SuiteSparse:GraphBLAS

SPDX-License-Identifier: Apache-2.0

For the GraphBLAS/GraphBLAS Octave/MATLAB interface only: SPDX-License-Identifier: GPL-3.0-or-later (see below for a discussion of the licensing of this package).

VERSION 6.2.4, Mar 8, 2022

SuiteSparse:GraphBLAS is a complete implementation of the GraphBLAS standard, which defines a set of sparse matrix operations on an extended algebra of semirings using an almost unlimited variety of operators and types. When applied to sparse adjacency matrices, these algebraic operations are equivalent to computations on graphs. GraphBLAS provides a powerful and expressive framework for creating graph algorithms based on the elegant mathematics of sparse matrix operations on a semiring.

SuiteSparse:GraphBLAS is used heavily in production. It appears as the underlying graph engine in the RedisGraph database by Redis, and as the built-in sparse matrix multiply in MATLAB R2021a, where C=A*B is now up to 30x faster than in prior versions of MATLAB (on my 20-core NVIDIA DGX Station).

The development of this package is supported by Intel, NVIDIA (including the donation of the 20-core DGX Station), Redis, MIT Lincoln Lab, IBM, and Julia Computing.

See the user guide in Doc/GraphBLAS_UserGuide.pdf for documentation on the SuiteSparse implementation of GraphBLAS, and how to use it in your applications.

See http://graphblas.org for more information on GraphBLAS, including the GraphBLAS C API. See https://github.com/GraphBLAS/GraphBLAS-Pointers for additional resources on GraphBLAS.

QUICK START: To compile and install, do these commands in this directory:

make
sudo make install

Please be patient; some files can take several minutes to compile. Requires an ANSI C11 compiler, so cmake will fail if your compiler is not C11 compliant. See the User Guide PDF in Doc/ for directions on how to use another compiler.

For faster compilation, do this instead of just "make", which uses 32 parallel threads to compile the package:

make JOBS=32

The output of the demo programs will be compared with their expected output.

To remove all compiled files:

make clean

To compile the demos:

make all

See the GraphBLAS/ subfolder for the Octave/MATLAB interface, which contains a README.md file with further details.

Files and folders in this GraphBLAS directory:

CMakeLists.txt: cmake instructions to compile GraphBLAS

Config: version-dependent files used by CMake

Demo: a set of demos on how to use GraphBLAS

Doc: SuiteSparse:GraphBLAS User Guide and license

GraphBLAS: the @GrB Octave/MATLAB interface, including its test suite and demos. This folder is called 'GraphBLAS' so that typing 'help graphblas' or 'doc graphblas' in the Octave or MATLAB Command Window can locate the Contents.m file. Note that this folder and all its contents are under the GNU GPLv3 (or later), not Apache-2.0. All other components of GraphBLAS (in particular, all code in libgraphblas.so, and the entire CUDA folder) are licensed as Apache-2.0.

Include: user-accessible include file, GraphBLAS.h

Makefile: to compile the SuiteSparse:GraphBLAS library and demos

README.md: this file

Source: source files of the SuiteSparse:GraphBLAS library.

Tcov: test coverage, requires Octave or MATLAB. See the log...txt files, which certify 100% test coverage.

Test: Extensive tests, not meant for general usage. To compile and run, go to this directory and type make;testall in Octave or MATLAB. Requires Octave or MATLAB

build: build directory for CMake, initially empty

alternative: an alternative to CMake; edit the alternative/Makefile and do "make" or "make run" in the 'alternative' directory.

CUDA: GPU interface, a work in progress. This is being developed in collaboration with Joe Eaton, Corey Nolet and others at NVIDIA, with support from NVIDIA. It appears in this release but the CUDA folder is a draft that isn't ready to use yet.

