TK1入门教程软件篇-安装cuDNN

TK1入门教程软件篇-安装cuDNN

说明：

• 介绍如何TK1上安装cuDNNR1或cuDNNR2版本，以及在Caffe上使用。

cuDNN ：

• cuDNN是NVIDIA的GPU加速库，用于深层神经网络和机器学习
• 它被集成到更高级别的机器学习框架，如加州大学伯克利分校的Caffe软件，现在可用于Jetson TK1
• cuDNN版本1是当前的稳定版本，而cuDNN v2是发布版本3
• Caffe使用cuDNN发行版1.所有库和文档均可从NVIDIA网站获取，需要免费开发者注册才能下载

安装cuDNN1：

• 建立目录

mkdir ~/cuDNNR1

• 下载软件包：

  • 软件包名称：cudnn-6.5-linux-armv7-R1.tgz
  • 下载地址：https://developer.nvidia.com/cuDNN

• 下载安装脚本：

cd ~/cuDNNR1
git clone https://gist.github.com/jetsonhacks/fa9f4ff89006607359ea
cp fa9f4ff89006607359ea/installcuDNNR1.sh ../installcuDNNR1.sh
vim installcuDNNR1.sh

• 脚本内容：

#!/bin/sh
# Install cuDNN R1 on NVIDIA Jetson TK1
# Register as a NVIDIA developer and download the cuDNN package
# Package is named cudnn-6.5-linux-armv7-R1.tgz
# This script and the cuDNN package should be in the same directory, usually ~/Downloads
# This script places the library and include files for cudnn in the CUDA 6.5 directories
# Make sure this is executing as root
if [ $(id -u) != 0 ]; then
  echo "This script requires root permissions"
  echo "$ sudo "$0""
  exit
fi
# unzip the archive
tar -zxvf cudnn-6.5-linux-armv7-R1.tgz
cd cudnn-6.5-linux-armv7-R1
# copy the include file
cp cudnn.h /usr/local/cuda-6.5/include
cp libcudnn* /usr/local/cuda-6.5/lib

• 也可以自行复制张贴对应内容
• 分配权限，执行安装

chmod +x installcuDNNR1.sh
./installcuDNNR1.sh

• 完成安装

安装cuDNNR2：

• 建立目录

mkdir ~/cuDNNR2

• 下载软件包：

  • 软件包名称：cudnn-6.5-linux-ARMv7-R2-rc1.tgz
  • 下载地址：https://developer.nvidia.com/cuDNN

• 下载安装脚本：

cd ~/cuDNNR2
git clone https://gist.github.com/jetsonhacks/5257f2352d444c2f3240
cp  5257f2352d444c2f3240/installcuDNNR2.sh ../installcuDNNR2.sh
vim installcuDNNR2.sh

• 脚本内容：

#!/bin/sh
# Install cuDNN R2 on NVIDIA Jetson TK1
# Register as a NVIDIA developer and download the cuDNN package
# Package is named cudnn-6.5-linux-ARMv7-R2-rc1.tgz
# This script and the cuDNN package should be in the same directory, usually ~/Downloads
# This script places the library and include files for cudnn in the CUDA 6.5 directories
# Make sure this is executing as root
if [ $(id -u) != 0 ]; then
  echo "This script requires root permissions"
  echo "$ sudo "$0""
  exit
fi
# unzip the archive
tar -zxvf cudnn-6.5-linux-ARMv7-R2-rc1.tgz
cd cudnn-6.5-linux-ARMv7-R2-rc1
# copy the include file
cp cudnn.h /usr/local/cuda-6.5/include
cp libcudnn* /usr/local/cuda-6.5/lib

• 也可以自行复制张贴对应内容
• 分配权限，执行安装

chmod +x installcuDNNR2.sh
./installcuDNNR2.sh

• 完成安装

Caffe使用cuDNN

• 修改Makefile.config

# Makefile.config (only relevant cuDNN configuration items shown)
USE_CUDNN := 1
INCLUDE_DIRS := /usr/local/cuda-6.5/include
LIBRARY_DIRS := /usr/local/cuda-6.5/lib

• 重新编译Caffe才生效

参考：

• https://devblogs.nvidia.com/parallelforall/embedded-machine-learning-cudnn-deep-neural-network-library-jetson-tk1/
• http://www.jetsonhacks.com/2015/01/20/nvidia-jetson-tk1-cudnn-install-caffe-example/
• https://gist.github.com/jetsonhacks/fa9f4ff89006607359ea
• https://gist.github.com/jetsonhacks/5257f2352d444c2f3240