Skip to content

Latest commit

 

History

History
1256 lines (848 loc) · 43.3 KB

README.md

File metadata and controls

1256 lines (848 loc) · 43.3 KB
Raw

SuiteSparse: A Suite of Sparse matrix packages at http://suitesparse.com

Oct 10, 2024, SuiteSparse VERSION 7.8.3

SuiteSparse is a set of sparse-matrix-related packages written or co-authored by Tim Davis, available at https://github.com/DrTimothyAldenDavis/SuiteSparse .

Primary author of SuiteSparse (codes and algorithms, excl. METIS): Tim Davis

Code co-authors, in alphabetical order (not including METIS or LAGraph): Patrick Amestoy, Mohsen Aznaveh, David Bateman, Jinhao Chen, Yanqing Chen, Iain Duff, Joe Eaton, Les Foster, William Hager, Raye Kimmerer, Scott Kolodziej, Chris Lourenco, Stefan Larimore, Lorena Mejia Domenzain, Erick Moreno-Centeno, Markus Mützel, Corey Nolel, Ekanathan Palamadai, Sivasankaran Rajamanickam, Sanjay Ranka, Wissam Sid-Lakhdar, and Nuri Yeralan.

LAGraph has been developed by the highest number of developers of any of the packages in SuiteSparse and deserves its own list. The list also appears in LAGraph/Contibutors.txt:

Janos B. Antal,    Budapest University of Technology and Economics, Hungary
Mohsen Aznaveh,    Texas A&M University
David A. Bader     New Jersey Institute of Technology
Aydin Buluc,       Lawrence Berkeley National Lab
Jinhao Chen,       Texas A&M University
Tim Davis,         Texas A&M University
Florentin Dorre,   Technische Univeritat Dresden, Neo4j
Marton Elekes,     Budapest University of Technology and Economics, Hungary
Balint Hegyi,      Budapest University of Technology and Economics, Hungary
Tanner Hoke,       Texas A&M University
James Kitchen,     Anaconda
Scott Kolodziej,   Texas A&M University
Pranav Konduri,    Texas A&M University
Roi Lipman,        Redis Labs (now FalkorDB)
Tze Meng Low,      Carnegie Mellon University
Tim Mattson,       Intel
Scott McMillan,    Carnegie Mellon University
Markus Muetzel
Michel Pelletier,  Graphegon
Gabor Szarnyas,    CWI Amsterdam, The Netherlands
Erik Welch,        Anaconda, NVIDIA
Carl Yang,         University of California at Davis, Waymo
Yongzhe Zhang,     SOKENDAI, Japan

METIS is authored by George Karypis.

Additional algorithm designers: Esmond Ng and John Gilbert.

Refer to each package for license, copyright, and author information.

Documentation

Refer to each package for the documentation on each package, typically in the Doc subfolder.

SuiteSparse branches

  • dev: the default branch, with recent updates of features to appear in the next stable release. The intent is to keep this branch in fully working order at all times, but the features will not be finalized at any given time.
  • stable: the most recent stable release.
  • dev2: working branch. All submitted PRs should made to this branch. This branch might not always be in working order.

SuiteSparse Packages

Packages in SuiteSparse, and files in this directory:

  • AMD

    approximate minimum degree ordering. This is the built-in AMD function in MATLAB.

    authors: Tim Davis, Patrick Amestoy, Iain Duff

  • bin

    where programs are placed when compiled, for make local

  • BTF

    permutation to block triangular form

    authors: Tim Davis, Ekanathan Palamadai

  • build

    folder for default build tree

  • CAMD

    constrained approximate minimum degree ordering

    authors: Tim Davis, Patrick Amestoy, Iain Duff, Yanqing Chen

  • CCOLAMD

    constrained column approximate minimum degree ordering

    authors: Tim Davis, Sivasankaran Rajamanickam, Stefan Larimore.

    Algorithm design collaborators: Esmond Ng, John Gilbert (for COLAMD)

  • ChangeLog

    a summary of changes to SuiteSparse. See */Doc/ChangeLog for details for each package.

  • CHOLMOD

    sparse Cholesky factorization. Requires AMD, COLAMD, CCOLAMD, the BLAS, and LAPACK. Optionally uses METIS. This is chol and x=A\b in MATLAB.

    author for all modules: Tim Davis

    CHOLMOD/Modify module authors: Tim Davis and William W. Hager

    CHOLMOD/SuiteSparse_metis: a modified version of METIS, embedded into the CHOLMOD library. See the README.txt files for details. author: George Karypis. This is a slightly modified copy included with SuiteSparse via the open-source license provided by George Karypis. SuiteSparse cannot use an unmodified copy of METIS.

  • CITATION.bib

    citations for SuiteSparse packages, in bibtex format.

  • CMakeLists.txt

    optional, to compile all of SuiteSparse. See below.

  • CODE_OF_CONDUCT.md

    community guidelines

  • COLAMD

    column approximate minimum degree ordering. This is the built-in COLAMD function in MATLAB.

    authors (of the code): Tim Davis and Stefan Larimore

    Algorithm design collaborators: Esmond Ng, John Gilbert

  • Contents.m

    a list of contents for 'help SuiteSparse' in MATLAB.

  • CONTRIBUTING.md

    how to contribute to SuiteSparse

  • CONTRIBUTOR-LICENSE.txt

    required contributor agreement

  • CSparse

    a concise sparse matrix package, developed for my book, "Direct Methods for Sparse Linear Systems", published by SIAM. Intended primarily for teaching. Note that the code is (c) Tim Davis, as stated in the book.

