jlblancoc/suitesparse-metis-for-windows

CMake scripts for painless usage of Tim Davis' SuiteSparse (CHOLMOD,UMFPACK,AMD,LDL,SPQR,...) and METIS from Visual Studio and the rest of Windows/Linux/OSX IDEs supported by CMake. The project includes precompiled BLAS/LAPACK DLLs for easy use with Visual C++. Licensed under BSD 3-Clause License.

The goal is using one single CMake code to build against SuiteSparse in standard Linux package systems (e.g. libsuitesparse-dev) and in manual compilations under Windows.

Credits: Jose Luis Blanco (Universidad de Almeria); Jerome Esnault (INRIA); @NeroBurner

1. Instructions

• (1) Install CMake.

• (2) Only for Linux/Mac: Install LAPACK & BLAS. In Debian/Ubuntu: sudo apt-get install liblapack-dev libblas-dev

• (3) Clone or download this project (latest release) and extract it somewhere in your disk, say SP_ROOT.

  • (OPTIONAL) CMake will download SuiteSparse sources automatically for you (skip to step 4), but you may do it manually if preferred:
    • Populate the directories within SP_ROOT with the original sources from each project:
      • SuiteSparse:
        • Download SuiteSparse-X.Y.Z.tar.gz from Tim Davis' webpage.
        • Extract it.
        • Merge (e.g. copy and paste from Windows Explorer) the tree SuiteSparse/* into SP_ROOT/SuiteSparse/*.
        • Make sure of looking for patches in the original webpage and apply them to prevent build errors.
      • METIS: (Optional, only if need METIS for partitioning)
        • Download metis-X.Y.Z.tar.gz.
        • Extract it.
        • Merge the tree metis-X.Y.Z/* into SP_ROOT/metis/*.
        • Add the command cmake_policy(SET CMP0022 NEW) right after the line project(METIS) in metis/CMakeLists.txt.

• (4) Run CMake (cmake-gui), then:

  • Set the "Source code" directory to SP_ROOT
  • Set the "Build" directory to any empty directory, typically SP_ROOT/build
  • Press "Configure", change anything (if needed)
  • Important: I recommend setting the CMAKE_INSTALL_PREFIX to some other directory different than "Program Files" or "/usr/local" so the INSTALL command does not require Admin privileges. By default it will point to SP_ROOT/build/install.
  • If you have an error like: "Cannot find source file: GKlib/conf/check_thread_storage.c", then manually adjust GKLIB_PATH to the correct path SP_ROOT/metis/GKlib.
  • HAVE_COMPLEX is OFF by default to avoid errors related to complex numbers in some compilers.
  • Press "Generate".

• (5) Compile and install:

  • In Visual Studio, open SuiteSparseProject.sln and build the INSTALL project in Debug and Release. You may get hundreds of warnings, but it's ok.
  • In Unix: Just execute make install or sudo make install if you did set the install prefix to /usr/*

• (6) Notice that a file SuiteSparseConfig.cmake should be located in your install directory. It will be required for your programs to correctly build and link against SuiteSparse.

• (7) Only for Windows: You will have to append CMAKE_INSTALL_PREFIX\lib*\lapack_blas_windows\ and CMAKE_INSTALL_PREFIX\lib* to the environment variable PATH before executing any program, for Windows to localize the required BLAS/Fortran libraries (.DLLs).

1.2 Using OpenBLAS as alternative to generic BLAS/LAPACK

Starting with v0.3.21 OpenBLAS ships with an f2c-converted copy of LAPACK v3.9.0 which is used if no fortran compiler is available. This means we can compile OpenBLAS to provide BLAS and LAPACK functionality with a regular C++ compiler like MSVC or GCC. This includes compiling everything as a static library. If that library is later used to compile an executable the executable can run without any external dependencies (like fortran runtimes or lapack dlls in the generic fortran based BLAS and LAPACK implementation).

When building OpenBLAS yourself be sure to use BUILD_WITHOUT_LAPACK=NO and NOFORTAN=1 when configuring OpenBLAS with CMake to get a pure C++ OpenBLAS library (with the benefits described above).

To tell SuiteSparse to build against this OpenBLAS implementation set the option WITH_OPENBLAS=ON.

2. Test program

Example CMake programs are provided for testing, based on Tim Davis' code in his manual:

• example-projects

An example to test CUDA support can be found here.

3. Integration in your code (unique code for Windows/Linux)

• Add a block like this to your CMake code (see complete example), and set the CMake variable SuiteSparse_DIR to SP_INSTALL_DIR/lib/cmake/suitesparse-5.4.0/ or the equivalent place where SuiteSparse-config.cmake was installed.

  # ------------------------------------------------------------------
  # Detect SuiteSparse libraries:
  # If not found automatically, set SuiteSparse_DIR in CMake to the
  # directory where SuiteSparse-config.cmake was installed.
  # ------------------------------------------------------------------
  find_package(SuiteSparse CONFIG REQUIRED)
  target_link_libraries(MY_PROGRAM PRIVATE ${SuiteSparse_LIBRARIES})
  
  # or directly add only the libs you need:
  target_link_libraries(MY_PROGRAM PRIVATE
    	SuiteSparse::suitesparseconfig
    	SuiteSparse::amd
    	SuiteSparse::btf
    	SuiteSparse::camd
    	SuiteSparse::ccolamd
    	SuiteSparse::colamd
    	SuiteSparse::cholmod
    	SuiteSparse::cxsparse
    	SuiteSparse::klu
    	SuiteSparse::ldl
    	SuiteSparse::umfpack
    	SuiteSparse::spqr
    	SuiteSparse::metis
    )
  

• In Windows, if you have a CMake error like:

  Could not find a package configuration file provided by "LAPACK" with any
   of the following names:
  
     LAPACKConfig.cmake
     lapack-config.cmake
  

  set the CMake variable LAPACK_DIR to SP_ROOT/lapack_windows/x64/ (or x32 for 32bit builds).

4. Why did you create this project?

Porting SuiteSparse to CMake wasn't trivial because this package makes extensive usage of a one-source-multiple-objects philosophy by compiling the same sources with different compiler flags, and this ain't directly possible to CMake's design.

My workaround to this limitation includes auxiliary Python scripts and dummy source files, so the whole thing became large enough to be worth publishing online so many others may benefit.