Combining process splitting and merging transformations for Polyhedral Process Networks

We use the polyhedral process network (PPN) model of computation to program and map streaming media applications onto embedded Multi-Processor Systems on Chip (MPSoCs) platforms. In previous works, it has been shown how to apply different process network transformations in isolation. In this work, we present a holistic approach combining the process splitting and merging transformations and show that it is necessary to use both transformations in combination to achieve the best performance results, which cannot be achieved using only one transformation. We solve the problem of ordering both transformation and, in addition, relieve the designer from the task to select the processes on which the transformation should be applied. Thus, our approach combines both transformations exploiting the data-level parallelism available in a PPN as much as possible, even in cases where the parallelism is restricted by topological cycles and stateful processes in the PPN.