 2*t5 + t6 + 2) >> 2;
    SRC(6,0)=SRC(7,2)= (t5 + t6 + 1) >> 1;
    SRC(7,1)= (t5 + 2*t6 + t7 + 2) >> 2;
    SRC(7,0)= (t6 + t7 + 1) >> 1;
}
static void pred8x8l_horizontal_down_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_TOP;
    PREDICT_8x8_LOAD_LEFT;
    PREDICT_8x8_LOAD_TOPLEFT;
    SRC(0,7)= (l6 + l7 + 1) >> 1;
    SRC(1,7)= (l5 + 2*l6 + l7 + 2) >> 2;
    SRC(0,6)=SRC(2,7)= (l5 + l6 + 1) >> 1;
    SRC(1,6)=SRC(3,7)= (l4 + 2*l5 + l6 + 2) >> 2;
    SRC(0,5)=SRC(2,6)=SRC(4,7)= (l4 + l5 + 1) >> 1;
    SRC(1,5)=SRC(3,6)=SRC(5,7)= (l3 + 2*l4 + l5 + 2) >> 2;
    SRC(0,4)=SRC(2,5)=SRC(4,6)=SRC(6,7)= (l3 + l4 + 1) >> 1;
    SRC(1,4)=SRC(3,5)=SRC(5,6)=SRC(7,7)= (l2 + 2*l3 + l4 + 2) >> 2;
    SRC(0,3)=SRC(2,4)=SRC(4,5)=SRC(6,6)= (l2 + l3 + 1) >> 1;
    SRC(1,3)=SRC(3,4)=SRC(5,5)=SRC(7,6)= (l1 + 2*l2 + l3 + 2) >> 2;
    SRC(0,2)=SRC(2,3)=SRC(4,4)=SRC(6,5)= (l1 + l2 + 1) >> 1;
    SRC(1,2)=SRC(3,3)=SRC(5,4)=SRC(7,5)= (l0 + 2*l1 + l2 + 2) >> 2;
    SRC(0,1)=SRC(2,2)=SRC(4,3)=SRC(6,4)= (l0 + l1 + 1) >> 1;
    SRC(1,1)=SRC(3,2)=SRC(5,3)=SRC(7,4)= (lt + 2*l0 + l1 + 2) >> 2;
    SRC(0,0)=SRC(2,1)=SRC(4,2)=SRC(6,3)= (lt + l0 + 1) >> 1;
    SRC(1,0)=SRC(3,1)=SRC(5,2)=SRC(7,3)= (l0 + 2*lt + t0 + 2) >> 2;
    SRC(2,0)=SRC(4,1)=SRC(6,2)= (t1 + 2*t0 + lt + 2) >> 2;
    SRC(3,0)=SRC(5,1)=SRC(7,2)= (t2 + 2*t1 + t0 + 2) >> 2;
    SRC(4,0)=SRC(6,1)= (t3 + 2*t2 + t1 + 2) >> 2;
    SRC(5,0)=SRC(7,1)= (t4 + 2*t3 + t2 + 2) >> 2;
    SRC(6,0)= (t5 + 2*t4 + t3 + 2) >> 2;
    SRC(7,0)= (t6 + 2*t5 + t4 + 2) >> 2;
}
static void pred8x8l_vertical_left_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_TOP;
    PREDICT_8x8_LOAD_TOPRIGHT;
    SRC(0,0)= (t0 + t1 + 1) >> 1;
    SRC(0,1)= (t0 + 2*t1 + t2 + 2) >> 2;
    SRC(0,2)=SRC(1,0)= (t1 + t2 + 1) >> 1;
    SRC(0,3)=SRC(1,1)= (t1 + 2*t2 + t3 + 2) >> 2;
    SRC(0,4)=SRC(1,2)=SRC(2,0)= (t2 + t3 + 1) >> 1;
    SRC(0,5)=SRC(1,3)=SRC(2,1)= (t2 + 2*t3 + t4 + 2) >> 2;
    SRC(0,6)=SRC(1,4)=SRC(2,2)=SRC(3,0)= (t3 + t4 + 1) >> 1;
    SRC(0,7)=SRC(1,5)=SRC(2,3)=SRC(3,1)= (t3 + 2*t4 + t5 + 2) >> 2;
    SRC(1,6)=SRC(2,4)=SRC(3,2)=SRC(4,0)= (t4 + t5 + 1) >> 1;
