GitLab

Implemented an optimized real-to-complex FFT

OpenCL FFT supports real-to-complex transforms

Fixed a performance regression on AMD GPUs

More fixes to OpenCL with CPU device

Fixed errors using OpenCL with CPU device

Parallelized extraction of inverse matrix for CCMA

Fixed a race condition in OpenCL

Simplification to reference platform tests

Fixed an error in doxygen comments

Fix HAngles with Amber prmtop files

Adds a test for it as well.

files.

CMAPTorsionForce implements updateParametersInContext()

Modified travis logic to only fail if retested tests fail.

Fixing multiple bugs in dcdfile / dcdreporter.

Fix very subtle bug in Amber prmtop with NBFIX

Some topology files have some of the nonbonded pairs in the Lennard-Jones matrix
pointing to a 0 entry in the HBOND_ACOEF and HBOND_BCOEF arrays. This commit
simply skips over those entries, having previously checked that all entries in
the 10-12 sections are 0.  There is actually an example of this already in the
test suite (see ff14ipq.parm7), but this fortuitously did not cause any
difference in those test results.

The reason this bug is not evident in that test case is because the A- and
B-coefficients pulled when the index was -1 (later decremented to -2) was also
coincidentally 0. However, if this was not the case (as it does not *have* to
be), then who knows what could happen.

This also adds more of a check against 10-12 topology files. ParmEd actually
supports using the 10-12 potential in OpenMM when specified in the Amber prmtop,
but it is sufficiently rare that it is not worth porting to OpenMM, IMO.