Angle-aware filtering with optional GPU acceleration

[CHrlS98]

Hey!
New script for computing angle-aware bilateral filtering (my ISMRM 2022 abstract). The algorithm can be applied sequentially, across multiple CPU threads or on the GPU (given that pyopencl is installed).

I didn't add pyopencl to the requirements to avoid introducing an additional dependency, but maybe it is something we should consider at some point, as GPU acceleration greatly increases compute time; on fODF from a full brain (HCP with Tractoflow), it takes 12 minutes versus 90 minutes to process using default parameters.

The PR also contains changes to scil_execute_asymmetric_filtering.py and its associated modules to avoid confusion. Also, for both methods, it is now possible to output the asymmetric and the symmetric signal during the same execution.

I put the OpenCL kernel file directly in the denoise module with the python files. Let me know how you feel about this.

I'll also add tests really soon, for the sequential case at least. You can test it with the FiberCup, it runs pretty fast.

To do:

• Add tests;
• Documentation for opencl_utils.py.

Thank you.

[scil-admin]

Build Failed 💥
https://jenkins.scil.usherbrooke.ca/job/scilpy/job/PR-525/1/

[scil-admin]

Build passed ! Good Job 🍻 !

[scil-admin]

Build passed ! Good Job 🍻 !

[scil-admin]

Build passed ! Good Job 🍻 !

[arnaudbore]

A couple of remaining docstring missing otherwise LGTM.

[arnaudbore]

Please add some docstring when needed

[scil-admin]

Build passed ! Good Job 🍻 !

[scil-admin]

Build passed ! Good Job 🍻 !

[mdesco]

Testing this PR at the moment on my Linux Desktop.

1. It completely freezes my display as it runs if I use --use_gpu. Is this expected? Essentially, it would be impossible for me to work on my machine at the same time.
2. Is there a way that we could run the pip install pyopencl for the user?
3. What is the htop or top way to see what the GPU is doing? (I'm novel to this)

Otherwise, it seems to work...

[CHrlS98]

@mdesco

1. It does slow down the PC, particularly for softwares also using the GPU. However, I never experienced any of my PCs completely freezing. I guess it could happen if all of your GPU memory were taken by the process.
2. It isn't added to the requirements.txt as I didn't want to introduce any additional dependencies. We could imagine having an optional requirements file that the user could manually install with pip install -r optional.txt but since there would only be one dependency inside, it's not that better than simply running pip install pyopencl.
3. For a NVIDIA graphics card, the command is watch nvidia-smi.

What is the size of the dataset you used for testing? Does the output look okay even though it freezes everything?

[CHrlS98]

@mdesco
It is possible that your PC froze because the wrong GPU was selected for the task (if you have more than one GPU, that is).

Therefore, I added a check for selecting the most suitable GPU based on the available memory. I will need to test on a PC with 2 or more GPUs to be sure it works, though.

[scil-admin]

Build passed ! Good Job 🍻 !

[mdesco]

All good! Good job @CHrlS98