|
|
;/* *************************************************************************;** INTEL Corporation Proprietary Information;**;** This listing is supplied under the terms of a license;** agreement with INTEL Corporation and may not be copied;** nor disclosed except in accordance with the terms of;** that agreement.;**;** Copyright (c) 1996 Intel Corporation.;** All Rights Reserved.;**;** *************************************************************************;*/;/* *************************************************************************;** $Header: S:\h26x\src\dec\d3mmc.asv 1.1 14 Mar 1996 14:34:54 AGUPTA2 $;** $Log: S:\h26x\src\dec\d3mmc.asv $;// ;// Rev 1.1 14 Mar 1996 14:34:54 AGUPTA2;// ;// Added alignment directives.;// ;// Rev 1.0 14 Mar 1996 14:32:58 AGUPTA2;// Initial revision.;** *************************************************************************;*/.586.model flatOPTION PROLOGUE:NoneOPTION EPILOGUE:None
.xlistinclude iammx.inc.list
MMXCODE1 SEGMENT PARA USE32 PUBLIC 'CODE'MMXCODE1 ENDS
MMXDATA1 SEGMENT PARA USE32 PUBLIC 'DATA'MMXDATA1 ENDS
MMXDATA1 SEGMENTALIGN 8C0101010101010101H DD 001010101H, 001010101HCfefefefefefefefeH DD 0fefefefeH, 0fefefefeHCfcfcfcfcfcfcfcfcH DD 0fcfcfcfcH, 0fcfcfcfcHC0303030303030303H DD 003030303H, 003030303HTWO DD 002020202H, 002020202HMMXDATA1 ENDS
PITCH TEXTEQU <384>
MMXCODE1 SEGMENT; @MMX_Interpolate_Int_Half; This routine computes interpolated pels shown by 'x' for a an 8x8 block; of pels. 'x' is computed by the formula (A+B+1)/2. The input and output; pitch is assumed to be 384 (PITCH).; A . . . . . . .; x x x x x x x x; B . . . . . . .; The basic instruction sequence is:; movq V0, A; movq V2, B; movq V1, V0; por V1, V2; pand V1, 0x0101010101010101; pand V0, 0xfefefefefefefefe; psrlq V0, 1; pand V2, 0xfefefefefefefefe; psrlq V2, 1; paddb V0, V1; paddb V0, V2; movq dest, V0; The instruction sequence for line 0 is 12 instructions. The instruction; sequence for line 1 should be 12 instructions but is not because some of; the values needed for line 1 have already been computed for line 0.;; Registers used for lines 0-7 are:; line 0: mm0, mm1, mm2; line 1: mm2, mm3, mm4; line 2: mm4, mm5, mm0; line 3: mm0, mm1, mm2; line 4: mm2, mm3, mm4; line 5: mm4, mm5, mm0; line 6: mm0, mm1, mm2; line 7: mm2, mm3, mm4; Constants 0x0101010101010101 and 0xfefefefefefefefe