Files
lammps/doc/src/Build_basics.rst
Axel Kohlmeyer f2890e39ef update to use new intel fortran compiler name
Co-authored-by: Richard Berger <richard.berger@temple.edu>
2021-03-15 20:53:49 -04:00

553 lines
24 KiB
ReStructuredText

Basic build options
===================
The following topics are covered on this page, for building both with
CMake and make:
* :ref:`Serial vs parallel build <serial>`
* :ref:`Choice of compiler and compile/link options <compile>`
* :ref:`Build the LAMMPS executable and library <exe>`
* :ref:`Including and removing debug support <debug>`
* :ref:`Install LAMMPS after a build <install>`
----------
.. _serial:
Serial vs parallel build
------------------------
LAMMPS is written to use the ubiquitous `MPI (Message Passing Interface)
<https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ library API
for distributed memory parallel computation. You need to have such a
library installed for building and running LAMMPS in parallel using a
domain decomposition parallelization. It is compatible with the MPI
standard version 2.x and later. LAMMPS can also be built into a
"serial" executable for use with a single processor using the bundled
MPI STUBS library.
Independent of the distributed memory MPI parallelization, parts of
LAMMPS are also written with support for shared memory parallelization
using the `OpenMP <https://en.wikipedia.org/wiki/OpenMP>`_ threading
standard. A more detailed discussion of that is below.
.. tabs::
.. tab:: CMake build
.. code-block:: bash
-D BUILD_MPI=value # yes or no, default is yes if CMake finds MPI, else no
-D BUILD_OMP=value # yes or no, default is yes if a compatible compiler is detected
-D LAMMPS_MACHINE=name # name = mpi, serial, mybox, titan, laptop, etc
# no default value
The executable created by CMake (after running make) is named
``lmp`` unless the ``LAMMPS_MACHINE`` option is set. When setting
``LAMMPS_MACHINE=name`` the executable will be called
``lmp_name``. Using ``BUILD_MPI=no`` will enforce building a
serial executable using the MPI STUBS library.
.. tab:: Traditional make
The build with traditional makefiles has to be done inside the source folder ``src``.
.. code-block:: bash
make mpi # parallel build, produces lmp_mpi using Makefile.mpi
make serial # serial build, produces lmp_serial using Makefile/serial
make mybox # uses Makefile.mybox to produce lmp_mybox
Any ``make machine`` command will look up the make settings from a
file ``Makefile.machine`` in the folder ``src/MAKE`` or one of its
sub-directories ``MINE``, ``MACHINES``, or ``OPTIONS``, create a
folder ``Obj_machine`` with all objects and generated files and an
executable called ``lmp_machine``\ . The standard parallel build
with ``make mpi`` assumes a standard MPI installation with MPI
compiler wrappers where all necessary compiler and linker flags to
get access and link with the suitable MPI headers and libraries
are set by the wrapper programs. For other cases or the serial
build, you have to adjust the make file variables ``MPI_INC``,
``MPI_PATH``, ``MPI_LIB`` as well as ``CC`` and ``LINK``\ . To
enable OpenMP threading usually a compiler specific flag needs to
be added to the compile and link commands. For the GNU compilers,
this is ``-fopenmp``\ , which can be added to the ``CC`` and
``LINK`` makefile variables.
For the serial build the following make variables are set (see src/MAKE/Makefile.serial):
.. code-block:: make
CC = g++
LINK = g++
MPI_INC = -I../STUBS
MPI_PATH = -L../STUBS
MPI_LIB = -lmpi_stubs
You also need to build the STUBS library for your platform before
making LAMMPS itself. A ``make serial`` build does this for you
automatically, otherwise, type ``make mpi-stubs`` from the src
directory, or ``make`` from the ``src/STUBS`` dir. If the build
fails, you may need to edit the ``STUBS/Makefile`` for your
platform. The stubs library does not provide MPI/IO functions
required by some LAMMPS packages, e.g. ``MPIIO`` or ``USER-LB``,
and thus is not compatible with those packages.
.. note::
The file ``src/STUBS/mpi.cpp`` provides a CPU timer function
called ``MPI_Wtime()`` that calls ``gettimeofday()``. If your
operating system does not support ``gettimeofday()``, you will
need to insert code to call another timer. Note that the
ANSI-standard function ``clock()`` rolls over after an hour or
so, and is therefore insufficient for timing long LAMMPS
simulations.
