#include "stdafx.h"

#include "BlobStorage.h"
#include <wmmintrin.h>

#ifdef _DEBUG
#define new DEBUG_NEW
#endif

#include "SISAssem.h"
#include "AssemParam.h"

int CSISAssem::ConvPixelVertS_Intrin(BYTE *pImgBuff, __int64 Threshold, __int64 maxDefect, __int64 nDefect
							   , __int64 left, __int64 right, __int64 top, __int64 bottom
							   , int* DefectX, int* DefectY, short* DefectT, short* DefectValue
							   , int* pGraySrc, int* pGrayCmp, int* pZoneID, int* pZoneTh
							   , __int64 frameWidth, __int64 iPitch16, __int64 fSPR, __int64 sSPR
							   , short* ZoneThTable, short* ZoneIdTable, __int64 ZoneTableWidth, void* pMatchTable
							   , CDefectPair *pDefectdPair, __int64 StartPair, __int64 MaxPair, __int64 PairSize
							   , __int64 dxPattern, __int64 dyPattern, __int64 yZoneByte, __int64 bReverseFilter
							   , __int64 ThresholdSlide, __int64 ThresholdSuppress, __int64 bUseUnpair, short* pConBuff
							   , int* endLine, __int64 ExTh
							   )
{
	const __int64 SprShift = 4;
	const __int64 SprMulti = 16;

	__m128i	xxTemp;

	__m128i	xxZTh;

	__m128i	xxSrc1;
	__m128i xxSrc2;

	__m128i	xxCmp1;
	__m128i xxCmp2;
	__m128i xxCmp3;

	__m128i	xxPeakS;
	__m128i	xxPeakAbs;

	__m128i	xxfSpr;
	__m128i	xxsSpr;

	__m128i xxDStart;
	__m128i xxDSlide;

	__int64 rax, rbx, rcx, rdx;

	__int64 ImgAddrOri	= (__int64)pImgBuff;
	__int64 ConAddr = (__int64)pConBuff;
	__int64 c8Pixel, rrNoD = 0, rrImgAddr, nPair = StartPair;

	__m128i xxSupp = _mm_set1_epi16((short)ThresholdSuppress);	

	xxDStart = _mm_set1_epi16((short)Threshold);
	xxDSlide = _mm_set1_epi16((short)ThresholdSlide);

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	short sSrc, sRef, sThres, sDiff;
	__m128i	xxZThB, xxZThW, xx0xff = _mm_set1_epi16(-1);
	__m128i xxDStartB(xxDStart);
	__m128i xxDSlideB(xxDSlide);

	if(m_Param.s_nBWB_Main > 0)
	{
		xxDStartB = _mm_set1_epi16((short)m_Param.s_nBWB_Main);
		xxDSlideB = _mm_set1_epi16((short)m_Param.s_nBWB_Slide);
	}
	else
	{
		m_Param.s_nBWB_Main = (int)Threshold;
		m_Param.s_nBWB_Slide = (int)ThresholdSlide;
	}
	// ----------------------------------- //

	rcx = iPitch16;
	rcx &= 0x0f;
	xxsSpr = _mm_set1_epi16((short)rcx);

	rax = SprMulti;
	rax -= rcx;
	xxfSpr = _mm_set1_epi16((short)rax);

	__int64 iPitch = iPitch16 / SprMulti;
	__int64 iPitchHUp = iPitch;
	if(rcx >= SprMulti / 2)
		iPitchHUp++;

	bottom--;
	__int64 ConBottom = bottom - iPitch - 1;
	__int64 pairBottom = ConBottom - iPitch;

	// -------------- à» Filtering -------------- //
	CRect rtDeadZone[MAX_DEAD_ZONE_COUNT];
	CPoint ptDefect;
	memcpy(rtDeadZone, m_Param.s_RectDeadZone, sizeof(CRect) * MAX_DEAD_ZONE_COUNT);
	int nDeadZoneCount = 0;
	for(int i = 0; i < MAX_DEAD_ZONE_COUNT; i++)
	{
		if(rtDeadZone[i].IsRectEmpty() == TRUE)
		{
			nDeadZoneCount = i;
			break;
		}
	}
	// ------------------------------------------ //

	__int64 pairTop = top + iPitch;
	__int64 ConvPitch = iPitch * 0x20;

	__int64 ix = left;
	__int64	iy = top;

	right = (right - left) / 16 * 16 + left;

	__m128i xxZero = _mm_setzero_si128();

	CDefectPair* PairAddr = pDefectdPair + nPair;

//LOOP_1:
	goto LOOP_1_CHECK;

LOOP_1_TAIL:
	goto PAIRING_VERT;

_return_PAIRING_VERT:
	ix += 0x10;

LOOP_1_CHECK:
	if(ix >= right)
		goto END_FUNC;

//LOOP_1_BODY:
	iy = top;
	rrImgAddr = ImgAddrOri + frameWidth * iy + ix;
	goto MAKE_CONBUFF_HEAD;

_return_MAKE_CONBUFF:
	iy = top;
	rrImgAddr = ConAddr;
//LOOP_2:
	goto LOOP_2_CHECK;

LOOP_2_TAIL:
	iy++;
	rrImgAddr += 0x20;

LOOP_2_CHECK:
	if(iy >= ConBottom)
		goto LOOP_1_TAIL;

//LOOP_2_BODY:
	goto XMM_MAKE;
_return_XMM_MAKE:
	goto XMM_COMPARE;
_return_XMM_COMPARE:
	goto LOOP_2_TAIL;

MAKE_CONBUFF_HEAD:
	rax = ConAddr;

MAKE_CONBUFF:
	xxSrc1 = _mm_loadu_si128((__m128i*) rrImgAddr);
	_mm_storeu_si128((__m128i*)rax, _mm_unpacklo_epi8(xxSrc1, xxZero));
	_mm_storeu_si128((__m128i*)(rax + 0x10), _mm_unpackhi_epi8(xxSrc1, xxZero));
	rrImgAddr += frameWidth;
	rax += 0x20;
	iy++;
	if(iy <= bottom)
		goto MAKE_CONBUFF;
	goto _return_MAKE_CONBUFF;

XMM_MAKE:
	xxSrc1 = _mm_loadu_si128((__m128i*) rrImgAddr);
	xxSrc2 = _mm_loadu_si128((__m128i*) (rrImgAddr + 0x10));

	rax = rrImgAddr + ConvPitch;

	xxCmp1 = _mm_loadu_si128((__m128i*) rax);
	xxCmp2 = _mm_loadu_si128((__m128i*) (rax + 0x10));
	xxCmp3 = _mm_loadu_si128((__m128i*) (rax + 0x20));
	xxTemp = _mm_loadu_si128((__m128i*) (rax + 0x30));

	xxCmp1 = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(xxCmp1, xxfSpr), _mm_mullo_epi16(xxCmp3, xxsSpr)), SprShift);
	xxCmp2 = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(xxCmp2, xxfSpr), _mm_mullo_epi16(xxTemp, xxsSpr)), SprShift);
	goto _return_XMM_MAKE;

XMM_COMPARE:
	c8Pixel = 0;
	xxSrc1 = _mm_min_epu16(xxSrc1, xxSupp);
	xxCmp1 = _mm_min_epu16(xxCmp1, xxSupp);
	goto _XMM_COMPARE_8PIXEL;

