What is the status of CUDA on Windows?

I’m trying to determine the status of CUDA on Windows. This page https://llvm.org/docs/CompileCudaWithLLVM.html mentions PR38811, but that bug was fixed almost a year ago.

That page also mentions the current version of clang is 7.0.0, but we’re about to release 10.0.0 so thats almost 3 versions behind.

I’d appreciate any information anyone has.

Thanks,

I’m trying to determine the status of CUDA on Windows. This page https://llvm.org/docs/CompileCudaWithLLVM.html mentions PR38811, but that bug was fixed almost a year ago.

That page also mentions the current version of clang is 7.0.0, but we’re about to release 10.0.0 so thats almost 3 versions behind.

The docs are terribly out of date. I’ll update them soon.

In general, there’s no owner for CUDA compilation with clang on windows, no testing infstastructure, etc, so things tend to bitrot.

At the moment clang-10 or newer should be able to compile CUDA on windows. It’s not something that’s been tested widely, but I did manage to compile a trivial kernel recently.
The likely source of the problems will be with the code which replies on using standard C++ library in the device-side code. Clang on linux has a set of header wrappers which works for a subset of libc++ and libstdc++. MSVC’s standard library is quite different and I’d expect code that uses something like std::complex<…> to fail in interesting ways.

–Artem