Memory Optimizations of Embedded Applications for Energy Efficiency 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

The current embedded processors often do not satisfy increasingly demanding computation requirements of embedded applications within acceptable energy efficiency, whereas application-specific integrated circuits require excessive design costs. In the Stanford Elm project, it was identified that instruction and data delivery, not computation, dominate the energy consumption of embedded processors. Consequently, the energy efficiency of delivering instructions and data must be sufficiently improved to close the efficiency gap between application-specific integrated circuits and programmable embedded processors. This dissertation demonstrates that the compiler and run-time system can play a crucial role in improving the energy efficiency of delivering instructions and data. Regarding instruction delivery, I present a compiler algorithm that manages L0 instruction scratch-pad memories that reside between processor cores and L1 caches. Despite the lack of tags, the scratch-pad memories with our algorithm can achieve lower miss rates than caches with the same capacities, saving significant instruction delivery energy. Regarding data delivery, I present methods that minimize memory-space requirements for parallelizing stream applications, applications that are commonly found in the embedded domain. When stream applications are parallelized in pipelining, large enough buffers are required between pipeline stages to sustain the throughput (e.g., double buffering). For static stream applications where production and consumption rates of stages are close to compile-time constants, a compiler analysis is presented, which computes the minimum buffer capacity that maximizes the throughput. Based on this analysis, a new static streamscheduling algorithm is developed, which yields considerable speed-up and data delivery energy saving compared to a previous algorithm. For dynamic stream applications, I present a dynamically-sized array-based queue design that achieves speed-up and data delivery energy saving compared to a linked-list based queue design.