Designing domain-specific processors

Abstract

We present a semi-automated method for the detection and exploitation of application domain specific instruction set extensions for embedded (VLIW) processors. It consists of three steps: the first step detects frequently occurring operation patterns, in the second step, the patterns are grouped and implemented in a number of Special Function Units (SFUs) and the third step incorporates the custom operations into the code generation process. Experiments show that the SFUs generated and exploited with our methodology can result in architectures that perform up to 30% better than architectures of the same cost without SFUs.