A set of scripts to (cross-)build the Tensorflow C lib for various architectures / OS

The goal is to provide packages containing the libtensorflow_c.so file as well as a corresponding pc file so that is .so can be used with pkg-config (one use of this is linking tensorflow with rust programs, see the bonus at the end)

We're using these scripts on Ubuntu and Archlinux, the commands given below assume that you're on a fairly recent Ubuntu box.

Theses script are a WIP but a good starting point on how to cross compile tensorflow. Expect to have to edit and tweak them to fit your needs.

Supported OS / Arch:

Before runing these scripts, you need to have bazel

$ echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee /etc/apt/sources.list.d/bazel.list
$ curl https://bazel.build/bazel-release.pub.gpg | sudo apt-key add -
$ sudo apt-get update
$ sudo apt-get install bazel

You'll also need numpy

$ sudo apt-get install python3-numpy

Here is a small description of interesting scripts in this repo

$ cd debian
$ ./create-deb-native.sh # or ./create-deb-armhf.sh then go grab a cup of coffee
$ sudo dpkg -i target/libtensorflow_1.0.0-snips-5_amd64.deb

$ cd archlinux
$ makepkg # then go grab a cup of coffee
$ sudo pacman -S libtensorflow-1.0.0-1-x86_64.pkg.tar.xz

You can use this repo as a brew tap

$ brew tap snipsco/tensorflow https://github.com/snipsco/tensorflow-build 
$ brew install snipsco/tensorflow/tensorflow

You can use tensorflow with your rust projects using the tensorflow crate, its dependency tensorflow-sys will seek the libtensorflow_c.so using pkgconfig and build it if it doesn't find it, which can take some time...

Install the package then cargo build should find the lib \o/

For cross compilation, you need a few more steps, the examples below are for a building for a Raspberry Pi 2/3 from a linux box

First, you need the rust target for the Pi

$ rustup target install armv7-unknown-linux-gnueabihf

Then you need a toolchain for the pi, let's use the one provided by RaspberryPi

$ git clone https://github.com/raspberrypi/tools

Configure cargo to use the toochain by adding this snippet to you ~/.cargo/config

[target.armv7-unknown-linux-gnueabihf]
linker = "/path/to/raspberrypi/tools/arm-bcm2708/gcc-linaro-arm-linux-gnueabihf-raspbian-x64/bin/arm-linux-gnueabihf-gcc"

We now need the .so file of tensorflow for the target arch, let's build the 1.0.0

$ ./compile-arm.sh v1.0.0

The .so is located in target/tensorflow/bazel-bin/tensorflow/libtensorflow.so you may want to move it somewhere else

Let's then generate the .pc file

$ ./create-pkgconfig.sh tensorflow_c /folder/where/the/so/is 1.0.0 "Tensorflow C Library" > /path/to/pc/file

Now that all is done, we can use cargo to crossbuild the app

$ PKG_CONFIG_PATH=/folder/where/the/pcfile/is PKG_CONFIG_ALLOW_CROSS=1 cargo build

Enjoy!