Magic-Mirror/gecko-dev
1
0
Fork 0
mirror of https://github.com/mozilla/gecko-dev.git synced 2024-11-07 04:05:49 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity
4a62decf3b
gecko-dev/media/libjpeg/simd/jf3dnflt.asm
Bobby Holley 2d0b90011e Bug 584894 - Move libjpeg to media/libjpeg. r=khuey 
--HG--
rename : jpeg/MOZCHANGES => media/libjpeg/MOZCHANGES
rename : jpeg/Makefile.in => media/libjpeg/Makefile.in
rename : jpeg/README => media/libjpeg/README
rename : jpeg/README-turbo.txt => media/libjpeg/README-turbo.txt
rename : jpeg/cderror.h => media/libjpeg/cderror.h
rename : jpeg/cdjpeg.h => media/libjpeg/cdjpeg.h
rename : jpeg/jaricom.c => media/libjpeg/jaricom.c
rename : jpeg/jcapimin.c => media/libjpeg/jcapimin.c
rename : jpeg/jcapistd.c => media/libjpeg/jcapistd.c
rename : jpeg/jcarith.c => media/libjpeg/jcarith.c
rename : jpeg/jccoefct.c => media/libjpeg/jccoefct.c
rename : jpeg/jccolor.c => media/libjpeg/jccolor.c
rename : jpeg/jcdctmgr.c => media/libjpeg/jcdctmgr.c
rename : jpeg/jchuff.c => media/libjpeg/jchuff.c
rename : jpeg/jchuff.h => media/libjpeg/jchuff.h
rename : jpeg/jcinit.c => media/libjpeg/jcinit.c
rename : jpeg/jcmainct.c => media/libjpeg/jcmainct.c
rename : jpeg/jcmarker.c => media/libjpeg/jcmarker.c
rename : jpeg/jcmaster.c => media/libjpeg/jcmaster.c
rename : jpeg/jcomapi.c => media/libjpeg/jcomapi.c
rename : jpeg/jconfig.h => media/libjpeg/jconfig.h
rename : jpeg/jconfig.h.in => media/libjpeg/jconfig.h.in
rename : jpeg/jcparam.c => media/libjpeg/jcparam.c
rename : jpeg/jcphuff.c => media/libjpeg/jcphuff.c
rename : jpeg/jcprepct.c => media/libjpeg/jcprepct.c
rename : jpeg/jcsample.c => media/libjpeg/jcsample.c
rename : jpeg/jdapimin.c => media/libjpeg/jdapimin.c
rename : jpeg/jdapistd.c => media/libjpeg/jdapistd.c
rename : jpeg/jdarith.c => media/libjpeg/jdarith.c
rename : jpeg/jdatadst.c => media/libjpeg/jdatadst.c
rename : jpeg/jdatasrc.c => media/libjpeg/jdatasrc.c
rename : jpeg/jdcoefct.c => media/libjpeg/jdcoefct.c
rename : jpeg/jdcolor.c => media/libjpeg/jdcolor.c
rename : jpeg/jdct.h => media/libjpeg/jdct.h
rename : jpeg/jddctmgr.c => media/libjpeg/jddctmgr.c
rename : jpeg/jdhuff.c => media/libjpeg/jdhuff.c
rename : jpeg/jdhuff.h => media/libjpeg/jdhuff.h
rename : jpeg/jdinput.c => media/libjpeg/jdinput.c
rename : jpeg/jdmainct.c => media/libjpeg/jdmainct.c
rename : jpeg/jdmarker.c => media/libjpeg/jdmarker.c
rename : jpeg/jdmaster.c => media/libjpeg/jdmaster.c
rename : jpeg/jdmerge.c => media/libjpeg/jdmerge.c
rename : jpeg/jdphuff.c => media/libjpeg/jdphuff.c
rename : jpeg/jdpostct.c => media/libjpeg/jdpostct.c
