How to use the Android NDK to build Arm NN

Introduction
Initial Setup
Download the Android NDK and make a standalone toolchain
Install Cmake
Build Flatbuffers
Download Arm NN
Get And Build TFLite
Build Arm Compute Library
Build Arm NN
Build Standalone Sample Dynamic Backend
Run the Arm NN unit tests on an Android device

Introduction

These are step-by-step instructions for using the Android NDK to build Arm NN. They have been tested on a clean installation of Ubuntu 18.04 and 20.04, and should also work with other OS versions. The instructions show how to build the Arm NN core library and its dependencies.

For ease of use there is a shell script version of this guide located in the scripts directory called build_android_ndk_guide.sh. Run the script with a -h flag to see the command line parameters.

The shell script version of this guide (build_android_ndk_guide.sh) also provides user the option to use the Arm NN and ComputeLibrary available in your BASE_DIR, instead of downloading a new version. BASE_DIR is path to the script file, which is armnn/scripts/.

Initial Setup

First, we need to specify the Android version and the directories you want to build Arm NN in and to install some applications required to build Arm NN and its dependencies.

export ANDROID_API=30
export WORKING_DIR=$HOME/armnn-devenv
export NDK_DIR=$WORKING_DIR/android-ndk-r25
export NDK_TOOLCHAIN_ROOT=$NDK_DIR/toolchains/llvm/prebuilt/linux-x86_64
export PATH=$NDK_TOOLCHAIN_ROOT/bin/:$PATH

You may want to append the above export variables commands to your ~/.bashrc (or ~/.bash_profile in macOS).

The ANDROID_API variable should be set to the Android API version number you are using. For example, "30" for Android R. The WORKING_DIR can be any directory you have write permissions to.

Required Applications

Git is required to obtain Arm NN. If this has not been already installed then install it using:

sudo apt install git

Arm Compute Library requires SCons. If this has not been already installed then install it using:

sudo apt install scons

CMake is required to build Arm NN and its dependencies. If this has not been already installed then install it using:

sudo apt install cmake

Download the Android NDK and make a standalone toolchain

Download the Android NDK from the official website:

mkdir -p $WORKING_DIR
cd $WORKING_DIR
# For Mac OS, change the NDK download link accordingly.
wget https://dl.google.com/android/repository/android-ndk-r25-linux.zip
unzip android-ndk-r25-linux.zip

With Android NDK-25, you no longer need to use the make_standalone_toolchain script to create a toolchain for a specific version of Android. Android's current preference is for you to just specify the architecture and operating system while setting the compiler and just use the ndk directory.

Install Cmake

Cmake 3.19rc3 or later is required to build Arm NN. If you are using Ubuntu 20.04 the command given in Initial Setup should install a usable version. If you're using Ubuntu 18.04 you may need to compile cmake yourself.

cd $WORKING_DIR
sudo apt-get install libssl-dev
wget https://github.com/Kitware/CMake/releases/download/v3.19.0-rc3/cmake-3.19.0-rc3.tar.gz
tar -zxvf cmake-3.19.0-rc3.tar.gz
cd cmake-3.19.0-rc3
./bootstrap --prefix=$WORKING_DIR/cmake/install
make all install
cd..

Build Flatbuffers

Download Flatbuffers:

cd $WORKING_DIR
wget https://github.com/google/flatbuffers/archive/v23.5.26.tar.gz
tar xf v23.5.26.tar.gz

Build Flatbuffers for x86:

cd $WORKING_DIR/flatbuffers-23.5.26

rm -f CMakeCache.txt

rm -rf build-x86
mkdir build-x86
cd build-x86

rm -rf $WORKING_DIR/flatbuffers-x86
mkdir $WORKING_DIR/flatbuffers-x86

CXXFLAGS="-fPIC" $CMAKE .. \
    -DFLATBUFFERS_BUILD_FLATC=1 \
    -DCMAKE_INSTALL_PREFIX:PATH=$WORKING_DIR/flatbuffers-x86

make all install -j16

Note: -fPIC is added to allow users to use the libraries in shared objects.

