lfmeadow wrote: > > That's right. There are cases (such as RCCL) where the device code doesn't > > have a corresponding host shadow HIP module. HSA is used to introspect the > > device side modules and drain the counters. > > Note that for linux, we don't even need the HIP path with this PR. Windows > > doesn't have HSA so we still need the HIP path there. > > This seems like a huge mess of likely 100% AI generated code. Could we at > least split this up? I see ifdefs on Win32, which wouldn't apply to the HSA > case (as far as I know, I do think there was some work for HSA on Windows at > some point).
I'll rewrite it by hand. https://github.com/llvm/llvm-project/pull/203056 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
