JCudnn 7


#1

Thank you for making JCudnn available. I’ve finally started using it and it is working smoothly. I’d like to know if it would be difficult for you to update it for JCudnn 7.


#2

Hello

It’s not conceptually difficult. There already is a pending pull request that is related to JCudnn 7, at https://github.com/jcuda/jcudnn/pull/1

The point is (that I also mentioned there) that I’d actually like to do this togehter with the update for CUDA 9. But I also have to admit that I did not yet have a closer look at what’s new in either CUDA 9 or cuDNN 7. (CUDA 9 is currently only available as a developer preview, anyhow).

Maybe it makes sense to update JCudnn independently. It would likely not be an own maven release, but at least a state that supports the latest cuDNN features, regardless of the underlying CUDA version.

I’ll try to check out cuDNN 7 as soon as possible (maybe I can allocate some time for this today).


#3

Well, it’s still “today” in most parts of the world :wink:

The main update for cuDNN 7.0.2 is in the commit at https://github.com/jcuda/jcudnn/commit/2ba861a92d69d99e544a976a2696b62ae1a02b69

As I already mentioned in the (now closed) PR:

Until now, I only tested it with the original MNIST sample, and a small sample that uses RNN.
Some of the new functionality that was introduced in cuDNN 6 and 7 is thus not yet tested thoroughly.
If you encounter any problems, or see any wrong method signatures or such, please let me know.

The functions of cuDNN can be intimidating (>25 (!) parameters!), so it’s really hard (or nearly impossible) to create a sample from scratch, but I’ll have a look at the new samples that NVIDIA included in their Linux packages - maybe some of them can be ported to JCuda.

(Right now, there is only a port of the RNN sample that I’ll hopefully be able to upload soon. Further contributions are, of course, always welcome :wink: )


#4

Thanks! I will try it out today on my CNN code and let you know. I’ll attempt to implement RNN within the next few weeks.