Skip to content

Latest commit

 

History

History
139 lines (127 loc) · 5.89 KB

README.md

File metadata and controls

139 lines (127 loc) · 5.89 KB

Project Description

This is a forked repository, explaining how to build CNTK 2.5 on NVIDIA's Jetson TX1. The original CNTK repository from Microsoft is here.

How to Build CNTK 2.5 on TX1

It is assumed that JetPack 3.2 is installed on TX1. It is also recommended to install additional storage device on TX1 and setup swap space to speed up build process.

Firstly, we will start by installing prerequisites in the following list. Most of the guides for installing these prerequisites are from the Microsoft CNTK Document "Setup CNTK on Linux".

  • OpenBLAS 0.2.20
  • libzip 1.1.2
  • Boost 1.60.0
  • CUB 1.7.4
  • Protobuf 3.1.0
  • OpenCV 3.1.0

After these prerequisites are setup, we will create some symbolic links, and then finally build CNTK 2.5 itself.

OpenBLAS 0.2.20

$ git clone https://github.com/xianyi/OpenBLAS.git
$ cd OpenBLAS
$ git checkout v0.2.20
$ sudo apt install gfortran
$ make -j4
$ sudo make install PREFIX=/usr/local/OpenBLAS

libzip 1.1.2

$ wget http://nih.at/libzip/libzip-1.1.2.tar.gz
$ tar xzf ./libzip-1.1.2.tar.gz
$ cd libzip-1.1.2
$ ./configure
$ make -j4
$ sudo make install

Boost 1.60.0

$ wget -q -O - https://sourceforge.net/projects/boost/files/boost/1.60.0/boost_1_60_0.tar.gz/download | tar -xzf -
$ cd boost_1_60_0
$ ./bootstrap.sh --prefix=/usr/local/boost-1.60.0
$ sudo apt install libbz2-dev python-dev
$ sudo ./b2 -d0 -j4 install

CUB 1.7.4

$ wget https://github.com/NVlabs/cub/archive/1.7.4.zip
$ unzip ./1.7.4.zip
$ sudo cp -r cub-1.7.4 /usr/local

Protobuf 3.1.0

$ wget https://github.com/google/protobuf/archive/v3.1.0.tar.gz
$ tar xzf v3.1.0.tar.gz
$ cd protobuf-3.1.0
$ sudo apt install curl
$ ./autogen.sh
$ ./configure CFLAGS=-fPIC CXXFLAGS=-fPIC --disable-shared --prefix=/usr/local/protobuf-3.1.0
$ make -j4
$ sudo make install

OpenCV 3.1.0

$ sudo apt install cmake libgtk2.0-dev pkg-config libavcodec-dev libavformat-dev libswscale-dev
$ wget https://github.com/Itseez/opencv/archive/3.1.0.zip
$ unzip 3.1.0.zip
$ cd opencv-3.1.0
$ mkdir release
$ cd release
$ cmake -D WITH_CUDA=OFF -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local/opencv-3.1.0 -D BUILD_PERF_TESTS=OFF -D BUILD_TESTS=OFF ..
$ make all
$ sudo make install

Create Some Symbolic Links

# create symbolic links for CUDNN
$ sudo mkdir /usr/local/cudnn-7.0/cuda/include -p
$ sudo ln -s /usr/lib/aarch64-linux-gnu /usr/local/cudnn-7.0/lib
$ sudo ln -s /usr/include/aarch64-linux-gnu/cudnn_v7.h /usr/local/cudnn-7.0/cuda/include/cudnn.h

# create symbolic link for NVML
$ sudo ln -s /usr/local/cuda-9.0/targets/aarch64-linux/include/nvml.h /usr/local/include

# create symbolic link for Open MPI C++ wrapper compiler
$ sudo mkdir /usr/lib/openmpi/bin
$ sudo ln -s /usr/bin/mpic++ /usr/lib/openmpi/bin/mpic++

Build CNTK 2.5

$ git clone https://github.com/nglee/CNTK.git
$ cd CNTK
$ git checkout v2.5_tx1
$ mkdir build/release -p
$ cd build/release
$ ../../configure --asgd=no \
                  --cuda=yes \
                  --with-openblas=/usr/local/OpenBLAS \
                  --with-boost=/usr/local/boost-1.60.0 \
                  --with-cudnn=/usr/local/cudnn-7.0 \
                  --with-protobuf=/usr/local/protobuf-3.1.0 \
                  --with-mpi=/usr/lib/openmpi \
                  --with-gdk-include=/usr/local/include \
                  --with-gdk-nvml-lib=/usr/local/cuda-9.0/targets/aarch64-linux/lib/stubs
$ make -C ../../ \
       BUILD_TOP=$PWD \
       SSE_FLAGS='' \
       GENCODE_FLAGS='-gencode arch=compute_53,code=\"sm_53,compute_53\"' \
       all \
       -j 4

After a successful build, we'll find CNTK 2.5 libraries and binaries placed under lib and bin directoy. Those directories are created at [CNTK_SOURCE_BASE]/build/release.

How to Test the Build

To ensure that CNTK is working properly, we can follow the test steps described here. However, TX1 needs some extra steps regarding NVML(NVIDIA Management Library). As mentioned here, NVML does not support TX1, and it just provides a stub library. Since the binary called cntk that is used for the test needs NVML, we will setup the stub library for cntk.

$ export LD_LIBRARY_PATH=/usr/local/cuda-9.0/targets/aarch64-linux/lib/stubs:$LD_LIBRARY_PATH
$ sudo ln -s /usr/local/cuda-9.0/targets/aarch64-linux/lib/stubs/libnvidia-ml.so /usr/local/cuda-9.0/targets/aarch64-linux/lib/stubs/libnvidia-ml.so.1

After setting up the stub library, let's try CNTK with CPU.

$ export PATH=$HOME/Repos/cntk/build/release/bin:$PATH
$ cd [CNTK_SOURCE_BASE]/Tutorials/HelloWorld-LogisticRegression
$ cntk configFile=lr_bs.cntk makeMode=false

To try CNTK with GPU:

$ cntk configFile=lr_bs.cntk makeMode=false deviceId=auto

Contribute

Contributions are always welcome. You can just start a pull request.

Credits

Microsoft CNTK Document "Setup CNTK on Linux"
Why do we need gfortran when building OpenBLAS
NVML does not support Tegra platform

License

The original CNTK library from Microsoft is released under MIT license. So is this port.