abi::__cxa_demangle

[nickjbrowning]

OK, could we just remove the depencency on the GCC-specific abi::__cxa_demangle? And use the mangled type here instead?

Originally posted by @Luthaf in #131 (comment)

[nickjbrowning]

We need the unmangled types to construct e.g my_kernel<float> or my_kernel<double, int>. See here:

sphericart/sphericart/src/cuda_base.cpp

Line 112 in 028b274

std::string kernel_name = getKernelName<scalar_t>("spherical_harmonics_kernel");

Do you have a better idea of doing this? The way I'm doing it is exactly the same way that NVRTC does it in their backend API...

[Luthaf]

Hmm, alright. Then it would be nice to move this code into a demangle_symbol function, with the appropriate #ifdef (like for dlopen & co). In the current state, I am a bit worried that we are making it a lot harder to get a windows build down the line.

[nickjbrowning]

windows has an equivalent function thats very similar so not too concerning. We can create a wrapper and ifdef the calls no problem.

[Luthaf]

I saw that the calls are already in an ifdef, so this should be doable later down the line. I would still create a wrapper to handle platform differences, but this should be good for now!