Title :
On combining temporal partitioning and sharing of functional units in compilation for reconfigurable architectures
Author :
Cardoso, João M P
Author_Institution :
Fac. of Sci. & Technol., Univ. do Algarve, Faro, Portugal
Abstract :
Resource virtualization on FPGA devices, achievable due to its dynamic reconfiguration capabilities, provides an attractive solution to save silicon area. Architectural synthesis for dynamically reconfigurable FPGA-based digital systems needs to consider the case of reducing the number of temporal partitions (reconfigurations) by enabling sharing of some functional units in the same temporal partition. This paper proposes a novel algorithm for automated datapath design from behavioral input descriptions (represented by an acyclic dataflow graph), which simultaneously performs temporal partitioning and sharing of functional units. The proposed algorithm attempts to minimize both the number of temporal partitions and the execution latency of the generated solution. Temporal partitioning, resource sharing, scheduling, and a simple form of allocation and binding are all integrated in a single task. The algorithm is based on heuristics and on a new concept of construction by gradually enlarging timing slots. Results show the efficiency and effectiveness of the algorithm when compared to existent approaches.
Keywords :
field programmable gate arrays; processor scheduling; reconfigurable architectures; resource allocation; allocation; automated datapath design; behavioral input descriptions; binding; compilation; dynamic reconfiguration; dynamically reconfigurable FPGA-based digital systems; execution latency; functional unit sharing; heuristics; reconfigurable architectures; resource sharing; resource virtualization; scheduling; temporal partitioning; timing slots; Algorithm design and analysis; Delay; Digital systems; Field programmable gate arrays; Partitioning algorithms; Reconfigurable architectures; Resource management; Resource virtualization; Silicon; Timing;
Journal_Title :
Computers, IEEE Transactions on
DOI :
10.1109/TC.2003.1234532