Microcode generation for the control of a massively parallel computer

Abstract

Software organization for the control of a single-instruction multiple-data (SIMD) machine is explained. It is argued that the bulk of an algorithm adapted to a SIMD architecture should be executed on a microsequencing controller which is directly connected to the SIMD machine. Typical systems have a main control unit which calls low-level primitives on a controller which is directly connected to an SIMD machine. The results show that an algorithm can be autonomously run on a microsequencer directly connected to an SIMD machine. Two algorithms are applied to the two major SIMD controller configurations and their performances are examined. To demonstrate how the software organization can be applied to a system, a compiler is written for the GAM pyramid that generates microcode from a high-level language with parallel data structures. The microcode executes on a microcode sequencer unit which is directly connected to a pyramid architecture SIMD machine. >