Adaptive mapping to resource availability for dynamic wavelet-based applications

Abstract

Platforms have to cope with unpredictably varying system resource requirements, because of inter-task level dynamism. To deal with this, they have to be at least partially reconfigurable. It is then important for applications to optimally exploit the memory hierarchy under varying memory availability. Moreover, in the case of intra-task dynamism, additional unpredictability is inserted and the exploration of the optimal memory hierarchy depends on data dependent application behavior. This paper presents a mapping strategy for a wavelet-based video application: depending on the encountered resource availability, switching to different memory optimized instantiations (i.e. localizations) of the application offers up to 25% energy gains in memory accesses, for a representative test sequence. We observe that it is possible to exploit the input motion characteristics in detail (that reflect the intra-task dynamism) by enabling motion-triggered switching, and further increase the achieved gains.