comparison motion_est.c @ 853:eacc2dd8fd9d libavcodec

* using DSPContext - so each codec could use its local (sub)set of CPU extension

852:c01c98206ee6

	pix2 += line_size - 16;
    }
    return s;
}

static void no_motion_search(MpegEncContext * s,
			     int *mx_ptr, int *my_ptr)
{
    *mx_ptr = 16 * s->mb_x;
    *my_ptr = 16 * s->mb_y;
}

    dmin = 0x7fffffff;
    mx = 0;
    my = 0;
    for (y = y1; y <= y2; y++) {
	for (x = x1; x <= x2; x++) {
	    d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
			     s->linesize);
	    if (d < dmin ||
		(d == dmin &&
		 (abs(x - xx) + abs(y - yy)) <
		 (abs(mx - xx) + abs(my - yy)))) {

    my = 0;

    do {
	for (y = y1; y <= y2; y += range) {
	    for (x = x1; x <= x2; x += range) {
		d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
		if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
		    dmin = d;
		    mx = x;
		    my = y;
		}

        dminx = 0x7fffffff;
        dminy = 0x7fffffff;

	lastx = x;
	for (x = x1; x <= x2; x += range) {
	    d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
	    if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
		dminx = d;
		mx = x;
	    }
	}

	x = lastx;
	for (y = y1; y <= y2; y += range) {
	    d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
	    if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
		dminy = d;
		my = y;
	    }
	}

#define CHECK_MV(x,y)\
{\
    const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
    const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
    if(map[index]!=key){\
        d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
        d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
        COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
        map[index]= key;\
        score_map[index]= d;\
    }\

#define CHECK_MV4(x,y)\
{\
    const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
    const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
    if(map[index]!=key){\
        d = pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
        d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
        COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
        map[index]= key;\
        score_map[index]= d;\
    }\

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

    dmin = pix_abs16x16(new_pic, old_pic, pic_stride);
    map[0]= map_generation;
    score_map[0]= dmin;

    /* first line */
    if ((s->mb_y == 0 || s->first_slice_line)) {

#endif
//check(best[0],best[1],0, b0)
    if(s->me_method==ME_EPZS)
        dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, pix_abs16x16);
    else
        dmin=         cross_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, pix_abs16x16);
//check(best[0],best[1],0, b1)
    *mx_ptr= best[0];
    *my_ptr= best[1];    

//    printf("%d %d %d \n", best[0], best[1], dmin);

    dmin = 1000000;
//printf("%d %d %d %d //",xmin, ymin, xmax, ymax); 
    /* first line */
    if ((s->mb_y == 0 || s->first_slice_line) && block<2) {
        CHECK_MV4(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
        CHECK_MV4(P_LAST[0]>>shift, P_LAST[1]>>shift)
        CHECK_MV4(P_MV1[0]>>shift, P_MV1[1]>>shift)
    }else{
        CHECK_MV4(P_MV1[0]>>shift, P_MV1[1]>>shift)
        //FIXME try some early stop

    }

    if(s->me_method==ME_EPZS)
        dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, pix_abs8x8);
    else
        dmin=         cross_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, pix_abs8x8);

    *mx_ptr= best[0];
    *my_ptr= best[1];    

//    printf("%d %d %d \n", best[0], best[1], dmin);

        P_MV1[1]= my;

        dmin4 = epzs_motion_search4(s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, ref_picture);

        dmin4= fast_halfpel_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
                                   pred_x4, pred_y4, ref_picture, pix_abs8x8_x2, 
                                   pix_abs8x8_y2, pix_abs8x8_xy2, block);
 
        s->motion_val[ s->block_index[block] ][0]= mx4;
        s->motion_val[ s->block_index[block] ][1]= my4;
        dmin_sum+= dmin4;
    }

