16-October-1999 1

Geometry at Work:Open Issues Encountered in

Real Applications using BRL-CADTM

Michael John MuussThe U. S. Army Research Laboratory

4th CGC Workshop onComputational Geometry

16-October-1999 2

Why We Model

• Storytellers communicate feelings to people.+ “Skin-deep” models are fine for movies.

• We are predicting or matching physical phenomena:+ Energy levels received by a sensor.+ Damage statistics of live-fire tests.

16-October-1999 3

Modeling is OnlyOne Part of the Process

Experiment

IntegratedSurvivability/Lethality

Analysis ProductsAnalyze

Model

Today’s topic is modeling and simulation.

16-October-1999 4

Modeling Means Different Things...

• Goal: Re-creating the real-world in simulation:+ Re-creating individual laboratory tests.

• Science & Engineering community starts here.+ Re-creating real proving grounds.+ Re-creating training centers and actual exercises.+ Re-creating combat locations and scenarios.

• Training community & wargamers start here.

16-October-1999 5

The Simulation Challenge

16-October-1999 6

Meeting the Simulation Challenge

• Engineering-level geometric detail.• Physics-based simulation.• Realistic 3-D atmosphere, ground, and sea models.• Fast: Hardware-in-the-loop, man-in-the-loop.

+ Real-time, near-real-time, Web, and offline.• Common geometry.• Common software.• Massively parallel processing.

16-October-1999 7

Two Types of Simulation

• Image Generation

• “If you can’t see it, you can’t shoot it.”

• Vulnerability/Lethality Analysis

• “Will the bullet bounce off?”

16-October-1999 8

OUTLINE

• I. BRL-CADTM and Targets• II. Shooting Bullets• III. Making Pictures

16-October-1999 9

I. BRL-CADTM and Targets

16-October-1999 10

BRL-CADTM Primitive Solids

sphere spheroid ellipsoid right circularcylinder

right ellipticalcylinder

truncated rightcircular cone

truncatedgeneral cone

truncatedelliptical cone

topo. cubic6-hedron

edge-contractedtopo. cubic 6-hedron right triangular

prismquadrilateral

pyramidtetrahedron

intersectionof halfspaces

torus

elliptical-ringtorus right parabolic

cylinderright hyperbolic

cylinderelliptical

paraboloidelliptical

hyperboloid waterline-basedpolyhedron

Path and bend convex hullof two spheres

extrudedbit map

revolvedplane curve

extrudedplane curve

halfspacevoxel data

general polyhedron

trimmed NURBS

16-October-1999 11

BRL-CAD Primitive Solids

16-October-1999 12

wedge

cylinder

block

(wedge block) cylinder

wedge block cylinder

block (wedge cylinder)

CSG Boolean Operations

16-October-1999 13

Hierarchical Database Organization

tank

turrethull suspension

gunturret_armor turret_interior

crew

breech gun_tube bore_evacuator

cylinder_1.s cylinder_2.s cylinder_3.s

Directed

Acyclic

Graph

16-October-1999 14

A Medium-Resolution BRL-CAD Database

16-October-1999 15

Corps Command Post

16-October-1999 16

Library of Existing BRL-CAD™ Geometry

16-October-1999 17

One Geometry,Multiple Uses

• To compute ballistic penetration & vulnerability:+ Need 3-D solid geometry and material information.

• The same targets are also useful for:+ Signatures: Radar, MMW, IR, X-ray, etc.+ Smoke & Obscurants simulation.+ Chem./Bio agent infiltration.+ Electro-Magnetic Interference.

• BRL-CADTM is the basis for all our simulations.

16-October-1999 18

Ray Tracing

Startingpoint

distance,obliquity,normal,

curvature,etc.

16-October-1999 19

Evaluating Boolean Expressions in CSG

A

B

C

100 110 010 011 010

A B – C

ABC:

Segments:

16-October-1999 20

II. Shooting Bullets

16-October-1999 21

physics,penetration models, ...

Vulnerability/LethalityAnalysis Process

Initialthreat/targetconditions

Componentdamage

Systemcapability

Systemutility

engineering,criticality analysis, ...

operations research,missions, scenarios, ...

Level 1

Level 2

Level 3

Level 4

16-October-1999 22

Computing Component Damage(Level 1 to Level 2 Mapping)

CSG model of vehicle

Damage Results

Specification ofmunition performance

Shotlines representingPenetrator+spall paths

Vulnerability model

Ray tracerSpall

16-October-1999 23

Ray-tracing Through a Target

glacisarmor

armor-piercingrounds

HEround

firewall

enginestarter

transmissionsump

fanrear

armor

16-October-1999 24

0 900 mm

Perforationinto internal volume

Residual penetrationinside internal volume

Penetration Results

16-October-1999 25

Behind-Armor Debris(Flash X-Ray)

16-October-1999 26

Spall: a Secondary Damage Mechanism

• Experimental Data:

+ Perpendicular jet.

