Could it be possible to use cuDNN inside CUDA kernels?

I have several simulations that run inside CUDA kernels parallelly, and I want to perform some neural network computation, so the first thing I can think of is cuDNN. But after studied for a while, I realized that cuDNN has no CUDA device interface, which means it can only be called in the CPU code.

Is there a way to use cuDNN without leaving the CUDA kernel? Thank you!


I think it’s not possible, since cuDNN APIs are created on top of CUDA to provide abstraction.