#include "stdafx.h"

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

#ifdef _DEBUG
#define new DEBUG_NEW
#endif

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

//Assem->Intrinsic Convert(2by2 Vertical DTh)
int CSISAssem::ConvFloatVert_BOE_Intrin(BYTE* pImgBuff, short* pConBuff, __int64 maxDefect, __int64 nDefect
										  , __int64 left, __int64 right, __int64 top, __int64 bottom
										  , int* DefectX, int* DefectY, short* DefectT, short* DefectValue
										  , CDefectPair *pDefectdPair, __int64 nPair, __int64 MaxPair, __int64 PairSize
										  , __int64 frameWidth, __int64 iPitch16, __int64 Threshold, __int64 Suppress
										  , int* pGraySrc, int* pGrayCmp, int* pZoneID, int* pZoneTh
										  , __int64 i2Pitch16, __int64 bReverseFilter, __int64 ThresholdSlide, int* endLine
										  )
{
	const __int64 ConvNSprShift = 6;
	const __int64 SprShift = 4;
	const __int64 ConvShift = 2;
	const __int64 SprMulti = 16;
	const __int64 ConvMulti = 4;

	__int64 ix = left;
	__int64	iy = top;

	__int64	rrBW = frameWidth;
	__int64 rrNoD = nDefect;
	CDefectPair* PairAddr = pDefectdPair + nPair;

	__int64 F_Y2Insp;

	__int64 ImgAddr	= (__int64)pImgBuff;
	__int64	ConAddr	= __int64(pConBuff + 0x10);
	__int64	maxPixelD = maxDefect;

	__m128i	xxSrc;
	__m128i	xxCmp;

	__m128i	xxPeak;
	__m128i	xxPeakEx;
	__m128i xxSrc1;
	__m128i xxSrc2;

	__m128i xxCmp1;
	__int64 rax, rbx, rcx, rdx;

	__m128i xxZero = _mm_setzero_si128();
	__m128i xxThres = _mm_set1_epi16((short)Threshold * SprMulti);
	__m128i	xxDStart = _mm_set1_epi16((short)Threshold / ConvMulti);
	__m128i	xxDSlide = _mm_set1_epi16((short)ThresholdSlide);
	__m128i xxSupp = _mm_set1_epi16((short)Suppress * SprMulti * ConvMulti);

	// ---- 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;
	}
	// ----------------------------------- //

	__int64 iPitch = iPitch16 / SprMulti;
	__int64 iPitchHUp = iPitch;
	__int64 iSpr = iPitch16 & 0xF;
	if(iSpr >= SprMulti / 2)
		iPitchHUp++;

	__m128i	xxsSPR = _mm_set1_epi16((short)iSpr);
	__m128i	xxfSPR = _mm_set1_epi16((short)SprMulti - (short)iSpr);

	__int64 i2Pitch = i2Pitch16 / SprMulti;
	__int64 i2Spr = i2Pitch16 & 0xF;

	bottom--;
	__int64 ConBottom = bottom;

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

	__int64 ImgAddrLeft = ImgAddr + top * rrBW + left;
	__int64 ConPitch = iPitch * 0x20;
	__int64 ConPitch2 = i2Pitch * 0x20;
	__int64 pairTop = top + iPitch;
	bottom = bottom - iPitch - 1;
	__int64 pairBottom = bottom - 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 rrImgAddr = ImgAddrLeft;
	__int64 rrConAddr = ConAddr;

	goto LOOP_X_CHECK;

LOOP_X_TAIL:
	goto PAIRING_VERT;

PAIRING_VERT_RETURN:
	ImgAddrLeft += 0x10;
	rrImgAddr = ImgAddrLeft;
	rrConAddr = ConAddr;
	iy = top;

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

	//LOOP_X_BODY:
	goto MAKE_CONBUFF_BODY_TAIL_CHECK;

MAKE_CONBUFF_RETURN:
	iy = top;
	rrConAddr = ConAddr;
	F_Y2Insp = 0;
	goto LOOP_Y_CHECK;

LOOP_Y_TAIL:
	iy++;
	rrImgAddr += rrBW;
	rrConAddr += 0x20;

LOOP_Y_CHECK:
	if(rrNoD >= maxPixelD)
		goto PAIRING_VERT;
	if(iy < bottom)
		goto LOOP_Y_BODY;
	ix += 8;
	if(F_Y2Insp >= 1)
		goto LOOP_X_TAIL;

