How to Cross-Compile ArmNN on x86_64 for arm64

Introduction

These are the step by step instructions on Cross-Compiling ArmNN under an x86_64 system to target an Arm64 system. This build flow has been tested with Ubuntu 16.04. The instructions show how to build the ArmNN core library and the Boost, Protobuf, Caffe and Compute Libraries necessary for compilation.

Cross-compiling ToolChain

  • Install the standard cross-compilation libraries for arm64:
    sudo apt install crossbuild-essential-arm64

Build and install Google's Protobuf library

  • Get protobuf-all-3.5.1.tar.gz from here: https://github.com/protocolbuffers/protobuf/releases/tag/v3.5.1
  • Extract:
    tar -zxvf protobuf-all-3.5.1.tar.gz
cd protobuf-3.5.1
  • Build a native (x86_64) version of the protobuf libraries and compiler (protoc): (Requires cUrl, autoconf, llibtool, and other build dependencies if not previously installed: sudo apt install curl autoconf libtool build-essential g++)
    mkdir x86_64_build
cd x86_64_build
../configure --prefix=$HOME/armnn-devenv/google/x86_64_pb_install
make install -j16
cd ..
  • Build the arm64 version of the protobuf libraries:
    mkdir arm64_build
cd arm64_build
CC=aarch64-linux-gnu-gcc \
CXX=aarch64-linux-gnu-g++ \
../configure --host=aarch64-linux \
--prefix=$HOME/armnn-devenv/google/arm64_pb_install \
--with-protoc=$HOME/armnn-devenv/google/x86_64_pb_install/bin/protoc
make install -j16
cd ..

Build Caffe for x86_64

  • Ubuntu 16.04 installation. These steps are taken from the full Caffe installation documentation at: http://caffe.berkeleyvision.org/install_apt.html
  • Install dependencies:
    sudo apt-get install libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev
sudo apt-get install --no-install-recommends libboost-all-dev
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
sudo apt-get install libopenblas-dev
sudo apt-get install libatlas-base-dev
  • Download Caffe-Master from: https://github.com/BVLC/caffe
    git clone https://github.com/BVLC/caffe.git
cd caffe
cp Makefile.config.example Makefile.config
  • Adjust Makefile.config as necessary for your environment, for example:
    #CPU only version:
CPU_ONLY := 1

#Add hdf5 and protobuf include and library directories (Replace $HOME with explicit /home/username dir):
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial/ $HOME/armnn-devenv/google/x86_64_pb_install/include/
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnu/hdf5/serial/ $HOME/armnn-devenv/google/x86_64_pb_install/lib/
  • Setup environment:
    export PATH=$HOME/armnn-devenv/google/x86_64_pb_install/bin/:$PATH
export LD_LIBRARY_PATH=$HOME/armnn-devenv/google/x86_64_pb_install/lib/:$LD_LIBRARY_PATH
  • Compilation with Make:
    make all
make test
make runtest
    These should all run without errors
  • caffe.pb.h and caffe.pb.cc will be needed when building ArmNN's Caffe Parser

Build Boost library for arm64

  • Build Boost library for arm64 Download Boost version 1.64 from http://www.boost.org/doc/libs/1_64_0/more/getting_started/unix-variants.html Using any version of Boost greater than 1.64 will fail to build ArmNN, due to different dependency issues.
    tar -zxvf boost_1_64_0.tar.gz
cd boost_1_64_0
echo "using gcc : arm : aarch64-linux-gnu-g++ ;" > user_config.jam
./bootstrap.sh --prefix=$HOME/armnn-devenv/boost_arm64_install
./b2 install toolset=gcc-arm link=static cxxflags=-fPIC --with-filesystem --with-test --with-log --with-program_options -j32 --user-config=user_config.jam

Build Compute Library

  • Building the Arm Compute Library:
    git clone https://github.com/ARM-software/ComputeLibrary.git
cd ComputeLibrary/
scons arch=arm64-v8a neon=1 opencl=1 embed_kernels=1 extra_cxx_flags="-fPIC" -j8 internal_only=0

