Example code and instructions on getting Tensorflow Lite running on a Xilinx Zynq

The first step is building a compiler for the Zynq and getting a Linux system up and running. Xilinx has some good resources on how to do that. Although building a custom toolchain and development system with Buildroot is also fairly straightforward.

Download buildroot

git clone git:https://git.busybox.net/buildroot
cd buildroot
make zynq_zed_defconfig
make nconfig

This will bring up a window that allows you to configure buildroot. You can now configure buildroot for your needs. I've checked in my buildroot config located in the buildroot directory. Just copy the config file to .config in the buildroot directory.

While still in the buildroot directory type the following

make -jN

Where N is the number of jobs, generally the number of cores you have. Now go to lunch or bed or something. When you get back Buildroot will have created the output directory. If you look in the output directory you will find the following directories: build host images staging target

The host directory will contain your compiler, libraries, etc. images contains the Xilinx fsbl (first level bootloader), u-boot, the rootfs, and the device tree blobs. It doesn't contain an FPGA image. So you may want to grab a prebuild one and put it on the SD card, otherwise u-boot will fail to boot.

This takes a bit of gynmastics to get working. I thought it should be straightforward, but I was unfamilar with Bazel and the intricacies.

First checkout Tensorflow. The easy method is to copy the tensorflow directory from this repo into your tensorflow repo. Otherwise follow along below

I copied the iOS Simple example and created a command line stand alone program. It is located in the tensorflow/tensorflow/contrib/lite/examples/simplelite directory. It uses OpenCV to load and resize the image before feeding it into the TFLITE.

I took the CROSSTOOL.tpl and made a copy of it. It is located in the tensorflow/third_party/toolchains/cpus/arm/ directory. I couldn't figure out how to get the %{ARM_COMPILER_PATH}% to point to the buildroot compilers. I'm sure there is a more elegant approach, but I simple replaced %{ARM_COMPILER_PATH}% with the hardcoded path to the buildroot compiler.

tool_path { name: "ar" path: "/home/nlbutts/projects/buildroot/output/host/bin/arm-buildroot-linux-gnueabihf-ar" }

I then updated the cxx_builtin_include_directory and some of the compiler flags to force the floating point ABI to use hard.

I did have to add the following line to the BUILD file in the tensorflow/third_party/toolchains/cpus/arm/ directory. This seems like an issue that needs to be fixed.

licenses(["notice"])  # Apache 2.0

The last change I needed to make was to TFLITE's floating point ABI. In tensorflow/tensorflow/contrib/lite/kernels/internal/BUILD they override compiler flags based on the target CPU.

    ":armv7a": [
        "-O3",
        "-mfpu=neon",
        "-mfloat-abi=softfp",
    ],

Android and iOS must use softfp. But in my experience you pay a slight performance penality. Therefore I compiled my Buildroot system to use -mfloat-abi=hard. Therefore this line needs to be changed to:

    ":armv7a": [
        "-O3",
        "-mfpu=neon",
        "-mfloat-abi=hard",
    ],

Now we can cross compile TFLITE for the zynq. Navigate to your Tensorflow directory.

bazel build --crosstool_top=third_party/toolchains/cpus/arm:toolchain --cpu=armv7a tensorflow/contrib/lite/examples/simplelite:simplelite

If everything worked well you should have a statically linked simplelite file in the bazel-bin/tensorflow/contrib/lite/examples/simplelite

Copy that to your Zynq target long with the files in the zynq_target_files directory. Log into your Zynq target and run it with the following commands and you will get the response shown below. I don't know why I get the link error with libneuralnetworks.so. Although it still works.

./simplelite mobilenet.lite labels.txt grace_hopper.jpg
Using graph mobilenet.lite with labels labels.txt and image grace_hopper.jpg
nnapi error: unable to open library libneuralnetworks.so
Loaded model mobilenet.lite
resolved reporter
Getting the input tensor
Resizing input tensor
input image size: 517x606x3
output image size: 224x224x3
Requesting output node
TFLite took 2855029us.
Predictions: 653 0.797  military uniform

As you can see it takes about 2.8 seconds to run the MobileNet_v1. Specifically I used the MobileNet_v1_1.0_224, which should take 569 million MACs per inference. The Zynq is running at 666 MHz. On an Intel i7-6700HQ, this inference takes ~100 ms.