are in mm6 and mm7,; respectively.; Parameters:; The source block parameter should be in ecx and the destination block; parameter should be in edx; i.e. it uses fastcall calling convention.; (I am not aware of a way to declare a MASM function of type __fastcall.); Performance:; 41 cycles ignoring unaligned memory accesses; 68 cycles if all loads are unaligned (41+9*3); stores should always be; aligned.ALIGN 4@MMX_Interpolate_Int_Half@8 PROC EXTRACTLOWBIT TEXTEQU <mm6> CLEARLOWBIT TEXTEQU <mm7> movq mm0, [ecx] ;0 ; movq mm2, [ecx+PITCH] ;0 movq mm1, mm0 ;0 movq mm6, C0101010101010101H ; movq mm3, mm2 ;1 movq mm7, CfefefefefefefefeH ; por mm1, mm2 ;0 pand mm0, CLEARLOWBIT ;0 pand mm2, CLEARLOWBIT ;0 psrlq mm0, 1 ;0 pand mm1, EXTRACTLOWBIT ;0 movq mm4, [ecx+2*PITCH] ;1 psrlq mm2, 1 ;0 paddb mm0, mm1 ;0 movq mm5, mm4 ;2 paddb mm0, mm2 ;0 por mm3, mm4 ;1 pand mm4, CLEARLOWBIT ;1 pand mm3, EXTRACTLOWBIT ;1 movq [edx+0*PITCH], mm0 ;0 psrlq mm4, 1 ;1 movq mm0, [ecx+3*PITCH] ;2 paddb mm2, mm3 ;1 movq mm1, mm0 ;3 paddb mm2, mm4 ;1 por mm5, mm0 ;2 pand mm0, CLEARLOWBIT ;2 movq [edx+1*PITCH], mm2 ;1 psrlq mm0, 1 ;2 paddb mm4, mm0 ;2 pand mm5, EXTRACTLOWBIT ;2 movq mm2, [ecx+4*PITCH] ;3 paddb mm4, mm5 ;2 por mm1, mm2 ;3 movq mm3, mm2 ;4 movq [edx+2*PITCH],mm4 ;2 pand mm2, CLEARLOWBIT ;3 psrlq mm2, 1 ;3 pand mm1, EXTRACTLOWBIT ;3 movq mm4, [ecx+5*PITCH] ;4 paddb mm0, mm1 ;3 movq mm5, mm4 ;5 paddb mm0, mm2 ;3 por mm3, mm4 ;4 pand mm4, CLEARLOWBIT ;4 movq [edx+3*PITCH],mm0 ;3 pand mm3, EXTRACTLOWBIT ;4 movq mm0, [ecx+6*PITCH] ;5 psrlq mm4, 1 ;4 movq mm1, mm0 ;6 paddb mm2, mm3 ;4 paddb mm2, mm4 ;4 por mm5, mm0 ;5 pand mm0, CLEARLOWBIT ;5 pand mm5, EXTRACTLOWBIT ;5 movq [edx+4*PITCH], mm2 ;4 psrlq mm0, 1 ;5 movq mm2, [ecx+7*PITCH] ;6 paddb mm4, mm5 ;5 movq mm3, mm2 ;7 paddb mm4, mm0 ;5 por mm1, mm2 ;6 pand mm2, CLEARLOWBIT ;6 movq [edx+5*PITCH], mm4 ;5 pand mm1, EXTRACTLOWBIT ;6 movq mm4, [ecx+8*PITCH] ;7 psrlq mm2, 1 ;6 por mm3, mm4 ;7 paddb mm0, mm1 ;6 paddb mm0, mm2 ;6 pand mm3, EXTRACTLOWBIT ;7 pand mm4, CLEARLOWBIT ;7 paddb mm3, mm2 ;7 movq [edx+6*PITCH], mm0 ;6 psrlq mm4, 1 ;7 paddb mm3, mm4 ;7 ; ; ; movq [edx+7*PITCH], mm3 ;7 ret EXTRACTLOWBIT TEXTEQU <> CLEARLOWBIT TEXTEQU <>@MMX_Interpolate_Int_Half@8 endp
