1 A Novel Page-Based Data Structure for Interactive Walkthroughs Behzad Sajadi Yan Huang Pablo Diaz-Gutierrez Sung-Eui Yoon M. Gopi.

1

A Novel Page-Based Data Structure for Interactive

Walkthroughs

Behzad SajadiYan Huang

Pablo Diaz-GutierrezSung-Eui Yoon

M. Gopi

2

Outline

Page-Based Rendering Data Fetching Algorithm 2-Factor Data Layout Implementation and Results Summary Conclusion and Future Work

3

Traditional Rendering Workflow

2

365

7

1

4

8

4

Page Based Rendering Workflow

2

3

1

4 65

78

5

Page Format

Self contained information Less bytes for vertex indices Effective compression

# vertices# vertices # faces# faces List of vertices with attributes

List of vertices with attributes

Indexed triangle list with attributes

Indexed triangle list with attributes

Eliminating fragment accesses

6

Traditional K-d tree Page based K-d tree

Storage

OScache

Processheap space

High data management

cost

No data management

cost

Low data management cost

7

Major Steps of the Workflows

Traditional rendering workflow

Page based rendering workflow

SceneData

structureLayout

LayoutApproximated

sceneData

structure

8

Advantages

Less space required for the data structure K-d tree on pages instead of triangles

Independence of the layout and data structure Any layout can be converted to the page

based format

9

Outline

Page-Based Rendering Data Fetching Algorithm 2-Factor Data Layout Implementation and Results Summary Conclusion and Future Work

10

Rendering Flow

K-d treestructure

RenderFetchdata

Page numbersafter VF Culling

Triangles

2

3

1

4

11

Sorting the Page Numbers

Access pattern is based on storage

…

..

.

K-d tree without sorting

…

..

.

1,72,5 4,12 5,7 9,14 8,16

1, 2, 4, 5, 7, 8, 9, 12, 14, 16

K-d tree after sorting

1,72,5 4,12 5,7 9,14 8,16

12

Rendering Flow

K-d treestructure

Sort pagenumbers

RenderFetchdata

Page numbersafter VF Culling

Triangles

1, 2, 4, 5

2

3

1

4

13

Backface Culling

14

Rendering Flow

K-d treestructure

Sort pagenumbers

RenderFetchdata

BackfaceCulling

Page numbersafter VF Culling

Triangles

1, 2, 4, 5

2

3

1

4

15

Outline

Page-Based Rendering Data Fetching Algorithm 2-Factor Data Layout Implementation and Results Summary Conclusion and Future Work

16

Related Works

Cache oblivious layout by Yoon et al. [2006]

Space filling curves Morton layout (Z-Order) Hilbert layout

17

Graph Construction

Nodes: Group of triangles Edge weights: Distance between the

primitives of the nodes

23

2.5

41.8

6.7

5

5.3

85

18

Steps of the Method

Use GLA (Generalized Lloyd's Algorithm) to partition the primitives

Use [Diaz-Gutierrez and Gopi 2005] to get a 2-factor

Iterate in a hierarchical manner

19

The Layout Hierarchy

Ordering using2-Factor Layouts

Partitioning usingGLA clustering

1 23 4 5 6 7

Final Layout: 2 1 6 7 4 5 3

20

Advantages

Operates globally Scalable Amenable for multiple proximity

measures

21

Outline

Page-Based Rendering Data Fetching Algorithm 2-Factor Data Layout Implementation and Results Summary Conclusion and Future Work

22

The City Model

Floor planned using role playing city map generator 5.40

23

The City Model (Contd.)

3D models of houses and trees were replaced

Cars were added in the streets

24

The City Model (Contd.)

Specifications 110 million triangles 90 million vertices Spans 4528 MB Around 10,000 objects 115 million vertices in page format Spans 3814 MB in page format

25

Texture Based Simplification

[Aliaga and Lastra 1997]

26

Statistics

Frame rates Raster scan layout: 20 FPS Cache oblivious layout by Yoon et al.

[2006]: 27 fps 2-factor layout: 28 FPS

Rendering statistics (our layout) Page size: 4 KB Average disk-page per frame: 2365 Average triangle per frame: 240 K

27

Statistics (Contd.)

Computation times Page format conversion: 15 min (Offline) K-d tree construction: 5 min (Offline) Billboard rendering: 20 hour (Offline) Program initialization: 20 sec

Consumed memory Billboards: 640 MB K-d tree: 96 MB (including bounding

boxes and normal cones)

28

29

Outline

Page-Based Rendering Data Fetching Algorithm 2-Factor Data Layout Implementation and Results Summary Conclusion and Future Work

30

Before Rendering (Offline) Steps

Given layoutConvert to thepage format

Construct thedisk-page hierarchy

2

3

1

4

Rendering (Online) Steps

K-d treestructure

Sort pagenumbers

RenderFetchdata

BackfaceCulling

Page numbersafter VF Culling

Triangles

1, 2, 4, 5

2

3

1

4

32

Outline

Page-Based Rendering Data Fetching Algorithm 2-Factor Data Layout Implementation and Results Summary Conclusion and Future Work

33

Conclusion

A simple data structure on the disk pages Simplicity High Performance Generality

A new cache oblivious layout

34

Future Work

Analyzing the disk-page hierarchy on other data structures

Exploring other applications that can make use of this data structure

Including simplification techniques Adding a cache management system Analyzing the number cache hits and

misses

35

Questions?