rename : jpeg/jdsample.c => media/libjpeg/jdsample.c
rename : jpeg/jdtrans.c => media/libjpeg/jdtrans.c
rename : jpeg/jerror.c => media/libjpeg/jerror.c
rename : jpeg/jerror.h => media/libjpeg/jerror.h
rename : jpeg/jfdctflt.c => media/libjpeg/jfdctflt.c
rename : jpeg/jfdctfst.c => media/libjpeg/jfdctfst.c
rename : jpeg/jfdctint.c => media/libjpeg/jfdctint.c
rename : jpeg/jidctflt.c => media/libjpeg/jidctflt.c
rename : jpeg/jidctfst.c => media/libjpeg/jidctfst.c
rename : jpeg/jidctint.c => media/libjpeg/jidctint.c
rename : jpeg/jidctred.c => media/libjpeg/jidctred.c
rename : jpeg/jinclude.h => media/libjpeg/jinclude.h
rename : jpeg/jmemmgr.c => media/libjpeg/jmemmgr.c
rename : jpeg/jmemnobs.c => media/libjpeg/jmemnobs.c
rename : jpeg/jmemsys.h => media/libjpeg/jmemsys.h
rename : jpeg/jmorecfg.h => media/libjpeg/jmorecfg.h
rename : jpeg/jpegcomp.h => media/libjpeg/jpegcomp.h
rename : jpeg/jpegint.h => media/libjpeg/jpegint.h
rename : jpeg/jpeglib.h => media/libjpeg/jpeglib.h
rename : jpeg/jquant1.c => media/libjpeg/jquant1.c
rename : jpeg/jquant2.c => media/libjpeg/jquant2.c
rename : jpeg/jsimd.h => media/libjpeg/jsimd.h
rename : jpeg/jsimd_none.c => media/libjpeg/jsimd_none.c
rename : jpeg/jsimddct.h => media/libjpeg/jsimddct.h
rename : jpeg/jutils.c => media/libjpeg/jutils.c
rename : jpeg/jversion.h => media/libjpeg/jversion.h
rename : jpeg/simd/Makefile.in => media/libjpeg/simd/Makefile.in
rename : jpeg/simd/jcclrmmx.asm => media/libjpeg/simd/jcclrmmx.asm
rename : jpeg/simd/jcclrss2-64.asm => media/libjpeg/simd/jcclrss2-64.asm
rename : jpeg/simd/jcclrss2.asm => media/libjpeg/simd/jcclrss2.asm
rename : jpeg/simd/jccolmmx.asm => media/libjpeg/simd/jccolmmx.asm
rename : jpeg/simd/jccolss2-64.asm => media/libjpeg/simd/jccolss2-64.asm
rename : jpeg/simd/jccolss2.asm => media/libjpeg/simd/jccolss2.asm
rename : jpeg/simd/jcolsamp.inc => media/libjpeg/simd/jcolsamp.inc
rename : jpeg/simd/jcqnt3dn.asm => media/libjpeg/simd/jcqnt3dn.asm
rename : jpeg/simd/jcqntmmx.asm => media/libjpeg/simd/jcqntmmx.asm
rename : jpeg/simd/jcqnts2f-64.asm => media/libjpeg/simd/jcqnts2f-64.asm
rename : jpeg/simd/jcqnts2f.asm => media/libjpeg/simd/jcqnts2f.asm
rename : jpeg/simd/jcqnts2i-64.asm => media/libjpeg/simd/jcqnts2i-64.asm
rename : jpeg/simd/jcqnts2i.asm => media/libjpeg/simd/jcqnts2i.asm
rename : jpeg/simd/jcqntsse.asm => media/libjpeg/simd/jcqntsse.asm
rename : jpeg/simd/jcsammmx.asm => media/libjpeg/simd/jcsammmx.asm
rename : jpeg/simd/jcsamss2-64.asm => media/libjpeg/simd/jcsamss2-64.asm
rename : jpeg/simd/jcsamss2.asm => media/libjpeg/simd/jcsamss2.asm
