KIPA Game Engine Seminars

1
KIPA Game Engine Seminars
Day 8

• Jonathan Blow
• Ajou University
• December 4, 2002

2
Small issues left over from yesterday

• Discuss cross-platformness of Keyboardevent
• Why I converted the 5km map example to
  centimeters
• Memory allocation and block loading
• Why block loading is kind of funny

3
Filter bandwidth

• When talking about the frame-time independent
  filters, I have been speaking of decay rate and
  not filter bandwidth
• When a question was asked in terms of bandwidth,
  it made me wonder why I do this
• Answer it has to do with uniform vs nonuniform
  sampling

4
Event Queues

• Show example of a big networking switch like we
  discussed yesterday
• Interesting structure where things can be sent
  noncentralized, but must be received centralized
• An indicator that the structure should be
  questioned?
• Example of in-game object destroy/create,
  hide/unhide
• Another reason to perhaps want that history query
  feature

5
Mesh Load/Save

• Looking through the code for the triangle strip
  mesh

6
Why the Tcp_Connection class

• Interface with Web services, etc

7
Looking at the networkprotocol code

• We talked about UDP vs TCP kinds of things, but
  never looked at the various classes
• Talk about the prevent-a-copy trick for
  wrapped_send() that I dont actually do yet
• Two-level structure of Net_Manager and
  Network

8
Rendering
9
BRDF Rendering

• (Show demo again)
• Discussion of the basic idea
• Two reasons to do this
• Better simulation of the real world
• Data-driven shaders

10
BRDFsAcquiring the data

• Gonioreflectometer measurements
• Virtual gonioreflectometer
• Raytracer and some microgeometry
• Analytic function
• Ward model

11
The Singular Value Decomposition (SVD)

• Factor A into VSW
• V and W are unitary
• S is diagonal, positive entries on
  diagonal,highest magnitude first
• Same as weighted sum of vectors

12
Using the SVDon BRDF data

• Our starting matrix is 4D, not 2D
• Our V, W are 2D matrices
• These matrices can be stored as texture maps, and
  multiplied together
• Just 2 texture units per piece of the
  approximation
• The parameterization of V and W are determined by
  the initial parameterization of A

13
SVD cost

• SVD is very expensive so we can use an
  approximation instead
• Kautz and McCool use the Normalized
  Decomposition (ND)

14
Dynamic Rangeon conventional hardware?

• Need to compute VskW and store that in the frame
  buffer
• Clamping in texture units causes problems!
• s too high?
• So we need to scale the s values to therange 0,
  1
• Then we suffer a lot of banding in the texture
  units
• This scaling needs to be the same for all objects
  in the scene
• One bright object makes the darker objects suffer

15
Solved on newer cards

• Radeon 9700, Geforce 5 FX
• Linear-light frame buffer (gamma 1), floating
  point pixels, exponentiate pixels when its time
  to page flip
• or let the CLUT / RAMDAC do the right thing
• Still not perfect
• Limited handling of fp pixels is available (no
  antialiasing, alpha blending)

16
Why shaders are not cool

• A shader is treated as a monolithic program
• Can only call subroutines within that program (no
  subroutines until pixel shader 2.0)
• There can be a combinatoric explosion of shaders
  based on what you want to do
• Example of different BRDF parameterizations