Combined Data Partitioning and Loop Nest Splitting for Energy Consumption Minimization

Abstract

For mobile embedded systems, the energy consumption is a limiting factor because of today’s battery capacities. Besides the processor, memory accesses consume a high amount of energy. The use of additional less power hungry memories like caches or scratchpads is thus common. This paper presents a combined approach for energy consumption minimization consisting of two complementary and phase-coupled optimizations, viz. data partitioning and loop nest splitting. In a first step, data partitioning partitions large arrays found in typical embedded software into smaller ones which are placed onto an on-chip scratchpad memory. Although being effective w. r. t. energy dissipation, this optimization adds overhead to the code since the correct part of a partitioned array has to be selected at runtime. Therefore, the control flow is optimized as a second step in our framework. In this phase, loop nests containing if-statements are split using genetic algorithms leading to minimized if-statement executions. However, loop nest splitting leads to an increase in code size and can potentially annul the program layout achieved by the first step. Consequently, the proposed approach iteratively applies these optimizations till a local optimum is found.

The proposed framework of combined memory and control flow optimization leads to considerable energy savings for a representative set of typical embedded software routines. Using an accurate energy model for the ARM7 processor, energy savings between 20.3% and 43.3% were measured.