Jan Stephan

`prepareForAsyncCopy` needs to be changed internally, too. I guess that it should be enough to check for the number of CUDA devices in the system and only act if that...

@BenjaminW3 Do you still think we should investigate this?

> isn't this exactly what the Fibers backend is supposed to do ? Yes, Coroutines and Boost.fiber are quite similar. However, three of our developers failed to set up the...

> I really find it ... disarming, for lack of a better word, how much you dislike having dependencies on boost (or anything else, for that matter). That mostly affects...

> if "HPC modules are notoriously out of date" why do you care so much about "a C++ standard that is two years old" (i.e. c++20) which so far is...

Also, if we want to discuss Boost specifically, please continue here: #481.

I think this is related to @psychocoderHPC's idea of accelerator tags which would somewhat decouple the interface from the back-ends: #1246. Do you think that approach would work for you?

#1766 adds clang-14 as CUDA compiler to the CI. Once that is merged this should be next.

While addressing the requested changes I noticed that full support for clang-cuda and `CUDA_SEPARABLE_COMPILATION` is only available in CMake 3.19. I can either do evil things and try to override...

I'm considering to shift this PR to 0.9 because of #1509. @SimeonEhrig @psychocoderHPC Do we require this feature for 0.8?