Orion Granatir Omar Rodriguez GDC 3/12/10 Don’t Dread Threads.

Orion Granatir Omar Rodriguez

GDC 3/12/10

Don’t Dread Threads

2

Agenda

• Threading is worthwhile

• Data decomposition is a good place to start

• Think tasks!!

• Intel tools help make things easy

3

Threading is important!!

4

Multi-core Needs Parallel Applications

Threading is required to maximize performance

GHz Era Multi-core Era

APP PERFORMANCE

TIME

PLATFORM POTENTIAL

PERF

ORM

ANCE

Parallel

Serial

33 FPS in our demo

104 FPS in our demo

5

Follow these steps to add threading…

1.Use data decomposition

2.Use tasks

6

Functional decomposition is limited

Core Core Core Core

7


Core Core Core Core

8


Core Core Core Core

• Potential latency with pipelining

• Poor load balancing

• Doesn’t scale on varying core counts

9

Data decomposition can scale to n-cores

Core Core Core Core

10

Big loops are ideal cases for data decomposition// Loop through each AIfor( int Index = 0; Index < g_NumAI; Index++ ){ // Update each AI for this frame g_AI[ Index ].Update();}

11

Minimize interactions// Loop through each AIfor( int Index = 0; Index < g_NumAI; Index++ ){ // Update each AI for this frame g_AI[ Index ].Update();}

AI 0 AI 1

Set m_HP to 10

12

Minimize interactions// Loop through each AIfor( int Index = 0; Index < g_NumAI; Index++ ){ // Update each AI for this frame g_AI[ Index ].Update();}

AI 0 AI 1

Set m_HP to 10

13

Avoid locking// Loop through each AIfor( int Index = 0; Index < g_NumAI; Index++ ){ // Update each AI for this frame g_AI[ Index ].Update();}

AI 0 AI 1

Set m_HP to 10

14

Read global data, don’t write// Loop through each AIfor( int Index = 0; Index < g_NumAI; Index++ ){ // Update each AI for this frame g_AI[ Index ].Update();}

15

OpenMP is a great way to get started// Loop through each AI#pragma omp parallel forfor( int Index = 0; Index < g_NumAI; Index++ ){ // Update each AI for this frame g_AI[ Index ].Update();}

Serial 6 Core

1.00x 2.31x

Algorithm

~12.0x

16

The next step is to use tasks

Core Core Core Core

17


Core Core Core Core

18


Core Core Core Core

19


Core Core Core Core

• Needed for load balancing (avoid oversubscription)

• Support large chucks of work

• Better utilization of cache

20

Task can be used to parallelize complex problems

Texture Lookup

Data Parallelism

ProcessingSetup

21

Tasks can be arranged in a dependency graph

Texture Lookup

Data Parallelism

ProcessingSetup

22

Dependency graph can be mapped to a thread pool

23

Dependency graph can be mapped to a thread pool

Core

Core

Core

Core

24

Think of a task as a unit of work

A task is a unit of work• It’s run on a thread pool

• It runs to completion

• It has heavy penalties for blocking

• It’s an efficient way to avoid oversubscription

• They adapt to any number of threads/cores … regardless of CPU topology

25

// Update all AIvoid UpdateAI( float DeltaTime ){

for( int Index = 0; Index < g_NumAI; Index++ ) { // Update each AI for this frame g_AI[ Index ].Update(); }}

Data decomposition makes defining tasks easy

26

// Update all AIvoid UpdateAI( float DeltaTime ){ // Determine the number of AI tasks we want to create unsigned int AIGroups = g_NumAI / MAX_AI_PER_GROUP;

for( unsigned int Index = 0; Index < AIGroups; Index++ ) { // Build the task specific data AITaskData* pData = new AITaskData(); pData->m_Start = Index * MAX_AI_PER_GROUP; pData->m_DeltaTime = DeltaTime;

// Submit task SubmitTask( Task_UpdateAI, (void*)pData ); }}

Data decomposition makes defining tasks easy

27

void Task_UpdateAI( void* pTaskData ){ // Read data AITaskData* pData = (AITaskData*)pTaskData; unsigned int Start = pData->m_Start; unsigned int End = pData->m_Start + MAX_AI_PER_GROUP;

// Gap End with max number of AI End = ( End > g_NumAI ) ? g_NumAI : End;

// Loop through all of our AI and update for( unsigned int Index = Start; Index < End; Index++ ) { g_AI[ Index ].Update(); }

// Cleanup delete pData;}

Individual task are run by the thread pool

28

Intel Threading Building Blocks is a good for tasksIntel® Threading Building Blocks (Intel® TBB) has a low-level API to create and process trees of work – each node is a task.

