[llvm] [AMDGPU][Attributor] Make `AAAMDFlatWorkGroupSize` honor existing attribute (PR #114357)
Shilei Tian via llvm-commits
llvm-commits at lists.llvm.org
Wed Jan 15 08:35:15 PST 2025
shiltian wrote:
No, vanilla LLVM doesn't have device-libs. That is only available in AMD's fork. For this particular case, you will have to build device-libs because the device-libs in your system have half-baked values that can cause performance regression. That's why I suggested to simply just remove the attributes from those `/opt/rocm/amdgcn/bitcode/*.bc` files.
You can still build vanilla LLVM, and use that to build device-libs. The only thing you want out of AMD's fork is the device-libs, which sits in `amd/device-libs`.
To make it simple, can you first verify if `amdgpu-flat-work-group-size` or `amdgpu-waves-per-eu` exist in any `/opt/rocm/amdgcn/bitcode/*.bc` files? If there is no, then you don't need to build device-libs. That would be a separate issue.
https://github.com/llvm/llvm-project/pull/114357
More information about the llvm-commits
mailing list