• Simulation Results:

+ Oblique impact.

Thousands of fragments to track!

Each generates another ray.

Behind-Armor-Debris is BAD.

16-October-1999 27

Main Armament Fault Tree

Main Armament Subsystem

Fault Trees map component failure

to subsystem capability.

Mapping from Damageto Capability (L2->L3)

Subsystems per vehicle: 50-100

16-October-1999 28

Is a Ray aGood Approximation

for a Fragment?• Sensor pixel. 0.01mm diameter -- OK.• Rifle bullet. 5.56mm diameter -- Maybe.

• Tank bullet. + 30-120mm diameter -- No.

16-October-1999 29

In General: No!

• Real particles have non-zero cross-section.+ A 0-thickness ray is not the best approximation.

• A real fragment will hit wires that the ray will miss.• Most damage is done by spall cloud.

+ Has a large total surface area.+ We sample density distribution with 1000’s of rays.+ This greatly under-samples the target geometry.

16-October-1999 30

Beam or Cone Tracing?

• Obvious solution: + Model particle path as a cylindrical beam.+ Model light ray as a cone.

• Solve cylinder-vs-object or cone-vs-object intersections.+ Such intersections yield complex volumes.

16-October-1999 31

A Ray Slipping ThroughComplex Geometry

Object on Centerline

Ray

16-October-1999 32

A Beam throughComplex Geometry

r > 0

Object on Centerline

16-October-1999 33

Objects in the Beam

Object on Centerline

16-October-1999 34

A Simplified Viewof the Relationships

Object on Centerline

But it isn’t this simple!

16-October-1999 35

Relations Along a Ray

appears-before

entirely-precedes occults

16-October-1999 36

Difficulties withCone-tracing

• Intersection with more general solids is expensive.+ E.g. height field, or t-NURBS.

• Representation of the results is difficult.+ No exact representation of the volume.

• At best, result could be some kind of B-rep.+ No convenient abstraction of volumetric result.

• Partially ordered sets!• Utilizing the results is difficult.

+ How to compute ricochet -vs- penetration?

16-October-1999 37

A Plea!

• Are there any good representations for these intersection volumes?

• Are there any good abstractions for these intersection volumes?

16-October-1999 38

Our Short-Term Strategy

• Fire additional rays distributed around main ray.+ Tightly coupled with space partitioning, for high

performance.+ User-selected patterns.

• Peripheral rays intersected only when main ray does not intersect geometry.

• Heuristics for choosing a single interval as “most representative” of material in that region.

• Reduces missing small objects in beam path.

16-October-1999 39

III. Making Pictures

16-October-1999 40

What is PST?

• PST = PTN and SWISS, Together!+ PTN = Paint-the-Night

• Real-time polygon rendering• From CECOM/NVESD

+ SWISS = Synthetic Wide-band Imaging Spectra-photometer and Environmental Simulation• Ray-traced BRL-CAD™ CSG geometry• From ARL/SLAD

16-October-1999 41

Application of PST

• The image generator is just one component of a larger simulation. E.g. MFS3, or missile simulation.

PSTPST ATR6 DoF

Flight DynamicsImages

Motion_t

Full Environment SimulationFull Platform Simulation

or HWIL

Control Decisions

Full Platform Simulation

or HWIL

16-October-1999 42

Paint-the-Night

• 8-12 micron IR image generator.• SGI Performer based.• Uses outboard image processor for sensor effects.• A large highly tuned monolithic application

+ With exceptionally high performance.+ Highest polygon rates seen on a real application.

• Individually drawn trees (2 perpendicular polygons)• Individually drawn boulders.

16-October-1999 43

SWISS

• A physics-based synthetic wide-band imaging spectrophotometer+ A camera-like sensor + Looks at any frequency of energy.

• A set of physics-based virtual worlds for it to look at:+ Atmosphere, clouds, smoke, targets, trees,

vegetation, high-resolution terrain.• A dynamic world; everything moves & changes.

16-October-1999 44

A Grand-ChallengeComputing Problem

• Real targets, enormous scene complexity, > 10Km2.• Physics-based hyper-spectral image generation.• Nano-atmospherics, smoke, and obscurants.• Ray-traced image generation, exact CSG geometry.

+ Near-real-time (6fps).• Fully scalable algorithms.• Network distributed MIMD parallel HPC.• Image delivery to desktop via ATM networks.

16-October-1999 45

Ray-Tracing forImage Synthesis

16-October-1999 46

Advantages of a Ray-Tracing SIG

• Allows reflection, refraction:

+ Windshields, glints.

