UC Riverside

As FPGAs become more common in mainstream general-purpose computing platforms, capturing and distributing high-performance implementations of applications on FPGAs will become increasingly important. Even in the presence of C-based synthesis tools for FPGAs, designers continue to implement applications as circuits, due in large part to allow for capture of clever spatial, circuit-level implementation features leading to superior performance and efficiency. We demonstrate the feasibility of a spatial form of FPGA application capture that offers portability advantages for FPGA applications unseen with current FPGA binary formats. We demonstrate the portability of such a distribution by developing a fast on-chip emulation framework that performs transparent optimizations, allowing spatially-captured FPGA applications to immediately run on FPGA platforms without costly and hard-to-use synthesis/mapping tool flows, and sometimes faster than PC-based execution. We develop several dynamic and transparent optimization techniques, including just-in-time compilation, bytecode acceleration, and just-in-time synthesis that take advantage of a platform's available resources, resulting in orders of magnitude performance improvement over normal emulation techniques and PC-based execution.