Skip to content

Latest commit

 

History

History
92 lines (73 loc) · 3.5 KB

build_cpp_utils.md

File metadata and controls

92 lines (73 loc) · 3.5 KB
Raw

Build C++ utility libraries

This document describes how to build and install C++ libraries, headers and executables that can be used for C++ standalone inference and training. The following instruction demonstrates the build procedure on Ubuntu 16.04, but we successfully build C++ libraries on macOS and Windows (manual) too in a similar way (the differences lie in the installation of some dependencies). We may add build instructions on another platform in the future.

Most of the procedure is the same as Build Python Package from Source.

Requirements

  • G++: sudo apt-get install build-essential
  • CMake>=3.1: sudo apt-get install cmake
  • Python: sudo apt-get install python python-pip(Used by code generator)
    • Python packages: PyYAML and MAKO: sudo -H pip install pyyaml mako
  • LibArchive: sudo apt-get install libarchive-dev
  • HDF5 (Optional): sudo apt-get install libhdf5-dev
  • Protobuf >=3: See below.

Installing protobuf3 C++ libraries and tools

Install from source.

Unlike Python Package compilation which requires protoc compiler only, the NNabla C++ utility library requires protobuf C++ library too. The following snippet running on your terminal will build and install protobuf-3.1.0 from source.

curl -L https://github.com/google/protobuf/archive/v3.1.0.tar.gz -o protobuf-v3.1.0.tar.gz
tar xvf protobuf-v3.1.0.tar.gz
cd protobuf-3.1.0
mkdir build && cd build
cmake -DCMAKE_POSITION_INDEPENDENT_CODE=ON -Dprotobuf_BUILD_TESTS=OFF ../cmake
make
sudo make install  # See a note below if you want your system clean.

NOTE: Installing protobuf on your system with sudo may harm a protobuf library previously installed on your system. It is recommended to install a protobuf to a user folder. Specify a cmake option -DCMAKE_INSTALL_PREFIX=<installation path> to set your preferred installation path (e.g. $HOME/nnabla_build_deps), and run make install without sudo. In the NNabla C++ library compilation described below, to help CMake system find the protobuf, pass the installation path to an environment variable CMAKE_FIND_ROOT_PATH. NNabla's CMake script is written to use the value (multiple paths can be passed with a delimiter ;, e.g. CMAKE_FIND_ROOT_PATH="<path A>;<path B>";${CMAKE_FIND_ROOT_PATH}) as search paths to find some packages (e.g. protobuf, hdf5, libarchive) on your system. Please don't forget to set environment variables such as a executable path and library paths (e.g. PATH, LD_LIBRARY_PATH etc) if you have runtime dependencies at your custom installation path.

Install from PPA package.

Here is the procedure using an informal PPA package. If you can not trust unofficial packages, please use the procedure to build from the source shown above.

sudo add-apt-repository ppa:maarten-fonville/protobuf
sudo apt install protobuf-compiler libprotoc-dev libprotobuf-dev

Build

git clone https://github.com/sony/nnabla
mkdir -p nnabla/build && cd nnabla/build
cmake .. -DBUILD_CPP_UTILS=ON -DBUILD_PYTHON_PACKAGE=OFF -DNNABLA_UTILS_WITH_HDF5=ON
make

Some optional arguments for cmake:

  • -DNNABLA_UTILS_WITH_HDF5=OFF to turn off HDF5 feature if you stacked in HDF5 installation.
  • -DNNABLA_UTILS_WITH_NPY=ON to turn on NPY feature if you want to use *.npy cache files.
  • -DBUILD_PYTHON_PACKAGE=ON to build Python package too.

The following command will install the libraries, the command line executables and the include header files to your system.

sudo make install