Gunrock^[1] is a CUDA library for graph-processing designed specifically for the GPU. It uses a high-level, bulk-synchronous/asynchronous, data-centric abstraction focused on operations on vertex or edge frontiers. Gunrock achieves a balance between performance and expressiveness by coupling high-performance GPU computing primitives and optimization strategies, particularly in the area of fine-grained load balancing, with a high-level programming model that allows programmers to quickly develop new graph primitives that scale from one to many GPUs on a node with small code size and minimal GPU programming knowledge.

Branch	Purpose	Stability
main	Default branch, ported from gunrock/essentials, serves as the official release branch.
develop	Development feature branch, ported from gunrock/essentials.
hip-develop	Development feature branch with a working ROCm HIP port, ported from gunrock/essentials.
master	Previous release branch for gunrock/gunrock version 1.x.x interface, preserves all commit history.
dev	Previous development branch for gunrock/gunrock. All changes now merged in master.

Before building Gunrock make sure you have CUDA Toolkit^[2] installed on your system. Other external dependencies such as NVIDIA/thrust, NVIDIA/cub, etc. are automatically fetched using cmake.

git clone https://github.com/gunrock/gunrock.git
cd gunrock
mkdir build && cd build
cmake .. 
make sssp # or for all algorithms, use: make -j$(nproc)
bin/sssp ../datasets/chesapeake/chesapeake.mtx

For a detailed explanation, please see the full documentation. The following example shows simple APIs using Gunrock's data-centric, bulk-synchronous programming model, we implement Breadth-First Search on GPUs. This example skips the setup phase of creating a problem_t and enactor_t struct and jumps straight into the actual algorithm.

We first prepare our frontier with the initial source vertex to begin push-based BFS traversal. A simple f->push_back(source) places the initial vertex we will use for our first iteration.

void prepare_frontier(frontier_t* f,
                      gcuda::multi_context_t& context) override {
  auto P = this->get_problem();
  f->push_back(P->param.single_source);
}

We then begin our iterative loop, which iterates until a convergence condition has been met. If no condition has been specified, the loop converges when the frontier is empty.

void loop(gcuda::multi_context_t& context) override {
  auto E = this->get_enactor();   // Pointer to enactor interface.
  auto P = this->get_problem();   // Pointer to problem (data) interface.
  auto G = P->get_graph();        // Graph that we are processing.

  auto single_source = P->param.single_source;  // Initial source node.
  auto distances = P->result.distances;         // Distances array for BFS.
  auto visited = P->visited.data().get();       // Visited map.
  auto iteration = this->iteration;             // Iteration we are on.

  // Following lambda expression is applied on every source,
  // neighbor, edge, weight tuple during the traversal.
  // Our intent here is to find and update the minimum distance when found.
  // And return which neighbor goes in the output frontier after traversal.
  auto search = [=] __host__ __device__(
                      vertex_t const& source,    // ... source
                      vertex_t const& neighbor,  // neighbor
                      edge_t const& edge,        // edge
                      weight_t const& weight     // weight (tuple).
                      ) -> bool {
    auto old_distance =
      math::atomic::min(&distances[neighbor], iteration + 1);
    return (iteration + 1 < old_distance);
  };

  // Execute advance operator on the search lambda expression.
  // Uses load_balance_t::block_mapped algorithm (try others for perf. tuning.)
  operators::advance::execute<operators::load_balance_t::block_mapped>(
    G, E, search, context);
}

include/gunrock/algorithms/bfs.hxx

Thank you for citing our work.

@article{Wang:2017:GGG,
  author =	 {Yangzihao Wang and Yuechao Pan and Andrew Davidson
                  and Yuduo Wu and Carl Yang and Leyuan Wang and
                  Muhammad Osama and Chenshan Yuan and Weitang Liu and
                  Andy T. Riffel and John D. Owens},
  title =	 {{G}unrock: {GPU} Graph Analytics},
  journal =	 {ACM Transactions on Parallel Computing},
  year =	 2017,
  volume =	 4,
  number =	 1,
  month =	 aug,
  pages =	 {3:1--3:49},
  doi =		 {10.1145/3108140},
  ee =		 {http://arxiv.org/abs/1701.01170},
  acmauthorize = {http://dl.acm.org/authorize?N45082},
  url =		 {http://escholarship.org/uc/item/9gj6r1dj},
  code =	 {https://github.com/gunrock/gunrock},
  ucdcite =	 {a115},
}

@InProceedings{Osama:2022:EOP,
  author =	 {Muhammad Osama and Serban D. Porumbescu and John D. Owens},
  title =	 {Essentials of Parallel Graph Analytics},
  booktitle =	 {Proceedings of the Workshop on Graphs,
                  Architectures, Programming, and Learning},
  year =	 2022,
  series =	 {GrAPL 2022},
  month =	 may,
  pages =	 {314--317},
  doi =		 {10.1109/IPDPSW55747.2022.00061},
  url =          {https://escholarship.org/uc/item/2p19z28q},
}

Gunrock is copyright The Regents of the University of California. The library, examples, and all source code are released under Apache 2.0.

1. This repository has been moved from gunrock/gunrock and the previous history is preserved with tags and under master branch. Read more about gunrock and essentials in our vision paper: Essentials of Parallel Graph Analytics.
2. Recommended CUDA v11.5.1 or higher due to support for stream ordered memory allocators.