    For production, use CXSparse instead. In particular, both CSparse and CXSparse have the same include filename: cs.h. This package is used for the built-in DMPERM in MATLAB.

    author: Tim Davis

  • CXSparse

    CSparse Extended. Includes support for complex matrices and both int or long integers. Use this instead of CSparse for production use; it creates a libcsparse.so (or dylib on the Mac) with the same name as CSparse. It is a superset of CSparse. Any code that links against CSparse should also be able to link against CXSparse instead.

    author: Tim Davis, David Bateman

  • Example

    a simple package that relies on almost all of SuiteSparse

  • .github

    workflows for CI testing on GitHub.

  • GraphBLAS

    graph algorithms in the language of linear algebra.

    https://graphblas.org

    authors: Tim Davis, Joe Eaton, Corey Nolet

  • include

    make install places user-visible include files for each package here, after make local.

  • KLU

    sparse LU factorization, primarily for circuit simulation. Requires AMD, COLAMD, and BTF. Optionally uses CHOLMOD, CAMD, CCOLAMD, and METIS.

    authors: Tim Davis, Ekanathan Palamadai

  • LAGraph

    a graph algorithms library based on GraphBLAS. See also https://github.com/GraphBLAS/LAGraph

    Authors: many.

  • LDL

    a very concise LDL' factorization package

    author: Tim Davis

  • lib

    make install places shared libraries for each package here, after make local.

  • LICENSE.txt

    collected licenses for each package.

  • Makefile

    optional, to compile all of SuiteSparse using make, which is used as a simple wrapper for cmake in each subproject.

    • make

      compiles SuiteSparse libraries. Subsequent make install will install in CMAKE_INSTALL_PATH (might default to /usr/local/lib on Linux or Mac).

    • make local

      compiles SuiteSparse. Subsequent make install will install in ./lib, ./include. Does not install in CMAKE_INSTALL_PATH.

    • make global

      compiles SuiteSparse libraries. Subsequent make install will install in /usr/local/lib (or whatever the configured CMAKE_INSTALL_PREFIX is). Does not install in ./lib and ./include.

    • make install

      installs in the current directory (./lib, ./include), or in /usr/local/lib and /usr/local/include, (the latter defined by CMAKE_INSTALL_PREFIX) depending on whether make, make local, or make global has been done.

    • make uninstall

      undoes make install.

    • make distclean

      removes all files not in distribution, including ./bin, ./share, ./lib, and ./include.

    • make purge

      same as make distclean.

    • make clean

      removes all files not in distribution, but keeps compiled libraries and demos, ./lib, ./share, and ./include.

    Each individual subproject also has each of the above make targets.

    Things you don't need to do:

    • make docs

      creates user guides from LaTeX files

    • make cov

      runs statement coverage tests (Linux only)

  • MATLAB_Tools

    various m-files for use in MATLAB

    author: Tim Davis (all parts)

    for spqr_rank: author Les Foster and Tim Davis

    • Contents.m

      list of contents

    • dimacs10

      loads matrices for DIMACS10 collection

    • Factorize

      object-oriented x=A\b for MATLAB

    • find_components

      finds connected components in an image

    • GEE

      simple Gaussian elimination

    • getversion.m

      determine MATLAB version

    • gipper.m

      create MATLAB archive

    • hprintf.m

      print hyperlinks in command window

    • LINFACTOR

      predecessor to Factorize package

    • MESHND

      nested dissection ordering of regular meshes

    • pagerankdemo.m

      illustrates how PageRank works

    • SFMULT

      C=S*F where S is sparse and F is full

    • shellgui

      display a seashell

    • sparseinv

      sparse inverse subset

    • spok

      check if a sparse matrix is valid

    • spqr_rank

      SPQR_RANK package. MATLAB toolbox for rank deficient sparse matrices: null spaces, reliable factorizations, etc. With Leslie Foster, San Jose State Univ.

    • SSMULT

      C=A*B where A and B are both sparse. This was the basis for the built-in C=A*B in MATLAB, until it was superseded by GraphBLAS in MATLAB R2021a.

    • SuiteSparseCollection

      for the SuiteSparse Matrix Collection

    • waitmex

      waitbar for use inside a mexFunction

  • Mongoose

    graph partitioning.

    authors: Nuri Yeralan, Scott Kolodziej, William Hager, Tim Davis

  • ParU

    a parallel unsymmetric pattern multifrontal method.

    Currently a pre-release.

    authors: Mohsen Aznaveh and Tim Davis

  • RBio

    read/write sparse matrices in Rutherford/Boeing format

    author: Tim Davis

  • README.md

    this file

  • SPEX

    solves sparse linear systems in exact arithmetic.

    Requires the GNU GMP and MPRF libraries.

    This will be soon replaced by a more general package, SPEX v3 that includes this method (exact sparse LU) and others (sparse exact Cholesky, and sparse exact update/downdate). The API of v3 will be changing significantly.

    authors: Chris Lourenco, Jinhao Chen, Erick Moreno-Centeno, Lorena Lorena Mejia Domenzain, and Tim Davis.

    See https://github.com/clouren/SPEX for the latest version.

  • SPQR

    sparse QR factorization. This the built-in qr and x=A\b in MATLAB. Also called SuiteSparseQR.