	F_Y2Insp = 1;
	iy = top;
	rrConAddr = ConAddr + 0x10;
	goto LOOP_Y_BODY;

MAKE_CONBUFF_BODY_TAIL_CHECK:
	xxSrc1 = _mm_loadu_si128((__m128i*) rrImgAddr);
	xxCmp1 = _mm_loadu_si128((__m128i*)(rrImgAddr + 1));

	_mm_storeu_si128((__m128i*)(rrConAddr), _mm_add_epi16(_mm_unpacklo_epi8(xxSrc1, xxZero), _mm_unpacklo_epi8(xxCmp1, xxZero)));
	_mm_storeu_si128((__m128i*)(rrConAddr + 0x10), _mm_add_epi16(_mm_unpackhi_epi8(xxSrc1, xxZero), _mm_unpackhi_epi8(xxCmp1, xxZero)));

	rrImgAddr += rrBW;
	rrConAddr += 0x20;
	iy++;
	if(iy <= ConBottom)
		goto MAKE_CONBUFF_BODY_TAIL_CHECK;
	goto MAKE_CONBUFF_RETURN;

LOOP_Y_BODY:
	xxSrc1 = _mm_loadu_si128((__m128i*) rrConAddr);
	xxSrc2 = _mm_loadu_si128((__m128i*) (rrConAddr + 0x20));
	xxSrc = _mm_slli_epi16(_mm_add_epi16(xxSrc1, xxSrc2), SprShift);
	xxSrc = _mm_subs_epu16(xxSrc, _mm_subs_epu16(xxSrc, xxSupp));

	rax = rrConAddr + ConPitch;
	xxCmp = _mm_add_epi16(_mm_add_epi16(_mm_mullo_epi16(_mm_loadu_si128((__m128i*)(rax)), xxfSPR), _mm_slli_epi16(_mm_loadu_si128((__m128i*)(rax + 0x20)), SprShift))
		, _mm_mullo_epi16(_mm_loadu_si128((__m128i*)(rax + 0x40)), xxsSPR));

	xxCmp = _mm_subs_epu16(xxCmp, _mm_subs_epu16(xxCmp, xxSupp));

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
	//¿ø·¡ÄÚµå: xxThres = _mm_slli_epi16(_mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_srli_epi16(
	//						_mm_max_epi16(xxSrc, xxCmp), ConvNSprShift), xxDSlide), 8), xxDStart), ConvNSprShift);
	xxZThW = _mm_slli_epi16(_mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_srli_epi16(
		_mm_max_epi16(xxSrc, xxCmp), ConvNSprShift), xxDSlide), 8), xxDStart), ConvNSprShift);
	xxZThB = _mm_slli_epi16(_mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_srli_epi16(
		_mm_max_epi16(xxSrc, xxCmp), ConvNSprShift), xxDSlideB), 8), xxDStartB), ConvNSprShift);

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

	//INSPECT_BLACK:
	xxPeakEx = _mm_subs_epu16(xxCmp, xxSrc);
	xxPeak = _mm_cmpgt_epi16(xxPeakEx, xxThres);
	if(_mm_movemask_epi8(xxPeak) == 0)
		goto INSPECT_WHITE;
	rbx = 0;
	goto DEFECT_00;

INSPECT_WHITE:
	xxPeakEx = _mm_subs_epu16(xxSrc, xxCmp);
	xxPeak = xxPeakEx;
	xxPeak = _mm_cmpgt_epi16(xxPeak, xxThres);
	if(_mm_movemask_epi8(xxPeak) == 0)
		goto LOOP_Y_TAIL;
	rbx = 1;
	goto DEFECT_00;

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

	rax = _mm_extract_epi16(xxPeakEx, 0x00);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x00);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x00);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x00);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakEx, 0x01);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x01);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x01);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x01);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakEx, 0x02);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x02);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x02);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x02);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakEx, 0x03);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x03);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x03);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x03);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakEx, 0x04);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x04);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x04);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x04);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakEx, 0x05);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x05);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x05);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x05);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakEx, 0x06);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x06);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x06);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x06);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

DEFECT_07:
	ix++;
	if(_mm_extract_epi16(xxPeak, 0x07) == 0)
		goto DEFECT_END;

	rax = _mm_extract_epi16(xxPeakEx, 0x07);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x07);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x07);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x07);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

DEFECT_END:
	ix -= 7;
	if(rbx == 0)
		goto INSPECT_WHITE;
	goto LOOP_Y_TAIL;

PAIRING_VERT:
	iy = -1;

