high performance computing on graphics processing units: hgpu.org

Although specialized hardware promises orders of magnitude performance gains, their uptake has been limited by how challenging it is to program them. Hardware accelerators present challenges programmers are not used to, exposing details of the hardware that are often hidden and requiring new programming styles to use them effectively. Existing programming models often involve learning complex and hardware-specific APIs, using Domain Specific Languages (DSLs), or programming in customized assembly languages. These programming models for hardware accelerators present a significant challenge to uptake: a steep, unforgiving, and untransferable learning curve. However, programming hardware accelerators using traditional programming models presents a challenge: mapping code not written with hardware accelerators in mind to accelerators with restricted behaviour. This thesis presents these challenges in the context of the acceleration equation, and it presents solutions to it in three different contexts: for regular expression accelerators, for API-programmable accelerators (with Fourier Transforms as a key case-study) and for heterogeneous coarse-grained reconfigurable arrays (CGRAs). This thesis shows that automatically morphing software written in traditional manners to fit hardware accelerators is possible with no programmer effort and that huge potential speedups are available.