    Includes two GPU libraries: SPQR/GPUQREngine and SPQR/SuiteSparse_GPURuntime.

    author of the CPU code: Tim Davis

    author of GPU modules: Tim Davis, Nuri Yeralan, Wissam Sid-Lakhdar, Sanjay Ranka

  • ssget

    MATLAB interface to the SuiteSparse Matrix Collection

    author: Tim Davis

  • SuiteSparse_config

    library with common functions and configuration for all the above packages. CSparse, GraphBLAS, LAGraph, and MATLAB_Tools do not use SuiteSparse_config.

    author: Tim Davis

  • SuiteSparse_demo.m

    a demo of SuiteSparse for MATLAB

  • SuiteSparse_install.m

    install SuiteSparse for MATLAB

  • SuiteSparse_paths.m

    set paths for SuiteSparse MATLAB mexFunctions

  • SuiteSparse_test.m

    exhaustive test for SuiteSparse in MATLAB

  • UMFPACK

    sparse LU factorization. Requires AMD and the BLAS.

    This is the built-in lu and x=A\b in MATLAB.

    author: Tim Davis

    algorithm design collaboration: Iain Duff

Refer to each package for license, copyright, and author information. All codes are authored or co-authored by Timothy A. Davis (email: [email protected]), except for METIS (by George Karypis), GraphBLAS/cpu_features (by Google), GraphBLAS/lz4, zstd, and xxHash (by Yann Collet, now at Facebook), and GraphBLAS/CUDA/jitify.hpp (by NVIDIA). Parts of GraphBLAS/CUDA are Copyright (c) by NVIDIA. Please refer to each of these licenses.

For distro maintainers (Linux, homebrew, spack, R, Octave, Trilinos, ...):

Thanks for packaging SuiteSparse! Here are some suggestions:

  • GraphBLAS takes a long time to compile because it creates many fast "FactoryKernels" at compile-time. If you want to reduce the compile time and library size, enable the GRAPHBLAS_COMPACT mode, but keep the JIT compiler enabled. Then GraphBLAS will compile the kernels it needs at run-time, via its JIT compiler. Performance will be the same as the FactoryKernels once the JIT kernels are compiled. User compiled kernels are placed in ~/.SuiteSparse, by default. You do not need to distribute the source for GraphBLAS to enable the JIT compiler: just libgraphblas.so and GraphBLAS.h is enough.

  • GraphBLAS needs OpenMP! It's fundamentally a parallel code so please distribute it with OpenMP enabled. Performance will suffer otherwise.

  • CUDA acceleration: CHOLMOD and SPQR can benefit from their CUDA kernels. If you do not have CUDA or do not want to include it in your distro, this version of SuiteSparse skips the building of the CHOLMOD_CUDA and SPQR_CUDA libraries, and does not link against the GPUQREngine and SuiteSparse_GPURuntime libraries.

How to cite the SuiteSparse meta-package and its component packages:

SuiteSparse is a meta-package of many packages, each with their own published papers. To cite the whole collection, use the URLs:

Please also cite the specific papers for the packages you use. This is a long list; if you want a shorter list, just cite the most recent "Algorithm XXX:" papers in ACM TOMS, for each package.

  • For the MATLAB x=A\b, see below for AMD, COLAMD, CHOLMOD, UMFPACK, and SuiteSparseQR (SPQR).

  • for GraphBLAS, and C=AB in MATLAB (sparse-times-sparse):

    T. A. Davis. Algorithm 1037: SuiteSparse:GraphBLAS: Parallel Graph Algorithms in the Language of Sparse Linear Algebra. ACM Trans. Math. Softw. 49, 3, Article 28 (September 2023), 30 pages. https://doi.org/10.1145/3577195

    T. Davis, Algorithm 1000: SuiteSparse:GraphBLAS: graph algorithms in the language of sparse linear algebra, ACM Trans on Mathematical Software, vol 45, no 4, Dec. 2019, Article No 44. https://doi.org/10.1145/3322125.

  • for LAGraph:

    G. Szárnyas et al., "LAGraph: Linear Algebra, Network Analysis Libraries, and the Study of Graph Algorithms," 2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), Portland, OR, USA, 2021, pp. 243-252. https://doi.org/10.1109/IPDPSW52791.2021.00046.

  • for CSparse/CXSParse:

    T. A. Davis, Direct Methods for Sparse Linear Systems, SIAM Series on the Fundamentals of Algorithms, SIAM, Philadelphia, PA, 2006. https://doi.org/10.1137/1.9780898718881

  • for SuiteSparseQR (SPQR): (also cite AMD, COLAMD):

    T. A. Davis, Algorithm 915: SuiteSparseQR: Multifrontal multithreaded rank-revealing sparse QR factorization, ACM Trans. on Mathematical Software, 38(1), 2011, pp. 8:1--8:22. https://doi.org/10.1145/2049662.2049670

  • for SuiteSparseQR/GPU:

    Sencer Nuri Yeralan, T. A. Davis, Wissam M. Sid-Lakhdar, and Sanjay Ranka. 2017. Algorithm 980: Sparse QR Factorization on the GPU. ACM Trans. Math. Softw. 44, 2, Article 17 (June 2018), 29 pages. https://doi.org/10.1145/3065870

  • for CHOLMOD: (also cite AMD, COLAMD):

    Y. Chen, T. A. Davis, W. W. Hager, and S. Rajamanickam, Algorithm 887: CHOLMOD, supernodal sparse Cholesky factorization and update/downdate, ACM Trans. on Mathematical Software, 35(3), 2008, pp. 22:1--22:14. https://dl.acm.org/doi/abs/10.1145/1391989.1391995

