Using eddy_cuda when applying dwifslpreproc in MicaPipe

[PaulBautin]

What is the current behaviour?

• From my understanding, MicaPipe disables by default the use of CUDA with line export CUDA_VISIBLE_DEVICES=''" in $MICAPIPE/functions/utilities.sh script.

What is the motivation / use case for changing the behaviour?

• Certain MRtrix dwifslpreproc eddy functions only work on GPU (ex: slice-to-volume correction).

Solution:

• The micapipe FSL version requires cudatoolkit version 9.1. This can be installed in the micapipe conda environement with the command: conda install numba::cudatoolkit (latest version of cudatoolkit on numba is v9.1).
• A pointer to toolkit in the conda environment must be addded with command export LD_LIBRARY_PATH="$CONDA_PREFIX/lib:$LD_LIBRARY_PATH"
• Comment this line export CUDA_VISIBLE_DEVICES=''" in $MICAPIPE/functions/utilities.sh script.

Other information
I think the solution could be implemented on the local MicaPipe env (keeping cudatoolkit version to eddy_cuda/FSL requirements). The use of CUDA could continue to be disabled by default and a different dwifslpreproc command could be used when flag -use_cuda is enabled.

• without flag -use_cuda:
  dwifslpreproc "$dwi_4proc" "$dwi_corr" $opt -pe_dir "$pe_dir" -readout_time "$ReadoutTime" -eddy_options " --data_is_shelled --slm=linear --repol " -nthreads "$threads" -nocleanup -scratch "$tmp" -force

• with flag -use_cuda:
  dwifslpreproc "$dwi_4proc" "$dwi_corr" $opt -pe_dir "$pe_dir" -readout_time "$ReadoutTime" -eddy_options " --data_is_shelled --slm=linear --repol --mporder=# --nvoxhp=4000" -nthreads "$threads" -nocleanup -scratch "$tmp" -force