Clustered Shading : Assigning arbitrarily shaped convex light volumes using conservative rasterization

Abstract: Context. In this thesis, a GPU-based light culling technique performed with conservative rasterization is presented. Accurate lighting calculations are expensive in real-time applications and the number of lights used in a typical virtual scene increases as real-time applications become more advanced. Performing light culling prior to shading a scene has in recent years become a vital part of any high-end rendering pipeline. Existing light culling techniques suffer from a variety of problems which clustered shading tries to address. Objectives. The main goal of this thesis is to explore the use of the rasterizer to efficiently assign convex light shapes to clusters. Being able to accurately represent and assign light volumes to clusters is a key objective in this thesis. Methods. This method is designed for real-time applications that use large amounts of dynamic and arbitrarily shaped convex lights. By using using conservative rasterization to assign convex light volumes to a 3D cluster structure, a more suitable light volume approximation can be used. This thesis implements a novel light culling technique in DirectX 12 by taking advantage of the hardware conservative rasterization provided by the latest consumer grade Nvidia GPUs. Experiments are conducted to prove the efficiency of the implementation and comparisons with AMD´s Forward+ tiled light culling are provided to relate the implementation to existing techniques. Results. The results from analyzing the algorithm shows that most problems with existing light culling techniques are addressed and the light assignment is of high quality and allows for easy integration of new convex light types. Assigning the lights and shading the CryTek Sponza scene with 2000 point lights and 2000 spot lights takes 2.92ms on a GTX970. Conclusions. The conclusion shows that the main goal of the thesis has been reached to the extent that all existing problems with current light culling techniques have been solved, at the cost of using more memory. The technique is novel and a lot of future work is outlined and would benefit the validity of the implementation if further researched.