Add expanded gaussian pair for hpmc

[josephburkhart]

Description

I have created a new expanded gaussian pair potential for HPMC. The implementation of this potential closely follows the implementation of the lennard-jones pair potential for HPMC, which has led to some minor code duplication that may warrant later refactoring. The mathematical definition of the new pair potential should be equivalent to that of the expanded gaussian in the MD module. As part of this change, the following files were created:

• hoomd/hpmc/PairPotentialExpandedGaussian.h - templated off of PairPotentialLennardJones.h
• hoomd/hpmc/PairPotentialExpandedGaussian.cc - templated off of PairPotentialLennardJones.cc
• hoomd/hpmc/pair/expanded_gaussian.py - templated off of lennard_jones.py
• hoomd/hpmc/pair/pytest/test_pair_expanded_gaussian.py - templated off of test_pair_lennard_jones.py

Motivation and context

Adding this pair potential will allow users to stochastically simulate systems with interactions that are defined by gaussian models, which previously has only been possible in MD. This may provide benefits, for example, in very large systems.

How has this been tested?

• [x] HOOMD compiled locally without errors
• [x] ctest completed locally without errors
• [x] pytest completed locally without errors
• [x] sphinx documentation built locally without errors and appears correct
• [x] an example simulation appeared visually correct (tested with both repulsive and attractive parameters)

Change log

Added:

* ``hpmc.pair.ExpandedGaussian`` computes the expanded Gaussian pair potential in HPMC
  (`#1817 <https://github.com/glotzerlab/hoomd-blue/pull/1817>`__).

Checklist:

• [x] I have reviewed the Contributor Guidelines.
• [x] I agree with the terms of the HOOMD-blue Contributor Agreement.
• [x] My name is on the list of contributors (sphinx-doc/credits.rst) in the pull request source branch.

[josephburkhart]

pre-commit.ci autofix

[josephburkhart]

pre-commit.ci autofix

[josephburkhart]

In an offline conversation, @SchoeniPhlippsn and I were discussing some revisions to the pre-computed values in PairPotentialExpandedGaussian.h, namely, removing the lines that pre-compute r_cut_squared and r_on_squared, since I just take the square root of them in PairPotentialExpandedGaussian.cc.

I pointed out that in PairPotentialExpandedGaussian.cc, there are some calculations that could actually be moved to the header file - many of the lines that calculate shifting/smoothing functions when the mode is xplore or shift do not actually depend on the current value of r at all, so they could just be pre-computed in the header.

I left the code this way because it's how it's structured in the HPMC LJ Pair Potential, and I wanted to keep things consistent. But moving the computation to the header file would make things more efficient. What do others think?

[joaander]

Yes, you can precompute additional constant terms and store them in the ParamType data structure. This adds memory per type pair, but saves a few cycles per interaction. On the CPU this is probably a performance win.

And this is a good observation on r_cut_squared. It may be best to keep this one consistent with LJ - if we move to a generic base class then they will need to be consistent. Using the pre-computed value avoids the need to compute sqrt(r_cut_squared) for every particle pair interaction when mode == 'shift'.

[joaander]

We can make an issue and plan to refactor the precomputing in a future PR.

On the GPU, memory is expensive and floating point operations are cheap. This is why the HOOMD MD code does not precompute these quantities. That behavior was copied into HPMC. However, these HPMC potentials are CPU only where the balance is different. I expect a small (but measurable) performance gain by precomputing these on the CPU (especially when sqrt is required).

This is another reason to wait a bit on refactoring common isotropic pair methods into a HPMC base class. We have not yet recognized all of the common elements (e.g. precomputation) that could be combined.