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。

static inline int snake_search(MpegEncContext * s, int *best, int dmin,
                                       UINT8 *new_pic, UINT8 *old_pic, int pic_stride,
                                       int pred_x, int pred_y, UINT16 *mv_penalty, int quant,
                                       int xmin, int ymin, int xmax, int ymax, int shift,
                                       uint32_t *map, uint16_t *score_map,int map_generation,
                                       op_pixels_abs_func pix_abs)
{
    int dir=0;
    int c=1;
    static int x_dir[8]= {1,1,0,-1,-1,-1, 0, 1};
    static int y_dir[8]= {0,1,1, 1, 0,-1,-1,-1};
    int fails=0;
    int last_d[2]={dmin, dmin};

/*static int good=0;
static int bad=0;
static int point=0;

point++;
if(256*256*256*64%point==0)
{
    printf("%d %d %d\n", good, bad, point);
}*/

    for(;;){
        int x= best[0];
        int y= best[1];
        int d;
        x+=x_dir[dir];
        y+=y_dir[dir];
        if(x>=xmin && x<=xmax && y>=ymin && y<=ymax){
            const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;
            const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
            if(map[index]!=key){
                d = pix_abs(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);
                d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;
                map[index]=key;
                score_map[index]=d;
            }else
                d= dmin+1;
        }else{
            d = dmin + 10000; //FIXME smarter boundary handling
        }
        if(d<dmin){
            best[0]=x;
            best[1]=y;
            dmin=d;

            if(last_d[1] - last_d[0] > last_d[0] - d) c= -c;
            dir+=c;

            fails=0;
//good++;
            last_d[1]=last_d[0];
            last_d[0]=d;
        }else{
//bad++;
            if(fails){
                if(fails>=SNAKE_1+1) return dmin;
            }else{
                if(dir&1) dir-= c*3;
                else      c= -c;
//                c= -c;
            }
            dir+=c*SNAKE_2;
            fails++;
        }
        dir&=7;
    }
}

static inline int cross_search(MpegEncContext * s, int *best, int dmin,
                                       UINT8 *new_pic, UINT8 *old_pic, int pic_stride,
                                       int pred_x, int pred_y, UINT16 *mv_penalty, int quant,
                                       int xmin, int ymin, int xmax, int ymax, int shift,
                                       uint32_t *map, uint16_t *score_map,int map_generation,
                                       op_pixels_abs_func pix_abs)
{
    static int x_dir[4]= {-1, 0, 1, 0};
    static int y_dir[4]= { 0,-1, 0, 1};
    int improvement[2]={100000, 100000};
    int dirs[2]={2, 3};
    int dir;
    int last_dir= -1;
    
    for(;;){
        dir= dirs[ improvement[0] > improvement[1] ? 0 : 1 ];
        if(improvement[dir&1]==-1) return dmin;
        
        {
            const int x= best[0] + x_dir[dir];
            const int y= best[1] + y_dir[dir];
            const int key= (y<<ME_MAP_MV_BITS) + x + map_generation;
            const int index= ((y<<ME_MAP_SHIFT) + x)&(ME_MAP_SIZE-1);
            int d;
            if(x>=xmin && x<=xmax && y>=ymin && y<=ymax){
                if(map[index]!=key){
                    d = pix_abs(new_pic, old_pic + x + y*pic_stride, pic_stride);
                    d += (mv_penalty[(x<<shift)-pred_x] + mv_penalty[(y<<shift)-pred_y])*quant;
                    map[index]=key;
                    score_map[index]=d;
                    if(d<dmin){
                        improvement[dir&1]= dmin-d;
                        improvement[(dir&1)^1]++;
                        dmin=d;
                        best[0]= x;
                        best[1]= y;
                        last_dir=dir;
                        continue;
                    }
                }else{
                    d= score_map[index];
                }
            }else{
                d= dmin + 1000; //FIXME is this a good idea?
            }
            /* evaluated point was cached or checked and worse */

            if(last_dir==dir){
                improvement[dir&1]= -1;
            }else{
                improvement[dir&1]= d-dmin;
                last_dir= dirs[dir&1]= dir^2;
            }
        }
    }
}

static inline int update_map_generation(MpegEncContext * s)
{
    s->me_map_generation+= 1<<(ME_MAP_MV_BITS*2);
    if(s->me_map_generation==0){
        s->me_map_generation= 1<<(ME_MAP_MV_BITS*2);
        memset(s->me_map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
    }
    return s->me_map_generation;
}

static int epzs_motion_search(MpegEncContext * s,
                             int *mx_ptr, int *my_ptr,
                             int P[10][2], int pred_x, int pred_y,
                             int xmin, int ymin, int xmax, int ymax, uint8_t * ref_picture)
{
    int best[2]={0, 0};
    int d, dmin; 
    UINT8 *new_pic, *old_pic;
    const int pic_stride= s->linesize;
    const int pic_xy= (s->mb_y*pic_stride + s->mb_x)*16;
    UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
    int quant= s->qscale; // qscale of the prev frame
    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me_map;
    uint16_t *score_map= s->me_score_map;
    int map_generation;

    new_pic = s->new_picture[0] + pic_xy;
    old_pic = ref_picture + pic_xy;
    
    map_generation= update_map_generation(s);

    dmin = s->dsp.pix_abs16x16(new_pic, old_pic, pic_stride);
    map[0]= map_generation;
    score_map[0]= dmin;

