motion_est.c

Trolltech公司发布的图形界面操作系统。可在qt-embedded-2.3.7平台上编译为嵌入式图形界面操作系统。

                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
            }
            pred_x= P_LEFT[0];
            pred_y= P_LEFT[1];
        }
        dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, ref_picture);
 
        break;
    }
    
    dmin= fast_halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
                                pred_x, pred_y, ref_picture, s->dsp.pix_abs16x16_x2, s->dsp.pix_abs16x16_y2,
                                s->dsp.pix_abs16x16_xy2, 0);
//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 mb_x, int mb_y,
                   int motion_fx, int motion_fy,
                   int motion_bx, int motion_by,
                   int pred_fx, int pred_fy,
                   int pred_bx, int pred_by)
{
    //FIXME optimize?
    //FIXME direct mode penalty
    UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
    uint8_t *dest_y = s->me_scratchpad;
    uint8_t *ptr;
    int dxy;
    int src_x, src_y;
    int fbmin;

    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->qscale;

    dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
    src_x = mb_x * 16 + (motion_fx >> 1);
    src_y = mb_y * 16 + (motion_fy >> 1);
    src_x = clip(src_x, -16, s->width);
    if (src_x == s->width)
        dxy&= 2;
    src_y = clip(src_y, -16, s->height);
    if (src_y == s->height)
        dxy&= 1;

    ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
    s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

    fbmin += (mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->qscale;

    dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
    src_x = mb_x * 16 + (motion_bx >> 1);
    src_y = mb_y * 16 + (motion_by >> 1);
    src_x = clip(src_x, -16, s->width);
    if (src_x == s->width)
        dxy&= 2;
    src_y = clip(src_y, -16, s->height);
    if (src_y == s->height)
        dxy&= 1;

    ptr = s->next_picture[0] + (src_y * s->linesize) + src_x;
    s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

    fbmin += s->dsp.pix_abs16x16(s->new_picture[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
    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_width + 2;
    const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
    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, mb_x, mb_y, 
                          motion_fx, motion_fy,
                          motion_bx, motion_by,
                          pred_fx, pred_fy,
                          pred_bx, pred_by);

   return fbmin;
}

static inline int direct_search(MpegEncContext * s,
                                int mb_x, int mb_y)
{
    int P[10][2];
    const int mot_stride = s->mb_width + 2;
    const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
    int dmin, dmin2;
    int motion_fx, motion_fy, motion_bx, motion_by, motion_bx0, motion_by0;
    int motion_dx, motion_dy;
    const int motion_px= s->p_mv_table[mot_xy][0];
    const int motion_py= s->p_mv_table[mot_xy][1];
    const int time_pp= s->pp_time;
    const int time_pb= s->pb_time;
    const int time_bp= time_pp - time_pb;
    int bx, by;
    int mx, my, mx2, my2;
    uint8_t *ref_picture= s->me_scratchpad - (mb_x - 1 + (mb_y - 1)*s->linesize)*16;
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
/*    uint16_t *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; */ // f_code of the prev frame

    /* thanks to iso-mpeg the rounding is different for the zero vector, so we need to handle that ... */
    motion_fx= (motion_px*time_pb)/time_pp;
    motion_fy= (motion_py*time_pb)/time_pp;
    motion_bx0= (-motion_px*time_bp)/time_pp;
    motion_by0= (-motion_py*time_bp)/time_pp;
    motion_dx= motion_dy=0;
    dmin2= check_bidir_mv(s, mb_x, mb_y, 
                          motion_fx, motion_fy,
                          motion_bx0, motion_by0,
                          motion_fx, motion_fy,
                          motion_bx0, motion_by0) - s->qscale;

    motion_bx= motion_fx - motion_px;
    motion_by= motion_fy - motion_py;
    for(by=-1; by<2; by++){
        for(bx=-1; bx<2; bx++){
            uint8_t *dest_y = s->me_scratchpad + (by+1)*s->linesize*16 + (bx+1)*16;
            uint8_t *ptr;
            int dxy;
            int src_x, src_y;
            const int width= s->width;
            const int height= s->height;

            dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
            src_x = (mb_x + bx) * 16 + (motion_fx >> 1);
            src_y = (mb_y + by) * 16 + (motion_fy >> 1);
            src_x = clip(src_x, -16, width);
            if (src_x == width) dxy &= ~1;
            src_y = clip(src_y, -16, height);
            if (src_y == height) dxy &= ~2;

            ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
            s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

            dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
            src_x = (mb_x + bx) * 16 + (motion_bx >> 1);
            src_y = (mb_y + by) * 16 + (motion_by >> 1);
            src_x = clip(src_x, -16, width);
            if (src_x == width) dxy &= ~1;
            src_y = clip(src_y, -16, height);
            if (src_y == height) dxy &= ~2;

	    s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
        }
    }

    P_LAST[0]        = mv_table[mot_xy    ][0];
    P_LAST[1]        = mv_table[mot_xy    ][1];
    P_LEFT[0]        = mv_table[mot_xy - 1][0];
    P_LEFT[1]        = mv_table[mot_xy - 1][1];
    P_LAST_RIGHT[0]  = mv_table[mot_xy + 1][0];
    P_LAST_RIGHT[1]  = mv_table[mot_xy + 1][1];
    P_LAST_BOTTOM[0] = mv_table[mot_xy + mot_stride][0];
    P_LAST_BOTTOM[1] = mv_table[mot_xy + mot_stride][1];
/*
    if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
    if(P_LAST_RIGHT[0] < (rel_xmin<<shift)) P_LAST_RIGHT[0] = (rel_xmin<<shift);
    if(P_LAST_BOTTOM[1]< (rel_ymin<<shift)) P_LAST_BOTTOM[1]= (rel_ymin<<shift);
*/
    /* special case for first line */
    if ((mb_y == 0 || s->first_slice_line)) {
    } else {
        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];
    
        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 = epzs_motion_search(s, &mx, &my, P, 0, 0, -16, -16, 15, 15, ref_picture);
    if(mx==0 && my==0) dmin=99999999; // not representable, due to rounding stuff
    if(dmin2<dmin){ 
        dmin= dmin2;
        mx=0;
        my=0;
    }
#if 1
    mx2= mx= mx*2; 
    my2= my= my*2;
    for(by=-1; by<2; by++){
        if(my2+by < -32) continue;
        for(bx=-1; bx<2; bx++){
            if(bx==0 && by==0) continue;
            if(mx2+bx < -32) continue;
            dmin2= check_bidir_mv(s, mb_x, mb_y, 
                          mx2+bx+motion_fx, my2+by+motion_fy,
                          mx2+bx+motion_bx, my2+by+motion_by,
                          mx2+bx+motion_fx, my2+by+motion_fy,
                          motion_bx, motion_by) - s->qscale;
            
            if(dmin2<dmin){
                dmin=dmin2;
                mx= mx2 + bx;
                my= my2 + by;
            }
        }
    }
#else
    mx*=2; my*=2;
#endif
    if(mx==0 && my==0){
        motion_bx= motion_bx0;
        motion_by= motion_by0;
    }

    s->b_direct_mv_table[mot_xy][0]= mx;
    s->b_direct_mv_table[mot_xy][1]= my;
    s->b_direct_forw_mv_table[mot_xy][0]= motion_fx + mx;
    s->b_direct_forw_mv_table[mot_xy][1]= motion_fy + my;
    s->b_direct_back_mv_table[mot_xy][0]= motion_bx + mx;
    s->b_direct_back_mv_table[mot_xy][1]= motion_by + my;
    return dmin;
}

void ff_estimate_b_frame_motion(MpegEncContext * s,
                             int mb_x, int mb_y)
{
    const int quant= s->qscale;
    int fmin, bmin, dmin, fbmin;
    int type=0;
    
    dmin= direct_search(s, mb_x, mb_y);

    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, s->last_picture[0], s->f_code);
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, s->next_picture[0], s->b_code) - quant;
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

    fbmin= bidir_refine(s, mb_x, mb_y);

    {
        int score= dmin;
        type=MB_TYPE_DIRECT;
        
        if(fmin<score){
            score=fmin;
            type= MB_TYPE_FORWARD; 
        }
        if(bmin<score){
            score=bmin;
            type= MB_TYPE_BACKWARD; 
        }
        if(fbmin<score){
            score=fbmin;
            type= MB_TYPE_BIDIR;
        }
        score= ((unsigned)(score*score + 128*256))>>16;
        s->mc_mb_var_sum += score;
        s->mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSD
    }

    if(s->flags&CODEC_FLAG_HQ){
        type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
    }

/*
{
static int count=0;
static int sum=0;
if(type==MB_TYPE_DIRECT){
  int diff= ABS(s->b_forw_mv_table)
}
}*/

    s->mb_type[mb_y*s->mb_width + mb_x]= type;
/*    if(mb_y==0 && mb_x==0) printf("\n");
    if(mb_x==0) printf("\n");
    printf("%d", av_log2(type));
*/
}

