Python MPB interface

[ChristopherHogan]

Adds MPB-based mode-solver support from Python.

Eventually, we might want to refactor this into MPB itself, but that will require refactoring libmeepgeom into a meep-independent part, so for now we will require users to install Meep in order to get the Python MPB interface.

Incidentally, this PR fixes #180.

[ChristopherHogan]

Description of Changes by File

.travis.yml

Add h5py and scipy as dependencies. Scipy is only used in mpb-tutorial.py as replacements for some of the optimization functions in libctl/base/math-utils.scm. These packages probably need to be checked in configure.

Makefile.am

Build libpympb if we have MPB and libctlgeom.

configure.ac

• Make a WITH_MPB variable available in automake.
• Make a @LIBCTLGEOM_LIBS@ variable available in automake.

libmeepgeom/material_data.hpp

• Convert susceptibility_list, medium_struct, and material_data into C++ style structs with default values set in a constructor. This is what fixes meep segfaults #180.

libmeepgeom/meepgeom.cpp

• Removed vacuum_medium because the default constructor gives us that.
• Remove static keyword from a couple functions so they can be called from libpympb.

libpympb/Makefile.am

Builds libpympb.so, which is the C++ portion of MPB that python calls into.

libpympb/pympb.cpp and pympb.hpp

Contains the mode_solver class, which holds all the MPB state. This class is wrapped by SWIG, and all MPB state is accessed from Python through an instance of this class. Most of the code in here is taken from MPB.

• dielectric_function is what's passed to set_maxwell_dielectric.
• mode_solver::init corresponds to init_params in mpb.c. It's called in the constructor though, so the Python API just creates a mode_solver and calls solve_kpoint for each kpoint. The get_Xfield functions all take arrays and a size so we can pass a numpy array to them as an output argument. The numpy arrays are created in the python API layer with the help of get_dims.

python /Makefile.am

• Build the MPB interface if WITH_MPB is defined.

python/examples/mpb-tutorial.py

A port of tutorial.ctl from MPB. This demonstrates all the functionality currently supported.

python/geom.py

Add the Matrix and Lattice Python classes, along with geometric_object_lattice_duplicates for the "Point defect state" tutorial.

python/meep.i

I put the common typemap functions used by both pymeep and pympb in typemap_utils.cpp. The typemap helpers specific to each module are defined in their respectiv ".i" file.

python/mpb.i

• Wraps the mode_solver class
• Defines some typemap helpers specific to pympb
• Adds typmaps for getting data from/to the mode_solver.

python/solver.py

Defines the main Python object, ModeSolver, that wraps the mode_solver. Analogous to simulation.py and Simulation. Most of the code is ported from mpb.scm. The output_*_field functions were ported from output_field_to_file in mpb/mpb/fields.c.

python/tests/data/tutorial-*.h5

Reference data collected from scheme runs (setting deterministic=true and passing a NULL mean_epsilon_func to set_maxwell_dielectric). This data is read by the unit tests and compared to the python results.

python/tests/geom.py

Add test case for #180

python/tests/mpb.py

Unit tests for all MPB functionality. There is a test for every section of the tutorial in mpb-tutorial.py comparing results obtained from scheme. The setUp function registers a function to clean up all the generated h5 files after the tests complete.

The only test that's questionable is test_output_efield_z. It doesn't have the accuracy that the rest do. I've spent hours in the debugger trying to track it down, but haven't been successful.

python/typemap_utils.cpp

Refactored to include the functions that are used in both meep.i and mpb.i.

@stevengj @oskooi @HomerReid