1

2 OpenMP Application
3 Program Interface

4 Version 3.0 May 2008

5 Copyright © 1997-2008 OpenMP Architecture Review Board.

6 Permission to copy without fee all or part of this material is granted,
7 provided the OpenMP Architecture Review Board copyright notice and the
8 title of this document appear. Notice is given that copying is by permission
9 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 ..............................2
5 1.2.2 OpenMP language terminology ......................2
6 1.2.3 Tasking Terminology ..............................8
7 1.2.4 Data Terminology .................................9
8 1.2.5 Implementation Terminology . . . . . . . . . . . . . . . . . . . . . . . . 10
9 1.3 Execution Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
10 1.4 Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
11 1.4.1 Structure of the OpenMP Memory Model . . . . . . . . . . . . . . . 13
12 1.4.2 The Flush Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
13 1.4.3 OpenMP Memory Consistency . . . . . . . . . . . . . . . . . . . . . . 16
14 1.5 OpenMP Compliance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
15 1.6 Normative References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
16 1.7 Organization of this document . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

17 2. Directives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
18 2.1 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 . . . . 26
23 2.2.2 Free Source Form Conditional Compilation Sentinel . . . . . . 27
24 2.3 Internal Control Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
25 2.3.1 ICV Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
26 2.3.2 Modifying and Retrieving ICV Values . . . . . . . . . . . . . . . . . . 29
27 2.3.3 How the Per-task ICVs Work . . . . . . . . . . . . . . . . . . . . . . . . 30
28 2.3.4 ICV Override Relationships . . . . . . . . . . . . . . . . . . . . . . . . . 30
29 2.4 parallel Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

30 i
1 2.4.1 Determining the Number of Threads for a
2 parallel Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 2.5 Worksharing Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4 2.5.1 Loop Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5 2.5.1.1 Determining the Schedule of a
6 Worksharing Loop . . . . . . . . . . . . . . . . . . . . . . . . 45
7 2.5.2 sections Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8 2.5.3 single Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
9 2.5.4 workshare Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
10 2.6 Combined Parallel Worksharing Constructs . . . . . . . . . . . . . . . . . . . 54
11 2.6.1 Parallel Loop construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
12 2.6.2 parallel sections Construct . . . . . . . . . . . . . . . . . . . . . 56
13 2.6.3 parallel workshare Construct . . . . . . . . . . . . . . . . . . . . 58
14 2.7 task Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
15 2.7.1 Task Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
16 2.8 Master and Synchronization Constructs . . . . . . . . . . . . . . . . . . . . . . 63
17 2.8.1 master Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
18 2.8.2 critical Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
19 2.8.3 barrier Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
20 2.8.4 taskwait Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
21 2.8.5 atomic Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
22 2.8.6 flush Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
23 2.8.7 ordered Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
24 2.9 Data Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
25 2.9.1 Data-sharing Attribute Rules . . . . . . . . . . . . . . . . . . . . . . . . 77
26 2.9.1.1 Data-sharing Attribute Rules for Variables
27 Referenced in a Construct . . . . . . . . . . . . . . . . . . 78
28 2.9.1.2 Data-sharing Attribute Rules for Variables
29 Referenced in a Region but not in a Construct . . 80
30 2.9.2 threadprivate Directive . . . . . . . . . . . . . . . . . . . . . . . . . 81

31 ii OpenMP ABI • Version 3.0 May 2008

1 2.9.3 Data-Sharing Attribute Clauses . . . . . . . . . . . . . . . . . . . . . . 85
2 2.9.3.1 default clause . . . . . . . . . . . . . . . . . . . . . . . . . 86
3 2.9.3.2 shared clause . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4 2.9.3.3 private clause . . . . . . . . . . . . . . . . . . . . . . . . . 89
5 2.9.3.4 firstprivate clause . . . . . . . . . . . . . . . . . . . . 92
6 2.9.3.5 lastprivate clause . . . . . . . . . . . . . . . . . . . . . 94
7 2.9.3.6 reduction clause . . . . . . . . . . . . . . . . . . . . . . . 96
8 2.9.4 Data Copying Clauses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
9 2.9.4.1 copyin clause . . . . . . . . . . . . . . . . . . . . . . . . . 101
10 2.9.4.2 copyprivate clause . . . . . . . . . . . . . . . . . . . . 102
11 2.10 Nesting of Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

12 3. Runtime Library Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 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 . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
19 3.2.5 omp_get_num_procs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
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 . . . . . . . . . . . . . . . . . . . . . . . . . 125
28 3.2.14 omp_set_max_active_levels . . . . . . . . . . . . . . . . . . . . 126
29 3.2.15 omp_get_max_active_levels . . . . . . . . . . . . . . . . . . . . 127

30 iii
1 3.2.16 omp_get_level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
2 3.2.17 omp_get_ancestor_thread_num . . . . . . . . . . . . . . . . . . 130
3 3.2.18 omp_get_team_size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4 3.2.19 omp_get_active_level . . . . . . . . . . . . . . . . . . . . . . . . . 133
5 3.3 Lock Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6 3.3.1 omp_init_lock and omp_init_nest_lock . . . . . . . . . . 136
7 3.3.2 omp_destroy_lock and omp_destroy_nest_lock . . . 137
8 3.3.3 omp_set_lock and omp_set_nest_lock . . . . . . . . . . . . 138
9 3.3.4 omp_unset_lock and omp_unset_nest_lock . . . . . . . . 140
10 3.3.5 omp_test_lock and omp_test_nest_lock . . . . . . . . . . 141
11 3.4 Timing Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 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 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
16 4.2 OMP_NUM_THREADS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
17 4.3 OMP_DYNAMIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
18 4.4 OMP_NESTED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
19 4.5 OMP_STACKSIZE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
20 4.6 OMP_WAIT_POLICY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
21 4.7 OMP_MAX_ACTIVE_LEVELS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
22 4.8 OMP_THREAD_LIMIT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

23 A. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
24 A.1 A Simple Parallel Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
25 A.2 The OpenMP Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
26 A.3 Conditional Compilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
27 A.4 Internal Control Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
28 A.5 The parallel Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
29 A.6 The num_threads Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

30 iv OpenMP ABI • Version 3.0 May 2008

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 . . . . . . . . . . . . . . . . . . . . . . . 174
6 A.12 The single Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7 A.13 Tasking Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8 A.14 The workshare Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9 A.15 The master Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
10 A.16 The critical Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
11 A.17 worksharing Constructs Inside a critical Construct . . . . . . . . . . 199
12 A.18 Binding of barrier Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
13 A.19 The atomic Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
14 A.20 Restrictions on the atomic Construct . . . . . . . . . . . . . . . . . . . . . . . 205
15 A.21 The flush Construct with a List . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
16 A.22 The flush Construct without a List . . . . . . . . . . . . . . . . . . . . . . . . . 211
17 A.23 Placement of flush, barrier, and taskwait Directives . . . . . . . 214
18 A.24 The ordered Clause and the ordered Construct . . . . . . . . . . . . . 215
19 A.25 The threadprivate Directive . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
20 A.26 Parallel Random Access Iterator Loop . . . . . . . . . . . . . . . . . . . . . . . 226
21 A.27 Fortran Restrictions on shared and private Clauses with
22 Common Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
23 A.28 The default(none) Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
24 A.29 Race Conditions Caused by Implied Copies of Shared Variable
25 in Fortran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
26 A.30 The private Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
27 A.31 Reprivatization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
28 A.32 Fortran Restrictions on Storage Association with the
29 private Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
30 A.33 C/C++ Arrays in a firstprivate Clause . . . . . . . . . . . . . . . . . . . 240

31 v
1 A.34 The lastprivate Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
2 A.35 The reduction Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
3 A.36 The copyin Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
4 A.37 The copyprivate Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5 A.38 Nested Loop Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6 A.39 Restrictions on Nesting of Regions . . . . . . . . . . . . . . . . . . . . . . . . . . 258
7 A.40 The omp_set_dynamic and
8 omp_set_num_threads Routines . . . . . . . . . . . . . . . . . . . . . . . . . 265
9 A.41 The omp_get_num_threads Routine . . . . . . . . . . . . . . . . . . . . . . 266
10 A.42 The omp_init_lock Routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
11 A.43 Ownership of Locks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
12 A.44 Simple Lock Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
13 A.45 Nestable Lock Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274

14 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 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

18 C.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
19 C.2 Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292

20 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 . . . . . . . . . . . . . 304
23 D.3 Example of a Fortran 90 Interface Declaration module . . . . . . . . . . 306
24 D.4 Example of a Generic Interface for a Library Routine . . . . . . . . . . . . 310

25 E. Implementation Defined Behaviors in OpenMP . . . . . . . . . . . . . . . . . . . 311

26 F. Changes from Version 2.5 to Version 3.0 . . . . . . . . . . . . . . . . . . . . . . . . 315

27 vi OpenMP ABI • Version 3.0 May 2008

1
2 CHAPTER 1

3 Introduction
4

5 This document specifies a collection of compiler directives, library routines, and

6 environment variables that can be used to specify shared-memory parallelism in C, C++
7 and Fortran programs. This functionality collectively defines the specification of the
8 OpenMP Application Program Interface (OpenMP API). This specification provides a
9 model for parallel programming that is portable across shared memory architectures
10 from different vendors. Compilers from numerous vendors support the OpenMP API.
11 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
14 allow users to create and manage parallel programs while permitting portability. The
15 directives extend the C, C++ and Fortran base languages with single program multiple
16 data (SPMD) constructs, tasking constructs, worksharing constructs, and
17 synchronization constructs, and they provide support for sharing and privatizing data.
18 The functionality to control the runtime environment is provided by library routines and
19 environment variables. Compilers that support the OpenMP API often include a
20 command line option to the compiler that activates and allows interpretation of all
21 OpenMP directives.

22

23 1.1 Scope
24 The OpenMP API covers only user-directed parallelization, wherein the user explicitly
25 specifies the actions to be taken by the compiler and runtime system in order to execute
26 the program in parallel. OpenMP-compliant implementations are not required to check
27 for data dependencies, data conflicts, race conditions, or deadlocks, any of which may
28 occur in conforming programs. In addition, compliant implementations are not required
29 to check for code sequences that cause a program to be classified as non-conforming.

30 1
1 The user is responsible for using OpenMP in his application to produce a conforming
2 program. OpenMP does not cover compiler-generated automatic parallelization and
3 directives to the compiler to assist such parallelization.

5 1.2 Glossary

6 1.2.1 Threading Concepts

7 thread An execution entity with a stack and associated static memory, called
8 threadprivate memory.

9 OpenMP thread A thread that is managed by the OpenMP runtime system.

10 thread-safe routine A routine that performs the intended function even when executed
11 concurrently (by more than one thread).

12 1.2.2 OpenMP language terminology

13 base language A programming language that serves as the foundation of the OpenMP
14 specification.

15 COMMENT: See Section 1.6 on page 17 for a listing of current base

16 languages for OpenMP.

17 base program A program written in a base language.

18 structured block For C/C++, an executable statement, possibly compound, with a single entry
19 at the top and a single exit at the bottom, or an OpenMP construct.

20 For Fortran, a block of executable statements with a single entry at the top and
21 a single exit at the bottom, or an OpenMP construct.

22 COMMENTS:

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,
6 or try block is considered to be a structured block if the
7 corresponding compound statement obtained by enclosing it in {
8 and } would be a structured block.

9 For Fortran:

10 • STOP statements are allowed in a structured block.

11 directive In C/C++, a #pragma, and in Fortran, a comment, that specifies OpenMP

12 program behavior.

13 COMMENT: See Section 2.1 on page 22 for a description of OpenMP

14 directive syntax.

15 white space A non-empty sequence of space and/or horizontal tab characters.

16 OpenMP program A program that consists of a base program, annotated with OpenMP directives
17 and runtime library routines.

18 conforming program An OpenMP program that follows all the rules and restrictions of the
19 OpenMP specification.

20 declarative directive An OpenMP directive that may only be placed in a declarative context. A
21 declarative directive has no associated executable user code, but instead has
22 one or more associated user declarations.

23 COMMENT: Only the threadprivate directive is a declarative directive.

24 executable directive An OpenMP directive that is not declarative; i.e., it may be placed in an
25 executable context.

26 COMMENT: All directives except the threadprivate directive are

27 executable directives.

28 stand-alone directive An OpenMP executable directive that has no associated executable user code.

29 COMMENT: Only the barrier, flush, and taskwait directives are

30 stand-alone directives.

31 Chapter 1 Introduction 3
1 simple directive An OpenMP executable directive whose associated user code must be a
2 simple (single, non-compound) executable statement.

3 COMMENT: Only the atomic directive is a simple directive.

4 loop directive An OpenMP executable directive whose associated user code must be a loop
5 nest that is a structured block.
6 COMMENTS:

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
9 are loop directives.

10 associated loop(s) The loop(s) controlled by a loop directive.

11 COMMENT: If the loop directive contains a collapse clause then there
12 may be more than one associated loop.

13 construct An OpenMP executable directive (and for Fortran, the paired end directive, if
14 any) and the associated statement, loop or structured block, if any, not
15 including the code in any called routines; i.e., the lexical extent of an
16 executable directive.

17 region All code encountered during a specific instance of the execution of a given
18 construct or of an OpenMP library routine. A region includes any code in
19 called routines as well as any implicit code introduced by the OpenMP
20 implementation. The generation of a task at the point where a task directive
21 is encountered is a part of the region of the encountering thread, but the
22 explicit task region associated with the task directive is not.

23 COMMENTS:

24 A region may also be thought of as the dynamic or runtime extent of a

25 construct or of an OpenMP library routine.

26 During the execution of an OpenMP program, a construct may give

27 rise to many regions.

28 active parallel region A parallel region that is executed by a team consisting of more than one
29 thread.

30 inactive parallel
31 region A parallel region that is executed by a team of only one thread.

32 4 OpenMP API • Version 3.0 May 2008

1 sequential part All code encountered during the execution of an OpenMP program that is not
2 part of a parallel region corresponding to a parallel construct or a
3 task region corresponding to a task construct.

4 COMMENTS:

5 The sequential part executes as if it were enclosed by an inactive

6 parallel region.

7 Executable statements in called routines may be in both the sequential

8 part and any number of explicit parallel regions at different points
9 in the program execution.

10 master thread The thread that encounters a parallel construct, creates a team, generates
11 a set of tasks, then executes one of those tasks as thread number 0.

12 parent thread The thread that encountered the parallel construct and generated a
13 parallel region is the parent thread of each of the threads in the team of
14 that parallel region. The master thread of a parallel region is the
15 same thread as its parent thread with respect to any resources associated with
16 an OpenMP thread.

17 ancestor thread For a given thread, its parent thread or one of its parent thread’s ancestor
18 threads.

19 team A set of one or more threads participating in the execution of a parallel

20 region.

21 COMMENTS:

22 For an active parallel region, the team comprises the master thread
23 and at least one additional thread.

24 For an inactive parallel region, the team comprises only the master
25 thread.

26 initial thread The thread that executes the sequential part.

27 implicit parallel
28 region The inactive parallel region that encloses the sequential part of an OpenMP
29 program.

30 nested construct A construct (lexically) enclosed by another construct.

31 nested region A region (dynamically) enclosed by another region; i.e., a region encountered
32 during the execution of another region.

33 COMMENT: Some nestings are conforming and some are not. See
34 Section 2.10 on page 104 for the restrictions on nesting.

35 Chapter 1 Introduction 5
1 closely nested region A region nested inside another region with no parallel region nested
2 between them.

3 all threads All OpenMP threads participating in the OpenMP program.

4 current team All threads in the team executing the innermost enclosing parallel region
5 encountering thread For a given region, the thread that encounters the corresponding construct.

6 all tasks All tasks participating in the OpenMP program.

7 current team tasks All tasks encountered during the execution of the innermost enclosing
8 parallel region by the threads of the corresponding team. Note that the
9 implicit tasks constituting the parallel region and any descendant tasks
10 encountered during the execution of these implicit tasks are included in this
11 binding task set.

12 generating task For a given region the task whose execution by a thread generated the region.

13 binding thread set The set of threads that are affected by, or provide the context for, the
14 execution of a region.

15 The binding thread set for a given region can be all threads, the current team,
16 or the encountering thread.

17 COMMENT: The binding thread set for a particular region is described in its
18 corresponding subsection of this specification.

19 binding task set The set of tasks that are affected by, or provide the context for, the execution
20 of a region.

21 The binding task set for a given region can be all tasks, the current team
22 tasks, or the generating task.

23 COMMENT: The binding task set for a particular region (if applicable) is
24 described in its corresponding subsection of this specification.

25 6 OpenMP API • Version 3.0 May 2008

1 binding region The enclosing region that determines the execution context and limits the
2 scope of the effects of the bound region is called the binding region.

3 Binding region is not defined for regions whose binding thread set is all
4 threads or the encountering thread, nor is it defined for regions whose binding
5 task set is all tasks.

6 COMMENTS:
7 The binding region for an ordered region is the innermost enclosing
8 loop region.
9 The binding region for a taskwait region is the innermost enclosing
10 task region.

11 For all other regions for which the binding thread set is the current
12 team or the binding task set is the current team tasks, the binding
13 region is the innermost enclosing parallel region.

14 For regions for which the binding task set is the generating task, the
15 binding region is the region of the generating task.
16 A parallel region need not be active nor explicit to be a binding
17 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

20 parallel region.

21 orphaned construct A construct that gives rise to a region whose binding thread set is the current
22 team, but that is not nested within another construct giving rise to the binding
23 region.

24 worksharing
25 construct A construct that defines units of work, each of which is executed exactly once
26 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

29 workshare.

30 sequential loop A loop that is not associated with any OpenMP loop directive.

31 barrier A point in the execution of a program encountered by a team of threads,

32 beyond which no thread in the team may execute until all threads in the team
33 have reached the barrier and all explicit tasks generated by the team have
34 executed to completion.

35 Chapter 1 Introduction 7
1 1.2.3 Tasking Terminology
2 task A specific instance of executable code and its data environment, generated
3 when a thread encounters a task construct or a parallel construct.

4 COMMENT: When a thread executes a task, it produces a task region.

5 task region A region consisting of all code encountered during the execution of a task.

6 COMMENT: A parallel region consists of one or more implicit task

7 regions.

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
10 parallel construct is encountered during execution.

11 initial task The implicit task associated with the implicit parallel region.

12 current task For a given thread, the task corresponding to the task region in which it is
13 executing.

14 child task A task is a child task of the region of its generating task. A child task region
15 is not part of its generating task region.

16 descendant task A task that is the child task of a task region or of one of its descendant task
17 regions.

18 task completion Task completion occurs when the end of the structured block associated with
19 the construct that generated the task is reached.

20 COMMENT: Completion of the initial task occurs at program exit.

21 task scheduling point A point during the execution of the current task region at which it can be
22 suspended to be resumed later; or the point of task completion, after which the
23 executing thread may switch to a different task region.

24 COMMENT:

25 Within tied task regions, task scheduling points only appear in the
26 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

32 task switching The act of a thread switching from the execution of one task to another task.

33 8 OpenMP API • Version 3.0 May 2008

1 tied task A task that, when its task region is suspended, can be resumed only by the
2 same thread that suspended it; that is, the task is tied to that thread.

3 untied task A task that, when its task region is suspended, can be resumed by any thread
4 in the team; that is, the task is not tied to any thread.

5 task synchronization
6 construct A taskwait or a barrier construct.

7 1.2.4 Data Terminology

8 variable A named data storage block, whose value can be defined and redefined during
9 the execution of a program.

10 Array sections and substrings are not considered variables.

11 private variable With respect to a given set of task regions that bind to the same parallel
12 region, a variable whose name provides access to a different block of storage
13 for each task region.

14 A variable which is part of another variable (as an array or structure element)

15 cannot be made private independently of other components.

16 shared variable With respect to a given set of task regions that bind to the same parallel
17 region, a variable whose name provides access to the same block of storage
18 for each task region.

19 A variable which is part of another variable (as an array or structure element)

20 cannot be shared independently of the other components, except for static data
21 members of C++ classes.

22 threadprivate
23 variable A variable that is replicated, one instance per thread, by the OpenMP
24 implementation, so that its name provides access to a different block of
25 storage for each thread.

26 A variable which is part of another variable (as an array or structure element)

27 cannot be made threadprivate independently of the other components, except
28 for static data members of C++ classes.

29 threadprivate
30 memory The set of threadprivate variables associated with each thread.

31 Chapter 1 Introduction 9
1 data environment All the variables associated with the execution of a given task. The data
2 environment for a given task is constructed from the data environment of the
3 generating task at the time the task is generated.

4 defined 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
9 having a valid value.

10 For variables of non-POD class type, the property of having been constructed
11 but not subsequently destructed.

12 For Fortran:

13 For the contents of variables, the property of having a valid value. For the
14 allocation or association status of variables, the property of having a valid
15 status.

16 COMMENT: Programs that rely upon variables that are not defined are non-
17 conforming programs.

18 class type For C++: Variables declared with one of the class, struct, or union keywords.

19 1.2.5 Implementation Terminology

20 supporting n levels of
21 parallelism Implies allowing an active parallel region to be enclosed by n-1 active
22 parallel regions.

23 supporting OpenMP Supporting at least one level of parallelism.

24 supporting nested
25 parallelism Supporting more than one level of parallelism.

