high performance computing on graphics processing units: hgpu.org

This bachelor thesis carries out a case study describing the performance and productivity of modern vector processors such as graphics processing units (GPUs) and central processing units (CPUs) based on three different computational routines arising from a magnetoencephalography application. I apply different programming paradigms to these routines targeting either the CPU or the GPU. Furthermore, I investigate the performance and productivity of programming paradigms such as OpenMP with respect to its auto-vectorization capabilities, Intel intrinsic AVX and Intel OpenCL for the CPU. Moreover, I examine NVIDIA’s CUDA and OpenCL APIs for GPU-sided applications. The results of the performed case study yield roughly the same performances for the CPU and GPU implementations, but favour the OpenMP paradigm (i.e. the CPU) with respect to productivity.