Iterative magnetic resonance imaging reconstruction

In this chapter we present the key steps for parallelizing and optimizing a loop-intensive application, iterative reconstruction of magnetic resonance imaging images, from its sequential form. We start with the appropriate organization of parallelization: the scatter approach versus the gather approach. We show that transforming from a scatter approach to a gather approach is key to avoiding atomic operations, which can significantly reduce the performance of parallel execution. We discuss the practical techniques, that is, loop fission and loop interchange, that are needed to enable the gather approach to parallelization. We further present the application of optimization techniques such as promoting array elements into registers, using constant memory/cache for input elements, and using hardware functions to improve the performance of the parallel kernel.