high performance computing on graphics processing units: hgpu.org

Graphics processing units (GPUs) traditionally have been used to accelerate only parts of the graphics pipelines. The emergence of the new age GPUs as highly parallel, multi-threaded and many core processor systems with the ability to perform general purpose computations has opened doors for new form of heterogeneous computing where the GPU and CPU can be used together in accelerating the underlying computations. General-purpose computing on graphics processing unit (GPGPU, also referred to as GP2U) techniques can be used to perform highly data parallel computations and to accelerate some critical sections of an application. Accelerating the computation of fMRI analysis on a graphics processing unit is mainly attractive when used in a clinical environment. In this thesis, I discuss methods which try to exploit the capabilities provided by GPUs to accelerate the analysis of time varying data acquired during fMRI experiments for identifying regions of activity/inactivity. Static activation maps are obtained by inspecting voxels independently with the help of statistical methods in parallel using CUDA (Compute unified device architecture) threads. I provide an efficient strategy for mapping each individual time varying voxels to GPU kernel threads for data parallel analysis of fMRI data and present GPU version of methods used in the fMRI analysis pipeline based on voxel to thread mapping technique. Also, an efficient method for octree based hierarchical clustering of voxels on a GPU and using a combination of GPU and CPU for enhanced clustering speedup is discussed. A comparison between the data parallel methods implemented on GPU and the corresponding CPU implementations and overall speed up achieved using combined GPU and CPU implementations in octree based hierarchical clustering is discussed.