motion_est_template.c

从FFMPEG转换而来的H264解码程序,VC下编译..

/*
 * Motion estimation
 * Copyright (c) 2002-2004 Michael Niedermayer
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg 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 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg 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 FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 *
 */

/**
 * @file motion_est_template.c
 * Motion estimation template.
 */

//lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
#define LOAD_COMMON\
    uint32_t av_unused * const score_map= c->score_map;\
    const int av_unused xmin= c->xmin;\
    const int av_unused ymin= c->ymin;\
    const int av_unused xmax= c->xmax;\
    const int av_unused ymax= c->ymax;\
    uint8_t *mv_penalty= c->current_mv_penalty;\
    const int pred_x= c->pred_x;\
    const int pred_y= c->pred_y;\

#define CHECK_HALF_MV(dx, dy, x, y)\
{\
    const int hx= 2*(x)+(dx);\
    const int hy= 2*(y)+(dy);\
    d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\
    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 hpel_motion_search)(MpegEncContext * s,
                                  int *mx_ptr, int *my_ptr, int dmin,
                                  uint8_t *ref_data[3],
                                  int size)
{
    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= c->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];
    }
    cmpf= s->dsp.me_cmp[size];
    chroma_cmpf= 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(c->skip){ //FIXME somehow move up (benchmark)
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }

    if(c->avctx->me_cmp != c->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 hpel_motion_search(MpegEncContext * s,
                                  int *mx_ptr, int *my_ptr, int dmin,
                                  int src_index, int ref_index,
                                  int size, int h)
{
    MotionEstContext * const c= &s->me;
    const int mx = *mx_ptr;
    const int my = *my_ptr;
    const int penalty_factor= c->sub_penalty_factor;
    me_cmp_func cmp_sub, chroma_cmp_sub;
    int bx=2*mx, by=2*my;

    LOAD_COMMON
    int flags= c->sub_flags;

 //FIXME factorize

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

    if(c->skip){ //FIXME move out of hpel?
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }

    if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
        dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
        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])*c->penalty_factor;
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)]
                     + (mv_penalty[bx-2 - pred_x] + mv_penalty[by   - pred_y])*c->penalty_factor;
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)]
                     + (mv_penalty[bx+2 - pred_x] + mv_penalty[by   - pred_y])*c->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])*c->penalty_factor;

#if 1
        int key;
        int map_generation= c->map_generation;
#ifndef NDEBUG
        uint32_t *map= c->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 no_sub_motion_search(MpegEncContext * s,
          int *mx_ptr, int *my_ptr, int dmin,
                                  int src_index, int ref_index,
                                  int size, int h)
{
    (*mx_ptr)<<=1;
    (*my_ptr)<<=1;
    return dmin;
}

inline int ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index,
                               int ref_index, int size, int h, int add_rate)
{
//    const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
    MotionEstContext * const c= &s->me;
    const int penalty_factor= c->mb_penalty_factor;
    const int flags= c->mb_flags;
    const int qpel= flags & FLAG_QPEL;
    const int mask= 1+2*qpel;
    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(!c->skip);
//    assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp);

    d= cmp(s, mx>>(qpel+1), my>>(qpel+1), mx&mask, my&mask, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
    //FIXME check cbp before adding penalty for (0,0) vector
    if(add_rate && (mx || my || size>0))
        d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;

    return d;
}

#define CHECK_QUARTER_MV(dx, dy, x, y)\
{\
    const int hx= 4*(x)+(dx);\
    const int hy= 4*(y)+(dy);\
    d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
    d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
    COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
}

static int qpel_motion_search(MpegEncContext * s,
                                  int *mx_ptr, int *my_ptr, int dmin,
                                  int src_index, int ref_index,
                                  int size, int h)
{
    MotionEstContext * const c= &s->me;
    const int mx = *mx_ptr;
    const int my = *my_ptr;
    const int penalty_factor= c->sub_penalty_factor;
    const int map_generation= c->map_generation;
    const int subpel_quality= c->avctx->me_subpel_quality;
    uint32_t *map= c->map;
    me_cmp_func cmpf, chroma_cmpf;
    me_cmp_func cmp_sub, chroma_cmp_sub;

    LOAD_COMMON
    int flags= c->sub_flags;

    cmpf= s->dsp.me_cmp[size];
    chroma_cmpf= 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(c->skip){ //FIXME somehow move up (benchmark)
        *mx_ptr = 0;
        *my_ptr = 0;
        return dmin;
    }

    if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
        dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
        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++){
                    //FIXME this could overflow (unlikely though)
                    const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
                    const int64_t c2= nx*nx*( r +  l - 2*c) + 4*nx*( r- l) + 32*c;
                    const int64_t 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 + 512)>>10;
                    int i;

                    if((nx&3)==0 && (ny&3)==0) continue;

                    score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;

//                    if(nx&1) score-=1024*c->penalty_factor;
//                    if(ny&1) score-=1024*c->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;
            //FIXME this could overflow (unlikely though)
            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)];
            }else{
                tl= cmp(s, mx-1, my-1, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);//FIXME wrong if chroma me is different
            }

            cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;

            assert(16*cx2 + 4*cx + 32*c == 32*r);
            assert(16*cx2 - 4*cx + 32*c == 32*l);
            assert(16*cy2 + 4*cy + 32*c == 32*b);
            assert(16*cy2 - 4*cy + 32*c == 32*t);
            assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);

            for(ny= -3; ny <= 3; ny++){
                for(nx= -3; nx <= 3; nx++){
                    //FIXME this could overflow (unlikely though)
                    int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
                    int i;

                    if((nx&3)==0 && (ny&3)==0) continue;

                    score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
//                    if(nx&1) score-=32*c->penalty_factor;
  //                  if(ny&1) score-=32*c->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;
                        }
                    }
                }
            }
        }
        for(i=0; i<subpel_quality; i++){
            nx= best_pos[i][0];
            ny= best_pos[i][1];
            CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
        }

#if 0
            const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];