Efficient GPU Path Rendering Using Scanline Rasterization

摘要

We introduce a novel GPU path rendering method based on scanlinernrasterization, which is highly work-efficient but traditionallyrnconsidered as GPU hostile. Our method is parallelized over boundaryrnfragments, i.e., pixels directly intersecting the path boundary.rnNon-boundary pixels are processed in bulk as horizontal spans likernin CPU scanline rasterizers, which saves a significant amount ofrnwinding number computation workload. The distinction also allowsrnthe majority of our algorithmic steps to focus on boundaryrnfragments only, which leads to highly balanced workload amongrnthe GPU threads. In addition, we develop a ray shooting patternrnthat minimizes the global data dependency when computing windingrnnumbers at anti-aliasing samples. This allows us to shift the majorityrnof winding-number-related workload to the same kernel thatrnconsumes its result, which saves a significant amount of GPU memoryrnbandwidth. Experiments show that our method gives a consistentrn2:5× speedup over state-of-the-art alternatives for high-qualityrnrendering at Ultra HD resolution, which can increase to more thanrn30× in extreme cases. We can also get a consistent 10u0002 speeduprnon animated input.