A Performance-Oriented Algorithm with Consideration on Communication Cost for Dynamically Reconfigurable FPGA Partitioning

Abstract

Dynamically reconfigurable FPGAs (DRFPGAs) have high logic utilization because of time-multiplexed interconnects and logic. In this article, we propose a performance-oriented algorithm for the DRFPGA partitioning problem. This algorithm partitions a given circuit system into stages such that the upper bound of the execution times of subcircuits is minimized. The communication cost is taken into consideration in the process of searching for the optimal solution. A graph is first constructed to represent the precedence constraints and calculate the number of buffers needed in a partitioning. This algorithm includes three phases. The first phase reduces the problem size by clustering the gates into subsystems that have only one output. Such a subsystem has a large number of intraconnections because the fan-outs of all vertices except for the one output are fed to the vertices inside the subsystem. This phase significantly reduces the computational complexity of partitioning. The second phase finds a partition with optimal performance. Finally, the third phase minimizes the communication cost by using an iterative improvement approach. Experimental results based on the Xilinx architecture show that our algorithm yields better partitioning solutions than traditional approaches.