comparison motion_est.c @ 2015:3ab8f3e2ae6a libavcodec

moving motion estimation specific variables from MpegEncContext -> MotionEstContext

2014:15c885db82a8

            c->ref[ref_index][i]= ref2[i] + offset[i];
        }
    }
}

static int get_flags(MpegEncContext *s, int direct, int chroma){
    return   ((s->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
           + (direct ? FLAG_DIRECT : 0) 
           + (chroma ? FLAG_CHROMA : 0);
}

static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,

    }
}

void ff_init_me(MpegEncContext *s){
    MotionEstContext * const c= &s->me;

    ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
    ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
    ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
    ff_set_cmp(&s->dsp, s->dsp.mb_cmp, s->avctx->mb_cmp);
    
    c->flags    = get_flags(s, 0, s->avctx->me_cmp    &FF_CMP_CHROMA);
    c->sub_flags= get_flags(s, 0, s->avctx->me_sub_cmp&FF_CMP_CHROMA);
    c->mb_flags = get_flags(s, 0, s->avctx->mb_cmp    &FF_CMP_CHROMA);

/*FIXME s->no_rounding b_type*/
    if(s->flags&CODEC_FLAG_QPEL){
        c->sub_motion_search= qpel_motion_search;
        c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
        if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
        else               c->qpel_put= s->dsp.put_qpel_pixels_tab;
    }else{
        if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
            c->sub_motion_search= hpel_motion_search;
        else if(   s->avctx->me_sub_cmp == FF_CMP_SAD 
                && s->avctx->    me_cmp == FF_CMP_SAD 
                && s->avctx->    mb_cmp == FF_CMP_SAD)
            c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
        else
            c->sub_motion_search= hpel_motion_search;
        c->hpel_avg= s->dsp.avg_pixels_tab;
        if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;

 */
static inline void get_limits(MpegEncContext *s, int x, int y)
{
    MotionEstContext * const c= &s->me;
/*
    if(s->avctx->me_range) c->range= s->avctx->me_range >> 1;
    else                   c->range= 16;
*/
    if (s->unrestricted_mv) {
        c->xmin = - x - 16;
        c->ymin = - y - 16;

        c->xmax = - x + s->mb_width *16 - 16;
        c->ymax = - y + s->mb_height*16 - 16;
    }
}

static inline void init_mv4_ref(MpegEncContext *s){
    MotionEstContext * const c= &s->me;
    const int stride= s->linesize;

    c->ref[1][0] = c->ref[0][0] + 8;
    c->ref[2][0] = c->ref[0][0] + 8*stride;
    c->ref[3][0] = c->ref[2][0] + 8;
    c->src[1][0] = c->src[0][0] + 8;

    const int h=8;
    int block;
    int P[10][2];
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
    int same=1;
    const int stride= s->linesize;
    const int uvstride= s->uvlinesize;
    uint8_t *mv_penalty= c->current_mv_penalty;

    init_mv4_ref(s);
    
    for(block=0; block<4; block++){
        int mx4, my4;
        int pred_x4, pred_y4;
        int dmin4;

    
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
        dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
    }
    
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
        int dxy;
        int mx, my;
        int offset;

        mx= ff_h263_round_chroma(mx4_sum);

    }
    
    c->pred_x= mx;
    c->pred_y= my;

    switch(s->avctx->mb_cmp&0xFF){
    /*case FF_CMP_SSE:
        return dmin_sum+ 32*s->qscale*s->qscale;*/
    case FF_CMP_RD:
        return dmin_sum;
    default:

    c->uvstride>>=1;

    if(same)
        return INT_MAX;
    
    switch(s->avctx->mb_cmp&0xFF){
    /*case FF_CMP_SSE:
        return dmin_sum+ 32*s->qscale*s->qscale;*/
    case FF_CMP_RD:
        return dmin_sum;
    default:

        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
        c->stride<<=1;
        c->uvstride<<=1;
        
        if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
            av_log(s->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
            return -1;
        }

        if(USES_LIST(mb_type, 0)){
            int field_select0= p->ref_index[0][xy ];

