A multi-GPU implementation of diffusion operators within 3D volumes

Diffusion operators have been used extensively in engineering disciplines as well as in image processing. The numerical solution of the diffusion partial differential equation is an iterative process that can require from a few hundred iterations to several thousand iterations depending on the specific application. Furthermore, large volumetric datasets need to be processed, especially in medical image processing and in X-ray microtomography. This work presents a multi-GPU implementation of a diffusion operator using OpenCL and compares it to a multi-threaded CPU-based algorithm. The algorithm theory and development is initially presented for the general diffusion model within a volumetric dataset. This will be specialised for the Perona-Malik diffusion filter that is frequently used as a preprocessing step when analysing MRI and CT scans.