/* find best f_code for ME which do unlimited searches */
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
{
    if(s->me_method>=ME_EPZS){
        int score[8];
        int i, y;
        UINT8 * fcode_tab= s->fcode_tab;
        int best_fcode=-1;
        int best_score=-10000000;

        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i); //FIXME *2 and all other too so its the same but nicer

        for(y=0; y<s->mb_height; y++){
            int x;
            int xy= (y+1)* (s->mb_width+2) + 1;
            i= y*s->mb_width;
            for(x=0; x<s->mb_width; x++){
                if(s->mb_type[i] & type){
                    int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
                                     fcode_tab[mv_table[xy][1] + MAX_MV]);
                    int j;
                    
                    for(j=0; j<fcode && j<8; j++){
                        if(s->pict_type==B_TYPE || s->mc_mb_var[i] < s->mb_var[i])
                            score[j]-= 170;
                    }
                }
                i++;
                xy++;
            }
        }
        
        for(i=1; i<8; i++){
            if(score[i] > best_score){
                best_score= score[i];
                best_fcode= i;
            }
//            printf("%d %d\n", i, score[i]);
        }

//    printf("fcode: %d type: %d\n", i, s->pict_type);
        return best_fcode;
/*        for(i=0; i<=MAX_FCODE; i++){
            printf("%d ", mv_num[i]);
        }
        printf("\n");*/
    }else{
        return 1;
    }
}

void ff_fix_long_p_mvs(MpegEncContext * s)
{
    const int f_code= s->f_code;
    int y;
    UINT8 * fcode_tab= s->fcode_tab;
//int clip=0;
//int noclip=0;
    /* clip / convert to intra 16x16 type MVs */
    for(y=0; y<s->mb_height; y++){
        int x;
        int xy= (y+1)* (s->mb_width+2)+1;
        int i= y*s->mb_width;
        for(x=0; x<s->mb_width; x++){
            if(s->mb_type[i]&MB_TYPE_INTER){
                if(   fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code
                   || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0
                   || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code
                   || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){
                    s->mb_type[i] &= ~MB_TYPE_INTER;
                    s->mb_type[i] |= MB_TYPE_INTRA;
                    s->p_mv_table[xy][0] = 0;
                    s->p_mv_table[xy][1] = 0;
//clip++;
                }
//else
//  noclip++;
            }
            xy++;
            i++;
        }
    }
//printf("%d no:%d %d//\n", clip, noclip, f_code);
    if(s->flags&CODEC_FLAG_4MV){
        const int wrap= 2+ s->mb_width*2;

        /* clip / convert to intra 8x8 type MVs */
        for(y=0; y<s->mb_height; y++){
            int xy= (y*2 + 1)*wrap + 1;
            int i= y*s->mb_width;
            int x;

            for(x=0; x<s->mb_width; x++){
                if(s->mb_type[i]&MB_TYPE_INTER4V){
                    int block;
                    for(block=0; block<4; block++){
                        int off= (block& 1) + (block>>1)*wrap;
                        int mx= s->motion_val[ xy + off ][0];
                        int my= s->motion_val[ xy + off ][1];

                        if(   fcode_tab[mx + MAX_MV] > f_code
                           || fcode_tab[mx + MAX_MV] == 0
                           || fcode_tab[my + MAX_MV] > f_code
                           || fcode_tab[my + MAX_MV] == 0 ){
                            s->mb_type[i] &= ~MB_TYPE_INTER4V;
                            s->mb_type[i] |= MB_TYPE_INTRA;
                        }
                    }
                }
                xy+=2;
                i++;
            }
        }
    }
}

void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
{
    int y;
    UINT8 * fcode_tab= s->fcode_tab;

    /* clip / convert to intra 16x16 type MVs */
    for(y=0; y<s->mb_height; y++){
        int x;
        int xy= (y+1)* (s->mb_width+2)+1;
        int i= y*s->mb_width;
        for(x=0; x<s->mb_width; x++){
            if(   fcode_tab[mv_table[xy][0] + MAX_MV] > f_code
               || fcode_tab[mv_table[xy][0] + MAX_MV] == 0){
                if(mv_table[xy][0]>0) mv_table[xy][0]=  (16<<f_code)-1;
                else                  mv_table[xy][0]= -(16<<f_code);
            }
            if(   fcode_tab[mv_table[xy][1] + MAX_MV] > f_code
               || fcode_tab[mv_table[xy][1] + MAX_MV] == 0){
                if(mv_table[xy][1]>0) mv_table[xy][1]=  (16<<f_code)-1;
                else                  mv_table[xy][1]= -(16<<f_code);
            }
            xy++;
            i++;
        }
    }
}