        }
        c->stride>>=1;
        c->uvstride>>=1;
    }else if(IS_8X8(mb_type)){
        if(!(s->flags & CODEC_FLAG_4MV)){
            av_log(s->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
            return -1;
        }
        cmpf= s->dsp.sse[1];
        chroma_cmpf= s->dsp.sse[1];
        init_mv4_ref(s);
        for(i=0; i<4; i++){
            xy= s->block_index[i];
            x= p->motion_val[0][xy][0]; 
            y= p->motion_val[0][xy][1];
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);

    assert(s->quarter_sample==0 || s->quarter_sample==1);
    assert(s->linesize == c->stride);
    assert(s->uvlinesize == c->uvstride);

    c->penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
    c->sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
    c->mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

    get_limits(s, 16*mb_x, 16*mb_y);
    c->skip=0;

    sum = s->dsp.pix_sum(pix, s->linesize);
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;

    pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
    pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
    s->mb_var_sum_temp += varc;

    if(s->avctx->me_threshold){
        vard= (check_input_motion(s, mb_x, mb_y, 1)+128)>>8;
        
        if(vard<s->avctx->me_threshold){
            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
            s->mc_mb_var_sum_temp += vard;
            if (vard <= 64 || vard < varc) { //FIXME
                s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
            }else{
                s->scene_change_score+= s->qscale;
            }
            return;
        }
        if(vard<s->avctx->mb_threshold)
            mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
    }

    switch(s->me_method) {
    case ME_ZERO:

        
    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;

    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
    s->mc_mb_var_sum_temp += vard;
    
#if 0
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
	   varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
#endif
    if(mb_type){
        if (vard <= 64 || vard < varc)
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        else
            s->scene_change_score+= s->qscale;

        if(mb_type == CANDIDATE_MB_TYPE_INTER){
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
            set_p_mv_tables(s, mx, my, 1);
        }else{

            set_p_mv_tables(s, mx, my, 0);
        }
        if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
            interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
        }
    }else if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
        if (vard <= 64 || vard < varc)
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        else
            s->scene_change_score+= s->qscale;

        if (vard*2 + 200 > varc)
            mb_type|= CANDIDATE_MB_TYPE_INTRA;
        if (varc*2 + 200 > vard){
            mb_type|= CANDIDATE_MB_TYPE_INTER;

    }else{
        int intra_score, i;
        mb_type= CANDIDATE_MB_TYPE_INTER;

        dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !c->skip)
            dmin= get_mb_score(s, mx, my, 0, 0);

        if((s->flags&CODEC_FLAG_4MV)
           && !c->skip && varc>50 && vard>10){
            int dmin4= h263_mv4_search(s, mx, my, shift);

                
//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
        set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);

        /* get intra luma score */
        if((s->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
            intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
        }else{
            int mean= (sum+128)>>8;
            mean*= 0x01010101;
            

            intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
        }
#if 0 //FIXME
        /* get chroma score */
        if(s->avctx->mb_cmp&FF_CMP_CHROMA){
            for(i=1; i<3; i++){
                uint8_t *dest_c;
                int mean;
                
                if(s->out_format == FMT_H263){

            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
        }else
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
        
        if (vard <= 64 || vard < varc) { //FIXME
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        }else{
            s->scene_change_score+= s->qscale;
        }
    }

    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
}

    const int xy= mb_x + mb_y*s->mb_stride;
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
    
    assert(s->quarter_sample==0 || s->quarter_sample==1);

    c->pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

    get_limits(s, 16*mb_x, 16*mb_y);
    c->skip=0;

    const int mot_stride = s->mb_stride;
    const int mot_xy = mb_y*mot_stride + mb_x;
    uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
    int mv_scale;
        
    c->penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
    c->sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
    c->mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
    c->current_mv_penalty= mv_penalty;

    get_limits(s, 16*mb_x, 16*mb_y);

    switch(s->me_method) {

        break;
    }
    
    dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
                                   
    if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !c->skip)
        dmin= get_mb_score(s, mx, my, 0, ref_index);

//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
    mv_table[mot_xy][0]= mx;

    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*c->mb_penalty_factor
           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*c->mb_penalty_factor
           + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
           
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
    }
    //FIXME CHROMA !!!
           
    return fbmin;
}

    if(c->sub_flags&FLAG_QPEL) 
        dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
    else
        dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
    
    if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !c->skip)
        dmin= get_mb_score(s, mx, my, 0, 0);
    
    get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed

    s->b_direct_mv_table[mot_xy][0]= mx;

    const int xy = mb_y*s->mb_stride + mb_x;
    init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);

    
    c->skip=0;
    if(s->avctx->me_threshold){
        int vard= (check_input_motion(s, mb_x, mb_y, 0)+128)>>8;
        
        if(vard<s->avctx->me_threshold){
//            pix = c->src[0][0];
//            sum = s->dsp.pix_sum(pix, s->linesize);
//            varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
        
