OpenMP 3.0 specification

1
2
3
OpenMP Application
Program Interface
4
Version 3.0 May 2008
5
6
7
8
9
Copyright © 1997-2008 OpenMP Architecture Review Board.
Permission to copy without fee all or part of this material is granted,
provided the OpenMP Architecture Review Board copyright notice and the
title of this document appear. Notice is given that copying is by permission
of OpenMP Architecture Review Board.
10
1
1.
Introduction
...............................................1
2
1.1
Scope
................................................1
3
1.2
Glossary
..............................................2
4
1.2.1
Threading Concepts
5
1.2.2
OpenMP language terminology
6
1.2.3
Tasking Terminology
7
1.2.4
Data Terminology
8
1.2.5
Implementation Terminology
9
1.3
Execution Model
10
1.4
Memory Model
..............................2
......................2
..............................8
.................................9
. . . . . . . . . . . . . . . . . . . . . . . . 10
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
11
1.4.1
Structure of the OpenMP Memory Model
12
1.4.2
The Flush Operation
13
1.4.3
OpenMP Memory Consistency
. . . . . . . . . . . . . . . 13
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
. . . . . . . . . . . . . . . . . . . . . . 16
14
1.5
OpenMP Compliance
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
15
1.6
Normative References
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
16
1.7
Organization of this document
17
18
2.
Directives
2.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Directive Format
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
19
2.1.1
Fixed Source Form Directives
. . . . . . . . . . . . . . . . . . . . . . . 23
20
2.1.2
Free Source Form Directives
. . . . . . . . . . . . . . . . . . . . . . . . 24
21
2.2
Conditional Compilation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
22
2.2.1
Fixed Source Form Conditional Compilation Sentinels
23
2.2.2
Free Source Form Conditional Compilation Sentinel
. . . . . . 27
Internal Control Variables
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
25
2.3.1
ICV Descriptions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
26
2.3.2
Modifying and Retrieving ICV Values
27
2.3.3
How the Per-task ICVs Work
28
2.3.4
ICV Override Relationships
24
29
30
2.3
. . . . 26
2.4
parallel Construct
. . . . . . . . . . . . . . . . . . 29
. . . . . . . . . . . . . . . . . . . . . . . . 30
. . . . . . . . . . . . . . . . . . . . . . . . . 30
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
i
2.4.1
1
parallel Region
2
3
Determining the Number of Threads for a
2.5
4
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Worksharing Constructs
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Loop Construct
2.5.1.1
5
Determining the Schedule of a
Worksharing Loop
6
7
2.5.2
sections Construct
8
2.5.3
single Construct
9
2.5.4
workshare Construct
10
2.6
. . . . . . . . . . . . . . . . . . . . . . . . 45
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Combined Parallel Worksharing Constructs
11
2.6.1
Parallel Loop construct
12
2.6.2
parallel sections Construct
13
2.6.3
parallel workshare Construct
14
2.7
2.7.1
15
16
task Construct
2.8
Task Scheduling
Master and Synchronization Constructs
master Construct
18
2.8.2
critical Construct
19
2.8.3
barrier Construct
20
2.8.4
taskwait Construct
21
2.8.5
atomic Construct
22
2.8.6
flush Construct
23
2.8.7
ordered Construct
25
Data Environment
2.9.1
2.9.1.2
28
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
. . . . . . . . . . . . . . . . . . . . . . . . 77
Data-sharing Attribute Rules for Variables
. . . . . . . . . . . . . . . . . . 78
Data-sharing Attribute Rules for Variables
Referenced in a Region but not in a Construct
29
31 ii
. . . . . . . . . . . . . . . . . . . . . . 63
Referenced in a Construct
27
30
. . . . . . . . . . . . . . . . . . . . 58
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Data-sharing Attribute Rules
2.9.1.1
26
. . . . . . . . . . . . . . . . . . . . . 56
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.8.1
2.9
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
17
24
. . . . . . . . . . . . . . . . . . . 54
2.9.2
threadprivate Directive
OpenMP ABI • Version 3.0 May 2008
. . 80
. . . . . . . . . . . . . . . . . . . . . . . . . 81
2.9.3
1
Data-Sharing Attribute Clauses
. . . . . . . . . . . . . . . . . . . . . . 85
2
2.9.3.1
default clause
3
2.9.3.2
shared clause
4
2.9.3.3
private clause
5
2.9.3.4
firstprivate clause
6
2.9.3.5
lastprivate clause
7
2.9.3.6
reduction clause
2.9.4
8
Data Copying Clauses
. . . . . . . . . . . . . . . . . . . . . . . . . 86
. . . . . . . . . . . . . . . . . . . . . . . . . . 88
. . . . . . . . . . . . . . . . . . . . . . . . . 89
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
2.9.4.1
copyin clause
10
2.9.4.2
copyprivate clause
2.10
12
3.
Nesting of Regions
Runtime Library Routines
. . . . . . . . . . . . . . . . . . . . . 94
. . . . . . . . . . . . . . . . . . . . . . . 96
9
11
. . . . . . . . . . . . . . . . . . . . 92
. . . . . . . . . . . . . . . . . . . . . . . . . 101
. . . . . . . . . . . . . . . . . . . . 102
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
13
3.1
Runtime Library Definitions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
14
3.2
Execution Environment Routines
. . . . . . . . . . . . . . . . . . . . . . . . . . . 109
15
3.2.1
omp_set_num_threads
. . . . . . . . . . . . . . . . . . . . . . . . . . 110
16
3.2.2
omp_get_num_threads
. . . . . . . . . . . . . . . . . . . . . . . . . . 111
17
3.2.3
omp_get_max_threads
. . . . . . . . . . . . . . . . . . . . . . . . . . 112
18
3.2.4
omp_get_thread_num
19
3.2.5
omp_get_num_procs
20
3.2.6
omp_in_parallel
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
21
3.2.7
omp_set_dynamic
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
22
3.2.8
omp_get_dynamic
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
23
3.2.9
omp_set_nested
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
24
3.2.10 omp_get_nested
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
25
3.2.11 omp_set_schedule
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
26
3.2.12 omp_get_schedule
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
27
3.2.13 omp_get_thread_limit
28
3.2.14 omp_set_max_active_levels
. . . . . . . . . . . . . . . . . . . . 126
29
3.2.15 omp_get_max_active_levels
. . . . . . . . . . . . . . . . . . . . 127
30
. . . . . . . . . . . . . . . . . . . . . . . . . . . 113
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
. . . . . . . . . . . . . . . . . . . . . . . . . 125
iii
1
3.2.16 omp_get_level
2
3.2.17 omp_get_ancestor_thread_num
3
3.2.18 omp_get_team_size
4
3.2.19 omp_get_active_level
3.3
5
Lock Routines
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
. . . . . . . . . . . . . . . . . . 130
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
. . . . . . . . . . . . . . . . . . . . . . . . . 133
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6
3.3.1
omp_init_lock and omp_init_nest_lock
7
3.3.2
omp_destroy_lock and omp_destroy_nest_lock
8
3.3.3
omp_set_lock and omp_set_nest_lock
9
3.3.4
omp_unset_lock and omp_unset_nest_lock
10
3.3.5
omp_test_lock and omp_test_nest_lock
3.4
11
Timing Routines
. . . . . . . . . . 136
. . . 137
. . . . . . . . . . . . 138
. . . . . . . . 140
. . . . . . . . . . 141
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
12
3.4.1
omp_get_wtime
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
13
3.4.2
omp_get_wtick
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
14
4.
Environment Variables
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
15
4.1
OMP_SCHEDULE
16
4.2
OMP_NUM_THREADS
17
4.3
OMP_DYNAMIC
18
4.4
OMP_NESTED
19
4.5
OMP_STACKSIZE
20
4.6
OMP_WAIT_POLICY
21
4.7
OMP_MAX_ACTIVE_LEVELS
22
4.8
OMP_THREAD_LIMIT
23
A. Examples
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
24
A.1
A Simple Parallel Loop
25
A.2
The OpenMP Memory Model
26
A.3
Conditional Compilation
27
A.4
Internal Control Variables
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
28
A.5
The parallel Construct
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
29
A.6
The num_threads Clause
30 iv
OpenMP ABI • Version 3.0 May 2008
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
1
A.7
Fortran Restrictions on the do Construct
. . . . . . . . . . . . . . . . . . . . . 167
2
A.8
Fortran Private Loop Iteration Variables
. . . . . . . . . . . . . . . . . . . . . . 169
3
A.9
The nowait clause
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
4
A.10 The collapse clause
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
5
A.11 The parallel sections Construct
6
A.12 The single Construct
7
A.13 Tasking Constructs
8
A.14 The workshare Construct
9
A.15 The master Construct
10
A.16 The critical Construct
11
A.17 worksharing Constructs Inside a critical Construct
12
A.18 Binding of barrier Regions
13
A.19 The atomic Construct
14
A.20 Restrictions on the atomic Construct
15
A.21 The flush Construct with a List
16
A.22 The flush Construct without a List
17
A.23 Placement of flush, barrier, and taskwait Directives
18
A.24 The ordered Clause and the ordered Construct
19
A.25 The threadprivate Directive
20
A.26 Parallel Random Access Iterator Loop
21
A.27 Fortran Restrictions on shared and private Clauses with
22
Common Blocks
. . . . . . . . . . . . . . . . . . . . . . . 174
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
. . . . . . . . . . 199
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
. . . . . . . . . . . . . . . . . . . . . . . 205
. . . . . . . . . . . . . . . . . . . . . . . . . . . 208
. . . . . . . . . . . . . . . . . . . . . . . . . 211
. . . . . . . 214
. . . . . . . . . . . . . 215
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
. . . . . . . . . . . . . . . . . . . . . . . 226
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
23
A.28 The default(none) Clause
24
A.29 Race Conditions Caused by Implied Copies of Shared Variable
25
in Fortran
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
26
A.30 The private Clause
27
A.31 Reprivatization
28
A.32 Fortran Restrictions on Storage Association with the
29
30
31
private Clause
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
A.33 C/C++ Arrays in a firstprivate Clause
. . . . . . . . . . . . . . . . . . . 240
v
1
A.34 The lastprivate Clause
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
2
A.35 The reduction Clause
3
A.36 The copyin Clause
4
A.37 The copyprivate Clause
5
A.38 Nested Loop Constructs
6
A.39 Restrictions on Nesting of Regions
7
A.40 The omp_set_dynamic and
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
omp_set_num_threads Routines
8
. . . . . . . . . . . . . . . . . . . . . . . . . . 258
. . . . . . . . . . . . . . . . . . . . . . . . . 265
9
A.41 The omp_get_num_threads Routine
10
A.42 The omp_init_lock Routine
11
A.43 Ownership of Locks
12
A.44 Simple Lock Routines
13
A.45 Nestable Lock Routines
14
. . . . . . . . . . . . . . . . . . . . . . 266
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
B. Stubs for Runtime Library Routines
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
15
B.1
C/C++ Stub Routines
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
16
B.2
Fortran Stub Routines
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
17
C. OpenMP C and C++ Grammar
18
C.1
Notation
19
C.2
Rules
20
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
D. Interface Declarations
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
21
D.1
Example of the omp.h Header File
. . . . . . . . . . . . . . . . . . . . . . . . . 302
22
D.2
Example of an Interface Declaration include File
23
D.3
Example of a Fortran 90 Interface Declaration module
24
D.4
Example of a Generic Interface for a Library Routine
25
E. Implementation Defined Behaviors in OpenMP
26
F. Changes from Version 2.5 to Version 3.0
27 vi
OpenMP ABI • Version 3.0 May 2008
. . . . . . . . . . . . . 304
. . . . . . . . . . 306
. . . . . . . . . . . . 310
. . . . . . . . . . . . . . . . . . . 311
. . . . . . . . . . . . . . . . . . . . . . . . 315
1
2
CHAPTER
1
Introduction
3
4
11
This document specifies a collection of compiler directives, library routines, and
environment variables that can be used to specify shared-memory parallelism in C, C++
and Fortran programs. This functionality collectively defines the specification of the
OpenMP Application Program Interface (OpenMP API). This specification provides a
model for parallel programming that is portable across shared memory architectures
from different vendors. Compilers from numerous vendors support the OpenMP API.
More information about OpenMP can be found at the following web site:
12
http://www.openmp.org
13
The directives, library routines, and environment variables defined in this document
allow users to create and manage parallel programs while permitting portability. The
directives extend the C, C++ and Fortran base languages with single program multiple
data (SPMD) constructs, tasking constructs, worksharing constructs, and
synchronization constructs, and they provide support for sharing and privatizing data.
The functionality to control the runtime environment is provided by library routines and
environment variables. Compilers that support the OpenMP API often include a
command line option to the compiler that activates and allows interpretation of all
OpenMP directives.
5
6
7
8
9
10
14
15
16
17
18
19
20
21
22
23
1.1
Scope
29
The OpenMP API covers only user-directed parallelization, wherein the user explicitly
specifies the actions to be taken by the compiler and runtime system in order to execute
the program in parallel. OpenMP-compliant implementations are not required to check
for data dependencies, data conflicts, race conditions, or deadlocks, any of which may
occur in conforming programs. In addition, compliant implementations are not required
to check for code sequences that cause a program to be classified as non-conforming.
30
1
24
25
26
27
28
The user is responsible for using OpenMP in his application to produce a conforming
program. OpenMP does not cover compiler-generated automatic parallelization and
directives to the compiler to assist such parallelization.
1
2
3
4
5
1.2
Glossary
6
1.2.1
Threading Concepts
thread
7
8
OpenMP thread
9
thread-safe routine
10
11
12
1.2.2
13
A thread that is managed by the OpenMP runtime system.
A routine that performs the intended function even when executed
concurrently (by more than one thread).
OpenMP language terminology
base language
14
A programming language that serves as the foundation of the OpenMP
specification.
COMMENT: See Section 1.6 on page 17 for a listing of current base
languages for OpenMP.
15
16
17
base program
18
structured block
19
An execution entity with a stack and associated static memory, called
threadprivate memory.
A program written in a base language.
For C/C++, an executable statement, possibly compound, with a single entry
at the top and a single exit at the bottom, or an OpenMP construct.
21
For Fortran, a block of executable statements with a single entry at the top and
a single exit at the bottom, or an OpenMP construct.
22
COMMENTS:
20
23
For all base languages,
24
•
Access to the structured block must not be the result of a branch.
25
•
The point of exit cannot be a branch out of the structured block.
26 2
OpenMP API • Version 3.0 May 2008
1
For C/C++:
2
•
The point of entry must not be a call to setjmp().
3
•
longjmp() and throw() must not violate the entry/exit criteria.
4
•
Calls to exit() are allowed in a structured block.
5
•
An expression statement, iteration statement, selection statement,
or try block is considered to be a structured block if the
corresponding compound statement obtained by enclosing it in {
and } would be a structured block.
6
7
8
9
For Fortran:
10
•
11
directive
12
STOP statements are allowed in a structured block.
In C/C++, a #pragma, and in Fortran, a comment, that specifies OpenMP
program behavior.
COMMENT: See Section 2.1 on page 22 for a description of OpenMP
directive syntax.
13
14
15
white space
16
OpenMP program
17
A non-empty sequence of space and/or horizontal tab characters.
A program that consists of a base program, annotated with OpenMP directives
and runtime library routines.
conforming program
An OpenMP program that follows all the rules and restrictions of the
OpenMP specification.
declarative directive
22
An OpenMP directive that may only be placed in a declarative context. A
declarative directive has no associated executable user code, but instead has
one or more associated user declarations.
23
COMMENT: Only the threadprivate directive is a declarative directive.
18
19
20
21
24
executable directive
25
COMMENT: All directives except the threadprivate directive are
executable directives.
26
27
28
29
30
31
An OpenMP directive that is not declarative; i.e., it may be placed in an
executable context.
stand-alone directive
An OpenMP executable directive that has no associated executable user code.
COMMENT: Only the barrier, flush, and taskwait directives are
stand-alone directives.
Chapter 1
Introduction
3
2
An OpenMP executable directive whose associated user code must be a
simple (single, non-compound) executable statement.
3
COMMENT: Only the atomic directive is a simple directive.
1
simple directive
5
An OpenMP executable directive whose associated user code must be a loop
nest that is a structured block.
6
COMMENTS:
4
loop directive
7
For C/C++, only the for directive is a loop directive.
8
For Fortran, only the do directive and the optional end do directive
are loop directives.
9
10
associated loop(s)
The loop(s) controlled by a loop directive.
COMMENT: If the loop directive contains a collapse clause then there
may be more than one associated loop.
11
12
construct
An OpenMP executable directive (and for Fortran, the paired end directive, if
any) and the associated statement, loop or structured block, if any, not
including the code in any called routines; i.e., the lexical extent of an
executable directive.
region
22
All code encountered during a specific instance of the execution of a given
construct or of an OpenMP library routine. A region includes any code in
called routines as well as any implicit code introduced by the OpenMP
implementation. The generation of a task at the point where a task directive
is encountered is a part of the region of the encountering thread, but the
explicit task region associated with the task directive is not.
23
COMMENTS:
13
14
15
16
17
18
19
20
21
A region may also be thought of as the dynamic or runtime extent of a
construct or of an OpenMP library routine.
24
25
During the execution of an OpenMP program, a construct may give
rise to many regions.
26
27
28
active parallel region
29
30
31
32 4
inactive parallel
region
A parallel region that is executed by a team consisting of more than one
thread.
A parallel region that is executed by a team of only one thread.
OpenMP API • Version 3.0 May 2008
3
All code encountered during the execution of an OpenMP program that is not
part of a parallel region corresponding to a parallel construct or a
task region corresponding to a task construct.
4
COMMENTS:
1
sequential part
2
The sequential part executes as if it were enclosed by an inactive
parallel region.
5
6
Executable statements in called routines may be in both the sequential
part and any number of explicit parallel regions at different points
in the program execution.
7
8
9
10
master thread
The thread that encounters a parallel construct, creates a team, generates
a set of tasks, then executes one of those tasks as thread number 0.
parent thread
The thread that encountered the parallel construct and generated a
parallel region is the parent thread of each of the threads in the team of
that parallel region. The master thread of a parallel region is the
same thread as its parent thread with respect to any resources associated with
an OpenMP thread.
11
12
13
14
15
16
17
ancestor thread
18
20
A set of one or more threads participating in the execution of a parallel
region.
21
COMMENTS:
19
team
For a given thread, its parent thread or one of its parent thread’s ancestor
threads.
For an active parallel region, the team comprises the master thread
and at least one additional thread.
22
23
For an inactive parallel region, the team comprises only the master
thread.
24
25
26
initial thread
27
implicit parallel
region
28
29
30
nested construct
31
nested region
32
33
34
35
The thread that executes the sequential part.
The inactive parallel region that encloses the sequential part of an OpenMP
program.
A construct (lexically) enclosed by another construct.
A region (dynamically) enclosed by another region; i.e., a region encountered
during the execution of another region.
COMMENT: Some nestings are conforming and some are not. See
Section 2.10 on page 104 for the restrictions on nesting.
Chapter 1
Introduction
5
1
closely nested region
2
3
all threads
4
current team
5
encountering thread
6
all tasks
7
current team tasks
8
9
10
11
12
generating task
13
binding thread set
14
16
18
21
22
23
24
25 6
All threads in the team executing the innermost enclosing parallel region
For a given region, the thread that encounters the corresponding construct.
All tasks participating in the OpenMP program.
All tasks encountered during the execution of the innermost enclosing
parallel region by the threads of the corresponding team. Note that the
implicit tasks constituting the parallel region and any descendant tasks
encountered during the execution of these implicit tasks are included in this
binding task set.
For a given region the task whose execution by a thread generated the region.
The set of threads that are affected by, or provide the context for, the
execution of a region.
COMMENT: The binding thread set for a particular region is described in its
corresponding subsection of this specification.
17
20
All OpenMP threads participating in the OpenMP program.
The binding thread set for a given region can be all threads, the current team,
or the encountering thread.
15
19
A region nested inside another region with no parallel region nested
between them.
binding task set
The set of tasks that are affected by, or provide the context for, the execution
of a region.
The binding task set for a given region can be all tasks, the current team
tasks, or the generating task.
COMMENT: The binding task set for a particular region (if applicable) is
described in its corresponding subsection of this specification.
OpenMP API • Version 3.0 May 2008
1
binding region
2
The enclosing region that determines the execution context and limits the
scope of the effects of the bound region is called the binding region.
5
Binding region is not defined for regions whose binding thread set is all
threads or the encountering thread, nor is it defined for regions whose binding
task set is all tasks.
6
COMMENTS:
3
4
The binding region for an ordered region is the innermost enclosing
loop region.
7
8
The binding region for a taskwait region is the innermost enclosing
task region.
9
10
For all other regions for which the binding thread set is the current
team or the binding task set is the current team tasks, the binding
region is the innermost enclosing parallel region.
11
12
13
For regions for which the binding task set is the generating task, the
binding region is the region of the generating task.
14
15
17
A parallel region need not be active nor explicit to be a binding
region.
18
A task region need not be explicit to be a binding region.
19
A region never binds to any region outside of the innermost enclosing
parallel region.
16
20
21
orphaned construct
22
23
24
worksharing
construct
A construct that gives rise to a region whose binding thread set is the current
team, but that is not nested within another construct giving rise to the binding
region.
26
A construct that defines units of work, each of which is executed exactly once
by one of the threads in the team executing the construct.
27
For C, worksharing constructs are for, sections, and single.
28
For Fortran, worksharing constructs are do, sections, single and
workshare.
25
29
30
sequential loop
31
barrier
32
33
34
35
A loop that is not associated with any OpenMP loop directive.
A point in the execution of a program encountered by a team of threads,
beyond which no thread in the team may execute until all threads in the team
have reached the barrier and all explicit tasks generated by the team have
executed to completion.
Chapter 1
Introduction
7
1
1.2.3
Tasking Terminology
3
A specific instance of executable code and its data environment, generated
when a thread encounters a task construct or a parallel construct.
4
COMMENT: When a thread executes a task, it produces a task region.
2
5
task
task region
A region consisting of all code encountered during the execution of a task.
COMMENT: A parallel region consists of one or more implicit task
regions.
6
7
8
explicit task
A task generated when a task construct is encountered during execution.
9
implicit task
A task generated by the implicit parallel region or generated when a
parallel construct is encountered during execution.
10
11
initial task
12
current task
13
14
child task
15
The implicit task associated with the implicit parallel region.
For a given thread, the task corresponding to the task region in which it is
executing.
A task is a child task of the region of its generating task. A child task region
is not part of its generating task region.
descendant task
A task that is the child task of a task region or of one of its descendant task
regions.
task completion
19
Task completion occurs when the end of the structured block associated with
the construct that generated the task is reached.
20
COMMENT: Completion of the initial task occurs at program exit.
16
17
18
23
A point during the execution of the current task region at which it can be
suspended to be resumed later; or the point of task completion, after which the
executing thread may switch to a different task region.
24
COMMENT:
21
task scheduling point
22
26
Within tied task regions, task scheduling points only appear in the
following:
27
•
encountered task constructs
28
•
encountered taskwait constructs
29
•
encountered barrier directives
30
•
implicit barrier regions
31
•
at the end of the tied task region
25
32
33 8
task switching
The act of a thread switching from the execution of one task to another task.
OpenMP API • Version 3.0 May 2008
tied task
1
2
untied task
3
4
5
6
7
task synchronization
construct
1.2.4
A task that, when its task region is suspended, can be resumed only by the
same thread that suspended it; that is, the task is tied to that thread.
A task that, when its task region is suspended, can be resumed by any thread
in the team; that is, the task is not tied to any thread.
A taskwait or a barrier construct.
Data Terminology
9
A named data storage block, whose value can be defined and redefined during
the execution of a program.
10
Array sections and substrings are not considered variables.
8
11
variable
private variable
12
13
A variable which is part of another variable (as an array or structure element)
cannot be made private independently of other components.
14
15
16
shared variable
17
18
20
21
23
threadprivate
variable
24
25
27
28
30
31
A variable that is replicated, one instance per thread, by the OpenMP
implementation, so that its name provides access to a different block of
storage for each thread.
A variable which is part of another variable (as an array or structure element)
cannot be made threadprivate independently of the other components, except
for static data members of C++ classes.
26
29
With respect to a given set of task regions that bind to the same parallel
region, a variable whose name provides access to the same block of storage
for each task region.
A variable which is part of another variable (as an array or structure element)
cannot be shared independently of the other components, except for static data
members of C++ classes.
19
22
With respect to a given set of task regions that bind to the same parallel
region, a variable whose name provides access to a different block of storage
for each task region.
threadprivate
memory
The set of threadprivate variables associated with each thread.
Chapter 1
Introduction
9
data environment
1
2
3
defined
4
All the variables associated with the execution of a given task. The data
environment for a given task is constructed from the data environment of the
generating task at the time the task is generated.
For variables, the property of having a valid value.
5
For C:
6
For the contents of variables, the property of having a valid value.
7
For C++:
8
For the contents of variables of POD (plain old data) type, the property of
having a valid value.
9
11
For variables of non-POD class type, the property of having been constructed
but not subsequently destructed.
12
For Fortran:
13
For the contents of variables, the property of having a valid value. For the
allocation or association status of variables, the property of having a valid
status.
10
14
15
COMMENT: Programs that rely upon variables that are not defined are nonconforming programs.
16
17
class type
18
19
20
21
1.2.5
Implementation Terminology
supporting n levels of
parallelism
22
23
supporting OpenMP
24
supporting nested
parallelism
25
26
27
28
29
30
31 10
For C++: Variables declared with one of the class, struct, or union keywords.
internal control
variable
Implies allowing an active parallel region to be enclosed by n-1 active
parallel regions.
Supporting at least one level of parallelism.
Supporting more than one level of parallelism.
A conceptual variable that specifies run-time behavior of a set of threads or
tasks in an OpenMP program.
COMMENT: The acronym ICV is used interchangeably with the term internal
control variable in the remainder of this specification.
OpenMP API • Version 3.0 May 2008
compliant
implementation
1
2
3
An implementation of the OpenMP specification that compiles and executes
any conforming program as defined by the specification.
COMMENT: A compliant implementation may exhibit unspecified behavior
when compiling or executing a non-conforming program.
4
5
7
A behavior or result that is not specified by the OpenMP specification or not
known prior to the compilation or execution of an OpenMP program.
8
Such unspecified behavior may result from:
9
10
• Issues documented by the OpenMP specification as having unspecified
behavior.
11
• A non-conforming program.
12
• A conforming program exhibiting an implementation defined behavior.
6
unspecified behavior
implementation
defined
13
14
15
16
Behavior that must be documented by the implementation, and which is
allowed to vary among different compliant implementations. An
implementation is allowed to define this behavior as unspecified.
COMMENT: All features that have implementation defined behavior are
documented in Appendix E.
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
1.3
Execution Model
The OpenMP API uses the fork-join model of parallel execution. Multiple threads of
execution perform tasks defined implicitly or explicitly by OpenMP directives. OpenMP
is intended to support programs that will execute correctly both as parallel programs
(multiple threads of execution and a full OpenMP support library) and as sequential
programs (directives ignored and a simple OpenMP stubs library). However, it is
possible and permitted to develop a program that executes correctly as a parallel
program but not as a sequential program, or that produces different results when
executed as a parallel program compared to when it is executed as a sequential program.
Furthermore, using different numbers of threads may result in different numeric results
because of changes in the association of numeric operations. For example, a serial
addition reduction may have a different pattern of addition associations than a parallel
reduction. These different associations may change the results of floating-point addition.
Chapter 1
Introduction
11
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42 12
An OpenMP program begins as a single thread of execution, called the initial thread.
The initial thread executes sequentially, as if enclosed in an implicit task region, called
the initial task region, that is defined by an implicit inactive parallel region
surrounding the whole program.
When any thread encounters a parallel construct, the thread creates a team of itself
and zero or more additional threads and becomes the master of the new team. A set of
implicit tasks, one per thread, is generated. The code for each task is defined by the code
inside the parallel construct. Each task is assigned to a different thread in the team
and becomes tied; that is, it is always executed by the thread to which it is initially
assigned. The task region of the task being executed by the encountering thread is
suspended, and each member of the new team executes its implicit task. There is an
implicit barrier at the end of the parallel construct. Beyond the end of the
parallel construct, only the master thread resumes execution, by resuming the task
region that was suspended upon encountering the parallel construct. Any number of
parallel constructs can be specified in a single program.
parallel regions may be arbitrarily nested inside each other. If nested parallelism is
disabled, or is not supported by the OpenMP implementation, then the new team that is
created by a thread encountering a parallel construct inside a parallel region
will consist only of the encountering thread. However, if nested parallelism is supported
and enabled, then the new team can consist of more than one thread.
When any team encounters a worksharing construct, the work inside the construct is
divided among the members of the team, and executed cooperatively instead of being
executed by every thread. There is an optional barrier at the end of each worksharing
construct. Redundant execution of code by every thread in the team resumes after the
end of the worksharing construct.
When any thread encounters a task construct, a new explicit task is generated.
Execution of explicitly generated tasks is assigned to one of the threads in the current
team, subject to the thread's availability to execute work. Thus, execution of the new
task could be immediate, or deferred until later. Threads are allowed to suspend the
current task region at a task scheduling point in order to execute a different task. If the
suspended task region is for a tied task, the initially assigned thread later resumes
execution of the suspended task region. If the suspended task region is for an untied
task, then any thread may resume its execution. In untied task regions, task scheduling
points may occur at implementation defined points anywhere in the region. In tied task
regions, task scheduling points may occur only in task, taskwait, explicit or
implicit barrier constructs, and at the completion point of the task. Completion of all
explicit tasks bound to a given parallel region is guaranteed before the master thread
leaves the implicit barrier at the end of the region. Completion of a subset of all explicit
tasks bound to a given parallel region may be specified through the use of task
synchronization constructs. Completion of all explicit tasks bound to the implicit
parallel region is guaranteed by the time the program exits.
OpenMP API • Version 3.0 May 2008
Synchronization constructs and library routines are available in OpenMP to coordinate
tasks and data access in parallel regions. In addition, library routines and
environment variables are available to control or to query the runtime environment of
OpenMP programs.
1
2
3
4
OpenMP makes no guarantee that input or output to the same file is synchronous when
executed in parallel. In this case, the programmer is responsible for synchronizing input
and output statements (or routines) using the provided synchronization constructs or
library routines. For the case where each thread accesses a different file, no
synchronization by the programmer is necessary.
5
6
7
8
9
10
11
1.4
Memory Model
12
1.4.1
Structure of the OpenMP Memory Model
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
OpenMP provides a relaxed-consistency, shared-memory model. All OpenMP threads
have access to a place to store and to retrieve variables, called the memory. In addition,
each thread is allowed to have its own temporary view of the memory. The temporary
view of memory for each thread is not a required part of the OpenMP memory model,
but can represent any kind of intervening structure, such as machine registers, cache, or
other local storage, between the thread and the memory. The temporary view of memory
allows the thread to cache variables and thereby to avoid going to memory for every
reference to a variable. Each thread also has access to another type of memory that must
not be accessed by other threads, called threadprivate memory.
A directive that accepts data-sharing attribute clauses determines two kinds of access to
variables used in the directive’s associated structured block: shared and private. Each
variable referenced in the structured block has an original variable, which is the variable
by the same name that exists in the program immediately outside the construct. Each
reference to a shared variable in the structured block becomes a reference to the original
variable. For each private variable referenced in the structured block, a new version of
the original variable (of the same type and size) is created in memory for each task that
contains code associated with the directive. Creation of the new version does not alter
the value of the original variable. However, the impact of attempts to access the original
variable during the region associated with the directive is unspecified; see
Section 2.9.3.3 on page 89 for additional details. References to a private variable in the
structured block refer to the current task’s private version of the original variable. The
Chapter 1
Introduction
13
relationship between the value of the original variable and the initial or final value of the
private version depends on the exact clause that specifies it. Details of this issue, as well
as other issues with privatization, are provided in Section 2.9 on page 77.
1
2
3
The minimum size at which memory accesses by multiple threads without
synchronization, either to the same variable or to different variables that are part of the
same variable (as array or structure elements), are atomic with respect to each other, is
implementation defined. Any additional atomicity restrictions, such as alignment, are
implementation defined.
4
5
6
7
8
A single access to a variable may be implemented with multiple load or store
instructions, and hence is not guaranteed to be atomic with respect to other accesses to
the same variable. Accesses to variables smaller than the implementation-defined
minimum size or to C or C++ bit-fields may be implemented by reading, modifying, and
rewriting a larger unit of memory, and may thus interfere with updates of variables or
fields in the same unit of memory.
9
10
11
12
13
14
If multiple threads write without synchronization to the same memory unit, including
cases due to atomicity considerations as described above, then a data race occurs.
Similarly, if at least one thread reads from a memory unit and at least one thread writes
without synchronization to that same memory unit, including cases due to atomicity
considerations as described above, then a data race occurs. If a data race occurs then the
result of the program is unspecified.
15
16
17
18
19
20
A private variable in a task region that eventually generates an inner nested parallel
region is permitted to be made shared by implicit tasks in the inner parallel region.
A private variable in a task region can be shared by an explicit task region generated
during its execution. However, it is the programmer’s responsibility to ensure through
synchronization that the lifetime of the variable does not end before completion of the
explicit task region sharing it. Any other access by one task to the private variables of
another task results in unspecified behavior.
21
22
23
24
25
26
27
28
1.4.2
29
30
31
32
33
34
35 14
The Flush Operation
The memory model has relaxed-consistency because a thread’s temporary view of
memory is not required to be consistent with memory at all times. A value written to a
variable can remain in the thread’s temporary view until it is forced to memory at a later
time. Likewise, a read from a variable may retrieve the value from the thread’s
temporary view, unless it is forced to read from memory. The OpenMP flush operation
enforces consistency between the temporary view and memory.
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
The flush operation is applied to a set of variables called the flush-set. The flush
operation restricts reordering of memory operations that an implementation might
otherwise do. Implementations must not reorder the code for a memory operation for a
given variable, or the code for a flush operation for the variable, with respect to a flush
operation that refers to the same variable.
If a thread has performed a write to its temporary view of a shared variable since its last
flush of that variable, then when it executes another flush of the variable, the flush does
not complete until the value of the variable has been written to the variable in memory.
If a thread performs multiple writes to the same variable between two flushes of that
variable, the flush ensures that the value of the last write is written to the variable in
memory. A flush of a variable executed by a thread also causes its temporary view of the
variable to be discarded, so that if its next memory operation for that variable is a read,
then the thread will read from memory when it may again capture the value in the
temporary view. When a thread executes a flush, no later memory operation by that
thread for a variable involved in that flush is allowed to start until the flush completes.
The completion of a flush of a set of variables executed by a thread is defined as the
point at which all writes to those variables performed by the thread before the flush are
visible in memory to all other threads and that thread’s temporary view of all variables
involved is discarded.
25
The flush operation provides a guarantee of consistency between a thread’s temporary
view and memory. Therefore, the flush operation can be used to guarantee that a value
written to a variable by one thread may be read by a second thread. To accomplish this,
the programmer must ensure that the second thread has not written to the variable since
its last flush of the variable, and that the following sequence of events happens in the
specified order:
26
1. The value is written to the variable by the first thread.
27
2. The variable is flushed by the first thread.
28
3. The variable is flushed by the second thread.
29
4. The value is read from the variable by the second thread.
20
21
22
23
24
30
31
32
33
34
Note – OpenMP synchronization operations, described in Section 2.8 on page 63 and in
Section 3.3 on page 134, are recommended for enforcing this order. Synchronization
through variables is possible; however, it is not recommended since proper timing of
flushes is difficult as shown in Section A.2 on page 154.
35
36
Chapter 1
Introduction
15
1
1.4.3
OpenMP Memory Consistency
The type of relaxed memory consistency provided by OpenMP is similar to weak
ordering as described in S. V. Adve and K. Gharachorloo, “Shared Memory Consistency
Models: A Tutorial”, IEEE Computer, 29(12), pp.66-76, December 1996. Weak ordering
requires that some memory operations be defined as synchronization operations and that
these be ordered with respect to each other. In the context of OpenMP, two flushes of the
same variable are synchronization operations. OpenMP does not apply any other
restriction to the reordering of memory operations executed by a single thread. The
OpenMP memory model is slightly weaker than weak ordering since flushes are not
ordered with respect to each other if their flush-sets have an empty intersection.
2
3
4
5
6
7
8
9
10
12
The restrictions in Section 1.4.2 on page 14 on reordering with respect to flush
operations guarantee the following:
13
• If the intersection of the flush-sets of two flushes performed by two different threads
14
is non-empty, then the two flushes must be completed as if in some sequential order,
seen by all threads.
11
15
• If the intersection of the flush-sets of two flushes performed by one thread is non-
16
empty, then the two flushes must appear to be completed in that thread’s program
order.
17
18
• If the intersection of the flush-sets of two flushes is empty, the threads can observe
19
these flushes in any order.
20
The flush operation can be specified using the flush directive, and is also implied at
various locations in an OpenMP program: see Section 2.8.6 on page 72 for details. For
an example illustrating the memory model, see Section A.2 on page 154.
21
22
23
24
25
1.5
26
27
28
29
30
31
32
33
34
35
36
37 16
OpenMP Compliance
An implementation of the OpenMP API is compliant if and only if it compiles and
executes all conforming programs according to the syntax and semantics laid out in
Chapters 1, 2, 3 and 4. Appendices A, B, C, D, E and F and sections designated as Notes
(see Section 1.7 on page 18) are for information purposes only and are not part of the
specification.
The OpenMP API defines constructs that operate in the context of the base language that
is supported by an implementation. If the base language does not support a language
construct that appears in this document, a compliant OpenMP implementation is not
required to support it, with the exception that for Fortran, the implementation must
allow case insensitivity for directive and API routines names, and must allow identifiers
of more than six characters.
OpenMP API • Version 3.0 May 2008
All library, intrinsic and built-in routines provided by the base language must be threadsafe in a compliant implementation. In addition, the implementation of the base
language must also be thread-safe (e.g., ALLOCATE and DEALLOCATE statements must
be thread-safe in Fortran). Unsynchronized concurrent use of such routines by different
threads must produce correct results (although not necessarily the same as serial
execution results, as in the case of random number generation routines).
1
2
3
4
5
6
In both Fortran 90 and Fortran 95, variables with explicit initialization have the SAVE
attribute implicitly. This is not the case in Fortran 77. However, a compliant OpenMP
Fortran implementation must give such a variable the SAVE attribute, regardless of the
underlying base language version.
7
8
9
10
Appendix E lists certain aspects of the OpenMP API that are implementation-defined. A
compliant implementation is required to define and document its behavior for each of
the items in Appendix E.
11
12
13
14
15
1.6
Normative References
16
• ISO/IEC 9899:1990, Information Technology - Programming Languages - C.
17
This OpenMP API specification refers to ISO/IEC 9899:1990 as C90.
18
• ISO/IEC 9899:1999, Information Technology - Programming Languages - C.
19
This OpenMP API specification refers to ISO/IEC 9899:1999 as C99.
20
• ISO/IEC 14882:1998, Information Technology - Programming Languages - C++.
21
This OpenMP API specification refers to ISO/IEC 14882:1998 as C++.
22
• ISO/IEC 1539:1980, Information Technology - Programming Languages - Fortran.
23
This OpenMP API specification refers to ISO/IEC 1539:1980 as Fortran 77.
24
• ISO/IEC 1539:1991, Information Technology - Programming Languages - Fortran.
25
This OpenMP API specification refers to ISO/IEC 1539:1991 as Fortran 90.
26
• ISO/IEC 1539-1:1997, Information Technology - Programming Languages - Fortran.
27
Chapter 1
Introduction
17
1
This OpenMP API specification refers to ISO/IEC 1539-1:1997 as Fortran 95.
2
Where this OpenMP API specification refers to C, C++ or Fortran, reference is made to
the base language supported by the implementation.
3
4
5
1.7
Organization of this document
6
The remainder of this document is structured as follows:
7
• Chapter 2: Directives
8
• Chapter 3: Runtime Library Routines
9
• Chapter 4: Environment Variables
10
• Appendix A: Examples
11
• Appendix B: Stubs for Runtime Library Routines
12
• Appendix C: OpenMP C and C++ Grammar
13
• Appendix D: Interface Declarations
14
• Appendix E: Implementation Defined Behaviors in OpenMP
15
• Appendix F: Changes from Version 2.5 to Version 3.0
16
Some sections of this document only apply to programs written in a certain base
language. Text that applies only to programs whose base language is C or C++ is shown
as follows:
17
18
19
C/C++
20
C/C++ specific text....
C/C++
21
Text that applies only to programs whose base language is Fortran is shown as follows:
22
Fortran
23
Fortran specific text......
24
Fortran
25
Where an entire page consists of, for example, Fortran specific text, a marker is shown
at the top of the page like this:
26
Fortran (cont.)
27
28
29 18
Some text is for information only, and is not part of the normative specification. Such
text is designated as a note, like this:
OpenMP API • Version 3.0 May 2008
1
2
Note – Non-normative text....
3
4
Chapter 1
Introduction
19
1
20
OpenMP API • Version 3.0 May 2008
1
2
3
CHAPTER
2
Directives
4
6
This chapter describes the syntax and behavior of OpenMP directives, and is divided
into the following sections:
7
• The language-specific directive format (Section 2.1 on page 22)
8
• Mechanisms to control conditional compilation (Section 2.2 on page 26)
9
• Control of OpenMP API ICVs (Section 2.3 on page 28)
10
• Details of each OpenMP directive (Section 2.4 on page 32 to Section 2.10 on page
5
11
104)
12
C/C++
13
14
In C/C++, OpenMP directives are specified by using the #pragma mechanism provided
by the C and C++ standards.
C/C++
15
Fortran
16
17
18
In Fortran, OpenMP directives are specified by using special comments that are
identified by unique sentinels. Also, a special comment form is available for conditional
compilation.
19
Fortran
20
Compilers can therefore ignore OpenMP directives and conditionally compiled code if
support of OpenMP is not provided or enabled. A compliant implementation must
provide an option or interface that ensures that underlying support of all OpenMP
directives and OpenMP conditional compilation mechanisms is enabled. In the
remainder of this document, the phrase OpenMP compilation is used to mean a
compilation with these OpenMP features enabled.
21
22
23
24
25
26
21
1
Fortran
2
Restrictions
3
The following restriction applies to all OpenMP directives:
4
• OpenMP directives may not appear in PURE or ELEMENTAL procedures.
5
Fortran
6
7
2.1
Directive Format
8
C/C++
9
10
11
OpenMP directives for C/C++ are specified with the pragma preprocessing directive.
The syntax of an OpenMP directive is formally specified by the grammar in
Appendix C, and informally as follows:
12
13
#pragma omp directive-name [clause[ [,] clause]...] new-line
18
Each directive starts with #pragma omp. The remainder of the directive follows the
conventions of the C and C++ standards for compiler directives. In particular, white
space can be used before and after the #, and sometimes white space must be used to
separate the words in a directive. Preprocessing tokens following the #pragma omp
are subject to macro replacement.
19
Directives are case-sensitive.
20
An OpenMP executable directive applies to at most one succeeding statement, which
must be a structured block.
14
15
16
17
21
C/C++
22
Fortran
23
OpenMP directives for Fortran are specified as follows:
24
25
26
27
28
29
30
31 22
sentinel directive-name [clause[[,] clause]...]
All OpenMP compiler directives must begin with a directive sentinel. The format of a
sentinel differs between fixed and free-form source files, as described in Section 2.1.1
on page 23 and Section 2.1.2 on page 24.
Directives are case-insensitive. Directives cannot be embedded within continued
statements, and statements cannot be embedded within directives.
OpenMP API • Version 3.0 May 2008
In order to simplify the presentation, free form is used for the syntax of OpenMP
directives for Fortran in the remainder of this document, except as noted.
1
2
3
Fortran
4
Only one directive-name can be specified per directive (note that this includes combined
directives, see Section 2.6 on page 54). The order in which clauses appear on directives
is not significant. Clauses on directives may be repeated as needed, subject to the
restrictions listed in the description of each clause.
5
6
7
Some data-sharing attribute clauses (Section 2.9.3 on page 85), data copying clauses
(Section 2.9.4 on page 100), the threadprivate directive (Section 2.9.2 on page 81)
and the flush directive (Section 2.8.6 on page 72) accept a list. A list consists of a
comma-separated collection of one or more list items.
8
9
10
11
12
C/C++
A list item is a variable name, subject to the restrictions specified in each of the sections
describing clauses and directives for which a list appears.
13
14
C/C++
15
Fortran
A list item is a variable name or a common block name (enclosed in slashes), subject to
the restrictions specified in each of the sections describing clauses and directives for
which a list appears.
16
17
18
19
Fortran
20
Fortran
21
22
2.1.1
Fixed Source Form Directives
The following sentinels are recognized in fixed form source files:
23
24
25
26
27
28
29
30
31
32
33
34
!$omp | c$omp | *$omp
Sentinels must start in column 1 and appear as a single word with no intervening
characters. Fortran fixed form line length, white space, continuation, and column rules
apply to the directive line. Initial directive lines must have a space or zero in column 6,
and continuation directive lines must have a character other than a space or a zero in
column 6.
Comments may appear on the same line as a directive. The exclamation point initiates a
comment when it appears after column 6. The comment extends to the end of the source
line and is ignored. If the first non-blank character after the directive sentinel of an
initial or continuation directive line is an exclamation point, the line is ignored.
Chapter 2
Directives
23
1
Fortran (cont.)
2
4
Note – in the following example, the three formats for specifying the directive are
equivalent (the first line represents the position of the first 9 columns):
5
c23456789
6
!$omp parallel do shared(a,b,c)
7
c$omp parallel do
8
c$omp+shared(a,b,c)
9
c$omp paralleldoshared(a,b,c)
3
10
11
2.1.2
12
Free Source Form Directives
The following sentinel is recognized in free form source files:
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30 24
!$omp
The sentinel can appear in any column as long as it is preceded only by white space
(spaces and tab characters). It must appear as a single word with no intervening
character. Fortran free form line length, white space, and continuation rules apply to the
directive line. Initial directive lines must have a space after the sentinel. Continued
directive lines must have an ampersand as the last nonblank character on the line, prior
to any comment placed inside the directive. Continuation directive lines can have an
ampersand after the directive sentinel with optional white space before and after the
ampersand.
Comments may appear on the same line as a directive. The exclamation point initiates a
comment. The comment extends to the end of the source line and is ignored. If the first
nonblank character after the directive sentinel is an exclamation point, the line is
ignored.
One or more blanks or horizontal tabs must be used to separate adjacent keywords in
directives in free source form, except in the following cases, where white space is
optional between the given pair of keywords:
OpenMP API • Version 3.0 May 2008
1
end critical
2
end do
3
end master
4
end ordered
5
end parallel
6
end sections
7
end single
8
end task
9
end workshare
10
parallel do
11
parallel sections
12
parallel workshare
13
15
Note – in the following example the three formats for specifying the directive are
equivalent (the first line represents the position of the first 9 columns):
16
!23456789
14
17
18
19
20
21
!$omp parallel do &
!$omp shared(a,b,c)
!$omp parallel &
!$omp&do shared(a,b,c)
!$omp paralleldo shared(a,b,c)
22
23
24
25
Fortran
Chapter 2
Directives
25
1
2
2.2
Conditional Compilation
In implementations that support a preprocessor, the _OPENMP macro name is defined to
have the decimal value yyyymm where yyyy and mm are the year and month designations
of the version of the OpenMP API that the implementation supports.