    SRC(1,7)=SRC(2,5)=SRC(3,3)=SRC(4,1)= (t4 + 2*t5 + t6 + 2) >> 2;
    SRC(2,6)=SRC(3,4)=SRC(4,2)=SRC(5,0)= (t5 + t6 + 1) >> 1;
    SRC(2,7)=SRC(3,5)=SRC(4,3)=SRC(5,1)= (t5 + 2*t6 + t7 + 2) >> 2;
    SRC(3,6)=SRC(4,4)=SRC(5,2)=SRC(6,0)= (t6 + t7 + 1) >> 1;
    SRC(3,7)=SRC(4,5)=SRC(5,3)=SRC(6,1)= (t6 + 2*t7 + t8 + 2) >> 2;
    SRC(4,6)=SRC(5,4)=SRC(6,2)=SRC(7,0)= (t7 + t8 + 1) >> 1;
    SRC(4,7)=SRC(5,5)=SRC(6,3)=SRC(7,1)= (t7 + 2*t8 + t9 + 2) >> 2;
    SRC(5,6)=SRC(6,4)=SRC(7,2)= (t8 + t9 + 1) >> 1;
    SRC(5,7)=SRC(6,5)=SRC(7,3)= (t8 + 2*t9 + t10 + 2) >> 2;
    SRC(6,6)=SRC(7,4)= (t9 + t10 + 1) >> 1;
    SRC(6,7)=SRC(7,5)= (t9 + 2*t10 + t11 + 2) >> 2;
    SRC(7,6)= (t10 + t11 + 1) >> 1;
    SRC(7,7)= (t10 + 2*t11 + t12 + 2) >> 2;
}
static void pred8x8l_horizontal_up_c(uint8_t *src, int has_topleft, int has_topright, int stride)
{
    PREDICT_8x8_LOAD_LEFT;
    SRC(0,0)= (l0 + l1 + 1) >> 1;
    SRC(1,0)= (l0 + 2*l1 + l2 + 2) >> 2;
    SRC(0,1)=SRC(2,0)= (l1 + l2 + 1) >> 1;
    SRC(1,1)=SRC(3,0)= (l1 + 2*l2 + l3 + 2) >> 2;
    SRC(0,2)=SRC(2,1)=SRC(4,0)= (l2 + l3 + 1) >> 1;
    SRC(1,2)=SRC(3,1)=SRC(5,0)= (l2 + 2*l3 + l4 + 2) >> 2;
    SRC(0,3)=SRC(2,2)=SRC(4,1)=SRC(6,0)= (l3 + l4 + 1) >> 1;
    SRC(1,3)=SRC(3,2)=SRC(5,1)=SRC(7,0)= (l3 + 2*l4 + l5 + 2) >> 2;
    SRC(0,4)=SRC(2,3)=SRC(4,2)=SRC(6,1)= (l4 + l5 + 1) >> 1;
    SRC(1,4)=SRC(3,3)=SRC(5,2)=SRC(7,1)= (l4 + 2*l5 + l6 + 2) >> 2;
    SRC(0,5)=SRC(2,4)=SRC(4,3)=SRC(6,2)= (l5 + l6 + 1) >> 1;
    SRC(1,5)=SRC(3,4)=SRC(5,3)=SRC(7,2)= (l5 + 2*l6 + l7 + 2) >> 2;
    SRC(0,6)=SRC(2,5)=SRC(4,4)=SRC(6,3)= (l6 + l7 + 1) >> 1;
    SRC(1,6)=SRC(3,5)=SRC(5,4)=SRC(7,3)= (l6 + 3*l7 + 2) >> 2;
    SRC(0,7)=SRC(1,7)=SRC(2,6)=SRC(2,7)=SRC(3,6)=
    SRC(3,7)=SRC(4,5)=SRC(4,6)=SRC(4,7)=SRC(5,5)=
    SRC(5,6)=SRC(5,7)=SRC(6,4)=SRC(6,5)=SRC(6,6)=
    SRC(6,7)=SRC(7,4)=SRC(7,5)=SRC(7,6)=SRC(7,7)= l7;
}
#undef PREDICT_8x8_LOAD_LEFT
#undef PREDICT_8x8_LOAD_TOP
#undef PREDICT_8x8_LOAD_TOPLEFT
#undef PREDICT_8x8_LOAD_TOPRIGHT
#undef PREDICT_8x8_DC
#undef PTR
#undef PT
#undef PL
#undef SRC