rename : jpeg/simd/jdclrmmx.asm => media/libjpeg/simd/jdclrmmx.asm
rename : jpeg/simd/jdclrss2-64.asm => media/libjpeg/simd/jdclrss2-64.asm
rename : jpeg/simd/jdclrss2.asm => media/libjpeg/simd/jdclrss2.asm
rename : jpeg/simd/jdcolmmx.asm => media/libjpeg/simd/jdcolmmx.asm
rename : jpeg/simd/jdcolss2-64.asm => media/libjpeg/simd/jdcolss2-64.asm
rename : jpeg/simd/jdcolss2.asm => media/libjpeg/simd/jdcolss2.asm
rename : jpeg/simd/jdct.inc => media/libjpeg/simd/jdct.inc
rename : jpeg/simd/jdmermmx.asm => media/libjpeg/simd/jdmermmx.asm
rename : jpeg/simd/jdmerss2-64.asm => media/libjpeg/simd/jdmerss2-64.asm
rename : jpeg/simd/jdmerss2.asm => media/libjpeg/simd/jdmerss2.asm
rename : jpeg/simd/jdmrgmmx.asm => media/libjpeg/simd/jdmrgmmx.asm
rename : jpeg/simd/jdmrgss2-64.asm => media/libjpeg/simd/jdmrgss2-64.asm
rename : jpeg/simd/jdmrgss2.asm => media/libjpeg/simd/jdmrgss2.asm
rename : jpeg/simd/jdsammmx.asm => media/libjpeg/simd/jdsammmx.asm
rename : jpeg/simd/jdsamss2-64.asm => media/libjpeg/simd/jdsamss2-64.asm
rename : jpeg/simd/jdsamss2.asm => media/libjpeg/simd/jdsamss2.asm
rename : jpeg/simd/jf3dnflt.asm => media/libjpeg/simd/jf3dnflt.asm
rename : jpeg/simd/jfmmxfst.asm => media/libjpeg/simd/jfmmxfst.asm
rename : jpeg/simd/jfmmxint.asm => media/libjpeg/simd/jfmmxint.asm
rename : jpeg/simd/jfss2fst-64.asm => media/libjpeg/simd/jfss2fst-64.asm
rename : jpeg/simd/jfss2fst.asm => media/libjpeg/simd/jfss2fst.asm
rename : jpeg/simd/jfss2int-64.asm => media/libjpeg/simd/jfss2int-64.asm
rename : jpeg/simd/jfss2int.asm => media/libjpeg/simd/jfss2int.asm
rename : jpeg/simd/jfsseflt-64.asm => media/libjpeg/simd/jfsseflt-64.asm
rename : jpeg/simd/jfsseflt.asm => media/libjpeg/simd/jfsseflt.asm
rename : jpeg/simd/ji3dnflt.asm => media/libjpeg/simd/ji3dnflt.asm
rename : jpeg/simd/jimmxfst.asm => media/libjpeg/simd/jimmxfst.asm
rename : jpeg/simd/jimmxint.asm => media/libjpeg/simd/jimmxint.asm
rename : jpeg/simd/jimmxred.asm => media/libjpeg/simd/jimmxred.asm
rename : jpeg/simd/jiss2flt-64.asm => media/libjpeg/simd/jiss2flt-64.asm
rename : jpeg/simd/jiss2flt.asm => media/libjpeg/simd/jiss2flt.asm
rename : jpeg/simd/jiss2fst-64.asm => media/libjpeg/simd/jiss2fst-64.asm
rename : jpeg/simd/jiss2fst.asm => media/libjpeg/simd/jiss2fst.asm
rename : jpeg/simd/jiss2int-64.asm => media/libjpeg/simd/jiss2int-64.asm
rename : jpeg/simd/jiss2int.asm => media/libjpeg/simd/jiss2int.asm
rename : jpeg/simd/jiss2red-64.asm => media/libjpeg/simd/jiss2red-64.asm
rename : jpeg/simd/jiss2red.asm => media/libjpeg/simd/jiss2red.asm
