Software-Controlled Scratchpad Mapping Strategies for Wavelet-Based Applications

Abstract

A comparison is performed between a simple hardware-controlled cache mapping and a systematic, software-controlled scratchpad mapping of a two-dimensional, lifting-based wavelet transform, in the context of dynamic applications. In this context it is important for the application to optimally exploit the memory hierarchy under the dynamically varying memory availability. Contrary to expectations, the comparison shows the cache-based mapping to exhibit a lower miss rate. The reasons for these different results are examined and improvements for the systematic scratchpad mapping are proposed which would give it a 30% lower missrate performance for localized wavelet-based applications