lucas-rami wrote: > IIUC that is because the flat workgroup size. Waves per EU must yield to the > value computed from flat workgroup size, and if it is absent, we must assume > it can be 1024.
AFAIU the intent of the existing implementation is that the default minimum waves/EU is set so that all the waves of a workgroup of maximum size can fit concurrently on a single CU. I am not proposing we change that, what I would like the "amdgpu-waves-per-eu" attribute to do is be able to lower than minimum at the user's request in cases where higher occupancies are not thought to be beneficial. https://github.com/llvm/llvm-project/pull/138284 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