    /* first line */
    if ((s->mb_y == 0 || s->first_slice_line)) {
        CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
        CHECK_MV(P_LAST[0]>>shift, P_LAST[1]>>shift)
    }else{
        if(dmin<256 && ( P_LEFT[0]    |P_LEFT[1]
                        |P_TOP[0]     |P_TOP[1]
                        |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
            *mx_ptr= 0;
            *my_ptr= 0;
            s->skip_me=1;
            return dmin;
        }
        CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
        if(dmin>256*2){
            CHECK_MV(P_LAST[0]    >>shift, P_LAST[1]    >>shift)
            CHECK_MV(P_LEFT[0]    >>shift, P_LEFT[1]    >>shift)
            CHECK_MV(P_TOP[0]     >>shift, P_TOP[1]     >>shift)
            CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
        }
    }
    if(dmin>256*4){
        CHECK_MV(P_LAST_RIGHT[0] >>shift, P_LAST_RIGHT[1] >>shift)
        CHECK_MV(P_LAST_BOTTOM[0]>>shift, P_LAST_BOTTOM[1]>>shift)
    }
#if 0 //doest only slow things down
    if(dmin>512*3){
        int step;
        dmin= score_map[0];
        best[0]= best[1]=0;
        for(step=128; step>0; step>>=1){
            const int step2= step;
            int y;
            for(y=-step2+best[1]; y<=step2+best[1]; y+=step){
                int x;
                if(y<ymin || y>ymax) continue;

                for(x=-step2+best[0]; x<=step2+best[0]; x+=step){
                    if(x<xmin || x>xmax) continue;
                    if(x==best[0] && y==best[1]) continue;
                    CHECK_MV(x,y)
                }
            }
        }
    }
#endif
//check(best[0],best[1],0, b0)
    if(s->me_method==ME_EPZS)
        dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
				   shift, map, score_map, map_generation, s->dsp.pix_abs16x16);
    else
        dmin=         cross_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, s->dsp.pix_abs16x16);
//check(best[0],best[1],0, b1)
    *mx_ptr= best[0];
    *my_ptr= best[1];    

//    printf("%d %d %d \n", best[0], best[1], dmin);
    return dmin;
}

static int epzs_motion_search4(MpegEncContext * s, int block,
                             int *mx_ptr, int *my_ptr,
                             int P[10][2], int pred_x, int pred_y,
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
{
    int best[2]={0, 0};
    int d, dmin; 
    UINT8 *new_pic, *old_pic;
    const int pic_stride= s->linesize;
    const int pic_xy= ((s->mb_y*2 + (block>>1))*pic_stride + s->mb_x*2 + (block&1))*8;
    UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
    int quant= s->qscale; // qscale of the prev frame
    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me_map;
    uint16_t *score_map= s->me_score_map;
    int map_generation;

    new_pic = s->new_picture[0] + pic_xy;
    old_pic = ref_picture + pic_xy;

    map_generation= update_map_generation(s);

    dmin = 1000000;
//printf("%d %d %d %d //",xmin, ymin, xmax, ymax); 
    /* first line */
    if ((s->mb_y == 0 || s->first_slice_line) && block<2) {
	CHECK_MV4(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
        CHECK_MV4(P_LAST[0]>>shift, P_LAST[1]>>shift)
        CHECK_MV4(P_MV1[0]>>shift, P_MV1[1]>>shift)
    }else{
        CHECK_MV4(P_MV1[0]>>shift, P_MV1[1]>>shift)
        //FIXME try some early stop
        if(dmin>64*2){
            CHECK_MV4(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
            CHECK_MV4(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
            CHECK_MV4(P_TOP[0]>>shift, P_TOP[1]>>shift)
            CHECK_MV4(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
            CHECK_MV4(P_LAST[0]>>shift, P_LAST[1]>>shift)
        }
    }
    if(dmin>64*4){
        CHECK_MV4(P_LAST_RIGHT[0]>>shift, P_LAST_RIGHT[1]>>shift)
        CHECK_MV4(P_LAST_BOTTOM[0]>>shift, P_LAST_BOTTOM[1]>>shift)
    }

    if(s->me_method==ME_EPZS)
        dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
				   shift, map, score_map, map_generation, s->dsp.pix_abs8x8);
    else
        dmin=         cross_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, s->dsp.pix_abs8x8);