26 internal control
27 variable A conceptual variable that specifies run-time behavior of a set of threads or
28 tasks in an OpenMP program.

29 COMMENT: The acronym ICV is used interchangeably with the term internal
30 control variable in the remainder of this specification.

31 10 OpenMP API • Version 3.0 May 2008

1 compliant
2 implementation An implementation of the OpenMP specification that compiles and executes
3 any conforming program as defined by the specification.

4 COMMENT: A compliant implementation may exhibit unspecified behavior

5 when compiling or executing a non-conforming program.

6 unspecified behavior A behavior or result that is not specified by the OpenMP specification or not
7 known prior to the compilation or execution of an OpenMP program.

8 Such unspecified behavior may result from:

9 • Issues documented by the OpenMP specification as having unspecified

10 behavior.
11 • A non-conforming program.
12 • A conforming program exhibiting an implementation defined behavior.

13 implementation
14 defined Behavior that must be documented by the implementation, and which is
15 allowed to vary among different compliant implementations. An
16 implementation is allowed to define this behavior as unspecified.

17 COMMENT: All features that have implementation defined behavior are

18 documented in Appendix E.

19

20 1.3 Execution Model

21 The OpenMP API uses the fork-join model of parallel execution. Multiple threads of
22 execution perform tasks defined implicitly or explicitly by OpenMP directives. OpenMP
23 is intended to support programs that will execute correctly both as parallel programs
24 (multiple threads of execution and a full OpenMP support library) and as sequential
25 programs (directives ignored and a simple OpenMP stubs library). However, it is
26 possible and permitted to develop a program that executes correctly as a parallel
27 program but not as a sequential program, or that produces different results when
28 executed as a parallel program compared to when it is executed as a sequential program.
29 Furthermore, using different numbers of threads may result in different numeric results
30 because of changes in the association of numeric operations. For example, a serial
31 addition reduction may have a different pattern of addition associations than a parallel
32 reduction. These different associations may change the results of floating-point addition.

33 Chapter 1 Introduction 11
1 An OpenMP program begins as a single thread of execution, called the initial thread.
2 The initial thread executes sequentially, as if enclosed in an implicit task region, called
3 the initial task region, that is defined by an implicit inactive parallel region
4 surrounding the whole program.
5 When any thread encounters a parallel construct, the thread creates a team of itself
6 and zero or more additional threads and becomes the master of the new team. A set of
7 implicit tasks, one per thread, is generated. The code for each task is defined by the code
8 inside the parallel construct. Each task is assigned to a different thread in the team
9 and becomes tied; that is, it is always executed by the thread to which it is initially
10 assigned. The task region of the task being executed by the encountering thread is
11 suspended, and each member of the new team executes its implicit task. There is an
12 implicit barrier at the end of the parallel construct. Beyond the end of the
13 parallel construct, only the master thread resumes execution, by resuming the task
14 region that was suspended upon encountering the parallel construct. Any number of
15 parallel constructs can be specified in a single program.
16 parallel regions may be arbitrarily nested inside each other. If nested parallelism is
17 disabled, or is not supported by the OpenMP implementation, then the new team that is
18 created by a thread encountering a parallel construct inside a parallel region
19 will consist only of the encountering thread. However, if nested parallelism is supported
20 and enabled, then the new team can consist of more than one thread.
21 When any team encounters a worksharing construct, the work inside the construct is
22 divided among the members of the team, and executed cooperatively instead of being
23 executed by every thread. There is an optional barrier at the end of each worksharing
24 construct. Redundant execution of code by every thread in the team resumes after the
25 end of the worksharing construct.
26 When any thread encounters a task construct, a new explicit task is generated.
27 Execution of explicitly generated tasks is assigned to one of the threads in the current
28 team, subject to the thread's availability to execute work. Thus, execution of the new
29 task could be immediate, or deferred until later. Threads are allowed to suspend the
30 current task region at a task scheduling point in order to execute a different task. If the
31 suspended task region is for a tied task, the initially assigned thread later resumes
32 execution of the suspended task region. If the suspended task region is for an untied
33 task, then any thread may resume its execution. In untied task regions, task scheduling
34 points may occur at implementation defined points anywhere in the region. In tied task
35 regions, task scheduling points may occur only in task, taskwait, explicit or
36 implicit barrier constructs, and at the completion point of the task. Completion of all
37 explicit tasks bound to a given parallel region is guaranteed before the master thread
38 leaves the implicit barrier at the end of the region. Completion of a subset of all explicit
39 tasks bound to a given parallel region may be specified through the use of task
40 synchronization constructs. Completion of all explicit tasks bound to the implicit
41 parallel region is guaranteed by the time the program exits.

42 12 OpenMP API • Version 3.0 May 2008

1 Synchronization constructs and library routines are available in OpenMP to coordinate
2 tasks and data access in parallel regions. In addition, library routines and
3 environment variables are available to control or to query the runtime environment of
4 OpenMP programs.
5 OpenMP makes no guarantee that input or output to the same file is synchronous when
6 executed in parallel. In this case, the programmer is responsible for synchronizing input
7 and output statements (or routines) using the provided synchronization constructs or
8 library routines. For the case where each thread accesses a different file, no
9 synchronization by the programmer is necessary.

10

11 1.4 Memory Model

12 1.4.1 Structure of the OpenMP Memory Model

13 OpenMP provides a relaxed-consistency, shared-memory model. All OpenMP threads
14 have access to a place to store and to retrieve variables, called the memory. In addition,
15 each thread is allowed to have its own temporary view of the memory. The temporary
16 view of memory for each thread is not a required part of the OpenMP memory model,
17 but can represent any kind of intervening structure, such as machine registers, cache, or
18 other local storage, between the thread and the memory. The temporary view of memory
19 allows the thread to cache variables and thereby to avoid going to memory for every
20 reference to a variable. Each thread also has access to another type of memory that must
21 not be accessed by other threads, called threadprivate memory.
22 A directive that accepts data-sharing attribute clauses determines two kinds of access to
23 variables used in the directive’s associated structured block: shared and private. Each
24 variable referenced in the structured block has an original variable, which is the variable
25 by the same name that exists in the program immediately outside the construct. Each
26 reference to a shared variable in the structured block becomes a reference to the original
27 variable. For each private variable referenced in the structured block, a new version of
28 the original variable (of the same type and size) is created in memory for each task that
29 contains code associated with the directive. Creation of the new version does not alter
30 the value of the original variable. However, the impact of attempts to access the original
31 variable during the region associated with the directive is unspecified; see
32 Section 2.9.3.3 on page 89 for additional details. References to a private variable in the
33 structured block refer to the current task’s private version of the original variable. The

34 Chapter 1 Introduction 13
1 relationship between the value of the original variable and the initial or final value of the
2 private version depends on the exact clause that specifies it. Details of this issue, as well
3 as other issues with privatization, are provided in Section 2.9 on page 77.
4 The minimum size at which memory accesses by multiple threads without
5 synchronization, either to the same variable or to different variables that are part of the
6 same variable (as array or structure elements), are atomic with respect to each other, is
7 implementation defined. Any additional atomicity restrictions, such as alignment, are
8 implementation defined.
9 A single access to a variable may be implemented with multiple load or store
10 instructions, and hence is not guaranteed to be atomic with respect to other accesses to
11 the same variable. Accesses to variables smaller than the implementation-defined
12 minimum size or to C or C++ bit-fields may be implemented by reading, modifying, and
13 rewriting a larger unit of memory, and may thus interfere with updates of variables or
14 fields in the same unit of memory.
15 If multiple threads write without synchronization to the same memory unit, including
16 cases due to atomicity considerations as described above, then a data race occurs.
17 Similarly, if at least one thread reads from a memory unit and at least one thread writes
18 without synchronization to that same memory unit, including cases due to atomicity
19 considerations as described above, then a data race occurs. If a data race occurs then the
20 result of the program is unspecified.
21 A private variable in a task region that eventually generates an inner nested parallel
22 region is permitted to be made shared by implicit tasks in the inner parallel region.
23 A private variable in a task region can be shared by an explicit task region generated
24 during its execution. However, it is the programmer’s responsibility to ensure through
25 synchronization that the lifetime of the variable does not end before completion of the
26 explicit task region sharing it. Any other access by one task to the private variables of
27 another task results in unspecified behavior.

28 1.4.2 The Flush Operation

29 The memory model has relaxed-consistency because a thread’s temporary view of
30 memory is not required to be consistent with memory at all times. A value written to a
31 variable can remain in the thread’s temporary view until it is forced to memory at a later
32 time. Likewise, a read from a variable may retrieve the value from the thread’s
33 temporary view, unless it is forced to read from memory. The OpenMP flush operation
34 enforces consistency between the temporary view and memory.

35 14 OpenMP API • Version 3.0 May 2008

1 The flush operation is applied to a set of variables called the flush-set. The flush
2 operation restricts reordering of memory operations that an implementation might
3 otherwise do. Implementations must not reorder the code for a memory operation for a
4 given variable, or the code for a flush operation for the variable, with respect to a flush
5 operation that refers to the same variable.
6 If a thread has performed a write to its temporary view of a shared variable since its last
7 flush of that variable, then when it executes another flush of the variable, the flush does
8 not complete until the value of the variable has been written to the variable in memory.
9 If a thread performs multiple writes to the same variable between two flushes of that
10 variable, the flush ensures that the value of the last write is written to the variable in
11 memory. A flush of a variable executed by a thread also causes its temporary view of the
12 variable to be discarded, so that if its next memory operation for that variable is a read,
13 then the thread will read from memory when it may again capture the value in the
14 temporary view. When a thread executes a flush, no later memory operation by that
15 thread for a variable involved in that flush is allowed to start until the flush completes.
16 The completion of a flush of a set of variables executed by a thread is defined as the
17 point at which all writes to those variables performed by the thread before the flush are
18 visible in memory to all other threads and that thread’s temporary view of all variables
19 involved is discarded.
20 The flush operation provides a guarantee of consistency between a thread’s temporary
21 view and memory. Therefore, the flush operation can be used to guarantee that a value
22 written to a variable by one thread may be read by a second thread. To accomplish this,
23 the programmer must ensure that the second thread has not written to the variable since
24 its last flush of the variable, and that the following sequence of events happens in the
25 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.

30

31 Note – OpenMP synchronization operations, described in Section 2.8 on page 63 and in

32 Section 3.3 on page 134, are recommended for enforcing this order. Synchronization
33 through variables is possible; however, it is not recommended since proper timing of
34 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
2 The type of relaxed memory consistency provided by OpenMP is similar to weak
3 ordering as described in S. V. Adve and K. Gharachorloo, “Shared Memory Consistency
4 Models: A Tutorial”, IEEE Computer, 29(12), pp.66-76, December 1996. Weak ordering
5 requires that some memory operations be defined as synchronization operations and that
6 these be ordered with respect to each other. In the context of OpenMP, two flushes of the
7 same variable are synchronization operations. OpenMP does not apply any other
8 restriction to the reordering of memory operations executed by a single thread. The
9 OpenMP memory model is slightly weaker than weak ordering since flushes are not
10 ordered with respect to each other if their flush-sets have an empty intersection.
11 The restrictions in Section 1.4.2 on page 14 on reordering with respect to flush
12 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,
15 seen by all threads.
16 • If the intersection of the flush-sets of two flushes performed by one thread is non-
17 empty, then the two flushes must appear to be completed in that thread’s program
18 order.
19 • If the intersection of the flush-sets of two flushes is empty, the threads can observe
20 these flushes in any order.
21 The flush operation can be specified using the flush directive, and is also implied at
22 various locations in an OpenMP program: see Section 2.8.6 on page 72 for details. For
23 an example illustrating the memory model, see Section A.2 on page 154.

24

25 1.5 OpenMP Compliance

26 An implementation of the OpenMP API is compliant if and only if it compiles and
27 executes all conforming programs according to the syntax and semantics laid out in
28 Chapters 1, 2, 3 and 4. Appendices A, B, C, D, E and F and sections designated as Notes
29 (see Section 1.7 on page 18) are for information purposes only and are not part of the
30 specification.
31 The OpenMP API defines constructs that operate in the context of the base language that
32 is supported by an implementation. If the base language does not support a language
33 construct that appears in this document, a compliant OpenMP implementation is not
34 required to support it, with the exception that for Fortran, the implementation must
35 allow case insensitivity for directive and API routines names, and must allow identifiers
36 of more than six characters.

37 16 OpenMP API • Version 3.0 May 2008

1 All library, intrinsic and built-in routines provided by the base language must be thread-
2 safe in a compliant implementation. In addition, the implementation of the base
3 language must also be thread-safe (e.g., ALLOCATE and DEALLOCATE statements must
4 be thread-safe in Fortran). Unsynchronized concurrent use of such routines by different
5 threads must produce correct results (although not necessarily the same as serial
6 execution results, as in the case of random number generation routines).
7 In both Fortran 90 and Fortran 95, variables with explicit initialization have the SAVE
8 attribute implicitly. This is not the case in Fortran 77. However, a compliant OpenMP
9 Fortran implementation must give such a variable the SAVE attribute, regardless of the
10 underlying base language version.
11 Appendix E lists certain aspects of the OpenMP API that are implementation-defined. A
12 compliant implementation is required to define and document its behavior for each of
13 the items in Appendix E.

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
3 the base language supported by the implementation.

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
17 language. Text that applies only to programs whose base language is C or C++ is shown
18 as follows:
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
26 at the top of the page like this:
Fortran (cont.)

27 Some text is for information only, and is not part of the normative specification. Such
28 text is designated as a note, like this:

29 18 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 CHAPTER 2

3 Directives
4

5 This chapter describes the syntax and behavior of OpenMP directives, and is divided
6 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
11 104)
12
C/C++
13 In C/C++, OpenMP directives are specified by using the #pragma mechanism provided
14 by the C and C++ standards.
C/C++
15
Fortran
16 In Fortran, OpenMP directives are specified by using special comments that are
17 identified by unique sentinels. Also, a special comment form is available for conditional
18 compilation.
19
Fortran
20 Compilers can therefore ignore OpenMP directives and conditionally compiled code if
21 support of OpenMP is not provided or enabled. A compliant implementation must
22 provide an option or interface that ensures that underlying support of all OpenMP
23 directives and OpenMP conditional compilation mechanisms is enabled. In the
24 remainder of this document, the phrase OpenMP compilation is used to mean a
25 compilation with these OpenMP features enabled.

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

7 2.1 Directive Format

8
C/C++
9 OpenMP directives for C/C++ are specified with the pragma preprocessing directive.
10 The syntax of an OpenMP directive is formally specified by the grammar in
11 Appendix C, and informally as follows:
12
13 #pragma omp directive-name [clause[ [,] clause]...] new-line

14 Each directive starts with #pragma omp. The remainder of the directive follows the
15 conventions of the C and C++ standards for compiler directives. In particular, white
16 space can be used before and after the #, and sometimes white space must be used to
17 separate the words in a directive. Preprocessing tokens following the #pragma omp
18 are subject to macro replacement.
19 Directives are case-sensitive.
20 An OpenMP executable directive applies to at most one succeeding statement, which
21 must be a structured block.
C/C++
22
Fortran
23 OpenMP directives for Fortran are specified as follows:
24
25 sentinel directive-name [clause[[,] clause]...]

26 All OpenMP compiler directives must begin with a directive sentinel. The format of a
27 sentinel differs between fixed and free-form source files, as described in Section 2.1.1
28 on page 23 and Section 2.1.2 on page 24.

29 Directives are case-insensitive. Directives cannot be embedded within continued

30 statements, and statements cannot be embedded within directives.

31 22 OpenMP API • Version 3.0 May 2008

1 In order to simplify the presentation, free form is used for the syntax of OpenMP
2 directives for Fortran in the remainder of this document, except as noted.
3
Fortran
4 Only one directive-name can be specified per directive (note that this includes combined
5 directives, see Section 2.6 on page 54). The order in which clauses appear on directives
6 is not significant. Clauses on directives may be repeated as needed, subject to the
7 restrictions listed in the description of each clause.
8 Some data-sharing attribute clauses (Section 2.9.3 on page 85), data copying clauses
9 (Section 2.9.4 on page 100), the threadprivate directive (Section 2.9.2 on page 81)
10 and the flush directive (Section 2.8.6 on page 72) accept a list. A list consists of a
11 comma-separated collection of one or more list items.
12
C/C++
13 A list item is a variable name, subject to the restrictions specified in each of the sections
14 describing clauses and directives for which a list appears.
C/C++
15
Fortran
16 A list item is a variable name or a common block name (enclosed in slashes), subject to
17 the restrictions specified in each of the sections describing clauses and directives for
18 which a list appears.
19
Fortran

20
Fortran
21 2.1.1 Fixed Source Form Directives
22 The following sentinels are recognized in fixed form source files:
23
24 !$omp | c$omp | *$omp

25 Sentinels must start in column 1 and appear as a single word with no intervening
26 characters. Fortran fixed form line length, white space, continuation, and column rules
27 apply to the directive line. Initial directive lines must have a space or zero in column 6,
28 and continuation directive lines must have a character other than a space or a zero in
29 column 6.

30 Comments may appear on the same line as a directive. The exclamation point initiates a
31 comment when it appears after column 6. The comment extends to the end of the source
32 line and is ignored. If the first non-blank character after the directive sentinel of an
33 initial or continuation directive line is an exclamation point, the line is ignored.

34 Chapter 2 Directives 23
1
Fortran (cont.)

3 Note – in the following example, the three formats for specifying the directive are
4 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)
10

11 2.1.2 Free Source Form Directives

12 The following sentinel is recognized in free form source files:
13
14 !$omp

15 The sentinel can appear in any column as long as it is preceded only by white space
16 (spaces and tab characters). It must appear as a single word with no intervening
17 character. Fortran free form line length, white space, and continuation rules apply to the
18 directive line. Initial directive lines must have a space after the sentinel. Continued
19 directive lines must have an ampersand as the last nonblank character on the line, prior
20 to any comment placed inside the directive. Continuation directive lines can have an
21 ampersand after the directive sentinel with optional white space before and after the
22 ampersand.

23 Comments may appear on the same line as a directive. The exclamation point initiates a
24 comment. The comment extends to the end of the source line and is ignored. If the first
25 nonblank character after the directive sentinel is an exclamation point, the line is
26 ignored.

27 One or more blanks or horizontal tabs must be used to separate adjacent keywords in
28 directives in free source form, except in the following cases, where white space is
29 optional between the given pair of keywords:

30 24 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

14 Note – in the following example the three formats for specifying the directive are
15 equivalent (the first line represents the position of the first 9 columns):
16 !23456789
17 !$omp parallel do &
18 !$omp shared(a,b,c)

19 !$omp parallel &

20 !$omp&do shared(a,b,c)

21 !$omp paralleldo shared(a,b,c)

22

23
24
Fortran

25 Chapter 2 Directives 25
1

2 2.2 Conditional Compilation

3 In implementations that support a preprocessor, the _OPENMP macro name is defined to
4 have the decimal value yyyymm where yyyy and mm are the year and month designations
5 of the version of the OpenMP API that the implementation supports.

6 If this macro is the subject of a #define or a #undef preprocessing directive, the

7 behavior is unspecified.

8 For examples of conditional compilation, see Section A.3 on page 161.

9
Fortran
10 The OpenMP API requires Fortran lines to be compiled conditionally, as described in
11 the following sections.

12 2.2.1 Fixed Source Form Conditional Compilation

13 Sentinels
14 The following conditional compilation sentinels are recognized in fixed form source
15 files:
16
17 !$ | *$ | c$

18 To enable conditional compilation, a line with a conditional compilation sentinel must

19 satisfy the following criteria:
20 • The sentinel must start in column 1 and appear as a single word with no intervening
21 white space.
22 • After the sentinel is replaced with two spaces, initial lines must have a space or zero
23 in column 6 and only white space and numbers in columns 1 through 5.
24 • After the sentinel is replaced with two spaces, continuation lines must have a
25 character other than a space or zero in column 6 and only white space in columns 1
26 through 5.

27 If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not
28 met, the line is left unchanged.

29 26 OpenMP API • Version 3.0 May 2008

1
Fortran (cont.)

3 Note – in the following example, the two forms for specifying conditional compilation
4 in fixed source form are equivalent (the first line represents the position of the first 9
5 columns):
6 c23456789
7 !$ 10 iam = omp_get_thread_num() +
8 !$ & index

9 #ifdef _OPENMP
10 10 iam = omp_get_thread_num() +
11 & index
12 #endif
13

14 2.2.2 Free Source Form Conditional Compilation

15 Sentinel
16 The following conditional compilation sentinel is recognized in free form source files:
17
18 !$

19 To enable conditional compilation, a line with a conditional compilation sentinel must

20 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,
25 prior to any comment appearing on the conditionally compiled line. (Continued lines
26 can have an ampersand after the sentinel, with optional white space before and after
27 the ampersand.)
28 If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not
29 met, the line is left unchanged.

