This document provides an overview of the technical details and design considerations for a graphical MUD game called Felwyrld. Key points include: - The game supports up to 256 players per server on a large 50km x 50km virtual world with static and dynamically placed content. - The renderer uses OpenGL, deferred shading, occlusion queries, and LOD techniques to optimize performance for trees, terrain, clouds, and other game elements across different distances. - The server uses prediction, compression, and quantization to minimize bandwidth while the dynamic spawning system balances CPU load as players explore the virtual world. - Technical challenges addressed include visibility testing, paging of terrain and tree assets, level-

1. Felwyrld
Alex Nankervis
anankervis@gmail.com

2. Overview
• Graphical MUD tech-demo
– Targets 256 players per server machine
• Large world (50km x 50km)
– Static and dynamically placed content
– Lots of work for the renderer
• OpenGL, OpenAL + Ogg/Vorbis,
OpenSSL, zlib, WinSock/enet, DirectInput

5. Server
• AI, player records, world state, chat system
• Low bandwidth
– Client side prediction (extrapolation)
– Server side prediction of client’s predicted state
• Updates clients only when potentially out of sync
– Client views
– Compression
– Quantization
• Coordinate system relative to client view
• CPU usage
– Dynamic spawning around clients
• Unpopulated areas inactive
• Encrypted login

6. Render Overview
• Z pre-pass
– Draw only to depth buffer, no color output
– Reduces overdraw, used for occlusion queries
• Issue occlusion queries
• CPU rendering work
– Kill time while waiting on occlusion queries
– Visibility, detail meshes, clouds, animation
• Retrieve occlusion queries
– Finish visibility testing
• Draw models, trees, terrain, sky dome, particles
• Post processing
– Lighting
• UI and overlays

7. Visibility Testing
• Quad tree
– Rejects large groups of trees and terrain quickly
– Cached to disk
• Load time and memory allocation optimization
• Occlusion queries
– Reject hidden objects
• Quad tree blocks
• Terrain blocks
• Tree bounding boxes
– Full detail trees, only
– A win under heavy loads
• Requires careful scheduling to maintain pipelining of GPU
– Mix CPU/GPU work, retrieve query results after some time
– Or, delay reading the query result until a later frame

8. Trees
• Several tree templates
– Instanced as 1,000,000 total trees
– Placed during terrain generation
• 7km draw distance
• Up to 60,000 drawn per frame

9. Tree Generation
• Tweak parameters for
different tree types
– Overall shape
– Branch twisting
– Leaf distribution
– Random variation
• Start at trunk(s), work
recursively to leaves

10. Tree Rendering
• Branches
– One vertex buffer
• Stored in order of importance
• LOD by truncating index list
– Wind animation
• Branches bend and sway in the wind
• Leaves
– One vertex buffer, no index buffer
• Stored in shuffled random order
• LOD by truncating removes random leaves
– Camera-aligned quads
– Wind animation
• Quads rotated in vertex program
– Static shadowing and color variation
• Darker near center and bottom of tree

11. Tree LOD
• Branches and leaves removed over
distance
• Billboards
– Texture contains tree from 16 angles
– Lighting
– Transition draws mix of tree geometry
and billboard
– Immediate vs. system array vs. VBO
• Similar performance
• Immediate is easiest and fastest
• VBO is future proof

12. Terrain
• 50km x 50km at ~12 meter resolution
– 4097x4097 heightfield
– Unlimited draw distance
• Hierarchy of 33x33 patches
– Matches quad tree layout
– Vertices, texture list, two blend textures
– Topology same for every block
• One set of indices

13. Terrain Creation
• Rivers, mountain ranges, deserts
• Driven by simple climate map
• Perlin Noise
– Modified by image filters
• Created as a pre-process
– ~20 minutes
• Outputs
– Heightfield
– Normal map
– Texture blend maps
– Tree placement

14. Terrain Paging
• Single vs. multiple files
• Eviction policy
– Unused blocks deallocated over time
– Video memory
– System memory

15. Terrain Coordinates
• Relative coordinate system
– Most geometry drawn relative to camera
• Double precision camera
– Counters precision errors
• Jittery vertices or texture sampling

