Performance Study and Optimization of an MRI Motion-Compensated Reconstruction Program

Motivation: MRI motion-compensated reconstruction programs rely on several computationally intensive algorithms. For clinical use, they need to use efficiently the computational resources of compute nodes to achieve a good performance. Goal(s): Evaluate the performance of a program across architectures and optimize its execution without any code modification. Approach: We investigate the different parallelization paradigms. We use the roofline model, and the performance profiling tools to derive the architectural efficiency on CPU. A Matlab GPU implementation of the reconstruction kernel is used to draw comparisons. Results: The optimal parallel mapping reduces considerably the reconstruction time. Moreover, GPUs don’t outperform CPUs in every reconstruction problem. Impact: We demonstrate the importance of micro-architectural study in the overall behavior of MRI motion-compensated reconstruction codes. This study provides insights about how to select the most suitable architecture to a given code, depending on its hardware limitations.