GitLab

The differences in the Ewald energies between the Reference and Cuda platforms were traced to roundoff error in the 
computation of the Vdw and Coulomb ixns for the short-range energy contribution (verified using double precision in the sum); the vdw and Coulomb terms are now accumulated separately and then added to the total energy 

In Ewald SR portion of code, the LJ contribution to dEdR was missing a factor of (inverseR*inverseR)

All energies existing in Preview Release 3 now run on the GPU, new Ewald energy terms still use Reference implementation

Forced computation of dot product used in computing angles to be done in double precision; acos() is very steep for angles near pi and single precison was giving disagreement w/ corresponding Gromacs values (discrepancies of the order of ~1.0)  

Fixed some MSVC conversion warnings and moved method-local inclusion of pme.h up to start of LJ file -- it wouldn't compile when included locally due to attempt to match extern binding with anonymous struct type.

No commit message

- MingW, where "WIN32" is set by CMake but the compiler is not Microsoft's (should use _MSC_VER) in that case, and 
- in gcc 4.4 which is fussy about some issues that earlier versions were silent about, such as passing a string constant to a "char*" rather than a "const char*".
I also fixed a syntax error in CMakeLists that showed up in the OpenMM workshop for some unknown reason.

Brook is not currently building correctly; I think it has not yet been fully converted to the getTime() interface needed for variable step integrators.

Optimization to constraint algorithm: skip the last iteration as soon as we see that all constraints have converged.

Eliminate some VC++ build problems via the usual expedient of explicit casts from unsigned to signed and from double to float. TNT Arrays apparently can't be indexed with unsigned because VC++ finds something ambiguous then above their operator[] methods.