MPI and OpenMP support in LAMMPS
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
If you are installing MPI yourself to build a parallel LAMMPS
executable, we recommend either MPICH or OpenMPI which are regularly
used and tested with LAMMPS by the LAMMPS developers. MPICH can be
downloaded from the `MPICH home page <https://www.mpich.org>`_ and
OpenMPI can be downloaded correspondingly from the `OpenMPI home page
<https://www.open-mpi.org>`_. Other MPI packages should also work. No
specific vendor provided and standard compliant MPI library is currently
known to be incompatible with LAMMPS. If you are running on a large
parallel machine, your system admins or the vendor should have already
installed a version of MPI, which is likely to be faster than a
self-installed MPICH or OpenMPI, so you should study the provided
documentation to find out how to build and link with it.
The majority of OpenMP (threading) support in LAMMPS is provided by the
``USER-OMP`` package; see the :doc:`Speed_omp`
page for details. The ``USER-INTEL`` package also includes OpenMP
threading (it is compatible with ``USER-OMP`` and will usually fall
back on styles from that package, if a ``USER-INTEL`` does not exist)
and adds vectorization support when compiled with compatible compilers,
in particular the Intel compilers on top of OpenMP. Also, the ``KOKKOS``
package can be compiled to include OpenMP threading.
In addition, there are a few commands in LAMMPS that have native OpenMP
support included as well. These are commands in the ``MPIIO``,
``SNAP``, ``USER-DIFFRACTION``, and ``USER-DPD`` packages. In addition
some packages support OpenMP threading indirectly through the libraries
they interface to: e.g. ``LATTE``, ``KSPACE``, and ``USER-COLVARS``.
See the :doc:`Packages details <Packages_details>` page for more
info on these packages and the pages for their respective commands
for OpenMP threading info.
For CMake, if you use ``BUILD_OMP=yes``, you can use these packages
and turn on their native OpenMP support and turn on their native OpenMP
support at run time, by setting the ``OMP_NUM_THREADS`` environment
variable before you launch LAMMPS.
For building via conventional make, the ``CCFLAGS`` and ``LINKFLAGS``
variables in Makefile.machine need to include the compiler flag that
enables OpenMP. For GNU compilers it is ``-fopenmp``\ . For (recent) Intel
compilers it is ``-qopenmp``\ . If you are using a different compiler,
please refer to its documentation.
.. _default-none-issues:
OpenMP Compiler compatibility
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Some compilers do not fully support the ``default(none)`` directive and
others (e.g. GCC version 9 and beyond, Clang version 10 and later) may
implement strict OpenMP 4.0 and later semantics, which are incompatible
with the OpenMP 3.1 semantics used in LAMMPS for maximal compatibility
with compiler versions in use. If compilation with OpenMP enabled fails
because of your compiler requiring strict OpenMP 4.0 semantics, you can
change the behavior by adding ``-D LAMMPS_OMP_COMPAT=4`` to the
``LMP_INC`` variable in your makefile, or add it to the command line
while configuring with CMake. LAMMPS will auto-detect a suitable setting
for most GNU, Clang, and Intel compilers.
----------
.. _compile:
Choice of compiler and compile/link options
---------------------------------------------------------
The choice of compiler and compiler flags can be important for maximum
performance. Vendor provided compilers for a specific hardware can
produce faster code than open-source compilers like the GNU compilers.
On the most common x86 hardware most popular C++ compilers are quite
similar in performance of C/C++ code at high optimization levels. When
using the ``USER-INTEL`` package, there is a distinct advantage in using
the `Intel C++ compiler <intel_>`_ due to much improved vectorization
through SSE and AVX instructions on compatible hardware as the source
code includes changes and Intel compiler specific directives to enable
high degrees of vectorization. This may change over time as equivalent
vectorization directives are included into OpenMP standard revisions and
other compilers adopt them.
.. _intel: https://software.intel.com/en-us/intel-compilers
On parallel clusters or supercomputers which use "environment modules"
for their compile/link environments, you can often access different
compilers by simply loading the appropriate module before building
LAMMPS.
.. tabs::
.. tab:: CMake build
By default CMake will use the compiler it finds according to
internal preferences and it will add optimization flags
appropriate to that compiler and any :doc:`accelerator packages
<Speed_packages>` you have included in the build. CMake will
check if the detected or selected compiler is compatible with the
C++ support requirements of LAMMPS and stop with an error, if this
is not the case.
