Has anyone here tried to use OpenCV Cuda APIs on the kepler-powered shieldtablet?
GpuMat srcMatGPU(srcMat);
How exactly did these fail? What status was returned, what error reported? How exactly did the app “crash”? Did it terminate with a segmentation fault? Is it possible that the app is simply trying to allocate too much memory, and continues despite failed allocations? Based on your description that seems like a plausible hypothesis. I am not sure whether there are potential kernel timeout issues. On regular Linux with X a watchdog timer that kills CUDA kernels to keep the GUI alive certainly exists.
As far as I understand OpenCV is not an NVIDIA product, but an open source project. This means that it is not supported by NVIDIA in the sense that NVIDIA is responsible for making it work on particular platforms and that you could file bugs at the NVIDIA bug reporting page. You might want to inquire about the issue on a forum dedicated to OpenCV as well, I would assume there is a sizable community of OpenCV users since the software is quite popular from what I can tell.