1 Plan for today Computer Graphics as Virtual Photography Small - - PDF document

1

Ray Tracing Optimizations

Paper Summaries

• Any takers?

Assignments

• Checkpoint 1

– All graded – Didn’t get a grade? See me!

• Checkpoint 2

– Due Thursday(?)

Projects

• Proposals due tonight

Logistics

• Job related news

– Co-op orientation

• Friday, Sept 16th 12:00 – 1:30pm 70-1400
• Friday, Oct 14th 1:00 – 2:30pm Eastman Aud.

– Career Fair

• Wednesday, Sept 28th
• 11am – 4pm
• Gordon Field House

– EA Job offer

One more announcement

• Electronic Arts (EA) is coming to RIT.

– Looking for a few (actually more than a few) good gaming programmers – EA Company Presentation

• Tuesday, October 4th
• 6-7:30pm
• Golisano auditorium

– Interviews

• Wednesday, October 5th (Career Center)

2 Plan for today

• Small change in plans

– Today:

• Ray tracing optimization
• Intro to Material Properties/BRDFs/Illumination

Models

– Thursday:

• Practical considerations for Illumination Modeling
• Raytracer – Checkpoint 3

Computer Graphics as Virtual Photography

camera (captures light) synthetic image camera model (focuses simulated lighting)

processing

photo processing tone reproduction real scene 3D models Photography: Computer Graphics: Photographic print

Ray Tracing

• Integrated aspects of light and object

interaction that had formerly been handled by separate algorithms:

– Hidden surface removal – Reflection – Refraction – Shadows – Global specular interaction

Turner Whitted

Ray Tracing - Problems

• Object - ray intersection
• Computation time
• Ray traced images are point sampled

– “Too sharp” (super real) – “wrong image” – Sharp shadows – Sharp Reflection/Refraction

• Multiple reflections especially are too sharp

– Aliasing

• Doesn’t handle major light transport functions

– Diffuse interaction – Scattering of light – Caustics

Ray Tracing: Intersection

• We only talked about ray intersection with spheres

and planes

• What about?

– Boxes – Cylinders – Cones – Triangles – Spline surfaces – Irregular shapes

Ray Tracing - Object - Ray Intersection

• For each ray, intersection test needs to be

made for each object

– This could costly if you have many

• bjects…consider if object has n polygons, n

very large! – Solutions

• Bounding Volume
• Spatial Subdivision

3

Ray Tracing - Bounding Volumes

• Place simple objects (i.e., sphere or box)

around complex objects

– Suitability depends on object being enclosed!

• Do initial intersection tests on bounding
• bjects.
• If ray intersects bounding volume, then test

complex bounded object

• Can be nested

Ray Tracing - Bounding Volumes

[Watt/Watt,235]

Ray Tracing - Bounding Volumes

[Watt/Watt,236]

Ray Tracing - Spatial Subdivision

• Motivation

– Without spatial subdivision, for each ray, you will need to query all objects/polygons and test for intersection – With spatial subdivision, you know which

• bjects are in each volume so you only test
• bjects that are in the volumes where a ray is

traveling.

Ray Tracing - Spatial Subdivision

• Subdivide your scene volume into hierarchical

regions

– Octrees – BSP Trees

• Create a tree structure that indicates for each region:

– if the region is empty – the object present at that particular region

• Test ray intersection with region volume

Ray Tracing - Octrees

• Recursively subdivide volume into equal

regions.

• If subregion is empty or contains a single
• bject, then stop
• Otherwise, further subdivide.
• Continue until each subregion is empty or

contains a single object.

4

Ray Tracing - Octrees

[Foley/Van Dam]

Ray Tracing – Oct(quad)trees

• Subdivisions represented as a tree

[Watt/Watt,244]

Ray Tracing - Oct(quad)trees

• Quadtree applet

http://njord.umiacs.umd.edu:1601/users/brabec/qu adtree/points/pointquad.html

Ray Tracing -- Octrees

• Images from

http://www.flipcode.com/tutorials/tut_octrees.shtml

Ray Tracing -- Octrees Ray Tracing -- Octrees

5 Ray Tracing -- Octrees

Shows viewing frustum and which “octants” need to be checked.