3
4
5
7
If this macro is the subject of a #define or a #undef preprocessing directive, the
behavior is unspecified.
8
For examples of conditional compilation, see Section A.3 on page 161.
6
9
Fortran
The OpenMP API requires Fortran lines to be compiled conditionally, as described in
the following sections.
10
11
12
2.2.1
13
14
15
Fixed Source Form Conditional Compilation
Sentinels
The following conditional compilation sentinels are recognized in fixed form source
files:
16
17
!$ | *$ | c$
19
To enable conditional compilation, a line with a conditional compilation sentinel must
satisfy the following criteria:
20
• The sentinel must start in column 1 and appear as a single word with no intervening
18
21
22
23
24
25
26
27
28
29 26
white space.
• After the sentinel is replaced with two spaces, initial lines must have a space or zero
in column 6 and only white space and numbers in columns 1 through 5.
• After the sentinel is replaced with two spaces, continuation lines must have a
character other than a space or zero in column 6 and only white space in columns 1
through 5.
If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not
met, the line is left unchanged.
OpenMP API • Version 3.0 May 2008
1
Fortran (cont.)
2
5
Note – in the following example, the two forms for specifying conditional compilation
in fixed source form are equivalent (the first line represents the position of the first 9
columns):
6
c23456789
7
!$ 10 iam = omp_get_thread_num() +
8
!$
9
#ifdef _OPENMP
3
4
&
index
10 iam = omp_get_thread_num() +
10
11
&
12
#endif
index
13
15
Free Source Form Conditional Compilation
Sentinel
16
The following conditional compilation sentinel is recognized in free form source files:
14
2.2.2
17
18
!$
20
To enable conditional compilation, a line with a conditional compilation sentinel must
satisfy the following criteria:
21
• The sentinel can appear in any column but must be preceded only by white space.
22
• The sentinel must appear as a single word with no intervening white space.
23
• Initial lines must have a space after the sentinel.
24
• Continued lines must have an ampersand as the last nonblank character on the line,
19
25
26
27
28
29
30
prior to any comment appearing on the conditionally compiled line. (Continued lines
can have an ampersand after the sentinel, with optional white space before and after
the ampersand.)
If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not
met, the line is left unchanged.
Chapter 2
Directives
27
1
4
Note – in the following example, the two forms for specifying conditional compilation
in free source form are equivalent (the first line represents the position of the first 9
columns):
5
c23456789
2
3
6
!$ iam = omp_get_thread_num() +
7
!$&
&
index
#ifdef _OPENMP
8
iam = omp_get_thread_num() +
9
&
index
10
#endif
11
12
13
14
Fortran
15
16
2.3
Internal Control Variables
An OpenMP implementation must act as if there were internal control variables (ICVs)
that control the behavior of an OpenMP program. These ICVs store information such as
the number of threads to use for future parallel regions, the schedule to use for
worksharing loops and whether nested parallelism is enabled or not. The ICVs are given
values at various times (described below) during the execution of the program. They are
initialized by the implementation itself and may be given values through OpenMP
environment variables and through calls to OpenMP API routines. The program can
retrieve the values of these ICVs only through OpenMP API routines.
17
18
19
20
21
22
23
24
For purposes of exposition, this document refers to the ICVs by certain names, but an
implementation is not required to use these names or to offer any way to access the
variables other than through the ways shown in Section 2.3.2 on page 29.
25
26
27
28
2.3.1
29
30 28
ICV Descriptions
The following ICVs store values that affect the operation of parallel regions.
OpenMP API • Version 3.0 May 2008
• dyn-var - controls whether dynamic adjustment of the number of threads is enabled
1
for encountered parallel regions. There is one copy of this ICV per task.
2
• nest-var - controls whether nested parallelism is enabled for encountered parallel
3
regions. There is one copy of this ICV per task.
4
• nthreads-var - controls the number of threads requested for encountered parallel
5
regions. There is one copy of this ICV per task.
6
• thread-limit-var - controls the maximum number of threads participating in the
7
OpenMP program. There is one copy of this ICV for the whole program.
8
• max-active-levels-var - controls the maximum number of nested active parallel
9
regions. There is one copy of this ICV for the whole program.
10
11
The following ICVs store values that affect the operation of loop regions.
12
• run-sched-var - controls the schedule that the runtime schedule clause uses for
loop regions. There is one copy of this ICV per task.
13
• def-sched-var - controls the implementation defined default scheduling of loop
14
15
regions. There is one copy of this ICV for the whole program.
16
The following ICVs store values that affect the program execution.
17
• stacksize-var - controls the stack size for threads that the OpenMP implementation
creates. There is one copy this ICV for the whole program.
18
• wait-policy-var - controls the desired behavior of waiting threads. There is one copy
19
of this ICV for the whole program.
20
21
2.3.2
Modifying and Retrieving ICV Values
23
The following table shows the methods for retrieving the values of the ICVs as well as
their initial values:
24
25
ICV
Scope Ways to modify value
Way to retrieve value
Initial value
26
dyn-var
task
OMP_DYNAMIC
omp_set_dynamic()
omp_get_dynamic()
See comments
below
27
nest-var
task
OMP_NESTED
omp_set_nested()
omp_get_nested()
false
28
nthreads-var
task
OMP_NUM_THREADS
omp_set_num_threads()
omp_get_max_threads()
Implementation
defined
29
run-sched-var
task
OMP_SCHEDULE
omp_set_schedule()
omp_get_schedule()
Implementation
defined
30
def-sched-var
global
(none)
(none)
Implementation
defined
31
stacksize-var
global
OMP_STACKSIZE
(none)
Implementation
defined
22
32
Chapter 2
Directives
29
1
2
ICV
Scope Ways to modify value
Way to retrieve value
Initial value
3
wait-policy-var
global
OMP_WAIT_POLICY
(none)
Implementation
defined
4
thread-limit-var global
OMP_THREAD_LIMIT
omp_get_thread_limit() Implementation
defined
5
max-activelevels-var
OMP_MAX_ACTIVE_LEVELS
omp_set_max_active_
levels()
omp_get_max_active_
levels()
global
See comments
below
6
Comments:
7
• The initial value of dyn-var is implementation defined if the implementation supports
8
dynamic adjustment of the number of threads; otherwise, the initial value is false.
9
• The initial value of max-active-levels-var is the number of levels of parallelism that
the implementation supports. See the definition of supporting n levels of parallelism
in Section 1.2.5 on page 10 for further details.
10
11
15
After the initial values are assigned, but before any OpenMP construct or OpenMP API
routine executes, the values of any OpenMP environment variables that were set by the
user are read and the associated ICVs are modified accordingly. After this point, no
changes to any OpenMP environment variables will affect the ICVs.
16
Clauses on OpenMP constructs do not modify the values of any of the ICVs.
12
13
14
17
2.3.3
Each task region has its own copies of the internal variables dyn-var, nest-var,
nthreads-var, and run-sched-var. When a task construct or parallel construct is
encountered during the execution of a task, the child task(s) generated inherit the values
of these ICVs from the generating task's ICV values.
18
19
20
21
Calls to omp_set_num_threads(), omp_set_dynamic(),
omp_set_nested(), and omp_set_schedule() modify only the ICVs
corresponding to their binding task region.
22
23
24
When encountering a loop worksharing region with schedule(runtime), all
implicit task regions that constitute the binding parallel region must have the same
value for run-sched-var. Otherwise, the behavior is unspecified.
25
26
27
28
How the Per-task ICVs Work
2.3.4
29
30
31 30
ICV Override Relationships
The override relationships among various construct clauses, OpenMP API routines,
environment variables, and the initial values of ICVs are shown in the following table:
OpenMP API • Version 3.0 May 2008
1
2
3
4
construct clause, if
used
5
overrides call to API routine
overrides setting of
environment
variable
overrides initial
value of
(none)
omp_set_dynamic()
OMP_DYNAMIC
dyn-var
6
(none)
omp_set_nested()
OMP_NESTED
nest-var
7
num_threads
omp_set_num_threads()
OMP_NUM_THREADS
nthreads-var
8
schedule
omp_set_schedule()
OMP_SCHEDULE
run-sched-var
9
schedule
(none)
(none)
def-sched-var
10
(none)
(none)
OMP_STACKSIZE
stacksize-var
11
(none)
(none)
OMP_WAIT_POLICY
wait-policy-var
12
(none)
(none)
OMP_THREAD_LIMIT
thread-limit-var
13
(none)
omp_set_max_active_
levels()
OMP_MAX_ACTIVE_
LEVELS
max-active-levels-var
14
Cross References:
15
• parallel construct, see Section 2.4 on page 32.
16
• num_threads clause, see Section 2.4.1 on page 35.
17
• schedule clause, see Section 2.5.1.1 on page 45.
18
• Loop construct, see Section 2.5.1 on page 38.
19
• omp_set_num_threads routine, see Section 3.2.1 on page 110.
20
• omp_get_max_threads routine, see Section 3.2.3 on page 112.
21
• omp_set_dynamic routine, see Section 3.2.7 on page 117.
22
• omp_get_dynamic routine, see Section 3.2.8 on page 118.
23
• omp_set_nested routine, see Section 3.2.9 on page 119.
24
• omp_get_nested routine, see Section 3.2.10 on page 120.
25
• omp_set_schedule routine, see Section 3.2.11 on page 121.
26
• omp_get_schedule routine, see Section 3.2.12 on page 123.
27
• omp_get_thread_limit routine, see Section 3.2.13 on page 125.
28
• omp_set_max_active_levels routine, see Section 3.2.14 on page 126.
29
• omp_get_max_active_levels routine, see Section 3.2.15 on page 127.
30
• OMP_SCHEDULE environment variable, see Section 4.1 on page 146.
31
• OMP_NUM_THREADS environment variable, see Section 4.2 on page 147.
32
• OMP_DYNAMIC environment variable, see Section 4.3 on page 148.
33
Chapter 2
Directives
31
1
• OMP_NESTED environment variable, see Section 4.4 on page 148.
2
• OMP_STACKSIZE environment variable, see Section 4.5 on page 149.
3
• OMP_WAIT_POLICY environment variable, see Section 4.6 on page 150.
4
• OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.7 on page 150.
5
• OMP_THREAD_LIMIT environment variable, see Section 4.8 on page 151.
6
7
2.4
parallel Construct
8
Summary
9
10
This fundamental construct starts parallel execution. See Section 1.3 on page 11 for a
general description of the OpenMP execution model.
11
Syntax
12
C/C++
13
The syntax of the parallel construct is as follows:
14
15
16
17
#pragma omp parallel [clause[ [, ]clause] ...] new-line
structured-block
where clause is one of the following:
18
if(scalar-expression)
19
num_threads(integer-expression)
20
default(shared | none)
21
private(list)
22
firstprivate(list)
23
shared(list)
24
copyin(list)
25
reduction(operator: list)
26
C/C++
27 32
OpenMP API • Version 3.0 May 2008
1
Fortran
2
The syntax of the parallel construct is as follows:
3
4
5
!$omp parallel [clause[[,] clause]...]
6
!$omp end parallel
7
structured-block
where clause is one of the following:
8
if(scalar-logical-expression)
9
num_threads(scalar-integer-expression)
10
default(private | firstprivate | shared | none)
11
private(list)
12
firstprivate(list)
13
shared(list)
14
copyin(list)
15
reduction({operator|intrinsic_procedure_name}:list)
16
The end parallel directive denotes the end of the parallel construct.
17
Fortran
18
Binding
19
20
The binding thread set for a parallel region is the encountering thread. The
encountering thread becomes the master thread of the new team.
21
Description
22
When a thread encounters a parallel construct, a team of threads is created to
execute the parallel region (see Section 2.4.1 on page 35 for more information about
how the number of threads in the team is determined, including the evaluation of the if
and num_threads clauses). The thread that encountered the parallel construct
becomes the master thread of the new team, with a thread number of zero for the
duration of the new parallel region. All threads in the new team, including the
master thread, execute the region. Once the team is created, the number of threads in the
team remains constant for the duration of that parallel region.
23
24
25
26
27
28
29
30
Chapter 2
Directives
33
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
Within a parallel region, thread numbers uniquely identify each thread. Thread
numbers are consecutive whole numbers ranging from zero for the master thread up to
one less than the number of threads in the team. A thread may obtain its own thread
number by a call to the omp_get_thread_num library routine.
A set of implicit tasks, equal in number to the number of threads in the team, is
generated by the encountering thread. The structured block of the parallel construct
determines the code that will be executed in each implicit task. Each task is assigned to
a different thread in the team and becomes tied. The task region of the task being
executed by the encountering thread is suspended and each thread in the team executes
its implicit task. Each thread can execute a path of statements that is different from that
of the other threads.
The implementation may cause any thread to suspend execution of its implicit task at a
task scheduling point, and switch to execute any explicit task generated by any of the
threads in the team, before eventually resuming execution of the implicit task (for more
details see Section 2.7 on page 59).
There is an implied barrier at the end of a parallel region. After the end of a
parallel region, only the master thread of the team resumes execution of the
enclosing task region.
If a thread in a team executing a parallel region encounters another parallel
directive, it creates a new team, according to the rules in Section 2.4.1 on page 35, and
it becomes the master of that new team.
If execution of a thread terminates while inside a parallel region, execution of all
threads in all teams terminates. The order of termination of threads is unspecified. All
work done by a team prior to any barrier that the team has passed in the program is
guaranteed to be complete. The amount of work done by each thread after the last
barrier that it passed and before it terminates is unspecified.
28
For an example of the parallel construct, see Section A.5 on page 164. For an
example of the num_threads clause, see Section A.6 on page 166.
29
Restrictions
30
Restrictions to the parallel construct are as follows:
31
• A program that branches into or out of a parallel region is non-conforming.
32
• A program must not depend on any ordering of the evaluations of the clauses of the
27
33
parallel directive, or on any side effects of the evaluations of the clauses.
34
• At most one if clause can appear on the directive.
35
• At most one num_threads clause can appear on the directive. The num_threads
36
37 34
expression must evaluate to a positive integer value.
OpenMP API • Version 3.0 May 2008
1
C/C++
• A throw executed inside a parallel region must cause execution to resume
2
within the same parallel region, and the same thread that threw the exception
must catch it.
3
4
C/C++
5
Fortran
• Unsynchronized use of Fortran I/O statements by multiple threads on the same unit
6
has unspecified behavior.
7
8
Fortran
9
Cross References
10
• default, shared, private, firstprivate, and reduction clauses, see
Section 2.9.3 on page 85.
11
12
• copyin clause, see Section 2.9.4 on page 100.
13
• omp_get_thread_num routine, see Section 3.2.4 on page 113.
14
15
16
17
18
19
20
21
22
23
24
25
26
27
2.4.1
Determining the Number of Threads for a
parallel Region
When execution encounters a parallel directive, the value of the if clause or
num_threads clause (if any) on the directive, the current parallel context, and the
values of the nthreads-var, dyn-var, thread-limit-var, max-active-level-var, and nest-var
ICVs are used to determine the number of threads to use in the region.
Note that using a variable in an if or num_threads clause expression of a
parallel construct causes an implicit reference to the variable in all enclosing
constructs. The if clause expression and the num_threads clause expression are
evaluated in the context outside of the parallel construct, and no ordering of those
evaluations is specified. It is also unspecified whether, in what order, or how many times
any side-effects of the evaluation of the num_threads or if clause expressions occur.
When a thread encounters a parallel construct, the number of threads is determined
according to Algorithm 2.1.
28
29
Algorithm 2.1
30
let ThreadsBusy be the number of OpenMP threads currently executing;
31
let ActiveParRegions be the number of enclosing active parallel regions;
32
if an if clause exists
33
Chapter 2
Directives
35
1
2
Algorithm 2.1
3
then let IfClauseValue be the value of the if clause expression;
4
else let IfClauseValue = true;
5
if a num_threads clause exists
6
7
then let ThreadsRequested be the value of the num_threads clause
expression;
8
else let ThreadsRequested = nthreads-var;
9
let ThreadsAvailable = (thread-limit-var - ThreadsBusy + 1);
10
if (IfClauseValue = false)
11
then number of threads = 1;
12
else if (ActiveParRegions >= 1) and (nest-var = false)
13
then number of threads = 1;
14
else if (ActiveParRegions = max-active-levels-var)
15
then number of threads = 1;
16
else if (dyn-var = true) and (ThreadsRequested <= ThreadsAvailable)
17
then number of threads = [ 1 : ThreadsRequested ];
18
else if (dyn-var = true) and (ThreadsRequested > ThreadsAvailable)
19
then number of threads = [ 1 : ThreadsAvailable ];
20
else if (dyn-var = false) and (ThreadsRequested <= ThreadsAvailable)
21
then number of threads = ThreadsRequested;
22
else if (dyn-var = false) and (ThreadsRequested > ThreadsAvailable)
23
then behavior is implementation defined;
24
25
26
27
Note – Since the initial value of the dyn-var ICV is implementation defined, programs
that depend on a specific number of threads for correct execution should explicitly
disable dynamic adjustment of the number of threads.
28
29 36
OpenMP API • Version 3.0 May 2008
1
Cross References
2
• nthreads-var, dyn-var, thread-limit-var, max-active-level-var, and nest-var ICVs, see
Section 2.3 on page 28.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
2.5
Worksharing Constructs
A worksharing construct distributes the execution of the associated region among the
members of the team that encounters it. Threads execute portions of the region in the
context of the implicit tasks each one is executing. If the team consists of only one
thread then the worksharing region is not executed in parallel.
A worksharing region has no barrier on entry; however, an implied barrier exists at the
end of the worksharing region, unless a nowait clause is specified. If a nowait
clause is present, an implementation may omit the barrier at the end of the worksharing
region. In this case, threads that finish early may proceed straight to the instructions
following the worksharing region without waiting for the other members of the team to
finish the worksharing region, and without performing a flush operation (see
Section A.9 on page 170 for an example.)
18
OpenMP defines the following worksharing constructs, and these are described in the
sections that follow:
19
• loop construct
20
• sections construct
21
• single construct
22
• workshare construct
23
Restrictions
24
The following restrictions apply to worksharing constructs:
25
• Each worksharing region must be encountered by all threads in a team or by none at
17
26
27
28
29
all.
• The sequence of worksharing regions and barrier regions encountered must be the
same for every thread in a team.
Chapter 2
Directives
37
1
2.5.1
Loop Construct
2
Summary
3
6
The loop construct specifies that the iterations of one or more associated loops will be
executed in parallel by threads in the team in the context of their implicit tasks. The
iterations are distributed across threads that already exist in the team executing the
parallel region to which the loop region binds.
7
Syntax
4
5
8
C/C++
9
The syntax of the loop construct is as follows:
10
11
12
#pragma omp for [clause[[,] clause] ... ] new-line
for-loops
13
where clause is one of the following:
14
private(list)
15
firstprivate(list)
16
lastprivate(list)
17
reduction(operator: list)
18
schedule(kind[, chunk_size])
19
collapse(n)
20
ordered
21
nowait
22 38
OpenMP API • Version 3.0 May 2008
1
2
3
4
C/C++ (cont.)
The for directive places restrictions on the structure of all associated for-loops.
Specifically, all associated for-loops must have the following canonical form:
5
6
for (init-expr; test-expr; incr-expr) structured-block
7
init-expr
12
test-expr
One of the following:
var relational-op b
b relational-op var
incr-expr
One of the following:
13
14
15
++var
var++
--var
var-var += incr
var -= incr
var = var + incr
var = incr + var
var = var - incr
16
17
18
19
20
21
22
23
24
25
var
One of the following:
A variable of a signed or unsigned integer type.
For C++, a variable of a random access iterator type.
For C, a variable of a pointer type.
If this variable would otherwise be shared, it is implicitly made
private in the loop construct. This variable must not be
modified during the execution of the for-loop other than in increxpr. Unless the variable is specified lastprivate on the
loop construct, its value after the loop is unspecified.
relational-op
One of the following:
<
<=
>
>=
lb and b
Loop invariant expressions of a type compatible with the type
of var.
incr
A loop invariant integer expression.
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
One of the following:
var = lb
integer-type var = lb
random-access-iterator-type var = lb
pointer-type var = lb
8
9
10
11
Chapter 2
Directives
39
3
The canonical form allows the iteration count of all associated loops to be computed
before executing the outermost loop. The computation is performed for each loop in an
integer type. This type is derived from the type of var as follows:
4
• If var is of an integer type, then the type is the type of var.
5
• For C++, if var is of a random access iterator type, then the type is the type that
1
2
6
would be used by std::distance applied to variables of the type of var.
7
• For C, if var is of a pointer type, then the type is ptrdiff_t.
8
The behavior is unspecified if any intermediate result required to compute the iteration
count cannot be represented in the type determined above.
9
10
11
12
There is no implied synchronization during the evaluation of the lb, b, or incr
expressions. It is unspecified whether, in what order, or how many times any side effects
within the lb, b, or incr expressions occur.
13
14
15
16
17
Note – Random access iterators are required to support random access to elements in
constant time. Other iterators are precluded by the restrictions since they can take linear
time or offer limited functionality. It is therefore advisable to use tasks to parallelize
those cases.
18
19
C/C++
20
Fortran
21
The syntax of the loop construct is as follows:
22
23
24
25
!$omp do [clause[[,] clause] ... ]
do-loops
[!$omp end do [nowait] ]
26
where clause is one of the following:
27
private(list)
28
firstprivate(list)
29
lastprivate(list)
30
reduction({operator|intrinsic_procedure_name}:list)
31 40
OpenMP API • Version 3.0 May 2008
1
schedule(kind[, chunk_size])
2
collapse(n)
3
ordered
4
5
6
7
8
9
10
11
12
13
If an end do directive is not specified, an end do directive is assumed at the end of the
do-loop.
All associated do-loops must be do-constructs as defined by the Fortran standard. If an
end do directive follows a do-construct in which several loop statements share a DO
termination statement, then the directive can only be specified for the outermost of these
DO statements. See Section A.7 on page 167 for examples.
If any of the loop iteration variables would otherwise be shared, they are implicitly
made private on the loop construct. See Section A.8 on page 169 for examples. Unless
the loop iteration variables are specified lastprivate on the loop construct, their
values after the loop are unspecified.
14
Fortran
15
Binding
16
19
The binding thread set for a loop region is the current team. A loop region binds to the
innermost enclosing parallel region. Only the threads of the team executing the
binding parallel region participate in the execution of the loop iterations and
(optional) implicit barrier of the loop region.
20
Description
21
The loop construct is associated with a loop nest consisting of one or more loops that
follow the directive.
17
18
22
23
24
25
26
27
28
29
30
31
32
33
There is an implicit barrier at the end of a loop construct unless a nowait clause is
specified.
The collapse clause may be used to specify how many loops are associated with the
loop construct. The parameter of the collapse clause must be a constant positive
integer expression. If no collapse clause is present, the only loop that is associated
with the loop construct is the one that immediately follows the construct.
If more than one loop is associated with the loop construct, then the iterations of all
associated loops are collapsed into one larger iteration space which is then divided
according to the schedule clause. The sequential execution of the iterations in all
associated loops determines the order of the iterations in the collapsed iteration space.
Chapter 2
Directives
41
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
The iteration count for each associated loop is computed before entry to the outermost
loop. If execution of any associated loop changes any of the values used to compute any
of the iteration counts then the behavior is unspecified.
The integer type (or kind, for Fortran) used to compute the iteration count for the
collapsed loop is implementation defined.
A worksharing loop has logical iterations numbered 0,1,...,N-1 where N is the number of
loop iterations, and the logical numbering denotes the sequence in which the iterations
would be executed if the associated loop(s) were executed by a single thread. The
schedule clause specifies how iterations of the associated loops are divided into
contiguous non-empty subsets, called chunks, and how these chunks are distributed
among threads of the team. Each thread executes its assigned chunk(s) in the context of
its implicit task. The chunk_size expression is evaluated using the original list items of
any variables that are made private in the loop construct. It is unspecified whether, in
what order, or how many times, any side-effects of the evaluation of this expression
occur. The use of a variable in a schedule clause expression of a loop construct
causes an implicit reference to the variable in all enclosing constructs.
Different loop regions with the same schedule and iteration count, even if they occur in
the same parallel region, can distribute iterations among threads differently. The only
exception is for the static schedule as specified in Table 2-1. Programs that depend
on which thread executes a particular iteration under any other circumstances are nonconforming.
23
See Section 2.5.1.1 on page 45 for details of how the schedule for a worksharing loop is
determined.
24
The schedule kind can be one of those specified in Table 2-1.
22
25 42
OpenMP API • Version 3.0 May 2008
1
TABLE 2-1
schedule clause kind values
2
3
4
5
static
When schedule(static, chunk_size) is specified, iterations are divided
into chunks of size chunk_size, and the chunks are assigned to the threads in
the team in a round-robin fashion in the order of the thread number.
6
7
8
When no chunk_size is specified, the iteration space is divided into chunks that
are approximately equal in size, and at most one chunk is distributed to each
thread. Note that the size of the chunks is unspecified in this case.
9
10
11
12
13
14
15
16
17
A compliant implementation of static schedule must ensure that the same
assignment of logical iteration numbers to threads will be used in two loop
regions if the following conditions are satisfied: 1) both loop regions have the
same number of loop iterations, 2) both loop regions have the same value of
chunk_size specified, or both loop regions have no chunk_size specified, and 3)
both loop regions bind to the same parallel region. A data dependence between
the same logical iterations in two such loops is guaranteed to be satisfied
allowing safe use of the nowait clause (see Section A.9 on page 170 for
examples).
18
19
20
21
dynamic
When schedule(dynamic, chunk_size) is specified, the iterations are
distributed to threads in the team in chunks as the threads request them. Each
thread executes a chunk of iterations, then requests another chunk, until no
chunks remain to be distributed.
22
23
Each chunk contains chunk_size iterations, except for the last chunk to be
distributed, which may have fewer iterations.
24
When no chunk_size is specified, it defaults to 1.
25
26
27
28
guided
When schedule(guided, chunk_size) is specified, the iterations are
assigned to threads in the team in chunks as the executing threads request
them. Each thread executes a chunk of iterations, then requests another chunk,
until no chunks remain to be assigned.
29
30
31
32
33
34
For a chunk_size of 1, the size of each chunk is proportional to the
number of unassigned iterations divided by the number of threads in the team,
decreasing to 1. For a chunk_size with value k (greater than 1), the
size of each chunk is determined in the same way, with the restriction
that the chunks do not contain fewer than k iterations (except for the last chunk
to be assigned, which may have fewer than k iterations).
35
When no chunk_size is specified, it defaults to 1.
36
37
38
39
40
auto
When schedule(auto) is specified, the decision regarding scheduling is
delegated to the compiler and/or runtime system. The programmer gives the
implementation the freedom to choose any possible mapping of iterations to
threads in the team.
Chapter 2
Directives
43
1
2
3
4
runtime
When schedule(runtime) is specified, the decision regarding scheduling
is deferred until run time, and the schedule and chunk size are taken from the
run-sched-var ICV. If the ICV is set to auto, the schedule is implementation
defined.
5
6
7
8
9
10
11
12
13
14
15
16
Note – For a team of p threads and a loop of n iterations, let n ⁄ p be the integer q
that satisfies n = p*q - r, with 0 <= r < p. One compliant implementation of the
static schedule (with no specified chunk_size) would behave as though chunk_size
had been specified with value q. Another compliant implementation would assign q
iterations to the first p-r threads, and q-1 iterations to the remaining r threads. This
illustrates why a conforming program must not rely on the details of a particular
implementation.
A compliant implementation of the guided schedule with a chunk_size value of k
would assign q = n ⁄ p iterations to the first available thread and set n to the larger of
n-q and p*k. It would then repeat this process until q is greater than or equal to the
number of remaining iterations, at which time the remaining iterations form the final
chunk. Another compliant implementation could use the same method, except with q =
n ⁄ ( 2p ) , and set n to the larger of n-q and 2*p*k.
17
18
Restrictions
19
Restrictions to the loop construct are as follows:
20
• All loops associated with the loop construct must be perfectly nested; that is, there
21
must be no intervening code nor any OpenMP directive between any two loops.
22
• The values of the loop control expressions of the loops associated with the loop
23
directive must be the same for all the threads in the team.
24
• Only one schedule clause can appear on a loop directive.
25
• Only one collapse clause can appear on a loop directive.
26
• chunk_size must be a loop invariant integer expression with a positive value.
27
• The value of the chunk_size expression must be the same for all threads in the team.
28
• The value of the run-sched-var ICV must be the same for all threads in the team.
29
• When schedule(runtime) or schedule(auto) is specified, chunk_size must
30
not be specified.
31
• Only a single ordered clause can appear on a loop directive.
32
• The ordered clause must be present on the loop construct if any ordered region
33
34 44
ever binds to a loop region arising from the loop construct.
OpenMP API • Version 3.0 May 2008
1
2
• The loop iteration variable may not appear in a threadprivate directive.
3
• The associated for-loops must be structured blocks.
4
• Only an iteration of the innermost associated loop may be curtailed by a continue
C/C++
statement.
5
6
• No statement can branch to any associated for statement.
7
• Only one nowait clause can appear on a for directive.
8
• If relational-op is < or <= then incr-expr must cause var to increase on each iteration
9
of the loop. Conversely, if relational-op is > or >= then incr-expr must cause var to
decrease on each iteration of the loop.
10
• A throw executed inside a loop region must cause execution to resume within the
11
same iteration of the loop region, and the same thread that threw the exception must
catch it.
12
13
C/C++
14
Fortran
15
• The associated do-loops must be structured blocks.
16
• Only an iteration of the innermost associated loop may be curtailed by a CYCLE
statement.
17
• No statement in the associated loops other than the DO statements can cause a branch
18
out of the loops.
19
20
• The do-loop iteration variable must be of type integer.
21
• The do-loop cannot be a DO WHILE or a DO loop without loop control.
22
Fortran
23
Cross References
24
• private, firstprivate, lastprivate, and reduction clauses, see
Section 2.9.3 on page 85.
25
26
• OMP_SCHEDULE environment variable, see Section 4.1 on page 146.
27
• ordered construct, see Section 2.8.7 on page 75.
28
29
30
31
32
33
2.5.1.1
Determining the Schedule of a Worksharing Loop
When execution encounters a loop directive, the schedule clause (if any) on the
directive, and the run-sched-var and def-sched-var ICVs are used to determine how loop
iterations are assigned to threads. See Section 2.3 on page 28 for details of how the
values of the ICVs are determined. If the loop directive does not have a schedule
Chapter 2
Directives
45
1
5
clause then the current value of the def-sched-var ICV determines the schedule. If the
loop directive has a schedule clause that specifies the runtime schedule kind then
the current value of the run-sched-var ICV determines the schedule. Otherwise, the
value of the schedule clause determines the schedule. Figure 2-1 describes how the
schedule for a worksharing loop is determined.
6
Cross References
7
• ICVs, see Section 2.3 on page 28.
2
3
4
8
START
9
10
11
12
schedule
clause present?
No
Use def-sched-var schedule kind
13
Yes
14
15
16
17
schedule kind
value is runtime?
No
Use schedule kind specified in
schedule clause
18
Yes
19
20
21
22 46
Use run-sched-var schedule kind
FIGURE 2-1
Determining the schedule for a worksharing loop.
OpenMP API • Version 3.0 May 2008
1
2.5.2
sections Construct
2
Summary
3
6
The sections construct is a noniterative worksharing construct that contains a set of
structured blocks that are to be distributed among and executed by the threads in a team.
Each structured block is executed once by one of the threads in the team in the context
of its implicit task.
7
Syntax
4
5
8
C/C++
9
The syntax of the sections construct is as follows:
10
11
12
13
14
15
16
17
18
#pragma omp sections [clause[[,] clause] ...] new-line
{
[#pragma omp section new-line]
structured-block
[#pragma omp section new-line
structured-block ]
...
}
19
where clause is one of the following:
20
private(list)
21
firstprivate(list)
22
lastprivate(list)
23
reduction(operator: list)
24
nowait
25
C/C++
26
Chapter 2
Directives
47
1
Fortran
2
The syntax of the sections construct is as follows:
3
4
5
6
7
8
9
10
!$omp sections [clause[[,] clause] ...]
[!$omp section]
structured-block
[!$omp section
structured-block ]
...
!$omp end sections [nowait]
11
where clause is one of the following:
12
private(list)
13
firstprivate(list)
14
lastprivate(list)
15
reduction({operator|intrinsic_procedure_name}:list)
16
17
Fortran
18
Binding
19
22
The binding thread set for a sections region is the current team. A sections
region binds to the innermost enclosing parallel region. Only the threads of the team
executing the binding parallel region participate in the execution of the structured
blocks and (optional) implicit barrier of the sections region.
23
Description
24
Each structured block in the sections construct is preceded by a section directive
except possibly the first block, for which a preceding section directive is optional.
20
21
25
26
27
28
29
30 48
The method of scheduling the structured blocks among the threads in the team is
implementation defined.
There is an implicit barrier at the end of a sections construct unless a nowait
clause is specified.
OpenMP API • Version 3.0 May 2008
1
Restrictions
2
Restrictions to the sections construct are as follows:
3
• Orphaned section directives are prohibited; i.e., the section directives must
appear within the sections construct and may not be encountered elsewhere in the
sections region.
4
5
6
• The code enclosed in a sections construct must be a structured block.
7
8
• Only a single nowait clause can appear on a sections directive.
9
• A throw executed inside a sections region must cause execution to resume within
C/C++
the same section of the sections region, and the same thread that threw the
exception must catch it.
10
11
C/C++
12
Cross References
13
• private, firstprivate, lastprivate, and reduction clauses, see
Section 2.9.3 on page 85.
14
15
2.5.3
single Construct
16
Summary
17
20
The single construct specifies that the associated structured block is executed by only
one of the threads in the team (not necessarily the master thread), in the context of its
implicit task. The other threads in the team, which do not execute the block, wait at an
implicit barrier at the end of the single construct unless a nowait clause is specified.
21
Syntax
18
19
22
C/C++
23
The syntax of the single construct is as follows:
24
25
26
27
#pragma omp single [clause[[,] clause] ...] new-line
structured-block
Chapter 2
Directives
49
1
where clause is one of the following:
2
private(list)
3
firstprivate(list)
4
copyprivate(list)
5
nowait
6
C/C++
7
Fortran
8
The syntax of the single construct is as follows:
9
10
11
12
!$omp single [clause[[,] clause] ...]
structured-block
!$omp end single [end_clause[[,] end_clause] ...]
13
where clause is one of the following:
14
private(list)
15
firstprivate(list)
16
and end_clause is one of the following:
17
copyprivate(list)
18
nowait
19
20
Fortran
21
Binding
22
The binding thread set for a single region is the current team. A single region
binds to the innermost enclosing parallel region. Only the threads of the team
executing the binding parallel region participate in the execution of the structured
block and the (optional) implicit barrier of the single region.
23
24
25
26 50
OpenMP API • Version 3.0 May 2008
1
Description
2
4
The method of choosing a thread to execute the structured block is implementation
defined. There is an implicit barrier at the end of the single construct unless a
nowait clause is specified.
5
For an example of the single construct, see Section A.12 on page 176.
6
Restrictions
7
Restrictions to the single construct are as follows:
8
• The copyprivate clause must not be used with the nowait clause.
9
10
• At most one nowait clause can appear on a single construct.
11
• A throw executed inside a single region must cause execution to resume within the
12
same single region, and the same thread that threw the exception must catch it.
3
C/C++
C/C++
13
Cross References
14
• private and firstprivate clauses, see Section 2.9.3 on page 85.
15
• copyprivate clause, see Section 2.9.4.2 on page 102.
16
Fortran
17
2.5.4
workshare Construct
18
Summary
19
The workshare construct divides the execution of the enclosed structured block into
separate units of work, and causes the threads of the team to share the work such that
each unit is executed only once by one thread, in the context of its implicit task.
20
21
22
Chapter 2
Directives
51
1
Fortran (cont.)
2
Syntax
3
The syntax of the workshare construct is as follows:
4
5
6
7
!$omp workshare
structured-block
!$omp end workshare [nowait]
8
The enclosed structured block must consist of only the following:
9
• array assignments
10
• scalar assignments
11
• FORALL statements
12
• FORALL constructs
13
• WHERE statements
14
• WHERE constructs
15
• atomic constructs
16
• critical constructs
17
• parallel constructs
18
19
Statements contained in any enclosed critical construct are also subject to these
restrictions. Statements in any enclosed parallel construct are not restricted.
20
Binding
21
24
The binding thread set for a workshare region is the current team. A workshare
region binds to the innermost enclosing parallel region. Only the threads of the team
executing the binding parallel region participate in the execution of the units of
work and the (optional) implicit barrier of the workshare region.
25
Description
26
There is an implicit barrier at the end of a workshare construct unless a nowait
clause is specified.
22
23
27
28
29
30
31
32
33 52
An implementation of the workshare construct must insert any synchronization that is
required to maintain standard Fortran semantics. For example, the effects of one
statement within the structured block must appear to occur before the execution of
succeeding statements, and the evaluation of the right hand side of an assignment must
appear to complete prior to the effects of assigning to the left hand side.
OpenMP API • Version 3.0 May 2008
1
Fortran (cont.)
2
The statements in the workshare construct are divided into units of work as follows:
3
• For array expressions within each statement, including transformational array
4
intrinsic functions that compute scalar values from arrays:
5
• Evaluation of each element of the array expression, including any references to
ELEMENTAL functions, is a unit of work.
6
7
8
• Evaluation of transformational array intrinsic functions may be freely subdivided
into any number of units of work.
9
• For an array assignment statement, the assignment of each element is a unit of work.
10
• For a scalar assignment statement, the assignment operation is a unit of work.
11
• For a WHERE statement or construct, the evaluation of the mask expression and the
12
13
14
15
masked assignments are each a unit of work.
• For a FORALL statement or construct, the evaluation of the mask expression,
expressions occurring in the specification of the iteration space, and the masked
assignments are each a unit of work.
16
• For an atomic construct, the update of the scalar variable is a unit of work.
17
• For a critical construct, the construct is a single unit of work.
18
• For a parallel construct, the construct is a unit of work with respect to the
19
20
21
22
23
24
25
26
27
28
29
30
31
32
workshare construct. The statements contained in the parallel construct are
executed by a new thread team.
• If none of the rules above apply to a portion of a statement in the structured block,
then that portion is a unit of work.
The transformational array intrinsic functions are MATMUL, DOT_PRODUCT, SUM,
PRODUCT, MAXVAL, MINVAL, COUNT, ANY, ALL, SPREAD, PACK, UNPACK,
RESHAPE, TRANSPOSE, EOSHIFT, CSHIFT, MINLOC, and MAXLOC.
It is unspecified how the units of work are assigned to the threads executing a
workshare region.
If an array expression in the block references the value, association status, or allocation
status of private variables, the value of the expression is undefined, unless the same
value would be computed by every thread.
If an array assignment, a scalar assignment, a masked array assignment, or a FORALL
assignment assigns to a private variable in the block, the result is unspecified.
34
The workshare directive causes the sharing of work to occur only in the workshare
construct, and not in the remainder of the workshare region.
35
For examples of the workshare construct, see Section A.14 on page 191.
33
36
Chapter 2
Directives
53
1
Restrictions
2
The following restrictions apply to the workshare directive:
3
• The construct must not contain any user defined function calls unless the function is
ELEMENTAL.
4
5
Fortran
6
7
2.6
Combined Parallel Worksharing
Constructs
8
Combined parallel worksharing constructs are shortcuts for specifying a worksharing
construct nested immediately inside a parallel construct. The semantics of these
directives are identical to that of explicitly specifying a parallel construct containing
one worksharing construct and no other statements.
9
10
11
12
16
The combined parallel worksharing constructs allow certain clauses that are permitted
both on parallel constructs and on worksharing constructs. If a program would have
different behavior depending on whether the clause were applied to the parallel
construct or to the worksharing construct, then the program’s behavior is unspecified.
17
The following sections describe the combined parallel worksharing constructs:
18
• The parallel loop construct.
19
• The parallel sections construct.
20
• The parallel workshare construct.
13
14
15
21
2.6.1
Parallel Loop construct
22
Summary
23
The parallel loop construct is a shortcut for specifying a parallel construct
containing one loop construct and no other statements.
24
25 54
OpenMP API • Version 3.0 May 2008
1
Syntax
2
C/C++
3
The syntax of the parallel loop construct is as follows:
4
5
6
7
8
#pragma omp parallel for [clause[[,] clause] ...] new-line
for-loop
where clause can be any of the clauses accepted by the parallel or for directives,
except the nowait clause, with identical meanings and restrictions.
C/C++
9
Fortran
10
The syntax of the parallel loop construct is as follows:
11
12
13
14
15
16
17
18
19
!$omp parallel do [clause[[,] clause] ...]
do-loop
[!$omp end parallel do]
where clause can be any of the clauses accepted by the parallel or do directives,
with identical meanings and restrictions.
If an end parallel do directive is not specified, an end parallel do directive is
assumed at the end of the do-loop. nowait may not be specified on an end
parallel do directive.
20
21
Fortran
Description
22
C/C++
23
24
The semantics are identical to explicitly specifying a parallel directive immediately
followed by a for directive.
C/C++
25
Fortran
26
27
28
29
30
The semantics are identical to explicitly specifying a parallel directive immediately
followed by a do directive, and an end do directive immediately followed by an end
parallel directive.
Fortran
Chapter 2
Directives
55
1
Restrictions
2
The restrictions for the parallel construct and the loop construct apply.
3
Cross References
4
• parallel construct, see Section 2.4 on page 32.
5
• loop construct, see Section 2.5.1 on page 38.
6
• Data attribute clauses, see Section 2.9.3 on page 85.
7
2.6.2
parallel sections Construct
8
Summary
9
10
The parallel sections construct is a shortcut for specifying a parallel
construct containing one sections construct and no other statements.
11
Syntax
12
C/C++
13
The syntax of the parallel sections construct is as follows:
14
15
16
17
18
19
20
21
22
23
24
#pragma omp parallel sections [clause[[,] clause] ...] new-line
{
[#pragma omp section new-line]
structured-block
[#pragma omp section new-line
structured-block ]
...
}
where clause can be any of the clauses accepted by the parallel or sections
directives, except the nowait clause, with identical meanings and restrictions.
C/C++
25 56
OpenMP API • Version 3.0 May 2008
1
Fortran
2
The syntax of the parallel sections construct is as follows:
3
4
5
6
7
8
9
10
11
12
13
14
!$omp parallel sections [clause[[,] clause] ...]
[!$omp section]
structured-block
[!$omp section
structured-block ]
...
!$omp end parallel sections
where clause can be any of the clauses accepted by the parallel or sections
directives, with identical meanings and restrictions.
The last section ends at the end parallel sections directive. nowait cannot be
specified on an end parallel sections directive.
15
16
Fortran
Description
17
C/C++
18
19
The semantics are identical to explicitly specifying a parallel directive immediately
followed by a sections directive.
C/C++
20
Fortran
21
22
23
The semantics are identical to explicitly specifying a parallel directive immediately
followed by a sections directive, and an end sections directive immediately
followed by an end parallel directive.
24
Fortran
25
For an example of the parallel sections construct, see Section A.11 on page 174.
26
Restrictions
27
The restrictions for the parallel construct and the sections construct apply.
28
Cross References:
29
• parallel construct, see Section 2.4 on page 32.
30
Chapter 2
Directives
57
1
• sections construct, see Section 2.5.2 on page 47.
2
• Data attribute clauses, see Section 2.9.3 on page 85.
3
Fortran
4
2.6.3
parallel workshare Construct
5
Summary
6
7
The parallel workshare construct is a shortcut for specifying a parallel
construct containing one workshare construct and no other statements.
8
Syntax
9
The syntax of the parallel workshare construct is as follows:
10
11
12
13
!$omp parallel workshare [clause[[,] clause] ...]
structured-block
!$omp end parallel workshare
16
where clause can be any of the clauses accepted by the parallel directive, with
identical meanings and restrictions. nowait may not be specified on an end
parallel workshare directive.
17
Description
18
20
The semantics are identical to explicitly specifying a parallel directive immediately
followed by a workshare directive, and an end workshare directive immediately
followed by an end parallel directive.
21
Restrictions
22
The restrictions for the parallel construct and the workshare construct apply.
23
Cross References
24
• parallel construct, see Section 2.4 on page 32.
25
• workshare construct, see Section 2.5.4 on page 51.
14
15
19
26 58
OpenMP API • Version 3.0 May 2008
• Data attribute clauses, see Section 2.9.3 on page 85.
1
2
3
Fortran
4
5
2.7
task Construct
6
Summary
7
The task construct defines an explicit task.
8
Syntax
9
C/C++
10
The syntax of the task construct is as follows:
11
12
13
14
#pragma omp task [clause[[,] clause] ...] new-line
structured-block
where clause is one of the following:
15
if(scalar-expression)
16
untied
17
default(shared | none)
18
private(list)
19
firstprivate(list)
20
shared(list)
21
C/C++
22
Chapter 2
Directives
59
1
Fortran
2
The syntax of the task construct is as follows:
3
4
5
6
!$omp task [clause[[,] clause] ...]
structured-block
!$omp end task
7
where clause is one of the following:
8
if(scalar-logical-expression)
9
untied
10
default(private | firstprivate | shared | none)
11
private(list)
12
firstprivate(list)
13
shared(list)
14
15
Fortran
16
Binding
17
18
The binding thread set of the task region is the current parallel team. A task region
binds to the innermost enclosing parallel region.
19
Description
20
When a thread encounters a task construct, a task is generated from the code for the
associated structured block. The data environment of the task is created according to the
data-sharing attribute clauses on the task construct and any defaults that apply.
21
22
23
24
25
26
27
28
29
30
31 60
The encountering thread may immediately execute the task, or defer its execution. In the
latter case, any thread in the team may be assigned the task. Completion of the task can
be guaranteed using task synchronization constructs. A task construct may be nested
inside an outer task, but the task region of the inner task is not a part of the task
region of the outer task.
When an if clause is present on a task construct and the if clause expression
evaluates to false, the encountering thread must suspend the current task region and
begin execution of the generated task immediately, and the suspended task region may
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
9
10
11
12
13
not be resumed until the generated task is completed. The task still behaves as a distinct
task region with respect to data environment, lock ownership, and synchronization
constructs. Note that the use of a variable in an if clause expression of a task
construct causes an implicit reference to the variable in all enclosing constructs.
A thread that encounters a task scheduling point within the task region may
temporarily suspend the task region. By default, a task is tied and its suspended task
region can only be resumed by the thread that started its execution. If the untied
clause is present on a task construct, any thread in the team can resume the task
region after a suspension.
The task construct includes a task scheduling point in the task region of its generating
task, immediately following the generation of the explicit task. Each explicit task
region includes a task scheduling point at its point of completion. An implementation
may add task scheduling points anywhere in untied task regions.
14
15
16
17
Note – When storage is shared by an explicit task region, it is the programmer's
responsibility to ensure, by adding proper synchronization, that the storage does not
reach the end of its lifetime before the explicit task region completes its execution.
18
19
Restrictions
20
Restrictions to the task construct are as follows:
21
• A program that branches into or out of a task region is non-conforming.
22
• A program must not depend on any ordering of the evaluations of the clauses of the
23
task directive, or on any side effects of the evaluations of the clauses.
24
25
• At most one if clause can appear on the directive.
26
• A throw executed inside a task region must cause execution to resume within the
C/C++
27
same task region, and the same thread that threw the exception must catch it.
C/C++
28
Fortran
29
30
31
32
• Unsynchronized use of Fortran I/O statements by multiple tasks on the same unit has
unspecified behavior.
