MAGELLAN: Multiway hardware-software partitioning and scheduling for latency minimization of hierarchical control-dataflow task graphs

The paper presents MAGELLAN, a heuristic technique for mapping hierarchical control-dataflow task graph specifications on heterogeneous architecture templates. The architecture can consist of multiple hardware and software processing elements as specified by the user. The objective of the technique is to minimize the worst case latency of the task graph subject to the area constraints on the architecture. The technique uses an iterative approach consisting of closely linked hardware-software partitioner and scheduler. Both the partitioner and scheduler operate on the task graph in a hierarchical top down manner. The technique optimizes deterministic loop constructs by applying clustering, unrolling and pipelining. The technique considers speculative execution for conditional constructs. The number of actual hardware/software implementations of a function in the task graph are also optimized by the technique. The effectiveness of the technique is demonstrated by a case study of an image compression algori thm.