|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; jdcolext.asm - colorspace conversion (MMX)
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB</ossman@cendio.se>
|
|
shun-iwasawa |
82a8f5 |
; Copyright (C) 2016, D. R. Commander.
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; Based on the x86 SIMD extension for IJG JPEG library
|
|
shun-iwasawa |
82a8f5 |
; Copyright (C) 1999-2006, MIYASAKA Masaru.
|
|
shun-iwasawa |
82a8f5 |
; For conditions of distribution and use, see copyright notice in jsimdext.inc
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; This file should be assembled with NASM (Netwide Assembler),
|
|
shun-iwasawa |
82a8f5 |
; can *not* be assembled with Microsoft's MASM or any compatible
|
|
shun-iwasawa |
82a8f5 |
; assembler (including Borland's Turbo Assembler).
|
|
shun-iwasawa |
82a8f5 |
; NASM is available from http://nasm.sourceforge.net/ or
|
|
shun-iwasawa |
82a8f5 |
; http://sourceforge.net/project/showfiles.php?group_id=6208
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%include "jcolsamp.inc"
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
; --------------------------------------------------------------------------
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; Convert some rows of samples to the output colorspace.
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; GLOBAL(void)
|
|
shun-iwasawa |
82a8f5 |
; jsimd_ycc_rgb_convert_mmx(JDIMENSION out_width, JSAMPIMAGE input_buf,
|
|
shun-iwasawa |
82a8f5 |
; JDIMENSION input_row, JSAMPARRAY output_buf,
|
|
shun-iwasawa |
82a8f5 |
; int num_rows)
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%define out_width(b) (b) + 8 ; JDIMENSION out_width
|
|
shun-iwasawa |
82a8f5 |
%define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf
|
|
shun-iwasawa |
82a8f5 |
%define input_row(b) (b) + 16 ; JDIMENSION input_row
|
|
shun-iwasawa |
82a8f5 |
%define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf
|
|
shun-iwasawa |
82a8f5 |
%define num_rows(b) (b) + 24 ; int num_rows
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%define original_ebp ebp + 0
|
|
shun-iwasawa |
82a8f5 |
%define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
; mmword wk[WK_NUM]
|
|
shun-iwasawa |
82a8f5 |
%define WK_NUM 2
|
|
shun-iwasawa |
82a8f5 |
%define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
align 32
|
|
shun-iwasawa |
82a8f5 |
GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_mmx)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
EXTN(jsimd_ycc_rgb_convert_mmx):
|
|
shun-iwasawa |
82a8f5 |
push ebp
|
|
shun-iwasawa |
82a8f5 |
mov eax, esp ; eax = original ebp
|
|
shun-iwasawa |
82a8f5 |
sub esp, byte 4
|
|
shun-iwasawa |
82a8f5 |
and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
|
|
shun-iwasawa |
82a8f5 |
mov [esp], eax
|
|
shun-iwasawa |
82a8f5 |
mov ebp, esp ; ebp = aligned ebp
|
|
shun-iwasawa |
82a8f5 |
lea esp, [wk(0)]
|
|
shun-iwasawa |
82a8f5 |
pushpic eax ; make a room for GOT address
|
|
shun-iwasawa |
82a8f5 |
push ebx
|
|
shun-iwasawa |
82a8f5 |
