GitLab

Tabulated function parameters are hardcoded in the kernel instead of being stored in an array.  This makes the code simpler and may help performance slightly.

Converted the routine for identifying molecules from a recursion to a loop.  This avoids stack overflows with very long linear molecules.

Platform specific header files get installed.  This allows plugins to be built with just an OpenMM installation, not a full source tree.

Exported some symbols from Cuda plugin, for use by external NML plugin.
  cudaOpenMMInitializeIntegration, gpuSetConstants, CudaPlatform
  (but I couldn't figure out how to export kGenerateRandoms)
Also turned off TestRandomCuda on Windows, since it always fails and makes the dashboard look bad.
Renamed InitializeIntegration method to cudaOpenMMInitializeIntegration, now that symbol is being exported.

Added a version of Platform::getPlatform() that looks up platforms by name.  Added a Context constructor that allows platform-specific properties to be specified.

Restructured how force and energy computations are done to make it easier to write plugins for the CUDA platform

Created a CudaPlatform property for specifying whether CUDA should sync or spin loop while waiting for results

(2) Changed erfc.cpp to a header file which defines missing erf() and erfc() as static functions when using VC++ but otherwise does nothing.
(3) Replaced all nonstandard uses of real() and imag() methods as lvalues; instead work directly with the complex values.
(4) Fixed a bug in TestFindExclusions which was caught by VC++'s extensive STL debug mode self-examination.

There are still problems in hilbert.c and not all CUDA tests pass.