Fortran
Chapter 2
Directives
61
1
2.7.1
Task Scheduling
4
Whenever a thread reaches a task scheduling point, the implementation may cause it to
perform a task switch, beginning or resuming execution of a different task bound to the
current team. Task scheduling points are implied at the following locations:
5
• the point immediately following the generation of an explicit task
6
• after the last instruction of a task region
7
• in taskwait regions
8
• in implicit and explicit barrier regions.
9
In addition, implementations may insert task scheduling points in untied tasks anywhere
that they are not specifically prohibited in this specification.
2
3
10
12
When a thread encounters a task scheduling point it may do one of the following,
subject to the Task Scheduling Constraints (below):
13
• begin execution of a tied task bound to the current team.
14
• resume any suspended task region, bound to the current team, to which it is tied.
15
• begin execution of an untied task bound to the current team.
16
• resume any suspended untied task region bound to the current team.
17
18
If more than one of the above choices is available, it is unspecified as to which will be
chosen.
19
Task Scheduling Constraints
20
1. An explicit task whose construct contained an if clause whose if clause expression
evaluated to false is executed immediately after generation of the task.
11
21
25
2. Other scheduling of new tied tasks is constrained by the set of task regions that are
currently tied to the thread, and that are not suspended in a barrier region. If this set
is empty, any new tied task may be scheduled. Otherwise, a new tied task may be
scheduled only if it is a descendant of every task in the set.
26
A program relying on any other assumption about task scheduling is non-conforming.
22
23
24
27
28
29
30
31
32
33
34
35 62
Note – Task scheduling points dynamically divide task regions into parts. Each part is
executed uninterruptedly from start to end. Different parts of the same task region are
executed in the order in which they are encountered. In the absence of task
synchronization constructs, the order in which a thread executes parts of different
schedulable tasks is unspecified.
A correct program must behave correctly and consistently with all conceivable
scheduling sequences that are compatible with the rules above.
OpenMP API • Version 3.0 May 2008
For example, if threadprivate storage is accessed (explicitly in the source code or
implicitly in calls to library routines) in one part of a task region, its value cannot be
assumed to be preserved into the next part of the same task region if another schedulable
task exists that modifies it (see Example A.13.7c on page 186, Example A.13.7f on page
186, Example A.13.8c on page 187 and Example A.13.8f on page 187).
1
2
3
4
5
As another example, if a lock acquire and release happen in different parts of a task
region, no attempt should be made to acquire the same lock in any part of another task
that the executing thread may schedule; otherwise, a deadlock is possible. A similar
situation can occur when a critical region spans multiple parts of a task and another
schedulable task contains a critical region with the same name (see Example A.13.9c on
page 188, Example A.13.9f on page 189, Example A.13.10c on page 190 and Example
A.13.10f on page 191).
6
7
8
9
10
11
12
The use of threadprivate variables and the use of locks or critical sections in an explicit
task with an if clause must take into account that when the if clause evaluates to
false, the task is executed immediately, without regard to Task Scheduling Constraint 2.
13
14
15
16
17
18
2.8
Master and Synchronization Constructs
19
The following sections describe :
20
• the master construct.
21
• the critical construct.
22
• the barrier construct.
23
• the taskwait construct.
24
• the atomic construct.
25
• the flush construct.
26
• the ordered construct.
27
2.8.1
master Construct
28
Summary
29
The master construct specifies a structured block that is executed by the master thread
of the team.
30
31
Chapter 2
Directives
63
1
Syntax
2
C/C++
3
The syntax of the master construct is as follows:
4
5
6
#pragma omp master new-line
structured-block
7
C/C++
8
Fortran
9
The syntax of the master construct is as follows:
10
11
12
13
!$omp master
structured-block
!$omp end master
14
15
Fortran
16
Binding
17
20
The binding thread set for a master region is the current team. A master region
binds to the innermost enclosing parallel region. Only the master thread of the team
executing the binding parallel region participates in the execution of the structured
block of the master region.
21
Description
22
23
Other threads in the team do not execute the associated structured block. There is no
implied barrier either on entry to, or exit from, the master construct.
24
For an example of the master construct, see Section A.15 on page 195.
25
26
Restrictions
27
• A throw executed inside a master region must cause execution to resume within the
28
same master region, and the same thread that threw the exception must catch it.
18
19
C/C++
C/C++
29 64
OpenMP API • Version 3.0 May 2008
1
2.8.2
critical Construct
2
Summary
3
4
The critical construct restricts execution of the associated structured block to a
single thread at a time.
5
Syntax
6
C/C++
7
The syntax of the critical construct is as follows:
8
9
10
#pragma omp critical [(name)] new-line
structured-block
11
C/C++
12
Fortran
13
The syntax of the critical construct is as follows:
14
15
16
17
!$omp critical [(name)]
structured-block
!$omp end critical [(name)]
18
19
Fortran
20
Binding
21
23
The binding thread set for a critical region is all threads. Region execution is
restricted to a single thread at a time among all the threads in the program, without
regard to the team(s) to which the threads belong.
24
Description
25
An optional name may be used to identify the critical construct. All critical
constructs without a name are considered to have the same unspecified name. A thread
waits at the beginning of a critical region until no thread is executing a critical
22
26
27
28
Chapter 2
Directives
65
region with the same name. The critical construct enforces exclusive access with
respect to all critical constructs with the same name in all threads, not just those
threads in the current team.
1
2
3
4
C/C++
Identifiers used to identify a critical construct have external linkage and are in a
name space that is separate from the name spaces used by labels, tags, members, and
ordinary identifiers.
5
6
7
C/C++
8
Fortran
The names of critical constructs are global entities of the program. If a name
conflicts with any other entity, the behavior of the program is unspecified.
9
10
11
Fortran
12
For an example of the critical construct, see Section A.16 on page 197.
13
14
Restrictions
15
• A throw executed inside a critical region must cause execution to resume within
16
the same critical region, and the same thread that threw the exception must catch
it.
C/C++
17
C/C++
18
Fortran
19
The following restrictions apply to the critical construct:
20
• If a name is specified on a critical directive, the same name must also be
specified on the end critical directive.
21
• If no name appears on the critical directive, no name can appear on the end
22
critical directive.
23
24
25
Fortran
2.8.3
barrier Construct
26
Summary
27
The barrier construct specifies an explicit barrier at the point at which the construct
appears.
28
29 66
OpenMP API • Version 3.0 May 2008
1
Syntax
2
C/C++
3
The syntax of the barrier construct is as follows:
4
5
6
7
8
9
10
11
#pragma omp barrier new-line
Because the barrier construct does not have a C language statement as part of its
syntax, there are some restrictions on its placement within a program. The barrier
directive may be placed only at a point where a base language statement is allowed. The
barrier directive may not be used in place of the statement following an if, while, do,
switch, or label. See Appendix C for the formal grammar. The examples in Section A.23
on page 214 illustrate these restrictions.
C/C++
12
Fortran
13
The syntax of the barrier construct is as follows:
14
15
!$omp barrier
16
17
Fortran
18
Binding
19
21
The binding thread set for a barrier region is the current team. A barrier region
binds to the innermost enclosing parallel region. See Section A.18 on page 200 for
examples.
22
Description
23
All threads of the team executing the binding parallel region must execute the
barrier region and complete execution of all explicit tasks generated in the binding
parallel region up to this point before any are allowed to continue execution beyond
the barrier.
20
24
25
26
27
28
29
The barrier region includes an implicit task scheduling point in the current task
region.
Chapter 2
Directives
67
1
Restrictions
2
The following restrictions apply to the barrier construct:
3
• Each barrier region must be encountered by all threads in a team or by none at all.
4
• The sequence of worksharing regions and barrier regions encountered must be the
same for every thread in a team.
5
6
2.8.4
taskwait Construct
7
Summary
8
9
The taskwait construct specifies a wait on the completion of child tasks generated
since the beginning of the current task.
10
Syntax
11
C/C++
12
The syntax of the taskwait construct is as follows:
13
14
15
16
17
18
19
20
#pragma omp taskwait newline
Because the taskwait construct does not have a C language statement as part of its
syntax, there are some restrictions on its placement within a program. The taskwait
directive may be placed only at a point where a base language statement is allowed. The
taskwait directive may not be used in place of the statement following an if, while,
do, switch, or label. See Appendix C for the formal grammar. The examples in
Section A.23 on page 214 illustrate these restrictions.
C/C++
21
Fortran
22
The syntax of the taskwait construct is as follows:
23
24
!$omp taskwait
25
26
27 68
Fortran
OpenMP API • Version 3.0 May 2008
1
Binding
2
3
A taskwait region binds to the current task region. The binding thread set of the
taskwait region is the current team.
4
Description
5
The taskwait region includes an implicit task scheduling point in the current task
region. The current task region is suspended at the task scheduling point until execution
of all its child tasks generated before the taskwait region are completed.
6
7
8
2.8.5
atomic Construct
9
Summary
10
11
The atomic construct ensures that a specific storage location is updated atomically,
rather than exposing it to the possibility of multiple, simultaneous writing threads.
12
Syntax
13
C/C++
14
The syntax of the atomic construct is as follows:
15
16
17
18
#pragma omp atomic new-line
expression-stmt
where expression-stmt is an expression statement with one of the following forms:
19
20
x binop= expr
21
x++
22
++x
23
x--
24
--x
25
In the preceding expressions:
26
• x is an lvalue expression with scalar type.
27
Chapter 2
Directives
69
1
2
• expr is an expression with scalar type, and it does not reference the variable
designated by x.
3
• binop is one of +, *, -, /, &, ^, |, <<, or >>.
4
• binop, binop=, ++, and -- are not overloaded operators.
C/C++
5
Fortran
6
The syntax of the atomic construct is as follows:
7
8
9
!$omp atomic
statement
10
where statement has one of the following forms:
11
x = x operator expr
12
x = expr operator x
13
x = intrinsic_procedure_name (x, expr_list)
14
x = intrinsic_procedure_name (expr_list, x)
15
In the preceding statements:
16
• x is a scalar variable of intrinsic type.
17
• expr is a scalar expression that does not reference x.
18
• expr_list is a comma-separated, non-empty list of scalar expressions that do not
19
20
reference x. When intrinsic_procedure_name refers to IAND, IOR, or IEOR, exactly
one expression must appear in expr_list.
21
• intrinsic_procedure_name is one of MAX, MIN, IAND, IOR, or IEOR.
22
• operator is one of +, *, -, /, .AND., .OR., .EQV., or .NEQV. .
23
• The operators in expr must have precedence equal to or greater than the precedence
25
of operator, x operator expr must be mathematically equivalent to x operator (expr),
and expr operator x must be mathematically equivalent to (expr) operator x.
26
• intrinsic_procedure_name must refer to the intrinsic procedure name and not to other
24
27
program entities.
28
• operator must refer to the intrinsic operator and not to a user-defined operator.
29
• The assignment must be intrinsic assignment.
30
31 70
Fortran
OpenMP API • Version 3.0 May 2008
1
Binding
2
5
The binding thread set for an atomic region is all threads. atomic regions enforce
exclusive access with respect to other atomic regions that update the same storage
location x among all the threads in the program without regard to the teams to which the
threads belong.
6
Description
7
Only the load and store of the variable designated by x are atomic; the evaluation of
expr is not atomic. No task scheduling points are allowed between the load and the store
of the variable designated by x. To avoid race conditions, all updates of the location that
could potentially occur in parallel must be protected with an atomic directive.
atomic regions do not enforce exclusive access with respect to any critical or
ordered regions that access the same storage location x. However, other OpenMP
synchronization can ensure the desired exclusive access. For example, a barrier
following a series of atomic updates to x guarantees that subsequent accesses do not
form a race with the atomic accesses.
3
4
8
9
10
11
12
13
14
15
18
A compliant implementation may enforce exclusive access between atomic regions
that update different storage locations. The circumstances under which this occurs are
implementation defined.
19
For an example of the atomic construct, see Section A.19 on page 202.
20
Restrictions
16
17
21
C/C++
22
The following restriction applies to the atomic construct:
23
• All atomic references to the storage location x throughout the program are required to
24
have a compatible type. See Section A.20 on page 205 for examples.
C/C++
25
Fortran
26
The following restriction applies to the atomic construct:
27
• All atomic references to the storage location of variable x throughout the program are
28
29
30
31
required to have the same type and type parameters. See Section A.20 on page 205
for examples.
Fortran
Chapter 2
Directives
71
1
Cross References
2
• critical construct, see Section 2.8.2 on page 65.
3
2.8.6
flush Construct
4
Summary
5
8
The flush construct executes the OpenMP flush operation. This operation makes a
thread’s temporary view of memory consistent with memory, and enforces an order on
the memory operations of the variables explicitly specified or implied. See the memory
model description in Section 1.4 on page 13 for more details.
9
Syntax
6
7
10
C/C++
11
The syntax of the flush construct is as follows:
12
13
14
15
16
17
18
19
#pragma omp flush [(list)] new-line
Note that because the flush construct does not have a C language statement as part of
its syntax, there are some restrictions on its placement within a program. The
taskwait directive may be placed only at a point where a base language statement is
allowed. The taskwait directive may not be used in place of the statement following
an if, while, do, switch, or label. See Appendix C for the formal grammar. See
Section A.23 on page 214 for an example that illustrates these placement restrictions.
C/C++
20
Fortran
21
The syntax of the flush construct is as follows:
22
23
!$omp flush [(list)]
24
25
26 72
Fortran
OpenMP API • Version 3.0 May 2008
1
Binding
2
6
The binding thread set for a flush region is the encountering thread. Execution of a
flush region affects the memory and the temporary view of memory of only the thread
that executes the region. It does not affect the temporary view of other threads. Other
threads must themselves execute a flush operation in order to be guaranteed to observe
the effects of the encountering thread’s flush operation.
7
Description
8
A flush construct with a list applies the flush operation to the items in the list, and
does not return until the operation is complete for all specified list items. A flush
construct without a list, executed on a given thread, operates as if the whole threadvisible data state of the program, as defined by the base language, is flushed.
3
4
5
9
10
11
12
C/C++
13
14
If a pointer is present in the list, the pointer itself is flushed, not the memory block to
which the pointer refers.
C/C++
15
Fortran
16
17
18
19
20
If the list item or a subobject of the list item has the POINTER attribute, the allocation
or association status of the POINTER item is flushed, but the pointer target is not. If the
list item is a Cray pointer, the pointer is flushed, but the object to which it points is not.
If the list item has the ALLOCATABLE attribute and the list item is allocated, the
allocated array is flushed; otherwise the allocation status is flushed.
21
Fortran
22
For examples of the flush construct, see Section A.21 on page 208 and Section A.22
on page 211.
23
24
Chapter 2
Directives
73
1
2
3
4
5
6
Note – the following examples illustrate the ordering properties of the flush operation.
In the following incorrect pseudocode example, the programmer intends to prevent
simultaneous execution of the critical section by the two threads, but the program does
not work properly because it does not enforce the proper ordering of the operations on
variables a and b.
7
8
9
10
11
12
13
14
15
16
17
Incorrect example:
a = b = 0
thread 1
thread 2
b = 1
flush(b)
flush(a)
if (a == 0) then
critical section
end if
a = 1
flush(a)
flush(b)
if (b == 0) then
critical section
end if
18
19
20
21
22
23
24
25
26
27
28
The problem with this example is that operations on variables a and b are not ordered
with respect to each other. For instance, nothing prevents the compiler from moving the
flush of b on thread 1 or the flush of a on thread 2 to a position completely after the
critical section (assuming that the critical section on thread 1 does not reference b and the
critical section on thread 2 does not reference a). If either re-ordering happens, both
threads can simultaneously executed the critical section.
The following correct pseudocode example correctly ensures that the critical section is
executed by not more than one of the two threads at any one time. Notice that execution of
the critical section by neither thread is considered correct in this example. This occurs if
both flushes complete prior to either thread executing its if statement.
29
30
31
32
33
34
35
36
37
38
Correct example:
a = b = 0
thread 1
b = 1
flush(a,b)
if (a == 0) then
critical section
end if
39
40 74
OpenMP API • Version 3.0 May 2008
thread 2
a = 1
flush(a,b)
if (b == 0) then
critical section
end if
The compiler is prohibited from moving the flush at all for either thread, ensuring that the
respective assignment is complete and the data is flushed before the if statement is
executed.
1
2
3
4
5
A flush region without a list is implied at the following locations:
6
• During a barrier region.
7
• At entry to and exit from parallel, critical, and ordered regions.
8
• At exit from worksharing regions unless a nowait is present.
9
• At entry to and exit from combined parallel worksharing regions.
10
• During omp_set_lock and omp_unset_lock regions.
11
• During omp_test_lock, omp_set_nest_lock, omp_unset_nest_lock
12
and omp_test_nest_lock regions, if the region causes the lock to be set or
unset.
13
14
• Immediately before and immediately after every task scheduling point.
15
A flush region with a list is implied at the following locations:
16
• At entry to and exit from atomic regions, where the list contains only the variable
updated in the atomic construct.
17
18
19
Note – A flush region is not implied at the following locations:
20
• At entry to worksharing regions.
21
• At entry to or exit from a master region.
22
23
2.8.7
ordered Construct
24
Summary
25
The ordered construct specifies a structured block in a loop region that will be
executed in the order of the loop iterations. This sequentializes and orders the code
within an ordered region while allowing code outside the region to run in parallel.
26
27
28
Chapter 2
Directives
75
1
Syntax
2
C/C++
3
The syntax of the ordered construct is as follows:
4
5
6
#pragma omp ordered new-line
structured-block
7
C/C++
Fortran
8
The syntax of the ordered construct is as follows:
9
10
11
12
!$omp ordered
structured-block
!$omp end ordered
13
14
Fortran
15
Binding
16
18
The binding thread set for an ordered region is the current team. An ordered region
binds to the innermost enclosing loop region. ordered regions that bind to different
loop regions execute independently of each other.
19
Description
20
25
The threads in the team executing the loop region execute ordered regions
sequentially in the order of the loop iterations. When the thread executing the first
iteration of the loop encounters an ordered construct, it can enter the ordered
region without waiting. When a thread executing any subsequent iteration encounters an
ordered region, it waits at the beginning of that ordered region until execution of
all the ordered regions belonging to all previous iterations have completed.
26
For examples of the ordered construct, see Section A.24 on page 215.
27
Restrictions
28
Restrictions to the ordered construct are as follows:
29
• The loop region to which an ordered region binds must have an ordered clause
17
21
22
23
24
30
31 76
specified on the corresponding loop (or parallel loop) construct.
OpenMP API • Version 3.0 May 2008
• During execution of an iteration of a loop or a loop nest within a loop region, a
1
thread must not execute more than one ordered region that binds to the same loop
region.
2
3
4
C/C++
• A throw executed inside a ordered region must cause execution to resume within
5
the same ordered region, and the same thread that threw the exception must catch
it.
6
7
C/C++
8
Cross References
9
• loop construct, see Section 2.5.1 on page 38.
10
• parallel loop construct, see Section 2.6.1 on page 54.
11
12
2.9
Data Environment
14
This section presents a directive and several clauses for controlling the data environment
during the execution of parallel, task, and worksharing regions.
15
• Section 2.9.1 on page 77 describes how the data-sharing attributes of variables
16
referenced in parallel, task, and worksharing regions are determined.
13
• The threadprivate directive, which is provided to create threadprivate memory,
17
is described in Section 2.9.2 on page 81.
18
19
• Clauses that may be specified on directives to control the data-sharing attributes of
20
variables referenced in parallel, task, or worksharing constructs are described
in Section 2.9.3 on page 85.
21
• Clauses that may be specified on directives to copy data values from private or
22
threadprivate variables on one thread to the corresponding variables on other threads
in the team are described in Section 2.9.4 on page 100.
23
24
25
2.9.1
Data-sharing Attribute Rules
28
This section describes how the data-sharing attributes of variables referenced in
parallel, task, and worksharing regions are determined. The following two cases
are described separately:
29
• Section 2.9.1.1 on page 78 describes the data-sharing attribute rules for variables
26
27
30
31
referenced in a construct.
Chapter 2
Directives
77
• Section 2.9.1.2 on page 80 describes the data-sharing attribute rules for variables
1
referenced in a region, but outside any construct.
2
3
2.9.1.1
4
5
6
7
8
9
Data-sharing Attribute Rules for Variables Referenced in a
Construct
The data-sharing attributes of variables that are referenced in a construct may be one of
the following: predetermined, explicitly determined, or implicitly determined.
Specifying a variable on a firstprivate, lastprivate, or reduction clause
of an enclosed construct causes an implicit reference to the variable in the enclosing
construct. Such implicit references are also subject to the following rules.
10
11
The following variables have predetermined data-sharing attributes:
12
• Variables appearing in threadprivate directives are threadprivate.
13
• Variables with automatic storage duration that are declared in a scope inside the
C/C++
14
construct are private.
15
• Variables with heap allocated storage are shared.
16
• Static data members are shared.
17
• The loop iteration variable(s) in the associated for-loop(s) of a for or parallel
18
for construct is(are) private.
19
• Variables with const-qualified type having no mutable member are shared.
20
• Static variables which are declared in a scope inside the construct are shared.
C/C++
21
Fortran
22
23
24
25
26
27
• Variables and common blocks appearing in threadprivate directives are
threadprivate.
• The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do
construct is(are) private.
• A loop iteration variable for a sequential loop in a parallel or task construct is
private in the innermost such construct that encloses the loop.
28
• implied-do and forall indices are private.
29
• Cray pointees inherit the data-sharing attribute of the storage with which their Cray
30
pointers are associated.
31
32 78
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
Variables with predetermined data-sharing attributes may not be listed in data-sharing
attribute clauses, except for the cases listed below. For these exceptions only, listing a
predetermined variable in a data-sharing attribute clause is allowed and overrides the
variable’s predetermined data-sharing attributes.
C/C++
6
7
• The loop iteration variable(s) in the associated for-loop(s) of a for or parallel
for construct may be listed in a private or lastprivate clause.
C/C++
8
Fortran
9
10
11
12
13
14
15
• The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do
construct may be listed in a private or lastprivate clause.
• Variables used as loop iteration variables in sequential loops in a parallel
construct or a task construct may be listed in private, firstprivate,
lastprivate, shared, or reduction clauses on the parallel or task
construct, and on enclosed constructs, subject to other restrictions.
• Assumed-size arrays may be listed in a shared clause.
16
17
18
19
20
Fortran
Additional restrictions on the variables which may appear in individual clauses are
described with each clause in Section 2.9.3 on page 85.
Variables with explicitly determined data-sharing attributes are those that are referenced
in a given construct and are listed in a data-sharing attribute clause on the construct.
23
Variables with implicitly determined data-sharing attributes are those that are referenced
in a given construct, do not have predetermined data-sharing attributes, and are not
listed in a data-sharing attribute clause on the construct.
24
Rules for variables with implicitly determined data-sharing attributes are as follows:
25
• In a parallel or task construct, the data-sharing attributes of these variables are
21
22
26
determined by the default clause, if present (see Section 2.9.3.1 on page 86).
27
• In a parallel construct, if no default clause is present, these variables are
28
29
30
shared.
• For constructs other than task, if no default clause is present, these variables
inherit their data-sharing attributes from the enclosing context.
31
• In a task construct, if no default clause is present, a variable that is determined
32
33
to be shared in all enclosing constructs, up to and including the innermost enclosing
parallel construct, is shared.
34
• In a task construct, if no default clause is present, a variable whose data-sharing
35
36
attribute is not determined by the rule above is firstprivate.
Chapter 2
Directives
79
Additional restrictions on the variables whose data-sharing attributes cannot be
implicitly determined in a task construct are described in the Restrictions section of
the firstprivate clause (Section 2.9.3.4 on page 92).
1
2
3
4
2.9.1.2
5
6
7
8
Data-sharing Attribute Rules for Variables Referenced in a
Region but not in a Construct
The data-sharing attributes of variables that are referenced in a region, but not in a
construct, are determined as follows:
C/C++
9
• Static variables declared in called routines in the region are shared.
10
• Variables with const-qualified type having no mutable member, and that are declared
11
12
13
in called routines, are shared.
• File-scope or namespace-scope variables referenced in called routines in the region
are shared unless they appear in a threadprivate directive.
14
• Variables with heap-allocated storage are shared.
15
• Static data members are shared unless they appear in a threadprivate directive.
16
• Formal arguments of called routines in the region that are passed by reference inherit
17
18
the data-sharing attributes of the associated actual argument.
• Other variables declared in called routines in the region are private.
C/C++
19
Fortran
20
21
22
23
• Local variables declared in called routines in the region and that have the save
attribute, or that are data initialized, are shared unless they appear in a
threadprivate directive.
• Variables belonging to common blocks, or declared in modules, and referenced in
25
called routines in the region are shared unless they appear in a threadprivate
directive.
26
• Dummy arguments of called routines in the region that are passed by reference inherit
24
27
28
29
30
31
the data-sharing attributes of the associated actual argument.
• Cray pointees inherit the data-sharing attribute of the storage with which their Cray
pointers are associated.
• implied-do indices, forall indices, and other local variables declared in called
routines in the region are private.
32
33 80
Fortran
OpenMP API • Version 3.0 May 2008
1
2.9.2
threadprivate Directive
2
Summary
3
4
The threadprivate directive specifies that variables are replicated, with each thread
having its own copy.
5
Syntax
6
C/C++
7
The syntax of the threadprivate directive is as follows:
8
9
10
11
#pragma omp threadprivate(list) new-line
where list is a comma-separated list of file-scope, namespace-scope, or static blockscope variables that do not have incomplete types.
C/C++
12
Fortran
13
The syntax of the threadprivate directive is as follows:
14
15
16
17
!$omp threadprivate(list)
where list is a comma-separated list of named variables and named common blocks.
Common block names must appear between slashes.
18
Fortran
19
Description
20
Each copy of a threadprivate variable is initialized once, in the manner specified by the
program, but at an unspecified point in the program prior to the first reference to that
copy. The storage of all copies of a threadprivate variable is freed according to how
static variables are handled in the base language, but at an unspecified point in the
program.
21
22
23
24
25
26
27
A program in which a thread references another thread’s copy of a threadprivate variable
is non-conforming.
Chapter 2
Directives
81
1
2
3
4
5
6
7
8
The content of a threadprivate variable can change across a task scheduling point if the
executing thread switches to another schedulable task that modifies the variable. For
more details on task scheduling, see Section 1.3 on page 11 and Section 2.7 on page 59.
In parallel regions, references by the master thread will be to the copy of the
variable in the thread which encountered the parallel region.
During the sequential part references will be to the initial thread’s copy of the variable.
The values of data in the initial thread’s copy of a threadprivate variable are guaranteed
to persist between any two consecutive references to the variable in the program.
11
The values of data in the threadprivate variables of non-initial threads are guaranteed to
persist between two consecutive active parallel regions only if all the following
conditions hold:
12
• Neither parallel region is nested inside another explicit parallel region.
13
• The number of threads used to execute both parallel regions is the same.
14
• The value of the dyn-var internal control variable in the enclosing task region is false
9
10
15
16
17
18
at entry to both parallel regions.
If these conditions all hold, and if a threadprivate variable is referenced in both regions,
then threads with the same thread number in their respective regions will reference the
same copy of that variable.
19
C/C++
20
21
22
23
24
25
26
If the above conditions hold, the storage duration, lifetime, and value of a thread’s copy
of a threadprivate variable that does not appear in any copyin clause on the second
region will be retained. Otherwise, the storage duration, lifetime, and value of a thread’s
copy of the variable in the second region is unspecified.
If the value of a variable referenced in an explicit initializer of a threadprivate variable
is modified prior to the first reference to any instance of the threadprivate variable, then
the behavior is unspecified.
27
28
29
30
Note – The order in which any constructors for different threadprivate variables of class
type are called is unspecified. The order in which any destructors for different
threadprivate variables of class type are called is unspecified.
31
32
C/C++
33
Fortran
34
35
36 82
A variable is affected by a copyin clause if the variable appears in the copyin clause
or it is in a common block that appears in the copyin clause.
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
If the above conditions hold, the definition, association, or allocation status of a thread’s
copy of a threadprivate variable or a variable in a threadprivate common block, that is
not affected by any copyin clause that appears on the second region, will be retained.
Otherwise, the definition and association status of a thread’s copy of the variable in the
second region is undefined, and the allocation status of an allocatable array will be
implementation defined.
If a common block, or a variable that is declared in the scope of a module, appears in a
threadprivate directive, it implicitly has the SAVE attribute.
12
If a threadprivate variable or a variable in a threadprivate common block is not affected
by any copyin clause that appears on the first parallel region in which it is
referenced, the variable or any subobject of the variable is initially defined or undefined
according to the following rules:
13
• If it has the ALLOCATABLE attribute, each copy created will have an initial
9
10
11
14
15
allocation status of not currently allocated.
• If it has the POINTER attribute:
18
• if it has an initial association status of disassociated, either through explicit
initialization or default initialization, each copy created will have an association
status of disassociated;
19
• otherwise, each copy created will have an association status of undefined.
16
17
20
• If it does not have either the POINTER or the ALLOCATABLE attribute:
21
22
• if it is initially defined, either through explicit initialization or default
initialization, each copy created is so defined;
23
• otherwise, each copy created is undefined.
24
Fortran
25
For examples of the threadprivate directive, see Section A.25 on page 220.
26
Restrictions
27
The restrictions to the threadprivate directive are as follows:
28
• A threadprivate variable must not appear in any clause except the copyin,
29
copyprivate, schedule, num_threads, and if clauses.
30
31
• A program in which an untied task accesses threadprivate storage is non-conforming.
32
• A variable which is part of another variable (as an array or structure element) cannot
C/C++
33
34
35
appear in a threadprivate clause unless it is a static data member of a C++
class.
Chapter 2
Directives
83
1
• A threadprivate directive for file-scope variables must appear outside any
2
definition or declaration, and must lexically precede all references to any of the
variables in its list.
3
4
• A threadprivate directive for static class member variables must appear in the
5
class definition, in the same scope in which the member variables are declared, and
must lexically precede all references to any of the variables in its list.
6
7
• A threadprivate directive for namespace-scope variables must appear outside
8
9
any definition or declaration other than the namespace definition itself, and must
lexically precede all references to any of the variables in its list.
10
• Each variable in the list of a threadprivate directive at file, namespace, or class
12
scope must refer to a variable declaration at file, namespace, or class scope that
lexically precedes the directive.
13
• A threadprivate directive for static block-scope variables must appear in the
11
15
scope of the variable and not in a nested scope. The directive must lexically precede
all references to any of the variables in its list.
16
• Each variable in the list of a threadprivate directive in block scope must refer to
17
18
a variable declaration in the same scope that lexically precedes the directive. The
variable declaration must use the static storage-class specifier.
19
• If a variable is specified in a threadprivate directive in one translation unit, it
14
20
21
must be specified in a threadprivate directive in every translation unit in which
it is declared.
22
• The address of a threadprivate variable is not an address constant.
23
• A threadprivate variable must not have an incomplete type or a reference type.
24
• A threadprivate variable with class type must have:
25
26
27
28
29
30
• an accessible, unambiguous default constructor in case of default initialization
without a given initializer;
• an accessible, unambiguous constructor accepting the given argument in case of
direct initialization;
• an accessible, unambiguous copy constructor in case of copy initialization with an
explicit initializer.
C/C++
31
Fortran
32
33
• A variable which is part of another variable (as an array or structure element) cannot
appear in a threadprivate clause.
34
• The threadprivate directive must appear in the declaration section of a scoping
35
unit in which the common block or variable is declared. Although variables in
common blocks can be accessed by use association or host association, common
36
37 84
OpenMP API • Version 3.0 May 2008
block names cannot. This means that a common block name specified in a
threadprivate directive must be declared to be a common block in the same
scoping unit in which the threadprivate directive appears.
1
2
3
• If a threadprivate directive specifying a common block name appears in one
4
program unit, then such a directive must also appear in every other program unit that
contains a COMMON statement specifying the same name. It must appear after the last
such COMMON statement in the program unit.
5
6
7
8
• A blank common block cannot appear in a threadprivate directive.
9
• A variable can only appear in a threadprivate directive in the scope in which it
is declared. It must not be an element of a common block or appear in an
EQUIVALENCE statement.
10
11
• A variable that appears in a threadprivate directive and is not declared in the
12
scope of a module must have the SAVE attribute.
13
14
Fortran
15
Cross References:
16
• dyn-var ICV, see Section 2.3 on page 28.
17
• number of threads used to execute a parallel region, see Section 2.4.1 on page 35.
18
• copyin clause, see Section 2.9.4.1 on page 101.
19
20
21
22
23
24
25
26
27
28
29
30
31
2.9.3
Data-Sharing Attribute Clauses
Several constructs accept clauses that allow a user to control the data-sharing attributes
of variables referenced in the construct. Data-sharing attribute clauses apply only to
variables whose names are visible in the construct on which the clause appears.
Not all of the clauses listed in this section are valid on all directives. The set of clauses
that is valid on a particular directive is described with the directive.
Most of the clauses accept a comma-separated list of list items (see Section 2.1 on page
22). All list items appearing in a clause must be visible, according to the scoping rules
of the base language. With the exception of the default clause, clauses may be
repeated as needed. A list item that specifies a given variable may not appear in more
than one clause on the same directive, except that a variable may be specified in both
firstprivate and lastprivate clauses.
Chapter 2
Directives
85
1
C/C++
If a variable referenced in a data-sharing attribute clause has a type derived from a
template, and there are no other references to that variable in the program, then any
behavior related to that variable is unspecified.
2
3
4
C/C++
5
Fortran
A named common block may be specified in a list by enclosing the name in slashes.
When a named common block appears in a list, it has the same meaning as if every
explicit member of the common block appeared in the list. An explicit member of a
common block is a variable that is named in a COMMON statement that specifies the
common block name and is declared in the same scoping unit in which the clause
appears.
6
7
8
9
10
11
Although variables in common blocks can be accessed by use association or host
association, common block names cannot. This means that a common block name
specified in a data-sharing attribute clause must be declared to be a common block in the
same scoping unit in which the data-sharing attribute clause appears.
12
13
14
15
When a named common block appears in a private, firstprivate,
lastprivate, or shared clause of a directive, none of its members may be declared
in another data-sharing attribute clause in that directive (see Section A.27 on page 227
for examples). When individual members of a common block appear in a private,
firstprivate, lastprivate, or reduction clause of a directive, the storage of
the specified variables is no longer associated with the storage of the common block
itself (see Section A.32 on page 237 for examples).
16
17
18
19
20
21
22
23
24
Fortran
2.9.3.1
default clause
25
Summary
26
The default clause allows the user to control the data-sharing attributes of variables that
are referenced in a parallel or task construct, and whose data-sharing attributes are
implicitly determined (see Section 2.9.1.1 on page 78).
27
28
29 86
OpenMP API • Version 3.0 May 2008
1
Syntax
2
C/C++
3
The syntax of the default clause is as follows:
4
5
default(shared | none)
6
C/C++
7
Fortran
8
The syntax of the default clause is as follows:
9
10
default(private | firstprivate | shared | none)
11
12
Fortran
13
Description
14
The default(shared) clause causes all variables referenced in the construct that
have implicitly determined data-sharing attributes to be shared.
15
16
Fortran
17
18
19
20
The default(firstprivate) clause causes all variables in the construct that have
implicitly determined data-sharing attributes to be firstprivate.
The default(private) clause causes all variables referenced in the construct that
have implicitly determined data-sharing attributes to be private.
21
Fortran
22
25
The default(none) clause requires that each variable that is referenced in the
construct, and that does not have a predetermined data-sharing attribute, must have its
data-sharing attribute explicitly determined by being listed in a data-sharing attribute
clause. See Section A.28 on page 229 for examples.
26
Restrictions
27
The restrictions to the default clause are as follows:
28
• Only a single default clause may be specified on a parallel or task directive.
23
24
29
Chapter 2
Directives
87
1
2.9.3.2
shared clause
2
Summary
3
4
The shared clause declares one or more list items to be shared by tasks generated by
a parallel or task construct.
5
Syntax
6
The syntax of the shared clause is as follows:
7
8
shared(list)
9
Description
10
All references to a list item within a task refer to the storage area of the original variable
at the point the directive was encountered.
11
12
13
14
It is the programmer's responsibility to ensure, by adding proper synchronization, that
storage shared by an explicit task region does not reach the end of its lifetime before
the explicit task region completes its execution.
15
Fortran
16
17
18
19
20
21
22
23
24
The association status of a shared pointer becomes undefined upon entry to and on exit
from the parallel or task construct if it is associated with a target or a subobject of
a target that is in a private, firstprivate, lastprivate, or reduction
clause inside the construct.
Under certain conditions, passing a shared variable to a non-intrinsic procedure may
result in the value of the shared variable being copied into temporary storage before the
procedure reference, and back out of the temporary storage into the actual argument
storage after the procedure reference. It is implementation defined when this situation
occurs. See Section A.29 on page 231 for an example of this behavior.
25
26
27
Note – This situation may occur when the following three conditions hold regarding an
actual argument in a reference to a non-intrinsic procedure:
28
a. The actual argument is one of the following:
29
• A shared variable.
30
• A subobject of a shared variable.
31 88
OpenMP API • Version 3.0 May 2008
1
• An object associated with a shared variable.
2
• An object associated with a subobject of a shared variable.
3
b. The actual argument is also one of the following:
4
• An array section.
5
• An array section with a vector subscript.
6
• An assumed-shape array.
7
• A pointer array.
8
c. The associated dummy argument for this actual argument is an explicit-shape array
or an assumed-size array.
9
This effectively results in references to, and definitions of, the temporary storage during
the procedure reference. Any references to (or definitions of) the shared storage that is
associated with the dummy argument by any other task must be synchronized with the
procedure reference to avoid possible race conditions.
10
11
12
13
14
15
16
17
Fortran
2.9.3.3
private clause
18
Summary
19
The private clause declares one or more list items to be private to a task.
20
Syntax
21
The syntax of the private clause is as follows:
22
23
private(list)
24
Description
25
Each task that references a list item that appears in a private clause in any statement
in the construct receives a new list item whose language-specific attributes are derived
from the original list item. Inside the construct, all references to the original list item are
replaced by references to the new list item. In the rest of the region, it is unspecified
whether references are to the new list item or the original list item. Therefore, if an
26
27
28
29
30
Chapter 2
Directives
89
3
attempt is made to reference the original item, its value after the region is also
unspecified. If a task does not reference a list item that appears in a private clause, it
is unspecified whether that task receives a new list item.
4
The value and/or allocation status of the original list item will change only:
5
• if accessed and modified via pointer,
6
• if (possibly) accessed in the region but outside of the construct, or
7
• as a side effect of directives or clauses.
8
List items that appear in a private, firstprivate, or reduction clause in a
parallel construct may also appear in a private clause in an enclosed parallel,
task, or workshare construct. List items that appear in a private or
firstprivate clause in a task construct may also appear in a private clause in
an enclosed parallel or task construct. List items that appear in a private,
firstprivate, lastprivate, or reduction clause in a workshare construct
may also appear in a private clause in an enclosed parallel or task construct.
See Section A.31 on page 235 for an example.
1
2
9
10
11
12
13
14
15
16
C/C++
17
18
19
20
21
22
23
24
25
A new list item of the same type, with automatic storage duration, is allocated for the
construct. The storage and thus lifetime of these list items lasts until the block in which
they are created exits. The size and alignment of the new list item are determined by the
type of the variable. This allocation occurs once for each task generated by the
construct, if the task references the list item in any statement.
The new list item is initialized, or has an undefined initial value, as if it had been locally
declared without an initializer. The order in which any default constructors for different
private variables of class type are called is unspecified. The order in which any
destructors for different private variables of class type are called is unspecified.
C/C++
26
Fortran
31
A new list item of the same type is allocated once for each implicit task in the
parallel region, or for each task generated by a task construct, if the construct
references the list item in any statement. The initial value of the new list item is
undefined. Within a parallel, worksharing, or task region, the initial status of a
private pointer is undefined.
32
For a list item with the ALLOCATABLE attribute:
33
• if the list item is "not currently allocated", the new list item will have an initial state
27
28
29
30
34
35
36
37 90
of "not currently allocated";
• if the list item is allocated, the new list item will have an initial state of allocated
with the same array bounds.
OpenMP API • Version 3.0 May 2008
4
A list item that appears in a private clause may be storage-associated with other
variables when the private clause is encountered. Storage association may exist
because of constructs such as EQUIVALENCE or COMMON. If A is a variable appearing
in a private clause and B is a variable that is storage-associated with A, then:
5
• The contents, allocation, and association status of B are undefined on entry to the
1
2
3
6
7
8
9
10
11
parallel or task region.
• Any definition of A, or of its allocation or association status, causes the contents,
allocation, and association status of B to become undefined.
• Any definition of B, or of its allocation or association status, causes the contents,
allocation, and association status of A to become undefined.
For examples, see Section A.32 on page 237.
12
Fortran
13
For examples of the private clause, see Section A.30 on page 232.
14
Restrictions
15
The restrictions to the private clause are as follows:
16
• A variable which is part of another variable (as an array or structure element) cannot
17
18
appear in a private clause.
C/C++
19
20
21
22
23
24
• A variable of class type (or array thereof) that appears in a private clause requires
an accessible, unambiguous default constructor for the class type.
• A variable that appears in a private clause must not have a const-qualified type
unless it is of class type with a mutable member.
• A variable that appears in a private clause must not have an incomplete type or a
reference type.
C/C++
25
Fortran
26
27
• A variable that appears in a private clause must either be definable, or an
allocatable array.
28
• Assumed-size arrays may not appear in a private clause.
29
• Variables that appear in namelist statements, in variable format expressions, and in
30
31
32
expressions for statement function definitions, may not appear in a private clause.
Fortran
Chapter 2
Directives
91
1
2.9.3.4
firstprivate clause
2
Summary
3
5
The firstprivate clause declares one or more list items to be private to a task, and
initializes each of them with the value that the corresponding original item has when the
construct is encountered.
6
Syntax
7
The syntax of the firstprivate clause is as follows:
4
8
9
firstprivate(list)
10
Description
11
The firstprivate clause provides a superset of the functionality provided by the
private clause.
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30 92
A list item that appears in a firstprivate clause is subject to the private clause
semantics described in Section 2.9.3.3 on page 89. In addition, the new list item is
initialized from the original list item existing before the construct. The initialization of
the new list item is done once for each task that references the list item in any statement
in the construct. The initialization is done prior to the execution of the construct.
For a firstprivate clause on a parallel or task construct, the initial value of
the new list item is the value of the original list item that exists immediately prior to the
construct in the task region where the construct is encountered. For a firstprivate
clause on a worksharing construct, the initial value of the new list item for each implicit
task of the threads that execute the worksharing construct is the value of the original list
item that exists in the implicit task immediately prior to the point in time that the
worksharing construct is encountered.
To avoid race conditions, concurrent updates of the original list item must be
synchronized with the read of the original list item that occurs as a result of the
firstprivate clause.
If a list item appears in both firstprivate and lastprivate clauses, the update
required for lastprivate occurs after all the initializations for firstprivate.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
3
4
5
6
7
For variables of non-array type, the initialization occurs by copy assignment. For a
(possibly multi-dimensional) array of elements of non-array type, each element is
initialized as if by assignment from an element of the original array to the corresponding
element of the new array. For variables of class type, a copy constructor is invoked to
perform the initialization. The order in which copy constructors for different variables of
class type are called is unspecified.
C/C++
8
Fortran
9
The initialization of the new list items occurs as if by assignment.
10
Fortran
11
Restrictions
12
The restrictions to the firstprivate clause are as follows:
13
• A variable which is part of another variable (as an array or structure element) cannot
14
15
16
17
18
19
20
21
22
appear in a firstprivate clause.
• A list item that is private within a parallel region must not appear in a
firstprivate clause on a worksharing construct if any of the worksharing
regions arising from the worksharing construct ever bind to any of the parallel
regions arising from the parallel construct.
• A list item that appears in a reduction clause of a parallel construct must not
appear in a firstprivate clause on a worksharing or task construct if any of
the worksharing or task regions arising from the worksharing or task construct
ever bind to any of the parallel regions arising from the parallel construct.
23
• A list item that appears in a reduction clause in a worksharing construct must not
24
appear in a firstprivate clause in a task construct encountered during execution
of any of the worksharing regions arising from the worksharing construct.
25
26
C/C++
27
28
29
30
31
32
• A variable of class type (or array thereof) that appears in a firstprivate clause
requires an accessible, unambiguous copy constructor for the class type.
• A variable that appears in a firstprivate clause must not have a const-
qualified type unless it is of class type with a mutable member.
• A variable that appears in a firstprivate clause must not have an incomplete
type or a reference type.
C/C++
33
Fortran
34
35
• A variable that appears in a firstprivate clause must be definable.
Chapter 2
Directives
93
• Fortran pointers, Cray pointers, and assumed-size arrays may not appear in a
1
firstprivate clause.
2
• Variables that appear in namelist statements, in variable format expressions, and in
3
expressions for statement function definitions, may not appear in a firstprivate
clause.
4
5
6
7
Fortran
2.9.3.5
lastprivate clause
8
Summary
9
11
The lastprivate clause declares one or more list items to be private to an implicit
task, and causes the corresponding original list item to be updated after the end of the
region.
12
Syntax
13
The syntax of the lastprivate clause is as follows:
10
14
15
lastprivate(list)
16
Description
17
The lastprivate clause provides a superset of the functionality provided by the
private clause.
18
19
20
21
22
23
A list item that appears in a lastprivate clause is subject to the private clause
semantics described in Section 2.9.3.3 on page 89. In addition, when a lastprivate
clause appears on the directive that identifies a worksharing construct, the value of each
new list item from the sequentially last iteration of the associated loops, or the lexically
last section construct, is assigned to the original list item.
24
C/C++
25
26
For a (possibly multi-dimensional) array of elements of non-array type, each element is
assigned to the corresponding element of the original array.
C/C++
27
28
29
30 94
List items that are not assigned a value by the sequentially last iteration of the loops, or
by the lexically last section construct, have unspecified values after the construct.
Unassigned subcomponents also have unspecified values after the construct.
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
The original list item becomes defined at the end of the construct if there is an implicit
barrier at that point. To avoid race conditions, concurrent reads or updates of the original
list item must be synchronized with the update of the original list item that occurs as a
result of the lastprivate clause.
If the lastprivate clause is used on a construct to which nowait is applied,
accesses to the original list item may create a data race. To avoid this, synchronization
must be inserted to ensure that the sequentially last iteration or lexically last section
construct has stored and flushed that list item.
10
If a list item appears in both firstprivate and lastprivate clauses, the update
required for lastprivate occurs after all initializations for firstprivate.
11
For an example of the lastprivate clause, see Section A.34 on page 241.
12
Restrictions
13
The restrictions to the lastprivate clause are as follows:
14
• A variable which is part of another variable (as an array or structure element) cannot
9
15
16
17
18
19
20
appear in a lastprivate clause.
• A list item that is private within a parallel region, or that appears in the
reduction clause of a parallel construct, must not appear in a lastprivate
clause on a worksharing construct if any of the corresponding worksharing regions
ever binds to any of the corresponding parallel regions.
C/C++
21
22
23
24
• A variable of class type (or array thereof) that appears in a lastprivate clause
requires an accessible, unambiguous default constructor for the class type, unless the
list item is also specified in a firstprivate clause.
• A variable of class type (or array thereof) that appears in a lastprivate clause
27
requires an accessible, unambiguous copy assignment operator for the class type.The
order in which copy assignment operators for different variables of class type are
called is unspecified.
