Installation notes for ABINIT¶
This page provides an introduction to the installation of the ABINIT package and the compilation of the executables. It also discusses how to test whether the compilation was successful by running the internal test suite. Finally, it gives related complements for the developers.
Any comment or suggestion to improve the procedure or this page is welcome! Simply contact the ABINIT group on the discourse forum.
For the vast majority of people willing to use ABINIT (simple users -not developers-, with Unix/Linux or MacOS working in the terminal), the installation/compilation steps are:
- Prerequisite: you need a Fortran compiler, a C compiler, the Python interpreter (>= 2.7.5), some mandatory libraries (BLAS/LAPACK, FFT, NetCDF4, HDF5 and LibXC), possibly some recommended libraries (MPI) and other optional libraries such as Wannier90. The libraries can be installed with the help of the “fallback” procedure, see below for more info on this step. Alternatively, you may want to install everything from source using the procedure detailed in this tutorial.
- Get the latest version of the ABINIT package (abinit-x.y.z.tar.gz) from the abinit web site. More information are available here.
- Prepare a configuration file named “hostname”.ac9, that contains the information about libraries and compilation options. This step is not mandatory as the ABINIT configure script will try to detect the installation location if options are not provided but the automatic detection procedure can fail. See below for more information, as well as this section.
./configure(or, even better, first create a tmp directory for the build then
cd tmpand finally run
../configurewithin the build directory). For further details, consult this link.
make -jNfor compiling with N processors, e.g.
make -j4to use four processors). This step might take dozen of minutes depending on the compilation options. More information are available here.
- Optionally, issue
make installto install the package (root privileges are needed if the installation direction is not specified via
Note that the details of step 1 and 3 might vary significantly depending on the operating system. Further details are provided by this tutorial that cover the scenario in which you want to build everything from source and install libraries in your $HOME directory. There are also other pages focusing on macOS, CentOS and Ubuntu that discuss how to bypass the compilation of the external libraries using homebrew or MacPorts (for macOS), dnf (for CentOS) or apt (for Ubuntu).
Examples of configuration files to compile Abinit on clusters are available in the abiconfig package on github (specifically the directory for ABINITv9), while the configuration files used for our buildbot testfarm are available in the autoconf_examples section. The ABINIT Wiki also has a build abinit section, that might be useful. In particular, the current documentation for the fallback procedure is available here while configuration options are documented in this page.
If you want to have a much better handling on the ABINIT source code than normal users, or if you downloaded ABINIT from Gitlab or GitHub anyhow, then consult the section For developers.
How to get a version of ABINIT?¶
We assume that you have a F90 compiler under UNIX/Linux or macOS X and that you want to compile the source files, and, perhaps, modify and/or add new files. This is the typical scenario for most users and system administrators. In what follows, x.y.z represents the ABINIT version.
In order to get the ABINIT package, you have first to download the file abinit-x.y.z.tar.gz from the packages webpage of the ABINIT Web site, then issue:
gunzip abinit-x.y.z.tar.gz | tar -xvf -
Alternatively, you may use
wget https://www.abinit.org/sites/default/files/packages/abinit-x.y.z.tar.gz tar xzf abinit-x.y.z.tar.gz
If correctly done, a main directory, denoted ~abinit in the present document (usually, its real name will be abinit-x.y.z) and a whole set of subdirectories should have been created, including:
- the source files inside the src directory;
- a tarball with the fallbacks and a script to create libraries in the directory fallbacks;
- the documentation in the doc directory;
- the complete set of tests and the pseudopotentials needed for the automatic tests in the directory tests;
- all the scripts and configuration options needed to generate configure and Makefiles in the config directory.
Obviously, the package does not contain object files and binary executables that will be built by make. Note also that pseudopotentials for production runs are not shipped with the package. The tarball contains several dozen pseudopotentials for testing purposes in ~abinit/tests/Psps_for_tests. The largest set of pseudopotentials and PAW atomic datafiles for production runs can be found on the ABINIT Atomic data pseudopotentials and PAW datasets Web page. Other pseudopotentials have been generated by many different users, and might be shared but you might have to contact them.
The web site https://www.abinit.org contains many other resources including links to the forum, the mailing list, the ABINIT events etc and, last but not least, tutorials and documentation pages.
How to compile the executables?¶
We now suppose that you have a F90 compiler and that you want to compile the source files.
In most cases, you will have to provide to the configure utility some information: the location of the F90 compiler (and sometimes even the C compiler), the adequate compiler options, and, if you want to produce the parallel binaries, the location of the MPI library.
