Compile Tensorflow on ARMv8 aarch64 arm64

compile-tensorflow-arm64-aarch64.md

Intro

This wiki explains how to build Tensorflow 1.13.1 for ARMv8 platform (aarch64 arm64)

Build

Preparation

Start a1.4xlarge instance with Ubuntu 16.04 OS and 64GB disk

sudo apt update
sudo apt upgrade
sudo reboot

sudo apt install build-essential

sudo apt install python3 python3-dev

sudo apt install pkg-config zip zlib1g-dev unzip curl tmux wget vim git htop

sudo apt install openjdk-8-jdk

sudo apt install liblapack3 libblas3 libhdf5-dev

sudo update-alternatives --install /usr/bin/python python /usr/bin/python3 10

curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
sudo python3 get-pip.py

sudo pip3 install six mock numpy grpcio
sudo pip3 install h5py==2.10.0
sudo pip3 install keras_applications --no-deps
sudo pip3 install keras_preprocessing --no-deps

Bazel

wget https://github.com/bazelbuild/bazel/releases/download/0.19.2/bazel-0.19.2-dist.zip
mkdir bazel-0.19.2
unzip bazel-0.19.2-dist.zip -d bazel-0.19.2
cd bazel-0.19.2

# open tmux session
tmux
./compile.sh

# Copy bazel to $PATH
sudo cp output/bazel /usr/local/bin/

cd ..

Tensorflow

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git checkout v1.13.1

# run configure. It will ask several questions
# change python path to /usr/bin/python3 
# hit enter for all other questions (default settings)

./configure

# Monolithic option will produce 90 MB wheel file. Without monolithic - 60MB
# --config=monolithic

# Make sure you are still in tmux session. Build will take about 1.5 hours

bazel build -c opt \
--copt=-O3 \
--copt=-std=c++11 \
--copt=-funsafe-math-optimizations \
--copt=-ftree-vectorize \
--copt=-fomit-frame-pointer \
--copt=-DRASPBERRY_PI \
--host_copt=-DRASPBERRY_PI \
--verbose_failures \
--config=noaws \
--config=nogcp \
--config=nohdfs \
--config=noignite \
--config=nokafka \
--config=nonccl \
//tensorflow/tools/pip_package:build_pip_package

# Build a universal wheel.
BDIST_OPTS="--universal" bazel-bin/tensorflow/tools/pip_package/build_pip_package ~/tensorflow_pkg

sudo pip3 install ~/tensorflow_pkg/tensorflow-1.13.1-cp35-cp35m-linux_aarch64.whl

Test

import tensorflow as tf
tf.__version__
tf.__path__

a = tf.placeholder(tf.float32, shape=[2, 2], name="input1")
b = tf.placeholder(tf.float32, shape=[2, 2], name="input2")
ab = tf.matmul(a, b, name="output")

inp1 = [[4., 7.], [2., 6.]]
inp2 = [[0.6, -0.7], [-0.2, 0.4]]

with tf.Session() as sess:
    res = sess.run(ab, feed_dict={a: inp1, b: inp2})

print(res)

Links

https://www.tensorflow.org/install/source https://collaborate.linaro.org/display/BDTS/Building+and+Installing+Tensorflow+on+AArch64 https://github.com/lhelontra/tensorflow-on-arm/releases

For everyone trying to follow this guide: It works basically. Small (and some obvious) changes needed:

• m6g.xlarge (4 CPU, 16GB RAM). 8 GB was not enough as some compile jobs used over 6GB.
• Ubuntu 20.04 (different version gives you different GLIBC versions)
• TF 2.8
• Bazel 4.2.2 (from wget https://github.com/bazelbuild/bazel/releases/download/4.2.2/bazel-4.2.2-linux-arm64)
• I removed the nonccl, nokafka, noignite and nohdfs options
• When using Debian for test purposes, I used openjdk-11-jdk instead of 8 and it seemed to make no difference
• I installed additionally those 2 Python packages: wheel and packaging as otherwise I got errors during the package build process at the end

I then ran into unrelated CPython issues (wrong CPython version on my target system, and too-new GLIBC), but beside this, all worked fine.