jidctint.c

jpeg编解码器

	pmaddwd		mm4, fix_n256_n196		;p1(17,14)

	movq		mm7, mm6				;p1(17,8)
	pmaddwd		mm5, fix_307n256_n196	;p1(17,17)
	
	paddd		mm6, [esi + 8*18]		;p1(17,9)
	
	paddd		mm6, const_0x2xx8		;p1(17,10)
	
	psrad		mm6, 9					;p1(17,11)
	paddd		mm4, [esi+8*25]			;p1(17,18)

	paddd		mm4, mm5				;p1(17,19)

	movq		mm5, mm4				;p1(17,20)
	
	paddd		mm4, [esi + 8*19]		;p1(17,21)
	
	paddd		mm4, const_0x2xx8		;p1(17,22)
	
	psrad		mm4, 9					;p1(17,23)

	packssdw	mm6, mm4				;p1(17,24)
	
	movq		[esi + 8*2], mm6		;p1(17,25)
	
	movq		mm4, [esi + 8*18]		;p1(17,26)

	movq		mm6, [esi + 8*19]		;p1(17,30)
	psubd		mm4, mm7				;p1(17,27)
	
	paddd		mm4, const_0x2xx8		;p1(17,28)
	psubd		mm6, mm5				;p1(17,31)

	psrad		mm4, 9					;p1(17,29)
	paddd		mm6, const_0x2xx8		;p1(17,32)

	psrad		mm6, 9					;p1(17,33)
	movq		mm7, mm2				;p1(18,3)

	packssdw	mm4, mm6				;p1(17,34)
	movq		[esi + 8*12], mm4		;p1(17,35)

	movq		mm6, mm1				;p1(18,0)
	punpcklwd	mm7, mm2				;p1(18,4)

	punpcklwd	mm6, mm3				;p1(18,1)
	pmaddwd		mm7, fix_205_n256n039	;p1(18,5)

	pmaddwd		mm6, fix_n039_n256		;p1(18,2)
	movq		mm4, mm1				;p1(18,12)

	paddd		mm6, [esi+8*24]			;p1(18,6)
	punpckhwd	mm4, mm3				;p1(18,13)

	paddd		mm6, mm7				;p1(18,7)
	pmaddwd		mm4, fix_n039_n256		;p1(18,14)

	movq		mm7, mm6				;p1(18,8)
	movq		mm5, mm2				;p1(18,15)

	paddd		mm6, [esi + 8*20]		;p1(18,9)
	punpckhwd	mm5, mm2				;p1(18,16)
	
	paddd		mm6, const_0x2xx8		;p1(18,10)
	
	psrad		mm6, 9					;p1(18,11)
	pmaddwd		mm5, fix_205_n256n039	;p1(18,17)

	paddd		mm4, [esi+8*25]			;p1(18,18)

	paddd		mm4, mm5				;p1(18,19)

	movq		mm5, mm4				;p1(18,20)
	
	paddd		mm4, [esi + 8*21]		;p1(18,21)
	
	paddd		mm4, const_0x2xx8		;p1(18,22)
	
	psrad		mm4, 9					;p1(18,23)

	packssdw	mm6, mm4				;p1(18,24)
	
	movq		[esi + 8*4], mm6		;p1(18,25)

	movq		mm4, [esi + 8*20]		;p1(18,26)

	psubd		mm4, mm7				;p1(18,27)
	
	paddd		mm4, const_0x2xx8		;p1(18,28)
	movq		mm7, mm0				;p1(19,3)
	
	psrad		mm4, 9					;p1(18,29)
	movq		mm6, [esi + 8*21]		;p1(18,30)
	
	psubd		mm6, mm5				;p1(18,31)
	punpcklwd	mm7, mm0				;p1(19,4)
	
	paddd		mm6, const_0x2xx8		;p1(18,32)
	
	psrad		mm6, 9					;p1(18,33)
	pmaddwd		mm7, fix_029_n089n196	;p1(19,5)

	packssdw	mm4, mm6				;p1(18,34)

	movq		[esi + 8*10], mm4		;p1(18,35)
	movq		mm6, mm3				;p1(19,0)

	punpcklwd	mm6, mm1				;p1(19,1)	
	movq		mm5, mm0				;p1(19,15)
	
	pmaddwd		mm6, fix_n196_n089		;p1(19,2)
	punpckhwd	mm5, mm0				;p1(19,16)


