NeuQuant (SSE2)

[netmaestro]

Can anyone spot what I'm doing wrong here? If the code runs with the current commenting, quality is good but the colors are wrong. Switch the commenting in the inxSearch loop and the colors become right but the quality goes way down. Here's the code, it'll download the sample image one time:

Code: Select all

IncludeFile "NeuQuant.pbi"

Macro GetPixel32BitImage (x, y) 
  PeekL(bitsptr_in + sz_colorbits_in  + ((x) << 2) - ((y) + 1) * wb_in)
EndMacro

Macro SetPixel8BitImage (x, y, col) 
  PokeB(bitsptr_out + (h*wb_out) + (x) - ((y) + 1) * wb_out, col)
EndMacro

Procedure ImageTo8bit(image_number)
  If ImageDepth(image_number)<>32
    this = CreateImage(#PB_Any, ImageWidth(image_number), ImageHeight(image_number), 32)
    StartDrawing(ImageOutput(this))
      DrawImage(ImageID(image_number), 0, 0)
    StopDrawing()
    FreeImage(image_number)
    image_number = this
    hImageIn = ImageID(this)
  Else
    hImageIn = ImageID(image_number)
  EndIf
  
  *palette = AllocateMemory(256*SizeOf(RGBQUAD)) 
  NeuQuant::InitNet()
  NeuQuant::Learn(image_number)
  NeuQuant::UnbiasNet(*palette)
  NeuQuant::InxBuild()
  
  GetObject_(hImageIn,SizeOf(BITMAP),bmp32.BITMAP)
  With bmp32
    w =               \bmWidth 
    h =               \bmHeight 
    bitsptr_in =      \bmBits
    sz_colorbits_in = \bmWidthBytes * bmp32\bmHeight
    wb_in =           \bmWidthBytes
  EndWith
  
  With bmi8.BITMAPINFO 
    \bmiHeader\biSize     = SizeOf(BITMAPINFOHEADER) 
    \bmiHeader\biWidth    = w 
    \bmiHeader\biHeight   = h 
    \bmiHeader\biPlanes   = 1 
    \bmiHeader\biBitCount = 8 
  EndWith
  
  hdcDest = CreateCompatibleDC_(0)
  
  hImageOut = CreateDIBSection_(hdcDest, @bmi8, #DIB_PAL_COLORS, @ppvbits, 0, 0) 
  SelectObject_(hdcDest, hImageOut)
  GdiFlush_()
  SetDIBColorTable_(hdcDest,0,256,*palette)    
  DeleteDC_(hdcDest)
  
  GetObject_(hImageOut, SizeOf(BITMAP), @bmp8.BITMAP)
  With bmp8
    wb_out =      \bmWidthBytes
    bitsptr_out = \bmBits
  EndWith

  StartDrawing(ImageOutput(image_number))
  max_x = w - 1
  max_y = h - 1
  For y = 0 To max_y
    For x = 0 To max_x
      index.a = NeuQuant::InxSearch(NeuQuant::PointOrdDith(x, y)) ; comment this line out if uncommenting the next 2
      ;thiscolor = GetPixel32BitImage(x, y)
      ;index.a = NeuQuant::InxSearch(thiscolor)
      SetPixel8BitImage(x, y, index)
    Next
  Next
  StopDrawing()
  
  FreeMemory(*palette) 
  ProcedureReturn hImageOut 
  
EndProcedure

;////////////////////////////////////////////////////
;                    TEST CODE
;////////////////////////////////////////////////////

file$ = GetTemporaryDirectory()+"girl.jpg"

If FileSize(file$) = -1
  InitNetwork()
  If Not ReceiveHTTPFile("http://www.lloydsplace.com/girl.jpg", file$)
    MessageRequester("oops!","Problem getting image... quitting")
  EndIf
EndIf

UseJPEGImageDecoder() 

image_number = LoadImage(#PB_Any, "girl.jpg") 

w=ImageWidth(image_number)
h=ImageHeight(image_number)

result = ImageTo8bit(image_number) 

; Let's take a look at the results...