Although the present building tools should be powerful enough to succeed to build the binaries without you giving such information, on a significant number of platforms it has been observed that it is still better to specify options explicitly in order to avoid suboptimal executables or downgraded capabilities.
Supposing that you are in the lucky case in which the build system is able detects all the requirements, then building ABINIT is rather simple. Issue:
./configure(or first create a tmp directory, then
cd tmp, then
make -jNfor compiling with N processors on a SMP machine, e.g.
make -j4for four processors). Might take two minutes.
Well, it might also be that only one additional information is needed, in which case something like:
configure FC=gfortran make
might work. In both cases, let’s explain a bit what is done, and the further possibilities.
The configure step produces the set of Makefile files (among other things), taking into account information about your machine and the “hostname”.ac9 file. It takes three minute or less. The make step compiles everything, according to the Makefile files produced in the previous step. The executables will be located in the subdirectory ~abinit/src/98_main, if you have chosen to issue ./configure in the ~abinit directory. If you have issued ./configure in another directory, it will be placed accordingly.
The time required to build everything is highly dependent on the compiler and platform. On a 2.8 GHz quad-proc machine (using make -j4), the whole compilation is about 10 minutes. On other platforms, with only one processor, it might be more than one hour.
The make command accepts the name of the target(s) as optional argument. To compile only one of the executable after the configure step, issue
where name_of_the_binary can be abinit, cut3d, anaddb, etc.
will install the package in the /usr/local directory.
How to write the “hostname”.ac9 file?¶
Let’s come back to the case where the build system needs some more
information. This information should be stored in a file named “hostname”.ac9,
where “hostname” is the result of executing the command
Note that the command
hostname by default returns the fully qualified domain name (FQDN),
while only the first word of the returned chain of character is needed, e.g. abiref.
This is the reasone why we had to use
There is a template for such “hostname”.ac9 file, located in ~abinit/doc/config. Its name is config-template.ac9. Examples of such files, that are used for testing the package on our testfarm, can be found in ~abinit/doc/build/config-examples, or equivalently in the autoconf_examples section. Additional examples of configuration files for clusters are provided by the abiconfig project and are available here.
Most of the examples provided in the ~abinit/doc/build/config-examples directory contain five important variables: location of the F90/C compilers, F90 and C compilation options, location of MPI library (if enabled). On the other hand, there are many other additional control flags (“with_XYZ”), needed for advanced use.
Your hostname.ac9 file might be placed in your home directory inside a new directory named ~/.abinit/build. If you opt for this solution, every time you install a new version of ABINIT, the configuration file will be automatically used by the configure script so you do not have to care anymore about this file after the first installation.
On the other hand, if you want to play with several configurations, you can place the hostname.ac9 file in the ~abinit directory, where such a hostname.ac9 file will be also seen by the build system (and preferred over the one located in ~/.abinit/build) or inside your build directory (like ~abinit/tmp). As mentioned above, you might even input the options contained in the hostname.ac9 file directly on the command line.
Note the order of precedence for the location of the hostname.ac9 file (or command-line interface), in case more than one possibility is used (decreasing order of precedence):
- Command line (overcome all other information)
- Your build directory (~abinit/tmp)
- The ABINIT top source directory (~abinit)
When the hostname.ac9 file is ready, you can come back to the configure/make sequence.
How to run the internal tests¶
The abinit code has several small internal tests (three basic ones, called fast, v1 and v5, and then one for each of the libraries bigdft, etsf_io, libxc, wannier90), that can be issued automatically, and that check automatically whether the results are correct. These tests are available whether you have got the package from the Web or from the ABINIT archive. Of course, you need to have compiled abinit in order to run the internal tests. Moreover, the simple implementation procedure assumes that the executable is located in ~abinit/src/98_main (the standard location after issuing make).
You can begin with the fast suite. Simply issue the command:
It will run for a few seconds. It should print:
Status file, reporting on built-in test fast ==> The run finished cleanly. Moreover, comparison of the total energy, and other (few) relevant quantities with reference values has been successful. This does not mean that no problem is present, however. Please run the complete set of ABINIT tests to gain a better confidence in your installation.
This means that the internal fast suite ran successfully. If you do not get this message, then the executables were not properly generated, or there is a problem with the makefile that drives the internal test. In this case, after having tried to solve the problem by yourself, you should contact somebody in the ABINIT group.
In addition to this small message, you can have access to all generated files, that are located inside the tests/built-in/Input subdirectory.
Supposing the fast tests are OK, then you might issue the command:
The fast tests will be executed once more, followed by the other internal tests. Again, we hope that you will get the positive diagnostics for the other tests. Of course, the bigdft, etsf_io, libxc, and wannier90 needs the appropriate library to be installed in order to work properly.
