hq4x32.asm

linux下的任天堂模拟器代码。供大家参考。

%macro PIXEL31_83 0
    Interp8 [ecx+ebx+4],[c8],[c4]
%endmacro

%macro PIXEL32_0 0
    mov [ecx+ebx+8],eax
%endmacro

%macro PIXEL32_10 0
    Interp1 [ecx+ebx+8],eax,[c9]
%endmacro

%macro PIXEL32_12 0
    Interp1 [ecx+ebx+8],eax,[c8]
%endmacro

%macro PIXEL32_14 0
    Interp1 [ecx+ebx+8],[c8],eax
%endmacro

%macro PIXEL32_21 0
    Interp2 [ecx+ebx+8],[c8],eax,[c6]
%endmacro

%macro PIXEL32_31 0
    Interp3 [ecx+ebx+8],[c6]
%endmacro

%macro PIXEL32_50 0
    Interp5 [ecx+ebx+8],[c8],eax
%endmacro

%macro PIXEL32_60 0
    Interp6 [ecx+ebx+8],[c8],[c6]
%endmacro

%macro PIXEL32_61 0
    Interp6 [ecx+ebx+8],[c8],[c9]
%endmacro

%macro PIXEL32_82 0
    Interp8 [ecx+ebx+8],eax,[c8]
%endmacro

%macro PIXEL32_83 0
    Interp8 [ecx+ebx+8],[c8],[c6]
%endmacro

%macro PIXEL33_0 0
    mov [ecx+ebx+12],eax
%endmacro

%macro PIXEL33_11 0
    Interp1 [ecx+ebx+12],eax,[c6]
%endmacro

%macro PIXEL33_12 0
    Interp1 [ecx+ebx+12],eax,[c8]
%endmacro

%macro PIXEL33_20 0
    Interp2 [ecx+ebx+12],eax,[c8],[c6]
%endmacro

%macro PIXEL33_50 0
    Interp5 [ecx+ebx+12],[c8],[c6]
%endmacro

%macro PIXEL33_80 0
    Interp8 [ecx+ebx+12],eax,[c9]
%endmacro

%macro PIXEL33_81 0
    Interp8 [ecx+ebx+12],eax,[c6]
%endmacro

%macro PIXEL33_82 0
    Interp8 [ecx+ebx+12],eax,[c8]
%endmacro

NEWSYM hq4x_32b
    cmp byte[curblank],40h
    jne .startcopy
    ret
.startcopy
    pushad
    mov ax,ds
    mov es,ax
    mov esi,[vidbuffer]
    mov edi,[WinVidMemStart]
    add esi,16*2+256*2+32*2
    mov ecx,[vidbufferofsb]
    mov [deltaptr],ecx
    cmp byte[FilteredGUI],0
    jne .filtergui
    cmp byte[GUIOn2],1
    je  nointerp
.filtergui
    cmp byte[MMXSupport],0
    je  nointerp
    cmp byte[hqFilter],0
    jne near hq4x

;----------------------------;
nointerp:
    mov dl,[resolutn]
    mov [lineleft],dl
    mov ebx,[NumBytesPerLine]
.loopy
    mov ecx,256
.loopx
    movzx eax,word[esi]
    mov edx,[BitConv32Ptr]
    mov eax,[edx+eax*4]
    mov [edi],eax
    mov [edi+4],eax
    mov [edi+8],eax
    mov [edi+12],eax
    mov [edi+ebx],eax
    mov [edi+ebx+4],eax
    mov [edi+ebx+8],eax
    mov [edi+ebx+12],eax
    mov edx,edi
    add edx,ebx
    add edx,ebx
    mov [edx],eax
    mov [edx+4],eax
    mov [edx+8],eax
    mov [edx+12],eax
    mov [edx+ebx],eax
    mov [edx+ebx+4],eax
    mov [edx+ebx+8],eax
    mov [edx+ebx+12],eax
    add esi,2
    add edi,16
    dec ecx
    jnz .loopx
    add edi,[AddEndBytes]
    add edi,ebx
    add edi,ebx
    add edi,ebx
    add esi,64
    dec byte[lineleft]
    jnz near .loopy
    popad
    ret

