Cross-compiling Arm NN for the Raspberry Pi and TensorFlow

Overview

This guide covers what we must do to cross-compile Arm NN using an x86_64 system to target a Raspberry Pi. Cross-compiling Arm NN allows us to work around the limited memory of the Raspberry Pi. Read the guide to find out how to build the Compute Library, Boost, Protobuf, TensorFlow, and Arm NN core libraries that you need for compilation. When we finish, we will be able to compile and run programs that use Arm NN on our Raspberry Pis.

Arm estimates that you will take 90-120 minutes to complete this guide.

Before you begin

This guide assumes:

• You have a Raspberry Pi board. Arm has tested these instructions on a Raspberry Pi 2 Model B V1.1 that runs Raspian 9, a Rasberry Pi 3 Model B+ that runs Raspian 8, and a Raspberry Pi 4 Model B that runs Raspian 9.
• You compile on a Linux virtual machine. Arm has tested these instructions on an Ubuntu 16.04 virtual machines that runs on MacOS and Windows 10 and an Ubuntu 18.04 machine.

You must install the following software tools on your virtual machine:

Git A version control system software developers use for source code management.

Scons An open-source software construction tool.

GNU C and C++ compilers for the armhf architecture The Raspberry Pi uses the armhf Arm architecture.

Curl A tool for transferring data to or from a Linux or Unix-like server.

GNU Autoconf A tool for producing configure scripts for building, installing and packaging software.

GNU libtool A generic library support script.

CMake An open-source and cross-platform family of tools for building, testing, and packaging software.

To install the tools you require, open a command window and enter the following commands:

sudo apt-get install git
sudo apt-get install scons
sudo apt-get install gcc-arm-linux-gnueabihf
sudo apt-get install g++-arm-linux-gnueabihf 
sudo apt-get install curl 
sudo apt-get install autoconf
sudo apt-get install libtool 
sudo apt-get install cmake

Downloading the repositories and bundles

Create a directory on your virtual machine to build your Arm NN distribution for your Raspberry Pi. 

To create a directory and build your distribution:

1. Enter the following commands:

   mkdir armnn-pi && cd armnn-pi
   export BASEDIR=`pwd`

2. Download the Compute Library, Boost, Protobuf, TensorFlow, and Arm NN git repositories and source bundles. To download the repositories and bundles, enter the following commands:

   git clone https://github.com/Arm-software/ComputeLibrary.git 
   git clone https://github.com/Arm-software/armnn
   wget https://dl.bintray.com/boostorg/release/1.64.0/source/boost_1_64_0.tar.bz2
   tar xf boost_1_64_0.tar.bz2
   git clone -b v3.5.0 https://github.com/google/protobuf.git
   git clone https://github.com/tensorflow/tensorflow.git
   cd tensorflow/
   git checkout a0043f9262dc1b0e7dc4bdf3a7f0ef0bebc4891e

Building the Compute Library

Build the Compute Library on your virtual machine. To build the library, enter the following command:

cd $BASEDIR/ComputeLibrary
scons extra_cxx_flags="-fPIC" Werror=0 debug=0 asserts=0 neon=1 opencl=0 os=linux arch=armv7a examples=1

Arm estimates that your virtual machine will take approximately 15-20 minutes to execute this command.

Building the Boost library for your Raspberry Pi

To build the Boost library for your Raspberry Pi:

1. Enter the following commands:

   cd $BASEDIR/boost_1_64_0/tools/build
   ./bootstrap.sh
   ./b2 install --prefix=$BASEDIR/boost.build
   export PATH=$BASEDIR/boost.build/bin:$PATH
   

2. Create a project-config.jam file by copying the user-config.jam file. To copy the user-config.jam file, enter the following command:

   cp $BASEDIR/boost_1_64_0/tools/build/example/user-config.jam $BASEDIR/boost_1_64_0/project-config.jam

3. Go to the $BASEDIR/boost_1_64_0/ directory and open the project-config.jam file in a text editor. In the GCC Configuration section, add the following line:

   using gcc : arm : arm-linux-gnueabihf-g++ ;

4. Save the project-config.jam file in the $BASEDIR/boost_1_64_0/ directory.

5. To complete the build, enter the following commands:

   cd $BASEDIR/boost_1_64_0
   b2 --build-dir=$BASEDIR/boost_1_64_0/build toolset=gcc-arm link=static cxxflags=-fPIC --with- 
   filesystem --with-test --with-log --with-program_options install --prefix=$BASEDIR/boost
   

Arm estimates that your virtual machine will take approximately 15 minutes to execute these commands.