28
• A variable that appears in a lastprivate clause must not have a const-qualified
25
26
29
30
31
type unless it is of class type with a mutable member.
• A variable that appears in a lastprivate clause must not have an incomplete type
or a reference type.
C/C++
32
Fortran
33
34
• A variable that appears in a lastprivate clause must be definable.
Chapter 2
Directives
95
1
• An original list item with the ALLOCATABLE attribute must be in the allocated state
2
at entry to the construct containing the lastprivate clause. The list item in the
sequentially last iteration or lexically last section must be in the allocated state upon
exit from that iteration or section with the same bounds as the corresponding original
list item.
3
4
5
• Fortran pointers, Cray pointers, and assumed-size arrays may not appear in a
6
lastprivate clause.
7
• Variables that appear in namelist statements, in variable format expressions, and in
8
expressions for statement function definitions, may not appear in a lastprivate
clause.
9
10
11
12
Fortran
2.9.3.6
reduction clause
13
Summary
14
17
The reduction clause specifies an operator and one or more list items. For each list
item, a private copy is created in each implicit task, and is initialized appropriately for
the operator. After the end of the region, the original list item is updated with the values
of the private copies using the specified operator.
18
Syntax
15
16
19
C/C++
20
The syntax of the reduction clause is as follows:
21
reduction(operator:list)
22
24
The following table lists the operators that are valid and their initialization values. The
actual initialization value depends on the data type of the reduction list item.
25
26
Operator
Initialization value
27
+
0
28
*
1
29
-
0
23
30 96
OpenMP API • Version 3.0 May 2008
1
2
&
~0
3
|
0
4
^
0
5
&&
1
6
||
0
7
C/C++
8
Fortran
9
The syntax of the reduction clause is as follows:
10
reduction({operator | intrinsic_procedure_name}:list)
11
14
The following table lists the operators and intrinsic_procedure_names that are valid and
their initialization values. The actual initialization value depends on the data type of the
reduction list item.
15
16
17
Operator/
Intrinsic
Initialization value
18
+
0
19
*
1
20
-
0
21
.and.
.true.
22
.or.
.false.
23
.eqv.
.true.
24
.neqv.
.false.
25
26
max
Most negative representable number in the
reduction list item type
27
28
min
Largest representable number in the reduction
list item type
29
iand
All bits on
30
ior
0
31
ieor
0
12
13
32
33
34
Fortran
Chapter 2
Directives
97
1
Description
2
The reduction clause can be used to perform some forms of recurrence calculations
(involving mathematically associative and commutative operators) in parallel.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
A private copy of each list item is created, one for each implicit task, as if the private
clause had been used. The private copy is then initialized to the initialization value for
the operator, as specified above. At the end of the region for which the reduction
clause was specified, the original list item is updated by combining its original value
with the final value of each of the private copies, using the operator specified. (The
partial results of a subtraction reduction are added to form the final value.)
If nowait is not used, the reduction computation will be complete at the end of the
construct; however, if the reduction clause is used on a construct to which nowait is
also applied, accesses to the original list item will create a race and, thus, have
unspecified effect unless synchronization ensures that they occur after all threads have
executed all of their iterations or section constructs, and the reduction computation
has completed and stored the computed value of that list item. This can most simply be
ensured through a barrier synchronization.
The order in which the values are combined is unspecified. Therefore, comparing
sequential and parallel runs, or comparing one parallel run to another (even if the
number of threads used is the same), there is no guarantee that bit-identical results will
be obtained or that side effects (such as floating point exceptions) will be identical.
To avoid race conditions, concurrent reads or updates of the original list item must be
synchronized with the update of the original list item that occurs as a result of the
reduction computation.
24
25
26
Note – List items specified in a reduction clause are typically used in the enclosed
region in certain forms.
27
C/C++
28
A reduction is typically specified for statements of the form:
29
30
31
32
33
34
35
36
37 98
x = x op expr
x binop= expr
x = expr op x
x++
++x
x---x
(except for subtraction)
OpenMP API • Version 3.0 May 2008
1
2
3
where expr has scalar type and does not reference x, op is not an overloaded operator,
but one of +, *, -, &, ^, |, &&, or ||, and binop is not an overloaded operator, but one
of +, *, -, &, ^, or |.
C/C++
4
Fortran
5
A reduction using an operator is typically specified for statements of the form:
6
7
8
x = x op expr
x = expr op x (except for subtraction)
10
where op is +, *, -, .and., .or., .eqv., or .neqv., the expression does not involve
x, and the reduction op is the last operation performed on the right hand side.
11
A reduction using an intrinsic is typically specified for statements of the form:
9
12
13
14
15
16
x = intr(x,expr_list)
x = intr(expr_list, x)
where intr is max, min, iand, ior, or ieor and expr_list is a comma separated list of
expressions not involving x.
17
18
Fortran
For examples, see Section A.35 on page 242.
19
20
Restrictions
21
The restrictions to the reduction clause are as follows:
22
• A list item that appears in a reduction clause of a worksharing construct must be
23
shared in the parallel regions to which any of the worksharing regions arising
from the worksharing construct bind.
24
25
26
27
28
29
• A list item that appears in a reduction clause of the innermost enclosing
worksharing or parallel construct may not be accessed in an explicit task.
• Any number of reduction clauses can be specified on the directive, but a list item
can appear only once in the reduction clause(s) for that directive.
C/C++
30
31
32
33
34
• The type of a list item that appears in a reduction clause must be valid for the
reduction operator.
• Aggregate types (including arrays), pointer types and reference types may not appear
in a reduction clause.
Chapter 2
Directives
99
1
• A list item that appears in a reduction clause must not be const-qualified.
2
• The operator specified in a reduction clause cannot be overloaded with respect to
the list items that appear in that clause.
3
C/C++
4
Fortran
• The type of a list item that appears in a reduction clause must be valid for the
5
reduction operator or intrinsic.
6
7
• A list item that appears in a reduction clause must be definable.
8
• A list item that appears in a reduction clause must be a named variable of
intrinsic type.
9
10
• An original list item with the ALLOCATABLE attribute must be in the allocated state
11
at entry to the construct containing the reduction clause. Additionally, the list item
must not be deallocated and/or allocated within the region.
12
• Fortran pointers, Cray pointers and assumed-size arrays may not appear in a
13
reduction clause.
14
• Operators specified must be intrinsic operators and any intrinsic_procedure_name
15
must refer to one of the allowed intrinsic procedures. Assignment to the reduction list
items must be via intrinsic assignment. See Section A.35 on page 242 for examples.
16
17
18
19
Fortran
2.9.4
20
21
22
23
24
25
26
27
28
29
30 100
Data Copying Clauses
This section describes the copyin clause (valid on the parallel directive and
combined parallel worksharing directives) and the copyprivate clause (valid on the
single directive).
These clauses support the copying of data values from private or threadprivate variables
on one implicit task or thread to the corresponding variables on other implicit tasks or
threads in the team.
The clauses accept a comma-separated list of list items (see Section 2.1 on page 22). All
list items appearing in a clause must be visible, according to the scoping rules of the
base language. Clauses may be repeated as needed, but a list item that specifies a given
variable may not appear in more than one clause on the same directive.
OpenMP API • Version 3.0 May 2008
1
2.9.4.1
copyin clause
2
Summary
3
5
The copyin clause provides a mechanism to copy the value of the master thread’s
threadprivate variable to the threadprivate variable of each other member of the team
executing the parallel region.
6
Syntax
7
The syntax of the copyin clause is as follows:
4
8
9
10
copyin(list)
Description
11
C/C++
12
13
14
15
16
17
18
The copy is done after the team is formed and prior to the start of execution of the
associated structured block. For variables of non-array type, the copy occurs by copy
assignment. For a (possibly multi-dimensional) array of elements of non-array type,
each element is copied as if by assignment from an element of the master thread’s array
to the corresponding element of the other thread’s array. For class types, the copy
assignment operator is invoked. The order in which copy assignment operators for
different variables of class type are called is unspecified.
C/C++
19
Fortran
20
21
The copy is done, as if by assignment, after the team is formed and prior to the start of
execution of the associated structured block.
24
On entry to any parallel region, each thread’s copy of a variable that is affected by
a copyin clause for the parallel region will acquire the allocation, association, and
definition status of the master thread’s copy, according to the following rules:
25
• If it has the POINTER attribute:
22
23
27
• if the master thread’s copy is associated with a target that each copy can become
associated with, each copy will become associated with the same target;
28
• if the master thread’s copy is disassociated, each copy will become disassociated;
29
• otherwise, each copy will have an undefined association status.
26
30
Chapter 2
Directives
101
• If it does not have the POINTER attribute, each copy becomes defined with the value
1
of the master thread’s copy as if by intrinsic assignment.
2
3
Fortran
4
For an example of the copyin clause, see Section A.36 on page 248.
5
Restrictions
6
7
The restrictions to the copyin clause are as follows:
8
• A list item that appears in a copyin clause must be threadprivate.
9
• A variable of class type (or array thereof) that appears in a copyin clause requires
C/C++
an accessible, unambiguous copy assignment operator for the class type.
10
C/C++
11
Fortran
• A list item that appears in a copyin clause must be threadprivate. Named variables
12
appearing in a threadprivate common block may be specified: it is not necessary to
specify the whole common block.
13
14
15
• A common block name that appears in a copyin clause must be declared to be a
16
common block in the same scoping unit in which the copyin clause appears.
• An array with the ALLOCATABLE attribute must be in the allocated state. Each
17
thread's copy of that array must be allocated with the same bounds.
18
19
20
Fortran
2.9.4.2
copyprivate clause
21
Summary
22
The copyprivate clause provides a mechanism to use a private variable to broadcast
a value from the data environment of one implicit task to the data environments of the
other implicit tasks belonging to the parallel region.
23
24
25
26
27
28 102
To avoid race conditions, concurrent reads or updates of the list item must be
synchronized with the update of the list item that occurs as a result of the
copyprivate clause.
OpenMP API • Version 3.0 May 2008
1
Syntax
2
The syntax of the copyprivate clause is as follows:
3
4
copyprivate(list)
5
Description
6
The effect of the copyprivate clause on the specified list items occurs after the
execution of the structured block associated with the single construct (see
Section 2.5.3 on page 49), and before any of the threads in the team have left the barrier
at the end of the construct.
7
8
9
10
C/C++
11
12
13
14
15
16
17
18
19
In all other implicit tasks belonging to the parallel region, each specified list item
becomes defined with the value of the corresponding list item in the implicit task whose
thread executed the structured block. For variables of non-array type, the definition
occurs by copy assignment. For a (possibly multi-dimensional) array of elements of nonarray type, each element is copied by copy assignment from an element of the array in
the data environment of the implicit task whose thread executed the structured block to
the corresponding element of the array in the data environment of the other implicit
tasks. For class types, a copy assignment operator is invoked. The order in which copy
assignment operators for different variables of class type are called is unspecified.
C/C++
20
Fortran
21
22
23
24
25
26
27
28
If a list item is not a pointer, then in all other implicit tasks belonging to the parallel
region, the list item becomes defined (as if by assignment) with the value of the
corresponding list item in the implicit task whose thread executed the structured block.
If the list item is a pointer, then in all other implicit tasks belonging to the parallel
region, the list item becomes pointer associated (as if by pointer assignment) with the
corresponding list item in the implicit task whose thread executed the structured block.
Fortran
For examples of the copyprivate clause, see Section A.37 on page 250.
29
30
31
32
Note – The copyprivate clause is an alternative to using a shared variable for the
value when providing such a shared variable would be difficult (for example, in a
recursion requiring a different variable at each level).
33
34
Chapter 2
Directives
103
1
Restrictions
2
The restrictions to the copyprivate clause are as follows:
3
• All list items that appear in the copyprivate clause must be either threadprivate
or private in the enclosing context.
4
• A list item that appears in a copyprivate clause may not appear in a private or
5
firstprivate clause on the single construct.
6
7
C/C++
8
• A variable of class type (or array thereof) that appears in a copyprivate clause
9
requires an accessible unambiguous copy assignment operator for the class type.
C/C++
10
Fortran
11
• A common block that appears in a copyprivate clause must be threadprivate.
12
• Assumed-size arrays may not appear in a copyprivate clause.
13
• An array with the ALLOCATABLE attribute must be in the allocated state with the
same bounds in all threads affected by the copyprivate clause.
14
15
Fortran
16
17
2.10
Nesting of Regions
19
This section describes a set of restrictions on the nesting of regions. The restrictions on
nesting are as follows:
20
• A worksharing region may not be closely nested inside a worksharing, explicit task,
18
21
22
23
24
25
26
27
28
29
critical, ordered, or master region.
• A barrier region may not be closely nested inside a worksharing, explicit task,
critical, ordered, or master region.
• A master region may not be closely nested inside a worksharing or explicit task
region.
• An ordered region may not be closely nested inside a critical or explicit task
region.
• An ordered region must be closely nested inside a loop region (or parallel loop
region) with an ordered clause.
30
• A critical region may not be nested (closely or otherwise) inside a critical
31
region with the same name. Note that this restriction is not sufficient to prevent
deadlock.
32
33 104
OpenMP API • Version 3.0 May 2008
1
2
3
For examples illustrating these rules, see Section A.17 on page 199, Section A.38 on
page 255, Section A.39 on page 258, and Section A.13 on page 177.
Chapter 2
Directives
105
1
106
OpenMP API • Version 3.0 May 2008
1
2
3
CHAPTER
3
Runtime Library Routines
4
6
This chapter describes the OpenMP API runtime library routines and is divided into the
following sections:
7
• Runtime library definitions (Section 3.1 on page 108).
8
• Execution environment routines that can be used to control and query the parallel
5
9
10
11
execution environment (Section 3.2 on page 109).
• Lock routines that can be used to synchronize access to data (Section 3.3 on page
134).
12
• Portable timer routines (Section 3.4 on page 142).
13
Throughout this chapter, true and false are used as generic terms to simplify the
description of the routines.
14
15
C/C++
16
true means a nonzero integer value and false means an integer value of zero.
C/C++
17
Fortran
18
true means a logical value of .TRUE. and false means a logical value of .FALSE..
19
Fortran
20
Fortran
21
Restrictions
22
The following restriction applies to all OpenMP runtime library routines:
23
• OpenMP runtime library routines may not be called from PURE or ELEMENTAL
24
25
26
procedures.
Fortran
107
1
2
3.1
3
4
5
6
7
Runtime Library Definitions
For each base language, a compliant implementation must supply a set of definitions for
the OpenMP API runtime library routines and the special data types of their parameters.
The set of definitions must contain a declaration for each OpenMP API runtime library
routine and a declaration for the simple lock, nestable lock and schedule data types. In
addition, each set of definitions may specify other implementation specific values.
8
C/C++
9
The library routines are external functions with “C” linkage.
10
11
Prototypes for the C/C++ runtime library routines described in this chapter shall be
provided in a header file named omp.h. This file defines the following:
12
• The prototypes of all the routines in the chapter.
13
• The type omp_lock_t.
14
• The type omp_nest_lock_t.
15
• The type omp_sched_t.
16
See Section D.1 on page 302 for an example of this file.
C/C++
17
Fortran
18
19
The OpenMP Fortran API runtime library routines are external procedures. The return
values of these routines are of default kind, unless otherwise specified.
23
Interface declarations for the OpenMP Fortran runtime library routines described in this
chapter shall be provided in the form of a Fortran include file named omp_lib.h or
a Fortran 90 module named omp_lib. It is implementation defined whether the
include file or the module file (or both) is provided.
24
These files define the following:
25
• The interfaces of all of the routines in this chapter.
26
• The integer parameter omp_lock_kind.
27
• The integer parameter omp_nest_lock_kind.
28
• The integer parameter omp_sched_kind.
29
• The integer parameter openmp_version with a value yyyymm where yyyy
30
and mm are the year and month designations of the version of the OpenMP Fortran
API that the implementation supports. This value matches that of the C preprocessor
macro _OPENMP, when a macro preprocessor is supported (see Section 2.2 on page
26).
20
21
22
31
32
33
34 108
OpenMP API • Version 3.0 May 2008
1
See Section D.2 on page 304 and Section D.3 on page 306 for examples of these files.
2
It is implementation defined whether any of the OpenMP runtime library routines that
take an argument are extended with a generic interface so arguments of different KIND
type can be accommodated. See Appendix D.4 for an example of such an extension.
3
4
5
Fortran
6
7
3.2
Execution Environment Routines
9
The routines described in this section affect and monitor threads, processors, and the
parallel environment.
10
• the omp_set_num_threads routine.
11
• the omp_get_num_threads routine.
12
• the omp_get_max_threads routine.
13
• the omp_get_thread_num routine.
14
• the omp_get_num_procs routine.
15
• the omp_in_parallel routine.
16
• the omp_set_dynamic routine.
17
• the omp_get_dynamic routine.
18
• the omp_set_nested routine.
19
• the omp_get_nested routine.
20
• the omp_set_schedule routine.
21
• the omp_get_schedule routine.
22
• the omp_get_thread_limit routine.
23
• the omp_set_max_active_levels routine.
24
• the omp_get_max_active_levels routine.
25
• the omp_get_level routine.
26
• the omp_get_ancestor_thread_num routine.
27
• the omp_get_team_size routine.
28
• the omp_get_active_level routine.
8
29
Chapter 3
Runtime Library Routines
109
1
3.2.1
omp_set_num_threads
2
Summary
3
5
The omp_set_num_threads routine affects the number of threads to be used for
subsequent parallel regions that do not specify a num_threads clause, by setting
the value of the nthreads-var ICV.
6
Format
4
C/C++
7
8
void omp_set_num_threads(int num_threads);
9
C/C++
Fortran
10
11
12
subroutine omp_set_num_threads(num_threads)
integer num_threads
13
14
Fortran
15
Constraints on Arguments
16
17
The value of the argument passed to this routine must evaluate to a positive integer, or
else the behavior of this routine is implementation defined.
18
Binding
19
The binding task set for an omp_set_num_threads region is the generating task.
20
Effect
21
The effect of this routine is to set the value of the nthreads-var ICV to the value
specified in the argument.
22
23
24
25 110
See Section 2.4.1 on page 35 for the rules governing the number of threads used to
execute a parallel region.
OpenMP API • Version 3.0 May 2008
2
For an example of the omp_set_num_threads routine, see Section A.40 on page
265.
3
Cross References
4
• nthreads-var ICV, see Section 2.3 on page 28.
5
• OMP_NUM_THREADS environment variable, see Section 4.2 on page 147.
6
• omp_get_max_threads routine, see Section 3.2.3 on page 112.
7
• parallel construct, see Section 2.4 on page 32.
8
• num_threads clause, see Section 2.4 on page 32.
1
9
3.2.2
omp_get_num_threads
10
Summary
11
12
The omp_get_num_threads routine returns the number of threads in the current
team.
13
Format
C/C++
14
15
int omp_get_num_threads(void);
16
C/C++
Fortran
17
18
integer function omp_get_num_threads()
19
20
Fortran
21
Binding
22
23
The binding region for an omp_get_num_threads region is the innermost enclosing
parallel region.
24
Chapter 3
Runtime Library Routines
111
1
Effect
2
The omp_get_num_threads routine returns the number of threads in the team
executing the parallel region to which the routine region binds. If called from the
sequential part of a program, this routine returns 1. For examples, see Section A.41 on
page 266.
3
4
5
7
See Section 2.4.1 on page 35 for the rules governing the number of threads used to
execute a parallel region.
8
Cross References
9
• parallel construct, see Section 2.4 on page 32.
10
• omp_set_num_threads routine, see Section 3.2.1 on page 110.
11
• OMP_NUM_THREADS environment variable, see Section 4.2 on page 147.
6
12
3.2.3
omp_get_max_threads
13
Summary
14
16
The omp_get_max_threads routine returns an upper bound on the number of
threads that could be used to form a new team if a parallel region without a
num_threads clause were encountered after execution returns from this routine.
17
Format
15
C/C++
18
19
int omp_get_max_threads(void);
20
C/C++
Fortran
21
22
integer function omp_get_max_threads()
23
24
25 112
Fortran
OpenMP API • Version 3.0 May 2008
1
Binding
2
The binding task set for an omp_get_max_threads region is the generating task.
3
Effect
4
The value returned by omp_get_max_threads is the value of the nthreads-var ICV.
This value is also an upper bound on the number of threads that could be used to form a
new team if a parallel region without a num_threads clause were encountered after
execution returns from this routine.
5
6
7
See Section 2.4.1 on page 35 for the rules governing the number of threads used to
execute a parallel region.
8
9
10
Note – The return value of the omp_get_max_threads routine can be used to
dynamically allocate sufficient storage for all threads in the team formed at the
subsequent active parallel region.
11
12
13
14
15
Cross References
16
• nthreads-var ICV, see Section 2.3 on page 28.
17
• parallel construct, see Section 2.4 on page 32.
18
• num_threads clause, see Section 2.4 on page 32.
19
• omp_set_num_threads routine, see Section 3.2.1 on page 110.
20
• OMP_NUM_THREADS environment variable, see Section 4.2 on page 147.
21
3.2.4
omp_get_thread_num
22
Summary
23
The omp_get_thread_num routine returns the thread number, within the current
team, of the thread executing the implicit or explicit task region from which
omp_get_thread_num is called.
24
25
26
Chapter 3
Runtime Library Routines
113
1
Format
C/C++
2
3
int omp_get_thread_num(void);
4
C/C++
Fortran
5
6
integer function omp_get_thread_num()
7
8
Fortran
9
Binding
10
12
The binding thread set for an omp_get_thread_num region is the current team. The
binding region for an omp_get_thread_num region is the innermost enclosing
parallel region.
13
Effect
14
The omp_get_thread_num routine returns the thread number of the current thread,
within the team executing the parallel region to which the routine region binds. The
thread number is an integer between 0 and one less than the value returned by
omp_get_num_threads, inclusive. The thread number of the master thread of the
team is 0. The routine returns 0 if it is called from the sequential part of a program.
11
15
16
17
18
19
20
21
22
Note – The thread number may change at any time during the execution of an untied
task. The value returned by omp_get_thread_num is not generally useful during the
execution of such a task region.
23
24
Cross References
25
• omp_get_num_threads routine, see Section 3.2.2 on page 111.
26 114
OpenMP API • Version 3.0 May 2008
1
3.2.5
omp_get_num_procs
2
Summary
3
4
The omp_get_num_procs routine returns the number of processors available to the
program.
5
Format
C/C++
6
7
int omp_get_num_procs(void);
8
C/C++
Fortran
9
10
integer function omp_get_num_procs()
11
12
Fortran
13
Binding
14
16
The binding thread set for an omp_get_num_procs region is all threads. The effect
of executing this routine is not related to any specific region corresponding to any
construct or API routine.
17
Effect
18
22
The omp_get_num_procs routine returns the number of processors that are available
to the program at the time the routine is called. Note that this value may change between
the time that it is determined by the omp_get_num_procs routine and the time that it
is read in the calling context due to system actions outside the control of the OpenMP
implementation.
23
Chapter 3
15
19
20
21
Runtime Library Routines
115
1
3.2.6
omp_in_parallel
2
Summary
3
4
The omp_in_parallel routine returns true if the call to the routine is enclosed by an
active parallel region; otherwise, it returns false.
5
Format
C/C++
6
7
int omp_in_parallel(void);
8
C/C++
Fortran
9
10
logical function omp_in_parallel()
11
12
Fortran
13
Binding
14
16
The binding thread set for an omp_in_parallel region is all threads. The effect of
executing this routine is not related to any specific parallel region but instead
depends on the state of all enclosing parallel regions.
17
Effect
18
omp_in_parallel returns true if any enclosing parallel region is active. If the
routine call is enclosed by only inactive parallel regions (including the implicit
parallel region), then it returns false.
15
19
20
21 116
OpenMP API • Version 3.0 May 2008
1
3.2.7
omp_set_dynamic
2
Summary
3
5
The omp_set_dynamic routine enables or disables dynamic adjustment of the
number of threads available for the execution of subsequent parallel regions by
setting the value of the dyn-var ICV.
6
Format
4
C/C++
7
8
void omp_set_dynamic(int dynamic_threads);
9
C/C++
Fortran
10
11
12
subroutine omp_set_dynamic (dynamic_threads)
logical dynamic_threads
13
14
Fortran
15
Binding
16
The binding task set for an omp_set_dynamic region is the generating task.
17
Effect
18
22
For implementations that support dynamic adjustment of the number of threads, if the
argument to omp_set_dynamic evaluates to true, dynamic adjustment is enabled;
otherwise, dynamic adjustment is disabled. For implementations that do not support
dynamic adjustment of the number of threads this routine has no effect: the value of
dyn-var remains false.
23
For an example of the omp_set_dynamic routine, see Section A.40 on page 265.
24
See Section 2.4.1 on page 35 for the rules governing the number of threads used to
execute a parallel region.
19
20
21
25
26
Chapter 3
Runtime Library Routines
117
1
Cross References:
2
• dyn-var ICV, see Section 2.3 on page 28.
3
• omp_get_num_threads routine, see Section 3.2.2 on page 111.
4
• omp_get_dynamic routine, see Section 3.2.8 on page 118.
5
• OMP_DYNAMIC environment variable, see Section 4.3 on page 148.
6
3.2.8
omp_get_dynamic
7
Summary
8
9
The omp_get_dynamic routine returns the value of the dyn-var ICV, which
determines whether dynamic adjustment of the number of threads is enabled or disabled.
10
Format
C/C++
11
12
int omp_get_dynamic(void);
13
C/C++
Fortran
14
15
logical function omp_get_dynamic()
16
17
Fortran
18
Binding
19
The binding task set for an omp_get_dynamic region is the generating task.
20
Effect
21
This routine returns true if dynamic adjustment of the number of threads is enabled; it
returns false, otherwise. If an implementation does not support dynamic adjustment of
the number of threads, then this routine always returns false.
22
23
24 118
OpenMP API • Version 3.0 May 2008
2
See Section 2.4.1 on page 35 for the rules governing the number of threads used to
execute a parallel region.
3
Cross References
4
• dyn-var ICV, see Section 2.3 on page 28.
5
• omp_set_dynamic routine, see Section 3.2.7 on page 117.
6
• OMP_DYNAMIC environment variable, see Section 4.3 on page 148.
1
7
3.2.9
omp_set_nested
8
Summary
9
10
The omp_set_nested routine enables or disables nested parallelism, by setting the
nest-var ICV.
11
Format
C/C++
12
13
void omp_set_nested(int nested);
14
C/C++
Fortran
15
16
17
subroutine omp_set_nested (nested)
logical nested
18
19
Fortran
20
Binding
21
The binding task set for an omp_set_nested region is the generating task.
22
Chapter 3
Runtime Library Routines
119
1
Effect
2
For implementations that support nested parallelism, if the argument to
omp_set_nested evaluates to true, nested parallelism is enabled; otherwise, nested
parallelism is disabled. For implementations that do not support nested parallelism, this
routine has no effect: the value of nest-var remains false.
3
4
5
7
See Section 2.4.1 on page 35 for the rules governing the number of threads used to
execute a parallel region.
8
Cross References
9
• nest-var ICV, see Section 2.3 on page 28.
10
• omp_set_max_active_levels routine, see Section 3.2.14 on page 126.
11
• omp_get_max_active_levels routine, see Section 3.2.15 on page 127.
12
• omp_get_nested routine, see Section 3.2.10 on page 120.
13
• OMP_NESTED environment variable, see Section 4.4 on page 148.
6
14
3.2.10
omp_get_nested
15
Summary
16
17
The omp_get_nested routine returns the value of the nest-var ICV, which
determines if nested parallelism is enabled or disabled.
18
Format
C/C++
19
20
int omp_get_nested(void);
21
C/C++
Fortran
22
23
logical function omp_get_nested()
24
25
26 120
Fortran
OpenMP API • Version 3.0 May 2008
1
Binding
2
The binding task set for an omp_get_nested region is the generating task.
3
Effect
4
This routine returns true if nested parallelism is enabled; it returns false, otherwise. If an
implementation does not support nested parallelism, this routine always returns false.
5
7
See Section 2.4.1 on page 35 for the rules governing the number of threads used to
execute a parallel region.
8
Cross References
9
• nest-var ICV, see Section 2.3 on page 28.
10
• omp_set_nested routine, see Section 3.2.9 on page 119.
11
• OMP_NESTED environment variable, see Section 4.4 on page 148.
6
12
3.2.11
omp_set_schedule
13
Summary
14
15
The omp_set_schedule routine affects the schedule that is applied when runtime
is used as schedule kind, by setting the value of the run-sched-var ICV.
16
Format
17
C/C++
18
19
void omp_set_schedule(omp_sched_t kind, int modifier);
20
C/C++
21
Chapter 3
Runtime Library Routines
121
1
Fortran
2
3
4
5
subroutine omp_set_schedule(kind, modifier)
integer (kind=omp_sched_kind) kind
integer modifier
6
7
Fortran
8
Constraints on Arguments
9
The first argument passed to this routine can be one of the valid OpenMP schedule kinds
(except for runtime) or any implementation specific schedule. The C/C++ header file
(omp.h) and the Fortran include file (omp_lib.h) and/or Fortran 90 module file
(omp_lib) define the valid constants. The valid constants must include the following,
which can be extended with implementation specific values:
10
11
12
13
14
C/C++
15
16
17
18
19
20
21
typedef enum omp_sched_t {
omp_sched_static = 1,
omp_sched_dynamic = 2,
omp_sched_guided = 3,
omp_sched_auto = 4
} omp_sched_t;
22
23
C/C++
24
Fortran
25
26
27
28
29
integer(kind=omp_sched_kind),
integer(kind=omp_sched_kind),
integer(kind=omp_sched_kind),
integer(kind=omp_sched_kind),
parameter
parameter
parameter
parameter
30
31
32
33 122
Fortran
OpenMP API • Version 3.0 May 2008
::
::
::
::
omp_sched_static = 1
omp_sched_dynamic = 2
omp_sched_guided = 3
omp_sched_auto = 4
1
Binding
2
The binding task set for an omp_set_schedule region is the generating task.
3
Effect
4
11
The effect of this routine is to set the value of the run-sched-var ICV to the values
specified in the two arguments. The schedule is set to the schedule type specified by the
first argument kind. It can be any of the standard schedule types or any other
implementation specific one. For the schedule types static, dynamic, and guided
the chunk_size is set to the value of the second argument, or to the default chunk_size if
the value of the second argument is less than 1; for the schedule type auto the second
argument has no meaning; for implementation specific schedule types, the values and
associated meanings of the second argument are implementation defined.
12
Cross References
13
• run-sched-var ICV, see Section 2.3 on page 28.
14
• omp_get_schedule routine, see Section 3.2.12 on page 123.
15
• OMP_SCHEDULE environment variable, see Section 4.1 on page 146.
16
• Determining the schedule of a worksharing loop, see Section 2.5.1.1 on page 45.
5
6
7
8
9
10
17
3.2.12
omp_get_schedule
18
Summary
19
The omp_get_schedule routine returns the schedule that is applied when the
runtime schedule is used.
20
21
Chapter 3
Runtime Library Routines
123
1
Format
2
C/C++
3
4
void omp_get_schedule(omp_sched_t * kind, int * modifier );
5
C/C++
6
Fortran
7
8
9
10
subroutine omp_get_schedule(kind, modifier)
integer (kind=omp_sched_kind) kind
integer modifier
11
12
Fortran
13
Binding
14
The binding task set for an omp_get_schedule region is the generating task.
15
Effect
16
20
This routine returns the run-sched-var ICV in the team executing the parallel region
to which the routine binds. The first argument kind returns the schedule to be used. It
can be any of the standard schedule types as defined in Section 3.2.11 on page 121, or
any implementation specific schedule type. The second argument is interpreted as in the
omp_set_schedule call, defined in Section 3.2.11 on page 121.
21
Cross References
22
• run-sched-var ICV, see Section 2.3 on page 28.
23
• omp_set_schedule routine, see Section 3.2.11 on page 121.
24
• OMP_SCHEDULE environment variable, see Section 4.1 on page 146.
25
• Determining the schedule of a worksharing loop, see Section 2.5.1.1 on page 45.
17
18
19
26 124
OpenMP API • Version 3.0 May 2008
1
3.2.13
omp_get_thread_limit
2
Summary
3
4
The omp_get_thread_limit routine returns the maximum number of OpenMP
threads available to the program.
5
Format
6
C/C++
7
8
int omp_get_thread_limit(void)
9
C/C++
10
Fortran
11
12
integer function omp_get_thread_limit()
13
14
Fortran
15
Binding
16
18
The binding thread set for an omp_get_thread_limit region is all threads. The
effect of executing this routine is not related to any specific region corresponding to any
construct or API routine.
19
Effect
20
The omp_get_thread_limit routine returns the maximum number of OpenMP
threads available to the program as stored in the ICV thread-limit-var.
17
21
22
Chapter 3
Runtime Library Routines
125
1
Cross References
2
• thread-limit-var ICV, see Section 2.3 on page 28.
3
• OMP_THREAD_LIMIT environment variable, see Section 4.8 on page 151.
4
3.2.14
omp_set_max_active_levels
5
Summary
6
7
The omp_set_max_active_levels routine limits the number of nested active
parallel regions, by setting the max-active-levels-var ICV.
8
Format
9
C/C++
10
11
void omp_set_max_active_levels (int max_levels)
12
C/C++
13
Fortran
14
15
16
subroutine omp_set_max_active_levels (max_levels)
integer max_levels
17
18
Fortran
19
Constraints on Arguments
20
The value of the argument passed to this routine must evaluate to a non-negative integer,
or else the behavior of this routine is implementation defined.
21
22 126
OpenMP API • Version 3.0 May 2008
1
Binding
2
5
When called from the sequential part of the program, the binding thread set for an
omp_set_max_active_levels region is the encountering thread. When called
from within any explicit parallel region, the binding thread set (and binding region, if
required) for the omp_set_max_active_levels region is implementation defined.
6
Effect
7
The effect of this routine is to set the value of the max-active-levels-var ICV to the value
specified in the argument.
3
4
8
If the number of parallel levels requested exceeds the number of levels of parallelism
supported by the implementation, the value of the max-active-levels-var ICV will be set
to the number of parallel levels support by the implementation.
9
10
11
14
This routine has the described effect only when called from the sequential part of the
program. When called from within an explicit parallel region, the effect of this
routine is implementation defined.
15
Cross References
16
• thread-limit-var ICV, see Section 2.3 on page 28.
17
• omp_get_max_active_levels routine, see Section 3.2.15 on page 127.
18
• OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.7 on page 150.
12
13
19
3.2.15
omp_get_max_active_levels
20
Summary
21
The omp_get_max_active_levels routine returns the value of the max-activelevels-var ICV, which determines the maximum number of nested active parallel regions.
22
23
Chapter 3
Runtime Library Routines
127
1
Format
2
C/C++
3
4
int omp_get_max_active_levels(void)
5
C/C++
6
Fortran
7
8
integer function omp_get_max_active_levels()
9
10
Fortran
11
Binding
12
15
When called from the sequential part of the program, the binding thread set for an
omp_get_max_active_levels region is the encountering thread. When called
from within any explicit parallel region, the binding thread set (and binding region, if
required) for the omp_get_max_active_levels region is implementation defined.
16
Effect
17
18
The omp_get_max_active_levels routine returns the value of the max-activelevels-var ICV, which determines the maximum number of nested active parallel regions.
19
Cross References
20
• thread-limit-var ICV, see Section 2.3 on page 28.
21
• omp_set_max_active_levels routine, see Section 3.2.14 on page 126.
22
• OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.7 on page 150.
13
14
23 128
OpenMP API • Version 3.0 May 2008
1
3.2.16
omp_get_level
2
Summary
3
4
The omp_get_level routine returns the number of nested parallel regions
enclosing the task that contains the call.
5
Format
6
C/C++
7
8
int omp_get_level(void)
9
C/C++
10
Fortran
11
12
integer function omp_get_level()
13
14
Fortran
15
Binding
16
18
The binding task set for an omp_get_level region is the generating task. The
binding region for an omp_get_level region is the innermost enclosing parallel
region.
19
Effect
20
23
The omp_get_level routine returns the number of nested parallel regions
(whether active or inactive) enclosing the task that contains the call, not including the
implicit parallel region. The routine always returns a non-negative integer, and returns 0
if it is called from the sequential part of the program.
24
Chapter 3
17
21
22
Runtime Library Routines
129
1
Cross References
2
• omp_get_active_level routine, see Section 3.2.19 on page 133.
3
• OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.7 on page 150.
4
3.2.17
omp_get_ancestor_thread_num
5
Summary
6
7
The omp_get_ancestor_thread_num routine returns, for a given nested level of
the current thread, the thread number of the ancestor or the current thread.
8
Format
9
C/C++
10
11
int omp_get_ancestor_thread_num(int level)
12
C/C++
13
Fortran
14
15
16
integer function omp_get_ancestor_thread_num(level)
integer level
17
18
Fortran
19
Binding
20
The binding thread set for an omp_get_ancestor_thread_num region is the
encountering thread. The binding region for an omp_get_ancestor_thread_num
region is the innermost enclosing parallel region.
21
22
23 130
OpenMP API • Version 3.0 May 2008
1
Effect
2
The omp_get_ancestor_thread_num routine returns the thread number of the
ancestor at a given nest level of the current thread or the thread number of the current
thread. If the requested nest level is outside the range of 0 and the nest level of the
current thread, as returned by the omp_get_level routine, the routine returns -1.
3
4
5
6
Note – When the omp_get_ancestor_thread_num routine is called with a value
of level=0, the routine always returns 0. If level=omp_get_level(), the routine
has the same effect as the omp_get_thread_num routine.
7
8
9
10
11
Cross References
12
• omp_get_level routine, see Section 3.2.16 on page 129.
13
• omp_get_thread_num routine, see Section 3.2.4 on page 113.
14
• omp_get_team_size routine, see Section 3.2.18 on page 131.
15
3.2.18
omp_get_team_size
16
Summary
17
18
The omp_get_team_size routine returns, for a given nested level of the current
thread, the size of the thread team to which the ancestor or the current thread belongs.
19
Chapter 3
Runtime Library Routines
131
1
Format
2
C/C++
3
4
int omp_get_team_size(int level)
5
C/C++
6
Fortran
7
8
9
integer function omp_get_team_size(level)
integer level
10
11
Fortran
12
Binding
13
15
The binding thread set for an omp_get_team_size region is the encountering
thread. The binding region for an omp_get_team_size region is the innermost
enclosing parallel region.
16
Effect
17
The omp_get_team_size routine returns the size of the thread team to which the
ancestor or the current thread belongs. If the requested nested level is outside the range
of 0 and the nested level of the current thread, as returned by the omp_get_level
routine, the routine returns -1. Inactive parallel regions are regarded like active parallel
regions executed with one thread.
14
18
19
20
21
22
23
24
25
Note – When the omp_get_team_size routine is called with a value of level=0,
the routine always returns 1. If level=omp_get_level(), the routine has the same
effect as the omp_get_num_threads routine.
26
27 132
OpenMP API • Version 3.0 May 2008
1
Cross References
2
• omp_get_num_threads routine, see Section 3.2.2 on page 111.
3
• omp_get_level routine, see Section 3.2.16 on page 129.
4
• omp_get_ancestor_thread_num routine, see Section 3.2.17 on page 130.
5
3.2.19
omp_get_active_level
6
Summary
7
8
The omp_get_active_level routine returns the number of nested, active
parallel regions enclosing the task that contains the call.
9
Format
10
C/C++
11
12
int omp_get_active_level(void)
13
C/C++
14
Fortran
15
16
integer function omp_get_active_level()
17
18
Fortran
19
Binding
20
The binding task set for the an omp_get_active_level region is the generating
task. The binding region for an omp_get_active_level region is the innermost
enclosing parallel region.
21
22
23
Chapter 3
Runtime Library Routines
133
1
Effect
2
5
The omp_get_active_level routine returns the number of nested, active parallel
regions enclosing the task that contains the call. The routine always returns a nonnegative integer, and always returns 0 if it is called from the sequential part of the
program.
6
Cross References
7
• omp_get_level routine, see Section 3.2.16 on page 129.
3
4
8
9
3.3
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35 134
Lock Routines
The OpenMP runtime library includes a set of general-purpose lock routines that can be
used for synchronization. These general-purpose lock routines operate on OpenMP locks
that are represented by OpenMP lock variables. An OpenMP lock variable must be
accessed only through the routines described in this section; programs that otherwise
access OpenMP lock variables are non-conforming.
An OpenMP lock may be in one of the following states: uninitialized, unlocked, or
locked. If a lock is in the unlocked state, a task may set the lock, which changes its state
to locked. The task which sets the lock is then said to own the lock. A task which owns
a lock may unset that lock, returning it to the unlocked state. A program in which a task
unsets a lock that is owned by another task is non-conforming.
Two types of locks are supported: simple locks and nestable locks. A nestable lock may
be set multiple times by the same task before being unset; a simple lock may not be set
if it is already owned by the task trying to set it. Simple lock variables are associated
with simple locks and may only be passed to simple lock routines. Nestable lock
variables are associated with nestable locks and may only be passed to nestable lock
routines.
Constraints on the state and ownership of the lock accessed by each of the lock routines
are described with the routine. If these constraints are not met, the behavior of the
routine is unspecified.
The OpenMP lock routines access a lock variable in such a way that they always read
and update the most current value of the lock variable. The lock routines include a flush
with no list; the read and update to the lock variable must be implemented as if they are
atomic with the flush. Therefore, it is not necessary for an OpenMP program to include
explicit flush directives to ensure that the lock variable’s value is consistent among
different tasks.
OpenMP API • Version 3.0 May 2008
2
See Section A.44 on page 271 and Section A.45 on page 274, for examples of using the
simple and the nestable lock routines, respectively.
3
Binding
4
6
The binding thread set for all lock routine regions is all threads. As a consequence, for
each OpenMP lock, the lock routine effects relate to all tasks that call the routines,
without regard to which team(s) the threads executing the tasks belong.
7
Simple Lock Routines
1
5
8
C/C++
9
10
11
The type omp_lock_t is an data type capable of representing a simple lock. For the
following routines, a simple lock variable must be of omp_lock_t type. All simple
lock routines require an argument that is a pointer to a variable of type omp_lock_t.
C/C++
12
Fortran
13
14
15
For the following routines, a simple lock variable must be an integer variable of
kind=omp_lock_kind.
Fortran
16
The simple lock routines are as follows:
17
• The omp_init_lock routine initializes a simple lock.
18
• The omp_destroy_lock routine uninitializes a simple lock.
19
• The omp_set_lock routine waits until a simple lock is available, and then sets it.
20
• The omp_unset_lock routine unsets a simple lock.
21
• The omp_test_lock routine tests a simple lock, and sets it if it is available.
22
Chapter 3
Runtime Library Routines
135
Nestable Lock Routines:
1
2
C/C++
The type omp_nest_lock_t is an data type capable of representing a nestable lock.
For the following routines, a nested lock variable must be of omp_nest_lock_t type.
All nestable lock routines require an argument that is a pointer to a variable of type
omp_nest_lock_t.
3
4
5
6
C/C++
7
Fortran
For the following routines, a nested lock variable must be an integer variable of
kind=omp_nest_lock_kind.
8
9
10
Fortran
11
The nestable lock routines are as follows:
12
• The omp_init_nest_lock routine initializes a nestable lock.
13
• The omp_destroy_nest_lock routine uninitializes a nestable lock.
14
• The omp_set_nest_lock routine waits until a nestable lock is available, and then
sets it.
15
16
• The omp_unset_nest_lock routine unsets a nestable lock.
17
• The omp_test_nest_lock routine tests a nestable lock, and sets it if it is
available.
18
19
3.3.1
omp_init_lock and omp_init_nest_lock
20
Summary
21
These routines provide the only means of initializing an OpenMP lock.
22 136
OpenMP API • Version 3.0 May 2008
Format
1
C/C++
2
void omp_init_lock(omp_lock_t *lock);
void omp_init_nest_lock(omp_nest_lock_t *lock);
3
4
5
C/C++
Fortran
6
subroutine omp_init_lock(svar)
integer (kind=omp_lock_kind) svar
7
8
subroutine omp_init_nest_lock(nvar)
integer (kind=omp_nest_lock_kind) nvar
9
10
11
12
Fortran
13
Constraints on Arguments
14
15
A program that accesses a lock that is not in the uninitialized state through either routine
is non-conforming.
16
Effect
17
18
The effect of these routines is to initialize the lock to the unlocked state (that is, no task
owns the lock). In addition, the nesting count for a nestable lock is set to zero.
19
For an example of the omp_init_lock routine, see Section A.42 on page 269.
21
omp_destroy_lock and
omp_destroy_nest_lock
22
Summary
23
These routines ensure that the OpenMP lock is uninitialized.
20
24
3.3.2
Chapter 3
Runtime Library Routines
137
Format
1
C/C++
2
void omp_destroy_lock(omp_lock_t *lock);
void omp_destroy_nest_lock(omp_nest_lock_t *lock);
3
4
5
C/C++
Fortran
6
subroutine omp_destroy_lock(svar)
integer (kind=omp_lock_kind) svar
7
8
subroutine omp_destroy_nest_lock(nvar)
integer (kind=omp_nest_lock_kind) nvar
9
10
11
12
Fortran
13
Constraints on Arguments
14
15
A program that accesses a lock that is not in the unlocked state through either routine is
non-conforming.
16
Effect
17
The effect of these routines is to change the state of the lock to uninitialized.
18
3.3.3
omp_set_lock and omp_set_nest_lock
19
Summary
20
These routines provide a means of setting an OpenMP lock. The calling task region is
suspended until the lock is set.
21
22 138
OpenMP API • Version 3.0 May 2008
1
Format
C/C++
2
3
4
void omp_set_lock(omp_lock_t *lock);
void omp_set_nest_lock(omp_nest_lock_t *lock);
5
C/C++
Fortran
6
7
8
9
10
subroutine omp_set_lock(svar)
integer (kind=omp_lock_kind) svar
subroutine omp_set_nest_lock(nvar)
integer (kind=omp_nest_lock_kind) nvar
11
12
Fortran
13
Constraints on Arguments
14
16
A program that accesses a lock that is in the uninitialized state through either routine is
non-conforming. A simple lock accessed by omp_set_lock that is in the locked state
must not be owned by the task that contains the call or deadlock will result.
17
Effect
18
Each of these routines causes suspension of the task executing the routine until the
specified lock is available and then sets the lock.
15
19
20
21
A simple lock is available if it is unlocked. Ownership of the lock is granted to the task
executing the routine.
24
A nestable lock is available if it is unlocked or if it is already owned by the task
executing the routine. The task executing the routine is granted, or retains, ownership of
the lock, and the nesting count for the lock is incremented.
25
Chapter 3
22
23
Runtime Library Routines
139
1
3.3.4
omp_unset_lock and omp_unset_nest_lock
2
Summary
3
These routines provide the means of unsetting an OpenMP lock.
4
Format
C/C++
5
6
7
void omp_unset_lock(omp_lock_t *lock);
void omp_unset_nest_lock(omp_nest_lock_t *lock);
8
C/C++
Fortran
9
10
11
12
13
subroutine omp_unset_lock(svar)
integer (kind=omp_lock_kind) svar
subroutine omp_unset_nest_lock(nvar)
integer (kind=omp_nest_lock_kind) nvar
14
15
Fortran
16
Constraints on Arguments
17
18
A program that accesses a lock that is not in the locked state or that is not owned by the
task that contains the call through either routine is non-conforming.
19
Effect
20
For a simple lock, the omp_unset_lock routine causes the lock to become unlocked.
21
For a nestable lock, the omp_unset_nest_lock routine decrements the nesting
count, and causes the lock to become unlocked if the resulting nesting count is zero.
22
23
24
25
26 140
For either routine, if the lock becomes unlocked, and if one or more tasks regions were
suspended because the lock was unavailable, the effect is that one task is chosen and
given ownership of the lock.
OpenMP API • Version 3.0 May 2008
1
3.3.5
omp_test_lock and omp_test_nest_lock
2
Summary
3
4
These routines attempt to set an OpenMP lock but do not suspend execution of the task
executing the routine.
5
Format
C/C++
6
7
8
int omp_test_lock(omp_lock_t *lock);
int omp_test_nest_lock(omp_nest_lock_t *lock);
9
C/C++
Fortran
10
11
12
13
14
logical function omp_test_lock(svar)
integer (kind=omp_lock_kind) svar
integer function omp_test_nest_lock(nvar)
integer (kind=omp_nest_lock_kind) nvar
15
16
Fortran
17
Constraints on Arguments
18
20
A program that accesses a lock that is in the uninitialized state through either routine is
non-conforming. The behavior is unspecified if a simple lock accessed by
omp_test_lock that is in the locked state is owned by the task that contains the call.
21
Effect
22
These routines attempt to set a lock in the same manner as omp_set_lock and
omp_set_nest_lock, except that they do not suspend execution of the task
executing the routine.
19
23
24
26
For a simple lock, the omp_test_lock routine returns true if the lock is successfully
set; otherwise, it returns false.
27
Chapter 3
25
Runtime Library Routines
141
For a nestable lock, the omp_test_nest_lock routine returns the new nesting count
if the lock is successfully set; otherwise, it returns zero.
1
2
3
4
3.4
Timing Routines
5
The routines described in this section support a portable wall clock timer.
6
• the omp_get_wtime routine.
7
• the omp_get_wtick routine.
8
3.4.1
omp_get_wtime
9
Summary
10
The omp_get_wtime routine returns elapsed wall clock time in seconds.
11
Format
C/C++
12
13
double omp_get_wtime(void);
14
C/C++
Fortran
15
16
double precision function omp_get_wtime()
17
18
Fortran
19
Binding
20
The binding thread set for an omp_get_wtime region is the encountering thread. The
routine’s return value is not guaranteed to be consistent across any set of threads.
21
22 142
OpenMP API • Version 3.0 May 2008
1
Effect
2
The omp_get_wtime routine returns a value equal to the elapsed wall clock time in
seconds since some “time in the past”. The actual “time in the past” is arbitrary, but it is
guaranteed not to change during the execution of the application program. The times
returned are “per-thread times”, so they are not required to be globally consistent across
all the threads participating in an application.
3
4
5
6
7
8
9
Note – It is anticipated that the routine will be used to measure elapsed times as shown
in the following example:
C/C++
10
11
12
13
14
15
16
double start;
double end;
start = omp_get_wtime();
... work to be timed ...
end = omp_get_wtime();
printf("Work took %f seconds\n", end - start);
17
C/C++
Fortran
18
19
20
21
22
23
24
25
DOUBLE PRECISION START, END
START = omp_get_wtime()
... work to be timed ...
END = omp_get_wtime()
PRINT *, "Work took", END - START, "seconds"
Fortran
26
27
28
Chapter 3
Runtime Library Routines
143
1
3.4.2
omp_get_wtick
2
Summary
3
4
The omp_get_wtick routine returns the precision of the timer used by
omp_get_wtime.
5
Format
C/C++
6
7
double omp_get_wtick(void);
8
C/C++
Fortran
9
10
double precision function omp_get_wtick()
11
12
Fortran
13
Binding
14
15
The binding thread set for an omp_get_wtick region is the encountering thread. The
routine’s return value is not guaranteed to be consistent across any set of threads.
16
Effect
17
The omp_get_wtick routine returns a value equal to the number of seconds between
successive clock ticks of the timer used by omp_get_wtime.
18
19 144
OpenMP API • Version 3.0 May 2008
1
2
3
CHAPTER
4
Environment Variables
4
12
This chapter describes the OpenMP environment variables that specify the settings of
the ICVs that affect the execution of OpenMP programs (see Section 2.3 on page 28).
The names of the environment variables must be upper case. The values assigned to the
environment variables are case insensitive and may have leading and trailing white
space. Modifications to the environment variables after the program has started, even if
modified by the program itself, are ignored by the OpenMP implementation. However,
the settings of some of the ICVs can be modified during the execution of the OpenMP
program by the use of the appropriate directive clauses or OpenMP API routines.
13
The environment variables are as follows:
14
• OMP_SCHEDULE sets the run-sched-var ICV for the runtime schedule type and
5
6
7
8
9
10
11
15
16
17
18
19
20
chunk size. It can be set to any of the valid OpenMP schedule types (i.e., static,
dynamic, guided, and auto).
• OMP_NUM_THREADS sets the nthreads-var ICV for the number of threads to use for
parallel regions.
• OMP_DYNAMIC sets the dyn-var ICV for the dynamic adjustment of threads to use
for parallel regions.
21
• OMP_NESTED sets the nest-var ICV to enable or to disable nested parallelism.
22
• OMP_STACKSIZE sets the stacksize-var ICV that specifies the size of the stack for
23
24
25
26
27
28
29
threads created by the OpenMP implementation.
• OMP_WAIT_POLICY sets the wait-policy-var ICV that controls the desired behavior
of waiting threads.
• OMP_MAX_ACTIVE_LEVELS sets the max-active-levels-var ICV that controls the
maximum number of nested active parallel regions.
• OMP_THREAD_LIMIT sets the thread-limit-var ICV that controls the maximum
number of threads participating in the OpenMP program.
32
The examples in this chapter only demonstrate how these variables might be set in Unix
C shell (csh) environments. In Korn shell (ksh) and DOS environments the actions are
similar, as follows:
33
145
30
31
• csh:
1
2
setenv OMP_SCHEDULE "dynamic"
3
• ksh:
4
5
export OMP_SCHEDULE="dynamic"
6
• DOS:
7
8
set OMP_SCHEDULE=dynamic
9
10
11
4.1
OMP_SCHEDULE
14
The OMP_SCHEDULE environment variable controls the schedule type and chunk size
of all loop directives that have the schedule type runtime, by setting the value of the
run-sched-var ICV.
15
The value of this environment variable takes the form:
16
type[,chunk]
17
where
18
• type is one of static, dynamic, guided, or auto
19
• chunk is an optional positive integer that specifies the chunk size
20
If chunk is present, there may be white space on either side of the “,”. See Section 2.5.1
on page 38 for a detailed description of the schedule types.
12
13
21
22
23
The behavior of the program is implementation defined if the value of OMP_SCHEDULE
does not conform to the above format.
26
Implementation specific schedules cannot be specified in OMP_SCHEDULE. They can
only be specified by calling omp_set_schedule, described in Section 3.2.11 on page
121.
27
Example:
24
25
28
29
30
31 146
setenv OMP_SCHEDULE "guided,4"
setenv OMP_SCHEDULE "dynamic"
OpenMP API • Version 3.0 May 2008
1
Cross References
2
• run-sched-var ICV, see Section 2.3 on page 28.
3
• Loop construct, see Section 2.5.1 on page 38.
4
• Parallel loop construct, see Section 2.6.1 on page 54.
5
• omp_set_schedule routine, see Section 3.2.11 on page 121.
6
• omp_get_schedule routine, see Section 3.2.12 on page 123.
7
8
9
10
11
12
13
4.2
OMP_NUM_THREADS
The OMP_NUM_THREADS environment variable sets the number of threads to use for
parallel regions by setting the initial value of the nthreads-var ICV. See Section 2.3
for a comprehensive set of rules about the interaction between the OMP_NUM_THREADS
environment variable, the num_threads clause, the omp_set_num_threads
library routine and dynamic adjustment of threads.
17
The value of this environment variable must be a positive integer. The behavior of the
program is implementation defined if the requested value of OMP_NUM_THREADS is
greater than the number of threads an implementation can support, or if the value is not
a positive integer.
18
Example:
14
15
16
19
20
setenv OMP_NUM_THREADS 16
21
Cross References:
22
• nthreads-var ICV, see Section 2.3 on page 28.
23
• num_threads clause, Section 2.4 on page 32.
24
• omp_set_num_threads routine, see Section 3.2.1 on page 110.
25
• omp_get_num_threads routine, see Section 3.2.2 on page 111.
26
• omp_get_max_threads routine, see Section 3.2.3 on page 112.
27
• omp_get_team_size routine, see Section 3.2.18 on page 131.
28
Chapter 4
Environment Variables
147
1
2
4.3
OMP_DYNAMIC
10
The OMP_DYNAMIC environment variable controls dynamic adjustment of the number
of threads to use for executing parallel regions by setting the initial value of the
dyn-var ICV. The value of this environment variable must be true or false. If the
environment variable is set to true, the OpenMP implementation may adjust the
number of threads to use for executing parallel regions in order to optimize the use
of system resources. If the environment variable is set to false, the dynamic
adjustment of the number of threads is disabled. The behavior of the program is
implementation defined if the value of OMP_DYNAMIC is neither true nor false.
11
Example:
3
4
5
6
7
8
9
12
setenv OMP_DYNAMIC true
13
14
Cross References:
15
• dyn-var ICV, see Section 2.3 on page 28.
16
• omp_set_dynamic routine, see Section 3.2.7 on page 117.
17
• omp_get_dynamic routine, see Section 3.2.8 on page 118.
18
19
4.4
OMP_NESTED
24
The OMP_NESTED environment variable controls nested parallelism by setting the
initial value of the nest-var ICV. The value of this environment variable must be true
or false. If the environment variable is set to true, nested parallelism is enabled; if
set to false, nested parallelism is disabled. The behavior of the program is
implementation defined if the value of OMP_NESTED is neither true nor false.
25
Example:
20
21
22
23
26
27
setenv OMP_NESTED false
28
Cross References
29
• nest-var ICV, see Section 2.3 on page 28.
30
• omp_set_nested routine, see Section 3.2.9 on page 119.
31 148
OpenMP API • Version 3.0 May 2008
• omp_get_nested routine, see Section 3.2.18 on page 131.
1
2
3
4.5
OMP_STACKSIZE
6
The OMP_STACKSIZE environment variable controls the size of the stack for threads
created by the OpenMP implementation, by setting the value of the stacksize-var ICV.
The environment variable does not control the size of the stack for the initial thread.
7
The value of this environment variable takes the form:
8
size | sizeB | sizeK | sizeM | sizeG
9
where:
10
• size is a positive integer that specifies the size of the stack for threads that are created
4
5
11
12
13
14
15
16
by the OpenMP implementation.
• B, K, M, and G are letters that specify whether the given size is in Bytes, Kilobytes,
Megabytes, or Gigabytes, respectively. If one of these letters is present, there may be
white space between size and the letter.
If only size is specified and none of B, K, M, or G is specified, then size is assumed to be
in Kilobytes.
19
The behavior of the program is implementation defined if OMP_STACKSIZE does not
conform to the above format, or if the implementation cannot provide a stack with the
requested size.
20
Examples:
17
18
21
22
23
24
25
26
27
28
setenv
setenv
setenv
setenv
setenv
setenv
setenv
OMP_STACKSIZE
OMP_STACKSIZE
OMP_STACKSIZE
OMP_STACKSIZE
OMP_STACKSIZE
OMP_STACKSIZE
OMP_STACKSIZE
2000500B
"3000 k "
10M
" 10 M "
"20 m "
" 1G"
20000
29
Cross References
30
• stacksize-var ICV, see Section 2.3 on page 28.
31
Chapter 4
Environment Variables
149
1
2
4.6
OMP_WAIT_POLICY
6
The OMP_WAIT_POLICY environment variable provides a hint to an OpenMP
implementation about the desired behavior of waiting threads by setting the wait-policyvar ICV. A compliant OpenMP implementation may or may not abide by the setting of
the environment variable.
7
The value of this environment variable takes the form:
8
ACTIVE | PASSIVE
9
The ACTIVE value specifies that waiting threads should mostly be active, i.e., consume
processor cycles, while waiting. An OpenMP implementation may, for example, make
waiting threads spin.
3
4
5
10
11
14
The PASSIVE value specifies that waiting threads should mostly be passive, i.e., not
consume processor cycles, while waiting. An OpenMP implementation, may for
example, make waiting threads yield the processor to other threads or go to sleep.
15
The details of the ACTIVE and PASSIVE behaviors are implementation defined.
16
Examples:
12
13
17
setenv
setenv
setenv
setenv
18
19
20
21
OMP_WAIT_POLICY
OMP_WAIT_POLICY
OMP_WAIT_POLICY
OMP_WAIT_POLICY
ACTIVE
active
PASSIVE
passive
22
Cross References
23
• wait-policy-var ICV, see Section 2.3 on page 24.
24
25
4.7
26
27
28
29 150
OMP_MAX_ACTIVE_LEVELS
The OMP_MAX_ACTIVE_LEVELS environment variable controls the maximum number
of nested active parallel regions by setting the initial value of the max-active-levels-var
ICV.
OpenMP API • Version 3.0 May 2008
5
The value of this environment variable must be a non-negative integer. The behavior of
the program is implementation defined if the requested value of
OMP_MAX_ACTIVE_LEVELS is greater than the maximum number of nested active
parallel levels an implementation can support, or if the value is not a non-negative
integer.
6
Cross References
7
• max-active-levels-var ICV, see Section 2.3 on page 28.
8
• omp_set_max_active_levels routine, see Section 3.2.14 on page 126.
9
• omp_get_max_active_levels routine, see Section 3.2.15 on page 127.
1
2
3
4
10
11
12
13
4.8
OMP_THREAD_LIMIT
The OMP_THREAD_LIMIT environment variable sets the number of OpenMP threads
to use for the whole OpenMP program by setting the thread-limit-var ICV.
17
The value of this environment variable must be a positive integer. The behavior of the
program is implementation defined if the requested value of OMP_THREAD_LIMIT is
greater than the number of threads an implementation can support, or if the value is not
a positive integer.
18
Cross References
19
• thread-limit-var ICV, see Section 2.3 on page 28.
20
• omp_get_thread_limit routine
14
15
16
21
Chapter 4
Environment Variables
151
1
152
OpenMP API • Version 3.0 May 2008
1
2
APPENDIX
A
Examples
3
4
The following are examples of the constructs and routines defined in this document.
5
6
C/C++
A statement following a directive is compound only when necessary, and a noncompound statement is indented with respect to a directive preceding it.
7
8
C/C++
9
10
11
12
13
A.1
A Simple Parallel Loop
The following example demonstrates how to parallelize a simple loop using the parallel
loop construct (Section 2.6.1 on page 54). The loop iteration variable is private by
default, so it is not necessary to specify it explicitly in a private clause.
14
C/C++
15
Example A.1.1c
16
17
18
void a1(int n, float *a, float *b)
{
int i;
19
20
21
22
#pragma omp parallel for
for (i=1; i<n; i++) /* i is private by default */
b[i] = (a[i] + a[i-1]) / 2.0;
}
C/C++
23
153
1
Fortran
Example A.1.1f
2
3
SUBROUTINE A1(N, A, B)
4
5
INTEGER I, N
REAL B(N), A(N)
6
7
8
9
10
!$OMP PARALLEL DO !I is private by default
DO I=2,N
B(I) = (A(I) + A(I-1)) / 2.0
ENDDO
!$OMP END PARALLEL DO
11
END SUBROUTINE A1
12
Fortran
13
14
A.2
15
16
17
18
19
20
21 154
The OpenMP Memory Model
In the following example, at Print 1, the value of x could be either 2 or 5, depending on
the timing of the threads, and the implementation of the assignment to x. There are two
reasons that the value at Print 1 might not be 5. First, Print 1 might be executed before
the assignment to x is executed. Second, even if Print 1 is executed after the assignment,
the value 5 is not guaranteed to be seen by thread 1 because a flush may not have been
executed by thread 0 since the assignment.
OpenMP API • Version 3.0 May 2008
1
2
3
The barrier after Print 1 contains implicit flushes on all threads, as well as a thread
synchronization, so the programmer is guaranteed that the value 5 will be printed by
both Print 2 and Print 3.
4
C/C++
5
Example A.2.1c
6
7
#include <stdio.h>
#include <omp.h>
8
9
10
11
12
13
int main(){
int x;
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
x = 2;
#pragma omp parallel num_threads(2) shared(x)
{
if (omp_get_thread_num() == 0) {
x = 5;
} else {
/* Print 1: the following read of x has a race */
printf("1: Thread# %d: x = %d\n", omp_get_thread_num(),x );
}
#pragma omp barrier
if (omp_get_thread_num() == 0) {
/* Print 2 */
printf("2: Thread# %d: x = %d\n", omp_get_thread_num(),x );
} else {
/* Print 3 */
printf("3: Thread# %d: x = %d\n", omp_get_thread_num(),x );
}
}
return 0;
}
C/C++
33
Appendix A
Examples
155
1
Fortran
2
Example A.2.1f
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
PROGRAM A2
INCLUDE "omp_lib.h"
INTEGER X
28
END PROGRAM A2
! or USE OMP_LIB
X = 2
!$OMP PARALLEL NUM_THREADS(2) SHARED(X)
IF (OMP_GET_THREAD_NUM() .EQ. 0) THEN
X = 5
ELSE
! PRINT 1: The following read of x has a race
PRINT *,"1: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X
ENDIF
!$OMP BARRIER
IF (OMP_GET_THREAD_NUM() .EQ. 0) THEN
! PRINT 2
PRINT *,"2: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X
ELSE
! PRINT 3
PRINT *,"3: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X
ENDIF
!$OMP END PARALLEL
29
Fortran
30
The following example demonstrates why synchronization is difficult to perform
correctly through variables. The value of flag is undefined in both prints on thread 1 and
the value of data is only well-defined in the second print.
31
32
33 156
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.2.2c
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
#include <omp.h>
#include <stdio.h>
int main()
{
int data;
int flag=0;
#pragma omp parallel
{
if (omp_get_thread_num()==0)
{
/* Write to the data buffer that will be
read by thread */
data = 42;
/* Flush data to thread 1 and strictly order
the write to data
relative to the write to the flag */
#pragma omp flush(flag, data)
/* Set flag to release thread 1 */
flag = 1;
/* Flush flag to ensure that thread 1 sees
the change */
#pragma omp flush(flag)
}
else if(omp_get_thread_num()==1)
{
/* Loop until we see the update to the flag */
#pragma omp flush(flag, data)
while (flag < 1)
{
#pragma omp flush(flag, data)
}
/* Values of flag and data are undefined */
printf("flag=%d data=%d\n", flag, data);
#pragma omp flush(flag, data)
/* Values data will be 42, value of flag
still undefined */
printf("flag=%d data=%d\n", flag, data);
}
}
}
C/C++
43
Appendix A
Examples
157
1
Fortran
2
Example A.2.2f
3
4
5
6
PROGRAM
INCLUDE
INTEGER
INTEGER
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
!$OMP PARALLEL
IF(OMP_GET_THREAD_NUM() .EQ. 0) THEN
! Write to the data buffer that will be read by thread 1
DATA = 42
! Flush DATA to thread 1 and strictly order the write to DATA
! relative to the write to the FLAG
!$OMP FLUSH(FLAG, DATA)
! Set FLAG to release thread 1
FLAG = 1;
! Flush FLAG to ensure that thread 1 sees the change */
!$OMP FLUSH(FLAG)
ELSE IF(OMP_GET_THREAD_NUM() .EQ. 1) THEN
! Loop until we see the update to the FLAG
!$OMP FLUSH(FLAG, DATA)
DO WHILE(FLAG .LT. 1)
!$OMP FLUSH(FLAG, DATA)
ENDDO
EXAMPLE
"omp_lib.h" ! or USE OMP_LIB
DATA
FLAG
24
25
26
! Values of FLAG and DATA are undefined
PRINT *, 'FLAG=', FLAG, ' DATA=', DATA
!$OMP FLUSH(FLAG, DATA)
27
28
29
30
31
!Values DATA will be 42, value of FLAG still undefined */
PRINT *, 'FLAG=', FLAG, ' DATA=', DATA
ENDIF
!$OMP END PARALLEL
END
32
Fortran
33
This example demonstrates why synchronization is difficult to perform correctly through
variables. The statements on thread 1 and thread 2 may execute in either order.
34
35 158
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.2.3c
3
4
5
6
7
#include <omp.h>
#include <stdio.h>
int main()
{
int flag=0;
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
#pragma omp parallel
{
if(omp_get_thread_num()==0)
{
/* Set flag to release thread 1 */
#pragma omp atomic
flag++;
/* Flush of flag is implied by the atomic directive */
}
else if(omp_get_thread_num()==1)
{
/* Loop until we see that flag reaches 1*/
#pragma omp flush(flag)
while(flag < 1)
{
#pragma omp flush(flag)
}
printf("Thread 1 awoken\n");
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
/* Set flag to release thread 2 */
#pragma omp atomic
flag++;
/* Flush of flag is implied by the atomic directive */
}
else if(omp_get_thread_num()==2)
{
/* Loop until we see that flag reaches 2 */
#pragma omp flush(flag)
while(flag < 2)
{
#pragma omp flush(flag)
}
printf("Thread 2 awoken\n");
}
}
}
C/C++
43
Appendix A
Examples
159
1
Fortran
2
Example A.2.3f
3
4
5
PROGRAM EXAMPLE
INCLUDE "omp_lib.h" ! or USE OMP_LIB
INTEGER FLAG
6
7
8
9
10
11
12
13
14
15
16
17
!$OMP PARALLEL
IF(OMP_GET_THREAD_NUM() .EQ. 0) THEN
! Set flag to release thread 1
!$OMP ATOMIC
FLAG = FLAG + 1
!Flush of FLAG is implied by the atomic directive
ELSE IF(OMP_GET_THREAD_NUM() .EQ. 1) THEN
! Loop until we see that FLAG reaches 1
!$OMP FLUSH(FLAG, DATA)
DO WHILE(FLAG .LT. 1)
!$OMP FLUSH(FLAG, DATA)
ENDDO
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
PRINT *, 'Thread 1 awoken'
! Set FLAG to release thread 2
!$OMP ATOMIC
FLAG = FLAG + 1
!Flush of FLAG is implied by the atomic directive
ELSE IF(OMP_GET_THREAD_NUM() .EQ. 2) THEN
! Loop until we see that FLAG reaches 2
!$OMP FLUSH(FLAG, DATA)
DO WHILE(FLAG .LT. 2)
!$OMP FLUSH(FLAG,
DATA)
ENDDO
PRINT *, 'Thread 2 awoken'
ENDIF
!$OMP END PARALLEL
END
33
34 160
Fortran
OpenMP API • Version 3.0 May 2008
1
2
A.3
Conditional Compilation
3
C/C++
4
5
6
The following example illustrates the use of conditional compilation using the OpenMP
macro _OPENMP (Section 2.2 on page 26). With OpenMP compilation, the _OPENMP
macro becomes defined.
7
C/C++
8
Example A.3.1c
9
#include <stdio.h>
10
11
int main()
{
12
13
14
# ifdef _OPENMP
printf("Compiled by an OpenMP-compliant implementation.\n");
# endif
15
16
}
return 0;
C/C++
17
Fortran
18
19
20
21
The following example illustrates the use of the conditional compilation sentinel (see
Section 2.2 on page 26). With OpenMP compilation, the conditional compilation
sentinel !$ is recognized and treated as two spaces. In fixed form source, statements
guarded by the sentinel must start after column 6.
22
Fortran
23
Example A.3.1f
24
PROGRAM A3
25
26
27
28
29
C234567890
!$
PRINT *, "Compiled by an OpenMP-compliant implementation."
END PROGRAM A3
Fortran
Appendix A
Examples
161
1
2
A.4
3
4
5
6
7
8
9
10
11
12
13 162
Internal Control Variables
According to Section 2.3 on page 28, an OpenMP implementation must act as if there
are ICVs that control the behavior of the program. This example illustrates two ICVs,
nthreads-var and max_active-levels-var. The nthreads-var ICV controls the number of
threads requested for encountered parallel regions; there is one copy of this ICV per
task. The max-active-levels-var ICV controls the maximum number of nested active
parallel regions; there is one copy of this ICV for the whole program.
In the following example, the value of the nthreads-var ICV is changed via a call to
omp_set_num_threads. The new value of nthreads-var applies only to the implicit
tasks that execute the parallel region and make the call to omp_set_num_threads.
The max-active-levels-var ICV is global; so its value is the same for all tasks.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.4.1c
3
4
#include <stdio.h>
#include <omp.h>
5
6
7
8
9
10
11
12
13
int main (void)
{
omp_set_nested(1);
omp_set_max_active_levels(8);
omp_set_dynamic(0);
omp_set_num_threads(2);
#pragma omp parallel
{
omp_set_num_threads(3);
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
#pragma omp parallel
{
omp_set_num_threads(4);
#pragma omp single
{
/*
* The following should print:
* Inner: max_act_lev=8, num_thds=3, max_thds=4
* Inner: max_act_lev=8, num_thds=3, max_thds=4
*/
printf ("Inner: max_act_lev=%d, num_thds=%d, max_thds=%d\n",
omp_get_max_active_levels(), omp_get_num_threads(),
omp_get_max_threads());
}
}
29
30
31
32
33
34
35
36
37
38
39
40
41
#pragma omp barrier
#pragma omp single
{
/*
* The following should print:
* Outer: max_act_lev=8, num_thds=2, max_thds=3
*/
printf ("Outer: max_act_lev=%d, num_thds=%d, max_thds=%d\n",
omp_get_max_active_levels(), omp_get_num_threads(),
omp_get_max_threads());
}
}
}
C/C++
42
Appendix A
Examples
163
1
Fortran
Example A.4.1f
2
3
4
program icv
use omp_lib
5
6
7
8
call
call
call
call
omp_set_nested(.true.)
omp_set_max_active_levels(8)
omp_set_dynamic(.false.)
omp_set_num_threads(2)
9
10
!$omp parallel
call omp_set_num_threads(3)
11
12
13
14
15
16
17
18
19
20
21
!$omp parallel
call omp_set_num_threads(4)
!$omp single!
!
The following should print:
!
Inner: max_act_lev= 8 , num_thds= 3 , max_thds= 4
!
Inner: max_act_lev= 8 , num_thds= 3 , max_thds= 4
print *, ("Inner: max_act_lev=", omp_get_max_active_levels(),
&
", num_thds=", omp_get_num_threads(),
&
", max_thds=", omp_get_max_threads())
!$omp end single
!$omp end parallel
22
23
24
25
26
27
28
29
30
31
!$omp barrier
!$omp single
!
The following should print:
!
Outer: max_act_lev= 8 , num_thds= 2 , max_thds= 3
print *, ("Outer: max_act_lev=", omp_get_max_active_levels(),
&
", num_thds=", omp_get_num_threads(),
&
", max_thds=", omp_get_max_threads())
!$omp end single
!$omp end parallel
end
32
Fortran
33
34
A.5
35
36
37
38 164
The parallel Construct
The parallel construct (Section 2.4 on page 32) can be used in coarse-grain parallel
programs. In the following example, each thread in the parallel region decides what
part of the global array x to work on, based on the thread number:
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.5.1c
3
#include <omp.h>
4
5
6
void subdomain(float *x, int istart, int ipoints)
{
int i;
7
8
9
}
10
11
12
void sub(float *x, int npoints)
{
int iam, nt, ipoints, istart;
13
14
15
16
17
18
19
20
21
22
23
#pragma omp parallel default(shared) private(iam,nt,ipoints,istart)
{
iam = omp_get_thread_num();
nt = omp_get_num_threads();
ipoints = npoints / nt;
/* size of partition */
istart = iam * ipoints; /* starting array index */
if (iam == nt-1)
/* last thread may do more */
ipoints = npoints - istart;
subdomain(x, istart, ipoints);
}
}
24
25
26
int main()
{
float array[10000];
for (i = 0; i < ipoints; i++)
x[istart+i] = 123.456;
27
28
29
sub(array, 10000);
return 0;
}
C/C++
30
Appendix A
Examples
165
1
Fortran
Example A.5.1f
2
3
4
5
SUBROUTINE SUBDOMAIN(X, ISTART, IPOINTS)
INTEGER ISTART, IPOINTS
REAL X(*)
6
7
8
9
10
INTEGER I
100
DO 100 I=1,IPOINTS
X(ISTART+I) = 123.456
CONTINUE
11
END SUBROUTINE SUBDOMAIN
12
13
14
15
16
17
SUBROUTINE SUB(X, NPOINTS)
INCLUDE "omp_lib.h"
! or USE OMP_LIB
REAL X(*)
INTEGER NPOINTS
INTEGER IAM, NT, IPOINTS, ISTART
18
19
20
21
22
23
24
25
26
27
!$OMP PARALLEL DEFAULT(PRIVATE) SHARED(X,NPOINTS)
28
29
!$OMP END PARALLEL
END SUBROUTINE SUB
IAM = OMP_GET_THREAD_NUM()
NT = OMP_GET_NUM_THREADS()
IPOINTS = NPOINTS/NT
ISTART = IAM * IPOINTS
IF (IAM .EQ. NT-1) THEN
IPOINTS = NPOINTS - ISTART
ENDIF
CALL SUBDOMAIN(X,ISTART,IPOINTS)
30
31
32
33
PROGRAM A5
REAL ARRAY(10000)
CALL SUB(ARRAY, 10000)
END PROGRAM A5
34
Fortran
35
36
A.6
37
38
39 166
The num_threads Clause
The following example demonstrates the num_threads clause (Section 2.4 on page
32). The parallel region is executed with a maximum of 10 threads.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.6.1c
3
4
5
6
#include <omp.h>
int main()
{
omp_set_dynamic(1);
7
8
9
10
11
12
#pragma omp parallel num_threads(10)
{
/* do work here */
}
return 0;
}
C/C++
13
Fortran
Example A.6.1f
14
15
16
17
PROGRAM A6
INCLUDE "omp_lib.h"
! or USE OMP_LIB
CALL OMP_SET_DYNAMIC(.TRUE.)
18
19
20
21
!$OMP
PARALLEL NUM_THREADS(10)
! do work here
!$OMP
END PARALLEL
END PROGRAM A6
22
Fortran
23
Fortran
24
25
26
27
28
29
30
A.7
Fortran Restrictions on the do Construct
If an end do directive follows a do-construct in which several DO statements share a
DO termination statement, then a do directive can only be specified for the first (i.e.
outermost) of these DO statements. For more information, see Section 2.5.1 on page 38.
The following example contains correct usages of loop constructs:
Appendix A
Examples
167
1
Example A.7.1f
2
3
4
SUBROUTINE WORK(I, J)
INTEGER I,J
END SUBROUTINE WORK
5
6
7
SUBROUTINE A7_GOOD()
INTEGER I, J
REAL A(1000)
8
9
10
11
12
!$OMP
100
DO 100 I = 1,10
DO
DO 100 J = 1,10
CALL WORK(I,J)
CONTINUE
! !$OMP ENDDO implied here
13
14
15
16
!$OMP
17
18
19
20
21
22
23
!$OMP
24
25
The following example is non-conforming because the matching do directive for the
end do does not precede the outermost loop:
26
Example A.7.2f
200
!$OMP
DO
DO 200 J = 1,10
A(I) = I + 1
ENDDO
DO
DO 300 I = 1,10
DO 300 J = 1,10
CALL WORK(I,J)
300
CONTINUE
!$OMP
ENDDO
END SUBROUTINE A7_GOOD
27
28
29
SUBROUTINE WORK(I, J)
INTEGER I,J
END SUBROUTINE WORK
30
31
SUBROUTINE A7_WRONG
INTEGER I, J
32
33
34
35
36
37
38
DO 100 I = 1,10
DO
DO 100 J = 1,10
CALL WORK(I,J)
100
CONTINUE
!$OMP
ENDDO
END SUBROUTINE A7_WRONG
!$OMP
39
40 168
Fortran
OpenMP API • Version 3.0 May 2008
1
Fortran
2
3
A.8
Fortran Private Loop Iteration Variables
7
In general loop iteration variables will be private, when used in the do-loop of a do and
parallel do construct or in sequential loops in a parallel construct (see
Section 2.5.1 on page 38 and Section 2.9.1 on page 77). In the following example of a
sequential loop in a parallel construct the loop iteration variable I will be private.
8
Example A.8.1f
9
10
SUBROUTINE A8_1(A,N)
INCLUDE "omp_lib.h"
11
12
REAL A(*)
INTEGER I, MYOFFSET, N
13
14
15
16
17
18
!$OMP PARALLEL PRIVATE(MYOFFSET)
MYOFFSET = OMP_GET_THREAD_NUM()*N
DO I = 1, N
A(MYOFFSET+I) = FLOAT(I)
ENDDO
!$OMP END PARALLEL
19
END SUBROUTINE A8_1
4
5
6
20
! or USE OMP_LIB
Appendix A
Examples
169
2
In exceptional cases, loop iteration variables can be made shared, as in the following
example:
3
Example A.8.2f
4
5
6
SUBROUTINE A8_2(A,B,N,I1,I2)
REAL A(*), B(*)
INTEGER I1, I2, N
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
!$OMP PARALLEL SHARED(A,B,I1,I2)
!$OMP SECTIONS
!$OMP SECTION
DO I1 = I1, N
IF (A(I1).NE.0.0) EXIT
ENDDO
!$OMP SECTION
DO I2 = I2, N
IF (B(I2).NE.0.0) EXIT
ENDDO
!$OMP END SECTIONS
!$OMP SINGLE
IF (I1.LE.N) PRINT *, 'ITEMS IN A UP TO ', I1, 'ARE ALL ZERO.'
IF (I2.LE.N) PRINT *, 'ITEMS IN B UP TO ', I2, 'ARE ALL ZERO.'
!$OMP END SINGLE
!$OMP END PARALLEL
23
END SUBROUTINE A8_2
24
Note however that the use of shared loop iteration variables can easily lead to race
conditions.
1
25
26
Fortran
27
28
A.9
29
30
31
32 170
The nowait clause
If there are multiple independent loops within a parallel region, you can use the
nowait clause (see Section 2.5.1 on page 38) to avoid the implied barrier at the end of
the loop construct, as follows:
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.9.1c
3
#include <math.h>
4
5
6
7
8
9
10
11
12
13
14
15
16
17
void a9(int n, int m, float *a, float *b, float *y, float *z)
{
int i;
#pragma omp parallel
{
#pragma omp for nowait
for (i=1; i<n; i++)
b[i] = (a[i] + a[i-1]) / 2.0;
#pragma omp for nowait
for (i=0; i<m; i++)
y[i] = sqrt(z[i]);
}
}
C/C++
18
Fortran
19
Example A.9.1f
20
SUBROUTINE A9(N, M, A, B, Y, Z)
21
22
INTEGER N, M
REAL A(*), B(*), Y(*), Z(*)
23
INTEGER I
24
!$OMP PARALLEL
25
26
27
28
29
!$OMP DO
30
31
32
33
34
!$OMP DO
35
!$OMP END PARALLEL
36
37
38
DO I=2,N
B(I) = (A(I) + A(I-1)) / 2.0
ENDDO
!$OMP END DO NOWAIT
DO I=1,M
Y(I) = SQRT(Z(I))
ENDDO
!$OMP END DO NOWAIT
END SUBROUTINE A9
Fortran
Appendix A
Examples
171
1
The following examples show the use of nowait with static scheduling.
2
Example A.9.2c
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
#include <math.h>
void a92(int n, float *a, float *b, float *c, float *y, float *z)
{
int i;
#pragma omp parallel
{
#pragma omp for schedule(static) nowait
for (i=0; i<n; i++)
c[i] = (a[i] + b[i]) / 2.0;
#pragma omp for schedule(static) nowait
for (i=0; i<n; i++)
z[i] = sqrt(c[i]);
#pragma omp for schedule(static) nowait
for (i=1; i<=n; i++)
y[i] = z[i-1] + a[i];
}
}
C/C++
C/C++
Fortran
20
Example A.9.2f
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
SUBROUTINE A92(N, A, B, C, Y, Z)
INTEGER N
REAL A(*), B(*), C(*), Y(*), Z(*)
INTEGER I
!$OMP PARALLEL
!$OMP DO SCHEDULE(STATIC)
DO I=1,N
C(I) = (A(I) + B(I)) / 2.0
ENDDO
!$OMP END DO NOWAIT
!$OMP DO SCHEDULE(STATIC)
DO I=1,N
Z(I) = SQRT(C(I))
ENDDO
!$OMP END DO NOWAIT
!$OMP DO SCHEDULE(STATIC)
DO I=2,N+1
Y(I) = Z(I-1) + A(I)
ENDDO
!$OMP END DO NOWAIT
!$OMP END PARALLEL
END SUBROUTINE A92
43
44 172
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
A.10
The collapse clause
In the next example, the loops over k and j are collapsed and their iteration space is
executed by all threads of the current team.
5
Fortran
6
Example A.10.1f
7
8
9
10
11
12
13
14
15
16
17
subroutine sub()
!$omp do collapse(2) private(i,j,k)
do k = kl, ku, ks
do j = jl, ju, js
do i = il, iu, is
call bar(a,i,j,k)
enddo
enddo
enddo
!$omp end do
end subroutine
18
19
20
Fortran
In the next example, the loops over k and j are collapsed and their iteration space is
executed by all threads of the current team. The example prints: 2 3.
21
Fortran
22
Example A.10.2f
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
program test
!$omp parallel
!$omp do private(j,k) collapse(2) lastprivate(jlast, klast)
do k = 1,2
do j = 1,3
jlast=j
klast=k
enddo
enddo
!$omp end do
!$omp single
print *, klast, jlast
!$omp end single
!$omp end parallel
end program test
38
39
Fortran
Appendix A
Examples
173
5
The next example illustrates use of the ordered construct with the collapse
construct. Since both loops are collapsed into one, the ordered construct has to be
inside all loops associated with the do construct. Since an iteration may not execute
more than one ordered region this program would be wrong without the
collapse(2) clause. The code prints
6
011
7
012
8
021
9
122
10
131
11
132
1
2
3
4
12
Fortran
Example A.10.3f
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
program test
include 'omp_lib.h'
parallel num_threads(2)
do collapse(2) ordered private(j,k) schedule(static,3)
do k = 1,3
do j = 1,2
ordered
print *, omp_get_thread_num(), k, j
end ordered
call work(a,j,k)
enddo
enddo
end do
end parallel
end program test
29
Fortran
30
31
A.11
32
33
34
35 174
The parallel sections Construct
In the following example (for Section 2.5.2 on page 47) routines xaxis, yaxis, and zaxis
can be executed concurrently. The first section directive is optional. Note that all
section directives need to appear in the parallel sections construct.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.11.1c
3
4
5
void XAXIS();
void YAXIS();
void ZAXIS();
6
7
8
9
10
11
void a11()
{
#pragma omp parallel sections
{
#pragma omp section
XAXIS();
12
13
14
15
16
17
18
#pragma omp section
YAXIS();
#pragma omp section
ZAXIS();
}
}
C/C++
19
Fortran
20
21
Example A.11.1f
SUBROUTINE A11()
22
23
24
!$OMP PARALLEL SECTIONS
!$OMP SECTION
CALL XAXIS()
25
26
!$OMP SECTION
CALL YAXIS()
27
28
!$OMP SECTION
CALL ZAXIS()
29
!$OMP END PARALLEL SECTIONS
30
31
32
END SUBROUTINE A11
Fortran
Appendix A
Examples
175
1
2
A.12
3
4
5
6
7
8
9
The single Construct
The following example demonstrates the single construct (Section 2.5.3 on page 49).
In the example, only one thread prints each of the progress messages. All other threads
will skip the single region and stop at the barrier at the end of the single construct
until all threads in the team have reached the barrier. If other threads can proceed
without waiting for the thread executing the single region, a nowait clause can be
specified, as is done in the third single construct in this example. The user must not
make any assumptions as to which thread will execute a single region.
10
C/C++
11
Example A.12.1c
12
#include <stdio.h>
13
14
void work1() {}
void work2() {}
15
16
17
18
19
20
void a12()
{
#pragma omp parallel
{
#pragma omp single
printf("Beginning work1.\n");
21
work1();
22
23
#pragma omp single
printf("Finishing work1.\n");
24
25
#pragma omp single nowait
printf("Finished work1 and beginning work2.\n");
26
27
28
work2();
}
}
C/C++
29 176
OpenMP API • Version 3.0 May 2008
1
Fortran
Example A.12.1f
2
3
4
5
6
7
8
9
10
PROGRAM A12
!$OMP PARALLEL
11
12
13
!$OMP SINGLE
print *, "Beginning work1."
!$OMP END SINGLE
SUBROUTINE WORK1()
END SUBROUTINE WORK1
SUBROUTINE WORK2()
END SUBROUTINE WORK2
14
CALL WORK1()
15
16
17
!$OMP SINGLE
print *, "Finishing work1."
!$OMP END SINGLE
18
19
20
!$OMP SINGLE
print *, "Finished work1 and beginning work2."
!$OMP END SINGLE NOWAIT
21
CALL WORK2()
22
!$OMP END PARALLEL
23
END PROGRAM A12
24
25
26
27
28
29
30
31
32
Fortran
A.13
Tasking Constructs
The following example shows how to traverse a tree-like structure using explicit tasks.
Note that the traverse function should be called from within a parallel region for the
different specified tasks to be executed in parallel. Also, note that the tasks will be
executed in no specified order because there are no synchronization directives. Thus,
assuming that the traversal will be done in post order, as in the sequential code, is
wrong.
Appendix A
Examples
177
1
C/C++
2
Example A.13.1c
3
4
5
6
7
8
9
10
11
12
13
14
15
16
struct node {
struct node *left;
struct node *right;
};
extern void process(struct node *);
void traverse( struct node *p ) {
if (p->left)
#pragma omp task
// p is firstprivate by default
traverse(p->left);
if (p->right)
#pragma omp task
// p is firstprivate by default
traverse(p->right);
process(p);
}
C/C++
17
Fortran
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Example A.13.1f
RECURSIVE SUBROUTINE traverse ( P )
TYPE Node
TYPE(Node), POINTER :: left, right
END TYPE Node
TYPE(Node) :: P
IF (associated(P%left)) THEN
!$OMP TASK
! P is firstprivate by default
call traverse(P%left)
!$OMP END TASK
ENDIF
IF (associated(P%right)) THEN
!$OMP TASK
! P is firstprivate by default
call traverse(P%right)
!$OMP END TASK
ENDIF
CALL process ( P )
END SUBROUTINE
36
37 178
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
In the next example, we force a postorder traversal of the tree by adding a taskwait
directive. Now, we can safely assume that the left and right sons have been executed
before we process the current node.
4
C/C++
5
Example A.13.2c
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
struct node {
struct node *left;
struct node *right;
};
extern void process(struct node *);
void postorder_traverse( struct node *p ) {
if (p->left)
#pragma omp task
// p is firstprivate by default
postorder_traverse(p->left);
if (p->right)
#pragma omp task
// p is firstprivate by default
postorder_traverse(p->right);
#pragma omp taskwait
process(p);
}
C/C++
21
Fortran
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
Example A.13.2f
RECURSIVE SUBROUTINE traverse ( P )
TYPE Node
TYPE(Node), POINTER :: left, right
END TYPE Node
TYPE(Node) :: P
IF (associated(P%left)) THEN
!$OMP TASK
! P is firstprivate by default
call traverse(P%left)
!$OMP END TASK
ENDIF
IF (associated(P%right)) THEN
!$OMP TASK
! P is firstprivate by default
call traverse(P%right)
!$OMP END TASK
ENDIF
!$OMP TASKWAIT
CALL process ( P )
END SUBROUTINE
Fortran
Appendix A
Examples
179
1
2
3
The following example demonstrates how to use the task construct to process elements
of a linked list in parallel. The pointer p is firstprivate by default on the task construct
so it is not necessary to specify it in a firstprivate clause.
4
C/C++
5
Example A.13.3c
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
typedef struct node node;
struct node {
int data;
node * next;
};
void process(node * p)
{
/* do work here */
}
void increment_list_items(node * head)
{
#pragma omp parallel
{
#pragma omp single
{
node * p = head;
while (p) {
#pragma omp task
// p is firstprivate by default
process(p);
p = p->next;
}
}
}
}
C/C++
32 180
OpenMP API • Version 3.0 May 2008
1
Fortran
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Example A.13.3f
MODULE LIST
TYPE NODE
INTEGER :: PAYLOAD
TYPE (NODE), POINTER :: NEXT
END TYPE NODE
CONTAINS
SUBROUTINE PROCESS(p)
TYPE (NODE), POINTER :: P
! do work here
END SUBROUTINE
SUBROUTINE INCREMENT_LIST_ITEMS (HEAD)
TYPE (NODE), POINTER :: HEAD
TYPE (NODE), POINTER :: P
!$OMP PARALLEL PRIVATE(P)
!$OMP SINGLE
P => HEAD
DO
!$OMP TASK
! P is firstprivate by default
CALL PROCESS(P)
!$OMP END TASK
P => P%NEXT
IF ( .NOT. ASSOCIATED (P) ) EXIT
END DO
!$OMP END SINGLE
!$OMP END PARALLEL
END SUBROUTINE
END MODULE
Fortran
Appendix A
Examples
181
1
2
3
This example calculates a Fibonacci number. If a call to this function is encountered by
a single thread in a parallel region, a nested task region will be spawned to carry out the
computation in parallel.
4
C/C++
5
Example A.13.4c
6
7
8
9
10
11
12
13
14
15
16
17
18
int fib(int n) {
int i, j;
if (n<2)
return n;
else {
#pragma omp task shared(i)
i=fib(n-1);
#pragma omp task shared(j)
j=fib(n-2);
#pragma omp taskwait
return i+j;
}
}
C/C++
19
Fortran
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Example A.13.4f
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
RECURSIVE INTEGER FUNCTION fib(n)
INTEGER n, i, j
IF ( n .LT. 2) THEN
fib = n
ELSE
TASK SHARED(i)
i = fib( n-1 )
END TASK
TASK SHARED(j)
j = fib( n-2 )
END TASK
END TASKWAIT
fib = i+j
END IF
END FUNCTION
36
Fortran
37
Note: Fibonacci number computation is a classic computer science example for showing
recursion, but it is also a classic example showing that a simple algorithm may not be
efficient. A more efficient method is through the calculation of exponentiation over an
integer matrix. Here we used the classic recursion algorithm for illustrative purposes.
38
39
40
41 182
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
The following example demonstrates a way to generate a large number of tasks with one
thread and execute them with the threads in the parallel team. While generating these
tasks, the implementation may reach its limit on unassigned tasks. If it does, the
implementation is allowed to cause the thread executing the task generating loop to
suspend its task at the task scheduling point in the task directive, and start executing
unassigned tasks. Once the number of unassigned tasks is sufficiently low, the thread
may resume execution of the task generating loop.
8
C/C++
9
Example A.13.5c
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
#define LARGE_NUMBER 10000000
double item[LARGE_NUMBER];
extern void process(double);
int main() {
#pragma omp parallel
{
#pragma omp single
{
int i;
for (i=0; i<LARGE_NUMBER; i++)
#pragma omp task
// i is firstprivate, item is shared
process(item[i]);
}
}
}
C/C++
26
Fortran
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
Example A.13.5f
real*8 item(10000000)
integer i
!$omp parallel
!$omp single ! loop iteration variable i is private
do i=1,10000000
!$omp task
! i is firstprivate, item is shared
call process(item(i))
!$omp end task
end do
!$omp end single
!$omp end parallel
end
Fortran
Appendix A
Examples
183
1
2
3
4
5
6
7
8
9
10
11
The following example is the same as the previous one, except that the tasks are
generated in an untied task. While generating the tasks, the implementation may reach
its limit on unassigned tasks. If it does, the implementation is allowed to cause the
thread executing the task generating loop to suspend its task at the task scheduling point
in the task directive, and start executing unassigned tasks. If that thread begins
execution of a task that takes a long time to complete, the other threads may complete
all the other tasks before it is finished.
