Set up guide

Currently, Chainer compiler does not provide any install scripts. You need to try it at the top of the chainer-compiler repository or manually set up paths by yourself.

Prerequisites

If you feel lucky, you can proceed to building chainer compiler and come back to this section if your build fails.

Setting up toolchains

$ apt-get install git curl wget build-essential cmake libopenblas-dev

Setting up CUDA

Download and install the latest CUDA Toolkit from NVIDIA's website.

• https://developer.nvidia.com/cuda-downloads

Non-GPU environment

Chainer compiler can be built on the non-GPU environment because it just requires a few methods in CUDA Runtime API (cuda_runtime.h), and CMake's FindCUDA only requires NVCC.

There are two ways to build Chainer compiler without CUDA.

Specifying CHAINER_COMPILER_ENABLE_CUDA

You can enable CUDA by specifying CHAINER_COMPILER_ENABLE_CUDA=ON.

Using stub driver

NVIDIA provides external package repositories for many Linux distributions, and you can introduce the specific package for non-GPU environment via the package manager instead of installing all components including the non-stub driver.

For example, you only need cuda-cudart-dev-10-0 and cuda-nvcc-10-0 on Ubuntu 18.04:

$ apt-get update
$ apt-get install --no-install-recommends gnupg2 curl ca-certificates
$ curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub | apt-key add
$ echo "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64 /" > /etc/apt/sources.list.d/cuda.list
$ echo "deb https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64 /" > /etc/apt/sources.list.d/nvidia-ml.list
$ apt-get update
$ apt-get install --no-install-recommends cuda-cudart-dev-10-0 cuda-nvcc-10-0
Reading package lists... Done
Building dependency tree
Reading state information... Done
The following additional packages will be installed:
  cuda-cudart-10-0 cuda-driver-dev-10-0 cuda-license-10-0
  cuda-misc-headers-10-0
The following NEW packages will be installed:
  cuda-cudart-10-0 cuda-cudart-dev-10-0 cuda-driver-dev-10-0 cuda-license-10-0
  cuda-misc-headers-10-0 cuda-nvcc-10-0
0 upgraded, 6 newly installed, 0 to remove and 10 not upgraded.
Need to get 21.2 MB of archives.
After this operation, 72.7 MB of additional disk space will be used.
Do you want to continue? [Y/n] y
...

The detailed instruction is described in the official document, and Dockerfiles for nvidia/cuda image are also helpful (see base and devel images).

Ref. Installing CUDA as a non-root user with no GPU - Stack Overflow

Note: cuda-driver-dev-* only installs a stub driver (/usr/local/cuda-10.0/lib64/stubs/libcuda.so) instead of the actual implementation.

Setting up Protocol Buffers

Chainer compiler requires libprotobuf.so and protoc command at build time. You can install them by using the package manager or building yourself.

$ apt-get install libprotobuf-dev protobuf-compiler

Setting up Python and the dependent Python packages

Chainer compiler requires Python and some libraries to build.

Check out the source code if you didn't:

$ git clone https://github.com/pfnet-research/chainer-compiler.git
$ cd chainer-compiler

And run:

$ apt-get install python3 python3-pip
$ git submodule update --init
$ ONNX_ML=1 pip3 install gast==0.3.2 numpy pytest packaging onnx

You need to install Chainer in a submodule directory (third_party/chainer).

$ CHAINER_BUILD_CHAINERX=1 pip3 install third_party/chainer   # install ChainerX without cuda
or
$ CHAINER_BUILD_CHAINERX=1 CHAINERX_BUILD_CUDA=1\
  CUDNN_ROOT_DIR=/path/to/cudnn_root pip3 install third_party/chainer   # install ChainerX with cuda

You need to install third_party/chainer to run its Python interface which requires ABI compatibility.

Building Chainer compiler from source

You may build Chainer compiler from source or install a python package via pip. In this section, we explain how to build from source.

1. Check out Chainer compiler repository to your local environment:

$ git clone https://github.com/pfnet-research/chainer-compiler.git

2. Build Chainer compiler

$ mkdir -p build
$ cd build

$ cmake -DCHAINER_COMPILER_ENABLE_CUDA=ON -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-10.0 ..
or
$ cmake -DCHAINER_COMPILER_ENABLE_CUDA=OFF ..

$ make

Setting up OpenCV (optional)

Install OpenCV via the package manager or by building the source code.

$ apt-get install libopencv-dev

You can enable tools/train_imagenet by adding -DCHAINER_COMPILER_ENABLE_OPENCV=ON.

Other CMake variables

There is a few more CMake variables. Try

$ grep '^option' CMakeLists.txt

to see the list of supported options.

TODO(hamaji): Document some of them. Notably,

1. CHAINER_COMPILER_ENABLE_CUDNN is important for EspNet.
2. CHAINER_COMPILER_ENABLE_PYTHON is necessary for Python interface.

Installing Chainer compiler via pip (optional)

You can install Chainer compiler as a python package. In this case, you do not need to build Chainer compiler from source.

$ CUDA_PATH=/path/to/cuda CUDNN_ROOT_DIR=/path/to/cudnn_root pip3 install .

Run tests

$ cd build
$ make test
$ cd ..
$ ./scripts/runtests.py
$ PYTHONPATH=. pytest tests  # If you set -DCHAINER_COMPILER_ENABLE_PYTHON=ON

Now you can proceed to example usage.