;----------------------------;
hq4x:
    mov dl,[resolutn]
    mov [lineleft],dl
    mov dword[prevline],0
    mov dword[nextline],576
    mov ebx,hirestiledat+1
    cmp byte[GUIOn],1
    je .loopy
    cmp byte[newengen],0
    je .loopy
    mov ebx,SpecialLine+1
.loopy
    mov [InterPtr],ebx
    cmp byte[ebx],1
    jbe .nohires
    call HighResProc
    mov edx,[deltaptr]
    mov ecx,128
    mov eax,0xAAAAAAAA
.a
    mov [edx],eax
    add edx,4
    dec ecx
    jnz .a
    mov [deltaptr],edx
    jmp .nexty
.nohires
    mov     dword[xcounter],254   ; x={Xres-2, Xres-1} are special cases.
    ; x=0 - special case
    mov     edx,[deltaptr]
    mov     ecx,[prevline]
    mov     eax,[nextline]
    movq    mm2,[esi+ecx]
    movq    mm3,[esi]
    movq    mm4,[esi+eax]
    movq    mm5,mm2
    movq    mm6,mm3
    movq    mm7,mm4
    pcmpeqw mm2,[edx+ecx]
    pcmpeqw mm3,[edx]
    pcmpeqw mm4,[edx+eax]
    pand    mm2,mm3
    pand    mm2,mm4
    movd    eax,mm2
    inc     eax
    jz      near .loopx_end
    movd    eax,mm5
    movzx   edx,ax
    mov     [w1],edx
    mov     [w2],edx
    shr     eax,16
    mov     [w3],eax
    movd    eax,mm6
    movzx   edx,ax
    mov     [w4],edx
    mov     [w5],edx
    shr     eax,16
    mov     [w6],eax
    movd    eax,mm7
    movzx   edx,ax
    mov     [w7],edx
    mov     [w8],edx
    shr     eax,16
    mov     [w9],eax
    jmp     .flags
.loopx
    mov     edx,[deltaptr]
    mov     ecx,[prevline]
    mov     eax,[nextline]
    movq    mm2,[esi+ecx-2]
    movq    mm3,[esi-2]
    movq    mm4,[esi+eax-2]
    movq    mm5,mm2
    movq    mm6,mm3
    movq    mm7,mm4
    pcmpeqw mm2,[edx+ecx-2]
    pcmpeqw mm3,[edx-2]
    pcmpeqw mm4,[edx+eax-2]
    pand    mm2,mm3
    pand    mm2,mm4
    movd    ebx,mm2
    psrlq   mm2,32
    movd    eax,mm2
    cwde
    and     eax,ebx
    inc     eax
    jz      near .loopx_end
    movd    eax,mm5
    mov     [edx+ecx-2],ax
    movzx   edx,ax
    mov     [w1],edx
    shr     eax,16
    mov     [w2],eax
    psrlq   mm5,32
    movd    eax,mm5
    movzx   edx,ax
    mov     [w3],edx
    movd    eax,mm6
    movzx   edx,ax
    mov     [w4],edx
    shr     eax,16
    mov     [w5],eax
    psrlq   mm6,32
    movd    eax,mm6
    movzx   edx,ax
    mov     [w6],edx
    movd    eax,mm7
    movzx   edx,ax
    mov     [w7],edx
    shr     eax,16
    mov     [w8],eax
    psrlq   mm7,32
    movd    eax,mm7
    movzx   edx,ax
    mov     [w9],edx
.flags
    mov     ebx,[RGBtoYUVPtr]
    mov     eax,[w5]
    xor     ecx,ecx
    movd    mm5,[ebx+eax*4]
    mov     dword[cross],0

