motion_est_template.c

arm平台下的H264编码和解码源代码

    for(i=0; i<minima_count; i++){
        const int x= minima[i].x;
        const int y= minima[i].y;
        int d;
        
        if(minima[i].checked) continue;
        
        if(   x >= xmax || x <= xmin
           || y >= ymax || y <= ymin)
           continue;

        SAB_CHECK_MV(x-1, y)
        SAB_CHECK_MV(x+1, y)
        SAB_CHECK_MV(x  , y-1)
        SAB_CHECK_MV(x  , y+1)
        
        minima[i].checked= 1;
    }
    
    best[0]= minima[0].x;
    best[1]= minima[0].y;
    dmin= minima[0].height;
    
    if(   best[0] < xmax && best[0] > xmin
       && best[1] < ymax && best[1] > ymin){
        int d;
        //ensure that the refernece samples for hpel refinement are in the map
        CHECK_MV(best[0]-1, best[1])
        CHECK_MV(best[0]+1, best[1])
        CHECK_MV(best[0], best[1]-1)
        CHECK_MV(best[0], best[1]+1)
    }
    return dmin;    
}

static inline int RENAME(var_diamond_search)(MpegEncContext * s, int *best, int dmin,
                                       uint8_t *src_data[3],
                                       uint8_t *ref_data[3], int stride, int uvstride,
                                       int const pred_x, int const pred_y, int const penalty_factor,
                                       int const shift,
                                       uint32_t *map, int map_generation, int size, int h, uint8_t * const mv_penalty
                                       )
{
    me_cmp_func cmp, chroma_cmp;
    int dia_size;
    LOAD_COMMON
    
    cmp= s->dsp.me_cmp[size];
    chroma_cmp= s->dsp.me_cmp[size+1];

    for(dia_size=1; dia_size<=s->me.dia_size; dia_size++){
        int dir, start, end;
        const int x= best[0];
        const int y= best[1];

        start= FFMAX(0, y + dia_size - ymax);
        end  = FFMIN(dia_size, xmax - x + 1);
        for(dir= start; dir<end; dir++){
            int d;

//check(x + dir,y + dia_size - dir,0, a0)
            CHECK_MV(x + dir           , y + dia_size - dir);
        }

        start= FFMAX(0, x + dia_size - xmax);
        end  = FFMIN(dia_size, y - ymin + 1);
        for(dir= start; dir<end; dir++){
            int d;

//check(x + dia_size - dir, y - dir,0, a1)
            CHECK_MV(x + dia_size - dir, y - dir           );
        }

        start= FFMAX(0, -y + dia_size + ymin );
        end  = FFMIN(dia_size, x - xmin + 1);
        for(dir= start; dir<end; dir++){
            int d;

//check(x - dir,y - dia_size + dir,0, a2)
            CHECK_MV(x - dir           , y - dia_size + dir);
        }

        start= FFMAX(0, -x + dia_size + xmin );
        end  = FFMIN(dia_size, ymax - y + 1);
        for(dir= start; dir<end; dir++){
            int d;

//check(x - dia_size + dir, y + dir,0, a3)
            CHECK_MV(x - dia_size + dir, y + dir           );
        }

        if(x!=best[0] || y!=best[1])
            dia_size=0;
#if 0
{
int dx, dy, i;
static int stats[8*8];
dx= ABS(x-best[0]);
dy= ABS(y-best[1]);
stats[dy*8 + dx] ++;
if(256*256*256*64 % (stats[0]+1)==0){
    for(i=0; i<64; i++){
        if((i&7)==0) printf("\n");
        printf("%6d ", stats[i]);
    }
    printf("\n");
}
}
#endif
    }
    return dmin;    
}

static int RENAME(epzs_motion_search)(MpegEncContext * s,
                             int *mx_ptr, int *my_ptr,
                             int P[10][2], int pred_x, int pred_y, uint8_t *src_data[3], 
                             uint8_t *ref_data[3], int stride, int uvstride, int16_t (*last_mv)[2], 
                             int ref_mv_scale, uint8_t * const mv_penalty)
{
    int best[2]={0, 0};
    int d, dmin; 
    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me.map;
    int map_generation;
    const int penalty_factor= s->me.penalty_factor;
    const int size=0;
    const int h=16;
    const int ref_mv_stride= s->mb_stride; //pass as arg  FIXME
    const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
    me_cmp_func cmp, chroma_cmp;
    LOAD_COMMON
    
    cmp= s->dsp.me_cmp[size];
    chroma_cmp= s->dsp.me_cmp[size+1];
    
    map_generation= update_map_generation(s);

    CMP(dmin, 0, 0, size);
    map[0]= map_generation;
    score_map[0]= dmin;

