Programming Massively Parallel Processors, 4th Edition

Chapter 18

Electrostatic potential map

With special contributions from John Stone

Abstract

This chapter presents a series of decisions and tradeoffs in parallelizing the calculation of electrostatic potential energy in a regularly spaced energy grid. We show that parallelizing a highly optimized sequential C implementation of the direct Coulomb summation method leads to a slow scatter kernel that requires heavy use of atomic operations. We then show that one can parallelize a less optimized sequential C code into a gather kernel that has much higher parallel execution speed. We also show that through thread coarsening, we can reclaim much of the efficiency of the optimized sequential execution. We further demonstrate that by carefully choosing ...