Experimental Evaluation of Memory Optimizations on an Embedded GPU Platform

This paper presents the experimental evaluation of data mapping techniques in the shared memory of an embedded GPU. The evaluated technique, previously presented in the literature in other contexts, aims at partitioning an array across the shared memory physical banks, so as to increase parallel accesses, resulting in appreciable gains in terms of both performance and energy efficiency. The paper presents the experimental setup used for characterizing physically the behaviour of the platform, allowing a validation and a closer understanding of the evaluated memory mapping technique.