;-----------------------------------------------------------------------------
; Monolith 15.  Gouraud 16 bit Z Buffered (LE or GT) 888
;
; This monolith selects code to draw either left or right.
;
;   Globals (ATTENTION.  Need to put in stack memory so can multi-process)
;   iZXorMask   - Copy of iZXorMask since ebx (pCtx) is used for other
;                   purposes.
;
;   uSpans  - Count containing the number of spans.
;   StackPos - Saves stack position.
;
;   Register Usage:
;
;   eax - Number of pixels to draw
;   ebx - Zbuffer value / temp to determine Z fail
;   ecx - Prim pointer.  Same as in non-monolithic loop code. (i.e. C codes pCtx->pPrim)
;   edx - Tri Z value in 16 bits.
;   esi - Pointer to Z Buffer
;   edi - Pointer to Screen Buffer.
;   ebp - Rast Span Pointer.  Same as in non-monolithic loop code. (i.e. pS in C code)
;   mm0 - contains four color components.
;   mm1 - contains delta color components.
;   mm2 - Lower DWORD stores Delta Z value
;   mm3 - Temp for color conversion.
;   mm4 - Lower DWORD stores Z value.
;   mm5 - temp for color conversion.
;
;   Both Left to Right and Right to Left are very similar.
;
;   Psuedo Code
;
;   loop:
;       Do LE/GR Z test (load ebx with Z buffer value, copy mm4 to edx then do aritmetic compare of ebx and edx)
;       if (Z test passed)
;       {
;           Write Z (still in edx)
;           copy mm0 to mm5
;           convert mm5 to 888.
;           write color
;       }
;       if (done drawing) {exit loop} (dec eax / jz NoMorePixelsLtoR)
;       increment colors  (mm0+=mm1)
;       increment Z buffer pointer (add esi, 2)
;       increment screen pointer (add edi, 4)
;       jump to top of loop
;
;-----------------------------------------------------------------------------

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

; Big seperating lines seperate code into span code
; and loop code.  If span and loop are not going to
; end up being combined then it will be easy to
; seperate the code.

.data
; It would help if these are close together in cache lines.
iZXorMask               dq  ?

; Need externs for all of the variables that are needed for various beads

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

; TBD. I think that I want to do 0xffff instead of 0xff.  This will
; have to be checked.  There is a value very similiar to this in
; buf write.
EXTERN SetAlphato0xffff:MMWORD
EXTERN SetAlphato0xff:MMWORD

; TODO This equate are identical to the ones in texread.mas.  Maybe they should be in a common .inc file.
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  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_15
_MMXMLRast_15:
    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

    xor        edi, edi

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

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

    ; Need to generate qword iZXorMask since it in pCtx

    movd    mm2, [ebx+RASTCTX_iZXorMask]
    punpckldq   mm2, mm2
    movq    MMWORD PTR iZXorMask, mm2

    ; ebx is free after this since gouraud does not need information from pCtx.

    ;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


        mov     edi, dword ptr [ebp+RASTSPAN_pSurface]
        mov     esi, [ebp+RASTSPAN_pZ]
        movzx   eax, word ptr [ebp+RASTSPAN_uPix]

    ;if (pP->uFlags & D3DI_RASTPRIM_X_DEC)
    ;{
        mov     edx, [ecx+RASTPRIM_uFlags]
        and     edx, D3DI_RASTPRIM_X_DEC
        test    edx, edx
        jz      LeftToRightSpan

; SCREWED UP RIGHT TO LEFT CASE
        movq    mm0, MMWORD PTR [ebp+RASTSPAN_uB]
        movd    mm4, dword ptr [ebp+RASTSPAN_uZ]
        movd    mm2, dword ptr [ecx+RASTPRIM_iDZDX]
        movq    mm1, MMWORD PTR [ecx+RASTPRIM_iDBDX]

beginingpixelsRtoL:
        ;WritePixel
        ; Ztestcode
        ; edx is uZ
        ; ebx is uZB
        ; 16 bit unsigned format
        ;UINT16 uZ = (UINT16)(pS->uZ>>15);
        ;UINT16 uZB = *((UINT16*)pS->pZ);
        movd      edx, mm4
        shr       edx, 15
        movzx     ebx, word ptr [esi]

        ;pS->uZ += pP->iDZDX;
        ;if ((pCtx->iZXorMask)^(uZ > uZB))
        sub     ebx, edx
        paddd   mm4, mm2
        xor     ebx, dword ptr iZXorMask
        test    ebx, ebx
        js      ZFailLabelRtoL1

        mov       word ptr [esi], dx

        ; Convert color.
        movq    mm5, mm0
        psrlw   mm5, 8          ; Convert color1 from 8.8 two 0.8
    packuswb    mm5, mm7        ; pack one color1
        movd    edx, mm5
        and     edx, 000ffffffh
        mov     [edi], edx

ZFailLabelRtoL1:
        dec     eax             ; Reduce Pixel count
        jz      NoMorePixelsRtoL

        sub     edi, 4          ; decrease destination pointer
        sub     esi, 2          ; decrease Z Buffer Pointer.

        ;pS->uB += pP->iDBDX; pS->uG += pP->iDGDX;
        ;pS->uR += pP->iDRDX; pS->uA += pP->iDADX;
        paddw   mm0, mm1

        jmp     beginingpixelsRtoL

NoMorePixelsRtoL:
        jmp     DoneSpanDirif
    ;else
    ;{
LeftToRightSpan:


; NORMAL LEFT TO RIGHT CASE
        movq    mm0, MMWORD PTR [ebp+RASTSPAN_uB]
        movd    mm4, dword ptr [ebp+RASTSPAN_uZ]
        movd    mm2, dword ptr [ecx+RASTPRIM_iDZDX]
        movq    mm1, MMWORD PTR [ecx+RASTPRIM_iDBDX]

beginingpixelsLtoR:
        ;WritePixel
        ; Ztestcode
        ; edx is uZ
        ; ebx is uZB
        ; 16 bit unsigned format
        ;UINT16 uZ = (UINT16)(pS->uZ>>15);
        ;UINT16 uZB = *((UINT16*)pS->pZ);
        movd      edx, mm4
        shr       edx, 15
        movzx     ebx, word ptr [esi]

        ;pS->uZ += pP->iDZDX;
        ;if ((pCtx->iZXorMask)^(uZ > uZB))
        sub     ebx, edx
        paddd   mm4, mm2
        xor     ebx, dword ptr iZXorMask
        test    ebx, ebx
        js      ZFailLabelLtoR1

        ; Write Z
        mov       word ptr [esi], dx

        ; Convert color.
        movq    mm5, mm0
        psrlw   mm5, 8          ; Convert color1 from 8.8 two 0.8
    packuswb    mm5, mm7        ; pack one color1
        movd    edx, mm5
        and     edx, 000ffffffh
        mov     [edi], edx

ZFailLabelLtoR1:
        dec     eax             ; Reduce Pixel count
        jz      NoMorePixelsLtoR

        add     edi, 4          ; Increase destination pointer.
        add     esi, 2          ; Increase Z Buffer Pointer.

        ;pS->uB += pP->iDBDX; pS->uG += pP->iDGDX;
        ;pS->uR += pP->iDRDX; pS->uA += pP->iDADX;
        paddw   mm0, mm1

        jmp     beginingpixelsLtoR

NoMorePixelsLtoR:

    ;}
DoneSpanDirif:

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


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