high performance computing on graphics processing units: hgpu.org

In recent years, graphics processing units (GPUs) became more and more popular as high performance processing units. Due to the availability of hundreds of cores, code fragments speed up significantly when they are transformed from CPU functions to GPU kernels. The transformation process is non-trivial and therefore error prone. Developing correct and efficient GPU accelerated programs is time consuming. One of the aspects that has to be considered is the logically and physically separation of a GPU’s address space from the CPU. In order to enable kernel execution on the GPU, data transfers between the CPU and GPU need to be explicitly scheduled in the code. Since the CPU-GPU communication bandwidth is relatively low, compared to the computational power of the GPU, the number of transferred bytes should be minimised. A GPU is equipped with an advanced hierarchy of scratchpad memories, that is different from CPU memory hierarchies. When these differences are not taken into during the development of the GPU kernel, the computational power of the GPU is not fully exploited. Memory access patterns need to be adapted to take full advantage of this memory hierarchy. In order to decrease development time, there is an increasing interest in tools that can assist in the transformation process from CPU programs to GPU accelerated programs. In this thesis, we propose a set of analysis tools and code optimisations to aid the transformation of CPU code fragments to GPU kernels. The tools focus on optimising the CPU-GPU communication strategy and optimising the kernels for the GPU’s advanced memory hierarchy.