	paddd		mm6, [esi+8*24]			;p1(19,6)
	movq		mm4, mm3				;p1(19,12)

	paddd		mm6, mm7				;p1(19,7)
	punpckhwd	mm4, mm1				;p1(19,13)


	movq		mm7, mm6				;p1(19,8)
	pmaddwd		mm4, fix_n196_n089  	;p1(19,14)

	paddd		mm6, [esi + 8*22]		;p1(19,9)

	pmaddwd		mm5, fix_029_n089n196	;p1(19,17)
	
	paddd		mm6, const_0x2xx8		;p1(19,10)
	
	psrad		mm6, 9					;p1(19,11)
	paddd		mm4, [esi+8*25]			;p1(19,18)

	paddd		mm4, mm5				;p1(19,19)

	movq		mm5, mm4				;p1(19,20)
	paddd		mm4, [esi + 8*23]		;p1(19,21)
	paddd		mm4, const_0x2xx8		;p1(19,22)
	psrad		mm4, 9					;p1(19,23)

	packssdw	mm6, mm4				;p1(19,24)
	movq		[esi + 8*6], mm6		;p1(19,25)

	movq		mm4, [esi + 8*22]		;p1(19,26)
	
	psubd		mm4, mm7				;p1(19,27)
	movq		mm6, [esi + 8*23]		;p1(19,30)
	
	paddd		mm4, const_0x2xx8		;p1(19,28)
	psubd		mm6, mm5				;p1(19,31)

	psrad		mm4, 9					;p1(19,29)
	paddd		mm6, const_0x2xx8		;p1(19,32)

	psrad		mm6, 9					;p1(19,33)

	packssdw	mm4, mm6				;p1(19,34)
	movq		[esi + 8*8], mm4		;p1(19,35)

//************************************************************  

	add		edi, 8
	add		ebx, 8
	add		esi, 8
  
//************************************************************  
	/* Pass 1. */
    
	movq		mm0, [ebx + 8*4]	;p1(1,0)
	pmullw		mm0, [edi + 8*4]	;p1(1,1)
	
    movq		mm1, [ebx + 8*12]	;p1(2,0)
	pmullw		mm1, [edi + 8*12]	;p1(2,1)

	movq		mm6, [ebx + 8*0]	;p1(5,0)
	pmullw		mm6, [edi + 8*0]	;p1(5,1)

	movq		mm2, mm0				;p1(3,0)
	movq		mm7, [ebx + 8*8]	;p1(6,0)

	punpcklwd	mm0, mm1				;p1(3,1)
	pmullw		mm7, [edi + 8*8]	;p1(6,1)

	movq		mm4, mm0				;p1(3,2)
	punpckhwd	mm2, mm1				;p1(3,4)

	pmaddwd		mm0, fix_054n184_054	;p1(3,3)
	movq		mm5, mm2				;p1(3,5)

	pmaddwd		mm2, fix_054n184_054	;p1(3,6)
	pxor		mm1, mm1	;p1(7,0)
    
	pmaddwd		mm4, fix_054_054p076		;p1(4,0)
	punpcklwd   mm1, mm6	;p1(7,1)

	pmaddwd		mm5, fix_054_054p076		;p1(4,1)
	psrad		mm1, 3		;p1(7,2)

	pxor		mm3, mm3	;p1(7,3)

	punpcklwd	mm3, mm7	;p1(7,4)
	
	psrad		mm3, 3		;p1(7,5)
	
	paddd		mm1, mm3	;p1(7,6)
	
	movq		mm3, mm1	;p1(7,7)	

	paddd		mm1, mm4	;p1(7,8)
	
	psubd		mm3, mm4	;p1(7,9)

	movq		[esi + 8*16], mm1	;p1(7,10)
	pxor		mm4, mm4	;p1(7,12)

	movq		[esi + 8*22], mm3	;p1(7,11)
	punpckhwd	mm4, mm6	;p1(7,13)

	psrad		mm4, 3		;p1(7,14)
	pxor		mm1, mm1	;p1(7,15)
	
	punpckhwd	mm1, mm7	;p1(7,16)
	
	psrad		mm1, 3		;p1(7,17)
	
	paddd		mm4, mm1	;p1(7,18)
	
	movq		mm3, mm4	;p1(7,19)	
	pxor		mm1, mm1	;p1(8,0)

