Profiler import failed

Hi all,
using nvprof (CUDA 9.0) I profiled an MPI+OpenACC code running on a Power8+P100 system in this way:

mpirun -np 4 --map-by ppr:2:socket --report-bindings nvprof -f -o mycode.%q{OMPI_COMM_WORLD_RANK}.nvprof ./mycode

It seems to correctly produce the output files, but when trying to import them on an x86 system using computeprof (or pgprof), importing them as multiple processes, I get this error:

Unable to import nvprof generated profile data.
OpenACC activities parsed out of order.

What does this mean? Should I add some flags when profiling the application?
Importing just one file it actually works, but it is not possible to see all the processes in the same view.

I also noticed that from time to time, when using nvprof, it crashes with “address not mapped” errors, while the code always runs fine without the profiler.




This is a known issue with OpenACC data in the profiler. You will need to import and view each process separately.