oat tatarchuk froblins(siggraph2008)

8/2/2019 Oat Tatarchuk Froblins(Siggraph2008)

http://slidepdf.com/reader/full/oat-tatarchuk-froblinssiggraph2008 1/105

Simulation and Rendering MassiveSimulation and Rendering Massive

Creatures on GPU

Creatures on GPU

Christopher Oat

Natal a Tatarchuk

AMD Graphics Products Group



•o va on

• Crowd movement simulation on GPU

– Global and local navigation and avoidance

•

• Crowd rendering and scene management on

• Conclusions

All slides © 2008 Advanced Micro Devices, Inc. Used with permission.



• ur as s o pus e enve ope

for interactive visual experience

“ –

interactive experiences?”

• Be ond ust the rett face

– We know we can do that… We’ve

done it!

– Can we make them think?




• a ng e gamep ay more un: game s a

series of choices...

– It helps when choices are interesting!

• Artificial intelligence is ubiquitous in currentgames

– Non- la er and la er characters re uire a form of AI

– Path finding, obstacle avoidance, decisions, etc.




• can spen > me on pa n ng

computations alone in games

– Limits possible decision complexity

– Thus we frequently see “zombie-like” NPCs

– Gameplay suffers

• many characters


–



• oa : mu a e an ren er mass ve crow s omu a e an ren er mass ve crow s o

characters on GPU at a scale that's bothcharacters on GPU at a scale that's both

rea a ng an c a eng ngrea a ng an c a eng ng

• Direct3D® 10.1

• Tessellation

• HD resolution and HDR rendering




• ynam c pa n ng compu a ons on

• Massive crowd rendering with LOD management

• Tessellation for high quality close-ups and stable performance

• HDR lighting and post-processing effects with-

• Terrain system

-

• Advanced global illumination system




• o va on



•


• Conclusions




D namic and En a in WorldD namic and En a in WorldThrough AIThrough AI

• om ne g o a pa n ng w

local avoidance and individual

• Crowd movement is more stable

and realistic solved on a lobal scale

– Crowd dynamics similar to fluids

– A ents “flow” towards the closest oal

along the path of least resistance

– Froblins follow correct paths and do not “get stuck”




• Based on Continuum Crowds [TCP06]• Convert motion lannin into an o timization

problem

–

– Computational algorithms from optics

• Smooth flow-like crowd movement

– Con estion avoidance and emer ent behavior




ndnd

aa computational iterative approach (eikonalcomputational iterative approach (eikonalsolver solver))

– – RepresentRepresent environment as a costenvironment as a cost fieldfield

– – rougroug scre za on o e e ona equa onscre za on o e e ona equa on

•• A licable A licable to man eneral al orithms andto man eneral al orithms and

areasareas




Example PathsExample Paths




• o ve or trave t me as a unct on o potent a• Potential = inte rated cost F alon shortest

path to goal

= – o ow nega ve gra en

• Set potential = 0 at goal, eikonal equation

elsewhere




Eikonal Solver via Fast MarchinEikonal Solver via Fast MarchinMethodMethod

• mu a es s ra s s or es pa a gor m

• A finite difference approximation to the continuous

eikonal equation

– Start from known otential =0 at oal and ro a ate to

neighbors until convergence

– Compute potential for neighbor cells based on known cells

• Serial algorithm; requires an ordered data structure




• s ng as era ve e o

• Use upwind finite difference approximation but no ordered

a a s ruc ures

• We use small grid resolutions (1282

-2562

)• Solve for four goals at once

– Taking advantage of vectorization – use RGBA FP16 texture

• Convergence determined empirically




• on nuum o e mo e s env ronmen as a

positive cost function

– Can’t be zero otherwise agents would move at infinite

speed; add some small base amount• Incorporate information about terrain and

obstacles (static and dynamic)

– Splat dynamic obstacles, including agents into the cost

function




=+ +




•

– Performance & No need to avoid agent far away

• p ate ve oc ty ase on pos t ons an ve oc t es

of nearby agents/obstacles – Based on Velocity Obstacle formulation [Fiorini and

Shiller 1998]

– See course notes for more details & Jeremy Shopf’stalk on in Beyond Programmable Shading: In Action




• Evaluate a fixed set of possible directions

relative to global navigation direction




• Need to query nearby agent positions and

velocities

• Spatial data structure

“ ”

