Using the stereo disparity example provided for VPI 2.3.9, I would like to run this estimation for multiple images at the same time so as to saturate the GPUs and maximize throughput. When I use the example individually everything works as expected, but when the use the same code wrapped with multiprocessing I get VPI initialization errors. If I use GNU parallel to launch the example program, I can run it maximum two jobs, if I use greater than 2 jobs I get similar VPI initialization error.
While GNU Parallel the error is as follows
VPI_ERROR_INVALID_OPERATION: PVA is not available and may be oversubscribed in the system: PvaError_DeviceUnavailable
I am using CUDA as the backend and not PVA. Also, the workaround you shared is for code using C++. I am using python. Could you suggest what should I change in this [code](VPI - Vision Programming Interface: Stereo Disparity [its the sample code for stereo that nvidia provides), that will help me enforce CUDA as backend flag while creating images?