    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 = pix_sum(pix, s->linesize);
    
    varc = (pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
    vard = (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;

        if (vard*2 + 200 > varc)
            mb_type|= MB_TYPE_INTRA;
        if (varc*2 + 200 > vard){
            mb_type|= MB_TYPE_INTER;
            if(s->me_method >= ME_EPZS)
                fast_halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
                                           pred_x, pred_y, ref_picture, pix_abs16x16_x2, pix_abs16x16_y2, 
                                           pix_abs16x16_xy2, 0);
            else
                halfpel_motion_search(     s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
                                           pred_x, pred_y, ref_picture, pix_abs16x16_x2, pix_abs16x16_y2, 
                                           pix_abs16x16_xy2, 0);                                           
        }else{
            mx <<=1;
            my <<=1;
        }
        if((s->flags&CODEC_FLAG_4MV)

        if (vard <= 64 || vard < varc) {
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
            mb_type|= MB_TYPE_INTER;
            if (s->me_method != ME_ZERO) {
                if(s->me_method >= ME_EPZS)
                    dmin= fast_halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
                                           pred_x, pred_y, ref_picture, pix_abs16x16_x2, pix_abs16x16_y2, 
                                           pix_abs16x16_xy2, 0);
                else
                    dmin= halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
                                           pred_x, pred_y, ref_picture, pix_abs16x16_x2, pix_abs16x16_y2, 
                                           pix_abs16x16_xy2, 0);
                if((s->flags&CODEC_FLAG_4MV)
                   && !s->skip_me && varc>50 && vard>10){
                    int dmin4= mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
                    if(dmin4 + 128 <dmin)
                        mb_type= MB_TYPE_INTER4V;

        dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, ref_picture);
 
        break;
    }
    
    dmin= fast_halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
                                pred_x, pred_y, ref_picture, pix_abs16x16_x2, pix_abs16x16_y2, 
                                pix_abs16x16_xy2, 0);
//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;

    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;
    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);
    src_x = mb_x * 16 + (motion_bx >> 1);
    src_y = mb_y * 16 + (motion_by >> 1);

    if (src_x == s->width)
        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;
    avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
    
    fbmin += 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,

            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;
            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);
            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;

            avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
        }
    }

    P_LAST[0]        = mv_table[mot_xy    ][0];
    P_LAST[1]        = mv_table[mot_xy    ][1];

853:eacc2dd8fd9d

	pix2 += line_size - 16;
    }
    return s;
}

static inline void no_motion_search(MpegEncContext * s,
				    int *mx_ptr, int *my_ptr)
{
    *mx_ptr = 16 * s->mb_x;
    *my_ptr = 16 * s->mb_y;
}

    dmin = 0x7fffffff;
    mx = 0;
    my = 0;
    for (y = y1; y <= y2; y++) {
	for (x = x1; x <= x2; x++) {
	    d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
			     s->linesize);
	    if (d < dmin ||
		(d == dmin &&
		 (abs(x - xx) + abs(y - yy)) <
		 (abs(mx - xx) + abs(my - yy)))) {

    my = 0;

    do {
	for (y = y1; y <= y2; y += range) {
	    for (x = x1; x <= x2; x += range) {
		d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
		if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
		    dmin = d;
		    mx = x;
		    my = y;
		}

        dminx = 0x7fffffff;
        dminy = 0x7fffffff;

	lastx = x;
	for (x = x1; x <= x2; x += range) {
	    d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
	    if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
		dminx = d;
		mx = x;
	    }
	}

	x = lastx;
	for (y = y1; y <= y2; y += range) {
	    d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
	    if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
		dminy = d;
		my = y;
	    }
	}

#define CHECK_MV(x,y)\
{\
    const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
    const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
    if(map[index]!=key){\
        d = s->dsp.pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
        d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
        COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
        map[index]= key;\
        score_map[index]= d;\
    }\

#define CHECK_MV4(x,y)\
{\
    const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
    const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
    if(map[index]!=key){\
        d = s->dsp.pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
        d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
        COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
        map[index]= key;\
        score_map[index]= d;\
    }\

    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);
    map[0]= map_generation;
    score_map[0]= dmin;

    /* first line */
    if ((s->mb_y == 0 || s->first_slice_line)) {

#endif
//check(best[0],best[1],0, b0)
    if(s->me_method==ME_EPZS)
        dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
				   shift, map, score_map, map_generation, s->dsp.pix_abs16x16);
    else
        dmin=         cross_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, s->dsp.pix_abs16x16);
//check(best[0],best[1],0, b1)
    *mx_ptr= best[0];
    *my_ptr= best[1];    

