Other than what’s in the user’s guide, I don’t have a list of restrictions. Other than the normal device code restrictions, there not much difference than what can be done in the “global” versus what can be done in “device”. Is there a specific issue that you’re encountering?
For examples, we have many general CUDA Fortran examples in “$PGI/
pgi/linux86-64/2015/examples/CUDA-Fortran/”. Though, it appears that we don’t have much for “device” routines except for the trivial example in “SDK/sgemm/sgemm.cuf”.