You can tell CMake to look for a specific compiler with setting
CMake variables (listed below) during configuration. For a few
common choices, there are also presets in the ``cmake/presets``
folder. For convenience, there is a ``CMAKE_TUNE_FLAGS`` variable
that can be set to apply global compiler options (applied to
compilation only), to be used for adding compiler or host specific
optimization flags in addition to the "flags" variables listed
below. You may also specify the corresponding ``CMAKE_*_FLAGS``
variables individually, if you want to experiment with alternate
optimization flags. You should specify all 3 compilers, so that
the (few) LAMMPS source files written in C or Fortran are built
with a compiler consistent with the one used for the C++ files:
.. code-block:: bash
-D CMAKE_CXX_COMPILER=name # name of C++ compiler
-D CMAKE_C_COMPILER=name # name of C compiler
-D CMAKE_Fortran_COMPILER=name # name of Fortran compiler
-D CMAKE_CXX_FLAGS=string # flags to use with C++ compiler
-D CMAKE_C_FLAGS=string # flags to use with C compiler
-D CMAKE_Fortran_FLAGS=string # flags to use with Fortran compiler
A few example command lines are:
.. code-block:: bash
# Building with GNU Compilers:
cmake ../cmake -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_Fortran_COMPILER=gfortran
# Building with Intel Compilers:
cmake ../cmake -DCMAKE_C_COMPILER=icc -DCMAKE_CXX_COMPILER=icpc -DCMAKE_Fortran_COMPILER=ifort
# Building with Intel oneAPI Compilers:
cmake ../cmake -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DCMAKE_Fortran_COMPILER=ifx
# Building with LLVM/Clang Compilers:
cmake ../cmake -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_Fortran_COMPILER=flang
# Building with PGI/Nvidia Compilers:
cmake ../cmake -DCMAKE_C_COMPILER=pgcc -DCMAKE_CXX_COMPILER=pgc++ -DCMAKE_Fortran_COMPILER=pgfortran
For compiling with the Clang/LLVM compilers a CMake preset is
provided that can be loaded with
`-C ../cmake/presets/clang.cmake`. Similarly,
`-C ../cmake/presets/intel.cmake` should switch the compiler
toolchain to the legacy Intel compilers, `-C ../cmake/presets/oneapi.cmake`
will switch to the LLVM based oneAPI Intel compilers,
and `-C ../cmake/presets/pgi.cmake`
will switch the compiler to the PGI compilers.
In addition you can set ``CMAKE_TUNE_FLAGS`` to specifically add
compiler flags to tune for optimal performance on given hosts. By
default this variable is empty.
.. note::
When the cmake command completes, it prints a summary to the
screen which compilers it is using and what flags and settings
will be used for the compilation. Note that if the top-level
compiler is mpicxx, it is simply a wrapper on a real compiler.
The underlying compiler info is what CMake will try to
determine and report. You should check to confirm you are
using the compiler and optimization flags you want.
.. tab:: Makefile.machine settings for traditional make
The "compiler/linker settings" section of a Makefile.machine lists
compiler and linker settings for your C++ compiler, including
optimization flags. For a parallel build it is recommended to use
``mpicxx`` or ``mpiCC``, since these compiler wrappers will
include a variety of settings appropriate for your MPI
installation and thus avoiding the guesswork of finding the right
flags.
Parallel build (see ``src/MAKE/Makefile.mpi``):
.. code-block:: bash
CC = mpicxx
CCFLAGS = -g -O3
LINK = mpicxx
LINKFLAGS = -g -O
Serial build with GNU gcc (see ``src/MAKE/Makefile.serial``):
.. code-block:: make
CC = g++
CCFLAGS = -g -O3
LINK = g++
LINKFLAGS = -g -O
.. note::
If compilation stops with a message like the following:
.. code-block::
g++ -g -O3 -DLAMMPS_GZIP -DLAMMPS_MEMALIGN=64 -I../STUBS -c ../main.cpp
In file included from ../pointers.h:24:0,
from ../input.h:17,
from ../main.cpp:16:
../lmptype.h:34:2: error: #error LAMMPS requires a C++11 (or later) compliant compiler. Enable C++11 compatibility or upgrade the compiler.
then you have either an unsupported (old) compiler or you have
to turn on C++11 mode. The latter applies to GCC 4.8.x shipped
with RHEL 7.x and CentOS 7.x or GCC 5.4.x shipped with Ubuntu16.04.
