1. Installation

1.1. Dependencies

CheMPS2 can be built with CMake and depends on

• BLAS
• LAPACK
• GSL (GNU Scientific library)
• HDF5 (Hierarchical Data Format Release 5)

It is parallelized for shared memory architectures with the Open Multi-Processing (OpenMP) API and for distributed memory architectures with the Message Passing Interface (MPI). A hybrid combination of both parallelization strategies is supported.

1.2. Download

It is advised to clone the CheMPS2 git repository from github. In your terminal, run:

$ cd /sourcefolder
$ git clone 'https://github.com/sebwouters/chemps2'
$ cd chemps2

That way, future updates and bug fixes can be easily pulled in:

$ cd /sourcefolder/chemps2
$ git pull

1.3. Build the chemps2 library and binary

The files

/sourcefolder/chemps2/CMakeLists.txt
/sourcefolder/chemps2/CheMPS2/CMakeLists.txt
/sourcefolder/chemps2/tests/CMakeLists.txt
/sourcefolder/chemps2/sphinx/CMakeLists.txt

provide a minimal compilation. In your terminal, run:

$ cd /sourcefolder/chemps2
$ mkdir build
$ cd build

CMake generates makefiles based on the user’s specifications:

$ CXX=option1 cmake .. -DMKL=option2 -DCMAKE_INSTALL_PREFIX=/option3 -DWITH_MPI=option4

1. Option1 is the c++ compiler; typically g++, icpc, or clang++ on Linux. It is advised to use the intel compiler if available.
2. Option2 can be ON or OFF and is used to switch on the intel math kernel library.
3. /option3 is the prefix of the installation directory; typically /usr or /usr/local on Linux. On my computer, libchemps2 is then installed in /option3/lib/x86_64-linux-gnu, the headers in /option3/include/chemps2, and the binary in /option3/bin/chemps2.
4. Option4 can be ON or OFF and is used to switch on the possibility to compile with MPI. Please note that the compiler should then provide mpi.h. Option1 should hence be the mpic++ compiler; typically mpic++ or mpiicpc on Linux. It is advised to use the intel compiler if available.

If one or more of the required libraries are not found, use the command

$ CMAKE_INCLUDE_PATH=option5 CMAKE_LIBRARY_PATH=option6 CXX=option1 cmake .. -DMKL=option2 -DCMAKE_INSTALL_PREFIX=/option3 -DWITH_MPI=option4

instead, where option5 and option6 are respectively the missing colon-separated include and library paths:

CMAKE_INCLUDE_PATH=/my_libs/lib1/include:/my_libs/lib2/include
CMAKE_LIBRARY_PATH=/my_libs/lib1/lib:/my_libs/lib2/lib

For debian/sid, the HDF5 headers are located in the folder /usr/include/hdf5/serial. If CMake complains about the HDF5 headers, try to pass it with the option -DHDF5_INCLUDE_DIRS=/usr/include/hdf5/serial.

To compile, run:

$ make

To install, run:

$ make install

For non-standard installation directories, please remember to append the library path to LD_LIBRARY_PATH in your .bashrc.

1.4. Test libchemps2

To test libchemps2 for compilation without MPI, run:

$ cd /sourcefolder/chemps2/build
$ make test

To test libchemps2 for compilation with MPI, run:

$ cd /sourcefolder/chemps2/build/tests
$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./test1
$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./test2
$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./test3
$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./test4
$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./test5
$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./test7
$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./test10

YYY specifies the number of threads per process and ZZZ the number of processes. The tests only require a very limited amount of memory (order 100 MB). Note that the tests are too small to see (near) linear scaling with the number of cores, although improvement should still be noticeable.

1.5. Test the chemps2 binary

To test the chemps2 binary for compilation without MPI, run:

$ man /sourcefolder/chemps2/manpage.1
$ cd /sourcefolder/chemps2/build/CheMPS2
$ ./chemps2 --help
$ ./chemps2 --fcidump=/sourcefolder/chemps2/tests/matrixelements/H2O.631G.FCIDUMP \
            --group=5 \
            --sweep_d=200,1000 \
            --sweep_econv=1e-8,1e-8 \
            --sweep_maxit=2,10 \
            --sweep_noise=0.05,0.0 \
            --twodmfile=2dm.out \
            --print_corr \
            --reorder=6,5,4,3,2,1,0,7,8,9,10,11,12

To test the chemps2 binary for compilation with MPI, prepend the binary with:

$ OMP_NUM_THREADS=YYY mpirun -np ZZZ ./chemps2 [OPTIONS]

1.6. Build PyCheMPS2

PyCheMPS2 is a python interface to libchemps2, for compilation without MPI. It can be built with Cython. The installation is independent of CMake and assumes that you have installed the CheMPS2 library with make install. For non-standard installation directories of CheMPS2, please remember to append the library path to LD_LIBRARY_PATH in your .bashrc. In addition, the include path should be appended to CPATH:

$ export CPATH=${CPATH}:/option3/include

where /option3 is the option provided to CMake with -DCMAKE_INSTALL_PREFIX=/option3 above. For debian/sid, the HDF5 headers are located in the folder /usr/include/hdf5/serial. If it was explicitly passed to CMake, it should also be appended to CPATH:

$ export CPATH=${CPATH}:/option3/include:/usr/include/hdf5/serial

The python wrapper can be installed with:

$ cd /sourcefolder/chemps2/PyCheMPS2
$ python setup.py build_ext -L ${LD_LIBRARY_PATH}
$ python setup.py install --prefix=/option3

On my machine, the python wrapper is installed to the folder /option3/lib/python2.7/site-packages, but the folder lib and the distribution of python can vary.

Compilation of PyCheMPS2 occurs by linking to the c++ library in the installation directory. The installation of PyCheMPS2 will fail if that library is not properly installed. If you have pulled a newer version of CheMPS2, please remember to reinstall the c++ library first, before reinstalling PyCheMPS2!

1.7. Test PyCheMPS2

When libchemps2 has been compiled without MPI, PyCheMPS2 can be tested by running (remember that the python site-packages folder can vary):

$ cd /sourcefolder/chemps2/PyCheMPS2/tests
$ export PYTHONPATH=${PYTHONPATH}:/option3/lib/python2.7/site-packages
$ python test1.py
$ python test2.py
$ python test3.py
$ python test4.py
$ python test5.py
$ python test6.py
$ python test7.py
$ python test8.py
$ python test9.py
$ python test10.py

If you compiled the c++ library with -DMKL=ON, you might get the error

Intel MKL FATAL ERROR: Cannot load libmkl_avx.so or libmkl_def.so.

This issue of using Intel’s MKL inside python is known and reported. To get the python tests to run, you can set the variable LD_PRELOAD in order to preload libmkl_rt.so. On my system, this is done with

$ export LD_PRELOAD=/opt/intel/mkl/lib/intel64/libmkl_rt.so

The python tests do exactly the same thing as the c++ tests above, and illustrate the usage of the python interface to libchemps2. The tests should end with a line stating whether or not they succeeded. They only require a very limited amount of memory (order 100 MB). Note that the tests are too small to see (near) linear scaling with the number of cores, although improvement should still be noticeable.