; @MMX_Interpolate_Half_Int; This routine computes interpolated pels shown by 'x' for a an 8x8 block; of pels. 'x' is computed by the formula (A+B+1)/2. The input and output; pitch is assumed to be 384 (PITCH).; A X B X . X . X . X . X . X . X; The basic instruction sequence is:; movq V0, A; movq V2, B; movq V1, V0; por V1, V2; pand V1, 0x0101010101010101; pand V0, 0xfefefefefefefefe; psrlq V0, 1; pand V2, 0xfefefefefefefefe; psrlq V2, 1; paddb V0, V1; paddb V0, V2; movq dest, V0; The instruction sequence for all lines is 12 instructions.;; Registers used for lines 0-7 are:; line 0: mm0, mm1, mm2; line 1: mm3, mm4, mm5; line 2: mm0, mm1, mm2; line 3: mm3, mm4, mm5; line 4: mm0, mm1, mm2; line 5: mm3, mm4, mm5; line 6: mm0, mm1, mm2; line 7: mm3, mm4, mm5; Constants 0x0101010101010101 and 0xfefefefefefefefe are in mm6 and mm7,; respectively.; Parameters:; The source block parameter should be in ecx and the destination block; parameter should be in edx; i.e. it uses fastcall calling convention.; Performance:; 51 cycles ignoring unaligned memory accesses; 99 cycles if all loads are unaligned (51+8*6); stores should always be; aligned.ALIGN 4@MMX_Interpolate_Half_Int@8 proc EXTRACTLOWBIT TEXTEQU <mm6> CLEARLOWBIT TEXTEQU <mm7> movq mm0, [ecx] ;0 mm0,mm1=left pels ; ; mm2 =right pels movq mm2, [ecx+1] ;0 mm1 =interp pels movq mm1, mm0 ;0 movq mm7, CfefefefefefefefeH ; por mm1, mm2 ;0 movq mm6, C0101010101010101H ; pand mm0, CLEARLOWBIT ;0 pand mm2, CLEARLOWBIT ;0 psrlq mm0, 1 ;0 psrlq mm2, 1 ;0 pand mm1, EXTRACTLOWBIT ;0 movq mm3, [ecx+1*PITCH] ;1 mm3,mm4=left pels paddb mm1, mm0 ;0 mm5 =right pels movq mm5, [ecx+1*PITCH+1] ;1 mm4 =interp pels paddb mm1, mm2 ;0 movq mm4, mm3 ;1 pand mm3, CLEARLOWBIT ;1 movq [edx], mm1 ;0 por mm4, mm5 ;1 psrlq mm3, 1 ;1 pand mm5, CLEARLOWBIT ;1 psrlq mm5, 1 ;1 pand mm4, EXTRACTLOWBIT ;1 movq mm0, [ecx+2*PITCH] ;2 mm0,mm1=left pels paddb mm4, mm3 ;1 mm2 =right pels movq mm2, [ecx+2*PITCH+1] ;2 mm1 =interp pels paddb mm4, mm5 ;1 movq mm1, mm0 ;2 pand mm0, CLEARLOWBIT ;2 movq [edx+1*PITCH], mm4 ;1 por mm1, mm2 ;2 psrlq mm0, 1 ;2 pand mm2, CLEARLOWBIT ;2 psrlq mm2, 1 ;2 pand mm1, EXTRACTLOWBIT ;2 movq mm3, [ecx+3*PITCH] ;3 mm3,mm4=left pels paddb mm1, mm0 ;2 mm5 =right pels movq mm5, [ecx+3*PITCH+1] ;3 mm4 =interp pels paddb mm1, mm2 ;2 movq mm4, mm3 ;3 pand mm3, CLEARLOWBIT ;3 movq [edx+2*PITCH], mm1 ;2 por mm4, mm5 ;3 psrlq mm3, 1 ;3 pand mm5, CLEARLOWBIT ;3 psrlq mm5, 1 ;3 pand