[Mocha] Where does libcudnn go on linux boxes?

question

#1

I am trying to use the Mocha package, and I’ve downloaded the Deep Neural Network cuda library from nVidia. I put a copy of it in /libs and another in ~/.julia/v0.5/Mocha/Deps. No luck running Pkg.test(“Mocha”). If I turn off the GPU backend it works fine, so I know all other parts of my setup are correct. The Mocha documentation skips the part of what to do once you download the shared library.

I know this is off topic but I couldn’t find an appropriate discourse like site for Mocha. If there is one, please redirect me there. Thanks!


#2

Mocha seems to use regular Base.find_library, so depending on your library add something to eg. LD_LIBRARY_PATH (Linux) or edit the relevant find_library calls (ie. the ones discovering CUDA and cuDNN in respectively cuda.jl and cudnn.jl) to include your CUDA toolkit library directory in the (currently empty) array of directories as second argument.

Their build system should probably discover your library, or provide a way to point towards it (eg. CUDA_HOME/CUDA_ROOT env var).


#3

OK that is sound advice! I will give it a try later today. The machine I am using it on here at work has two Titan cards in it, and the option pricing black scholes demo calculated 25 billion prices a second. So I am keenly interested to see if I can get similar computing power with my large data sets.