high performance computing on graphics processing units: hgpu.org

Image reconstruction, a very compute-intense process in general, can often be reduced to large linear equation systems represented as sparse under-determined matrices. Solvers for these equation systems (not restricted to image reconstruction) spend most of their time in sparse matrix-vector multiplications (SpMV). In this paper we will present a GPU-accelerated scheme for a Conjugate Gradient (CG) solver, with focus on the SpMV. We will discuss and quantify the optimizations employed to achieve a soft-real time constraint as well as alternative solutions relying on FPGAs, the Cell Broadband Engine, a highly optimized SSE-based software implementation, and other GPU SpMV implementations.