The use of graphics processing units (GPUs) for scientific computing has found broad acceptance, because they enable certain applications to run faster for a given power or money budget. On the other hand, programming GPUs is often considered hard, because inherent complexities of massively parallel algorithms and hardware need to be addressed. Also, established sequential algorithms are no longer competitive with parallel alternatives, which requires a redesign of the algorithm stack and data structures in complicated applications. To simplify the use of GPUs, an overview of GPU-accelerated functionality in the free open source linear algebra library ViennaCL is provided. The focus of the discussion is on high-performance implementations of iterative solvers and massively parallel preconditioners. Techniques for preserving high performance across hardware from different vendors are presented, underlining that a solid understanding of the underlying hardware is the key for good software implementations.

Karl Rupp is a postdoctoral researcher at the Institute for Microelectronics, TU Wien, and leads the development of the ViennaCL library. He is also active in the development of improved support for GPUs in the large-scale solver library PETSc. His scientific interests are highly efficient simulations of semiconductor nanodevices, which requires continuous interaction and exchange with engineers, physicists, mathematicians, and computer scientists.