Highly Scalable Character Rendering

1
Highly Scalable Character Rendering
Tom ForsythGraphics coderMuckyfoot Productions
Ltd
TomF_at_muckyfoot.com
2
The Problem

• Crowd scenes
• Close-up faces
• Detailed scenery
• Consistent level of detail for world
• Clutter
• Lots of incidental stuff
• Smooth transitions

3
Detailed people worlds

• High poly count
• Highly animated - lots of bones
• Multiple materials textures
• Lots of rendering passes
• VS PS effects everywhere

4
Crowd scenes clutter

• Low poly count per object
• Few bones
• Single material
• Single rendering pass
• Maintains good batching
• Allows sort-by-object-type
• But not necessarily simple shaders

5
Methods

• Tessellation displacement
• VIPM
• But VIPM cant do everything
• Displacement vertices
• Update of Meltdown 2000 talk
• Progressive bones
• Scalable shaders
• Lose layers effects gradually
• Friendly to older hardware

6
Tessellation displacement

• Author-time phase
• N-patches with creases
• Or whatever your artists want
• Then displace by bumpmap
• Artist preview tools very important
• Consistent detail for whole world
• Reduces artist load e.g. vehicle wheels
• Copes with obsolescence disasters

7
VIPM

• Using sliding window method
• Very low LoD change cost
• No editing of index or vertex buffers
• Zero instance memory cost
• No editing means its all shared
• Some global memory cost
• Most stored in memory-mapped files
• Most used for high detail objects
• See Games Programming Gems 2

8
Displacement Vertices

• Base mesh defined
• Certain (low) level of VIPM mesh
• Base mesh animated lit fully
• Rest of vertices interpolated
• 3 base vertex indices
• I, J, displacement
• Norm N1 I(N2-N1) J(N3-N1)
• Pos P1 I(P2-P1) J(P3-P1) dispNorm

9
Displacement Vertices (2)

• Reduced vertex bandwidth
• 8 bytes per vertex
• Only base mesh needs processing
• Usually 100 verts
• Number changes with LoD level
• Loaded into VS constant space
• Can require mesh partitioning
• Dot3, BDRF, etc processing done here

10
Displacement Vertices (3)

• Displacements from hi-rez mesh
• Use U,V coords to find I,J values
• Get closest disp value possible
• Anneal one of I,J,disp for a few goes
• Easy linear equations
• Usually converges very quickly
• Changing I,J changes U,V
• Usually doesnt matter
• Artists post-warp textures where it does

11
Progressive Bones

• Extremely detailed bone model
• Collapse bones together
• Child-parent (e.g. foot - leg)
• Child-sibling (e.g. finger - finger)
• Collapses according to LoD
• Same LoD value as used for VIPM
• Uses sliding window levels
• Each animation reduces further
• e.g. different sorts of wave

12
Progressive Bones (2)

• Bones stored in collapse order
• Remap from vertex bones-real bones
• Animations further modify remap list
• Can collapse one bone to a lesser bone
• Anim blending causes conflicts
• Do each remap in pre-blend copy
• Do neither when rendering mesh
• Rare - only happens in transitions
• Area-specific anims work fine

13
Material Effects

• Vertex calcs done on base mesh
• Usually by CPU
• Use software VS processing for easy SIMD
• Linearly interpolated for non-base verts
• Hardware VS could do some calcs
• More per-vertex data - bandwidth cost
• If VS not already saturated, raise LoD!
• Multiple materials at high LoD
• Transition to single material at low LoD
• Transitions at VIPM level change

14
Material Effects (2)

• Dot 3 bumpmapping
• With VIPM, better lighting than vertex
• Do for first n lights - rest are vertex
• n based on LoD of course
• Do specular properly
• Power done by dependent texture lookup
• Or can cheat
• Use envmap specular trick
• Modulate by bumpmap Z
• Looks fine very cheap
• Works on more hardware

15
Material Effects (3)

• Emboss bumpmapping
• Lower texture bandwidth than Dot3
• Vertex calcs almost identical to Dot3
• Works on old hardware
• Better than no bumpmapping
• Similar specular/envmap cheat
• EMBM
• Another good cheat
• Better specular than emboss

16
Material Effects (4)

• Parallax texturing
• Read diffuse map twice, once with shifted UV
  coords
  
• Same UV calcs as emboss bumpmap
• Uses view vector instead of light vector
• Blend by heightfield in alpha channel
• Same data as emboss bumpmap!
• Vertex displacement modulates blend
• Fake parallax - real parallax with VIPM
• Cheap version of fur

17
Material Effects (5)

• Emissive map
• Special effects (equipment, power ups)
• Fade with LoD
• Component still retained in diffuse map
• Glossmap
• Env map or RTT cube-map
• Modulate by alpha channel of diffuse
• Fade with LoD
• Rise ambient or specular to compensate

18
Material Effects (6)

• Self-shadowing bumpmaps
• Peter-Pike Sloans method
• Multiple textures - more accurate
• My method
• Single 3D texture - quicker
• Less hardware support
• Both need multiple passes
• Both use texkill / alpha-test / stencil
• But shadows only significant up close
• Both still to be published

19
Material Effects (7)

• ID-based Shadowmaps
• Variable number of ID sections
• Each segment shadows others
• Based on same LoD value yet again
• VIPM-style list of IDs
• Do merges, etc as for VIPM
• Remember different LoD often wanted when
  rendering shadowmap
  
• A bit hard to smoothly change number
• Cant really fade the shadow in or anything

20
Summary

• Scalability is king
• Copes with changing platform
• Copes with widely varying scene types
• Maximum available detail at all times
• Simple artist tools
• Full automation not required or wanted
• Use what works
• Perfect algorithms are expensive
• Hacks look pretty good for far less cost

21
Questions...

• www.muckyfoot.com
• TomF_at_muckyfoot.com

22
Main References

• Charles Bloom (www.cbloom.com)
• Excellent discussion of VIPM methods
• Peter-Pike Sloan
• Self-shadowing bumpmapping
• More details on website
• www.muckyfoot.com - downloads section