1

Schedule
Part 1 - 2:00pm Introduction to OpenVDB, Ken Museth (DWA) - 2:45pm Overview of toolset, Mihai Alden (DWA) - 3:15pm Break Part 2

SIGGRAPH2013

OpenVDB

- 3:30pm Adoption at DreamWorks Animation, Jeff Budsberg (DWA) - 4:00pm Adoption at Digital Domain, John Johansson (DD) - 4:30pm Adoption in Houdini, Edward Lam (SideFX Software) - 5:00pm Concluding remarks and questions, Ken Museth (DWA)
2

Course Material

SIGGRAPH2013

OpenVDB

Main site: http://www.openvdb.org - Course slides and hip les - Coding cookbook - FAQ - Google group: OpenVDB Forum Technical paper on VDB (http://ken.museth.org)
3

History of VDB
2000
2001

SIGGRAPH2013

OpenVDB

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

DT-Grid H-RLE DB-Grid DB+Grid VDB OpenVDB

4

Motivating VDB

SIGGRAPH2013

OpenVDB

Consider extremes
- DT-Grid / H-RLE (level sets only) - DB-Grid (doesnt scale and non-hierarchical) - Octrees (slow tree traversal)

Dichotomy - Static versus dynamic data structure

5

Facts
DWA feature lms using VDB - Puss in Boots, Rise of the Guardians, The Croods, Turbo OpenVDB developers - K. Museth, P. Cucka, M. Aldn and D. Hill Availability - http://www.openvdb.org - https://github.com/dreamworksanimation/openvdb_dev License - Mozilla Public License version 2 and CLA

SIGGRAPH2013

OpenVDB

Library Versioning
Patch: Minor:
major.minor.patch

SIGGRAPH2013

OpenVDB

- No change to API, le format or ABI of Grid or its member classes - Change to API but not Grid ABI; backward-compatible le format Major: - Change to ABI of Grid or non-backward-compatible le format No release guarantees complete ABI compatibility! - Library is namespaced on the complete version number
7

SideEffects Houdini

SIGGRAPH2013

OpenVDB

Scott Keating & Jeff Wagner, SideFX

Commercial Renderers

SIGGRAPH2013

OpenVDB

SideFX Mantra

Terminology
Voxel
- Smallest addressable unit of index space - Resides at the leaf node level

SIGGRAPH2013

OpenVDB

Tile
- Larger constant domain of index space - Resides at the upper (non-leaf) tree levels

Active state
- All values (voxels and tiles) have a binary state - Interpretation of state is application-dened
10

VDB Data Structure

Child pointers Root node (unbounded) Tile values with active/inactive states

SIGGRAPH2013

OpenVDB

Internal Node 1

Active Mask Child Mask Tile values / Child pointers

Internal Node 2

Leaf Node

Active Mask Voxels

Bit Mask Compression

11

Grid/Tree Class Structure

Tree
Root Node Internal Node Leaf Node Voxels Accessor Registry Background Bit Mask Bit Mask Tiles Tiles

SIGGRAPH2013

OpenVDB

Grid

Meta Map Transform

Linear Maps Frustum

12

Sparsity and Efciency

Memory efciency - Sparse, hierarchical tree structure - Allocation on insert - Custom compression schemes on I/O Computational efciency - Hierarchical algorithms - Sparse computations - C++ template metaprogramming - Bit tricks, multithreading and SIMD

SIGGRAPH2013

OpenVDB

13

Grid Types
Fully compile-time congurable - Value type - Node size, Tree depth, e.g. tile-grid, Octree, N-tree etc.
typedef LeafNode<float,3> N0; typedef InternalNode<N0,4> N1; typedef InternalNode<N1,5> N2; typedef RootNode<N2> RootType;

SIGGRAPH2013

OpenVDB

typedef Tree<RootType> TreeType; // FloatTree typedef Grid<TreeType> GridType; // FloatGrid

14

Common Scalar Grid Categories

SIGGRAPH2013

OpenVDB

Fog or density volumes - Scalar normalized density [01] - Background zero - Interior voxels have value 1 - Active tiles SDFs or narrow band level sets - Scalar truncated signed distance ] 2, 2[ - Background is the positive half-width of the narrow band ( 2 ) - Interior is the negative half-width of the narrow band ( 2 ) - No active tiles
15

Fundamentals of Level Sets

SIGGRAPH2013

OpenVDB

16

Narrow Band Level Sets

SIGGRAPH2013

OpenVDB

InternalNodes: 10243 InternalNodes: 643 LeafNodes: 43 Voxels: 13 228 million sparse voxel vs 69 billion dense voxels

Active voxel span: 7897 x 1504 x 5774

>1GB memory footprint vs 1/4 TB dense grid

17

Properties of Level Sets

Closest signed distance Interface normals Closest Distance Transform Mean curvature Surface deformations
( x ) = 1