mm4, EXTRACTLOWBIT ;3 movq mm0, [ecx+4*PITCH] ;4 mm0,mm1=left pels paddb mm4, mm3 ;3 mm2 =right pels movq mm2, [ecx+4*PITCH+1] ;4 mm1 =interp pels paddb mm4, mm5 ;3 movq mm1, mm0 ;4 pand mm0, CLEARLOWBIT ;4 movq [edx+3*PITCH], mm4 ;3 por mm1, mm2 ;4 psrlq mm0, 1 ;4 pand mm2, CLEARLOWBIT ;4 psrlq mm2, 1 ;4 pand mm1, EXTRACTLOWBIT ;4 movq mm3, [ecx+5*PITCH] ;5 mm3,mm4=left pels paddb mm1, mm0 ;4 mm5 =right pels movq mm5, [ecx+5*PITCH+1] ;5 mm4 =interp pels paddb mm1, mm2 ;4 movq mm4, mm3 ;5 pand mm3, CLEARLOWBIT ;5 movq [edx+4*PITCH], mm1 ;4 por mm4, mm5 ;5 psrlq mm3, 1 ;5 pand mm5, CLEARLOWBIT ;5 psrlq mm5, 1 ;5 pand mm4, EXTRACTLOWBIT ;5 movq mm0, [ecx+6*PITCH] ;6 mm0,mm1=left pels paddb mm4, mm3 ;5 mm2 =right pels movq mm2, [ecx+6*PITCH+1] ;6 mm1 =interp pels paddb mm4, mm5 ;5 movq mm1, mm0 ;6 pand mm0, CLEARLOWBIT ;6 movq [edx+5*PITCH], mm4 ;5 por mm1, mm2 ;6 psrlq mm0, 1 ;6 pand mm2, CLEARLOWBIT ;6 psrlq mm2, 1 ;6 pand mm1, EXTRACTLOWBIT ;6 movq mm3, [ecx+7*PITCH] ;7 mm3,mm4=left pels paddb mm1, mm0 ;6 mm5 =right pels movq mm5, [ecx+7*PITCH+1] ;7 mm4 =interp pels paddb mm1, mm2 ;6 movq mm4, mm3 ;7 pand mm3, CLEARLOWBIT ;7 por mm4, mm5 ;7 psrlq mm3, 1 ;7 pand mm4, EXTRACTLOWBIT ;7 pand mm5, CLEARLOWBIT ;7 psrlq mm5, 1 ;7 paddb mm4, mm3 ;7 movq [edx+6*PITCH], mm1 ;6 paddb mm4, mm5 ;7 ; ; movq [edx+7*PITCH], mm4 ;7 ret EXTRACTLOWBIT TEXTEQU <> CLEARLOWBIT TEXTEQU <>@MMX_Interpolate_Half_Int@8 endp
; @MMX_Interpolate_Half_Half; This routine computes interpolated pels shown by 'X' for a an 8x8 block; of pels. 'x' is computed by the formula (A+B+C+D+2)/4. The input and; output pitch is assumed to be 384 (PITCH).; A B; X; C D; The value (A+B+C+D+2)/4 is computed as (A'+B'+C'+D')+((A*+B*+C*+D*+2)/4); where A = 4*A' + A*, etc.; Parameters:; The source block parameter should be in ecx and the destination block; parameter should be in edx; i.e. it uses fastcall calling convention.; Performance:; 84 cycles ignoring unaligned memory accesses; 138 cycles if all loads are unaligned (84+9*2*3); stores should always be; aligned. Average cycle count will be less than 138.ALIGN 4@MMX_Interpolate_Half_Half@8 proc EXTRACTLOWBITS TEXTEQU <mm6> CLEARLOWBITS TEXTEQU <mm7> movq mm0, [ecx] ;0 A(mm0,mm1) B(mm4,mm5) ; 0 movq mm7, CfcfcfcfcfcfcfcfcH ; C(mm2,mm3) D(mm4,mm5) movq mm1, mm0 ;0 movq mm4, [ecx+1] ;0 pand mm0, CLEARLOWBITS ;0 movq mm6, C0303030303030303H ; movq mm5, mm4 ;0 pand mm4, CLEARLOWBITS ;0 pand mm1, EXTRACTLOWBITS ;0 psrlq mm0, 2 ;0 pand mm5, EXTRACTLOWBITS ;0 psrlq mm4, 2 ;0 paddb mm1, mm5 ;0 (A+B) low movq mm2, [ecx+1*PITCH] ;0 paddb mm0, mm4 ;0 (A+B)/4 high movq mm4, [ecx+1*PITCH+1] ;0 movq mm3, mm2 ;0 pand mm3, EXTRACTLOWBITS ;0 movq mm5, mm4 ;0 pand mm5, EXTRACTLOWBITS ;0 pand mm2, CLEARLOWBITS ;0 pand mm4, CLEARLOWBITS ;0 paddb mm3, mm5 ;0 (C+D) low paddb mm3, TWO ;0 (C+D+2) low = mm3 psrlq mm2, 2 ;0 paddb mm1, mm3 ;0 (A+B+C+D+2) low psrlq mm4, 2 ;0 paddb mm2, mm4 ;0 (C+D)/4 high = mm2 psrlq mm1, 2 ;0 (A+B+C+D+2)/4 low dirty paddb mm0, mm2 ;0 (A+B+C+D)/4 high pand mm1, EXTRACTLOWBITS ;0 (A+B+C+D+2)/4 low clean movq mm4, [ecx+2*PITCH] ;1 high(mm2) low(mm3) paddb mm0, mm1 ;0 1 movq mm1, [ecx+2*PITCH+1] ;1 C(mm4,mm5) D(mm0,mm1) movq mm5, mm4 ;1 movq [edx], mm0 ;0 movq mm0, mm1 ;1 pand mm0, CLEARLOWBITS ;1 pand mm4, CLEARLOWBITS ;1 psrlq mm0, 2 ;1 pand mm1, EXTRACTLOWBITS ;1 psrlq mm4, 2 ;1 pand mm5, EXTRACTLOWBITS ;1 paddb mm0, mm4 ;1 (C+D)/4 high = mm0 paddb mm1, mm5 ;1 (C+D) low paddb mm2, mm0 ;1 (A+B+C+D)/4 high paddb mm3, mm1 ;1 (A+B+C+D+2) low movq mm4, [ecx+3*PITCH] ;2 psrlq mm3, 2 ;1 (A+B+C+D+2)/4 low dirty movq mm5, mm4 ;2 high(mm0) low(mm1) pand mm3, EXTRACTLOWBITS ;1 2 paddb mm2, mm3 ;1 C(mm4,mm5) D(mm2,mm3) pand mm5, EXTRACTLOWBITS ;2 movq mm3, [ecx+3*PITCH+1] ;2 pand mm4, CLEARLOWBITS ;2 movq [edx+1*PITCH], mm2 ;1 movq mm2, mm3 ;2 pand mm3, EXTRACTLOWBITS ;2 pand mm2, CLEARLOWBITS ;2 psrlq mm4, 2 ;2 paddb mm3, mm5 ;2 paddb mm3, TWO ;2 (C+D+2) low = mm3 psrlq mm2, 2 ;2 paddb mm1, mm3 ;2 (A+B+C+D+2) low paddb mm2, mm4 ;2 (C+D)/4 hign = mm2 psrlq mm1, 2 ;2 (A+B+C+D+2)/4 low dirty paddb mm0, mm2 ;2 (A+B+C+D)/4 high movq mm4, [ecx+4*PITCH] ;3 high(mm2) low(mm3) pand mm1, EXTRACTLOWBITS ;2 3 movq mm5, mm4 ;3 C(mm4,mm5) D(mm0,mm1) paddb mm0, mm1 ;2 movq mm1, [ecx+4*PITCH+1] ;3 pand mm4, CLEARLOWBITS ;3 movq [edx+2*PITCH], mm0 ;2 movq mm0, mm1 ;3 pand mm0, CLEARLOWBITS ;3 pand mm1, EXTRACTLOWBITS ;3 psrlq mm0, 2 ;3 pand mm5, EXTRACTLOWBITS ;3 psrlq mm4, 2 ;3 paddb mm1, mm5 ;3 (C+D) low = mm1 paddb mm0, mm4 ;3 (C+D)/4 high = mm0 paddb mm3, mm1 ;3 (A+B+C+D+2) low paddb mm2, mm0 ;3 (A+B+C+D)/4 high psrlq mm3, 2 ;3 (A+B+C+D+2)/4 low dirty movq mm4, [ecx+5*PITCH] ;4 pand mm3, EXTRACTLOWBITS ;3 (A+B+C+D+2)/4 low