#include <stdio.h>
#include <stdlib.h>
#include <windows.h>

BYTE src1[640*480];
BYTE src2[640*480];
BYTE dst1[640*480];
BYTE dst2[640*480];

void main(void)
 {
	int i, j;

	srand(timeGetTime( ));

	for(i = 0; i < 640*480/4; i ++)
	 {
		int t = rand( ) & 0x0f;
		DWORD masktmp[16] = {
				0x00000000, 0x000000ff, 0x0000ff00, 0x0000ffff,
				0x00ff0000, 0x00ff00ff, 0x00ffff00, 0x00ffffff,
				0xff000000, 0xff0000ff, 0xff00ff00, 0xff00ffff,
				0xffff0000, 0xffff00ff, 0xffffff00, 0xffffffff };
		DWORD tmp = rand( );
		tmp &= masktmp[t];
		for(j = 0; j < 4; j ++)
		 {
			dst1[i*4+j] = src1[i*4+j] = (BYTE)(tmp & 0xff);
			tmp >>= 8;
		 }
	 }

	for(i = 0; i < 640*480; i ++)
		src2[i] = (BYTE)(rand( ) & 0xff);

	for(i = 0; i < 640*480; i ++)
		dst2[i] = (src1[i] == 0) ? src2[i] : src1[i];

	DWORD srcAddress = (DWORD)src1;
	DWORD dstAddress = (DWORD)dst1;

	DWORD spr = (DWORD)dst1;
	DWORD bak = (DWORD)src2;

	_asm
	 {
		mov ecx, 640*480/4
		mov edx, spr
		mov ebx, bak

		shr ecx, 1
		sub edx, 8
	 LP:	mov eax, [edx + 8]
			add edx, 8
		// 2
			mov edi, [edx + 4]
			and eax, 0x7f7f7f7f
		// 3
			and edi, 0x7f7f7f7f
			add eax, 0x7f7f7f7f
		// 4
			add edi, 0x7f7f7f7f
			mov esi, [edx]
		// 5
			or  eax, esi
			mov esi, [edx + 4]
		// 6
			or  edi, esi
			and eax, 0x80808080
		// 7
			shr eax, 7
			and edi, 0x80808080
		// 8
			shr edi, 7
			add eax, 0x7f7f7f7f
		// 9
			xor eax, 0x80808080
			mov esi, [ebx]
		// 10
			and eax, esi
			mov esi, [edx]
		// 11
			or  eax, esi
			add edi, 0x7f7f7f7f
		// 12
			xor edi, 0x80808080
			mov esi, [ebx+4]
		// 13
			and edi, esi
			mov esi, [edx+4]
		// 14
			or  edi, esi
			mov [edx], eax
		// 15
			mov [edx+4], edi
			add ebx, 8
		// 16
			dec ecx
			jnz short LP
	 }

	DWORD *pt1 = (DWORD*)src1;
	DWORD *pt2 = (DWORD*)src2;
	DWORD *pt3 = (DWORD*)dst1;
	DWORD *pt4 = (DWORD*)dst2;

	for(i = 0; i < 640*480/4; i ++)
	 {
		if(*pt3 != *pt4)
			printf("%8x:%8x -> %8x %8x\n", *pt1, *pt2, *pt3, *pt4);
		pt1 ++;
		pt2 ++;
		pt3 ++;
		pt4 ++;
	 }

 }