rename : jpeg/simd/jisseflt.asm => media/libjpeg/simd/jisseflt.asm
rename : jpeg/simd/jsimd.h => media/libjpeg/simd/jsimd.h
rename : jpeg/simd/jsimd_i386.c => media/libjpeg/simd/jsimd_i386.c
rename : jpeg/simd/jsimd_x86_64.c => media/libjpeg/simd/jsimd_x86_64.c
rename : jpeg/simd/jsimdcfg.inc => media/libjpeg/simd/jsimdcfg.inc
rename : jpeg/simd/jsimdcfg.inc.h => media/libjpeg/simd/jsimdcfg.inc.h
rename : jpeg/simd/jsimdcpu.asm => media/libjpeg/simd/jsimdcpu.asm
rename : jpeg/simd/jsimdext.inc => media/libjpeg/simd/jsimdext.inc
rename : jpeg/transupp.h => media/libjpeg/transupp.h
2011-10-17 21:25:53 -07:00

321 lines
8.7 KiB
NASM
Raw Blame History

;
; jf3dnflt.asm - floating-point FDCT (3DNow!)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
;
; Based on
; x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
; For conditions of distribution and use, see copyright notice in jsimdext.inc
;
; This file should be assembled with NASM (Netwide Assembler),
; can *not* be assembled with Microsoft's MASM or any compatible
; assembler (including Borland's Turbo Assembler).
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208
;
; This file contains a floating-point implementation of the forward DCT
; (Discrete Cosine Transform). The following code is based directly on
; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
;
; [TAB8]
%include "jsimdext.inc"
%include "jdct.inc"
; --------------------------------------------------------------------------
SECTION SEG_CONST
alignz 16
global EXTN(jconst_fdct_float_3dnow)
EXTN(jconst_fdct_float_3dnow):
PD_0_382 times 2 dd 0.382683432365089771728460
PD_0_707 times 2 dd 0.707106781186547524400844
PD_0_541 times 2 dd 0.541196100146196984399723
PD_1_306 times 2 dd 1.306562964876376527856643
alignz 16
; --------------------------------------------------------------------------
SECTION SEG_TEXT
BITS 32
;
; Perform the forward DCT on one block of samples.
;
; GLOBAL(void)
; jsimd_fdct_float_3dnow (FAST_FLOAT * data)
;
%define data(b) (b)+8 ; FAST_FLOAT * data
%define original_ebp ebp+0
%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
%define WK_NUM 2
align 16
global EXTN(jsimd_fdct_float_3dnow)
EXTN(jsimd_fdct_float_3dnow):
push ebp
mov eax,esp ; eax = original ebp
sub esp, byte 4
and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
mov [esp],eax
mov ebp,esp ; ebp = aligned ebp
lea esp, [wk(0)]
pushpic ebx
; push ecx ; need not be preserved
; push edx ; need not be preserved
; push esi ; unused
; push edi ; unused
get_GOT ebx ; get GOT address
; ---- Pass 1: process rows.
