Hierarchy-Aware mapping of pipelined applications

This paper presents hierarchy-aware mapping of a pipelined application to a given heterogeneous platform. The applications we target are executed in a pipelined manner and can be captured with a Model of Computation (MoC) with explicit representations of pipelined execution. Our optimization goal is to minimize execution time. It is crucial to balance the pipeline stages in the given MoC, where each stage consists of a set of hierarchical tasks, each is decomposed into either over which state transitions depending on data are defined or which run in parallel. Previous works have not considered such complex hierarchy in a general MoC; They assumed that each iteration of the execution of the application is represented with an acyclic directed graph and that a stage can be decomposed only into one or more parallel tasks. Therefore, the tasks tended to be coarse-grained since a task with complex hierarchy can be hardly partitioned more. Our approach partition each hierarchically described task and may map the partition separately to balance the stages. The case study is performed with randomly generated platforms and streaming applications: Canny Edge Detector and JPEG. We compare our Hierarchy-Aware mapping to the hierarchy-unaware mapping found through an exhaustive search. Hierarchy-Aware mapping decreases execution time on average by 23.3%.