📄 2xsaimmx.inc
字号:
pand mm1, mm3
por mm3, mm4
pxor mm7, mm7
pand mm2, mm4
pcmpeqw mm3, mm7
por mm1, mm2
pand mm0, mm3
por mm0, mm1
punpcklwd mm5, mm0
punpckhwd mm6, mm0
%ifdef __DJGPP__
movq [fs:edx], mm5
movq [fs:edx+8], mm6
%else
movq [es:edx], mm5
movq [es:edx+8], mm6
%endif
;------------------------------------------------
; Create the Nextline
;------------------------------------------------
;3 ;if ((colorA == colorD) && (colorB != colorC) && (colorA == colorG) && (colorC == colorO)
movq mm0, [eax+ebx+colorA] ;mm0 and mm1 contain colorA
movq mm2, [eax+ebx+ebx+colorC] ;mm2 and mm3 contain colorC
movq mm1, mm0
movq mm3, mm2
push eax
add eax, ebx
pcmpeqw mm0, [eax+ebx+colorD]
pcmpeqw mm1, [eax+colorG]
pcmpeqw mm2, [eax+ebx+ebx+colorO]
pcmpeqw mm3, [eax+colorB]
pop eax
pand mm0, mm1
pxor mm1, mm1
pand mm0, mm2
pcmpeqw mm3, mm1
pand mm0, mm3 ;result in mm0
;if ((colorA == colorB) && (colorG != colorC) && (colorA == colorH) && (colorC == colorM)
movq mm4, [eax+ebx+colorA] ;mm4 and mm5 contain colorA
movq mm6, [eax+ebx+ebx+colorC] ;mm6 and mm7 contain colorC
movq mm5, mm4
movq mm7, mm6
push eax
add eax, ebx
pcmpeqw mm4, [eax+ebx+colorH]
pcmpeqw mm5, [eax+colorB]
pcmpeqw mm6, [eax+ebx+ebx+colorM]
pcmpeqw mm7, [eax+colorG]
pop eax
pand mm4, mm5
pxor mm5, mm5
pand mm4, mm6
pcmpeqw mm7, mm5
pand mm4, mm7 ;result in mm4
por mm0, mm4 ;combine the masks
movq [Mask1], mm0
;--------------------------------------------
;4
;if ((colorB == colorC) && (colorA != colorD) && (colorC == colorH) && (colorA == colorF)
movq mm0, [eax+ebx+ebx+colorC] ;mm0 and mm1 contain colorC
movq mm2, [eax+ebx+colorA] ;mm2 and mm3 contain colorA
movq mm1, mm0
movq mm3, mm2
pcmpeqw mm0, [eax+ebx+colorB]
pcmpeqw mm1, [eax+ebx+ebx+colorH]
pcmpeqw mm2, [eax+colorF]
pcmpeqw mm3, [eax+ebx+ebx+colorD]
pand mm0, mm1
pxor mm1, mm1
pand mm0, mm2
pcmpeqw mm3, mm1
pand mm0, mm3 ;result in mm0
;if ((colorC == colorG) && (colorC == colorD) && (colorA != colorH) && (colorA == colorI)
movq mm4, [eax+ebx+ebx+colorC] ;mm4 and mm5 contain colorC
movq mm6, [eax+ebx+colorA] ;mm6 and mm7 contain colorA
movq mm5, mm4
movq mm7, mm6
pcmpeqw mm4, [eax+ebx+ebx+colorD]
pcmpeqw mm5, [eax+ebx+colorG]
pcmpeqw mm6, [eax+colorI]
pcmpeqw mm7, [eax+ebx+ebx+colorH]
pand mm4, mm5
pxor mm5, mm5
pand mm4, mm6
pcmpeqw mm7, mm5
pand mm4, mm7 ;result in mm4
por mm0, mm4 ;combine the masks
movq [Mask2], mm0
;----------------------------------------------
;interpolate colorA and colorC
movq mm0, [eax+ebx+colorA]
movq mm1, [eax+ebx+ebx+colorC]
movq mm2, mm0
movq mm3, mm1
pand mm0, [colorMask]
pand mm1, [colorMask]
psrlw mm0, 1
psrlw mm1, 1
pand mm3, [lowPixelMask]
paddw mm0, mm1
pand mm3, mm2
paddw mm0, mm3 ;mm0 contains the interpolated values
;-------------
;assemble the pixels
movq mm1, [eax+ebx+colorA]
movq mm2, [eax+ebx+ebx+colorC]
movq mm3, [Mask1]
movq mm4, [Mask2]
pand mm1, mm3
pand mm2, mm4
por mm3, mm4
