I have successfully tested a small C++ program which uses the NVIDIA cuDNN API to run code on a GPU. The cuDNN API level is 9.1.1. The program compiles, links and runs OK.
What puzzles me is that according to the official NVIDIA API documentation (https://docs.nvidia.com/deeplearning/cudnn/latest/api/overview.html) a large number of the cuDNN functions are marked as deprecated. E.g. one such function is cudnnActivationForward()
, which seems to be at the core of what the cuDNN library is for.
Nowhere can I seem to find notes about *why *these functions are going away, or what *replaces *them. It is of course OK that an API is being modernized, but normally then some information on what to use instead, or a migration path is provided by the API developer.
Does anyone have any pointers or insight why these functions are being deprecated, and what else to use instead?