Skip to content
This repository has been archived by the owner. It is now read-only.
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

⚠️ [ARCHIVED] libgdf: GPU Dataframes

All development has moved to the cuDF repo effective October 28th 2018

The contents of this repo and the README have been archived for reference. Future development for libgdf will take place in the /libgdf folder of the cuDF repo.

Outstanding PRs

With the refactoring of moving all files to the folder /libgdf in this repo, updating to master should reduce the merge conflicts when trying to merge with master on cuDF. The entire commit history of libgdf has been merged into cuDF to also assist in this transition.

Outstanding Issues

All issues will be copied and migrated to the cuDF repo.

Build Status

libgdf is a C library for implementing common functionality for a GPU Data Frame. For more project details, see the wiki.

Development Setup

The following instructions are tested on Linux and OSX systems.

Compiler requirement:

  • g++ 4.8 or 5.4
  • cmake 3.12+

CUDA requirement:

  • CUDA 9.0+

You can obtain CUDA from

Get dependencies

Note: This repo uses submodules. Make sure you cloned recursively:

git clone --recurse-submodules

Or, after cloning:

cd libgdf
git submodule update --init --recursive

Since cmake will download and build Apache Arrow (version 0.7.1 or 0.8+) you may need to install Boost C++ (version 1.58) before running cmake:

# Install Boost C++ 1.58 for Ubuntu 16.04
$ sudo apt-get install libboost-all-dev


# Install Boost C++ 1.58 for Conda (you will need a Python 3.3 environment)
$ conda install -c omnia boost=1.58.0=py33_0

Libgdf supports Apache Arrow versions 0.7.1 and 0.8+ (0.10.0 is default) that use different metadata versions in IPC. So, it is important to specify which Apache arrow version will be used during building libgdf. To select required Apache Arrow version, define the following environment variables (using Arrow version 0.10.0 as an example):

$ export ARROW_VERSION=0.10.0

where the latter is used by libgdf cmake configuration files. Note that when using libgdf, defining the above environment variables is not necessary.

You can install Boost C++ 1.58 from sources as well:

To run the python tests it is recommended to setup a conda environment for the dependencies.

# create the conda environment (assuming in build directory)
$ conda env create --name libgdf_dev --file ../conda_environments/dev_py35.yml
# activate the environment
$ source activate libgdf_dev
# when not using default arrow version 0.10.0, run
$ conda install pyarrow=$ARROW_VERSION -c conda-forge

This installs the required cmake and pyarrow into the libgdf_dev conda environment and activates it.

For additional information, the python cffi wrapper code requires cffi and pytest. The testing code requires numba and cudatoolkit as an additional dependency. All these are installed from the previous commands.

The environment can be updated from ../conda_environments/dev_py35.yml as development includes/changes the depedencies. To do so, run:

conda env update --name libgdf_dev --file ../conda_environments/dev_py35.yml

Note that dev_py35.yml uses the latest version of pyarrow. Reinstall pyarrow if needed using conda install pyarrow=$ARROW_VERSION -c conda-forge.

Configure and build

This project uses cmake for building the C/C++ library. To configure cmake, run:

$ mkdir build   # create build directory for out-of-source build
$ cd build      # enter the build directory
$ cmake ..      # configure cmake (will download and build Apache Arrow and Google Test)

If installing libgdf to conda environment is desired, then replace the last command with


To build the C/C++ code, run make. This should produce a shared library named or libgdf.dylib.

If you run into compile errors about missing header files:

cub/device/device_segmented_radix_sort.cuh: No such file or directory

See the note about submodules in the Get dependencies section above.

Link python files into the build directory

To make development and testing more seamless, the python files and tests can be symlinked into the build directory by running make copy_python. With that, any changes to the python files are reflected in the build directory. To rebuild the libgdf, run make again.

Run tests

Currently, all tests are written in python with py.test. A make target is available to trigger the test execution. In the build directory (and with the conda environment activated), run below to exceute test:

$ make pytest   # this auto trigger target "copy_python"