Research On Streamlization Feature And Streamlizing Method Of Application

As a novel high performance microprocessor, stream processors are widely applied in media processing, signal processing, scientific computation, etc. As new video encoding standard and high definition image take much more computation , processors need to be improved. It is a valuable subject to study how to use the stream processing technique efficiently.This paper introduces the background at the beginning, trying to find out the essence of the highly efficient stream processing. We put forward a method to analyze streamlization features and the corresponding metrics (inter-kernel metrics and intra-kernel metrics). Then we validate the metrics by comparing the features with the performance of a set of typical benchmark. We proved that the metrics can reflect the performance of the stream application.We are encountered with a lot of problems during implementing stream program. This paper summarizes the methods of streamlization and optimization of the typical structure rather than discussing every possible application. Instead of being designed for some certain platforms, these methods can be commonly applied.At last, we apply the methods to analyze H264 and YGX2 which represents the media processing and scientific computation respectively. It shows that the metrics of H264 perform excellent, while the metrics of YGX2 are not satisfactory, especially poor in inter-kernel metrics. We can get a good performance H264 but a bad YGX2 after streamlizing them. By analyzing NPB, we find there are a lot of small loops in it, which will affect the performance deeply.