Achieving a better balance between productivity and performance on FPGAs through Heterogeneous Extensible Multiprocessor Systems

Field Programmable Gate Arrays (FPGAs) were first introduced circa 1980, and they held the promise of delivering performance levels associated with customized circuits, but with productivity levels more closely associated with software development. Achieving both performance and productivity objectives has been a long standing challenge problem for the reconfigurable computing community and remains unsolved today. On one hand, Vendor supplied design flows have tended towards achieving the high levels of performance through gate level customization, but at the cost of very low productivity. On the other hand, FPGA densities are following Moore\u27s law and and can now support complete multiprocessor system architectures. Thus FPGAs can be turned into an architecture with programmable processors which brings productivity but sacrifices the peak performance advantages of custom circuits. In this thesis we explore how the two use cases can be combined to achieve the best from both. The flexibility of the FPGAs to host a heterogeneous multiprocessor system with different types of programmable processors and custom accelerators allows the software developers to design a platform that matches the unique performance needs of their application. However, currently no automated approaches are publicly available to create such heterogeneous architectures as well as the software support for these platforms. Creating base architectures, configuring multiple tool chains, and repetitive engineering design efforts can and should be automated. This thesis introduces Heterogeneous Extensible Multiprocessor System (HEMPS) template approach which allows an FPGA to be programmed with productivity levels close to those associated with parallel processing, and with performance levels close to those associated with customized circuits. The work in this thesis introduces an ArchGen script to automate the generation of HEMPS systems as well as a library of portable and self tuning polymorphic functions. These tools will abstract away the HW/SW co-design details and provide a transparent programming language to capture different levels of parallelisms, without sacrificing productivity or portability