; push ecx ; need not be preserved
|
|
shun-iwasawa |
82a8f5 |
; push edx ; need not be preserved
|
|
shun-iwasawa |
82a8f5 |
push esi
|
|
shun-iwasawa |
82a8f5 |
push edi
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
get_GOT ebx ; get GOT address
|
|
shun-iwasawa |
82a8f5 |
movpic POINTER [gotptr], ebx ; save GOT address
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
mov ecx, JDIMENSION [out_width(eax)] ; num_cols
|
|
shun-iwasawa |
82a8f5 |
test ecx, ecx
|
|
shun-iwasawa |
82a8f5 |
jz near .return
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
push ecx
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
mov edi, JSAMPIMAGE [input_buf(eax)]
|
|
shun-iwasawa |
82a8f5 |
mov ecx, JDIMENSION [input_row(eax)]
|
|
shun-iwasawa |
82a8f5 |
mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
|
|
shun-iwasawa |
82a8f5 |
mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
|
|
shun-iwasawa |
82a8f5 |
mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
|
|
shun-iwasawa |
82a8f5 |
lea esi, [esi+ecx*SIZEOF_JSAMPROW]
|
|
shun-iwasawa |
82a8f5 |
lea ebx, [ebx+ecx*SIZEOF_JSAMPROW]
|
|
shun-iwasawa |
82a8f5 |
lea edx, [edx+ecx*SIZEOF_JSAMPROW]
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
pop ecx
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
mov edi, JSAMPARRAY [output_buf(eax)]
|
|
shun-iwasawa |
82a8f5 |
mov eax, INT [num_rows(eax)]
|
|
shun-iwasawa |
82a8f5 |
test eax, eax
|
|
shun-iwasawa |
82a8f5 |
jle near .return
|
|
shun-iwasawa |
82a8f5 |
alignx 16, 7
|
|
shun-iwasawa |
82a8f5 |
.rowloop:
|
|
shun-iwasawa |
82a8f5 |
push eax
|
|
shun-iwasawa |
82a8f5 |
push edi
|
|
shun-iwasawa |
82a8f5 |
push edx
|
|
shun-iwasawa |
82a8f5 |
push ebx
|
|
shun-iwasawa |
82a8f5 |
push esi
|
|
shun-iwasawa |
82a8f5 |
push ecx ; col
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
mov esi, JSAMPROW [esi] ; inptr0
|
|
shun-iwasawa |
82a8f5 |
mov ebx, JSAMPROW [ebx] ; inptr1
|
|
shun-iwasawa |
82a8f5 |
mov edx, JSAMPROW [edx] ; inptr2
|
|
shun-iwasawa |
82a8f5 |
mov edi, JSAMPROW [edi] ; outptr
|
|
shun-iwasawa |
82a8f5 |
movpic eax, POINTER [gotptr] ; load GOT address (eax)
|
|
shun-iwasawa |
82a8f5 |
alignx 16, 7
|
|
shun-iwasawa |
82a8f5 |
.columnloop:
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mm5, MMWORD [ebx] ; mm5=Cb(01234567)
|
|
shun-iwasawa |
82a8f5 |
movq mm1, MMWORD [edx] ; mm1=Cr(01234567)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
pcmpeqw mm4, mm4
|
|
shun-iwasawa |
82a8f5 |
pcmpeqw mm7, mm7
|
|
shun-iwasawa |
82a8f5 |
psrlw mm4, BYTE_BIT
|
|
shun-iwasawa |
82a8f5 |
psllw mm7, 7 ; mm7={0xFF80 0xFF80 0xFF80 0xFF80}
|
|
shun-iwasawa |
82a8f5 |
movq mm0, mm4 ; mm0=mm4={0xFF 0x00 0xFF 0x00 ..}
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
pand mm4, mm5 ; mm4=Cb(0246)=CbE
|
|
shun-iwasawa |
82a8f5 |
psrlw mm5, BYTE_BIT ; mm5=Cb(1357)=CbO
|
|
shun-iwasawa |
82a8f5 |
pand mm0, mm1 ; mm0=Cr(0246)=CrE
|
|
shun-iwasawa |
82a8f5 |
psrlw mm1, BYTE_BIT ; mm1=Cr(1357)=CrO
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
paddw mm4, mm7
|
|
