mot_est_mb.c

来自「实现在linux下的mpeg4编解码」· C语言 代码 · 共 846 行 · 第 1/2 页

					y_curr+y+rel_ori_y), act_block, 
					(vop_width/*+2*edge*/), sad_min);

			if (sad<sad_min)
			{
				sad_min=sad;
				mv_x=x;
				mv_y=y;
			}
			else if (sad==sad_min)
			if((ABS(x)+ABS(y)) < (ABS(mv_x)+ABS(mv_y)))
			{
				sad_min=sad;
				mv_x=x;
				mv_y=y;
			}
		}									  /*for*/
		#else								  /* PIH (NTU) Fast ME 08/08/99 */
		sad = SAD_Macroblock(prev+INDEX_BIG(x_curr+rel_ori_x,
			y_curr+rel_ori_y), act_block, (vop_width/*+2*edge*/), MV_MAX_ERROR)
			- (128 + 1);
		if (sad<sad_min)
		{
			sad_min=sad;
			mv_x = mv_y = 0;
		}
		do
		{
			check_pts=total_check_pts;

			do
			{
				check_pt_x = diamond[d_type].point[pt_nmbr].x + d_centre_x;
				check_pt_y = diamond[d_type].point[pt_nmbr].y + d_centre_y;

				/* Restrict the search to the searching window ; Note: This constraint can be removed */
				if ( check_pt_x < int_mv_x_min || check_pt_x > int_mv_x_max || check_pt_y < int_mv_y_min || check_pt_y > int_mv_y_max)
				{
					sad = MV_MAX_ERROR;
				}
				else
				{
					sad=SAD_Macroblock(prev+INDEX_BIG(x_curr+check_pt_x+rel_ori_x,
						y_curr+check_pt_y+rel_ori_y), act_block, 
						(vop_width/*+2*edge*/), sad_min);
					#ifdef _SAD_EXHAUS_
					fprintf(stdout,"+o+ [%2d,%2d] sad16(%3d,%3d)=%4d\n",i,j,x,y,sad);
					#endif
				}
				if (sad<sad_min)
				{
					sad_min=sad;
					mv_x=check_pt_x;
					mv_y=check_pt_y;
					mot_dirn=pt_nmbr;
				}
				else if (sad==sad_min)
				if((ABS(check_pt_x)+ABS(check_pt_y)) < (ABS(mv_x)+ABS(mv_y)))
				{
					sad_min=sad;
					mv_x=check_pt_x;
					mv_y=check_pt_y;
					mot_dirn=pt_nmbr;
				}

				pt_nmbr+=1;
				if((pt_nmbr)>= 8) pt_nmbr-=8;
				check_pts-=1;
			}
			while(check_pts>0);

			if( d_type == 0)
			{
				stop_flag = 1;

			}
			else
			{
				if( (mv_x == d_centre_x) && (mv_y == d_centre_y) )
				{
					d_type=0;
					pt_nmbr=0;
					total_check_pts = 4;
				}
				else
				{
					if((mv_x==d_centre_x) ||(mv_y==d_centre_y))
						total_check_pts=5;
					else
						total_check_pts=3;
					pt_nmbr=diamond[d_type].point[mot_dirn].start_nmbr;
					d_centre_x = mv_x;
					d_centre_y = mv_y;

				}
			}
		}
		while(stop_flag!=1);
		#endif

		flags[0]=flags[0] && (mv_x==int_mv_x_min);
		flags[1]=flags[1] && (mv_x==int_mv_x_max);
		flags[2]=flags[2] && (mv_y==int_mv_y_min);
		flags[3]=flags[3] && (mv_y==int_mv_y_max);
	}
	else
	{
		mv_x=mv_y=0;
		sad_min=MV_MAX_ERROR;
	}

	/* Store result */

	out |=((mv_x==2000)||(mv_y==2000));
	if(out)
	{
		mv_x=mv_y=0;
		sad_min=MV_MAX_ERROR;
	}

	pos16=2*j*2*mvm_width + 2*i;
	mv16_w[pos16]=   mv_x; mv16_h[pos16]=   mv_y;
	mv16_w[pos16+1]= mv_x; mv16_h[pos16+1]= mv_y;
	pos16+=2*mvm_width;
	mv16_w[pos16]=   mv_x; mv16_h[pos16]=   mv_y;
	mv16_w[pos16+1]= mv_x; mv16_h[pos16+1]= mv_y;
	min_error16 = sad_min;

	*min_error = min_error16;

	/* Compute 8x8 MVs */

	if(enable_8x8_mv==1)
	{
		if(!out)
		{
			for (block=0;block<4;block++)
			{
				/* Obtain range */
				if(block==0)
				{
					hb=vb=0;
				}
				else if (block==1)
				{
					hb=1;vb=0;
				}
				else if (block==2)
				{
					hb=0;vb=1;
				}
				else
				{
					hb=vb=1;
				}

