[torrance]

If your project depends heavily on general purpose GPU programming, you might start one in C++.

This was the case for a project I am working on that was started in the last year. The interop features in rust (and other languages) are simply not as reliable as writing kernels directly in CUDA or HIP or even DPC++. You _can_ attempt to write the GPU code in C++ and call to this from $LANG via ffi, but if you want to preserve data structures and methods to work on both the host and device, its still easier to write it once in C++.

[the__alchemist]

I concur re not being able to share data structures. I've been using `cudarc` for FFI (Shaders in C++; CPU code in rust), but that would be a nice perk.