wine/gecko
0
0
Fork 0
mirror of https://gitlab.winehq.org/wine/wine-gecko.git synced 2024-09-13 09:24:08 -07:00
Code Issues Packages Projects Releases Wiki Activity
e6ec1a706e
gecko/media/libjpeg/simd/jfmmxfst.asm
Bobby Holley f21bd429e5 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

398 lines
11 KiB
NASM
Raw Blame History

;
; jfmmxfst.asm - fast integer FDCT (MMX)
;
; 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 fast, not so accurate integer implementation of
; the forward DCT (Discrete Cosine Transform). The following code is
; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
; for more details.
;
; [TAB8]
%include "jsimdext.inc"
%include "jdct.inc"
; --------------------------------------------------------------------------
%define CONST_BITS 8 ; 14 is also OK.
%if CONST_BITS == 8
F_0_382 equ 98 ; FIX(0.382683433)
F_0_541 equ 139 ; FIX(0.541196100)
F_0_707 equ 181 ; FIX(0.707106781)
F_1_306 equ 334 ; FIX(1.306562965)
%else
; NASM cannot do compile-time arithmetic on floating-point constants.
%define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
F_0_382 equ DESCALE( 410903207,30-CONST_BITS) ; FIX(0.382683433)
F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100)
F_0_707 equ DESCALE( 759250124,30-CONST_BITS) ; FIX(0.707106781)
F_1_306 equ DESCALE(1402911301,30-CONST_BITS) ; FIX(1.306562965)
%endif
; --------------------------------------------------------------------------
SECTION SEG_CONST
; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
%define PRE_MULTIPLY_SCALE_BITS 2
%define CONST_SHIFT (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
alignz 16
global EXTN(jconst_fdct_ifast_mmx)
EXTN(jconst_fdct_ifast_mmx):
PW_F0707 times 4 dw F_0_707 << CONST_SHIFT
PW_F0382 times 4 dw F_0_382 << CONST_SHIFT
PW_F0541 times 4 dw F_0_541 << CONST_SHIFT
PW_F1306 times 4 dw F_1_306 << CONST_SHIFT
alignz 16
; --------------------------------------------------------------------------
SECTION SEG_TEXT
BITS 32
;
; Perform the forward DCT on one block of samples.
;
; GLOBAL(void)
; jsimd_fdct_ifast_mmx (DCTELEM * data)
;
%define data(b) (b)+8 ; DCTELEM * 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_ifast_mmx)
EXTN(jsimd_fdct_ifast_mmx):
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)] ; (DCTELEM *)
mov ecx, DCTSIZE/4
alignx 16,7
.rowloop:
movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
movq mm2, MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)]
movq mm3, MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)]
; mm0=(20 21 22 23), mm2=(24 25 26 27)
; mm1=(30 31 32 33), mm3=(34 35 36 37)
movq mm4,mm0 ; transpose coefficients(phase 1)
punpcklwd mm0,mm1 ; mm0=(20 30 21 31)
punpckhwd mm4,mm1 ; mm4=(22 32 23 33)
movq mm5,mm2 ; transpose coefficients(phase 1)
punpcklwd mm2,mm3 ; mm2=(24 34 25 35)
punpckhwd mm5,mm3 ; mm5=(26 36 27 37)
movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)]
movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)]
; mm6=(00 01 02 03), mm1=(04 05 06 07)
; mm7=(10 11 12 13), mm3=(14 15 16 17)
movq MMWORD [wk(0)], mm4 ; wk(0)=(22 32 23 33)
