Untitled

Abstract

Loop unrolling is a well-established code transformation technique that can improve the performance of a program at runtime. The key benefit of unrolling a loop is that it often requires fewer instruction executions than the original loop. However, determining the optimal number of loop unrolling is a critical concern. This paper presents a novel method for predicting the optimal unroll factor for a given program. Specifically, a dataset is constructed that includes the execution times of several programs with varying loop unroll factors. The programs are sourced from different benchmarks, such as Ploybench, Shooutout, and other programs. Similarity measures between the unseen program and the existing programs are computed, and the three most similar programs are identified. The unroll factor that led to the greatest reduction in execution time for the most similar programs is selected as the candidate for the unseen program. Experimental results demonstrate that the proposed method can enhance the performance of training programs for unroll factors of 2, 4, 6, and 8 by approximately 13%, 18%, 19%, and 21%, respectively. For the unseen programs, the speedup rate is approximately 37.7% for five programs.