Boris Fomitchev

A month ago I have taken a stab at refactoring Convolution classes : a very outdated version is here https://github.com/borisfom/cudnn.torch/tree/R5_exp (FullConvolution changed fundamentally since then). However, it may give you...

I found some neat examples of taking advantage of Lua binding of C functions by name in the new RNN class contributed by NVidia folks here: https://github.com/borisfom/cudnn.torch/blob/R5/RNN.lua (createDescriptors etc).

@marcel1991 : Can you post a snippet of your code where this is happening ? What is your card, exactly? What does cutorch.hasFastHalfInstructions() return ? Also: try using CUDNN6. A...

Option may make sense, however: What is your exact use-case? Do you have CUDNN installed in location you do not have in your LD_LIBRARY_PATH at the time of install ?

Do you have existing CUDNN 5.1 in your LD_LIBRARY_PATH? It works for me. Also, does it do it more than once for you ?

@soumith: it's trivial to add an option, and I am not even against it. I am just wondering what the use-case is. How do you know before installing Torch for...

@dasguptar Older cards won't be an issue. I have put 5.1 requirement in as there were certain important bug fixes and there are no cases where 5.0 would do better....

Right, I have changed required version to '5' which would fit both cases - with and without headers.

Yes there are some cases where cudnnFindConvolutionForwardAlgorithm may fail. You may try set cudnn.useFindEx. Also make sure you upgrade to CUDNN 5.1.

Your installation is just old - please update the package from Nvidia repo: bfomitchev@borisf-desktop:~/git/caffe$ sudo dpkg -L libcudnn4 /. /usr /usr/lib /usr/lib/x86_64-linux-gnu /usr/lib/x86_64-linux-gnu/libcudnn.so.4.0.3 /usr/share /usr/share/doc /usr/share/doc/libcudnn4 /usr/share/doc/libcudnn4/copyright /usr/share/doc/libcudnn4/changelog.gz /usr/share/lintian /usr/share/lintian/overrides...