Hazard-Aware Performance Prediction for Automatic Instruction-Set Selection

Abstract

Recent research in the area of application specific instruction set processors (ASIPs) has focused on the automatic selection of a custom instruction set based on a high level description of the application. Existing methods perform instruction selection under the assumption that data hazards can be ignored due to functional unit forwarding. This paper addresses data hazards in the ASIP flow when functional unit to functional unit forwarding is too expensive. This is accomplished by devising a "hazard-aware" predictor for measuring the impact of custom instructions on performance. Results show that our predictor reduces prediction error from 50% to 15% compared to the existing simple predictor and with a fraction of the run-time of rescheduling. When incorporated into an instruction enumeration and selection algorithm, our predictor reduces the total schedule length by as much as 8.4%