    T. A. Davis and W. W. Hager, Dynamic supernodes in sparse Cholesky update/downdate and triangular solves, ACM Trans. on Mathematical Software, 35(4), 2009, pp. 27:1--27:23. https://doi.org/10.1145/1462173.1462176

  • for CHOLMOD/Modify Module: (also cite AMD, COLAMD):

    T. A. Davis and William W. Hager, Row Modifications of a Sparse Cholesky Factorization SIAM Journal on Matrix Analysis and Applications 2005 26:3, 621-639. https://doi.org/10.1137/S089547980343641X

    T. A. Davis and William W. Hager, Multiple-Rank Modifications of a Sparse Cholesky Factorization SIAM Journal on Matrix Analysis and Applications 2001 22:4, 997-1013. https://doi.org/10.1137/S0895479899357346

    T. A. Davis and William W. Hager, Modifying a Sparse Cholesky Factorization, SIAM Journal on Matrix Analysis and Applications 1999 20:3, 606-627. https://doi.org/10.1137/S0895479897321076

  • for CHOLMOD/GPU Modules:

    Steven C. Rennich, Darko Stosic, Timothy A. Davis, Accelerating sparse Cholesky factorization on GPUs, Parallel Computing, Vol 59, 2016, pp 140-150. https://doi.org/10.1016/j.parco.2016.06.004

  • for AMD and CAMD:

    P. Amestoy, T. A. Davis, and I. S. Duff, Algorithm 837: An approximate minimum degree ordering algorithm, ACM Trans. on Mathematical Software, 30(3), 2004, pp. 381--388. https://dl.acm.org/doi/abs/10.1145/1024074.1024081

    P. Amestoy, T. A. Davis, and I. S. Duff, An approximate minimum degree ordering algorithm, SIAM J. Matrix Analysis and Applications, 17(4), 1996, pp. 886--905. https://doi.org/10.1137/S0895479894278952

  • for COLAMD, SYMAMD, CCOLAMD, and CSYMAMD:

    T. A. Davis, J. R. Gilbert, S. Larimore, E. Ng, Algorithm 836: COLAMD, an approximate column minimum degree ordering algorithm, ACM Trans. on Mathematical Software, 30(3), 2004, pp. 377--380. https://doi.org/10.1145/1024074.1024080

    T. A. Davis, J. R. Gilbert, S. Larimore, E. Ng, A column approximate minimum degree ordering algorithm, ACM Trans. on Mathematical Software, 30(3), 2004, pp. 353--376. https://doi.org/10.1145/1024074.1024079

  • for UMFPACK: (also cite AMD and COLAMD):

    T. A. Davis, Algorithm 832: UMFPACK - an unsymmetric-pattern multifrontal method with a column pre-ordering strategy, ACM Trans. on Mathematical Software, 30(2), 2004, pp. 196--199. https://dl.acm.org/doi/abs/10.1145/992200.992206

    T. A. Davis, A column pre-ordering strategy for the unsymmetric-pattern multifrontal method, ACM Trans. on Mathematical Software, 30(2), 2004, pp. 165--195. https://dl.acm.org/doi/abs/10.1145/992200.992205

    T. A. Davis and I. S. Duff, A combined unifrontal/multifrontal method for unsymmetric sparse matrices, ACM Trans. on Mathematical Software, 25(1), 1999, pp. 1--19. https://doi.org/10.1145/305658.287640

    T. A. Davis and I. S. Duff, An unsymmetric-pattern multifrontal method for sparse LU factorization, SIAM J. Matrix Analysis and Computations, 18(1), 1997, pp. 140--158. https://doi.org/10.1137/S0895479894246905

  • for the FACTORIZE m-file:

    T. A. Davis, Algorithm 930: FACTORIZE, an object-oriented linear system solver for MATLAB, ACM Trans. on Mathematical Software, 39(4), 2013, pp. 28:1-28:18. https://doi.org/10.1145/2491491.2491498

  • for KLU and BTF (also cite AMD and COLAMD):

    T. A. Davis and Ekanathan Palamadai Natarajan. 2010. Algorithm 907: KLU, A Direct Sparse Solver for Circuit Simulation Problems. ACM Trans. Math. Softw. 37, 3, Article 36 (September 2010), 17 pages. https://dl.acm.org/doi/abs/10.1145/1824801.1824814

  • for LDL:

    T. A. Davis. Algorithm 849: A concise sparse Cholesky factorization package. ACM Trans. Math. Softw. 31, 4 (December 2005), 587–591. https://doi.org/10.1145/1114268.1114277

  • for ssget and the SuiteSparse Matrix Collection:

    T. A. Davis and Yifan Hu. 2011. The University of Florida sparse matrix collection. ACM Trans. Math. Softw. 38, 1, Article 1 (November 2011), 25 pages. https://doi.org/10.1145/2049662.2049663

    Kolodziej et al., (2019). The SuiteSparse Matrix Collection Website Interface. Journal of Open Source Software, 4(35), 1244. https://doi.org/10.21105/joss.01244

  • for spqr_rank:

    Leslie V. Foster and T. A. Davis. 2013. Algorithm 933: Reliable calculation of numerical rank, null space bases, pseudoinverse solutions, and basic solutions using suitesparseQR. ACM Trans. Math. Softw. 40, 1, Article 7 (September 2013), 23 pages. https://doi.org/10.1145/2513109.2513116