    *mx_ptr= best[0];
    *my_ptr= best[1];    

//    printf("%d %d %d \n", best[0], best[1], dmin);
    return dmin;
}

#define CHECK_HALF_MV(suffix, x, y) \
{\
    d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*quant;\
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
}

    
/* The idea would be to make half pel ME after Inter/Intra decision to 
   save time. */
static inline int halfpel_motion_search(MpegEncContext * s,
				  int *mx_ptr, int *my_ptr, int dmin,
				  int xmin, int ymin, int xmax, int ymax,
                                  int pred_x, int pred_y, uint8_t *ref_picture,
                                  op_pixels_abs_func pix_abs_x2, 
                                  op_pixels_abs_func pix_abs_y2, op_pixels_abs_func pix_abs_xy2, int n)
{
    UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
    const int quant= s->qscale;
    int mx, my, xx, yy, dminh;
    UINT8 *pix, *ptr;

    if(s->skip_me){
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }

    xx = 16 * s->mb_x + 8*(n&1);
    yy = 16 * s->mb_y + 8*(n>>1);
    pix =  s->new_picture[0] + (yy * s->linesize) + xx;

    mx = *mx_ptr;
    my = *my_ptr;
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
    
    dminh = dmin;

    if (mx > xmin && mx < xmax && 
        my > ymin && my < ymax) {
        int dx=0, dy=0;
        int d, pen_x, pen_y; 

        mx<<=1;
        my<<=1;
        
        pen_x= pred_x + mx;
        pen_y= pred_y + my;

        ptr-= s->linesize;
        CHECK_HALF_MV(xy2, -1, -1)
        CHECK_HALF_MV(y2 ,  0, -1)
        CHECK_HALF_MV(xy2, +1, -1)
        
        ptr+= s->linesize;
        CHECK_HALF_MV(x2 , -1,  0)
        CHECK_HALF_MV(x2 , +1,  0)
        CHECK_HALF_MV(xy2, -1, +1)
        CHECK_HALF_MV(y2 ,  0, +1)
        CHECK_HALF_MV(xy2, +1, +1)

        mx+=dx;
        my+=dy;
    }else{
        mx<<=1;
        my<<=1;
    }

    *mx_ptr = mx;
    *my_ptr = my;
    return dminh;
}

static inline int fast_halfpel_motion_search(MpegEncContext * s,
				  int *mx_ptr, int *my_ptr, int dmin,
				  int xmin, int ymin, int xmax, int ymax,
                                  int pred_x, int pred_y, uint8_t *ref_picture,
                                  op_pixels_abs_func pix_abs_x2, 
                                  op_pixels_abs_func pix_abs_y2, op_pixels_abs_func pix_abs_xy2, int n)
{
    UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
    uint16_t *score_map= s->me_score_map;
    const int quant= s->qscale;
    int mx, my, xx, yy, dminh;
    UINT8 *pix, *ptr;

    if(s->skip_me){
//    printf("S");
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }
//    printf("N");
        
    xx = 16 * s->mb_x + 8*(n&1);
    yy = 16 * s->mb_y + 8*(n>>1);
    pix =  s->new_picture[0] + (yy * s->linesize) + xx;

    mx = *mx_ptr;
    my = *my_ptr;
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
    
    dminh = dmin;

    if (mx > xmin && mx < xmax && 
        my > ymin && my < ymax) {
        int dx=0, dy=0;
        int d, pen_x, pen_y; 
        const int index= (my<<ME_MAP_SHIFT) + mx;
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
        mx<<=1;
        my<<=1;

        
        pen_x= pred_x + mx;
        pen_y= pred_y + my;

        ptr-= s->linesize;
        if(t<=b){
            CHECK_HALF_MV(y2 ,  0, -1)
            if(l<=r){
                CHECK_HALF_MV(xy2, -1, -1)
                if(t+r<=b+l){
                    CHECK_HALF_MV(xy2, +1, -1)
                    ptr+= s->linesize;
                }else{
                    ptr+= s->linesize;
                    CHECK_HALF_MV(xy2, -1, +1)
                }
                CHECK_HALF_MV(x2 , -1,  0)
            }else{
                CHECK_HALF_MV(xy2, +1, -1)
                if(t+l<=b+r){
                    CHECK_HALF_MV(xy2, -1, -1)
                    ptr+= s->linesize;
                }else{
                    ptr+= s->linesize;
                    CHECK_HALF_MV(xy2, +1, +1)
                }
                CHECK_HALF_MV(x2 , +1,  0)
            }