Cupy: How to use cupy's cudnn Interface?

Created on 25 Jun 2017  路  3Comments  路  Source: cupy/cupy

The cudnn.py file doesn't have comments hence it is very difficult to understand and use it. So can you please provide me a simple hello world example on how to use cupy's cudnn interface.

document feature

Most helpful comment

Current cudnn.py provides undocumented API. These are only for Chainer, because the cuDNN APIs are very different in each version.
We will refactor cudnn.py and open these API.

Which API do you want use?

All 3 comments

Current cudnn.py provides undocumented API. These are only for Chainer, because the cuDNN APIs are very different in each version.
We will refactor cudnn.py and open these API.

Which API do you want use?

Thanks for the reply.

Actually, I'm writing a simple tool (with static computational graphs) similar to Tensorflow, just for learning deep learning system development process. I will be using cupy as my CUDA library to communicate with Nvidia GPUs.

Appreciate if you could explain to me how to use cuDNN convolutional operations exposed by cudnn.py.

Thanks,
Upul

For now, the best approach might be referring to Chainer's implementation.
For instance, chainer/functions/connection/convolution_nd.py might be a good example.
To understand the meaning and specification of function arugments, you can check NVIDIA's official cuDNN reference.

(https://github.com/chainer/chainer/blob/master/chainer/functions/connection/convolution_nd.py)

Was this page helpful?
0 / 5 - 0 ratings