Technique of vectorial loop peeling using cost analysis
HE Yue-qiang · Jisuanji gongcheng yu sheji · 2011
To vectorize a loop more sufficiently in the compile time,a loop peeling technique for vectorization based on a cost analysis is presented.Firstly,the fragments that do not have dependency are marked in the loop.Then a simple and cost effective model is designed to evaluate the CPU clock cycles in the vectorized or non-vectorized loop fragments.Finally,the final results from the cost analysis determined whether to vectorize the fragments or not,thus the vectorization is applied to short loop fragments.The experimental results demonstrate effectiveness of the technique and show an obvious optimization in loops have a long dependence distance.