Power efficiency through application-specific instruction memory transformations

The instruction memory communication path constitutes a significant amount of power consumption in embedded processors. We propose an encoding technique that exploits application information to reduce the associated power consumption. The microarchitectural support enables reprogrammability of the encoding transformations so as to track code particularities effectively. The restriction to functional transformations enables effective coding while delivering major power savings, in the process obviating furthermore the necessity to rely on dictionary lookup, one of the major shortcomings of prior approaches. The frugal functional transformation, reliant on a single bit logic gate, introduces no impact to the critical fetch stage of the processor pipeline while delivering fully all the theoretically achievable power savings. The reprogrammable hardware implementation enables flexible and inexpensive switches between the transformations. Extensive experimental results on numerical and DSP codes confirm the theoretically expected magnitude of power savings, evincing reductions that range up to half of the original transitions.