– Bin Counter : Color buffer

– Bin Array : Depth Buffer Array




Agent Positions Bin Counter Bin Array

2

3

51

0 1 0 0

1 0 0 1

0 0 0 1

3

1 5

26

4 0 0 0 26

4

• or space pos t on mappe to n ex

• Bin Counter: color buffer, tracks bin loads

• ID Array: depth array, binned agent IDs





0 1 0 0

1 0 0 1

0 0 0 1

3

1 5

2

0 0 0 26

4





0 1 0 0

1 0 0 1

0 0 0 1

3

1 5

2

•

0 0 0 26

4





0 1 0 0

1 0 0 1

0 0 0 1

3

1 5

2

•

0 0 0 26

4

• Fetch load from Bin Counter (color buffer)





0 1 0 0

1 0 0 1

0 0 0 1

3

1 5

2

•

0 0 0 26

4

• Fetch load from Bin Counter (color buffer)

• Fetch agent IDs from Bin Array (depth array)


–



3 0 1 0 0 3

64

2

51 1 0 0 1

0 0 0 10 0 0 2

6

1 5

24

• Binning is a multi-pass algorithm

– Once binned, agents removed from working set

– Algorithm repeats until working set is empty

– Overflow is ossible & detectable




• Initialize data structure –

– Clear bin array (depth array) to MAX_DEPTH

– Working set is array of all agent IDs

• No VB necessar




• n n coun er rs s ce o n array

• Draw working set as point primitives

• Vertex shader:

–

– Set point’s depth to normalized agent ID

• Pixel shader: output “1”

• Depth test: LESS_THAN




……

• ex s ce o n array oun as ep u er

• VS: Sample ID from previous slice of bin array

– Reject points less than or equal to previous

• -

• PS: write pass number

• Depth test: LESS_THAN




....

• ’

been binned get streamed-out and rasterized

• ept test ensures t e po nt w t owest

gets binned• Results in points binned in sorted order

–

• Stream-out buffer becomes new working set




• We want to halt the algorithm once all points

are binned

– Do not query size of stream out buffer each pass

–

• Would like to hand the whole thing off to the

o con ro execu on




– Number of bin array slices

• Make all the draw calls for the max number

of iterations – Use cascading predicated draw calls

“ ” – raw u o

– Predicate on number of stream-out elements




• GPU terminates algorithm once all agents

are binned

– Once working set (stream out buffer) goes to 0

–

• The algorithm either terminates or overflows




•

– Efficient queries

• Early-out on empty bins, known bin load, sorted order

– Overflow detection – Stream-out reduction of working set

–

• Many applications beyond local avoidance




d

– Dominant directional light color: Lc

• Left over residual lighting environment: Le

• s

• Surface normal: N

( ) ),SH_Eval(**0,max escd L N V L L N +⋅




• Dynamic characters cast shadows on terrain

– Shadow map shadows mingle with SHLM shadows

–




• Results in double shadows…

– Characters in shadow still cast shadows




Detectin Direct Sun Li ht




• Shadow map only used on non-occluded

regions of terrain




• o va on



•


• Conclusions




• Don’t want to render thousands million-poly

c arac ers

– Wasteful if details are unseen

• CPU side character management is impractical

when doing GPU simulation

– Requires a read-back


–



• e use rec . ea ures or scene

management and level of detail management

• Render froblins as an army of instanced

characters• 3 discrete LODs

• Use tessellation to get maximum amount of details with stable performance for close-ups




• Act on instances

• A set of oint rimitives instance data as

input

-

– Discard the rest

– Use DrawAuto for multiple tests or combine into asingle GS invocation for efficiency




. en er e occ u er geome ry

2. Construct the Hi-Z map

3. Run all characters through the view frustum

cullin filter

4. Results are run through the occlusion culling

. Run through a series of LOD selection filters6. Render each LOD




• s ng e er ng per orm v ew rus um

culling

– Using standard methods

• VS checks for intersection between character bounding volume and the view frustum

– Re ular methods a l

• If the test passes, the character is visible: emit it


– erw se s cu e



• Render all occluders prior to rendering

characters

• Determine which characters are occluded by

– Use resulting depth buffer

–Cull against arbitrary, dynamic occluders

• Novel formulation: Hi-Z map




--

• A hierarchical depth image

– Uses the Z buffer information

• A mip-mapped, screen-resolution image

– ac exe a eve = max a exe s a eve -

• Does not require a separate depth pass




--

-

• Examine the depth information for pixels covered

by the object’s projected bounding sphere

– Compare the max fetch depth to the projected depth of

the point on the sphere closest to the camera

• Conservative culling

– Does not result in false culling

•




--

–

– DepthStencil view (with MS DB in Direct3D® 10.1)