				/* VM 2.*: restrict the search range based on the current 16x16 MV
				 *         inside a window around it
				 */

				pmv_x=mv16_w[pos16]; pmv_y=mv16_h[pos16];

											  /* UB 990215 added search range */
				Obtain_Range(f_code, sr, MBM_INTER8,
					pmv_x, pmv_y, &mv_x_min,
											  /*UB 990215 added quarter pel support */
					&mv_x_max, &mv_y_min, &mv_y_max, 0 /*GetVopQuarterPel(curr_vop)*/);

				RangeInSearchArea(i,j, block+1, prev_x, prev_y,
					vop_width, vop_height, br_x, br_y, edge,f_code,
											  /*UB 990215 added quarter pel support */
					&mv_x_min, &mv_x_max, &mv_y_min,&mv_y_max,&out);

				if(!out)
				{
					int_mv_x_min=(int)ceil((double)mv_x_min);
					int_mv_y_min=(int)ceil((double)mv_y_min);
					int_mv_x_max=(int)floor((double)mv_x_max);
					int_mv_y_max=(int)floor((double)mv_y_max);
					flags[4+block*4]=ARE_EQUAL(int_mv_x_min,mv_x_min);
					flags[4+block*4+1]=ARE_EQUAL(int_mv_x_max,mv_x_max);
					flags[4+block*4+2]=ARE_EQUAL(int_mv_y_min,mv_y_min);
					flags[4+block*4+3]=ARE_EQUAL(int_mv_y_max,mv_y_max);

					sad_min=MV_MAX_ERROR;
					mv_x = mv_y = 2000;		  /* A very large MV */

					for (y=int_mv_y_min; y<=int_mv_y_max; y++)
						for (x=int_mv_x_min; x<=int_mv_x_max; x++)
					{
						sad=SAD_Block(prev+
							INDEX_BIG(x_curr + x + 8*(block==1||block==3)+rel_ori_x,
							y_curr + y + 8*(block==2||block==3)+rel_ori_y),
							act_block+INDEX_NOR(8*(block==1||block==3),
							8*(block==2||block==3)),
							(vop_width /*+2*edge*/), sad_min);

						if (sad<sad_min)
						{
							sad_min=sad;
							mv_x=x;
							mv_y=y;
						}
						else if (sad==sad_min)
						if((ABS(x)+ABS(y)) < (ABS(mv_x)+ABS(mv_y)))
						{
							sad_min=sad;
							mv_x=x;
							mv_y=y;
						}
					}						  /*for*/

					flags[4+block*4]   = flags[4+block*4]   && (mv_x==int_mv_x_min);
					flags[4+block*4+1] = flags[4+block*4+1] && (mv_x==int_mv_x_max);
					flags[4+block*4+2] = flags[4+block*4+2] && (mv_y==int_mv_y_min);
					flags[4+block*4+3] = flags[4+block*4+3] && (mv_y==int_mv_y_max);
				}
				else
				{
					mv_x=mv_y=0;
					sad_min=MV_MAX_ERROR;
					/* punish the OUTER blocks with MV_MAX_ERROR in order to
					   be INTRA CODED */
				}

				/* Store result */

				if(block==0)
				{
					pos8=2*j*2*mvm_width + 2*i;
					min_error8 += sad_min;
				}
				else if (block==1)
				{
					pos8=2*j*2*mvm_width + 2*i+1;
					min_error8 += sad_min;
				}
				else if (block==2)
				{
					pos8=(2*j+1)*2*mvm_width + 2*i;
					min_error8 += sad_min;
				}
				else
				{
					pos8=(2*j+1)*2*mvm_width + 2*i+1;
					min_error8 += sad_min;
				}

				/* Store result: absolute coordinates (note that in restricted mode
				   pmv is (0,0)) */
				mv8_w[pos8]=mv_x;
				mv8_h[pos8]=mv_y;
			}									  /*for block*/

			if (min_error8 < *min_error)
				*min_error = min_error8;
		}
		else
		{										  /* all the four blocks are out */
			pos8=2*j*2*mvm_width + 2*i;      mv8_w[pos8]=mv8_h[pos8]=0.0;
			pos8=2*j*2*mvm_width + 2*i+1;    mv8_w[pos8]=mv8_h[pos8]=0.0;
			pos8=(2*j+1)*2*mvm_width + 2*i;  mv8_w[pos8]=mv8_h[pos8]=0.0;
			pos8=(2*j+1)*2*mvm_width + 2*i+1;mv8_w[pos8]=mv8_h[pos8]=0.0;
			min_error8 = MV_MAX_ERROR;
		}
	}											  /* enable_8x8_mv*/

}


/***********************************************************CommentBegin******
 *
 * -- FindSubPel -- Computes MV with half-pixel accurary	
 *
 * Purpose :
 *      Computes MV with sub-pixel accurary for a 16x16 or 8x8 block
 *
 * Description :
 *    1) The preference for the 16x16 null motion vector must be applied
 *       again (now, the SAD is recomputed for the recontructed interpolated
 *       reference)
 *
 ***********************************************************CommentEnd********/