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

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

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

	rbx = iy;
	rcx = *(DefectY + iy);
	rax = rcx + iPitchHUp;
	rrBW = *(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(rrBW != rcx)
		goto PAIR_TOP;

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

	rax = (__int64)*(DefectT + iy);
	if(rax == rrConAddr)
		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 / 4) * 64;
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

	*(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)rrConAddr;
	PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_TOP_UNPAIR:
	if(m_Param.m_bUseUnpair == FALSE)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)(rrBW);
	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 / 4) * 64;
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

	*(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 / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);

	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(rrBW != rcx)
		goto PAIR_CENTER;

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

	rax = (__int64)*(DefectT + iy);
	if(rax == rrConAddr)
		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 / 4) * 64;
		else
			sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

		*(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)rrConAddr;
	PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_BOTTOM:
	if(m_Param.m_bUseUnpair == FALSE)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rrBW;
	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 / 4) * 64;
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

	*(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)rrBW;
	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 / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

LOOP_PAIR_NUM_END:
	rrNoD = 0;
	rrBW = frameWidth;
	goto PAIRING_VERT_RETURN;

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

	return (int)nPair;
}


//Assem->Intrinsic Convert(2by2 Vertical DTh DPC)
int CSISAssem::ConvFloatVTD_BOE_Intrin(BYTE* pImgBuff, short* pConBuff, __int64 maxDefect, __int64 nDefect
										 , __int64 left, __int64 right, __int64 top, __int64 bottom
										 , int* DefectX, int* DefectY, short* DefectT, short* DefectValue
										 , CDefectPair *pDefectdPair, __int64 nPair, __int64 MaxPair, __int64 PairSize
										 , __int64 frameWidth, __int64 iPitch16, __int64 Threshold, __int64 Suppress
										 , short*  i2Pitch16, __int64 tilt16, int* PitchBuff, __int64 ThresholdSlide
										 , int* pGraySrc, int* pGrayCmp, int* pZoneID, int* pZoneTh
										 , int* endLine
										 )
{
	const __int64 ConvNSprShift = 6;
	const __int64 SprShift = 4;
	const __int64 ConvShift = 2;
	const __int64 SprMulti = 16;
	const __int64 ConvMulti = 4;

	__int64 ix = left;
	__int64	iy = top;

	__int64 rrNoD = nDefect;
	CDefectPair* PairAddr = pDefectdPair + nPair;

	__int64 F_Y2Insp;

	__int64 ImgAddr	= (__int64)pImgBuff;
	__int64	ConAddr	= __int64(pConBuff + 0x10);
	__int64	ConAddr2	= __int64(i2Pitch16 + 0x10);
	__int64	maxPixelD = maxDefect;
	__int64	iPitchLast = 0;

	__m128i	xxTemp;
	__m128i	xxSrc;
	__m128i	xxCmp;

	__m128i	xxPeak;
	__m128i	xxPeakAbs;
	__m128i xxSrc1;
	__m128i xxCmp1;

	__m128i xxfSprT;
	__m128i xxsSprT;

	__int64 rax, rbx, rcx, rdx;

	__m128i xxZero = _mm_setzero_si128();
	__m128i xxThres = _mm_set1_epi16((short)Threshold * SprMulti);
	__m128i	xxDStart = _mm_set1_epi16((short)Threshold / ConvMulti);
	__m128i	xxDSlide = _mm_set1_epi16((short)ThresholdSlide);
	__m128i xxSupp = _mm_set1_epi16((short)Suppress * SprMulti * ConvMulti);

	// ---- 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 = tilt16;
	rcx &= 0x0f;

	xxsSprT = _mm_set1_epi16((short)rcx);

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

	tilt16 /= 16;

	__int64 iPitch = iPitch16 / SprMulti;
	__int64 iPitchHUp = iPitch;
	__int64 iSpr = iPitch16 & 0xF;
	if(iSpr >= SprMulti / 2)
		iPitchHUp++;

	__m128i	xxsSPR = _mm_set1_epi16((short)iSpr);
	__m128i	xxfSPR = _mm_set1_epi16((short)SprMulti - (short)iSpr);

	bottom--;
	__int64 ConBottom = bottom;

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

	__int64 ImgAddrLeft = ImgAddr + top * frameWidth + left;
	__int64 ConPitch = iPitch * 0x20;

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

	__int64 rrImgAddr = ImgAddrLeft;
	__int64 rrConAddr = ConAddr;
	__int64 rrConAddr2 = ConAddr2;

	goto LOOP_X_CHECK;