shun-iwasawa |
82a8f5 |
paddw mm5, mm7
|
|
shun-iwasawa |
82a8f5 |
paddw mm0, mm7
|
|
shun-iwasawa |
82a8f5 |
paddw mm1, mm7
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
; (Original)
|
|
shun-iwasawa |
82a8f5 |
; R = Y + 1.40200 * Cr
|
|
shun-iwasawa |
82a8f5 |
; G = Y - 0.34414 * Cb - 0.71414 * Cr
|
|
shun-iwasawa |
82a8f5 |
; B = Y + 1.77200 * Cb
|
|
shun-iwasawa |
82a8f5 |
;
|
|
shun-iwasawa |
82a8f5 |
; (This implementation)
|
|
shun-iwasawa |
82a8f5 |
; R = Y + 0.40200 * Cr + Cr
|
|
shun-iwasawa |
82a8f5 |
; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
|
|
shun-iwasawa |
82a8f5 |
; B = Y - 0.22800 * Cb + Cb + Cb
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mm2, mm4 ; mm2=CbE
|
|
shun-iwasawa |
82a8f5 |
movq mm3, mm5 ; mm3=CbO
|
|
shun-iwasawa |
82a8f5 |
paddw mm4, mm4 ; mm4=2*CbE
|
|
shun-iwasawa |
82a8f5 |
paddw mm5, mm5 ; mm5=2*CbO
|
|
shun-iwasawa |
82a8f5 |
movq mm6, mm0 ; mm6=CrE
|
|
shun-iwasawa |
82a8f5 |
movq mm7, mm1 ; mm7=CrO
|
|
shun-iwasawa |
82a8f5 |
paddw mm0, mm0 ; mm0=2*CrE
|
|
shun-iwasawa |
82a8f5 |
paddw mm1, mm1 ; mm1=2*CrO
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
pmulhw mm4, [GOTOFF(eax,PW_MF0228)] ; mm4=(2*CbE * -FIX(0.22800))
|
|
shun-iwasawa |
82a8f5 |
pmulhw mm5, [GOTOFF(eax,PW_MF0228)] ; mm5=(2*CbO * -FIX(0.22800))
|
|
shun-iwasawa |
82a8f5 |
pmulhw mm0, [GOTOFF(eax,PW_F0402)] ; mm0=(2*CrE * FIX(0.40200))
|
|
shun-iwasawa |
82a8f5 |
pmulhw mm1, [GOTOFF(eax,PW_F0402)] ; mm1=(2*CrO * FIX(0.40200))
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
paddw mm4, [GOTOFF(eax,PW_ONE)]
|
|
shun-iwasawa |
82a8f5 |
paddw mm5, [GOTOFF(eax,PW_ONE)]
|
|
shun-iwasawa |
82a8f5 |
psraw mm4, 1 ; mm4=(CbE * -FIX(0.22800))
|
|
shun-iwasawa |
82a8f5 |
psraw mm5, 1 ; mm5=(CbO * -FIX(0.22800))
|
|
shun-iwasawa |
82a8f5 |
paddw mm0, [GOTOFF(eax,PW_ONE)]
|
|
shun-iwasawa |
82a8f5 |
paddw mm1, [GOTOFF(eax,PW_ONE)]
|
|
shun-iwasawa |
82a8f5 |
psraw mm0, 1 ; mm0=(CrE * FIX(0.40200))
|
|
shun-iwasawa |
82a8f5 |
psraw mm1, 1 ; mm1=(CrO * FIX(0.40200))
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
paddw mm4, mm2
|
|
shun-iwasawa |
82a8f5 |
paddw mm5, mm3
|
|
shun-iwasawa |
82a8f5 |
paddw mm4, mm2 ; mm4=(CbE * FIX(1.77200))=(B-Y)E
|
|
shun-iwasawa |
82a8f5 |
paddw mm5, mm3 ; mm5=(CbO * FIX(1.77200))=(B-Y)O
|
|
shun-iwasawa |
82a8f5 |
paddw mm0, mm6 ; mm0=(CrE * FIX(1.40200))=(R-Y)E
|
|
shun-iwasawa |
82a8f5 |
paddw mm1, mm7 ; mm1=(CrO * FIX(1.40200))=(R-Y)O
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [wk(0)], mm4 ; wk(0)=(B-Y)E
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [wk(1)], mm5 ; wk(1)=(B-Y)O
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mm4, mm2
|
|
shun-iwasawa |
82a8f5 |
movq mm5, mm3
|
|
shun-iwasawa |
82a8f5 |
punpcklwd mm2, mm6
|
|
shun-iwasawa |
82a8f5 |
punpckhwd mm4, mm6
|
|
shun-iwasawa |
82a8f5 |
pmaddwd mm2, [GOTOFF(eax,PW_MF0344_F0285)]
|
|
shun-iwasawa |
82a8f5 |
pmaddwd mm4, [GOTOFF(eax,PW_MF0344_F0285)]
|
|
shun-iwasawa |
82a8f5 |
punpcklwd mm3, mm7
|
|
shun-iwasawa |
82a8f5 |
punpckhwd mm5, mm7
|
|
