;-------------------------------------------------------------------------
; cxm12161 -- This function performs YUV12-to-RGB16 color conversion for H26x.
;             It handles any format in which there are three fields, the low
;             order field being B and fully contained in the low order byte, the
;             second field being G and being somewhere in bits 4 through 11,
;             and the high order field being R and fully contained in the high
;             order byte.
;
;             The YUV12 input is planar, 8 bits per pel.  The Y plane may have
;             a pitch of up to 768.  It may have a width less than or equal
;             to the pitch.  It must be DWORD aligned, and preferably QWORD
;             aligned.  Pitch and Width must be a multiple of four.  For best
;             performance, Pitch should not be 4 more than a multiple of 32.
;             Height may be any amount, but must be a multiple of two.  The U
;             and V planes may have a different pitch than the Y plane, subject
;             to the same limitations.
;

;include iammx.inc
;include locals.inc


RGB_formats:
    dd  RGB565
    dd  RGB555
    dd  RGB664
    dd  RGB655

Minusg:             dd   00800080h, 00800080h
Yadd:               dd   10101010h, 10101010h
VtR:                dd   00660066h, 00660066h ;01990199h,01990199h
VtG:                dd   00340034h, 00340034h ;00d000d0h,00d000d0h
UtG:                dd   00190019h, 00190019h ;00640064h,00640064h
UtB:                dd   00810081h, 00810081h ;02050205h,02050205h
Ymul:               dd   004a004ah, 004a004ah ;012a012ah,012a012ah
UVtG:               dd   00340019h, 00340019h ;00d00064h,00d00064h
VtRUtB:             dd   01990205h, 01990205h
fourbitu:           dd  0f0f0f0f0h, 0f0f0f0f0h
fivebitu:           dd  0e0e0e0e0h, 0e0e0e0e0h
sixbitu:            dd  0c0c0c0c0h, 0c0c0c0c0h

%assign LocalFrameSize  156
%assign RegisterStorageSize  16

; Arguments:
%assign YPlane                    LocalFrameSize + RegisterStorageSize +  4
%assign UPlane                    LocalFrameSize + RegisterStorageSize +  8
%assign VPlane                    LocalFrameSize + RegisterStorageSize + 12
%assign FrameWidth                LocalFrameSize + RegisterStorageSize + 16
%assign FrameHeight               LocalFrameSize + RegisterStorageSize + 20
%assign YPitch                    LocalFrameSize + RegisterStorageSize + 24
%assign ChromaPitch               LocalFrameSize + RegisterStorageSize + 28
%assign AspectAdjustmentCount     LocalFrameSize + RegisterStorageSize + 32
%assign ColorConvertedFrame       LocalFrameSize + RegisterStorageSize + 36
%assign DCIOffset                 LocalFrameSize + RegisterStorageSize + 40
%assign CCOffsetToLine0           LocalFrameSize + RegisterStorageSize + 44
%assign CCOPitch                  LocalFrameSize + RegisterStorageSize + 48
%assign CCType                    LocalFrameSize + RegisterStorageSize + 52
%assign EndOfArgList              LocalFrameSize + RegisterStorageSize + 56

; Locals (on local stack frame)
%assign CCOCursor                   0
%assign CCOSkipDistance             4
%assign ChromaLineLen               8
%assign YCursor                    12
%assign DistanceFromVToU           16
%assign EndOfChromaLine            20
%assign AspectCount                24
%assign AspectBaseCount            28
%assign tmpYCursorEven             32
%assign tmpYCursorOdd              36
%assign tmpCCOPitch                40
%assign temp_mmx                   44  ; note it is 48 bytes
%assign RLeftShift                 92
%assign GLeftShift                100
%assign RRightShift               108
%assign GRightShift               116
%assign BRightShift               124
%assign RUpperLimit               132
%assign GUpperLimit               140
%assign BUpperLimit               148


; extern void C MMX_YUV12ToRGB16 (
;                                     U8* YPlane,
;                                     U8* UPlane,
;                                     U8* VPlane,
;                                     UN  FrameWidth,
;                                     UN  FrameHeight,
;                                     UN  YPitch,
;                                     UN  VPitch,
;                                     UN  AspectAdjustmentCount,
;                                     U8* ColorConvertedFrame,
;                                     U32 DCIOffset,
;                                     U32 CCOffsetToLine0,
;                                     IN  CCOPitch,
;                                     IN  CCType)
;
;  The local variables are on the stack,
;  The tables are in the one and only data segment.
;
;  CCOffsetToLine0 is relative to ColorConvertedFrame.
;  CCType  used by RGB color convertors to determine the exact conversion type.
;    RGB565 = 0 
;    RGB555 = 1
;    RGB664 = 2
;    RGB655 = 3

global yuv_2_rgb

yuv_2_rgb:
  push       esi
  push       edi

  push       ebp
  push       ebx

  sub        esp, LocalFrameSize
  mov        eax, [esp+CCType]
  cmp        eax,4
  jae        near finish

  jmp        [RGB_formats+eax*4]