//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
/*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
            s->mb_var_sum_temp    += varc;*/
            s->mc_mb_var_sum_temp += vard;
/*            if (vard <= 64 || vard < varc) {
                s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
            }else{
                s->scene_change_score+= s->qscale;
            }*/
            return;
        }
        if(vard<s->avctx->mb_threshold){
            type= s->mb_type[mb_y*s->mb_stride + mb_x];
            if(type == CANDIDATE_MB_TYPE_DIRECT){
                direct_search(s, mb_x, mb_y);
            }
            if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){

            score=bimin;
            type= CANDIDATE_MB_TYPE_BACKWARD_I;
        }
        
        score= ((unsigned)(score*score + 128*256))>>16;
        s->mc_mb_var_sum_temp += score;
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
    }

    if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
        type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
        if(fimin < INT_MAX)
            type |= CANDIDATE_MB_TYPE_FORWARD_I;
        if(bimin < INT_MAX)
            type |= CANDIDATE_MB_TYPE_BACKWARD_I;

    }
}

void ff_fix_long_p_mvs(MpegEncContext * s)
{
    const int f_code= s->f_code;
    int y, range;
    assert(s->pict_type==P_TYPE);

    range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
    
    if(s->msmpeg4_version) range= 16;
    
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
    
//printf("%d no:%d %d//\n", clip, noclip, f_code);
    if(s->flags&CODEC_FLAG_4MV){
        const int wrap= s->b8_stride;

 * @param truncate 1 for truncation, 0 for using intra
 */
void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select, 
                     int16_t (*mv_table)[2], int f_code, int type, int truncate)
{
    int y, h_range, v_range;

    // RAL: 8 in MPEG-1, 16 in MPEG-4
    int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);

    if(s->msmpeg4_version) range= 16;
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;

    h_range= range;
    v_range= field_select_table ? range>>1 : range;

    /* clip / convert to intra 16x16 type MVs */

2015:3ab8f3e2ae6a

            c->ref[ref_index][i]= ref2[i] + offset[i];
        }
    }
}

static int get_flags(MotionEstContext *c, int direct, int chroma){
    return   ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
           + (direct ? FLAG_DIRECT : 0) 
           + (chroma ? FLAG_CHROMA : 0);
}

static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,

    }
}

void ff_init_me(MpegEncContext *s){
    MotionEstContext * const c= &s->me;
    c->avctx= s->avctx;

    ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
    ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
    ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
    ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
    
    c->flags    = get_flags(c, 0, c->avctx->me_cmp    &FF_CMP_CHROMA);
    c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
    c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp    &FF_CMP_CHROMA);

/*FIXME s->no_rounding b_type*/
    if(s->flags&CODEC_FLAG_QPEL){
        c->sub_motion_search= qpel_motion_search;
        c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
        if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
        else               c->qpel_put= s->dsp.put_qpel_pixels_tab;
    }else{
        if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
            c->sub_motion_search= hpel_motion_search;
        else if(   c->avctx->me_sub_cmp == FF_CMP_SAD 
                && c->avctx->    me_cmp == FF_CMP_SAD 
                && c->avctx->    mb_cmp == FF_CMP_SAD)
            c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
        else
            c->sub_motion_search= hpel_motion_search;
        c->hpel_avg= s->dsp.avg_pixels_tab;
        if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;

 */
static inline void get_limits(MpegEncContext *s, int x, int y)
{
    MotionEstContext * const c= &s->me;
/*
    if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
    else                   c->range= 16;
*/
    if (s->unrestricted_mv) {
        c->xmin = - x - 16;
        c->ymin = - y - 16;

        c->xmax = - x + s->mb_width *16 - 16;
        c->ymax = - y + s->mb_height*16 - 16;
    }
}

static inline void init_mv4_ref(MotionEstContext *c){
    const int stride= c->stride;

    c->ref[1][0] = c->ref[0][0] + 8;
    c->ref[2][0] = c->ref[0][0] + 8*stride;
    c->ref[3][0] = c->ref[2][0] + 8;
    c->src[1][0] = c->src[0][0] + 8;

    const int h=8;
    int block;
    int P[10][2];
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
    int same=1;
    const int stride= c->stride;
    const int uvstride= c->uvstride;
    uint8_t *mv_penalty= c->current_mv_penalty;

    init_mv4_ref(c);
    
    for(block=0; block<4; block++){
        int mx4, my4;
        int pred_x4, pred_y4;
        int dmin4;

