Data-parallel Micropolygon Rasterization

Abstract

Abstract We implement a tile based sort-middle rasterizer in CUDA and study its performance characteristics when used as a backend for adaptive tessellation down to micropolygons. Tessellation and bucketing map very well to the data-parallel paradigm of CUDA, and the majority of time is spent with rasterization. Despite this, our fastest implementation is able to reach 30-50% of the hardware rasterization performance of an Nvidia GTX 280. Overall we are able to rasterize 4 M textured and Phong shaded microquads into a 1600x1200 framebuffer at 10-12 fps.Abstract We implement a tile based sort-middle rasterizer in CUDA and study its performance characteristics when used as a backend for adaptive tessellation down to micropolygons. Tessellation and bucketing map very well to the data-parallel paradigm of CUDA, and the majority of time is spent with rasterization. Despite this, our fastest implementation is able to reach 30-50% of the hardware rasterization performance of an Nvidia GTX 280. Overall we are able to rasterize 4 M textured and Phong shaded microquads into a 1600x1200 framebuffer at 10-12 fps.