mov edx, POINTER [data(eax)] ; (FAST_FLOAT *)
mov ecx, DCTSIZE/2
alignx 16,7
.rowloop:
movq mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
movq mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
movq mm2, MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)]
movq mm3, MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)]
; mm0=(00 01), mm1=(10 11), mm2=(06 07), mm3=(16 17)
movq mm4,mm0 ; transpose coefficients
punpckldq mm0,mm1 ; mm0=(00 10)=data0
punpckhdq mm4,mm1 ; mm4=(01 11)=data1
movq mm5,mm2 ; transpose coefficients
punpckldq mm2,mm3 ; mm2=(06 16)=data6
punpckhdq mm5,mm3 ; mm5=(07 17)=data7
movq mm6,mm4
movq mm7,mm0
pfsub mm4,mm2 ; mm4=data1-data6=tmp6
pfsub mm0,mm5 ; mm0=data0-data7=tmp7
pfadd mm6,mm2 ; mm6=data1+data6=tmp1
pfadd mm7,mm5 ; mm7=data0+data7=tmp0
movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
movq mm2, MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)]
movq mm5, MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)]
; mm1=(02 03), mm3=(12 13), mm2=(04 05), mm5=(14 15)
movq MMWORD [wk(0)], mm4 ; wk(0)=tmp6
movq MMWORD [wk(1)], mm0 ; wk(1)=tmp7
movq mm4,mm1 ; transpose coefficients
punpckldq mm1,mm3 ; mm1=(02 12)=data2
punpckhdq mm4,mm3 ; mm4=(03 13)=data3
movq mm0,mm2 ; transpose coefficients
punpckldq mm2,mm5 ; mm2=(04 14)=data4
punpckhdq mm0,mm5 ; mm0=(05 15)=data5
movq mm3,mm4
movq mm5,mm1
pfadd mm4,mm2 ; mm4=data3+data4=tmp3
pfadd mm1,mm0 ; mm1=data2+data5=tmp2
pfsub mm3,mm2 ; mm3=data3-data4=tmp4
pfsub mm5,mm0 ; mm5=data2-data5=tmp5
; -- Even part
movq mm2,mm7
movq mm0,mm6
pfsub mm7,mm4 ; mm7=tmp13
pfsub mm6,mm1 ; mm6=tmp12
pfadd mm2,mm4 ; mm2=tmp10
pfadd mm0,mm1 ; mm0=tmp11
pfadd mm6,mm7
pfmul mm6,[GOTOFF(ebx,PD_0_707)] ; mm6=z1
movq mm4,mm2
movq mm1,mm7
pfsub mm2,mm0 ; mm2=data4
pfsub mm7,mm6 ; mm7=data6
pfadd mm4,mm0 ; mm4=data0
pfadd mm1,mm6 ; mm1=data2
movq MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)], mm2
movq MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)], mm7
movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4
movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], mm1
; -- Odd part
movq mm0, MMWORD [wk(0)] ; mm0=tmp6
movq mm6, MMWORD [wk(1)] ; mm6=tmp7
pfadd mm3,mm5 ; mm3=tmp10
pfadd mm5,mm0 ; mm5=tmp11
pfadd mm0,mm6 ; mm0=tmp12, mm6=tmp7
pfmul mm5,[GOTOFF(ebx,PD_0_707)] ; mm5=z3
movq mm2,mm3 ; mm2=tmp10
pfsub mm3,mm0
pfmul mm3,[GOTOFF(ebx,PD_0_382)] ; mm3=z5
pfmul mm2,[GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610)
pfmul mm0,[GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296)
pfadd mm2,mm3 ; mm2=z2
pfadd mm0,mm3 ; mm0=z4
movq mm7,mm6
pfsub mm6,mm5 ; mm6=z13
pfadd mm7,mm5 ; mm7=z11
movq mm4,mm6
movq mm1,mm7
pfsub mm6,mm2 ; mm6=data3
pfsub mm7,mm0 ; mm7=data7
pfadd mm4,mm2 ; mm4=data5
pfadd mm1,mm0 ; mm1=data1
movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], mm6
movq MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)], mm7
movq MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)], mm4
movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1
add edx, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
dec ecx
jnz near .rowloop
; ---- Pass 2: process columns.
