Bit late with this one, but I didn't notice initially.
It seems there is some timing/sync issue on R9285 with dual instance enabled
after

c59628d11b134fc016388a170880f7646e100d6f
st/va: enable dual instances encode by sync surface

Testing with large rawvideo/higher bitrates lucked me out of noticing initially
as visually these tend to be OK, though making say 20 and md5summing them will
show inconsistencies. I can change the number of "bads" in some tests by
flipping my cpus between cpufreq on_demand and perf.

At lower sizes/bitrates/transcoding it's possible to get corruption, either in
the form of some runs giving abnormally low bitrate with vbr, or sometimes with
cbr there is a chance of an out of order frame around an IDR frame.

Testing this with gstreamer.