30 Chapter 2 Directives 27
1

2 Note – in the following example, the two forms for specifying conditional compilation
3 in free source form are equivalent (the first line represents the position of the first 9
4 columns):
5 c23456789
6 !$ iam = omp_get_thread_num() + &
7 !$& index

8 #ifdef _OPENMP
9 iam = omp_get_thread_num() + &
10 index
11 #endif
12

13
14
Fortran

15

16 2.3 Internal Control Variables

17 An OpenMP implementation must act as if there were internal control variables (ICVs)
18 that control the behavior of an OpenMP program. These ICVs store information such as
19 the number of threads to use for future parallel regions, the schedule to use for
20 worksharing loops and whether nested parallelism is enabled or not. The ICVs are given
21 values at various times (described below) during the execution of the program. They are
22 initialized by the implementation itself and may be given values through OpenMP
23 environment variables and through calls to OpenMP API routines. The program can
24 retrieve the values of these ICVs only through OpenMP API routines.
25 For purposes of exposition, this document refers to the ICVs by certain names, but an
26 implementation is not required to use these names or to offer any way to access the
27 variables other than through the ways shown in Section 2.3.2 on page 29.

28 2.3.1 ICV Descriptions

29 The following ICVs store values that affect the operation of parallel regions.

30 28 OpenMP API • Version 3.0 May 2008

1 • dyn-var - controls whether dynamic adjustment of the number of threads is enabled
2 for encountered parallel regions. There is one copy of this ICV per task.
3 • nest-var - controls whether nested parallelism is enabled for encountered parallel
4 regions. There is one copy of this ICV per task.
5 • nthreads-var - controls the number of threads requested for encountered parallel
6 regions. There is one copy of this ICV per task.
7 • thread-limit-var - controls the maximum number of threads participating in the
8 OpenMP program. There is one copy of this ICV for the whole program.
9 • max-active-levels-var - controls the maximum number of nested active parallel
10 regions. There is one copy of this ICV for the whole program.
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
13 loop regions. There is one copy of this ICV per task.
14 • def-sched-var - controls the implementation defined default scheduling of loop
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
18 creates. There is one copy this ICV for the whole program.
19 • wait-policy-var - controls the desired behavior of waiting threads. There is one copy
20 of this ICV for the whole program.

21 2.3.2 Modifying and Retrieving ICV Values

22 The following table shows the methods for retrieving the values of the ICVs as well as
23 their initial values:
24
25 ICV Scope Ways to modify value Way to retrieve value Initial value

26 dyn-var task OMP_DYNAMIC omp_get_dynamic() See comments

omp_set_dynamic() below

27 nest-var task OMP_NESTED omp_get_nested() false

omp_set_nested()

28 nthreads-var task OMP_NUM_THREADS omp_get_max_threads() Implementation

omp_set_num_threads() defined

29 run-sched-var task OMP_SCHEDULE omp_get_schedule() Implementation

omp_set_schedule() defined

30 def-sched-var global (none) (none) Implementation

defined

31 stacksize-var global OMP_STACKSIZE (none) Implementation

defined

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-active- global OMP_MAX_ACTIVE_LEVELS omp_get_max_active_ See comments

levels-var omp_set_max_active_ levels()
levels() 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
10 the implementation supports. See the definition of supporting n levels of parallelism
11 in Section 1.2.5 on page 10 for further details.
12 After the initial values are assigned, but before any OpenMP construct or OpenMP API
13 routine executes, the values of any OpenMP environment variables that were set by the
14 user are read and the associated ICVs are modified accordingly. After this point, no
15 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.

17 2.3.3 How the Per-task ICVs Work

18 Each task region has its own copies of the internal variables dyn-var, nest-var,
19 nthreads-var, and run-sched-var. When a task construct or parallel construct is
20 encountered during the execution of a task, the child task(s) generated inherit the values
21 of these ICVs from the generating task's ICV values.
22 Calls to omp_set_num_threads(), omp_set_dynamic(),
23 omp_set_nested(), and omp_set_schedule() modify only the ICVs
24 corresponding to their binding task region.
25 When encountering a loop worksharing region with schedule(runtime), all
26 implicit task regions that constitute the binding parallel region must have the same
27 value for run-sched-var. Otherwise, the behavior is unspecified.

28 2.3.4 ICV Override Relationships

29 The override relationships among various construct clauses, OpenMP API routines,
30 environment variables, and the initial values of ICVs are shown in the following table:

31 30 OpenMP API • Version 3.0 May 2008

1
2 overrides setting of
3 construct clause, if environment overrides initial
4 used overrides call to API routine variable value of

5 (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_ OMP_MAX_ACTIVE_ max-active-levels-var

levels() LEVELS

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.

7 2.4 parallel Construct

8 Summary
9 This fundamental construct starts parallel execution. See Section 1.3 on page 11 for a
10 general description of the OpenMP execution model.

11 Syntax
12
C/C++
13 The syntax of the parallel construct is as follows:
14
15 #pragma omp parallel [clause[ [, ]clause] ...] new-line
16 structured-block

17 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 !$omp parallel [clause[[,] clause]...]
5 structured-block
6 !$omp end parallel

7 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 The binding thread set for a parallel region is the encountering thread. The
20 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
23 execute the parallel region (see Section 2.4.1 on page 35 for more information about
24 how the number of threads in the team is determined, including the evaluation of the if
25 and num_threads clauses). The thread that encountered the parallel construct
26 becomes the master thread of the new team, with a thread number of zero for the
27 duration of the new parallel region. All threads in the new team, including the
28 master thread, execute the region. Once the team is created, the number of threads in the
29 team remains constant for the duration of that parallel region.

30 Chapter 2 Directives 33
1 Within a parallel region, thread numbers uniquely identify each thread. Thread
2 numbers are consecutive whole numbers ranging from zero for the master thread up to
3 one less than the number of threads in the team. A thread may obtain its own thread
4 number by a call to the omp_get_thread_num library routine.
5 A set of implicit tasks, equal in number to the number of threads in the team, is
6 generated by the encountering thread. The structured block of the parallel construct
7 determines the code that will be executed in each implicit task. Each task is assigned to
8 a different thread in the team and becomes tied. The task region of the task being
9 executed by the encountering thread is suspended and each thread in the team executes
10 its implicit task. Each thread can execute a path of statements that is different from that
11 of the other threads.
12 The implementation may cause any thread to suspend execution of its implicit task at a
13 task scheduling point, and switch to execute any explicit task generated by any of the
14 threads in the team, before eventually resuming execution of the implicit task (for more
15 details see Section 2.7 on page 59).
16 There is an implied barrier at the end of a parallel region. After the end of a
17 parallel region, only the master thread of the team resumes execution of the
18 enclosing task region.
19 If a thread in a team executing a parallel region encounters another parallel
20 directive, it creates a new team, according to the rules in Section 2.4.1 on page 35, and
21 it becomes the master of that new team.
22 If execution of a thread terminates while inside a parallel region, execution of all
23 threads in all teams terminates. The order of termination of threads is unspecified. All
24 work done by a team prior to any barrier that the team has passed in the program is
25 guaranteed to be complete. The amount of work done by each thread after the last
26 barrier that it passed and before it terminates is unspecified.
27 For an example of the parallel construct, see Section A.5 on page 164. For an
28 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
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 expression must evaluate to a positive integer value.

37 34 OpenMP API • Version 3.0 May 2008

1
C/C++
2 • A throw executed inside a parallel region must cause execution to resume
3 within the same parallel region, and the same thread that threw the exception
4 must catch it.
C/C++
5
Fortran
6 • Unsynchronized use of Fortran I/O statements by multiple threads on the same unit
7 has unspecified behavior.
8
Fortran

9 Cross References
10 • default, shared, private, firstprivate, and reduction clauses, see
11 Section 2.9.3 on page 85.
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 2.4.1 Determining the Number of Threads for a

15 parallel Region
16 When execution encounters a parallel directive, the value of the if clause or
17 num_threads clause (if any) on the directive, the current parallel context, and the
18 values of the nthreads-var, dyn-var, thread-limit-var, max-active-level-var, and nest-var
19 ICVs are used to determine the number of threads to use in the region.
20 Note that using a variable in an if or num_threads clause expression of a
21 parallel construct causes an implicit reference to the variable in all enclosing
22 constructs. The if clause expression and the num_threads clause expression are
23 evaluated in the context outside of the parallel construct, and no ordering of those
24 evaluations is specified. It is also unspecified whether, in what order, or how many times
25 any side-effects of the evaluation of the num_threads or if clause expressions occur.
26 When a thread encounters a parallel construct, the number of threads is determined
27 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 then let ThreadsRequested be the value of the num_threads clause
7 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 Note – Since the initial value of the dyn-var ICV is implementation defined, programs
26 that depend on a specific number of threads for correct execution should explicitly
27 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
3 Section 2.3 on page 28.

5 2.5 Worksharing Constructs

6 A worksharing construct distributes the execution of the associated region among the
7 members of the team that encounters it. Threads execute portions of the region in the
8 context of the implicit tasks each one is executing. If the team consists of only one
9 thread then the worksharing region is not executed in parallel.
10 A worksharing region has no barrier on entry; however, an implied barrier exists at the
11 end of the worksharing region, unless a nowait clause is specified. If a nowait
12 clause is present, an implementation may omit the barrier at the end of the worksharing
13 region. In this case, threads that finish early may proceed straight to the instructions
14 following the worksharing region without waiting for the other members of the team to
15 finish the worksharing region, and without performing a flush operation (see
16 Section A.9 on page 170 for an example.)
17 OpenMP defines the following worksharing constructs, and these are described in the
18 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
26 all.
27 • The sequence of worksharing regions and barrier regions encountered must be the
28 same for every thread in a team.

29 Chapter 2 Directives 37
1 2.5.1 Loop Construct

2 Summary
3 The loop construct specifies that the iterations of one or more associated loops will be
4 executed in parallel by threads in the team in the context of their implicit tasks. The
5 iterations are distributed across threads that already exist in the team executing the
6 parallel region to which the loop region binds.

7 Syntax
8
C/C++
9 The syntax of the loop construct is as follows:
10
11 #pragma omp for [clause[[,] clause] ... ] new-line
12 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
C/C++ (cont.)
3 The for directive places restrictions on the structure of all associated for-loops.
4 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 One of the following:
8 var = lb
9 integer-type var = lb
10 random-access-iterator-type var = lb
11 pointer-type var = lb
12 test-expr One of the following:
13 var relational-op b
14 b relational-op var
15 incr-expr One of the following:
16 ++var
17 var++
18 --var
19 var--
20 var += incr
21 var -= incr
22 var = var + incr
23 var = incr + var
24 var = var - incr
25 var One of the following:
26 A variable of a signed or unsigned integer type.
27 For C++, a variable of a random access iterator type.
28 For C, a variable of a pointer type.
29 If this variable would otherwise be shared, it is implicitly made
30 private in the loop construct. This variable must not be
31 modified during the execution of the for-loop other than in incr-
32 expr. Unless the variable is specified lastprivate on the
33 loop construct, its value after the loop is unspecified.
34 relational-op One of the following:
35 <
36 <=
37 >
38 >=
39 lb and b Loop invariant expressions of a type compatible with the type
40 of var.
41 incr A loop invariant integer expression.

42 Chapter 2 Directives 39
1 The canonical form allows the iteration count of all associated loops to be computed
2 before executing the outermost loop. The computation is performed for each loop in an
3 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
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
9 count cannot be represented in the type determined above.
10 There is no implied synchronization during the evaluation of the lb, b, or incr
11 expressions. It is unspecified whether, in what order, or how many times any side effects
12 within the lb, b, or incr expressions occur.
13

14 Note – Random access iterators are required to support random access to elements in
15 constant time. Other iterators are precluded by the restrictions since they can take linear
16 time or offer limited functionality. It is therefore advisable to use tasks to parallelize
17 those cases.
18

19
C/C++

20
Fortran
21 The syntax of the loop construct is as follows:
22
23 !$omp do [clause[[,] clause] ... ]
24 do-loops
25 [!$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 If an end do directive is not specified, an end do directive is assumed at the end of the
5 do-loop.

6 All associated do-loops must be do-constructs as defined by the Fortran standard. If an

7 end do directive follows a do-construct in which several loop statements share a DO
8 termination statement, then the directive can only be specified for the outermost of these
9 DO statements. See Section A.7 on page 167 for examples.
10 If any of the loop iteration variables would otherwise be shared, they are implicitly
11 made private on the loop construct. See Section A.8 on page 169 for examples. Unless
12 the loop iteration variables are specified lastprivate on the loop construct, their
13 values after the loop are unspecified.
14
Fortran

15 Binding
16 The binding thread set for a loop region is the current team. A loop region binds to the
17 innermost enclosing parallel region. Only the threads of the team executing the
18 binding parallel region participate in the execution of the loop iterations and
19 (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
22 follow the directive.
23 There is an implicit barrier at the end of a loop construct unless a nowait clause is
24 specified.
25 The collapse clause may be used to specify how many loops are associated with the
26 loop construct. The parameter of the collapse clause must be a constant positive
27 integer expression. If no collapse clause is present, the only loop that is associated
28 with the loop construct is the one that immediately follows the construct.
29 If more than one loop is associated with the loop construct, then the iterations of all
30 associated loops are collapsed into one larger iteration space which is then divided
31 according to the schedule clause. The sequential execution of the iterations in all
32 associated loops determines the order of the iterations in the collapsed iteration space.

33 Chapter 2 Directives 41
1 The iteration count for each associated loop is computed before entry to the outermost
2 loop. If execution of any associated loop changes any of the values used to compute any
3 of the iteration counts then the behavior is unspecified.
4 The integer type (or kind, for Fortran) used to compute the iteration count for the
5 collapsed loop is implementation defined.
6 A worksharing loop has logical iterations numbered 0,1,...,N-1 where N is the number of
7 loop iterations, and the logical numbering denotes the sequence in which the iterations
8 would be executed if the associated loop(s) were executed by a single thread. The
9 schedule clause specifies how iterations of the associated loops are divided into
10 contiguous non-empty subsets, called chunks, and how these chunks are distributed
11 among threads of the team. Each thread executes its assigned chunk(s) in the context of
12 its implicit task. The chunk_size expression is evaluated using the original list items of
13 any variables that are made private in the loop construct. It is unspecified whether, in
14 what order, or how many times, any side-effects of the evaluation of this expression
15 occur. The use of a variable in a schedule clause expression of a loop construct
16 causes an implicit reference to the variable in all enclosing constructs.
17 Different loop regions with the same schedule and iteration count, even if they occur in
18 the same parallel region, can distribute iterations among threads differently. The only
19 exception is for the static schedule as specified in Table 2-1. Programs that depend
20 on which thread executes a particular iteration under any other circumstances are non-
21 conforming.
22 See Section 2.5.1.1 on page 45 for details of how the schedule for a worksharing loop is
23 determined.

24 The schedule kind can be one of those specified in Table 2-1.

25 42 OpenMP API • Version 3.0 May 2008

1 TABLE 2-1 schedule clause kind values
2
3 static When schedule(static, chunk_size) is specified, iterations are divided
4 into chunks of size chunk_size, and the chunks are assigned to the threads in
5 the team in a round-robin fashion in the order of the thread number.

6 When no chunk_size is specified, the iteration space is divided into chunks that
7 are approximately equal in size, and at most one chunk is distributed to each
8 thread. Note that the size of the chunks is unspecified in this case.

9 A compliant implementation of static schedule must ensure that the same

10 assignment of logical iteration numbers to threads will be used in two loop
11 regions if the following conditions are satisfied: 1) both loop regions have the
12 same number of loop iterations, 2) both loop regions have the same value of
13 chunk_size specified, or both loop regions have no chunk_size specified, and 3)
14 both loop regions bind to the same parallel region. A data dependence between
15 the same logical iterations in two such loops is guaranteed to be satisfied
16 allowing safe use of the nowait clause (see Section A.9 on page 170 for
17 examples).
18 dynamic When schedule(dynamic, chunk_size) is specified, the iterations are
19 distributed to threads in the team in chunks as the threads request them. Each
20 thread executes a chunk of iterations, then requests another chunk, until no
21 chunks remain to be distributed.

22 Each chunk contains chunk_size iterations, except for the last chunk to be
23 distributed, which may have fewer iterations.

24 When no chunk_size is specified, it defaults to 1.

25 guided When schedule(guided, chunk_size) is specified, the iterations are
26 assigned to threads in the team in chunks as the executing threads request
27 them. Each thread executes a chunk of iterations, then requests another chunk,
28 until no chunks remain to be assigned.

29 For a chunk_size of 1, the size of each chunk is proportional to the

30 number of unassigned iterations divided by the number of threads in the team,
31 decreasing to 1. For a chunk_size with value k (greater than 1), the
32 size of each chunk is determined in the same way, with the restriction
33 that the chunks do not contain fewer than k iterations (except for the last chunk
34 to be assigned, which may have fewer than k iterations).

35 When no chunk_size is specified, it defaults to 1.

36 auto When schedule(auto) is specified, the decision regarding scheduling is
37 delegated to the compiler and/or runtime system. The programmer gives the
38 implementation the freedom to choose any possible mapping of iterations to
39 threads in the team.

40 Chapter 2 Directives 43
1 runtime When schedule(runtime) is specified, the decision regarding scheduling
2 is deferred until run time, and the schedule and chunk size are taken from the
3 run-sched-var ICV. If the ICV is set to auto, the schedule is implementation
4 defined.
5

6
Note – For a team of p threads and a loop of n iterations, let n ⁄ p be the integer q
7 that satisfies n = p*q - r, with 0 <= r < p. One compliant implementation of the
8 static schedule (with no specified chunk_size) would behave as though chunk_size
9 had been specified with value q. Another compliant implementation would assign q
10 iterations to the first p-r threads, and q-1 iterations to the remaining r threads. This
11 illustrates why a conforming program must not rely on the details of a particular
12 implementation.

13 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
14 n-q and p*k. It would then repeat this process until q is greater than or equal to the
15 number of remaining iterations, at which time the remaining iterations form the final
16 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 ever binds to a loop region arising from the loop construct.

34 44 OpenMP API • Version 3.0 May 2008

1 • The loop iteration variable may not appear in a threadprivate directive.
2
C/C++
3 • The associated for-loops must be structured blocks.
4 • Only an iteration of the innermost associated loop may be curtailed by a continue
5 statement.
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
10 decrease on each iteration of the loop.
11 • A throw executed inside a loop region must cause execution to resume within the
12 same iteration of the loop region, and the same thread that threw the exception must
13 catch it.
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
17 statement.
18 • No statement in the associated loops other than the DO statements can cause a branch
19 out of the loops.
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
25 Section 2.9.3 on page 85.
26 • OMP_SCHEDULE environment variable, see Section 4.1 on page 146.
27 • ordered construct, see Section 2.8.7 on page 75.

28 2.5.1.1 Determining the Schedule of a Worksharing Loop

29 When execution encounters a loop directive, the schedule clause (if any) on the
30 directive, and the run-sched-var and def-sched-var ICVs are used to determine how loop
31 iterations are assigned to threads. See Section 2.3 on page 28 for details of how the
32 values of the ICVs are determined. If the loop directive does not have a schedule

33 Chapter 2 Directives 45
1
clause then the current value of the def-sched-var ICV determines the schedule. If the
2 loop directive has a schedule clause that specifies the runtime schedule kind then
3 the current value of the run-sched-var ICV determines the schedule. Otherwise, the
4 value of the schedule clause determines the schedule. Figure 2-1 describes how the
5 schedule for a worksharing loop is determined.

6 Cross References
7 • ICVs, see Section 2.3 on page 28.

8
9 START

10
11

12 schedule No
clause present? Use def-sched-var schedule kind

13
14 Yes
15
16

17 schedule kind No Use schedule kind specified in

value is runtime? schedule clause

18
19 Yes
20 Use run-sched-var schedule kind

21 FIGURE 2-1 Determining the schedule for a worksharing loop.

22 46 OpenMP API • Version 3.0 May 2008

1 2.5.2 sections Construct

2 Summary
3 The sections construct is a noniterative worksharing construct that contains a set of
4 structured blocks that are to be distributed among and executed by the threads in a team.
5 Each structured block is executed once by one of the threads in the team in the context
6 of its implicit task.

7 Syntax
8
C/C++
9 The syntax of the sections construct is as follows:
10
11 #pragma omp sections [clause[[,] clause] ...] new-line
12 {
13 [#pragma omp section new-line]
14 structured-block
15 [#pragma omp section new-line
16 structured-block ]
17 ...
18 }

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 !$omp sections [clause[[,] clause] ...]
5 [!$omp section]
6 structured-block
7 [!$omp section
8 structured-block ]
9 ...
10 !$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 The binding thread set for a sections region is the current team. A sections
20 region binds to the innermost enclosing parallel region. Only the threads of the team
21 executing the binding parallel region participate in the execution of the structured
22 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
25 except possibly the first block, for which a preceding section directive is optional.

26 The method of scheduling the structured blocks among the threads in the team is
27 implementation defined.

28 There is an implicit barrier at the end of a sections construct unless a nowait

29 clause is specified.

30 48 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
4 appear within the sections construct and may not be encountered elsewhere in the
5 sections region.
6 • The code enclosed in a sections construct must be a structured block.
7 • Only a single nowait clause can appear on a sections directive.
8
C/C++
9 • A throw executed inside a sections region must cause execution to resume within
10 the same section of the sections region, and the same thread that threw the
11 exception must catch it.
C/C++

12 Cross References
13 • private, firstprivate, lastprivate, and reduction clauses, see
14 Section 2.9.3 on page 85.

15 2.5.3 single Construct

16 Summary
17 The single construct specifies that the associated structured block is executed by only
18 one of the threads in the team (not necessarily the master thread), in the context of its
19 implicit task. The other threads in the team, which do not execute the block, wait at an
20 implicit barrier at the end of the single construct unless a nowait clause is specified.

