mgehre-amd wrote:

> In that case, I think we need to fix the wheels, since it is provided by us, 
> and it breaks the convention.

>From reading the comments, I think we are actually aligned. I hear you saying 
>that the compiler should not rely on ROCM_PATH to find its device libs, but 
>instead find them in the resource directory.
This is exactly the behavior we want for the python wheels. Unfortunately, some 
other tools in ROCm/external build systems depend on ROCM_PATH to find other 
parts of ROCm, so even with the python wheels, we sometimes need to set 
ROCM_PATH.

But setting ROCM_PATH currently keeps the compiler from looking in the resource 
directory, and then it fails to find the device libs. This is the issue this 
PRs tries to address.

This PR is a small step to make the compiler less reliant on ROCM_PATH - it can 
still find the device libs in the resource directory even when the ROCM_PATH 
environment variable is set and doesn't point to a place where the device libs 
are.

https://github.com/llvm/llvm-project/pull/177346
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to