support for nvidia Turing GPU

Hi,

I’m trying to use clang 7.0.1 to compile a simple program as shown on https://llvm.org/docs/CompileCudaWithLLVM.html for nvidia RTX2080 (compute capability 7.5). clang complains about not supporting sm_75. I then try the llvm repository head 7a5cc00 from the git mirror and the simple program compiles and runs fine. But then I found that for more complicated programs, I often got this runtime error:

cudaDeviceSynchronize() error( cudaErrorIllegalInstruction): an illegal instruction was encountered

cuda-gdb traces this back to __cuda_sm70_warpsync()

Can someone tell me that if the Turing GPU is currently supported or not?

Thanks,
Char

Those of us who are the main compiler developers don’t have any Turing GPUs in hand, so we haven’t tested the compiler on Turing. It’s “supported” today only inasmuch as “it compiles”.

I am surprised to see this error and I’m not sure how that’s happening. If you can provide steps to reproduce we might be able to assist you, although we might need to buy some hardware first…