GPGPU separation of opaque and transparent mesh polygons

Abstract: Context: By doing a depth-prepass in a tiled forward renderer, pixels can be prevented from being shaded more than once. More aggressive culling of lights that might contribute to tiles can also be performed. In order to produce artifact-free rendering, only meshes containing fully opaque polygons can be included in the depth-prepass. This limits the benefit of the depth-prepass for scenes containing large, mostly opaque, meshes that has some portions of transparency in them. Objectives: The objective of this thesis was to classify the polygons of a mesh as either opaque or transparent using the GPU. Then to separate the polygons into two different vertex buffers depending on the classification. This allows all opaque polygons in a scene to be used in the depth-prepass, potentially increasing render performance. Methods: An implementation was performed using OpenCL, which was then used to measure the time it took to separate the polygons in meshes of different complexity. The polygon separation times were then compared to the time it took to load the meshes into the game. What effect the polygon separation had on rendering times was also investigated. Results: The results showed that polygon separation times were highly dependent on the number of polygons and the texture resolution. It took roughly 350ms to separate a mesh with 100k polygons and a 2048x2048 texture, while the same mesh with a 1024x1024 texture took a quarter of the time. In the test scene used the rendering times differed only slightly. Conclusions: If the polygon separation should be performed when loading the mesh or when exporting it depends on the game. For games with a lower geometrical and textural detail level it may be feasible to separate the polygons each time the mesh is loaded, but for most game it would be recommended to perform it once when exporting the mesh.