University of Utah, We examine a rendering system that interactively ray traces an image on a conventional multiprocessor. The implementation is “brute force ” in that it explicitly traces rays through every screen pixel, yet pays careful attention to system resources for acceleration. The design of the system is described, along with issues related to material models, lighting and shadows, and frameless rendering. The system is demonstrated for several different types of input scenes.

Simulating realistic lighting and rendering complex scenes are usually considered separate problems with incompatible solutions. Accurate lighting calculations are typically performed using ray tracing algorithms, which require that the entire scene database reside in memory to perform well. Conversely, most systems capable of rendering complex scenes use scan-conversion algorithms that access memory coherently, but are unable to incorporate sophisticated illumination. We have developed algorithms that use caching and lazy creation of texture and geometry to manage scene complexity. To improve cache performance, we increase locality of reference by dynamically reordering the rendering computation based on the contents of the cache. We have used these algorithms to compute images of scenes containing millions of primitives, while storing ten percent of the scene description in memory. Thus, a machine of a given memory capacity can render realistic scenes that are an order of magnitude more complex than was previously possible.

Many disciplines must handle the creation, visualization, and manipulation of huge and complex 3D environments. Examples include large structural and mechanical engineering projects dealing with entire cars, ships, buildings, and processing plants. The complexity of such models is usually far beyond the interactive rendering capabilities of todays 3D graphics hardware. Previous approaches relied on costly preprocessing for reducing the number of polygons that need to be rendered per frame but suffered from excessive precomputation times --- often several days or even weeks.

Global illumination is an area of research which tries to develop algorithms and methods to render images of

Figure 1: Examples of interactively rendering complex and dynamic scenes with a ray-tracing-based renderer. The scenes show a pre-lighted theatre, robots moving through a city, large numbers of moving trees with sharp shadows, as well as the integration of volumes, lightfields, and procedural shading in an office environment. These examples run interactively at a resolution of 640 × 480 using four to eight dual PCs. Ray-tracing is well-known as a general and flexible rendering algorithm that generates high-quality images. But in the past, raytracing implementations were too slow to be used in an interactive context. Recently, the performance of ray-tracing has been increased by over an order of magnitude, making it interesting as an alternative to rasterization-based rendering. We present a new rendering engine for interactive 3D graphics based on a fast, scalable, and distributed ray-tracer. It offers an extended OpenGL-like API, supports interactive modifications of the scene, handles complex scenes with millions of polygons, and scales efficiently to many client machines. We demonstrate that the new renderer provides more flexibility, more rendering features, and higher performance for complex scenes than current rasterization hardware. Its flexibility enables new types of applications including a system for interactive global illumination.

The achievement of parallel application performance on non-dedicated workstation clusters requires careful attention to the scheduling of tasks and communication on the underlying platform. In the literature, application scheduling policies are usually chosen by matching the resource requirements of an application with the performance characteristics of the target platform. However, when clusters of workstations are shared with other users, platform performance is non-uniform and varies over time. As a result, the performance of distinct scheduling policies may also vary depending on dynamic system state and particular characteristics of the job being run. Our experimental work focuses on a master/slave parallel ray-tracing application executing on a set of workstation clusters at UCSD and the San Diego Supercomputer Center. The experiments show that two di erent scheduling strategies, one static and one dynamic, exhibit very di erent performance sensitivities to variabilities in resou...

Parallelising ray tracing with a data parallel approach allows rendering of arbitrarily large models, but the inherent load imbalances may lead to severe inefficiencies. To compensate for the uneven load distribution, demand-driven tasks may be split off and scheduled to processors that are less busy. We propose a hybrid scheduling algorithm which brings tasks and data together according to coherence between rays. The amount of demand-driven versus dataparallel tasks is a function of the coherence between rays and the amount of imbalance in the basic data-parallel load. Processing power, communication and memory are three resources which should be evenly used. Our current implementation is assessed against these requirements, showing good scalability and very little communication at the cost of a slightly larger memory overhead. CR Categories: D.1.3 [Programming Techniques]: Parallel Programming; I.3.0 [Computer Graphics]: General I.3.7 [Computer Graphics]: Three-Dimensional Graphics...

Photon mapping is a global illumination algorithm for generating and visualizing a sparse representation of the incident radiance on surfaces. Photon mapping places an enormous burden on the memory hierarchy. A 512×512 image using the standard kd-tree data structure requires more than 196GB of raw bandwidth to access the photon map. This bandwidth is a major obstacle to our long term goal of designing hardware capable of real time photon mapping. This paper investigates two approaches for reducing the required bandwidth: 1) reordering the kNN searches; and 2) cache conscious data structures. Using a Hilbert curve reordering, we demonstrate an approximate lower bound of 15MB of bandwidth. This improvement of four orders of magnitude requires a prohibitive amount of intermediate storage. We then demonstrate two more costeffective algorithms that reduce the bandwidth by one order of magnitude to 24GB with 1MB of storage. We explain why the choice of data structure can not, by itself, achieve this reduction. Irradiance caching, a popular technique that reduces the number of required kNN searches, receives the same proportional benefit as the higher quality photon gathers.

Parallelising ray tracing with a data parallel approach allows rendering of arbitrarily large models, but the inherent load imbalances may lead to severe inefficiencies. To compensate for the uneven load distribution, demand-driven tasks may be split off and scheduled to processors that are less busy. We propose a hybrid scheduling algorithm which brings tasks and data together according to coherence between rays. Coherent tasks are scheduled demand driven and the remainder is executed data parallel. This method removes the worst hot-spots from the data parallel component and reschedules those as demand driven tasks, thereby evening out the workload. Processing power, communication and memory are three resources which should be evenly used. Our current implementation is assessed against these requirements. Related issues, such as the distribution of the workload over space and the resulting requirements for the distribution objects over the processors, are investigated as well. Final...

We present recent results in the application of distributed shared memory to image parallel ray tracing on clusters. Image parallel rendering is traditionally limited to scenes that are small enough to be replicated in the memory of each node, because any processor may require access to any piece of the scene. We solve this problem by making all of a cluster’s memory available through software distributed shared memory layers. With gigabit ethernet connections, this mechanism is sufficiently fast for interactive rendering of multi-gigabyte datasets. Object- and page-based distributed shared memories are compared, and optimizations for efficient memory use are discussed. Key words: scientific visualization, out-of-core rendering, distributed shared memory, ray tracing, cache miss reduction 1