Will we ever see libnvvm be updated to use llvm 12/13? Using LLVM 7 is pretty difficult considering most compilers nowadays use a more modern version of LLVM. New versions also bring more optimizations that may help kernels.
You may want to compile your CUDA code directly using the latest clang version. It uses a modern LVVM engine and has a PTX backend for CUDA. The only downside is that it does not always support the latest CUDA toolkits right away.
I am not compiling cuda, i am writing a backend for a compiler which already uses LLVM 12. Because of that, it would be nice if libnvvm could use llvm 12 bitcode, but it cannot from my knowledge. I can’t really use the LLVM PTX backend for a variety of reasons.