LOOP_X_TAIL:
	goto PAIRING_VERT;

PAIRING_VERT_RETURN:
	ImgAddrLeft += 0x10;
	rrImgAddr = ImgAddrLeft;
	rrConAddr = ConAddr;
	rrConAddr2 = ConAddr2;
	iy = top;

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

//LOOP_X_BODY:
	iy = top;
	rrImgAddr = ImgAddrLeft;
	rrConAddr = ConAddr;
	rrConAddr2 = ConAddr2;
	goto MAKE_CONBUFF_BODY_TAIL_CHECK;

MAKE_CONBUFF_RETURN:
	iy = top;
	rrImgAddr = ImgAddrLeft;
	rrConAddr = ConAddr;
	rrConAddr2 = ConAddr2;
	goto MAKE_CONBUFF2_HEAD;

MAKE_CONBUFF2_RETURN:
	iy = top;
	rrImgAddr = ImgAddrLeft;
	rrConAddr = ConAddr;
	rrConAddr2 = ConAddr2;
	F_Y2Insp = 0;
	goto LOOP_Y_CHECK;

DPC_READ:
	if(PitchBuff == NULL)
		goto DPC_READ_RETURN;

	rax = *(PitchBuff + iy / 16);
	if(rax == iPitchLast)
		goto DPC_READ_RETURN;

	iPitchLast = rax;
	rbx = rax;
	rbx &= 0x0f;
	rax /= 16;
	rax *= 32;
	ConPitch = rax;
	xxsSPR = _mm_set1_epi16((short)rbx);

	rdx = SprMulti;
	rdx -= rbx;
	xxfSPR = _mm_set1_epi16((short)rdx);
	goto DPC_READ_RETURN;

LOOP_Y_TAIL:
	iy++;
	rrImgAddr += frameWidth;
	rrConAddr += 0x20;
	rrConAddr2 += 0x20;

LOOP_Y_CHECK:
	if(rrNoD >= maxPixelD)
		goto PAIRING_VERT;
	if(iy < bottom)
		goto LOOP_Y_BODY;
	ix += 8;
	if(F_Y2Insp >= 1)
		goto LOOP_X_TAIL;

	F_Y2Insp = 1;
	iy = top;
	rrConAddr = ConAddr;
	rrConAddr += 0x10;
	rrConAddr2 = ConAddr2;
	rrConAddr2 += 0x10;
	goto LOOP_Y_BODY;

MAKE_CONBUFF_BODY_TAIL_CHECK:
	xxSrc = _mm_loadu_si128((__m128i*) rrImgAddr);
	xxSrc1 = xxSrc;
	xxSrc = _mm_unpacklo_epi8(xxSrc, xxZero);
	xxSrc1 = _mm_unpackhi_epi8(xxSrc1, xxZero);

	xxCmp = _mm_loadu_si128((__m128i*)(rrImgAddr + 1));
	xxCmp1 = xxCmp;
	xxCmp = _mm_unpacklo_epi8(xxCmp, xxZero);
	xxCmp1 = _mm_unpackhi_epi8(xxCmp1, xxZero);

	xxSrc = _mm_add_epi16(xxSrc, xxCmp);
	xxSrc1 = _mm_add_epi16(xxSrc1, xxCmp1);

	_mm_storeu_si128((__m128i*)(rrConAddr), xxSrc);
	_mm_storeu_si128((__m128i*)(rrConAddr + 0x10), xxSrc1);

	rrImgAddr += frameWidth;
	rrConAddr += 0x20;
	iy++;
	if(iy <= ConBottom)
		goto MAKE_CONBUFF_BODY_TAIL_CHECK;
	goto MAKE_CONBUFF_RETURN;

MAKE_CONBUFF2_HEAD:
	rax = rrImgAddr;
	rax += tilt16;
	rrConAddr2 = ConAddr2;

