Architecture for transparent binary acceleration of loops with memory accesses

This paper presents an extension to a hardware/software system architecture in which repetitive instruction traces, called Megablocks, Reconfigurable Processing Unit (RPU). This scheme is supported by a custom toolchain able to automatically generate a RPU tailored for the execution of one or more Megablocks detected offline. Switching between hardware and software execution is done transparently, without modifications to source code or executable binaries. Our approach has been evaluated using an architecture with a MicroBlaze General Purpose Processor (GPP) softcore. By using a memory sharing mechanism, the RPU can access the GPP's data memory, allowing the acceleration of Megablocks with load/store operations. For a set of 21 embedded benchmarks, an average speedup of 1.43× is achieved, and a potential speedup of 2.09× is predicted for an implementation using a low overhead interface for communication between GPP and RPU.