Accelerating ray tracing engine of BLENDER on the new Sunway architecture

With the increasing popularity of high-resolution displays, there is a growing demand for more realistic rendered images. Ray tracing has become the most effective algorithm for image rendering, but its complexity and large amount of computing data require sophisticated HPC solutions. In this article, we present our efforts to port the ray tracing engine CYCLES of Blender to the new generation of Sunway supercomputers. We propose optimizations that are tailored to the new hardware architecture, including a multi-level parallel scheme that efficiently maps and scales Blender onto the novel Sunway architecture, strategies to address memory bottlenecks, a revised task dispatching method that achieves excellent load balancing, and a pipeline approach that maximizes computation and communication overlap. By combining all these optimizations, we achieve a significant reduction in rendering time for a single-frame image, from 2260 s using the single-core serial version to 71 s using 48 processes, which is a speedup of about 128x. Accelerating the ray tracing engine CYCLES of Blender in the new generation of Sunway supercomputers.