shun-iwasawa |
82a8f5 |
pmaddwd mm3, [GOTOFF(eax,PW_MF0344_F0285)]
|
|
shun-iwasawa |
82a8f5 |
pmaddwd mm5, [GOTOFF(eax,PW_MF0344_F0285)]
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
paddd mm2, [GOTOFF(eax,PD_ONEHALF)]
|
|
shun-iwasawa |
82a8f5 |
paddd mm4, [GOTOFF(eax,PD_ONEHALF)]
|
|
shun-iwasawa |
82a8f5 |
psrad mm2, SCALEBITS
|
|
shun-iwasawa |
82a8f5 |
psrad mm4, SCALEBITS
|
|
shun-iwasawa |
82a8f5 |
paddd mm3, [GOTOFF(eax,PD_ONEHALF)]
|
|
shun-iwasawa |
82a8f5 |
paddd mm5, [GOTOFF(eax,PD_ONEHALF)]
|
|
shun-iwasawa |
82a8f5 |
psrad mm3, SCALEBITS
|
|
shun-iwasawa |
82a8f5 |
psrad mm5, SCALEBITS
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
packssdw mm2, mm4 ; mm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
|
|
shun-iwasawa |
82a8f5 |
packssdw mm3, mm5 ; mm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
|
|
shun-iwasawa |
82a8f5 |
psubw mm2, mm6 ; mm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
|
|
shun-iwasawa |
82a8f5 |
psubw mm3, mm7 ; mm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mm5, MMWORD [esi] ; mm5=Y(01234567)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
pcmpeqw mm4, mm4
|
|
shun-iwasawa |
82a8f5 |
psrlw mm4, BYTE_BIT ; mm4={0xFF 0x00 0xFF 0x00 ..}
|
|
shun-iwasawa |
82a8f5 |
pand mm4, mm5 ; mm4=Y(0246)=YE
|
|
shun-iwasawa |
82a8f5 |
psrlw mm5, BYTE_BIT ; mm5=Y(1357)=YO
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
paddw mm0, mm4 ; mm0=((R-Y)E+YE)=RE=(R0 R2 R4 R6)
|
|
shun-iwasawa |
82a8f5 |
paddw mm1, mm5 ; mm1=((R-Y)O+YO)=RO=(R1 R3 R5 R7)
|
|
shun-iwasawa |
82a8f5 |
packuswb mm0, mm0 ; mm0=(R0 R2 R4 R6 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
packuswb mm1, mm1 ; mm1=(R1 R3 R5 R7 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
paddw mm2, mm4 ; mm2=((G-Y)E+YE)=GE=(G0 G2 G4 G6)
|
|
shun-iwasawa |
82a8f5 |
paddw mm3, mm5 ; mm3=((G-Y)O+YO)=GO=(G1 G3 G5 G7)
|
|
shun-iwasawa |
82a8f5 |
packuswb mm2, mm2 ; mm2=(G0 G2 G4 G6 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
packuswb mm3, mm3 ; mm3=(G1 G3 G5 G7 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
paddw mm4, MMWORD [wk(0)] ; mm4=(YE+(B-Y)E)=BE=(B0 B2 B4 B6)
|
|
shun-iwasawa |
82a8f5 |
paddw mm5, MMWORD [wk(1)] ; mm5=(YO+(B-Y)O)=BO=(B1 B3 B5 B7)
|
|
shun-iwasawa |
82a8f5 |
packuswb mm4, mm4 ; mm4=(B0 B2 B4 B6 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
packuswb mm5, mm5 ; mm5=(B1 B3 B5 B7 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%if RGB_PIXELSIZE == 3 ; ---------------
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
punpcklbw mmA, mmC ; mmA=(00 10 02 12 04 14 06 16)
|
|
shun-iwasawa |
82a8f5 |
punpcklbw mmE, mmB ; mmE=(20 01 22 03 24 05 26 07)
|
|
shun-iwasawa |
82a8f5 |
punpcklbw mmD, mmF ; mmD=(11 21 13 23 15 25 17 27)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mmG, mmA
|
|
shun-iwasawa |
82a8f5 |
movq mmH, mmA
|
|
shun-iwasawa |
82a8f5 |
punpcklwd mmA, mmE ; mmA=(00 10 20 01 02 12 22 03)
|
|
shun-iwasawa |
82a8f5 |
punpckhwd mmG, mmE ; mmG=(04 14 24 05 06 16 26 07)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
