motion_est_template.c

arm平台下的H264编码和解码源代码

/*
 * Motion estimation 
 * Copyright (c) 2002-2004 Michael Niedermayer
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 *
 */
 
/**
 * @file motion_est_template.c
 * Motion estimation template.
 */
//FIXME ref2_y next_pic?
//lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
//Note, the last line is there to kill these ugly unused var warnings
#define LOAD_COMMON\
    uint32_t * const score_map= s->me.score_map;\
    const int time_pp= s->pp_time;\
    const int time_pb= s->pb_time;\
    const int xmin= s->me.xmin;\
    const int ymin= s->me.ymin;\
    const int xmax= s->me.xmax;\
    const int ymax= s->me.ymax;\
    uint8_t * const src_y= src_data[0];\
    uint8_t * const src_u= src_data[1];\
    uint8_t * const src_v= src_data[2];\
    uint8_t * const ref_y= ref_data[0];\
    uint8_t * const ref_u= ref_data[1];\
    uint8_t * const ref_v= ref_data[2];\
    op_pixels_func (*hpel_put)[4];\
    op_pixels_func (*hpel_avg)[4]= &s->dsp.avg_pixels_tab[size];\
    op_pixels_func (*chroma_hpel_put)[4];\
    qpel_mc_func (*qpel_put)[16];\
    qpel_mc_func (*qpel_avg)[16]= &s->dsp.avg_qpel_pixels_tab[size];\
    const __attribute__((unused)) int unu= time_pp + time_pb + (size_t)src_u + (size_t)src_v + (size_t)ref_u + (size_t)ref_v\
                                           + (size_t)hpel_avg + (size_t)qpel_avg + (size_t)score_map\
                                           + xmin + xmax + ymin + ymax;\
    if(s->no_rounding /*FIXME b_type*/){\
        hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];\
        chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];\
        qpel_put= &s->dsp.put_no_rnd_qpel_pixels_tab[size];\
    }else{\
        hpel_put=& s->dsp.put_pixels_tab[size];\
        chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];\
        qpel_put= &s->dsp.put_qpel_pixels_tab[size];\
    }


#ifdef CMP_HPEL
    
#define CHECK_HALF_MV(dx, dy, x, y)\
{\
    const int hx= 2*(x)+(dx);\
    const int hy= 2*(y)+(dy);\
    CMP_HPEL(d, dx, dy, x, y, size);\
    d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
    COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
}

#if 0
static int RENAME(hpel_motion_search)(MpegEncContext * s,
				  int *mx_ptr, int *my_ptr, int dmin,
                                  int pred_x, int pred_y, uint8_t *ref_data[3], 
                                  int size, uint8_t * const mv_penalty)
{
    const int xx = 16 * s->mb_x + 8*(n&1);
    const int yy = 16 * s->mb_y + 8*(n>>1);
    const int mx = *mx_ptr;
    const int my = *my_ptr;
    const int penalty_factor= s->me.sub_penalty_factor;
    
    LOAD_COMMON
    
 //   INIT;
 //FIXME factorize
    me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub;

    if(s->no_rounding /*FIXME b_type*/){
        hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];
        chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];
    }else{
        hpel_put=& s->dsp.put_pixels_tab[size];
        chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];
    }
    cmp= s->dsp.me_cmp[size];
    chroma_cmp= s->dsp.me_cmp[size+1];
    cmp_sub= s->dsp.me_sub_cmp[size];
    chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];

    if(s->me.skip){ //FIXME somehow move up (benchmark)
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }
        
    if(s->avctx->me_cmp != s->avctx->me_sub_cmp){
        CMP_HPEL(dmin, 0, 0, mx, my, size);
        if(mx || my)
            dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
    }
        
    if (mx > xmin && mx < xmax && 
        my > ymin && my < ymax) {
        int bx=2*mx, by=2*my;
        int d= dmin;
        
        CHECK_HALF_MV(1, 1, mx-1, my-1)
        CHECK_HALF_MV(0, 1, mx  , my-1)        
        CHECK_HALF_MV(1, 1, mx  , my-1)
        CHECK_HALF_MV(1, 0, mx-1, my  )
        CHECK_HALF_MV(1, 0, mx  , my  )
        CHECK_HALF_MV(1, 1, mx-1, my  )
        CHECK_HALF_MV(0, 1, mx  , my  )        
        CHECK_HALF_MV(1, 1, mx  , my  )

        assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2);