For further information on these internal tests, see the ~abinit/tests/built-in/README file.
You might now read the new user’s guide, in order to learn how to use the code, and then follow the four basic tutorials, see the entry page for the tutorials. This is useful if you consider that the installation has been successful. Or you might continue to read the present Web page, and try to perform the speed tests, as well as the other tests.
How to execute the automatic tests¶
The workhorse script to run the test suite is called runtests.py. It is very flexible. A reasonable set of tests (those contained in the fast and v”x” directories), can be executed automatically by issuing inside the ~abinit/tests directory the command:
to run the test suite in parallel with 4 cores.
This is the recommended procedure for developers. In order to execute these tests, you will need a larger disk space than for the simple installation of the code (the total additional disk space required is on the order of 1GB).
The command line options of runtests.py are numerous. Please issue
in order to access the documentation with the examples or consult the video below.
Let us now examine the different subdirectories of the ~abinit/tests directory.
~ABINIT/tests/fast is the simplest, and its content will be described in some detail below. For tests of the parallel version see the directory tests/paral. For tests of the response function features of abinit, and for tests of mrgddb and anaddb, see the subdirectories tests/v2. The other directories tests/v3, tests/v4, etc present further tests of recently implemented features of ABINIT.
1) tests/fast (for the sequential version only)
This subdirectory contains a basic set of tests of the code, aimed at testing whether the code is coherent in time (successive versions), and exercising several parts of the code. However, they do not examine its accuracy on physical problems, mainly because the number of plane waves is too small, and some tests are not run to self-consistent convergence. 32 MB of memory should be enough for these tests.
The input files for each of the tests can be found in the ~abinit/tests/fast/Input directory. At the bottom of each input file, there is a section with metadata and parameters defining the test. Such metadata mentions the executable to be used, the output files to be analyzed, the admitted tolerances with respect to reference output files, the author of the test and a brief description of the test.
To run only the tests in this directory, simply issue
The script will create a directory named Test_suite. All the results will be stored in that directory. The output files are be automatically compared to a set of reference files (in ~abinit/tests/fast/Refs) thanks to the GNU ‘diff’ utility. The corresponding difference files are prefixed by ‘diff.’.
In addition to ‘diff’, there are two other levels of automatic analysis: one based on a python diff tool called ‘fldiff’, producing ‘fldiff.report’ files, and another where the output of ‘fldiff’ is further analyzed to produce a brief report called ‘report’. The latter step is only performed in case all the tests cases of one directory are performed (including the case where tests of different directories are performed).
The one-line summaries produced by fldiff (see later) are compared with the tolerances indicated in the input file (metadata added at the end of the input file). This procedure produces a file called “report”, in which there is a one line assessment of the behaviour of each test: succeeded (everything is OK), passed (the test is OK for users in production), passed marginally (the test is within 1.5 of the usually accepted deviation, which is likely OK for most applications - still to be improved by the development team, though), failed (there is a problem, the deviation is usually not accepted). This is by far the most convenient tool to analyze the automatic tests of abinit.
The vast majority of tests cases succeed or pass on all platforms that are used by the developer team in Louvain-la-neuve. Some problems are mentioned in the file ~abinit/KNOWN_PROBLEMS. Additionally, there might be specific problems for some test case for some platforms, also mentioned in ~abinit/KNOWN_PROBLEMS. So, apart from the known problems, mentioned in this file, the “report” file should mention, for each test case, only “succeeded” or “passed”.
The comparing tool ‘fldiff’ -for ‘floating diff’- performs in a more detailed way the comparison of floating numbers between the output files and the reference files than in the case of a ‘diff’ command. As used presently, for each run inside one directory, one single file, called ‘fldiff.report’, will be produced, and gather the analysis for all tests in that directory.
If for one test case, the two files differ by the number of lines, the ‘fldiff.report’ file will report that it cannot compare the two files. Usually this problem will be seen at the level of ‘command signs’ appearing sometimes in the first column of the output files, so a typical error message (announcing something went wrong) will be:
Case_1 22 The diff analysis cannot be pursued: the command sign differ.
By contrast, it will identify the floating numbers and ignore their differences if they are within some prescribed tolerance, or if the difference is not relevant. For example, it is able to ignore the differences in timings. If everything goes fine for a test, fldiff should identify only the differences in:
- the dates of execution (possibly);
- the version numbers (possibly);
- the platform description (possibly);
- the overall execution time (this is ALWAYS printed, even without differences).
