RACBVHs: Random-Accessible Compressed Bounding Volume

HierarchiesTae-Joon Kim Bochang Moon Duksu Kim Sung-Eui Yoon

KAIST (Korea Advanced Institute of Science and Technology)

To appear at IEEE TVCG 09

Goal

Design a compact representation for interactive geometric and graphics applications of massive models Support random access

Support various applications

Improve performance of applications

Motivation – Massive Models

Take high disk/memory spaces

Long data access time and low I/O performance

St. Matthew model(372M triangles)

Power plant with Hugo model(82M triangles)

Low Growth Rate of Data Access Speed

Ran

dom

Seq

uent

ial

RA

M

CP

U

GP

U

Disk access access speed speed

1

10

100

1000

1.3X 4.5X

10X

56X 192X Accumulated

growth rate during 1999~2009

(log scale)

Reducing data access time is critical!

accessspeeddisk access speed

Large Scale Applications

Ray tracing

Collision detection

Visibility queries

Dynamic simulation

Motion planning

Large Scale Applications

Common geometric data structures Meshes

Acceleration hierarchies

k-d trees

Bounding volume hierarchies (BVHs)

Usually require random access on meshes and hierarchies

Supporting random access is important!

Our Approach

Compress bounding volume hierarchies (BVHs) to reduce expensive data access time Support random access

Support fast runtime decompression for performance improvements

Provide general API for various applications

Random-Accessible Compressed BVHs

Ray Tracing of St. Matthew Model

128M triangles (4GB)

256M BVH (8GB)

9.6:1 compression ratios for the BVH over original uncompressed BVH

4.4:1 runtime performance improvement

Related Work

Mesh compression

Compression and random access

Tree and BVH compression

Mesh Compression

Designed to achieve a maximum compression ratio or efficient transmission [Touma and Gotsman 98, Devillers and Gandoin 00]

Encode vertices [Alliez and Desbrun 01, Kälberer et al. 05]

Encode edges [Isenburg and Snoeyink 00]

Encode faces [Gumhold and Strasser 98, Rossignac 99, Lee et al. 02]

Do not directly support random access

Compression and Random Access

Quite common in video & audio encoding E.g., MPEG video

Single or multi-resolution mesh compression [Choe et al. 04, Yoon and Lindstrom 07, Choe et al. 09]

[Gobbetti et al. 06, Kim et al. 06]

Regular volumetric grids and images [Ihm and Park 99, Rodler 99, Lefebvre and Hoppe 06]

Tree and BVH Compression

Tree compression [Zaks 80, Zerling 85, Katajainen and Makinen 90]

Do not support random access

Encode bounding volumes Quantization: [Cline et al. 06, Mahovsky 05, Terdiman 03]

Hierarchical encoding: [Rusinkiewicz and Levoy 00, Hubo et al. 06]

Encode tree structures Assume a particular tree structure: [Lauterbach et al. 07 and 08]

Assume a particular layout for the tree: [Lefebvre and Hoppe 07]

Outline

Overview

Compression at preprocessing

Runtime data access framework

Results

Outline

Overview

Compression at preprocessing

Runtime data access framework

Results

Overview - Compression at Preprocessing

Decompose the BVHs into set of clusters

Compress each cluster separately Each cluster serves as an

access point C1 C2 C3

BVH

Overview - Runtime BVH Access Framework

Main memory

Applications

External drive

Compressed data pool

Cached dataCluster c0

Cluster c1

Cluster c2

Cluster c3

Cluster c4

Cluster c5

…Cluster cn

…

012345…n

Offset table

Requestdata

Requested data

Decompress using CPU power

Overview - Runtime BVH Access Framework

Main memory

Applications

Compressed data pool

Cached dataCluster c0

Cluster c1

Cluster c2

Cluster c3

Cluster c4

Cluster c5

…Cluster cn

…

012345…n

Offset table

Requestdata

Requested data

Preloaded data

Outline

Overview

Compression at preprocessing

Runtime data access framework

Results

Bounding Volume Hierarchies (BVHs)

A node of BVHs has: Indices of children nodes

A bounding volume (BV)

We use the axis-aligned bounding box (AABB) and a binary BVH Due to its simplicity and the wide use

