A high-performance GPU aggregation kernel for GNNs.

Downloading Repo

This project uses Professor Scott Beamer's GAP Benchmark Suite. To clone this repo, use this command.

git clone --recurse-submodules [email protected]:karimsaraipour/cuAgg.git

If you already cloned the repo without the submodules, don't worry! Just run this command.

git submodule update --init --recursive

Installing CMake

. cmake_install.sh

You may need to run chmod u+x cmake_install.sh to give execute permission to the script.

Building Project

Make & enter a build folder for CMake.

mkdir build && cd build

Run CMake build generation. Re-run this when new source files get added.

cmake ..

Build project.

make

The excutables can be found in the build folder.

Generating Graphs

To use Kronecker graph generator, run the generate_graph binary.

./generate_graph -g <scale> -k <degree> [-f out_file_name.g]

Such that the graph will have 2^{scale} number of nodes (before pruning) and an average degere of degree. By default, it'll create a unique output file name, but there's the option name it yourself (with the .g extension).

Contributing to Project

The compile commands are all described in CMakeLists.txt.

To add your own executable (.cpp or .cu file with a int main(int, char**) function), add the following line.

add_executable(<binary name> <path/to/source.cpp or .cu>)

If it uses graph or aggregation related code, also add this command (graph for graph; agg for aggregation).

target_link_libraries(<binary name> graph agg)

To add code related to the graph or aggregation code, modify this line (lines 18 and/or 19).

add_library(<graph or agg> ... <your new source file1> <file2> <etc>)