; Copyright (c) Microsoft Corporation.  All rights reserved.


; Custom Build Step, including a listing file placed in intermediate directory
; but without Source Browser information
; debug:
; ml -c -Zi "-Fl$(IntDir)\$(InputName).lst" "-Fo$(IntDir)\$(InputName).obj" "$(InputPath)"
; release:
; ml -c "-Fl$(IntDir)\$(InputName).lst" "-Fo$(IntDir)\$(InputName).obj" "$(InputPath)"
; outputs:
; $(IntDir)\$(InputName).obj

; Custom Build Step, including a listing file placed in intermediate directory
; and Source Browser information also placed in intermediate directory
; debug:
; ml -c -Zi "-Fl$(IntDir)\$(InputName).lst" "-FR$(IntDir)\$(InputName).sbr" "-Fo$(IntDir)\$(InputName).obj" "$(InputPath)"
; release:
; ml -c "-Fl$(IntDir)\$(InputName).lst" "-FR$(IntDir)\$(InputName).sbr" "-Fo$(IntDir)\$(InputName).obj" "$(InputPath)"
; outputs:
; $(IntDir)\$(InputName).obj
; $(IntDir)\$(InputName).sbr

;.386

;.MODEL FLAT, C

PBYTE	TYPEDEF		PTR BYTE
PWORD	TYPEDEF		PTR	WORD
PDWORD	TYPEDEF		PTR DWORD

.data


.code

; Euclid's algorithm finds the greatest common divisor by repeatedly
; subtracting the smaller number from the larger number until zero
; is reached. The number remaining is the greatest common divisor.

ImgAddr		equ		r14

ByteAddr	equ		ecx
ShortAddr	equ		edx
nByte		equ		r8

iByte		equ		rsi

Byte_To_Short PROC

	mov		iByte, 16
	pxor	xmm0, xmm0

	cmp		nByte, 16
	jl		START_EXTRA

LOOP_SIMD:
	
	movdqu	xmm1, [ByteAddr]
	movdqa	xmm2, xmm1

	punpcklbw	xmm1, xmm0
	punpckhbw	xmm2, xmm0

	movntdq	[ShortAddr], xmm1
	movntdq	[ShortAddr+16], xmm2

	add		ByteAddr, 16
	add		ShortAddr, 32

	add		iByte, 16
	cmp		iByte, nByte
	jle		LOOP_SIMD
	sub		iByte, 16

START_EXTRA:
	xor		rax, rax

LOOP_EXTRA:

	mov		al, [ByteAddr]
	mov		[ShortAddr], ax


	inc		ByteAddr
	add		ShortAddr, 2
	inc		iByte
	
	cmp		iByte, nByte
	jl		LOOP_EXTRA


RET
Byte_To_Short ENDP




;==========================================================

iPitch		equ		rcx
SPR			equ		rdx
fSPR		equ		r8
sSPR		equ		r9

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

pData		equ		qword ptr[rbp+ 50h]
BuffWidth	equ		qword ptr[rbp+ 58h]

ix			equ		rsi
iy			equ		rdi


xxSrc		equ		xmm4
xxSrc2		equ		xmm5
xxCmp		equ		xmm6
xxCmp2		equ		xmm7
xxSpr		equ		xmm8
xxOne		equ		xmm9
xxfSpr		equ		xmm10
xxsSpr		equ		xmm11
xxMinus		equ		xmm12
xxMinus2	equ		xmm13
xxRslt		equ		xmm14

rrImgAddr	equ		r15
rrPitch		equ		r10
rrPitch2	equ		r11




CheckPitchAsm	PROC
	local	ImgAddrLine		: qword
	local	ImgLineEnd		: qword
	local	ImgLineLength	: qword
	local	Result64[2]		: qword

	mov		rrPitch, iPitch
	mov		rrPitch2, rrPitch
	add		rrPitch2, 1h
	

	mov		rax, 1
	movd		xxOne, rax		; 1·Î ä¿î´Ù.(word)
	movdqa		xmm0, xxOne
	punpcklwd	xmm0, xxOne
	pshufd		xxOne, xmm0, 0
	
	mov		rax, SPR
	movd		xxSpr, rax
	movdqa		xmm0, xxSpr
	punpcklwd	xmm0, xxSpr
	pshufd		xxSpr, xmm0, 0

	mov		rax, fSpr
	movd		xxfSpr, rax
	movdqa		xmm0, xxfSpr
	punpcklwd	xmm0, xxfSpr
	pshufd		xxfSpr, xmm0, 0
	