    mov     edx,[w2]
    cmp     eax,edx
    je      .noflag2
    or      dword[cross],1
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag2
    or      ecx,2
.noflag2
    mov     edx,[w4]
    cmp     eax,edx
    je      .noflag4
    or      dword[cross],2
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag4
    or      ecx,8
.noflag4
    mov     edx,[w6]
    cmp     eax,edx
    je      .noflag6
    or      dword[cross],4
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag6
    or      ecx,16
.noflag6
    mov     edx,[w8]
    cmp     eax,edx
    je      .noflag8
    or      dword[cross],8
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag8
    or      ecx,64
.noflag8
    cmp     dword[cross],0
    jnz     .testflag1
    mov     ebx,[BitConv32Ptr]
    mov     eax,[ebx+eax*4]
    mov     ebx,[NumBytesPerLine]
    AUXADDRESS
    mov     [edi],eax
    mov     [edi+4],eax
    mov     [edi+8],eax
    mov     [edi+12],eax
    mov     [edi+ebx],eax
    mov     [edi+ebx+4],eax
    mov     [edi+ebx+8],eax
    mov     [edi+ebx+12],eax
    mov     [ecx],eax
    mov     [ecx+4],eax
    mov     [ecx+8],eax
    mov     [ecx+12],eax
    mov     [ecx+ebx],eax
    mov     [ecx+ebx+4],eax
    mov     [ecx+ebx+8],eax
    mov     [ecx+ebx+12],eax
    jmp     .loopx_end
.testflag1
    mov     edx,[w1]
    cmp     eax,edx
    je      .noflag1
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag1
    or      ecx,1
.noflag1
    mov     edx,[w3]
    cmp     eax,edx
    je      .noflag3
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag3
    or      ecx,4
.noflag3
    mov     edx,[w7]
    cmp     eax,edx
    je      .noflag7
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag7
    or      ecx,32
.noflag7
    mov     edx,[w9]
    cmp     eax,edx
    je      .noflag9
    movq    mm1,mm5
    movd    mm2,[ebx+edx*4]
    psubusb mm1,mm2
    psubusb mm2,mm5
    por     mm1,mm2
    psubusb mm1,[threshold]
    movd    edx,mm1
    test    edx,edx
    jz      .noflag9
    or      ecx,128
.noflag9
    mov     ebx,[BitConv32Ptr]
    mov     eax,[ebx+eax*4]
    mov     edx,[w2]
    mov     edx,[ebx+edx*4]
    mov     [c2],edx
    mov     edx,[w4]
    mov     edx,[ebx+edx*4]
    mov     [c4],edx
    mov     edx,[w6]
    mov     edx,[ebx+edx*4]
    mov     [c6],edx
    mov     edx,[w8]
    mov     edx,[ebx+edx*4]
    mov     [c8],edx
    test    ecx,0x005A
    jz      .switch
    mov     edx,[w1]
    mov     edx,[ebx+edx*4]
    mov     [c1],edx
    mov     edx,[w3]
    mov     edx,[ebx+edx*4]
    mov     [c3],edx
    mov     edx,[w7]
    mov     edx,[ebx+edx*4]
    mov     [c7],edx
    mov     edx,[w9]
    mov     edx,[ebx+edx*4]
    mov     [c9],edx
.switch
    mov     ebx,[NumBytesPerLine]
    jmp     [FuncTable+ecx*4]

