.data

astPair	STRUCT
	pos		dword ?
	pair	dword ?
	dtype	dword ?
	peak	dword ?
	x		dword ?
	y		dword ?
	src		dword ?
	ref		dword ?
	zone	dword ?
	Thre	dword ?	
astPair	ENDS


.code

InspLeft		equ		qword ptr[rbp+ 30h]
InspRight		equ		qword ptr[rbp+ 38h]
InspTop		equ		qword ptr[rbp+ 40h]
InspBottom		equ		qword ptr[rbp+ 48h]

DefectX		equ		qword ptr[rbp+ 50h]
DefectY		equ		qword ptr[rbp+ 58h]
DefectType	equ		qword ptr[rbp+ 60h]
DefectVal	equ		qword ptr[rbp+ 68h]


BuffWidth	equ		qword ptr[rbp+ 70h]
iPitch		equ		qword ptr[rbp+ 78h]
fSPR		equ		qword ptr[rbp+ 80h]
sSPR		equ		qword ptr[rbp+ 88h]

Threshold	equ		qword ptr[rbp+ 90h]
Suppress	equ		qword ptr[rbp+ 98h]
SPR			equ		qword ptr[rbp+ 0a0h]
ShiftCnt	equ		qword ptr[rbp+ 0a8h]

iPitch2	equ		qword ptr[rbp+ 0b0h]
fSPR2		equ		qword ptr[rbp+ 0b8h]
sSPR2		equ		qword ptr[rbp+ 0c0h]

ix			equ		rcx
iy			equ		rdx

rrImgAddr	equ		rsi
rrNDefect	equ		rdi


rrPeakFlag		equ		r8
rrBuffWidth	equ		r9
rrdx		equ		r10
rrdy		equ		r11
rrdt		equ		r12
rrdp		equ		r13


rrConAddr	equ		r15


xxTemp		equ		xmm1

xxPlus2	equ		xmm8
xxMinus2	equ		xmm9

xxSrc		equ		xmm4
xxCmp		equ		xmm5
xxMinus	equ		xmm6
xxPlus		equ		xmm7

xxPeak		equ		xmm2
xxPeak2	equ		xmm3

xxfSPR		equ		xmm10
xxsSPR		equ		xmm11
xxfSpr2	equ		xmm14
xxsSpr2	equ		xmm15
xxThres	equ		xmm12
xxSupp		equ		xmm13



lShiftCnt	equ		4

ConvVert2 PROC
	local	ImgAddr		: qword		; ÆĶó¹ÌÅÍ·Î ³Ñ¾î¿Â °ª
	local	ImgAddrLeft	: qword
	local	ImgConLine		: qword		; Conv Buffer ¸¸µé¶§ ½ÃÀÛ ÁÖ¼Ò
	local	F_Y2Insp		: qword
	local	ShftSize		: qword
	local	maxDefect		: qword
	local	ConPitch		: qword	; 8*iPitch
	local	ConPitch2		: qword
	local	ConvSize		: qword
	local	ConAddr		: qword
	local	ConOrigin		: qword
	local	InspTop1		: qword
	local	InspTop2		: qword
	


	push	rbp
	mov		rbp, rsp
	push	r10
	push	r11
	push	r12
	push	r13
	push	r14
	push	r15
	push	rdi
	push	rsi
	push	rbx
	ADD		rbp, 070h


	mov		ConvSize, 4
	mov		rax, ShiftCnt
	mov		ShftSize, 4

; ÆĶó¹ÌÅÍ·Î ³Ñ¾î¿Â register°ª ¿Å±â±â
	mov		ImgAddr		, rcx
	mov		ConOrigin		, rdx
	mov		maxDefect	, r8
	mov		rrNDefect	, r9
;	sub		maxDefect, 16


