high performance computing on graphics processing units: hgpu.org

This doctoral thesis focuses on GPU acceleration of medical image registration and sparse general matrix-matrix multiplication (SpGEMM). The comprehensive work presented here aims to enable new possibilities in Image Guided Surgery (IGS). IGS provides the surgeon with advanced navigation tools during surgery. Image registration, which is a part of IGS, is computationally demanding, therefore GPU acceleration is greatly desirable. spGEMM, which is an essential part in many scientific and data analytics applications, e.g., graph applications, is also a useful tool in biomechanical modeling and sparse vessel network registration. We present this work in two parts. The first part of this thesis describes the optimization of the most demanding part of non-rigid Free Form Deformation registration, i.e., B-spline interpolation. Our novel optimization technique minimizes the data movement between processing cores and memory and maximizes the utilization of the very fast register file. In addition, our approach re-formulates B-spline interpolation to fully utilize Fused Multiply Accumulation instructions for additional benefits in performance and accuracy. Our optimized B-spline interpolation provides significant speedup to image registration. The second part describes the optimization of spGEMM. Hardware manufacturers, with the aim of increasing the performance of deep-learning, created specialized dense matrix multiplication units, called Tensor Core Units (TCUs). However, until now, no work takes advantage of TCUs for sparse matrix multiplication. With this work we provide the first TCU implementation of spGEMM and prove its benefits over conventional GPU spGEMM.