MAKE_CONBUFF2_BODY:
	xxSrc = _mm_loadu_si128((__m128i*) rax);
	xxSrc1 = xxSrc;
	xxSrc = _mm_unpacklo_epi8(xxSrc, xxZero);
	xxSrc1 = _mm_unpackhi_epi8(xxSrc1, xxZero);
	xxSrc = _mm_mullo_epi16(xxSrc, xxfSprT);
	xxSrc1 = _mm_mullo_epi16(xxSrc1, xxfSprT);

	xxCmp = _mm_loadu_si128((__m128i*)(rax + 2));
	xxCmp1 = xxCmp;
	xxCmp = _mm_unpacklo_epi8(xxCmp, xxZero);
	xxCmp1 = _mm_unpackhi_epi8(xxCmp1, xxZero);
	xxCmp = _mm_mullo_epi16(xxCmp, xxsSprT);
	xxCmp1 = _mm_mullo_epi16(xxCmp1, xxsSprT);

	xxSrc = _mm_add_epi16(xxSrc, xxCmp);
	xxSrc1 = _mm_add_epi16(xxSrc1, xxCmp1);

	xxSrc = _mm_srli_epi16(xxSrc, SprShift);
	xxSrc1 = _mm_srli_epi16(xxSrc1, SprShift);

	xxCmp = _mm_loadu_si128((__m128i*)(rax + 1));
	xxCmp1 = xxCmp;
	xxCmp = _mm_unpacklo_epi8(xxCmp, xxZero);
	xxCmp1 = _mm_unpackhi_epi8(xxCmp1, xxZero);

	xxSrc = _mm_add_epi16(xxSrc, xxCmp);
	xxSrc1 = _mm_add_epi16(xxSrc1, xxCmp1);

	_mm_storeu_si128((__m128i*)(rrConAddr2), xxSrc);
	_mm_storeu_si128((__m128i*)(rrConAddr2 + 0x10), xxSrc1);

	rax += frameWidth;
	rrConAddr2 += 0x20;

	iy++;
	if(iy <= ConBottom)
		goto MAKE_CONBUFF2_BODY;
	goto MAKE_CONBUFF2_RETURN;

LOOP_Y_BODY:
	goto DPC_READ;

DPC_READ_RETURN:
	xxSrc1 = _mm_loadu_si128((__m128i*) rrConAddr);
	xxTemp = _mm_loadu_si128((__m128i*) (rrConAddr + 0x20));
	xxSrc = xxSrc1;
	xxSrc = _mm_add_epi16(xxSrc, xxTemp);
	xxSrc = _mm_slli_epi16(xxSrc, SprShift);
	xxTemp = xxSrc;
	xxTemp = _mm_subs_epu16(xxTemp, xxSupp);
	xxSrc = _mm_subs_epu16(xxSrc, xxTemp);

	rax = rrConAddr2 + ConPitch;
	xxCmp1 = _mm_loadu_si128((__m128i*)(rax));
	xxCmp = xxCmp1;
	xxCmp = _mm_mullo_epi16(xxCmp, xxfSPR);

	xxCmp1 = _mm_loadu_si128((__m128i*)(rax + 0x20));
	xxTemp = xxCmp1;
	xxTemp = _mm_slli_epi16(xxTemp, SprShift);
	xxCmp = _mm_add_epi16(xxCmp, xxTemp);

	xxTemp = _mm_loadu_si128((__m128i*)(rax + 0x40));
	xxTemp = _mm_mullo_epi16(xxTemp, xxsSPR);
	xxCmp = _mm_add_epi16(xxCmp, xxTemp);

	xxTemp = xxCmp;
	xxTemp = _mm_subs_epu16(xxTemp, xxSupp);
	xxCmp = _mm_subs_epu16(xxCmp, xxTemp);

	xxThres = xxSrc;
	xxThres = _mm_max_epi16(xxThres, xxCmp);
	xxThres = _mm_srli_epi16(xxThres, ConvNSprShift);
	xxThres = _mm_mullo_epi16(xxThres, xxDSlide);
	xxThres = _mm_srli_epi16(xxThres, 8);
	xxThres = _mm_add_epi16(xxThres, xxDStart);
	xxThres = _mm_slli_epi16(xxThres, ConvNSprShift);

	// ---- Èæ¹é°áÇÔ Â÷µî ¹®Åΰª Àû¿ë ---- //
/*¿ø·¡ÄÚµå:
	xxThres = xxSrc;
	xxThres = _mm_max_epi16(xxThres, xxCmp);
	xxThres = _mm_srli_epi16(xxThres, ConvNSprShift);
	xxThres = _mm_mullo_epi16(xxThres, xxDSlide);
	xxThres = _mm_srli_epi16(xxThres, 8);
	xxThres = _mm_add_epi16(xxThres, xxDStart);
	xxThres = _mm_slli_epi16(xxThres, ConvNSprShift);
	*/
	xxZThW = _mm_slli_epi16(_mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_srli_epi16(
		_mm_max_epi16(xxSrc, xxCmp), ConvNSprShift), xxDSlide), 8), xxDStart), ConvNSprShift);
	xxZThB = _mm_slli_epi16(_mm_add_epi16(_mm_srli_epi16(_mm_mullo_epi16(_mm_srli_epi16(
		_mm_max_epi16(xxSrc, xxCmp), ConvNSprShift), xxDSlideB), 8), xxDStartB), ConvNSprShift);