psrlq mmH, 2*BYTE_BIT ; mmH=(02 12 04 14 06 16 -- --)
|
|
shun-iwasawa |
82a8f5 |
psrlq mmE, 2*BYTE_BIT ; mmE=(22 03 24 05 26 07 -- --)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mmC, mmD
|
|
shun-iwasawa |
82a8f5 |
movq mmB, mmD
|
|
shun-iwasawa |
82a8f5 |
punpcklwd mmD, mmH ; mmD=(11 21 02 12 13 23 04 14)
|
|
shun-iwasawa |
82a8f5 |
punpckhwd mmC, mmH ; mmC=(15 25 06 16 17 27 -- --)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
psrlq mmB, 2*BYTE_BIT ; mmB=(13 23 15 25 17 27 -- --)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mmF, mmE
|
|
shun-iwasawa |
82a8f5 |
punpcklwd mmE, mmB ; mmE=(22 03 13 23 24 05 15 25)
|
|
shun-iwasawa |
82a8f5 |
punpckhwd mmF, mmB ; mmF=(26 07 17 27 -- -- -- --)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
punpckldq mmA, mmD ; mmA=(00 10 20 01 11 21 02 12)
|
|
shun-iwasawa |
82a8f5 |
punpckldq mmE, mmG ; mmE=(22 03 13 23 04 14 24 05)
|
|
shun-iwasawa |
82a8f5 |
punpckldq mmC, mmF ; mmC=(15 25 06 16 26 07 17 27)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st16
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+0*SIZEOF_MMWORD], mmA
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+1*SIZEOF_MMWORD], mmE
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+2*SIZEOF_MMWORD], mmC
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
jz short .nextrow
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
add esi, byte SIZEOF_MMWORD ; inptr0
|
|
shun-iwasawa |
82a8f5 |
add ebx, byte SIZEOF_MMWORD ; inptr1
|
|
shun-iwasawa |
82a8f5 |
add edx, byte SIZEOF_MMWORD ; inptr2
|
|
shun-iwasawa |
82a8f5 |
add edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; outptr
|
|
shun-iwasawa |
82a8f5 |
jmp near .columnloop
|
|
shun-iwasawa |
82a8f5 |
alignx 16, 7
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
.column_st16:
|
|
shun-iwasawa |
82a8f5 |
lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte 2*SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st8
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+0*SIZEOF_MMWORD], mmA
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+1*SIZEOF_MMWORD], mmE
|
|
shun-iwasawa |
82a8f5 |
movq mmA, mmC
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte 2*SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
add edi, byte 2*SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
jmp short .column_st4
|
|
shun-iwasawa |
82a8f5 |
.column_st8:
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st4
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+0*SIZEOF_MMWORD], mmA
|
|
shun-iwasawa |
82a8f5 |
movq mmA, mmE
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
add edi, byte SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
.column_st4:
|
|
shun-iwasawa |
82a8f5 |
movd eax, mmA
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_DWORD
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st2
|
|
shun-iwasawa |
82a8f5 |
mov dword [edi+0*SIZEOF_DWORD], eax
|
|
shun-iwasawa |
82a8f5 |
psrlq mmA, DWORD_BIT
|
|
shun-iwasawa |
82a8f5 |
movd eax, mmA
