motion_est.c

tcpmp播放器的flv插件

    int P[10][2];
    const int shift= 1+s->quarter_sample;
    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->lambda, s->lambda2, c->avctx->me_cmp);
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
    c->current_mv_penalty= mv_penalty;

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

    switch(s->me_method) {
    case ME_ZERO:
    default:
	no_motion_search(s, &mx, &my);
        dmin = 0;
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
#if 0
    case ME_FULL:
	dmin = full_motion_search(s, &mx, &my, range, ref_picture);
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
    case ME_LOG:
	dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
    case ME_PHODS:
	dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
#endif
    case ME_X1:
    case ME_EPZS:
       {
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
            P_LEFT[1]        = mv_table[mot_xy - 1][1];

            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);

            /* special case for first line */
            if (!s->first_slice_line) {
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
                if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
        
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
            }
            c->pred_x= P_LEFT[0];
            c->pred_y= P_LEFT[1];
        }
        
        if(mv_table == s->b_forw_mv_table){
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
        }else{
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
        }
        
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
 
        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= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);

//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;
    mv_table[mot_xy][1]= my;

    return dmin;
}

static inline int check_bidir_mv(MpegEncContext * s,
                   int motion_fx, int motion_fy,
                   int motion_bx, int motion_by,
                   int pred_fx, int pred_fy,
                   int pred_bx, int pred_by,
                   int size, int h)
{
    //FIXME optimize?
    //FIXME better f_code prediction (max mv & distance)
    //FIXME pointers
    MotionEstContext * const c= &s->me;
    uint8_t * const mv_penalty= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
    int stride= c->stride;
    uint8_t *dest_y = c->scratchpad;
    uint8_t *ptr;
    int dxy;
    int src_x, src_y;
    int fbmin;
    uint8_t **src_data= c->src[0];
    uint8_t **ref_data= c->ref[0];
    uint8_t **ref2_data= c->ref[2];

    if(s->quarter_sample){
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
        src_x = motion_fx >> 2;
        src_y = motion_fy >> 2;

        ptr = ref_data[0] + (src_y * stride) + src_x;
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);

        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
        src_x = motion_bx >> 2;
        src_y = motion_by >> 2;
    
        ptr = ref2_data[0] + (src_y * stride) + src_x;
        s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
    }else{
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
        src_x = motion_fx >> 1;
        src_y = motion_fy >> 1;

        ptr = ref_data[0] + (src_y * stride) + src_x;
        s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
        src_x = motion_bx >> 1;
        src_y = motion_by >> 1;
    
        ptr = ref2_data[0] + (src_y * stride) + src_x;
        s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
    }

    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;
}

/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
{
    const int mot_stride = s->mb_stride;
    const int xy = mb_y *mot_stride + mb_x;
    int fbmin;
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];

    //FIXME do refinement and add flag
    
    fbmin= check_bidir_mv(s, motion_fx, motion_fy,
                          motion_bx, motion_by,
                          pred_fx, pred_fy,
                          pred_bx, pred_by,
                          0, 16);

   return fbmin;
}

static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
{
    MotionEstContext * const c= &s->me;
    int P[10][2];
    const int mot_stride = s->mb_stride;
    const int mot_xy = mb_y*mot_stride + mb_x;
    const int shift= 1+s->quarter_sample;
    int dmin, i;
    const int time_pp= s->pp_time;
    const int time_pb= s->pb_time;
    int mx, my, xmin, xmax, ymin, ymax;
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
    
    c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
    ymin= xmin=(-32)>>shift;
    ymax= xmax=   31>>shift;

    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
        s->mv_type= MV_TYPE_8X8;
    }else{
        s->mv_type= MV_TYPE_16X16;
    }

    for(i=0; i<4; i++){
        int index= s->block_index[i];
        int min, max;
    
        c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
        c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
        c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
        c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
//        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
//        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);

        max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
        min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
        max+= 16*mb_x + 1; // +-1 is for the simpler rounding
        min+= 16*mb_x - 1;
        xmax= FFMIN(xmax, s->width - max);
        xmin= FFMAX(xmin, - 16     - min);

        max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
        min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
        max+= 16*mb_y + 1; // +-1 is for the simpler rounding
        min+= 16*mb_y - 1;
        ymax= FFMIN(ymax, s->height - max);
        ymin= FFMAX(ymin, - 16      - min);
        
        if(s->mv_type == MV_TYPE_16X16) break;
    }
    
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
    
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
        s->b_direct_mv_table[mot_xy][0]= 0;
        s->b_direct_mv_table[mot_xy][1]= 0;

        return 256*256*256*64;
    }
    
    c->xmin= xmin;
    c->ymin= ymin;
    c->xmax= xmax;
    c->ymax= ymax;
    c->flags     |= FLAG_DIRECT;
    c->sub_flags |= FLAG_DIRECT;
    c->pred_x=0;
    c->pred_y=0;

    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);

    /* special case for first line */
    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
    
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
    }
 
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
    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= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
    
    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;
    s->b_direct_mv_table[mot_xy][1]= my;
    c->flags     &= ~FLAG_DIRECT;
    c->sub_flags &= ~FLAG_DIRECT;

    return dmin;
}

void ff_estimate_b_frame_motion(MpegEncContext * s,
                             int mb_x, int mb_y)
{
    MotionEstContext * const c= &s->me;
    const int penalty_factor= c->mb_penalty_factor;
    int fmin, bmin, dmin, fbmin, bimin, fimin;
    int type=0;
    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);

    get_limits(s, 16*mb_x, 16*mb_y);
    
    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){
                c->skip=0;
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
            }
            if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
                c->skip=0;
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
            }
            if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
                c->skip=0;
                c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
                interlaced_search(s, 0,
                                        s->b_field_mv_table[0], s->b_field_select_table[0],
                                        s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
            }
            if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
                c->skip=0;
                c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
                interlaced_search(s, 2,
                                        s->b_field_mv_table[1], s->b_field_select_table[1],
                                        s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
            }
            return;
        }
    }

    if (s->codec_id == CODEC_ID_MPEG4)
        dmin= direct_search(s, mb_x, mb_y);
    else
        dmin= INT_MAX;
//FIXME penalty stuff for non mpeg4
    c->skip=0;
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
    
    c->skip=0;
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

    c->skip=0;
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
    
    if(s->flags & CODEC_FLAG_INTERLACED_ME){
//FIXME mb type penalty
        c->skip=0;
        c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;