Ray Tracing Animated Scenes using Coherent Grid Traversal

1
Ray Tracing Animated Scenes using Coherent Grid
Traversal

• Ingo Wald Thiago Ize Andrew
  Kensler
• Aaron Knoll Steven G. Parker
• SCI Institute, University of Utah

SIGGRAPH 2006
2
Abstract

• A new traversal techniques for uniform grids
• Achieves performance competitive with the fastest
  unknown kd-tree based ray tracer
• allows for per-frame rebuilds (dynamic scenes)

3
Outline

• Introduction
• Related Work
• Coherent Grid Traversal
• Experiments and Results
• Conclusion and Future Work

4
Introduction
5
Whats so special about grids?

• Since 70ies Lots of different RT data structures

6
Whats so special about grids?

• Since 70ies Lots of different RT data structures

BVH
7
Whats so special about grids?

• Since 70ies Lots of different RT data structures

BVH
8
Whats so special about grids?

• Since 70ies Lots of different RT data structures

BVH
9
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
10
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
11
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
12
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
13
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
14
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
15
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
16
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
Kd-tree
17
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
Kd-tree
18
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
Kd-tree
19
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
Kd-tree
20
Whats so special about grids?

• Since 70ies Lots of different RT data structures

Octree
BVH
Grid
Kd-tree
? Of all these, grid is only that is not
hierarchical !
21
Whats so special about grids?

• Grid is not hierarchical
• ? Much simpler to build
• Build-times in the paper 2.2M Soda Hall in 110
  ms
• ? Ideally suited for handling dynamic scenes
• Full rebuild every frame, no restrictions at all !

22
Related Work
23
Related Work

• All of the recent advancements of RT are for
  kd-trees !
• Wald 01 New concept ? coherent ray tracing
  (for kd-tree)
• Trace packets of coherent rays ? 10x faster
  than single rays
• Woop 05 First RT hardware prototype ? RPU
  (for kd-tree)
• Reshetov 05 New concept ? multilevel ray
  tracing (kd-tree)
• Trace packets using bounding frustum ? another
  10x faster than CRT !
• But kd-trees are too costly to build

24
Related Work

• SIGGRAPH 05 Dynamic Scenes huge problem
• Ray tracing has become very fast (MLRT 100fps)
• If ray tracing is to ever replace rasterization,
  it must support dynamic scenes (games)
• But All our fast RT algos are for kd-trees
• and kd-trees cant do dynamic scenes

25
Coherent Grid Traversal
26
Coherent Grid Traversal

• Grid too slow to traverse (vs kd-tree).
• Deliver to grids the same components that made
  kd-tree as fast as they are today
• Packets
• SIMD
• Frustum traversal

27
Issues with Packetized Grids

• Grid based ray tracers typically use 3D digital
  differential analyzers (3DDDA) to iterate through
  the voxels traversed by the ray.
• Chose only one cell at a time
• In which order to test the nodes ?
• ABCD or ABDC ?
• Split diverging packet ?
• Quickly degenerates to single-raytraversal
• Fix by re-merging packets ?
• Expensive cost

28
A Slice-based Packet Traversal for Grids

• First Transform all rays into the canonical grid
  coordinate system
• i.e., 0,0,0-Nx,Ny,Nz

29
A Slice-based Packet Traversal for Grids

• Pick the dominant component (the X, Y, Z axis)
  of the direction of the first ray.
• Major traversal axis

K
30
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays

K
31
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays

K
32
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell

K
33
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell

K
1
34
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell

K
1
2
35
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell

K
1
2
3
36
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell

K
1
2
3
4
37
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell
• For each slice, compute frustum/slice overlap

38
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell
• For each slice, compute frustum/slice overlap

K
U
V
39
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell
• For each slice, compute frustum/slice overlap
• One SIMD float-to-int truncation gives overlapped
  cell

40
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell
• For each slice, compute frustum/slice overlap
• One SIMD float-to-int truncation gives overlapped
  cell
• Intersect all cells in given slice

41
A Slice-based Packet Traversal for Grids

• Idea Consider only frustum, not set of rays
• Traverse slice by slice instead of cell to
  cell
• For each slice, compute frustum/slice overlap
• One SIMD float-to-int truncation gives overlapped
  cell
• Intersect all cells in given slice
• Loop incrementally compute nextslices overlap
  box
• One four-float SIMD addition

42
A Slice-based Packet Traversal for Grids

• The whole process of computing the next slices
  overlapped cell coordinates costs only two
  instructions
• One four-float SIMD addition
• One SIMD float-to-int truncation

43
Coherent Grid Traversal

• Grid usually less efficient than kd-tree

44
Coherent Grid Traversal

• Grid usually less efficient than kd-tree
• Cannot adapt to geometry as well ? more
  intersections test

Kd-tree
grid
45
Coherent Grid Traversal

• Grid usually less efficient than kd-tree
• Cannot adapt to geometry as well ? more
  intersections test
• Triangles straddle many cells ? re-intersection

46
Coherent Grid Traversal

• The overhead of these additional tests can be
  avoided using two well-known techniques
• SIMD Frustum Culling Dmitriev et al. 2004
• A SIMD 4-ray-triangle intersection test

Outside frustum ? cull
47
Coherent Grid Traversal

• The overhead of these additional tests can be
  avoided using two well-known techniques
• SIMD Frustum Culling Dmitriev et al. 2004
• Mailboxing Kirk and Arvo 1991
• Each packet is assigned a unique ID
• A triangle is tagged with that ID
• before the intersection test
• If a packet visits a triangle already
• tagged with its ID ? Skip

Mailbox detectsre-intersection
48
Impact of Mailboxing and Frustum Culling

• Compare with OpenRTs kd-tree system
• Mailboxing reduces tests by up to a factor of 2
• Frustum culling reduces tests by a factor of 4 to
  9

49
Experiments and Results
50
Impact of Packet Resolution
Best is 4x4 (green) or 8x8 (blue)
51
Performance for Static Scenes

• Compare to kd-tree
• To OpenRT 2x-8x faster
• To MLRT 3x slower
• Tests performed on kd-tree friendly models

52
Comparison to Single-Ray Grid Traversal

• Compare to an optimized single-ray 3DDDA
  implementation of a hierarchical grid
• Speedup 6.5x to 20.9x, usually 10x

53
Result

• Platform 10242, dual 3.2 GHz Xeon PC
• X/Y
• Xraycast only
• Yraytraceshadetextureshadows

Poser78K triangles15.8/7.8 fps
Hand16K triangles34.5/15.3 fps
Fairy174K triangles3.4/1.2 fps
Marbles8.8K triangles57.1/26.2 fps
Toys11K triangles29.3/10.2 fps
54
Conclusion and Future Work
55
Conclusion

• Have developed a new technique that
• Makes grid compatible with packets frusta
• Is competitive with BVHs and kd-trees
• Most general in handling dynamic scenes
• Future Work
• Hybrid approaches
• E.g., a kd-tree for static content and one
  separate grid for each animated character