Error running offload-template example program

[Colin-Brennan-ema]

Hello!

I am attempting to incorporate Celeritas within an existing Geant4 application. I'm running into some issues I was hoping to get some support with. Within my application, I am getting floating point exceptions upon first offloading tracks to the GPU - when taking a step back, I found that I get the same exceptions when building and running the offload-template example. I've attached the console output.

offload-template-error.txt

I think it may be due to some oddities with how I built. I used Spack to install Celeritas and all dependencies, but had to make some changes to the default configuration -

• For compatibility with my application, I had to use the compiler flag -GLIBCXX_USE_CXX_11_ABI (for VecGeom, I had to make a patch to pass this flag through nvcc in the CMAKE_CUDA_FLAGS)
• I had some combination of CMake and compiler errors (I wish I could recall, I was working on this about a month ago) trying to build the default versions of g4vg and VecGeom, so I switched to build g4vg@develop and vecgeom@1.2.11

My environment may also be odd... I'm running on an Ubuntu image through WSL on Windows. I have Cuda 12.9 with an RTX 4050, Celeritas was built with cuda_arch=89, which I think matches.

Thank you,

Colin

[sethrj]

Hey @Colin-Brennan-ema , that is a new one! Could you try to run with export CELER_DISABLE_DEVICE=1 (see environment variables) and see if that fails? That will at least tell us whether it's a mismatch of CUDA. Any additional information about your build environment, including the output from celer-sim --config, would be helpful. Thanks!

[Colin-Brennan-ema]

Hi Seth, thank you for help with this!

Error does not occur when running with CELER_DISABLE_DEVICE=1.

Output from celer-sim --config

{"_category":"system","_label":"build","config":{"build_type":"release,shared","core_geo":"VecGeom","core_rng":"xorwow","debug":false,"gpu_architectures":"89","hostname":"Ogore","openmp":"disabled","real_type":"double","units":"CGS","use":["cuda","geant4","vecgeom"],"vecgeom":"Scalar","versions":{"CLHEP":"2.4.7.1","CUDA":"12.9.41","G4VG":"1.0.4-4+main.d377c26","Geant4":"11.3.2","Thrust":"1.17.2.0","VecGeom":"1.2.11"}},"version":"0.6.0"}

[sethrj]

That all looks reasonable. Is there any chance of getting a debugger to give any insight into the error? (Floating point exception may be a garbage error message...) Also, do you have the configure output from Celeritas (or can you reconfigure?) to ensure you don't have a warning like 'CUDA architecture types betweeen VecGeom and Celeritas should probably match: runtime errors may result'?