21 Syntax
22
C/C++
23 The syntax of the single construct is as follows:
24
25 #pragma omp single [clause[[,] clause] ...] new-line
26 structured-block

27 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 !$omp single [clause[[,] clause] ...]
11 structured-block
12 !$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
23 binds to the innermost enclosing parallel region. Only the threads of the team
24 executing the binding parallel region participate in the execution of the structured
25 block and the (optional) implicit barrier of the single region.

26 50 OpenMP API • Version 3.0 May 2008

1 Description
2 The method of choosing a thread to execute the structured block is implementation
3 defined. There is an implicit barrier at the end of the single construct unless a
4 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 • At most one nowait clause can appear on a single construct.
10
C/C++
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.
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
20 separate units of work, and causes the threads of the team to share the work such that
21 each unit is executed only once by one thread, in the context of its implicit task.

22 Chapter 2 Directives 51
1
Fortran (cont.)
2 Syntax
3 The syntax of the workshare construct is as follows:
4
5 !$omp workshare
6 structured-block
7 !$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 Statements contained in any enclosed critical construct are also subject to these
19 restrictions. Statements in any enclosed parallel construct are not restricted.

20 Binding
21 The binding thread set for a workshare region is the current team. A workshare
22 region binds to the innermost enclosing parallel region. Only the threads of the team
23 executing the binding parallel region participate in the execution of the units of
24 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
27 clause is specified.

28 An implementation of the workshare construct must insert any synchronization that is

29 required to maintain standard Fortran semantics. For example, the effects of one
30 statement within the structured block must appear to occur before the execution of
31 succeeding statements, and the evaluation of the right hand side of an assignment must
32 appear to complete prior to the effects of assigning to the left hand side.

33 52 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
6 ELEMENTAL functions, is a unit of work.
7 • Evaluation of transformational array intrinsic functions may be freely subdivided
8 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 masked assignments are each a unit of work.
13 • For a FORALL statement or construct, the evaluation of the mask expression,
14 expressions occurring in the specification of the iteration space, and the masked
15 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 workshare construct. The statements contained in the parallel construct are
20 executed by a new thread team.
21 • If none of the rules above apply to a portion of a statement in the structured block,
22 then that portion is a unit of work.

23 The transformational array intrinsic functions are MATMUL, DOT_PRODUCT, SUM,

24 PRODUCT, MAXVAL, MINVAL, COUNT, ANY, ALL, SPREAD, PACK, UNPACK,
25 RESHAPE, TRANSPOSE, EOSHIFT, CSHIFT, MINLOC, and MAXLOC.

26 It is unspecified how the units of work are assigned to the threads executing a
27 workshare region.

28 If an array expression in the block references the value, association status, or allocation
29 status of private variables, the value of the expression is undefined, unless the same
30 value would be computed by every thread.

31 If an array assignment, a scalar assignment, a masked array assignment, or a FORALL

32 assignment assigns to a private variable in the block, the result is unspecified.

33 The workshare directive causes the sharing of work to occur only in the workshare
34 construct, and not in the remainder of the workshare region.

35 For examples of the workshare construct, see Section A.14 on page 191.

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
4 ELEMENTAL.
5
Fortran

7 2.6 Combined Parallel Worksharing

8 Constructs
9 Combined parallel worksharing constructs are shortcuts for specifying a worksharing
10 construct nested immediately inside a parallel construct. The semantics of these
11 directives are identical to that of explicitly specifying a parallel construct containing
12 one worksharing construct and no other statements.
13 The combined parallel worksharing constructs allow certain clauses that are permitted
14 both on parallel constructs and on worksharing constructs. If a program would have
15 different behavior depending on whether the clause were applied to the parallel
16 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.

21 2.6.1 Parallel Loop construct

22 Summary
23 The parallel loop construct is a shortcut for specifying a parallel construct
24 containing one loop construct and no other statements.

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 #pragma omp parallel for [clause[[,] clause] ...] new-line
6 for-loop

7 where clause can be any of the clauses accepted by the parallel or for directives,
8 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 !$omp parallel do [clause[[,] clause] ...]
13 do-loop
14 [!$omp end parallel do]

15 where clause can be any of the clauses accepted by the parallel or do directives,
16 with identical meanings and restrictions.

17 If an end parallel do directive is not specified, an end parallel do directive is

18 assumed at the end of the do-loop. nowait may not be specified on an end
19 parallel do directive.
20
Fortran

21 Description
22
C/C++
23 The semantics are identical to explicitly specifying a parallel directive immediately
24 followed by a for directive.
C/C++
25
Fortran
26 The semantics are identical to explicitly specifying a parallel directive immediately
27 followed by a do directive, and an end do directive immediately followed by an end
28 parallel directive.
29
Fortran

30 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 The parallel sections construct is a shortcut for specifying a parallel
10 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 #pragma omp parallel sections [clause[[,] clause] ...] new-line
16 {
17 [#pragma omp section new-line]
18 structured-block
19 [#pragma omp section new-line
20 structured-block ]
21 ...
22 }

23 where clause can be any of the clauses accepted by the parallel or sections
24 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 !$omp parallel sections [clause[[,] clause] ...]
5 [!$omp section]
6 structured-block
7 [!$omp section
8 structured-block ]
9 ...
10 !$omp end parallel sections

11 where clause can be any of the clauses accepted by the parallel or sections
12 directives, with identical meanings and restrictions.
13 The last section ends at the end parallel sections directive. nowait cannot be
14 specified on an end parallel sections directive.
15
Fortran

16 Description
17
C/C++
18 The semantics are identical to explicitly specifying a parallel directive immediately
19 followed by a sections directive.
C/C++
20
Fortran
21 The semantics are identical to explicitly specifying a parallel directive immediately
22 followed by a sections directive, and an end sections directive immediately
23 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 The parallel workshare construct is a shortcut for specifying a parallel
7 construct containing one workshare construct and no other statements.

8 Syntax
9 The syntax of the parallel workshare construct is as follows:
10
11 !$omp parallel workshare [clause[[,] clause] ...]
12 structured-block
13 !$omp end parallel workshare

14 where clause can be any of the clauses accepted by the parallel directive, with
15 identical meanings and restrictions. nowait may not be specified on an end
16 parallel workshare directive.

17 Description
18 The semantics are identical to explicitly specifying a parallel directive immediately
19 followed by a workshare directive, and an end workshare directive immediately
20 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.

26 58 OpenMP API • Version 3.0 May 2008

1 • Data attribute clauses, see Section 2.9.3 on page 85.
2
3
Fortran

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 #pragma omp task [clause[[,] clause] ...] new-line
13 structured-block

14 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 !$omp task [clause[[,] clause] ...]
5 structured-block
6 !$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 The binding thread set of the task region is the current parallel team. A task region
18 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
21 associated structured block. The data environment of the task is created according to the
22 data-sharing attribute clauses on the task construct and any defaults that apply.
23 The encountering thread may immediately execute the task, or defer its execution. In the
24 latter case, any thread in the team may be assigned the task. Completion of the task can
25 be guaranteed using task synchronization constructs. A task construct may be nested
26 inside an outer task, but the task region of the inner task is not a part of the task
27 region of the outer task.
28 When an if clause is present on a task construct and the if clause expression
29 evaluates to false, the encountering thread must suspend the current task region and
30 begin execution of the generated task immediately, and the suspended task region may

31 60 OpenMP API • Version 3.0 May 2008

1 not be resumed until the generated task is completed. The task still behaves as a distinct
2 task region with respect to data environment, lock ownership, and synchronization
3 constructs. Note that the use of a variable in an if clause expression of a task
4 construct causes an implicit reference to the variable in all enclosing constructs.
5 A thread that encounters a task scheduling point within the task region may
6 temporarily suspend the task region. By default, a task is tied and its suspended task
7 region can only be resumed by the thread that started its execution. If the untied
8 clause is present on a task construct, any thread in the team can resume the task
9 region after a suspension.
10 The task construct includes a task scheduling point in the task region of its generating
11 task, immediately following the generation of the explicit task. Each explicit task
12 region includes a task scheduling point at its point of completion. An implementation
13 may add task scheduling points anywhere in untied task regions.
14

15 Note – When storage is shared by an explicit task region, it is the programmer's

16 responsibility to ensure, by adding proper synchronization, that the storage does not
17 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 • At most one if clause can appear on the directive.
25
C/C++
26 • A throw executed inside a task region must cause execution to resume within the
27 same task region, and the same thread that threw the exception must catch it.
C/C++
28
Fortran
29 • Unsynchronized use of Fortran I/O statements by multiple tasks on the same unit has
30 unspecified behavior.
31
Fortran

32 Chapter 2 Directives 61
1 2.7.1 Task Scheduling
2 Whenever a thread reaches a task scheduling point, the implementation may cause it to
3 perform a task switch, beginning or resuming execution of a different task bound to the
4 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
10 that they are not specifically prohibited in this specification.
11 When a thread encounters a task scheduling point it may do one of the following,
12 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 If more than one of the above choices is available, it is unspecified as to which will be
18 chosen.
19 Task Scheduling Constraints
20 1. An explicit task whose construct contained an if clause whose if clause expression
21 evaluated to false is executed immediately after generation of the task.

22 2. Other scheduling of new tied tasks is constrained by the set of task regions that are
23 currently tied to the thread, and that are not suspended in a barrier region. If this set
24 is empty, any new tied task may be scheduled. Otherwise, a new tied task may be
25 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.
27

28 Note – Task scheduling points dynamically divide task regions into parts. Each part is
29 executed uninterruptedly from start to end. Different parts of the same task region are
30 executed in the order in which they are encountered. In the absence of task
31 synchronization constructs, the order in which a thread executes parts of different
32 schedulable tasks is unspecified.
33 A correct program must behave correctly and consistently with all conceivable
34 scheduling sequences that are compatible with the rules above.

35 62 OpenMP API • Version 3.0 May 2008

1 For example, if threadprivate storage is accessed (explicitly in the source code or
2 implicitly in calls to library routines) in one part of a task region, its value cannot be
3 assumed to be preserved into the next part of the same task region if another schedulable
4 task exists that modifies it (see Example A.13.7c on page 186, Example A.13.7f on page
5 186, Example A.13.8c on page 187 and Example A.13.8f on page 187).
6 As another example, if a lock acquire and release happen in different parts of a task
7 region, no attempt should be made to acquire the same lock in any part of another task
8 that the executing thread may schedule; otherwise, a deadlock is possible. A similar
9 situation can occur when a critical region spans multiple parts of a task and another
10 schedulable task contains a critical region with the same name (see Example A.13.9c on
11 page 188, Example A.13.9f on page 189, Example A.13.10c on page 190 and Example
12 A.13.10f on page 191).
13 The use of threadprivate variables and the use of locks or critical sections in an explicit
14 task with an if clause must take into account that when the if clause evaluates to
15 false, the task is executed immediately, without regard to Task Scheduling Constraint 2.
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
30 of the team.

31 Chapter 2 Directives 63
1 Syntax
2
C/C++
3 The syntax of the master construct is as follows:
4
5 #pragma omp master new-line
6 structured-block

7
C/C++
8
Fortran
9 The syntax of the master construct is as follows:
10
11 !$omp master
12 structured-block
13 !$omp end master

14
15
Fortran

16 Binding
17 The binding thread set for a master region is the current team. A master region
18 binds to the innermost enclosing parallel region. Only the master thread of the team
19 executing the binding parallel region participates in the execution of the structured
20 block of the master region.

21 Description
22 Other threads in the team do not execute the associated structured block. There is no
23 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 Restrictions
26
C/C++
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.
C/C++

29 64 OpenMP API • Version 3.0 May 2008

1 2.8.2 critical Construct

2 Summary
3 The critical construct restricts execution of the associated structured block to a
4 single thread at a time.

5 Syntax
6
C/C++
7 The syntax of the critical construct is as follows:
8
9 #pragma omp critical [(name)] new-line
10 structured-block

11
C/C++
12
Fortran
13 The syntax of the critical construct is as follows:
14
15 !$omp critical [(name)]
16 structured-block
17 !$omp end critical [(name)]

18
19
Fortran

20 Binding
21 The binding thread set for a critical region is all threads. Region execution is
22 restricted to a single thread at a time among all the threads in the program, without
23 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
26 constructs without a name are considered to have the same unspecified name. A thread
27 waits at the beginning of a critical region until no thread is executing a critical

28 Chapter 2 Directives 65
1 region with the same name. The critical construct enforces exclusive access with
2 respect to all critical constructs with the same name in all threads, not just those
3 threads in the current team.
4
C/C++
5 Identifiers used to identify a critical construct have external linkage and are in a
6 name space that is separate from the name spaces used by labels, tags, members, and
7 ordinary identifiers.
C/C++
8
Fortran
9 The names of critical constructs are global entities of the program. If a name
10 conflicts with any other entity, the behavior of the program is unspecified.
11
Fortran
12 For an example of the critical construct, see Section A.16 on page 197.

13 Restrictions
14
C/C++
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
17 it.
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
21 specified on the end critical directive.
22 • If no name appears on the critical directive, no name can appear on the end
23 critical directive.
24
Fortran

25 2.8.3 barrier Construct

26 Summary
27 The barrier construct specifies an explicit barrier at the point at which the construct
28 appears.

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 #pragma omp barrier new-line

6 Because the barrier construct does not have a C language statement as part of its
7 syntax, there are some restrictions on its placement within a program. The barrier
8 directive may be placed only at a point where a base language statement is allowed. The
9 barrier directive may not be used in place of the statement following an if, while, do,
10 switch, or label. See Appendix C for the formal grammar. The examples in Section A.23
11 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 The binding thread set for a barrier region is the current team. A barrier region
20 binds to the innermost enclosing parallel region. See Section A.18 on page 200 for
21 examples.

22 Description
23 All threads of the team executing the binding parallel region must execute the
24 barrier region and complete execution of all explicit tasks generated in the binding
25 parallel region up to this point before any are allowed to continue execution beyond
26 the barrier.
27 The barrier region includes an implicit task scheduling point in the current task
28 region.

29 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
5 same for every thread in a team.

6 2.8.4 taskwait Construct

7 Summary
8 The taskwait construct specifies a wait on the completion of child tasks generated
9 since the beginning of the current task.

10 Syntax
11
C/C++
12 The syntax of the taskwait construct is as follows:
13
14 #pragma omp taskwait newline

15 Because the taskwait construct does not have a C language statement as part of its
16 syntax, there are some restrictions on its placement within a program. The taskwait
17 directive may be placed only at a point where a base language statement is allowed. The
18 taskwait directive may not be used in place of the statement following an if, while,
19 do, switch, or label. See Appendix C for the formal grammar. The examples in
20 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
Fortran

27 68 OpenMP API • Version 3.0 May 2008

1 Binding
2 A taskwait region binds to the current task region. The binding thread set of the
3 taskwait region is the current team.

4 Description
5 The taskwait region includes an implicit task scheduling point in the current task
6 region. The current task region is suspended at the task scheduling point until execution
7 of all its child tasks generated before the taskwait region are completed.

8 2.8.5 atomic Construct

9 Summary
10 The atomic construct ensures that a specific storage location is updated atomically,
11 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 #pragma omp atomic new-line
17 expression-stmt

18 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 • expr is an expression with scalar type, and it does not reference the variable
2 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 !$omp atomic
9 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 reference x. When intrinsic_procedure_name refers to IAND, IOR, or IEOR, exactly
20 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
24 of operator, x operator expr must be mathematically equivalent to x operator (expr),
25 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
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
Fortran

31 70 OpenMP API • Version 3.0 May 2008

1 Binding
2 The binding thread set for an atomic region is all threads. atomic regions enforce
3 exclusive access with respect to other atomic regions that update the same storage
4 location x among all the threads in the program without regard to the teams to which the
5 threads belong.

6 Description
7 Only the load and store of the variable designated by x are atomic; the evaluation of
8 expr is not atomic. No task scheduling points are allowed between the load and the store
9 of the variable designated by x. To avoid race conditions, all updates of the location that
10 could potentially occur in parallel must be protected with an atomic directive.
11 atomic regions do not enforce exclusive access with respect to any critical or
12 ordered regions that access the same storage location x. However, other OpenMP
13 synchronization can ensure the desired exclusive access. For example, a barrier
14 following a series of atomic updates to x guarantees that subsequent accesses do not
15 form a race with the atomic accesses.
16 A compliant implementation may enforce exclusive access between atomic regions
17 that update different storage locations. The circumstances under which this occurs are
18 implementation defined.
19 For an example of the atomic construct, see Section A.19 on page 202.

20 Restrictions
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 required to have the same type and type parameters. See Section A.20 on page 205
29 for examples.
30
Fortran

31 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 The flush construct executes the OpenMP flush operation. This operation makes a
6 thread’s temporary view of memory consistent with memory, and enforces an order on
7 the memory operations of the variables explicitly specified or implied. See the memory
8 model description in Section 1.4 on page 13 for more details.

9 Syntax
10
C/C++
11 The syntax of the flush construct is as follows:
12
13 #pragma omp flush [(list)] new-line

14 Note that because the flush construct does not have a C language statement as part of
15 its syntax, there are some restrictions on its placement within a program. The
16 taskwait directive may be placed only at a point where a base language statement is
17 allowed. The taskwait directive may not be used in place of the statement following
18 an if, while, do, switch, or label. See Appendix C for the formal grammar. See
19 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
Fortran

26 72 OpenMP API • Version 3.0 May 2008

1 Binding
2 The binding thread set for a flush region is the encountering thread. Execution of a
3 flush region affects the memory and the temporary view of memory of only the thread
4 that executes the region. It does not affect the temporary view of other threads. Other
5 threads must themselves execute a flush operation in order to be guaranteed to observe
6 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
9 does not return until the operation is complete for all specified list items. A flush
10 construct without a list, executed on a given thread, operates as if the whole thread-
11 visible data state of the program, as defined by the base language, is flushed.
12
C/C++
13 If a pointer is present in the list, the pointer itself is flushed, not the memory block to
14 which the pointer refers.
C/C++
15
Fortran
16 If the list item or a subobject of the list item has the POINTER attribute, the allocation
17 or association status of the POINTER item is flushed, but the pointer target is not. If the
18 list item is a Cray pointer, the pointer is flushed, but the object to which it points is not.
19 If the list item has the ALLOCATABLE attribute and the list item is allocated, the
20 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
23 on page 211.

24 Chapter 2 Directives 73
1

2 Note – the following examples illustrate the ordering properties of the flush operation.
3 In the following incorrect pseudocode example, the programmer intends to prevent
4 simultaneous execution of the critical section by the two threads, but the program does
5 not work properly because it does not enforce the proper ordering of the operations on
6 variables a and b.

7
8 Incorrect example:
9
10 a = b = 0

11 thread 1 thread 2

12 b = 1 a = 1
13 flush(b) flush(a)
14 flush(a) flush(b)
15 if (a == 0) then if (b == 0) then
16 critical section critical section
17 end if end if
18

19 The problem with this example is that operations on variables a and b are not ordered
20 with respect to each other. For instance, nothing prevents the compiler from moving the
21 flush of b on thread 1 or the flush of a on thread 2 to a position completely after the
22 critical section (assuming that the critical section on thread 1 does not reference b and the
23 critical section on thread 2 does not reference a). If either re-ordering happens, both
24 threads can simultaneously executed the critical section.
25 The following correct pseudocode example correctly ensures that the critical section is
26 executed by not more than one of the two threads at any one time. Notice that execution of
27 the critical section by neither thread is considered correct in this example. This occurs if
28 both flushes complete prior to either thread executing its if statement.
29
30 Correct example:
31
32 a = b = 0

33 thread 1 thread 2

34 b = 1 a = 1
35 flush(a,b) flush(a,b)
36 if (a == 0) then if (b == 0) then
37 critical section critical section
38 end if end if
39

40 74 OpenMP API • Version 3.0 May 2008

1 The compiler is prohibited from moving the flush at all for either thread, ensuring that the
2 respective assignment is complete and the data is flushed before the if statement is
3 executed.
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
13 unset.
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
17 updated in the atomic construct.
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
26 executed in the order of the loop iterations. This sequentializes and orders the code
27 within an ordered region while allowing code outside the region to run in parallel.

28 Chapter 2 Directives 75
1 Syntax
2
C/C++
3 The syntax of the ordered construct is as follows:
4
5 #pragma omp ordered new-line
6 structured-block

7
C/C++
Fortran
8 The syntax of the ordered construct is as follows:
9
10 !$omp ordered
11 structured-block
12 !$omp end ordered

13
14
Fortran

15 Binding
16 The binding thread set for an ordered region is the current team. An ordered region
17 binds to the innermost enclosing loop region. ordered regions that bind to different
18 loop regions execute independently of each other.

19 Description
20 The threads in the team executing the loop region execute ordered regions
21 sequentially in the order of the loop iterations. When the thread executing the first
22 iteration of the loop encounters an ordered construct, it can enter the ordered
23 region without waiting. When a thread executing any subsequent iteration encounters an
24 ordered region, it waits at the beginning of that ordered region until execution of
25 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
30 specified on the corresponding loop (or parallel loop) construct.