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;
    const 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)*s->linesize;
    uint8_t * src_cb= pic->data[1] + (mx>>3) + (my>>3)*s->uvlinesize;
    uint8_t * src_cr= pic->data[2] + (mx>>3) + (my>>3)*s->uvlinesize;
    int extra_width= (s->flags&CODEC_FLAG_EMU_EDGE) ? 0 : 16; //FIXME increase edge?, IMHO not worth it
    int extra_height= extra_width;
    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;

    if(!pic->data[0])
        return;

    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*s->linesize, s->linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
            src_y= s->edge_emu_buffer + 2 + 2*s->linesize;
        emu=1;
    }

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

    if(s->flags&CODEC_FLAG_GRAY) return;

    if(emu){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_cb, s->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, s->uvlinesize, chroma_height, mx&7, my&7);

    if(emu){
        ff_emulated_edge_mc(s->edge_emu_buffer, src_cr, s->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, s->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*s->  linesize;
    dest_cb +=   x_offset +   y_offset*s->uvlinesize;
    dest_cr +=   x_offset +   y_offset*s->uvlinesize;
    x_offset += 8*s->mb_x;
    y_offset += 8*s->mb_y;

    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*s->  linesize;
    dest_cb +=   x_offset +   y_offset*s->uvlinesize;
    dest_cr +=   x_offset +   y_offset*s->uvlinesize;
    x_offset += 8*s->mb_x;
    y_offset += 8*s->mb_y;

    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 = tmp_cb + 8*s->uvlinesize;
        uint8_t *tmp_y  = tmp_cr + 8*s->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,  s->  linesize, 5, weight0, weight1, 0);
            chroma_weight_avg(dest_cb, tmp_cb, s->uvlinesize, 5, weight0, weight1, 0);
            chroma_weight_avg(dest_cr, tmp_cr, s->uvlinesize, 5, weight0, weight1, 0);
        }else{
            luma_weight_avg(dest_y, tmp_y, s->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, s->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, s->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, s->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, s->uvlinesize, h->chroma_log2_weight_denom,
                             h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
            chroma_weight_op(dest_cr, s->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 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= s->mb_x + s->mb_y*s->mb_stride;
    const int mb_type= s->current_picture.mb_type[mb_xy];

    assert(IS_INTER(mb_type));

    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*s->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*s->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*s->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*s->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));
                }
            }
        }
    }
}

static void decode_init_vlc(H264Context *h){
    static int done = 0;

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

        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, 1);

        for(i=0; i<4; 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, 1);
        }

        for(i=0; i<3; i++){
            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, 1);
        }
        for(i=0; i<15; i++){
            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, 1);
        }

        for(i=0; i<6; i++){
            init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
                     &run_len [i][0], 1, 1,
                     &run_bits[i][0], 1, 1, 1);
        }
        init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
                 &run_len [6][0], 1, 1,
                 &run_bits[6][0], 1, 1, 1);
    }
}

