Page History

 A. Local setup

Step 1. Install prerequisites

These are not required, but they are recommended practices. The next steps assume that Homebrew, the package manager manager for macOS, is installed in your device. Otherwise, it can easily be installed by running the following command:

/usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"

1.1. Install and configure git

brew install git
git config --global user.name "<USERNAME>"

1.2. Configure SSH for GitHub

Using SSH, you can connect to GitHub without having to input your credentials every time yo do so. This is not necessary, but it is convenient and time saving in the long run. The steps to configure SSH for this task can be found on GitHub's FAQ.

1.3. Install git-secrets

Git-secrets is a tool that prevents you from committing passwords or sensitive information to your repository. More information can be found on git-secrets' GitHub repo.

...

After installing git-secrets, you must install the hooks for every repo that you wish to use with git secrets install.

1.4. Install Xcode

Download and install Xcode, which contains a suite of software development tools by Apple. It also includes the Xcode IDE, although here we will use CLion as IDE instead. Newer versions of Xcode might not be compatible with your macOS version, you can check the requirements here.

...

sudo xcodebuild -license accept

Step 2. Get the MXNet codebase

Once you have installed the required tools, it is time to get the project's code. 

2.1. Fork the Apache MXNet repository

MXNet is hosted in a GitHub repository, incubator-mxnet. The best practice for development is to fork this repository and work on your local copy.

• Go to https://github.com/apache/incubator-mxnet 
• Click the fork button on the top right.
  
  

2.2. Clone your fork on your machine

git clone --recursive https://github.com/<YOUR GITHUB USERNAME>/incubator-mxnet

...

Next, set git secrets to track the repository with git secrets --install. 

cd incubator-mxnet
git secrets --install

Once the repo is being tracked, you can add different prohibited patterns for git-secrets to control. Check the documentation for further instructions on how to add patterns, exceptions and other utilities.

Step 3. Install additional tools 

Anchor
tools tools

Before installing CLion, there is a set of tools for compiling C++ projects which we need to install.

3.1. Install CMake

CMake is an application to manage builds in a compiler independent way. Installing it with Homebrew is straightforward:

brew install cmake

3.2. Install a compiler

MXNet uses OpenMP, an API for high level parallelism in C++ programs, which is not built in the default C++ compiler in recent versions of macOS. The simplest workaround is to install either LLVM or GCC as compilers.

Option 1. Install LLVM 

LLVM is a collection of compiler and toolchain technologies. It contains Clang, a C++ compiler which delivers fast compiles, as well as other useful tools such as clang-format. It can be installed via Homebrew (be aware that the build might take a long time).

brew install --with-toolchain llvm

Option 2. Install gcc

An alternative to LLVM is GCC, GNU's Compiler Collection. Although slower than Clang, it can also be used to compile MXNet. It can easily be installed with Homebrew.

brew install gcc

3.3. Install ccache

Quoting the documentation, ccache is a compiler cache which speeds up recompilation by caching previous compilations and detecting when the same compilation is being done again. 

brew install ccache

3.4. Install jemalloc

MXNet can use jemalloc for memory allocation, which is a general purpose malloc implementation that emphasizes fragmentation avoidance and scalable concurrency support. 

brew install jemalloc

Step 4. CLion setup 

Start by downloading CLion from JetBrains' website and installing it.

...

Next, we will modify the preferences to use the C++ compiler we previously downloaded, as well as CMake and ccache. Open up the preferences window from CLion > Preferences (or ⌘ + ,) and proceed to the next step.

4.1. Select toolchain

1.  In the preferences window (⌘ + ,), search for toolchains.
2. Click on the + sign to add a toolchain, name it and set the paths of the C and C++ compilers. These paths will depend on the compiler you chose on step 3.2. If you followed the steps above, they should be:

4.2. Build configuration 

Anchor
setup4.2 setup4.2

1. In the preferences window, search for CMake, under the Build, Execution, Deployment section.
2. Add a new profile using the + button on the bottom and give it a name of your choice.
3. Set the toolchain to the one we defined in 4.1.
4. Define the CMake options (explained in the next step, 4.3).

...

4.3. Set CMake options 

Anchor
setup4.3 setup4.3

The following options are used in order to be able to compile in a Mac:

For linux without CUDA you can use the following options:

...

-DUSE_CUDA=OFF
-DUSE_OPENMP=ON
-DUSE_OPENCV=ON
-DUSE_MKL_IF_AVAILABLE=OFF
-DCMAKE_VERBOSE_MAKEFILE=OFF
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache
-DCMAKE_C_COMPILER_LAUNCHER=ccache
-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache

For Linux with CUDA:

CMake options: 
-DUSE_CUDA=ON
-DUSE_OLDCMAKECUDA=OFF
-DUSE_OPENMP=ON
-DUSE_OPENCV=ON
-DUSE_MKL_IF_AVAILABLE=OFF
-DUSE_MKLDNN=OFF
-DUSE_MKLML_MKL=OFF
-DCMAKE_VERBOSE_MAKEFILE=OFF
-DUSE_SIGNAL_HANDLER=ON
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache
-DCMAKE_C_COMPILER_LAUNCHER=ccache
-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache
-DCMAKE_VERBOSE_MAKEFILE=ON

Environment:

CUDACXX=/usr/local/cuda/bin/nvcc

 4.4. Run tests

To make sure everything is working properly, select mxnet_unit_tests as run configuration and run it. 

...

To select specific unit tests, you can click Run → "edit configurations" and add "–gtest_filter=" And a pattern like "SUITE.Test" such as --gtest_filter="inv_khatri_rao.OneInputMatrixTransposed"

Step 5. Select a style file (optional)

CLion provides code generation and formatting tools which can come in handy when developing. Coding style can be defined in the preferences dialogue to have CLion stick to certain coding guidelines. These can be loaded from an external style file, and they can be applied at project level or globally, at IDE level.

5.1. Loading a style file

As a reference, you can use the following style file: mxnet-style.xml.

...

You can use Google C++ coding style as the default since it's the one we use in MXNet:

Step 6. Debugging MXNet backend code with Python frontend

We will first create a small python script that invokes MXNet. For this example we will create a Python script for testing the Quadratic operator from the Operator tutorial 

Image Added

To debug the C++ code, AFAIK you need to have a general idea of the call flow and setup a break point. Here I have setup a breakpoint to the MXListAllOpNames which gets called to generate the operators on the Python Frontend. 

we will setup an Application that calls the python binary and pass the python script. (Note: debugging as a python application won't invoke the debugger)

• To setup an Application for debugging, goto Edit Configurations and add a new Application, 
• Choose the Python binary that you want to execute(If you are compiling inside a Conda environment, choose that version)
• In Program Arguments, choose your python script
• In Working directory, choose your mxnet base path.
• In Environment Variables,

  • set PYTHONPATH to your base path/python(~/mxnet_cpp/python)

  • set MXNET_LIBRARY_PATH to the path of mxnet shared path(  ~/mxnet_cpp/cmake-build-debug-llvm/libmxnet.dylib)

Image Added

• Now you can set a break point in your C++ code and start debugging

Image Added

B. Remote host development

Since version 2018.3 (it is an experimental version as of the writing of this page), CLion will support remote projects. Using this feature, it is possible to code locally in CLion and build, run and test on your remote machine. In order for this feature to work, you need:

...