Terrain Rendering: Performance Issues

• Yon Clipping
  • the simplest way to improve performance: enable yon clipping and use fog to hide the hard visual edge of the clipping plant
  • can improve framerate dramatically by avoiding pixel fill
• Visibility (Occlusion Culling)
  • because terrain heightfield geometry is special, you can do visibility occlusion more easily than in the general case
  • paper on Visibility in Terrains by James Stewart
  • Hybrid dVPS
    • a commercial library which does several kinds of occlusion culling, independent of the scene graph library it is used with
    • licensing costs $1 for pure research uses, commercial licensing is done via a partnership with Criterion
    • you can download several demos, which include urban and large-area terrain scenes
    • tried the demo on my machine (P3-800, GeForce2) - interestingly, although the occlusion culling worked well as shown in wireframe rendering, it was not significantly faster 
• Fillrate
  • usually measured in millions of pixels per second (Mpixels)
  • how much fillrate is needed for an entirely-drawn window without overdraw?
    • 800x600 @ 60Hz = 29 Mpixels
    • 1024x768 @ 60Hz = 48 Mpixels
    • 1280x1024 @ 60 Hz = 79 Mpixels
  • so, in theory all that's needed for a scene with depth complexity of 1 is a card that does ~100 real-world megapixels
  • however, thanks to modern scenes having high depth complexity and multiple rendering passes, several hundred Mpixels are needed
• Floating-point accuracy and world scale
  • a major problem with drawing large scenes is that floating-point accuracy is not always sufficient
  • Sean O'Neil's Gamasutra article A Real-Time Procedural Universe, Part Three Matters of Scale describes the challenges, and some solutions, in great detail

• Parallel Rendering (high end)
  • IEEE Transactions on Visualization and Computer Graphics (TVCG), Vol. 2, No. 3 (Sept. 1996)
    RIVA: A Versatile Parallel Rendering System for Interactive Scientific Visualization
    P. Peggy Li, William H. Duquette, and David W. Curkendall
    Abstract is online, IEEE membership required to access paper
  • 5/99 studied the paper
  • implementation was run on two platforms: Cray T3D (256 CPUs) and a Paragon (512 CPUs)
  • despite the massive computational power, they only get 1-3 fps
  • can handle a whole planet (spherical data)