/**
 * Sets the intra prediction function pointers.
 */
static void init_pred_ptrs(H264Context *h){
//    MpegEncContext * const s = &h->s;

    h->pred4x4[VERT_PRED           ]= pred4x4_vertical_c;
    h->pred4x4[HOR_PRED            ]= pred4x4_horizontal_c;
    h->pred4x4[DC_PRED             ]= pred4x4_dc_c;
    h->pred4x4[DIAG_DOWN_LEFT_PRED ]= pred4x4_down_left_c;
    h->pred4x4[DIAG_DOWN_RIGHT_PRED]= pred4x4_down_right_c;
    h->pred4x4[VERT_RIGHT_PRED     ]= pred4x4_vertical_right_c;
    h->pred4x4[HOR_DOWN_PRED       ]= pred4x4_horizontal_down_c;
    h->pred4x4[VERT_LEFT_PRED      ]= pred4x4_vertical_left_c;
    h->pred4x4[HOR_UP_PRED         ]= pred4x4_horizontal_up_c;
    h->pred4x4[LEFT_DC_PRED        ]= pred4x4_left_dc_c;
    h->pred4x4[TOP_DC_PRED         ]= pred4x4_top_dc_c;
    h->pred4x4[DC_128_PRED         ]= pred4x4_128_dc_c;

    h->pred8x8l[VERT_PRED           ]= pred8x8l_vertical_c;
    h->pred8x8l[HOR_PRED            ]= pred8x8l_horizontal_c;
    h->pred8x8l[DC_PRED             ]= pred8x8l_dc_c;
    h->pred8x8l[DIAG_DOWN_LEFT_PRED ]= pred8x8l_down_left_c;
    h->pred8x8l[DIAG_DOWN_RIGHT_PRED]= pred8x8l_down_right_c;
    h->pred8x8l[VERT_RIGHT_PRED     ]= pred8x8l_vertical_right_c;
    h->pred8x8l[HOR_DOWN_PRED       ]= pred8x8l_horizontal_down_c;
    h->pred8x8l[VERT_LEFT_PRED      ]= pred8x8l_vertical_left_c;
    h->pred8x8l[HOR_UP_PRED         ]= pred8x8l_horizontal_up_c;
    h->pred8x8l[LEFT_DC_PRED        ]= pred8x8l_left_dc_c;
    h->pred8x8l[TOP_DC_PRED         ]= pred8x8l_top_dc_c;
    h->pred8x8l[DC_128_PRED         ]= pred8x8l_128_dc_c;

    h->pred8x8[DC_PRED8x8     ]= pred8x8_dc_c;
    h->pred8x8[VERT_PRED8x8   ]= pred8x8_vertical_c;
    h->pred8x8[HOR_PRED8x8    ]= pred8x8_horizontal_c;
    h->pred8x8[PLANE_PRED8x8  ]= pred8x8_plane_c;
    h->pred8x8[LEFT_DC_PRED8x8]= pred8x8_left_dc_c;
    h->pred8x8[TOP_DC_PRED8x8 ]= pred8x8_top_dc_c;
    h->pred8x8[DC_128_PRED8x8 ]= pred8x8_128_dc_c;

    h->pred16x16[DC_PRED8x8     ]= pred16x16_dc_c;
    h->pred16x16[VERT_PRED8x8   ]= pred16x16_vertical_c;
    h->pred16x16[HOR_PRED8x8    ]= pred16x16_horizontal_c;
    h->pred16x16[PLANE_PRED8x8  ]= pred16x16_plane_c;
    h->pred16x16[LEFT_DC_PRED8x8]= pred16x16_left_dc_c;
    h->pred16x16[TOP_DC_PRED8x8 ]= pred16x16_top_dc_c;
    h->pred16x16[DC_128_PRED8x8 ]= pred16x16_128_dc_c;
}

static void free_tables(H264Context *h){
    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);
    av_freep(&h->top_borders[1]);
    av_freep(&h->top_borders[0]);
    h->slice_table= NULL;

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

    av_freep(&h->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 * sizeof(uint8_t))
    CHECKED_ALLOCZ(h->top_borders[0]    , s->mb_width * (16+8+8) * sizeof(uint8_t))
    CHECKED_ALLOCZ(h->top_borders[1]    , s->mb_width * (16+8+8) * sizeof(uint8_t))
    CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))

    if( h->pps.cabac ) {
        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  * sizeof(uint8_t));
    h->slice_table= h->slice_table_base + s->mb_stride + 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;
}

static 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;

    init_pred_ptrs(h);

