comparison motion_est.c @ 903:22ee74da2cd3 libavcodec

cleanup adding AVVideoFrame moving quality, pict_type, key_frame, qscale_table, ... to AVVideoFrame removing obsolete variables in AVCodecContext skiping of MBs in b frames correctly initalizing AVCodecContext picture buffer cleanup

902:6acc8394960d

    if (y1 < ymin)
	y1 = ymin;
    y2 = yy + range - 1;
    if (y2 > ymax)
	y2 = ymax;
    pix = s->new_picture[0] + (yy * s->linesize) + xx;
    dmin = 0x7fffffff;
    mx = 0;
    my = 0;
    for (y = y1; y <= y2; y++) {
	for (x = x1; x <= x2; x++) {

    /* Lower limit */
    y2 = yy + range;
    if (y2 > ymax)
	y2 = ymax;

    pix = s->new_picture[0] + (yy * s->linesize) + xx;
    dmin = 0x7fffffff;
    mx = 0;
    my = 0;

    do {

    /* Lower limit */
    y2 = yy + range;
    if (y2 > ymax)
	y2 = ymax;

    pix = s->new_picture[0] + (yy * s->linesize) + xx;
    mx = 0;
    my = 0;

    x = xx;
    y = yy;

    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me_map;
    uint16_t *score_map= s->me_score_map;
    int map_generation;

    new_pic = s->new_picture[0] + pic_xy;
    old_pic = ref_picture + pic_xy;
    
    map_generation= update_map_generation(s);

    dmin = s->dsp.pix_abs16x16(new_pic, old_pic, pic_stride);

    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me_map;
    uint16_t *score_map= s->me_score_map;
    int map_generation;

    new_pic = s->new_picture[0] + pic_xy;
    old_pic = ref_picture + pic_xy;

    map_generation= update_map_generation(s);

    dmin = 1000000;

        return dmin;
    }

    xx = 16 * s->mb_x + 8*(n&1);
    yy = 16 * s->mb_y + 8*(n>>1);
    pix =  s->new_picture[0] + (yy * s->linesize) + xx;

    mx = *mx_ptr;
    my = *my_ptr;
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
    

    }
//    printf("N");
        
    xx = 16 * s->mb_x + 8*(n&1);
    yy = 16 * s->mb_y + 8*(n>>1);
    pix =  s->new_picture[0] + (yy * s->linesize) + xx;

    mx = *mx_ptr;
    my = *my_ptr;
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
    

static inline int mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
{
    int block;
    int P[10][2];
    uint8_t *ref_picture= s->last_picture[0];
    int dmin_sum=0;

    for(block=0; block<4; block++){
        int mx4, my4;
        int pred_x4, pred_y4;

    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
    int pred_x=0, pred_y=0;
    int P[10][2];
    const int shift= 1+s->quarter_sample;
    int mb_type=0;
    uint8_t *ref_picture= s->last_picture[0];

    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
    rel_xmin= xmin - mb_x*16;
    rel_xmax= xmax - mb_x*16;
    rel_ymin= ymin - mb_y*16;

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

    pix = s->new_picture[0] + (yy * s->linesize) + xx;
    /* At this point (mx,my) are full-pell and the relative displacement */
    ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
    
    sum = s->dsp.pix_sum(pix, s->linesize);
    
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
    // FIXME: MMX OPTIMIZE
    vard = (s->dsp.pix_norm(pix, ppix, s->linesize)+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);
    s->mb_var   [s->mb_width * mb_y + mb_x] = varc;
    s->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;
    s->mb_mean  [s->mb_width * mb_y + mb_x] = (sum+128)>>8;
    s->mb_var_sum    += varc;
    s->mc_mb_var_sum += 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);

        dxy&= 2;
    src_y = clip(src_y, -16, s->height);
    if (src_y == s->height)
        dxy&= 1;

    ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
    s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

    fbmin += (mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->qscale;

    dxy = ((motion_by & 1) << 1) | (motion_bx & 1);

        dxy&= 2;
    src_y = clip(src_y, -16, s->height);
    if (src_y == s->height)
        dxy&= 1;

    ptr = s->next_picture[0] + (src_y * s->linesize) + src_x;
    s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

    fbmin += s->dsp.pix_abs16x16(s->new_picture[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
    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,

            src_x = clip(src_x, -16, width);
            if (src_x == width) dxy &= ~1;
            src_y = clip(src_y, -16, height);
            if (src_y == height) dxy &= ~2;

            ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
            s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

            dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
            src_x = (mb_x + bx) * 16 + (motion_bx >> 1);
            src_y = (mb_y + by) * 16 + (motion_by >> 1);

    int fmin, bmin, dmin, fbmin;
    int type=0;
    
    dmin= direct_search(s, mb_x, mb_y);

    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, s->last_picture[0], s->f_code);
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, s->next_picture[0], s->b_code) - quant;
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

    fbmin= bidir_refine(s, mb_x, mb_y);

    {

        if(fbmin<score){
            score=fbmin;
            type= MB_TYPE_BIDIR;
        }
        score= ((unsigned)(score*score + 128*256))>>16;
        s->mc_mb_var_sum += score;
        s->mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSD
    }

    if(s->flags&CODEC_FLAG_HQ){
        type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
    }

