Issue on LUMI update

[StevE-Ong]

Hi, LUMI has undergone a major update.

ROCm 6.3.4 is now the system-installed and default ROCm version and ROCm 6.0.3 is no longer on the system.

I am trying to update the picongpu.profile as follow:

module load init-lumi/0.2
module load LUMI/25.03  partition/G
module load buildtools/25.03

export MPICH_GPU_SUPPORT_ENABLED=1
# Compiling with amdclang from rocm module
# Note, independent of the PrgEnv, all compilers are wrapped via cc (C-compiler) and CC (C++ compiler)
module load PrgEnv-amd
module load craype-accel-amd-gfx90a
module load rocm/6.3.4
module load Boost/1.88.0-cpeAMD-25.03

There are some error during the compilation, in which I am not sure if it is the problem of my loaded module, or any setting I miss out.

CMake Warning (dev) at /opt/rocm-6.3.4/lib/cmake/hip/hip-config-amd.cmake:91 (message):
   AMDGPU_TARGETS was not set, and system GPU detection was unsuccsesful.
   
   The amdgpu-arch tool failed:
   Error: 'Failed to get device count'
   Output: ''
   
   As a result, --offload-arch will not be set for subsuqent
   compilations, and the default architecture
   (gfx906 for dynamic build / gfx942 for static build) will be used

and also:

CMake Warning at /users/ongjianf/picongpu/thirdParty/alpaka/cmake/alpakaCommon.cmake:615 (message):
  HIP 6.3 is not official supported by alpaka.  Supported versions: 5.1 - 6.2

I am using the dev version of picongpu and compile the FoilLCT example.

compile (1).log

picongpu (1).profile.txt

[BrianMarre]

We have run into a similar problem and are still in the process of figuring out what goes wrong.
@chillenzer anything new on that front?

[chillenzer]

Yes, I'm already working on this one. It's a know problem and recent LLVM versions (starting from 21.8.0 persumably) have a patch for this: llvm/llvm-project#136133

Unfortunately, those ROCm versions are based on older LLVM versions that don't include it, so I'm currently trying to backport it into the mallocMC codebase directly.

[chillenzer]

Hi,
I've implemented a custom span to get rid of the standard library dependency. Unfortunately, I wasn't able to reproduce the issue locally. You could try from this branch to

• either copy the content of that branch into your PIConGPU's thirdParty/mallocMC directory and try to compile PIConGPU
• or run the tests of mallocMC (our custom GPU memory allocator) like this

export SRC_DIR="/wherever/you/want"
export BUILD_DIR="/somewhere/else"
git clone https://github.com/chillenzer/mallocMC "$SRC_DIR"
git -C "$SRC_DIR" checkout origin/tmp-fix-array-assert
cmake -S "$SRC_DIR"/test -B "$BUILD_DIR"
cd "$BUILD_DIR"
cmake --build . -j

If you should try, please report back, so we can bring this mainline if successful.

[StevE-Ong]

Hi, Thanks for the info. I have difficulty to clone the page today or any git action. Will come back again if I succeed.

[chillenzer]

There's a PR fixing this in mallocMC: alpaka-group/mallocMC#297
Once it's mainline in mallocMC, we'll pull it into PIConGPU. Latest dev should contain the fix in the very near future.

[chillenzer]

For the impatient, you could copy the content of that branch into thirdParty/mallocMC of your local PIConGPU clone and get going.

[chillenzer]

As there was some confusion about what needs fixing here:

• The CMake warnings pasted into the description by @StevE-Ong are misleading. We should probably fix them at some point but they are not failing the compilation. As @ikbuibui has pointed out, they are related to this issue.
• Looking into the logs attached to the issue, it's this problem that's actually causing the compilation error. That's what Update mallocMC #5618 is addressing.