; mmx °ª ¸¸µé±â
	mov			rax, Suppress
	mul			SPR
	mul			ConvSize
	movd		xxSupp, rax		; supress
	movdqa		xmm0, xxSupp
	punpcklwd	xmm0, xxSupp
	pshufd		xxSupp, xmm0, 0
	
	
	mov			rax, Threshold
	movd		xxThres, rax		; Threshold
	movdqa		xmm0, xxThres
	punpcklwd	xmm0, xxThres
	pshufd		xxThres, xmm0, 0

	movd		xxfSPR, fSpr		; fSpr
	movdqa		xmm0, xxfSPR
	punpcklwd	xmm0, xxfSPR
	pshufd		xxfSPR, xmm0, 0

	movd		xxsSPR, sSpr		; sSpr
	movdqa		xmm0, xxsSPR
	punpcklwd	xmm0, xxsSPR
	pshufd		xxsSPR, xmm0, 0


	movd		xxfSPR2, fSpr2		; fSpr
	movdqa		xmm0, xxfSPR2
	punpcklwd	xmm0, xxfSPR2
	pshufd		xxfSPR2, xmm0, 0

	movd		xxsSPR2, sSpr2		; sSpr
	movdqa		xmm0, xxsSPR2
	punpcklwd	xmm0, xxsSPR2
	pshufd		xxsSPR2, xmm0, 0


; register ÇÒ´ç.
	mov		ix, InspLeft
	mov		iy, InspTop

	mov		rrdx, DefectX
	mov		rrdy, DefectY
	mov		rrdt, DefectType
	mov		rrdp, DefectVal
	mov		rrBuffWidth, BuffWidth




; °Ë»ç ¿µ¿ª ¼³Á¤(width¸¦ pitch¸¸Å­ À߶ó¾ß ºñ±³ÇÑ´Ù)
; ÁÖ¼Ò°ª ¼³Á¤

	mov		rax, rrBuffWidth
	mul		InspTop
	add		rax, ImgAddr
	add		rax, ix
	mov		ImgAddrLeft, rax
	
	mov		rax, 20h
	mul		InspTop
	add		rax, ConOrigin
	mov		ConAddr, rax



	mov		rax, InspRight
	sub		rax, InspLeft
	sar		rax, 4
	sal		rax, 4
	add		rax, InspLeft
	mov		InspRight, rax		; ½ÇÁ¦ °Ë»ç ¿µ¿ª ³¡ÁöÁ¡ º¯°æ(8 byte align)

	mov		rax, 20h
	mul		iPitch
	mov		ConPitch, rax
	
	mov		rax, 20h
	mul		iPitch2
	mov		ConPitch2, rax

	mov		rax, InspTop
	add		rax, iPitch
	add		rax, 2
	mov		InspTop1, rax
	
	mov		rax, InspBottom
	sub		rax, iPitch
	sub		rax, 2
	mov		InspTop2, rax


	pxor	xmm0, xmm0	
	mov		iy, InspTop
	sub		ImgAddrLeft, 10h


LOOP_X:
	cmp		ix, InspRight
	jge		LOOP_X_END
	
	add		ImgAddrLeft, 10h
	mov		rrImgAddr, ImgAddrLeft
	mov		rrConAddr, ConAddr
	
	mov		iy, InspTop

; Àӽà Conbuff
;	mov		rrImgAddr, ImgAddr
;	add		rrImgAddr, ix
;	mov		rrConAddr, ConOrigin
;	mov		iy, 0
; Àӽà Conbuff

; Con ¹öÆÛ ¸¸µé±â
LOOP_Y_CON_BUFF:
	movdqu		xxSrc, [rrImgAddr]
	movdqa		xxCmp, xxSrc
	punpcklbw	xxSrc, xmm0
	punpckhbw	xxCmp, xmm0

	movdqu		xxPlus, [rrImgAddr+ 1]
	movdqa		xxMinus, xxPlus
	punpcklbw	xxPlus, xmm0
	punpckhbw	xxMinus, xmm0

	paddw		xxSrc, xxPlus
	paddw		xxCmp, xxMinus
	movdqu		[rrConAddr], xxSrc
	movdqu		[rrConAddr+ 10h], xxCmp
	