Build ArmNN

  • Compile ArmNN for arm64:

    git clone https://github.com/ARM-software/armnn.git
cd armnn
mkdir build
cd build

  • Use CMake to configure your build environment, update the following script and run it from the armnn/build directory to set up the armNN build:

    #!/bin/bash
CXX=aarch64-linux-gnu-g++ \
CC=aarch64-linux-gnu-gcc \
cmake .. \
-DARMCOMPUTE_ROOT=$HOME/armnn-devenv/ComputeLibrary \
-DARMCOMPUTE_BUILD_DIR=$HOME/armnn-devenv/ComputeLibrary/build/ \
-DBOOST_ROOT=$HOME/armnn-devenv/boost_arm64_install/ \
-DARMCOMPUTENEON=1 -DARMCOMPUTECL=1 -DARMNNREF=1 \
-DCAFFE_GENERATED_SOURCES=$HOME/armnn-devenv/caffe/build/src \
-DBUILD_CAFFE_PARSER=1 \
-DPROTOBUF_ROOT=$HOME/armnn-devenv/google/x86_64_pb_install/ \
-DPROTOBUF_LIBRARY_DEBUG=$HOME/armnn-devenv/google/arm64_pb_install/lib/libprotobuf.so.15.0.1 \
-DPROTOBUF_LIBRARY_RELEASE=$HOME/armnn-devenv/google/arm64_pb_install/lib/libprotobuf.so.15.0.1

  • If you want to include standalone sample dynamic backend tests, add the argument to enable the tests and the dynamic backend path to the CMake command:

    -DSAMPLE_DYNAMIC_BACKEND=1 \
-DDYNAMIC_BACKEND_PATHS=$SAMPLE_DYNAMIC_BACKEND_PATH

  • Run the build

    make -j32

Build Standalone Sample Dynamic Backend

  • The sample dynamic backend is located in armnn/src/dynamic/sample

    mkdir build
cd build

  • Use CMake to configure your build environment, update the following script and run it from the armnn/src/dynamic/sample/build directory to set up the armNN build:

    #!/bin/bash
CXX=aarch64-linux-gnu-g++ \
CC=aarch64-linux-gnu-gcc \
cmake .. \
-DCMAKE_CXX_FLAGS=--std=c++14 \
-DBOOST_ROOT=$HOME/armnn-devenv/boost_arm64_install/ \
-DBoost_SYSTEM_LIBRARY=$HOME/armnn-devenv/boost_arm64_install/lib/libboost_system.a \
-DBoost_FILESYSTEM_LIBRARY=$HOME/armnn-devenv/boost_arm64_install/lib/libboost_filesystem.a \
-DARMNN_PATH=$HOME/armnn-devenv/armnn/build/libarmnn.so

  • Run the build

    make

Run Unit Tests

  • Copy the build folder to an arm64 linux machine
  • Copy the libprotobuf.so.15.0.1 library file to the build folder
  • If you enable the standalone sample dynamic tests, also copy libArm_SampleDynamic_backend.so library file to the folder specified as $SAMPLE_DYNAMIC_BACKEND_PATH when you build ArmNN
  • cd to the build folder on your arm64 machine and set your LD_LIBRARY_PATH to its current location:
    cd build/
export LD_LIBRARY_PATH=`pwd`
  • Create a symbolic link to libprotobuf.so.15.0.1:
    ln -s libprotobuf.so.15.0.1 ./libprotobuf.so.15
  • Run the UnitTests:
    ./UnitTests
Running 567 test cases...

*** No errors detected

Troubleshooting and Errors:

Error adding symbols: File in wrong format

  • When building armNN:
    /usr/local/lib/libboost_log.a: error adding symbols: File in wrong format
collect2: error: ld returned 1 exit status
CMakeFiles/armnn.dir/build.make:4028: recipe for target 'libarmnn.so' failed
make[2]: *** [libarmnn.so] Error 1
CMakeFiles/Makefile2:105: recipe for target 'CMakeFiles/armnn.dir/all' failed
make[1]: *** [CMakeFiles/armnn.dir/all] Error 2
Makefile:127: recipe for target 'all' failed
make: *** [all] Error 2
  • Boost libraries are not compiled for the correct architecture, try recompiling for arm64

Virtual memory exhausted

  • When compiling the boost libraries:
    virtual memory exhausted: Cannot allocate memory
  • Not enough memory available to compile. Increase the amount of RAM or swap space available.

