112. Accelerated Particle-Grid Mapping
Authors: Ahmed Sanaullah (Boston University)
Abstract: Charge mapping is critical to electrostatic computations for Molecular Dynamics. It reduces the complexity of evaluating long-range coulombic forces by diffusing discrete particle charges to a regular grid. Efficient charge mapping on accelerators (GPUs, FPGAs) is non-trivial, with the compute and memory intensive nature of the algorithm limiting performance benefits in naive implementations. On FPGAs, resource constraints have only allowed low order (bicubic) interpolations. In our work, we explore methods for improving the performance on both platforms. These include application specific data structures and low complexity kernels for GPUs and deep pipelines with interleaved memory access for FPGAs. Our best case implementation shows > 62x speed-up over existing CPU codes and >30x speed-up over existing GPU codes. We also find that, when using the Altera Arria 10, high resource availability enables the building of a balanced accelerator for the entire long-range electrostatics computation on a single FPGA.
Two-page extended abstract: pdf