CONTRIBUTOR-LICENSE.txt: how to contribute to SuiteSparse:GraphBLAS

cpu_features: (c) Google.com, Apache 2.0 license.

logo: the (awesome!) GraphBLAS logo by Jakab Rokob, CC BY 4.0 license

rmm_wrap: Rapids Memory Manager, (c) NVIDIA, to use with CUDA. (draft; not yet in use)

GraphBLAS C API Specification:

Versions v5.2.0 and earlier conform to the version 1.3.0 (Sept 25, 2019) of the GraphBLAS C API Specification. Versions v6.0.0 and later conform to the version 2.0.0 (Nov, 2021) of the GraphBLAS C API Specification. This library also includes several additional functions and features as extensions to the spec.

All functions, objects, and macros with the prefix GxB are extensions to the spec. Functions, objects, and macros with prefix GB must not be accessed by user code. They are for internal use in GraphBLAS only.

About Benchmarking

Do not use the demos in GraphBLAS/Demos for benchmarking or in production. Those are simple methods for illustration only, and can be slow. Use LAGraph for benchmarking and production uses.

I have tested this package extensively on multicore single-socket systems, but have not yet optimized it for multi-socket systems with a NUMA architecture. That will be done in a future release. If you publish benchmarks with this package, please state the SuiteSparse:GraphBLAS version, and a caveat if appropriate. If you see significant performance issues when going from a single-socket to multi-socket system, I would like to hear from you so I can look into it.

Contact me at [email protected] for any questions about benchmarking SuiteSparse:GraphBLAS and LAGraph.

Contributing to SuiteSparse:GraphBLAS

To add an issue for a bug report (gasp!) or a feature request, you can use the issue tracker on github.com, at [https://github.com/DrTimothyAldenDavis/GraphBLAS/issues] (https://github.com/DrTimothyAldenDavis/GraphBLAS/issues) or [https://github.com/DrTimothyAldenDavis/SuiteSparse/issues] (https://github.com/DrTimothyAldenDavis/SuiteSparse/issues).

To contribute code, you can submit a pull request. To do so, you must first agree to the Contributor License Agreement CONTRIBUTOR-LICENSE.txt. Print a copy of the txt file (as a PDF), sign and date it, and email it to me at [email protected]. Pull requests will only be included into SuiteSparse after I receive your email with the signed PDF.

Licensing and supporting SuiteSparse:GraphBLAS

SuiteSparse:GraphBLAS is released primarily under the Apache-2.0 license, because of how the project is supported by many organizations (NVIDIA, Redis, MIT Lincoln Lab, Intel, IBM, and Julia Computing), primarily through gifts to the Texas A&M Foundation. Because of this support, and to facilitate the wide-spread use of GraphBLAS, the decision was made to give this library a permissive open-source license (Apache-2.0). Currently all source code required to create the C-callable library libgraphblas.so is licensed with Apache-2.0, and there are no plans to change this.

However, just because this code is free to use doesn't make it zero-cost to create. If you are using GraphBLAS in a commercial closed-source product and are not supporting its development, please consider supporting this project to ensure that it will continue to be developed and enhanced in the future.

The Octave/MATLAB interface is now licensed under the GNU GPLv3.0 (or later) license, so if any code from the GraphBLAS/GraphBLAS folder is needed in a commercial closed-source application, you will either need to license your commercial application as GNU GPLv3.0, or you can ask me for a non-GNU and non-Apache license.

This licensing change has no impact at all on the use of GraphBLAS in Octave, since Octave itself is already under the GNU GPLv3.0 license.

To support the development of GraphBLAS, contact the author ([email protected]) or the Texas A&M Foundation (True Brown, [email protected]; or Kevin McGinnis, [email protected]) for details.

Software Acknowledgements

SuiteSparse:GraphBLAS relies on the following packages (details in the LICENSE file, and in the GraphBLAS User Guide):

(1) the LZ4 compression package (c) by Yann Collet, appearing here under the BSD-2-clause license.

(2) cpu_features (c) Google, Apache 2.0 license with components (c) IBM and Intel (also Apache 2.0), and the cpu_featurer/ndk_compat component (c) The Android Open Source Project (BSD-2-clause)

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

DrTimothyAldenDavis / GraphBLAS