Building the Google Protobuf library

You use two versions of the Google Protobuf library. One version of the library runs on your virtual machine and the other runs on your Raspberry Pi.

Building the Google Protobuf library for your virtual machine

To build the Google Protobuf library for your virtual machine:

1. Enter the following commands:

   cd $BASEDIR/protobuf
   git submodule update --init --recursive
   ./autogen.sh
   ./configure --prefix=$BASEDIR/protobuf-host
   make

   Arm estimates that your virtual machine will take approximately 15 minutes to execute these commands.

2. Enter the following commands:

   make install
   make clean

Building the Google Protobuf library for your Raspberry Pi

To build the Google Protobuf library for your Raspberry Pi:

1. Enter the following commands:

   ./configure --prefix=$BASEDIR/protobuf-arm --host=arm-linux CC=arm-linux-gnueabihf-gcc CXX=arm-linux-gnueabihf-g++ --with-protoc=$BASEDIR/protobuf-host/bin/protoc
   make

   Arm estimates that your virtual machine will take approximately 15 minutes to execute these commands.

2. Enter the following command:

   make install

Generating the TensorFlow Protobuf library

To generate the TensorFlow library, enter the following command:

cd $BASEDIR/tensorflow
../armnn/scripts/generate_tensorflow_protobuf.sh ../tensorflow-protobuf ../protobuf-host

Building Arm NN

To build Arm NN:

1. Enter the following commands:

   cd $BASEDIR/armnn
   mkdir build
   cd build
   

2. Place the library files you require in the build directory. To place the library files, enter:

   cmake .. -DCMAKE_LINKER=/usr/bin/arm-linux-gnueabihf-ld \
   -DCMAKE_C_COMPILER=/usr/bin/arm-linux-gnueabihf-gcc \
   -DCMAKE_CXX_COMPILER=/usr/bin/arm-linux-gnueabihf-g++ \
   -DCMAKE_C_COMPILER_FLAGS=-fPIC \
   -DCMAKE_CXX_FLAGS=-mfpu=neon \
   -DARMCOMPUTE_ROOT=$BASEDIR/ComputeLibrary \
   -DARMCOMPUTE_BUILD_DIR=$BASEDIR/ComputeLibrary/build \
   -DBOOST_ROOT=$BASEDIR/boost \
   -DTF_GENERATED_SOURCES=$BASEDIR/tensorflow-protobuf \
   -DBUILD_TF_PARSER=1 \
   -DPROTOBUF_ROOT=$BASEDIR/protobuf-arm \
   -DARMCOMPUTENEON=1 \
   -DARMNNREF=1
   make
   

   Arm estimates that your virtual machine will take approximately 12 minutes to execute these commands.

Extracting Arm NN on your Raspberry Pi and running a sample program

To create an archive of cross-compiled libraries, binaries, and directories:

1. Copy the following libraries, binaries, and directories from your virtual machine. To copy these libraries, binaries, and directories enter the following commands:

   cd $BASEDIR
   mkdir armnn-dist
   mkdir armnn-dist/armnn
   mkdir armnn-dist/armnn/lib
   
   ARMNN_MAJOR_VERSION=19
   ARMNN_MINOR_VERSION=11
   ARMNN_PATCH_VERSION=00
   
   cp $BASEDIR/armnn/build/libarmnn.so.$ARMNN_MAJOR_VERSION.$ARMNN_MINOR_VERSION $BASEDIR/armnn-dist/armnn/lib
   ln -s libarmnn.so.$ARMNN_MAJOR_VERSION.$ARMNN_MINOR_VERSION libarmnn.so.$MAJOR_VERSION
   ln -s libarmnn.so.$ARMNN_MAJOR_VERSION libarmnn.so
   cp $BASEDIR/armnn/build/libarmnnTfParser.so.$ARMNN_MAJOR_VERSION.$ARMNN_MINOR_VERSION $BASEDIR/armnn-dist/armnn/lib
   ln -s libarmnnTfParser.so.$ARMNN_MAJOR_VERSION.$ARMNN_MINOR_VERSION libarmnnTfParser.so.$MAJOR_VERSION
   ln -s libarmnnTfParser.so.$ARMNN_MAJOR_VERSION libarmnnTfParser.so
   cp $BASEDIR/protobuf-arm/lib/libprotobuf.so.15.0.0 $BASEDIR/armnn-dist/armnn/lib/libprotobuf.so
   cp $BASEDIR/protobuf-arm/lib/libprotobuf.so.15.0.0 $BASEDIR/armnn-dist/armnn/lib/libprotobuf.so.15
   cp -r $BASEDIR/armnn/include $BASEDIR/armnn-dist/armnn/include
   cp -r $BASEDIR/boost $BASEDIR/armnn-dist/boost