	add		rrImgAddr, rrBuffWidth
	add		rrConAddr, 20h

	inc		iy
	cmp		iy, InspBottom
	jle		LOOP_Y_CON_BUFF


	mov		rrConAddr, ConAddr
	mov		iy, InspTop
	mov		F_Y2Insp, 0

LOOP_Y:
	cmp		iy, InspBottom
	jge		LOOP_Y_END

	cmp		rrNDefect, maxDefect
	jge		End_Buffer


; ¼Ò½º ¹öÆÛ
	movdqu	xxSrc, [rrConAddr]
	movdqu	xxTemp, [rrConAddr+ 20h]
	paddw	xxSrc, xxTemp
	psllw	xxSrc, lShiftCnt

	movdqa		xxTemp, xxSrc				; suppress (2*2*SPR ÇØÁØ suppress)
	psubusw	xxTemp, xxSupp
	psubusw	xxSrc, xxTemp


	cmp		iy, InspTop2
	jg		CaseCmpTop3

; ºñ±³ ¹öÆÛ ÇÕ ¸¸µé±â
	mov		rax, rrConAddr
	add		rax, ConPitch

	movdqu	xxCmp, [rax]
	pmullw	xxCmp, xxfSPR

	movdqu	xxTemp, [rax+ 20h]
	psllw	xxTemp, lShiftCnt
	paddw	xxCmp, xxTemp
	
	movdqu	xxTemp, [rax+ 40h]
	pmullw	xxTemp, xxsSPR
	paddw	xxCmp, xxTemp

	jmp		CaseCmpEnd

CaseCmpTop3:
	mov		rax, rrConAddr
	sub		rax, ConPitch
	
	movdqu	xxCmp, [rax+ 20h]
	pmullw	xxCmp, xxfSPR
	
	movdqu	xxTemp, [rax]
	psllw	xxTemp, lShiftCnt
	paddw	xxCmp, xxTemp
	
	movdqu	xxTemp, [rax- 20h]
	pmullw	xxTemp, xxsSPR
	paddw	xxCmp, xxTemp


CaseCmpEnd:
	movdqa		xxTemp, xxCmp				; suppress (2*2*SPR ÇØÁØ suppress)
	psubusw	xxTemp, xxSupp
	psubusw	xxCmp, xxTemp


	movdqa	xxMinus	, xxCmp
	psubusw	xxMinus	, xxSrc
	movdqa	xxPlus	, xxSrc
	psubusw	xxPlus	, xxCmp

Inspect_Black:
	movdqa		xxPeak, xxMinus
	pcmpgtw	xxPeak, xxThres
	pmovmskb	rax, xxPeak

Inspect_White:
	movdqa		xxPeak2, xxPlus
	pcmpgtw	xxPeak2, xxThres
	pmovmskb	rbx, xxPeak2
	add		rax, rbx
	cmp		rax, 0
	je		Defect_None

	cmp		iy, InspTop2
	jg		CaseCmp2Top3

	cmp		iy, InspTop1
	jg		CaseCmp2Top2
	
; °áÇÔ °ËÁõ

CaseCmp2Top1:
	mov		rax, rrConAddr
	add		rax, ConPitch2

	movdqu	xxCmp, [rax]
	pmullw	xxCmp, xxfSPR2

	movdqu	xxTemp, [rax+ 20h]
	psllw	xxTemp, lShiftCnt
	paddw	xxCmp, xxTemp

	movdqu	xxTemp, [rax+ 40h]
	pmullw	xxTemp, xxsSPR2
	paddw	xxCmp, xxTemp

	jmp		Inspect_Black2

CaseCmp2Top2:
	mov		rax, rrConAddr
	sub		rax, ConPitch

	movdqu	xxCmp, [rax+ 20h]
	pmullw	xxCmp, xxfSPR

	movdqu	xxTemp, [rax]
	psllw	xxTemp, lShiftCnt
	paddw	xxCmp, xxTemp

	movdqu	xxTemp, [rax- 20h]
	pmullw	xxTemp, xxsSPR
	paddw	xxCmp, xxTemp

	
	jmp		Inspect_Black2

CaseCmp2Top3:
	mov		rax, rrConAddr
	sub		rax, ConPitch2

	movdqu	xxCmp, [rax+ 20]
	pmullw	xxCmp, xxfSPR2

	movdqu	xxTemp, [rax]
	psllw	xxTemp, lShiftCnt
	paddw	xxCmp, xxTemp

	movdqu	xxTemp, [rax- 20h]
	pmullw	xxTemp, xxsSPR2
	paddw	xxCmp, xxTemp



Inspect_Black2:
	movdqa		xxTemp, xxCmp				; suppress (2*2*SPR ÇØÁØ suppress)
	psubusw	xxTemp, xxSupp
	psubusw	xxCmp, xxTemp