OpenWindow(0,0,0,w,h,"") 
ImageGadget(0,0,0,0,0,result)
Repeat:EventID = WaitWindowEvent():Until EventID = #PB_Event_CloseWindow

This code works fine with the luis port of Dekker's algorithm but this implementation is so much faster I'd like to get it working with 8bit output.

[wilbert]

@netmaestro,
I haven't checked your example but based on what you are saying, I think it might have to do with RGBA vs BGRA .
For Windows internal format, red and blue are switched compared to what Point or PointOrdDith return.

[netmaestro]

That sounds like exactly like a solution. You don't happen to have an asm version of the conversion kicking around somewhere, do you? It would save me an hour's worth of effort and a possible concussion, but it's something you either have already or could do inside of two minutes. If it wouldn't be too much trouble.

[Keya]

here's some i prepared earlier! i would love to know how it compares to wilberts hehe (cunningly give him a challenge where there's not much room for improvement to artificially inflate my ego )

Code: Select all

Procedure RGBtoBGRinplace(addrPixel) ;and vice versa
  CompilerIf #PB_Compiler_Processor = #PB_Processor_x64
  !mov rax, [p.v_addrPixel]
  !mov ecx, [rax]         ;/  Input  \
  !bswap ecx              ;33 22 11 FF -> FF 11 22 33
  !ror ecx, 8             ;FF 11 22 33 -> 11 22 33 FF
  !mov [rax], ecx         ;               \  Output /
  CompilerElse
  !mov eax, [p.v_addrPixel]
  !mov ecx, [eax]
  !bswap ecx
  !ror ecx, 8
  !mov [eax], ecx
  CompilerEndIf
EndProcedure
;eg. RGBtoBGRinplace(@rgbPixel.l)
;for a whole image it's easier to use this than the following version which would also require lots of PokeA's!

Procedure RGBAtoBGRA(pixel) ;and vice versa
  CompilerIf #PB_Compiler_Processor = #PB_Processor_x64
    !mov rax, [p.v_pixel]
  CompilerElse
    !mov eax, [p.v_pixel]
  CompilerEndIf
  !bswap eax
  !ror eax, 8
  ProcedureReturn ;eax/rax
EndProcedure
;eg. swappedrgbPixel.l = RGBAtoBGRA(rgbPixel.l)

[netmaestro]

Thanks to keya, I've implemented the swap.

This results in correct colors, bad quality:

Code: Select all

      color = RGBAtoBGRA(NeuQuant::PointOrdDith(x, y)) ; comment this line out if uncommenting the next 2
      index.a = NeuQuant::InxSearch(color)             ; comment this line out if uncommenting the next 2
      ;thiscolor = RGBAtoBGRA(GetPixel32BitImage(x, y))
      ;index.a = NeuQuant::InxSearch(thiscolor)
      SetPixel8BitImage(x, y, index)

and this results in fine quality, wrong colors:

Code: Select all

      ;color = RGBAtoBGRA(NeuQuant::PointOrdDith(x, y)) ; comment this line out if uncommenting the next 2
      ;index.a = NeuQuant::InxSearch(color)             ; comment this line out if uncommenting the next 2
      thiscolor = RGBAtoBGRA(GetPixel32BitImage(x, y))
      index.a = NeuQuant::InxSearch(thiscolor)
      SetPixel8BitImage(x, y, index)

[Keya]

make sure you're feeding the pixels to the quantizer in the right order too

[wilbert]

Any reason you just aren't using Point ?

Code: Select all

index.a = NeuQuant::InxSearch(Point(x, y))

[netmaestro]

Point(x,y) produces results equal to NeuQuant::PointOrdDith(x, y), nothing to gain there. One applies the dither and the other doesn't. Outcome is the same.

[netmaestro]

make sure you're feeding the pixels to the quantizer in the right order too

Doublechecked, inxSearch is getting the swapped version in both cases.

[wilbert]

You probably need a copy of the palette with red and blue swapped to use for SetDIBColorTable_ .