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

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

static 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->low_delay= 1;
    avctx->pix_fmt= PIX_FMT_YUV420P;

    decode_init_vlc(h);

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

    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);

    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? */
    if(!s->obmc_scratchpad)
        s->obmc_scratchpad = av_malloc(16*s->linesize + 2*8*s->uvlinesize);

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

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

    // 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[0]= h->top_borders[0][s->mb_x][15];
    for(i=1; i<17; i++){
        h->left_border[i]= src_y[15+i*  linesize];
    }

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

    if(!(s->flags&CODEC_FLAG_GRAY)){
        h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
        h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
        for(i=1; i<9; i++){
            h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
            h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
        }
        *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
        *(uint64_t*)(h->top_borders[0][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){
    MpegEncContext * const s = &h->s;
    int temp8, i;
    uint64_t temp64;
    int deblock_left = (s->mb_x > 0);
    int deblock_top  = (s->mb_y > 0);

    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<17; i++){
            XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
        }
    }

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

    if(!(s->flags&CODEC_FLAG_GRAY)){
        if(deblock_left){
            for(i = !deblock_top; i<9; i++){
                XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
                XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
            }
        }
        if(deblock_top){
            XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
            XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
        }
    }
}

static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
    MpegEncContext * const s = &h->s;
    int i;

    src_y  -= 2 *   linesize;
    src_cb -= 2 * uvlinesize;
    src_cr -= 2 * uvlinesize;

    // 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[0]= h->top_borders[0][s->mb_x][15];
    h->left_border[1]= h->top_borders[1][s->mb_x][15];
    for(i=2; i<34; i++){
        h->left_border[i]= src_y[15+i*  linesize];
    }

    *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
    *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
    *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
    *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);

    if(!(s->flags&CODEC_FLAG_GRAY)){
        h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
        h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
        h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
        h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
        for(i=2; i<18; i++){
            h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
            h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
        }
        *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
        *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
        *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
        *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
    }
}

static inline void xchg_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg){
    MpegEncContext * const s = &h->s;
    int temp8, i;
    uint64_t temp64;
    int deblock_left = (s->mb_x > 0);
    int deblock_top  = (s->mb_y > 0);

    tprintf("xchg_pair_border: src_y:%p src_cb:%p src_cr:%p ls:%d uvls:%d\n", src_y, src_cb, src_cr, linesize, uvlinesize);

    src_y  -= 2 *   linesize + 1;
    src_cb -= 2 * uvlinesize + 1;
    src_cr -= 2 * uvlinesize + 1;

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

    if(deblock_left){
        for(i = (!deblock_top)<<1; i<34; i++){
            XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
        }
    }

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

    if(!(s->flags&CODEC_FLAG_GRAY)){
        if(deblock_left){
            for(i = (!deblock_top) << 1; i<18; i++){
                XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
                XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
            }
        }
        if(deblock_top){
            XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
            XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
            XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
            XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
        }
    }
}

static void hl_decode_mb(H264Context *h){
    MpegEncContext * const s = &h->s;
    const int mb_x= s->mb_x;
    const int mb_y= s->mb_y;
    const int mb_xy= mb_x + mb_y*s->mb_stride;
    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 unsigned int bottom = mb_y & 1;
    const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass);
    void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
    void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);

    if(!s->decode)
        return;

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

    if (h->mb_field_decoding_flag) {
        linesize = s->linesize * 2;
        uvlinesize = s->uvlinesize * 2;
        block_offset = &h->block_offset[24];
        if(mb_y&1){ //FIXME move out of this func?
            dest_y -= s->linesize*15;
            dest_cb-= s->uvlinesize*7;
            dest_cr-= s->uvlinesize*7;
        }
    } else {
        linesize = s->linesize;
        uvlinesize = s->uvlinesize;
//        dct_offset = s->linesize * 16;
    }

    if(transform_bypass){
        idct_dc_add =
        idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
    }else if(IS_8x8DCT(mb_type)){
        idct_dc_add = s->dsp.h264_idct8_dc_add;
        idct_add = s->dsp.h264_idct8_add;
    }else{
        idct_dc_add = s->dsp.h264_idct_dc_add;
        idct_add = s->dsp.h264_idct_add;
    }

    if (IS_INTRA_PCM(mb_type)) {
        unsigned int x, y;