Root

Task

More

Callback

Spawn & Wait

Root

Task

More

Spawn

Wait

Blocking calls go down

Continuations go up

Root

29

Learn more about tasking…

… or get Game Engine Gems 1* and read Brad Werth’s article.

Task-based Multithreading – How to Program for 100 Cores

Presented by Ron Fosner

Friday, March 12 @ 4:30PMSouth 300

* Other names and brands may be claimed as the property of others.

30

Time to look at our example…

31

Hotspots are good candidates for threading

• Use tools like Intel® Vtune™ and Intel®Parallel Studio to locate hotspots.

32

Hotspots are good candidates for threading

• Use tools like Intel® Vtune™ and Intel®Parallel Studio to locate hotspots.

• Intel® Parallel Studio inspector shows that Flock() is the main bottleneck. This is a good place to investigate threading.

33

Validate threading results with Parallel Amplifier

1.

2.

34

Use Parallel Amplifier to validate concurrency

35


• We have “ideal” CPU utilization for Flocking. • Now we can start looking for other hotspots to optimize.

36


• We have “ideal” CPU utilization for Flocking. • Now we can start looking for other hotspots to optimize.• There is still a lot of serial code…

37

Use Parallel Inspector to find threading errors

38


39


• Have a lot of system memory

• Use a reduced data set

• Workload should be repeatable

40

Use other tools as needed… I like Intel® GPA• Intel® Graphics Performance Analyzer is designed for games.

• System Analyzer gives a complete view of system resources (CPU, GPU, Bus)

• Frame Analyzer allows you to dive into a DX frame • Platform View allow you to instrument code to analyze workload balance and execution time.

41

Conclusion

• Threading is required to maximize your game

• Use data decomposition to scale to n-cores

• Use tasks for load balancing and to be platform independent

• Use Intel tools to make your life easier

• Attend: “Task-based Multithreading – How to Program for 100 Cores” this Friday.

42

Email: [email protected] [email protected]

http://www.intel.com/software/gdc

See Intel at GDC: Intel Booth at Expo, North HallIntel Interactive Lounge

Contact Information

mailto:[email protected]

mailto:[email protected]

43

Other Sessions

A Visual Guide to Game and Task Performance on Mass-market PC Game Platforms

Thursday, March 11 @ 4:30PMNorth 122

Building Games for NetbooksFriday, March 12 @ 9AMSouth 310

Simpler Better Faster VectorFriday, March 12 @ 1:30PMNorth 122

44

Other Sessions

Tuning Your Game for Next Generation Intel Graphics


Task-based Multithreading – How to Program for 100 Cores


Please fill out an evaluation form… it’ll help us win a bet

Thank you

Legal Disclaimer INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL® PRODUCTS. NO

LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY INTELLECTUAL PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL’S TERMS AND CONDITIONS OF SALE FOR SUCH PRODUCTS, INTEL ASSUMES NO LIABILITY WHATSOEVER, AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO SALE AND/OR USE OF INTEL® PRODUCTS INCLUDING LIABILITY OR WARRANTIES RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT, COPYRIGHT OR OTHER INTELLECTUAL PROPERTY RIGHT. INTEL PRODUCTS ARE NOT INTENDED FOR USE IN MEDICAL, LIFE SAVING, OR LIFE SUSTAINING APPLICATIONS.

Intel may make changes to specifications and product descriptions at any time, without notice. All products, dates, and figures specified are preliminary based on current expectations, and

are subject to change without notice. Intel, processors, chipsets, and desktop boards may contain design defects or errors known as

errata, which may cause the product to deviate from published specifications. Current characterized errata are available on request.

Performance tests and ratings are measured using specific computer systems and/or components and reflect the approximate performance of Intel products as measured by those tests. Any difference in system hardware or software design or configuration may affect actual performance.

Intel, Intel Inside, and the Intel logo are trademarks of Intel Corporation in the United States and other countries.

Any software source code reprinted in this document is furnished under a software license and may only be used or copied in accordance with the terms of that license

*Other names and brands may be claimed as the property of others. Copyright © 2010 Intel Corporation.

Orion Granatir Omar Rodriguez GDC 3/12/10 Don’t Dread Threads.

Documents

index g

numai index

index aigrou

unsigned int index

number of ai tasks

tasks core

frame g

data decomposition loop