In OpenCV 3.0 (http://opencv.org/opencv-3-0.html), there is the so-called T-API introducted, where functions are accelerated, transparently for the user, on the GPU with OpenCL.
I just wonder whether that is safe, when i have several CPU threads, each calling a OpenCV function which under the hood is accelerated with OpenCL ?
So, in short, are the 'accelerated' OpenCV functions STILL CPU-THREAD-SAFE ? Are the OpenCV functions tested in a test suite with respect to CPU-thread-safety ?
I have my doubts, because e.g. for NVIDIA NPP functions i conservatively assume that they are NOT CPU-thread safe, because i cannot know whether they internall use stuff like constant memory, texture references which per se are not cpu-trhead-safe.
Regarding NPP and thread-safety, see e.g. https://devtalk.nvidia.com/default/topic/534966/can-npp-be-safely-used-in-multi-threaded-code-/
Maybe someone can provide more information, or a better guess.
As in our application stability goes over performance, if there are any doubts we would have to disable the T-API when building OpenCV 3.0 for us.
↧