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

	//INSPECT_BLACK:
	xxPeakAbs = xxCmp;
	xxPeakAbs = _mm_subs_epu16(xxPeakAbs, xxSrc);
	xxPeak = xxPeakAbs;
	xxPeak = _mm_cmpgt_epi16(xxPeak, xxThres);
	rax = _mm_movemask_epi8(xxPeak);
	if(rax == 0)
		goto INSPECT_WHITE;
	rbx = 0;
	goto DEFECT_00;

INSPECT_WHITE:
	xxPeakAbs = xxSrc;
	xxPeakAbs = _mm_subs_epu16(xxPeakAbs, xxCmp);
	xxPeak = xxPeakAbs;
	xxPeak = _mm_cmpgt_epi16(xxPeak, xxThres);
	rax = _mm_movemask_epi8(xxPeak);
	if(rax == 0)
		goto LOOP_Y_TAIL;
	rbx = 1;
	goto DEFECT_00;

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

	rax = _mm_extract_epi16(xxPeakAbs, 0x00);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x00);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x00);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x00);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakAbs, 0x01);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x01);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x01);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x01);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakAbs, 0x02);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x02);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x02);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x02);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakAbs, 0x03);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x03);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x03);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x03);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakAbs, 0x04);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x04);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x04);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x04);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakAbs, 0x05);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x05);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x05);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x05);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

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

	rax = _mm_extract_epi16(xxPeakAbs, 0x06);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x06);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x06);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x06);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

DEFECT_07:
	ix++;
	rax = _mm_extract_epi16(xxPeak, 0x07);
	if(rax == 0)
		goto DEFECT_END;

	rax = _mm_extract_epi16(xxPeakAbs, 0x07);
	rax =  rax >> ConvNSprShift;
	*(DefectX + rrNoD) = (int)ix;
	*(DefectY + rrNoD) = (int)iy;
	*(DefectT + rrNoD) = (int)rbx;
	*(DefectValue + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxSrc, 0x07);
	*(pGraySrc + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxCmp, 0x07);
	*(pGrayCmp + rrNoD) = (int)rax;

	rax = _mm_extract_epi16(xxThres, 0x07);
	*(pZoneTh + rrNoD) = (int)rax;

	rrNoD++;

DEFECT_END:
	ix -= 7;
	if(rbx == 0)
		goto INSPECT_WHITE;
	goto LOOP_Y_TAIL;

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;
	rrConAddr2 = *(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(rrConAddr2 != rcx)
		goto PAIR_TOP;

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

	rax = (__int64)*(DefectT + iy);
	if(rax == rrConAddr)
		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 / 4) * 64;
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

	*(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)rrConAddr;
	PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_TOP_UNPAIR:
	if(m_Param.m_bUseUnpair == FALSE)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)(rrConAddr2);
	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 / 4) * 64;
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

	*(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 / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);

	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(rrConAddr2 != rcx)
		goto PAIR_CENTER;

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

	rax = (__int64)*(DefectT + iy);
	if(rax == rrConAddr)
		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 / 4) * 64;
		else
			sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

		*(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)rrConAddr;
	PairAddr->s_fDefectPeak = (int)*(DefectValue + rbx);
	PairAddr->s_nGraySrc = (int)(sSrc / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);

	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

PAIR_BOTTOM:
	if(m_Param.m_bUseUnpair == FALSE)
		goto LOOP_PAIR_NUM;

	// -------------- à» Filtering -------------- //
	ptDefect.x = (LONG)rrConAddr2;
	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 / 4) * 64;
	else
		sThres = (short)(sRef * m_Param.s_nBWB_Slide / 256 + m_Param.s_nBWB_Main) * 64;

	*(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)rrConAddr2;
	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 / 64);
	PairAddr->s_nGrayRef = (int)(sRef / 64);
	PairAddr->s_nThresold = (int)(sThres / 64);
	PairAddr++;
	nPair++;
	goto LOOP_PAIR_NUM;

LOOP_PAIR_NUM_END:
	rrNoD = 0;
	goto PAIRING_VERT_RETURN;

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

	return (int)nPair;
}