_return_XMM_COMPARE_8PIXEL:
	if(rrNoD >= maxDefect)
		goto PAIRING_VERT;

	c8Pixel++;
	if(c8Pixel == 2)
		goto XMM_COMPARE_END;

	xxSrc1 = _mm_min_epu16(xxSrc2, xxSupp);
	xxCmp1 = _mm_min_epu16(xxCmp2, xxSupp);

_XMM_COMPARE_8PIXEL:
	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	//¿ø·¡ÄÚµå: xxZTh = _mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_max_epu16(xxSrc1, xxCmp1), xxDSlide), 8), xxDStart);
	xxZThW = _mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_max_epu16(xxSrc1, xxCmp1), xxDSlide), 8), xxDStart);
	xxZThB = _mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_max_epu16(xxSrc1, xxCmp1), xxDSlideB), 8), xxDStartB);

		xxZTh = _mm_min_epu16(xxZThW, xxZThB);
	// ----------------------------------- //

	//Â÷ÀÌ°ªÀ» ±¸ÇÏ°í
	xxPeakS = _mm_sub_epi16(xxSrc1, xxCmp1);
	xxPeakAbs = _mm_abs_epi16(xxPeakS);
	xxTemp = _mm_cmpgt_epi16(xxPeakAbs, xxZTh);
	if(_mm_movemask_epi8(xxTemp) != 0)
		goto DEFECT_00;

	ix += 8;
	goto _return_XMM_COMPARE_8PIXEL;

XMM_COMPARE_END:
	ix -= 0x10;
	goto _return_XMM_COMPARE;


DEFECT_00:
	if(_mm_extract_epi16(xxTemp, 0x00) == 0)
		goto DEFECT_01;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x00);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x00);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x00);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x00);
	if((short)_mm_extract_epi16(xxPeakS, 0x00)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_01:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x01) == 0)
		goto DEFECT_02;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x01);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x01);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x01);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x01);
	if((short)_mm_extract_epi16(xxPeakS, 0x01)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_02:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x02) == 0)
		goto DEFECT_03;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x02);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x02);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x02);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x02);
	if((short)_mm_extract_epi16(xxPeakS, 0x02)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_03:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x03) == 0)
		goto DEFECT_04;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x03);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x03);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x03);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x03);
	if((short)_mm_extract_epi16(xxPeakS, 0x03)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_04:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x04) == 0)
		goto DEFECT_05;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x04);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x04);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x04);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x04);
	if((short)_mm_extract_epi16(xxPeakS, 0x04)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_05:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x05) == 0)
		goto DEFECT_06;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x05);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x05);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x05);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x05);
	if((short)_mm_extract_epi16(xxPeakS, 0x05)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_06:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x06) == 0)
		goto DEFECT_07;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x06);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x06);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x06);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x06);
	if((short)_mm_extract_epi16(xxPeakS, 0x06)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_07:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x07) == 0)
		goto DEFECT_08;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x07);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x07);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x07);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x07);
	if((short)_mm_extract_epi16(xxPeakS, 0x07)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_08:
	ix++;
	goto _return_XMM_COMPARE_8PIXEL;

PAIRING_VERT:
	iy = -1;

LOOP_PAIR_NUM:
	iy++;
	if(iy >= rrNoD)
		goto LOOP_PAIR_NUM_END;

	rax = nPair;
	if(rax >= MaxPair)
		goto END_FUNC;

	if(*(DefectT + iy) >= 3)
		goto LOOP_PAIR_NUM;

	rbx = iy;
	rcx = *(DefectY + iy);
	rax = rcx + iPitchHUp;
	rrImgAddr = *(DefectX + iy);

	if(rcx > pairBottom)
		goto PAIR_BOTTOM;

	if(rcx >= pairTop)
		goto PAIR_CENTER;

PAIR_TOP:
	rbx++;
	if(rbx >= rrNoD)
		goto PAIR_TOP_UNPAIR;

	rdx = *(DefectY + rbx);
	if(rdx < rax)
		goto PAIR_TOP;
	else if(rdx > rax)
		goto PAIR_TOP_UNPAIR;

	rcx = *(DefectX + rbx);
	if(rrImgAddr != rcx)
		goto PAIR_TOP;

	rrImgAddr = *(DefectT + rbx);
	if(rrImgAddr >= 3)
		goto LOOP_PAIR_NUM;

	rax = (__int64)*(DefectT + iy);
	if(rax == rrImgAddr)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rcx;
	ptDefect.y = (LONG)rdx;
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	sSrc = (int)(*(pGraySrc + rbx));
	sRef = (int)(*(pGrayCmp + rbx));

	if(sSrc > sRef)
		sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

	*(pZoneTh + rbx) = sThres;

	sDiff = abs(sSrc - sRef);
	if(sDiff <= sThres)
		goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	*(DefectT + rbx) = 4;

	PairAddr->s_DefectPos = (DefectPos)1;
	PairAddr->s_DefectPair = (DefectPair)1;
	PairAddr->s_nDefectX = (int)rcx;
	PairAddr->s_nDefectY = (int)rdx;
	PairAddr->s_DefectType = (DefectType)rrImgAddr;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc);
	PairAddr->s_nGrayRef = (int)(sRef);
	PairAddr->s_nZone = (int)(*(pZoneID + rbx));
	PairAddr->s_nThresold = (int)(sThres);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_TOP_UNPAIR:
	if(bUseUnpair != 1)
		goto LOOP_PAIR_NUM;

	//rcx = (__int64)*(DefectValue + iy);
	//if(rcx <= ExTh)
	//	goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)(rrImgAddr);
	ptDefect.y = (LONG)(rax - iPitchHUp);
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	sSrc = (int)(*(pGraySrc + iy));
	sRef = (int)(*(pGrayCmp + iy));

	if(sSrc > sRef)
		sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

	*(pZoneTh + iy) = sThres;

	sDiff = abs(sSrc - sRef);
	if(sDiff <= sThres)
		goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	PairAddr->s_DefectPos = (DefectPos)1;
	PairAddr->s_DefectPair = (DefectPair)5;
	PairAddr->s_nDefectX = (int)ptDefect.x;
	PairAddr->s_nDefectY = (int)ptDefect.y;
	PairAddr->s_DefectType = (DefectType)*(DefectT + iy);
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + iy);
	PairAddr->s_nGraySrc = (int)(sSrc);
	PairAddr->s_nGrayRef = (int)(sRef);
	PairAddr->s_nZone = (int)(*(pZoneID + iy));
	PairAddr->s_nThresold = (int)(sThres);
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_CENTER:
	rbx++;
	if(rbx >= rrNoD)
		goto LOOP_PAIR_NUM;

	rdx = *(DefectY + rbx);
	if(rdx < rax)
		goto PAIR_CENTER;
	else if(rdx > rax)
		goto LOOP_PAIR_NUM;

	rcx = *(DefectX + rbx);
	if(rrImgAddr != rcx)
		goto PAIR_CENTER;

	rrImgAddr = *(DefectT + rbx);
	if(rrImgAddr >= 3)
		goto LOOP_PAIR_NUM;

	rax = (__int64)*(DefectT + iy);
	if(rax == rrImgAddr)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rcx;
	ptDefect.y = (LONG)rdx;
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
		sSrc = (int)(*(pGraySrc + rbx));
		sRef = (int)(*(pGrayCmp + rbx));

		if(sSrc > sRef)
			sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
		else
			sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

		*(pZoneTh + rbx) = sThres;

		sDiff = abs(sSrc - sRef);
		if(sDiff <= sThres)
			goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	*(DefectT + rbx) = 4;