Void
FindSubPel(
Int    x,										  /* <-- horizontal block coordinate in pixels            */
Int    y,										  /* <-- vertical blocl coordinate in pixels              */
SInt   *prev,                                     /* <-- interpolated reference vop                       */
SInt   *curr,                                	  /* <-- current block                                    */											  
Int    bs_x,              						  /* <-- hor. block size (8/16)                           */
Int    bs_y,									  /* <-- vert. block size (8/16)                          */
Int    comp,									  /* <-- block position in MB (0,1,2,3)                   */
Int    rel_x,									  /* <-- relative horiz. position between curr & prev vops*/
Int    rel_y,									  /* <-- relative verti. position between curr & prev vops*/
Int    pels,									  /* <-- vop width                                        */
Int    lines,									  /* <-- vop height                                       */
Int    edge,									  /* <-- edge                                             */
SInt   *flags,									  /* <-- flags                                            */
SInt   *curr_comp_mb,                             /* <-> motion compensated current mb                    */
Float  *mvx,									  /* <-> horizontal motion vector                         */
Float  *mvy,									  /* <-> vertical motion vector                           */
Int    *min_error								  /* --> prediction error                                 */
)
{
	static PixPoint	search[9] = 
	{	
		{0, 0}, {-1, -1}, {0, -1}, {1, -1},
		{-1, 0}, {1, 0}, {-1, 1}, {0, 1}, {1, 1}
	};
	/* searching map
		1 2 3
		4 0 5
		6 7 8
	*/

	Int			i, m, n;
	Int			new_x, new_y,
				lx, size_x;								  /* MW QPEL 07-JUL-1998 */
	Int			min_pos;
	Int			AE, AE_min;
	Int			flag_pos;
	SInt		*pRef, *pComp;

	int flag_search[9] = {1, 1, 1, 1, 1, 1, 1, 1, 1};

	Int SubDimension = 2;

	lx = 2*(pels /*+ 2*edge*/);

	new_x = (Int)((x + *mvx + rel_x)*(SubDimension));
	new_y = (Int)((y + *mvy + rel_y)*(SubDimension));
	new_x += ((comp&1)<<3)*SubDimension;
	new_y += ((comp&2)<<2)*SubDimension;

	size_x=16;

	/** in 1-pixel search we check if we are inside the range; so don't check
		it again
	**/

	/* avoids trying outside the reference image */
	/* we also check with flags if we are inside */
	/* search range (1==don't make 1/x search by */
	/* this side                                 */

	if (bs_x==8)								  
		flag_pos=4+comp*4;
	else										  /* ==16*/
		flag_pos=0;

	if (((new_x/SubDimension) <= 0/*-edge*/) || *(flags+flag_pos)) {
		flag_search[1] = flag_search[4] = flag_search[6] = 0;
	} else if  (((new_x/SubDimension) >= (pels - bs_x /*+ edge*/)) || *(flags+flag_pos+1)) {
		flag_search[3] = flag_search[5] = flag_search[8] = 0;
	};

	if (((new_y/SubDimension) <= 0/*-edge*/) || *(flags+flag_pos+2)) {
		flag_search[1] = flag_search[2] = flag_search[3] = 0;
	} else if  (((new_y/SubDimension) >= (lines- bs_y /*+ edge*/)) || *(flags+flag_pos+3)) {
		flag_search[6] = flag_search[7] = flag_search[8] = 0;
	};

	AE_min = MV_MAX_ERROR;
	min_pos = 0;
	for (i = 0; i < 9; i++)
	{
		if (flag_search[i])
		{
			AE = 0;
			/* estimate on the already interpolated half-pel image */
			pRef = prev + new_x + search[i].x + (new_y + search[i].y) * lx;
			pComp = curr;

			n = bs_y;
			while (n--) {
				m = bs_x;
				while (m--) {
					AE += abs((Int)*pRef - (Int)*pComp);
					pRef += 2;
					pComp ++;
				}
				pRef += 2 * lx - 2 * bs_x;
				pComp += size_x - bs_x;
			}

			if (i==0 && bs_y==16 && *mvx==0 && *mvy==0)
				AE -= (128 + 1);

			if (AE < AE_min)
			{
				AE_min = AE;
				min_pos = i;
			}
		}
		/* else don't look outside the reference */
	}

	/* Store optimal values */
	*min_error = AE_min;
	*mvx += ((Float)search[min_pos].x)/(Float)(SubDimension);
	*mvy += ((Float)search[min_pos].y)/(Float)(SubDimension);

	// generate comp mb data for the minimum point
	pRef = prev + new_x + search[min_pos].x + (new_y + search[min_pos].y) * lx;
	pComp = curr_comp_mb;

	n = bs_y;
	while (n--) {
		m = bs_x;
		while (m--) {
			*(pComp ++) = *pRef;
			pRef += 2;
		}
		pRef += 2 * lx - 2 * bs_x;
		pComp += 16 - bs_x;
	}

	return;
}