motion_est.c

arm平台下的H264编码和解码源代码

            const int mot_stride = s->b8_stride;
            const int mot_xy = s->block_index[0];

            P_LEFT[0]       = s->current_picture.motion_val[0][mot_xy - 1][0];
            P_LEFT[1]       = s->current_picture.motion_val[0][mot_xy - 1][1];

            if(P_LEFT[0]       > (s->me.xmax<<shift)) P_LEFT[0]       = (s->me.xmax<<shift);

            if(!s->first_slice_line) {
                P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][0];
                P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][1];
                P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
                P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
                if(P_TOP[1]      > (s->me.ymax<<shift)) P_TOP[1]     = (s->me.ymax<<shift);
                if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
                if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.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]);

                if(s->out_format == FMT_H263){
                    pred_x = P_MEDIAN[0];
                    pred_y = P_MEDIAN[1];
                }else { /* mpeg1 at least */
                    pred_x= P_LEFT[0];
                    pred_y= P_LEFT[1];
                }
            }else{
                pred_x= P_LEFT[0];
                pred_y= P_LEFT[1];
            }

        }
        dmin = s->me.motion_search[0](s, &mx, &my, P, pred_x, pred_y, 
                                      src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);
 
        break;
    }

    /* intra / predictive decision */
    xx = mb_x * 16;
    yy = mb_y * 16;

    pix = src_data[0];
    /* At this point (mx,my) are full-pell and the relative displacement */
    ppix = ref_data[0] + (my * s->linesize) + mx;
    
    sum = s->dsp.pix_sum(pix, s->linesize);
    
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;

//printf("%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
    pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
    pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
    s->mb_var_sum_temp    += varc;
    s->mc_mb_var_sum_temp += vard;
//printf("E%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
    
#if 0
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
	   varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
#endif
    if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
        if (vard <= 64 || vard < varc)
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        else
            s->scene_change_score+= s->qscale;

        if (vard*2 + 200 > varc)
            mb_type|= CANDIDATE_MB_TYPE_INTRA;
        if (varc*2 + 200 > vard){
            mb_type|= CANDIDATE_MB_TYPE_INTER;
            s->me.sub_motion_search(s, &mx, &my, dmin,
				   pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
            if(s->flags&CODEC_FLAG_MV0)
                if(mx || my)
                    mb_type |= CANDIDATE_MB_TYPE_SKIPED; //FIXME check difference
        }else{
            mx <<=shift;
            my <<=shift;
        }
        if((s->flags&CODEC_FLAG_4MV)
           && !s->me.skip && varc>50 && vard>10){
            if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
                mb_type|=CANDIDATE_MB_TYPE_INTER4V;

            set_p_mv_tables(s, mx, my, 0);
        }else
            set_p_mv_tables(s, mx, my, 1);
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
           && !s->me.skip){ //FIXME varc/d checks
            if(interlaced_search(s, src_data, ref_data, s->p_field_mv_table, s->p_field_select_table, s->f_code, mx, my) < INT_MAX)
                mb_type |= CANDIDATE_MB_TYPE_INTER_I;
        }
    }else{
        int intra_score, i;
        mb_type= CANDIDATE_MB_TYPE_INTER;

        dmin= s->me.sub_motion_search(s, &mx, &my, dmin,
                                    pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
            dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, src_data, ref_data, stride, uvstride, mv_penalty);

        if((s->flags&CODEC_FLAG_4MV)
           && !s->me.skip && varc>50 && vard>10){
            int dmin4= h263_mv4_search(s, mx, my, shift);
            if(dmin4 < dmin){
                mb_type= CANDIDATE_MB_TYPE_INTER4V;
                dmin=dmin4;
            }
        }
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
           && !s->me.skip){ //FIXME varc/d checks
            int dmin_i= interlaced_search(s, src_data, ref_data, s->p_field_mv_table, s->p_field_select_table, s->f_code, mx, my);
            if(dmin_i < dmin){
                mb_type = CANDIDATE_MB_TYPE_INTER_I;
                dmin= dmin_i;
            }
        }
                
//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
        set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);

        /* get intra luma score */
        if((s->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
            intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
        }else{
            int mean= (sum+128)>>8;
            mean*= 0x01010101;
            
            for(i=0; i<16; i++){
                *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 0]) = mean;
                *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 4]) = mean;
                *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 8]) = mean;
                *(uint32_t*)(&s->me.scratchpad[i*s->linesize+12]) = mean;
            }

            intra_score= s->dsp.mb_cmp[0](s, s->me.scratchpad, pix, s->linesize, 16);
        }
#if 0 //FIXME
        /* get chroma score */
        if(s->avctx->mb_cmp&FF_CMP_CHROMA){
            for(i=1; i<3; i++){
                uint8_t *dest_c;
                int mean;
                
                if(s->out_format == FMT_H263){
                    mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
                }else{
                    mean= (s->last_dc[i] + 4)>>3;
                }
                dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;
                
                mean*= 0x01010101;
                for(i=0; i<8; i++){
                    *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 0]) = mean;
                    *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 4]) = mean;
                }
                
                intra_score+= s->dsp.mb_cmp[1](s, s->me.scratchpad, dest_c, s->uvlinesize);
            }                
        }
#endif
        intra_score += s->me.mb_penalty_factor*16;
        
        if(intra_score < dmin){
            mb_type= CANDIDATE_MB_TYPE_INTRA;
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
        }else
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
        
        if (vard <= 64 || vard < varc) { //FIXME
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
        }else{
            s->scene_change_score+= s->qscale;
        }
    }