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte SIZEOF_DWORD
|
|
shun-iwasawa |
82a8f5 |
add edi, byte SIZEOF_DWORD
|
|
shun-iwasawa |
82a8f5 |
.column_st2:
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_WORD
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st1
|
|
shun-iwasawa |
82a8f5 |
mov word [edi+0*SIZEOF_WORD], ax
|
|
shun-iwasawa |
82a8f5 |
shr eax, WORD_BIT
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte SIZEOF_WORD
|
|
shun-iwasawa |
82a8f5 |
add edi, byte SIZEOF_WORD
|
|
shun-iwasawa |
82a8f5 |
.column_st1:
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_BYTE
|
|
shun-iwasawa |
82a8f5 |
jb short .nextrow
|
|
shun-iwasawa |
82a8f5 |
mov byte [edi+0*SIZEOF_BYTE], al
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%else ; RGB_PIXELSIZE == 4 ; -----------
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%ifdef RGBX_FILLER_0XFF
|
|
shun-iwasawa |
82a8f5 |
pcmpeqb mm6, mm6 ; mm6=(X0 X2 X4 X6 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
pcmpeqb mm7, mm7 ; mm7=(X1 X3 X5 X7 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
%else
|
|
shun-iwasawa |
82a8f5 |
pxor mm6, mm6 ; mm6=(X0 X2 X4 X6 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
pxor mm7, mm7 ; mm7=(X1 X3 X5 X7 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
%endif
|
|
shun-iwasawa |
82a8f5 |
; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
punpcklbw mmA, mmC ; mmA=(00 10 02 12 04 14 06 16)
|
|
shun-iwasawa |
82a8f5 |
punpcklbw mmE, mmG ; mmE=(20 30 22 32 24 34 26 36)
|
|
shun-iwasawa |
82a8f5 |
punpcklbw mmB, mmD ; mmB=(01 11 03 13 05 15 07 17)
|
|
shun-iwasawa |
82a8f5 |
punpcklbw mmF, mmH ; mmF=(21 31 23 33 25 35 27 37)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mmC, mmA
|
|
shun-iwasawa |
82a8f5 |
punpcklwd mmA, mmE ; mmA=(00 10 20 30 02 12 22 32)
|
|
shun-iwasawa |
82a8f5 |
punpckhwd mmC, mmE ; mmC=(04 14 24 34 06 16 26 36)
|
|
shun-iwasawa |
82a8f5 |
movq mmG, mmB
|
|
shun-iwasawa |
82a8f5 |
punpcklwd mmB, mmF ; mmB=(01 11 21 31 03 13 23 33)
|
|
shun-iwasawa |
82a8f5 |
punpckhwd mmG, mmF ; mmG=(05 15 25 35 07 17 27 37)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq mmD, mmA
|
|
shun-iwasawa |
82a8f5 |
punpckldq mmA, mmB ; mmA=(00 10 20 30 01 11 21 31)
|
|
shun-iwasawa |
82a8f5 |
punpckhdq mmD, mmB ; mmD=(02 12 22 32 03 13 23 33)
|
|
shun-iwasawa |
82a8f5 |
movq mmH, mmC
|
|
shun-iwasawa |
82a8f5 |
punpckldq mmC, mmG ; mmC=(04 14 24 34 05 15 25 35)
|
|
shun-iwasawa |
82a8f5 |
punpckhdq mmH, mmG ; mmH=(06 16 26 36 07 17 27 37)
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st16
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+0*SIZEOF_MMWORD], mmA
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+1*SIZEOF_MMWORD], mmD
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+2*SIZEOF_MMWORD], mmC
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+3*SIZEOF_MMWORD], mmH
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
jz short .nextrow
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
add esi, byte SIZEOF_MMWORD ; inptr0
|
|
shun-iwasawa |