Unrecognized command line option ‘-m64’

  • When compiling the boost libraries:
    aarch64-linux-gnu-g++: error: unrecognized command line option ‘-m64’
  • Clean the boost library directory before trying to build with a different architecture:
    sudo ./b2 clean
  • It should show the following for arm64:
    - 32-bit                   : no
- 64-bit                   : yes
- arm                      : yes

Missing libz.so.1

  • When compiling armNN:

    /usr/lib/gcc-cross/aarch64-linux-gnu/5/../../../../aarch64-linux-gnu/bin/ld: warning: libz.so.1, needed by /home/<username>/armNN/usr/lib64/libprotobuf.so.15.0.0, not found (try using -rpath or -rpath-link)

  • Missing arm64 libraries for libz.so.1, these can be added by adding a second architecture to dpkg and explicitly installing them:

    sudo dpkg --add-architecture arm64
sudo apt-get install zlib1g:arm64
sudo apt-get update
sudo ldconfig

  • If apt-get update returns 404 errors for arm64 repos refer to section 5 below.

  • Alternatively the missing arm64 version of libz.so.1 can be downloaded and installed from a .deb package here: https://launchpad.net/ubuntu/wily/arm64/zlib1g/1:1.2.8.dfsg-2ubuntu4

    sudo dpkg -i zlib1g_1.2.8.dfsg-2ubuntu4_arm64.deb

Unable to install arm64 packages after adding arm64 architecture

  • Using sudo apt-get update should add all of the required repos for arm64 but if it does not or you are getting 404 errors the following instructions can be used to add the repos manually:

  • From stackoverflow: https://askubuntu.com/questions/430705/how-to-use-apt-get-to-download-multi-arch-library/430718

  • Open /etc/apt/sources.list with your preferred text editor.

  • Mark all the current (default) repos as [arch=<current_os_arch>], e.g.

    deb [arch=amd64] http://archive.ubuntu.com/ubuntu/ xenial main restricted

  • Then add the following:

    deb [arch=arm64] http://ports.ubuntu.com/ xenial main restricted
deb [arch=arm64] http://ports.ubuntu.com/ xenial-updates main restricted
deb [arch=arm64] http://ports.ubuntu.com/ xenial universe
deb [arch=arm64] http://ports.ubuntu.com/ xenial-updates universe
deb [arch=arm64] http://ports.ubuntu.com/ xenial multiverse
deb [arch=arm64] http://ports.ubuntu.com/ xenial-updates multiverse
deb [arch=arm64] http://ports.ubuntu.com/ xenial-backports main restricted universe multiverse

  • Update and install again:

    sudo apt-get install zlib1g:arm64
sudo apt-get update
sudo ldconfig

Undefined references to google::protobuf:: functions

  • When compiling armNN there are multiple errors of the following type:
    libarmnnCaffeParser.so: undefined reference to `google::protobuf:*
  • Missing or out of date protobuf compilation libraries. Use the command ‘protoc --version’ to check which version of protobuf is available (version 3.5.1 is required). Follow the instructions above to install protobuf 3.5.1 Note this will require you to recompile Caffe for x86_64

Errors on strict-aliasing rules when compiling the Compute Library

  • When compiling the Compute Library there are multiple errors on strict-aliasing rules:
    cc1plus: error: unrecognized command line option ‘-Wno-implicit-fallthrough’ [-Werror]
  • Add Werror=0 to the scons command:
    scons arch=arm64-v8a neon=1 opencl=1 embed_kernels=1 extra_cxx_flags="-fPIC" -j8 internal_only=0 Werror=0