  • for Mongoose:

    T. A. Davis, William W. Hager, Scott P. Kolodziej, and S. Nuri Yeralan. 2020. Algorithm 1003: Mongoose, a Graph Coarsening and Partitioning Library. ACM Trans. Math. Softw. 46, 1, Article 7 (March 2020), 18 pages. https://doi.org/10.1145/3337792

  • for SPEX:

    Christopher Lourenco, Jinhao Chen, Erick Moreno-Centeno, and T. A. Davis. 2022. Algorithm 1021: SPEX Left LU, Exactly Solving Sparse Linear Systems via a Sparse Left-Looking Integer-Preserving LU Factorization. ACM Trans. Math. Softw. June 2022. https://doi.org/10.1145/3519024

About the BLAS and LAPACK libraries

NOTE: if you use OpenBLAS, be sure to use version 0.3.27 or later.

To select your BLAS/LAPACK, see the instructions in SuiteSparseBLAS.cmake in SuiteSparse_config/cmake_modules. If SuiteSparse_config finds a BLAS with 64-bit integers (such as the Intel MKL ilp64 BLAS), it configures SuiteSparse_config.h with the SUITESPARSE_BLAS_INT defined as int64_t. Otherwise, if a 32-bit BLAS is found, this type is defined as int32_t. If later on, UMFPACK, CHOLMOD, or SPQR are compiled and linked with a BLAS that has a different integer size, you must override the definition with -DBLAS64 (to assert the use of 64-bit integers in the BLAS) or -DBLAS32, (to assert the use of 32-bit integers in the BLAS).

The size of the BLAS integer has nothing to do with sizeof(void *).

When distributed in a binary form (such as a Debian, Ubuntu, Spack, or Brew package), SuiteSparse should probably be compiled to expect a 32-bit BLAS, since this is the most common case. The default is to use a 32-bit BLAS, but this can be changed by setting the cmake variable SUITESPARSE_USE_64BIT_BLAS to ON.

By default, SuiteSparse hunts for a suitable BLAS library. To enforce a particular BLAS library use either:

CMAKE_OPTIONS="-DBLA_VENDOR=OpenBLAS" make
cd Package ; cmake -DBLA_VENDOR=OpenBLAS .. make

To use the default (hunt for a BLAS), do not set BLA_VENDOR, or set it to an empty string. In this case, if SUITESPARSE_USE_64BIT_BLAS is ON, preference is given to a 64-bit BLAS, but a 32-bit BLAS library will be used if no 64-bit library is found. However, if both SUITESPARSE_USE_64BIT_BLAS and SUITESPARSE_USE_STRICT are ON, then only a 64-bit BLAS is considered.

When selecting a particular BLAS library, the SUITESPARSE_USE_64BIT_BLAS setting is strictly followed. If set to true, only a 64-bit BLAS library will be used. If false (the default), only a 32-bit BLAS library will be used. If no such BLAS is found, the build will fail.

QUICK START FOR THE C/C++ LIBRARIES:

Type the following in this directory (requires system priviledge to do the sudo make install):

    mkdir -p build && cd build
    cmake ..
    cmake --build .
    sudo cmake --install .

All libraries will be created and installed into the default system-wide folder (/usr/local/lib on Linux). All include files needed by the applications that use SuiteSparse are installed into /usr/local/include/suitesparse (on Linux).

To build only a subset of libraries, set SUITESPARSE_ENABLE_PROJECTS when configuring with CMake. E.g., to build and install CHOLMOD and CXSparse (including their dependencies), use the following commands:

    mkdir -p build && cd build
    cmake -DSUITESPARSE_ENABLE_PROJECTS="cholmod;cxsparse" ..
    cmake --build .
    sudo cmake --install .

For Windows (MSVC), import the CMakeLists.txt file into MS Visual Studio. Be sure to specify the build type as Release; for example, to build SuiteSparse on Windows in the command window, run:

    mkdir -p build && cd build
    cmake ..
    cmake --build . --config Release
    cmake --install .

Be sure to first install all required libraries: BLAS and LAPACK for UMFPACK, CHOLMOD, and SPQR, and GMP and MPFR for SPEX. Be sure to use the latest libraries; SPEX requires MPFR 4.0.2 and GMP 6.1.2 (these version numbers do NOT correspond to the X.Y.Z suffix of libgmp.so.X.Y.Z and libmpfr.so.X.Y.Z; see the SPEX user guide for details).

To compile the libraries and install them only in SuiteSparse/lib (not /usr/local/lib), do this instead in the top-level of SuiteSparse:

    mkdir -p build && cd build
    cmake -DCMAKE_INSTALL_PREFIX=.. ..
    cmake --build .
    cmake --install .

If you add /home/me/SuiteSparse/lib to your library search path (LD_LIBRARY_PATH in Linux), you can do the following (for example):

    S = /home/me/SuiteSparse
    cc myprogram.c -I$(S)/include/suitesparse -lumfpack -lamd -lcholmod -lsuitesparseconfig -lm

To change the C and C++ compilers, and to compile in parallel use:

    cmake -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER==g++ ..

for example, which changes the compiler to gcc and g++.

This will work on Linux/Unix and the Mac. It should automatically detect if you have the Intel compilers or not, and whether or not you have CUDA.

See SuiteSparse_config/cmake_modules/SuiteSparsePolicy.cmake to select your BLAS.

