I am wondering whether cuDNN has a device side api for dynamic parallelism (I want to call cuDNN kernels inside other kernels). I have found that cuBLAS has such an api: cuBlas but I could not find any information on whether cuDNN provides a similar api. I checked the cuDNN paper and it says that "The library exposes a host-callable C language API" but the paper is old and I want to know whether they added a device side api. I couldn't find any information in the cuDNN documentation. Is anyone aware of such an api?
There is no device API for cuDNN at this time. Currently there is also no device API for CUBLAS in recent CUDA releases. That functionality was deprecated and removed in the CUDA 10 timeframe.