Small: Collaborative Research: Transform-to-Perform: Languages, Algorithms, and Code Transformations for High-Performance FEM

Baylor University, Waco TX

Investigators

Abstract

Simulation of natural and engineering phenomena is a multi-layered technical task with high demands on mathematical sophistication and computational power. Producing a computer simulation code requires work at many different levels of detail. A computer program should represent a scientific problem in a language close to that used by domain specialists, but this differs greatly from low-level, hardware-specific details of computers. Bridging between the two requires several different links. This project sets up many intermediate software stages, called ?representations? modeling the domain knowledge of engineers, numerical analysts, and computer scientists by describing partial differential equations, the so-called weak forms needed for numerical methods, loop nests required to build discrete operations, and finally low-level code that can be executed by computers. ?Transformations? are then programs connecting these representations, injecting knowledge about algorithms and hardware. The key advance in this research is that, through this chain of transformations, domain knowledge about each level of detail, be it application-related, numerical, or computational, can be supplied at the appropriate level of detail. The tools developed in this project promote the advancement of science by both shortening the development time and increasing the resulting power of high-performance simulation codes used by scientists and engineers, enabling them to impact the world.