accelerating rigid body simulation on today’s...

1 | Accelerating Rigid Body Simulation on Today’s GPUs

Accelerating Rigid Body Simulation

on Today’s GPUs

Takahiro Harada

[email protected]


RIGID BODY SIMULATION PIPELINE

Broad phase collision detection

– Quick check using bounding volumes

Narrow phase collision detection

– Detailed check using geometry

Solve A

B

C

D

E

F


WHY USE GPU?

Large scale simulation requires a lot of bodies

– Destruction

Increase the cost of a simulation

GPU has high-

– Peak performance

– Memory bandwidth

GPU is different from CPU


GPU RIGID BODY TECHNIQUES

GPU Graphics GPU Compute

Bro

ad p

hase

N

arr

ow

phase

Solv

e



Rahul Sathe, Rigid Body Collision

Detection on the GPU, Sig

Poster(2006)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

Cube map

Sathe(2006)

Cube map

! "#"$%&' $( %) ' **"+"' , %- . /. 0/"' , %' , %/1. %2 34%! "#$%&' "(#)&*+"#$%,- ,. "(#) / 01()%,234 56&7 8"4 &9": )&*"8"4 ,(,%": ) / 01()%,234 5&

; "4 01/ 4"56 ($&

#

@)#%/ , 5)-6 "C#)+"#$%&41"6 #-' #)-, 2#-0#)B&4A"*)' #(%"#-, )"%' "*)-, 37#

N"):' #*&, ' -2"%#)B&4A"*)' #@#( , 2#O7#D-3/ %"#J#' +&B' #)+"#(13&%-)+6 #-, #(*)-&, 7#! "#' +&&)#%(9' #0%&6 #*", )%&-2#&0#&4A"*)#@#))+"#?"%)"E#&0#

&4A"*)#O7#! "#*(1*/ 1()"#)+"#2-' )( , *"#4")B"", #)+"' "#)B$&-, )' #P2JQ#-, #B&%12#' $(*"7#= ' -, 3#)+"#2-' )( , *"#*/ 4"56 ($#1&&R5/ $C#B"#*( , #0-, 2#

&/ )#+&B#0(%#&4A"*)#@:' #4&/ , 2(%9#-' #0%&6 #-)' #*", )%&-2#-, #)+"#

2-%"*)-&, #&0#)+"#?"%)"E#&0#&4A"*)#O#P2HQ7#! "#)+", #*&6 $(%"#2H#( , 2#2J7#G0#2H#ST#2JC#)+", #)+"#(*)/ (1#2-' )( , *"#4")B"", #)+"#)B$&-, )' #-' #

1"' ' #)+(, #)+"#2-' )( , *"#0%&6 #*", )%&-2#)-)' #4&/ , 2(%97#8+-' #6 "( , ' #

)+()#B+", #2H#ST#2J#)+"#&4A"*)' #(%"#-, )"%' "*)-, 37#! "#%"$"( )#)+-' #' )"$#

0&%# (11# )+"# ?"%)-*"' # ( , 2# )+", # B-)+# )+"# &4A"*)' # @# ( , 2# O#-, )"%*+(, 3"2C#/ , )-1#)+-' #*&, 2-)-&, #-' #)%/ "7##

#

! "

# " $

# " %

# " &

# " '

# $'

# $$

# $"

! $

# " "

, $, "

, $

, "-. /012345/0, 6, 78/963:6; 01/03. 6"

:13? 6! " 6@87. >68A@510B? 526433CB@2

, $ DE/@546, 78/96F0/G00. 6! " 65. , 6#$"

H710E/73. 63:6; 01/7E0863:6234=>3. 6$6:13? 6E0. /016! " 63:6234=>3. 6"

, "

, " I 6/0

#C0



Takahiro Harada et al. , Smoothed

Particle Hydrodynamics on GPUs,

Proc. of CGI(2007)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)

Cube map

Sathe(2006)

Cube map

Uniform

Grid

(Atomics)

Penalty



Takahiro Harada, Real-time Rigid

Body Simulation on GPUs, GPU

Gems 3 (2007)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)

Particle based rigid body

Harada(2007) Cube map

Sathe(2006)

Cube map Particles

Uniform

Grid

(Atomics)

Penalty



Scott Le Grand, Broad-phase

Collision Detection with CUDA,

GPU Gems3(2007)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007) Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)