+ Branch reflections, 3-5 µm.

• Atmospheric attenuation, scattering.

+ Individual path integrals.

• Accurate shadows:

+ Haze, clouds, smoke.

• Multiple light sources:

+ Sunlight, flare, spotlight.

2nd-Generation FLIR image, 8-12 µm

(Downsampled to 1/4 NTSC)

16-October-1999 47

CSG Rendering Advantages

• Ray-traced CSG is free from limitations of hardware polygon rendering:+ No approximate polygonal geometry.

• No seams, exact curvatures.+ Exact profile edges. Important for ATR!+ No level-of-detail switching, no “popping”.+ Full temperature range in Kelvins, not 0-255.+ Unlimited spectral resolution, not just 3 channels.

16-October-1999 48

Cruise Missile Shadow

Ridge Profile

Missile Shadow

Terrain Quantization

16-October-1999 49

Target Geometry Complexity

• Need at least 1cm resolvable features on targets.

16-October-1999 50

Complex Geometry Today

• < 1cm target features.• 1m terrain fence-post spacing• Three-dimensional trees:

+ Leaves.+ Bark.

• Procedural grass, other ground-cover.

• Boulders, other clutter.Current

Developmental

16-October-1999 51

Procedural Grass

16-October-1999 52

Ray-Traced Atmosphere

• Propagation easy in vacuum!

• Modeling four effects:

+ Absorption

+ Emission

+ In-scatter

+ Out-scatter

• Computer can’t do integrals.

+ Repeated summation

+ Discretized atmosphere

16-October-1999 53

The Blue Hills of Fort Hunter-Liggett

16-October-1999 54

Hyper-Spectral: The Power of a Single Pixel

16-October-1999 55

Open Issue:Representing Reflectance

• For each material used in the virtual world, need bi-directional reflectance distribution function (BRDF),

• Need BRDF as function of wavelength, too!• Seek a representation of this data which is:

+ Compact to store.+ Easy to locate important lobes.

16-October-1999 56

V/L Server

Terrain

Thermal Models

VehicleDynamics

Paint-the-NightPolygon Renderer

BRL-CAD™Ray Tracer

::

HLA

with

enh

ance

men

ts

Backplane Philosophy

• Standardized Slots (Interface).• Location independent

+ Except for performance.

Paint-the-NightPolygon Renderer

16-October-1999 57

PST Implementation Goals

• To have a software backplane:+ Allowing each function to run as separate process.+ Allowing easy reconfiguration.+ Allowing independent software development.+ Using common geometry throughout.+ Multiple Synthetic Image Generator (SIG) types.

• Keep simulation details out of the SIGs.

16-October-1999 58

Required Backplane Features

• Event Services+ Implement with HLA interactions.

• Query/Response Services+ HLA interactions with custom routing space.

• Continuous/Bulk Data+ Custom Distributed Shared Memory software.

• Auto-broadcast, optional subscriber notification.• Notification, subscriber polls for data update.

16-October-1999 59

PST Simulation with RT

RT

SIG

RT

SIG

DB

Solar

Load

Gen

Atmosphere

Ground Therm

Tree Therm

Target Therm

Monitor

Met

Input

Transducers

Entity

ControllersWorld

Simulations

Sensor

SimulationOutput

Transducers

ToD

MFS3

HW

FlyBox

Mapper

Mapper

MapperVehicle

Controller

Vehicle

Dynamics

MODSAF

I/F

Vehicle

Dynamics

Sensor

Controller

MODSAF

Intersect

Process

Magic

Carpet RTSYNC

Temp.

16-October-1999 60

PST Simulation with PTN

PTN

SIG

Data-cube

DB

Solar

Load

Gen

Atmosphere

Ground Therm

Tree Therm

Target Therm

Monitor

MetTextures

Input

Transducers

Entity

ControllersWorld

Simulations

Sensor

Simulation Output

Transducers

ToD

MFS3

HW

FlyBox

Mapper

Mapper

MapperVehicle

Controller

Vehicle

Dynamics

MODSAF

I/F

Vehicle

Dynamics

Sensor

Controller

MODSAF

Intersect

Process

Magic

Carpet

16-October-1999 61

Independent Time Scales

• Image generators need to run fast:+ 30 Hz for humans.+ 6 Hz is fastest acquisition rate of ATRs.+ 800 Hz for non-imaging sensors (Stinger rosette).

• Physics-based simulations can run slower:+ 90 sec/update for thermal & atmosphere models.

• Transient effects need to be added as a delta:+ Leaf flutter, explosions, smoke details.

16-October-1999 62

Hardware Environment

• Multiple CPUs per cabinet.• Multiple cabinets linked via OC-3 or OC-12 ATM.