movq MMWORD [wk(1)], mm2 ; wk(1)=(24 34 25 35)
movq mm4,mm6 ; transpose coefficients(phase 1)
punpcklwd mm6,mm7 ; mm6=(00 10 01 11)
punpckhwd mm4,mm7 ; mm4=(02 12 03 13)
movq mm2,mm1 ; transpose coefficients(phase 1)
punpcklwd mm1,mm3 ; mm1=(04 14 05 15)
punpckhwd mm2,mm3 ; mm2=(06 16 07 17)
movq mm7,mm6 ; transpose coefficients(phase 2)
punpckldq mm6,mm0 ; mm6=(00 10 20 30)=data0
punpckhdq mm7,mm0 ; mm7=(01 11 21 31)=data1
movq mm3,mm2 ; transpose coefficients(phase 2)
punpckldq mm2,mm5 ; mm2=(06 16 26 36)=data6
punpckhdq mm3,mm5 ; mm3=(07 17 27 37)=data7
movq mm0,mm7
movq mm5,mm6
psubw mm7,mm2 ; mm7=data1-data6=tmp6
psubw mm6,mm3 ; mm6=data0-data7=tmp7
paddw mm0,mm2 ; mm0=data1+data6=tmp1
paddw mm5,mm3 ; mm5=data0+data7=tmp0
movq mm2, MMWORD [wk(0)] ; mm2=(22 32 23 33)
movq mm3, MMWORD [wk(1)] ; mm3=(24 34 25 35)
movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
movq mm7,mm4 ; transpose coefficients(phase 2)
punpckldq mm4,mm2 ; mm4=(02 12 22 32)=data2
punpckhdq mm7,mm2 ; mm7=(03 13 23 33)=data3
movq mm6,mm1 ; transpose coefficients(phase 2)
punpckldq mm1,mm3 ; mm1=(04 14 24 34)=data4
punpckhdq mm6,mm3 ; mm6=(05 15 25 35)=data5
movq mm2,mm7
movq mm3,mm4
paddw mm7,mm1 ; mm7=data3+data4=tmp3
paddw mm4,mm6 ; mm4=data2+data5=tmp2
psubw mm2,mm1 ; mm2=data3-data4=tmp4
psubw mm3,mm6 ; mm3=data2-data5=tmp5
; -- Even part
movq mm1,mm5
movq mm6,mm0
psubw mm5,mm7 ; mm5=tmp13
psubw mm0,mm4 ; mm0=tmp12
paddw mm1,mm7 ; mm1=tmp10
paddw mm6,mm4 ; mm6=tmp11
paddw mm0,mm5
psllw mm0,PRE_MULTIPLY_SCALE_BITS
pmulhw mm0,[GOTOFF(ebx,PW_F0707)] ; mm0=z1
movq mm7,mm1
movq mm4,mm5
psubw mm1,mm6 ; mm1=data4
psubw mm5,mm0 ; mm5=data6
paddw mm7,mm6 ; mm7=data0
paddw mm4,mm0 ; mm4=data2
movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)], mm1
movq MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)], mm5
movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7
movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
; -- Odd part
movq mm6, MMWORD [wk(0)] ; mm6=tmp6
movq mm0, MMWORD [wk(1)] ; mm0=tmp7
paddw mm2,mm3 ; mm2=tmp10
paddw mm3,mm6 ; mm3=tmp11
paddw mm6,mm0 ; mm6=tmp12, mm0=tmp7
psllw mm2,PRE_MULTIPLY_SCALE_BITS
psllw mm6,PRE_MULTIPLY_SCALE_BITS
psllw mm3,PRE_MULTIPLY_SCALE_BITS
pmulhw mm3,[GOTOFF(ebx,PW_F0707)] ; mm3=z3
movq mm1,mm2 ; mm1=tmp10
psubw mm2,mm6
pmulhw mm2,[GOTOFF(ebx,PW_F0382)] ; mm2=z5
pmulhw mm1,[GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610)
pmulhw mm6,[GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296)
paddw mm1,mm2 ; mm1=z2
paddw mm6,mm2 ; mm6=z4
movq mm5,mm0
psubw mm0,mm3 ; mm0=z13
paddw mm5,mm3 ; mm5=z11
movq mm7,mm0
movq mm4,mm5
psubw mm0,mm1 ; mm0=data3
psubw mm5,mm6 ; mm5=data7
paddw mm7,mm1 ; mm7=data5
paddw mm4,mm6 ; mm4=data1
movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0
movq MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)], mm5
movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)], mm7
movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4
add edx, byte 4*DCTSIZE*SIZEOF_DCTELEM
dec ecx
jnz near .rowloop
; ---- Pass 2: process columns.
