Reconfigurable framework for high-bandwidth stream-oriented data processing

Alexander Mykyta, Dorin Patru, Eli Saber, Gene Roylance, Brad Larson
2012 2012 IEEE International SOC Conference  
ii Acknowledgements I would like to thank: Dr. Dorin Patru for giving me the opportunity to be a part of this research and for guidance throughout the thesis process; Brad Larson and Gene Roylance for their insight and continued support; Ryan Toukatly for his research efforts which lay the technical groundwork for this project; Abstract Designing a digital system that implements an assortment of specialized highperformance algorithms can be costly. Considerable non-recurring engineering costs
more » ... e required to develop an application specific integrated circuit (ASIC). Additionally, updating or adding features to a design requires the ASIC to be redesigned and refabricated. An alternative to using an ASIC is the field programmable gate array (FPGA). The modern FPGA's ability to be partially reconfigured at runtime allows for the device to have the flexibility normally associated with a processor, while also being able to implement digital logic like in an ASIC. This capability allows for multiple digital functions to be loaded into the device at runtime only as needed. This thesis focuses on developing a reconfigurable framework that enables stream-oriented applications to make more effective use of FPGA resources and to manage partial reconfiguration operations across multiple tasks. This multichannel framework addresses several shortcomings of past research that evaluated various dynamic partial reconfiguration techniques using a color space conversion (CSC) engine. This framework allows for multiple different computations to be performed simultaneously, further improving throughput and flexibility of applications implemented within it. Performance of the system is evaluated by comparing its computational throughput to previous efforts using the CSC engine as well as the performance gained from the flexible scheduling that the framework allows. Implementations using the CSC engine show that performance can be improved up to 5 times faster than previous works, as a result of exploiting parallelism. iv
doi:10.1109/socc.2012.6398391 dblp:conf/socc/MykytaPSRL12 fatcat:yasut6izkzbhzmjiu57tcjnxpa