	PairAddr->s_DefectPos = (DefectPos)2;
	PairAddr->s_DefectPair = (DefectPair)1;
	PairAddr->s_nDefectX = (int)rcx;
	PairAddr->s_nDefectY = (int)rdx;
	PairAddr->s_DefectType = (DefectType)rrImgAddr;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc);
	PairAddr->s_nGrayRef = (int)(sRef);
	PairAddr->s_nZone = (int)(*(pZoneID + rbx));
	PairAddr->s_nThresold = (int)(sThres);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_BOTTOM:
	if(bUseUnpair != 1)
		goto LOOP_PAIR_NUM;

	//rcx = (__int64)*(DefectValue + iy);
	//if(rcx <= ExTh)
	//	goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rrImgAddr;
	ptDefect.y = (LONG)rax;
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	sSrc = (int)(*(pGrayCmp + iy));
	sRef = (int)(*(pGraySrc + iy));

	if(sSrc > sRef)
		sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

	*(pZoneTh + iy) = sThres;

	sDiff = abs(sSrc - sRef);
	if(sDiff <= sThres)
		goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	PairAddr->s_DefectPos = (DefectPos)3;
	PairAddr->s_DefectPair = (DefectPair)5;
	PairAddr->s_nDefectX = (int)rrImgAddr;
	PairAddr->s_nDefectY = (int)rax;
	rdx = (*(DefectT + iy) + 1) & 1;
	PairAddr->s_DefectType = (DefectType)rdx;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + iy);
	PairAddr->s_nGraySrc = (int)sSrc;
	PairAddr->s_nGrayRef = (int)sRef;
	PairAddr->s_nZone = (int)(*(pZoneID + iy));
	PairAddr->s_nThresold = (int)sThres;
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

LOOP_PAIR_NUM_END:
	rrNoD = 0;
	goto _return_PAIRING_VERT;

END_FUNC:
	ix += 0x10;
	*endLine = (int)ix;

	return (int)nPair;
}

int CSISAssem::ConvPixelVTDS_Intrin(BYTE *pImgBuff, __int64 Threshold, __int64 maxDefect, __int64 nDefect
							, __int64 left, __int64 right, __int64 top, __int64 bottom
							, int* DefectX, int* DefectY, short* DefectT, short* DefectValue
							, int* pGraySrc, int* pGrayCmp, int* pZoneID, int* pZoneTh
							, __int64 frameWidth, __int64 iPitch16, int* piPitch16, __int64 sSPR
							, short* ZoneThTable, short* ZoneIdTable, __int64 ZoneTableWidth, void* pMatchTable
							, CDefectPair *pDefectdPair, __int64 StartPair, __int64 MaxPair, __int64 PairSize
							, __int64 dxPattern, __int64 dyPattern, __int64 yZoneByte, __int64 bReverseFilter
							, __int64 ThresholdSlide, __int64 ThresholdSuppress, __int64 bUseUnpair, short* pConBuff
							, int* endLine, __int64 ExTh
							)
{
	const __int64 SprShift = 4;
	const __int64 SprMulti = 16;

	__m128i	xxTemp;

	__m128i	xxZTh;

	__m128i	xxSrc1;
	__m128i xxSrc2;

	__m128i	xxCmp1;
	__m128i xxCmp2;
	__m128i xxCmp3;

	__m128i	xxPeakS;
	__m128i	xxPeakAbs;

	__m128i	xxfSpr;
	__m128i	xxsSpr;

	__m128i xxDStart;
	__m128i xxDSlide;

	__int64 rax, rbx, rcx, rdx;


	__int64 ImgAddrOri	= (__int64)pImgBuff;
	__int64 ConAddr = (__int64)pConBuff;
	__int64 c8Pixel, rrNoD = 0, rrImgAddr, nPair = StartPair, _piPitch16, cDpc;

	__m128i xxSupp = _mm_set1_epi16((short)ThresholdSuppress);
		
	xxDStart = _mm_set1_epi16((short)Threshold);
	xxDSlide = _mm_set1_epi16((short)ThresholdSlide);

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	short sSrc, sRef, sThres, sDiff;
	__m128i	xxZThB, xxZThW, xx0xff = _mm_set1_epi16(-1);
	__m128i xxDStartB(xxDStart);
	__m128i xxDSlideB(xxDSlide);

	if(m_Param.s_nBWB_Main > 0)
	{
		xxDStartB = _mm_set1_epi16((short)m_Param.s_nBWB_Main);
		xxDSlideB = _mm_set1_epi16((short)m_Param.s_nBWB_Slide);
	}
	else
	{
		m_Param.s_nBWB_Main = (int)Threshold;
		m_Param.s_nBWB_Slide = (int)ThresholdSlide;
	}
	// ----------------------------------- //

	rcx = iPitch16;
	rcx &= 0x0f;
	xxsSpr = _mm_set1_epi16((short)rcx);

	rax = SprMulti;
	rax -= rcx;
	xxfSpr = _mm_set1_epi16((short)rax);

	__int64 iPitch = iPitch16 / SprMulti;
	__int64 iPitchHUp = iPitch;
	if(rcx >= SprMulti / 2)
		iPitchHUp++;

	bottom--;
	__int64 ConBottom = bottom - iPitch - 1;
	__int64 pairBottom = ConBottom - iPitch;

	// -------------- à» Filtering -------------- //
	CRect rtDeadZone[MAX_DEAD_ZONE_COUNT];
	CPoint ptDefect;
	memcpy(rtDeadZone, m_Param.s_RectDeadZone, sizeof(CRect) * MAX_DEAD_ZONE_COUNT);
	int nDeadZoneCount = 0;
	for(int i = 0; i < MAX_DEAD_ZONE_COUNT; i++)
	{
		if(rtDeadZone[i].IsRectEmpty() == TRUE)
		{
			nDeadZoneCount = i;
			break;
		}
	}
	// ------------------------------------------ //

	__int64 pairTop = top + iPitch;
	__int64 ConvPitch = iPitch * 0x20;

	__int64 ix = left;
	__int64	iy = top;

	right = (right - left) / 16 * 16 + left;

	__m128i xxZero = _mm_setzero_si128();

	CDefectPair* PairAddr = pDefectdPair + nPair;

	//LOOP_1:
	goto LOOP_1_CHECK;

LOOP_1_TAIL:
	goto PAIRING_VERT;

_return_PAIRING_VERT:
	ix += 0x10;

LOOP_1_CHECK:
	if(ix >= right)
		goto END_FUNC;

	//LOOP_1_BODY:
	iy = top;
	rrImgAddr = ImgAddrOri + frameWidth * iy + ix;
	
	_piPitch16 = 0;
	if(piPitch16 == NULL)
		goto DPC_PASS;

	rax = iy;
	rax /= 16;
	rax *= 4;
	rax += __int64(piPitch16);
	_piPitch16 = rax;

DPC_PASS:	
	goto MAKE_CONBUFF_HEAD;

_return_MAKE_CONBUFF:
	iy = top;
	rrImgAddr = ConAddr;
	//LOOP_2:
	cDpc = 0x10;
	goto LOOP_2_CHECK;

LOOP_2_TAIL:
	iy++;
	rrImgAddr += 0x20;

LOOP_2_CHECK:
	if(iy >= ConBottom)
		goto LOOP_1_TAIL;

	//LOOP_2_BODY:
	if(_piPitch16 == 0)
		goto XMM_MAKE;
	goto DPC_MAKE;
_return_DPC_MAKE:
	cDpc++;
	goto XMM_MAKE;
_return_XMM_MAKE:
	goto XMM_COMPARE;
	_return_XMM_COMPARE:
	goto LOOP_2_TAIL;

