inline uint16 avg(const uint16 col1,const uint16 col2)
{
	//res = col1/2 + col2/2
	uint16 r1,r2;
	uint16 g1,g2,b1,b2;
	r1 = (col1 & 0xF800)>>11;
	r2 = (col2 & 0xF800)>>11;

	g1 = (col1 & 0x07E0)>>5;
	g2 = (col2 & 0x07E0)>>5;

	b1 = (col1 & 0x1F);
	b2 = (col2 & 0x1F);
	return (((r1+r2)>>1)<<11) + (((g1+g2)>>1)<<5)+ ( (b1+b2)>>1) ;
}

inline uint16 GetScale2xVal(const uint16 col,const uint16 near1,const uint16 near2,const uint16 far1,const uint16 far2)
{
	if ( (near1==near2) && (near1!=far1) && (near1 != far2))
		return near1; //this is the original scale2x code (http://scale2x.sourceforge.net)
	else if ((near1==near2) && (near1==far1) && (near1==far2) && (near1 !=col))
		return avg(col,near1); // remove dither patterns
	else if ((col != near1) && (col !=near2) && (col!= far1) && (col!=far2))
		return avg(col,avg(near1,near2)); //if col is different from all neighbours ==> blur it
	else return col;
}
void AdvMame2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
							 int width, int height) {
	unsigned int nextlineSrc = srcPitch / sizeof(uint16);
	const uint16 *p = (const uint16 *)srcPtr;

	unsigned int nextlineDst = dstPitch / sizeof(uint16);
	uint16 *q = (uint16 *)dstPtr;
	
	uint16 A, B, C;
	uint16 D, E, F;
	uint16 G, H, I;

	while (height--) {
		B = *(p - 1 - nextlineSrc);
		E = *(p - 1);
		H = *(p - 1 + nextlineSrc);
		C = *(p - nextlineSrc);
		F = *(p);
		I = *(p + nextlineSrc);
		for (int i = 0; i < width; ++i) {
			p++;
			A = B; B = C; C = *(p - nextlineSrc);
			D = E; E = F; F = *(p);
			G = H; H = I; I = *(p + nextlineSrc);

			*(q) = GetScale2xVal(E,D,B,F,H);
			*(q + 1) = GetScale2xVal(E,B,F,D,H);
			*(q + nextlineDst) = GetScale2xVal(E,D,H,B,F);
			*(q + nextlineDst + 1) = GetScale2xVal(E,H,F,D,B);
			q += 2;
		}
		p += nextlineSrc - width;
		q += (nextlineDst - width) << 1;
	}
}