windows-server-2003/multimedia/directx/dxg/swrast/rgb/mmxspan/ml17.mas

;-----------------------------------------------------------------------------
;
;   Monolith 17. Perspective Correct Nearest texturing
;               16 bit Zbuffer (LE or GR) X888
;
;   Exactly the same as monolith 3 except color input is 32 bits and
;   output is 32 bits
;
;-----------------------------------------------------------------------------

INCLUDE iammx.inc
INCLUDE offs_acp.inc



; Names are read LSB to MSB, so B5G6R5 means five bits of blue starting
; at the LSB, then six bits of green, then five bits of red.



;TBD check to see if this value is correct.
COLOR_SHIFT equ 8

.586
.model flat


.data

  EXTERN IncHighandLow16:MMWORD
  EXTERN UFracVFracMask:MMWORD
  EXTERN UV32to15Mask:MMWORD
  EXTERN Makelow16one:MMWORD
  EXTERN MaskKeepUValues:MMWORD
  EXTERN MaskKeepVValues:MMWORD
  EXTERN UFrac:MMWORD
  EXTERN VFrac:MMWORD
  EXTERN Zero:MMWORD
  EXTERN memD3DTFG_POINT:MMWORD
  EXTERN GiveUp:MMWORD
  EXTERN LastW:MMWORD
  EXTERN Val0x000a000a:MMWORD
  EXTERN Val0xffff:MMWORD
  EXTERN Val0x0000002000000020:MMWORD
  EXTERN Val0x0000ffff0000ffff:MMWORD

opt_MaskRed565to888 MMWORD 000000000000F800H

EXTERN MaskRed565to888:MMWORD
EXTERN MaskGreen565to888:MMWORD
EXTERN MaskBlue565to888:MMWORD

EXTERN MaskRed555to888:MMWORD
EXTERN MaskGreen555to888:MMWORD
EXTERN MaskBlue555to888:MMWORD

EXTERN MaskAlpha1555to8888:MMWORD
EXTERN MaskRed1555to8888:MMWORD
EXTERN MaskGreen1555to8888:MMWORD
EXTERN MaskBlue1555to8888:MMWORD


EXTERN SetAlphato0xffff:MMWORD
EXTERN SetAlphato0xff:MMWORD

RedShift565to888     equ 8
GreenShift565to888   equ 5
BlueShift565to888    equ 3

RedShift555to888     equ 9
GreenShift555to888   equ 6
BlueShift555to888    equ 3

AlphaShift1555to8888 equ 16
RedShift1555to8888   equ 9
GreenShift1555to8888 equ 6
BlueShift1555to8888  equ 3

EXTERN Zero:MMWORD


EXTERN DW_One_One:MMWORD


EXTERN MaskOffAlpha:MMWORD
EXTERN ShiftTA:MMWORD
EXTERN Val0x00ff00ff00ff00ff:MMWORD
EXTERN Val0x000000ff00ff00ff:MMWORD
EXTERN Val0X0000000001000000:MMWORD
EXTERN AlphaVal128:MMWORD
EXTERN RGBVal128:MMWORD


    EXTERN  g_uDitherValue:MMWORD
    EXTERN  SetAlphato0xff:MMWORD
    EXTERN  u888to565RedBlueMask:MMWORD
    EXTERN  u888to565GreenMask:MMWORD
    EXTERN  u888to565Multiplier:MMWORD
    EXTERN  uVal0x000007ff03ff07ff:MMWORD
    EXTERN  uVal0x0000078003c00780:MMWORD
    EXTERN  u888to555RedBlueMask:MMWORD
    EXTERN  u888to555GreenMask:MMWORD
    EXTERN  u888to555Multiplier:MMWORD
    EXTERN  uVal0x000007ff07ff07ff:MMWORD
    EXTERN  uVal0x0000078007800780:MMWORD

; Span Variables
StackPos    dd  ?
uSpans      dd  ?
;-----------------------------------------------------------------------------
; Loop Variables

iSurfaceStep    dd  ?
iZStep          dd  ?
uPix            dd  ?