You may also need to add SuiteSparse/lib to your path. If your copy of SuiteSparse is in /home/me/SuiteSparse, for example, then add this to your ~/.bashrc file:

LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/me/SuiteSparse/lib
export LD_LIBRARY_PATH

For the Mac, use this instead:

DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:/home/me/SuiteSparse/lib
export DYLD_LIBRARY_PATH

Default install location of files is below, where PACKAGE is one of the packages in SuiteSparse:

* `CMAKE_INSTALL_PREFIX/include/suitesparse/`: include files
* `CMAKE_INSTALL_PREFIX/lib/`: compiled libraries
* `CMAKE_INSTALL_PREFIX/lib/cmake/SuiteSparse/`: `*.cmake` scripts
    for all of SuiteSparse
* `CMAKE_INSTALL_PREFIX/lib/cmake/PACKAGE/`: `*Config.cmake` scripts for a
    specific package
* `CMAKE_INSTALL_PREFIX/lib/pkgconfig/PACKAGE.pc`: `.pc` scripts for
    a specific package pkgconfig

QUICK START FOR MATLAB USERS (Linux or Mac):

Suppose you place SuiteSparse in the /home/me/SuiteSparse folder.

Add the SuiteSparse/lib folder to your run-time library path. On Linux, add this to your ~/.bashrc script, assuming /home/me/SuiteSparse is the location of your copy of SuiteSparse:

    LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/me/SuiteSparse/lib
    export LD_LIBRARY_PATH

For the Mac, use this instead, in your ~/.zshrc script, assuming you place SuiteSparse in /Users/me/SuiteSparse:

    DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:/Users/me/SuiteSparse/lib
    export DYLD_LIBRARY_PATH

Compile all of SuiteSparse with make local.

Next, compile the GraphBLAS MATLAB library. In the system shell while in the SuiteSparse folder, type make gbmatlab if you want to install it system-wide with make install, or make gblocal if you want to use the library in your own SuiteSparse/lib.

Then in the MATLAB Command Window, cd to the SuiteSparse directory and type SuiteSparse_install. All packages will be compiled, and several demos will be run. To run a (long!) exhaustive test, do SuiteSparse_test.

Save your MATLAB path for future sessions with the MATLAB pathtool or savepath commands. If those methods fail because you don't have system-wide permission, add the new paths to your startup.m file, normally in Documents/MATLAB/startup.m. You can also use the SuiteSparse_paths m-file to set all your paths at the start of each MATLAB session.

Compilation options

You can set specific options for CMake with the command (for example):

    cmake -DCHOLMOD_PARTITION=OFF -DBUILD_STATIC_LIBS=OFF -DCMAKE_BUILD_TYPE=Debug ..

That command will compile all of SuiteSparse except for CHOLMOD/Partition Module (because of -DCHOLMOD_PARTITION=OFF). Debug mode will be used (the build type). The static libraries will not be built (since -DBUILD_STATIC_LIBS=OFF is set).

  • SUITESPARSE_ENABLE_PROJECTS:

    Semicolon separated list of projects to be built or all. Default: all in which case the following projects are built:

    suitesparse_config;mongoose;amd;btf;camd;ccolamd;colamd;cholmod;cxsparse;ldl;klu;umfpack;paru;rbio;spqr;spex;graphblas;lagraph

    Additionally, csparse can be included in that list to build CSparse.

  • CMAKE_BUILD_TYPE:

    Default: Release, use Debug for debugging.

  • SUITESPARSE_USE_STRICT:

    SuiteSparse has many user-definable settings of the form SUITESPARSE_USE_* or (package)_USE_* for some particular package. In general, these settings are not strict. For example, if SUITESPARSE_USE_OPENMP is ON then OpenMP is preferred, but SuiteSparse can be used without OpenMP so no error is generated if OpenMP is not found. However, if SUITESPARSE_USE_STRICT is ON then all *_USE_* settings are treated strictly and an error occurs if any are set to ON but the corresponding package or setting is not available. The *_USE_SYSTEM_* settings are always treated as strict. Default: OFF.

  • SUITESPARSE_USE_CUDA:

    If set to ON, CUDA is enabled for all of SuiteSparse. Default: ON,

    CUDA on Windows with MSVC appears to be working with this release, but it should be considered as a prototype and may not be fully functional. I have limited resources for testing CUDA on Windows. If you encounter issues, disable CUDA and post this as an issue on GitHub.

  • CHOLMOD_USE_CUDA:

    Default: ON. Both SUITESPARSE_USE_CUDA and CHOLMOD_USE_CUDA must be enabled to use CUDA in CHOLMOD.

  • SPQR_USE_CUDA:

    Default: ON. Both SUITESPARSE_USE_CUDA and SPQR_USE_CUDA must be enabled to use CUDA in SPQR.

  • CMAKE_INSTALL_PREFIX:

    Defines the install location (default on Linux is /usr/local). For example, this command while in a folder build in the top level SuiteSparse folder will set the install directory to /stuff, used by the subsequent sudo cmake --install .:

    cmake -DCMAKE_INSTALL_PREFIX=/stuff ..
    sudo cmake --install .

  • SUITESPARSE_PKGFILEDIR:

    Directory where CMake Config and pkg-config files will be installed. By default, CMake Config files will be installed in the subfolder cmake of the directory where the (static) libraries will be installed (e.g., lib). The .pc files for pkg-config will be installed in the subfolder pkgconfig of the directory where the (static) libraries will be installed.

