Language Support for Pipelining Wavefront Computations

Wavefront computations, characterized by a data dependent flow of computation across a data space, are receiving increasing attention as an important class of parallel computations. Though sophisticated compiler optimizations can often produce efficient pipelined implementations from sequential representations, we argue that a language-based approach to representing wavefront computations is a more practical technique. A language-based approach is simple for the programmer yet unambiguously parallel. In this paper we introduce simple array language extensions that directly support wavefront computations. We show how a programmer may reason about the extensions’ legality and performance; we describe their implementation and give performance data demonstrating the importance of parallelizing these codes.

• Type: Conference
• Year: 1999
• Where: LCPC
• Authors: Bradford L. Chamberlain, E. Christopher Lewis, Lawrence Snyder
• Keywords: Language-based Approach, LCPC 1999, Parallel Computations, System Software, Wavefront Computations