Rapid Estimation of Energy Consumption for Embedded Software

Sammanfattning: Energy consumption is an important parameter very early in the design phase for embedded systems, especially battery powered systems. To obtain a cycle-accurate estimation of a program’s energy consumption, the program must be compiled and simulated on the target architecture, but this requires a hardware specification and complete code which may not be available early in the design phase. Simulation is computationally and time intensive, and simulation must be done with proprietary tools for some hardware platforms. Previous work has investigated instruction-level power models for embedded systems, and other work has extended this to source code power analysis tools for RISC processors with near constant instantaneous power consumption per instruction, but no work has examined the effectiveness of source code estimation techniques on modern digital signal processing units with non-constant power consumption which varies per instruction. This thesis proposes and evaluates a technique for source code estimation of power and energy consumption on an Analog Devices Blackfin BF70x DSP, and compares this technique to other power estimation techniques realizable at different stages of the design cycle. The proposed method is also made available as an extension for the pure-python, open source Lizard code analysis tool, and the method can easily be extended for other processors. This Lizard based source code power estimation tool is directed at the use case in which machine generated programs must be evaluated for energy efficiency, and excels in the case where: 1. Speed of estimation is more important than absolute accuracy 2. Not all source code dependencies are available (compilation may not be possible at this stage)