The texture shading cache : an approach to asynchronous interactive illumination

Abstract

The introduction of algorithms simulating physically correct light transport in rendering has pushed the computational bottleneck in interactive visualisation applications towards pixel shading. Rendering is increasingly becoming bound by illumination processes and as a result, many interactive applications tend to make use of techniques that try to mimic the effects of light transport algorithms, sacrificing correctness while maintaining interactivity[18, 20]. With physically correct light transport algorithms, these effects are handled by the nature of the algorithm itself at the expense of computing power. This project explores the decoupling of physically correct shading computations from the traditional rasterization-based rendering process through the creation of a texture cache which is kept updated asynchronously at runtime. A simple scheduler determines which shading computations are computed asynchronously and which are effected in lock-step with frame rendering using the traditional approach. This is done with the aim of achieving a stable frame rate without dips or stuttering, whilst preserving image quality. A solution was implemented in C++ using OpenGL [12] for the traditional renderer. Moreover, a path tracer to simulate physically correct light transport was built using Nvidia Optix [17] for the asynchronous calculations. A texture atlas was used as the texture cache, where each geometric primitive in the scene is allocated a space within this atlas during runtime. This solution was then evaluated thoroughly to assess how effective the concept is with respect to both performance and image quality. Multiple modes were evaluated: Decoupling all the shader computations; decoupling only the texture and bounce lighting computations; decoupling only the bounce lighting calculations. The performance of each mode was evaluated, using metrics such as frames per second and memory usage. Furthermore, the quality of the images produced was also measured using HDR-VDP-2 and SSIM against high quality versions of the scenes rendered offline.