GitLab

Fixed potential energy discrepancies in large systems between CUDA and Reference platform. The energy calculation now uses the same method as that of reciprocalConvolution so it mirrors the reference implementation. OpenCL is fine.  

When converting to fixed point, multiply by 0x100000000 instead of 0xFFFFFFFF.  This should be (very very slightly) more accurate, since its reciprocal can be exactly represented in floating point.

Fixed error in updateNeighborListSize() when createInteractionKernel() has not been called (e.g. with AmoebaVdwForce)

CustomNonbondedForce.cpp, CudaNonbondedUtilities.cpp and CudaKernels.cpp are now fully reverted to 3478.

LPW attempted bugfixes: (1) Changed unit multipliers in AmoebaVdwForce::getParticleParameters (2) Sorted vectors in exclusionLists so it agrees with atomExclusions

OpenCLArray is no longer templatized and doesn't provide a host buffer.  This is in preparation for adding mixed/double precision support.

Began implementing new mixed precision model that does integration in double precision and force evaluation in single precision