//    printf("%d %d %d \n", best[0], best[1], dmin);

    dmin = 1000000;
//printf("%d %d %d %d //",xmin, ymin, xmax, ymax); 
    /* first line */
    if ((s->mb_y == 0 || s->first_slice_line) && block<2) {
	CHECK_MV4(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
        CHECK_MV4(P_LAST[0]>>shift, P_LAST[1]>>shift)
        CHECK_MV4(P_MV1[0]>>shift, P_MV1[1]>>shift)
    }else{
        CHECK_MV4(P_MV1[0]>>shift, P_MV1[1]>>shift)
        //FIXME try some early stop

    }

    if(s->me_method==ME_EPZS)
        dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
				   shift, map, score_map, map_generation, s->dsp.pix_abs8x8);
    else
        dmin=         cross_search(s, best, dmin, new_pic, old_pic, pic_stride, 
                                   pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, 
                                   shift, map, score_map, map_generation, s->dsp.pix_abs8x8);

    *mx_ptr= best[0];
    *my_ptr= best[1];    

//    printf("%d %d %d \n", best[0], best[1], dmin);

        P_MV1[1]= my;

        dmin4 = epzs_motion_search4(s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, ref_picture);

        dmin4= fast_halfpel_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
					  pred_x4, pred_y4, ref_picture, s->dsp.pix_abs8x8_x2,
					  s->dsp.pix_abs8x8_y2, s->dsp.pix_abs8x8_xy2, block);
 
        s->motion_val[ s->block_index[block] ][0]= mx4;
        s->motion_val[ s->block_index[block] ][1]= my4;
        dmin_sum+= dmin4;
    }

    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 = (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;

        if (vard*2 + 200 > varc)
            mb_type|= MB_TYPE_INTRA;
        if (varc*2 + 200 > vard){
            mb_type|= MB_TYPE_INTER;
            if(s->me_method >= ME_EPZS)
                fast_halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
					   pred_x, pred_y, ref_picture, s->dsp.pix_abs16x16_x2,
					   s->dsp.pix_abs16x16_y2, s->dsp.pix_abs16x16_xy2, 0);
            else
                halfpel_motion_search(     s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
				           pred_x, pred_y, ref_picture, s->dsp.pix_abs16x16_x2,
				           s->dsp.pix_abs16x16_y2, s->dsp.pix_abs16x16_xy2, 0);
        }else{
            mx <<=1;
            my <<=1;
        }
        if((s->flags&CODEC_FLAG_4MV)

        if (vard <= 64 || vard < varc) {
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
            mb_type|= MB_TYPE_INTER;
            if (s->me_method != ME_ZERO) {
                if(s->me_method >= ME_EPZS)
		    dmin= fast_halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
                                           pred_x, pred_y, ref_picture, s->dsp.pix_abs16x16_x2, s->dsp.pix_abs16x16_y2,
                                           s->dsp.pix_abs16x16_xy2, 0);
                else
                    dmin= halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
                                           pred_x, pred_y, ref_picture, s->dsp.pix_abs16x16_x2, s->dsp.pix_abs16x16_y2,
                                           s->dsp.pix_abs16x16_xy2, 0);
                if((s->flags&CODEC_FLAG_4MV)
                   && !s->skip_me && varc>50 && vard>10){
                    int dmin4= mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
                    if(dmin4 + 128 <dmin)
                        mb_type= MB_TYPE_INTER4V;

        dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, ref_picture);
 
        break;
    }
    
    dmin= fast_halfpel_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
                                pred_x, pred_y, ref_picture, s->dsp.pix_abs16x16_x2, s->dsp.pix_abs16x16_y2,
                                s->dsp.pix_abs16x16_xy2, 0);
//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;

    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);
    src_x = mb_x * 16 + (motion_bx >> 1);
    src_y = mb_y * 16 + (motion_by >> 1);

    if (src_x == s->width)
        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,

            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);
            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;

	    s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
        }
    }

    P_LAST[0]        = mv_table[mot_xy    ][0];
    P_LAST[1]        = mv_table[mot_xy    ][1];