Multiprocessor Game Loops: Lessons from Uncharted 2: Among Thieves

Multiprocessor Game Loops:
Lessons from

Jason Gregory
Naughty Dog, Inc.

Wednesday, January 27, 2010

Agenda

Game Forum Germany 2010 2

Agenda
• Goal:


Agenda
• Goal:
 Learn about modern multiprocessor game engine update loops...


Agenda
• Goal:
 Learn about modern multiprocessor game engine update loops...
 ... by investigating Naughty Dog’s train level.


Agenda


Agenda
• During our investigation, we’ll answer the following questions:


Agenda
 Why did we make the train itself dynamic (not static)?


Agenda
 How does the train move?


Agenda
 How does it tie into the update of other engine systems?


Agenda
 How do player mechanics and animation work on the train?


Agenda
 How do game object attachment hierarchies work?


Agenda
 How are ray and sphere casts used on the train?


Agenda
 How are ray and sphere casts used on the train?
 How do we utilize the PS3’s parallel computing resources?


Static or Dynamic?

5

Could the Train be Static?
• In the past, games with trains in them have used a static train
approach.
 Train is actually stationary.
 Background scrolls by to produce illusion of movement.
• This solves a lot of problems.
 Player mechanics, NPC locomotion, weapon mechanics, etc.
are all the same as in a “regular” non-moving game level.


Was That Good Enough for Us?


• No way!


• No way!
• Stationary train design imposes undue limitations:


• No way!
 Train must move in a perfectly straight line...


• No way!
 ... or along a broad circular arc (constant curvature).


• No way!
 No ups and downs allowed in the track either.


• No way!
 Not nearly as much fun to jump and shoot between trains.


• No way!
 Not nearly as much fun to jump and shoot between trains.

∴ We decided to go for a fully dynamic train.


Train Movement

9

Spline Tracking
• The Uncharted 2 train follows a spline.
 Catmull-Rom.
 2 trackers for realistic movement.


The Master Car
• Each train car is an independent game object (GO).
• One car is designated as the master.
 It moves without regard for the other cars.
 Every other car is a slave: it simply maintains proper spacing with
the car(s) in front of and/or behind it.


Spacing and Speed
• To maintain proper spacing, we must work in terms of arc length.
 Catmull-Roms are parameterized by a unitless quantity u.
 Arc length (s) is not the same thing as u.
 Careful—must use s for spacing and ds/dt for speed, not u, du/dt.

u = 0.5
u = 1
u = 0


Spacing and Speed
• To maintain proper spacing, we must work in terms of arc length.
 Catmull-Roms are parameterized by a unitless quantity u.
 Arc length (s) is not the same thing as u.
 Careful—must use s for spacing and ds/dt for speed, not u, du/dt.
s = 14
u = 0.5
u = 1
s = 0
u = 0 s = 7


Updating the Cars
• Train car game objects need to update in a speciﬁc order:
 Master ﬁrst.
 Then cars in front of master, from master to locomotive.
 Then cars behind master, from master to caboose.


Updating the Cars
 Master ﬁrst.

1


Updating the Cars
 Master ﬁrst.

1 4 ... n
2 3


Updating the Cars
 Master ﬁrst.

1 4 ... n
2 3
n + 1

n + 2 ... m


Updating the Cars
• In the Uncharted 2 engine, game objects are managed as a tree.
 Children update after their parent.
• For the train...
M

M+1

M+2

M+3

M−1

M−2

M−3


Teleporting the Train
• The train sometimes teleports in the game.
 e.g. To transition from a looped section to a straight-away.
 Typically hidden by a camera cut.


Teleporting the Train
• Problem: Teleport the train such that whichever car the player is
on moves to a predeﬁned location on the track.
 To implement, change the master to be the player’s car...
 ... and teleport it to the desired location.
 All other cars follow automatically.


Updating Large-Scale Engine Systems

17

Large-Scale Engine Systems
• Let’s deﬁne large-scale engine systems as:
 Engine components that operate on lots of data...
 ... and require careful performance optimization.