MAKE_CONBUFF_HEAD:
	rax = ConAddr;

MAKE_CONBUFF:
	xxSrc1 = _mm_loadu_si128((__m128i*) rrImgAddr);
	_mm_storeu_si128((__m128i*)rax, _mm_unpacklo_epi8(xxSrc1, xxZero));
	_mm_storeu_si128((__m128i*)(rax + 0x10), _mm_unpackhi_epi8(xxSrc1, xxZero));
	rrImgAddr += frameWidth;
	rax += 0x20;
	iy++;
	if(iy <= bottom)
		goto MAKE_CONBUFF;
	goto _return_MAKE_CONBUFF;

DPC_MAKE:	// verticla Àº 16¹ø¿¡ Çѹø¾¿¸¸ ·ÎµùÇÑ´Ù. cDpc Ä«¿îÆ®, µ¿ÀÏ Pitch ÆÇ´ÜÈÄ Pass ±â´É
	if(cDpc != 0x10)				//cmp		cDpc, 010h
		goto _return_DPC_MAKE;		//jne		_return_DPC_MAKE
	cDpc = 0;						//mov		cDpc, 0
	rax = _piPitch16;				//mov		rax, piPitch16
	rax = *(int*)(rax);				//movsxd	rax, dword ptr[rax]
	_piPitch16 +=4;					//add		piPitch16, 4

	if(iPitch16 == rax)				//cmp		iPitch16, rax
		goto _return_DPC_MAKE;		//je		_return_DPC_MAKE

	iPitch16 = rax;					//mov		iPitch16, rax	
	rcx = rax;						//mov		rcx, rax
	rcx &= 0x0f;					//and		rcx, 0fh

	rax /= 16;						//shr		rax, SprShift
	iPitch = rax;					//mov		iPitch, rax

	rax *= 32;						//sal		rax, 5
	ConvPitch = rax;				//mov		ConPitch, rax


	rdx = 0x10;						//mov		rdx, 010h
	rdx -= rcx;						//sub		rdx, rcx

	xxsSpr = _mm_set1_epi16((short)rcx);
	xxfSpr = _mm_set1_epi16((short)rdx);

	goto _return_DPC_MAKE;

XMM_MAKE:
	xxSrc1 = _mm_loadu_si128((__m128i*) rrImgAddr);
	xxSrc2 = _mm_loadu_si128((__m128i*) (rrImgAddr + 0x10));

	rax = rrImgAddr + ConvPitch;

	xxCmp1 = _mm_loadu_si128((__m128i*) rax);
	xxCmp2 = _mm_loadu_si128((__m128i*) (rax + 0x10));
	xxCmp3 = _mm_loadu_si128((__m128i*) (rax + 0x20));
	xxTemp = _mm_loadu_si128((__m128i*) (rax + 0x30));

	xxCmp1 = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(xxCmp1, xxfSpr), _mm_mullo_epi16(xxCmp3, xxsSpr)), SprShift);
	xxCmp2 = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(xxCmp2, xxfSpr), _mm_mullo_epi16(xxTemp, xxsSpr)), SprShift);
	goto _return_XMM_MAKE;

XMM_COMPARE:
	c8Pixel = 0;
	xxSrc1 = _mm_min_epu16(xxSrc1, xxSupp);
	xxCmp1 = _mm_min_epu16(xxCmp1, xxSupp);
	goto _XMM_COMPARE_8PIXEL;

_return_XMM_COMPARE_8PIXEL:
	if(rrNoD >= maxDefect)
		goto PAIRING_VERT;

	c8Pixel++;
	if(c8Pixel == 2)
		goto XMM_COMPARE_END;

	xxSrc1 = _mm_min_epu16(xxSrc2, xxSupp);
	xxCmp1 = _mm_min_epu16(xxCmp2, xxSupp);

_XMM_COMPARE_8PIXEL:
	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	//¿ø·¡ÄÚµå: xxZTh = _mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_max_epu16(xxSrc1, xxCmp1), xxDSlide), 8), xxDStart);
	xxZThW = _mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_max_epu16(xxSrc1, xxCmp1), xxDSlide), 8), xxDStart);
	xxZThB = _mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_max_epu16(xxSrc1, xxCmp1), xxDSlideB), 8), xxDStartB);

		xxZTh = _mm_min_epu16(xxZThW, xxZThB);
	// ----------------------------------- //

	//Â÷ÀÌ°ªÀ» ±¸ÇÏ°í
	xxPeakS = _mm_sub_epi16(xxSrc1, xxCmp1);
	xxPeakAbs = _mm_abs_epi16(xxPeakS);
	xxTemp = _mm_cmpgt_epi16(xxPeakAbs, xxZTh);
	if(_mm_movemask_epi8(xxTemp) != 0)
		goto DEFECT_00;

	ix += 8;
	goto _return_XMM_COMPARE_8PIXEL;

XMM_COMPARE_END:
	ix -= 0x10;
	goto _return_XMM_COMPARE;


DEFECT_00:
	if(_mm_extract_epi16(xxTemp, 0x00) == 0)
		goto DEFECT_01;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x00);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x00);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x00);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x00);
	if((short)_mm_extract_epi16(xxPeakS, 0x00)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_01:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x01) == 0)
		goto DEFECT_02;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x01);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x01);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x01);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x01);
	if((short)_mm_extract_epi16(xxPeakS, 0x01)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_02:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x02) == 0)
		goto DEFECT_03;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x02);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x02);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x02);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x02);
	if((short)_mm_extract_epi16(xxPeakS, 0x02)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_03:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x03) == 0)
		goto DEFECT_04;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x03);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x03);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x03);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x03);
	if((short)_mm_extract_epi16(xxPeakS, 0x03)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_04:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x04) == 0)
		goto DEFECT_05;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x04);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x04);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x04);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x04);
	if((short)_mm_extract_epi16(xxPeakS, 0x04)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_05:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x05) == 0)
		goto DEFECT_06;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x05);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x05);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x05);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x05);
	if((short)_mm_extract_epi16(xxPeakS, 0x05)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_06:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x06) == 0)
		goto DEFECT_07;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x06);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x06);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x06);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x06);
	if((short)_mm_extract_epi16(xxPeakS, 0x06)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_07:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x07) == 0)
		goto DEFECT_08;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x07);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x07);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZero, 0x07);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x07);
	if((short)_mm_extract_epi16(xxPeakS, 0x07)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_08:
	ix++;
	goto _return_XMM_COMPARE_8PIXEL;

PAIRING_VERT:
	iy = -1;

LOOP_PAIR_NUM:
	iy++;
	if(iy >= rrNoD)
		goto LOOP_PAIR_NUM_END;

	rax = nPair;
	if(rax >= MaxPair)
		goto END_FUNC;

	if(*(DefectT + iy) >= 3)
		goto LOOP_PAIR_NUM;

	rbx = iy;
	rcx = *(DefectY + iy);
	rax = rcx + iPitchHUp;
	rrImgAddr = *(DefectX + iy);

	if(rcx > pairBottom)
		goto PAIR_BOTTOM;

	if(rcx >= pairTop)
		goto PAIR_CENTER;