So, a successful execution of one test case may be announced as follows in the fldiff.report file:
Case_1 2 < Version 8.0.6 of ABINIT > Version 8.0.3 of ABINIT 5 < Starting date: Mon 23 May 2016. > Starting date: Mon 4 Apr 2016 202 < +overall time at end (sec): cpu= 7.1 wall= 8.0 > +Overall time at end (sec): cpu= 7.3 wall= 8.0 Summary Case_1: no significant difference has been found
The fldiff.report file will have one such section for each test_case that was run. It begins with the number of the test case, then includes a few blocks of three lines: the number of the line where something is happening, followed by the content of the two lines. Finally, there is a one-line summary for each test case.
More information on the fldiff script can be found in the ~abinit/tests/Scripts/fldiff.pl file.
This directory contains tests built in the same spirit as those in the test/fast directory, but that exercise other basic features, like the treatment of metals, the GGA, the new pseudopotentials, the multi-dataset mode, the cell parameters optimization, and the spatial symmetry groups database. These were developed during the development time of the version 1 of ABINIT. Of course, the automatic difference procedure only compares to recent runs of the ABINIT code. As for the ‘fast’ test cases, the fldiff.report and report files are also available. 64 MB of memory should be enough for these tests.
This directory contains tests built in the same spirit as those in the tests/fast/ or tests/v1 directory, but that exercise features not present in version 1 of the ABINIT package, mainly the response function features, and the use of the mrgddb and anaddb codes. Again, the automatic difference procedure only compares to recent runs of the ABINIT code. As for the ‘fast’ test cases, the fldiff.report and report files are also available. 64 MB of memory should be enough for these tests.
4) tests/v3, tests/v4, tests/v5, tests/v6, tests/v67mbpt, tests/v7, tests/v8, tests/v9, tests/bigdft, tests/etsf_io, tests/libxc, tests/wannier90
These directories contain tests built in the same spirit as those in the tests/fast/, tests/v1, tests/v2 directory, but that exercise features not present in the version 1 or 2 of the ABINIT package, noticeably the use of the GW code, the utilities Cut3d, AIM, .., the PAW … . Or the interfacing with fallbacks. Again, the automatic difference procedure only compares to recent runs of the ABINIT code. Like for the ‘fast’ test cases, the fldiff.report and report files are also available. 64 MB of memory should be enough for these tests.
5) tests/paral/ and tests/mpiio/ (need MPI support)
This directory contains tests built in the same spirit as those in the test/fast/ directory, but that exercise the parallel version of the ABINIT code.
The script runtests.py considers one of the different input files, and for this file, it will use the parallel code with one processing node, then perform different parallel runs with an increasing number of processing nodes, as specified in the metadata contained in the input file. As for the other series of test, the diff and the fldiff utilities are used automatically, and fldiff.report and report files are produced automatically.
Note the documentation available in the pages labelled Developers, as well as the developer’s corner of the Wiki The following sections are complements for the installation from gitlab, and the generation of the ABINIT distribution.
Git, autotools and makemake¶
If you want to have a full handle on the package (compilation, modification of files, writing of scripts, you need additional prerequisites, the (free) software applications git, automake, autoconf, libtools.
More explicitly, you need minimally (version numbers can be upgraded)
If you do not have these tools, please consult your local computer guru, and/or the following pages:
If you want to develop on a regular basis, please have a Git(lab) access created for you by contacting Jean-Michel Beuken, as described in these pages. If it is only very occasional, you might as well rely on the ABINIT Github Web site.
It is strongly advised to subscribe to the ABINIT discourse forum to receive the latest information concerning new developments.
After having installed git, and obtained a gitlab branch on the ABINIT internal server, create an autonomous copy of the source code, on top of which you have to make your development. This is explained in the ABINIT wiki gitlab: ABINIT specificities
For your gitlab branches on the internal server, you will have the permission not only to clone/fetch/pull, but also to commit/push your modifications. You might alternatively download other branches of the archives, but you will not be able to commit to these branches. So, do not start to modify these, you will not be able to include them afterwards in the archive.
git clone creates a local archive for your daily work, this
archive being linked to the main ABINIT archive. This very efficient technique
is recommended as it makes you more independent for the management of your
work (you will be able to create new branches). One big advantage of this
approach is that people working with a laptop can develop and commit safely
without a network connection.
Then, make sure you have the ABINIT prerequisite (compilers, libraries etc), and that you have set up an ac9 file.
At this stage, before being able to compile, cd to the newly created abinit directory, and issue:
This command initializes a whole set of files and scripts, needed for the autotools, as well as for the global work on ABINIT sources. This initialization might take up to two minutes. After this initialisation, you can proceed with the configure/make procedure as described in section 2.
How to generate the source package¶
If you want to produce the source package abinit-x.y.z.tar.gz, after having compiled and tested ABINIT, issue: