KD-Tree Acceleration Structures for a GPU Raytracer

1
KD-Tree Acceleration Structures for a GPU
Raytracer

• Tim Foley, Jeremy Sugerman
• Stanford University

2
Motivation

• Accelerated raytracing
• On commodity HW
• Production rendering
• Real-time applications?
• Performance trend
• 9800 XT 170M ray-triangle intersects/s
• X800 XT PE 350M ray-triangle intersects/s

3
GPU Raytracing

• Promising early results
• Simple scenes
• Uniform grid
• Problems with complex scenes
• Hierarchical accelerator (kd-tree)
• Improve scalability

4
Outline

• Background
• GPU Raytracing
• KD-Tree Algorithm
• KD-Restart, KD-Backtrack
• Results
• Future Work

5
Background

• RayEngine Carr et al. 2002
• Parallel ray-triangle intersection
• Host controls culling
• Purcell et al. 2002
• Entire raytracing pipeline
• Many rays required for efficiency
• Uniform Grid

6
Why not KD-Tree?

• Uniform grid acceleration structure
• Regular structure efficient traversal
• Regular structure poor partitioning
• KD-Trees
• Adapt to scene complexity
• Compact storage, efficient traversal
• Best for CPU raytracing Havran 2000

7
KD-Tree
tmin
Z
X
B
Y
D
C
A
tmax
8
KD-Tree Traversal
9
Per-Fragment Stacks

• Parallel (per-ray) push
• No indexed write in fragment program
• Per-ray stack storage
• Ernst et al. 2004
• Emulate push with extra passes
• Impractical, slow

10
Our Contribution

• Stackless kd-tree traversal algorithms
• KD-Restart
• KD-Backtrack

11
Observation
Current leafs tmax
Next leafs tmin

12
KD-Restart

• Standard traversal
• Omit stack operations
• Proceed to 1st leaf
• If no intersection
• Advance (tmin,tmax)
• Restart from root
• Proceed to next leaf

13
KD-Restart

• Restart traversal after each leaf
• m leaves
• Average depth d
• Cost O(md)
• Balanced tree of n nodes
• Upper bound O(n log(n))
• Standard algorithm O(n)
• Expected O( log(n) )

14
Observation
Ancestor of A is parent of Z
15
KD-Backtrack

• If no intersection
• Advance (tmin, tmax)
• Start backtracking
• If node intersects (tmin, tmax)
• Resume traversal
• Proceed to next leaf

16
KD-Backtrack

• Backtrack after leaf
• Revisits previous nodes
• At most twice from left, right
• Within constant factor of standard traversal
• Upper bound O(n)
• Expected O( log(n) )
• Requires additional storage
• Parent pointers
• Bounding boxes for internal nodes

17
Implementation

• Built GPU raytracer in Brook Buck et al.
• 4 intersection schemes
• Brute Force
• Uniform Grid
• KD-Restart
• KD-Backtrack

18
Scenes
Stanford Bunny 69451 triangles
Cornell Box 32 triangles
BART Robots 71708 triangles
BART Kitchen 110561 triangles
19
Results
Box
Bunny
Robots
Kitchen
12.9
Relative speedup over brute-force intersection.
20
Results
Ideal Restart Backtrack
Traverse 10.86M 21.80M 10.86M
Backtrack 0 0 7.78M
Intersect 5.91M 5.91M 5.91M
Rays in each state throughout traversal.
21
Discussion

• Absolute performance
• Trails best CPU implementations 5-6x
• Sources of inefficiency
• Load balancing
• Data reuse

22
Load Balancing

• Subset of rays intersecting, traversing
• Occlusion queries to select kernel
• Early-Z to cull inactive rays
• Approximately 5x overhead
• Query, kernel switch overhead
• Worse with fewer rays

23
Data Reuse

• Every kernel
• Loads ray origin/direction
• Load/Store traversal state
• Consumes streaming bandwidth
• We are bandwidth-limited
• CPU implementation stores these in registers

24
Branching

• Merge multiple passes into larger kernel
• Fragment branches for load balancing
• Avoid load/store of reused data
• Current branching has high overhead
• Shifts efficiency burden to HW

25
Conclusion

• Stackless Traversal
• Allows efficient GPU kd-tree
• Scales to larger, more complex scenes
• Future Work
• Changes in HW
• Alternative acceleration structures
• Out-of-core scenes
• Dynamic scenes

26
Acknowledgements

• Tim Purcell (NVIDA)
• Streaming raytracer
• Mark Segal (ATI)
• Demo machine
• NVIDIA, ATI HW
• DARPA, Rambus Funding

27
Questions