        *mx_ptr = bx;
        *my_ptr = by;
    }else{
        *mx_ptr =2*mx;
        *my_ptr =2*my;
    }

    return dmin;
}

#else
static int RENAME(hpel_motion_search)(MpegEncContext * s,
				  int *mx_ptr, int *my_ptr, int dmin,
                                  int pred_x, int pred_y, uint8_t *src_data[3], 
                                  uint8_t *ref_data[3], int stride, int uvstride,
                                  int size, int h, uint8_t * const mv_penalty)
{
    const int mx = *mx_ptr;
    const int my = *my_ptr;   
    const int penalty_factor= s->me.sub_penalty_factor;
    me_cmp_func cmp_sub, chroma_cmp_sub;
    int bx=2*mx, by=2*my;

    LOAD_COMMON
    
 //FIXME factorize

    cmp_sub= s->dsp.me_sub_cmp[size];
    chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];

    if(s->me.skip){ //FIXME move out of hpel?
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }
        
    if(s->avctx->me_cmp != s->avctx->me_sub_cmp){
        CMP_HPEL(dmin, 0, 0, mx, my, size);
        if(mx || my || size>0)
            dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
    }
        
    if (mx > xmin && mx < xmax && 
        my > ymin && my < ymax) {
        int d= dmin;
        const int index= (my<<ME_MAP_SHIFT) + mx;
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] 
                     + (mv_penalty[bx   - pred_x] + mv_penalty[by-2 - pred_y])*s->me.penalty_factor;
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)]
                     + (mv_penalty[bx-2 - pred_x] + mv_penalty[by   - pred_y])*s->me.penalty_factor;
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)]
                     + (mv_penalty[bx+2 - pred_x] + mv_penalty[by   - pred_y])*s->me.penalty_factor;
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
                     + (mv_penalty[bx   - pred_x] + mv_penalty[by+2 - pred_y])*s->me.penalty_factor;
    
#if 1
        int key;
        int map_generation= s->me.map_generation;
#ifndef NDEBUG
        uint32_t *map= s->me.map;
#endif
        key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
        assert(map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
        key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
        assert(map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
        key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation;
        assert(map[(index+1)&(ME_MAP_SIZE-1)] == key);
        key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;
        assert(map[(index-1)&(ME_MAP_SIZE-1)] == key);
#endif                
        if(t<=b){
            CHECK_HALF_MV(0, 1, mx  ,my-1)
            if(l<=r){
                CHECK_HALF_MV(1, 1, mx-1, my-1)
                if(t+r<=b+l){
                    CHECK_HALF_MV(1, 1, mx  , my-1)
                }else{
                    CHECK_HALF_MV(1, 1, mx-1, my  )
                }
                CHECK_HALF_MV(1, 0, mx-1, my  )
            }else{
                CHECK_HALF_MV(1, 1, mx  , my-1)
                if(t+l<=b+r){
                    CHECK_HALF_MV(1, 1, mx-1, my-1)
                }else{
                    CHECK_HALF_MV(1, 1, mx  , my  )
                }
                CHECK_HALF_MV(1, 0, mx  , my  )
            }
        }else{
            if(l<=r){
                if(t+l<=b+r){
                    CHECK_HALF_MV(1, 1, mx-1, my-1)
                }else{
                    CHECK_HALF_MV(1, 1, mx  , my  )
                }
                CHECK_HALF_MV(1, 0, mx-1, my)
                CHECK_HALF_MV(1, 1, mx-1, my)
            }else{
                if(t+r<=b+l){
                    CHECK_HALF_MV(1, 1, mx  , my-1)
                }else{
                    CHECK_HALF_MV(1, 1, mx-1, my)
                }
                CHECK_HALF_MV(1, 0, mx  , my)
                CHECK_HALF_MV(1, 1, mx  , my)
            }
            CHECK_HALF_MV(0, 1, mx  , my)
        }
        assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);
    }

