Anisotropic nonlinear diffusion for filtering 3D images on GPUs

Optimizing sophisticated PDE-based filtering methods, such as the Anisotropic Nonlinear Diffusion (AND), to GPUs is complicated and time consuming. In this work, we expressed AND as iterative multiple 3D-stencils, where each 3D-stencil is implemented into one kernel, and then we analyzed all possible kernel fusions on the GPU. We experimentally found that fusing dependent stencils with similar concurrency and lower on-chip pressure makes the optimal combination run 1, 52× faster than the next better one.