GPU++: an embedded GPU development system for general-purpose computations

Using the graphics processing unit (GPU) to accelerate general-purpose computations has become an important technique in scientific research. However, the development complexity is significantly higher than for CPU-based solutions, due to the mainly graphics-oriented concepts and development tools for GPU-programming. As a consequence, general-purpose computations on the GPU are mainly discussed in the academic domain and have not yet fully reached industrial software development. This thesis presents a novel contribution to general-purpose GPU programming – the Gpu development system. It features a seamless integration into the C programming language to address graphics hardware via a familiar syntax, an abstraction layer to efficiently hide the different computation frequencies, and a novel approach to relax the vector processor paradigm of the GPU. The Gpu development system enables software engineers to embed GPU-based development into their existing software engineering workflow, thus, largely dissolving the differences between main and graphics processor. The developed system has been evaluated by the implementation of well-known generalpurpose algorithms on both processor platforms – including general matrix multiplication, sorting of values, the fast Fourier transform, and tomographic reconstruction via filtered back-projection. The Gpu implementations outperform the CPU solutions by up to three orders of magnitude without a noticeable increase in source code complexity. These results clearly demonstrate that the novel Gpu development system significantly reduces the source code complexity and development time of general-purpose GPU applications to a level which is comparable to main processor implementations, while obtaining the tremendous performance advantages of today’s graphics hardware architectures.