    This option allows to install them at a location different from the (static) libraries. This allows to install multiple configurations of the SuiteSparse libraries at the same time (e.g., by also setting a different CMAKE_RELEASE_POSTFIX and CMAKE_INSTALL_LIBDIR for each of them). To pick up the respective configuration in downstream projects, set, e.g., CMAKE_PREFIX_PATH (for CMake) or PKG_CONFIG_PATH (for build systems using pkg-config) to the path containing the respective CMake Config files or pkg-config files.

  • SUITESPARSE_INCLUDEDIR_POSTFIX:

    Postfix for installation target of header from SuiteSparse. Default: suitesparse, so the default include directory is: CMAKE_INSTALL_PREFIX/include/suitesparse

  • BUILD_SHARED_LIBS:

    If ON, shared libraries are built. Default: ON.

  • BUILD_STATIC_LIBS:

    If ON, static libraries are built. Default: ON, except for GraphBLAS, which takes a long time to compile so the default for GraphBLAS is OFF unless BUILD_SHARED_LIBS is OFF.

  • SUITESPARSE_CUDA_ARCHITECTURES:

    A string, such as "all" or "35;50;75;80" that lists the CUDA architectures to use when compiling CUDA kernels with nvcc. The "all" option requires CMake 3.23 or later. Default: "52;75;80".

  • BLA_VENDOR:

    A string. Leave unset, or use "ANY" to select any BLAS library (the default). Or set to the name of a BLA_VENDOR defined by FindBLAS.cmake. See: https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors

  • SUITESPARSE_USE_64BIT_BLAS:

    If ON, look for a 64-bit BLAS. If OFF: 32-bit only. Default: OFF.

  • SUITESPARSE_USE_PYTHON:

    If ON, build Python interfaces for any package in SuiteSparse that has one (currently only SPEX). If OFF: do not build any Python interface. Default: ON.

  • SUITESPARSE_USE_OPENMP:

    If ON, OpenMP is used by default if it is available. Default: ON.

    GraphBLAS, LAGraph, and ParU will be vastly slower if OpenMP is not used. CHOLMOD will be somewhat slower without OpenMP (as long as it still has a parallel BLAS/LAPACK). Three packages (UMFPACK, CHOLMOD, and SPQR) rely heavily on parallel BLAS/LAPACK libraries and those libraries may use OpenMP internally. If you wish to disable OpenMP in an entire application, select a single-threaded BLAS/LAPACK, or a parallel BLAS/LAPACK that does not use OpenMP (such as the Apple Accelerate Framework). Using a single-threaded BLAS/LAPACK library will cause UMFPACK, CHOLMOD, and SPQR to be vastly slower.

    WARNING: GraphBLAS may not be thread-safe if built without OpenMP or pthreads (see the GraphBLAS User Guide for details).

  • SUITESPARSE_CONFIG_USE_OPENMP:

    If ON, SuiteSparse_config uses OpenMP if it is available. Default: SUITESPARSE_USE_OPENMP. It is not essential and only used to let SuiteSparse_time call omp_get_wtime.

  • CHOLMOD_USE_OPENMP:

    If ON, OpenMP is used in CHOLMOD if it is available. Default: SUITESPARSE_USE_OPENMP.

  • GRAPHBLAS_USE_OPENMP:

    If ON, OpenMP is used in GraphBLAS if it is available. Default: SUITESPARSE_USE_OPENMP.

  • LAGRAPH_USE_OPENMP:

    If ON, OpenMP is used in LAGraph if it is available. Default: SUITESPARSE_USE_OPENMP.

  • PARU_USE_OPENMP:

    If ON, OpenMP is used in ParU if it is available. Default: SUITESPARSE_USE_OPENMP.

  • SPEX_USE_OPENMP:

    If ON, OpenMP is used in SPEX if it is available. Default: SUITESPARSE_USE_OPENMP.

  • SUITESPARSE_DEMOS:

    If ON, build the demo programs for each package. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_BTF:

    If ON, use BTF libraries installed on the build system. If OFF, automatically build BTF as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_CHOLMOD:

    If ON, use CHOLMOD libraries installed on the build system. If OFF, automatically build CHOLMOD as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_AMD:

    If ON, use AMD libraries installed on the build system. If OFF, automatically build AMD as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_COLAMD:

    If ON, use COLAMD libraries installed on the build system. If OFF, automatically build COLAMD as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_CAMD:

    If ON, use CAMD libraries installed on the build system. If OFF, automatically build CAMD as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_CCOLAMD:

    If ON, use CCOLAMD libraries installed on the build system. If OFF, automatically build CCOLAMD as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_GRAPHBLAS:

    If ON, use GraphBLAS libraries installed on the build system. If OFF, automatically build GraphBLAS as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_SYSTEM_SUITESPARSE_CONFIG:

    If ON, use SuiteSparse_config libraries installed on the build system. If OFF, automatically build SuiteSparse_config as dependency if needed. Default: OFF.

  • SUITESPARSE_USE_FORTRAN

    If ON, use the Fortran compiler to determine how C calls Fortan, and to build several optional Fortran routines. If OFF, use SUITESPARSE_C_TO_FORTRAN to define how C calls Fortran. Default: ON.

