jrbyrnes wrote: > In production Moe kernels, controlling register placement yields >10% > performance gains.
Are you using the full feature coexec scheduler available here https://github.com/ROCm/llvm-project/tree/CoExecScheduler ? This scheduler should be able to reach your performance goals for gfx1250, and, if not, we need to understand and likely improve the scheduler behavior for this? Can you reach out to me offline ? I don't see your email address. https://github.com/llvm/llvm-project/pull/233773 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
