Abstract

Graphics processors are highly parallel computational units employing multiple levels of computational parallelism and memory hierarchy. Due to their high computational power they are increasingly used in scientific applications. However, optimizing algorithms for high performance on GPUs is not trivial.

We discuss optimizing sparse linear algebra on GPUs (specifically, sparse matrix-vector multiplication, SpMV, which is the most time-consuming step in many applications). We describe several known sparse matrix storage formats and present a new storage format that allows SpMV performance to be improved.

[This paper is to appear in HiPEAC’10]

About the speaker: Alexander Monakov is a PhD student at Moscow State University and an employee at Institute for System Programming of Russian Academy of Sciences (ISP-RAS), where he works on improving the GCC compiler. His interests include general-purpose GPU computing, compiler optimization technology and in particular using polyhedral model for parallelism and locality optimization.