Dynamic Parallelism and cuDNN

Accelerated Computing GPU-Accelerated Libraries
1

I am looking for the possibility to compute entirely on the device using the cuDNN library. Is this possible? And how would one go about doing this?