For those compilers, you need to add the ``-std=c++11`` flag.
If there is no compiler that supports this flag (or equivalent),
you would have to install a newer compiler that supports C++11;
either as a binary package or through compiling from source.
If you build LAMMPS with any :doc:`Speed_packages` included,
there may be specific compiler or linker flags that are either
required or recommended to enable required features and to
achieve optimal performance. You need to include these in the
CCFLAGS and LINKFLAGS settings above. For details, see the
documentation for the individual packages listed on the
:doc:`Speed_packages` page. Or examine these files in the
src/MAKE/OPTIONS directory. They correspond to each of the 5
accelerator packages and their hardware variants:
.. code-block:: bash
Makefile.opt # OPT package
Makefile.omp # USER-OMP package
Makefile.intel_cpu # USER-INTEL package for CPUs
Makefile.intel_coprocessor # USER-INTEL package for KNLs
Makefile.gpu # GPU package
Makefile.kokkos_cuda_mpi # KOKKOS package for GPUs
Makefile.kokkos_omp # KOKKOS package for CPUs (OpenMP)
Makefile.kokkos_phi # KOKKOS package for KNLs (OpenMP)
----------
.. _exe:
.. _library:
Build the LAMMPS executable and library
---------------------------------------
LAMMPS is always built as a library of C++ classes plus an executable.
The executable is a simple ``main()`` function that sets up MPI and then
creates a LAMMPS class instance from the LAMMPS library, which
will then process commands provided via a file or from the console
input. The LAMMPS library can also be called from another application
or a scripting language. See the :doc:`Howto couple <Howto_couple>` doc
page for more info on coupling LAMMPS to other codes. See the
:doc:`Python <Python_head>` page for more info on wrapping and
running LAMMPS from Python via its library interface.
.. tabs::
.. tab:: CMake build
For CMake builds, you can select through setting CMake variables
between building a shared or a static LAMMPS library and what kind
of suffix is added to them (in case you want to concurrently
install multiple variants of binaries with different settings). If
none are set, defaults are applied.
.. code-block:: bash
-D BUILD_SHARED_LIBS=value # yes or no (default)
-D LAMMPS_MACHINE=name # name = mpi, serial, mybox, titan, laptop, etc
# no default value
The compilation will always produce a LAMMPS library and an
executable linked to it. By default this will be a static library
named ``liblammps.a`` and an executable named ``lmp`` Setting
``BUILD_SHARED_LIBS=yes`` will instead produce a shared library
called ``liblammps.so`` (or ``liblammps.dylib`` or
``liblammps.dll`` depending on the platform) If
``LAMMPS_MACHINE=name`` is set in addition, the name of the
generated libraries will be changed to either ``liblammps_name.a``
or ``liblammps_name.so``\ , respectively and the executable will
be called ``lmp_name``.
.. tab:: Traditional make
With the traditional makefile based build process, the choice of
the generated executable or library depends on the "mode" setting.
Several options are available and ``mode=static`` is the default.
.. code-block:: bash
make machine # build LAMMPS executable lmp_machine
make mode=static machine # same as "make machine"
make mode=shared machine # build LAMMPS shared lib liblammps_machine.so instead
The "static" build will generate a static library called
``liblammps_machine.a`` and an executable named ``lmp_machine``\ ,
while the "shared" build will generate a shared library
``liblammps_machine.so`` instead and ``lmp_machine`` will be
linked to it. The build step will also create generic soft links,
named ``liblammps.a`` and ``liblammps.so``\ , which point to the
specific ``liblammps_machine.a/so`` files.
Additional information
^^^^^^^^^^^^^^^^^^^^^^
Note that for creating a shared library, all the libraries it depends on
must be compiled to be compatible with shared libraries. This should be
the case for libraries included with LAMMPS, such as the dummy MPI
library in ``src/STUBS`` or any package libraries in the ``lib``
directory, since they are always built in a shared library compatible
way using the ``-fPIC`` compiler switch. However, if an auxiliary
library (like MPI or FFTW) does not exist as a compatible format, the
shared library linking step may generate an error. This means you will
need to install a compatible version of the auxiliary library. The
build instructions for that library should tell you how to do this.
