Test routines using nvtx

This repository comprises a toy example of using the nvtx profiler to analyse Fortran code.

It largely follows the structure of an NVIDIA tutorial.
An updated nvtx module nvtx.cuf comes from maxcuda, with the original moved to nvtx.cuf_old.

module purge
module load rhel8/global rhel8/slurm
module use /usr/local/software/spack/spack-modules/a100-20210927/linux-centos8-zen2

nvfortran -o nvtx_test main.f90 nvtx.cuf -L/usr/local/software/cuda/12.1/lib64 -lnvToolsExt

./nvtx_test

We rename the nvtx mofule file to have an .f90 extension, otherwise gfortran can't interpret it.
We also need to load cuda as a module - it isn't loaded as a dependency of nvhpc and the executable needs the .so to be in the LD_LIBRARY_PATH which is done at load time.

module purge
module load rhel8/global rhel8/slurm
module use /usr/local/software/spack/spack-modules/a100-20210927/linux-centos8-zen2
module load cuda/11.4.0

gfortran -o nvtx_test nvtx.f90 main.f90 -L/usr/local/software/cuda/12.1/lib64 -lnvToolsExt

./nvtx_test