• Generate the mip chain levels with reduction

– Mind the dimensions – screen-space images don’t

– Use integer operations to avoid incorrect indexing




--

• Render into one mip level while sampling the

previous level

– Rendering into smaller mip reducing the larger one

• value

• Fetch additional texels on the odd-sizedboundary




• Use a discrete LOD scheme

– Each LOD is selected by character’s distance to

camera

• – Separate the characters into three disjoint sets

– LOD parameters easily specified for each set




• ompu e se ec on pos cu ng

– Only process visible characters

– Culling results are only computed once and re-used

• Render closest LOD usin tessellation and displacement

• Simplified geometry and shaders for furthest LOD




• ee ns ance coun or ssu ng e raw ca or

each LOD

• This requires a stream out stats query

– Can cause significant stall when results are used in thesame frame issuing the query

• Re-or anize the draw-calls to fill the a between

issuing the query and using the results –




• ra ona y ma r x pa e es are compu e on e

CPU per character

– Loaded into constant store

• Limitations on the number of characters handled inone draw-call with this approach

– Lar e crowds of characters re uire numerous draw calls

• GPU character management makes this tricky




• Agents have a set number of actions with

associated animation sequence




• t preprocess ng t me, a e an mat on ata

into texture arrays

– Slice number = animation sequence

–

• At rendering time, determine per-instance

animation sequence index and time offset – Durin simulation in our case




• amp e an nterpo ate an mat on ata

when rendering the character

– Sample the animation textures




• preprocess ng, a en e rans orm erarc y

• Compute bone transforms for each key frame

transforming that bone into object space

• Durin simulation assi n an animation se uenceto each character

–




…

→…

Matrix Row 0Bone 1

Matrix Row 2

Matrix Row 0

Matrix Row 1

Matrix Row 2

Bone 2

↓




• ur ng ren er ng e c es, n erpo a es, an

blends the key frames for each bone

– Using per-character information

• skinning in object space

• Transforms the result using its positionand orientation




• om ne w rec ns anc ng suppor

• Render using interpolative planar tessellation

– Fast evaluation

– -

• Control tessellation level per-draw call

– Use character location

• The same art assets as conventional rendering




• esse a on re uces memory oo pr n an

bandwidth

– Only store low resolution mesh

– Polygons Total Memory

Low resolution 5160 polygons VB/IB: 100K

Froblin model 2K x 2K 16 bit displacement map: 10MB

ZBrush High res

Froblin model

>15M polygons ~270MB VB and 180MB IB storage




• Scalability

• Stable and redictable erformance


Tessellation Performance AnalysisTessellation Performance Analysis



Fa ra w a yv iew

Close-u p

v iew

Render ing

Mode

Num faces :

ATIATI ATIATI ATIATI ATIATIRadeonRadeonHD 4870HD 4870

RadeonRadeonHD 2900 XTHD 2900 XT

RadeonRadeonHD 4870HD 4870

RadeonRadeonHD 2900 XTHD 2900 XT

Original l o w r e s 4,050852 fps852 fps 359 fps359 fps 850 fps850 fps 275275

NT = 411 x NL

Con t inuous lytesse l la tedmesh N

1.6 M triangles 232 fps232 fps 143 fps143 fps 213 fps213 fps 101101

Adap t i ve lytesse l la ted

mesh NA

Dynamic,1.6M > NA >4K triangles

845 fps845 fps 356 fps356 fps 574574 fpsfps 207207

*Configuration: AMD reference platform with AMD Athlon™ 64 X2 Dual-Core Processor 4600+, 2.40GHz, 2GB RAM.

Motherboard: ASUSTek M2R32-MVP. Memory: DDR2-800 400 MHz.

Operating System: Windows Vista® SP1.” All slides © 2008 Advanced Micro Devices, Inc. Used with permission.

