Visualization of Industrial Structures with Implicit GPU Primitives

Rodrigo de Toledo

Bruno Levy

Plan

• Motivation

• Previous work

• Our contributions– New GPU primitives– Implicit information on GPU– Study-cases

• Results

• Conclusions

Motivation• Problem statement: Interactive visualization of

massive models is very difficult– Too much primitives (triangles)

• Proposition: Whenever it is possible, use implicit GPU primitives rather than triangles– Example: power plant

13 M triangles

200K implicit primitives+ 1M triangles

(2x speed-up + quality)

Reverse engineering

Previous Work (1/3)

• Extended GPU primitives:– 2004, “Extending the graphic pipeline with

new GPU-accelerated primitives”– Quadrics:

GPU primitives• Ray cast• Z-buffer update• Freely combined with triangle meshes• Advantages:

– Quality• Computations by pixel• Silhouette, phong etc…

– Speed• Very simple fragment shaders• LOD behavior

– Memory• Represented by few parameters

Previous Work (2/3)

• ISVC 2007, “Iterative Methods for Visualization of Implicit Surfaces on GPU”– Cubics and Quartics– TORUS

Multiple Tori• GPU primitives are

faster• 16000 tori

– Newton at 50 fps– the others < 1fps

GeForce 7900

Previous Work (3/3)TMCE 2008, “Reverse Engineering for Industrial-Environment CAD Models”

– Industrial environments are mainly composed by tubular structures (90%)

– Topological approach

Quadrics Torus

Our contributions

• New GPU primitives– Billboard cylinder– Truncated cone– Torus slice

• Implicit information encoded on GPU memory– Floating-point texture

• Two study-cases– Power-plant– Oil platform

Cylinders

• Billboard– only 4 vertices– tight projection enclosing the cylinder

perspective orthographic

gl_RECT(-1,-1,1,1); orgl_RECT(-u,-v,u,v);

• View-dependent coordinate system

• Computing perspective

Cylinders

• Including caps on cylinders in perspective

Cylinders

• Truncated and complete cones

Cones

Torus slices

• Adaptive polyhedra

• Up to 180º

• Very useful for CAD

Grouping primitive parameters

• Floating-point texture (each texel 4 scalars)• Read by vertices• Double speed• Our biggest example only uses 10MB

(340k primitives)

Study cases

• Power plant13M triangles

• Oil platform27M triangles

Reverse EngineeringResults

Memory reductionPPlant (without UT) PPlant (with UT)

Converted triangles 409 KB 482 KB

Primitives 6.5 KB 79 KB

Reduction factor 98 % 83 %

Speed-up (1/3)

• Test settings:– Without any culling

• They would disturb our results• Surely, culling algorithms would increase all frame rates

– We have compared GPU primitives against triangles– Triangles rendering:

• Multiple VBO (Vertex Buffer Object)

– GPU primitives in two situations:• Exclusively• Combined with unrecognized triangles

– GeForce 7900

Speed-up (2/3)

Triangles 12.7 MΔ

Unrecovered 1.2 MΔ

Primitives 200 K

Speed-up (3/3)

Image Quality SilhouetteSilhouette

ContinuityContinuity

IntersectionIntersection

Conclusion• Topological reverse engineering shows good

efficiency– but it is restricted to CAD models (depends on

regularity)

• With GPU primitives rendering we combine more quality, more speed and less memory

• Future work: new implicit GPU primitives for CAD models– GPU implementation + Rev. Eng.– Ex: half-sphere, half-ellipsoid, sheared cylinder

Thank you!

For questions please email us:

Gmail: rodrigodetoledo