;-----------------------------------------------------------------------------

.code


    PUBLIC _MMXMLRast_17
_MMXMLRast_17:
    push    ebp
    mov     StackPos, esp
    mov     eax, esp
    sub     esp, 0Ch        ; This will need to change if stack frame size changes.
    push    ebx
    push    esi
    push    edi

    ; Put pCtx into ebx
    mov     ebx, [eax+8]

    ;PD3DI_RASTPRIM pP = pCtx->pPrim;
    mov     ecx, [ebx+RASTCTX_pPrim]

    ;while (pP)
    ;{
PrimLoop:
    cmp     ecx, 0
    je      ExitPrimLoop

    ;UINT16 uSpans = pP->uSpans;
    movzx   eax, word ptr [ecx+RASTPRIM_uSpans]
    mov     uSpans, eax

    ;PD3DI_RASTSPAN pS = (PD3DI_RASTSPAN)(pP + 1);
    mov     ebp, ecx
    add     ebp, SIZEOF_RASTPRIM


    ;while (uSpans-- > 0)
    ;{
SpanLoop:
    mov     edx, uSpans
    mov     eax, edx
    dec     eax
    mov     uSpans, eax
    test    edx, edx
    jle     ExitSpanLoop

    ;pCtx->pfnBegin(pCtx, pP, pS);

;-----------------------------------------------------------------------------
;  LoopAny code inserted here.  This is to get rid of an extra
;  jump.
;-----------------------------------------------------------------------------

; Setup Code begins

    ; get values to iterate

    ;uPix = pS->uPix;
    movzx   eax, word ptr [ebp+RASTSPAN_uPix]
    mov     uPix, eax


        ;pCtx->SI.iDW = 0x0;
        mov dword ptr [ebx+RASTCTX_SI+SPANITER_iDW], 0

        mov     esi, [ebp+RASTSPAN_iW]
        movq    mm5, MMWORD PTR [ebp+RASTSPAN_iUoW1]


    ;pCtx->SI.iUd_TexNum = d_WTimesUVoW(pS->iW,pS->iUoW1);
    ;pCtx->SI.iVd_TexNum = d_WTimesUVoW(pS->iW,pS->iVoW1);
    pslld   mm5, 8
    shl     esi, 4
    movd    eax, mm5
    psrlq   mm5, 32
    imul    esi
    mov     [ebx+RASTCTX_SI+SPANITER_iU1], edx
    movd    eax, mm5
    imul    esi
    mov     [ebx+RASTCTX_SI+SPANITER_iV1], edx


        ;if (pP->iDOoWDX > 0)
        ;{
        cmp dword ptr [ecx+RASTPRIM_iDOoWDX], 0
        jg  SpecialWLastMonTest
            ;// iSpecialW should be negative for the first 3 pixels of span
            ;pCtx->SI.iSpecialW = -3;
            mov word ptr [ebx+RASTCTX_SI+SPANITER_iSpecialW], -3
            jmp DoneSpecialWifMonTest
        ;}
        ;else
        ;{
SpecialWLastMonTest:
            ;// iSpecialW should be negative for the last 3 pixels of span
            ;pCtx->SI.iSpecialW = 0x7fff - uPix;
            mov     eax, 07fffh
            sub     eax, uPix
            ;pCtx->SI.iSpecialW += 5;        // this may wrap, but it should
            add     eax, 5
            mov     [ebx+RASTCTX_SI+SPANITER_iSpecialW], eax
        ;}
DoneSpecialWifMonTest:


    ;if (pP->uFlags & D3DI_RASTPRIM_X_DEC)
    ;{
    mov     eax, [ecx+RASTPRIM_uFlags]
    and     eax, D3DI_RASTPRIM_X_DEC
    test    eax, eax
    jz      LeftToRightSpan
        ;iZStep = -pCtx->iZStep;
        mov eax, [ebx+RASTCTX_iZStep]
        neg eax
        mov iZStep, eax
        ;iSurfaceStep = -pCtx->iSurfaceStep;
        mov eax, [ebx+RASTCTX_iSurfaceStep]
        neg eax
        mov iSurfaceStep, eax
    ;}
        jmp DoneSpanDirif
    ;else
    ;{
LeftToRightSpan:
        ;iZStep = pCtx->iZStep;
        mov eax, [ebx+RASTCTX_iZStep]
        mov iZStep, eax
        ;iSurfaceStep = pCtx->iSurfaceStep;
        mov eax, [ebx+RASTCTX_iSurfaceStep]
        mov iSurfaceStep, eax
    ;}
DoneSpanDirif:

