/************************************************ *  Color space conversion matrix coefficients  * ************************************************//* all components are scaled up by 15 bits to	*  * yield a minimum degree of precision while	* * still using pure integer math				*/#define Yr	(short)(0.257 * 32768 + 0.5)#define Yg	(short)(0.504 * 32768 + 0.5)#define Yb	(short)(0.098 * 32768 + 0.5)#define Ur	(short)(-0.148 * 32768 + 0.5)#define Ug	(short)(-0.291 * 32768 + 0.5)#define Ub	(short)(0.439  * 32768 + 0.5)#define Vr	(short)(0.439  * 32768 + 0.5)#define Vg	(short)(-0.368 * 32768 + 0.5)#define Vb	(short)(-0.071 * 32768 + 0.5)static const short Y_COEFF[4] = {	Yb, Yg, Yr, 0x00};static const short U_COEFF[4] = {	Ub, Ug, Ur, 0x00};static const short V_COEFF[4] = {	Vb, Vg, Vr, 0x00};/* SIMD (MMX™) version */int swciRgbToYuv(swcImage* image) {		dword_ptr	src;	byte_ptr	y;	byte_ptr	u;	byte_ptr	v;	int			w, h;	if (!image || !image->rgb) {		return SWC_ERR;	}	w	= image->cols;	h	= image->rows;	src = image->rgb;		y	= image->y;	u	= image->u;	v	= image->v;	__asm {		/* push all double word registers and init FPU */		pushad		finit		/* setup pointers */		mov			esi, src		mov			edi, y		mov			ecx, u		mov			edx, v				mov			ebx, hrow_loop0:		/* loop y */		 		push		ebx		mov			ebx, w		col_loop0:		/* loop x */		push		ebx		/* ------------------------------------------------------- */		/*	 two pixels at once are processed by these routines	   */		/* ------------------------------------------------------- */		/* load two pixels */		movq		mm0, [esi]		movq		mm1, mm0		pxor		mm2, mm2		/* mm0 = 0|0|0|R1|0|G1|0|B1 */		/* mm1 = 0|0|0|R2|0|G2|0|B2 */		punpcklbw	mm0, mm2		punpckhbw	mm1, mm2		/* save unpacked 16 rgb vectors */		movq		mm3, mm0		movq		mm4, mm1		/* compute dot product for two Y components */		movq		mm5, [Y_COEFF]		pmaddwd		mm0, mm5		pmaddwd		mm1, mm5		/* shift to 15 bit precision */		psrad		mm0, 15		psrad		mm1, 15				/* tricky here - need to use mm6 as accumulator */		movq		mm6, mm0		psrlq		mm6, 32				/* add intermediate results for final Y calculation */		paddd		mm0, mm6		movd		eax, mm0		add			eax, 16				/* add intermediate results for final Y calculation (second pixel) */		movq		mm6, mm1		psrlq		mm6, 32				paddd		mm1, mm6		movd		ebx, mm1		add			ebx, 16		/* combine results from eax and ebx to one word */		shl			ebx, 8		or			eax, ebx				/* save two Y components */		mov			[edi], ax		/* process U and v components - same as above */		/* restore unpacked 16 bit vectors */		movq		mm0, mm3		movq		mm1, mm4		movq		mm5, [U_COEFF]		pmaddwd		mm0, mm5		pmaddwd		mm1, mm5		psrad		mm0, 15		psrad		mm1, 15				movq		mm6, mm0		psrlq		mm6, 32						paddd		mm0, mm6		movd		eax, mm0		add			eax, 128		movq		mm6, mm1		psrlq		mm6, 32				paddd		mm1, mm6		movd		ebx, mm1		add			ebx, 128		/* add the two components */		add			eax, ebx		/* save intermediate result */		push		eax		/* restore unpacked 16 bit vectors */		movq		mm0, mm3		movq		mm1, mm4		movq		mm5, [V_COEFF]		pmaddwd		mm0, mm5		pmaddwd		mm1, mm5		psrad		mm0, 15		psrad		mm1, 15				movq		mm6, mm0		psrlq		mm6, 32						paddd		mm0, mm6		movd		eax, mm0		add			eax, 128		movq		mm6, mm1		psrlq		mm6, 32				paddd		mm1, mm6		movd		ebx, mm1		add			ebx, 128		/* add the two components */		add			eax, ebx		/* save intermediate result */		push		eax		mov			ebx, w		/* load two pixels  from second scanline */		movq		mm0, [esi + ebx*4]		movq		mm1, mm0		pxor		mm2, mm2				/* mm0 = 0|0|0|R1|0|G1|0|B1 */		/* mm1 = 0|0|0|R2|0|G2|0|B2 */		punpcklbw	mm0, mm2		punpckhbw	mm1, mm2		/* save unpacked 16 rgb vectors */		movq		mm3, mm0		movq		mm4, mm1		/* compute dot product for second Y component */		movq		mm5, [Y_COEFF]		pmaddwd		mm0, mm5		pmaddwd		mm1, mm5		psrad		mm0, 15		psrad		mm1, 15				movq		mm6, mm0		psrlq		mm6, 32				paddd		mm0, mm6		movd		eax, mm0		add			eax, 16				movq		mm6, mm1		psrlq		mm6, 32				paddd		mm1, mm6		movd		ebx, mm1		add			ebx, 16		shl			ebx, 8		or			eax, ebx				/* save Y component */		mov			ebx, w		mov			[edi + ebx], ax		/* restore unpacked 16 bit vectors */		movq		mm0, mm3		movq		mm1, mm4		/* IMPORTANT NOTE: need to process V first, since last value on stack */		/* is intermediate result from V components!						  */		movq		mm5, [V_COEFF]		pmaddwd		mm0, mm5		pmaddwd		mm1, mm5		psrad		mm0, 15		psrad		mm1, 15				movq		mm6, mm0		psrlq		mm6, 32						paddd		mm0, mm6		movd		eax, mm0		add			eax, 128		movq		mm6, mm1		psrlq		mm6, 32				paddd		mm1, mm6		movd		ebx, mm1		add			ebx, 128		/* add the two components */		add			eax, ebx		/* restore intermediate result from last scanline */		pop			ebx		/* add current values and average them */		add			eax, ebx		sar			eax, 2		/* store V component */		mov			[edx], al		/* restore unpacked 16 bit vectors */		movq		mm0, mm3		movq		mm1, mm4		movq		mm5, [U_COEFF]		pmaddwd		mm0, mm5		pmaddwd		mm1, mm5		psrad		mm0, 15		psrad		mm1, 15				movq		mm6, mm0		psrlq		mm6, 32						paddd		mm0, mm6		movd		eax, mm0		add			eax, 128		movq		mm6, mm1		psrlq		mm6, 32				paddd		mm1, mm6		movd		ebx, mm1		add			ebx, 128		/* add the two components */		add			eax, ebx		/* restore intermediate result from last scanline */		pop			ebx		/* add current values and average them */		add			eax, ebx		sar			eax, 2		/* store U component */		mov			[ecx], al		/* increment Y,U,V target pointers */		add			edi, 2		inc			ecx		inc			edx		/* increment source pointer */		add			esi, 8		/* endloop x */		pop			ebx		sub			ebx, 2		jnz			col_loop0		/* skip to next scanline */		mov			ebx, w		add			edi, ebx		shl			ebx, 2		add			esi, ebx		/* endloop y */		pop			ebx		sub			ebx, 2		jnz			row_loop0		emms		popad	}	return SWC_OK;}int swciYuvToRgb(swcImage* image) {	dword_ptr tmp;	int		  w, h;	byte_ptr  y, u, v;		if (!image || !image->rgb) {		/* invalid params */		return SWC_ERR;	}	/* copy image properties to local variables to make them accessable from	 * within our inline assembly core	*/	tmp = image->rgb;	w = image->cols;	h = image->rows;	y = image->y;	u = image->u;	v = image->v;	__asm {		pushad				/* save all dword registers */		finit				/* initialize FPU */		/* actual MMX implementation follows here */		xor	eax, eax		/* clear eax to zero */				mov	ebx, h			/* store image height in ebx */		mov	ecx, w			/* store image width in ecx */		mov	edx, y				mov	edi, u		mov	esi, vrow_loop:		push	ebx			/* save image height - we use ebx for many things */		mov		ebx, ecx	/* setup col count */col_loop:		push	ebx			/* save col count - we mess with ebx later */		xor		ebx, ebx	/* clear all 32 bits */		/* mm0 = Bu | Gu + Gv | Rv | 0000  */		mov			al, [edi]		mov			bl, [esi]		movq		mm0, [CoefRgbU + 8 * eax]		paddw		mm0, [CoefRgbV + 8 * ebx]		/* for each U,V pair there are four Y components to process (YUV420!) */		/* first scanline */		/* mm1 = By0 | Gy0 | Ry0 | 0000, mm1 = By1 | Gy1 | Ry1 | 0000 */		mov			al, [edx]		mov			bl, [edx+1]		movq		mm1, [CoefRgbY + 8 * eax]		movq		mm2, [CoefRgbY + 8 * ebx]				/* next scanline */		/* mm3 = By2 | Gy2 | Ry2 | 0000, mm4 = By3 | Gy3 | Ry3 | 0000 */		mov			al, [edx + ecx]		mov			bl, [edx + ecx + 1]		movq		mm3, [CoefRgbY + 8 * eax]		movq		mm4, [CoefRgbY + 8 * ebx]		/* process all four pixels at one iteration */		paddw		mm1, mm0		paddw		mm2, mm0		paddw		mm3, mm0		paddw		mm4, mm0		/* divide by 64 to scale to byte values */		psraw		mm1, 6		psraw		mm2, 6		psraw		mm3, 6		psraw		mm4, 6		/* now we need to clamp values - packuswb comes in handy here:		 *	values below zero are clamped to zero, values above 255 are set to 255		 */		packuswb	mm1, mm2		packuswb	mm3, mm4		/* now store the for pixels from mm1 and mm3 into target */		mov			ebx, tmp		movntq		[ebx], mm1		movq		[ebx + 4 * ecx], mm3		/* proceed to next pixel pair */		add			tmp, 8		add			edx, 2		add			edi, 1		add			esi, 1		/* loop */		pop			ebx		sub			ebx, 2		jnz			col_loop		/* skip to next row */		mov			ebx, tmp			lea			ebx, [ebx + 4*ecx]		add			edx, ecx		mov			tmp, ebx		pop			ebx		sub			ebx, 2		jnz			row_loop		emms				/* clear FPU state */		popad				/* pop all dword registers back from stack :) */	};	return SWC_OK;}