Install

SciFortran is available in the form of a static Fortran library libscifor.a and the related Fortran module SCIFOR. Our build system relies on CMake. Experimental support for linking Intel’s MKL is provided, although it may fail on some systems.

Building SciFortran

Clone the repo:

git clone https://github.com/SciFortran/SciFortran scifor

Optionally [1] define the fortran compiler:

export FC=mpif90/gfortran/ifort

From the repository directory (cd scifor) make a standard out-of-source CMake compilation:

GNU Make

Using GNU make is the default CMake workflow, with widest version support (CMake > 3.0). Note that parallel make execution is tested and working.

mkdir build
cd build
cmake ..
make -j

Ninja

Using ninja if a fortran-capable [2] version of ninja is available in your system (and CMake can [3] take advantage of it), you can use it to build the library at lightning, multi-threaded, speed.

mkdir build
cd build
cmake -GNinja ..
ninja

The CMake compilation can be customized using the following additional variables (default values between < >, optional in [ ]):

-DPREFIX = prefix directory   #<~/opt/scifor/VERSION/PLATFORM/[GIT_BRANCH]>

-DUSE_MPI = <yes>/no        #set use of MPI

-DVERBOSE = yes/<no>

-DBUILD_TYPE = <RELEASE>/TESTING/DEBUG/AGGRESSIVE  #compilation options. TESTING:mild or no optimization, DEBUG:relevant debugging options, AGGRESSIVE: all debug options of (might not compile on  some systems)

-DWITH_BLAS_LAPACK = yes/<no>   # skip search of preinstalled linear algebra libraries and enforce compilation from local source

-DWITH_SCALAPACK = <yes>/no       #search and link to available ScaLAPACK library

Install SciFortran

System-wide installation is completed after the build step using either:

make install

or

ninja install

To actually link the library to any of your project we provide different solutions:

• A generated environment module, installed to ~/.modules.d/scifor/<PLAT>/<VERSION>

• A generated bash script at <PREFIX>/bin/configvars.sh, to be sourced for permanent loading.

• A generated pkg-config file to, installed to ~/.pkgconfig.d/scifor.pc

which you can choose among by following the instructions printed on screen.

Uninstall

Although CMake does not officially provide uninstall procedures in the generated Make/Ninja files. Hence SciFortran supplies a homebrew method to remove the generated files by calling (from the relevant build folder):

make uninstall

or

ninja uninstall

Known issues

SciFortran has been tested with success on several Unix/Linux platforms. Support for Windows, through WSL, is still experimental, although few people reported successful installation with minimal efforts.

Some have reported issues concerning the wrong setup for the library pkg-config file, contained in $PREFIX/<PLAT>/<VERSION>/etc/scifor.pc. The variable Libs=-L${libdir} -lscifor <blas/lapack/scalapack> produced by cmake during the configuration and installation process can be not properly defined for the part corresponding to third parties libraries such as Blas/Lapack/Scalapack. This breaks compilation against scifor whenever pkg-config is used to generate the linking options.

Tip

Corrupted pkg-config file can be fixed manually. Edit the scifor.pc file overwriting the definition of the variable Libs, as appropriate for your system.

footnotes

[1]

In some cases CMake fails to find the MPI fortran compiler, even if it is effectively installed and loaded into the system. An easy fix is to setup and export the FC=mpif90 environment variable before invoking the cmake <options> .. command.

[2]

Ninja did not support fortran before version 1.10, although Kitware has long mantained a fortran-capable fork, which might be obtained easily as a Spack package. Nevertheless we note that as of fall 2022 pip install ninja –user ships Ninja v1.10.2, hence obtaining a suitable official Ninja release should be trivial.

[3]

This depends on your CMake version. Comparing this to this one would suggest that CMake started supporting Ninja’s fortran features only after v3.17 but we have verified that at least v3.16.3 (current version shipped by apt on Ubuntu 20.04 LTS) does indeed work. For more information you can take a look to the related issue.