RGB555:
  xor        eax, eax
  mov        ebx, 2   ; 10-8 for byte shift
  mov        [esp+RLeftShift], ebx
  mov        [esp+RLeftShift+4], eax
  mov        ebx, 5
  mov        [esp+GLeftShift], ebx
  mov        [esp+GLeftShift+4], eax
  mov        ebx, 9
  mov        [esp+RRightShift], ebx
  mov        [esp+RRightShift+4], eax
  mov        [esp+GRightShift], ebx
  mov        [esp+GRightShift+4], eax
  mov        [esp+BRightShift], ebx
  mov        [esp+BRightShift+4], eax
  movq       mm0, [fivebitu]
  movq       [esp+RUpperLimit], mm0
  movq       [esp+GUpperLimit], mm0
  movq       [esp+BUpperLimit], mm0
  jmp        RGBEND

RGB664:
  xor        eax, eax
  mov        ebx, 2   ; 8-6
  mov        [esp+RLeftShift], ebx
  mov        [esp+RLeftShift+4], eax
  mov        ebx, 4
  mov        [esp+GLeftShift], ebx
  mov        [esp+GLeftShift+4], eax
  mov        ebx, 8
  mov        [esp+RRightShift], ebx
  mov        [esp+RRightShift+4], eax
  mov        [esp+GRightShift], ebx
  mov        [esp+GRightShift+4], eax
  mov        ebx, 10
  mov        [esp+BRightShift], ebx
  mov        [esp+BRightShift+4], eax
  movq       mm0, [sixbitu]
  movq       [esp+RUpperLimit], mm0
  movq       [esp+GUpperLimit], mm0
  movq       mm0, [fourbitu]
  movq       [esp+BUpperLimit], mm0
  jmp        RGBEND

RGB655:
  xor        eax, eax
  mov        ebx, 2   ; 8-6
  mov        [esp+RLeftShift], ebx
  mov        [esp+RLeftShift+4], eax
  mov        ebx, 5
  mov        [esp+GLeftShift], ebx
  mov        [esp+GLeftShift+4], eax
  mov        ebx, 8
  mov        [esp+RRightShift], ebx
  mov        [esp+RRightShift+4], eax
  mov        ebx, 9
  mov        [esp+GRightShift], ebx
  mov        [esp+GRightShift+4], eax
  mov        [esp+BRightShift], ebx
  mov        [esp+BRightShift+4], eax
  movq       mm0, [sixbitu]
  movq       [esp+RUpperLimit], mm0
  movq       mm0, [fivebitu]
  movq       [esp+GUpperLimit], mm0
  movq       [esp+BUpperLimit], mm0
  jmp        RGBEND

RGB565:
  xor        eax, eax
  mov        ebx, 3   ; 8-5
  mov        [esp+RLeftShift], ebx
  mov        [esp+RLeftShift+4], eax
  mov        ebx, 5
  mov        [esp+GLeftShift], ebx
  mov        [esp+GLeftShift+4], eax
  mov        ebx, 9
  mov        [esp+RRightShift], ebx
  mov        [esp+RRightShift+4], eax
  mov        [esp+BRightShift], ebx
  mov        [esp+BRightShift+4], eax
  mov        ebx, 8
  mov        [esp+GRightShift], ebx
  mov        [esp+GRightShift+4], eax
  movq       mm0, [fivebitu]
  movq       [esp+RUpperLimit], mm0
  movq       [esp+BUpperLimit], mm0
  movq       mm0, [sixbitu]
  movq       [esp+GUpperLimit], mm0
;  jmp        RGBEND

RGBEND:
  mov        ebx, [esp+VPlane]
  mov        ecx, [esp+UPlane]
  sub        ecx, ebx
  mov        [esp+DistanceFromVToU], ecx

  mov        eax, [esp+ColorConvertedFrame]
  add        eax, [esp+DCIOffset]
  add        eax, [esp+CCOffsetToLine0]
  mov        [esp+CCOCursor], eax


  mov		ecx,[esp+YPitch]
  mov		ebx,[esp+FrameWidth]
  mov		eax,[esp+CCOPitch]
  sub        eax, ebx         ; CCOPitch-FrameWidth
  sub        eax, ebx         ; CCOPitch-2*FrameWidth
  sar        ebx, 1           ; FrameWidth/2
  mov		esi,[esp+YPlane]           ; Fetch cursor over luma plane.
  mov		[esp+ChromaLineLen],ebx    ; FrameWidth/2
  mov		[esp+ CCOSkipDistance],eax  ; CCOPitch-3*FrameWidth
  mov		[esp+YCursor],esi
  mov		edx,[esp+AspectAdjustmentCount]
  mov		esi,[esp+VPlane]

  cmp   edx,1
  je    near finish
  mov		[esp+AspectCount],edx
  mov		[esp+AspectBaseCount],edx
  xor        eax, eax

  mov		edi,[esp+ChromaLineLen]
  mov		[esp+EndOfChromaLine],edi
  mov		edi,[esp+CCOCursor]

  mov		edx,[esp+DistanceFromVToU]
  mov		ebp,[esp+YCursor]                       ; Fetch Y Pitch.
  mov		ebx,[esp+FrameWidth]

  add        ebp, ebx
  mov		[esp+tmpYCursorEven],ebp
  mov		eax,[esp+YPitch]
  add        ebp, eax
  mov		[esp+tmpYCursorOdd],ebp