	paddd		mm3, mm5	;p1(7,20)
	punpcklwd	mm1, mm6	;p1(8,1)

	psubd		mm4, mm5	;p1(7,21)
	psrad		mm1, 3		;p1(8,2)

	movq		[esi + 8*17], mm3	;p1(7,22)
	pxor		mm5, mm5	;p1(8,3)
	
	movq		[esi + 8*23], mm4	;p1(7,23)
	punpcklwd	mm5, mm7	;p1(8,4)

	psrad		mm5, 3		;p1(8,5)
	pxor		mm4, mm4	;p1(8,12)

	psubd		mm1, mm5	;p1(8,6)	
	punpckhwd	mm4, mm6	;p1(8,13)

	movq		mm3, mm1	;p1(8,7)
	psrad		mm4, 3		;p1(8,14)

	paddd		mm1, mm0	;p1(8,8)
	pxor		mm5, mm5	;p1(8,15)

	psubd		mm3, mm0	;p1(8,9)
	movq		mm0, [ebx + 8*14]	;p1(9,0)

	punpckhwd	mm5, mm7	;p1(8,16)
	pmullw		mm0, [edi + 8*14]	;p1(9,1)

	movq		[esi + 8*18], mm1	;p1(8,10)
	psrad		mm5, 3		;p1(8,17)

	movq		[esi + 8*20], mm3	;p1(8,11)
	psubd		mm4, mm5	;p1(8,18)

	movq		mm3, mm4	;p1(8,19)
	movq		mm1, [ebx + 8*6]	;p1(10,0)

	paddd		mm3, mm2	;p1(8,20)
	pmullw		mm1, [edi + 8*6]	;p1(10,1)

	psubd		mm4, mm2	;p1(8,21)
	movq		mm5, mm0			;p1(11,1)

	movq		[esi + 8*21], mm4	;p1(8,23)

	movq		[esi + 8*19], mm3	;p1(8,22)
	movq		mm4, mm0			;p1(11,0)

	punpcklwd	mm4, mm1			;p1(11,2)
	movq		mm2, [ebx + 8*10]	;p1(12,0)

	punpckhwd	mm5, mm1			;p1(11,4)	
	pmullw		mm2, [edi + 8*10]	;p1(12,1)

	movq		mm3, [ebx + 8*2]	;p1(13,0)

	pmullw		mm3, [edi + 8*2]	;p1(13,1)
	movq		mm6, mm2			;p1(14,0)

	pmaddwd		mm4, fix_117_117	;p1(11,3)
	movq		mm7, mm2			;p1(14,1)
	
	pmaddwd		mm5, fix_117_117	;p1(11,5)
	punpcklwd	mm6, mm3			;p1(14,2)

	pmaddwd		mm6, fix_117_117	;p1(14,3)
	punpckhwd	mm7, mm3			;p1(14,4)
	
	pmaddwd		mm7, fix_117_117	;p1(14,5)
	paddd		mm4, mm6			;p1(15,0)

	paddd		mm5, mm7			;p1(15,1)
   	movq		[esi+8*24], mm4		;p1(15,2)

	movq		[esi+8*25], mm5		;p1(15,3)
	movq		mm6, mm0				;p1(16,0)

	movq		mm7, mm3				;p1(16,3)
	punpcklwd	mm6, mm2				;p1(16,1)
	
	punpcklwd	mm7, mm3				;p1(16,4)
	pmaddwd		mm6, fix_n039_n089		;p1(16,2)

	pmaddwd		mm7, fix_150_n089n039	;p1(16,5)
	movq		mm4, mm0				;p1(16,12)

	paddd		mm6, [esi+8*24]			;p1(16,6)
	punpckhwd	mm4, mm2				;p1(16,13)

	paddd		mm6, mm7				;p1(16,7)
	pmaddwd		mm4, fix_n039_n089		;p1(16,14)

	movq		mm7, mm6				;p1(16,8)
	paddd		mm4, [esi+8*25]			;p1(16,18)

	movq		mm5, mm3				;p1(16,15)
	paddd		mm6, [esi + 8*16]		;p1(16,9)

	punpckhwd	mm5, mm3				;p1(16,16)
	paddd		mm6, const_0x2xx8		;p1(16,10)
	
	psrad		mm6, 9					;p1(16,11)
	pmaddwd		mm5, fix_150_n089n039	;p1(16,17)

