gecko/media/libjpeg/simd/jcqntsse.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

212 lines
5.7 KiB
NASM

;
; jcqntsse.asm - sample data conversion and quantization (SSE & 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
;
; [TAB8]
%include "jsimdext.inc"
%include "jdct.inc"
; --------------------------------------------------------------------------
SECTION SEG_TEXT
BITS 32
;
; Load data into workspace, applying unsigned->signed conversion
;
; GLOBAL(void)
; jsimd_convsamp_float_sse (JSAMPARRAY sample_data, JDIMENSION start_col,
; FAST_FLOAT * workspace);
;
%define sample_data ebp+8 ; JSAMPARRAY sample_data
%define start_col ebp+12 ; JDIMENSION start_col
%define workspace ebp+16 ; FAST_FLOAT * workspace
align 16
global EXTN(jsimd_convsamp_float_sse)
EXTN(jsimd_convsamp_float_sse):
push ebp
mov ebp,esp
push ebx
; push ecx ; need not be preserved
; push edx ; need not be preserved
push esi
push edi
pcmpeqw mm7,mm7
psllw mm7,7
packsswb mm7,mm7 ; mm7 = PB_CENTERJSAMPLE (0x808080..)
mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *)
mov eax, JDIMENSION [start_col]
mov edi, POINTER [workspace] ; (DCTELEM *)
mov ecx, DCTSIZE/2
alignx 16,7
.convloop:
mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
movq mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE]
movq mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE]
psubb mm0,mm7 ; mm0=(01234567)
psubb mm1,mm7 ; mm1=(89ABCDEF)
punpcklbw mm2,mm0 ; mm2=(*0*1*2*3)
punpckhbw mm0,mm0 ; mm0=(*4*5*6*7)
punpcklbw mm3,mm1 ; mm3=(*8*9*A*B)
punpckhbw mm1,mm1 ; mm1=(*C*D*E*F)
punpcklwd mm4,mm2 ; mm4=(***0***1)
punpckhwd mm2,mm2 ; mm2=(***2***3)
punpcklwd mm5,mm0 ; mm5=(***4***5)
punpckhwd mm0,mm0 ; mm0=(***6***7)
psrad mm4,(DWORD_BIT-BYTE_BIT) ; mm4=(01)
psrad mm2,(DWORD_BIT-BYTE_BIT) ; mm2=(23)
cvtpi2ps xmm0,mm4 ; xmm0=(01**)
cvtpi2ps xmm1,mm2 ; xmm1=(23**)
psrad mm5,(DWORD_BIT-BYTE_BIT) ; mm5=(45)
psrad mm0,(DWORD_BIT-BYTE_BIT) ; mm0=(67)
cvtpi2ps xmm2,mm5 ; xmm2=(45**)
cvtpi2ps xmm3,mm0 ; xmm3=(67**)
punpcklwd mm6,mm3 ; mm6=(***8***9)
punpckhwd mm3,mm3 ; mm3=(***A***B)
punpcklwd mm4,mm1 ; mm4=(***C***D)
punpckhwd mm1,mm1 ; mm1=(***E***F)
psrad mm6,(DWORD_BIT-BYTE_BIT) ; mm6=(89)
psrad mm3,(DWORD_BIT-BYTE_BIT) ; mm3=(AB)
cvtpi2ps xmm4,mm6 ; xmm4=(89**)
cvtpi2ps xmm5,mm3 ; xmm5=(AB**)
psrad mm4,(DWORD_BIT-BYTE_BIT) ; mm4=(CD)
psrad mm1,(DWORD_BIT-BYTE_BIT) ; mm1=(EF)
cvtpi2ps xmm6,mm4 ; xmm6=(CD**)
cvtpi2ps xmm7,mm1 ; xmm7=(EF**)
movlhps xmm0,xmm1 ; xmm0=(0123)
movlhps xmm2,xmm3 ; xmm2=(4567)
movlhps xmm4,xmm5 ; xmm4=(89AB)
movlhps xmm6,xmm7 ; xmm6=(CDEF)
movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm2
movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm4
movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
add esi, byte 2*SIZEOF_JSAMPROW
add edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
dec ecx
jnz near .convloop
emms ; empty MMX state
pop edi
pop esi
; pop edx ; need not be preserved
; pop ecx ; need not be preserved
pop ebx
pop ebp
ret
; --------------------------------------------------------------------------
;
; Quantize/descale the coefficients, and store into coef_block
;
; GLOBAL(void)
; jsimd_quantize_float_sse (JCOEFPTR coef_block, FAST_FLOAT * divisors,
; FAST_FLOAT * workspace);
;
%define coef_block ebp+8 ; JCOEFPTR coef_block
%define divisors ebp+12 ; FAST_FLOAT * divisors
%define workspace ebp+16 ; FAST_FLOAT * workspace
align 16
global EXTN(jsimd_quantize_float_sse)
EXTN(jsimd_quantize_float_sse):
push ebp
mov ebp,esp
; push ebx ; unused
; push ecx ; unused
; push edx ; need not be preserved
push esi
push edi
mov esi, POINTER [workspace]
mov edx, POINTER [divisors]
mov edi, JCOEFPTR [coef_block]
mov eax, DCTSIZE2/16
alignx 16,7
.quantloop:
movaps xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
movaps xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)]
mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
mulps xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
movaps xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
movaps xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)]
mulps xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
mulps xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
movhlps xmm4,xmm0
movhlps xmm5,xmm1
cvtps2pi mm0,xmm0
cvtps2pi mm1,xmm1
cvtps2pi mm4,xmm4
cvtps2pi mm5,xmm5
movhlps xmm6,xmm2
movhlps xmm7,xmm3
cvtps2pi mm2,xmm2
cvtps2pi mm3,xmm3
cvtps2pi mm6,xmm6
cvtps2pi mm7,xmm7
packssdw mm0,mm4
packssdw mm1,mm5
packssdw mm2,mm6
packssdw mm3,mm7
movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1
movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm2
movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm3
add esi, byte 16*SIZEOF_FAST_FLOAT
add edx, byte 16*SIZEOF_FAST_FLOAT
add edi, byte 16*SIZEOF_JCOEF
dec eax
jnz short .quantloop
emms ; empty MMX state
pop edi
pop esi
; pop edx ; need not be preserved
; pop ecx ; unused
; pop ebx ; unused
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