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

int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
                                    int mb_x, int mb_y)
{
    int mx, my, dmin;
    int pred_x=0, pred_y=0;
    int P[10][2];
    const int shift= 1+s->quarter_sample;
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
    const int xy= mb_x + mb_y*s->mb_stride;
    const int stride= s->linesize;
    const int uvstride= s->uvlinesize;
    uint8_t *src_data[3]= {
        s->new_picture.data[0] + 16*(mb_x + stride*mb_y),
        s->new_picture.data[1] + 8*(mb_x + uvstride*mb_y),
        s->new_picture.data[2] + 8*(mb_x + uvstride*mb_y)
    };
    uint8_t *ref_data[3]= {
        s->last_picture.data[0] + 16*(mb_x + stride*mb_y),
        s->last_picture.data[1] + 8*(mb_x + uvstride*mb_y),
        s->last_picture.data[2] + 8*(mb_x + uvstride*mb_y)
    };
    
    assert(s->quarter_sample==0 || s->quarter_sample==1);

    s->me.pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);

    get_limits(s, 16*mb_x, 16*mb_y);
    s->me.skip=0;

    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];

    if(P_LEFT[0]       < (s->me.xmin<<shift)) P_LEFT[0]       = (s->me.xmin<<shift);

    /* special case for first line */
    if (s->first_slice_line) {
        pred_x= P_LEFT[0];
        pred_y= P_LEFT[1];
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 
    } else {
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
        if(P_TOP[1]      < (s->me.ymin<<shift)) P_TOP[1]     = (s->me.ymin<<shift);
        if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift);
        if(P_TOPRIGHT[1] < (s->me.ymin<<shift)) P_TOPRIGHT[1]= (s->me.ymin<<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]);

        pred_x = P_MEDIAN[0];
        pred_y = P_MEDIAN[1];
    }
    dmin = s->me.pre_motion_search(s, &mx, &my, P, pred_x, pred_y, 
                                   src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);

    s->p_mv_table[xy][0] = mx<<shift;
    s->p_mv_table[xy][1] = my<<shift;
    
    return dmin;
}

static int ff_estimate_motion_b(MpegEncContext * s,
                       int mb_x, int mb_y, int16_t (*mv_table)[2], uint8_t *src_data[3],
                       uint8_t *ref_data[3], int stride, int uvstride, int f_code)
{
    int mx, my, dmin;
    int pred_x=0, pred_y=0;
    int P[10][2];
    const int shift= 1+s->quarter_sample;
    const int mot_stride = s->mb_stride;
    const int mot_xy = mb_y*mot_stride + mb_x;
    uint8_t * const ref_picture= ref_data[0] - 16*s->mb_x - 16*s->mb_y*s->linesize; //FIXME ugly
    uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
    int mv_scale;
        
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);

    get_limits(s, 16*mb_x, 16*mb_y);

    switch(s->me_method) {
    case ME_ZERO:
    default:
	no_motion_search(s, &mx, &my);
        dmin = 0;
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
#if 0
    case ME_FULL:
	dmin = full_motion_search(s, &mx, &my, range, ref_picture);
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
    case ME_LOG:
	dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
    case ME_PHODS:
	dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
        mx-= mb_x*16;
        my-= mb_y*16;
        break;
#endif
    case ME_X1:
    case ME_EPZS:
       {
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
            P_LEFT[1]        = mv_table[mot_xy - 1][1];

            if(P_LEFT[0]       > (s->me.xmax<<shift)) P_LEFT[0]       = (s->me.xmax<<shift);

            /* special case for first line */
            if (!s->first_slice_line) {
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
                if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1]= (s->me.ymax<<shift);
                if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
                if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.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]);
            }
            pred_x= P_LEFT[0];
            pred_y= P_LEFT[1];
        }
        
        if(mv_table == s->b_forw_mv_table){
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
        }else{
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
        }
        
        dmin = s->me.motion_search[0](s, &mx, &my, P, pred_x, pred_y, 
                                      src_data, ref_data, stride, uvstride, s->p_mv_table, mv_scale, mv_penalty);
 
        break;
    }
    
    dmin= s->me.sub_motion_search(s, &mx, &my, dmin,
				   pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
                                   
    if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
        dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, src_data, ref_data, stride, uvstride, mv_penalty);

//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
    mv_table[mot_xy][0]= mx;
    mv_table[mot_xy][1]= my;

    return dmin;
}

static inline int check_bidir_mv(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
                   int stride, int uvstride,
                   int motion_fx, int motion_fy,
                   int motion_bx, int motion_by,
                   int pred_fx, int pred_fy,
                   int pred_bx, int pred_by,
                   int size, int h)
{
    //FIXME optimize?
    //FIXME move into template?
    //FIXME better f_code prediction (max mv & distance)
    //FIXME pointers
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
    uint8_t *dest_y = s->me.scratchpad;
    uint8_t *ptr;
    int dxy;
    int src_x, src_y;
    int fbmin;

    if(s->quarter_sample){
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
        src_x = motion_fx >> 2;
        src_y = motion_fy >> 2;