mov edx, POINTER [data(eax)] ; (FAST_FLOAT *)
mov ecx, DCTSIZE/2
alignx 16,7
.columnloop:
movq mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
movq mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)]
movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)]
; mm0=(00 10), mm1=(01 11), mm2=(60 70), mm3=(61 71)
movq mm4,mm0 ; transpose coefficients
punpckldq mm0,mm1 ; mm0=(00 01)=data0
punpckhdq mm4,mm1 ; mm4=(10 11)=data1
movq mm5,mm2 ; transpose coefficients
punpckldq mm2,mm3 ; mm2=(60 61)=data6
punpckhdq mm5,mm3 ; mm5=(70 71)=data7
movq mm6,mm4
movq mm7,mm0
pfsub mm4,mm2 ; mm4=data1-data6=tmp6
pfsub mm0,mm5 ; mm0=data0-data7=tmp7
pfadd mm6,mm2 ; mm6=data1+data6=tmp1
pfadd mm7,mm5 ; mm7=data0+data7=tmp0
movq mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
movq mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
movq mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)]
movq mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)]
; mm1=(20 30), mm3=(21 31), mm2=(40 50), mm5=(41 51)
movq MMWORD [wk(0)], mm4 ; wk(0)=tmp6
movq MMWORD [wk(1)], mm0 ; wk(1)=tmp7
movq mm4,mm1 ; transpose coefficients
punpckldq mm1,mm3 ; mm1=(20 21)=data2
punpckhdq mm4,mm3 ; mm4=(30 31)=data3
movq mm0,mm2 ; transpose coefficients
punpckldq mm2,mm5 ; mm2=(40 41)=data4
punpckhdq mm0,mm5 ; mm0=(50 51)=data5
movq mm3,mm4
movq mm5,mm1
pfadd mm4,mm2 ; mm4=data3+data4=tmp3
pfadd mm1,mm0 ; mm1=data2+data5=tmp2
pfsub mm3,mm2 ; mm3=data3-data4=tmp4
pfsub mm5,mm0 ; mm5=data2-data5=tmp5
; -- Even part
movq mm2,mm7
movq mm0,mm6
pfsub mm7,mm4 ; mm7=tmp13
pfsub mm6,mm1 ; mm6=tmp12
pfadd mm2,mm4 ; mm2=tmp10
pfadd mm0,mm1 ; mm0=tmp11
pfadd mm6,mm7
pfmul mm6,[GOTOFF(ebx,PD_0_707)] ; mm6=z1
movq mm4,mm2
movq mm1,mm7
pfsub mm2,mm0 ; mm2=data4
pfsub mm7,mm6 ; mm7=data6
pfadd mm4,mm0 ; mm4=data0
pfadd mm1,mm6 ; mm1=data2
movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], mm2
movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], mm7
movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4
movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], mm1
; -- Odd part
movq mm0, MMWORD [wk(0)] ; mm0=tmp6
movq mm6, MMWORD [wk(1)] ; mm6=tmp7
pfadd mm3,mm5 ; mm3=tmp10
pfadd mm5,mm0 ; mm5=tmp11
pfadd mm0,mm6 ; mm0=tmp12, mm6=tmp7
pfmul mm5,[GOTOFF(ebx,PD_0_707)] ; mm5=z3
movq mm2,mm3 ; mm2=tmp10
pfsub mm3,mm0
pfmul mm3,[GOTOFF(ebx,PD_0_382)] ; mm3=z5
pfmul mm2,[GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610)
pfmul mm0,[GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296)
pfadd mm2,mm3 ; mm2=z2
pfadd mm0,mm3 ; mm0=z4
movq mm7,mm6
pfsub mm6,mm5 ; mm6=z13
pfadd mm7,mm5 ; mm7=z11
movq mm4,mm6
movq mm1,mm7
pfsub mm6,mm2 ; mm6=data3
pfsub mm7,mm0 ; mm7=data7
pfadd mm4,mm2 ; mm4=data5
pfadd mm1,mm0 ; mm1=data1
movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], mm6
movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], mm7
movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], mm4
movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1
add edx, byte 2*SIZEOF_FAST_FLOAT
dec ecx
jnz near .columnloop
femms ; empty MMX/3DNow! state
; pop edi ; unused
; pop esi ; unused
; pop edx ; need not be preserved
; pop ecx ; need not be preserved
poppic ebx
mov esp,ebp ; esp <- aligned ebp
pop esp ; esp <- original ebp
pop ebp
ret
; For some reason, the OS X linker does not honor the request to align the
; segment unless we do this.
align 16
Reference in New Issue View Git Blame Copy Permalink