Cube map Particles

Uniform

Grid

(Sort)

Uniform

Grid

(Atomics)

Penalty



Peter Kipfer, LCP Algorithms for

Collision Detection using CUDA,

GPU Gems 3(2007)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)

Cube map Geometry Particles

Uniform

Grid

(Sort)

Uniform

Grid

(Atomics)

Penalty



Takahiro Harada, Parallelizing the

Physics Pipeline, GDC(2009)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)


Constraint

Partially

Serial

Batching

Batching

Harada(2009)

Uniform

Grid

(Sort)

Uniform

Grid

(Atomics)

Penalty



Richard Tonge, PhysX GPU Rigid

Bodies in Batman, Game

Programming Gems 8(2010)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)


Constraint

Partially

Serial

Batching

Parallel

Batching

Constraint

Batching

Harada(2009) Batching

Tonge(2010)

Uniform

Grid

(Sort)

Uniform

Grid

(Atomics)

Penalty



Liu et al., Real-time Collision Culling

of a Million Bodies on Graphics

Processing Units, Siggraph

Asia(2010)


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

GPU Sweep & Prune

Liu(2010)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)


Uniform

Grid

(Atomics)

Penalty Constraint

Partially

Serial

Batching

Parallel

Batching

Constraint

Batching


Tonge(2010)

Uniform

Grid

(Sort)

Sweep &

Prune


Architecture & Algorithm


GPU ARCHITECTURE

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

Global Memory



Used GPU as a processor with

many ALUs


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

GPU Sweep & Prune

Liu(2010)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)


Constraint

Partially

Serial

Batching

Parallel

Batching

Constraint

Batching


Tonge(2010)

Uniform

Grid

(Sort)

Sweep &

Prune

Uniform

Grid

(Atomics)

Penalty


GPU ARCHITECTURE

Global Memory

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

ALU

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

AL

U

LDS

Compute Unit (CU)



LCP

– A CU processes a pair

– Synchronization

– LDS

Broad phase collision

– Radix sort


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

GPU Sweep & Prune

Liu(2010)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)

Cube map Particles

Constraint

Partially

Serial

Batching

Parallel

Batching

Constraint

Batching


Tonge(2010)

Geometry

Uniform

Grid

(Sort)

Sweep &

Prune

Uniform

Grid

(Atomics)

Penalty



Solver using CU

Approximated narrowphase using

CU

– Performance


Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

GPU Sweep & Prune

Liu(2010)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)


Constraint

Partially

Serial

Batching

Parallel

Batching

Constraint

Batching


Tonge(2010)

Uniform

Grid

(Sort)

Sweep &

Prune

Batching

Constraint

Representati

on for GPU

Uniform

Grid

(Atomics)

