1 of 1 people found this helpful
When collecting performance counters, the GPU profiler may have to replay the kernel more than once. There is a hardware limit on the number of counters that can be queried for a given kernel dispatch. In order to collect all counters, the GPU profiler will replay the kernel the required number of times. The profiler tracks the buffers used by kernels and will save and restore their state prior to re-dispatchingthe kernel. It does this to ensure that the replayed kernel behaves identically each time it is re-dispatched. If you are running into a case where this does not appear to be working correctly, the profiler team would be very interested in seeing your test case. Can you share the application where this does not appear to be working correctly so that we can investigate?
thx for the reply. I will try to compact my code sample to share the application where this does not appear to be working correctly...
Can you tell me what I have to do, to stop this behaviour above?