pictureb_sse.cpp

H.263的编码程序,加了CPU指令优化,VC版.

int  findbiSAD_sse(unsigned char *prev, unsigned char *next, unsigned char *curr, 
			   int lx_ipol,  int lx_curr, int blocksize)
{
	int sad = 0;
	__int64  mask = 0x00ff00ff00ff00ff;
	int distance = lx_ipol*2;

	if (16 == blocksize)
	{
		__asm
		{
			mov  esi, [prev]
			mov  edi, [next]
			mov  edx, [curr]
			mov  ecx, 16
			xor  eax, eax

loop028:    movq   mm6, [mask]
			movq   mm0, [edx]
			movq   mm2, [edx+8]
			movq   mm1, [esi]
			movq   mm4, [esi+8]
			movq   mm5, [edi]
			movq   mm7, [edi+8]
			pand   mm1, mm6
			pand   mm4, mm6
			pand   mm5, mm6
			pand   mm7, mm6
			packuswb mm1, mm4
			packuswb mm5, mm7
			pavgb  mm1, mm5
			movq   mm3, [esi+16]
			movq   mm4, [esi+24]
			movq   mm5, [edi+16]
			movq   mm7, [edi+24]
			pand   mm3, mm6
			pand   mm4, mm6
			pand   mm5, mm6
			pand   mm7, mm6
			packuswb mm3, mm4
			packuswb mm5, mm7
			pavgb  mm3, mm5
			psadbw mm0, mm1
			psadbw mm2, mm3
			paddw  mm0, mm2
			push   ebx
			movd   ebx, mm0
			add    eax, ebx
			pop    ebx

			add    esi, [distance]
			add    edi, [distance]
			add    edx, [lx_curr]

			dec    ecx
			jnz    loop028
			mov    [sad], eax

			emms
		}
	}
	else
	{
		__asm
		{
			mov  esi, [prev]
			mov  edi, [next]
			mov  edx, [curr]
			mov  ecx, 8
loop029:    movq   mm6, [mask]
			movq   mm0, [edx]
			movq   mm1, [esi]
			movq   mm4, [esi+8]
			movq   mm5, [edi]
			movq   mm7, [edi+8]
			pand   mm1, mm6
			pand   mm4, mm6
			pand   mm5, mm6
			pand   mm7, mm6
			packuswb mm1, mm4
			packuswb mm5, mm7
			pavgb  mm1, mm5
			psadbw mm0, mm1
    		push   ebx
			movd   ebx, mm0
			add    eax, ebx
			pop    ebx

			add    esi, [distance]
			add    edi, [distance]
			add    edx, [lx_curr]

			dec    ecx
			jnz    loop029
			mov    [sad], eax

			emms
		}
	}

	return sad;

}