clean movq mm5, mm4 ;4 paddb mm2, mm3 ;3 high(mm0) low(mm1) movq mm3, [ecx+5*PITCH+1] ;4 4 pand mm4, CLEARLOWBITS ;4 C(mm4,mm5) D(mm2,mm3) movq [edx+3*PITCH], mm2 ;3 movq mm2, mm3 ;4 pand mm2, CLEARLOWBITS ;4 pand mm5, EXTRACTLOWBITS ;4 psrlq mm4, 2 ;4 pand mm3, EXTRACTLOWBITS ;4 psrlq mm2, 2 ;4 paddb mm3, mm5 ;4 paddb mm3, TWO ;4 (C+D+2) low = mm3 paddb mm2, mm4 ;4 (C+D)/4 high = mm2 paddb mm1, mm3 ;4 (A+B+C+D+2) low paddb mm0, mm2 ;4 (A+B+C+D)/4 high movq mm4, [ecx+6*PITCH] ;5 psrlq mm1, 2 ;4 (A+B+C+D+2)/4 low dirty movq mm5, mm4 ;5 pand mm1, EXTRACTLOWBITS ;4 (A+B+C+D+2)/4 low clean paddb mm0, mm1 ;4 pand mm4, CLEARLOWBITS ;5 high(mm2) low(mm3) movq mm1, [ecx+6*PITCH+1] ;5 5 psrlq mm4, 2 ;5 C(mm4,mm5) D(mm0,mm1) movq [edx+4*PITCH], mm0 ;4 movq mm0, mm1 ;5 pand mm1, EXTRACTLOWBITS ;5 pand mm5, EXTRACTLOWBITS ;5 pand mm0, CLEARLOWBITS ;5 paddb mm1, mm5 ;5 (C+D) low = mm1 psrlq mm0, 2 ;5 paddb mm3, mm1 ;5 (A+B+C+D+2) low psrlq mm3, 2 ;5 (A+B+C+D+2)/4 low dirty paddb mm0, mm4 ;5 (C+D)/4 high = mm0 pand mm3, EXTRACTLOWBITS ;5 (A+B+C+D+2)/4 low clean paddb mm2, mm0 ;5 (A+B+C+D)/4 high movq mm4, [ecx+7*PITCH] ;6 high(mm0) low(mm1) paddb mm2, mm3 ;5 6 movq mm3, [ecx+7*PITCH+1] ;6 C(mm4,mm5) D(mm2,mm3) movq mm5, mm4 ;6 movq [edx+5*PITCH], mm2 ;5 movq mm2, mm3 ;6 pand mm5, EXTRACTLOWBITS ;6 pand mm3, EXTRACTLOWBITS ;6 pand mm2, CLEARLOWBITS ;6 paddb mm3, mm5 ;6 pand mm4, CLEARLOWBITS ;6 psrlq mm2, 2 ;6 paddb mm3, TWO ;6 (C+D+2) low = mm3 psrlq mm4, 2 ;6 paddb mm2, mm4 ;6 (C+D)/4 high = mm2 paddb mm1, mm3 ;6 (A+B+C+D+2) low paddb mm0, mm2 ;6 (A+B+C+D)/4 high psrlq mm1, 2 ;6 (A+B+C+D+2)/4 low dirty movq mm4, [ecx+8*PITCH] ;7 high(mm2) low(mm3) pand mm1, EXTRACTLOWBITS ;6 7 movq mm5, mm4 ;7 C(mm4,mm5) D(mm0,mm1) paddb mm0, mm1 ;6 movq mm1, [ecx+8*PITCH+1] ;7 pand mm4, CLEARLOWBITS ;7 movq [edx+6*PITCH], mm0 ;6 movq mm0, mm1 ;7 pand mm0, CLEARLOWBITS ;7 pand mm5, EXTRACTLOWBITS ;7 psrlq mm4, 2 ;7 pand mm1, EXTRACTLOWBITS ;7 psrlq mm0, 2 ;7 paddb mm1, mm5 ;7 (C+D) low paddb mm0, mm4 ;7 (C+D)/4 high paddb mm3, mm1 ;7 (A+B+C+D+2) low psrlq mm3, 2 ;7 (A+B+C+D+2)/4 low dirty paddb mm2, mm0 ;7 (A+B+C+D)/4 high pand mm3, EXTRACTLOWBITS ;7 (A+B+C+D+2)/4 low clean ; paddb mm2, mm3 ;7 ; ; ; movq [edx+7*PITCH], mm2 ;7 ret EXTRACTLOWBITS TEXTEQU <> CLEARLOWBITS TEXTEQU <>@MMX_Interpolate_Half_Half@8 endp
MMXCODE1 ENDS
END
|
