Paper Details

Communicating Process Architectures (CPA)
 Title: Code Specialisation of Auto-Generated GPU Kernels
 Conference: Communicating Process Architectures 2015
 Authors: Troels Blum, Brian Vinter
Niels Bohr Institute, University of Copenhagen
 Abstract: This paper explores and evaluates the effect of automatic code specialization on auto generated GPU kernels. When combining the high productivity coding environment of computational science with the Just-In-Time compilation nature of many GPU runtime systems, there is a clear cut opportunity for code optimization and specialization. We have developed a hybrid kernel generation method which is shown to be useful and competitive across very different use cases, and requires minimal knowledge of the overall structure of the program. Stencil codes which are commonly found at the core of computer simulations are ideal candidates for this type of code specialization. For exactly this type of application we are able to achive speedups of up to 2.5 times with the implemented strategy. 

BibTeX Entry

Full paper