How to use llvm as the backend for cuda?


I received an email from Nvidia, saying that llvm now supports ptx code generation. I am confused how to use llvm as my backend, as it seems clang’s support for cuda is not complete yet. My question is: given a cuda program, how can I use llvm to generate ptx code?


Bu Wu, we have contributed back our ptx backend implementation (NVPTX). You should be able to find it in the trunk now.

At the GTC next week, we will talk about our road maps and how to build different compilers using the software we currently provide. Here is a link to the session .

More details can be provided after GTC.