Output details
11 - Computer Science and Informatics
University of Edinburgh
Semi-automatic extraction and exploitation of hierarchical pipeline parallelism using profiling information
<08> Originality: Presents the first methodology to extract pipeline parallelism from legacy sequential applications. Profiling information replaces conservative static analysis. Pipeline stages are automatically identified and balanced and if safe and profitable replicated, functions split, and data parallel loops parallelised. User only checks final program.
Significance: Pipeline parallelism is at the heart of many streaming, multimedia, cryptographic and compression applications. The parallelisation tools from this research have potential to enhance performance of many applications on current and future multicores.
Rigour: Ideas implemented in commercial CoSy software framework, evaluated against large real-world applications. PACT'10 conference had just 17% acceptance rate.