This issue only occurs with OpenCV 2.4.11-2.4.x branch compiled with Cuda 7.5 and does not with Cuda 6.5.
Each stream and buffers are alloced and page-locked on the correct device with gpu::setDevice.
It runs well with Cuda 6.5 on multiple GPUs. However it fails on the first OpenCV kernel call on the second GPU (GPU device 1) with the following exception with Cuda 7.5:
"d:opencv-2.4modulesgpuincludeopencv2gpudevicedetail/transform_detail.hpp:386: error: (-217) invalid resource handlen". Strangely, no issue happens on the first GPU (GPU device 0)
When using directly NPP on Cuda 7.5 with multiple GPUs, no issue happens. Unfortunately, I could not managed to find out what's happening...
I can provide sample code if needed.
Thanks
I know this is an old thread. But i am facing same issue with openCV 3.4 and Cuda 9.0. Did u find any solution?
Most helpful comment
I know this is an old thread. But i am facing same issue with openCV 3.4 and Cuda 9.0. Did u find any solution?