Can be easily extended to other types of BVs and k-ary BVHs

BV

Index of child node

Layouts of BVHs

Order of stored BV nodes in memory Affect the performance of

applications

Our method preserves the original layouts Support any layouts

Use cache-efficient layouts for the best performance[Yoon and Manocha 06]

Layout of a BVH

Cluster-based Compression

Cluster has: Fixed number (e.g. 4K) of

consecutive nodes in the layouts

Compute clusters with cache-coherent nodes Implicitly performed by using

cache-coherent layouts of BVHs

C1 C2 C3

Encoding Bounding Volumes

Quantize min and max extents of BVs

Predict child BVs using their parent BV Use the simple median

prediction

Encode prediction errors Use a dictionary-based

encoder for fast decompression

Parent BVLeft BV

Right BVPrediction error

Front-based Tree Structure Encoding

Maintain a front of nodes, one of whose child nodes is uncompressed yet

Average size of front: 13 4K nodes with cache-efficient

layouts

Encode tree structures using only a few bits by connecting uncompressed nodes to a node in the front

: Not yet compressed nodes

: Compressed nodes

n6 n7 n8

n9

n10 n11 n13 n14

n12

n1

n3

n5

n4

n2

n0

n2 n3 n4 n5Front

0 1 2 3

n1

n3

n5

n4

n2

n0

Outline

Overview

Compression at preprocessing

Runtime data access framework

Results

Various applications can transparently access our representation using the following API:

Applications

BVH Access API

Main memory

External drive

Compressed data pool

Cached dataCluster c0

Cluster c1

Cluster c2

Cluster c3

Cluster c4

Cluster c5

…Cluster cn

…

012345…n

Offset table

Decompression

GetRootIndex (.)

GetBV (.)

IsLeaf (.)

GetLeftChildIndex (.)

GetRightChildIndex (.)

GetTriangleIndex (.)

Outline

Overview

Compression at preprocessing

Runtime data access framework

Results

Compression Results

Model Tri.(M)

St. Matthew 128Iso surface 102

Lucy 28Power plant 13CAD turbine 1.8

Comp. ratio over uncomp. BVH

9.6:18.8:1

8.7:112:18.3:1

Compression bits per node

Tree BV8.66 18.08.75 20.3

8.70 20.79.22 12.39.19 21.6

16 bits quantization for BVs

4K nodes per cluster

Size (MB) ofuncomp. BVH

7,8116,254

1,684778108

Benchmark Applications

Ray tracing Typically traverses large portions of BVHs

Collision detection Accesses smaller and more localized portions of BVHs than

ray tracing

Video for Benchmark Applications

Performance Improvement

Mainly due to higher I/O performance Reduce the data size by using compression→ Reduce or remove the expensive disk I/O time

Use CPU power to efficiently decompress the data

Parallel Random Access

Ray tracing of St. Matthew model (128M tri.) using only primary rays

1 2 3 40

0.51

1.52

2.53

3.54

4.5 Ideal, 4

Scalability of Ray Tracing

Number of Threads1 2 3 4

00.5

11.5

22.5

33.5

44.5 Ideal, 4

Original, 0.8

Scalability of Ray Tracing

Number of Threads1 2 3 4

00.5

11.5

22.5

33.5

44.5 Ideal, 4

Original, 0.8

Comp. Data, 3.6

Scalability of Ray Tracing

Number of Threads

Speedup

Limitation

Lower (by up to 3%) the performance with small models which can fit in main memory Overhead of identifying clusters

Less tight BVs due to a conservative quantization

Summary & Conclusion

Low storage requirement Achieve up to a 12:1 compression ratio

Improved performance with random accessibility Achieve a 4:1 runtime performance improvement

Wide applicability Allows various applications to access the compressed BVHs

High scalability Achieve a 3.6:1 performance improvement when using 4

threads over single thread

Future Work

Compact in-core representations

Apply to highly parallel architectures GPU

Larrabee

Apply to levels-of-detail (LOD) hierarchies

Acknowledgments

Members of SGLab. in KAIST

Model contributors

Funding agencies KAIST seed grant

Ministry of Knowledge Economy

Samsung

Microsoft Research Asia

Korea Research Foundation

Any Questions?

Q & AThank you!