Developments and experimental evaluation of partitioning algorithms for adaptive computing systems

Multi-FPGA systems offer the potential to deliver higher performance solutions than traditional computers for some low-level computing tasks. This requires a flexible hardware substrate and an automated mapping system. CHAMPION is an automated mapping system for implementing image processing applications in multi-FPGA systems under development at the University of Tennessee. CHAMPION will map applications in the Khoros Cantata graphical programming environment to hardware. The work described in this dissertation involves the automation of the CHAMPION backend design flow, which includes the partitioning problem, netlist to structural VHDL conversion, synthesis and placement and routing, and host code generation. The primary goal is to investigate the development and evaluation of three different k-way partitioning approaches. In the first and the second approaches, we discuss the development and implementation of two existing algorithms. The first approach is a hierarchical partitioning method based on topological ordering (HP). The second approach is a recursive algorithm based on the Fiduccia and Mattheyses bipartitioning heuristic (RP). We extend these algorithms to handle the multiple constraints imposed by adaptive computing systems. We also introduce a new recursive partitioning method based on topological ordering and levelization (RPL). In addition to handling the partitioning constraints, the new approach efficiently addresses the problem of minimizing the number of FPGAs used and the amount of computation, thereby overcoming some of the weaknesses of the HP and RP algorithms.