mov edx, POINTER [data(eax)] ; (DCTELEM *)
mov ecx, DCTSIZE/4
alignx 16,7
.columnloop:
movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
; mm0=(02 12 22 32), mm2=(42 52 62 72)
; mm1=(03 13 23 33), mm3=(43 53 63 73)
movq mm4,mm0 ; transpose coefficients(phase 1)
punpcklwd mm0,mm1 ; mm0=(02 03 12 13)
punpckhwd mm4,mm1 ; mm4=(22 23 32 33)
movq mm5,mm2 ; transpose coefficients(phase 1)
punpcklwd mm2,mm3 ; mm2=(42 43 52 53)
punpckhwd mm5,mm3 ; mm5=(62 63 72 73)
movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
movq mm1, MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
movq mm3, MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
; mm6=(00 10 20 30), mm1=(40 50 60 70)
; mm7=(01 11 21 31), mm3=(41 51 61 71)
movq MMWORD [wk(0)], mm4 ; wk(0)=(22 23 32 33)
movq MMWORD [wk(1)], mm2 ; wk(1)=(42 43 52 53)
movq mm4,mm6 ; transpose coefficients(phase 1)
punpcklwd mm6,mm7 ; mm6=(00 01 10 11)
punpckhwd mm4,mm7 ; mm4=(20 21 30 31)
movq mm2,mm1 ; transpose coefficients(phase 1)
punpcklwd mm1,mm3 ; mm1=(40 41 50 51)
punpckhwd mm2,mm3 ; mm2=(60 61 70 71)
movq mm7,mm6 ; transpose coefficients(phase 2)
punpckldq mm6,mm0 ; mm6=(00 01 02 03)=data0
punpckhdq mm7,mm0 ; mm7=(10 11 12 13)=data1
movq mm3,mm2 ; transpose coefficients(phase 2)
punpckldq mm2,mm5 ; mm2=(60 61 62 63)=data6
punpckhdq mm3,mm5 ; mm3=(70 71 72 73)=data7
movq mm0,mm7
movq mm5,mm6
psubw mm7,mm2 ; mm7=data1-data6=tmp6
psubw mm6,mm3 ; mm6=data0-data7=tmp7
paddw mm0,mm2 ; mm0=data1+data6=tmp1
paddw mm5,mm3 ; mm5=data0+data7=tmp0
movq mm2, MMWORD [wk(0)] ; mm2=(22 23 32 33)
movq mm3, MMWORD [wk(1)] ; mm3=(42 43 52 53)
movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
movq mm7,mm4 ; transpose coefficients(phase 2)
punpckldq mm4,mm2 ; mm4=(20 21 22 23)=data2
punpckhdq mm7,mm2 ; mm7=(30 31 32 33)=data3
movq mm6,mm1 ; transpose coefficients(phase 2)
punpckldq mm1,mm3 ; mm1=(40 41 42 43)=data4
punpckhdq mm6,mm3 ; mm6=(50 51 52 53)=data5
movq mm2,mm7
movq mm3,mm4
paddw mm7,mm1 ; mm7=data3+data4=tmp3
paddw mm4,mm6 ; mm4=data2+data5=tmp2
psubw mm2,mm1 ; mm2=data3-data4=tmp4
psubw mm3,mm6 ; mm3=data2-data5=tmp5
; -- Even part
movq mm1,mm5
movq mm6,mm0
psubw mm5,mm7 ; mm5=tmp13
psubw mm0,mm4 ; mm0=tmp12
paddw mm1,mm7 ; mm1=tmp10
paddw mm6,mm4 ; mm6=tmp11
paddw mm0,mm5
psllw mm0,PRE_MULTIPLY_SCALE_BITS
pmulhw mm0,[GOTOFF(ebx,PW_F0707)] ; mm0=z1
movq mm7,mm1
movq mm4,mm5
psubw mm1,mm6 ; mm1=data4
psubw mm5,mm0 ; mm5=data6
paddw mm7,mm6 ; mm7=data0
paddw mm4,mm0 ; mm4=data2
movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)], mm1
movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)], mm5
movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7
movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
; -- Odd part
movq mm6, MMWORD [wk(0)] ; mm6=tmp6
movq mm0, MMWORD [wk(1)] ; mm0=tmp7
paddw mm2,mm3 ; mm2=tmp10
paddw mm3,mm6 ; mm3=tmp11
paddw mm6,mm0 ; mm6=tmp12, mm0=tmp7
psllw mm2,PRE_MULTIPLY_SCALE_BITS
psllw mm6,PRE_MULTIPLY_SCALE_BITS
psllw mm3,PRE_MULTIPLY_SCALE_BITS
pmulhw mm3,[GOTOFF(ebx,PW_F0707)] ; mm3=z3
movq mm1,mm2 ; mm1=tmp10
psubw mm2,mm6
pmulhw mm2,[GOTOFF(ebx,PW_F0382)] ; mm2=z5
pmulhw mm1,[GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610)
pmulhw mm6,[GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296)
paddw mm1,mm2 ; mm1=z2
paddw mm6,mm2 ; mm6=z4
movq mm5,mm0
psubw mm0,mm3 ; mm0=z13
paddw mm5,mm3 ; mm5=z11
movq mm7,mm0
movq mm4,mm5
psubw mm0,mm1 ; mm0=data3
psubw mm5,mm6 ; mm5=data7
paddw mm7,mm1 ; mm7=data5
paddw mm4,mm6 ; mm4=data1
movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0
movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)], mm5
movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)], mm7
movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4
add edx, byte 4*SIZEOF_DCTELEM
dec ecx
jnz near .columnloop
emms ; empty MMX 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