Strategies for improving data locality in embedded applications

This paper introduces a dynamic layout optimization strategy to minimize the number of cycles spent in memory accesses in a cache-based memory environment. In this approach, a given multi-dimensional array may have different memory layouts in different segments of the same application if doing so improves data locality (cache behavior) beyond the static approaches that fix memory layouts at specific forms at compile-time. An important characteristic of this strategy is that different memory layouts that a given array will assume at run-time are determined statically at compile-time; however the layout modifications (transformations), themselves, occur dynamically during the course of execution. To test the effectiveness of our strategy, we used it in optimizing several array-dominated applications. Our preliminary results on an embedded MIPS processor core show that this dynamic strategy is very successful and outperforms previous approaches based on loop transformations, data transformations, or integrated loop/data transformations.