	paddd		mm4, mm5				;p1(16,19)

	movq		mm5, mm4				;p1(16,20)

	paddd		mm4, [esi + 8*17]		;p1(16,21)
	
	paddd		mm4, const_0x2xx8		;p1(16,22)
	
	psrad		mm4, 9					;p1(16,23)

	packssdw	mm6, mm4				;p1(16,24)
	
	movq		[esi + 8*0], mm6		;p1(16,25)
		
	movq		mm4, [esi + 8*16]		;p1(16,26)
	
	psubd		mm4, mm7				;p1(16,27)
	movq		mm6, [esi + 8*17]		;p1(16,30)
	
	paddd		mm4, const_0x2xx8		;p1(16,28)
	movq		mm7, mm1				;p1(17,3)

	psrad		mm4, 9					;p1(16,29)
	psubd		mm6, mm5				;p1(16,31)

	paddd		mm6, const_0x2xx8		;p1(16,32)
	punpcklwd	mm7, mm1				;p1(17,4)

	pmaddwd		mm7, fix_307n256_n196	;p1(17,5)
	psrad		mm6, 9					;p1(16,33)

	packssdw	mm4, mm6				;p1(16,34)
	movq		[esi + 8*14], mm4		;p1(16,35)

	movq		mm6, mm0				;p1(17,0)
	movq		mm4, mm0				;p1(17,12)

	punpcklwd	mm6, mm2				;p1(17,1)
	punpckhwd	mm4, mm2				;p1(17,13)
	
	pmaddwd		mm6, fix_n256_n196		;p1(17,2)
	movq		mm5, mm1				;p1(17,15)

	paddd		mm6, [esi+8*24]			;p1(17,6)
	punpckhwd	mm5, mm1				;p1(17,16)
	
	paddd		mm6, mm7				;p1(17,7)
	pmaddwd		mm4, fix_n256_n196		;p1(17,14)

	movq		mm7, mm6				;p1(17,8)
	pmaddwd		mm5, fix_307n256_n196	;p1(17,17)
	
	paddd		mm6, [esi + 8*18]		;p1(17,9)
	
	paddd		mm6, const_0x2xx8		;p1(17,10)
	
	psrad		mm6, 9					;p1(17,11)
	paddd		mm4, [esi+8*25]			;p1(17,18)

	paddd		mm4, mm5				;p1(17,19)

	movq		mm5, mm4				;p1(17,20)
	
	paddd		mm4, [esi + 8*19]		;p1(17,21)
	
	paddd		mm4, const_0x2xx8		;p1(17,22)
	
	psrad		mm4, 9					;p1(17,23)

	packssdw	mm6, mm4				;p1(17,24)
	
	movq		[esi + 8*2], mm6		;p1(17,25)
	
	movq		mm4, [esi + 8*18]		;p1(17,26)

	movq		mm6, [esi + 8*19]		;p1(17,30)
	psubd		mm4, mm7				;p1(17,27)
	
	paddd		mm4, const_0x2xx8		;p1(17,28)
	psubd		mm6, mm5				;p1(17,31)

	psrad		mm4, 9					;p1(17,29)
	paddd		mm6, const_0x2xx8		;p1(17,32)

	psrad		mm6, 9					;p1(17,33)
	movq		mm7, mm2				;p1(18,3)

	packssdw	mm4, mm6				;p1(17,34)
	movq		[esi + 8*12], mm4		;p1(17,35)

	movq		mm6, mm1				;p1(18,0)
	punpcklwd	mm7, mm2				;p1(18,4)

	punpcklwd	mm6, mm3				;p1(18,1)
	pmaddwd		mm7, fix_205_n256n039	;p1(18,5)

	pmaddwd		mm6, fix_n039_n256		;p1(18,2)
	movq		mm4, mm1				;p1(18,12)

	paddd		mm6, [esi+8*24]			;p1(18,6)
	punpckhwd	mm4, mm3				;p1(18,13)

	paddd		mm6, mm7				;p1(18,7)
	pmaddwd		mm4, fix_n039_n256		;p1(18,14)

	movq		mm7, mm6				;p1(18,8)
	movq		mm5, mm2				;p1(18,15)

	paddd		mm6, [esi + 8*20]		;p1(18,9)
	punpckhwd	mm5, mm2				;p1(18,16)
	
	paddd		mm6, const_0x2xx8		;p1(18,10)