high performance computing on graphics processing units: hgpu.org

These days the field of Artificial Intelligence (and its many subfields) is moving really fast, many new techniques are becoming available from various different subfields. However, many of these algorithms are only made to run on very powerful research workstations without considering how they can be used on real-world hardware, be it embedded hardware, powerful GPUs or even distributed systems. In this PhD we try to bridge this gap. We study how different algorithms, from fields like artificial intelligence, computer vision or image processing, can be adapted to different hardware that is accompanied by some restrictions or/and programming challenges. We look at hardware that is constrained in terms of resources (eg. embedded devices). The big challenge here is to make sure that the requirements of the algorithm, and the requirements of the contexts in which it is used (eg. being able to run in real time) are fulfilled using only these limited resources. Other than that, we also look at hardware that uses different programming paradigms which often makes them challenging to program (eg. GPUs). While these kinds of hardware are often not limited in terms of resources (often the opposite is true), exotic architectures mean that using this hardware to its full potential remains a challenge. In this thesis we examine these challenges based on many applications, and lay them out on top of a framework with constraints and trade-offs that outlines what they have in common and sets them apart.