Would like to add 12bit greyscale support, where to start?

[smlehbleh]

Hello,
I'd like to add 12bit greyscale support if its not already on the horizon. I'm still trying to figure out how difficult a task it is.
Any pointers on where to start?

Cheers
Russell

[MartinPulec]

Hi Russel,

no, unfortunately we do not plan this. As for you question, if you check https://www.w3.org/Graphics/JPEG/itu-t81.pdf and look up the phrase "12-bit", you'll see, what is different against 8-bit. GPUJPEG already supports one channel grayscale 8-bit. So, from my point of view, there will be needed to implement following for encoder:

1. preprocess - it depends on the packing of the source image, it may be needed to fit it to more "sparse" format (eg. 16-bit int) in order to allow DCT access the samples in more reasonable way (12-bit alignment is quite unsuitable for CUDA)
2. DCT - is computed in float, so there should be sufficient to read the sample from the format from postprocess
3. quantization - there will be perhaps more to do - there are more quantization tables, are different etc. While 8-bit source
4. also huffman encoder would need to be extended

For decoder it should be more or less conversely.

Honestly, I am not the original author so I am not as much familiar with JPEG but the difficulty may depend on your experiences with the concepts or if you need to study them - it is not so hard but it will require some work (it is not the case that eg. extending the preprocessor is enough). Knowledge of CUDA is only secondary.

Regards,
Martin