Tessellation Pre Direct3D® 11Tessellation Pre Direct3D® 11



•

– Supported on XboxTM

360 – a eon er es an eyon : suppor e

on all models

• Subset of Direct3D® 11 tessellation features

• Contact AMD ATI developer relations for detailson accessing tessellation


Preview and Prepare for Preview and Prepare for



Direct3D® 11Direct3D® 11•

for this feature

– Su ort on Microsoft® Windows® XP as well as Microsoft®

Windows® Vista

– PC versions can use XboxTM 360 native features

– Reach more players

• Developing artwork takes time

– Integrating tessellation early gives you and your artists time todesign and polish




Tessellator

Evaluatesurface

AddSuper Super--primprim

MeshMeshTessellated andDisplaced Mesh

TessellatedTessellated

MeshMesh

displacement


DisplacementDisplacement

MapMap

Combinin Normal Ma s andCombinin Normal Ma s and



Displacement MapsDisplacement Maps

• As we displace, we change the actual

displayed normal




se norma maps even w

displacement

• Generate displacement and normal maps

usin the same tan ent s ace – The same TS computations for preprocessing

– e norma enco e n norma map w ma c

the normal from displacement


Regularly drawn Froblin

N t ll ti d



No tessellation used


High Detail Froblin with

Di l t M i



Displacement Mapping


Froblin Close-up:

No tessellation



No tessellation


Froblin With Tessellation

Close up



Close-up




• ynam c num er o e a e c arac ers n v ew

– Simulation changes interactively, no a priori control

• Need stable frame rate in dense crowd

simulations

• Compute tessellation level as a function of

– Avoid polygonal count explosion




l = c amp max , , max

• N characters rendered with tessellation

• Bound the number of amplified triangles

• r m ve coun never excee s an e cosof M fully tessellated characters




transformation on control cage

Pass 1: Control

cage animation

n a pre-pass

– Allows for more complex per-

Stream Out Buffer Stream Out Buffer

vertex operations

• Pass 2:

(de)compression for reducingTessellate the

control cage




• Shader-based vertex compression /

decom ression

• Reduce stream out memory footprint

e ween e pre-pass an esse a epass

• Reduce fetch bandwidth for tessellated




• Pack transformed vertex positions into

128 bit format

–Lets us load each vertex with one fetch

–Gives as much as 30% speedup

–




**



– From 900 polygons -> 6K -> 1.6M at the closest tessellated level of details

– Up to 18M triangles per frame at fast interactive rates

– 6M-8M triangles on average at 20-25 fps

• – Rendering all objects into multiple shadow maps more than doubles

polygon count per frame

• en er ng an s mu a ng g qua y e a e ro ns

at 21 fps on average



GPU: ATI Radeon™ HD 4870 Grahpics. Motherboard: ASUSTekM2R32-MVP. Memory: DDR2-800 400 MHz.

Operating System: Windows Vista® SP1.”

**



• mo es ren er w an pos -

processing• mu a e e av or an ren er agen s a ps

• AI simulation for 65K agents alone – 45 fps

• Rendering and simulating 65K agents – 31 fps

• Rendering 9,800,000 polygons each frame

• AI simulation executes with high efficiency resulting in~1 teraflops!



GPU: ATI Radeon™ HD 4870 Grahpics. Motherboard: ASUSTekM2R32-MVP. Memory: DDR2-800 400 MHz.

Operating System: Windows Vista® SP1.”



• rac ca an e c en s mu a on an ren er ng o

large crowds of characters on GPU• New GPU algorithms for

– Global and local navigation and spatial data structures

– Scene management

–

• And a number of other advanced techniques!




• All shaders are included in the course notes

• Course notes are available for download:

– ACM Digital Library: http://portal.acm.org/dl.cfm

– rap cs ec n ca u ca ons we page:http://ati.amd.com/developer/techreports.html




Chain unChain unx genx gen

StudiosStudiosegor m cegor m c




• REUILLE, ., OOPER, ., AND OPOVI , . . on nuum

Crowds. ACM Trans. Graph. 25, 3 (Jul. 2006), pp. 1160-1168, Boston,

MA.

• [JEONGWHITAKER07B] JEONG, W.-K, AND WHITAKER, R.T. 2007. A Fast

Iterative Method for a Class of Hamilton-Jacobi Equations on Parallel

S stems. Universit of Utah School of Com utin Technical Re ort

UUCS-07-010.

• [T ATARCHUK08] T ATARCHUK, N. 2008. Advanced Topics in GPU

Tessellation: Algorithms and Lessons Learned. Presentation.

Gamefest 2008, Seattle, WA, July 2008.




Tradem ark A t t r i bu t i on

AMD, the AMD Arrow logo, AMD Opteron, AMD Athlon, AMD Turion, AMD Phenom, ATI, the ATI, , , ,

Inc. Microsoft, Windows, and Windows Vista are registered trademarks of Microsoft Corporation

in the United States and/or other jurisdictions. Other names are for identification purposesonly and may be trademarks of their respective owners.


©2008 Advanced Micro Devices, Inc. All rights reserved.

oat tatarchuk froblins(siggraph2008)

Documents