Power and memory optimization techniques in embedded systems design

Abstract

Embedded systems incur tight constraints on power consumption and memory (which impacts size) in addition to other constraints such as weight and cost. This dissertation addresses two key factors in embedded system design, namely minimization of power consumption and memory requirement. The first part of this dissertation considers the problem of optimizing power consumption (peak power as well as average power) in high-level synthesis (HLS). The second part deals with memory usage optimization mainly targeting a restricted class of computations expressed as loops accessing large data arrays that arises in scientific computing such as the coupled cluster and configuration interaction methods in quantum chemistry. First, a mixed-integer linear programming (MILP) formulation is presented for the scheduling problem in HLS using multiple supply-voltages in order to optimize peak power as well as average power and energy consumptions. For large designs, the MILP formulation may not be suitable; therefore, a two-phase iterative linear programming formulation and a power-resource-saving heuristic are presented to solve this problem. In addition, a new heuristic that uses an adaptation of the well-known force-directed scheduling heuristic is presented for the same problem. Next, this work considers the problem of module selection simultaneously with scheduling for minimizing peak and average power consumption. Then, the problem of power consumption (peak and average) in synchronous sequential designs is addressed. A solution integrating basic retiming and multiple-voltage scheduling (MVS) is proposed and evaluated. A two-stage algorithm namely power-oriented retiming followed by a MVS technique for peak and/or average power optimization is presented. Memory optimization is addressed next. Dynamic memory usage optimization during the evaluation of a special class of interdependent large data arrays is considered. Finally, this dissertation develops a novel integer-linear programming (ILP) formulation for static memory optimization using the well-known fusion technique by encoding of legality rules for loop fusion of a special class of loops using logical constraints over binary decision variables and a highly effective approximation of memory usage.