    
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
        dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
    }
    
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
        int dxy;
        int mx, my;
        int offset;

        mx= ff_h263_round_chroma(mx4_sum);

    }
    
    c->pred_x= mx;
    c->pred_y= my;

    switch(c->avctx->mb_cmp&0xFF){
    /*case FF_CMP_SSE:
        return dmin_sum+ 32*s->qscale*s->qscale;*/
    case FF_CMP_RD:
        return dmin_sum;
    default:

    c->uvstride>>=1;

    if(same)
        return INT_MAX;
    
    switch(c->avctx->mb_cmp&0xFF){
    /*case FF_CMP_SSE:
        return dmin_sum+ 32*s->qscale*s->qscale;*/
    case FF_CMP_RD:
        return dmin_sum;
    default:

        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
        c->stride<<=1;
        c->uvstride<<=1;
        
        if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
            av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
            return -1;
        }

        if(USES_LIST(mb_type, 0)){
            int field_select0= p->ref_index[0][xy ];

        }
        c->stride>>=1;
        c->uvstride>>=1;
    }else if(IS_8X8(mb_type)){
        if(!(s->flags & CODEC_FLAG_4MV)){
            av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
            return -1;
        }
        cmpf= s->dsp.sse[1];
        chroma_cmpf= s->dsp.sse[1];
        init_mv4_ref(c);
        for(i=0; i<4; i++){
            xy= s->block_index[i];
            x= p->motion_val[0][xy][0]; 
            y= p->motion_val[0][xy][1];
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);

    assert(s->quarter_sample==0 || s->quarter_sample==1);
    assert(s->linesize == c->stride);
    assert(s->uvlinesize == c->uvstride);

    c->penalty_factor    = get_penalty_factor(s, c->avctx->me_cmp);
    c->sub_penalty_factor= get_penalty_factor(s, c->avctx->me_sub_cmp);
    c->mb_penalty_factor = get_penalty_factor(s, c->avctx->mb_cmp);
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

    get_limits(s, 16*mb_x, 16*mb_y);
    c->skip=0;

    sum = s->dsp.pix_sum(pix, s->linesize);
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;

    pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
    pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
    c->mb_var_sum_temp += varc;

    if(c->avctx->me_threshold){
        vard= (check_input_motion(s, mb_x, mb_y, 1)+128)>>8;
        
        if(vard<c->avctx->me_threshold){
            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
            c->mc_mb_var_sum_temp += vard;
            if (vard <= 64 || vard < varc) { //FIXME
                c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
            }else{
                c->scene_change_score+= s->qscale;
            }
            return;
        }
        if(vard<c->avctx->mb_threshold)
            mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
    }

    switch(s->me_method) {
    case ME_ZERO:

        
    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;

    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
    c->mc_mb_var_sum_temp += vard;
    
#if 0
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
	   varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
#endif
    if(mb_type){
        if (vard <= 64 || vard < varc)
            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        else
            c->scene_change_score+= s->qscale;

        if(mb_type == CANDIDATE_MB_TYPE_INTER){
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
            set_p_mv_tables(s, mx, my, 1);
        }else{

            set_p_mv_tables(s, mx, my, 0);
        }
        if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
            interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
        }
    }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
        if (vard <= 64 || vard < varc)
            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        else
            c->scene_change_score+= s->qscale;

        if (vard*2 + 200 > varc)
            mb_type|= CANDIDATE_MB_TYPE_INTRA;
        if (varc*2 + 200 > vard){
            mb_type|= CANDIDATE_MB_TYPE_INTER;

    }else{
        int intra_score, i;
        mb_type= CANDIDATE_MB_TYPE_INTER;

        dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
        if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
            dmin= get_mb_score(s, mx, my, 0, 0);

        if((s->flags&CODEC_FLAG_4MV)
           && !c->skip && varc>50 && vard>10){
            int dmin4= h263_mv4_search(s, mx, my, shift);

                
//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
        set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);