16. Terrain Lighting
• Lighting
– Sun + sky lookup
– Normals in half-resolution texture map
• Bilinear filtering from half resolution gives better
results than vertex normals
• Much lower memory usage
– Compresses well
• Detail normal maps

17. Terrain Texturing
• Texture splatting
– Two blend textures
• Up to 8 detail textures per block
– Blend factors modified by noise
• Increases apparent detail
– Detail textures are sampled at two rates
• Near and far detail
– Result modulated by half-resolution color map
• Pre-rendered
• Breaks up uniformity
• Can control terrain appearance at arbitrary locations

18. Terrain LOD
• LOD by distance
– Draw higher levels of the quad tree
• Still 33x33 blocks, but each block contains the
area of 4, 16, 64, etc. lowest LOD blocks

19. Terrain LOD Seams
• Seams hidden by vertical 'skirt‘
– Drop edge vertices
– Rendering is simple and fast
– Not visible to user
– No complicated processing or logic

20. Bushes, Grass, Rocks
• Uniform grid of blocks around the camera
– Each block contains ~1000 detail meshes
– Meshes batched together into a single VBO
– Blocks generated from a seed number
• Created as needed
• Revisited blocks will be the same
• Evicted similar to terrain blocks
• Wind sway through vertex program
• Frustum culling at block level

21. Sky Dome
• Sky color lookup
– Time of day and angle
• Sun color lookup
• Sunset flare
• Matches fog and sky lighting for geometry
– Fogs seamlessly into the horizon

22. Clouds
• 32 unique clouds
– Instanced 8 times each
– Based on work by Mark Harris
– Lit particle system
• Lighting calculation
– Fast pre-process
• Once per unique cloud when light changes
• Spread over multiple frames
– Handles any number of lights plus ambient
• Felwyrld uses just one, the sun
– GPU vs. CPU
• Fast, special purpose CPU rasterizer best
– Improvement over original approach
– Allows real-time lighting updates

23. Cloud Rendering
• Full detail
– For close clouds
– Requires roughly sorted particles
• Update sorting only when significantly off
– Camera can be inside cloud
• Billboards
– For distant clouds
– Billboard updated as needed
• Angle to viewer has changed beyond a threshold
– Can’t tell the difference

24. Water
• Basic bump / specular
– Multiple detail normal maps based on
distance
• Water fogging
– Fades out around shoreline and objects
– Uses scene depth texture
• Distance to nearest object along view vector

25. Characters
• Skinned animation
– Extends to attached objects
• Death effect
– 3D texture animates a burn sequence
– Works with all game models
• No extra artist work

26. Particles
• Simple particle system
– Designed for speed
– Only moderate flexibility
• Particle rotation uses vertex program
– Less CPU work

27. UI
• Basic UI
– Controls inherit from template interfaces
• Basic element, value control
– Container, Text Box, List Box, Slider, Spin,
Button, Check Box, Icon, Tool-tip
– Maintains hierarchy of elements
• Draw order
• Input and updates

28. Deferred Rendering
• Lighting and fogging done as a post process
– Objects output color, normal, and depth (implicit)
– Full screen pass processes each stored pixel
– Fog color matches sky color
• Works well with unified lighting model
– One place to adjust lighting, easier to maintain
– All objects fit together
– Unique lighting models per-object harder
• ~15% speed up due to less overdraw

29. Deferred Issues 1
• GL_RGBA8 for best performance
– No noticeable quality hit for Felwyrld
• But not always good enough
– Normals, quality of faked AA
– Need 3 textures
• Color
• Eye space normal
– Can be two components with derived Z
• Depth (24-bit depth + 8-bit stencil)
– Not explicitly output, acts as hardware depth
buffer

30. Deferred Issues 2
• MRT best with smaller, same-width
formats
– Bandwidth
– Repeat fragment program execution
– Smaller formats also faster to sample
• MSAA issues
– Many potential solutions
• Involve tradeoffs in speed, simplicity, and
quality

31. Performance Measurement
• GPU / CPU counters and labels
– GL_EXT_timer_query