• Examples include:
 skeletal animation,
 collision detection,
 rigid body dynamics,
 rendering, ...


A Simple Approach (That Doesn’t Work)

• Most game programmers ﬁrst
learn about the game loop
from rendering tutorials.



while (!quit)
{
• Most game programmers ﬁrst ReadJoypad();
learn about the game loop UpdateScene();
from rendering tutorials. DrawScene();
FlipBuffers();
}



• Since we need to update our game objects anyway...
UpdateScene() becomes UpdateGameObjects()

• In the spirit of good object-oriented design, we should let
the game objects drive the large-scale engine systems.
(Right???)



while (!quit)
{
ReadJoypad();
UpdateGameObjects();
// DrawScene();
FlipBuffers();
}


void Tank::Update ()
{
MoveTank();
while (!quit) AimTurret();
{ FireIfNecessary();
ReadJoypad();
UpdateGameObjects(); Animate();
// DrawScene(); DetectCollisions();
FlipBuffers(); SimulatePhysics();
} UpdateAudio();
Draw();
}


• The only problem with this is...
it doesn’t work!
• For one thing, it’s just not feasible for some engine systems.
 e.g. Collision detection cannot be done (properly) one object
at a time.
• Need to solve iteratively, account for time of impact (TOI),
• optimize collision detection via spatial subdivision (e.g.
broadphase AABB prune and sweep),
• optimize dynamics by grouping into islands, etc.


• It’s also terribly inefﬁcient:


 potential for duplicated computations,


 possible reallocation of resources,


 poor data and instruction cache coherence,


 poor data and instruction cache coherence,
 not conducive to parallel computation.


Hardware in the Old Days
• On an Intel Pentium 286 (cira 1982):
 CPU speed was the bottleneck (6 MHz).
• ∴ Use if() tests to avoid unnecessary computations.
 Integer instructions faster than ﬂoating-point.
• ∴ Use ﬁxed-point numbers, or FPU.
 More compute power provided by adding transistors.


Modern Hardware
• On a Cell (PS3) or Xenon (Xbox 360):
 Memory access time is the bottleneck.
• L1 cache miss = ~60 cycles.
• L2 cache miss = ~500 cycles.
• ∴ Organize data in compact, contiguous blocks.
• Use struct of arrays rather than array of structs.
 Pipelined architecture = branching & data conversion are slow.
• ∴ Use duplicated computations to avoid if() tests.
 Compute power is provided via parallelism.


Optimizing Large-Scale Updates

Ineﬃcient
GameObject 1 GameObject 2 GameObject 3

Anim  Coll.  Draw  Anim  Coll.  Draw  Anim  Coll.  Draw    ...
Data Data Data Data Data Data Data Data Data



Much beEer!
Anima?on Engine Collision Engine Rendering Engine

GO 1  GO 2  GO 3  GO 1  GO 2  GO 3  GO 1  GO 2  GO 3 
Anim  Anim  Anim  ... Coll.  Coll.  Coll.  ... Draw  Draw  Draw  ...
Data Data Data Data Data Data Data Data Data


• Game objects do not own their large-scale system data...
 ... they point to data that is owned by the large-scale systems.
• Game objects should not mutate their large-scale system data...
 ... they should only request changes.
• That way, each large-scale system can:
 apply any requested changes as part of its batch update,
 organize its data in whatever manner is most efﬁcient.


Anima?on 
Engine
class Tank : public GameObject animControl15 skelInstance15
{
public: animControl42 skelInstance42
// interface...
animControl7 skelInstance7
private:
... ...
// components -- owned by engine subsystems
AnimControl* m_pAnimControl;
RigidBody* m_pRigidBody;
MeshInstance* m_pRenderable;
};

...



class Tank : public GameObject
{
public: Collision/Physics World
// interface...
rigidBody9 collidable9
private:
// components -- owned by engine subsystems rigidBody3 collidable3
RigidBody* m_pRigidBody; rigidBody41 collidable41
}; ... ...

