Tensorflow_iMX6Q

Tensorflow C++ source on Linux i.MX6 porting guide

1. Goals:

• Cross compile tensorflow static lib, test benchmark on i.MX6Q
• Compile an example(tensorflow/contribs/pi-example/label_image) using libtensorflow.a

2. Enviroment:

• Host: Ubuntu 16.04
• Target: Linux imx6qsabresd 3.14.52
• Cross Compiler: gcc-linaro-arm-linux-gnueabihf-4.8-2014.04_linux
  • Do not use gcc-4.9, it'll cause compile error __atomic_compare_exchange; and I also try gcc-4.6.2, result in --std=c++11 not found error.

3. Build libtensorflow-core.a and benchmark:

cd tensorflow-master
tensorflow/contrib/makefile/download_dependencies.sh

The dependencies locate on tensorflow/contrib/makefile/downloads/: only nsync and protoc need to be cross-compiled, other dependencies only need to use the head files.

• Compile protoc, reference to this guide:
  • The first step is building a native (x86) version of the protobuf libraries and compiler (protoc) since the native compiler is necessary to build the test applications even if you're compiling for a different architecture:
    • Run the ./configure script without argument from tensorflow/contrib/makefile/downloads/protobuf/ than run make from tensorflow/contrib/makefile/downloads/protobuf/src. If everything works you should find the protoc executable in tensorflow/contrib/makefile/downloads/protobuf/src/.libs
    • Copy the protoc executable somewhere (for example in /usr/bin)
  • build the protobuf libraries for ARM:
    • CC=arm-linux-gnueabihf-gcc CXX=arm-linux-gnueabihf-g++ ./configure --host=arm-linux --with-protoc=/usr/bin/protoc
    • rm -r src/.libs
    • Run make in src/
  • Copy useful libs to tensorflow root directory:
    • cd tensorflow-master/ && mkdir build
    • cp -r tensorflow/contrib/makefile/downloads/protobuf/src/.libs build/libs
• Build nsync:
  • Build a native(x86) version:

    export HOST_NSYNC_LIB=`tensorflow/contrib/makefile/compile_nsync.sh`
    

    The object files locate on tensorflow/contrib/makefile/downloads/nsync/builds/default.li nux.c++11/
  • Build a ARM version:

    export TARGET_NSYNC_LIB=`tensorflow/contrib/makefile/compile_nsync.sh -t linux -a armv7`
    

    this command will create a armv7.linux.c++11/ directory in tensorflow/contrib/makefile/downloads/nsync/builds/, and generate native(x86) object files, we need to recompile an arm version:

    cd tensorflow/contrib/makefile/downloads/nsync/builds/armv7.linux.c++11/
    make clean
    vim Makefile
    modify first line to "CC=arm-linux-gnueabihf-g++"
    make
    

• Build tensorflow:

  cd tensorflow-master/
  make -f tensorflow/contrib/makefile/Makefile HOST_OS=PI TARGET=PI OPTFLAGS="-Os" CXX="arm-linux-gnueabihf-g++ -march=armv7-a -mfloat-abi=hard -mfpu=neon -mtune=cortex-a9 --sysroot=/opt/fsl-imx-fb/3.14.52-1.1.0/sysroots/cortexa9hf-vfp-neon-poky-linux-gnueabi -L`pwd`/build/libs"
  

  Don't forget to add your --sysroot!
• Test benchmark: download data model cp tensorflow/contrib/makefile/gen/bin/benchmark and build/libs/* to arm board

  benchmark \
  --graph=./tensorflow_inception_graph.pb \
  --input_layer="input:0" \
  --input_layer_shape="1,224,224,3" \
  --input_layer_type="float" \
  --output_layer="output:0"
  

4. Build label_image:

• Replace tensorflow/contrib/pi_examples/label_image/Makefile with this repository.
• make -f tensorflow/contrib/pi_examples/label_image/Makefile CXX="arm-linux-gnueabihf-g++ -march=armv7-a -mfloat-abi=hard -mfpu=neon -mtune=cortex-a9 --sysroot=/opt/fsl-imx-fb/3.14.52-1.1.0/sysroots/cortexa9hf-vfp-neon-poky-linux-gnueabi -Lpwd/build/libs"
• Run label_image with this guide

Reference:

https://github.com/tensorflow/tensorflow/tree/r1.4/tensorflow/contrib/makefile
https://github.com/eurotech/edc-examples/wiki/Cross-compiling-protobuf-for-ARM-architecture
tensorflow/tensorflow#9259
https://releases.linaro.org/archive/14.04/components/toolchain/binaries/
https://github.com/tensorflow/tensorflow/tree/r1.4/tensorflow/contrib/pi_examples