GitLab

This change is intended to improve the behavior when multiple
CUDA contexts are created on a node with multiple CUDA GPUs,
each of which is set to process-exclusive mode.

Added checks for nan coordinates in CUDA and OpenCL

Add code supporting input with and without units for CharmmPsfFile.

Add relevant tests (that used to fail). This also fixes a subtle bug where
switching functions were not applied when NBFIXes were specified.

Removed some unnecessary double precision

Add permissive flag to permit undefined atomtypes when reading charmm parameter files

Fixed a bug in AmoebaMultipoleForce

Replace JOM with NMake in Appveyor build

Hard wall constraints for DrudeLangevinIntegrator

Further improved energy accuracy in mixed precision mode

More optimizations to CPU platform

Fix units of returned dipoles and quadrupoles from AmoebaMultipoleForce.

Now that I have my head on straight, that is.

Fix very irritating behavior on some Mac systems.

Don't try to build tests if BUILD_TESTING is off

Hard-code CC and CXX to clang for Macs, since gcc/g++ will *not* work with
Anaconda, despite the fact that distutils will try to use them.  System Python,
homebrew, and MacPorts on Macs will always use clang, so this hack should always
work and fix issues with users that have GCC installed from MacPorts or homebrew
*and* Anaconda.

Add a set of unit tests for unit handling in the core API.