Energy-Aware Compiler Optimization for VLIW-DSP Cores

Abstract

VLIW-DSP processor cores are widely used in embedded SoCs. Improving the energy efficiency becomes one of the key issues in designing a VLIW-DSP core. This paper proposes compiler optimization algorithms to reduce the register file power in a VLIW-DSP processor. The optimization is targeted to VLIW processors in which each execution slot is associated with a low-powered local register file. Instruction scheduling and register allocation algorithms are proposed to direct operand accesses to the local register files. We propose energy-aware list scheduling algorithm to reduce cross-slot data dependencies without affecting the program execution time. Constrained by the instruction scheduling result, energy-aware register allocation is performed through weighted graph coloring. Evaluation with MiBench benchmark suite shows that our approach reduces over 50% of data transfer energy with low hardware cost. This research shows a cost-effective way to design an energy-efficient VLIW- DSP processor.Keywordsenergy-aware instruction schedulingweighted graph coloringregister allocationVLIW-DSP processor