31 76 OpenMP API • Version 3.0 May 2008

1 • During execution of an iteration of a loop or a loop nest within a loop region, a
2 thread must not execute more than one ordered region that binds to the same loop
3 region.
4
C/C++
5 • A throw executed inside a ordered region must cause execution to resume within
6 the same ordered region, and the same thread that threw the exception must catch
7 it.
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

13 This section presents a directive and several clauses for controlling the data environment
14 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.
17 • The threadprivate directive, which is provided to create threadprivate memory,
18 is described in Section 2.9.2 on page 81.
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
21 in Section 2.9.3 on page 85.
22 • Clauses that may be specified on directives to copy data values from private or
23 threadprivate variables on one thread to the corresponding variables on other threads
24 in the team are described in Section 2.9.4 on page 100.

25 2.9.1 Data-sharing Attribute Rules

26 This section describes how the data-sharing attributes of variables referenced in
27 parallel, task, and worksharing regions are determined. The following two cases
28 are described separately:
29 • Section 2.9.1.1 on page 78 describes the data-sharing attribute rules for variables
30 referenced in a construct.

31 Chapter 2 Directives 77
1 • Section 2.9.1.2 on page 80 describes the data-sharing attribute rules for variables
2 referenced in a region, but outside any construct.

3 2.9.1.1 Data-sharing Attribute Rules for Variables Referenced in a

4 Construct
5 The data-sharing attributes of variables that are referenced in a construct may be one of
6 the following: predetermined, explicitly determined, or implicitly determined.
7 Specifying a variable on a firstprivate, lastprivate, or reduction clause
8 of an enclosed construct causes an implicit reference to the variable in the enclosing
9 construct. Such implicit references are also subject to the following rules.
10 The following variables have predetermined data-sharing attributes:
11
C/C++
12 • Variables appearing in threadprivate directives are threadprivate.
13 • Variables with automatic storage duration that are declared in a scope inside the
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 • Variables and common blocks appearing in threadprivate directives are
23 threadprivate.
24 • The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do
25 construct is(are) private.
26 • A loop iteration variable for a sequential loop in a parallel or task construct is
27 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
Fortran

32 78 OpenMP API • Version 3.0 May 2008

1 Variables with predetermined data-sharing attributes may not be listed in data-sharing
2 attribute clauses, except for the cases listed below. For these exceptions only, listing a
3 predetermined variable in a data-sharing attribute clause is allowed and overrides the
4 variable’s predetermined data-sharing attributes.
5
C/C++
6 • The loop iteration variable(s) in the associated for-loop(s) of a for or parallel
7 for construct may be listed in a private or lastprivate clause.
C/C++
8
Fortran
9 • The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do
10 construct may be listed in a private or lastprivate clause.
11 • Variables used as loop iteration variables in sequential loops in a parallel
12 construct or a task construct may be listed in private, firstprivate,
13 lastprivate, shared, or reduction clauses on the parallel or task
14 construct, and on enclosed constructs, subject to other restrictions.
15 • Assumed-size arrays may be listed in a shared clause.
16
Fortran
17 Additional restrictions on the variables which may appear in individual clauses are
18 described with each clause in Section 2.9.3 on page 85.
19 Variables with explicitly determined data-sharing attributes are those that are referenced
20 in a given construct and are listed in a data-sharing attribute clause on the construct.
21 Variables with implicitly determined data-sharing attributes are those that are referenced
22 in a given construct, do not have predetermined data-sharing attributes, and are not
23 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
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 shared.
29 • For constructs other than task, if no default clause is present, these variables
30 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 to be shared in all enclosing constructs, up to and including the innermost enclosing
33 parallel construct, is shared.
34 • In a task construct, if no default clause is present, a variable whose data-sharing
35 attribute is not determined by the rule above is firstprivate.

36 Chapter 2 Directives 79
1 Additional restrictions on the variables whose data-sharing attributes cannot be
2 implicitly determined in a task construct are described in the Restrictions section of
3 the firstprivate clause (Section 2.9.3.4 on page 92).

4 2.9.1.2 Data-sharing Attribute Rules for Variables Referenced in a

5 Region but not in a Construct
6 The data-sharing attributes of variables that are referenced in a region, but not in a
7 construct, are determined as follows:
8
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 in called routines, are shared.
12 • File-scope or namespace-scope variables referenced in called routines in the region
13 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 the data-sharing attributes of the associated actual argument.
18 • Other variables declared in called routines in the region are private.
C/C++
19
Fortran
20 • Local variables declared in called routines in the region and that have the save
21 attribute, or that are data initialized, are shared unless they appear in a
22 threadprivate directive.
23 • Variables belonging to common blocks, or declared in modules, and referenced in
24 called routines in the region are shared unless they appear in a threadprivate
25 directive.
26 • Dummy arguments of called routines in the region that are passed by reference inherit
27 the data-sharing attributes of the associated actual argument.
28 • Cray pointees inherit the data-sharing attribute of the storage with which their Cray
29 pointers are associated.
30 • implied-do indices, forall indices, and other local variables declared in called
31 routines in the region are private.
32
Fortran

33 80 OpenMP API • Version 3.0 May 2008

1 2.9.2 threadprivate Directive

2 Summary
3 The threadprivate directive specifies that variables are replicated, with each thread
4 having its own copy.

5 Syntax
6
C/C++
7 The syntax of the threadprivate directive is as follows:
8
9 #pragma omp threadprivate(list) new-line

10 where list is a comma-separated list of file-scope, namespace-scope, or static block-

11 scope variables that do not have incomplete types.
C/C++
12
Fortran
13 The syntax of the threadprivate directive is as follows:
14
15 !$omp threadprivate(list)

16 where list is a comma-separated list of named variables and named common blocks.
17 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
21 program, but at an unspecified point in the program prior to the first reference to that
22 copy. The storage of all copies of a threadprivate variable is freed according to how
23 static variables are handled in the base language, but at an unspecified point in the
24 program.
25 A program in which a thread references another thread’s copy of a threadprivate variable
26 is non-conforming.

27 Chapter 2 Directives 81
1 The content of a threadprivate variable can change across a task scheduling point if the
2 executing thread switches to another schedulable task that modifies the variable. For
3 more details on task scheduling, see Section 1.3 on page 11 and Section 2.7 on page 59.
4 In parallel regions, references by the master thread will be to the copy of the
5 variable in the thread which encountered the parallel region.
6 During the sequential part references will be to the initial thread’s copy of the variable.
7 The values of data in the initial thread’s copy of a threadprivate variable are guaranteed
8 to persist between any two consecutive references to the variable in the program.
9 The values of data in the threadprivate variables of non-initial threads are guaranteed to
10 persist between two consecutive active parallel regions only if all the following
11 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
15 at entry to both parallel regions.
16 If these conditions all hold, and if a threadprivate variable is referenced in both regions,
17 then threads with the same thread number in their respective regions will reference the
18 same copy of that variable.
19
C/C++
20 If the above conditions hold, the storage duration, lifetime, and value of a thread’s copy
21 of a threadprivate variable that does not appear in any copyin clause on the second
22 region will be retained. Otherwise, the storage duration, lifetime, and value of a thread’s
23 copy of the variable in the second region is unspecified.
24 If the value of a variable referenced in an explicit initializer of a threadprivate variable
25 is modified prior to the first reference to any instance of the threadprivate variable, then
26 the behavior is unspecified.
27

28 Note – The order in which any constructors for different threadprivate variables of class
29 type are called is unspecified. The order in which any destructors for different
30 threadprivate variables of class type are called is unspecified.
31

32
C/C++

33
Fortran
34 A variable is affected by a copyin clause if the variable appears in the copyin clause
35 or it is in a common block that appears in the copyin clause.

36 82 OpenMP API • Version 3.0 May 2008

1 If the above conditions hold, the definition, association, or allocation status of a thread’s
2 copy of a threadprivate variable or a variable in a threadprivate common block, that is
3 not affected by any copyin clause that appears on the second region, will be retained.
4 Otherwise, the definition and association status of a thread’s copy of the variable in the
5 second region is undefined, and the allocation status of an allocatable array will be
6 implementation defined.
7 If a common block, or a variable that is declared in the scope of a module, appears in a
8 threadprivate directive, it implicitly has the SAVE attribute.
9 If a threadprivate variable or a variable in a threadprivate common block is not affected
10 by any copyin clause that appears on the first parallel region in which it is
11 referenced, the variable or any subobject of the variable is initially defined or undefined
12 according to the following rules:
13 • If it has the ALLOCATABLE attribute, each copy created will have an initial
14 allocation status of not currently allocated.
15 • If it has the POINTER attribute:
16 • if it has an initial association status of disassociated, either through explicit
17 initialization or default initialization, each copy created will have an association
18 status of disassociated;
19 • otherwise, each copy created will have an association status of undefined.
20 • If it does not have either the POINTER or the ALLOCATABLE attribute:
21 • if it is initially defined, either through explicit initialization or default
22 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 • A program in which an untied task accesses threadprivate storage is non-conforming.
31
C/C++
32 • A variable which is part of another variable (as an array or structure element) cannot
33 appear in a threadprivate clause unless it is a static data member of a C++
34 class.

35 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
3 variables in its list.
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
6 must lexically precede all references to any of the variables in its list.
7 • A threadprivate directive for namespace-scope variables must appear outside
8 any definition or declaration other than the namespace definition itself, and must
9 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
11 scope must refer to a variable declaration at file, namespace, or class scope that
12 lexically precedes the directive.
13 • A threadprivate directive for static block-scope variables must appear in the
14 scope of the variable and not in a nested scope. The directive must lexically precede
15 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 a variable declaration in the same scope that lexically precedes the directive. The
18 variable declaration must use the static storage-class specifier.
19 • If a variable is specified in a threadprivate directive in one translation unit, it
20 must be specified in a threadprivate directive in every translation unit in which
21 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 • an accessible, unambiguous default constructor in case of default initialization
26 without a given initializer;
27 • an accessible, unambiguous constructor accepting the given argument in case of
28 direct initialization;
29 • an accessible, unambiguous copy constructor in case of copy initialization with an
30 explicit initializer.
C/C++
31
Fortran
32 • A variable which is part of another variable (as an array or structure element) cannot
33 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
36 common blocks can be accessed by use association or host association, common

37 84 OpenMP API • Version 3.0 May 2008

1 block names cannot. This means that a common block name specified in a
2 threadprivate directive must be declared to be a common block in the same
3 scoping unit in which the threadprivate directive appears.
4 • If a threadprivate directive specifying a common block name appears in one
5 program unit, then such a directive must also appear in every other program unit that
6 contains a COMMON statement specifying the same name. It must appear after the last
7 such COMMON statement in the program unit.
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
10 is declared. It must not be an element of a common block or appear in an
11 EQUIVALENCE statement.
12 • A variable that appears in a threadprivate directive and is not declared in the
13 scope of a module must have the SAVE attribute.
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 2.9.3 Data-Sharing Attribute Clauses

20 Several constructs accept clauses that allow a user to control the data-sharing attributes
21 of variables referenced in the construct. Data-sharing attribute clauses apply only to
22 variables whose names are visible in the construct on which the clause appears.
23 Not all of the clauses listed in this section are valid on all directives. The set of clauses
24 that is valid on a particular directive is described with the directive.
25 Most of the clauses accept a comma-separated list of list items (see Section 2.1 on page
26 22). All list items appearing in a clause must be visible, according to the scoping rules
27 of the base language. With the exception of the default clause, clauses may be
28 repeated as needed. A list item that specifies a given variable may not appear in more
29 than one clause on the same directive, except that a variable may be specified in both
30 firstprivate and lastprivate clauses.

31 Chapter 2 Directives 85
1
C/C++
2 If a variable referenced in a data-sharing attribute clause has a type derived from a
3 template, and there are no other references to that variable in the program, then any
4 behavior related to that variable is unspecified.
C/C++
5
Fortran
6 A named common block may be specified in a list by enclosing the name in slashes.
7 When a named common block appears in a list, it has the same meaning as if every
8 explicit member of the common block appeared in the list. An explicit member of a
9 common block is a variable that is named in a COMMON statement that specifies the
10 common block name and is declared in the same scoping unit in which the clause
11 appears.
12 Although variables in common blocks can be accessed by use association or host
13 association, common block names cannot. This means that a common block name
14 specified in a data-sharing attribute clause must be declared to be a common block in the
15 same scoping unit in which the data-sharing attribute clause appears.
16 When a named common block appears in a private, firstprivate,
17 lastprivate, or shared clause of a directive, none of its members may be declared
18 in another data-sharing attribute clause in that directive (see Section A.27 on page 227
19 for examples). When individual members of a common block appear in a private,
20 firstprivate, lastprivate, or reduction clause of a directive, the storage of
21 the specified variables is no longer associated with the storage of the common block
22 itself (see Section A.32 on page 237 for examples).
23
Fortran

24 2.9.3.1 default clause

25 Summary
26 The default clause allows the user to control the data-sharing attributes of variables that
27 are referenced in a parallel or task construct, and whose data-sharing attributes are
28 implicitly determined (see Section 2.9.1.1 on page 78).

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
15 have implicitly determined data-sharing attributes to be shared.
16
Fortran
17 The default(firstprivate) clause causes all variables in the construct that have
18 implicitly determined data-sharing attributes to be firstprivate.
19 The default(private) clause causes all variables referenced in the construct that
20 have implicitly determined data-sharing attributes to be private.
21
Fortran
22 The default(none) clause requires that each variable that is referenced in the
23 construct, and that does not have a predetermined data-sharing attribute, must have its
24 data-sharing attribute explicitly determined by being listed in a data-sharing attribute
25 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.

29 Chapter 2 Directives 87
1 2.9.3.2 shared clause

2 Summary
3 The shared clause declares one or more list items to be shared by tasks generated by
4 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
11 at the point the directive was encountered.
12 It is the programmer's responsibility to ensure, by adding proper synchronization, that
13 storage shared by an explicit task region does not reach the end of its lifetime before
14 the explicit task region completes its execution.
15
Fortran
16 The association status of a shared pointer becomes undefined upon entry to and on exit
17 from the parallel or task construct if it is associated with a target or a subobject of
18 a target that is in a private, firstprivate, lastprivate, or reduction
19 clause inside the construct.
20 Under certain conditions, passing a shared variable to a non-intrinsic procedure may
21 result in the value of the shared variable being copied into temporary storage before the
22 procedure reference, and back out of the temporary storage into the actual argument
23 storage after the procedure reference. It is implementation defined when this situation
24 occurs. See Section A.29 on page 231 for an example of this behavior.
25

26 Note – This situation may occur when the following three conditions hold regarding an
27 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
9 or an assumed-size array.
10 This effectively results in references to, and definitions of, the temporary storage during
11 the procedure reference. Any references to (or definitions of) the shared storage that is
12 associated with the dummy argument by any other task must be synchronized with the
13 procedure reference to avoid possible race conditions.
14

15
16
Fortran

17 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
26 in the construct receives a new list item whose language-specific attributes are derived
27 from the original list item. Inside the construct, all references to the original list item are
28 replaced by references to the new list item. In the rest of the region, it is unspecified
29 whether references are to the new list item or the original list item. Therefore, if an

30 Chapter 2 Directives 89
1 attempt is made to reference the original item, its value after the region is also
2 unspecified. If a task does not reference a list item that appears in a private clause, it
3 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
9 parallel construct may also appear in a private clause in an enclosed parallel,
10 task, or workshare construct. List items that appear in a private or
11 firstprivate clause in a task construct may also appear in a private clause in
12 an enclosed parallel or task construct. List items that appear in a private,
13 firstprivate, lastprivate, or reduction clause in a workshare construct
14 may also appear in a private clause in an enclosed parallel or task construct.
15 See Section A.31 on page 235 for an example.
16
C/C++
17 A new list item of the same type, with automatic storage duration, is allocated for the
18 construct. The storage and thus lifetime of these list items lasts until the block in which
19 they are created exits. The size and alignment of the new list item are determined by the
20 type of the variable. This allocation occurs once for each task generated by the
21 construct, if the task references the list item in any statement.
22 The new list item is initialized, or has an undefined initial value, as if it had been locally
23 declared without an initializer. The order in which any default constructors for different
24 private variables of class type are called is unspecified. The order in which any
25 destructors for different private variables of class type are called is unspecified.
C/C++
26
Fortran
27 A new list item of the same type is allocated once for each implicit task in the
28 parallel region, or for each task generated by a task construct, if the construct
29 references the list item in any statement. The initial value of the new list item is
30 undefined. Within a parallel, worksharing, or task region, the initial status of a
31 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
34 of "not currently allocated";
35 • if the list item is allocated, the new list item will have an initial state of allocated
36 with the same array bounds.

37 90 OpenMP API • Version 3.0 May 2008

1 A list item that appears in a private clause may be storage-associated with other
2 variables when the private clause is encountered. Storage association may exist
3 because of constructs such as EQUIVALENCE or COMMON. If A is a variable appearing
4 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
6 parallel or task region.
7 • Any definition of A, or of its allocation or association status, causes the contents,
8 allocation, and association status of B to become undefined.
9 • Any definition of B, or of its allocation or association status, causes the contents,
10 allocation, and association status of A to become undefined.
11 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 appear in a private clause.
18
C/C++
19 • A variable of class type (or array thereof) that appears in a private clause requires
20 an accessible, unambiguous default constructor for the class type.
21 • A variable that appears in a private clause must not have a const-qualified type
22 unless it is of class type with a mutable member.
23 • A variable that appears in a private clause must not have an incomplete type or a
24 reference type.
C/C++
25
Fortran
26 • A variable that appears in a private clause must either be definable, or an
27 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 expressions for statement function definitions, may not appear in a private clause.
31
Fortran

32 Chapter 2 Directives 91
1 2.9.3.4 firstprivate clause

2 Summary
3 The firstprivate clause declares one or more list items to be private to a task, and
4 initializes each of them with the value that the corresponding original item has when the
5 construct is encountered.

6 Syntax
7 The syntax of the firstprivate clause is as follows:
8
9 firstprivate(list)

10 Description
11 The firstprivate clause provides a superset of the functionality provided by the
12 private clause.
13 A list item that appears in a firstprivate clause is subject to the private clause
14 semantics described in Section 2.9.3.3 on page 89. In addition, the new list item is
15 initialized from the original list item existing before the construct. The initialization of
16 the new list item is done once for each task that references the list item in any statement
17 in the construct. The initialization is done prior to the execution of the construct.
18 For a firstprivate clause on a parallel or task construct, the initial value of
19 the new list item is the value of the original list item that exists immediately prior to the
20 construct in the task region where the construct is encountered. For a firstprivate
21 clause on a worksharing construct, the initial value of the new list item for each implicit
22 task of the threads that execute the worksharing construct is the value of the original list
23 item that exists in the implicit task immediately prior to the point in time that the
24 worksharing construct is encountered.
25 To avoid race conditions, concurrent updates of the original list item must be
26 synchronized with the read of the original list item that occurs as a result of the
27 firstprivate clause.
28 If a list item appears in both firstprivate and lastprivate clauses, the update
29 required for lastprivate occurs after all the initializations for firstprivate.

30 92 OpenMP API • Version 3.0 May 2008

1
C/C++
2 For variables of non-array type, the initialization occurs by copy assignment. For a
3 (possibly multi-dimensional) array of elements of non-array type, each element is
4 initialized as if by assignment from an element of the original array to the corresponding
5 element of the new array. For variables of class type, a copy constructor is invoked to
6 perform the initialization. The order in which copy constructors for different variables of
7 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 appear in a firstprivate clause.
15 • A list item that is private within a parallel region must not appear in a
16 firstprivate clause on a worksharing construct if any of the worksharing
17 regions arising from the worksharing construct ever bind to any of the parallel
18 regions arising from the parallel construct.
19 • A list item that appears in a reduction clause of a parallel construct must not
20 appear in a firstprivate clause on a worksharing or task construct if any of
21 the worksharing or task regions arising from the worksharing or task construct
22 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
25 of any of the worksharing regions arising from the worksharing construct.
26
C/C++
27 • A variable of class type (or array thereof) that appears in a firstprivate clause
28 requires an accessible, unambiguous copy constructor for the class type.
29 • A variable that appears in a firstprivate clause must not have a const-
30 qualified type unless it is of class type with a mutable member.
31 • A variable that appears in a firstprivate clause must not have an incomplete
32 type or a reference type.
C/C++
33
Fortran
34 • A variable that appears in a firstprivate clause must be definable.

35 Chapter 2 Directives 93
1 • Fortran pointers, Cray pointers, and assumed-size arrays may not appear in a
2 firstprivate clause.
3 • Variables that appear in namelist statements, in variable format expressions, and in
4 expressions for statement function definitions, may not appear in a firstprivate
5 clause.
6
Fortran

7 2.9.3.5 lastprivate clause

8 Summary
9 The lastprivate clause declares one or more list items to be private to an implicit
10 task, and causes the corresponding original list item to be updated after the end of the
11 region.