82a8f5 |
add ebx, byte SIZEOF_MMWORD ; inptr1
|
|
shun-iwasawa |
82a8f5 |
add edx, byte SIZEOF_MMWORD ; inptr2
|
|
shun-iwasawa |
82a8f5 |
add edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; outptr
|
|
shun-iwasawa |
82a8f5 |
jmp near .columnloop
|
|
shun-iwasawa |
82a8f5 |
alignx 16, 7
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
.column_st16:
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_MMWORD/2
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st8
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+0*SIZEOF_MMWORD], mmA
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+1*SIZEOF_MMWORD], mmD
|
|
shun-iwasawa |
82a8f5 |
movq mmA, mmC
|
|
shun-iwasawa |
82a8f5 |
movq mmD, mmH
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte SIZEOF_MMWORD/2
|
|
shun-iwasawa |
82a8f5 |
add edi, byte 2*SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
.column_st8:
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_MMWORD/4
|
|
shun-iwasawa |
82a8f5 |
jb short .column_st4
|
|
shun-iwasawa |
82a8f5 |
movq MMWORD [edi+0*SIZEOF_MMWORD], mmA
|
|
shun-iwasawa |
82a8f5 |
movq mmA, mmD
|
|
shun-iwasawa |
82a8f5 |
sub ecx, byte SIZEOF_MMWORD/4
|
|
shun-iwasawa |
82a8f5 |
add edi, byte 1*SIZEOF_MMWORD
|
|
shun-iwasawa |
82a8f5 |
.column_st4:
|
|
shun-iwasawa |
82a8f5 |
cmp ecx, byte SIZEOF_MMWORD/8
|
|
shun-iwasawa |
82a8f5 |
jb short .nextrow
|
|
shun-iwasawa |
82a8f5 |
movd dword [edi+0*SIZEOF_DWORD], mmA
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
%endif ; RGB_PIXELSIZE ; ---------------
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
alignx 16, 7
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
.nextrow:
|
|
shun-iwasawa |
82a8f5 |
pop ecx
|
|
shun-iwasawa |
82a8f5 |
pop esi
|
|
shun-iwasawa |
82a8f5 |
pop ebx
|
|
shun-iwasawa |
82a8f5 |
pop edx
|
|
shun-iwasawa |
82a8f5 |
pop edi
|
|
shun-iwasawa |
82a8f5 |
pop eax
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
add esi, byte SIZEOF_JSAMPROW
|
|
shun-iwasawa |
82a8f5 |
add ebx, byte SIZEOF_JSAMPROW
|
|
shun-iwasawa |
82a8f5 |
add edx, byte SIZEOF_JSAMPROW
|
|
shun-iwasawa |
82a8f5 |
add edi, byte SIZEOF_JSAMPROW ; output_buf
|
|
shun-iwasawa |
82a8f5 |
dec eax ; num_rows
|
|
shun-iwasawa |
82a8f5 |
jg near .rowloop
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
emms ; empty MMX state
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
.return:
|
|
shun-iwasawa |
82a8f5 |
pop edi
|
|
shun-iwasawa |
82a8f5 |
pop esi
|
|
shun-iwasawa |
82a8f5 |
; pop edx ; need not be preserved
|
|
shun-iwasawa |
82a8f5 |
; pop ecx ; need not be preserved
|
|
shun-iwasawa |
82a8f5 |
pop ebx
|
|
shun-iwasawa |
82a8f5 |
mov esp, ebp ; esp <- aligned ebp
|
|
shun-iwasawa |
82a8f5 |
pop esp ; esp <- original ebp
|
|
shun-iwasawa |
82a8f5 |
pop ebp
|
|
shun-iwasawa |
82a8f5 |
ret
|
|
shun-iwasawa |
82a8f5 |
|
|
shun-iwasawa |
82a8f5 |
; For some reason, the OS X linker does not honor the request to align the
|
|
shun-iwasawa |
82a8f5 |
; segment unless we do this.
|
|
shun-iwasawa |
82a8f5 |
align 32
|