Interactive program improvement via EAVE: An expert adviser for vectorization

Abstract

An example of a state-of-the-art high-end mainframe is the IBM 3090 VF; its associated compiler, the IBM VS Fortran (version 2) compiler, incorporates some of the latest techniques in automatic vectorization and code optimization. Advances in compiler technology not withstanding, a potential limitation is the "knowledge gap" which exists between the typical end user and the compiler/machine sub-system. In particular, the user often does not know how to write source code which will result in generation of efficient, high performance object code. In this paper, we address the issue of providing interactive aids for program development and transformation as a means to bridge this knowledge gap. We show how heuristic program changes, guided by the knowledge of a particular compiler-machine pair, help approach achieveable peak performance through enhanced vectorization/parallelization. We present our techniques and results in terms of an implemented expert system, called EAVF, which has been designed to help users tune their programs for enhanced performance on the IBM 3090 VF.