12 Syntax
13 The syntax of the lastprivate clause is as follows:
14
15 lastprivate(list)

16 Description
17 The lastprivate clause provides a superset of the functionality provided by the
18 private clause.
19 A list item that appears in a lastprivate clause is subject to the private clause
20 semantics described in Section 2.9.3.3 on page 89. In addition, when a lastprivate
21 clause appears on the directive that identifies a worksharing construct, the value of each
22 new list item from the sequentially last iteration of the associated loops, or the lexically
23 last section construct, is assigned to the original list item.
24
C/C++
25 For a (possibly multi-dimensional) array of elements of non-array type, each element is
26 assigned to the corresponding element of the original array.
C/C++
27 List items that are not assigned a value by the sequentially last iteration of the loops, or
28 by the lexically last section construct, have unspecified values after the construct.
29 Unassigned subcomponents also have unspecified values after the construct.

30 94 OpenMP API • Version 3.0 May 2008

1 The original list item becomes defined at the end of the construct if there is an implicit
2 barrier at that point. To avoid race conditions, concurrent reads or updates of the original
3 list item must be synchronized with the update of the original list item that occurs as a
4 result of the lastprivate clause.
5 If the lastprivate clause is used on a construct to which nowait is applied,
6 accesses to the original list item may create a data race. To avoid this, synchronization
7 must be inserted to ensure that the sequentially last iteration or lexically last section
8 construct has stored and flushed that list item.
9 If a list item appears in both firstprivate and lastprivate clauses, the update
10 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
15 appear in a lastprivate clause.
16 • A list item that is private within a parallel region, or that appears in the
17 reduction clause of a parallel construct, must not appear in a lastprivate
18 clause on a worksharing construct if any of the corresponding worksharing regions
19 ever binds to any of the corresponding parallel regions.
20
C/C++
21 • A variable of class type (or array thereof) that appears in a lastprivate clause
22 requires an accessible, unambiguous default constructor for the class type, unless the
23 list item is also specified in a firstprivate clause.
24 • A variable of class type (or array thereof) that appears in a lastprivate clause
25 requires an accessible, unambiguous copy assignment operator for the class type.The
26 order in which copy assignment operators for different variables of class type are
27 called is unspecified.
28 • A variable that appears in a lastprivate clause must not have a const-qualified
29 type unless it is of class type with a mutable member.
30 • A variable that appears in a lastprivate clause must not have an incomplete type
31 or a reference type.
C/C++
32
Fortran
33 • A variable that appears in a lastprivate clause must be definable.

34 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
3 sequentially last iteration or lexically last section must be in the allocated state upon
4 exit from that iteration or section with the same bounds as the corresponding original
5 list item.
6 • Fortran pointers, Cray pointers, and assumed-size arrays may not appear in a
7 lastprivate clause.
8 • Variables that appear in namelist statements, in variable format expressions, and in
9 expressions for statement function definitions, may not appear in a lastprivate
10 clause.
11
Fortran

12 2.9.3.6 reduction clause

13 Summary
14 The reduction clause specifies an operator and one or more list items. For each list
15 item, a private copy is created in each implicit task, and is initialized appropriately for
16 the operator. After the end of the region, the original list item is updated with the values
17 of the private copies using the specified operator.

18 Syntax
19
C/C++
20 The syntax of the reduction clause is as follows:
21
22 reduction(operator:list)

23 The following table lists the operators that are valid and their initialization values. The
24 actual initialization value depends on the data type of the reduction list item.

25
26 Operator Initialization value

27 + 0

28 * 1

29 - 0

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
11 reduction({operator | intrinsic_procedure_name}:list)

12 The following table lists the operators and intrinsic_procedure_names that are valid and
13 their initialization values. The actual initialization value depends on the data type of the
14 reduction list item.
15
16 Operator/
17 Intrinsic Initialization value

18 + 0

19 * 1

20 - 0

21 .and. .true.

22 .or. .false.

23 .eqv. .true.

24 .neqv. .false.
25 max Most negative representable number in the
26 reduction list item type
27 min Largest representable number in the reduction
28 list item type
29 iand All bits on
30 ior 0
31 ieor 0

32
33
Fortran

34 Chapter 2 Directives 97
1 Description
2 The reduction clause can be used to perform some forms of recurrence calculations
3 (involving mathematically associative and commutative operators) in parallel.
4 A private copy of each list item is created, one for each implicit task, as if the private
5 clause had been used. The private copy is then initialized to the initialization value for
6 the operator, as specified above. At the end of the region for which the reduction
7 clause was specified, the original list item is updated by combining its original value
8 with the final value of each of the private copies, using the operator specified. (The
9 partial results of a subtraction reduction are added to form the final value.)
10 If nowait is not used, the reduction computation will be complete at the end of the
11 construct; however, if the reduction clause is used on a construct to which nowait is
12 also applied, accesses to the original list item will create a race and, thus, have
13 unspecified effect unless synchronization ensures that they occur after all threads have
14 executed all of their iterations or section constructs, and the reduction computation
15 has completed and stored the computed value of that list item. This can most simply be
16 ensured through a barrier synchronization.
17 The order in which the values are combined is unspecified. Therefore, comparing
18 sequential and parallel runs, or comparing one parallel run to another (even if the
19 number of threads used is the same), there is no guarantee that bit-identical results will
20 be obtained or that side effects (such as floating point exceptions) will be identical.
21 To avoid race conditions, concurrent reads or updates of the original list item must be
22 synchronized with the update of the original list item that occurs as a result of the
23 reduction computation.
24

25 Note – List items specified in a reduction clause are typically used in the enclosed
26 region in certain forms.
27
C/C++
28 A reduction is typically specified for statements of the form:
29
30 x = x op expr
31 x binop= expr
32 x = expr op x (except for subtraction)
33 x++
34 ++x
35 x--
36 --x

37 98 OpenMP API • Version 3.0 May 2008

1 where expr has scalar type and does not reference x, op is not an overloaded operator,
2 but one of +, *, -, &, ^, |, &&, or ||, and binop is not an overloaded operator, but one
3 of +, *, -, &, ^, or |.
C/C++
4
Fortran
5 A reduction using an operator is typically specified for statements of the form:
6
7 x = x op expr
8 x = expr op x (except for subtraction)

9 where op is +, *, -, .and., .or., .eqv., or .neqv., the expression does not involve
10 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:
12
13 x = intr(x,expr_list)
14 x = intr(expr_list, x)

15 where intr is max, min, iand, ior, or ieor and expr_list is a comma separated list of
16 expressions not involving x.
17
Fortran
18 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
24 from the worksharing construct bind.
25 • A list item that appears in a reduction clause of the innermost enclosing
26 worksharing or parallel construct may not be accessed in an explicit task.
27 • Any number of reduction clauses can be specified on the directive, but a list item
28 can appear only once in the reduction clause(s) for that directive.
29
C/C++
30 • The type of a list item that appears in a reduction clause must be valid for the
31 reduction operator.
32 • Aggregate types (including arrays), pointer types and reference types may not appear
33 in a reduction clause.

34 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
3 the list items that appear in that clause.
C/C++
4
Fortran
5 • The type of a list item that appears in a reduction clause must be valid for the
6 reduction operator or intrinsic.
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
9 intrinsic type.
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
12 must not be deallocated and/or allocated within the region.
13 • Fortran pointers, Cray pointers and assumed-size arrays may not appear in a
14 reduction clause.
15 • Operators specified must be intrinsic operators and any intrinsic_procedure_name
16 must refer to one of the allowed intrinsic procedures. Assignment to the reduction list
17 items must be via intrinsic assignment. See Section A.35 on page 242 for examples.
18
Fortran

19 2.9.4 Data Copying Clauses

20 This section describes the copyin clause (valid on the parallel directive and
21 combined parallel worksharing directives) and the copyprivate clause (valid on the
22 single directive).
23 These clauses support the copying of data values from private or threadprivate variables
24 on one implicit task or thread to the corresponding variables on other implicit tasks or
25 threads in the team.
26 The clauses accept a comma-separated list of list items (see Section 2.1 on page 22). All
27 list items appearing in a clause must be visible, according to the scoping rules of the
28 base language. Clauses may be repeated as needed, but a list item that specifies a given
29 variable may not appear in more than one clause on the same directive.

30 100 OpenMP API • Version 3.0 May 2008

1 2.9.4.1 copyin clause

2 Summary
3 The copyin clause provides a mechanism to copy the value of the master thread’s
4 threadprivate variable to the threadprivate variable of each other member of the team
5 executing the parallel region.

6 Syntax
7 The syntax of the copyin clause is as follows:
8
9 copyin(list)

10 Description
11
C/C++
12 The copy is done after the team is formed and prior to the start of execution of the
13 associated structured block. For variables of non-array type, the copy occurs by copy
14 assignment. For a (possibly multi-dimensional) array of elements of non-array type,
15 each element is copied as if by assignment from an element of the master thread’s array
16 to the corresponding element of the other thread’s array. For class types, the copy
17 assignment operator is invoked. The order in which copy assignment operators for
18 different variables of class type are called is unspecified.
C/C++
19
Fortran
20 The copy is done, as if by assignment, after the team is formed and prior to the start of
21 execution of the associated structured block.
22 On entry to any parallel region, each thread’s copy of a variable that is affected by
23 a copyin clause for the parallel region will acquire the allocation, association, and
24 definition status of the master thread’s copy, according to the following rules:
25 • If it has the POINTER attribute:
26 • if the master thread’s copy is associated with a target that each copy can become
27 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.

30 Chapter 2 Directives 101

1 • If it does not have the POINTER attribute, each copy becomes defined with the value
2 of the master thread’s copy as if by intrinsic assignment.
3
Fortran
4 For an example of the copyin clause, see Section A.36 on page 248.

5 Restrictions
6 The restrictions to the copyin clause are as follows:
7
C/C++
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
10 an accessible, unambiguous copy assignment operator for the class type.
C/C++
11
Fortran
12 • A list item that appears in a copyin clause must be threadprivate. Named variables
13 appearing in a threadprivate common block may be specified: it is not necessary to
14 specify the whole common block.
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.
17 • An array with the ALLOCATABLE attribute must be in the allocated state. Each
18 thread's copy of that array must be allocated with the same bounds.
19
Fortran

20 2.9.4.2 copyprivate clause

21 Summary
22 The copyprivate clause provides a mechanism to use a private variable to broadcast
23 a value from the data environment of one implicit task to the data environments of the
24 other implicit tasks belonging to the parallel region.
25 To avoid race conditions, concurrent reads or updates of the list item must be
26 synchronized with the update of the list item that occurs as a result of the
27 copyprivate clause.

28 102 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
7 execution of the structured block associated with the single construct (see
8 Section 2.5.3 on page 49), and before any of the threads in the team have left the barrier
9 at the end of the construct.
10
C/C++
11 In all other implicit tasks belonging to the parallel region, each specified list item
12 becomes defined with the value of the corresponding list item in the implicit task whose
13 thread executed the structured block. For variables of non-array type, the definition
14 occurs by copy assignment. For a (possibly multi-dimensional) array of elements of non-
15 array type, each element is copied by copy assignment from an element of the array in
16 the data environment of the implicit task whose thread executed the structured block to
17 the corresponding element of the array in the data environment of the other implicit
18 tasks. For class types, a copy assignment operator is invoked. The order in which copy
19 assignment operators for different variables of class type are called is unspecified.
C/C++
20
Fortran
21 If a list item is not a pointer, then in all other implicit tasks belonging to the parallel
22 region, the list item becomes defined (as if by assignment) with the value of the
23 corresponding list item in the implicit task whose thread executed the structured block.
24 If the list item is a pointer, then in all other implicit tasks belonging to the parallel
25 region, the list item becomes pointer associated (as if by pointer assignment) with the
26 corresponding list item in the implicit task whose thread executed the structured block.
27
Fortran
28 For examples of the copyprivate clause, see Section A.37 on page 250.
29

30 Note – The copyprivate clause is an alternative to using a shared variable for the
31 value when providing such a shared variable would be difficult (for example, in a
32 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
4 or private in the enclosing context.
5 • A list item that appears in a copyprivate clause may not appear in a private or
6 firstprivate clause on the single construct.
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
14 same bounds in all threads affected by the copyprivate clause.
15
Fortran

16

17 2.10 Nesting of Regions

18 This section describes a set of restrictions on the nesting of regions. The restrictions on
19 nesting are as follows:
20 • A worksharing region may not be closely nested inside a worksharing, explicit task,
21 critical, ordered, or master region.
22 • A barrier region may not be closely nested inside a worksharing, explicit task,
23 critical, ordered, or master region.
24 • A master region may not be closely nested inside a worksharing or explicit task
25 region.
26 • An ordered region may not be closely nested inside a critical or explicit task
27 region.
28 • An ordered region must be closely nested inside a loop region (or parallel loop
29 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
32 deadlock.

33 104 OpenMP API • Version 3.0 May 2008

1 For examples illustrating these rules, see Section A.17 on page 199, Section A.38 on
2 page 255, Section A.39 on page 258, and Section A.13 on page 177.

3 Chapter 2 Directives 105

1 106 OpenMP API • Version 3.0 May 2008
1
2 CHAPTER 3

3 Runtime Library Routines

4

5 This chapter describes the OpenMP API runtime library routines and is divided into the
6 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
9 execution environment (Section 3.2 on page 109).
10 • Lock routines that can be used to synchronize access to data (Section 3.3 on page
11 134).
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
14 description of the routines.
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 procedures.
25
Fortran

26 107
1

2 3.1 Runtime Library Definitions

3 For each base language, a compliant implementation must supply a set of definitions for
4 the OpenMP API runtime library routines and the special data types of their parameters.
5 The set of definitions must contain a declaration for each OpenMP API runtime library
6 routine and a declaration for the simple lock, nestable lock and schedule data types. In
7 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 Prototypes for the C/C++ runtime library routines described in this chapter shall be
11 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 The OpenMP Fortran API runtime library routines are external procedures. The return
19 values of these routines are of default kind, unless otherwise specified.

20 Interface declarations for the OpenMP Fortran runtime library routines described in this
21 chapter shall be provided in the form of a Fortran include file named omp_lib.h or
22 a Fortran 90 module named omp_lib. It is implementation defined whether the
23 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
31 API that the implementation supports. This value matches that of the C preprocessor
32 macro _OPENMP, when a macro preprocessor is supported (see Section 2.2 on page
33 26).

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
3 take an argument are extended with a generic interface so arguments of different KIND
4 type can be accommodated. See Appendix D.4 for an example of such an extension.
5
Fortran

7 3.2 Execution Environment Routines

8 The routines described in this section affect and monitor threads, processors, and the
9 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.

29 Chapter 3 Runtime Library Routines 109

1 3.2.1 omp_set_num_threads

2 Summary
3 The omp_set_num_threads routine affects the number of threads to be used for
4 subsequent parallel regions that do not specify a num_threads clause, by setting
5 the value of the nthreads-var ICV.

6 Format
C/C++
7
8 void omp_set_num_threads(int num_threads);
9
C/C++
Fortran
10
11 subroutine omp_set_num_threads(num_threads)
12 integer num_threads
13
14
Fortran

15 Constraints on Arguments
16 The value of the argument passed to this routine must evaluate to a positive integer, or
17 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
22 specified in the argument.

23 See Section 2.4.1 on page 35 for the rules governing the number of threads used to
24 execute a parallel region.

25 110 OpenMP API • Version 3.0 May 2008

1 For an example of the omp_set_num_threads routine, see Section A.40 on page
2 265.

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.

9 3.2.2 omp_get_num_threads

10 Summary
11 The omp_get_num_threads routine returns the number of threads in the current
12 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 The binding region for an omp_get_num_threads region is the innermost enclosing
23 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
3 executing the parallel region to which the routine region binds. If called from the
4 sequential part of a program, this routine returns 1. For examples, see Section A.41 on
5 page 266.

6 See Section 2.4.1 on page 35 for the rules governing the number of threads used to
7 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.

12 3.2.3 omp_get_max_threads

13 Summary
14 The omp_get_max_threads routine returns an upper bound on the number of
15 threads that could be used to form a new team if a parallel region without a
16 num_threads clause were encountered after execution returns from this routine.

17 Format
C/C++
18
19 int omp_get_max_threads(void);
20
C/C++
Fortran
21
22 integer function omp_get_max_threads()
23
24
Fortran

25 112 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.
5 This value is also an upper bound on the number of threads that could be used to form a
6 new team if a parallel region without a num_threads clause were encountered after
7 execution returns from this routine.

8 See Section 2.4.1 on page 35 for the rules governing the number of threads used to
9 execute a parallel region.
10

11 Note – The return value of the omp_get_max_threads routine can be used to

12 dynamically allocate sufficient storage for all threads in the team formed at the
13 subsequent active parallel region.
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
24 team, of the thread executing the implicit or explicit task region from which
25 omp_get_thread_num is called.

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 The binding thread set for an omp_get_thread_num region is the current team. The
11 binding region for an omp_get_thread_num region is the innermost enclosing
12 parallel region.

13 Effect
14 The omp_get_thread_num routine returns the thread number of the current thread,
15 within the team executing the parallel region to which the routine region binds. The
16 thread number is an integer between 0 and one less than the value returned by
17 omp_get_num_threads, inclusive. The thread number of the master thread of the
18 team is 0. The routine returns 0 if it is called from the sequential part of a program.
19

20 Note – The thread number may change at any time during the execution of an untied
21 task. The value returned by omp_get_thread_num is not generally useful during the
22 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 The omp_get_num_procs routine returns the number of processors available to the
4 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 The binding thread set for an omp_get_num_procs region is all threads. The effect
15 of executing this routine is not related to any specific region corresponding to any
16 construct or API routine.

17 Effect
18 The omp_get_num_procs routine returns the number of processors that are available
19 to the program at the time the routine is called. Note that this value may change between
20 the time that it is determined by the omp_get_num_procs routine and the time that it
21 is read in the calling context due to system actions outside the control of the OpenMP
22 implementation.

23 Chapter 3 Runtime Library Routines 115

1 3.2.6 omp_in_parallel

2 Summary
3 The omp_in_parallel routine returns true if the call to the routine is enclosed by an
4 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 The binding thread set for an omp_in_parallel region is all threads. The effect of
15 executing this routine is not related to any specific parallel region but instead
16 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
19 routine call is enclosed by only inactive parallel regions (including the implicit
20 parallel region), then it returns false.

21 116 OpenMP API • Version 3.0 May 2008

1 3.2.7 omp_set_dynamic

2 Summary
3 The omp_set_dynamic routine enables or disables dynamic adjustment of the
4 number of threads available for the execution of subsequent parallel regions by
5 setting the value of the dyn-var ICV.

6 Format
C/C++
7
8 void omp_set_dynamic(int dynamic_threads);
9
C/C++
Fortran
10
11 subroutine omp_set_dynamic (dynamic_threads)
12 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 For implementations that support dynamic adjustment of the number of threads, if the
19 argument to omp_set_dynamic evaluates to true, dynamic adjustment is enabled;
20 otherwise, dynamic adjustment is disabled. For implementations that do not support
21 dynamic adjustment of the number of threads this routine has no effect: the value of
22 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
25 execute a parallel region.

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 The omp_get_dynamic routine returns the value of the dyn-var ICV, which
9 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
22 returns false, otherwise. If an implementation does not support dynamic adjustment of
23 the number of threads, then this routine always returns false.

24 118 OpenMP API • Version 3.0 May 2008

1 See Section 2.4.1 on page 35 for the rules governing the number of threads used to
2 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.

7 3.2.9 omp_set_nested

8 Summary
9 The omp_set_nested routine enables or disables nested parallelism, by setting the
10 nest-var ICV.

11 Format
C/C++
12
13 void omp_set_nested(int nested);
14
C/C++
Fortran
15
16 subroutine omp_set_nested (nested)
17 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
3 omp_set_nested evaluates to true, nested parallelism is enabled; otherwise, nested
4 parallelism is disabled. For implementations that do not support nested parallelism, this
5 routine has no effect: the value of nest-var remains false.

6 See Section 2.4.1 on page 35 for the rules governing the number of threads used to
7 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.

14 3.2.10 omp_get_nested

15 Summary
16 The omp_get_nested routine returns the value of the nest-var ICV, which
17 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
Fortran

26 120 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
5 implementation does not support nested parallelism, this routine always returns false.

6 See Section 2.4.1 on page 35 for the rules governing the number of threads used to
7 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.

12 3.2.11 omp_set_schedule

13 Summary
14 The omp_set_schedule routine affects the schedule that is applied when runtime
15 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 subroutine omp_set_schedule(kind, modifier)
4 integer (kind=omp_sched_kind) kind
5 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
10 (except for runtime) or any implementation specific schedule. The C/C++ header file
11 (omp.h) and the Fortran include file (omp_lib.h) and/or Fortran 90 module file
12 (omp_lib) define the valid constants. The valid constants must include the following,
13 which can be extended with implementation specific values:

14
C/C++
15
16 typedef enum omp_sched_t {
17 omp_sched_static = 1,
18 omp_sched_dynamic = 2,
19 omp_sched_guided = 3,
20 omp_sched_auto = 4
21 } omp_sched_t;
22
23
C/C++
24
Fortran
25
26 integer(kind=omp_sched_kind), parameter :: omp_sched_static = 1
27 integer(kind=omp_sched_kind), parameter :: omp_sched_dynamic = 2
28 integer(kind=omp_sched_kind), parameter :: omp_sched_guided = 3
29 integer(kind=omp_sched_kind), parameter :: omp_sched_auto = 4
30
31
32
Fortran

33 122 OpenMP API • Version 3.0 May 2008

1 Binding
2 The binding task set for an omp_set_schedule region is the generating task.

3 Effect
4 The effect of this routine is to set the value of the run-sched-var ICV to the values
5 specified in the two arguments. The schedule is set to the schedule type specified by the
6 first argument kind. It can be any of the standard schedule types or any other
7 implementation specific one. For the schedule types static, dynamic, and guided
8 the chunk_size is set to the value of the second argument, or to the default chunk_size if
9 the value of the second argument is less than 1; for the schedule type auto the second
10 argument has no meaning; for implementation specific schedule types, the values and
11 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.

