Over the past couple of weeks further optimizations were applied to the AGBNP3 algorithm. These mainly focused on the pairwise-descreening calculation for Born radii. Prior to this, the descreening kernel employed an interpolation procedure which utilizes two C-spline calculations. The new algorithm utilizes only one. The benefits are substantial, especially for the larger systems, as illustrated below (see May results vs. April results):
These results complete the main phase of the AGBNP3 serial performance optimization project. We nearly tripled the speed of the code relative to AGBNP2 without non-bonded cutoffs and doubled it relative to AGBNP2 with cutoffs (for system sizes illustrated above). Future efforts will focus on parallelization on multi-core and GPU architectures.