PAIR_TOP:
	rbx++;
	if(rbx >= rrNoD)
		goto PAIR_TOP_UNPAIR;

	rdx = *(DefectY + rbx);
	if(rdx < rax)
		goto PAIR_TOP;
	else if(rdx > rax)
		goto PAIR_TOP_UNPAIR;

	rcx = *(DefectX + rbx);
	if(rrImgAddr != rcx)
		goto PAIR_TOP;

	rrImgAddr = *(DefectT + rbx);
	if(rrImgAddr >= 3)
		goto LOOP_PAIR_NUM;

	rax = (__int64)*(DefectT + iy);
	if(rax == rrImgAddr)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rcx;
	ptDefect.y = (LONG)rdx;
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //
	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //	
	sSrc = (int)(*(pGraySrc + rbx));
	sRef = (int)(*(pGrayCmp + rbx));

	if(sSrc > sRef)
		sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

	*(pZoneTh + rbx) = sThres;

	sDiff = abs(sSrc - sRef);
	if(sDiff <= sThres)
		goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	*(DefectT + rbx) = 4;

	PairAddr->s_DefectPos = (DefectPos)1;
	PairAddr->s_DefectPair = (DefectPair)1;
	PairAddr->s_nDefectX = (int)rcx;
	PairAddr->s_nDefectY = (int)rdx;
	PairAddr->s_DefectType = (DefectType)rrImgAddr;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc);
	PairAddr->s_nGrayRef = (int)(sRef);
	PairAddr->s_nZone = (int)(*(pZoneID + rbx));
	PairAddr->s_nThresold = (int)(sThres);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_TOP_UNPAIR:
	if(bUseUnpair != 1)
		goto LOOP_PAIR_NUM;

	//rcx = (__int64)*(DefectValue + iy);
	//if(rcx <= ExTh)
	//	goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)(rrImgAddr);
	ptDefect.y = (LONG)(rax - iPitchHUp);
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	sSrc = (int)(*(pGraySrc + iy));
	sRef = (int)(*(pGrayCmp + iy));

	if(sSrc > sRef)
		sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

	*(pZoneTh + iy) = sThres;

	sDiff = abs(sSrc - sRef);
	if(sDiff <= sThres)
		goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	PairAddr->s_DefectPos = (DefectPos)1;
	PairAddr->s_DefectPair = (DefectPair)5;
	PairAddr->s_nDefectX = (int)ptDefect.x;
	PairAddr->s_nDefectY = (int)ptDefect.y;
	PairAddr->s_DefectType = (DefectType)*(DefectT + iy);
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + iy);
	PairAddr->s_nGraySrc = (int)(sSrc);
	PairAddr->s_nGrayRef = (int)(sRef);
	PairAddr->s_nZone = (int)(*(pZoneID + iy));
	PairAddr->s_nThresold = (int)(sThres);
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_CENTER:
	rbx++;
	if(rbx >= rrNoD)
		goto LOOP_PAIR_NUM;

	rdx = *(DefectY + rbx);
	if(rdx < rax)
		goto PAIR_CENTER;
	else if(rdx > rax)
		goto LOOP_PAIR_NUM;

	rcx = *(DefectX + rbx);
	if(rrImgAddr != rcx)
		goto PAIR_CENTER;

	rrImgAddr = *(DefectT + rbx);
	if(rrImgAddr >= 3)
		goto LOOP_PAIR_NUM;

	rax = (__int64)*(DefectT + iy);
	if(rax == rrImgAddr)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rcx;
	ptDefect.y = (LONG)rdx;
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
		sSrc = (int)(*(pGraySrc + rbx));
		sRef = (int)(*(pGrayCmp + rbx));

		if(sSrc > sRef)
			sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
		else
			sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

		*(pZoneTh + rbx) = sThres;

		sDiff = abs(sSrc - sRef);
		if(sDiff <= sThres)
			goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	*(DefectT + rbx) = 4;

	PairAddr->s_DefectPos = (DefectPos)2;
	PairAddr->s_DefectPair = (DefectPair)1;
	PairAddr->s_nDefectX = (int)rcx;
	PairAddr->s_nDefectY = (int)rdx;
	PairAddr->s_DefectType = (DefectType)rrImgAddr;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc);
	PairAddr->s_nGrayRef = (int)(sRef);
	PairAddr->s_nZone = (int)(*(pZoneID + rbx));
	PairAddr->s_nThresold = (int)(sThres);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_BOTTOM:
	if(bUseUnpair != 1)
		goto LOOP_PAIR_NUM;

	//rcx = (__int64)*(DefectValue + iy);
	//if(rcx <= ExTh)
	//	goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rrImgAddr;
	ptDefect.y = (LONG)rax;
	for(int i = 0; i < nDeadZoneCount; i++)
	{
		if(rtDeadZone[i].PtInRect(ptDefect) )
			goto LOOP_PAIR_NUM;
	}
	// ------------------------------------------ //

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	sSrc = (int)(*(pGrayCmp + iy));
	sRef = (int)(*(pGraySrc + iy));

	if(sSrc > sRef)
		sThres = (short)(sSrc * ThresholdSlide / 256 + Threshold);
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main);

	*(pZoneTh + iy) = sThres;

	sDiff = abs(sSrc - sRef);
	if(sDiff <= sThres)
		goto LOOP_PAIR_NUM;
	// ----------------------------------- //

	PairAddr->s_DefectPos = (DefectPos)3;
	PairAddr->s_DefectPair = (DefectPair)5;
	PairAddr->s_nDefectX = (int)rrImgAddr;
	PairAddr->s_nDefectY = (int)rax;
	rdx = (*(DefectT + iy) + 1) & 1;
	PairAddr->s_DefectType = (DefectType)rdx;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + iy);
	PairAddr->s_nGraySrc = (int)sSrc;
	PairAddr->s_nGrayRef = (int)sRef;
	PairAddr->s_nZone = (int)(*(pZoneID + iy));
	PairAddr->s_nThresold = (int)sThres;
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

LOOP_PAIR_NUM_END:
	rrNoD = 0;
	goto _return_PAIRING_VERT;

END_FUNC:
	ix += 0x10;
	*endLine = (int)ix;

	return (int)nPair;
}

int CSISAssem::ConvPixel_Intrin(BYTE *pImgBuff, __int64 Threshold, __int64 maxDefect, __int64 nDefect
								  , __int64 left, __int64 right, __int64 top, __int64 bottom
								  , int* DefectX, int* DefectY, short* DefectT, short* DefectValue
								  , int* pGraySrc, int* pGrayCmp, int* pZoneID, int* pZoneTh
								  , __int64 frameWidth, __int64 iPitch, __int64 fSPR, __int64 sSPR
								  , short* ZoneThTable, short* ZoneIdTable, __int64 ZoneTableWidth, void* pMatchTable
								  , CDefectPair *pDefectdPair, __int64 StartPair, __int64 MaxPair, __int64 PairSize
								  , __int64 dxPattern, __int64 dyPattern, __int64 yZoneByte,  __int64 bReverseFilter
								  , __int64 ThresholdSlide, __int64 ThresholdSuppress, __int64 bUseUnpair, int* endLine
								  )
{
	const __int64 SprShift = 4;
	const __int64 SprMulti = 16;

