lammps/examples/COUPLE/plugin

This directory has a simple C code that shows how
LAMMPS can be linked to a driver application as a library. The purpose
is to illustrate how another code could perform computations while
using LAMMPS to perform MD on all or a subset of the processors, or
how an umbrella code or script could call both LAMMPS and some other
code to perform a coupled calculation.

simple.c           is the C driver
liblammpsplugin.c  is the LAMMPS library plugin loader

The 3 codes do the same thing, so you can compare them to see how to
drive LAMMPS from each language.  See lammps/python/example/simple.py
to do something similar from Python.  The Fortran driver requires an
additional wrapper library that interfaces the C interface of the
LAMMPS library to Fortran and also translates the MPI communicator
from Fortran to C.

First build LAMMPS as a library (see examples/COUPLE/README), e.g. 

cd $HOME/lammps/src
make mode=shlib mpi

or 

cd $HOME/lammps
mkdir build-shared
cd build-shared
cmake -D BUILD_LIB=on -D BUILD_SHARED_LIBS=on ../cmake
make

You can then build any of the driver codes with compile lines like
these, which include paths to the LAMMPS library interface, and
linking with FFTW (only needed if you built LAMMPS as a library with
its PPPM solver).

mpicc -c simple.c
mpicc simple.o -llammps -lfftw -o simpleC


You then run simpleC on a parallel machine
on some number of processors Q with 3 arguments:

% mpirun -np Q simpleC P in.lj $HOME/lammps/src/liblammps.so

or

% mpirun -np Q simpleC P in.lj $HOME/lammps/build-shared/liblammps.so

P is the number of procs you want LAMMPS to run on (must be <= Q) and
in.lj is a LAMMPS input script and the last argument is the path to
the LAMMPS shared library. This either has to be an absolute path, or
liblammps.so has to be in a folder that is included in the environment
variable LD_LIBRARY_PATH so it will be found by the dynamic object loader.

The driver will launch LAMMPS on P procs, read the input script a line
at a time, and pass each command line to LAMMPS.  The final line of
the script is a "run" command, so LAMMPS will run the problem.

The driver then requests all the atom coordinates from LAMMPS, moves
one of the atoms a small amount "epsilon", passes the coordinates back
to LAMMPS, and runs LAMMPS again.  If you look at the output, you
should see a small energy change between runs, due to the moved atom.

The C driver is calling C-style routines in the src/library.cpp file
of LAMMPS through the function pointers in the liblammpsplugin_t struct.
This has the benefit that your binary is not linked to liblammps.so directly
and thus you can change the name of the shared library (e.g. to have 
different variants compiled, or to load a different LAMMPS versions without
having to update your executable).