In this case, since the loop is in an untied task, any other thread is eligible to resume the
task generating loop. In the previous examples, the other threads would be forced to
wait idly until the generating thread finishes its long task, since the task generating loop
was in a tied task.
12
C/C++
13
Example A.13.6c
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
#define LARGE_NUMBER 10000000
double item[LARGE_NUMBER];
extern void process(double);
int main() {
#pragma omp parallel
{
#pragma omp single
{
int i;
#pragma omp task untied
// i is firstprivate, item is shared
{
for (i=0; i<LARGE_NUMBER; i++)
#pragma omp task
process(item[i]);
}
}
}
}
C/C++
33 184
OpenMP API • Version 3.0 May 2008
1
Fortran
2
Example A.13.6f
3
4
5
6
7
8
9
10
11
12
13
14
15
16
real*8 item(10000000)
!$omp parallel
!$omp single
!$omp task untied
! loop iteration variable i is private
do i=1,10000000
!$omp task ! i is firstprivate, item is shared
call process(item(i))
!$omp end task
end do
!$omp end task
!$omp end single
!$omp end parallel
end
17
Fortran
18
The following two examples demonstrate how the scheduling rules illustrated in
Section 2.7.1 on page 62 affect the usage of threadprivate variables in tasks. The value
of a threadprivate variable will change across task scheduling points if the executing
thread executes a part of another schedulable task that modifies the variable. In tied
tasks, the user can control where task scheduling points appear in the code.
19
20
21
22
23
24
25
26
A single thread may execute both of the task regions that modify tp. The parts of these
task regions in which tp is modified may be executed in any order so the resulting value
of var can be either 1 or 2.
Appendix A
Examples
185
1
C/C++
2
Example A.13.7c
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
int tp;
#pragma omp threadprivate(tp)
int var;
void work()
{
#pragma omp task
{
/* do work here */
#pragma omp task
{
tp = 1;
/* do work here */
#pragma omp task
{
/* no modification of tp */
}
var = tp; //value of tp can be 1 or 2
}
tp = 2;
}
}
C/C++
Fortran
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
Example A.13.7f
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
module example
integer tp
threadprivate(tp)
integer var
contains
subroutine work
use globals
task
! do work here
task
tp = 1
! do work here
task
! no modification of tp
end task
var = tp
! value of var can be 1 or 2
end task
tp = 2
end task
end subroutine
end module
46
47 186
Fortran
OpenMP API • Version 3.0 May 2008
3
In this example, scheduling rules prohibit a single thread from scheduling a new task
that modifies tp while another such task region is suspended. Therefore, the value
written will persist across the task scheduling point.
4
Example A.13.8c
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
#include <omp.h>
int tp;
#pragma omp threadprivate(tp)
int var;
void work()
{
#pragma omp parallel
{
/* do work here */
#pragma omp task
{
tp++;
/* do work here */
#pragma omp task
{
/* do work here but don't modify tp */
}
var = tp; //Value does not change after write above
}
}
}
1
2
C/C++
C/C++
Fortran
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
Example A.13.8f
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
module example
integer tp
threadprivate(tp)
integer var
contains
subroutine work
parallel
! do work here
task
tp = tp + 1
! do work here
task
! do work here but don't modify tp
end task
var = tp
! value does not change after write above
end task
end parallel
end subroutine
end module
Fortran
Appendix A
Examples
187
1
2
3
4
5
6
7
The following two examples demonstrate how the scheduling rules illustrated in
Section 2.7.1 on page 62 affect the usage of locks and critical sections in tasks. If a lock
is held across a task scheduling point, no attempt should be made to acquire the same
lock in any code that may be interleaved. Otherwise, a deadlock is possible.
In the example below, suppose the thread executing task 1 defers task 2. When it
encounters the task scheduling point at task 3, it could suspend task 1 and begin task 2
which will result in a deadlock when it tries to enter critical region 1.
8
C/C++
9
Example A.13.9c
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
void work()
{
#pragma omp task
{ //Task 1
#pragma omp task
{ //Task 2
#pragma omp critical //Critical region 1
{/*do work here */ }
}
#pragma omp critical //Critical Region 2
{
//Capture data for the following task
#pragma omp task
{ /* do work here */ } //Task 3
}
}
}
C/C++
27 188
OpenMP API • Version 3.0 May 2008
1
Fortran
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Example A.13.9f
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
!$omp
module example
contains
subroutine work
task
! Task 1
task
! Task 2
critical
! Critical region 1
! do work here
end critical
end task
critical
! Critical region 2
! Capture data for the following task
task
!Task 3
! do work here
end task
end critical
end task
end subroutine
end module
Fortran
Appendix A
Examples
189
1
2
3
4
In the following example, lock is held across a task scheduling point. However,
according to the scheduling restrictions outlined in Section 2.7.1 on page 62, the
executing thread can't begin executing one of the non-descendant tasks that also acquires
lock before the task region is complete. Therefore, no deadlock is possible.
5
C/C++
6
Example A.13.10c
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
#include <omp.h>
void work() {
omp_lock_t lock;
#pragma omp parallel
{
int i;
#pragma omp for
for (i = 0; i < 100; i++) {
#pragma omp task
{
// lock is shared by default in the task
omp_set_lock(&lock);
// Capture data for the following task
#pragma omp task
// Task Scheduling Point 1
{ /* do work here */ }
omp_unset_lock(&lock);
}
}
}
}
C/C++
28 190
OpenMP API • Version 3.0 May 2008
1
Fortran
2
Example A.13.10f
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
module example
include 'omp_lib.h'
integer (kind=omp_lock_kind) lock
integer i
contains
subroutine work
!$omp parallel
!$omp do
do i=1,100
!$omp task
! Outer task
call omp_set_lock(lock)
! lock is shared by
! default in the task
! Capture data for the following task
!$omp task
! Task Scheduling Point 1
! do work here
!$omp end task
call omp_unset_lock(lock)
!$omp end task
end do
!$omp end parallel
end subroutine
end module
26
Fortran
27
Fortran
28
29
30
31
32
A.14
The workshare Construct
The following are examples of the workshare construct (see Section 2.5.4 on page
51).
Appendix A
Examples
191
1
Fortran (cont.)
4
In the following example, workshare spreads work across the threads executing the
parallel region, and there is a barrier after the last statement. Implementations must
enforce Fortran execution rules inside of the workshare block.
5
Example A.14.1f
2
3
6
7
8
9
10
11
12
13
14
15
SUBROUTINE A14_1(AA, BB, CC, DD, EE, FF, N)
INTEGER N
REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N), EE(N,N), FF(N,N)
!$OMP
!$OMP
!$OMP
!$OMP
16
PARALLEL
WORKSHARE
AA = BB
CC = DD
EE = FF
END WORKSHARE
END PARALLEL
END SUBROUTINE A14_1
19
In the following example, the barrier at the end of the first workshare region is
eliminated with a nowait clause. Threads doing CC = DD immediately begin work on
EE = FF when they are done with CC = DD.
20
Example A.14.2f
17
18
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35 192
SUBROUTINE A14_2(AA, BB, CC, DD, EE, FF, N)
INTEGER N
REAL AA(N,N), BB(N,N), CC(N,N)
REAL DD(N,N), EE(N,N), FF(N,N)
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
PARALLEL
WORKSHARE
AA = BB
CC = DD
END WORKSHARE NOWAIT
WORKSHARE
EE = FF
END WORKSHARE
END PARALLEL
END SUBROUTINE A14_2
OpenMP API • Version 3.0 May 2008
1
Fortran (cont.)
3
The following example shows the use of an atomic directive inside a workshare
construct. The computation of SUM(AA) is workshared, but the update to R is atomic.
4
Example A.14.3f
2
5
6
7
8
9
10
11
12
13
14
15
16
17
18
SUBROUTINE A14_3(AA, BB, CC, DD, N)
INTEGER N
REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N)
REAL R
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
R=0
PARALLEL
WORKSHARE
AA = BB
ATOMIC
R = R + SUM(AA)
CC = DD
END WORKSHARE
END PARALLEL
END SUBROUTINE A14_3
22
Fortran WHERE and FORALL statements are compound statements, made up of a control
part and a statement part. When workshare is applied to one of these compound
statements, both the control and the statement parts are workshared. The following
example shows the use of a WHERE statement in a workshare construct.
23
Each task gets worked on in order by the threads:
24
AA = BB then
25
CC = DD then
26
EE .ne. 0 then
27
FF = 1 / EE then
28
GG = HH
19
20
21
29
Appendix A
Examples
193
1
2
Fortran (cont.)
Example A.14.4f
3
4
5
6
7
SUBROUTINE A14_4(AA, BB, CC, DD, EE, FF, GG, HH, N)
INTEGER N
REAL AA(N,N), BB(N,N), CC(N,N)
REAL DD(N,N), EE(N,N), FF(N,N)
REAL GG(N,N), HH(N,N)
8
9
10
11
12
13
14
15
16
17
!$OMP
!$OMP
18
19
In the following example, an assignment to a shared scalar variable is performed by one
thread in a workshare while all other threads in the team wait.
20
Example A.14.5f
!$OMP
!$OMP
PARALLEL
WORKSHARE
AA = BB
CC = DD
WHERE (EE .ne. 0) FF = 1 / EE
GG = HH
END WORKSHARE
END PARALLEL
END SUBROUTINE A14_4
21
22
23
SUBROUTINE A14_5(AA, BB, CC, DD, N)
INTEGER N
REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N)
24
INTEGER SHR
25
26
27
28
29
30
31
32
33
!$OMP
!$OMP
34
The following example contains an assignment to a private scalar variable, which is
performed by one thread in a workshare while all other threads wait. It is nonconforming because the private scalar variable is undefined after the assignment
statement.
35
36
37
38 194
!$OMP
!$OMP
PARALLEL SHARED(SHR)
WORKSHARE
AA = BB
SHR = 1
CC = DD * SHR
END WORKSHARE
END PARALLEL
END SUBROUTINE A14_5
OpenMP API • Version 3.0 May 2008
Example A.14.6f
1
2
3
4
SUBROUTINE A14_6_WRONG(AA, BB, CC, DD, N)
INTEGER N
REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N)
5
INTEGER PRI
6
7
8
9
10
11
12
!$OMP
!$OMP
!$OMP
!$OMP
13
PARALLEL PRIVATE(PRI)
WORKSHARE
AA = BB
PRI = 1
CC = DD * PRI
END WORKSHARE
END PARALLEL
END SUBROUTINE A14_6_WRONG
17
Fortran execution rules must be enforced inside a workshare construct. In the
following example, the same result is produced in the following program fragment
regardless of whether the code is executed sequentially or inside an OpenMP program
with multiple threads:
18
Example A.14.7f
14
15
16
19
20
21
SUBROUTINE A14_7(AA, BB, CC, N)
INTEGER N
REAL AA(N), BB(N), CC(N)
22
23
24
25
26
27
!$OMP
!$OMP
!$OMP
!$OMP
28
PARALLEL
WORKSHARE
AA(1:50) = BB(11:60)
CC(11:20) = AA(1:10)
END WORKSHARE
END PARALLEL
END SUBROUTINE A14_7
29
Fortran
30
31
32
33
34
35
A.15
The master Construct
The following example demonstrates the master construct (Section 2.8.1 on page 63). In
the example, the master keeps track of how many iterations have been executed and
prints out a progress report. The other threads skip the master region without waiting.
Appendix A
Examples
195
1
C/C++
2
Example A.15.1c
3
#include <stdio.h>
4
extern float average(float,float,float);
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
void a15( float* x, float* xold, int n, float tol )
{
int c, i, toobig;
float error, y;
c = 0;
#pragma omp parallel
{
do{
#pragma omp for private(i)
for( i = 1; i < n-1; ++i ){
xold[i] = x[i];
}
#pragma omp single
{
toobig = 0;
}
#pragma omp for private(i,y,error) reduction(+:toobig)
for( i = 1; i < n-1; ++i ){
y = x[i];
x[i] = average( xold[i-1], x[i], xold[i+1] );
error = y - x[i];
if( error > tol || error < -tol ) ++toobig;
}
#pragma omp master
{
++c;
printf( "iteration %d, toobig=%d\n", c, toobig );
}
}while( toobig > 0 );
}
}
C/C++
36 196
OpenMP API • Version 3.0 May 2008
1
Fortran
Example A.15.1f
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
34
SUBROUTINE A15( X, XOLD, N, TOL )
REAL X(*), XOLD(*), TOL
INTEGER N
INTEGER C, I, TOOBIG
REAL ERROR, Y, AVERAGE
EXTERNAL AVERAGE
C = 0
TOOBIG = 1
PARALLEL
DO WHILE( TOOBIG > 0 )
DO PRIVATE(I)
DO I = 2, N-1
XOLD(I) = X(I)
ENDDO
SINGLE
TOOBIG = 0
END SINGLE
DO PRIVATE(I,Y,ERROR), REDUCTION(+:TOOBIG)
DO I = 2, N-1
Y = X(I)
X(I) = AVERAGE( XOLD(I-1), X(I), XOLD(I+1) )
ERROR = Y-X(I)
IF( ERROR > TOL .OR. ERROR < -TOL ) TOOBIG = TOOBIG+1
ENDDO
MASTER
C = C + 1
PRINT *, 'Iteration ', C, 'TOOBIG=', TOOBIG
END MASTER
ENDDO
END PARALLEL
END SUBROUTINE A15
Fortran
35
36
37
38
39
40
41
42
43
A.16
The critical Construct
The following example includes several critical constructs (Section 2.8.2 on page
65). The example illustrates a queuing model in which a task is dequeued and worked
on. To guard against multiple threads dequeuing the same task, the dequeuing operation
must be in a critical region. Because the two queues in this example are
independent, they are protected by critical constructs with different names, xaxis
and yaxis.
Appendix A
Examples
197
1
C/C++
2
Example A.16.1c
3
4
int dequeue(float *a);
void work(int i, float *a);
5
6
7
void a16(float *x, float *y)
{
int ix_next, iy_next;
8
9
10
11
12
#pragma omp parallel shared(x, y) private(ix_next, iy_next)
{
#pragma omp critical (xaxis)
ix_next = dequeue(x);
work(ix_next, x);
13
14
15
16
17
#pragma omp critical (yaxis)
iy_next = dequeue(y);
work(iy_next, y);
}
}
C/C++
18
Fortran
19
20
21
22
Example A.16.1f
SUBROUTINE A16(X, Y)
REAL X(*), Y(*)
INTEGER IX_NEXT, IY_NEXT
23
!$OMP PARALLEL SHARED(X, Y) PRIVATE(IX_NEXT, IY_NEXT)
24
25
26
27
!$OMP CRITICAL(XAXIS)
CALL DEQUEUE(IX_NEXT, X)
!$OMP END CRITICAL(XAXIS)
CALL WORK(IX_NEXT, X)
28
29
30
31
!$OMP CRITICAL(YAXIS)
CALL DEQUEUE(IY_NEXT,Y)
!$OMP END CRITICAL(YAXIS)
CALL WORK(IY_NEXT, Y)
32
!$OMP END PARALLEL
33
END SUBROUTINE A16
34
35 198
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
A.17
worksharing Constructs Inside a
critical Construct
The following example demonstrates using a worksharing construct inside a critical
construct (see Section 2.8.2 on page 65). This example is conforming because the
single region and the critical region are not closely nested (see Section 2.10 on
page 104).
8
C/C++
9
Example A.17.1c
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
void a17()
{
int i = 1;
#pragma omp parallel sections
{
#pragma omp section
{
#pragma omp critical (name)
{
#pragma omp parallel
{
#pragma omp single
{
i++;
}
}
}
}
}
}
C/C++
30
Appendix A
Examples
199
1
Fortran
Example A.17.1f
2
3
SUBROUTINE A17()
4
5
INTEGER I
I = 1
6
7
8
9
10
11
12
13
14
15
16
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
PARALLEL SECTIONS
SECTION
CRITICAL (NAME)
PARALLEL
SINGLE
I = I + 1
END SINGLE
END PARALLEL
END CRITICAL (NAME)
END PARALLEL SECTIONS
END SUBROUTINE A17
17
Fortran
18
19
A.18
20
21
22
23
24
25
26
27
28
29
30 200
Binding of barrier Regions
The binding rules call for a barrier region to bind to the closest enclosing
parallel region (see Section 2.8.3 on page 66).
In the following example, the call from the main program to sub2 is conforming because
the barrier region (in sub3) binds to the parallel region in sub2. The call from
the main program to sub1 is conforming because the barrier region binds to the
parallel region in subroutine sub2.
The call from the main program to sub3 is conforming because the barrier region
binds to the implicit inactive parallel region enclosing the sequential part. Also note
that the barrier region in sub3 when called from sub2 only synchronizes the team of
threads in the enclosing parallel region and not all the threads created in sub1.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.18.1c
3
void work(int n) {}
4
5
6
7
8
9
void sub3(int n)
{
work(n);
#pragma omp barrier
work(n);
}
10
11
12
13
14
void sub2(int k)
{
#pragma omp parallel shared(k)
sub3(k);
}
15
16
17
18
19
20
21
22
23
24
void sub1(int n)
{
int i;
#pragma omp parallel private(i) shared(n)
{
#pragma omp for
for (i=0; i<n; i++)
sub2(i);
}
}
25
26
27
28
29
30
31
int main()
{
sub1(2);
sub2(2);
sub3(2);
return 0;
}
C/C++
32
Appendix A
Examples
201
1
Fortran
Example A.18.1f
2
3
4
5
SUBROUTINE WORK(N)
INTEGER N
END SUBROUTINE WORK
6
7
8
9
10
11
SUBROUTINE SUB3(N)
INTEGER N
CALL WORK(N)
!$OMP
BARRIER
CALL WORK(N)
END SUBROUTINE SUB3
12
13
14
15
16
17
SUBROUTINE SUB2(K)
INTEGER K
!$OMP
PARALLEL SHARED(K)
CALL SUB3(K)
!$OMP
END PARALLEL
END SUBROUTINE SUB2
18
19
20
21
22
23
24
25
26
27
SUBROUTINE SUB1(N)
INTEGER N
INTEGER I
!$OMP
PARALLEL PRIVATE(I) SHARED(N)
!$OMP
DO
DO I = 1, N
CALL SUB2(I)
END DO
!$OMP
END PARALLEL
END SUBROUTINE SUB1
28
29
30
31
32
PROGRAM A18
CALL SUB1(2)
CALL SUB2(2)
CALL SUB3(2)
END PROGRAM A18
33
Fortran
34
35
A.19
36
37
38 202
The atomic Construct
The following example avoids race conditions (simultaneous updates of an element of x
by multiple threads) by using the atomic construct (Section 2.8.5 on page 69).
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
The advantage of using the atomic construct in this example is that it allows updates
of two different elements of x to occur in parallel. If a critical construct (see
Section 2.8.2 on page 65) were used instead, then all updates to elements of x would be
executed serially (though not in any guaranteed order).
Note that the atomic directive applies only to the statement immediately following it.
As a result, elements of y are not updated atomically in this example.
7
C/C++
8
Example A.19.1c
9
10
11
12
float work1(int i)
{
return 1.0 * i;
}
13
14
15
16
float work2(int i)
{
return 2.0 * i;
}
17
18
19
void a19(float *x, float *y, int *index, int n)
{
int i;
20
21
22
23
24
25
26
}
27
28
29
30
31
32
int main()
{
float x[1000];
float y[10000];
int index[10000];
int i;
33
34
35
36
37
38
39
40
41
#pragma omp parallel for shared(x, y, index, n)
for (i=0; i<n; i++) {
#pragma omp atomic
x[index[i]] += work1(i);
y[i] += work2(i);
}
for (i = 0; i < 10000; i++) {
index[i] = i % 1000;
y[i]=0.0;
}
for (i = 0; i < 1000; i++)
x[i] = 0.0;
a19(x, y, index, 10000);
return 0;
}
C/C++
42
Appendix A
Examples
203
1
Fortran
2
Example A.19.1f
3
4
5
6
7
REAL FUNCTION WORK1(I)
INTEGER I
WORK1 = 1.0 * I
RETURN
END FUNCTION WORK1
8
9
10
11
12
REAL FUNCTION WORK2(I)
INTEGER I
WORK2 = 2.0 * I
RETURN
END FUNCTION WORK2
13
14
15
SUBROUTINE SUBA19(X, Y, INDEX, N)
REAL X(*), Y(*)
INTEGER INDEX(*), N
16
17
18
19
20
21
22
INTEGER I
!$OMP
!$OMP
PARALLEL DO SHARED(X, Y, INDEX, N)
DO I=1,N
ATOMIC
X(INDEX(I)) = X(INDEX(I)) + WORK1(I)
Y(I) = Y(I) + WORK2(I)
ENDDO
23
END SUBROUTINE SUBA19
24
25
26
27
28
29
30
31
32
33
34
35
36
PROGRAM A19
REAL X(1000), Y(10000)
INTEGER INDEX(10000)
INTEGER I
37
38
39
DO I=1,10000
INDEX(I) = MOD(I, 1000) + 1
Y(I) = 0.0
ENDDO
DO I = 1,1000
X(I) = 0.0
ENDDO
CALL SUBA19(X, Y, INDEX, 10000)
END PROGRAM A19
40
41 204
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
A.20
Restrictions on the atomic Construct
The following examples illustrate the restrictions on the atomic construct. For more
information, see Section 2.8.5 on page 69.
5
C/C++
6
7
8
All atomic references to the storage location of each variable that appears on the lefthand side of an atomic assignment statement throughout the program are required to
have a compatible type.
C/C++
9
Fortran
10
11
12
All atomic references to the storage location of each variable that appears on the lefthand side of an atomic assignment statement throughout the program are required to
have the same type and type parameters.
13
14
Fortran
The following are some non-conforming examples:
15
C/C++
16
Example A.20.1c
17
18
19
void a20_1_wrong ()
{
union {int n; float x;} u;
20
21
22
23
#pragma omp parallel
{
#pragma omp atomic
u.n++;
24
25
#pragma omp atomic
u.x += 1.0;
26
27
28
29
/* Incorrect because the atomic constructs reference the same location
through incompatible types */
}
}
C/C++
30
Appendix A
Examples
205
1
Fortran
2
Example A.20.1f
3
4
5
6
7
8
9
10
11
12
13
14
15
SUBROUTINE A20_1_WRONG()
INTEGER:: I
REAL:: R
EQUIVALENCE(I,R)
!$OMP
!$OMP
PARALLEL
ATOMIC
I = I + 1
!$OMP
ATOMIC
R = R + 1.0
! incorrect because I and R reference the same location
! but have different types
!$OMP
END PARALLEL
END SUBROUTINE A20_1_WRONG
16
Fortran
17
C/C++
18
Example A.20.2c
19
20
21
22
23
void a20_2_wrong ()
{
int x;
int *i;
float
*r;
24
25
i = &x;
r = (float *)&x;
26
27
28
29
#pragma omp parallel
{
#pragma omp atomic
*i += 1;
30
31
#pragma omp atomic
*r += 1.0;
32
33
/* Incorrect because the atomic constructs reference the same location
through incompatible types */
34
35
}
}
C/C++
36 206
OpenMP API • Version 3.0 May 2008
1
Fortran
3
The following example is non-conforming because I and R reference the same location
but have different types.
4
Example A.20.2f
2
5
6
7
8
9
10
SUBROUTINE SUB()
COMMON /BLK/ R
REAL R
!$OMP
11
12
13
SUBROUTINE A20_2_WRONG()
COMMON /BLK/ I
INTEGER I
14
!$OMP
15
16
17
18
19
!$OMP
20
ATOMIC
R = R + 1.0
END SUBROUTINE SUB
PARALLEL
ATOMIC
I = I + 1
CALL SUB()
!$OMP
END PARALLEL
END SUBROUTINE A20_2_WRONG
Appendix A
Examples
207
2
Although the following example might work on some implementations, this is also nonconforming:
3
Example A.20.3f
1
4
5
6
7
SUBROUTINE A20_3_WRONG
INTEGER:: I
REAL:: R
EQUIVALENCE(I,R)
8
9
10
11
12
13
!$OMP
!$OMP
14
15
16
17
18
19
!$OMP
!$OMP
PARALLEL
ATOMIC
I = I + 1
! incorrect because I and R reference the same location
! but have different types
!$OMP
END PARALLEL
PARALLEL
ATOMIC
R = R + 1.0
! incorrect because I and R reference the same location
! but have different types
!$OMP
END PARALLEL
20
END SUBROUTINE A20_3_WRONG
21
Fortran
22
23
A.21
24
25
The flush Construct with a List
The following example uses the flush construct (see Section 2.8.6 on page 72) for
point-to-point synchronization of specific variables between pairs of threads:
26
C/C++
27
Example A.21.1c
28
29
#include <omp.h>
#define NUMBER_OF_THREADS 256
30
31
32
int
synch[NUMBER_OF_THREADS];
float work[NUMBER_OF_THREADS];
float result[NUMBER_OF_THREADS];
33
34
35
float fn1(int i)
{
return i*2.0;
36 208
OpenMP API • Version 3.0 May 2008
1
}
2
3
4
5
float fn2(float a, float b)
{
return a + b;
}
6
7
8
int main()
{
int iam, neighbor;
9
10
11
12
#pragma omp parallel private(iam,neighbor) shared(work,synch)
{
iam = omp_get_thread_num();
synch[iam] = 0;
13
14
15
#pragma omp barrier
/*Do computation into my portion of work array */
work[iam] = fn1(iam);
16
17
18
19
/* Announce that I am done with my work. The first flush
* ensures that my work is made visible before synch.
* The second flush ensures that synch is made visible.
*/
20
21
22
#pragma omp flush(work,synch)
synch[iam] = 1;
#pragma omp flush(synch)
23
24
25
26
27
/* Wait for neighbor. The first flush ensures that synch is read
* from memory, rather than from the temporary view of memory.
* The second flush ensures that work is read from memory, and
* is done so after the while loop exits.
*/
28
29
30
31
32
33
34
35
36
37
neighbor = (iam>0 ? iam : omp_get_num_threads()) - 1;
while (synch[neighbor] == 0) {
#pragma omp flush(synch)
}
}
38
/* output result here */
39
40
#pragma omp flush(work,synch)
/* Read neighbor’s values of work array */
result[iam] = fn2(work[neighbor], work[iam]);
return 0;
}
C/C++
41
Appendix A
Examples
209
1
Fortran
2
Example A.21.1f
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
REAL FUNCTION FN1(I)
INTEGER I
FN1 = I * 2.0
RETURN
END FUNCTION FN1
REAL FUNCTION FN2(A, B)
REAL A, B
FN2 = A + B
RETURN
END FUNCTION FN2
PROGRAM A21
INCLUDE "omp_lib.h"
INTEGER ISYNC(256)
REAL
WORK(256)
REAL
RESULT(256)
INTEGER IAM, NEIGHBOR
!$OMP
PARALLEL PRIVATE(IAM, NEIGHBOR) SHARED(WORK, ISYNC)
IAM = OMP_GET_THREAD_NUM() + 1
ISYNC(IAM) = 0
!$OMP BARRIER
C Do computation into my portion of work array
WORK(IAM) = FN1(IAM)
C Announce that I am done with my work.
C The first flush ensures that my work is made visible before
C synch. The second flush ensures that synch is made visible.
!$OMP
FLUSH(WORK,ISYNC)
ISYNC(IAM) = 1
!$OMP
FLUSH(ISYNC)
C Wait until neighbor is done. The first flush ensures that
C synch is read from memory, rather than from the temporary
C view of memory. The second flush ensures that work is read
C from memory, and is done so after the while loop exits.
IF (IAM .EQ. 1) THEN
NEIGHBOR = OMP_GET_NUM_THREADS()
ELSE
NEIGHBOR = IAM - 1
ENDIF
!$OMP
!$OMP
!$OMP
DO WHILE (ISYNC(NEIGHBOR) .EQ. 0)
FLUSH(ISYNC)
END DO
FLUSH(WORK, ISYNC)
RESULT(IAM) = FN2(WORK(NEIGHBOR), WORK(IAM))
END PARALLEL
END PROGRAM A21
48
49 210
! or USE OMP_LIB
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
A.22
The flush Construct without a List
The following example (for Section 2.8.6 on page 72) distinguishes the shared variables
affected by a flush construct with no list from the shared objects that are not affected:
5
C/C++
6
Example A.22.1c
7
int x, *p = &x;
8
9
10
11
12
13
14
15
void f1(int *q)
{
*q = 1;
#pragma omp flush
/* x, p, and *q are flushed */
/* because they are shared and accessible */
/* q is not flushed because it is not shared. */
}
16
17
18
19
20
void f2(int *q)
{
#pragma omp barrier
*q = 2;
#pragma omp barrier
21
22
23
24
25
}
26
27
28
29
30
31
32
int g(int n)
{
int i = 1, j, sum = 0;
*p = 1;
#pragma omp parallel reduction(+: sum) num_threads(10)
{
f1(&j);
/*
/*
/*
/*
a barrier implies a flush */
x, p, and *q are flushed */
because they are shared and accessible */
q is not flushed because it is not shared. */
33
34
35
36
/* i, n and sum were not flushed */
/* because they were not accessible in f1 */
/* j was flushed because it was accessible */
sum += j;
37
f2(&j);
38
39
40
/* i, n, and sum were not flushed */
/* because they were not accessible in f2 */
/* j was flushed because it was accessible */
41
Appendix A
Examples
211
1
2
3
4
}
5
6
7
8
9
int main()
{
int result = g(7);
return result;
}
sum += i + j + *p + n;
}
return sum;
C/C++
10
Fortran
11
Example A.22.1f
12
13
14
15
16
17
18
19
20
21
SUBROUTINE F1(Q)
COMMON /DATA/ X, P
INTEGER, TARGET :: X
INTEGER, POINTER :: P
INTEGER Q
!$OMP
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43 212
Q = 1
FLUSH
! X, P and Q are flushed
! because they are shared and accessible
END SUBROUTINE F1
SUBROUTINE F2(Q)
COMMON /DATA/ X, P
INTEGER, TARGET :: X
INTEGER, POINTER :: P
INTEGER Q
!$OMP
BARRIER
Q = 2
!$OMP
BARRIER
! a barrier implies a flush
! X, P and Q are flushed
! because they are shared and accessible
END SUBROUTINE F2
INTEGER FUNCTION G(N)
COMMON /DATA/ X, P
INTEGER, TARGET :: X
INTEGER, POINTER :: P
INTEGER N
INTEGER I, J, SUM
I = 1
SUM = 0
P = 1
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
9
10
11
12
!$OMP
!$OMP
PARALLEL REDUCTION(+: SUM) NUM_THREADS(10)
CALL F1(J)
! I, N and SUM were not flushed
!
because they were not accessible in F1
! J was flushed because it was accessible
SUM = SUM + J
CALL F2(J)
! I, N, and SUM were not flushed
!
because they were not accessible in f2
! J was flushed because it was accessible
SUM = SUM + I + J + P + N
END PARALLEL
13
14
G = SUM
END FUNCTION G
15
16
17
18
19
PROGRAM A22
COMMON /DATA/ X, P
INTEGER, TARGET :: X
INTEGER, POINTER :: P
INTEGER RESULT, G
20
21
22
23
P => X
RESULT = G(7)
PRINT *, RESULT
END PROGRAM A22
24
25
Fortran
Appendix A
Examples
213
1
C/C++
2
A.23
3
Placement of flush, barrier, and
taskwait Directives
6
The following example is non-conforming, because the flush, barrier, and
taskwait directives cannot be the immediate substatement of an if statement. See
Section 2.8.3 on page 66, Section 2.8.6 on page 72, and Section 2.8.4 on page 68.
7
Example A.23.1c
8
9
10
void a23_wrong()
{
int a = 1;
11
12
13
14
15
16
#pragma omp parallel
{
if (a != 0)
#pragma omp flush(a)
/* incorrect as flush cannot be immediate substatement
of if statement */
17
18
19
20
if (a != 0)
#pragma omp barrier
/* incorrect as barrier cannot be immediate substatement
of if statement */
21
22
23
24
if (a != 0)
#pragma omp taskwait
/* incorrect as taskwait cannot be immediate substatement
of if statement */
25
26
}
4
5
27
28
29 214
}
The following version of the above example is conforming because the flush,
barrier, and taskwait directives are enclosed in a compound statement.
OpenMP API • Version 3.0 May 2008
1
Example A.23.2c
2
3
4
void a23()
{
int a = 1;
5
6
7
8
9
10
11
12
13
14
15
16
17
#pragma omp
{
if (a !=
#pragma omp
}
if (a !=
#pragma omp
}
if (a !=
#pragma omp
}
}
parallel
0) {
flush(a)
0) {
barrier
0) {
taskwait
}
C/C++
18
19
20
21
22
23
24
A.24
The ordered Clause and the ordered
Construct
Ordered constructs (Section 2.8.7 on page 75) are useful for sequentially ordering the
output from work that is done in parallel. The following program prints out the indices
in sequential order:
Appendix A
Examples
215
1
C/C++
2
Example A.24.1c
3
#include <stdio.h>
4
5
6
7
8
void work(int k)
{
#pragma omp ordered
printf(" %d\n", k);
}
9
10
11
void a24(int lb, int ub, int stride)
{
int i;
12
13
14
15
}
16
17
18
19
20
int main()
{
a24(0, 100, 5);
return 0;
}
#pragma omp parallel for ordered schedule(dynamic)
for (i=lb; i<ub; i+=stride)
work(i);
C/C++
21 216
OpenMP API • Version 3.0 May 2008
1
Fortran
2
3
4
5
6
7
Example A.24.1f
SUBROUTINE WORK(K)
INTEGER k
!$OMP ORDERED
WRITE(*,*) K
!$OMP END ORDERED
8
END SUBROUTINE WORK
9
10
11
SUBROUTINE SUBA24(LB, UB, STRIDE)
INTEGER LB, UB, STRIDE
INTEGER I
12
13
14
15
16
!$OMP PARALLEL DO ORDERED SCHEDULE(DYNAMIC)
DO I=LB,UB,STRIDE
CALL WORK(I)
END DO
!$OMP END PARALLEL DO
17
END SUBROUTINE SUBA24
18
19
20
PROGRAM A24
CALL SUBA24(1,100,5)
END PROGRAM A24
21
Fortran
22
It is possible to have multiple ordered constructs within a loop region with the
ordered clause specified. The first example is non-conforming because all iterations
execute two ordered regions. An iteration of a loop must not execute more than one
ordered region:
23
24
25
26
Appendix A
Examples
217
1
C/C++
2
Example A.24.2c
3
void work(int i) {}
4
5
6
7
8
9
10
11
12
13
14
15
16
void a24_wrong(int n)
{
int i;
#pragma omp for ordered
for (i=0; i<n; i++) {
/* incorrect because an iteration may not execute more than one
ordered region */
#pragma omp ordered
work(i);
#pragma omp ordered
work(i+1);
}
}
C/C++
17
Fortran
18
Example A.24.2f
19
20
21
SUBROUTINE WORK(I)
INTEGER I
END SUBROUTINE WORK
22
23
SUBROUTINE A24_WRONG(N)
INTEGER N
24
25
26
27
28
29
30
31
32
33
34
35
36
INTEGER I
DO ORDERED
DO I = 1, N
! incorrect because an iteration may not execute more than one
! ordered region
!$OMP
ORDERED
CALL WORK(I)
!$OMP
END ORDERED
!$OMP
!$OMP
ORDERED
CALL WORK(I+1)
!$OMP
END ORDERED
END DO
END SUBROUTINE A24_WRONG
37
38 218
Fortran
OpenMP API • Version 3.0 May 2008
1
2
The following is a conforming example with more than one ordered construct. Each
iteration will execute only one ordered region:
3
C/C++
4
Example A.24.3c
5
6
7
8
void work(int i) {}
void a24_good(int n)
{
int i;
9
10
11
12
13
14
#pragma omp for ordered
for (i=0; i<n; i++) {
if (i <= 10) {
#pragma omp ordered
work(i);
}
15
16
17
18
19
20
if (i > 10) {
#pragma omp ordered
work(i+1);
}
}
}
C/C++
21
Fortran
22
Example A.24.3f
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
SUBROUTINE A24_GOOD(N)
INTEGER N
!$OMP
!$OMP
!$OMP
DO ORDERED
DO I = 1,N
IF (I <= 10) THEN
ORDERED
CALL WORK(I)
END ORDERED
ENDIF
IF (I > 10) THEN
ORDERED
CALL WORK(I+1)
!$OMP
END ORDERED
ENDIF
ENDDO
END SUBROUTINE A24_GOOD
!$OMP
Fortran
Appendix A
Examples
219
1
2
A.25
3
4
The threadprivate Directive
The following examples demonstrate how to use the threadprivate directive
(Section 2.9.2 on page 81) to give each thread a separate counter.
5
C/C++
6
Example A.25.1c
7
8
int counter = 0;
#pragma omp threadprivate(counter)
9
10
11
12
13
int increment_counter()
{
counter++;
return(counter);
}
C/C++
14
Fortran
15
Example A.25.1f
16
17
18
INTEGER FUNCTION INCREMENT_COUNTER()
COMMON/A25_COMMON/COUNTER
!$OMP
THREADPRIVATE(/A25_COMMON/)
19
20
21
22
COUNTER = COUNTER +1
INCREMENT_COUNTER = COUNTER
RETURN
END FUNCTION INCREMENT_COUNTER
23
Fortran
24
C/C++
25
The following example uses threadprivate on a static variable:
26
Example A.25.2c
27
28
29
30
31
32
33
int increment_counter_2()
{
static int counter = 0;
#pragma omp threadprivate(counter)
counter++;
return(counter);
}
34 220
OpenMP API • Version 3.0 May 2008
3
The following example illustrates how modifying a variable that appears in an initializer
can cause unspecified behavior, and also how to avoid this problem by using an
auxiliary variable and a copy-constructor.
4
Example A.25.3c
5
6
7
8
9
10
class T {
public:
int val;
T (int);
T (const T&);
};
11
12
13
T :: T (int v){
val = v;
}
14
15
16
T :: T (const T& t) {
val = t.val;
}
17
18
19
void g(T a, T b){
a.val += b.val;
}
20
21
22
23
24
int x = 1;
T a(x);
const T b_aux(x); /* Capture value of x = 1 */
T b(b_aux);
#pragma omp threadprivate(a, b)
25
26
27
28
29
30
31
void f(int n) {
x++;
#pragma omp parallel for
/* In each thread:
* a is constructed from x (with value 1 or 2?)
* b is copy-constructed from b_aux
*/
1
2
32
33
34
35
for (int i=0; i<n; i++) {
g(a, b); /* Value of a is unspecified. */
}
}
C/C++
36
Appendix A
Examples
221
1
Fortran
2
3
4
The following examples show non-conforming uses and correct uses of the
threadprivate directive. For more information, see Section 2.9.2 on page 81 and
Section 2.9.4.1 on page 101.
6
The following example is non-conforming because the common block is not declared
local to the subroutine that refers to it:
7
Example A.25.2f
5
8
9
10
11
12
13
14
15
16
MODULE A25_MODULE
COMMON /T/ A
END MODULE A25_MODULE
SUBROUTINE A25_2_WRONG()
USE A25_MODULE
!$OMP
THREADPRIVATE(/T/)
!non-conforming because /T/ not declared in A25_4_WRONG
END SUBROUTINE A25_2_WRONG
18
The following example is also non-conforming because the common block is not
declared local to the subroutine that refers to it:
19
Example A.25.3f
20
21
22
SUBROUTINE A25_3_WRONG()
COMMON /T/ A
!$OMP
THREADPRIVATE(/T/)
23
24
25
26
27
28
29
CONTAINS
SUBROUTINE A25_3S_WRONG()
!$OMP
PARALLEL COPYIN(/T/)
!non-conforming because /T/ not declared in A35_5S_WRONG
!$OMP
END PARALLEL
END SUBROUTINE A25_3S_WRONG
END SUBROUTINE A25_3_WRONG
17
30 222
OpenMP API • Version 3.0 May 2008
1
Fortran (cont.)
2
The following example is a correct rewrite of the previous example:
3
Example A.25.4f
4
5
6
7
8
9
10
11
12
13
14
15
!$OMP
!$OMP
!$OMP
!$OMP
SUBROUTINE A25_4_GOOD()
COMMON /T/ A
THREADPRIVATE(/T/)
CONTAINS
SUBROUTINE A25_4S_GOOD()
COMMON /T/ A
THREADPRIVATE(/T/)
PARALLEL COPYIN(/T/)
END PARALLEL
END SUBROUTINE A25_4S_GOOD
END SUBROUTINE A25_4_GOOD
Appendix A
Examples
223
1
Fortran (cont.)
2
The following is an example of the use of threadprivate for local variables:
3
Example A.25.5f
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
PROGRAM A25_5_GOOD
INTEGER, ALLOCATABLE, SAVE :: A(:)
INTEGER, POINTER, SAVE :: PTR
INTEGER, SAVE :: I
INTEGER, TARGET :: TARG
LOGICAL :: FIRSTIN = .TRUE.
!$OMP
THREADPRIVATE(A, I, PTR)
ALLOCATE (A(3))
A = (/1,2,3/)
PTR => TARG
I = 5
!$OMP
!$OMP
PARALLEL COPYIN(I, PTR)
CRITICAL
IF (FIRSTIN) THEN
TARG = 4
! Update target of ptr
I = I + 10
IF (ALLOCATED(A)) A = A + 10
FIRSTIN = .FALSE.
END IF
IF (ALLOCATED(A)) THEN
PRINT *, 'a = ', A
ELSE
PRINT *, 'A is not allocated'
END IF
31
32
33
PRINT *, 'ptr = ', PTR
PRINT *, 'i = ', I
PRINT *
34
35
36
!$OMP
!$OMP
37
38
The above program, if executed by two threads, will print one of the following two sets
of output:
39
40
41
a = 11 12 13
ptr = 4
i = 15
42
A is not allocated
43 224
END CRITICAL
END PARALLEL
END PROGRAM A25_5_GOOD
OpenMP API • Version 3.0 May 2008
1
2
ptr = 4
i = 5
3
or
4
5
6
A is not allocated
ptr = 4
i = 15
7
8
9
a = 1 2 3
ptr = 4
i = 5
10
The following is an example of the use of threadprivate for module variables:
11
Example A.25.6f
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
MODULE A25_MODULE6
REAL, POINTER :: WORK(:)
SAVE WORK
!$OMP
THREADPRIVATE(WORK)
END MODULE A25_MODULE6
37
SUBROUTINE SUB1(N)
USE A25_MODULE6
!$OMP
PARALLEL PRIVATE(THE_SUM)
ALLOCATE(WORK(N))
CALL SUB2(THE_SUM)
WRITE(*,*)THE_SUM
!$OMP
END PARALLEL
END SUBROUTINE SUB1
SUBROUTINE SUB2(THE_SUM)
USE A25_MODULE6
WORK(:) = 10
THE_SUM=SUM(WORK)
END SUBROUTINE SUB2
PROGRAM A25_6_GOOD
N = 10
CALL SUB1(N)
END PROGRAM A25_6_GOOD
Fortran
38
C/C++
39
40
41
42
The following example illustrates initialization of threadprivate variables for class-type
T. t1 is default constructed, t2 is constructed taking a constructor accepting one
argument of integer type, t3 is copy constructed with argument f():
Appendix A
Examples
225
1
Example A.25.4c
2
3
4
5
6
7
static T t1;
#pragma omp threadprivate(t1)
static T t2( 23 );
#pragma omp threadprivate(t2)
static T t3 = f();
#pragma omp threadprivate(t3)
8
10
The following example illustrates the use of threadprivate for static class members. The
threadprivate directive for a static class member must be placed inside the class
definition.
11
Example A.25.5c
12
13
14
15
16
class T {
public:
static int i;
#pragma omp threadprivate(i)
};
9
17
C/C++
18
C/C++
19
20
A.26
Parallel Random Access Iterator Loop
21
The following example shows a parallel Random access iterator loop.
22
Example A.26.1c
23
24
25
26
27
28
29
30
31
32
33
#include <vector>
void iterator_example()
{
std::vector<int> vec(23);
std::vector<int>::iterator it;
#pragma omp parallel for default(none) shared(vec)
for (it = vec.begin(); it < vec.end(); it++)
{
// do work with *it //
}
}
C/C++
34 226
OpenMP API • Version 3.0 May 2008
1
Fortran
2
3
4
A.27
Fortran Restrictions on shared and
private Clauses with Common Blocks
8
When a named common block is specified in a private, firstprivate, or
lastprivate clause of a construct, none of its members may be declared in another
data-sharing attribute clause on that construct. The following examples illustrate this
point. For more information, see Section 2.9.3 on page 85.
9
The following example is conforming:
10
Example A.27.1f
5
6
7
11
12
13
SUBROUTINE A27_1_GOOD()
COMMON /C/ X,Y
REAL X, Y
14
15
16
!$OMP
17
18
19
20
!$OMP
21
!$OMP
PARALLEL PRIVATE (/C/)
! do work here
END PARALLEL
PARALLEL SHARED (X,Y)
! do work here
!$OMP
END PARALLEL
END SUBROUTINE A27_1_GOOD
Appendix A
Examples
227
1
Fortran (cont.)
2
The following example is also conforming:
3
Example A.27.2f
4
5
6
SUBROUTINE A27_2_GOOD()
COMMON /C/ X,Y
REAL X, Y
7
INTEGER I
8
9
10
11
12
13
14
15
16
17
18
19
20
21
!$OMP
!$OMP
22
The following example is conforming:
23
Example A.27.3f
!$OMP
!
!$OMP
DO PRIVATE(X)
DO I=1,1000
! do work here
ENDDO
!$OMP
END DO
!$OMP
END PARALLEL
END SUBROUTINE A27_2_GOOD
24
25
SUBROUTINE A27_3_GOOD()
COMMON /C/ X,Y
26
27
28
!$OMP
29
30
31
32
!$OMP
33 228
PARALLEL
DO PRIVATE(/C/)
DO I=1,1000
! do work here
ENDDO
END DO
!$OMP
PARALLEL PRIVATE (/C/)
! do work here
END PARALLEL
PARALLEL SHARED (/C/)
! do work here
!$OMP
END PARALLEL
END SUBROUTINE A27_3_GOOD
OpenMP API • Version 3.0 May 2008
1
The following example is non-conforming because x is a constituent element of c:
2
Example A.27.4f
3
4
5
6
7
8
9
SUBROUTINE A27_4_WRONG()
COMMON /C/ X,Y
! Incorrect because X is a constituent element of C
!$OMP
PARALLEL PRIVATE(/C/), SHARED(X)
! do work here
!$OMP
END PARALLEL
END SUBROUTINE A27_4_WRONG
10
11
The following example is non-conforming because a common block may not be
declared both shared and private:
12
Example A.27.5f
13
14
15
16
17
18
19
SUBROUTINE A27_5_WRONG()
COMMON /C/ X,Y
! Incorrect: common block C cannot be declared both
! shared and private
!$OMP
PARALLEL PRIVATE (/C/), SHARED(/C/)
! do work here
!$OMP
END PARALLEL
20
END SUBROUTINE A27_5_WRONG
21
Fortran
22
23
24
25
26
27
A.28
The default(none) Clause
The following example distinguishes the variables that are affected by the
default(none) clause from those that are not. For more information on the
default clause, see Section 2.9.3.1 on page 86.
