Setup-Torch-cuDNN-on-Jetson-TK1.md

Setup cuDNN library for Torch on Jetson Tegra K1

This tutorial assumes that NVIDIA toolkit (less than 6.5) and Torch are already installed on the board.

Set up basic configurations on TK1 device

Now it is time to update CUDA compiler (>= 6.5). Download the package from NVIDIA website on the board and install.

wget http://developer.download.nvidia.com/embedded/L4T/r21_Release_v3.0/cuda-repo-l4t-r21.3-6-5-prod_6.5-42_armhf.deb
sudo dpkg -i cuda-repo-l4t-r21.3-6-5-prod_6.5-42_armhf.deb
sudo apt-get update
sudo apt-get install cuda-toolkit-6-5
sudo usermod -a -G video ubuntu
echo ' ' >> ~/.bashrc
echo '# Cuda dependencies' >> ~/.bashrc
echo 'export PATH=/usr/local/cuda-6.5/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda-6.5/lib:$LD_LIBRARY_PATH' >> ~/.bashrc

Then, set up basic configurations at your preference on TK1 device. First, allow community-maintained open-source softwares.

sudo add-apt-repository universe

Installing Torch and packages

Tutorial here. You can easily install Torch with ezinstall script.

To install the CUDNN libraries and package, check the tutorial here.

Test

Run test scripts in https://github.com/soumith/cudnn.torch or

require("cudnn")

local net = nn.Sequential()
net:add(cudnn.SpatialConvolution(3, 48, 5, 5, 1, 1, 0, 0))
net:add(cudnn.ReLU())
net:add(cudnn.SpatialMaxPooling(2, 2, 2, 2))
net = net:cuda()

local input = torch.Tensor(128, 3, 231, 231):cuda()
net:forward(input)
cutorch.synchronize()