motion_est.c

arm平台下的H264编码和解码源代码

        ptr = ref_data[0] + (src_y * stride) + src_x;
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);

        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
        src_x = motion_bx >> 2;
        src_y = motion_by >> 2;
    
        ptr = ref_data[3] + (src_y * stride) + src_x;
        s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
    }else{
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
        src_x = motion_fx >> 1;
        src_y = motion_fy >> 1;

        ptr = ref_data[0] + (src_y * stride) + src_x;
        s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
        src_x = motion_bx >> 1;
        src_y = motion_by >> 1;
    
        ptr = ref_data[3] + (src_y * stride) + src_x;
        s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
    }

    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
           + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
           
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
    }
    //FIXME CHROMA !!!
           
    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, uint8_t *src_data[3], uint8_t *ref_data[6],
                                  int stride, int uvstride,
                                  int mb_x, int mb_y)
{
    const int mot_stride = s->mb_stride;
    const int xy = mb_y *mot_stride + mb_x;
    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, src_data, ref_data, stride, uvstride,
                          motion_fx, motion_fy,
                          motion_bx, motion_by,
                          pred_fx, pred_fy,
                          pred_bx, pred_by,
                          0, 16);

   return fbmin;
}

static inline int direct_search(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
                                int stride, int uvstride,
                                int mb_x, int mb_y)
{
    int P[10][2];
    const int mot_stride = s->mb_stride;
    const int mot_xy = mb_y*mot_stride + mb_x;
    const int shift= 1+s->quarter_sample;
    int dmin, i;
    const int time_pp= s->pp_time;
    const int time_pb= s->pb_time;
    int mx, my, xmin, xmax, ymin, ymax;
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
    uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
    
    ymin= xmin=(-32)>>shift;
    ymax= xmax=   31>>shift;

    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
        s->mv_type= MV_TYPE_8X8;
    }else{
        s->mv_type= MV_TYPE_16X16;
    }

    for(i=0; i<4; i++){
        int index= s->block_index[i];
        int min, max;
    
        s->me.co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
        s->me.co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
        s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
        s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
//        s->me.direct_basis_mv[1][i][0]= s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
//        s->me.direct_basis_mv[1][i][1]= s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);

        max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
        min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
        max+= 16*mb_x + 1; // +-1 is for the simpler rounding
        min+= 16*mb_x - 1;
        xmax= FFMIN(xmax, s->width - max);
        xmin= FFMAX(xmin, - 16     - min);

        max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
        min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
        max+= 16*mb_y + 1; // +-1 is for the simpler rounding
        min+= 16*mb_y - 1;
        ymax= FFMIN(ymax, s->height - max);
        ymin= FFMAX(ymin, - 16      - min);
        
        if(s->mv_type == MV_TYPE_16X16) break;
    }
    
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
    
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
        s->b_direct_mv_table[mot_xy][0]= 0;
        s->b_direct_mv_table[mot_xy][1]= 0;

        return 256*256*256*64;
    }
    
    s->me.xmin= xmin;
    s->me.ymin= ymin;
    s->me.xmax= xmax;
    s->me.ymax= ymax;

    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);

    /* special case for first line */
    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
    
        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]);
    }
 
    //FIXME direct_search  ptr in context!!! (needed for chroma anyway or this will get messy)   
    if(s->flags&CODEC_FLAG_QPEL){
        dmin = simple_direct_qpel_epzs_motion_search(s, &mx, &my, P, 0, 0, 
                                                     src_data, ref_data, stride, uvstride, mv_table, 1<<14, mv_penalty);
        dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin,
                                                0, 0, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
        
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
            dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, src_data, ref_data, stride, uvstride, mv_penalty);
    }else{
        dmin = simple_direct_hpel_epzs_motion_search(s, &mx, &my, P, 0, 0, 
                                                     src_data, ref_data, stride, uvstride, mv_table, 1<<15, mv_penalty);
        dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin,
                                                0, 0, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);

