Best practices for OpenCV (GPU module) in multi-threaded applicaiton

edit

Dear all,

What are best practices/advice for running a multi-threaded application using OpenCV GPU module.

I want to analyse mutliple video streams on GPU in parallel. At the moment, I am creating a separate thread for analysing each file in parallel.

My application works for 2 threads. When I try to run more than 2 threads, I start getting exceptions

Can anybody share his/her experience of developing a multi-threaded application using OpenCV (GPU module) + CUDA? What are the best practices? What to avoid?

What are alternates to using pthreads for multi-threading on GPU (I am unable to use cv::gpu::stream)?

NOTE: It does not appear to be a resource issue. My application is only using 12% of CPU processing and 11% of the GPU meory

Regards,