ViennaCL - A High Level Linear Algebra Library for GPUs and Multi-Core CPUs

The vast computing resources in graphics processing units (GPUs) have become very attractive for general purpose scientific computing over the past years. Moreover, central processing units (CPUs) consist of an increasing number of individual cores. Most applications today still make use of a single core only, because standard data types and algorithms in wide-spread procedural languages such as C++ make use of a single core only. A customized adaption of existing algorithms to parallel architecture requires a considerable amount of effort both from algorithmic and programming point of view. Taking this additional amount of work hours required for an adaption to GPUs starting from scratch into account, the use of GPUs may not pay off on the overall. The Vienna Computing Library (ViennaCL), which is presented in this work, aims at providing standard data types for linear algebra operations on GPUs and multi-core CPUs. It is based on OpenCL, which provides unified access to both GPUs and multi-core CPUs. The ViennaCL API following existing programming and interface conventions established with uBLAS, which is part of the peer-reviewed Boost library. Thus, the open source library can be easily integrated into existing C++ implementations and therefore reduces the necessary code changes in existing software to a minimum. In addition, algorithms provided with ViennaCL can directly be used with uBLAS types due to the common interface. The algorithmic focus of ViennaCL is on iterative solvers, which are often used for the solution of large systems of linear equations typically encountered in the discretization of partial differential equations using e.g. finite element methods. Benchmark results given in this work show that the performance gain of ViennaCL over uBLAS is on both GPUs and multi-core CPUs up up to an order of magnitude. For small amounts of data, the use of ViennaCL may not pay off due to an OpenCL management overhead associated with the launch of compute kernels.