A two-level scheduling method: an effective parallelizing technique for uniform nested loops on a DSP multiprocessor

摘要

A digital signal processor (DSP), which is a special-purpose microprocessor, is designed to achieve higher performance on DSP applications. Because most DSP applications contain many nested loops and permit a very high degree of parallelism, the DSP multiprocessor has a suitable architecture to execute these applications. Unfortunately, conventional scheduling methods used on DSP multiprocessors allocate only one operation to each DSP every time unit, even if the DSP includes several function units that can operate in parallel. Obviously they cannot achieve full function unit utilization. Hence, in this paper, we propose a two-level scheduling method (TSM) to overcome this common failing. TSM contains two approaches, which integrates unimodular transformations, loop tiling technique, and conventional methods used on single DSP. Besides introducing algorithm, we also use an analytic module to analyze its preliminary performance. Based on our analyses the TSM can achieve shorter execution time and more scalable speedup results. In addition, the TSM causes less memory access and synchronization overheads, which are usually negligible in the DSP multiprocessor architecture.