        // The pixels are stored in h->mb array in the same order as levels,
        // copy them in output in the correct order.
        for(i=0; i<16; i++) {
            for (y=0; y<4; y++) {
                for (x=0; x<4; x++) {
                    *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
                }
            }
        }
        for(i=16; i<16+4; i++) {
            for (y=0; y<4; y++) {
                for (x=0; x<4; x++) {
                    *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
                }
            }
        }
        for(i=20; i<20+4; i++) {
            for (y=0; y<4; y++) {
                for (x=0; x<4; x++) {
                    *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
                }
            }
        }
    } else {
        if(IS_INTRA(mb_type)){
            if(h->deblocking_filter) {
                if (h->mb_aff_frame) {
                    if (!bottom)
                        xchg_pair_border(h, dest_y, dest_cb, dest_cr, s->linesize, s->uvlinesize, 1);
                } else {
                    xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1);
                }
            }

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

            if(IS_INTRA4x4(mb_type)){
                if(!s->encoding){
                    if(IS_8x8DCT(mb_type)){
                        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] ];
                            const int nnz = h->non_zero_count_cache[ scan8[i] ];
                            h->pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
                                                   (h->topright_samples_available<<(i+1))&0x8000, 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
                    for(i=0; i<16; i++){
                        uint8_t * const ptr= dest_y + block_offset[i];
                        uint8_t *topright;
                        const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
                        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->pred4x4[ dir ](ptr, topright, linesize);
                        nnz = h->non_zero_count_cache[ scan8[i] ];
                        if(nnz){
                            if(s->codec_id == CODEC_ID_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->pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
                if(s->codec_id == CODEC_ID_H264){
                    if(!transform_bypass)
                        h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[IS_INTRA(mb_type) ? 0:3][s->qscale][0]);
                }else
                    svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
            }
            if(h->deblocking_filter) {
                if (h->mb_aff_frame) {
                    if (bottom) {
                        uint8_t *pair_dest_y  = s->current_picture.data[0] + ((mb_y-1) * 16* s->linesize  ) + mb_x * 16;
                        uint8_t *pair_dest_cb = s->current_picture.data[1] + ((mb_y-1) * 8 * s->uvlinesize) + mb_x * 8;
                        uint8_t *pair_dest_cr = s->current_picture.data[2] + ((mb_y-1) * 8 * s->uvlinesize) + mb_x * 8;
                        s->mb_y--;
                        xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
                        s->mb_y++;
                    }
                } else {
                    xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
                }
            }
        }else if(s->codec_id == CODEC_ID_H264){
            hl_motion(h, dest_y, dest_cb, dest_cr,
                      s->dsp.put_h264_qpel_pixels_tab, s->dsp.put_h264_chroma_pixels_tab,
                      s->dsp.avg_h264_qpel_pixels_tab, s->dsp.avg_h264_chroma_pixels_tab,
                      s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
        }


        if(!IS_INTRA4x4(mb_type)){
            if(s->codec_id == CODEC_ID_H264){
                if(IS_INTRA16x16(mb_type)){
                    for(i=0; i<16; i++){
                        if(h->non_zero_count_cache[ scan8[i] ])
                            idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
                        else if(h->mb[i*16])
                            idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
                    }
                }else{
                    const int di = IS_8x8DCT(mb_type) ? 4 : 1;
                    for(i=0; i<16; i+=di){
                        int nnz = h->non_zero_count_cache[ scan8[i] ];
                        if(nnz){
                            if(nnz==1 && h->mb[i*16])
                                idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
                            else
                                idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
                        }
                    }
                }
            }else{
                for(i=0; i<16; i++){
                    if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
                        uint8_t * const ptr= dest_y + block_offset[i];
                        svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
                    }
                }
            }
        }

        if(!(s->flags&CODEC_FLAG_GRAY)){
            uint8_t *dest[2] = {dest_cb, dest_cr};