Open source OpenCL jpeg 2000 library

[boxerab]

Hello List,

I am developing an open source OpenCL jpeg 2000 library.

https://github.com/OpenCodec/ThousandthChicken

Its named after Seymour Cray's famousline:


"If you were plowing a field, which would you rather use: Two strong oxen or 1024 chickens?"

Currently I am focusing on decoding.
If anyone is interested in helping out, please let me know.

Thanks,
Aaron

[Piotr Tarsa]

Some files looks like they're written in CUDA, not OpenCL.

Do you have some benchmarks? Do you have some sample decoded images to compare with reference decoders?

[boxerab]

Test

[boxerab]

Hi Piotr,
Yes, this is a port of a CUDA library, and there is still some unconverted CUDA code.
I have not managed to encode or decode any images yet, so no benchmarks or comparisons.

I am hoping to interest some hackers in helping with the coding. If I just do it myself, it should take another 3 months
before I have a working decoder.

Aaron

[SvenBent]

What I'm curious about is quality. It seems that a lot of cuda/opencl implementations of lossy compressions, gets a even more lossy at same output size.
If quality/size cant keep up, the speed advantage is not really important (well to me at least).

[Piotr Tarsa]

3 months is not that much. There're very few people that can program in OpenCL or CUDA and are free to volunteer so probably you'll finish your project before someone volunteers.

I have some little OpenCL experience but I'm busy with my SortingAlgorithmsBenchmark.

[boxerab]

@Sven: I am not sure I understand. The algorithm should be the same, whether on gpu or cpu.

[boxerab]

@Piotr You're right. But, I've been hacking on this project for 3 months now, so the foundation gpu code has been built already. Right now, it is mostly
getting the j2k algorithm correct. And OpenCL is just C99 with a few extensions, so it is not that difficult to learn. Once the codec is correct,
the really hard part is optimizing it for a particular platform.

[boxerab]

By the way, Piotr, I would like to create an open benchmark for different j2k codecs when I am done.
Nothing like this exists at the moment, so it is hard to tell which codecs provide the best performance and quality.

[SvenBent]

some info on jp2k encoders benchmarks here
http://compression.ru/video/codec_co...arison_en.html
Seems like the acdsee and lurawave deals the best in psnr values.

[boxerab]

Thanks, Sven. Very interesting. Their test images will be useful.
It would be interesting to extend the test to cover larger images such as 4K digital cinema frames.

[cade]

Yes, this is a port of a CUDA library, and there is still some unconverted CUDA code.

Is there any new part, or is this project only to convert an existing CUDA library into OpenCL?

[boxerab]

@cade - first step is to convert existing CUDA library and get correct encode/decode. Then, next step is to optimize for various platforms such as Intel and AMD (nVidia has terrible OpenCL drivers, for obvious reasons, so can't target them unfortunately). I would like to see if I can encode 4k DCP in real time on inexpensive hardware.

Then, I will experiment with different ideas being proposed in papers for parallelizing some of the pieces of the j2k algorithm. For example, papers by Jiri Matela.