...



class Tank : public GameObject
{ Scene Graph
public:
meshInstance8
// interface...
meshInstance5
private:
// components -- owned by engine subsystems meshInstance37
RigidBody* m_pRigidBody; ...
};

...


while (!quit)
{
ReadJoypad();

g_gameWorld.UpdateAllGameObjects();
//for (each GameObject* pGo)
// pGo->Update();

g_animationEngine.Update();
g_collisionWorld.DetectCollisions();
g_physicsWorld.Simulate();
g_audioEngine.Update();
g_renderingEngine.DrawAndFlipBuffers();
}


Player Mechanics and Animation

30

Player Mechanics on a Moving Object
• Player’s position and orientation represented as an attached frame
of reference:
 reference to parent game object,
 local transform (relative to parent),
 cached world-space transform (with dirty ﬂag).
• Not quite as simple as it may sound!
 Just switching to attached frames got us ~60% of the way there.
 Lots of bugs to ﬁx.
 Special-case handling of transitions between attached frames.


Animation Pipeline Review
1.Animation Update
 Update clocks, trigger keyframed events, detect end-of-animation
conditions, take transitions to new animation states.
2.Pose Blending Phase
 Extract poses from anim clips, blend individual joint poses.
• Generates local joint transforms (parent-relative).


Animation Pipeline Review
3.Global Pose Phase
 Walk hierarchy, calculate global joint transforms (model-space or
world-space) and matrix palette (input to renderer).
4.Post-Processing
 Apply IK, procedural animation, etc.
• Generates new local joint transforms.
• Recalculate global poses if necessary (re-run phase 3).


Game Object Hooks
• It’s convenient to provide game objects with hooks into the
various phases of the animation update.
• What we do at Naughty Dog:
 GameObject::Update()
• Regular GO update; runs before animation.
 GameObject::PostAnimUpdate()
• Runs after Animation Phase 1.
• Game objects may respond to animation events, force
transitions to new animation states, etc.


Game Object Hooks
 GameObject::PostAnimBlending()
• Game objects can apply procedural animation in local space.
 GameObject::PostJointUpdate()
• This is the ﬁrst time during the frame that the global transform
of every joint is known.
• Game objects can apply IK, or use the global space joint
transforms in other ways.


while (!quit)
{
// ...
g_gameWorld.UpdateAllGameObjects();

g_animationEngine.RunPhase1();
g_gameWorld.CallPostAnimUpdateHooks();

g_gameWorld.CallPostAnimBlendingHooks();

g_gameWorld.CallPostJointUpdateHooks();
// ...


Object Attachment Hierarchy

Bucketed Updates

38

Game Object Attachment Hierarchy


Bucketed Game Object Updates
• Can implement this by updating game objects in buckets.
 Game object dependencies represented by a dependency tree.


• Can implement this by updating game objects in buckets.
 Game object dependencies represented by a dependency tree.
 Each bucket corresponds to one level of the tree.
Bucket 0 Bucket 1 Bucket 2 Bucket 3

trainCar7
box12 weapon3

trainCar8 crate21 playerCharacter weapon17

npc65
trainCar9
... ... ... ...



• Simpler to group objects into pre-determined buckets.

Vehicle Bucket Object Bucket Character Bucket Prop Bucket

trainCar7
box12 weapon3

trainCar8 crate21 playerCharacter weapon17

npc65
trainCar9
... ... ... ...


while (!quit)
{
// ...
for (each bucket)
{
g_gameObjectMgr.UpdateObjects(bucket);
AnimateBucket(bucket);
}
g_collphysWorld.CollideAndSimulate(dt);
g_audioEngine.Update(dt);
g_renderingEngine.DrawAndFlipBuffers(dt);
}


void AnimateBucket(bucket)
{
g_animationEngine.UpdateClocks(bucket);
for (each GameObject* pGo in bucket)
pGo->PostAnimUpdate();

g_animationEngine.CalcLocalPoses(bucket);
pGo->PostAnimBlending();

g_animationEngine.CalcGlobalPoses(bucket);
pGo->PostJointUpdate();
}


Ray and Sphere Casts

44

Ray and Sphere Casts
• A collision cast is an instantaneous collision query.
 Input:
• Snapshot of collision geometry in game world at time t.
• One or more rays / moving spheres (capsules) to cast.
 Output:
• Would any of the rays/spheres strike anything?
• If so, what?


