This is a playground for getting nvjpeg2k decoding into numcodecs to benchmark if tile loading for deeplearning can be significantly sped up when decoding directly on the GPU.
Currently, this is a bit of a playground for testing feasibility and establishing good benchmarks. But feel free to open issues in the issue tracker in case you would like to help and or contribute ❤️
Currently no wheels or conda packages. To get started you need cuda and a cupy installation. Then clone the repo and run:
cd vendor
./get_nvjpeg2k.shafter that you should be able to run the following in the repository root:
pip install .Please use pre-commit unless it interferes with your workflow too much.
I'm happy to get contributions, so don't let linting be in the way of that.
Build with love by Andreas Poehlmann.
nvjpeg2k_numcodecs: copyright 2022 Andreas Poehlmann, licensed under MIT