; Setup Code Ends
; ----------------------------------------------------------------------------------------------------------------
; Loop Code Begins

PixelLoop:
    ; Ztestcode
    ; edx is uZ
    ; eax is uZB
    ; 16 bit unsigned format
    ;UINT16 uZ = (UINT16)(pS->uZ>>15);
    ;UINT16 uZB = *((UINT16*)pS->pZ);
    mov       edx, [ebp+RASTSPAN_uZ]
    movd      mm4, edx
    mov       esi, [ebp+RASTSPAN_pZ]
    shr       edx, 15
    movzx     eax, word ptr [esi]

    ;pS->uZ += pP->iDZDX;
    ;if ((pCtx->iZXorMask)^(uZ > uZB))
    ; !(uZ > uZB) <==>
    ; (uZ <= uZB) <==>
    ; (uZ < uZB+1) <==>
    ;
    sub     eax, edx
    paddd   mm4, [ecx+RASTPRIM_iDZDX]
    movd    [ebp+RASTSPAN_uZ], mm4
    xor     eax, [ebx+RASTCTX_iZXorMask]
    test    eax, eax
    js     FailLabel

    mov       word ptr [esi], dx
    ; texturecode

    mov   esi, [ebx+RASTCTX_pTexture]

    movq        mm5, MMWORD PTR Val0x000a000a  ; This is TEX_FINAL_SHIFT - 6 = 10.

    movd        mm4, [esi+SPANTEX_iShiftU]
    psubw       mm5, mm4
    movq        mm4, mm5
    pand        mm5, MMWORD PTR Val0xffff

    psrld       mm4, 16

    movd        mm1, [ebx+RASTCTX_SI+SPANITER_iU1]
    psrad       mm1, mm5
    movd        mm2, [ebx+RASTCTX_SI+SPANITER_iV1]
    psrad       mm2, mm4

    punpckldq   mm1, mm2

    movzx   edx, word ptr [esi+SPANTEX_iShiftPitch]
    add     edx, 16
    movd    mm2, edx
    movq    mm5, MMWORD ptr Makelow16one
    pslld   mm5, mm2

    por       mm5, MMWORD ptr Makelow16one

    psrad     mm1, 6

    packssdw  mm1, mm1  ; Value needs to be packed since all wrap/mirror

    movd      mm0, [esi+SPANTEX_uMaskU]     ; Load U and V mask

    movq      mm7, mm1
    movd      mm4, [esi+SPANTEX_iFlipMaskU]

    pand      mm7, mm4

    pcmpeqw   mm7, MMWORD PTR Zero
    pandn     mm7, mm0
    pand      mm1, mm0
    pxor      mm1, mm7

    pmaddwd   mm1, mm5
    mov       edi, [esi+SPANTEX_pBits]
    movd      eax, mm1

    mov     edx, dword ptr [edi+4*eax]
    and     edx, 000ffffffh         ; Have to make alpha 0x00 in 32 bit cases.
    mov     edi, [ebp+RASTSPAN_pSurface]
    mov     [edi], edx


FailLabel:

        dec   uPix

        jle   ExitPixelLoop

    movq    mm1, [ebp+RASTSPAN_uB]
    paddw   mm1, [ecx+RASTPRIM_iDBDX]
    movq    [ebp+RASTSPAN_uB], mm1


    movq    mm5, [ebp+RASTSPAN_iUoW1]
    paddd   mm5, [ecx+RASTPRIM_iDUoW1DX]
    movq    [ebp+RASTSPAN_iUoW1], mm5

    xor     eax, eax
    mov     ax, [ebp+RASTSPAN_iLOD]
    add     ax, [ebp+RASTSPAN_iDLOD]
    mov     [ebp+RASTSPAN_iLOD], ax


