I am observing the phenomenon that repeated clicking on the compile-button produces different statistics for the same kernel.
Is there a random-based optimization going on ?
E.g. for the 6xxx series I get values between 430 and 520 MThreads/s for a simple kernel.
Who could explain this?
Identified the circumstances the Analyzer shows the described behaviour.
Whenever I am using the cl_device in parallel (even without running kernels on it, just created it) to using SKA, SKA's statistics start to fluctuate.
System: Windows7-64b, Radeon 4650.
Can you post the post. I am not sure i understood completely what you are trying to do.
Any how, did you try the profiler and does the timestamps fluctuate there too?
I managed to trigger this behaviour again and made 3 screenshots of it. It is always the same kernel, and every click on the "Compile" button produces a slightly different statistic.
I am running the SKA for a long time, and when the statistics start to fluctuate I have to close and restart it. Then the statistics are stable again.
I aplologize for the image size, but I cannot insert links ( I omitted a 3rd shot which again shows different value) . Please have look at the throughput column especially.
Thats really weird.
I tried to reproduce it on my system but failed. It is always stable. So i will forward this to Developer team.
Can you please post the SDK and Catalyst version you are using. Also mention the profiler version.
Also I think you have currently disabled the 5xxx devices and firestream devices. Are performance values stable for them?