Build Flatbuffers for Android:

cd $WORKING_DIR/flatbuffers-23.5.26

rm -f CMakeCache.txt

rm -rf build-android
mkdir build-android
cd build-android

rm -rf $WORKING_DIR/flatbuffers-android
mkdir $WORKING_DIR/flatbuffers-android

CC=/usr/bin/aarch64-linux-gnu-gcc CXX=/usr/bin/aarch64-linux-gnu-g++ \
CXXFLAGS="-fPIC" \
cmake .. \
    -DCMAKE_ANDROID_NDK=$NDK_DIR \
    -DCMAKE_SYSTEM_NAME=Android \
    -DCMAKE_SYSTEM_VERSION=$ANDROID_API \
    -DCMAKE_ANDROID_ARCH_ABI=arm64-v8a \
    -DCMAKE_CXX_FLAGS=--std=c++14 \
    -DFLATBUFFERS_BUILD_FLATC=OFF \
    -DCMAKE_BUILD_TYPE=Release \
    -DFLATBUFFERS_BUILD_TESTS=OFF \
    -DCMAKE_INSTALL_PREFIX=$WORKING_DIR/flatbuffers-android

make all install -j16

Download Arm NN

Clone Arm NN:

cd $WORKING_DIR
git clone https://github.com/ARM-software/armnn.git

Checkout the Arm NN branch:

cd armnn
git checkout <branch_name>
git pull

For example, if you want to check out the 23.02 release branch:

cd armnn
git checkout branches/armnn_23_02
git pull

Get And Build TFLite

This optional step is only required if you intend to build the TFLite delegate or parser for Arm NN.

First clone Tensorflow manually and check out the version Arm NN was tested with:

cd $WORKING_DIR
git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git fetch && git checkout v2.15.0

Or use the script that Arm NN provides:

git fetch && git checkout $(../armnn/scripts/get_tensorflow.sh -p)

Next, set variable TFLITE_ROOT_DIR and build Tensorflow Lite:

export TFLITE_ROOT_DIR=$WORKING_DIR/tensorflow/tensorflow/lite
cd $WORKING_DIR
mkdir -p tflite-out/android
cd tflite-out/android

CMARGS="-DTFLITE_ENABLE_XNNPACK=OFF \
        -DFLATBUFFERS_BUILD_FLATC=OFF \
        -DBUILD_SHARED_LIBS=OFF \
        -DBUILD_TESTING=OFF"

CMARGS="$CMARGS -DCMAKE_TOOLCHAIN_FILE=$NDK_DIR/build/cmake/android.toolchain.cmake \
    -DANDROID_ABI=arm64-v8a \
    -DANDROID_PLATFORM=$ANDROID_API"

cmake $CMARGS $TFLITE_ROOT_DIR

cd $WORKING_DIR
cmake --build tflite-out/android -j 16

Now generate the Tensorflow Lite Schema for the TFLite parser:

cd $WORKING_DIR
mkdir -p $WORKING_DIR/tflite-out/tensorflow/tensorflow/lite/schema

SCHEMA_LOCATION=$WORKING_DIR/tensorflow/tensorflow/lite/schema/schema.fbs
cp $SCHEMA_LOCATION $WORKING_DIR/tflite-out/tensorflow/tensorflow/lite/schema

cd $WORKING_DIR/tflite-out/tensorflow/tensorflow/lite/schema
$WORKING_DIR/flatbuffers-x86/bin/flatc -c --gen-object-api --reflect-types --reflect-names schema.fbs

Build Arm Compute Library

Clone Arm Compute Library:

cd $WORKING_DIR
git clone https://github.com/ARM-software/ComputeLibrary.git

Checkout Arm Compute Library release tag:

cd ComputeLibrary
git checkout <tag_name>

For example, if you want to check out the 23.02 release tag:

cd ComputeLibrary
git checkout v23.02

Arm NN and Arm Compute Library are developed closely together. To use a particular version of Arm NN you will need a compatible version of ACL. Arm NN provides a script that downloads the version of Arm Compute Library that Arm NN was tested with:

git checkout $(../armnn/scripts/get_compute_library.sh -p)

