Performance improvements¶
Improved performance of systems that are inhomogeneous along z¶
The non-bonded kernels got very inefficient with systems that are have a very inhomogeneous atom distribution along the z-dimension. Now all three dimensions are considered when estimating atom density.
Reduced host-side blocking in the GPU update and constraints setup¶
The mass and temperature-coupling group data that the GPU update and constraints upload on every neighbour-search step is now held in pinned host memory and copied asynchronously, rather than with blocking copies. This cuts the host-side cost of the GPU update and constraints setup by roughly a factor of 2.6, with the largest benefit for runs with frequent pair-list updates.