17 3.2.12 omp_get_schedule

18 Summary
19 The omp_get_schedule routine returns the schedule that is applied when the
20 runtime schedule is used.

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 subroutine omp_get_schedule(kind, modifier)
9 integer (kind=omp_sched_kind) kind
10 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 This routine returns the run-sched-var ICV in the team executing the parallel region
17 to which the routine binds. The first argument kind returns the schedule to be used. It
18 can be any of the standard schedule types as defined in Section 3.2.11 on page 121, or
19 any implementation specific schedule type. The second argument is interpreted as in the
20 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.

26 124 OpenMP API • Version 3.0 May 2008

1 3.2.13 omp_get_thread_limit

2 Summary
3 The omp_get_thread_limit routine returns the maximum number of OpenMP
4 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 The binding thread set for an omp_get_thread_limit region is all threads. The
17 effect of executing this routine is not related to any specific region corresponding to any
18 construct or API routine.

19 Effect
20 The omp_get_thread_limit routine returns the maximum number of OpenMP
21 threads available to the program as stored in the ICV thread-limit-var.

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 The omp_set_max_active_levels routine limits the number of nested active
7 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 subroutine omp_set_max_active_levels (max_levels)
16 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,
21 or else the behavior of this routine is implementation defined.

22 126 OpenMP API • Version 3.0 May 2008

1 Binding
2 When called from the sequential part of the program, the binding thread set for an
3 omp_set_max_active_levels region is the encountering thread. When called
4 from within any explicit parallel region, the binding thread set (and binding region, if
5 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
8 specified in the argument.
9 If the number of parallel levels requested exceeds the number of levels of parallelism
10 supported by the implementation, the value of the max-active-levels-var ICV will be set
11 to the number of parallel levels support by the implementation.
12 This routine has the described effect only when called from the sequential part of the
13 program. When called from within an explicit parallel region, the effect of this
14 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.

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-active-
22 levels-var ICV, which determines the maximum number of nested active parallel regions.

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 When called from the sequential part of the program, the binding thread set for an
13 omp_get_max_active_levels region is the encountering thread. When called
14 from within any explicit parallel region, the binding thread set (and binding region, if
15 required) for the omp_get_max_active_levels region is implementation defined.

16 Effect
17 The omp_get_max_active_levels routine returns the value of the max-active-
18 levels-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.

23 128 OpenMP API • Version 3.0 May 2008

1 3.2.16 omp_get_level

2 Summary
3 The omp_get_level routine returns the number of nested parallel regions
4 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 The binding task set for an omp_get_level region is the generating task. The
17 binding region for an omp_get_level region is the innermost enclosing parallel
18 region.

19 Effect
20 The omp_get_level routine returns the number of nested parallel regions
21 (whether active or inactive) enclosing the task that contains the call, not including the
22 implicit parallel region. The routine always returns a non-negative integer, and returns 0
23 if it is called from the sequential part of the program.

24 Chapter 3 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 The omp_get_ancestor_thread_num routine returns, for a given nested level of
7 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 integer function omp_get_ancestor_thread_num(level)
16 integer level

17
18
Fortran

19 Binding
20 The binding thread set for an omp_get_ancestor_thread_num region is the
21 encountering thread. The binding region for an omp_get_ancestor_thread_num
22 region is the innermost enclosing parallel region.

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
3 ancestor at a given nest level of the current thread or the thread number of the current
4 thread. If the requested nest level is outside the range of 0 and the nest level of the
5 current thread, as returned by the omp_get_level routine, the routine returns -1.
6

7 Note – When the omp_get_ancestor_thread_num routine is called with a value

8 of level=0, the routine always returns 0. If level=omp_get_level(), the routine
9 has the same effect as the omp_get_thread_num routine.
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 The omp_get_team_size routine returns, for a given nested level of the current
18 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 integer function omp_get_team_size(level)
9 integer level

10
11
Fortran

12 Binding
13 The binding thread set for an omp_get_team_size region is the encountering
14 thread. The binding region for an omp_get_team_size region is the innermost
15 enclosing parallel region.

16 Effect
17 The omp_get_team_size routine returns the size of the thread team to which the
18 ancestor or the current thread belongs. If the requested nested level is outside the range
19 of 0 and the nested level of the current thread, as returned by the omp_get_level
20 routine, the routine returns -1. Inactive parallel regions are regarded like active parallel
21 regions executed with one thread.
22

23 Note – When the omp_get_team_size routine is called with a value of level=0,

24 the routine always returns 1. If level=omp_get_level(), the routine has the same
25 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 The omp_get_active_level routine returns the number of nested, active
8 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
21 task. The binding region for an omp_get_active_level region is the innermost
22 enclosing parallel region.

23 Chapter 3 Runtime Library Routines 133

1 Effect
2 The omp_get_active_level routine returns the number of nested, active parallel
3 regions enclosing the task that contains the call. The routine always returns a non-
4 negative integer, and always returns 0 if it is called from the sequential part of the
5 program.

6 Cross References
7 • omp_get_level routine, see Section 3.2.16 on page 129.

9 3.3 Lock Routines

10 The OpenMP runtime library includes a set of general-purpose lock routines that can be
11 used for synchronization. These general-purpose lock routines operate on OpenMP locks
12 that are represented by OpenMP lock variables. An OpenMP lock variable must be
13 accessed only through the routines described in this section; programs that otherwise
14 access OpenMP lock variables are non-conforming.

15 An OpenMP lock may be in one of the following states: uninitialized, unlocked, or

16 locked. If a lock is in the unlocked state, a task may set the lock, which changes its state
17 to locked. The task which sets the lock is then said to own the lock. A task which owns
18 a lock may unset that lock, returning it to the unlocked state. A program in which a task
19 unsets a lock that is owned by another task is non-conforming.

20 Two types of locks are supported: simple locks and nestable locks. A nestable lock may
21 be set multiple times by the same task before being unset; a simple lock may not be set
22 if it is already owned by the task trying to set it. Simple lock variables are associated
23 with simple locks and may only be passed to simple lock routines. Nestable lock
24 variables are associated with nestable locks and may only be passed to nestable lock
25 routines.

26 Constraints on the state and ownership of the lock accessed by each of the lock routines
27 are described with the routine. If these constraints are not met, the behavior of the
28 routine is unspecified.
29 The OpenMP lock routines access a lock variable in such a way that they always read
30 and update the most current value of the lock variable. The lock routines include a flush
31 with no list; the read and update to the lock variable must be implemented as if they are
32 atomic with the flush. Therefore, it is not necessary for an OpenMP program to include
33 explicit flush directives to ensure that the lock variable’s value is consistent among
34 different tasks.

35 134 OpenMP API • Version 3.0 May 2008

1 See Section A.44 on page 271 and Section A.45 on page 274, for examples of using the
2 simple and the nestable lock routines, respectively.

3 Binding

4 The binding thread set for all lock routine regions is all threads. As a consequence, for
5 each OpenMP lock, the lock routine effects relate to all tasks that call the routines,
6 without regard to which team(s) the threads executing the tasks belong.

7 Simple Lock Routines

8
C/C++
9 The type omp_lock_t is an data type capable of representing a simple lock. For the
10 following routines, a simple lock variable must be of omp_lock_t type. All simple
11 lock routines require an argument that is a pointer to a variable of type omp_lock_t.
C/C++
12
Fortran
13 For the following routines, a simple lock variable must be an integer variable of
14 kind=omp_lock_kind.
15
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

1 Nestable Lock Routines:
2
C/C++
3 The type omp_nest_lock_t is an data type capable of representing a nestable lock.
4 For the following routines, a nested lock variable must be of omp_nest_lock_t type.
5 All nestable lock routines require an argument that is a pointer to a variable of type
6 omp_nest_lock_t.
C/C++
7
Fortran
8 For the following routines, a nested lock variable must be an integer variable of
9 kind=omp_nest_lock_kind.
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
15 sets it.
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
18 available.

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

1 Format
C/C++
2
3 void omp_init_lock(omp_lock_t *lock);
4 void omp_init_nest_lock(omp_nest_lock_t *lock);
5
C/C++
Fortran
6
7 subroutine omp_init_lock(svar)
8 integer (kind=omp_lock_kind) svar

9 subroutine omp_init_nest_lock(nvar)
10 integer (kind=omp_nest_lock_kind) nvar
11
12
Fortran

13 Constraints on Arguments
14 A program that accesses a lock that is not in the uninitialized state through either routine
15 is non-conforming.

16 Effect
17 The effect of these routines is to initialize the lock to the unlocked state (that is, no task
18 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.

20 3.3.2 omp_destroy_lock and

21 omp_destroy_nest_lock

22 Summary
23 These routines ensure that the OpenMP lock is uninitialized.

24 Chapter 3 Runtime Library Routines 137

1 Format
C/C++
2
3 void omp_destroy_lock(omp_lock_t *lock);
4 void omp_destroy_nest_lock(omp_nest_lock_t *lock);
5
C/C++
Fortran
6
7 subroutine omp_destroy_lock(svar)
8 integer (kind=omp_lock_kind) svar

9 subroutine omp_destroy_nest_lock(nvar)
10 integer (kind=omp_nest_lock_kind) nvar
11
12
Fortran

13 Constraints on Arguments
14 A program that accesses a lock that is not in the unlocked state through either routine is
15 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
21 suspended until the lock is set.

22 138 OpenMP API • Version 3.0 May 2008

1 Format
C/C++
2
3 void omp_set_lock(omp_lock_t *lock);
4 void omp_set_nest_lock(omp_nest_lock_t *lock);
5
C/C++
Fortran
6
7 subroutine omp_set_lock(svar)
8 integer (kind=omp_lock_kind) svar

9 subroutine omp_set_nest_lock(nvar)
10 integer (kind=omp_nest_lock_kind) nvar
11
12
Fortran

13 Constraints on Arguments
14 A program that accesses a lock that is in the uninitialized state through either routine is
15 non-conforming. A simple lock accessed by omp_set_lock that is in the locked state
16 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
19 specified lock is available and then sets the lock.
20 A simple lock is available if it is unlocked. Ownership of the lock is granted to the task
21 executing the routine.
22 A nestable lock is available if it is unlocked or if it is already owned by the task
23 executing the routine. The task executing the routine is granted, or retains, ownership of
24 the lock, and the nesting count for the lock is incremented.

25 Chapter 3 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 void omp_unset_lock(omp_lock_t *lock);
7 void omp_unset_nest_lock(omp_nest_lock_t *lock);
8
C/C++
Fortran
9
10 subroutine omp_unset_lock(svar)
11 integer (kind=omp_lock_kind) svar

12 subroutine omp_unset_nest_lock(nvar)
13 integer (kind=omp_nest_lock_kind) nvar
14
15
Fortran

16 Constraints on Arguments
17 A program that accesses a lock that is not in the locked state or that is not owned by the
18 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
22 count, and causes the lock to become unlocked if the resulting nesting count is zero.
23 For either routine, if the lock becomes unlocked, and if one or more tasks regions were
24 suspended because the lock was unavailable, the effect is that one task is chosen and
25 given ownership of the lock.

26 140 OpenMP API • Version 3.0 May 2008

1 3.3.5 omp_test_lock and omp_test_nest_lock

2 Summary
3 These routines attempt to set an OpenMP lock but do not suspend execution of the task
4 executing the routine.

5 Format
C/C++
6
7 int omp_test_lock(omp_lock_t *lock);
8 int omp_test_nest_lock(omp_nest_lock_t *lock);
9
C/C++
Fortran
10
11 logical function omp_test_lock(svar)
12 integer (kind=omp_lock_kind) svar

13 integer function omp_test_nest_lock(nvar)

14 integer (kind=omp_nest_lock_kind) nvar
15
16
Fortran

17 Constraints on Arguments
18 A program that accesses a lock that is in the uninitialized state through either routine is
19 non-conforming. The behavior is unspecified if a simple lock accessed by
20 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
23 omp_set_nest_lock, except that they do not suspend execution of the task
24 executing the routine.
25 For a simple lock, the omp_test_lock routine returns true if the lock is successfully
26 set; otherwise, it returns false.

27 Chapter 3 Runtime Library Routines 141

1 For a nestable lock, the omp_test_nest_lock routine returns the new nesting count
2 if the lock is successfully set; otherwise, it returns zero.

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
21 routine’s return value is not guaranteed to be consistent across any set of threads.

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
3 seconds since some “time in the past”. The actual “time in the past” is arbitrary, but it is
4 guaranteed not to change during the execution of the application program. The times
5 returned are “per-thread times”, so they are not required to be globally consistent across
6 all the threads participating in an application.

8 Note – It is anticipated that the routine will be used to measure elapsed times as shown
9 in the following example:
C/C++
10
11 double start;
12 double end;
13 start = omp_get_wtime();
14 ... work to be timed ...
15 end = omp_get_wtime();
16 printf("Work took %f seconds\n", end - start);

17
C/C++
Fortran

18
19 DOUBLE PRECISION START, END
20 START = omp_get_wtime()
21 ... work to be timed ...
22 END = omp_get_wtime()
23 PRINT *, "Work took", END - START, "seconds"
24
25
Fortran
26
27

28 Chapter 3 Runtime Library Routines 143

1 3.4.2 omp_get_wtick

2 Summary
3 The omp_get_wtick routine returns the precision of the timer used by
4 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 The binding thread set for an omp_get_wtick region is the encountering thread. The
15 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
18 successive clock ticks of the timer used by omp_get_wtime.

19 144 OpenMP API • Version 3.0 May 2008

1
2 CHAPTER 4

3 Environment Variables
4

5 This chapter describes the OpenMP environment variables that specify the settings of
6 the ICVs that affect the execution of OpenMP programs (see Section 2.3 on page 28).
7 The names of the environment variables must be upper case. The values assigned to the
8 environment variables are case insensitive and may have leading and trailing white
9 space. Modifications to the environment variables after the program has started, even if
10 modified by the program itself, are ignored by the OpenMP implementation. However,
11 the settings of some of the ICVs can be modified during the execution of the OpenMP
12 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
15 chunk size. It can be set to any of the valid OpenMP schedule types (i.e., static,
16 dynamic, guided, and auto).
17 • OMP_NUM_THREADS sets the nthreads-var ICV for the number of threads to use for
18 parallel regions.
19 • OMP_DYNAMIC sets the dyn-var ICV for the dynamic adjustment of threads to use
20 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 threads created by the OpenMP implementation.
24 • OMP_WAIT_POLICY sets the wait-policy-var ICV that controls the desired behavior
25 of waiting threads.
26 • OMP_MAX_ACTIVE_LEVELS sets the max-active-levels-var ICV that controls the
27 maximum number of nested active parallel regions.
28 • OMP_THREAD_LIMIT sets the thread-limit-var ICV that controls the maximum
29 number of threads participating in the OpenMP program.
30 The examples in this chapter only demonstrate how these variables might be set in Unix
31 C shell (csh) environments. In Korn shell (ksh) and DOS environments the actions are
32 similar, as follows:

33 145
1 • csh:
2
3 setenv OMP_SCHEDULE "dynamic"

4 • ksh:
5
6 export OMP_SCHEDULE="dynamic"

7 • DOS:
8
9 set OMP_SCHEDULE=dynamic

10

11 4.1 OMP_SCHEDULE
12 The OMP_SCHEDULE environment variable controls the schedule type and chunk size
13 of all loop directives that have the schedule type runtime, by setting the value of the
14 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
21 on page 38 for a detailed description of the schedule types.
22 The behavior of the program is implementation defined if the value of OMP_SCHEDULE
23 does not conform to the above format.
24 Implementation specific schedules cannot be specified in OMP_SCHEDULE. They can
25 only be specified by calling omp_set_schedule, described in Section 3.2.11 on page
26 121.
27 Example:
28
29 setenv OMP_SCHEDULE "guided,4"
30 setenv OMP_SCHEDULE "dynamic"

31 146 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.

8 4.2 OMP_NUM_THREADS
9 The OMP_NUM_THREADS environment variable sets the number of threads to use for
10 parallel regions by setting the initial value of the nthreads-var ICV. See Section 2.3
11 for a comprehensive set of rules about the interaction between the OMP_NUM_THREADS
12 environment variable, the num_threads clause, the omp_set_num_threads
13 library routine and dynamic adjustment of threads.
14 The value of this environment variable must be a positive integer. The behavior of the
15 program is implementation defined if the requested value of OMP_NUM_THREADS is
16 greater than the number of threads an implementation can support, or if the value is not
17 a positive integer.
18 Example:
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
3 The OMP_DYNAMIC environment variable controls dynamic adjustment of the number
4 of threads to use for executing parallel regions by setting the initial value of the
5 dyn-var ICV. The value of this environment variable must be true or false. If the
6 environment variable is set to true, the OpenMP implementation may adjust the
7 number of threads to use for executing parallel regions in order to optimize the use
8 of system resources. If the environment variable is set to false, the dynamic
9 adjustment of the number of threads is disabled. The behavior of the program is
10 implementation defined if the value of OMP_DYNAMIC is neither true nor false.
11 Example:
12
13 setenv OMP_DYNAMIC true

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
20 The OMP_NESTED environment variable controls nested parallelism by setting the
21 initial value of the nest-var ICV. The value of this environment variable must be true
22 or false. If the environment variable is set to true, nested parallelism is enabled; if
23 set to false, nested parallelism is disabled. The behavior of the program is
24 implementation defined if the value of OMP_NESTED is neither true nor false.
25 Example:
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

1 • omp_get_nested routine, see Section 3.2.18 on page 131.

3 4.5 OMP_STACKSIZE
4 The OMP_STACKSIZE environment variable controls the size of the stack for threads
5 created by the OpenMP implementation, by setting the value of the stacksize-var ICV.
6 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
11 by the OpenMP implementation.
12 • B, K, M, and G are letters that specify whether the given size is in Bytes, Kilobytes,
13 Megabytes, or Gigabytes, respectively. If one of these letters is present, there may be
14 white space between size and the letter.
15 If only size is specified and none of B, K, M, or G is specified, then size is assumed to be
16 in Kilobytes.
17 The behavior of the program is implementation defined if OMP_STACKSIZE does not
18 conform to the above format, or if the implementation cannot provide a stack with the
19 requested size.
20 Examples:
21
22 setenv OMP_STACKSIZE 2000500B
23 setenv OMP_STACKSIZE "3000 k "
24 setenv OMP_STACKSIZE 10M
25 setenv OMP_STACKSIZE " 10 M "
26 setenv OMP_STACKSIZE "20 m "
27 setenv OMP_STACKSIZE " 1G"
28 setenv OMP_STACKSIZE 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
3 The OMP_WAIT_POLICY environment variable provides a hint to an OpenMP
4 implementation about the desired behavior of waiting threads by setting the wait-policy-
5 var ICV. A compliant OpenMP implementation may or may not abide by the setting of
6 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
10 processor cycles, while waiting. An OpenMP implementation may, for example, make
11 waiting threads spin.
12 The PASSIVE value specifies that waiting threads should mostly be passive, i.e., not
13 consume processor cycles, while waiting. An OpenMP implementation, may for
14 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:
17
18 setenv OMP_WAIT_POLICY ACTIVE
19 setenv OMP_WAIT_POLICY active
20 setenv OMP_WAIT_POLICY PASSIVE
21 setenv OMP_WAIT_POLICY passive

22 Cross References
23 • wait-policy-var ICV, see Section 2.3 on page 24.

24

25 4.7 OMP_MAX_ACTIVE_LEVELS
26 The OMP_MAX_ACTIVE_LEVELS environment variable controls the maximum number
27 of nested active parallel regions by setting the initial value of the max-active-levels-var
28 ICV.

29 150 OpenMP API • Version 3.0 May 2008

1 The value of this environment variable must be a non-negative integer. The behavior of
2 the program is implementation defined if the requested value of
3 OMP_MAX_ACTIVE_LEVELS is greater than the maximum number of nested active
4 parallel levels an implementation can support, or if the value is not a non-negative
5 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.

10

11 4.8 OMP_THREAD_LIMIT
12 The OMP_THREAD_LIMIT environment variable sets the number of OpenMP threads
13 to use for the whole OpenMP program by setting the thread-limit-var ICV.
14 The value of this environment variable must be a positive integer. The behavior of the
15 program is implementation defined if the requested value of OMP_THREAD_LIMIT is
16 greater than the number of threads an implementation can support, or if the value is not
17 a positive integer.

