high performance computing on graphics processing units: hgpu.org

With modern processing hardware converging on the physical barrier in terms of transistor size and speed per single core, hardware manufacturers have shifted their focus to improve performance from raw clock power towards parallelization. Solutions to utilize the computation power of GPUs are published and supported by graphics card manufacturers. While there exist solutions for arbitrary precision integer arithmetics on the CPU there has been little adoption of these libraries to the GPU. This thesis presents an approach to map arbitrary precision integer operations to single threads on the GPU. This novel computation mapping technique is benchmarked and compared to a library that runs these computations on the CPU. Furthermore the novel parallelization technique is compared to an alternative mapping scheme proposed by Langer et al [Lan15]. It is shown that mapping computations to single threads outperforms both the CPU and the approach by Langer. This thesis also explored the feasibility of rational number operations on the GPU and shows that this is in fact practically usable by providing benchmarks.