Ray Tracing - BSP Trees

• Binary Space Partitioning Trees (BSP

Trees)

– Like Octrees but divides space into a pair of subregions – Subregions need not be equally spaced – Planes separating regions can be placed at

• bject boundaries.

Ray Tracing - Spatial Subdivision

• Octrees vs BSP Trees

[Watt/Watt,246]

Ray Tracing - BSP Trees

• BSP Trees – Each non-terminal node represents a single

partitioning plane that divides space in two

[Watt/Watt,247]

Ray Tracing - BSP Trees

• BSPTree applet

http://symbolcraft.com/graphics/bsp/index.html

Ray Tracing - Spatial Subdivision

• Advantages

– Efficient means for finding objects within your space – Compact representation

• Disadvantages

– Preprocessing required – If scene changes, must rebuild your tree – Not foolproof

6

Ray Tracing - Spatial Subdivision

• Octrees vs BSP trees

– BSP Trees are generally more balanced than Octrees – BSP Traversal more efficient – BSP Trees have additional storage overhead as must store subdivision planes

Parallel Computing

• What is it?

– Solving a problem by dividing it into tasks to be handled by separate processors

• Why use it?

– Performance – Goal linear speedup

Parallel Hardware

• SIMD

– Same instruction executed in lockstep fashion on different data

• MIMD

– Different instructions executed

• n different data
• MIMD-SM
• MIMD-DM

Parallel vs. Distributed

Parallel Distributed

Issues in Parallel Computing: Communication

• More processors means more organization

is required (Consider working on a team of 10 people vs. 1000 people)

• For performance, each processor must have

a lot of work to do in comparison with any need to communicate to another processor.

• Keep all processors busy (load balancing)

RIT Ray Tracing 1991

975 objects 3 light sources 2048 X 2048 pixels Sun 4/490 – 1991 1.5 hours CPU time

7 RIT Fractals and Ray Tracing 1991

87,000 objects 5 light sources 1024 X 1024 Sun 4/490 – 1991 18 hours CPU time

Ray Tracing

• Ray tracing lends itself nicely to parallel

programming because the work can easily be sub-divided

– Per pixel calculation, i.e., not dependent on

• ther pixels

– Potentially large calculation time

• How to distribute work?

Image Space Partitioning

Strip = Section of Viewing Plane per Processor

Data Parallel strip division Processor 1 Processor 2 Processor 3 Processor 4 Processor 1 Processor 2 Processor 3 Processor 4 Data Parallel: Each processor does the rows marked with its color

Object Space Partitioning

Voxel = Section of scene data for processor

8 Object Partitioning

Processor 3 Processor 2 Processor 1

Reinhard’s Parallel Example

Adaptive Partitioning to Load Balance

[Reinhard99 – thesis]

Partitioning for Ray Tracing

• Object Space Partitioning

– May be necessary if very complex scene! – Less local processor memory needed for scene data – High delay when passing rays between processors means crossing machines

• Object Partitioning

– Same as object space (just divided differently)

• Image Space Partitioning

– Each local processor must have complete scene data – Messages are only sent to divide and reassemble, not while doing the work

Partitioning for Parallel Ray Tracing

• Image space partitioning is the best choice
• When might we use the others?

– Object Space Partitioning: when there is not enough local processor memory for the scene data – Object Partitioning: same as above when object complexity is similar

[Reinhard99 – thesis] [Reinhard98]

Image space partitioning won’t work here!

Ray Tracing Optimization

• Summary

– Object-Ray Intersection

• Bounding Volume
• Spatial Subdivision

– Parallel Ray Tracing – break