Memory-efficient Data Structures for Systolic Programs

Generating imperative programs from systolic speciications implies the use of a memory-eecient model for representing data used in the original, single-assignment problem. We present a new method of generating data structures , designed for use in systolic program compilers for distributed-memory parallel computers. Our method is based on the mathematical properties of systolic programs and consists of a set of algebraically deened elementary transformations. Using these transformations, the total size of data structures can be optimized, allowing very memory-eecient code to be produced. The method has been implemented as a part of an experimental compiler of the Alpha language, currently under development. Generating memory-eecient imperative data structures from systolic programs R esum e : Le g en eration de programmes imp eratifs a partir de sp eciications systoliques exige un mod ele de m emorisation eecace pour repr esenter les don-n ees utilis ees dans la version a assignation unique du programme source. Nous d ecrivons une nouvelle m ethode pour la g en eration de structures de donn ees, destin ee a la compilation de programmes systoliques pour des architectures pa-rall eles a m emoire distribu ee. Notre m ethode est bas ee sur les propri et es math e-matiques des programmes systoliques, et consiste en un ensemble de transformations alg ebriques el ementaires. Avec ces transformations, la taille totale des structures de donn ees peut ^ etre optimis ee, ce qui permet de g en erer un code tr es eecace en taille m emoire. La m ethode a et e impl ement ee dans un compilateur exp erimental pour le language Alpha, actuellement en cours de d eveloppement. Mots-cl e : programmation parall ele, architectures systoliques, architectures a m emoire distribu ee Memory-eecient data structures for systolic programs