  • SUITESPARSE_C_TO_FORTRAN

    A string that defines how C calls Fortran (i.e., functions exported by the BLAS library). This setting is used if no working Fortran compiler could be detected or SUITESPARSE_USE_FORTRAN is set to OFF. This string is to be read as the argument list and the body of a preprocessor macro. The first argument to that macro is any Fortran function name in lowercase letters. The second argument is the same function name in uppercase letters. The body defines by which function name Fortran functions are called. This is necessary because Fortran is case-insensitive, and different Fortran compilers use different name mangling conventions. If a MSVC C/C++ compiler is used, this defaults to "(name,NAME) name" (i.e., lower case without trailing underscore). That is the name mangling convention for the Intel Fortran compiler on Windows. If any other C/C++ compilers are used, this defaults to "(name,NAME) name##_" (i.e., lower case with trailing underscore). That is the name mangling convention for most of the commonly used Fortran compilers (like ifx on platforms other than Windows, gfortran, flang, ...). The latter name mangling convention is also used by default by OpenBLAS (independent on the platform or the compiler used to build OpenBLAS). You might need to configure with -DSUITESPARSE_C_TO_FORTRAN="(name,NAME) name##_" if you'd like to build SuiteSparse using a MSVC compiler and link to OpenBLAS.

Additional options are available for specific packages:

  • UMFPACK_USE_CHOLMOD:

    If ON, UMFPACK uses CHOLMOD for additional (optional) ordering options. Default: ON.

  • KLU_USE_CHOLMOD:

    If ON, KLU uses CHOLMOD for additional (optional) ordering options. Default: ON.

CHOLMOD is composed of a set of Modules that can be independently selected; all options default to ON:

  • CHOLMOD_GPL

    If OFF, do not build any GPL-licensed module (MatrixOps, Modify, Supernodal, and GPU modules)

  • CHOLMOD_CHECK

    If OFF, do not build the Check module.

  • CHOLMOD_MATRIXOPS

    If OFF, do not build the MatrixOps module.

  • CHOLMOD_CHOLESKY If OFF, do not build the Cholesky module. This also disables the Supernodal and Modify modules.

  • CHOLMOD_MODIFY

    If OFF, do not build the Modify module.

  • CHOLMOD_CAMD

    If OFF, do not link against CAMD and CCOLAMD. This also disables the Partition module.

  • CHOLMOD_PARTITION

    If OFF, do not build the Partition module.

  • CHOLMOD_SUPERNODAL

    If OFF, do not build the Supernodal module.

Possible build/install issues

One common issue can affect all packages: getting the right #include files that match the current libraries being built. It's possible that your Linux distro has an older copy of SuiteSparse headers in /usr/include or /usr/local/include, or that Homebrew has installed its suite-sparse bundle into /opt/homebrew/include or other places. Old libraries can appear in in /usr/local/lib, /usr/lib, etc. When building a new copy of SuiteSparse, the cmake build system is normally (or always?) able to avoid these, and use the right header for the right version of each library.

As an additional guard against this possible error, each time one SuiteSparse package #include's a header from another one, it checks the version number in the header file, and reports an #error to the compiler if a stale version is detected. In addition, the Example package checks both the header version and the library version (by calling a function in each library). If the versions mismatch in any way, the Example package reports an error at run time.

For example, CHOLMOD 5.1.0 requires AMD 3.3.0 or later. If it detects an older one in amd.h, it will report an #error:

    #include "amd.h"
    #if ( ... AMD version is stale ... )
    #error "CHOLMOD 5.1.0 requires AMD 3.3.0 or later"
    #endif

and the compilation will fail. The Example package makes another check, by calling amd_version and comparing it with the versions from the amd.h header file.

If this error or one like it occurs, check to see if you have an old copy of SuiteSparse, and uninstall it before compiling your new copy of SuiteSparse.

There are other many possible build/install issues that are covered by the corresponding user guides for each package, such as finding the right BLAS, OpenMP, and other libraries, and how to compile on the Mac when using GraphBLAS inside MATLAB, and so on. Refer to the User Guides for more details.

Interfaces to SuiteSparse

MATLAB/Octave/R/Mathematica interfaces:

Many built-in methods in MATLAB and Octave rely on SuiteSparse, including C=A*B x=A\b, L=chol(A), [L,U,P,Q]=lu(A), R=qr(A), dmperm(A), p=amd(A), p=colamd(A), ... See also Mathematica, R, and many many more. The list is too long.

Julia interface:

https://github.com/JuliaSparse/SparseArrays.jl

Python interface to GraphBLAS by Anaconda and NVIDIA:

https://pypi.org/project/python-graphblas

Intel's Go interface to GraphBLAS:

https://pkg.go.dev/github.com/intel/forGraphBLASGo

See scikit-sparse and scikit-umfpack for the Python interface via SciPy:

https://github.com/scikit-sparse/scikit-sparse https://github.com/scikit-umfpack/scikit-umfpack

See math.js by Jos de Jong for a JavaScript port of CSparse:

https://github.com/josdejong/mathjs

See russell for a Rust interface:

https://github.com/cpmech/russell

Acknowledgements

Markus Mützel contributed the most recent update of the SuiteSparse build system for all SuiteSparse packages, extensively porting it and modernizing it.

I would also like to thank François Bissey, Sebastien Villemot, Erik Welch, Jim Kitchen, and Fabian Wein for their valuable feedback on the SuiteSparse build system and how it works with various Linux / Python distros and other package managers. If you are a maintainer of a SuiteSparse packaging for a Linux distro, conda-forge, R, spack, brew, vcpkg, etc, please feel free to contact me if there's anything I can do to make your life easier. I would also like to thank Raye Kimmerer for adding support for 32-bit row/column indices in SPQR v4.2.0.

See also the various Acknowledgements within each package.