        /* get intra luma score */
        if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
            intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
        }else{
            int mean= (sum+128)>>8;
            mean*= 0x01010101;
            

            intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
        }
#if 0 //FIXME
        /* get chroma score */
        if(c->avctx->mb_cmp&FF_CMP_CHROMA){
            for(i=1; i<3; i++){
                uint8_t *dest_c;
                int mean;
                
                if(s->out_format == FMT_H263){

            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
        }else
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
        
        if (vard <= 64 || vard < varc) { //FIXME
            c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        }else{
            c->scene_change_score+= s->qscale;
        }
    }

    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
}

    const int xy= mb_x + mb_y*s->mb_stride;
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
    
    assert(s->quarter_sample==0 || s->quarter_sample==1);

    c->pre_penalty_factor    = get_penalty_factor(s, c->avctx->me_pre_cmp);
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

    get_limits(s, 16*mb_x, 16*mb_y);
    c->skip=0;

    const int mot_stride = s->mb_stride;
    const int mot_xy = mb_y*mot_stride + mb_x;
    uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
    int mv_scale;
        
    c->penalty_factor    = get_penalty_factor(s, c->avctx->me_cmp);
    c->sub_penalty_factor= get_penalty_factor(s, c->avctx->me_sub_cmp);
    c->mb_penalty_factor = get_penalty_factor(s, c->avctx->mb_cmp);
    c->current_mv_penalty= mv_penalty;

    get_limits(s, 16*mb_x, 16*mb_y);

    switch(s->me_method) {

        break;
    }
    
    dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
                                   
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
        dmin= get_mb_score(s, mx, my, 0, ref_index);

//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
    mv_table[mot_xy][0]= mx;

    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*c->mb_penalty_factor
           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*c->mb_penalty_factor
           + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
           
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
    }
    //FIXME CHROMA !!!
           
    return fbmin;
}

    if(c->sub_flags&FLAG_QPEL) 
        dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
    else
        dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
    
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
        dmin= get_mb_score(s, mx, my, 0, 0);
    
    get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed

    s->b_direct_mv_table[mot_xy][0]= mx;

    const int xy = mb_y*s->mb_stride + mb_x;
    init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);

    
    c->skip=0;
    if(c->avctx->me_threshold){
        int vard= (check_input_motion(s, mb_x, mb_y, 0)+128)>>8;
        
        if(vard<c->avctx->me_threshold){
//            pix = c->src[0][0];
//            sum = s->dsp.pix_sum(pix, s->linesize);
//            varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
        
//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
/*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
            c->mb_var_sum_temp    += varc;*/
            c->mc_mb_var_sum_temp += vard;
/*            if (vard <= 64 || vard < varc) {
                c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
            }else{
                c->scene_change_score+= s->qscale;
            }*/
            return;
        }
        if(vard<c->avctx->mb_threshold){
            type= s->mb_type[mb_y*s->mb_stride + mb_x];
            if(type == CANDIDATE_MB_TYPE_DIRECT){
                direct_search(s, mb_x, mb_y);
            }
            if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){

            score=bimin;
            type= CANDIDATE_MB_TYPE_BACKWARD_I;
        }
        
        score= ((unsigned)(score*score + 128*256))>>16;
        c->mc_mb_var_sum_temp += score;
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
    }

    if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
        type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
        if(fimin < INT_MAX)
            type |= CANDIDATE_MB_TYPE_FORWARD_I;
        if(bimin < INT_MAX)
            type |= CANDIDATE_MB_TYPE_BACKWARD_I;

    }
}

void ff_fix_long_p_mvs(MpegEncContext * s)
{
    MotionEstContext * const c= &s->me;
    const int f_code= s->f_code;
    int y, range;
    assert(s->pict_type==P_TYPE);

    range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
    
    if(s->msmpeg4_version) range= 16;
    
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
    
//printf("%d no:%d %d//\n", clip, noclip, f_code);
    if(s->flags&CODEC_FLAG_4MV){
        const int wrap= s->b8_stride;

 * @param truncate 1 for truncation, 0 for using intra
 */
void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select, 
                     int16_t (*mv_table)[2], int f_code, int type, int truncate)
{
    MotionEstContext * const c= &s->me;
    int y, h_range, v_range;

    // RAL: 8 in MPEG-1, 16 in MPEG-4
    int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);

    if(s->msmpeg4_version) range= 16;
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;

    h_range= range;
    v_range= field_select_table ? range>>1 : range;

    /* clip / convert to intra 16x16 type MVs */