Penalty


Compute Unit Aware

Rigid Body Simulation


PIPELINE

Copy body and pair buffer

GPU allocates big buffers

– Contact

– Constraints

Narrow phase and solve is done on the GPU

Don’t have to read back big buffers

Bo

dy

Pai

r

Bo

dy

Co

nta

ct

Co

nst

rain

t

Pai

r

Merge

Dispatch Logic

CPU

Broad phase Collision

GPU

NP Collision

Solve

Body, Pair

Body Copy

Copy

Copy

Copy


NARROWPHASE – CU LEVEL PARALLELIZATION

Collision of a pair is independent

Use a SIMD lane for a pair

– Not enough resource for a SIMD lane

Use a CU for a pair

A CU processes and “append” colliding pairs

– HW accelerated append operation on AMD GPU

Load balancing

– CU can fetch a pair from queue

– Explicit pair split

CU level parallelization

Contact buffer

Pair buffer

Narrowphase

CU CU CU CU CU CU CU CU


NARROWPHASE – SIMD LANE PARALLELIZATION

A collision has to be split into parallel works

Parallel collision detection

Support arbitrary convex shapes

Too many contact points

– Increase cost of solver

Redundancy

– Use sync and LDS for vector wide operation

– Eliminate redundant contacts

Contact buffer

Pair buffer

Narrowphase

Read

Reduce

Write

Collide


SHAPE REPRESENTATION


SOLVER

Constraint based solver uses a Gauss Seidel method

– Velocity is input and output

– Penalty method

Sequential solve

Parallel solve is possible only if

– Bodies are not shared among constraints

When a body is colliding to several bodies, parallel

solve cannot be used

Solution

– Split constraints into batches

– Constraints in a batch doesn’t share bodies


BATCHING

2 level batching

– Global batching

Split into independent sections

– Local batching

Batch in each section

Contact buffer

Contact buffer

Local batching


Global batching


Contact buffer


ITERATIVE LOCAL BATCHING

A CU processes a section of contact buffer

– Local batching Global batching

– Extract independent pairs in a set

Stream processing pair buffer

– Read to local, reorder

Pairs are localized

– X Parallel batching

– X Serial batching

– O Iterative batching

Contact buffer

Local batching

Read

Extract Batch

Write

Contact buffer


SOLVER

A CU processes a section of contact buffer

GPU dispatches works by itself

– Read constraints

– Fill batch buffer

– Solve

– If the batch is done, go to the next batch

Previous work

– CPU had to dispatch a kernel per batch

Our method

– CPU dispatches some

– GPU dispatches works by itself

Solver

Read

Solve

Write

Contact buffer

Fill local batch


DEMO

OpenCL

Radeon HD5870

20K objects

About 30fps

GPU collide and solve ≈ 30ms

(Inc. data transfer)


DEMO

OpenCL

Radeon HD5870

12K objects

About 30fps

GPU collide and solve ≈ 30ms

(Inc. data transfer)


OFFLINE DEMO

400K objects

GPU collide and solve < 1s


LIMITATIONS

GPU prefers uniform work granularity

Some works are better to use CPUs


HETEROGENEOUS ERA

GPU Graphics Heterogeneous GPU Compute

Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

GPU Sweep & Prune

Liu(2010)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)


Uniform Grid

(Atomics)

Penalty Constraint

Partially

Serial

Batching

Parallel

Batching

Constraint

Batching


Tonge(2010)

Uniform

Grid

(Sort)

Sweep &

Prune

Batching?

Constraint?

Representati

on for GPU


HETEROGENEOUS ERA

GPU Graphics Heterogeneous GPU Compute

Bro

ad p

hase

N

arr

ow

phase

Solv

e

SPH

Harada(2007)


Harada(2007)

LCP

Kipfer(2007)

GPU Sweep & Prune

Liu(2010)

Cube map

Sathe(2006)

Sort based grid

Le Grand(2007)


Uniform Grid

(Atomics)

Penalty Constraint

Partially

Serial

Batching

Parallel

Batching

Constraint

Batching


Tonge(2010)

Uniform

Grid

(Sort)

Sweep &

Prune

Batching?

Constraint?

Representati

on for GPU

AMD Fusion

– Llano, Zacate

Advantages

– Shared memory

– CPU & GPU can work together


HETEROGENEOUS PARTICLE BASED SIMULATION G

PU

CP

U0

CP

U1

CP

U2

Bu

ild

Ac

c. S

truc

ture

SS

Co

llisio

n

S

Integ.

LS

Co

llisio

n

LS

Co

llisio

n

LS

Co

llisio

n

Synchronization

Merge Merge Merge

LL Coll.

L

Integ.

Synchronization

S S

ortin

g

S S

ortin

g

S S

ortin

g

Synchronization

Takahiro Harada, Physics Simulation on Fusion Architecture, AMD Fusion

Developer Summit(2011)


REFERENCES

Peter Kipfer, LCP Algorithms for Collision Detection using CUDA, GPU Gems 3(2007)

Takahiro Harada et al. , Smoothed Particle Hydrodynamics on GPUs, Proc. of CGI(2007)

Takahiro Harada, Real-time Rigid Body Simulation on GPUs, GPU Gems 3 (2007)

Liu et al., Real-time Collision Culling of a Million Bodies on Graphics Processing Units, Siggraph

Asia(2010)

Takahiro Harada, Parallelizing the Physics Pipeline, GDC(2009)

Richard Tonge, PhysX GPU Rigid Bodies in Batman, Game Programming Gems 8(2010)

Rahul Sathe, Rigid Body Collision Detection on the GPU, Sig Poster(2006)

Scott Le Grand, Broad-phase Collision Detection with CUDA, GPU Gems3(2007)

Takahiro Harada, Physics Simulation on Fusion Architecture, AMD Fusion Developer Summint(2011)

accelerating rigid body simulation on today’s...

Documents