gecko/media/libjpeg/simd/jdclrss2.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

506 lines
16 KiB
NASM

;
; jdclrss2.asm - colorspace conversion (SSE2)
;
; 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
;
; [TAB8]
%include "jcolsamp.inc"
; --------------------------------------------------------------------------
SECTION SEG_TEXT
BITS 32
;
; Convert some rows of samples to the output colorspace.
;
; GLOBAL(void)
; jsimd_ycc_rgb_convert_sse2 (JDIMENSION out_width,
; JSAMPIMAGE input_buf, JDIMENSION input_row,
; JSAMPARRAY output_buf, int num_rows)
;
%define out_width(b) (b)+8 ; JDIMENSION out_width
%define input_buf(b) (b)+12 ; JSAMPIMAGE input_buf
%define input_row(b) (b)+16 ; JDIMENSION input_row
%define output_buf(b) (b)+20 ; JSAMPARRAY output_buf
%define num_rows(b) (b)+24 ; int num_rows
%define original_ebp ebp+0
%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
%define WK_NUM 2
%define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr
align 16
global EXTN(jsimd_ycc_rgb_convert_sse2)
EXTN(jsimd_ycc_rgb_convert_sse2):
push ebp
mov eax,esp ; eax = original ebp
sub esp, byte 4
and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
mov [esp],eax
mov ebp,esp ; ebp = aligned ebp
lea esp, [wk(0)]
pushpic eax ; make a room for GOT address
push ebx
; push ecx ; need not be preserved
; push edx ; need not be preserved
push esi
push edi
get_GOT ebx ; get GOT address
movpic POINTER [gotptr], ebx ; save GOT address
mov ecx, JDIMENSION [out_width(eax)] ; num_cols
test ecx,ecx
jz near .return
push ecx
mov edi, JSAMPIMAGE [input_buf(eax)]
mov ecx, JDIMENSION [input_row(eax)]
mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
lea esi, [esi+ecx*SIZEOF_JSAMPROW]
lea ebx, [ebx+ecx*SIZEOF_JSAMPROW]
lea edx, [edx+ecx*SIZEOF_JSAMPROW]
pop ecx
mov edi, JSAMPARRAY [output_buf(eax)]
mov eax, INT [num_rows(eax)]
test eax,eax
jle near .return
alignx 16,7
.rowloop:
push eax
push edi
push edx
push ebx
push esi
push ecx ; col
mov esi, JSAMPROW [esi] ; inptr0
mov ebx, JSAMPROW [ebx] ; inptr1
mov edx, JSAMPROW [edx] ; inptr2
mov edi, JSAMPROW [edi] ; outptr
movpic eax, POINTER [gotptr] ; load GOT address (eax)
alignx 16,7
.columnloop:
movdqa xmm5, XMMWORD [ebx] ; xmm5=Cb(0123456789ABCDEF)
movdqa xmm1, XMMWORD [edx] ; xmm1=Cr(0123456789ABCDEF)
pcmpeqw xmm4,xmm4
pcmpeqw xmm7,xmm7
psrlw xmm4,BYTE_BIT
psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
movdqa xmm0,xmm4 ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
pand xmm4,xmm5 ; xmm4=Cb(02468ACE)=CbE
psrlw xmm5,BYTE_BIT ; xmm5=Cb(13579BDF)=CbO
pand xmm0,xmm1 ; xmm0=Cr(02468ACE)=CrE
psrlw xmm1,BYTE_BIT ; xmm1=Cr(13579BDF)=CrO
paddw xmm4,xmm7
paddw xmm5,xmm7
paddw xmm0,xmm7
paddw xmm1,xmm7
; (Original)
; R = Y + 1.40200 * Cr
; G = Y - 0.34414 * Cb - 0.71414 * Cr
; B = Y + 1.77200 * Cb
;
; (This implementation)
; R = Y + 0.40200 * Cr + Cr
; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
; B = Y - 0.22800 * Cb + Cb + Cb
movdqa xmm2,xmm4 ; xmm2=CbE
movdqa xmm3,xmm5 ; xmm3=CbO
paddw xmm4,xmm4 ; xmm4=2*CbE
paddw xmm5,xmm5 ; xmm5=2*CbO
movdqa xmm6,xmm0 ; xmm6=CrE
movdqa xmm7,xmm1 ; xmm7=CrO
paddw xmm0,xmm0 ; xmm0=2*CrE
paddw xmm1,xmm1 ; xmm1=2*CrO
pmulhw xmm4,[GOTOFF(eax,PW_MF0228)] ; xmm4=(2*CbE * -FIX(0.22800))
pmulhw xmm5,[GOTOFF(eax,PW_MF0228)] ; xmm5=(2*CbO * -FIX(0.22800))
pmulhw xmm0,[GOTOFF(eax,PW_F0402)] ; xmm0=(2*CrE * FIX(0.40200))
pmulhw xmm1,[GOTOFF(eax,PW_F0402)] ; xmm1=(2*CrO * FIX(0.40200))
paddw xmm4,[GOTOFF(eax,PW_ONE)]
paddw xmm5,[GOTOFF(eax,PW_ONE)]
psraw xmm4,1 ; xmm4=(CbE * -FIX(0.22800))
psraw xmm5,1 ; xmm5=(CbO * -FIX(0.22800))