As an example, here is how to build and install the `MPICH library
<mpich_>`_, a popular open-source version of MPI, as a shared library
in the default /usr/local/lib location:
.. _mpich: https://www.mpich.org
.. code-block:: bash
./configure --enable-shared
make
make install
You may need to use ``sudo make install`` in place of the last line if
you do not have write privileges for ``/usr/local/lib`` or use the
``--prefix`` configuration option to select an installation folder,
where you do have write access. The end result should be the file
``/usr/local/lib/libmpich.so``. On many Linux installations the folder
``${HOME}/.local`` is an alternative to using ``/usr/local`` and does
not require superuser or sudo access. In that case the configuration
step becomes:
.. code-block:: bash
./configure --enable-shared --prefix=${HOME}/.local
Avoiding to use "sudo" for custom software installation (i.e. from source
and not through a package manager tool provided by the OS) is generally
recommended to ensure the integrity of the system software installation.
----------
.. _debug:
Including or removing debug support
-----------------------------------
By default the compilation settings will include the *-g* flag which
instructs the compiler to include debug information (e.g. which line of
source code a particular instruction correspond to). This can be
extremely useful in case LAMMPS crashes and can help to provide crucial
information in :doc:`tracking down the origin of a crash <Errors_debug>`
and help the LAMMPS developers fix bugs in the source code. However,
this increases the storage requirements for object files, libraries, and
the executable 3-5 fold.
If this is a concern, you can change the compilation settings or remove
the debug information from the LAMMPS executable:
- **Traditional make**: edit your ``Makefile.<machine>`` to remove the
*-g* flag from the ``CCFLAGS`` and ``LINKFLAGS`` definitions
- **CMake**: use ``-D CMAKE_BUILD_TYPE=Release`` or explicitly reset
the applicable compiler flags (best done using the text mode or
graphical user interface).
- **Remove debug info**: If you are only concerned about the executable
being too large, you can use the ``strip`` tool (e.g. ``strip
lmp_serial``) to remove the debug information from the executable file.
Do not strip libraries or object files, as that will render them unusable.
----------
.. _tools:
Build LAMMPS tools
------------------------------
Some tools described in :doc:`Auxiliary tools <Tools>` can be built directly
using CMake or Make.
.. tabs::
.. tab:: CMake build
.. code-block:: bash
-D BUILD_TOOLS=value # yes or no (default)
-D BUILD_LAMMPS_SHELL=value # yes or no (default)
The generated binaries will also become part of the LAMMPS installation
(see below).
.. tab:: Traditional make
.. code-block:: bash
cd lammps/tools
make all # build all binaries of tools
make binary2txt # build only binary2txt tool
make chain # build only chain tool
make micelle2d # build only micelle2d tool
make thermo_extract # build only thermo_extract tool
cd lammps/tools/lammps-shell
make # build LAMMPS shell
----------
.. _install:
Install LAMMPS after a build
------------------------------------------
After building LAMMPS, you may wish to copy the LAMMPS executable of
library, along with other LAMMPS files (library header, doc files) to
a globally visible place on your system, for others to access. Note
that you may need super-user privileges (e.g. sudo) if the directory
you want to copy files to is protected.
.. tabs::
.. tab:: CMake build
.. code-block:: bash
cmake -D CMAKE_INSTALL_PREFIX=path [options ...] ../cmake
make # perform make after CMake command
make install # perform the installation into prefix
During the installation process CMake will by default remove any runtime
path settings for loading shared libraries. Because of this you may
have to set or modify the ``LD_LIBRARY_PATH`` (or ``DYLD_LIBRARY_PATH``)
environment variable, if you are installing LAMMPS into a non-system
location and/or are linking to libraries in a non-system location that
depend on such runtime path settings.
As an alternative you may set the CMake variable ``LAMMPS_INSTALL_RPATH``
to ``on`` and then the runtime paths for any linked shared libraries
and the library installation folder for the LAMMPS library will be
embedded and thus the requirement to set environment variables is avoided.
The ``off`` setting is usually preferred for packaged binaries or when
setting up environment modules, the ``on`` setting is more convenient
for installing software into a non-system or personal folder.
.. tab:: Traditional make
There is no "install" option in the ``src/Makefile`` for LAMMPS.
If you wish to do this you will need to first build LAMMPS, then
manually copy the desired LAMMPS files to the appropriate system
directories.