    *mx_ptr = bx;
    *my_ptr = by;
    
    return dmin;
}
#endif

static int RENAME(hpel_get_mb_score)(MpegEncContext * s, int mx, int my, int pred_x, int pred_y, uint8_t *src_data[3], 
                                  uint8_t *ref_data[3], int stride, int uvstride,
                                  uint8_t * const mv_penalty)
{
//    const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
    const int size= 0;
    const int h= 16;
    const int penalty_factor= s->me.mb_penalty_factor;
    me_cmp_func cmp_sub, chroma_cmp_sub;
    int d;

    LOAD_COMMON
    
 //FIXME factorize

    cmp_sub= s->dsp.mb_cmp[size];
    chroma_cmp_sub= s->dsp.mb_cmp[size+1];
    
    assert(!s->me.skip);
    assert(s->avctx->me_sub_cmp != s->avctx->mb_cmp);

    CMP_HPEL(d, mx&1, my&1, mx>>1, my>>1, size);
    //FIXME check cbp before adding penalty for (0,0) vector
    if(mx || my || size>0)
        d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
        
    return d;
}

#endif /* CMP_HPEL */



#ifdef CMP_QPEL

#define CHECK_QUARTER_MV(dx, dy, x, y)\
{\
    const int hx= 4*(x)+(dx);\
    const int hy= 4*(y)+(dy);\
    CMP_QPEL(d, dx, dy, x, y, size);\
    d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
    COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
}

static int RENAME(qpel_motion_search)(MpegEncContext * s,
				  int *mx_ptr, int *my_ptr, int dmin,
                                  int pred_x, int pred_y, uint8_t *src_data[3], 
                                  uint8_t *ref_data[3], int stride, int uvstride,                                  
                                  int size, int h, uint8_t * const mv_penalty)
{
    const int mx = *mx_ptr;
    const int my = *my_ptr;   
    const int penalty_factor= s->me.sub_penalty_factor;
    const int map_generation= s->me.map_generation;
    const int subpel_quality= s->avctx->me_subpel_quality;
    uint32_t *map= s->me.map;
    me_cmp_func cmp, chroma_cmp;
    me_cmp_func cmp_sub, chroma_cmp_sub;

    LOAD_COMMON
    
    cmp= s->dsp.me_cmp[size];
    chroma_cmp= s->dsp.me_cmp[size+1]; //factorize FIXME
 //FIXME factorize

    cmp_sub= s->dsp.me_sub_cmp[size];
    chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];

    if(s->me.skip){ //FIXME somehow move up (benchmark)
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }
        
    if(s->avctx->me_cmp != s->avctx->me_sub_cmp){
        CMP_QPEL(dmin, 0, 0, mx, my, size);
        if(mx || my || size>0)
            dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
    }
        
    if (mx > xmin && mx < xmax && 
        my > ymin && my < ymax) {
        int bx=4*mx, by=4*my;
        int d= dmin;
        int i, nx, ny;
        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)];
        const int c= score_map[(index                    )&(ME_MAP_SIZE-1)];
        int best[8];
        int best_pos[8][2];
        
        memset(best, 64, sizeof(int)*8);
#if 1
        if(s->me.dia_size>=2){        
            const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
            const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
            const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
            const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];

            for(ny= -3; ny <= 3; ny++){
                for(nx= -3; nx <= 3; nx++){
                    const int t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
                    const int c2= nx*nx*( r +  l - 2*c) + 4*nx*( r- l) + 32*c;
                    const int b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
                    int score= ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2;
                    int i;
                    
                    if((nx&3)==0 && (ny&3)==0) continue;
                    
                    score += 1024*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
                    
//                    if(nx&1) score-=1024*s->me.penalty_factor;
//                    if(ny&1) score-=1024*s->me.penalty_factor;
                    
                    for(i=0; i<8; i++){
                        if(score < best[i]){
                            memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
                            memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
                            best[i]= score;
                            best_pos[i][0]= nx + 4*mx;
                            best_pos[i][1]= ny + 4*my;
                            break;
                        }
                    }
                }
            }
        }else{
            int tl;
            const int cx = 4*(r - l);
            const int cx2= r + l - 2*c; 
            const int cy = 4*(b - t);
            const int cy2= b + t - 2*c;
            int cxy;
              
            if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME
                tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];