2. Copy the following dynamic backend related files from your virtual machine. To copy these files, enter the following commands:

   mkdir -p $BASEDIR/armnn-dist/src/backends/backendsCommon/test/
   cp -r $BASEDIR/armnn/build/src/backends/backendsCommon/test/testSharedObject $BASEDIR/armnn-dist/src/backends/backendsCommon/test/testSharedObject/
   
   cp -r $BASEDIR/armnn/build/src/backends/backendsCommon/test/testDynamicBackend/ $BASEDIR/armnn-dist/src/backends/backendsCommon/test/testDynamicBackend/
   cp -r $BASEDIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath1/ $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath1/
   
   mkdir -p $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath2
   cp $BASEDIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath2/
   
   ln -s Arm_CpuAcc_backend.so $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so.1
   ln -s Arm_CpuAcc_backend.so.1 $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so.1.2
   ln -s Arm_CpuAcc_backend.so.1.2 $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so.1.2.3
   cp $BASEDIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath2/Arm_GpuAcc_backend.so $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath2/
   ln -s nothing $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath2/Arm_no_backend.so
   
   mkdir -p $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath3
   
   cp -r $BASEDIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath5/ $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath5
   cp -r $BASEDIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath6/ $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath6
   
   mkdir -p $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath7
   
   cp -r $BASEDIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath9/ $BASEDIR/armnn-dist/src/backends/backendsCommon/test/backendsTestPath9
   
   mkdir -p $BASEDIR/armnn-dist/src/backends/dynamic/reference
   cp $BASEDIR/armnn/build/src/backends/dynamic/reference/Arm_CpuRef_backend.so $BASEDIR/armnn-dist/src/backends/dynamic/reference/

3. Copy the Unit Tests and a sample Arm NN program. To copy this test and program, enter the following commands:

   cp $BASEDIR/armnn/build/UnitTests $BASEDIR/armnn-dist
   cp $BASEDIR/armnn/samples/SimpleSample.cpp $BASEDIR/armnn-dist

4. Create the archive for your Raspberry Pi. To create the archive, enter the following command:

   tar -czf $BASEDIR/armnn-dist.tar.gz armnn-dist

Extract the libraries, binaries, and directories to your Raspberry Pi

Extract the libraries, binaries, and directories to your Raspberry Pi. To extract the libraries, binaries and directories enter the following commands:

cd /home/pi
tar -xzf /home/pi/armnn-dist.tar.gz

Compiling and running a sample Arm NN program on your Raspberry Pi

To compile and run a sample C++ program that uses Arm NN on your Raspberry Pi:

1. Enter the following commands:

   export LD_LIBRARY_PATH=/home/pi/armnn-dist/armnn/lib
   cd /home/pi/armnn-dist

   To compile the program, enter the following commands:

   g++ SimpleSample.cpp -I/home/pi/armnn-dist/armnn/include -I/home/pi/armnn-dist/boost/include -L/home/pi/armnn-dist/armnn/lib -larmnn -larmnnTfParser -lprotobuf -o SimpleSample

2. To run the program, enter the following command:

   ./SimpleSample

   The console returns the following:

   Please enter a number:

   Enter your number. For example:

   345

   The console returns the following:

   Your number was 345

Testing your build

To test that your build of Arm NN works correctly, run the Unit Tests. To run the Unit Tests, enter the following:

export LD_LIBRARY_PATH=/home/pi/armnn-dist/armnn/lib
cd /home/pi/armnn-dist
./UnitTests

If the tests are successful they output the following to the console:

*** No errors detected

If some of the tests are unsuccessful, go back through the steps and check that all the commands have been entered correctly.

Next steps

You are now ready to compile and run programs that use Arm NN on your Raspberry Pi.

Find out how to deploy Caffe and TensorFlow models.