Model-guided strip size selection for minimal execution time on imagine stream processor
Jing Du, Yuhua Tang, Fujiang Ao, Tao Tang, Xuejun Yang · 2008
Strip-mining is a critical optimization for improving the effectiveness of memory hierarchy of Imagine. In this paper, we present an efficient compiler algorithm for selecting the optimal strip size to minimize the execution time of stream programs. First, we build a graceful analytical model that characterizes the effect of strip size on key performance factors. Then, we design a novel algorithm for selecting optimal strip size according to the model analysis and apply it to some stream programs. Furthermore, we implement the algorithm in the stream compiler. The experimental results show that when the algorithm is used, the execution time is close to the experimentally best. It is certain that our algorithm can efficiently exploit the tremendous potential of Imagine.