    mov     eax, [ebp+RASTSPAN_iOoW]
    add     eax, [ecx+RASTPRIM_iDOoWDX]
    mov     [ebp+RASTSPAN_iOoW], eax


    mov     edx, [ebp+RASTSPAN_iW]
    mov     LastW, edx          ; Save iW to calc iDW for next time.
    add     edx, [ebx+RASTCTX_SI+SPANITER_iDW]

    xor       edi, edi
    cmp       di, word ptr [ebx+RASTCTX_SI+SPANITER_iSpecialW]
    jle       DontDoSpecialW1


        cmp     edx, edi
        jl      WOutOfRange1

            mov edx, LastW
            sar edx, 1
WOutOfRange1:

        mov   GiveUp, 8    ; Pre decrementing instead of post decrementing.

SpecW1Loop1:

        dec   GiveUp
        jz    ExitSpecWLoop1

            mov     esi, (1 SHL 16)
            mov     edi, edx
            imul    edx
            sub     esi, edx

SpecW1Loop2:
            test    esi, esi
            jns     SpecW1ExitLoop2   ; This jump should be predicted correctly most of the time.
                add     esi, (1 SHL 15)
                sar     esi, 1
                jmp     SpecW1Loop2

SpecW1ExitLoop2:

            mov     eax, edi

            shl     eax, 5        ; 1.15.16 << 5  = 1.10.21  TBD  Can I shift off upper bits??
            shl     esi, 12       ; 4.15   << 12 = 4.27     ;

            mul    esi
            sub     edi, edx

            mov     eax, edi
            sar     eax, 31
            xor     edi, eax
            sub     edi, eax

            cmp     edi, 020h                   ;Assuming that loop will only happen once.
            jbe     ExitSpecWLoop1

            mov     eax, [ebp+RASTSPAN_iOoW]
            jmp     SpecW1Loop1

DontDoSpecialW1:
        mov     esi, (1 SHL 16)
        mov     edi, edx
        mul    edx
        sub     esi, edx
        shl     esi, 15
        mov     eax, esi

        mul    edi                       ; 0.2.30 * 1.15.16 = 1.17.46 >> 32 = 1.17.14
        shl     edx, 2    ; 1.17.14 << 2 = 1.15.16

ExitSpecWLoop1:

    mov     [ebp+RASTSPAN_iW], edx
    mov     esi, edx      ; Save W for multiplying by UoW and VoW
    sub     edx, LastW
    mov     [ebx+RASTCTX_SI+SPANITER_iDW], edx


    inc     word ptr [ebx+RASTCTX_SI+SPANITER_iSpecialW]

    pslld   mm5, 8
    shl     esi, 4
    movd    eax, mm5
    psrlq   mm5, 32
    imul    esi
    mov     [ebx+RASTCTX_SI+SPANITER_iU1], edx
    movd    eax, mm5
    imul    esi
    mov     [ebx+RASTCTX_SI+SPANITER_iV1], edx

        mov   eax, dword ptr [ebp+RASTSPAN_pZ]
        mov   edx, dword ptr [ebp+RASTSPAN_pSurface]

        add   eax, iZStep
        add   edx, iSurfaceStep

        mov   dword ptr [ebp+RASTSPAN_pZ], eax
        mov   dword ptr [ebp+RASTSPAN_pSurface], edx

    jmp   PixelLoop


ExitPixelLoop:
; Loop code ends

;-----------------------------------------------------------------------------
;  LoopAny code ends here
;-----------------------------------------------------------------------------

    ;pS++;
    add     ebp, SIZEOF_RASTSPAN

    ;}
    jmp     SpanLoop
ExitSpanLoop:
    ;pP = pP->pNext;
    mov     ecx, [ecx+RASTPRIM_pNext]
    ;}
    jmp     PrimLoop

ExitPrimLoop:
    ;_asm{
    emms
    ;}

    ;return S_OK;
    xor     eax, eax
;}
    pop     edi
    pop     esi
    pop     ebx
    mov     esp, StackPos
    pop     ebp
    ret

END