Section: New Results

Pipeline-aware Scheduling of Polyhedral Process Networks

Participants : Christophe Alias, Julien Rudeau.

The polyhedral model is a well known framework to develop accurate and optimal automatic parallelizers for high-performance computing kernels. It is progressively migrating to high-level synthesis through polyhedral process networks (PPN), a dataflow model of computation which serves as intermediate representation for high-level synthesis. Many locks must be overcome before having a fully working polyhedral HLS tool, both from a front-end (C $\to$ PPN) and back-end (PPN $\to$ FPGA) perspective. In this work [15], we propose a front-end scheduling algorithm which reorganizes the computation of processes to maximize the pipeline efficiency of the processes' arithmetic operators. We show that our approach improve significantly the overall latency as well as the pipeline efficiency.