Uses for Collision Casting
• Player and NPCs use downward casts to determine what surface
they are standing on.
• NPCs use ray casts to answer line of sight questions.
• Weapons use ray casts to determine bullet impacts.
• and the list goes on...


• Weapons use ray casts to determine bullet impacts.


Inter-Game-Object Queries and Synchronization
• Synchronization problems can arise whenever:
 game object A...
 ... queries game object B.
• Not just limited to ray and sphere casts.
 Any kind of inter-object query can be affected.


Game Object State Vectors
• Can think of a game object as a heterogeneous “state vector.”
• The state of the ith game object is a vector function of time t.

Si(t) = [ ri(t),
vi(t),
mi, ...,
healthi(t),
ammoi(t), ... ]

Bucketing and Game Object State Vectors
• Theoretically, the state vectors of all game objects are updated
instantaneously and in parallel.
t1
SA

SB

SC

• Theoretically, the state vectors of all game objects are updated
instantaneously and in parallel.
t1 t2 = t1 + ∆t
SA SA

SB SB

SC SC

• In practice, updates take time—object states can be inconsistent
during the update.
t1
SA

SB

SC

during the update.
t1 t2 = t1 + ∆t
SA SA

SB

SC

during the update.
t1 t2 = t1 + ∆t
SA SA

SB SB

SC

during the update.
t1 t2 = t1 + ∆t
SA SA

SB SB

SC SC

• Problems arise when we query the state of object A during the
update of object B or C.
t1 t2 = t1 + ∆t
SA SA

SB

SC

• Problems arise when we query the state of object A during the
update of object B or C.
t1 t2 = t1 + ∆t
SA SA

SB
query
SC

• Major contributor to the ubiquitous “one frame off bug.”
• Update ordering via bucketing helps, but only when the following
rule is adhered to:
An object in bucket b may only read the state
of objects in buckets (b - 1), (b - 2), ...
• You can’t reliably read the state of objects in your own bucket!


One Frame Off Bugs in Collision Queries
• Another complication arises because:
 Collision and physics run after the game objects have updated.
 So, the location of all collision geometry is one frame old when
we cast our rays/spheres.



collision geo
is leR behind



collision geo ray cast
is leR behind misses its target


• Problem:
 What we really want is to update each game object’s collision
geometry with its bucket.
• Impossible, because coll/phys update is monolithic—happens
after all game objects have been updated.


• Problem:
 What we really want is to update each game object’s collision
geometry with its bucket.
• Impossible, because coll/phys update is monolithic—happens
after all game objects have been updated.
 ... or is it?


• Observation:
 Ray and sphere casts don’t care about the full collision/physics
update.
• Don’t need contact information, velocity, etc.
• Only need to know where the collision geo will be.
• Solution:
 All we need to do is update the broadphase AABBs between
buckets.


Solving One Frame Off Bugs in Collision Queries
• Instead, we’ll update the AABBs early, in between game object
bucket updates.
 This allows our collision queries to return correct results.


bucket updates.

collision geo
is leR behind


bucket updates.

broadphase
AABB update


bucket updates.

broadphase
AABB update
ray cast
hits


Achieving Parallelism

57

Game Console Architecture: Xbox 360
PowerPC PowerPC PowerPC
Core 0 Core 1 Core 2
L1  L1  L1  L1  L1  L1 
Data Instr. Data Instr. Data Instr.