Build the Arm Compute Library:

scons arch=arm64-v8a os=android toolchain_prefix=llvm- compiler_prefix=aarch64-linux-android$ANDROID_API- \
    neon=1 opencl=1 embed_kernels=1 extra_cxx_flags="-fPIC" \
    benchmark_tests=0 validation_tests=0 -j16

Build Arm NN

Build Arm NN:

mkdir $WORKING_DIR/armnn/build
cd $WORKING_DIR/armnn/build
CXX=aarch64-linux-android$ANDROID_API-clang++ \
CC=aarch64-linux-android$ANDROID_API-clang \
CXX_FLAGS="-fPIE -fPIC" \
cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_ANDROID_NDK=$NDK_DIR \
    -DNDK_VERSION=r25 \
    -DCMAKE_SYSTEM_NAME=Android \
    -DCMAKE_SYSTEM_VERSION=$ANDROID_API \
    -DCMAKE_ANDROID_ARCH_ABI=arm64-v8a \
    -DCMAKE_SYSROOT=$WORKING_DIR/android-ndk-r25/toolchains/llvm/prebuilt/linux-x86_64/sysroot \
    -DARMCOMPUTE_ROOT=$WORKING_DIR/ComputeLibrary \
    -DARMCOMPUTE_BUILD_DIR=$WORKING_DIR/ComputeLibrary/build \
    -DARMCOMPUTENEON=1 -DARMCOMPUTECL=1 -DARMNNREF=1 \
    -DFLATBUFFERS_INCLUDE_PATH=$WORKING_DIR/flatbuffers-x86/include \
    -DFLATBUFFERS_ROOT=$WORKING_DIR/flatbuffers-android \
    -DFLATC_DIR=$WORKING_DIR/flatbuffers-x86 \
    -DBUILD_UNIT_TESTS=1 \
    -DBUILD_TESTS=1 \
    -fexception \

To include the Arm NN TFLite delegate add these arguments to the above list:

    -DBUILD_ARMNN_TFLITE_DELEGATE=1 \
    -DTENSORFLOW_ROOT=$WORKING_DIR/tensorflow \
    -DTFLITE_LIB_ROOT=$WORKING_DIR/tflite-out/android \
    -DTFLITE_ROOT_DIR=$WORKING_DIR/tensorflow/tensorflow/lite \

To include the Arm NN TFLite Parser add these arguments to the above list:

    -DBUILD_TF_LITE_PARSER=1 \
    -DTF_LITE_GENERATED_PATH=$WORKING_DIR/tflite-out/tensorflow/tensorflow/lite/schema \
    -DTENSORFLOW_ROOT=$WORKING_DIR/tensorflow \
    -DTFLITE_LIB_ROOT=$WORKING_DIR/tflite-out/android \

To include standalone sample dynamic backend tests, add these arguments to enable the tests and the dynamic backend path to the CMake command:

    -DSAMPLE_DYNAMIC_BACKEND=1 \
    -DDYNAMIC_BACKEND_PATHS=$SAMPLE_DYNAMIC_BACKEND_PATH
# Where $SAMPLE_DYNAMIC_BACKEND_PATH is the path where libArm_SampleDynamic_backend.so library file is pushed

Run the build

make -j16

Build Standalone Sample Dynamic Backend

This step is optional. The sample dynamic backend is located in armnn/src/dynamic/sample

mkdir build
cd build

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

#!/bin/bash
CXX=aarch64-linux-android$ANDROID_API-clang++ \
CC=aarch64-linux-android$ANDROID_API-clang \
CXX_FLAGS="-fPIE -fPIC" \
cmake \
    -DCMAKE_C_COMPILER_WORKS=TRUE \
    -DCMAKE_CXX_COMPILER_WORKS=TRUE \
    -DCMAKE_ANDROID_NDK=$NDK_DIR \
    -DCMAKE_SYSTEM_NAME=Android \
    -DCMAKE_SYSTEM_VERSION=$ANDROID_API \
    -DCMAKE_ANDROID_ARCH_ABI=arm64-v8a \
    -DCMAKE_SYSROOT=$WORKING_DIR/android-ndk-r25/toolchains/llvm/prebuilt/linux-x86_64/sysroot \
    -DCMAKE_CXX_FLAGS=--std=c++14 \
    -DCMAKE_EXE_LINKER_FLAGS="-pie -llog" \
    -DCMAKE_MODULE_LINKER_FLAGS="-llog" \
    -DARMNN_PATH=$WORKING_DIR/armnn/build/libarmnn.so ..

