Optimizing Triangle Strips for Fast Rendering

1
Optimizing Triangle Strips for Fast Rendering
Francine Evans, Steven Skiena, and Amitabh
VarshneyState University of New York at Stony
Brook
Arthur LouieOctober 27, 2000
2
Outline

• Introduction
• Related work the SGI algorithm
• The STRIPE algorithm
• Results
• Impact of buffer size
• Summary and conclusions

3
Introduction

• Surfaces are represented as triangles.
• Rather than send all three vertices of each
  triangle down the pipeline, its far better to
  send triangle strips.
• This is the strip (1, 2, 3, 4, 5, 6, 7, 8), where
  triangle i is described by the vertices i, i 1,
  and i 2.

4
Generalized triangle strips

• The previous example used a FIFO queuing
  technique.
• For greater freedom, introduce a SWAP command.
• Example (1, 2, 3, SWAP, 4, 5, 6)
• SWAP is supported in Iris GL, but notOpenGL.
• Simulate SWAP (1, 2, 3, 2, 4, 5, 6)

5
The problem

• How do we construct good triangle strips from
  polygonal models?
• First step is triangulation.
• Triangulating a polygonal model for optimal
  strips is NP-complete!
• Thus we must focus on heuristic techniques.

6
Goals of the algorithm

• Maximize the length of each strip
• Minimize swaps
• Minimize the number of singleton strips

7
Related work the SGI algorithm

• A greedy algorithm the next triangle in a strip
  is the one with the least number of neighbors.
• Starts with lowest degree and extends outward.
• In case of first tie go one level deeper.
• In case of second tie choose arbitrarily.
• This is a local approach.

8
The STRIPE algorithm

• Exploits both local and global structure.
• Global technique patchification
• Try to find large patches consisting only of
  quadrilaterals.
• Triangulate patches along rows or columns.Cost
  either 3 swaps or 2 vertices per turn.
• Enforce a patch cutoff size
• Patchification is O(pn) wherep is the number of
  patchesfound.

9
The STRIPE algorithm

• Two approaches for exploiting the coherence of
  large patches
• Generate one strip along each row or column of
  the patch
• Minimizes number of swaps
• Generate one strip for the entire patch
• 3 swaps per turn
• In both cases, we removethe patches and use a
  localalgorithm for the remainingtriangulation.

10
The STRIPE algorithm

• Local algorithm 2 approaches
• Dynamic whole-face triangulation
• Completely triangulate each face
• Dynamic partial-face triangulation
• Ability to triangulate and walk only part of a
  face before exiting it
• Fewer swaps

11
The STRIPE algorithm

• Five ways to break ties
• Arbitrary
• Use the first face found among low-adjacency
  faces
• Look-ahead
• Same as SGI
• Alternate
• Alternate directions (because sequential strips
  alternate)
• Random
• Sequential
• Choose the next face that wont produce a swap
  (if none random)

12
Results

• Best approach for the OpenGL cost model
• global row/column strips
• patch cutoff size 5
• local whole-face triangulation

13
Results
14
Results
15
Impact of buffer size

• Typical buffer size is two vertices.
• Would it be worthwhile to increase?

16
(No Transcript)
17
Summary and conclusions

• Theres no practical need to develop better
  algorithmsSTRIPE is very close to being optimal.
• STRIPE is 10-30 better than the SGI algorithm.
• The better performance is due more to reducing
  the number of swaps than the number of strips.
• Increasing the buffer size helps, but the
  improvements diminish very quickly.
• A buffer size of about 8 is recommended.