        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
            dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, src_data, ref_data, stride, uvstride, mv_penalty);
    }
    
    get_limits(s, 16*mb_x, 16*mb_y); //restore s->me.?min/max, maybe not needed

    s->b_direct_mv_table[mot_xy][0]= mx;
    s->b_direct_mv_table[mot_xy][1]= my;
    return dmin;
}

void ff_estimate_b_frame_motion(MpegEncContext * s,
                             int mb_x, int mb_y)
{
    const int penalty_factor= s->me.mb_penalty_factor;
    int fmin, bmin, dmin, fbmin, bimin, fimin;
    int type=0;
    const int stride= s->linesize;
    const int uvstride= s->uvlinesize;
    uint8_t *src_data[3]= {
        s->new_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
        s->new_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
        s->new_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y)
    };
    uint8_t *ref_data[6]= {
        s->last_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
        s->last_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
        s->last_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y),
        s->next_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
        s->next_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
        s->next_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y)
    };
    
    s->me.skip=0;
    if (s->codec_id == CODEC_ID_MPEG4)
        dmin= direct_search(s, src_data, ref_data, stride, uvstride, mb_x, mb_y);
    else
        dmin= INT_MAX;
//FIXME penalty stuff for non mpeg4
    s->me.skip=0;
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, src_data, 
                               ref_data, stride, uvstride, s->f_code) + 3*penalty_factor;
    
    s->me.skip=0;
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, src_data, 
                               ref_data+3, stride, uvstride, s->b_code) + 2*penalty_factor;
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

    s->me.skip=0;
    fbmin= bidir_refine(s, src_data, ref_data, stride, uvstride, mb_x, mb_y) + penalty_factor;
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
    
    if(s->flags & CODEC_FLAG_INTERLACED_ME){
        const int xy = mb_y*s->mb_stride + mb_x;

//FIXME mb type penalty
        s->me.skip=0;
        fimin= interlaced_search(s, src_data, ref_data  , 
                                 s->b_field_mv_table[0], s->b_field_select_table[0], s->f_code,
                                 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
        bimin= interlaced_search(s, src_data, ref_data+3, 
                                 s->b_field_mv_table[1], s->b_field_select_table[1], s->b_code,
                                 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1]);
    }else
        fimin= bimin= INT_MAX;

    {
        int score= fmin;
        type = CANDIDATE_MB_TYPE_FORWARD;
        
        if (dmin <= score){
            score = dmin;
            type = CANDIDATE_MB_TYPE_DIRECT;
        }
        if(bmin<score){
            score=bmin;
            type= CANDIDATE_MB_TYPE_BACKWARD; 
        }
        if(fbmin<score){
            score=fbmin;
            type= CANDIDATE_MB_TYPE_BIDIR;
        }
        if(fimin<score){
            score=fimin;
            type= CANDIDATE_MB_TYPE_FORWARD_I;
        }
        if(bimin<score){
            score=bimin;
            type= CANDIDATE_MB_TYPE_BACKWARD_I;
        }
        
        score= ((unsigned)(score*score + 128*256))>>16;
        s->mc_mb_var_sum_temp += score;
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
    }

    if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
        type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
        if(fimin < INT_MAX)
            type |= CANDIDATE_MB_TYPE_FORWARD_I;
        if(bimin < INT_MAX)
            type |= CANDIDATE_MB_TYPE_BACKWARD_I;
        if(fimin < INT_MAX && bimin < INT_MAX){
            type |= CANDIDATE_MB_TYPE_BIDIR_I;
        }
         //FIXME something smarter
        if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
#if 0        
        if(s->out_format == FMT_MPEG1)
            type |= CANDIDATE_MB_TYPE_INTRA;
#endif
    }

    s->mb_type[mb_y*s->mb_stride + mb_x]= 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_t * 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);

        for(y=0; y<s->mb_height; y++){
            int x;
            int xy= y*s->mb_stride;
            for(x=0; x<s->mb_width; x++){
                if(s->mb_type[xy] & 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->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
                            score[j]-= 170;
                    }
                }
                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, range;
    assert(s->pict_type==P_TYPE);

    range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
    
    if(s->msmpeg4_version) range= 16;
    
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
    
//printf("%d no:%d %d//\n", clip, noclip, f_code);
    if(s->flags&CODEC_FLAG_4MV){
        const int wrap= s->b8_stride;

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

            for(x=0; x<s->mb_width; x++){
                if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
                    int block;
                    for(block=0; block<4; block++){
                        int off= (block& 1) + (block>>1)*wrap;
                        int mx= s->current_picture.motion_val[0][ xy + off ][0];
                        int my= s->current_picture.motion_val[0][ xy + off ][1];

                        if(   mx >=range || mx <-range
                           || my >=range || my <-range){
                            s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
                            s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
                            s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
                        }
                    }
                }
                x