    /* first line */
    if (s->first_slice_line) {
        CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
        CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                        (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
    }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->me.skip=1;
            return dmin;
        }
        CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
        if(dmin>256*2){
            CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                            (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
            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){
        if(s->me.pre_pass){
            CHECK_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16, 
                            (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
            if(!s->first_slice_line)
                CHECK_CLIPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 
                                (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
        }else{
            CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, 
                            (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
            if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line
                CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 
                                (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
        }
    }

    if(s->avctx->last_predictor_count){
        const int count= s->avctx->last_predictor_count;
        const int xstart= FFMAX(0, s->mb_x - count);
        const int ystart= FFMAX(0, s->mb_y - count);
        const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
        const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
        int mb_y;

        for(mb_y=ystart; mb_y<yend; mb_y++){
            int mb_x;
            for(mb_x=xstart; mb_x<xend; mb_x++){
                const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
                int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
                int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;

                if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
                CHECK_MV(mx,my)
            }
        }
    }

//check(best[0],best[1],0, b0)
    if(s->me.dia_size==-1)
        dmin= RENAME(funny_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else if(s->me.dia_size<-1)
        dmin= RENAME(sab_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else if(s->me.dia_size<2)
        dmin= RENAME(small_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else
        dmin= RENAME(var_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);

//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;
}

#ifndef CMP_DIRECT /* no 4mv search needed in direct mode */
static int RENAME(epzs_motion_search4)(MpegEncContext * s,
                             int *mx_ptr, int *my_ptr,
                             int P[10][2], int pred_x, int pred_y,
                             uint8_t *src_data[3], 
                             uint8_t *ref_data[3], int stride, int uvstride, int16_t (*last_mv)[2], 
                             int ref_mv_scale, uint8_t * const mv_penalty)
{
    int best[2]={0, 0};
    int d, dmin; 
    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me.map;
    int map_generation;
    const int penalty_factor= s->me.penalty_factor;
    const int size=1;
    const int h=8;
    const int ref_mv_stride= s->mb_stride;
    const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
    me_cmp_func cmp, chroma_cmp;
    LOAD_COMMON
    
    cmp= s->dsp.me_cmp[size];
    chroma_cmp= s->dsp.me_cmp[size+1];

    map_generation= update_map_generation(s);

    dmin = 1000000;
//printf("%d %d %d %d //",xmin, ymin, xmax, ymax); 
    /* first line */
    if (s->first_slice_line) {
	CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
        CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                        (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
        CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
    }else{
        CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
        //FIXME try some early stop
        if(dmin>64*2){
            CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[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)
            CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                            (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
        }
    }
    if(dmin>64*4){
        CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, 
                        (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
        if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line
            CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 
                            (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
    }

    if(s->me.dia_size==-1)
        dmin= RENAME(funny_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else if(s->me.dia_size<-1)
        dmin= RENAME(sab_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else if(s->me.dia_size<2)
        dmin= RENAME(small_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else
        dmin= RENAME(var_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);


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

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

//try to merge with above FIXME (needs PSNR test)
static int RENAME(epzs_motion_search2)(MpegEncContext * s,
                             int *mx_ptr, int *my_ptr,
                             int P[10][2], int pred_x, int pred_y,
                             uint8_t *src_data[3], 
                             uint8_t *ref_data[3], int stride, int uvstride, int16_t (*last_mv)[2], 
                             int ref_mv_scale, uint8_t * const mv_penalty)
{
    int best[2]={0, 0};
    int d, dmin; 
    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me.map;
    int map_generation;
    const int penalty_factor= s->me.penalty_factor;
    const int size=0; //FIXME pass as arg
    const int h=8;
    const int ref_mv_stride= s->mb_stride;
    const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
    me_cmp_func cmp, chroma_cmp;
    LOAD_COMMON
    
    cmp= s->dsp.me_cmp[size];
    chroma_cmp= s->dsp.me_cmp[size+1];

    map_generation= update_map_generation(s);

    dmin = 1000000;
//printf("%d %d %d %d //",xmin, ymin, xmax, ymax); 
    /* first line */
    if (s->first_slice_line) {
	CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
        CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                        (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
        CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
    }else{
        CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
        //FIXME try some early stop
        if(dmin>64*2){
            CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[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)
            CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, 
                            (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
        }
    }
    if(dmin>64*4){
        CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, 
                        (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
        if(s->mb_y+1<s->end_mb_y)  //FIXME replace at least with last_slice_line
            CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, 
                            (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
    }

    if(s->me.dia_size==-1)
        dmin= RENAME(funny_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else if(s->me.dia_size<-1)
        dmin= RENAME(sab_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else if(s->me.dia_size<2)
        dmin= RENAME(small_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);
    else
        dmin= RENAME(var_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride,
                                   pred_x, pred_y, penalty_factor, 
				   shift, map, map_generation, size, h, mv_penalty);


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

//    printf("%d %d %d \n", best[0], best[1], dmin);
    return dmin;
}
#endif /* !CMP_DIRECT */