OpenCL Evaluation for Numerical Linear Algebra Library Development

With the help of of CUDA [7], [6], many applications improved their performance by using GPUs. In our project called Matrix Algebra on GPU and Multicore Architectures (MAGMA) [10], we mainly focus on dense linear algebra routines similar to those from LAPACK [1]. Other than CUDA, there exist other frameworks that allow platformindependent programming for GPUs. The main three frameworks are: 1) DirectCompute from Mircosoft, 2) OpenGL Shading Language (GLSL), and 3) OpenCL The first one allows access to graphics cards from multiple vendors. However, it is specific to Microsoft Windows and therefore it is not portable between host Operating Systems (OS). OpenGL Shading language [8] is portable across both GPU hardware and the host OS. However, it is specifically geared towards programming new graphics effects – GLSL does not have the scientific focus. OpenCL [3] has been designed for general purpose computing on GPUs (GPGPU). It is an open standard maintained by the Khronos group with the backing of major graphics hardware vendors as well as large computer industry vendors interested in off-loading computations to GPUs. As a result, there exist working OpenCL implementations for graphical cards and, in addition, there is an implementation that works without a GPU by off-loading computations to multi-core processors. As a result, OpenCL offers portability across GPU hardware, OS software, as well as multicore processors. Therefore OpenCL is our choice of implementing a portable numerical linear algebra library.