search

FMRIB-Software-Library / support

Public FSL issue tracker and support portal
4 stars 0 forks source link

bedpostx_gpu #4

Open hanayik opened 3 years ago

hanayik commented 3 years ago

I have two questions related to bedpostx_gpu -

  1. I am preparing to run bedpostx_gpu and probtrackx2_gpu using an NVIDIA Titan V CUDA card on a Linux machine running openSUSE LEAP 15.2. I have the CUDA 11.0 graphics drivers and libraries installed. I also have the CUDA 10.1 libraries installed. I saw that bedpostx_gpu and probtrackx2_gpu are only compatible with CUDA versions 10.2 and below, so I downloaded the bedpostx_gpu and probtrackx2_gpu binaries for CUDA 10.1. However, now when I try to run bedpostx_gpu, I get the following error:
getopt: unrecognized option '--has_coprocessor=cuda'
No GPU configured/found

I've previously been able to run bedpostx_gpu on this machine but have not run it since the upgrade to CUDA 11.0. Is it a problem if I'm running the graphics drivers from CUDA 11.0?

  1. I have two different DTI datasets - one with a single shell and 64 directions, and the other with two shells (42 directions and 83 directions). I will be analyzing them separately for different studies. When running bedpostx_gpu on the first (single-shell) dataset, should I use -model=1? And when running on the second (two-shell) dataset, should I use -model=2?
hanayik commented 3 years ago

It looks like you may be using an internal version of the bedpostx_gpu script. You should be able to restore the previous script with:

cp $FSLDIR/src/fdt/CUDA/bedpostx_gpu $FSLDIR/bin/bedpostx_gpu

For the second question, you should definitely use model 2 for the multi-shell data.