MPI support for SGP construction

[YuuuXie]

Features

Added a new class ParallelSGP to build SGP from a large, fixed training data set using MPI, not on-the-fly construction. The construction of SGP consists of two processes:

• Load, distribute training data and compute descriptors, by method load_local_training_data
• Compute kernel matrices and vectors from the training set, by method compute_matrices
• Method build is a general interface, which calls load_local_training_data and compute_matrices

Added unit test: tests/test_parallel_sgp.cpp

• To run the unit test, try ./tests for serial version,
• and mpirun -n <N_procs> ./tests for parallel version, N_procs=1, 4, 9, 16,...

Added timing test of MPI: timing/mpi_construction.cpp

• To run the timing test, go into the folder build/timing, and run mpirun -n <N_procs> ./time_mpi <N_strucs>, N_strucs is the number of training structures, each structure will have 100 atoms and 5 sparse envs

Added .xyz file reader. Since the distmatrix does not support python binding, the ParallelSGP can not build python interface, thus python io like ASE can not be used to convert data in file into arrays. So a c++ reader for xyz file is provided

• src/flare_pp/utils.cpp
• flare_pp/utils.py
• test_utils.cpp

Dependencies

• OpenMPI. Below are the modules that need to be loaded before running cmake

  module load cmake/3.17.3-fasrc01
  module load gcc/9.3.0-fasrc01 openmpi/4.0.5-fasrc01
  module load intel-mkl/2019.5.281-fasrc01
  module load python/3.6.3-fasrc01
  module load eigen/3.3.7-fasrc01

• Distmatrix by Anders. Added to CMakeLists.txt

Todos

• [x] unit test: tests/test_parallel_sgp.cpp
• [x] Timing test of MPI: timing/mpi_construction.cpp
• [x] .xyz file reader.
• [x] multiple kernels/descriptors
• [x] Pull branch #20
• [x] Add likelihood and gradient evaluation
• [ ] Add more docs
  • [x] docs for python api
• [ ] Move some printings to the debug flag
• [ ] The distmatrix needs to be public @anjohan

Lower priority:

• Add update_matrix to support OTF training (theoretically we can still do OTF with current ParallelSGP, but each time it construct on the whole dataset, instead of update matrices with the new added data)
• The distmatrix class needs python binding