GPU Production Rendering

This chapter is not about realtime rendering. It’s about a world in which rendering takes seconds, minutes, or hours. It’s about users who require features that are extremely difficult to implement on modern graphics hardware. But we’d still like to take advantage of the tremendous horsepower that GPUs afford us. First, we’ll discuss the arcane world of production rendering for film (and other non-game high-end uses) — what people need from a film renderer, and how this is different from a game engine. Next, we’ll dive into the gory details about how GPUs were used for certain features in NVIDIA’s Gelato R © renderer. There are some clever techniques here, but it’s also an object lesson about how insanely tricky it is to get GPUs to render in ways they weren’t designed for, though occasionally if you do it right, you can get fantastic performance benefits. NVIDIA’s Gelato renderer is an offline (batch-oriented, non-realtime) renderer designed to render absurdly complex scenes at high enough quality to be used in feature film visual effects and animation. Unique among similarly capable renderers (as of this writing), it requires a GPU, which it uses to accelerate various of its algorithms. But it is not using graphics hardware for previews or reduced-quality rendering— it uses the GPU to help generate the highest quality final pixels. Indeed, part of the requirement of the product is that we use GPU where we can, but never allow hardware usage to sacrifice image quality, flexibility, or feature set. The bottom line is that we’ve spent the past few years learning how to use GPUs to make very high-quality imagery, often in ways quite different from how a game or any other realtime application would use the graphics hardware. For these notes, we’ll concentrate on just a small subset of Gelato’s functionality: hidden