openmp-tutorial

Description

IWOMP 20051An Introduction In to OpenMPRuud van der PasSenior Staff EngineerScalable Systems GroupSun MicrosystemsIWOMP 2005University of OregonEugene, Oregon, USAJune 1-4, 2005An Introduction Into OpenMP Copyright©2005 Sun MicrosystemsRvdP/V1.1IWOMP 2005Outline2❑The OpenMP Programming Model❑OpenMP Guided Tour❑OpenMP Overview● C lauses● Worksharing constructs● Synchronization constructs● Environment variables● Global Data● Runtime functions❑Wrap-upAn Introduction Into OpenMP Copyright©2005 Sun MicrosystemsRvdP/V1.1IWOMP 20053The OpenMP Programming ModelAn Introduction Into OpenMP Copyright©2005 Sun MicrosystemsRvdP/V1.1IWOMP 20054Memory0 1 PAn Introduction Into OpenMP Copyright©2005 Sun MicrosystemsRvdP/V1.1IWOMP 2005Shared Memory Model5Programming ModelTT✔All threads have access to the privatesame, globally shared, privatememory✔Data can be shared or privateShared ✔Shared data is accessible by Memory all threadsT✔Private data can be accessed private only by the threads that owns private itT T ✔Data transfer is transparent to the programmerprivate✔Synchronization takes place, but it is mostly implicitAn Introduction Into OpenMP Copyright©2005 Sun MicrosystemsRvdP/V1.1IWOMP 2005About Data6In a shared memory parallel program variables have a "label" attached to them:☞L abelled "Private" ⇨ Visible to one thread only✔C hange made in local data, is not seen by others✔Example - L ocal variables in a function that is ...

Informations

• Publié par: Mofeg
• Nombre de lectures: 19
• Langue: English

Extrait

IWOMP 2005

1

RvdP/V1.1

An Introduction Into OpenMP

Ruud van der Pas

Senior Staff Engineer Scalable Systems Group Sun Microsystems

IWOMP 2005 University of Oregon Eugene, Oregon, USA June 1-4, 2005

An Introduction Into OpenMP

Copyright©2005 Sun Microsystems

IWOMP 2005 2

RvdP/V1.1

Outline

❑ The OpenMP Programming Model

❑ OpenMP Guided Tour

❑ OpenMP Overview

● Clauses

● Worksharing constructs

● Synchronization constructs

● Environment variables

● Global Data

● Runtime functions

❑ Wrap-up

An Introduction Into OpenMP

Copyright©2005 Sun Microsystems

IWOMP 2005

3

RvdP/V1.1

The OpenMP Programming Model

An Introduction Into OpenMP

Copyright©2005 Sun Microsystems

IWOMP 2005

4

RvdP/V1.1

0

Memory

1

An Introduction Into OpenMP

P

Copyright©2005 Sun Microsystems

gnM maimorrgPelod ah sdaerht llA✔ ystems IW n Micros MO P0250 p ir dro S✔ah avet datared acce is b elbiss ht lla  y Pr s ✔ adre dae ativ eva ccse sott eh  same, globally rahs ,de omem D✔ yr a atn ca sbereha mmargorp eht  ot ntrepansra tisr  ec ,p alkasenot zatironiynch er ✔S t ybt eho  de  yln ace ssce cta ban t arsnef i   ✔ t aDat hat ownshreads t mp icili

T private

Shared Memory Model

t

private

private

T

An Introduction Into OpenMP

 tubi tiom sylts

T

T

RvdP/V1.1

Shared Memory

T

private

T

private

©2005 SuopyrightC 5

IWOMP2 005 6

RvdP/V1.1

About Data

 In a shared memory parallel program variables have a "label" attached to them:

☞ Labelled "Private"⇨Visible to one thread only

✔ Change made in local data, is not seen by others ✔ Example -Local variables in a function that is executed in parallel ☞ Labelled "Shared"⇨Visible to all threads

✔ Change made in global data, is seen by all others ✔ Example -Global data  

An Introduction Into OpenMP

Copyirhgt©0205S un Mircsoystems

The OpenMP execution model

RvdP/V1.1

Fork and Join Model

An Introduction Into OpenMP

Parallel region

Worker Threads

Synchronization

Parallel region

Master Thread

"threads"

7 oCiryp©thg I s OW tsme 005PM2   Sun2005rosy Mic

8 oCt©ghripyun S0520ysorciM  OWI smets 005MP 2

=

RvdP/V1.1

... etc ... An Introduction Into OpenMP

TID = 0 for (i=0,1,2,3,4) i = 0 sum =  b[i=0][j]*c[j] a[0] = sum i = 1 sum =  b[i=1][j]*c[j] a[1] = sum

#pragma omp parallel for default(none) \  private(i,j,sum) shared(m,n,a,b,c) f { or (i=0; i<m; i++) j  sum = 0.0;  for (j=0; j<n; j++)  sum += b[i][j]*c[j];  a[i] = sum;  }

Example - Matrix times vector

TID 1 = for (i=5,6,7,8,9) i = 5 sum =  b[i=5][j]*c[j] a[5] = sum i = 6 sum =  b[i=6][j]*c[j] a[6] = sum

*

i

IWOMP 2005

9

RvdP/V1.1

OpenMP Guided Tour

An Introduction Into OpenMP

Copyright©2005 Sun Microsystems

IWOMP 2005

10

RvdP/V1.1

http://www.openmp.org

http://www.compunity.org

An Introduction Into OpenMP

Copyright©2005 Sun Microsystems

WIMOP2 00 1 5 1

RvdP/V1.1

When to consider using OpenMP?

❑ The compiler may not be able to do the parallelization in   the way you like to see it:

● A loop is not parallelized ✔ data dependency analysis is not able toThe determine whether it is safe to parallelize or not ● The granularity is not high enough ✔ The compiler lacks information to parallelize at the highest possible level ❑ This is when explicit parallelization through OpenMP directives and functions comes into the picture

An Introduction Into OpenMP

Copyirhgt©2005 Sun Mircosytsems