Custom instruction search for application specific instruction-set processor using guided simulated annealing

Abstract

This paper presents an algorithm for searching for Custom Instructions (CI) for Application specific Instruction-set processors (ASIPs) using Simulated Annealing algorithm (SA). A key feature of ASIP design is the ability of defining application specific CI, which are assisting HW modules in the EX phase of the processor that handles patterns of instructions that if handled by HW can improve the performance of the processor. In a previous paper [4] we presented a framework for searching and synthesizing CIs using tree based search algorithm, but this algorithm although giving high improvement in runtime with tight values of I/O ports, the runtime increases exponentially when relaxing the restrictions on those variables. In this paper we propose a guided form of SA algorithm based on the CI characteristics which can avoid to certain extend scanning through the invalid search space and using properties of CIs like convexity and data flow to guide the SA algorithm to take smarter decisions on how to scan the search space. Results shown improvement in runtime, especially with relaxed restrictions while providing optimum or near optimum results.