I'm getting very desparate for performance tuning information on Linux, and at this point I'm sufficiently desparate to write something myself. Are the performance counters on the card documented so that I can write my own interface to them? Would there be any hope of guidance from AMD so that an open-source tuning application can be written? Something that just gets the raw counters from the device is a good first step. Where does support have to be written, at kernel level or at userspace?
At this point, I cannot further optimize my application without a profiler... and rather than just complaining I'm fully ready to do it myself with some guidance.