Thanks for the re-test. It's odd that I couldn't reproduce it any more. It may be that I was just lucky. However, it's worth noting that Daniel has a GCN 1.0 card and David has a GCN 1.1 card, i.e. running on the radeon kernel module and DDX. It's possible that an amdgpu-only change has randomly fixed this for my Tonga-based test setup.