A Dynamic Caching System for Rendering an Animated Crowd in Real-Time

Abstract

We present a method to accelerate the rendering of large crowds of animated characters. Recent trends have seen matrix-palette skinning become the prevalent approach due to its low memory overhead and fully dynamic geometry. However, the performance of skeletal animation remains modest in comparison to static rendering since neither temporal nor intra-frame coherency can be exploited. We cast crowd rendering as a memory-management problem and allocate a small geometry cache on the GPU within which animated characters can be stored. This serves to augment matrix-palette skinning with baked geometry and allows animation frames to be re-used by multi-pass rendering, between multiple agents and across multiple frames. Our method builds its cache dynamically and adapts to the current simulation state through use of the page-replacement algorithms traditionally employed by virtual-memory systems. In many cases this negates the need for skinning altogether and enables thousands of characters to be rendered in real-time, each independently animated and without loss of fidelity.