Interactive and offline rendering in Blender Cycles

In this paper, we describe interactive and offline rendering performed on multiple nodes of an HPC (High Performance Computing) cluster. Compute nodes can be equipped with Intel Xeon processors and also Intel Xeon Phi coprocessors. Rendering is utilized in Blender (open source 3D creation suite). We have modified the kernel of the Blender Cycles rendering engine and then extended its capabilities to support the HPC cluster. We call it CyclesPhi. The CyclesPhi supports hybrid MPI/OpenMP/Offload concepts. The original Blender Cycles engine has limited network rendering capabilities, which cannot be used on supercomputers. Our paper presents a basic algorithm for image rendering, decomposition tasks for application of parallel strategies, and basic collective communication routine using MPI methods. All of the presented techniques improve strong scalability of a cluster in both the offline and also interactive rendering.