SIGGRAPH2013

OpenVDB

18

Morphological Operations
Opening Closing

SIGGRAPH2013

OpenVDB

erosion

dilation

dilation

erosion

19

math::Transform
Transforms Index Space (coordinates)

SIGGRAPH2013

OpenVDB

World Space (locations)

- Vec3d Transform::indexToWorld(const Vec3d& xyz) const; - Vec3d Transform::indexToWorld(const Coord& ijk) const; - Vec3d Transform::worldToIndex(const Vec3d& xyz) const;

Supports - Afne Transforms (mix of scaling, rotation, shearing, translation) - Frustum Transforms - Updates with pre and post semantics Wraps a minimal representation BaseMap - Virtual hit in calling BaseMap
20

Maps: Reduced Representation

Distinct implementations for common transforms - MapType: Scale, Translation, Frustum, etc. - Minimal number of oating point operations For higher performance, template on MapType - No virtual hits - math::processTypedMap

SIGGRAPH2013

OpenVDB

Used in vector calculus operators (Gradient, Curl) in world space - Computes rst and second derivatives of the transform - Chain Rule-based logic implemented for various operators
21

Data Access
Sequential access - Fundamental to simulations - Access elements according to memory layout Random access - Coordinate-based access: getValue(x,y,z) - Spatial coherence Stencil access - Fundamental to nite differencing, ltering and interpolation - Often combined with sequential or random access

SIGGRAPH2013

OpenVDB

22

Iterators
Bit mask - Located in all nodes - Facilitates efcient iterators
Grid/Tree/NodeType::ValueOnIter Grid/Tree/NodeType::ValueOffIter Grid/Tree/NodeType::ValueAllIter

SIGGRAPH2013

OpenVDB

NodeType = {RootNode,InternalNode,LeafNode} Tree::LeafIter Tree::NodeIter

23

tree::LeafManager
Fundamental assumptions - Voxel processing only, so safe to ignore tiles - Static topology Linearize tree as an array of leaf nodes - Lightweight

SIGGRAPH2013

OpenVDB

- Very fast iteration and trivial split operator for multithreading Temporal value buffers for each leaf node - Solve time-dependent PDEs, ltering and convolution - Multithreaded swapping of buffers
24

tree::ValueAccessor
Random or coordinate-based access - Should never be truly random! - Virtually always spatially coherent Improved random access - Cache nodes - Inverted tree traversal - On average shorter path - Amortize overhead of slow dynamic RootNode

SIGGRAPH2013

OpenVDB

25

Stencil Operations
Finite differencing - Gradient, curvature, Laplacian, curl, etc. - Central-difference schemes (2nd6th order) - Upwind schemes (1st5th order) Interpolation - Nearest neighbor: tools::PointSampler - Trilinear: tools::BoxSampler - Triquadratic: tools::QuadraticSampler - Staggered variants for MAC grids

SIGGRAPH2013

OpenVDB

26

Optimization
Use ValueAccessor for random access Use sparse iterators for sequential access

SIGGRAPH2013

OpenVDB

Use LeafManager when applicable (i.e., no tiles and static topology) Disable asserts (-DNDEBUG) Use library tools and routines (dont reinvent the wheel!) If possible, partition algorithm into topology and value updates - Use topology copy constructors - Use setValueOnly Use multithreading
27

Thread Safety
Non-thread-safe operations - Insert new values (due to allocate-on-insert) - Change state of a value (due to bit masks)

SIGGRAPH2013

OpenVDB

- Read or write via ValueAccessor (due to internal node caching) Thread-safe operations - Random or sequential read of constant values and states - Modify values in existing nodes

28

Threading Strategies
Thread over values or nodes using sparse iterators Consider using tree::LeafManager if applicable Assign one tree::ValueAccessor to each thread Reuse ValueAccessors as much as possible

SIGGRAPH2013

OpenVDB

- Never allocate a new ValueAccessor per access operation - In extreme cases use thread-safe access methods on Tree

29

Threading Strategies
If the topology of the output grid is known

SIGGRAPH2013

OpenVDB

- Pre-allocate output grid with correct topology (use topologyCopy) - Use tools::foreach or tbb::parallel_for over output grid - Use setValueOnly If the topology of the output grid is unknown - Assign an empty grid to each computational thread - Use tools::transformValues or tbb::parallel_reduce - Use hierarchical merge or compositing in thread join - Use concurrent malloc, e.g., tbb_malloc or jemalloc
30

End

SIGGRAPH2013

OpenVDB

Questions?

31

Looking further ahead

Better support for accelerated volume rendering Out-of-core rendering support Accelerated particle storage and query Multiresolution tools and grids (e.g., MipMaps) Python bindings Level set morphing Sparse density advection More lters (sharpening, anisotropic lters) Particle skinning by metaballs (vs LS) C++11 support? (gcc 4.5+, icc11+) [What are we missing?]

SIGGRAPH2013

OpenVDB

32

End

SIGGRAPH2013

OpenVDB

Questions?

33