Scalarization on Short Vector Machines

Abstract

Scalarization is a process that converts array statements into loop nests so that they can run on a scalar machine. One technical difficulty of scalarization is that temporary storage often needs to be allocated in order to preserve the semantics of array syntax - "fetch before store". Many techniques have been developed to reduce the size of temporary storage requirement in order to improve the memory hierarchy performance. With the emergence of short vector units on modern microprocessors, it is interesting to see how to extend the preexisting scalarization methods so that the underlying vector infrastructure is fully utilized, while at the same time keep the temporary storage minimized. In this paper, we extend a loop alignment algorithm for scalarization on short vector machines. The revised algorithm not only achieves vector execution with minimum temporary storage, but also handles data alignment properly, which is very important for performance. Our experiments on two types of widely available architectures demonstrate the effectiveness of our strategy