Estimating potential parallelism for platform retargeting

Scientific, symbolic, and multimedia applications present diverse computing workloads with different types of inherent parallelism. Tomorrow´s processors will employ varying combinations of parallel execution mechanisms to efficiently harness this parallelism. The explosion of consumer products that incorporate high performance embedded computing will increase the stratification of the processor design space. However, existing code assets are limited to sequential expression of what should be highly parallel algorithms. Retargeting to parallel mechanisms is difficult, but can provide significant increases in efficiency. It is desirable to estimate potential parallelism before undertaking the expensive process of reverse engineering and retargeting. This paper presents a lightweight dynamic analysis technique for characterizing the types of parallelism that are inherent in a given program to estimate the potential benefit of retargeting. The technique is validated on Spec95 and MediaBench benchmarks widely used to evaluate processor performance. Results correlate well with previous experience in parallelizing these well-understood applications.