	__m128i	xxTemp;

	__m128i	xxZTh;
	__m128i	xxZId;

	__m128i	xxSrc1;
	__m128i xxSrc2;

	__m128i	xxCmp1;
	__m128i xxCmp2;
	__m128i xxCmp3;

	__m128i	xxPeakS;
	__m128i	xxPeakAbs;

	__m128i	xxfSpr;
	__m128i	xxsSpr;

	__m128i xxDStart;
	__m128i xxDSlide;

	__int64 rax, rbx, rcx, rdx;

	__int64 ImgAddrOri	= (__int64)pImgBuff;
	__int64 c8Pixel, rrNoD = 0, rrImgAddr, nPair = StartPair;
		
	__m128i xxSupp = _mm_set1_epi16((short)ThresholdSuppress);

	xxDStart = _mm_set1_epi16((short)Threshold);
	xxDSlide = _mm_set1_epi16((short)ThresholdSlide);

	xxZId = _mm_setzero_si128();

	rcx = iPitch;
	rcx &= 0x0f;
	xxsSpr = _mm_set1_epi16((short)rcx);

	rax = SprMulti;
	rax -= rcx;
	xxfSpr = _mm_set1_epi16((short)rax);

	iPitch = iPitch / SprMulti;
	__int64 iPitchHUp = iPitch;
	if(rcx >= SprMulti / 2)
		iPitchHUp++;

	right = (right - left - iPitch) / 16 * 16 + left;	

	__int64 pairRight = right - iPitch;
	__int64 pairLeft = left + iPitch;

	__int64 ix = left;
	__int64	iy = top;

	__m128i xxZero = _mm_setzero_si128();

	CDefectPair* PairAddr = pDefectdPair + nPair;

	//LOOP_1:
	goto LOOP_1_CHECK;

LOOP_1_TAIL:
	goto PAIRING;

_return_PAIRING:
	iy++;

LOOP_1_CHECK:
	if(iy >= bottom)
		goto END_FUNC;

	//LOOP_1_BODY:
	ix = left;
	rrImgAddr = ImgAddrOri + frameWidth * iy + ix;

	//LOOP_2:
	goto LOOP_2_CHECK;

LOOP_2_TAIL:
	rrImgAddr += 0x10;

LOOP_2_CHECK:
	if(ix >= right)
		goto LOOP_1_TAIL;

	//LOOP_2_BODY:
	goto XMM_MAKE;
_return_XMM_MAKE:
	goto XMM_COMPARE;
_return_XMM_COMPARE:
	goto LOOP_2_TAIL;


XMM_MAKE:
	xxSrc2 = _mm_loadu_si128((__m128i*) rrImgAddr);
	xxSrc1 = _mm_unpacklo_epi8(xxSrc2, xxZero);
	xxSrc2 = _mm_unpackhi_epi8(xxSrc2, xxZero);

	rax = rrImgAddr + iPitch;

	xxCmp2 = _mm_loadu_si128((__m128i*) rax);
	xxCmp1 = _mm_unpacklo_epi8(xxCmp2, xxZero);
	xxCmp2 = _mm_unpackhi_epi8(xxCmp2, xxZero);

	xxCmp1 = _mm_mullo_epi16(xxCmp1, xxfSpr);
	xxCmp2 = _mm_mullo_epi16(xxCmp2, xxfSpr);

	xxCmp3 = _mm_loadu_si128((__m128i*) (rax + 1));

	xxTemp = xxCmp3;
	xxTemp = _mm_unpacklo_epi8(xxTemp, xxZero);
	xxTemp = _mm_mullo_epi16(xxTemp, xxsSpr);
	xxCmp1 = _mm_add_epi16(xxCmp1, xxTemp);

	xxTemp = xxCmp3;
	xxTemp = _mm_unpackhi_epi8(xxTemp, xxZero);
	xxTemp = _mm_mullo_epi16(xxTemp, xxsSpr);
	xxCmp2 = _mm_add_epi16(xxCmp2, xxTemp);


	xxCmp1 = _mm_srli_epi16(xxCmp1, SprShift);
	xxCmp2 = _mm_srli_epi16(xxCmp2, SprShift);
	goto _return_XMM_MAKE;

XMM_COMPARE:
	c8Pixel = 0;
	goto _XMM_COMPARE_8PIXEL;

_return_XMM_COMPARE_8PIXEL:
	if(rrNoD >= maxDefect)
		goto PAIRING;

	ix++;

	c8Pixel++;
	if(c8Pixel == 2)
		goto _return_XMM_COMPARE;

	xxSrc1 = xxSrc2;
	xxCmp1 = xxCmp2;

_XMM_COMPARE_8PIXEL:
	xxSrc1 = _mm_min_epu16(xxSrc1, xxSupp);
	xxCmp1 = _mm_min_epu16(xxCmp1, xxSupp);

	xxZTh = _mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_max_epu16(xxSrc1, xxCmp1), xxDSlide), 8), xxDStart);
	xxPeakS = _mm_sub_epi16(xxSrc1, xxCmp1);
	xxPeakAbs = _mm_abs_epi16(xxPeakS);
	xxTemp = _mm_cmpgt_epi16(xxPeakAbs, xxZTh);
	if(_mm_movemask_epi8(xxTemp) != 0)
		goto DEFECT_00;

	ix += 7;
	goto _return_XMM_COMPARE_8PIXEL;

DEFECT_00:
	if(_mm_extract_epi16(xxTemp, 0x00) == 0)
		goto DEFECT_01;
	
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x00);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x00);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x00);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x00);
	if((short)_mm_extract_epi16(xxPeakS, 0x00)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_01:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x01) == 0)
		goto DEFECT_02;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x01);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x01);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x01);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x01);
	if((short)_mm_extract_epi16(xxPeakS, 0x01)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_02:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x02) == 0)
		goto DEFECT_03;
		
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x02);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x02);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x02);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x02);
	if((short)_mm_extract_epi16(xxPeakS, 0x02)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_03:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x03) == 0)
		goto DEFECT_04;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x03);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x03);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x03);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x03);
	if((short)_mm_extract_epi16(xxPeakS, 0x03)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_04:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x04) == 0)
		goto DEFECT_05;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x04);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x04);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x04);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x04);
	if((short)_mm_extract_epi16(xxPeakS, 0x04)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_05:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x05) == 0)
		goto DEFECT_06;
		
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x05);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x05);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x05);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x05);
	if((short)_mm_extract_epi16(xxPeakS, 0x05)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_06:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x06) == 0)
		goto DEFECT_07;
		
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x06);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x06);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x06);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x06);
	if((short)_mm_extract_epi16(xxPeakS, 0x06)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

DEFECT_07:
	ix++;
	if(_mm_extract_epi16(xxTemp, 0x07) == 0)
		goto _return_XMM_COMPARE_8PIXEL;

	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;

	*(pGraySrc + rrNoD) = _mm_extract_epi16(xxSrc1, 0x07);
	*(pGrayCmp + rrNoD) = _mm_extract_epi16(xxCmp1, 0x07);
	*(pZoneID + rrNoD) = _mm_extract_epi16(xxZId, 0x07);
	*(pZoneTh + rrNoD) = _mm_extract_epi16(xxZTh, 0x07);
	if((short)_mm_extract_epi16(xxPeakS, 0x07)  > 0)
		*(DefectT + rrNoD) = 1;
	else
		*(DefectT + rrNoD) = 0;

	rrNoD++;

	goto _return_XMM_COMPARE_8PIXEL;

PAIRING:
	ix = -1;

LOOP_PAIR_NUM:
	ix++;
	if(ix >= rrNoD)
		goto LOOP_PAIR_NUM_END;

	rax = nPair;
	if(rax >= MaxPair)
		goto END_FUNC;

	if(*(DefectT + ix) >= 3)
		goto LOOP_PAIR_NUM;

	rbx = ix;
	rdx = *(DefectX + ix);
	rcx = rdx;
	rcx = rcx + iPitchHUp;

	if(rdx < pairLeft)
		goto PAIR_LEFT;

	if(rdx <= pairRight)
		goto PAIR_CENTER;