Appendix A
Examples
229
1
C/C++
2
Example A.28.1c
3
4
5
#include <omp.h>
int x, y, z[1000];
#pragma omp threadprivate(x)
6
7
8
void a28(int a) {
const int c = 1;
int i = 0;
9
10
11
12
13
14
15
16
17
#pragma omp parallel default(none) private(a) shared(z)
{
int j = omp_get_num_threads();
/* O.K. - j is declared within parallel region */
a = z[j];
/* O.K. - a is listed in private clause */
/*
- z is listed in shared clause */
x = c;
/* O.K. - x is threadprivate */
/*
- c has const-qualified type */
z[i] = y;
/* Error - cannot reference i or y here */
18
19
20
21
22
23
24
25
#pragma omp for firstprivate(y)
for (i=0; i<10 ; i++) {
z[i] = y; /* O.K. - i is the loop iteration variable */
/* - y is listed in firstprivate clause */
}
z[i] = y;
/* Error - cannot reference i or y here */
}
}
C/C++
26 230
OpenMP API • Version 3.0 May 2008
1
Fortran
Example A.28.1f
2
3
4
5
6
SUBROUTINE A28(A)
INCLUDE "omp_lib.h"
! or USE OMP_LIB
INTEGER A
7
8
9
10
11
INTEGER X, Y, Z(1000)
COMMON/BLOCKX/X
COMMON/BLOCKY/Y
COMMON/BLOCKZ/Z
!$OMP THREADPRIVATE(/BLOCKX/)
12
13
INTEGER I, J
i = 1
14
15
16
17
18
19
20
!$OMP
21
22
23
24
25
!$OMP
26
27
28
!$OMP
29
PARALLEL DEFAULT(NONE) PRIVATE(A) SHARED(Z) PRIVATE(J)
J = OMP_GET_NUM_THREADS();
! O.K. - J is listed in PRIVATE clause
A = Z(J) ! O.K. - A is listed in PRIVATE clause
!
- Z is listed in SHARED clause
X = 1
! O.K. - X is THREADPRIVATE
Z(I) = Y ! Error - cannot reference I or Y here
DO firstprivate(y)
DO I = 1,10
Z(I) = Y ! O.K. - I is the loop iteration variable
! Y is listed in FIRSTPRIVATE clause
END DO
Z(I) = Y
! Error - cannot reference I or Y here
END PARALLEL
END SUBROUTINE A28
Fortran
30
Fortran
31
32
33
34
35
36
37
38
A.29
Race Conditions Caused by Implied
Copies of Shared Variables in Fortran
The following example contains a race condition, because the shared variable, which is
an array section, is passed as an actual argument to a routine that has an assumed-size
array as its dummy argument (see Section 2.9.3.2 on page 88). The subroutine call
passing an array section argument may cause the compiler to copy the argument into a
Appendix A
Examples
231
3
temporary location prior to the call and copy from the temporary location into the
original variable when the subroutine returns. This copying would cause races in the
parallel region.
4
Example A.29.1f
5
6
7
8
9
10
SUBROUTINE A29
11
!$OMP PARALLEL SHARED(A) PRIVATE(MYTHREAD)
1
2
INCLUDE "omp_lib.h"
! or USE OMP_LIB
REAL A(20)
INTEGER MYTHREAD
12
13
14
15
16
17
MYTHREAD = OMP_GET_THREAD_NUM()
IF (MYTHREAD .EQ. 0) THEN
CALL SUB(A(1:10)) ! compiler may introduce writes to A(6:10)
ELSE
A(6:10) = 12
ENDIF
18
!$OMP END PARALLEL
19
END SUBROUTINE A29
20
21
22
23
SUBROUTINE SUB(X)
REAL X(*)
X(1:5) = 4
END SUBROUTINE SUB
24
Fortran
25
26
A.30
27
28
29
30
31 232
The private Clause
In the following example, the values of original list items i and j are retained on exit
from the parallel region, while the private list items i and j are modified within the
parallel construct. For more information on the private clause, see
Section 2.9.3.3 on page 89.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.30.1c
3
4
#include <stdio.h>
#include <assert.h>
5
6
7
8
int main()
{
int i, j;
int *ptr_i, *ptr_j;
9
10
i = 1;
j = 2;
11
12
ptr_i = &i;
ptr_j = &j;
13
14
15
16
17
18
#pragma omp parallel private(i) firstprivate(j)
{
i = 3;
j = j + 2;
assert (*ptr_i == 1 && *ptr_j == 2);
}
19
assert(i == 1 && j == 2);
20
21
return 0;
}
C/C++
22
Fortran
23
Example A.30.1f
24
25
PROGRAM A30
INTEGER I, J
26
27
28
29
30
31
32
33
34
35
I = 1
J = 2
!$OMP
!$OMP
PARALLEL PRIVATE(I) FIRSTPRIVATE(J)
I = 3
J = J + 2
END PARALLEL
PRINT *, I, J
END PROGRAM A30
! I .eq. 1 .and. J .eq. 2
Fortran
Appendix A
Examples
233
1
2
3
In the following example, all uses of the variable a within the loop construct in the
routine f refer to a private list item a, while it is unspecified whether references to a in
the routine g are to a private list item or the original list item.
4
C/C++
5
Example A.30.2c
6
int a;
7
8
9
10
11
void g(int k) {
a = k; /* Accessed in the region but outside of the construct;
* therefore unspecified whether original or private list
* item is modified. */
}
12
13
14
15
16
17
18
19
20
void f(int n) {
int a = 0;
#pragma omp parallel for private(a)
for (int i=1; i<n; i++) {
a = i;
g(a*2);
/* Private copy of "a" */
}
}
C/C++
21 234
OpenMP API • Version 3.0 May 2008
1
Fortran
2
Example A.30.2f
3
4
MODULE A30_2
REAL A
5
CONTAINS
6
7
8
9
10
11
SUBROUTINE G(K)
REAL K
A = K ! Accessed in the region but outside of the
! construct; therefore unspecified whether
! original or private list item is modified.
END SUBROUTINE G
12
13
14
15
16
17
18
19
20
21
22
23
SUBROUTINE F(N)
INTEGER N
REAL A
!$OMP
!$OMP
24
INTEGER I
PARALLEL DO PRIVATE(A)
DO I = 1,N
A = I
CALL G(A*2)
ENDDO
END PARALLEL DO
END SUBROUTINE F
END MODULE A30_2
25
Fortran
26
27
28
29
30
31
A.31
Reprivatization
The following example demonstrates the reprivatization of variables (see Section 2.9.3.3
on page 89). Private variables can be marked private again in a nested construct.
They do not have to be shared in the enclosing parallel region.
Appendix A
Examples
235
1
C/C++
2
Example A.31.1c
3
4
5
6
#include <assert.h>
void a31()
{
int i, a;
7
8
9
10
11
12
13
14
15
16
17
#pragma omp parallel private(a)
{
a = 1;
#pragma omp parallel for private(a)
for (i=0; i<10; i++)
{
a = 2;
}
assert(a == 1);
}
}
C/C++
18
Fortran
19
Example A.31.1f
20
21
22
23
24
25
26
27
28
29
30
31
SUBROUTINE A31()
INTEGER I, A
!$OMP
PARALLEL PRIVATE(A)
A = 1
!$OMP
PARALLEL DO PRIVATE(A)
DO I = 1, 10
A = 2
END DO
!$OMP
END PARALLEL DO
PRINT *, A ! Outer A still has value 1
!$OMP
END PARALLEL
END SUBROUTINE A31
32
33 236
Fortran
OpenMP API • Version 3.0 May 2008
1
Fortran
2
3
4
A.32
Fortran Restrictions on Storage
Association with the private Clause
6
The following non-conforming examples illustrate the implications of the private
clause rules with regard to storage association (see Section 2.9.3.3 on page 89).
7
Example A.32.1f
5
8
9
10
11
SUBROUTINE SUB()
COMMON /BLOCK/ X
PRINT *,X
END SUBROUTINE SUB
12
13
14
15
16
17
18
19
PROGRAM A32_1
COMMON /BLOCK/ X
X = 1.0
!$OMP
PARALLEL PRIVATE (X)
X = 2.0
CALL SUB()
!$OMP
END PARALLEL
END PROGRAM A32_1
20
Example A.32.2f
21
22
23
24
25
26
27
28
PROGRAM A32_2
COMMON /BLOCK2/ X
X = 1.0
!$OMP
!$OMP
PARALLEL PRIVATE (X)
X = 2.0
CALL SUB()
END PARALLEL
CONTAINS
29
30
SUBROUTINE SUB()
COMMON /BLOCK2/ Y
31
32
33
PRINT *,X
PRINT *,Y
END SUBROUTINE SUB
34
35
! X is undefined
! X is undefined
! Y is undefined
END PROGRAM A32_2
Appendix A
Examples
237
1
2
3
Fortran (cont.)
Example A.32.3f
4
5
6
PROGRAM A32_3
EQUIVALENCE (X,Y)
X = 1.0
7
8
9
10
11
12
!$OMP
13
Example A.32.4f
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36 238
PARALLEL PRIVATE(X)
PRINT *,Y
Y = 10
PRINT *,X
!$OMP
END PARALLEL
END PROGRAM A32_3
! Y is undefined
! X is undefined
PROGRAM A32_4
INTEGER I, J
INTEGER A(100), B(100)
EQUIVALENCE (A(51), B(1))
!$OMP PARALLEL DO DEFAULT(PRIVATE) PRIVATE(I,J) LASTPRIVATE(A)
DO I=1,100
DO J=1,100
B(J) = J - 1
ENDDO
DO J=1,100
A(J) = J
ENDDO
! B becomes undefined at this point
DO J=1,50
B(J) = B(J) + 1 ! B is undefined
! A becomes undefined at this point
ENDDO
ENDDO
!$OMP END PARALLEL DO
! The LASTPRIVATE write for A has
! undefined results
PRINT *, B
! B is undefined since the LASTPRIVATE
! write of A was not defined
END PROGRAM A32_4
OpenMP API • Version 3.0 May 2008
1
Example A.32.5f
2
3
4
5
6
7
8
9
10
11
12
SUBROUTINE SUB1(X)
DIMENSION X(10)
13
14
PROGRAM A32_5
COMMON /BLOCK5/ A
!
!
!
!
!
This use of X does not conform to the
specification. It would be legal Fortran 90,
but the OpenMP private directive allows the
compiler to break the sequence association that
A had with the rest of the common block.
FORALL (I = 1:10) X(I) = I
END SUBROUTINE SUB1
15
16
DIMENSION B(10)
EQUIVALENCE (A,B(1))
17
18
! the common block has to be at least 10 words
A = 0
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
!$OMP
PARALLEL PRIVATE(/BLOCK5/)
!
!
!
!
!
!$OMP
!$OMP
!$OMP
Without the private clause,
we would be passing a member of a sequence
that is at least ten elements long.
With the private clause, A may no longer be
sequence-associated.
CALL SUB1(A)
MASTER
PRINT *, A
END MASTER
END PARALLEL
END PROGRAM A32_5
Fortran
Appendix A
Examples
239
1
C/C++
2
A.33
3
4
5
6
C/C++ Arrays in a firstprivate Clause
The following example illustrates the size and value of list items of array or pointer type
in a firstprivate clause (Section 2.9.3.4 on page 92). The size of new list items is
based on the type of the corresponding original list item, as determined by the base
language.
7
In this example:
8
• The type of A is array of two arrays of two ints.
9
• The type of B is adjusted to pointer to array of n ints, because it is a function parameter.
10
• The type of C is adjusted to pointer to int, because it is a function parameter.
11
• The type of D is array of two arrays of two ints.
12
• The type of E is array of n arrays of n ints.
13
Note that B and E involve variable length array types.
14
The new items of array type are initialized as if each integer element of the original
array is assigned to the corresponding element of the new array. Those of pointer type
are initialized as if by assignment from the original item to the new item.
15
16
17 240
OpenMP API • Version 3.0 May 2008
1
Example A.33.1c
2
#include <assert.h>
3
int A[2][2] = {1, 2, 3, 4};
4
5
6
7
void f(int n, int B[n][n], int C[])
{
int D[2][2] = {1, 2, 3, 4};
int E[n][n];
8
9
assert(n >= 2);
E[1][1] = 4;
10
11
12
13
14
15
#pragma omp parallel firstprivate(B, C, D, E)
{
assert(sizeof(B) == sizeof(int (*)[n]));
assert(sizeof(C) == sizeof(int*));
assert(sizeof(D) == 4 * sizeof(int));
assert(sizeof(E) == n * n * sizeof(int));
16
17
18
19
20
21
22
}
23
24
25
26
int main() {
f(2, A, A[0]);
return 0;
}
/* Private B and C have values of original B and C. */
assert(&B[1][1] == &A[1][1]);
assert(&C[3] == &A[1][1]);
assert(D[1][1] == 4);
assert(E[1][1] == 4);
}
C/C++
27
28
29
30
31
32
33
A.34
The lastprivate Clause
Correct execution sometimes depends on the value that the last iteration of a loop
assigns to a variable. Such programs must list all such variables in a lastprivate
clause (Section 2.9.3.5 on page 94) so that the values of the variables are the same as
when the loop is executed sequentially.
Appendix A
Examples
241
1
C/C++
2
Example A.34.1c
3
4
5
void a34 (int n, float *a, float *b)
{
int i;
6
7
8
9
10
11
#pragma omp parallel
{
#pragma omp for lastprivate(i)
for (i=0; i<n-1; i++)
a[i] = b[i] + b[i+1];
}
12
13
a[i]=b[i];
/* i == n-1 here */
}
C/C++
14
Fortran
Example A.34.1f
15
16
SUBROUTINE A34(N, A, B)
17
18
19
INTEGER N
REAL A(*), B(*)
INTEGER I
20
21
!$OMP PARALLEL
!$OMP DO LASTPRIVATE(I)
22
23
24
DO I=1,N-1
A(I) = B(I) + B(I+1)
ENDDO
25
!$OMP END PARALLEL
26
A(I) = B(I)
27
! I has the value of N here
END SUBROUTINE A34
28
Fortran
29
30
A.35
31
32
33 242
The reduction Clause
The following example demonstrates the reduction clause (Section 2.9.3.6 on page
96):
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.35.1c
3
4
5
6
void a35_1(float *x, int *y, int n)
{
int i, b;
float a;
7
8
9
10
11
a = 0.0;
b = 0;
#pragma omp parallel for private(i) shared(x, y, n) \
reduction(+:a) reduction(^:b)
for (i=0; i<n; i++) {
12
13
a += x[i];
b ^= y[i];
14
15
}
}
C/C++
16
Fortran
17
Example A.35.1f
18
SUBROUTINE A35_1(A, B, X, Y, N)
19
20
21
22
INTEGER N
REAL X(*), Y(*), A, B
!$OMP PARALLEL DO PRIVATE(I) SHARED(X, N) REDUCTION(+:A)
!$OMP& REDUCTION(MIN:B)
23
DO I=1,N
24
A = A + X(I)
25
B = MIN(B, Y(I))
26
27
28
29
30
31
32
!
!
!
Note that some reductions can be expressed in
other forms. For example, the MIN could be expressed as
IF (B > Y(I)) B = Y(I)
END DO
END SUBROUTINE A35_1
Fortran
Appendix A
Examples
243
1
2
A common implementation of the preceding example is to treat it as if it had been
written as follows:
3
C/C++
4
Example A.35.2c
5
6
7
8
void a35_2(float *x, int *y, int n)
{
int i, b, b_p;
float a, a_p;
9
10
a = 0.0;
b = 0;
11
12
13
14
15
#pragma omp parallel shared(a, b, x, y, n) \
private(a_p, b_p)
{
a_p = 0.0;
b_p = 0;
16
17
#pragma omp for private(i)
for (i=0; i<n; i++) {
18
19
a_p += x[i];
b_p ^= y[i];
20
}
21
22
23
24
25
#pragma omp critical
{
a += a_p;
b ^= b_p;
}
26
27
}
}
C/C++
28 244
OpenMP API • Version 3.0 May 2008
1
Fortran
2
Example A.35.2f
3
4
5
6
7
SUBROUTINE A35_2 (A, B, X, Y, N)
INTEGER N
REAL X(*), Y(*), A, B, A_P, B_P
!$OMP PARALLEL SHARED(X, Y, N, A, B) PRIVATE(A_P, B_P)
8
9
A_P = 0.0
B_P = HUGE(B_P)
10
11
12
13
14
15
!$OMP
16
17
18
19
!$OMP
20
!$OMP END PARALLEL
21
!$OMP
!$OMP
DO PRIVATE(I)
DO I=1,N
A_P = A_P + X(I)
B_P = MIN(B_P, Y(I))
ENDDO
END DO
CRITICAL
A = A + A_P
B = MIN(B, B_P)
END CRITICAL
END SUBROUTINE A35_2
23
The following program is non-conforming because the reduction is on the intrinsic
procedure name MAX but that name has been redefined to be the variable named MAX.
24
Example A.35.3f
22
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
PROGRAM A35_3_WRONG
MAX = HUGE(0)
M = 0
!$OMP PARALLEL DO REDUCTION(MAX: M)
! MAX is no longer the intrinsic so this is non-conforming
DO I = 1, 100
CALL SUB(M,I)
END DO
END PROGRAM A35_3_WRONG
SUBROUTINE SUB(M,I)
M = MAX(M,I)
END SUBROUTINE SUB
Appendix A
Examples
245
2
The following conforming program performs the reduction using the intrinsic procedure
name MAX even though the intrinsic MAX has been renamed to REN.
3
Example A.35.4f
4
5
6
MODULE M
INTRINSIC MAX
END MODULE M
7
8
9
10
11
12
13
14
PROGRAM A35_4
USE M, REN => MAX
N = 0
!$OMP PARALLEL DO REDUCTION(REN: N)
DO I = 1, 100
N = MAX(N,I)
END DO
END PROGRAM A35_4
15
16
The following conforming program performs the reduction using intrinsic procedure
name MAX even though the intrinsic MAX has been renamed to MIN.
17
Example A.35.5f
18
19
20
MODULE MOD
INTRINSIC MAX, MIN
END MODULE MOD
21
22
23
24
PROGRAM A35_5
USE MOD, MIN=>MAX, MAX=>MIN
REAL :: R
R = -HUGE(0.0)
25
26
27
28
29
30
!$OMP PARALLEL DO REDUCTION(MIN: R)
DO I = 1, 1000
R = MIN(R, SIN(REAL(I)))
END DO
PRINT *, R
END PROGRAM A35_5
1
31
32
33
34
35
36 246
! still does MAX
! still does MAX
Fortran
The following example is non-conforming because the initialization (a = 0) of the
original list item "a" is not synchronized with the update of "a" as a result of the
reduction computation in the for loop. Therefore, the example may print an incorrect
value for "a".
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
To avoid this problem, the initialization of the original list item "a" should complete
before any update of "a" as a result of the reduction clause. This can be achieved by
adding an explicit barrier after the assignment a = 0, or by enclosing the assignment a =
0 in a single directive (which has an implied barrier), or by initializing "a" before the
start of the parallel region.
6
C/C++
7
Example A.35.3c
8
#include <stdio.h>
9
10
11
int main (void)
{
int a, i;
12
13
14
15
#pragma omp parallel shared(a) private(i)
{
#pragma omp master
a = 0;
16
// To avoid race conditions, add a barrier here.
17
18
19
20
#pragma omp for reduction(+:a)
for (i = 0; i < 10; i++) {
a += i;
}
21
22
23
24
#pragma omp single
printf ("Sum is %d\n", a);
}
}
C/C++
25
Appendix A
Examples
247
1
Fortran
Example A.35.6f
2
3
INTEGER A, I
4
!$OMP PARALLEL SHARED(A) PRIVATE(I)
5
6
7
!$OMP MASTER
A = 0
!$OMP END MASTER
8
! To avoid race conditions, add a barrier here.
9
10
11
12
!$OMP DO REDUCTION(+:A)
DO I= 0, 9
A = A + I
END DO
13
14
15
!$OMP SINGLE
PRINT *, "Sum is ", A
!$OMP END SINGLE
16
17
!$OMP END PARALLEL
END
18
Fortran
19
20
A.36
21
22
23
24 248
The copyin Clause
The copyin clause (see Section 2.9.4.1 on page 101) is used to initialize threadprivate
data upon entry to a parallel region. The value of the threadprivate variable in the
master thread is copied to the threadprivate variable of each other team member.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.36.1c
3
#include <stdlib.h>
4
5
6
float* work;
int size;
float tol;
7
#pragma omp threadprivate(work,size,tol)
8
9
10
11
12
13
void build()
{
int i;
work = (float*)malloc( sizeof(float)*size );
for( i = 0; i < size; ++i ) work[i] = tol;
}
14
15
16
17
18
19
20
21
22
void a36( float t, int n )
{
tol = t;
size = n;
#pragma omp parallel copyin(tol,size)
{
build();
}
}
23
C/C++
24
Appendix A
Examples
249
1
Fortran
2
Example A.36.1f
3
4
5
6
7
8
MODULE M
REAL, POINTER, SAVE :: WORK(:)
INTEGER :: SIZE
REAL :: TOL
!$OMP
THREADPRIVATE(WORK,SIZE,TOL)
END MODULE M
9
10
11
12
13
14
15
16
17
18
SUBROUTINE A36( T, N )
USE M
REAL :: T
INTEGER :: N
TOL = T
SIZE = N
!$OMP
PARALLEL COPYIN(TOL,SIZE)
CALL BUILD
!$OMP
END PARALLEL
END SUBROUTINE A36
19
20
21
22
23
SUBROUTINE BUILD
USE M
ALLOCATE(WORK(SIZE))
WORK = TOL
END SUBROUTINE BUILD
24
Fortran
25
26
A.37
27
28
29
30
31
32
33
34
35 250
The copyprivate Clause
The copyprivate clause (see Section 2.9.4.2 on page 102) can be used to broadcast
values acquired by a single thread directly to all instances of the private variables in the
other threads. In this example, if the routine is called from the sequential part, its
behavior is not affected by the presence of the directives. If it is called from a
parallel region, then the actual arguments with which a and b are associated must be
private. After the input routine has been executed by one thread, no thread leaves the
construct until the private variables designated by a, b, x, and y in all threads have
become defined with the values read.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.37.1c
3
4
5
#include <stdio.h>
float x, y;
#pragma omp threadprivate(x, y)
6
7
8
9
10
11
void init(float a, float b ) {
#pragma omp single copyprivate(a,b,x,y)
{
scanf("%f %f %f %f", &a, &b, &x, &y);
}
}
C/C++
12
Fortran
13
Example A.37.1f
14
15
16
17
SUBROUTINE INIT(A,B)
REAL A, B
COMMON /XY/ X,Y
!$OMP
THREADPRIVATE (/XY/)
18
19
20
!$OMP
21
22
23
!$OMP
SINGLE
READ (11) A,B,X,Y
END SINGLE COPYPRIVATE (A,B,/XY/)
END SUBROUTINE INIT
Fortran
Appendix A
Examples
251
1
2
3
4
In contrast to the previous example, suppose the input must be performed by a particular
thread, say the master thread. In this case, the copyprivate clause cannot be used to
do the broadcast directly, but it can be used to provide access to a temporary shared
variable.
5
C/C++
6
Example A.37.2c
7
8
#include <stdio.h>
#include <stdlib.h>
9
10
11
float read_next( ) {
float * tmp;
float return_val;
12
13
14
15
#pragma omp single copyprivate(tmp)
{
tmp = (float *) malloc(sizeof(float));
} /* copies the pointer only */
16
17
18
19
#pragma omp master
{
scanf("%f", tmp);
}
20
21
22
#pragma omp barrier
return_val = *tmp;
#pragma omp barrier
23
24
25
26
#pragma omp single nowait
{
free(tmp);
}
27
28
return return_val;
}
C/C++
29 252
OpenMP API • Version 3.0 May 2008
1
Fortran
2
Example A.37.2f
3
4
REAL FUNCTION READ_NEXT()
REAL, POINTER :: TMP
5
6
7
!$OMP
8
9
10
!$OMP
11
12
13
!$OMP
14
15
16
17
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
SINGLE
ALLOCATE (TMP)
END SINGLE COPYPRIVATE (TMP)
MASTER
READ (11) TMP
END MASTER
BARRIER
READ_NEXT = TMP
BARRIER
SINGLE
DEALLOCATE (TMP)
END SINGLE NOWAIT
END FUNCTION READ_NEXT
18
19
20
21
! copies the pointer only
Fortran
Suppose that the number of lock variables required within a parallel region cannot
easily be determined prior to entering it. The copyprivate clause can be used to
provide access to shared lock variables that are allocated within that parallel region.
22
C/C++
23
Example A.37.3c
24
25
26
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
27
28
29
omp_lock_t *new_lock()
{
omp_lock_t *lock_ptr;
30
31
32
33
34
35
36
#pragma omp single copyprivate(lock_ptr)
{
lock_ptr = (omp_lock_t *) malloc(sizeof(omp_lock_t));
omp_init_lock( lock_ptr );
}
return lock_ptr;
}
C/C++
37
Appendix A
Examples
253
1
Fortran
2
Example A.37.3f
3
4
5
FUNCTION NEW_LOCK()
USE OMP_LIB
! or INCLUDE "omp_lib.h"
INTEGER(OMP_LOCK_KIND), POINTER :: NEW_LOCK
6
7
8
9
10
!$OMP
11
12
Note that the effect of the copyprivate clause on a variable with the allocatable
attribute is different than on a variable with the pointer attribute.
13
Example A.37.4f
SINGLE
ALLOCATE(NEW_LOCK)
CALL OMP_INIT_LOCK(NEW_LOCK)
!$OMP
END SINGLE COPYPRIVATE(NEW_LOCK)
END FUNCTION NEW_LOCK
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
SUBROUTINE S(N)
INTEGER N
REAL, DIMENSION(:), ALLOCATABLE :: A
REAL, DIMENSION(:), POINTER :: B
!$OMP
!$OMP
ALLOCATE (A(N))
SINGLE
ALLOCATE (B(N))
READ (11) A,B
END SINGLE COPYPRIVATE(A,B)
! Variable A is private and is
! assigned the same value in each thread
! Variable B is shared
!$OMP
!$OMP
BARRIER
SINGLE
DEALLOCATE (B)
!$OMP
END SINGLE NOWAIT
END SUBROUTINE S
32
33 254
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
A.38
Nested Loop Constructs
The following example of loop construct nesting (see Section 2.10 on page 104) is
conforming because the inner and outer loop regions bind to different parallel
regions:
6
C/C++
7
Example A.38.1c
8
void work(int i, int j) {}
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
void good_nesting(int n)
{
int i, j;
#pragma omp parallel default(shared)
{
#pragma omp for
for (i=0; i<n; i++) {
#pragma omp parallel shared(i, n)
{
#pragma omp for
for (j=0; j < n; j++)
work(i, j);
}
}
}
}
C/C++
25
Appendix A
Examples
255
1
Fortran
2
Example A.38.1f
3
4
5
SUBROUTINE WORK(I, J)
INTEGER I, J
END SUBROUTINE WORK
6
7
SUBROUTINE GOOD_NESTING(N)
INTEGER N
8
9
10
11
12
13
14
15
16
17
18
19
20
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
!$OMP
INTEGER I
PARALLEL DEFAULT(SHARED)
DO
DO I = 1, N
PARALLEL SHARED(I,N)
DO
DO J = 1, N
CALL WORK(I,J)
END DO
END PARALLEL
END DO
END PARALLEL
END SUBROUTINE GOOD_NESTING
21
22 256
Fortran
OpenMP API • Version 3.0 May 2008
1
The following variation of the preceding example is also conforming:
2
C/C++
3
Example A.38.2c
4
void work(int i, int j) {}
5
6
7
8
9
10
11
12
13
14
void work1(int i, int n)
{
int j;
#pragma omp parallel default(shared)
{
#pragma omp for
for (j=0; j<n; j++)
work(i, j);
}
}
15
16
17
18
19
20
21
22
23
24
void good_nesting2(int n)
{
int i;
#pragma omp parallel default(shared)
{
#pragma omp for
for (i=0; i<n; i++)
work1(i, n);
}
}
C/C++
25
Appendix A
Examples
257
1
Fortran
Example A.38.2f
2
3
4
5
SUBROUTINE WORK(I, J)
INTEGER I, J
END SUBROUTINE WORK
6
7
8
9
10
11
12
13
14
SUBROUTINE WORK1(I, N)
INTEGER J
!$OMP PARALLEL DEFAULT(SHARED)
!$OMP DO
DO J = 1, N
CALL WORK(I,J)
END DO
!$OMP END PARALLEL
END SUBROUTINE WORK1
15
16
17
18
19
20
21
22
23
SUBROUTINE GOOD_NESTING2(N)
INTEGER N
!$OMP PARALLEL DEFAULT(SHARED)
!$OMP DO
DO I = 1, N
CALL WORK1(I, N)
END DO
!$OMP END PARALLEL
END SUBROUTINE GOOD_NESTING2
24
Fortran
25
26
A.39
27
28
29 258
Restrictions on Nesting of Regions
The examples in this section illustrate the region nesting rules. For more information on
region nesting, see Section 2.10 on page 104.
OpenMP API • Version 3.0 May 2008
1
2
The following example is non-conforming because the inner and outer loop regions are
closely nested:
3
C/C++
4
Example A.39.1c
5
void work(int i, int j) {}
6
7
8
9
10
11
12
13
14
15
16
17
18
19
void wrong1(int n)
{
#pragma omp parallel default(shared)
{
int i, j;
#pragma omp for
for (i=0; i<n; i++) {
/* incorrect nesting of loop regions */
#pragma omp for
for (j=0; j<n; j++)
work(i, j);
}
}
}
C/C++
20
Fortran
21
Example A.39.1f
22
23
24
SUBROUTINE WORK(I, J)
INTEGER I, J
END SUBROUTINE WORK
25
26
SUBROUTINE WRONG1(N)
INTEGER N
27
28
29
30
31
32
33
34
35
36
37
38
39
INTEGER I,J
PARALLEL DEFAULT(SHARED)
DO
DO I = 1, N
!$OMP
DO
! incorrect nesting of loop regions
DO J = 1, N
CALL WORK(I,J)
END DO
END DO
!$OMP
END PARALLEL
END SUBROUTINE WRONG1
!$OMP
!$OMP
Fortran
Appendix A
Examples
259
1
The following orphaned version of the preceding example is also non-conforming:
2
Example A.39.2c
3
4
5
6
7
8
9
10
11
void work(int i, int j) {}
void work1(int i, int n)
{
int j;
/* incorrect nesting of loop regions */
#pragma omp for
for (j=0; j<n; j++)
work(i, j);
}
12
13
14
15
16
17
18
19
20
21
void wrong2(int n)
{
#pragma omp parallel default(shared)
{
int i;
#pragma omp for
for (i=0; i<n; i++)
work1(i, n);
}
}
C/C++
C/C++
22
Fortran
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
Example A.39.2f
!$OMP
!$OMP
!$OMP
!$OMP
SUBROUTINE WORK1(I,N)
INTEGER I, N
INTEGER J
DO
! incorrect nesting of loop regions
DO J = 1, N
CALL WORK(I,J)
END DO
END SUBROUTINE WORK1
SUBROUTINE WRONG2(N)
INTEGER N
INTEGER I
PARALLEL DEFAULT(SHARED)
DO
DO I = 1, N
CALL WORK1(I,N)
END DO
END PARALLEL
END SUBROUTINE WRONG2
42
43 260
Fortran
OpenMP API • Version 3.0 May 2008
1
2
The following example is non-conforming because the loop and single regions are
closely nested:
3
C/C++
4
Example A.39.3c
5
6
7
8
9
10
11
12
13
14
15
16
17
18
void work(int i, int j) {}
void wrong3(int n)
{
#pragma omp parallel default(shared)
{
int i;
#pragma omp for
for (i=0; i<n; i++) {
/* incorrect nesting of regions */
#pragma omp single
work(i, 0);
}
}
}
C/C++
19
Fortran
20
Example A.39.3f
21
22
23
24
25
26
27
28
29
30
31
32
33
34
SUBROUTINE WRONG3(N)
INTEGER N
INTEGER I
PARALLEL DEFAULT(SHARED)
DO
DO I = 1, N
!$OMP
SINGLE
! incorrect nesting of regions
CALL WORK(I, 1)
!$OMP
END SINGLE
END DO
!$OMP
END PARALLEL
END SUBROUTINE WRONG3
!$OMP
!$OMP
Fortran
Appendix A
Examples
261
1
2
The following example is non-conforming because a barrier region cannot be closely
nested inside a loop region:
3
C/C++
4
Example A.39.4c
5
6
7
void work(int i, int j) {}
void wrong4(int n)
{
8
9
10
11
12
13
14
15
16
17
18
19
#pragma omp parallel default(shared)
{
int i;
#pragma omp for
for (i=0; i<n; i++) {
work(i, 0);
/* incorrect nesting of barrier region in a loop region */
#pragma omp barrier
work(i, 1);
}
}
}
C/C++
20
Fortran
21
Example A.39.4f
22
23
24
25
26
27
28
29
30
31
32
33
34
SUBROUTINE WRONG4(N)
INTEGER N
INTEGER I
PARALLEL DEFAULT(SHARED)
DO
DO I = 1, N
CALL WORK(I, 1)
! incorrect nesting of barrier region in a loop region
!$OMP
BARRIER
CALL WORK(I, 2)
END DO
!$OMP
END PARALLEL
END SUBROUTINE WRONG4
!$OMP
!$OMP
35
36 262
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
The following example is non-conforming because the barrier region cannot be
closely nested inside the critical region. If this were permitted, it would result in
deadlock due to the fact that only one thread at a time can enter the critical region:
4
C/C++
5
Example A.39.5c
6
7
8
9
10
11
12
13
14
15
16
17
18
19
void work(int i, int j) {}
void wrong5(int n)
{
#pragma omp parallel
{
#pragma omp critical
{
work(n, 0);
/* incorrect nesting of barrier region in a critical region */
#pragma omp barrier
work(n, 1);
}
}
}
C/C++
20
Fortran
21
Example A.39.5f
22
23
24
25
26
27
28
29
30
31
32
33
34
SUBROUTINE WRONG5(N)
INTEGER N
!$OMP
!$OMP
PARALLEL DEFAULT(SHARED)
CRITICAL
CALL WORK(N,1)
! incorrect nesting of barrier region in a critical region
!$OMP
BARRIER
CALL WORK(N,2)
!$OMP
END CRITICAL
!$OMP
END PARALLEL
END SUBROUTINE WRONG5
Fortran
Appendix A
Examples
263
1
2
3
The following example is non-conforming because the barrier region cannot be
closely nested inside the single region. If this were permitted, it would result in
deadlock due to the fact that only one thread executes the single region:
4
C/C++
5
Example A.39.6c
6
7
8
9
10
11
12
13
14
15
16
17
18
19
void work(int i, int j) {}
void wrong6(int n)
{
#pragma omp parallel
{
#pragma omp single
{
work(n, 0);
/* incorrect nesting of barrier region in a single region */
#pragma omp barrier
work(n, 1);
}
}
}
C/C++
20
Fortran
21
Example A.39.6f
22
23
24
25
26
27
28
29
30
31
32
SUBROUTINE WRONG6(N)
INTEGER N
!$OMP
!$OMP
PARALLEL DEFAULT(SHARED)
SINGLE
CALL WORK(N,1)
! incorrect nesting of barrier region in a single region
!$OMP
BARRIER
CALL WORK(N,2)
!$OMP
END SINGLE
!$OMP
END PARALLEL
END SUBROUTINE WRONG6
33
34 264
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
A.40
The omp_set_dynamic and
omp_set_num_threads Routines
Some programs rely on a fixed, prespecified number of threads to execute correctly.
Because the default setting for the dynamic adjustment of the number of threads is
implementation defined, such programs can choose to turn off the dynamic threads
capability and set the number of threads explicitly to ensure portability. The following
example shows how to do this using omp_set_dynamic (Section 3.2.7 on page 117),
and omp_set_num_threads (Section 3.2.1 on page 110).
In this example, the program executes correctly only if it is executed by 16 threads. If
the implementation is not capable of supporting 16 threads, the behavior of this example
is implementation defined. Note that the number of threads executing a parallel
region remains constant during the region, regardless of the dynamic threads setting.
The dynamic threads mechanism determines the number of threads to use at the start of
the parallel region and keeps it constant for the duration of the region.
16
C/C++
17
Example A.40.1c
18
19
#include <omp.h>
#include <stdlib.h>
20
void do_by_16(float *x, int iam, int ipoints) {}
21
22
23
void a40(float *x, int npoints)
{
int iam, ipoints;
24
25
omp_set_dynamic(0);
omp_set_num_threads(16);
26
27
28
29
#pragma omp parallel shared(x, npoints) private(iam, ipoints)
{
if (omp_get_num_threads() != 16)
abort();
30
31
32
33
34
iam = omp_get_thread_num();
ipoints = npoints/16;
do_by_16(x, iam, ipoints);
}
}
C/C++
35
Appendix A
Examples
265
1
Fortran
Example A.40.1f
2
3
4
5
6
SUBROUTINE DO_BY_16(X, IAM, IPOINTS)
REAL X(*)
INTEGER IAM, IPOINTS
END SUBROUTINE DO_BY_16
7
SUBROUTINE SUBA40(X, NPOINTS)
8
INCLUDE "omp_lib.h"
9
10
INTEGER NPOINTS
REAL X(NPOINTS)
11
INTEGER IAM, IPOINTS
12
13
CALL OMP_SET_DYNAMIC(.FALSE.)
CALL OMP_SET_NUM_THREADS(16)
14
!$OMP
! or USE OMP_LIB
PARALLEL SHARED(X,NPOINTS) PRIVATE(IAM, IPOINTS)
15
16
17
IF (OMP_GET_NUM_THREADS() .NE. 16) THEN
STOP
ENDIF
18
19
20
IAM = OMP_GET_THREAD_NUM()
IPOINTS = NPOINTS/16
CALL DO_BY_16(X,IAM,IPOINTS)
21
!$OMP
22
END PARALLEL
END SUBROUTINE SUBA40
23
Fortran
24
25
A.41
26
27
28
29
30 266
The omp_get_num_threads Routine
In the following example, the omp_get_num_threads call (see Section 3.2.2 on
page 111) returns 1 in the sequential part of the code, so np will always be equal to 1. To
determine the number of threads that will be deployed for the parallel region, the
call should be inside the parallel region.
OpenMP API • Version 3.0 May 2008
1
C/C++
2
Example A.41.1c
3
4
#include <omp.h>
void work(int i);
5
6
7
void incorrect()
{
int np, i;
8
np = omp_get_num_threads();
9
10
11
12
#pragma omp parallel for schedule(static)
for (i=0; i < np; i++)
work(i);
/* misplaced */
}
C/C++
13
Fortran
14
Example A.41.1f
15
16
17
18
SUBROUTINE WORK(I)
INTEGER I
I = I + 1
END SUBROUTINE WORK
19
20
21
SUBROUTINE INCORRECT()
INCLUDE "omp_lib.h"
INTEGER I, NP
22
23
24
25
26
27
28
29
30
! or USE OMP_LIB
NP = OMP_GET_NUM_THREADS()
!misplaced: will return 1
PARALLEL DO SCHEDULE(STATIC)
DO I = 0, NP-1
CALL WORK(I)
ENDDO
!$OMP
END PARALLEL DO
END SUBROUTINE INCORRECT
!$OMP
Fortran
Appendix A
Examples
267
1
2
The following example shows how to rewrite this program without including a query for
the number of threads:
3
C/C++
4
Example A.41.2c
5
6
#include <omp.h>
void work(int i);
7
8
9
void correct()
{
int i;
10
11
12
13
14
15
#pragma omp parallel private(i)
{
i = omp_get_thread_num();
work(i);
}
}
C/C++
16
Fortran
17
Example A.41.2f
18
19
20
21
SUBROUTINE WORK(I)
INTEGER I
22
23
24
25
26
END SUBROUTINE WORK
27
28
29
30
31
I = I + 1
SUBROUTINE CORRECT()
INCLUDE "omp_lib.h"
INTEGER I
!$OMP
!$OMP
PARALLEL PRIVATE(I)
I = OMP_GET_THREAD_NUM()
CALL WORK(I)
END PARALLEL
END SUBROUTINE CORRECT
32
33 268
! or USE OMP_LIB
Fortran
OpenMP API • Version 3.0 May 2008
1
2
3
4
A.42
The omp_init_lock Routine
The following example demonstrates how to initialize an array of locks in a parallel
region by using omp_init_lock (Section 3.3.1 on page 136).
5
C/C++
6
Example A.42.1c
7
#include <omp.h>
8
9
10
11
omp_lock_t *new_locks()
{
int i;
omp_lock_t *lock = new omp_lock_t[1000];
12
13
14
15
16
17
18
#pragma omp parallel for private(i)
for (i=0; i<1000; i++)
{
omp_init_lock(&lock[i]);
}
return lock;
}
C/C++
19
Fortran
20
Example A.42.1f
21
22
23
FUNCTION NEW_LOCKS()
USE OMP_LIB
! or INCLUDE "omp_lib.h"
INTEGER(OMP_LOCK_KIND), DIMENSION(1000) :: NEW_LOCKS
24
25
26
27
28
29
30
31
32
33
INTEGER I
!$OMP
!$OMP
PARALLEL DO PRIVATE(I)
DO I=1,1000
CALL OMP_INIT_LOCK(NEW_LOCKS(I))
END DO
END PARALLEL DO
END FUNCTION NEW_LOCKS
Fortran
Appendix A
Examples
269
1
2
A.43
3
4
5
6
7
8
9
10
11
12
13
Ownership of Locks
Ownership of locks has changed from OpenMP 2.5 to OpenMP 3.0. In OpenMP 2.5,
locks are owned by threads; so a lock released by the omp_unset_lock routine must
be owned by the same thread executing the routine. In OpenMP 3.0, on the other hand,
locks are owned by task regions; so a lock released by the omp_unset_lock routine
in a task region must be owned by the same task region.
This change in ownership requires extra care when using locks. The following program
is conforming in OpenMP 2.5 because the thread that releases the lock lck in the parallel
region is the same thread that acquired the lock in the sequential part of the program
(master thread of parallel region and the initial thread are the same). However, it is not
conforming in OpenMP 3.0, because the task region that releases the lock lck is different
from the task region that acquires the lock.
14
C/C++
15
Example A.43.1c
16
17
18
#include <stdlib.h>
#include <stdio.h>
#include <omp.h>
19
20
21
22
23
24
25
26
int main()
{
int x;
omp_lock_t lck;
27
28
29
30
31
32
33
#pragma omp parallel shared (x)
{
#pragma omp master
{
x = x + 1;
omp_unset_lock (&lck);
}
34
35
36
37
38
omp_init_lock (&lck);
omp_set_lock (&lck);
x = 0;
/* Some more stuff. */
}
omp_destroy_lock (&lck);
}
C/C++
39 270
OpenMP API • Version 3.0 May 2008
1
Fortran
Example A.43.1f
2
3
4
5
6
program lock
use omp_lib
integer :: x
integer (kind=omp_lock_kind) :: lck
7
8
9
call omp_init_lock (lck)
call omp_set_lock(lck)
x = 0
10
11
12
13
14
15
16
17
!$omp parallel shared (x)
!$omp master
x = x + 1
call omp_unset_lock(lck)
!$omp end master
!
Some more stuff.
!$omp end parallel
18
19
call omp_destroy_lock(lck)
end
20
Fortran
21
22
23
24
25
26
27
A.44
Simple Lock Routines
In the following example (for Section 3.3 on page 134), the lock routines cause the
threads to be idle while waiting for entry to the first critical section, but to do other work
while waiting for entry to the second. The omp_set_lock function blocks, but the
omp_test_lock function does not, allowing the work in skip to be done.
Appendix A
Examples
271
1
C/C++
3
Note that the argument to the lock routines should have type omp_lock_t, and that
there is no need to flush it.
4
Example A.44.1c
5
6
#include <stdio.h>
#include <omp.h>
7
8
void skip(int i) {}
void work(int i) {}
9
10
11
12
int main()
{
omp_lock_t lck;
int id;
2
13
omp_init_lock(&lck);
14
15
16
#pragma omp parallel shared(lck) private(id)
{
id = omp_get_thread_num();
17
18
19
20
omp_set_lock(&lck);
/* only one thread at a time can execute this printf */
printf("My thread id is %d.\n", id);
omp_unset_lock(&lck);
21
22
23
24
while (! omp_test_lock(&lck)) {
skip(id);
/* we do not yet have the lock,
so we must do something else */
}
25
26
work(id);
27
28
29
30
31
32
/* we now have the lock
and can do the work */
omp_unset_lock(&lck);
}
omp_destroy_lock(&lck);
return 0;
}
C/C++
33 272
OpenMP API • Version 3.0 May 2008
1
Fortran
2
Note that there is no need to flush the lock variable.
3
Example A.44.1f
4
5
SUBROUTINE SKIP(ID)
END SUBROUTINE SKIP
6
7
SUBROUTINE WORK(ID)
END SUBROUTINE WORK
8
PROGRAM A44
9
INCLUDE "omp_lib.h"
10
11
INTEGER(OMP_LOCK_KIND) LCK
INTEGER ID
12
CALL OMP_INIT_LOCK(LCK)
13
14
15
16
17
!$OMP
! or USE OMP_LIB
PARALLEL SHARED(LCK) PRIVATE(ID)
ID = OMP_GET_THREAD_NUM()
CALL OMP_SET_LOCK(LCK)
PRINT *, 'My thread id is ', ID
CALL OMP_UNSET_LOCK(LCK)
18
19
20
21
DO WHILE (.NOT. OMP_TEST_LOCK(LCK))
CALL SKIP(ID)
! We do not yet have the lock
! so we must do something else
END DO
22
23
CALL WORK(ID)
24
CALL OMP_UNSET_LOCK( LCK )
25
26
27
28
29
!$OMP
! We now have the lock
! and can do the work
END PARALLEL
CALL OMP_DESTROY_LOCK( LCK )
END PROGRAM A44
Fortran
Appendix A
Examples
273
1
2
A.45
3
4
Nestable Lock Routines
The following example (for Section 3.3 on page 134) demonstrates how a nestable lock
can be used to synchronize updates both to a whole structure and to one of its members.
