Efficient Embedded Computing a Dissertation Submitted to the Department of Electrical Engineering and the Committee on Graduate Studies of Stanford University in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Embedded computer systems are ubiquitous, and contemporary embedded applications exhibit demanding computation and efficiency requirements. Meeting these demands presently requires the use of applicationspecific integrated circuits and collections of complex system-on-chip components. The design and implementation of application-specific integrated circuits is both expensive and time consuming. Most of the effort and expense arises from the non-recurring engineering activities required to manually lower high-level descriptions of systems to equivalent low-level descriptions that are better suited to hardware realization. Programmable systems, particularly those that can be targeted effectively using high-level programming languages, offer reduced development costs and faster design times. They also offer the flexibility required to upgrade previously deployed systems as new standards and applications are developed. However, programmable systems are less efficient than fixed-function hardware. This significantly limits the class of applications for which programamble processors are an acceptable alternatives to application-specific fixedfunction hardware, as efficiency demands often preclude the use of programmable hardware. With most contemporary computer systems limited by efficiency, improving the efficiency of programmable systems is a critical challenge and an active area of computer systems research. This dissertation describes Elm, an efficient programmable system for high-performance embedded applications. Elm is significantly more efficient than conventional embedded processors on compute-intensive kernels. Elm allows software to exploit parallelism to achieve performance while managing locality to achieve efficiency. Elm implements a novel distributed and hierarchical system organization that allows software to exploit the abundant parallelism, reuse, and locality that are present in embedded applications. Elm provides a variety of mechanisms to assist software in mapping applications efficiently to massively parallel systems. To improve efficiency, Elm allows software to explicitly schedule and orchestrate the movement and placement of instructions and data. This dissertation proposes and critically analyzes concepts that encompass the interaction of computer architecture, compiler technology, and VLSI circuits to increase performance and efficiency in modern embedded computer systems. A central theme of this dissertation is that the efficiency of programmable embedded systems can be improved significantly by exposing deep and distributed storage hierarchies to software. This allows software to exploit temporal and spatial reuse and locality at multiple levels in applications in order to reduce instruction and data movement.