Implementation of BM3D Filter on Intel Xeon Phi for Rendering in Blender Cycles

摘要

In this paper parallel implementation of Sparse 3D Transform-Domain Collaborative filter (BM3D) on the Intel Xeon Phi architecture is presented. Efficiency of the implementation in terms of speedup compared to serial implementation of the filter is demonstrated on denoising of rendered images. We also provide comparison with another parallel CPU version and show that ours performs better. Using the state-of-the-art image filters such as BM3D offers powerful denoising capability in the area of image filtering. To achieve the highest possible quality of the result, the filter has to perform multiple demanding tasks over a single image. Effective implementation of the filter is therefore very important. This is also the case, when filtering is used for image rendering. Rendering times can be significantly decreased by application of powerful time efficient denoising filters. Unfortunately the existing serial implementation of the BM3D filter is time consuming. In this paper we provide efficient parallel implementation of the BM3D filter, and we apply it as a noise reduction technique to the rendered images that reduces the rendering times. We also provide an optimized version of the filter for the Intel Xeon Phi and Intel Xeon architecture.