24.9 Internal compiler error

Immediately downloaded the new HPC SDK 24.9, thanks for the update. When installing a new compiler, I generally create the libraries I usually need for my development. These contain openmpi (currently 5.0.5). Although not really needed for nvc++ (it contains its own MPI), my script does so anyway, and this time I got an internal compiler error:

NVC+±S-0000-Internal compiler error. expand_smove/generate_ldst, unexp. dt 13 (…/…/…/openmpi-5.0.5/opal/datatype/opal_datatype_pack.c: 459)
nvc-Fatal-/opt/nvidia/hpc_sdk/Linux_x86_64/24.9/compilers/bin/tools/nvcpfe TERMINATED by signal 11

It is not a problem for me, as there is builtin-mpi, as stated, but I thought perhaps you are interested in an example generating an internal compiler error.

Ciao,
Manuel

Hi Manuel,

Very sorry you encountered this. Can you send me the options you passed to ./configure when building Open MPI 5.0.5? I need to see if I can reproduce the issue here.

Thanks,

+chris

Sure. Nothing very special:
…/openmpi-5.0.5/configure --enable-mpi-fortran=all --enable-static --prefix=/opt/libs/nvhpc-24.9/openmpi/5.0.5
This was on two different system, but both running openSUSE Leap 15.5. I’ll try to reproduce on a Rocky Linux 9 I have access to as well.
Thanks for looking into this!

Ciao,
Manuel

We‘ve ran into the same issue with Rocky 8, RHEL 9 (on aarch64) and Ubuntu 22.04, also with OpenMPI 5.0.3 for example. OpenMPI 4.1.6 works fine. You don‘t need any special configure flags (only the NVHPC compilers).
We planned to do a new EasyBuild toolchain on our HPC systems based on NVHPC 24.9 and OpenMPI 5.0.5, but we might need to use 24.7 instead. That‘s unfortunate, since several issues are resolved we ran into previously

I got the recommendation to use --enable-alt-short-float=no as a workaround when configuring OpenMPI. With this, NVHPC 24.9 can build OpenMP 5.0.5 on Ubuntu 22.04 (x86_64) just fine. I’m currently testing if it also works with EasyBuild on Rocky 8 and RHEL 9, but I expect that it works as well.

1 Like

I checked around with our developers, and this is a known issue with 24.9. A fix is in the works. Use the workaround for now.

Thanks,

+chris

Confirmed on openSUSE Leap 15.5, compiles successfully with said workaround. Not yet tested for functionality.