RACBVHs: Random-Accessible Compressed Bounding Volume Hierarchies Tae-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 Accumulated growth rate during 1999~2009 (log scale) Reducing data access time is critical! access speed disk 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 C1C1 C2C2 C3C3 BVH

Overview - Runtime BVH Access Framework Main memory Applications External drive Compressed data pool Cached data Cluster c 0 Cluster c 1 Cluster c 2 Cluster c 3 Cluster c 4 Cluster c 5 … Cluster c n … … n Offset table Request data Requested data Decompress using CPU power

Overview - Runtime BVH Access Framework Main memory Applications Compressed data pool Cached data Cluster c 0 Cluster c 1 Cluster c 2 Cluster c 3 Cluster c 4 Cluster c 5 … Cluster c n … … n Offset table Request data 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 C1C1 C2C2 C3C3

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 BV Left BV Right BV Prediction 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 n6n6 n6n6 n7n7 n7n7 n8n8 n8n8 n9n9 n9n9 n 10 n 11 n 13 n 14 n 12 n1n1 n1n1 n3n3 n3n3 n5n5 n5n5 n4n4 n4n4 n2n2 n2n2 n0n0 n0n0 n2n2 n2n2 n3n3 n3n3 n4n4 n4n4 n5n5 n5n5 Front n1n1 n1n1 n3n3 n3n3 n5n5 n5n5 n4n4 n4n4 n2n2 n2n2 n0n0 n0n0

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 data Cluster c 0 Cluster c 1 Cluster c 2 Cluster c 3 Cluster c 4 Cluster c 5 … Cluster c n … … 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. Matthew128 Iso surface102 Lucy28 Power plant13 CAD turbine1.8 Comp. ratio over uncomp. BVH 9.6:1 8.8:1 8.7:1 12:1 8.3:1 Compression bits per node TreeBV  16 bits quantization for BVs  4K nodes per cluster Size (MB) of uncomp. BVH 7,811 6,254 1,

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 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 & A Thank you!