paddw xmm0,[GOTOFF(eax,PW_ONE)]
paddw xmm1,[GOTOFF(eax,PW_ONE)]
psraw xmm0,1 ; xmm0=(CrE * FIX(0.40200))
psraw xmm1,1 ; xmm1=(CrO * FIX(0.40200))
paddw xmm4,xmm2
paddw xmm5,xmm3
paddw xmm4,xmm2 ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
paddw xmm5,xmm3 ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
paddw xmm0,xmm6 ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
paddw xmm1,xmm7 ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=(B-Y)E
movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(B-Y)O
movdqa xmm4,xmm2
movdqa xmm5,xmm3
punpcklwd xmm2,xmm6
punpckhwd xmm4,xmm6
pmaddwd xmm2,[GOTOFF(eax,PW_MF0344_F0285)]
pmaddwd xmm4,[GOTOFF(eax,PW_MF0344_F0285)]
punpcklwd xmm3,xmm7
punpckhwd xmm5,xmm7
pmaddwd xmm3,[GOTOFF(eax,PW_MF0344_F0285)]
pmaddwd xmm5,[GOTOFF(eax,PW_MF0344_F0285)]
paddd xmm2,[GOTOFF(eax,PD_ONEHALF)]
paddd xmm4,[GOTOFF(eax,PD_ONEHALF)]
psrad xmm2,SCALEBITS
psrad xmm4,SCALEBITS
paddd xmm3,[GOTOFF(eax,PD_ONEHALF)]
paddd xmm5,[GOTOFF(eax,PD_ONEHALF)]
psrad xmm3,SCALEBITS
psrad xmm5,SCALEBITS
packssdw xmm2,xmm4 ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
packssdw xmm3,xmm5 ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
psubw xmm2,xmm6 ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
psubw xmm3,xmm7 ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
movdqa xmm5, XMMWORD [esi] ; xmm5=Y(0123456789ABCDEF)
pcmpeqw xmm4,xmm4
psrlw xmm4,BYTE_BIT ; xmm4={0xFF 0x00 0xFF 0x00 ..}
pand xmm4,xmm5 ; xmm4=Y(02468ACE)=YE
psrlw xmm5,BYTE_BIT ; xmm5=Y(13579BDF)=YO
paddw xmm0,xmm4 ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
paddw xmm1,xmm5 ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
packuswb xmm0,xmm0 ; xmm0=R(02468ACE********)
packuswb xmm1,xmm1 ; xmm1=R(13579BDF********)
paddw xmm2,xmm4 ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
paddw xmm3,xmm5 ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
packuswb xmm2,xmm2 ; xmm2=G(02468ACE********)
packuswb xmm3,xmm3 ; xmm3=G(13579BDF********)
paddw xmm4, XMMWORD [wk(0)] ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
paddw xmm5, XMMWORD [wk(1)] ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
packuswb xmm4,xmm4 ; xmm4=B(02468ACE********)
packuswb xmm5,xmm5 ; xmm5=B(13579BDF********)
%if RGB_PIXELSIZE == 3 ; ---------------
; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
punpcklbw xmmE,xmmB ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
punpcklbw xmmD,xmmF ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
movdqa xmmG,xmmA
movdqa xmmH,xmmA
punpcklwd xmmA,xmmE ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
punpckhwd xmmG,xmmE ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
psrldq xmmH,2 ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
psrldq xmmE,2 ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
movdqa xmmC,xmmD
movdqa xmmB,xmmD
punpcklwd xmmD,xmmH ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
punpckhwd xmmC,xmmH ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
psrldq xmmB,2 ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
movdqa xmmF,xmmE
punpcklwd xmmE,xmmB ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
punpckhwd xmmF,xmmB ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
pshufd xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
movdqa xmmB,xmmE
punpckldq xmmA,xmmD ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
punpckldq xmmE,xmmH ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
punpckhdq xmmD,xmmB ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
pshufd xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
