Load-store reordering for low-power multimedia data transfers

We present a load-store reordering technique for low-power multimedia data transfers. The proposed technique is based on two common characteristics of multimedia applications: (1) many multimedia streams have a strong spatial locality and (2) output data generated by many multimedia operations are significantly different from the input data. In this paper, we propose a compiler-level technique which combines loop unrolling and load-store scheduling to minimize both the self and coupled transition power on a data bus. Experimental results show that the total amount of the transition activities is reduced on average by 21.3% over the original code for many image processing applications.