I-Corps: Mistos-Enabling Write-once Run-everywhere High Performance Software

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

Computer systems that combine traditional processors with powerful graphics processing units (GPUs) have become the standard in all platforms ranging from cloud servers, to laptops and tablets, and even cell phones. GPUs can yield performance or energy savings of up to 300x through their unique design. However, developing applications to exploit these systems is a difficult challenge, requiring specialized hardware expertise and tedious performance optimizations that make them unusable by a large fraction of software developers. To overcome this challenge, this project will explore the commercialization opportunities for a source-to-source optimization system that transparently generates high performance software for heterogeneous computer systems. This I-CORPS team has developed an automatic parallelization tool called Mistos that can help programmers write code for these high performance systems. With Mistos, programmers need to only focus on functionality, rather than hardware details and performance tuning. Mistos will enable users of all kinds of computing platforms, including mobile application developers, cloud services, and members of the scientific community, the ability to harness the computing power of heterogeneous GPU systems in a seamless and efficient manner.