movdqa xmmB,xmmF
punpckldq xmmG,xmmC ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
punpckldq xmmF,xmmH ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
punpckhdq xmmC,xmmB ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
punpcklqdq xmmA,xmmE ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
punpcklqdq xmmD,xmmG ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
punpcklqdq xmmF,xmmC ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
cmp ecx, byte SIZEOF_XMMWORD
jb short .column_st32
test edi, SIZEOF_XMMWORD-1
jnz short .out1
; --(aligned)-------------------
movntdq XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
movntdq XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
movntdq XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF
add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
jmp short .out0
.out1: ; --(unaligned)-----------------
pcmpeqb xmmH,xmmH ; xmmH=(all 1's)
maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [edi], xmmA
add edi, byte SIZEOF_XMMWORD ; outptr
maskmovdqu xmmD,xmmH ; movntdqu XMMWORD [edi], xmmD
add edi, byte SIZEOF_XMMWORD ; outptr
maskmovdqu xmmF,xmmH ; movntdqu XMMWORD [edi], xmmF
add edi, byte SIZEOF_XMMWORD ; outptr
.out0:
sub ecx, byte SIZEOF_XMMWORD
jz near .nextrow
add esi, byte SIZEOF_XMMWORD ; inptr0
add ebx, byte SIZEOF_XMMWORD ; inptr1
add edx, byte SIZEOF_XMMWORD ; inptr2
jmp near .columnloop
alignx 16,7
.column_st32:
pcmpeqb xmmH,xmmH ; xmmH=(all 1's)
lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE
cmp ecx, byte 2*SIZEOF_XMMWORD
jb short .column_st16
maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [edi], xmmA
add edi, byte SIZEOF_XMMWORD ; outptr
maskmovdqu xmmD,xmmH ; movntdqu XMMWORD [edi], xmmD
add edi, byte SIZEOF_XMMWORD ; outptr
movdqa xmmA,xmmF
sub ecx, byte 2*SIZEOF_XMMWORD
jmp short .column_st15
.column_st16:
cmp ecx, byte SIZEOF_XMMWORD
jb short .column_st15
maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [edi], xmmA
add edi, byte SIZEOF_XMMWORD ; outptr
movdqa xmmA,xmmD
sub ecx, byte SIZEOF_XMMWORD
.column_st15:
mov eax,ecx
xor ecx, byte 0x0F
shl ecx, 2
movd xmmB,ecx
psrlq xmmH,4
pcmpeqb xmmE,xmmE
psrlq xmmH,xmmB
psrlq xmmE,xmmB
punpcklbw xmmE,xmmH
; ----------------
mov ecx,edi
and ecx, byte SIZEOF_XMMWORD-1
jz short .adj0
add eax,ecx
cmp eax, byte SIZEOF_XMMWORD
ja short .adj0
and edi, byte (-SIZEOF_XMMWORD) ; align to 16-byte boundary
shl ecx, 3 ; pslldq xmmA,ecx & pslldq xmmE,ecx
movdqa xmmG,xmmA
movdqa xmmC,xmmE
pslldq xmmA, SIZEOF_XMMWORD/2
pslldq xmmE, SIZEOF_XMMWORD/2
movd xmmD,ecx
sub ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
jb short .adj1
movd xmmF,ecx
psllq xmmA,xmmF
psllq xmmE,xmmF
jmp short .adj0
.adj1: neg ecx
movd xmmF,ecx
psrlq xmmA,xmmF
psrlq xmmE,xmmF
psllq xmmG,xmmD
psllq xmmC,xmmD
por xmmA,xmmG
por xmmE,xmmC
.adj0: ; ----------------
maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA
%else ; RGB_PIXELSIZE == 4 ; -----------
%ifdef RGBX_FILLER_0XFF
pcmpeqb xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
pcmpeqb xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
%else
pxor xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
pxor xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
%endif
; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
punpcklbw xmmE,xmmG ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
punpcklbw xmmB,xmmD ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
punpcklbw xmmF,xmmH ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
movdqa xmmC,xmmA
punpcklwd xmmA,xmmE ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