Run the build

make

Run the Arm NN unit tests on an Android device

Push the build results to an Android device and make symbolic links for shared libraries: Currently adb version we have used for testing is 1.0.41.

adb push libarmnn.so /data/local/tmp/
adb push libtimelineDecoder.so /data/local/tmp/
adb push libtimelineDecoderJson.so /data/local/tmp/
adb push GatordMock /data/local/tmp/
adb push libarmnnBasePipeServer.so /data/local/tmp/
adb push libarmnnTestUtils.so /data/local/tmp/
adb push UnitTests /data/local/tmp/
adb push $NDK_DIR/sources/cxx-stl/llvm-libc++/libs/arm64-v8a/libc++_shared.so /data/local/tmp/

Push the files needed for the unit tests (they are a mix of files, directories and symbolic links):

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/testSharedObject
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/testSharedObject/* /data/local/tmp/src/backends/backendsCommon/test/testSharedObject/

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/testDynamicBackend
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/testDynamicBackend/* /data/local/tmp/src/backends/backendsCommon/test/testDynamicBackend/

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath1
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath1/* /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath1/

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath2
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath2/
adb shell ln -s Arm_CpuAcc_backend.so /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so.1
adb shell ln -s Arm_CpuAcc_backend.so.1 /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so.1.2
adb shell ln -s Arm_CpuAcc_backend.so.1.2 /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath2/Arm_CpuAcc_backend.so.1.2.3
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath2/Arm_GpuAcc_backend.so /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath2/
adb shell ln -s nothing /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath2/Arm_no_backend.so

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath3

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath5
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath5/* /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath5/

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath6
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath6/* /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath6/

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath7

adb shell mkdir -p /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath9
adb push -p $WORKING_DIR/armnn/build/src/backends/backendsCommon/test/backendsTestPath9/* /data/local/tmp/src/backends/backendsCommon/test/backendsTestPath9/

adb shell mkdir -p /data/local/tmp/src/backends/dynamic/reference
adb push -p $WORKING_DIR/armnn/build/src/backends/dynamic/reference/Arm_CpuRef_backend.so /data/local/tmp/src/backends/dynamic/reference/

If the standalone sample dynamic tests are enabled, also push libArm_SampleDynamic_backend.so library file to the folder specified as $SAMPLE_DYNAMIC_BACKEND_PATH when Arm NN is built. This is the example when $SAMPLE_DYNAMIC_BACKEND_PATH is specified as /data/local/tmp/dynamic/sample/:

adb shell mkdir -p /data/local/tmp/dynamic/sample/
adb push -p $WORKING_DIR/armnn/src/dynamic/sample/build/libArm_SampleDynamic_backend.so /data/local/tmp/dynamic/sample/

If the delegate was built, push the delegate unit tests too.

adb push $WORKING_DIR/armnn/build/delegate/DelegateUnitTests /data/local/tmp/
adb push $WORKING_DIR/armnn/build/delegate/libarmnnDelegate.so /data/local/tmp/

Run Arm NN unit tests:

adb shell 'LD_LIBRARY_PATH=/data/local/tmp:/vendor/lib64:/vendor/lib64/egl /data/local/tmp/UnitTests'

If the delegate was built run Arm Delegate NN unit tests:

adb shell 'LD_LIBRARY_PATH=/data/local/tmp:/vendor/lib64:/vendor/lib64/egl /data/local/tmp/DelegateUnitTests'

If libarmnnUtils.a is present in $WORKING_DIR/armnn/build/ and the unit tests run without failure then the build was successful.