5
C/C++
6
Example A.45.1c
7
8
9
10
#include <omp.h>
typedef struct {
int a,b;
omp_nest_lock_t lck; } pair;
11
12
13
14
15
16
17
18
19
20
21
22
int work1();
int work2();
int work3();
void incr_a(pair *p, int a)
{
/* Called only from incr_pair, no need to lock. */
p->a += a;
}
void incr_b(pair *p, int b)
{
/* Called both from incr_pair and elsewhere, */
/* so need a nestable lock. */
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
omp_set_nest_lock(&p->lck);
p->b += b;
omp_unset_nest_lock(&p->lck);
}
void incr_pair(pair *p, int a, int b)
{
omp_set_nest_lock(&p->lck);
incr_a(p, a);
incr_b(p, b);
omp_unset_nest_lock(&p->lck);
}
void a45(pair *p)
{
#pragma omp parallel sections
{
#pragma omp section
incr_pair(p, work1(), work2());
#pragma omp section
incr_b(p, work3());
}
}
C/C++
44 274
OpenMP API • Version 3.0 May 2008
1
Fortran
2
Example A.45.1f
3
4
5
6
7
8
9
10
MODULE DATA
USE OMP_LIB, ONLY: OMP_NEST_LOCK_KIND
TYPE LOCKED_PAIR
INTEGER A
INTEGER B
INTEGER (OMP_NEST_LOCK_KIND) LCK
END TYPE
END MODULE DATA
11
12
13
14
15
16
17
SUBROUTINE INCR_A(P, A)
! called only from INCR_PAIR, no need to lock
USE DATA
TYPE(LOCKED_PAIR) :: P
INTEGER A
P%A = P%A + A
END SUBROUTINE INCR_A
18
19
20
21
22
23
24
25
26
27
28
SUBROUTINE INCR_B(P, B)
! called from both INCR_PAIR and elsewhere,
! so we need a nestable lock
USE OMP_LIB
! or INCLUDE "omp_lib.h"
USE DATA
TYPE(LOCKED_PAIR) :: P
INTEGER B
CALL OMP_SET_NEST_LOCK(P%LCK)
P%B = P%B + B
CALL OMP_UNSET_NEST_LOCK(P%LCK)
END SUBROUTINE INCR_B
29
30
31
32
33
34
SUBROUTINE INCR_PAIR(P, A, B)
USE OMP_LIB
! or INCLUDE "omp_lib.h"
USE DATA
TYPE(LOCKED_PAIR) :: P
INTEGER A
INTEGER B
35
36
37
38
39
CALL OMP_SET_NEST_LOCK(P%LCK)
CALL INCR_A(P, A)
CALL INCR_B(P, B)
CALL OMP_UNSET_NEST_LOCK(P%LCK)
END SUBROUTINE INCR_PAIR
40
41
42
43
44
45
SUBROUTINE A45(P)
USE OMP_LIB
! or INCLUDE "omp_lib.h"
USE DATA
TYPE(LOCKED_PAIR) :: P
INTEGER WORK1, WORK2, WORK3
EXTERNAL WORK1, WORK2, WORK3
46
Appendix A
Examples
275
1
!$OMP
PARALLEL SECTIONS
2
3
4
5
6
!$OMP
SECTION
CALL INCR_PAIR(P, WORK1(), WORK2())
SECTION
CALL INCR_B(P, WORK3())
END PARALLEL SECTIONS
!$OMP
!$OMP
7
END SUBROUTINE A45
8
9
Fortran
276
OpenMP API • Version 3.0 May 2008
1
2
3
4
APPENDIX
B
Stubs for Runtime Library
Routines
5
6
7
8
9
10
11
12
13
14
15
16
17
This section provides stubs for the runtime library routines defined in the OpenMP API.
The stubs are provided to enable portability to platforms that do not support the
OpenMP API. On these platforms, OpenMP programs must be linked with a library
containing these stub routines. The stub routines assume that the directives in the
OpenMP program are ignored. As such, they emulate serial semantics.
Note that the lock variable that appears in the lock routines must be accessed
exclusively through these routines. It should not be initialized or otherwise modified in
the user program.
In an actual implementation the lock variable might be used to hold the address of an
allocated memory block, but here it is used to hold an integer value. Users should not
make assumptions about mechanisms used by OpenMP implementations to implement
locks based on the scheme used by the stub procedures.
18
Fortran
19
20
21
22
23
24
Note – In order to be able to compile the Fortran stubs file, the include file
omp_lib.h was split into two files: omp_lib_kinds.h and omp_lib.h and the
omp_lib_kinds.h file included where needed. There is no requirement for the
implementation to provide separate files.
Fortran
277
1
2
B.1
C/C++ Stub Routines
3
4
5
#include <stdio.h>
#include <stdlib.h>
#include "omp.h"
6
7
8
void omp_set_num_threads(int num_threads)
{
}
9
10
11
12
int omp_get_num_threads(void)
{
return 1;
}
13
14
15
16
int omp_get_max_threads(void)
{
return 1;
}
17
18
19
20
int omp_get_thread_num(void)
{
return 0;
}
21
22
23
24
int omp_get_num_procs(void)
{
return 1;
}
25
26
27
28
int omp_in_parallel(void)
{
return 0;
}
29
30
31
void omp_set_dynamic(int dynamic_threads)
{
}
32
33
34
35
int omp_get_dynamic(void)
{
return 0;
}
36
37
38
void omp_set_nested(int nested)
{
}
39 278
OpenMP API • Version 3.0 May 2008
1
2
3
4
int omp_get_nested(void)
{
return 0;
}
5
6
7
void omp_set_schedule(omp_sched_t kind, int modifier)
{
}
8
9
10
11
12
void omp_get_schedule(omp_sched_t *kind, int *modifier)
{
*kind = omp_sched_static;
*modifier = 0;
}
13
14
15
16
int omp_get_thread_limit(void)
{
return 1;
}
17
18
19
void omp_set_max_active_levels(int max_active_levels)
{
}
20
21
22
23
int omp_get_max_active_levels(void)
{
return 0;
}
24
25
26
27
int omp_get_level(void)
{
return 0;
}
28
29
30
31
32
33
34
35
36
37
38
int omp_get_ancestor_thread_num(int level)
{
if (level == 0)
{
return 0;
}
else
{
return -1;
}
}
39
Appendix B
Stubs for Runtime Library Routines
279
1
2
3
4
5
6
7
8
9
10
11
int omp_get_team_size(int level)
{
if (level == 0)
{
return 1;
}
else
{
return -1;
}
}
12
13
14
15
int omp_get_active_level(void)
{
return 0;
}
16
17
18
19
struct __omp_lock
{
int lock;
};
20
enum { UNLOCKED = -1, INIT, LOCKED };
21
22
23
24
25
void omp_init_lock(omp_lock_t *arg)
{
struct __omp_lock *lock = (struct __omp_lock *)arg;
lock->lock = UNLOCKED;
}
26
27
28
29
30
void omp_destroy_lock(omp_lock_t *arg)
{
struct __omp_lock *lock = (struct __omp_lock *)arg;
lock->lock = INIT;
}
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
void omp_set_lock(omp_lock_t *arg)
{
struct __omp_lock *lock = (struct __omp_lock *)arg;
if (lock->lock == UNLOCKED)
{
lock->lock = LOCKED;
}
else if (lock->lock == LOCKED)
{
fprintf(stderr,
"error: deadlock in using lock variable\n");
exit(1);
}
else
{
fprintf(stderr, "error: lock not initialized\n");
exit(1);
}
49 280
OpenMP API • Version 3.0 May 2008
1
}
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
void omp_unset_lock(omp_lock_t *arg)
{
struct __omp_lock *lock = (struct __omp_lock *)arg;
if (lock->lock == LOCKED)
{
lock->lock = UNLOCKED;
}
else if (lock->lock == UNLOCKED)
{
fprintf(stderr, "error: lock not set\n");
exit(1);
}
else
{
fprintf(stderr, "error: lock not initialized\n");
exit(1);
}
}
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
int omp_test_lock(omp_lock_t *arg)
{
struct __omp_lock *lock = (struct __omp_lock *)arg;
if (lock->lock == UNLOCKED)
{
lock->lock = LOCKED;
return 1;
}
else if (lock->lock == LOCKED)
{
return 0;
}
else
{
fprintf(stderr, "error: lock not initialized\n");
exit(1);
}
}
38
39
40
41
42
struct __omp_nest_lock
{
short owner;
short count;
};
43
enum { NOOWNER = -1, MASTER = 0 };
44
45
46
47
48
void omp_init_nest_lock(omp_nest_lock_t *arg)
{
struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
nlock->owner = NOOWNER;
nlock->count = 0;
49
Appendix B
Stubs for Runtime Library Routines
281
1
}
2
3
4
5
6
7
void omp_destroy_nest_lock(omp_nest_lock_t *arg)
{
struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
nlock->owner = NOOWNER;
nlock->count = UNLOCKED;
}
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
void omp_set_nest_lock(omp_nest_lock_t *arg)
{
struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
if (nlock->owner == MASTER && nlock->count >= 1)
{
nlock->count++;
}
else if (nlock->owner == NOOWNER && nlock->count == 0)
{
nlock->owner = MASTER;
nlock->count = 1;
}
else
{
fprintf(stderr,
"error: lock corrupted or not initialized\n");
exit(1);
}
}
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
void omp_unset_nest_lock(omp_nest_lock_t *arg)
{
struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
if (nlock->owner == MASTER && nlock->count >= 1)
{
nlock->count--;
if (nlock->count == 0)
{
nlock->owner = NOOWNER;
}
}
else if (nlock->owner == NOOWNER && nlock->count == 0)
{
fprintf(stderr, "error: lock not set\n");
exit(1);
}
else
{
fprintf(stderr,
"error: lock corrupted or not initialized\n");
exit(1);
48 282
OpenMP API • Version 3.0 May 2008
1
2
}
3
4
5
6
7
8
int omp_test_nest_lock(omp_nest_lock_t *arg)
{
struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
omp_set_nest_lock(arg);
return nlock->count;
}
9
10
11
12
13
14
15
16
double omp_get_wtime(void)
{
/* This function does not provide a working
* wallclock timer. Replace it with a version
* customized for the target machine.
*/
return 0.0;
}
17
18
19
20
21
22
23
24
double omp_get_wtick(void)
{
/* This function does not provide a working
* clock tick function. Replace it with
* a version customized for the target machine.
*/
return 365. * 86400.;
}
25
}
Appendix B
Stubs for Runtime Library Routines
283
1
2
B.2
3
4
5
6
7
Fortran Stub Routines
C23456
subroutine omp_set_num_threads(num_threads)
integer num_threads
return
end subroutine
8
9
10
11
integer function omp_get_num_threads()
omp_get_num_threads = 1
return
end function
12
13
14
15
integer function omp_get_max_threads()
omp_get_max_threads = 1
return
end function
16
17
18
19
integer function omp_get_thread_num()
omp_get_thread_num = 0
return
end function
20
21
22
23
integer function omp_get_num_procs()
omp_get_num_procs = 1
return
end function
24
25
26
27
logical function omp_in_parallel()
omp_in_parallel = .false.
return
end function
28
29
30
31
subroutine omp_set_dynamic(dynamic_threads)
logical dynamic_threads
return
end subroutine
32
33
34
35
logical function omp_get_dynamic()
omp_get_dynamic = .false.
return
end function
36
37
38
39
subroutine omp_set_nested(nested)
logical nested
return
end subroutine
40 284
OpenMP API • Version 3.0 May 2008
1
2
3
4
logical function omp_get_nested()
omp_get_nested = .false.
return
end function
5
6
7
8
9
10
subroutine omp_set_schedule(kind, modifier)
include 'omp_lib_kinds.h'
integer (kind=omp_sched_kind) kind
integer modifier
return
end subroutine
11
12
13
14
subroutine omp_get_schedule(kind, modifier)
include 'omp_lib_kinds.h'
integer (kind=omp_sched_kind) kind
integer modifier
15
16
17
18
kind = omp_sched_static
modifier = 0
return
end subroutine
19
20
21
22
integer function omp_get_thread_limit()
omp_get_thread_limit = 1
return
end function
23
24
25
subroutine omp_set_max_active_levels( level )
integer level
end subroutine
26
27
28
29
integer function omp_get_max_active_levels()
omp_get_max_active_levels = 0
return
end function
30
31
32
33
integer function omp_get_level()
omp_get_level = 0
return
end function
34
35
36
37
38
39
40
41
42
integer function omp_get_ancestor_thread_num( level )
integer level
if ( level .eq. 0 ) then
omp_get_ancestor_thread_num = 0
else
omp_get_ancestor_thread_num = -1
end if
return
end function
43
Appendix B
Stubs for Runtime Library Routines
285
1
2
3
4
5
6
7
8
9
integer function omp_get_team_size( level )
integer level
if ( level .eq. 0 ) then
omp_get_team_size = 1
else
omp_get_team_size = -1
end if
return
end function
10
11
12
13
integer function omp_get_active_level()
omp_get_active_level = 0
return
end function
14
15
16
17
18
19
subroutine omp_init_lock(lock)
! lock is 0 if the simple lock is not initialized
!
-1 if the simple lock is initialized but not set
!
1 if the simple lock is set
include 'omp_lib_kinds.h'
integer(kind=omp_lock_kind) lock
20
21
22
lock = -1
return
end subroutine
23
24
25
subroutine omp_destroy_lock(lock)
include 'omp_lib_kinds.h'
integer(kind=omp_lock_kind) lock
26
27
28
lock = 0
return
end subroutine
29
30
31
subroutine omp_set_lock(lock)
include 'omp_lib_kinds.h'
integer(kind=omp_lock_kind) lock
32
33
34
35
36
37
38
39
40
41
42
43 286
if (lock .eq. -1) then
lock = 1
elseif (lock .eq. 1) then
print *, 'error: deadlock in using lock variable'
stop
else
print *, 'error: lock not initialized'
stop
endif
return
end subroutine
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
9
10
11
12
subroutine omp_unset_lock(lock)
include 'omp_lib_kinds.h'
integer(kind=omp_lock_kind) lock
if (lock .eq. 1) then
lock = -1
elseif (lock .eq. -1) then
print *, 'error: lock not set'
stop
else
print *, 'error: lock not initialized'
stop
endif
13
14
return
end subroutine
15
16
17
logical function omp_test_lock(lock)
include 'omp_lib_kinds.h'
integer(kind=omp_lock_kind) lock
18
19
20
21
22
23
24
25
26
if (lock .eq. -1) then
lock = 1
omp_test_lock = .true.
elseif (lock .eq. 1) then
omp_test_lock = .false.
else
print *, 'error: lock not initialized'
stop
endif
27
28
return
end function
29
30
31
32
33
34
35
36
subroutine omp_init_nest_lock(nlock)
! nlock is
! 0 if the nestable lock is not initialized
! -1 if the nestable lock is initialized but not set
! 1 if the nestable lock is set
! no use count is maintained
include 'omp_lib_kinds.h'
integer(kind=omp_nest_lock_kind) nlock
37
nlock = -1
38
39
return
end subroutine
40
Appendix B
Stubs for Runtime Library Routines
287
1
2
3
4
subroutine omp_destroy_nest_lock(nlock)
include 'omp_lib_kinds.h'
integer(kind=omp_nest_lock_kind) nlock
nlock = 0
5
6
return
end subroutine
7
8
9
subroutine omp_set_nest_lock(nlock)
include 'omp_lib_kinds.h'
integer(kind=omp_nest_lock_kind) nlock
10
11
12
13
14
15
16
17
18
if (nlock .eq. -1)
nlock = 1
elseif (nlock .eq.
print *, 'error:
stop
else
print *, 'error:
stop
endif
then
0) then
nested lock not initialized'
deadlock using nested lock variable'
19
20
return
end subroutine
21
22
23
subroutine omp_unset_nest_lock(nlock)
include 'omp_lib_kinds.h'
integer(kind=omp_nest_lock_kind) nlock
24
25
26
27
28
29
30
31
32
33
34
35 288
if (nlock .eq. 1) then
nlock = -1
elseif (nlock .eq. 0) then
print *, 'error: nested lock not initialized'
stop
else
print *, 'error: nested lock not set'
stop
endif
return
end subroutine
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
9
10
11
12
integer function omp_test_nest_lock(nlock)
include 'omp_lib_kinds.h'
integer(kind=omp_nest_lock_kind) nlock
if (nlock .eq. -1) then
nlock = 1
omp_test_nest_lock = 1
elseif (nlock .eq. 1) then
omp_test_nest_lock = 0
else
print *, 'error: nested lock not initialized'
stop
endif
13
14
return
end function
15
16
17
18
double precision function omp_get_wtime()
! this function does not provide a working
! wall clock timer. replace it with a version
! customized for the target machine.
19
omp_get_wtime = 0.0d0
20
21
return
end function
22
23
24
25
26
27
double precision function omp_get_wtick()
! this function does not provide a working
! clock tick function. replace it with
! a version customized for the target machine.
double precision one_year
parameter (one_year=365.d0*86400.d0)
28
29
30
31
omp_get_wtick = one_year
return
end function
Appendix B
Stubs for Runtime Library Routines
289
1
290
OpenMP API • Version 3.0 May 2008
1
2
APPENDIX
C
OpenMP C and C++ Grammar
3
4
5
6
7
8
9
10
C.1
Notation
The grammar rules consist of the name for a non-terminal, followed by a colon,
followed by replacement alternatives on separate lines.
The syntactic expression termopt indicates that the term is optional within the
replacement.
12
The syntactic expression termoptseq is equivalent to term-seqopt with the following
additional rules:
13
term-seq :
14
term
15
term-seq term
16
term-seq , term
11
17
291
1
2
C.2
Rules
5
The notation is described in Section 6.1 of the C standard. This grammar appendix
shows the extensions to the base language grammar for the OpenMP C and C++
directives.
6
/* in C++ (ISO/IEC 14882:1998) */
7
statement-seq:
3
4
8
statement
9
openmp-directive
10
statement-seq statement
11
statement-seq openmp-directive
12
/* in C90 (ISO/IEC 9899:1990) */
13
statement-list:
14
statement
15
openmp-directive
16
statement-list statement
17
statement-list openmp-directive
18
/* in C99 (ISO/IEC 9899:1999) */
19
block-item:
20
declaration
21
statement
22
openmp-directive
23 292
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
statement:
/* standard statements */
openmp-construct
openmp-construct:
parallel-construct
6
for-construct
7
sections-construct
8
single-construct
9
parallel-for-construct
10
parallel-sections-construct
11
task-construct
12
master-construct
13
critical-construct
14
atomic-construct
15
ordered-construct
16
openmp-directive:
17
barrier-directive
18
taskwait-directive
19
flush-directive
20
structured-block:
21
statement
22
23
24
25
26
parallel-construct:
parallel-directive structured-block
parallel-directive:
# pragma omp parallel parallel-clauseoptseq new-line
Appendix C
OpenMP C and C++ Grammar
293
1
2
parallel-clause:
unique-parallel-clause
3
data-default-clause
4
data-privatization-clause
5
data-privatization-in-clause
6
data-sharing-clause
7
data-reduction-clause
8
9
unique-parallel-clause:
if ( expression )
10
num_threads ( expression )
11
copyin ( variable-list )
12
13
14
15
16
17
for-construct:
for-directive iteration-statement
for-directive:
# pragma omp for for-clauseoptseq new-line
for-clause:
unique-for-clause
18
data-privatization-clause
19
data-privatization-in-clause
20
data-privatization-out-clause
21
data-reduction-clause
22
nowait
23
unique-for-clause:
24
ordered
25
schedule ( schedule-kind )
26
schedule ( schedule-kind , expression )
27
collapse ( expression )
28 294
OpenMP API • Version 3.0 May 2008
1
schedule-kind:
2
static
3
dynamic
4
guided
5
auto
6
runtime
7
8
9
10
11
sections-construct:
sections-directive section-scope
sections-directive:
# pragma omp sections sections-clauseoptseq new-line
sections-clause:
12
data-privatization-clause
13
data-privatization-in-clause
14
data-privatization-out-clause
15
data-reduction-clause
16
nowait
17
section-scope:
18
19
20
21
22
23
24
25
26
27
28
{ section-sequence }
section-sequence:
section-directiveopt structured-block
section-sequence section-directive structured-block
section-directive:
# pragma omp section new-line
single-construct:
single-directive structured-block
single-directive:
# pragma omp single single-clauseoptseq new-line
Appendix C
OpenMP C and C++ Grammar
295
1
2
single-clause:
unique-single-clause
3
data-privatization-clause
4
data-privatization-in-clause
5
nowait
6
7
8
9
10
11
12
unique-single-clause:
copyprivate ( variable-list )
task-construct:
task-directive structured-block
task-directive:
# pragma omp task task-clauseoptseq new-line
task-clause:
13
unique-task-clause
14
data-default-clause
15
data-privatization-clause
16
data-privatization-in-clause
17
data-sharing-clause
18
unique-task-clause:
19
if ( scalar-expression )
20
untied
21
22
23
24
25
26
parallel-for-construct:
parallel-for-directive iteration-statement
parallel-for-directive:
# pragma omp parallel for parallel-for-clauseoptseq new-line
parallel-for-clause:
unique-parallel-clause
27
unique-for-clause
28
data-default-clause
29 296
OpenMP API • Version 3.0 May 2008
1
data-privatization-clause
2
data-privatization-in-clause
3
data-privatization-out-clause
4
data-sharing-clause
5
data-reduction-clause
6
parallel-sections-construct:
7
8
9
10
parallel-sections-directive section-scope
parallel-sections-directive:
# pragma omp parallel sections parallel-sections-clauseoptseq new-line
parallel-sections-clause:
11
unique-parallel-clause
12
data-default-clause
13
data-privatization-clause
14
data-privatization-in-clause
15
data-privatization-out-clause
16
data-sharing-clause
17
data-reduction-clause
18
19
20
21
22
23
24
25
26
27
28
master-construct:
master-directive structured-block
master-directive:
# pragma omp master new-line
critical-construct:
critical-directive structured-block
critical-directive:
# pragma omp critical region-phraseopt new-line
region-phrase:
( identifier )
Appendix C
OpenMP C and C++ Grammar
297
1
2
3
4
5
6
7
8
9
10
11
barrier-directive:
# pragma omp barrier new-line
taskwait-directive:
# pragma omp taskwait new-line
atomic-construct:
atomic-directive expression-statement
atomic-directive:
# pragma omp atomic new-line
flush-directive:
# pragma omp flush flush-varsopt new-line
flush-vars:
12
( variable-list )
13
ordered-construct:
14
15
16
17
18
19
20
21
22
ordered-directive structured-block
ordered-directive:
# pragma omp ordered new-line
declaration:
/* standard declarations */
threadprivate-directive
threadprivate-directive:
# pragma omp threadprivate ( variable-list ) new-line
data-default-clause:
23
default ( shared )
24
default ( none )
25
26
27
28
29 298
data-privatization-clause:
private ( variable-list )
data-privatization-in-clause:
firstprivate ( variable-list )
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
data-privatization-out-clause:
lastprivate ( variable-list )
data-sharing-clause:
shared ( variable-list )
data-reduction-clause:
reduction ( reduction-operator : variable-list )
reduction-operator:
One of: + * - & ^ | && ||
9
/* in C */
10
variable-list:
11
identifier
12
variable-list , identifier
13
/* in C++ */
14
variable-list:
15
16
17
id-expression
variable-list , id-expression
Appendix C
OpenMP C and C++ Grammar
299
1
300
OpenMP API • Version 3.0 May 2008
1
2
3
APPENDIX
D
Interface Declarations
4
7
This appendix gives examples of the C/C++ header file, the Fortran include file and
Fortran 90 module that shall be provided by implementations as specified in Chapter 3.
It also includes an example of a Fortran 90 generic interface for a library routine.
8
301
5
6
1
2
D.1
Example of the omp.h Header File
3
4
#ifndef _OMP_H_DEF
#define _OMP_H_DEF
5
6
7
8
/*
* define the lock data types
*/
typedef void *omp_lock_t;
9
typedef void *omp_nest_lock_t;
10
11
12
13
14
15
16
17
18
19
20
/*
* define the schedule kinds
*/
typedef enum omp_sched_t
{
omp_sched_static = 1,
omp_sched_dynamic = 2,
omp_sched_guided = 3,
omp_sched_auto = 4
/* , Add vendor specific schedule constants here */
} omp_sched_t;
21
22
23
24
25
26
27
/*
* exported OpenMP functions
*/
#ifdef __cplusplus
extern
"C"
{
#endif
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
extern
extern
extern
extern
extern
extern
extern
extern
extern
extern
extern
extern
extern
extern
extern
43 302
void
int
int
int
int
int
void
int
void
int
int
void
int
int
int
omp_set_num_threads(int num_threads);
omp_get_num_threads(void);
omp_get_max_threads(void);
omp_get_thread_num(void);
omp_get_num_procs(void);
omp_in_parallel(void);
omp_set_dynamic(int dynamic_threads);
omp_get_dynamic(void);
omp_set_nested(int nested);
omp_get_nested(void);
omp_get_thread_limit(void);
omp_set_max_active_levels(int max_active_levels);
omp_get_max_active_levels(void);
omp_get_level(void);
omp_get_ancestor_thread_num(int level);
OpenMP API • Version 3.0 May 2008
1
2
3
4
extern
extern
extern
extern
int
int
void
void
omp_get_team_size(int level);
omp_get_active_level(void);
omp_set_schedule(omp_sched_t kind, int modifier);
omp_get_schedule(omp_sched_t *kind, int *modifier);
5
6
7
8
9
extern
extern
extern
extern
extern
void
void
void
void
int
omp_init_lock(omp_lock_t *lock);
omp_destroy_lock(omp_lock_t *lock);
omp_set_lock(omp_lock_t *lock);
omp_unset_lock(omp_lock_t *lock);
omp_test_lock(omp_lock_t *lock);
10
11
12
13
14
extern
extern
extern
extern
extern
void
void
void
void
int
omp_init_nest_lock(omp_nest_lock_t *lock);
omp_destroy_nest_lock(omp_nest_lock_t *lock);
omp_set_nest_lock(omp_nest_lock_t *lock);
omp_unset_nest_lock(omp_nest_lock_t *lock);
omp_test_nest_lock(omp_nest_lock_t *lock);
15
16
extern double omp_get_wtime(void);
extern double omp_get_wtick(void);
17
18
19
#ifdef __cplusplus
}
#endif
20
#endif
21
Appendix D
Interface Declarations
303
1
2
D.2
3
4
5
6
7
8
Example of an Interface Declaration include
File
omp_lib_kinds.h:
integer
omp_lock_kind
parameter ( omp_lock_kind = 8 )
integer
omp_nest_lock_kind
parameter ( omp_nest_lock_kind = 8 )
9
10
integer
omp_sched_kind
parameter ( omp_sched_kind = 4)
11
12
13
14
15
16
17
18
integer (
parameter
integer (
parameter
integer (
parameter
integer (
parameter
19
20
21
22
omp_sched_kind ) omp_sched_static
( omp_sched_static = 1 )
omp_sched_kind ) omp_sched_dynamic
( omp_sched_dynamic = 2 )
omp_sched_kind ) omp_sched_guided
( omp_sched_guided = 3 )
omp_sched_kind ) omp_sched_auto
( omp_sched_auto = 4 )
omp_lib.h:
C
C
C
default integer type assumed below
default logical type assumed below
OpenMP Fortran API v3.0
23
24
25
include 'omp_lib_kinds.h'
integer
openmp_version
parameter ( openmp_version = 200805 )
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
external
external
integer
external
integer
external
integer
external
integer
external
logical
external
external
logical
external
external
logical
external
external
external
46 304
omp_set_num_threads
omp_get_num_threads
omp_get_num_threads
omp_get_max_threads
omp_get_max_threads
omp_get_thread_num
omp_get_thread_num
omp_get_num_procs
omp_get_num_procs
omp_in_parallel
omp_in_parallel
omp_set_dynamic
omp_get_dynamic
omp_get_dynamic
omp_set_nested
omp_get_nested
omp_get_nested
omp_set_schedule
omp_get_schedule
omp_get_thread_limit
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
6
7
8
9
10
11
12
integer omp_get_thread_limit
external omp_set_max_active_levels
external omp_get_max_active_levels
integer omp_get_max_active_levels
external omp_get_level
integer omp_get_level
external omp_get_ancestor_thread_num
integer omp_get_ancestor_thread_num
external omp_get_team_size
integer omp_get_team_size
external omp_get_active_level
integer omp_get_active_level
13
14
15
16
17
18
external
external
external
external
external
logical
omp_init_lock
omp_destroy_lock
omp_set_lock
omp_unset_lock
omp_test_lock
omp_test_lock
19
20
21
22
23
24
external
external
external
external
external
integer
omp_init_nest_lock
omp_destroy_nest_lock
omp_set_nest_lock
omp_unset_nest_lock
omp_test_nest_lock
omp_test_nest_lock
25
26
27
28
external omp_get_wtick
double precision omp_get_wtick
external omp_get_wtime
double precision omp_get_wtime
29
Appendix D
Interface Declarations
305
1
2
D.3
Example of a Fortran 90 Interface Declaration
3
module
4
5
!
the "!" of this comment starts in column 1
!23456
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
&
&
&
&
21
22
23
24
module omp_lib
use omp_lib_kinds
!
OpenMP Fortran API v3.0
integer, parameter :: openmp_version = 200805
25
26
27
28
29
30
module omp_lib_kinds
integer, parameter :: omp_integer_kind = 4
integer, parameter :: omp_logical_kind = 4
integer, parameter :: omp_lock_kind = 8
integer, parameter :: omp_nest_lock_kind = 8
integer, parameter :: omp_sched_kind = 4
integer(kind=omp_sched_kind), parameter ::
omp_sched_static = 1
integer(kind=omp_sched_kind), parameter ::
omp_sched_dynamic = 2
integer(kind=omp_sched_kind), parameter ::
omp_sched_guided = 3
integer(kind=omp_sched_kind), parameter ::
omp_sched_auto = 4
end module omp_lib_kinds
interface
&
subroutine omp_set_num_threads (number_of_threads_expr)
use omp_lib_kinds
integer (kind=omp_integer_kind), intent(in) ::
number_of_threads_expr
end subroutine omp_set_num_threads
31
32
33
34
function omp_get_num_threads ()
use omp_lib_kinds
integer (kind=omp_integer_kind) :: omp_get_num_threads
end function omp_get_num_threads
35
36
37
38
function omp_get_max_threads ()
use omp_lib_kinds
integer (kind=omp_integer_kind) :: omp_get_max_threads
end function omp_get_max_threads
39
40
41
42
function omp_get_thread_num ()
use omp_lib_kinds
integer (kind=omp_integer_kind) :: omp_get_thread_num
end function omp_get_thread_num
43 306
OpenMP API • Version 3.0 May 2008
1
2
3
4
function omp_get_num_procs ()
use omp_lib_kinds
integer (kind=omp_integer_kind) :: omp_get_num_procs
end function omp_get_num_procs
5
6
7
8
function omp_in_parallel ()
use omp_lib_kinds
logical (kind=omp_logical_kind) :: omp_in_parallel
end function omp_in_parallel
9
10
11
12
13
subroutine omp_set_dynamic (enable_expr)
use omp_lib_kinds
logical (kind=omp_logical_kind), intent(in) ::
enable_expr
end subroutine omp_set_dynamic
&
14
15
16
17
function omp_get_dynamic ()
use omp_lib_kinds
logical (kind=omp_logical_kind) :: omp_get_dynamic
end function omp_get_dynamic
18
19
20
21
22
subroutine omp_set_nested (enable_expr)
use omp_lib_kinds
logical (kind=omp_logical_kind), intent(in) ::
enable_expr
end subroutine omp_set_nested
&
23
24
25
26
function omp_get_nested ()
use omp_lib_kinds
logical (kind=omp_logical_kind) :: omp_get_nested
end function omp_get_nested
27
28
29
30
31
subroutine omp_set_schedule (kind, modifier)
use omp_lib_kinds
integer(kind=omp_sched_kind), intent(in) :: kind
integer(kind=omp_integer_kind), intent(in) :: modifier
end subroutine omp_set_schedule
32
33
34
35
36
subroutine omp_get_schedule (kind, modifier)
use omp_lib_kinds
integer(kind=omp_sched_kind), intent(out) :: kind
integer(kind=omp_integer_kind), intent(out)::modifier
end subroutine omp_get_schedule
37
38
39
40
function omp_get_thread_limit()
use omp_lib_kinds
integer (kind=omp_integer_kind) :: omp_get_thread_limit
end function omp_get_thread_limit
41
42
43
44
subroutine omp_set_max_active_levels(var)
use omp_lib_kinds
integer (kind=omp_integer_kind), intent(in) :: var
end subroutine omp_set_max_active_levels
45
Appendix D
Interface Declarations
307
1
2
3
4
5
&
function omp_get_max_active_levels()
use omp_lib_kinds
integer (kind=omp_integer_kind) ::
omp_get_max_active_levels
end function omp_get_max_active_levels
6
7
8
9
function omp_get_level()
use omp_lib_kinds
integer (kind=omp_integer_kind) :: omp_get_level
end function omp_get_level
10
11
12
13
14
15
16
function omp_get_ancestor_thread_num(level)
use omp_lib_kinds
integer (kind=omp_integer_kind), intent(in) ::
level
integer (kind=omp_integer_kind) ::
omp_get_ancestor_thread_num
end function omp_get_ancestor_thread_num
17
18
19
20
21
22
23
24
25
26
27
&
&
&
&
function omp_get_team_size(level)
use omp_lib_kinds
integer (kind=omp_integer_kind), intent(in) ::
level
integer (kind=omp_integer_kind) :: omp_get_team_size
end function omp_get_team_size
function omp_get_active_level()
use omp_lib_kinds
integer (kind=omp_integer_kind) ::
omp_get_active_level
end function omp_get_active_level
28
29
30
31
subroutine omp_init_lock (var)
use omp_lib_kinds
integer (kind=omp_lock_kind), intent(out) :: var
end subroutine omp_init_lock
32
33
34
35
subroutine omp_destroy_lock (var)
use omp_lib_kinds
integer (kind=omp_lock_kind), intent(inout) :: var
end subroutine omp_destroy_lock
36
37
38
39
subroutine omp_set_lock (var)
use omp_lib_kinds
integer (kind=omp_lock_kind), intent(inout) :: var
end subroutine omp_set_lock
40
41
42
43
subroutine omp_unset_lock (var)
use omp_lib_kinds
integer (kind=omp_lock_kind), intent(inout) :: var
end subroutine omp_unset_lock
44 308
OpenMP API • Version 3.0 May 2008
1
2
3
4
5
function omp_test_lock (var)
use omp_lib_kinds
logical (kind=omp_logical_kind) :: omp_test_lock
integer (kind=omp_lock_kind), intent(inout) :: var
end function omp_test_lock
6
7
8
9
subroutine omp_init_nest_lock (var)
use omp_lib_kinds
integer (kind=omp_nest_lock_kind), intent(out) :: var
end subroutine omp_init_nest_lock
10
11
12
13
subroutine omp_destroy_nest_lock (var)
use omp_lib_kinds
integer (kind=omp_nest_lock_kind), intent(inout) :: var
end subroutine omp_destroy_nest_lock
14
15
16
17
subroutine omp_set_nest_lock (var)
use omp_lib_kinds
integer (kind=omp_nest_lock_kind), intent(inout) :: var
end subroutine omp_set_nest_lock
18
19
20
21
subroutine omp_unset_nest_lock (var)
use omp_lib_kinds
integer (kind=omp_nest_lock_kind), intent(inout) :: var
end subroutine omp_unset_nest_lock
22
23
24
25
26
27
function omp_test_nest_lock (var)
use omp_lib_kinds
integer (kind=omp_integer_kind) :: omp_test_nest_lock
integer (kind=omp_nest_lock_kind), intent(inout) ::
var
end function omp_test_nest_lock
&
28
29
30
function omp_get_wtick ()
double precision :: omp_get_wtick
end function omp_get_wtick
31
32
33
function omp_get_wtime ()
double precision :: omp_get_wtime
end function omp_get_wtime
34
end interface
35
end module omp_lib
36
Appendix D
Interface Declarations
309
1
2
D.4
3
4
5
6
7
8
9
Example of a Generic Interface for a Library
Routine
Any of the OMP runtime library routines that take an argument may be extended with a
generic interface so arguments of different KIND type can be accommodated.
Assume an implementation supports both default INTEGER as KIND =
OMP_INTEGER_KIND and another INTEGER KIND, KIND = SHORT_INT. Then
OMP_SET_NUM_THREADS could be specified in the omp_lib module as the
following:
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
!
the "!" of this comment starts in column 1
interface omp_set_num_threads
subroutine omp_set_num_threads_1 ( number_of_threads_expr )
use omp_lib_kinds
integer ( kind=omp_integer_kind ), intent(in) :: &
&
number_of_threads_expr
end subroutine omp_set_num_threads_1
subroutine omp_set_num_threads_2 ( number_of_threads_expr )
use omp_lib_kinds
integer ( kind=short_int ), intent(in) :: &
&
number_of_threads_expr
end subroutine omp_set_num_threads_2
end interface omp_set_num_threads
27
28 310
OpenMP API • Version 3.0 May 2008
1
2
3
4
APPENDIX
E
Implementation Defined
Behaviors in OpenMP
5
9
This appendix summarizes the behaviors that are described as implementation defined in
this API. Each behavior is cross-referenced back to its description in the main
specification. An implementation is required to define and document its behavior in
these cases.
10
• Task scheduling points: it is implementation defined where task scheduling points
6
7
8
11
12
13
14
occur in untied task regions (see Section 1.3 on page 11).
• Memory model: it is implementation defined as to whether, and in what sizes,
memory accesses by multiple threads to the same variable without synchronization
are atomic with respect to each other (see Section 1.4.1 on page 13).
15
• Internal control variables: the initial values of nthreads-var, dyn-var, run-sched-var,
16
17
def-sched-var, stacksize-var, wait-policy-var, thread-limit-var, and max-active-levelsvar are implementation defined (see Section 2.3.2 on page 29).
18
• Dynamic adjustment of threads: it is implementation defined whether the ability to
19
20
21
22
dynamically adjust the number of threads is provided. Implementations are allowed
to deliver fewer threads (but at least one) than indicated in Algorithm 2-1 even if
dynamic adjustment is disabled (see Section 2.4.1 on page 35).
• Loop directive: the integer type or kind used to compute the iteration count of a
25
collapsed loop is implementation defined. The effect of the schedule(runtime)
clause when the run-sched-var ICV is set to auto is implementation defined. See
Section 2.5.1 on page 38.
26
• sections construct: the method of scheduling the structured blocks among threads
23
24
27
28
29
30
in the team is implementation defined (see Section 2.5.2 on page 47).
• single construct: the method of choosing a thread to execute the structured block
is implementation defined (see Section 2.5.3 on page 49).
311
1
2
3
4
5
6
7
8
• atomic construct: a compliant implementation may enforce exclusive access
between atomic regions which update different storage locations. The
circumstances under which this occurs are implementation defined (see Section 2.8.5
on page 69).
• omp_set_num_threads routine: if the argument is not a positive integer the
behavior is implementation defined (see Section 3.2.1 on page 110).
• omp_set_schedule routine: the behavior for implementation defined schedule
types is implementation defined (see Section 3.2.11 on page 121).
9
• omp_set_max_active_levels routine: when called from within any explicit
10
13
parallel region the binding thread set (and binding region, if required) for the
omp_set_max_active_levels region is implementation defined and the
behavior is implementation defined. If the argument is not a non-negative integer
then the behavior is implementation defined (see Section 3.2.14 on page 126).
14
• omp_get_max_active_levels routine: when called from within any explicit
11
12
15
16
17
parallel region the binding thread set (and binding region, if required) for the
omp_get_max_active_levels region is implementation defined (see
Section 3.2.15 on page 127).
18
• OMP_SCHEDULE environment variable: if the value of the variable does not
19
conform to the specified format then the result is implementation defined (see
Section 4.1 on page 146).
20
21
22
23
24
25
26
27
28
29
30
31
32
33
• OMP_NUM_THREADS environment variable: if the value of the variable is greater
than the number of threads the implementation can support or is not a positive integer
then the result is implementation defined (see Section 4.2 on page 147).
• OMP_DYNAMIC environment variable: if the value is neither true nor false the
behavior is implementation defined (see Section 4.3 on page 148).
• OMP_NESTED environment variable: if the value is neither true nor false the
behavior is implementation defined (see Section 4.4 on page 148).
• OMP_STACKSIZE environment variable: if the value does not conform to the
specified format or the implementation cannot provide a stack of the specified size
then the behavior is implementation defined (see Section 4.5 on page 149).
• OMP_WAIT_POLICY environment variable: the details of the ACTIVE and
PASSIVE behaviors are implementation defined (see Section 4.6 on page 150).
• OMP_MAX_ACTIVE_LEVELS environment variable: if the value is not a non-
35
negative integer or is greater than the number of parallel levels an impementation can
support then the behavior is implementation defined (see Section 4.7 on page 150).
36
• OMP_THREAD_LIMIT environment variable: if the requested value is greater than
37
the number of threads an implementation can support, or if the value is not a positive
integer, the behavior of the program is implementation defined (see Section 4.8 on
page 151).
34
38
39
40 312
OpenMP API • Version 3.0 May 2008
1
Fortran
2
• threadprivate directive: if the conditions for values of data in the threadprivate
3
5
objects of threads (other than the initial thread) to persist between two consecutive
active parallel regions do not all hold, the allocation status of an allocatable array in
the second region is implementation defined (see Section 2.9.2 on page 81).
6
• shared clause: passing a shared variable to a non-intrinsic procedure may result in
7
the value of the shared variable being copied into temporary storage before the
procedure reference, and back out of the temporary storage into the actual argument
storage after the procedure reference. Situations where this occurs other than those
specified are implementation defined (see Section 2.9.3.2 on page 88).
4
8
9
10
11
• Runtime library definitions: it is implementation defined whether the include file
12
omp_lib.h or the module omp_lib (or both) is provided. It is implementation
defined whether any of the OpenMP runtime library routines that take an argument
are extended with a generic interface so arguments of different KIND type can be
accommodated (see Section 3.1 on page 108).
13
14
15
16
17
Fortran
Appendix E
Implementation Defined Behaviors in OpenMP
313
1
314
OpenMP API • Version 3.0 May 2008
1
2
3
4
APPENDIX
F
Changes from Version 2.5 to
Version 3.0
5
7
This appendix summarizes the major changes between the OpenMP API Version 2.5
specification and the OpenMP API Version 3.0 specification.
8
• The concept of tasks has been added to the OpenMP execution model (see
6
9
10
11
12
13
14
Section 1.2.3 on page 8 and Section 1.3 on page 11).
• The task construct (see Section 2.7 on page 59) has been added, which provides a
mechanism for creating tasks explicitly.
• The taskwait construct (see Section 2.8.4 on page 68) has been added, which
causes a task to wait for all its child tasks to complete.
• The OpenMP memory model now covers atomicity of memory accesses (see
16
Section 1.4.1 on page 13). The description of the behavior of volatile in terms of
flush was removed.
17
• In Version 2.5, there was a single copy of of the nest-var, dyn-var, nthreads-var and
18
run-sched-var internal control variables (ICVs) for the whole program. In Version
3.0, there is one copy of these ICVs per task (see Section 2.3 on page 28). As a result,
the omp_set_num_threads, omp_set_nested and omp_set_dynamic
runtime library routines now have specified effects when called from inside a
parallel region (See Section 3.2.1 on page 110, Section 3.2.7 on page 117 and
Section 3.2.9 on page 119).
15
19
20
21
22
23
24
• The definition of active parallel region has been changed: in Version 3.0 a
26
parallel region is active if it is executed by a team consisting of more than one
thread (see Section 1.2.2 on page 2).
27
• The rules for determining the number of threads used in a parallel region have
25
28
29
30
31
been modified (see Section 2.4.1 on page 35).
• In Version 3.0, the assignment of iterations to threads in a loop construct with a
static schedule kind is deterministic (see Section 2.5.1 on page 38).
315
1
2
3
• In Version 3.0, a loop construct may be associated with more than one perfectly
nested loop. The number of associated loops may be controlled by the collapse
clause (see Section 2.5.1 on page 38).
4
• Random access iterators, and variables of unsigned integer type, may now be used as
5
loop iterators in loops associated with a loop construct (see Section 2.5.1 on page 38).
6
7
8
9
10
11
12
• The schedule kind auto has been added, which gives the implementation the
freedom to choose any possible mapping of iterations in a loop construct to threads in
the team (see Section 2.5.1 on page 38).
• Fortran assumed-size arrays now have predetermined data-sharing attributes (see
Section 2.9.1.1 on page 78).
• In Fortran, firstprivate is now permitted as an argument to the default
clause (see Section 2.9.3.1 on page 86).
13
• For list items in the private clause, implementations are no longer permitted to use
14
the storage of the original list item to hold the new list item on the master thread. If
no attempt is made to reference the original list item inside the parallel region, its
value is well defined on exit from the parallel region (see Section 2.9.3.3 on page
89).
15
16
17
18
19
20
21
22
23
24
25
• In Version 3.0, Fortran allocatable arrays may appear in private,
firstprivate, lastprivate, reduction, copyin and copyprivate
clauses. (see Section 2.9.2 on page 81, Section 2.9.3.3 on page 89, Section 2.9.3.4 on
page 92, Section 2.9.3.5 on page 94, Section 2.9.3.6 on page 96, Section 2.9.4.1 on
page 101 and Section 2.9.4.2 on page 102).
• In Version 3.0, static class members variables may appear in a threadprivate
directive (see Section 2.9.2 on page 81).
• Version 3.0 makes clear where, and with which arguments, constructors and
28
destructors of private and threadprivate class type variables are called (see
Section 2.9.2 on page 81, Section 2.9.3.3 on page 89, Section 2.9.3.4 on page 92,
Section 2.9.4.1 on page 101 and Section 2.9.4.2 on page 102)
29
• The runtime library routines omp_set_schedule and omp_get_schedule
26
27
30
31
32
have been added; these routines respectively set and retrieve the value of the
run_sched_var ICV (see Section 3.2.11 on page 121 and Section 3.2.12 on page 123).
• The thread-limit-var ICV has been added, which controls the maximium number of
36
threads participating in the OpenMP program. The value of this ICV can be set with
the OMP_THREAD_LIMIT environment variable and retrieved with the
omp_get_thread_limit runtime library routine (see Section 2.3.1 on page 28,
Section 3.2.13 on page 125 and Section 4.8 on page 151).
37
• The max-active-levels-var ICV has been added, which controls the number of nested
38
active parallel regions. The value of this ICV can be set with the
OMP_MAX_ACTIVE_LEVELS environment variable and the
omp_set_max_active_levels runtime library routine, and it can be retrieved
33
34
35
39
40
41 316
OpenMP API • Version 3.0 May 2008
3
with the omp_get_max_active_levels runtime library routine (see
Section 2.3.1 on page 28, Section 3.2.14 on page 126, Section 3.2.15 on page 127 and
Section 4.7 on page 150).
4
• The stacksize-var ICV has been added, which controls the stack size for threads that
1
2
5
6
7
8
the OpenMP implementation creates. The value of this ICV can be set with the
OMP_STACKSIZE environment variable (see Section 2.3.1 on page 28 and
Section 4.5 on page 149).
• The wait-policy-var ICV has been added, which controls the desired behavior of
10
waiting threads. The value of this ICV can be set with the OMP_WAIT_POLICY
environment variable (see Section 2.3.1 on page 28 and Section 4.6 on page 150).
11
• The omp_get_level runtime library routine has been added, which returns the
9
12
13
number of nested parallel regions enclosing the task that contains the call (see
Section 3.2.16 on page 129).
14
• The omp_get_ancestor_thread_num runtime library routine has been added,
15
16
which returns, for a given nested level of the current thread, the thread number of the
ancestor (see Section 3.2.17 on page 130).
17
• The omp_get_team_size runtime library routine has been added, which returns,
18
for a given nested level of the current thread, the size of the thread team to which the
ancestor belongs (see Section 3.2.18 on page 131).
19
20
21
22
23
24
25
• The omp_get_active_level runtime library routine has been added, which
returns the number of nested, active parallel regions enclosing the task that
contains the call (see Section 3.2.19 on page 133).
• In Version 3.0, locks are owned by tasks, not by threads (see Section 3.3 on page
134).
Appendix F
Changes from Version 2.5 to Version 3.0
317
1
318
OpenMP API • Version 3.0 May 2008