                    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->mc_mb_var[i] < s->mb_var[i])
                            score[j]-= 170;
                    }
                }
                i++;
                xy++;

903:22ee74da2cd3

    if (y1 < ymin)
	y1 = ymin;
    y2 = yy + range - 1;
    if (y2 > ymax)
	y2 = ymax;
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
    dmin = 0x7fffffff;
    mx = 0;
    my = 0;
    for (y = y1; y <= y2; y++) {
	for (x = x1; x <= x2; x++) {

    /* Lower limit */
    y2 = yy + range;
    if (y2 > ymax)
	y2 = ymax;

    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
    dmin = 0x7fffffff;
    mx = 0;
    my = 0;

    do {

    /* Lower limit */
    y2 = yy + range;
    if (y2 > ymax)
	y2 = ymax;

    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
    mx = 0;
    my = 0;

    x = xx;
    y = yy;

    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me_map;
    uint16_t *score_map= s->me_score_map;
    int map_generation;

    new_pic = s->new_picture.data[0] + pic_xy;
    old_pic = ref_picture + pic_xy;
    
    map_generation= update_map_generation(s);

    dmin = s->dsp.pix_abs16x16(new_pic, old_pic, pic_stride);

    const int shift= 1+s->quarter_sample;
    uint32_t *map= s->me_map;
    uint16_t *score_map= s->me_score_map;
    int map_generation;

    new_pic = s->new_picture.data[0] + pic_xy;
    old_pic = ref_picture + pic_xy;

    map_generation= update_map_generation(s);

    dmin = 1000000;

        return dmin;
    }

    xx = 16 * s->mb_x + 8*(n&1);
    yy = 16 * s->mb_y + 8*(n>>1);
    pix =  s->new_picture.data[0] + (yy * s->linesize) + xx;

    mx = *mx_ptr;
    my = *my_ptr;
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
    

    }
//    printf("N");
        
    xx = 16 * s->mb_x + 8*(n&1);
    yy = 16 * s->mb_y + 8*(n>>1);
    pix =  s->new_picture.data[0] + (yy * s->linesize) + xx;

    mx = *mx_ptr;
    my = *my_ptr;
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
    

static inline int mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
{
    int block;
    int P[10][2];
    uint8_t *ref_picture= s->last_picture.data[0];
    int dmin_sum=0;

    for(block=0; block<4; block++){
        int mx4, my4;
        int pred_x4, pred_y4;

    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
    int pred_x=0, pred_y=0;
    int P[10][2];
    const int shift= 1+s->quarter_sample;
    int mb_type=0;
    uint8_t *ref_picture= s->last_picture.data[0];
    Picture * const pic= &s->current_picture;

    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
    rel_xmin= xmin - mb_x*16;
    rel_xmax= xmax - mb_x*16;
    rel_ymin= ymin - mb_y*16;

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

    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
    /* At this point (mx,my) are full-pell and the relative displacement */
    ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
    
    sum = s->dsp.pix_sum(pix, s->linesize);
    
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
    // FIXME: MMX OPTIMIZE
    vard = (s->dsp.pix_norm(pix, ppix, s->linesize)+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_width * mb_y + mb_x] = varc;
    pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;
    pic->mb_mean  [s->mb_width * mb_y + mb_x] = (sum+128)>>8;
    pic->mb_var_sum    += varc;
    pic->mc_mb_var_sum += 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);

        dxy&= 2;
    src_y = clip(src_y, -16, s->height);
    if (src_y == s->height)
        dxy&= 1;

    ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
    s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

    fbmin += (mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->qscale;

    dxy = ((motion_by & 1) << 1) | (motion_bx & 1);

        dxy&= 2;
    src_y = clip(src_y, -16, s->height);
    if (src_y == s->height)
        dxy&= 1;

    ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
    s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

    fbmin += s->dsp.pix_abs16x16(s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
    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,

            src_x = clip(src_x, -16, width);
            if (src_x == width) dxy &= ~1;
            src_y = clip(src_y, -16, height);
            if (src_y == height) dxy &= ~2;

            ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
            s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

            dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
            src_x = (mb_x + bx) * 16 + (motion_bx >> 1);
            src_y = (mb_y + by) * 16 + (motion_by >> 1);

    int fmin, bmin, dmin, fbmin;
    int type=0;
    
    dmin= direct_search(s, mb_x, mb_y);

    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, s->last_picture.data[0], s->f_code);
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, s->next_picture.data[0], s->b_code) - quant;
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

    fbmin= bidir_refine(s, mb_x, mb_y);

    {

        if(fbmin<score){
            score=fbmin;
            type= MB_TYPE_BIDIR;
        }
        score= ((unsigned)(score*score + 128*256))>>16;
        s->current_picture.mc_mb_var_sum += score;
        s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSD
    }

    if(s->flags&CODEC_FLAG_HQ){
        type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
    }

                    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[i] < s->current_picture.mb_var[i])
                            score[j]-= 170;
                    }
                }
                i++;
                xy++;