Shared L2 Cache (1 MB)

Uniﬁed RAM
GPU
(512 MB)


Game Console Architecture: PLAYSTATION 3

PPU SPU0 SPU1 SPU6

L1  L1 
Data Instr. Local Store Local Store ... Local Store
L2 Cache (256 kB) (256 kB) (256 kB)
(512 kB)
EIB
DMAC

System RAM Video RAM
GPU
(256 MB) (256 MB)


Thinking About Modern Multiprocessor Hardware
• We’ll think in terms of multiple hardware threads.
• These could be provided by:
 hyperthreaded CPU,
 multiple cores,
 SPUs on PS3/Cell.


Ways to Achieve Parallelism: Thread per Subsystem

game collision/
animaUon rendering audio
loop physics


Ways to Achieve Parallelism: Fork/Join
set up


set up
fork

thread2
thread0 thread1 thread3


set up
fork

thread2
thread0 thread1 thread3

process 
join
results


Ways to Achieve Parallelism: Jobs
PPU SPU0 SPU1 SPU5
anim  ray cast anim 
pose pose
anim 
ray cast pose anim 
pose
game ...
broad  broad 
loop phase phase visibility

mesh 
visibility visibility proc


PPU SPU0 SPU1 SPU5
pose pose
anim 
kick ray cast pose anim 
pose
game ...
broad  broad 

mesh 


PPU SPU0 SPU1 SPU5
pose pose
anim 
kick ray cast pose anim 
pose
game ...
broad  broad 
wait
mesh 


• Job = [ (code + input data) → output data ]
 Kick job = request job to be scheduled on a HW thread.
 Must wait for job before processing its output data.
• If job is done, wait takes close to zero time.
• If job is not done, main thread (PPU) blocks until it is done.

• Job manager handles scheduling, allocates buffers in local store,
and coordinates DMAs.
 Programmer speciﬁes data sources & destinations via a DMA list.


Parallelism on the Train

while (!quit)
{
// ...

RayCastHandle hRayCast
= kickRayCast(...);

// do other useful work on PPU
// ...

waitRayCast(hRayCast);
processRayCastResults(hRayCast);

// ...
}


RayCastHandle hRayCast = INVALID;
while (!quit)
{
while (!quit)
// ...
{
// ...
RayCastHandle hRayCast
= kickRayCast(...); // wait and re-kick immediately
if (hRayCast.IsValid())
// do other useful work on PPU
{
// ...
}
hRayCast = kickRayCast(...);
// ...
// ...
}
}


• How does parallelism with jobs affect our train design?
 Large-scale engine system updates and ray/sphere casts are
largely asynchronous in U2:AT.
• Main game loop (PPU) kicks jobs on SPU.
• Other work can proceed on the PPU while jobs are running.
• Results picked up later this frame... or next frame.


• Data must be compact and contiguous so it can be DMA’d to the
SPUs.
 We’re doing this already, to maintain good cache coherency.
• The collision world must be locked in order to update broadphase
AABBs:
 Wait until all outstanding ray/sphere jobs are done.
 Lock.
 Update AABBs.
 Unlock.


• We can even do our broadphase AABB updates asynchronously.

PPU Update Bucket 0



Kick 
BP Job



Kick 
BP Job

Broadphase 
SPU AABB Job



Kick 
BP Job
Update Bucket 1

Broadphase 
SPU AABB Job



query/cast

Kick 
BP Job
Update Bucket 1

Broadphase 
SPU AABB Job


Conclusions

69

Conclusions
• The combination of modern hardware restrictions and the
problems generated by the train level...
 forced us to design our engine in a robust and efﬁcient manner.
• Results:
 U2:AT boasts near 100% hardware utilization on PS3
(PPU + all 6 SPUs).
 Way-cool train level as pay-off for all the hard work.
 Technology was the primary enabler for the convoy level as well.


Thanks For Listening!
• Free free to send questions to me at:
jason_gregory@naughtydog.com


Multiprocessor Game Loops: Lessons from Uncharted 2: Among Thieves

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (6)

Recently uploaded

Recently uploaded (20)

Multiprocessor Game Loops: Lessons from Uncharted 2: Among Thieves