GitLab

The forces were not correct for different uind/uinp as a result.

Non-PME OPT calculations now use a much more efficient Cartesian field algorithm for the dipole response force terms.

The QI direct space energy uses the energy functional

    U = 1/4 (E_p µ_d + E_d µ_p)

while the code it replaced uses the equivalent form

    U = 1/2 E_p µ_d

Therefore the reciprocal and self terms have been adapted to use the more symmetric form.  The bad code yielded incorrect energies when PME is used AND µ_d and µ_p are different.  The forces and torques were not affected.

o Replaced Cartesians in the reciprocal and self PME terms.
o Added non-PME spherical harmonics field generation and force evaluation routines.
o Added spherical harmonic potential evaluation, to allow removal of the Cartesian moments.

o Permanent and induced moment energies, forces and torques are implemented, for PME direct space.
o The code is activated by a preprocessor macro at the top of AmoebaReferenceMultipoleForce.h, and can be completely disabled.
o Currently, the Cartesian momenents are stored alongside the spherical harmonic moments, but these can eventually be removed when the reciprocal and self terms have been adapted to use spherical harmonics (in progress) and the non-PME routines have been coded.

stretch-bend test correctly.

The potential energy calculation done in the ReferenceStretchBendTest file was
updated to mirror that of the Reference kernel, but first I made sure that the
test passed *before* updating to the new energy expression, just to make sure
that the energies and forces didn't *change* with the new expression.

of just 1.  Backwards compatibility is provided by making the 2nd force constant
default to -1, which is reinterpreted as "copy the first force constant".

Updates both the reference and CUDA kernels.