René Widera

An objective should be oriented toward real-world use cases you try to solve. What is your use case, why you looked specifically into mixing native CUDA/HIP with alpaka?

> Right now this reuses a lot of collision code simply by copying files and also somewhat modifying them. We should unify this, for example, by generalizing the collision kernels....

note: the behaviour of `sycl::ext::oneapi::group_local_memory_for_overwrite` has changed with onapi 2025.2. Allocation of the same data type is providing aliased memory equal to the cuda behaviour. I found it in alpaka3...

Since we have a unique id anyway we can avoid using the undocumented forceinline which is most likly more stable. With 2025 2 we can therefore avoid using the lookup...

I will open an issue in alpaka3 about it and link it to you tests and comments. Currently all works in alpaka3 even with extension on nvidia and amd. Never...