Since the appearance of parallel processors and their rapid diversification across a broad spectrum, developers must phrase algorithms in a parallel manner using originally imperative and thus inappropriate high-level...
详细信息
ISBN:
(纸本)9781479986705
Since the appearance of parallel processors and their rapid diversification across a broad spectrum, developers must phrase algorithms in a parallel manner using originally imperative and thus inappropriate high-level languages. Language extensions as well as highly complex debugging methods (e.g., profilers) to handle concurrent and non-deterministic execution are therefore continuously developed. Most tools, however, suffer from inflexibility and platform dependencies. Moreover, binary-instrumenting profilers involve high overhead, influencing and thus deforming the runtime behavior. This may even hide critical behavior, thus developers still rely on their experience and often manually include measures in their software-code (in-line profiling). In this work, we propose a platform independent abstraction layer enabling a unified parallelization and runtime-flexible choice of the actual parallelization framework (e.g., OpenMP, TBB). based on a source-code aware point of view, we further introduce an automated in-line profiling methodology in order to allow an objective rating of the parallelization success. Moreover, we automatically extract runtime influencing aspects and exemplarily apply these methodologies to implementations of two different video-based driver-assistance algorithms considering two different processor types.
暂无评论