	movdqa	xxsSpr, xxSpr
	psubw	xxsSpr, xxfSpr



; À̹ÌÁö ÁÖ¼Ò ¸¸µé±â
	mov		rrImgAddr, pData
	mov		rax, BuffWidth	; ImgAddr= ImgAddr+ BuffWidth*InspTop + InspLeft;
	mul		InspTop
	add		rax, InspLeft
	add		rrImgAddr, rax
	mov		ImgAddrLine, rrImgAddr
	
	mov		rax, InspRight
	sub		rax, InspLeft
	mov		ImgLineLength, rax
	
	add		rax, rrImgAddr
	mov		ImgLineEnd, rax

	mov		iy, InspTop
	pxor	xxRslt, xxRslt
	pxor	xmm0, xmm0

Cmp_16:
										; 1. Src Buffer
	movdqu	xxSrc, [rrImgAddr]			; Load
	movdqa	xxSrc2, xxSrc	
	punpcklbw	xxSrc, xmm0				; Unpack
	punpckhbw	xxSrc2, xmm0	

	movdqu	xxMinus, [rrImgAddr+ 1]			; Load
	movdqa	xxMinus2, xxMinus	
	punpcklbw	xxMinus, xmm0				; Unpack
	punpckhbw	xxMinus2, xmm0	
	
	paddw	xxSrc, xxMinus
	paddw	xxSrc2, xxMinus2
	pmullw	xxSrc, xxSpr				; SPR
	pmullw	xxSrc2, xxSpr


										;	2. Compare Buffer
	movdqu	xxCmp, [rrImgAddr+ rrPitch]	; Load
	movdqa	xxCmp2, xxCmp
	punpcklbw	xxCmp, xmm0				; Unpack
	punpckhbw	xxCmp2, xmm0
	pmullw	xxCmp, xxfSpr				; fSpr
	pmullw	xxCmp2, xxfSpr
	
	psubw	xxSrc, xxCmp				; SPR - fSpr
	psubw	xxSrc2, xxCmp2

	movdqu	xxCmp, [rrImgAddr+ rrPitch2]
	movdqa	xxCmp2, xxCmp
	punpcklbw	xxCmp, xmm0
	punpckhbw	xxCmp2, xmm0
	pmullw	xxCmp, xxSpr
	pmullw	xxCmp2, xxSpr
	
	psubw	xxSrc, xxCmp
	psubw	xxSrc2, xxCmp2

	movdqu	xxCmp, [rrImgAddr+ rrPitch2+ 1]		; Load
	movdqa	xxCmp2, xxCmp
	punpcklbw	xxCmp, xmm0				; Unpack
	punpckhbw	xxCmp2, xmm0
	pmullw	xxCmp, xxsSpr				; sSpr
	pmullw	xxCmp2, xxsSpr

	psubw	xxSrc, xxCmp				; (SPR- fSpr) - sSpr
	psubw	xxSrc2, xxCmp2


	pabsw	xxSrc, xxSrc				; Àý´ë°ª
	psubusw	xxSrc, xxOne
	pmaddwd	xxSrc, xxOne				; multiply and add ( 16 -> 32)
	pabsw	xxSrc2, xxSrc2				; Àý´ë°ª
	psubusw	xxSrc2, xxOne
	pmaddwd	xxSrc2, xxOne				; multiply and add ( 16 -> 32)
	
	movdqa	xxCmp, xxSrc
	punpckldq	xxCmp, xmm0				; unpack 32 -> 64
	punpckhdq	xxSrc, xmm0
	
	paddq	xxRslt, xxSrc				; add to Result(64)
	paddq	xxRslt, xxCmp				; add to Result(64)

	movdqa	xxCmp, xxSrc2
	punpckldq	xxCmp, xmm0				; unpack 32 -> 64
	punpckhdq	xxSrc2, xmm0

	paddq	xxRslt, xxSrc2				; add to Result(64)
	paddq	xxRslt, xxCmp				; add to Result(64)

	
	
	add	rrImgAddr, 10h
	cmp	rrImgAddr, ImgLineEnd
	jl	Cmp_16
	
	inc	iy
	cmp	iy, InspBottom
	jge	End_Buff


	mov	rax, ImgAddrLine
	add	rax, BuffWidth
	mov	ImgAddrLine, rax
	mov	rrImgAddr, rax
	
	add	rax, ImgLineLength
	mov	ImgLineEnd, rax
	
	jmp	Cmp_16


End_Buff:

	lea	esi, Result64
	movdqu	[esi], xxRslt
	mov	rax, [esi]
	mov	rcx, [esi+8h]
	add	rax, rcx

RET
CheckPitchAsm	ENDP
end