pxor mm7, mm7
por mm1, mm2
pcmpeqw mm3, mm7
pand mm0, mm3
por mm0, mm1
movq [ACPixel], mm0
;////////////////////////////////
; Decide which "branch" to take
;--------------------------------
movq mm0, [eax+ebx+colorA]
movq mm1, [eax+ebx+colorB]
movq mm6, mm0
movq mm7, mm1
pcmpeqw mm0, [eax+ebx+ebx+colorD]
pcmpeqw mm1, [eax+ebx+ebx+colorC]
pcmpeqw mm6, mm7
movq mm2, mm0
movq mm3, mm0
pand mm0, mm1 ;colorA == colorD && colorB == colorC
pxor mm7, mm7
pcmpeqw mm2, mm7
pand mm6, mm0
pand mm2, mm1 ;colorA != colorD && colorB == colorC
pcmpeqw mm1, mm7
pand mm1, mm3 ;colorA == colorD && colorB != colorC
pxor mm0, mm6
por mm1, mm6
movq mm7, mm0
movq [Mask2], mm2
packsswb mm7, mm7
movq [Mask1], mm1
movd ecx, mm7
test ecx, ecx
jz near .SKIP_GUESS
;---------------------------------------------
; Map of the pixels: I|E F|J
; G|A B|K
; H|C D|L
; M|N O|P
movq mm6, mm0
movq mm4, [eax+ebx+colorA]
movq mm5, [eax+ebx+colorB]
pxor mm7, mm7
pand mm6, [ONE]
movq mm0, [eax+colorE]
movq mm1, [eax+ebx+colorG]
movq mm2, mm0
movq mm3, mm1
pcmpeqw mm0, mm4
pcmpeqw mm1, mm4
pcmpeqw mm2, mm5
pcmpeqw mm3, mm5
pand mm0, mm6
pand mm1, mm6
pand mm2, mm6
pand mm3, mm6
paddw mm0, mm1
paddw mm2, mm3
pxor mm3, mm3
pcmpgtw mm0, mm6
pcmpgtw mm2, mm6
pcmpeqw mm0, mm3
pcmpeqw mm2, mm3
pand mm0, mm6
pand mm2, mm6
paddw mm7, mm0
psubw mm7, mm2
movq mm0, [eax+colorF]
movq mm1, [eax+ebx+colorK]
movq mm2, mm0
movq mm3, mm1
pcmpeqw mm0, mm4
pcmpeqw mm1, mm4
pcmpeqw mm2, mm5
pcmpeqw mm3, mm5
pand mm0, mm6
pand mm1, mm6
pand mm2, mm6
pand mm3, mm6
paddw mm0, mm1
paddw mm2, mm3
pxor mm3, mm3
pcmpgtw mm0, mm6
pcmpgtw mm2, mm6
pcmpeqw mm0, mm3
pcmpeqw mm2, mm3
pand mm0, mm6
pand mm2, mm6
paddw mm7, mm0
psubw mm7, mm2
push eax
add eax, ebx
movq mm0, [eax+ebx+colorH]
movq mm1, [eax+ebx+ebx+colorN]
movq mm2, mm0
movq mm3, mm1
pcmpeqw mm0, mm4
pcmpeqw mm1, mm4
pcmpeqw mm2, mm5
pcmpeqw mm3, mm5
pand mm0, mm6
pand mm1, mm6
pand mm2, mm6
pand mm3, mm6
paddw mm0, mm1
paddw mm2, mm3
pxor mm3, mm3
pcmpgtw mm0, mm6
pcmpgtw mm2, mm6
pcmpeqw mm0, mm3
pcmpeqw mm2, mm3
pand mm0, mm6
pand mm2, mm6
paddw mm7, mm0
psubw mm7, mm2
movq mm0, [eax+ebx+colorL]
movq mm1, [eax+ebx+ebx+colorO]
movq mm2, mm0
movq mm3, mm1
pcmpeqw mm0, mm4
pcmpeqw mm1, mm4
pcmpeqw mm2, mm5
pcmpeqw mm3, mm5
pand mm0, mm6
pand mm1, mm6
pand mm2, mm6
pand mm3, mm6
paddw mm0, mm1
paddw mm2, mm3
pxor mm3, mm3
pcmpgtw mm0, mm6
pcmpgtw mm2, mm6
pcmpeqw mm0, mm3
pcmpeqw mm2, mm3
pand mm0, mm6
pand mm2, mm6
paddw mm7, mm0
psubw mm7, mm2
pop eax
movq mm1, mm7
pxor mm0, mm0
pcmpgtw mm7, mm0
pcmpgtw mm0, mm1
por mm7, [Mask1]
por mm1, [Mask2]
movq [Mask1], mm7
movq [Mask2], mm1
.SKIP_GUESS:
;----------------------------
;interpolate A, B, C and D
movq mm0, [eax+ebx+colorA]
movq mm1, [eax+ebx+colorB]
movq mm4, mm0
movq mm2, [eax+ebx+ebx+colorC]
movq mm5, mm1
movq mm3, [qcolorMask]
movq mm6, mm2
movq mm7, [qlowpixelMask]
pand mm0, mm3
pand mm1, mm3