	//PAIR_RIGHT:
	if(bUseUnpair == 0)
		goto LOOP_PAIR_NUM;

	//rcx = (__int64)*(DefectValue + iy);
	//if(rcx <= ExTh)
	//	goto LOOP_PAIR_NUM;

	PairAddr->s_DefectPos = (DefectPos)3;
	PairAddr->s_DefectPair = (DefectPair)5;
	PairAddr->s_nDefectX = (int)rcx;
	PairAddr->s_nDefectY = (int)(*(DefectY + ix));
	rcx = (*(DefectT + ix) + 1) & 1;
	PairAddr->s_DefectType = (DefectType)rcx;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + iy);
	PairAddr->s_nGraySrc = (int)(*(pGrayCmp + ix));
	PairAddr->s_nGrayRef = (int)(*(pGraySrc + ix));
	PairAddr->s_nZone = (int)(*(pZoneID + ix));
	PairAddr->s_nThresold = (int)(*(pZoneTh + ix));
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_LEFT:
	rbx++;
	if(rbx >= rrNoD)
		goto PAIR_LEFT_UNPAIR;

	rax = *(DefectX + rbx);
	if(rax < rcx)
		goto PAIR_LEFT;
	else if(rax > rcx)
		goto PAIR_LEFT_UNPAIR;

	rdx = *(DefectT + rbx);
	if(rdx >= 3)
		goto PAIR_LEFT;

	rax = (__int64)*(DefectT + ix);
	if(rax == rdx)
		goto PAIR_LEFT;

	*(DefectT + rbx) = 4;

	PairAddr->s_DefectPos = (DefectPos)1;
	PairAddr->s_DefectPair = (DefectPair)1;
	PairAddr->s_nDefectX = (int)rcx;
	PairAddr->s_nDefectY = (int)(*(DefectY + rbx));
	PairAddr->s_DefectType = (DefectType)rdx;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(*(pGraySrc + rbx));
	PairAddr->s_nGrayRef = (int)(*(pGrayCmp + rbx));
	PairAddr->s_nZone = (int)(*(pZoneID + rbx));
	PairAddr->s_nThresold = (int)(*(pZoneTh + rbx));

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_LEFT_UNPAIR:
	if(bUseUnpair == 0)
		goto LOOP_PAIR_NUM;

	//rcx = (__int64)*(DefectValue + iy);
	//if(rcx <= ExTh)
	//	goto LOOP_PAIR_NUM;

	PairAddr->s_DefectPos = (DefectPos)1;
	PairAddr->s_DefectPair = (DefectPair)5;
	PairAddr->s_nDefectX = (int)(*(DefectX + ix));
	PairAddr->s_nDefectY = (int)(*(DefectY + ix));
	PairAddr->s_DefectType = (DefectType)*(DefectT + ix);
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + iy);
	PairAddr->s_nGraySrc = (int)(*(pGraySrc + ix));
	PairAddr->s_nGrayRef = (int)(*(pGrayCmp + ix));
	PairAddr->s_nZone = (int)(*(pZoneID + ix));
	PairAddr->s_nThresold = (int)(*(pZoneTh + ix));
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_CENTER:
	rbx++;
	if(rbx >= rrNoD)
		goto LOOP_PAIR_NUM;

	rax = *(DefectX + rbx);
	if(rax < rcx)
		goto PAIR_CENTER;
	else if(rax > rcx)
		goto PAIR_CENTER_UNPAIR;

	rdx = *(DefectT + rbx);
	if(rdx >= 3)
		goto PAIR_CENTER;

	rax = (__int64)*(DefectT + ix);
	if(rax == rdx)
		goto PAIR_CENTER;

	*(DefectT + rbx) = 4;

	PairAddr->s_DefectPos = (DefectPos)2;
	PairAddr->s_DefectPair = (DefectPair)1;
	PairAddr->s_nDefectX = (int)rcx;
	PairAddr->s_nDefectY = (int)(*(DefectY + rbx));
	PairAddr->s_DefectType = (DefectType)rdx;
	//PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(*(pGraySrc + rbx));
	PairAddr->s_nGrayRef = (int)(*(pGrayCmp + rbx));
	PairAddr->s_nZone = (int)(*(pZoneID + rbx));
	PairAddr->s_nThresold = (int)(*(pZoneTh + rbx));

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;
PAIR_CENTER_UNPAIR:
	goto LOOP_PAIR_NUM;

LOOP_PAIR_NUM_END:
	rrNoD = 0;
	goto _return_PAIRING;

END_FUNC:
	iy ++;
	*endLine = (int)iy;

	return (int)nPair;
}

int CSISAssem::ConvPixel_Intrin_Twice(BYTE *pImgBuff, __int64 Threshold, __int64 maxDefect, __int64 nDefect
								  , __int64 left, __int64 right, __int64 top, __int64 bottom
								  , int* DefectX, int* DefectY, short* DefectT, short* DefectValue
								  , int* pGraySrc, int* pGrayCmp, int* pZoneID, int* pZoneTh
								  , __int64 frameWidth, __int64 iPitch, __int64 fSPR, __int64 sSPR
								  , short* ZoneThTable, short* ZoneIdTable, __int64 ZoneTableWidth, void* pMatchTable
								  , CDefectPair *pDefectdPair, __int64 StartPair, __int64 MaxPair, __int64 PairSize
								  , __int64 dxPattern, __int64 dyPattern, __int64 yZoneByte,  __int64 bReverseFilter
								  , __int64 ThresholdSlide, __int64 ThresholdSuppress, __int64 bUseUnpair, int* endLine
								  , __int64 iPitch2, __int64 Threshold2, __int64 ThresholdSlide2)
{
	//°Å´ë °áÇÔÀ» °ËÃâÇϱâ À§ÇØ ÀåÁÖ±â·Î Çѹø´õ ºñ±³ÇÑ´Ù.
	__int64 nPair1 = ConvPixel_Intrin(pImgBuff, Threshold, maxDefect, nDefect
		, left, right, top, bottom
		, DefectX, DefectY, DefectT, DefectValue
		, pGraySrc, pGrayCmp, pZoneID, pZoneTh
		, frameWidth, iPitch, fSPR, sSPR
		, ZoneThTable, ZoneIdTable, ZoneTableWidth, pMatchTable
		, pDefectdPair, StartPair, MaxPair, PairSize
		, dxPattern, dyPattern, yZoneByte,  bReverseFilter
		, ThresholdSlide, ThresholdSuppress, bUseUnpair, endLine
		);

	__int64 nPair2 = ConvPixel_Intrin(pImgBuff, Threshold2, maxDefect, nDefect
		, left, right, top, bottom
		, DefectX, DefectY, DefectT, DefectValue
		, pGraySrc, pGrayCmp, pZoneID, pZoneTh
		, frameWidth, iPitch2, fSPR, sSPR
		, ZoneThTable, ZoneIdTable, ZoneTableWidth, pMatchTable
		, pDefectdPair, nPair1, MaxPair, PairSize
		, dxPattern, dyPattern, yZoneByte,  bReverseFilter
		, ThresholdSlide2, ThresholdSuppress, bUseUnpair, endLine
		);

