Compilation problem with ModuleMapCuda.cu on aarch64 with CMAKE_CUDA_STANDARD 23

[elmsheus]

There is a compilation problem with Acts v47.0.0 in Plugins/Gnn/src/ModuleMapCuda.cu on the Athena,aarch64-el9-gcc15 platform using CMAKE_CUDA_STANDARD 23 and cmake 4.4.0 as done in the atlasexternals MR https://gitlab.cern.ch/atlas/atlasexternals/-/merge_requests/1420

It works fine for the platform Athena,x86_64-el9-gcc15 in the same MR.

There is not much in the CI pipeline logs at https://gitlab.cern.ch/atlas/atlasexternals/-/jobs/78870939/raw apart from:

2026-07-17T23:04:20.685409Z 01O Compilation aborted.
2026-07-17T23:04:20.685416Z 01O gmake[5]: *** [Plugins/Gnn/CMakeFiles/ActsPluginGnn.dir/build.make:164: Plugins/Gnn/CMakeFiles/ActsPluginGnn.dir/src/ModuleMapCuda.cu.o] Error 1

It's reproducible locally when building Acts in this atlasexternals setup.

The problem disappears when using CMAKE_CUDA_STANDARD 20 or -DACTS_GNN_ENABLE_MODULEMAP:BOOL=OFF with CMAKE_CUDA_STANDARD 23 - this is obviously just working around the problem.

Are we missing something in the setup for aarch64 in atlasexternals or is there some aarch64 bug in ModuleMapCuda.cu ? Or should we actually use -DACTS_GNN_ENABLE_MODULEMAP:BOOL=OFF on aarch64 ?

Can you please have a look ?

Cheers, Johannes

[andiwand]

tagging @benjaminhuth

[elmsheus]

The same error happens with Acts v47.2.0, and v47.4.0 - see again https://gitlab.cern.ch/atlas/atlasexternals/-/merge_requests/1420

Did you get a chance to look into this ? Should I use -DACTS_GNN_ENABLE_MODULEMAP:BOOL=OFF in atlasexternals ?

[andiwand]

tagging @benjaminhuth @paulgessinger

[paulgessinger]

As far as me and Claude are able to tell, this is a compiler bug/incompatibility. Are we able to turn the module map build off in just the aarch64 build?

[elmsheus]

Something like the following would switch it on/off depending on the architecture (untested so far):

diff --git a/External/Acts/CMakeLists.txt b/External/Acts/CMakeLists.txt
index e5c0da35..47d1b6be 100644
--- a/External/Acts/CMakeLists.txt
+++ b/External/Acts/CMakeLists.txt
@@ -74,11 +74,19 @@ if( ATLAS_ACTS_USE_GNN )
         -DACTS_GNN_ENABLE_ONNX:BOOL=ON
         -DACTS_GNN_ENABLE_TORCH:BOOL=OFF
         -DACTS_GNN_ENABLE_TENSORRT:BOOL=OFF
-        -DACTS_GNN_ENABLE_MODULEMAP:BOOL=ON
         -DMMG_BUILD_EXECUTABLES:BOOL=OFF
         -DCUDAToolkit_ROOT:PATH=${CUDAToolkit_LIBRARY_ROOT}
         -DCMAKE_CUDA_COMPILER:FILEPATH=${CMAKE_CUDA_COMPILER}
         -DCMAKE_CUDA_ARCHITECTURES:STRING=${CMAKE_CUDA_ARCHITECTURES} )
+
+   # Switch ACTS_GNN_ENABLE_MODULEMAP on/off depending on architecture
+   # https://github.com/acts-project/acts/issues/5705
+   if( "${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "x86_64" )
+     list( APPEND _extraOptions -DACTS_GNN_ENABLE_MODULEMAP:BOOL=ON )
+   elseif( "${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "aarch64" )
+     list( APPEND _extraOptions -DACTS_GNN_ENABLE_MODULEMAP:BOOL=OFF )
+   endif()
+
 endif()
 
 # Optional traccc package.

Is this something to try out ?

[elmsheus]

I've added the suggested cmake configuration update from the previous comment in switching off ACTS_GNN_ENABLE_MODULEMAP on aarch64 in MR https://gitlab.cern.ch/atlas/atlasexternals/-/merge_requests/1420 and the CI succeeds.

Could this lead to physics differences in ACTS production code between x86 and march64 or does this only affect current R&D activities ?

[paulgessinger]

We don't run any of the gnn pipeline in production configurations as far as I know, but I've recently found out that the gnn pipeline seems quite sensitive to hardware and architecture changes in General.