When 27x or 280 drivers are used instead of 26x ones there is huge increase in CPU consumption in my OpenCL builds.
Almost whole CPU core used. With 26x drivers CPU consumption quite low.
It was shown, that CUDA test can mimic such OpenCL app behavior on 280 drivers when synching mode will be changed from cudaDeviceBlockingSync to cudaDeviceScheduleYield. CUDA build begin to use whole CPU core too under this condition.
The question is: how to change used synching mode under OpenCL ? Is it possible? If not, why synching mode was changed in 26x to 27x and 280 drivers transition? There is no noticeable speedup in elapsed time for my application, just huge increase in CPU time and it looks like resource waste. I need CPU for useful computations too, not just as GPU servant…