Hello,
I'm trying to run eddy_cuda 8.0 on some older DTI scans using fieldmaps that were not generated by topup (they are the standard Seimens stock scans-2 magnitude, 1 phase). I am also submitting this to our remote cluster via slurm. I have run the eddy_cuda8.0 command on all subjects without the --field option and it ran fine on everyone. When I use the --field option I get the following error for some subjects:

CUDA error after call to EddyKernels::spline_interpolate
Error message: an illegal memory access was encountered

Do I just need to request more memory when I submit my job to our cluster or is this a different issue? I've posted the eddy command below for completeness:

this is not an indication that you are running out of memory, but that there is a problem with your fieldmap. The fieldmap is essentially a pixel-shift map for the PE-direction. If these pixel-shifts are large it may mean that the new pixel location ends up outside of the original FOV. That in itself will not cause this crash _unless_ the new location lies more than a whole FOV outside of the original FOV. That is a very large shift, and something that shouldn’t in reality ever happen, so it implies that the fieldmap has some extreme values or has been scaled wrongly.

Jesper

eddy_cuda8.0 --imain=../dti/${SUBJ}_dti --mask=../dti/${SUBJ}_dti_brain_mask --field=${SUBJ}_fieldmap_flirted_HZ --acqp=../../../scripts/acqparams.txt --index=../../../scripts/index.txt --bvecs=../dti/${SUBJ}.bvec --bvals=../dti/${SUBJ}.bval --repol --out=${SUBJ}_eddy_corr --mporder=6 --slspec=../../../scripts/slspec.txt --s2v_niter=5 --s2v_lambda=1 --s2v_interp=trilinear

Thanks!
Amanda

--

Amanda Rodrigue, PhD

Neurocognition, Neurocomputation, and Neurogenetics (n3) Division
Yale University School of Medicine
40 Temple Street, 6E
New Haven, CT 06510