punpckhwd xmmC,xmmE ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
movdqa xmmG,xmmB
punpcklwd xmmB,xmmF ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
punpckhwd xmmG,xmmF ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
movdqa xmmD,xmmA
punpckldq xmmA,xmmB ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
punpckhdq xmmD,xmmB ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
movdqa xmmH,xmmC
punpckldq xmmC,xmmG ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
punpckhdq xmmH,xmmG ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
cmp ecx, byte SIZEOF_XMMWORD
jb short .column_st32
test edi, SIZEOF_XMMWORD-1
jnz short .out1
; --(aligned)-------------------
movntdq XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
movntdq XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
movntdq XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC
movntdq XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH
add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
jmp short .out0
.out1: ; --(unaligned)-----------------
pcmpeqb xmmE,xmmE ; xmmE=(all 1's)
maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA
add edi, byte SIZEOF_XMMWORD ; outptr
maskmovdqu xmmD,xmmE ; movntdqu XMMWORD [edi], xmmD
add edi, byte SIZEOF_XMMWORD ; outptr
maskmovdqu xmmC,xmmE ; movntdqu XMMWORD [edi], xmmC
add edi, byte SIZEOF_XMMWORD ; outptr
maskmovdqu xmmH,xmmE ; movntdqu XMMWORD [edi], xmmH
add edi, byte SIZEOF_XMMWORD ; outptr
.out0:
sub ecx, byte SIZEOF_XMMWORD
jz near .nextrow
add esi, byte SIZEOF_XMMWORD ; inptr0
add ebx, byte SIZEOF_XMMWORD ; inptr1
add edx, byte SIZEOF_XMMWORD ; inptr2
jmp near .columnloop
alignx 16,7
.column_st32:
pcmpeqb xmmE,xmmE ; xmmE=(all 1's)
cmp ecx, byte SIZEOF_XMMWORD/2
jb short .column_st16
maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA
add edi, byte SIZEOF_XMMWORD ; outptr
maskmovdqu xmmD,xmmE ; movntdqu XMMWORD [edi], xmmD
add edi, byte SIZEOF_XMMWORD ; outptr
movdqa xmmA,xmmC
movdqa xmmD,xmmH
sub ecx, byte SIZEOF_XMMWORD/2
.column_st16:
cmp ecx, byte SIZEOF_XMMWORD/4
jb short .column_st15
maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA
add edi, byte SIZEOF_XMMWORD ; outptr
movdqa xmmA,xmmD
sub ecx, byte SIZEOF_XMMWORD/4
.column_st15:
cmp ecx, byte SIZEOF_XMMWORD/16
jb short .nextrow
mov eax,ecx
xor ecx, byte 0x03
inc ecx
shl ecx, 4
movd xmmF,ecx
psrlq xmmE,xmmF
punpcklbw xmmE,xmmE
; ----------------
mov ecx,edi
and ecx, byte SIZEOF_XMMWORD-1
jz short .adj0
lea eax, [ecx+eax*4] ; RGB_PIXELSIZE
cmp eax, byte SIZEOF_XMMWORD
ja short .adj0
and edi, byte (-SIZEOF_XMMWORD) ; align to 16-byte boundary
shl ecx, 3 ; pslldq xmmA,ecx & pslldq xmmE,ecx
movdqa xmmB,xmmA
movdqa xmmG,xmmE
pslldq xmmA, SIZEOF_XMMWORD/2
pslldq xmmE, SIZEOF_XMMWORD/2
movd xmmC,ecx
sub ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
jb short .adj1
movd xmmH,ecx
psllq xmmA,xmmH
psllq xmmE,xmmH
jmp short .adj0
.adj1: neg ecx
movd xmmH,ecx
psrlq xmmA,xmmH
psrlq xmmE,xmmH
psllq xmmB,xmmC
psllq xmmG,xmmC
por xmmA,xmmB
por xmmE,xmmG
.adj0: ; ----------------
maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA
%endif ; RGB_PIXELSIZE ; ---------------
alignx 16,7
.nextrow:
pop ecx
pop esi
pop ebx
pop edx
pop edi
pop eax
add esi, byte SIZEOF_JSAMPROW
add ebx, byte SIZEOF_JSAMPROW
add edx, byte SIZEOF_JSAMPROW
add edi, byte SIZEOF_JSAMPROW ; output_buf
dec eax ; num_rows
jg near .rowloop
sfence ; flush the write buffer
.return:
pop edi
pop esi
; pop edx ; need not be preserved
; pop ecx ; need not be preserved
pop 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