Quick OK Image format

[ebs]

I would expect any changes to the file format would require only minor code changes, which I'm sure the author would provide.

Right now it works. I'll wait until the format is finalized before using it in any projects.

[BarryG]

So how do we view the images? I tried ebs's code but it results in a ".qoi" image format.

[ebs]

You can't view QOI-format images directly, as there is no image viewer for that format (at least none that I know of).
The original author's C code (and my PureBasic translation) provides routines to encode raw bitmap data into a QOI-format file, and to decode a QOI-format file back into raw bitmap data.

The test stub I added shows how to convert between BMP image files and QOI-format files. The original author presented a similar program using PNG images. I used BMP images because they don't use compression. This made it easier to compare the original BMP image to one that had been converted to QOI-format and then back to BMP, to verify that my code was working properly. Using the test code, you can convert a BMP image to QOI-format and then back to BMP.

The code provides all the tools necessary to write your own QOI-format image viewer in PureBasic, if desired.

[infratec]

@ebs

I also started a PB version, but you were faster

You use always PeekA(), which is slow, because it is a procedure call with the usual additional stuff of push and pop some registers.
Faster is to access the memory without a call:

Code: Select all

Procedure.l qoi_read_32(*bytes.Ascii)
  
  Protected.a a, b, c, d
  
  a = *bytes\a
  *bytes + 1
  b = *bytes\a
  *bytes + 1
  c = *bytes\a
  *bytes + 1
  d = *bytes\a
  *bytes + 1
  
  ProcedureReturn (a << 24) | (b << 16) | (c << 8) | d
	
EndProcedure

I also eliminated the pos pointer p, because I directly increase the pointer itself.

I'm just finished with my version, but it is untested. If I have some time over the weekend, I will test it and write a LoadQOImage()
and a SaveQOIImage() procedure which is compatible wth PB.


Maybe this will work too:

Code: Select all

Structure rgba_Structure
  r.a
  g.a
  b.a
  a.a
EndStructure

Structure qoi_rgba_t
  StructureUnion
    rgba.rgba_Structure
    v.l
  EndStructureUnion
EndStructure

Procedure.l qoi_read_32(*bytes.qoi_rgba_t)
  
  Protected Result.l
  
  Result = *bytes\v
  *bytes + 4
  
  ProcedureReturn Result
	
EndProcedure

[ebs]

@infratec,

Your direct access to the bytes is much better than my Peek/Poke method.

I noticed that the 'qoi_read32()' and 'qoi_write_32()' procedures aren't really that important,
since they're used only to read/write the qoi image header. The real work is done in the
encode/decode routines. Your code should make those routines much faster.

Please post your code when it's done!

Regards,
Eric

[wilbert]

I would expect any changes to the file format would require only minor code changes

I don't know.
The code from the experimental branch is quite a bit different in some places and certainly not compatible with the original code.
https://github.com/phoboslab/qoi/tree/experimental
It depends on what will end up in the final specification.

[infratec]

I tested now my version and it works. (only channel 3 stuff up to now)
Your listing gives not the same result.

I build the original qoiconv.exe withe Pelles C.
Then I converted a small PNG file to OUI.

I load this OUI file with my test and it is shown as it should.
If I use your version, I see a wrong picture

Here is my version:

CODE REMOVED SINCE FINAL SPECS ARE NOW AVAILABLE

The saved OUI file is identical to the loaded file.
In my case the original PNG file is 11kB and the OUI file is 44kB.
But I reduced the PNG colors to 16 and use compression level 9.
In this case is the size definately larger.

[wilbert]

The saved OUI file is identical to the loaded file.
In my case the original PNG file is 11kB and the OUI file is 44kB.
But I reduced the PNG colors to 16 and use compression level 9.
In this case is the size definately larger.

It's not expected to beat PNG compression when it comes to the compressed size.
The advantage is mainly speed and the simplicity of the code.
For a 16 color PNG it can make quite a difference what colors are used. The color hash function from the original code is a simple xor of the red, green and blue values. Maybe there's a lot of collisions with your color palette.

[infratec]

LoadImageQOI()

CODE REMOVED SINCE FINAL SPECS ARE NOW AVAILABLE

[ebs]

@infratec,

Do you have somewhere that you can post the 1) EXE you built from the C code, 2) the original image you used, and 3) the QOI file you get from the EXE ?

I can encode/decode using my version, but that doesn't mean that it is working as the original code intended.

Thanks,
Eric

[ebs]

@infratec,

I found out why my image doesn't show correctly with your code. You mentioned it: "Unfortunately the internal format in windows is BGR."

Based on that, either of the two changes below should show a QOI image encoded with my code and decoded with your code correctly.

1. Replace your 'Plot()' calls with 'CopyMemory()':

Code: Select all

    CreateImage(0, desc\Width, desc\Height, 8 * desc\channels)
    If StartDrawing(ImageOutput(0))
        *pixel = *img
        Select desc\channels
          Case 3
            *ImageAddress = DrawingBuffer()
            CopyMemory(*pixel, *ImageAddress, desc\Width * desc\Height * desc\channels)
            
            ; For y = 0 To desc\Height - 1
              ; For x = 0 To desc\Width - 1
                ; Plot(x, y, RGB(*pixel\rgba\r, *pixel\rgba\g, *pixel\rgba\B))
                ; *pixel + 3
              ; Next x
            ; Next y

or
2. Change the order of the colors in your 'Plot()' calls:

Code: Select all

    CreateImage(0, desc\Width, desc\Height, 8 * desc\channels)
    If StartDrawing(ImageOutput(0))
        *pixel = *img
        Select desc\channels
          Case 3
            For y = 0 To desc\Height - 1
              For x = 0 To desc\Width - 1
                Plot(x, y, RGB(*pixel\rgba\B, *pixel\rgba\g, *pixel\rgba\r))
                ; Plot(x, y, RGB(*pixel\rgba\r, *pixel\rgba\g, *pixel\rgba\B))
                *pixel + 3
              Next x
            Next y

Since you did the comparison of your images with images generated from the original C code, your version must be the right one.

[infratec]

@ebs:
I think you did it the wrong way.
In the QOI file it is RGB.
So the Plot with R G B is correct, since I use the RGB macro.

Original (master) qoiconv:

https://www.von-der-salierburg.de/downl ... oiconv.exe

The original png file:

https://www.von-der-salierburg.de/downl ... IP_IMA.png

The qoi file:

https://www.von-der-salierburg.de/downl ... IP_IMA.qoi

[ebs]

Thanks - please see my message immediately above yours.

[infratec]

I read it already.
Look my first lines above.

I uploaded the complete quiconv stuff.
Simply download and install Pelles C V11
Then you can compile it by yourself.

https://www.von-der-salierburg.de/downl ... OIConv.zip

[ebs]

Our messages keep crossing on the internet! Yes, I agree that your code is the right way.

@ebs:
I think you did it the wrong way.
In the QOI file it is RGB.
So the Plot with R G B is correct, since I use the RGB macro.

I think that's the difference:
Your 'QOI_COLOR_HASH()' macro used the structure r/g/b/a elements, while mine used PB procedures 'Red()/Green()/Blue()/Alpha()', which probably accounts for the difference.

Thanks for working through this with me.

Regards,
Eric