OSS: Efficient Compiler Approach for Selecting Optimal Strip Size on the Imagine Stream Processor

Abstract

Strip-mining technique is critical for improving performance of large-scale scientific applications on Imagine. In this paper, we present a model-guided strip size selection approach (OSS) for finding the optimal strip size to minimize the execution time. Our strategy consists of a detailed analytical model that characterizes the effect of strip size on program behavior. Then according to the model analysis, we design a simple strip size selection strategy, that is, the optimal strip size is 512 words. Our experimental results show that when the optimal strip size is used, the execution time is close to the experimentally best. It is certain that our strategy can efficiently exploit the tremendous potential of Imagine.