updating nvidia-cudnn?

[M. Zhou]

Hi Andreas,

I can deal with the update later. Generally it is as easy as bumping
a version number as well as a checksum in the code.

On Thu, 2022-09-29 at 15:46 +0200, Andreas Beckmann wrote:
> Hi lumin,
> 
> nvidia-cudnn is currently at version 8.2.4.15~cuda11.4 while 
> nvidia-cuda-toolkit 11.5.2 recently made into sid.
> Does nvidia-cudnn need to be updated for the newer cuda toolkit? What 
> signficance does the enbedded version number have?

I will handle the update later.

> (Since there is no build dependency on nvidia-cuda-toolkit, cudnn falls 
> through my regular testing raster.)
> 
> Is there anything we could do to have some minimal superficial 
> autopkgtest for nvidia-cudnn (runnable on a host without nvidia gpu or 
> driver available)? E.g. compile something ...

Compilation will not detect problems. The real problems (e.g. library
incompatibility will be raised at runtime as cuda device error or something
alike. So I don't think there is any way to add a sensible test for a
builder without nvidia gpu.

> Since nvidia-cudnn is an downloader package and does not contain any 
> non-free bits (at least I assume that it only depends on them), 
> shouldn't that be rather in contrib instead of non-free?

In this sense it does not make much difference. In my fuzzy memory
the flashplayer plugin package (maybe ten years ago) was already
a downloader in non-free.

> Is cudnn available for other architectures besides amd64?
> 
> 

Yes, there are ppc64el and aarch64 versions
https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-850/support-matrix/index.html
I just need to build the package with qemu to test the corresponding architectures.

> Andreas