Minimization of memory access overhead for multidimensional DSPapplications via multilevel partitioning and scheduling

摘要

Massive uniform nested loops are broadly used in multidimensionalndigital signal processing (DSP) applications. Due to the large amount ofndata handled by such applications, the optimization of data accesses bynfully utilizing the local memory and minimizing communication overheadnis important in order to improve the overall system performance. Most ofnthe traditional partition strategies do not consider the effect of datanaccess on the computational performance. In this paper, a multilevelnpartitioning method, based on a static data scheduling technique knownnas carrot-hole data scheduling, is proposed to control the data trafficnbetween different levels of memory. Based on this data schedule, optimalnpartition vector, scheduling vector and the partition size are chosen innsuch a way to minimize communication overhead. Nonhomogeneous sizenpartitions are the final result of the partition scheme which produces ansignificant performance improvement. Experiments show that by using thisntechnique, local memory misses are significantly reduced as compared tonresults obtained from traditional methods. This method can be used innapplication specific DSP system design and compiler for DSP processors