	movdqa		xxMinus2	, xxCmp
	psubusw	xxMinus2	, xxSrc
	movdqa		xxPlus2	, xxSrc
	psubusw	xxPlus2	, xxCmp

	mov			bx, 0
	movdqa		xxTemp, xxMinus2
	pcmpgtw	xxTemp, xxThres
	pand		xxPeak, xxTemp
	movdqa		xxTemp, xxMinus
	pmovmskb	rax, xxPeak
	cmp		rax, 0
	jne		DEFECT_00

Inspect_White2:
	mov			bx, 1
	movdqa		xxPeak, xxPlus2
	pcmpgtw	xxPeak, xxThres
	pand		xxPeak, xxPeak2
	movdqa		xxTemp, xxPlus
	pmovmskb	rax, xxPeak
	cmp		rax, 0
	je		Defect_None

 

DEFECT_00:
	pextrw	rax, xxPeak, 0
	cmp		rax, 0h
	je		Defect_01
	
	pextrw	rax, xxTemp, 0
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

DEFECT_01:
	inc		ix
	pextrw	rax, xxPeak, 1
	cmp		rax, 0h
	je		Defect_02

	pextrw	rax, xxTemp, 1
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

DEFECT_02:
	inc		ix
	pextrw	rax, xxPeak, 2
	cmp		rax, 0h
	je		Defect_03
	
	pextrw	rax, xxTemp, 2
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

DEFECT_03:
	inc		ix
	pextrw	rax, xxPeak, 3
	cmp		rax, 0h
	je		Defect_04
	
	pextrw	rax, xxTemp, 3
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

DEFECT_04:
	inc		ix
	pextrw	rax, xxPeak, 4
	cmp		rax, 0h
	je		Defect_05
	
	pextrw	rax, xxTemp, 4
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

DEFECT_05:
	inc		ix
	pextrw	rax, xxPeak, 5
	cmp		rax, 0h
	je		Defect_06
	
	pextrw	rax, xxTemp, 5
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

DEFECT_06:
	inc		ix
	pextrw	rax, xxPeak, 6
	cmp		rax, 0h
	je		Defect_07
	
	pextrw	rax, xxTemp, 6
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

DEFECT_07:
	inc		ix
	pextrw	rax, xxPeak, 7
	cmp		rax, 0h
	je		Defect_End
	
	pextrw	rax, xxTemp, 7
	sar		rax, 6
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	mov		[rrdp+ 2*rrNDefect], ax
	inc		rrNDefect

Defect_End:
	sub		ix, 7
	cmp		bx, 0
	je		Inspect_White2

Defect_None:
	mov		[rrdx+ 4*rrNDefect], ecx
	mov		[rrdy+ 4*rrNDefect], edx
	mov		[rrdt+ 2*rrNDefect], bx
	
	inc		iy	
	add		rrConAddr, 20h
	jmp		LOOP_Y

LOOP_Y_END:
	cmp		F_Y2Insp, 1
	je		LOOP_Y_END2

	add		ix, 8
	mov		rrConAddr, ConAddr
	add		rrConAddr, 10h
	mov		iy, InspTop
	mov		F_Y2Insp, 1

	jmp		LOOP_Y

LOOP_Y_END2:

	add		ix, 8h
	jmp		LOOP_X
LOOP_X_END:




END_BUFFER:

	mov		rax, rrNDefect					; return rrNDefect
	
	pop	rbx
	pop	rsi
	pop	rdi
	pop	r15
	pop	r14
	pop	r13
	pop	r12
	pop	r11
	pop	r10
	pop		rbp

RET
ConvVert2 ENDP

end