Buffer memory optimization in DSP applications -- an evolutionary approach

In the context of digital signal processing, synchronous data flow (SDF) graphs [12] are widely used for specification. For these, so called single appearance schedules provide program memory-optimal uniprocessor implementations. Here, buffer memory minimized schedules are explored among these using an Evolutionary Algorithm (EA). Whereas for a restricted class of graphs, there exist optimal polynomial algorithms, these are not exact and may provide poor results when applied to arbitrary, i.e., randomly generated graphs. We show that a careful EA implementation may outperform these algorithms by sometimes orders of magnitude