Hardware virtualization on a coarse-grained reconfigurable processor

In this thesis, we propose to use a reconfigurable processor as main computation element in embedded systems for applications from the multi-media and communications domain. A reconfigurable processor integrates an embedded CPU core with a Reconfigurable Processing Unit (RPU). Many of our target applications require real-time signal-processing of data streams and expose a high computational demand. The key challenge in designing embedded systems for these applications is to find an implementation that satisfies the performance goals and is adaptable to new applications, while the system cost is minimized. Implementations that solely use an embedded CPU are likely to miss the performance goals. Application-Specific Integrated Circuit (ASIC)-based coprocessors can be used for some high-volume products with fixed functions, but fall short for systems with varying applications. We argue that a reconfigurable processor with a coarse-grained, dynamically reconfigurable array of modest size provides an attractive implementation platform for our application domain. The computational intensive application kernels are executed on the RPU, while the remaining parts of the application are executed on the CPU. Reconfigurable hardware allows for implementing application specific coprocessors with a high performance, while the function of the coprocessor can still be adapted due to the programmability. So far, reconfigurable technology is used in embedded systems primarily with static configurations, e.g., for implementing glue-logic, replacing ASICs, and for implementing fixed-function coprocessors. Changing the configuration at runtime enables a number of interesting application modes, e.g., on-demand loading of coprocessors and time-multiplexed execution of coprocessors, which is commonly denoted as hardware virtualization. While the use of static configurations is well understood and supported by design-tools, the role of dynamic reconfiguration is not well investigated yet. Current application specification methods and design-tools do not provide an end-to-end tool-flow that considers dynamic reconfiguration. A key idea of our approach is to reduce system cost by keeping the size of the reconfigurable array small and to use hardware virtualization techniques to compensate for the limited hardware resources. The main contribution of this thesis is the codesign of a reconfigurable processor architecture named ZIPPY, the corresponding hardware and software implementation tools, and an application specification model which explicitly considers hardware virtualization. The ZIPPY architecture is widely parametrized and allows for specifying a whole family of processor architectures. The implementation tools are also parametrized and can target any architectural variant. We evaluate the performance of the architecture with a system-level, cycle-accurate cosimulation framework. This framework enables us to perform design-space exploration for a variety of reconfigurable processor architectures. With two case studies, we demonstrate, that hardware virtualization on the Zippy architecture is feasible and enables us to trade-off performance for area in embedded systems. Finally, we present a novel method for optimal temporal partitioning of sequential circuits, which is an important form of hardware virtualization. The method based on Slowdown and Retiming allows us to decompose any sequential circuit into a number of smaller, communicating subcircuits that can be executed on a dynamically reconfigurable architecture.