..@flag0
..@flag1
..@flag4
..@flag32
..@flag128
..@flag5
..@flag132
..@flag160
..@flag33
..@flag129
..@flag36
..@flag133
..@flag164
..@flag161
..@flag37
..@flag165
    AUXADDRESS
    PIXEL00_20
    PIXEL01_60
    PIXEL02_60
    PIXEL03_20
    PIXEL10_60
    PIXEL11_70
    PIXEL12_70
    PIXEL13_60
    PIXEL20_60
    PIXEL21_70
    PIXEL22_70
    PIXEL23_60
    PIXEL30_20
    PIXEL31_60
    PIXEL32_60
    PIXEL33_20
    jmp .loopx_end
..@flag2
..@flag34
..@flag130
..@flag162
    AUXADDRESS
    PIXEL00_80
    PIXEL01_10
    PIXEL02_10
    PIXEL03_80
    PIXEL10_61
    PIXEL11_30
    PIXEL12_30
    PIXEL13_61
    PIXEL20_60
    PIXEL21_70
    PIXEL22_70
    PIXEL23_60
    PIXEL30_20
    PIXEL31_60
    PIXEL32_60
    PIXEL33_20
    jmp .loopx_end
..@flag16
..@flag17
..@flag48
..@flag49
    AUXADDRESS
    PIXEL00_20
    PIXEL01_60
    PIXEL02_61
    PIXEL03_80
    PIXEL10_60
    PIXEL11_70
    PIXEL12_30
    PIXEL13_10
    PIXEL20_60
    PIXEL21_70
    PIXEL22_30
    PIXEL23_10
    PIXEL30_20
    PIXEL31_60
    PIXEL32_61
    PIXEL33_80
    jmp .loopx_end
..@flag64
..@flag65
..@flag68
..@flag69
    AUXADDRESS
    PIXEL00_20
    PIXEL01_60
    PIXEL02_60
    PIXEL03_20
    PIXEL10_60
    PIXEL11_70
    PIXEL12_70
    PIXEL13_60
    PIXEL20_61
    PIXEL21_30
    PIXEL22_30
    PIXEL23_61
    PIXEL30_80
    PIXEL31_10
    PIXEL32_10
    PIXEL33_80
    jmp .loopx_end
..@flag8
..@flag12
..@flag136
..@flag140
    AUXADDRESS
    PIXEL00_80
    PIXEL01_61
    PIXEL02_60
    PIXEL03_20
    PIXEL10_10
    PIXEL11_30
    PIXEL12_70
    PIXEL13_60
    PIXEL20_10
    PIXEL21_30
    PIXEL22_70
    PIXEL23_60
    PIXEL30_80
    PIXEL31_61
    PIXEL32_60
    PIXEL33_20
    jmp .loopx_end
..@flag3
..@flag35
..@flag131
..@flag163
    AUXADDRESS
    PIXEL00_81
    PIXEL01_31
    PIXEL02_10
    PIXEL03_80
    PIXEL10_81
    PIXEL11_31
    PIXEL12_30
    PIXEL13_61
    PIXEL20_60
    PIXEL21_70
    PIXEL22_70
    PIXEL23_60
    PIXEL30_20
    PIXEL31_60
    PIXEL32_60
    PIXEL33_20
    jmp .loopx_end
..@flag6
..@flag38
..@flag134
..@flag166
    AUXADDRESS
    PIXEL00_80
    PIXEL01_10
    PIXEL02_32
    PIXEL03_82
    PIXEL10_61
    PIXEL11_30
    PIXEL12_32
    PIXEL13_82
    PIXEL20_60
    PIXEL21_70
    PIXEL22_70
    PIXEL23_60
    PIXEL30_20
    PIXEL31_60
    PIXEL32_60
    PIXEL33_20
    jmp .loopx_end
..@flag20
..@flag21
..@flag52
..@flag53
    AUXADDRESS
    PIXEL00_20
    PIXEL01_60
    PIXEL02_81
    PIXEL03_81
    PIXEL10_60
    PIXEL11_70
    PIXEL12_31
    PIXEL13_31
    PIXEL20_60
    PIXEL21_70
    PIXEL22_30
    PIXEL23_10
    PIXEL30_20
    PIXEL31_60
    PIXEL32_61
    PIXEL33_80
    jmp .loopx_end
..@flag144
..@flag145
..@flag176
..@flag177
    AUXADDRESS
    PIXEL00_20
    PIXEL01_60
    PIXEL02_61
    PIXEL03_80
    PIXEL10_60
    PIXEL11_70
    PIXEL12_30
    PIXEL13_10
    PIXEL20_60
    PIXEL21_70
    PIXEL22_32
    PIXEL23_32
    PIXEL30_20
    PIXEL31_60
    PIXEL32_82
    PIXEL33_82
    jmp .loopx_end
..@flag192
..@flag193
..@flag196
..@flag197
    AUXADDRESS
    PIXEL00_20
    PIXEL01_60
    PIXEL02_60
    PIXEL03_20
    PIXEL10_60
    PIXEL11_70
    PIXEL12_70
    PIXEL13_60
    PIXEL20_61
    PIXEL21_30
    PIXEL22_31
    PIXEL23_81
    PIXEL30_80
    PIXEL31_10
    PIXEL32_31
    PIXEL33_81
    jmp .loopx_end
..@flag96
..@flag97
..@flag100
..@flag101
    AUXADDRESS
    PIXEL00_20
    PIXEL01_60
    PIXEL02_60
    PIXEL03_20
    PIXEL10_60
    PIXEL11_70
    PIXEL12_70