[netmaestro]

Very good suggestion, but though implementing it did change the results, not for the better. I'll probably have to stick with the luis version, even though it's 3 times slower. I really can't see where I'm making a mistake here. If there was something wrong with my SetPixel8bitImage macro, it wouldn't look like a girl at all. Same with my GetPixel32bitImage macro. Has someone else successfully used this code to affect colors from a truecolor image to an 8bit image? If so, I hope they'll weigh in.

[wilbert]

Does this work for you ?

Code: Select all

IncludeFile "NeuQuant.pbi"

Procedure RGBAtoBGRA(pixel)
  !mov eax, [p.v_pixel]
  !bswap eax
  !ror eax, 8
  ProcedureReturn
EndProcedure

Procedure ImageTo8bit(image_number)
  Protected Dim palette.l(255)
  
  w = ImageWidth(image_number)
  h = ImageHeight(image_number)
  
  NeuQuant::InitNet()
  NeuQuant::Learn(image_number)
  NeuQuant::UnbiasNet(@palette())
  NeuQuant::InxBuild()
  
  For i = 0 To 255
    palette(i) = RGBAtoBGRA(palette(i))
  Next
 
  With bmi8.BITMAPINFO 
    \bmiHeader\biSize     = SizeOf(BITMAPINFOHEADER) 
    \bmiHeader\biWidth    = w
    \bmiHeader\biHeight   = h
    \bmiHeader\biPlanes   = 1 
    \bmiHeader\biBitCount = 8 
  EndWith
  
  hdcDest = CreateCompatibleDC_(0)
  
  hImageOut = CreateDIBSection_(hdcDest, @bmi8, #DIB_PAL_COLORS, @ppvbits, 0, 0) 
  SelectObject_(hdcDest, hImageOut)
  GdiFlush_()
  SetDIBColorTable_(hdcDest,0,256,@palette())    
  DeleteDC_(hdcDest)
  
  GetObject_(hImageOut, SizeOf(BITMAP), @bmp8.BITMAP)
  With bmp8
    nextline_offset = -\bmWidthBytes - w
    *bitsptr.Ascii = \bmBits + (h - 1) * \bmWidthBytes
  EndWith
  
  StartDrawing(ImageOutput(image_number))
  max_x = w - 1
  max_y = h - 1
  For y = 0 To max_y
    For x = 0 To max_x
      *bitsptr\a = NeuQuant::InxSearch(NeuQuant::PointOrdDith(x, y))
      *bitsptr + 1
    Next
    *bitsptr + nextline_offset  
  Next
  StopDrawing()

  ProcedureReturn hImageOut 
  
EndProcedure

;////////////////////////////////////////////////////
;                    TEST CODE
;////////////////////////////////////////////////////

file$ = GetTemporaryDirectory()+"girl.jpg"

If FileSize(file$) = -1
  InitNetwork()
  If Not ReceiveHTTPFile("http://www.lloydsplace.com/girl.jpg", file$)
    MessageRequester("oops!","Problem getting image... quitting")
  EndIf
EndIf

UseJPEGImageDecoder() 

image_number = LoadImage(#PB_Any, GetTemporaryDirectory()+"girl.jpg") 

w=ImageWidth(image_number)
h=ImageHeight(image_number)

result = ImageTo8bit(image_number) 

; Let's take a look at the results...

OpenWindow(0,0,0,w,h,"") 
ImageGadget(0,0,0,0,0,result)
Repeat:EventID = WaitWindowEvent():Until EventID = #PB_Event_CloseWindow

[netmaestro]

YES!!!!
Very tired, on the way to bed but I'll see if I can make sense of it anyway. You've never let me down Wilbert and I can't put into words how much I appreciate your help.

[netmaestro]

Looked through it and mostly understand it. I'm in awe. Very strong coding, thanks again.

[wilbert]

Looked through it and mostly understand it.

Good to hear

The important thing to know is that SetDIBColorTable requires BGRA and NeuQuant::InxSearch requires RGBA.
You can of course also use Point instead of PointOrdDith if you don't want to use dithering. Personally I like the combination of NeuQuant and ordered dithering.
It also works well for animated gifs. Dithering based on error diffusion like Sierra or Floyd-Steinberg doesn't look so good for animated gifs I think.