We have a license for the Allinea DDT debugger for CUDA. To debug GPU code, you need to be logged into (via a graphics capable connection) 'knot-gpu' which is a login node with two GPUs.
To use DDT, you must first include the Cuda libraries in your path (although you may already have it if you are using Cuda)
(bash users - the default) export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH or (CSH users) setenv LD_LIBRARY_PATH /usr/local/cuda/lib64:$LD_LIBRARY_PATH
You then start the program with /sw/bin/ddt
Uncheck the MPI box, and check the 'CUDA' box (if the cuda box is greyed out, make sure you're logged into Knot-gpu - and not Knot), then you can run your job.