+ Geographically distributed (Belvoir, APG, Knox).• Multi-vendor system, e.g.:

+ Cray vector machine for thermal mesh solution.+ SGI Origin 2000 for parallel ray-tracing.+ SGI Infinite Reality for polygon rendering.

• 100-200 processors participating.

16-October-1999 63

Ft. Knox Applicationof PST

• 1 RT SIG, 3 SGI SIGs, soldiers-in-the-loop.

DTV

DTV

DTV

DTV

PSTPST

RT

PTN

PTN

PTN

DREN

ATM

AT

M to D

-2 Video

Digital V

ideo to AT

M

Mapper

Mapper

Mapper

Mapper

DREN ATM

16-October-1999 64

The Ft. Knox Experiments

• One SWISS ray-traced image display.

+ For most sensitive tests.

• Multiple PTN polygon image displays.

+ Esp. to test teamwork.

• All in one shared environment: terrain, air:

+ Craters, smoke, haze…

+ Multiple players, threats.

• Ray-tracer in Maryland

+ >200 cpu Origin-2000

• Environment Sim in MD.

+ Cray T916, 16 cpus.

• Three PTN SIGs in VA.

+ Infinite Reality

• Man-in-the-Loop in KY.

+ Digital Image Display

OC-3 ATM

OC-3 ATM

16-October-1999 65

Real-Time Performance!

• Geometry has Spatial Coherence.+ We exploit that!

• A lot of the performance of our ray-tracer comes from extensive use of space partitioning.+ Muuss NUBSP tree, a flavor of Kd tree.

• Cost of ray/model intersection is proportional to local geometric complexity, not O(n). (good property)

16-October-1999 66

Data Structures versus The Hardware

• Kd tree • Memory Hierarchy

Processor

L1 Cache

L2 Cache

Translation Look-aside Buffer (TLB)

Memory Bus / Network

Bank Arbitration

RAM

Retrieve:

axis,

value,

pointer

O(3 log n)

trips to RAM

16-October-1999 67

Walking a Binary Tree

• Speed of walking a binary tree is limited by raw access time to main memory.+ No pipelining, no unrolling, no pre-fetch.+ Caches don’t help much.

• Binary tree’s memory-wait penalty accounts for 65% of runtime on R8000. Worse on Origin2000.

• Can’t live without space partitioning, alternative is massive O(n) search, >100X slower.

16-October-1999 68

A Plea to the Research Community

• Develop “cache friendly” algorithms & data structures• Universities teach that memory access time is uniform

across address space, not time varying. E.g. Knuth.+ Valid for old non-cached non-paged machines, 70s+ DEC PDP-11/45, Cray-X/MP.

• DEC VAX-11/780 added TLB, circa 1980.+ Extra trips to memory on TLB miss.

• And then came the caches…

16-October-1999 69

Example of an Alternative: NUGrid

• Non-uniform 3-dimensional grid of leaf nodes.• Index in several dozen instructions, 3D DDA.• Few loads required to get to intended volume.

+ Can double overall speed on large models.• M. Gigante, “Accelerated Ray Tracing using Non-

Uniform Grids”, Proceedings of Ausgraph ’90.

O(0) extra

memory

references

16-October-1999 70

Some Final Thoughts

• ARL really cares about Geometry and Algorithms.+ We appreciate your research!

• When you write an ARO grant proposal,+ Feel free to suggest me as a reviewer.+ Or Paul Tanenbaum.

16-October-1999 71

Acknowledgements

• Lee Butler (environment: smoke, grass, …)• Paul Tanenbaum (PO sets)• Max Lorenzo, CECOM

+ His team of SAIC contractors• Chris Johnson, Paladin Software• DoD HPCMO• OSD ATR• VPG

16-October-1999 72

Who is this MUUSS Fellow, Anyway?

Mike Muuss

Señor Scientist

U.S. Army Research Laboratory

APG, MD 21005-5068 U.S.A.

<Mike@ARL.MIL>

http://ftp.arl.mil/~mike/

16-October-1999 73

Peer Assessment of BRL-CAD

“an effective constructive solid modeling capability with highly efficient ray tracing”

“a computer-aided engineering (CAE) system, uniquely suited to survivability and lethality applications … in which physics-based simulations can build on an efficient ray-tracing engine”

“a platform for a ‘virtual test environment’ that could provide a powerful, cost-effective capability for survivability and lethality evaluation.”

1998 Assessment of the Army Research Laboratory,ARL Technical Assessment Board, National Research Council

CSG BREP Splines BREP Facets

radius(r)

vertex(x,y,z)

4 numbers

20 knot values

45 control pts(x,y,z)

45 weights

200 numbers

287 triangles

~287 vertices (x,y,z)

~861 numbers

CSG is most storage efficient

Data Storage Comparison