pand mm2, mm3
pand mm3, [eax+ebx+ebx+colorD]
psrlw mm0, 2
pand mm4, mm7
psrlw mm1, 2
pand mm5, mm7
psrlw mm2, 2
pand mm6, mm7
psrlw mm3, 2
pand mm7, [eax+ebx+ebx+colorD]
paddw mm0, mm1
paddw mm2, mm3
paddw mm4, mm5
paddw mm6, mm7
paddw mm4, mm6
paddw mm0, mm2
psrlw mm4, 2
pand mm4, [qlowpixelMask]
paddw mm0, mm4 ;mm0 contains the interpolated value of A, B, C and D
;\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\
;assemble the pixels
movq mm1, [Mask1]
movq mm2, [Mask2]
movq mm4, [eax+ebx+colorA]
movq mm5, [eax+ebx+colorB]
pand mm4, mm1
pand mm5, mm2
pxor mm7, mm7
por mm1, mm2
por mm4, mm5
pcmpeqw mm1, mm7
pand mm0, mm1
por mm4, mm0 ;mm4 contains the diagonal pixels
movq mm0, [ACPixel]
movq mm1, mm0
punpcklwd mm0, mm4
punpckhwd mm1, mm4
push edx
add edx, [ebp+dstPitch]
%ifdef __DJGPP__
movq [fs:edx], mm0
movq [fs:edx+8], mm1
%else
movq [es:edx], mm0
movq [es:edx+8], mm1
%endif
pop edx
.SKIP_PROCESS:
mov ecx, [ebp+deltaPtr]
add ecx, 8
mov [ebp+deltaPtr], ecx
add edx, 16
add eax, 8
pop ecx
sub ecx, 4
cmp ecx, 0
jg near .Loop
; Restore some stuff
popad
mov esp, ebp
pop ebp
emms
ret
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
NEWSYM Init_2xSaIMMX
; Store some stuff
push ebp
mov ebp, esp
push edx
;Damn thing doesn't work
; mov eax,1
; cpuid
; test edx, 0x00800000 ;test bit 23
; jz end2 ;bit not set => no MMX detected
mov eax, [ebp+8] ;PixelFormat
cmp eax, 555
jz Bits555
cmp eax, 565
jz Bits565
end2:
mov eax, 1
jmp end
Bits555:
mov edx, 0x7BDE7BDE
mov eax, colorMask
mov [eax], edx
mov [eax+4], edx
mov edx, 0x04210421
mov eax, lowPixelMask
mov [eax], edx
mov [eax+4], edx
mov edx, 0x739C739C
mov eax, qcolorMask
mov [eax], edx
mov [eax+4], edx
mov edx, 0x0C630C63
mov eax, qlowpixelMask
mov [eax], edx
mov [eax+4], edx
mov eax, 0
jmp end
Bits565:
mov edx, 0xF7DEF7DE
mov eax, colorMask
mov [eax], edx
mov [eax+4], edx
mov edx, 0x08210821
mov eax, lowPixelMask
mov [eax], edx
mov [eax+4], edx
mov edx, 0xE79CE79C
mov eax, qcolorMask
mov [eax], edx
mov [eax+4], edx
mov edx, 0x18631863
mov eax, qlowpixelMask
mov [eax], edx
mov [eax+4], edx
mov eax, 0
jmp end
end:
pop edx
mov esp, ebp
pop ebp
ret
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
SECTION .data ALIGN = 32
;Some constants
colorMask dd 0xF7DEF7DE, 0xF7DEF7DE
lowPixelMask dd 0x08210821, 0x08210821
qcolorMask dd 0xE79CE79C, 0xE79CE79C
qlowpixelMask dd 0x18631863, 0x18631863
FALSE dd 0x00000000, 0x00000000
TRUE dd 0xffffffff, 0xffffffff
ONE dd 0x00010001, 0x00010001
SECTION .bss ALIGN = 32
ACPixel resb 8
Mask1 resb 8
Mask2 resb 8
I56Pixel resb 8
I23Pixel resb 8
I5556Pixel resb 8
I2223Pixel resb 8
I5666Pixel resb 8
I2333Pixel resb 8
Mask26 resb 8
Mask35 resb 8
Mask26b resb 8
Mask35b resb 8
product1a resb 8
product1b resb 8
product2a resb 8
product2b resb 8
final1a resb 8
final1b resb 8
final2a resb 8
final2b resb 8
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -