| With current rendering algorithms, it is now possible to generate photorealistic images on a single desktop computer. At the core of many of these algorithms is a ray tracer, which is used to simulate physical light transport. All too often, the ray tracer is also the performance bottleneck, and as a result, photorealistic images often take hours to generate and/or employ approximations which detract from realism. Interactive content, by comparison, often appears dull and artificial.;The past decade has seen vast improvements to the core ray tracing algorithms, promising greater realism with Paster render times. A large body of work has been dedicated to so-called "coherent ray tracing" methods, where the coherence between neighboring rays is exploited to improve ray tracing performance. These algorithms map cleanly to modern processors' Single Instruction Multiple Data (SIMD) units, render primary visibility and point-light shadows at interactive to real-time rates, and provide over an order of magnitude performance improvement over traditional one-at-a-time ray tracers in some cases. These benefits do not come cheaply: we must sacrifice some of the generality of one-at-a-time ray tracing. Moreover, coherence is not guaranteed for more complex light transport, beyond primary visibility and point-light shadows. So it remains unclear as to how well these new methods extend to the more complex light transport effects, such as area lighting, reflections, refractions, depth of field, motion blur, and global illumination.;In this thesis, I propose three bodies of work which introduce and evaluate new algorithms for coherent ray tracing dedicated to accelerating complex light transport effects: a real-time beam tracer with application to exact soft shadows from area light sources, large ray packets for real-time Whitted ray tracing, and adaptive wavelet rendering for general high-dimensional effects.;I first introduce a highly optimized beam tracer which computes noise-free soft shadows in seconds and renders antialiased primary visibility and point-light shadows in real-time. Whereas bounding frusta have previously been used to cull away expensive per ray intersection tests for entire ray packets, I use the frustum as the atomic ray primitive. This results in faster visibility testing when the scene geometry is coherent, and also provides an exact visibility solution for efficient antialiasing and analytic soft shadows.;Then, to handle reflections and refractions, I construct an interactive Whitted ray tracer composed of new algorithms for large ray packets and frustum culling. Within this framework, I offer a thorough analysis of several coherent ray tracing algorithms, and observe strong benefits, albeit less than for primary visibility and point-light shadow rays. Even in situations of extreme incoherence, large ray packets tend to be 3x-6x faster than 4-wide SIMD rays.;Finally, I propose adaptive wavelet rendering for general high-dimensional effects, such as area lighting, depth of field, motion blur, and diffuse inter-reflections. This is an adaptive Monte Carlo algorithm, but rather than adapt to a per pixel measure of variance, this new algorithm adapts to variance in a multi-scale wavelet basis. Thus it adapts to smooth sources of variance, such as the blur from an out-of-focus camera, by sampling at an effectively lower resolution, while targeting edges with more focused sample distributions. The remaining fine-scale noise is removed by a novel wavelet reconstruction filter. One notable aspect of adaptive wavelet rendering is that it adaptively samples image regions rather than points, and so is well suited to coherent ray tracing techniques. This new algorithm often achieves near-reference quality images with general combinations of high-dimensional effects with an average of only 32 samples per pixel, far fewer than required by traditional means. Moreover, the algorithm is efficient, and maintains low overhead even when used with an optimized coherent ray tracer.;Together, these three works improve coherent ray tracing performance for a broad range of complex light transport effects that are vital for photorealistic rendering. They increase the quality of interactive content and decrease the render times of offline photorealistic content, bringing us two steps closer to interactive photorealistic images. |
