FPGA high-level synthesis and virtualization

点击翻译按钮获取中文摘要

There are two ways to perform computations: in software or in hardware. The software approach is widely used and the skills are readily available. However, the use of software running on a processor imposes significant overheads in terms of speed and energy efficiency, owing to the need to fetch and decode instructions, store results in memory, and so on. Hardware, on the other hand, can bring orders of magnitude improvement in speed and energy efficiency to computing. However, hardware design requires specialized skills, is time intensive, and error prone. The overarching aim of the proposed research is to bring the speed and energy benefits of hardware to those having solely software skills.** High-level synthesis (HLS) refers to the automated synthesis of a software program into a hardware circuit that performs the same function as the software. The faculty applicant leads the LegUp high-level synthesis project (http://legup.eecg.toronto.edu) at the University of Toronto, which automatically translates a C language program into a hardware circuit. The present grant aims to improve the quality (performance, power, and cost) of the automatically generated hardware to the point of being on par with human-expert-crafted hardware. A second HLS direction concerns defining a standard methodology for comparing one HLS tool with another. A third direction pertains to synthesizing hardware directly from an executable binary, rather than source code.** The second research package in the grant relates to "virtualization" of field-programmable gate arrays (FPGAs), which are specialized computer chips whose function can be set by the end user. FPGAs are widely used to implement custom computing accelerators, where, in conjunction with standard processors, can be leveraged to improve computational throughput and energy efficiency. FPGA virtualization seeks to make FPGAs easily useable within cloud/datacentre environments, where they can be flexibly used for a wide variety of computing tasks.