	//Æä¾î¸µµÈ °áÇÔÁß °°Àº ÁÂÇ¥ÀÇ Pair ´Â Áö¿î´Ù.
	__int64 iPair1, iPair2;
	CDefectPair *pPair1, *pPair2;

	for(iPair1= StartPair; iPair1 < nPair1; iPair1++)
	{
		pPair1= pDefectdPair+ iPair1;
		if(pPair1->s_DefectType == DEFTYPE_DELETE)
			continue;

		for(iPair2= nPair1; iPair2 < nPair2; iPair2++)
		{
			pPair2= pDefectdPair+ iPair2;
			if(pPair2->s_DefectType == DEFTYPE_DELETE)
				continue;

			if(pPair1->s_nDefectY < pPair2->s_nDefectY - 1)
				break;

			if(pPair1->s_nDefectX < pPair2->s_nDefectX - 1)
				continue;

			if(abs(pPair1->s_nDefectX - pPair2->s_nDefectX) < 2
				&& abs(pPair1->s_nDefectY - pPair2->s_nDefectY) < 2)
				pPair2->s_DefectType = DEFTYPE_DELETE;
		}
	}

	int	oldBlob = m_nBlob;
	m_nPair = (int)nPair1;
	CBlobStorage::BlobDefect();
	CBlobStorage::MergeDivision(oldBlob, m_Param.m_nDefectMergePixel);
	CBlobStorage::ShadowFilter(oldBlob, m_Param.GetRealConvPitch());

	int nThres = m_Param.s_nThreshold;
	int nDThSlide = m_Param.s_nDThSlide;
	double dPitchConvReal = m_Param.s_dPitchConvReal;
	double dPitchScanReal = m_Param.s_dPitchScanReal;

	m_Param.s_nThreshold = m_Param.s_n2ndThres;
	m_Param.s_nDThSlide = m_Param.s_n2ndThresSlide;
	m_Param.s_dPitchConvReal = m_Param.s_d2ndPitch;
	m_Param.s_dPitchScanReal = m_Param.s_d2ndPitch;

	oldBlob = m_nBlob;
	m_nPair = (int)nPair2;
	CBlobStorage::BlobDefect();
	CBlobStorage::MergeDivision(oldBlob, m_Param.m_nDefectMergePixel);
	CBlobStorage::ShadowFilter(oldBlob, m_Param.GetRealConvPitch());

	m_Param.s_nThreshold = nThres;
	m_Param.s_nDThSlide = nDThSlide;
	m_Param.s_dPitchConvReal = dPitchConvReal;
	m_Param.s_dPitchScanReal = dPitchScanReal;

	return (int)nPair2;
}

int CSISAssem::ConvPixelVertS_Intrin_Twice(BYTE *pImgBuff, __int64 Threshold, __int64 maxDefect, __int64 nDefect
											, __int64 left, __int64 right, __int64 top, __int64 bottom
											, int* DefectX, int* DefectY, short* DefectT, short* DefectValue
											, int* pGraySrc, int* pGrayCmp, int* pZoneID, int* pZoneTh
											, __int64 frameWidth, __int64 iPitch16, __int64 fSPR, __int64 sSPR
											, short* ZoneThTable, short* ZoneIdTable, __int64 ZoneTableWidth, void* pMatchTable
											, CDefectPair *pDefectdPair, __int64 StartPair, __int64 MaxPair, __int64 PairSize
											, __int64 dxPattern, __int64 dyPattern, __int64 yZoneByte, __int64 bReverseFilter
											, __int64 ThresholdSlide, __int64 ThresholdSuppress, __int64 bUseUnpair, short* pConBuff
											, int* endLine, __int64 ExTh, __int64 iPitch2, __int64 Threshold2
											, __int64 ThresholdSlide2
											)
{	
	//°Å´ë °áÇÔÀ» °ËÃâÇϱâ À§ÇØ ÀåÁÖ±â·Î Çѹø´õ ºñ±³ÇÑ´Ù.
	__int64 nPair1 = ConvPixelVertS_Intrin(pImgBuff, Threshold, maxDefect, nDefect
		, left, right, top, bottom
		, DefectX, DefectY, DefectT, DefectValue
		, pGraySrc, pGrayCmp, pZoneID, pZoneTh
		, frameWidth, iPitch16, fSPR, sSPR
		, ZoneThTable, ZoneIdTable, ZoneTableWidth, pMatchTable
		, pDefectdPair, StartPair, MaxPair, PairSize
		, dxPattern, dyPattern, yZoneByte, bReverseFilter
		, ThresholdSlide, ThresholdSuppress, bUseUnpair, pConBuff
		, endLine, ExTh);	

	__int64 nPair2 = ConvPixelVertS_Intrin(pImgBuff, Threshold2, maxDefect, nDefect
		, left, right, top, bottom
		, DefectX, DefectY, DefectT, DefectValue
		, pGraySrc, pGrayCmp, pZoneID, pZoneTh
		, frameWidth, iPitch2, fSPR, sSPR
		, ZoneThTable, ZoneIdTable, ZoneTableWidth, pMatchTable
		, pDefectdPair, nPair1, MaxPair, PairSize
		, dxPattern, dyPattern, yZoneByte, bReverseFilter
		, ThresholdSlide2, ThresholdSuppress, bUseUnpair, pConBuff
		, endLine, ExTh);

	//Æä¾î¸µµÈ °áÇÔÁß °°Àº ÁÂÇ¥ÀÇ Pair ´Â Áö¿î´Ù.
	__int64 iPair1, iPair2;
	CDefectPair *pPair1, *pPair2;

	for(iPair1= StartPair; iPair1 < nPair1; iPair1++)
	{
		pPair1= pDefectdPair+ iPair1;
		if(pPair1->s_DefectType == DEFTYPE_DELETE)
			continue;

		for(iPair2= nPair1; iPair2 < nPair2; iPair2++)
		{
			pPair2= pDefectdPair+ iPair2;
			if(pPair2->s_DefectType == DEFTYPE_DELETE)
				continue;

			if(pPair1->s_nDefectX < pPair2->s_nDefectX - 1)
				break;

			if(pPair1->s_nDefectY < pPair2->s_nDefectY - 1)
				continue;

			if(abs(pPair1->s_nDefectX - pPair2->s_nDefectX) < 2
				&& abs(pPair1->s_nDefectY - pPair2->s_nDefectY) < 2)
				pPair2->s_DefectType = DEFTYPE_DELETE;
		}
	}

	int	oldBlob = m_nBlob;
	m_nPair = (int)nPair1;
	CBlobStorage::BlobDefectVert();
	CBlobStorage::MergeDivision(oldBlob, m_Param.m_nDefectMergePixel);
	CBlobStorage::ShadowFilterVert(oldBlob, m_Param.GetRealScanPitch());

	int nThres = m_Param.s_nThreshold;
	int nDThSlide = m_Param.s_nDThSlide;
	double dPitchConvReal = m_Param.s_dPitchConvReal;
	double dPitchScanReal = m_Param.s_dPitchScanReal;

	m_Param.s_nThreshold = m_Param.s_n2ndThres;
	m_Param.s_nDThSlide = m_Param.s_n2ndThresSlide;
	m_Param.s_dPitchConvReal = m_Param.s_d2ndPitch;
	m_Param.s_dPitchScanReal = m_Param.s_d2ndPitch;

	oldBlob = m_nBlob;
	m_nPair = (int)nPair2;
	CBlobStorage::BlobDefectVert();
	CBlobStorage::MergeDivision(oldBlob, m_Param.m_nDefectMergePixel);
	CBlobStorage::ShadowFilterVert(oldBlob, m_Param.GetRealScanPitch());

	m_Param.s_nThreshold = nThres;
	m_Param.s_nDThSlide = nDThSlide;
	m_Param.s_dPitchConvReal = dPitchConvReal;
	m_Param.s_dPitchScanReal = dPitchScanReal;

	return (int)nPair2;
}