Thorsten Hater

So, here's what I did: - copy out the `g++` line that's failing - you'll get this by `make VERBOSE=1` - remove all the options like `-MT` and similar plus...

Hey, we have just merged #1961 which denies building with incompatible Pybind11 and this issue has work-around identified. I am closing this, if still problematic, please reopen.

Except #1987 this is done.

Re: True, but that was the motivator. Rest: all HPC systems we have access to are on CUDA11 now and we never liked GCC8 anyhow.

This should absolutely be part of 0.8, these compilers are starting to reek.

YES! It's been orverdue to do that.

Hi @Helveg, I think that's not entirely correct: The catalogue can be built with GPU support and used in a simulation linking a CPU-only libarbor. So, in theory, such a...

That's also not entirely correct ;) We have at least one system where one head node manages multiple backend, some without GPU, some with NV GPUs and some with AMD...

Ah, here comes an ugly truth: We can only ever support - one CPU backend (arch+SIMD=0/1) - one GPU backend (HIP/CUDA) so the a-b-c tool needs to be coupled to...

Thanks @w-klijn for the suggestions, I have read the CSA paper and looked into the NML-style population/projection based approach. To solve this issue for Arbor, I will suggest a DSL...