18 Cross References
19 • thread-limit-var ICV, see Section 2.3 on page 28.
20 • omp_get_thread_limit routine

21 Chapter 4 Environment Variables 151

1 152 OpenMP API • Version 3.0 May 2008
1
2 APPENDIX A

3 Examples
4

5 The following are examples of the constructs and routines defined in this document.
6
C/C++
7 A statement following a directive is compound only when necessary, and a non-
8 compound statement is indented with respect to a directive preceding it.
C/C++

10 A.1 A Simple Parallel Loop

11 The following example demonstrates how to parallelize a simple loop using the parallel
12 loop construct (Section 2.6.1 on page 54). The loop iteration variable is private by
13 default, so it is not necessary to specify it explicitly in a private clause.
14
C/C++
15 Example A.1.1c

16 void a1(int n, float *a, float *b)

17 {
18 int i;

19 #pragma omp parallel for

20 for (i=1; i<n; i++) /* i is private by default */
21 b[i] = (a[i] + a[i-1]) / 2.0;
22 }
C/C++

23 153
1
Fortran
2 Example A.1.1f

3 SUBROUTINE A1(N, A, B)

4 INTEGER I, N
5 REAL B(N), A(N)

6 !$OMP PARALLEL DO !I is private by default

7 DO I=2,N
8 B(I) = (A(I) + A(I-1)) / 2.0
9 ENDDO
10 !$OMP END PARALLEL DO

11 END SUBROUTINE A1
12
Fortran

13

14 A.2 The OpenMP Memory Model

15 In the following example, at Print 1, the value of x could be either 2 or 5, depending on
16 the timing of the threads, and the implementation of the assignment to x. There are two
17 reasons that the value at Print 1 might not be 5. First, Print 1 might be executed before
18 the assignment to x is executed. Second, even if Print 1 is executed after the assignment,
19 the value 5 is not guaranteed to be seen by thread 1 because a flush may not have been
20 executed by thread 0 since the assignment.

21 154 OpenMP API • Version 3.0 May 2008

1 The barrier after Print 1 contains implicit flushes on all threads, as well as a thread
2 synchronization, so the programmer is guaranteed that the value 5 will be printed by
3 both Print 2 and Print 3.
4
C/C++
5 Example A.2.1c

6 #include <stdio.h>
7 #include <omp.h>

8 int main(){
9 int x;
10
11 x = 2;
12 #pragma omp parallel num_threads(2) shared(x)
13 {

14 if (omp_get_thread_num() == 0) {
15 x = 5;
16 } else {
17 /* Print 1: the following read of x has a race */
18 printf("1: Thread# %d: x = %d\n", omp_get_thread_num(),x );
19 }
20
21 #pragma omp barrier
22
23 if (omp_get_thread_num() == 0) {
24 /* Print 2 */
25 printf("2: Thread# %d: x = %d\n", omp_get_thread_num(),x );
26 } else {
27 /* Print 3 */
28 printf("3: Thread# %d: x = %d\n", omp_get_thread_num(),x );
29 }
30 }
31 return 0;
32 }
C/C++

33 Appendix A Examples 155

1
Fortran
2 Example A.2.1f

3 PROGRAM A2
4 INCLUDE "omp_lib.h" ! or USE OMP_LIB
5 INTEGER X
6
7 X = 2
8 !$OMP PARALLEL NUM_THREADS(2) SHARED(X)
9
10 IF (OMP_GET_THREAD_NUM() .EQ. 0) THEN
11 X = 5
12 ELSE
13 ! PRINT 1: The following read of x has a race
14 PRINT *,"1: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X
15 ENDIF
16
17 !$OMP BARRIER
18
19 IF (OMP_GET_THREAD_NUM() .EQ. 0) THEN
20 ! PRINT 2
21 PRINT *,"2: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X
22 ELSE
23 ! PRINT 3
24 PRINT *,"3: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X
25 ENDIF
26
27 !$OMP END PARALLEL

28 END PROGRAM A2
29
Fortran
30 The following example demonstrates why synchronization is difficult to perform
31 correctly through variables. The value of flag is undefined in both prints on thread 1 and
32 the value of data is only well-defined in the second print.

33 156 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.2.2c

3 #include <omp.h>
4 #include <stdio.h>
5 int main()
6 {
7 int data;
8 int flag=0;
9 #pragma omp parallel
10 {
11 if (omp_get_thread_num()==0)
12 {
13 /* Write to the data buffer that will be
14 read by thread */
15 data = 42;
16 /* Flush data to thread 1 and strictly order
17 the write to data
18 relative to the write to the flag */
19 #pragma omp flush(flag, data)
20 /* Set flag to release thread 1 */
21 flag = 1;
22 /* Flush flag to ensure that thread 1 sees
23 the change */
24 #pragma omp flush(flag)
25 }
26 else if(omp_get_thread_num()==1)
27 {
28 /* Loop until we see the update to the flag */
29 #pragma omp flush(flag, data)
30 while (flag < 1)
31 {
32 #pragma omp flush(flag, data)
33 }
34 /* Values of flag and data are undefined */
35 printf("flag=%d data=%d\n", flag, data);
36 #pragma omp flush(flag, data)
37 /* Values data will be 42, value of flag
38 still undefined */
39 printf("flag=%d data=%d\n", flag, data);
40 }
41 }
42 }
C/C++

43 Appendix A Examples 157

1
Fortran
2 Example A.2.2f

3 PROGRAM EXAMPLE
4 INCLUDE "omp_lib.h" ! or USE OMP_LIB
5 INTEGER DATA
6 INTEGER FLAG

7 !$OMP PARALLEL
8 IF(OMP_GET_THREAD_NUM() .EQ. 0) THEN
9 ! Write to the data buffer that will be read by thread 1
10 DATA = 42
11 ! Flush DATA to thread 1 and strictly order the write to DATA
12 ! relative to the write to the FLAG
13 !$OMP FLUSH(FLAG, DATA)
14 ! Set FLAG to release thread 1
15 FLAG = 1;
16 ! Flush FLAG to ensure that thread 1 sees the change */
17 !$OMP FLUSH(FLAG)
18 ELSE IF(OMP_GET_THREAD_NUM() .EQ. 1) THEN
19 ! Loop until we see the update to the FLAG
20 !$OMP FLUSH(FLAG, DATA)
21 DO WHILE(FLAG .LT. 1)
22 !$OMP FLUSH(FLAG, DATA)
23 ENDDO

24 ! Values of FLAG and DATA are undefined

25 PRINT *, 'FLAG=', FLAG, ' DATA=', DATA
26 !$OMP FLUSH(FLAG, DATA)

27 !Values DATA will be 42, value of FLAG still undefined */

28 PRINT *, 'FLAG=', FLAG, ' DATA=', DATA
29 ENDIF
30 !$OMP END PARALLEL
31 END
32
Fortran
33 This example demonstrates why synchronization is difficult to perform correctly through
34 variables. The statements on thread 1 and thread 2 may execute in either order.

35 158 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.2.3c

3 #include <omp.h>
4 #include <stdio.h>
5 int main()
6 {
7 int flag=0;

8 #pragma omp parallel

9 {
10 if(omp_get_thread_num()==0)
11 {
12 /* Set flag to release thread 1 */
13 #pragma omp atomic
14 flag++;
15 /* Flush of flag is implied by the atomic directive */
16 }
17 else if(omp_get_thread_num()==1)
18 {
19 /* Loop until we see that flag reaches 1*/
20 #pragma omp flush(flag)
21 while(flag < 1)
22 {
23 #pragma omp flush(flag)
24 }
25 printf("Thread 1 awoken\n");

26 /* Set flag to release thread 2 */

27 #pragma omp atomic
28 flag++;
29 /* Flush of flag is implied by the atomic directive */
30 }
31 else if(omp_get_thread_num()==2)
32 {
33 /* Loop until we see that flag reaches 2 */
34 #pragma omp flush(flag)
35 while(flag < 2)
36 {
37 #pragma omp flush(flag)
38 }
39 printf("Thread 2 awoken\n");
40 }
41 }
42 }
C/C++

43 Appendix A Examples 159

1
Fortran
2 Example A.2.3f

3 PROGRAM EXAMPLE
4 INCLUDE "omp_lib.h" ! or USE OMP_LIB
5 INTEGER FLAG

6 !$OMP PARALLEL
7 IF(OMP_GET_THREAD_NUM() .EQ. 0) THEN
8 ! Set flag to release thread 1
9 !$OMP ATOMIC
10 FLAG = FLAG + 1
11 !Flush of FLAG is implied by the atomic directive
12 ELSE IF(OMP_GET_THREAD_NUM() .EQ. 1) THEN
13 ! Loop until we see that FLAG reaches 1
14 !$OMP FLUSH(FLAG, DATA)
15 DO WHILE(FLAG .LT. 1)
16 !$OMP FLUSH(FLAG, DATA)
17 ENDDO

18 PRINT *, 'Thread 1 awoken'

19 ! Set FLAG to release thread 2

20 !$OMP ATOMIC
21 FLAG = FLAG + 1
22 !Flush of FLAG is implied by the atomic directive
23 ELSE IF(OMP_GET_THREAD_NUM() .EQ. 2) THEN
24 ! Loop until we see that FLAG reaches 2
25 !$OMP FLUSH(FLAG, DATA)
26 DO WHILE(FLAG .LT. 2)
27 !$OMP FLUSH(FLAG, DATA)
28 ENDDO

29 PRINT *, 'Thread 2 awoken'

30 ENDIF
31 !$OMP END PARALLEL
32 END
33
Fortran

34 160 OpenMP API • Version 3.0 May 2008

1

2 A.3 Conditional Compilation

3
C/C++
4 The following example illustrates the use of conditional compilation using the OpenMP
5 macro _OPENMP (Section 2.2 on page 26). With OpenMP compilation, the _OPENMP
6 macro becomes defined.
7
C/C++
8 Example A.3.1c

9 #include <stdio.h>

10 int main()
11 {

12 # ifdef _OPENMP
13 printf("Compiled by an OpenMP-compliant implementation.\n");
14 # endif

15 return 0;
16 }
C/C++
17
Fortran
18 The following example illustrates the use of the conditional compilation sentinel (see
19 Section 2.2 on page 26). With OpenMP compilation, the conditional compilation
20 sentinel !$ is recognized and treated as two spaces. In fixed form source, statements
21 guarded by the sentinel must start after column 6.
22
Fortran
23 Example A.3.1f

24 PROGRAM A3

25 C234567890
26 !$ PRINT *, "Compiled by an OpenMP-compliant implementation."

27 END PROGRAM A3
28
Fortran

29 Appendix A Examples 161

1

2 A.4 Internal Control Variables

3 According to Section 2.3 on page 28, an OpenMP implementation must act as if there
4 are ICVs that control the behavior of the program. This example illustrates two ICVs,
5 nthreads-var and max_active-levels-var. The nthreads-var ICV controls the number of
6 threads requested for encountered parallel regions; there is one copy of this ICV per
7 task. The max-active-levels-var ICV controls the maximum number of nested active
8 parallel regions; there is one copy of this ICV for the whole program.
9 In the following example, the value of the nthreads-var ICV is changed via a call to
10 omp_set_num_threads. The new value of nthreads-var applies only to the implicit
11 tasks that execute the parallel region and make the call to omp_set_num_threads.
12 The max-active-levels-var ICV is global; so its value is the same for all tasks.

13 162 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.4.1c

3 #include <stdio.h>
4 #include <omp.h>

5 int main (void)

6 {
7 omp_set_nested(1);
8 omp_set_max_active_levels(8);
9 omp_set_dynamic(0);
10 omp_set_num_threads(2);
11 #pragma omp parallel
12 {
13 omp_set_num_threads(3);

14 #pragma omp parallel

15 {
16 omp_set_num_threads(4);
17 #pragma omp single
18 {
19 /*
20 * The following should print:
21 * Inner: max_act_lev=8, num_thds=3, max_thds=4
22 * Inner: max_act_lev=8, num_thds=3, max_thds=4
23 */
24 printf ("Inner: max_act_lev=%d, num_thds=%d, max_thds=%d\n",
25 omp_get_max_active_levels(), omp_get_num_threads(),
26 omp_get_max_threads());
27 }
28 }

29 #pragma omp barrier

30 #pragma omp single
31 {
32 /*
33 * The following should print:
34 * Outer: max_act_lev=8, num_thds=2, max_thds=3
35 */
36 printf ("Outer: max_act_lev=%d, num_thds=%d, max_thds=%d\n",
37 omp_get_max_active_levels(), omp_get_num_threads(),
38 omp_get_max_threads());
39 }
40 }
41 }
C/C++

42 Appendix A Examples 163

1
Fortran
2 Example A.4.1f

3 program icv
4 use omp_lib

5 call omp_set_nested(.true.)
6 call omp_set_max_active_levels(8)
7 call omp_set_dynamic(.false.)
8 call omp_set_num_threads(2)

9 !$omp parallel
10 call omp_set_num_threads(3)

11 !$omp parallel
12 call omp_set_num_threads(4)
13 !$omp single!
14 ! The following should print:
15 ! Inner: max_act_lev= 8 , num_thds= 3 , max_thds= 4
16 ! Inner: max_act_lev= 8 , num_thds= 3 , max_thds= 4
17 print *, ("Inner: max_act_lev=", omp_get_max_active_levels(),
18 & ", num_thds=", omp_get_num_threads(),
19 & ", max_thds=", omp_get_max_threads())
20 !$omp end single
21 !$omp end parallel

22 !$omp barrier
23 !$omp single
24 ! The following should print:
25 ! Outer: max_act_lev= 8 , num_thds= 2 , max_thds= 3
26 print *, ("Outer: max_act_lev=", omp_get_max_active_levels(),
27 & ", num_thds=", omp_get_num_threads(),
28 & ", max_thds=", omp_get_max_threads())
29 !$omp end single
30 !$omp end parallel
31 end
32
Fortran

33

34 A.5 The parallel Construct

35 The parallel construct (Section 2.4 on page 32) can be used in coarse-grain parallel
36 programs. In the following example, each thread in the parallel region decides what
37 part of the global array x to work on, based on the thread number:

38 164 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.5.1c

3 #include <omp.h>

4 void subdomain(float *x, int istart, int ipoints)

5 {
6 int i;

7 for (i = 0; i < ipoints; i++)

8 x[istart+i] = 123.456;
9 }

10 void sub(float *x, int npoints)

11 {
12 int iam, nt, ipoints, istart;

13 #pragma omp parallel default(shared) private(iam,nt,ipoints,istart)

14 {
15 iam = omp_get_thread_num();
16 nt = omp_get_num_threads();
17 ipoints = npoints / nt; /* size of partition */
18 istart = iam * ipoints; /* starting array index */
19 if (iam == nt-1) /* last thread may do more */
20 ipoints = npoints - istart;
21 subdomain(x, istart, ipoints);
22 }
23 }

24 int main()
25 {
26 float array[10000];

27 sub(array, 10000);

28 return 0;
29 }
C/C++

30 Appendix A Examples 165

1
Fortran
2 Example A.5.1f

3 SUBROUTINE SUBDOMAIN(X, ISTART, IPOINTS)

4 INTEGER ISTART, IPOINTS
5 REAL X(*)

6 INTEGER I
7
8 DO 100 I=1,IPOINTS
9 X(ISTART+I) = 123.456
10 100 CONTINUE

11 END SUBROUTINE SUBDOMAIN

12 SUBROUTINE SUB(X, NPOINTS)

13 INCLUDE "omp_lib.h" ! or USE OMP_LIB
14
15 REAL X(*)
16 INTEGER NPOINTS
17 INTEGER IAM, NT, IPOINTS, ISTART

18 !$OMP PARALLEL DEFAULT(PRIVATE) SHARED(X,NPOINTS)

19
20 IAM = OMP_GET_THREAD_NUM()
21 NT = OMP_GET_NUM_THREADS()
22 IPOINTS = NPOINTS/NT
23 ISTART = IAM * IPOINTS
24 IF (IAM .EQ. NT-1) THEN
25 IPOINTS = NPOINTS - ISTART
26 ENDIF
27 CALL SUBDOMAIN(X,ISTART,IPOINTS)

28 !$OMP END PARALLEL

29 END SUBROUTINE SUB

30 PROGRAM A5
31 REAL ARRAY(10000)
32 CALL SUB(ARRAY, 10000)
33 END PROGRAM A5
34
Fortran

35

36 A.6 The num_threads Clause

37 The following example demonstrates the num_threads clause (Section 2.4 on page
38 32). The parallel region is executed with a maximum of 10 threads.

39 166 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.6.1c

3 #include <omp.h>
4 int main()
5 {
6 omp_set_dynamic(1);

7 #pragma omp parallel num_threads(10)

8 {
9 /* do work here */
10 }
11 return 0;
12 }
C/C++
13
Fortran
14 Example A.6.1f

15 PROGRAM A6
16 INCLUDE "omp_lib.h" ! or USE OMP_LIB
17 CALL OMP_SET_DYNAMIC(.TRUE.)

18 !$OMP PARALLEL NUM_THREADS(10)

19 ! do work here
20 !$OMP END PARALLEL
21 END PROGRAM A6
22
Fortran
23
Fortran
24

25 A.7 Fortran Restrictions on the do Construct

26 If an end do directive follows a do-construct in which several DO statements share a
27 DO termination statement, then a do directive can only be specified for the first (i.e.
28 outermost) of these DO statements. For more information, see Section 2.5.1 on page 38.
29 The following example contains correct usages of loop constructs:

30 Appendix A Examples 167

1 Example A.7.1f

2 SUBROUTINE WORK(I, J)
3 INTEGER I,J
4 END SUBROUTINE WORK

5 SUBROUTINE A7_GOOD()
6 INTEGER I, J
7 REAL A(1000)

8 DO 100 I = 1,10
9 !$OMP DO
10 DO 100 J = 1,10
11 CALL WORK(I,J)
12 100 CONTINUE ! !$OMP ENDDO implied here

13 !$OMP DO
14 DO 200 J = 1,10
15 200 A(I) = I + 1
16 !$OMP ENDDO

17 !$OMP DO
18 DO 300 I = 1,10
19 DO 300 J = 1,10
20 CALL WORK(I,J)
21 300 CONTINUE
22 !$OMP ENDDO
23 END SUBROUTINE A7_GOOD

24 The following example is non-conforming because the matching do directive for the
25 end do does not precede the outermost loop:

26 Example A.7.2f

27 SUBROUTINE WORK(I, J)
28 INTEGER I,J
29 END SUBROUTINE WORK

30 SUBROUTINE A7_WRONG
31 INTEGER I, J

32 DO 100 I = 1,10
33 !$OMP DO
34 DO 100 J = 1,10
35 CALL WORK(I,J)
36 100 CONTINUE
37 !$OMP ENDDO
38 END SUBROUTINE A7_WRONG
39
Fortran

40 168 OpenMP API • Version 3.0 May 2008

1
Fortran
2

3 A.8 Fortran Private Loop Iteration Variables

4 In general loop iteration variables will be private, when used in the do-loop of a do and
5 parallel do construct or in sequential loops in a parallel construct (see
6 Section 2.5.1 on page 38 and Section 2.9.1 on page 77). In the following example of a
7 sequential loop in a parallel construct the loop iteration variable I will be private.

8 Example A.8.1f

9 SUBROUTINE A8_1(A,N)
10 INCLUDE "omp_lib.h" ! or USE OMP_LIB

11 REAL A(*)
12 INTEGER I, MYOFFSET, N

13 !$OMP PARALLEL PRIVATE(MYOFFSET)

14 MYOFFSET = OMP_GET_THREAD_NUM()*N
15 DO I = 1, N
16 A(MYOFFSET+I) = FLOAT(I)
17 ENDDO
18 !$OMP END PARALLEL

19 END SUBROUTINE A8_1

20 Appendix A Examples 169

1 In exceptional cases, loop iteration variables can be made shared, as in the following
2 example:

3 Example A.8.2f

4 SUBROUTINE A8_2(A,B,N,I1,I2)
5 REAL A(*), B(*)
6 INTEGER I1, I2, N

7 !$OMP PARALLEL SHARED(A,B,I1,I2)

8 !$OMP SECTIONS
9 !$OMP SECTION
10 DO I1 = I1, N
11 IF (A(I1).NE.0.0) EXIT
12 ENDDO
13 !$OMP SECTION
14 DO I2 = I2, N
15 IF (B(I2).NE.0.0) EXIT
16 ENDDO
17 !$OMP END SECTIONS
18 !$OMP SINGLE
19 IF (I1.LE.N) PRINT *, 'ITEMS IN A UP TO ', I1, 'ARE ALL ZERO.'
20 IF (I2.LE.N) PRINT *, 'ITEMS IN B UP TO ', I2, 'ARE ALL ZERO.'
21 !$OMP END SINGLE
22 !$OMP END PARALLEL

23 END SUBROUTINE A8_2

24 Note however that the use of shared loop iteration variables can easily lead to race
25 conditions.
26
Fortran

27

28 A.9 The nowait clause

29 If there are multiple independent loops within a parallel region, you can use the
30 nowait clause (see Section 2.5.1 on page 38) to avoid the implied barrier at the end of
31 the loop construct, as follows:

32 170 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.9.1c

3 #include <math.h>

4 void a9(int n, int m, float *a, float *b, float *y, float *z)
5 {
6 int i;
7 #pragma omp parallel
8 {
9 #pragma omp for nowait
10 for (i=1; i<n; i++)
11 b[i] = (a[i] + a[i-1]) / 2.0;
12
13 #pragma omp for nowait
14 for (i=0; i<m; i++)
15 y[i] = sqrt(z[i]);
16 }
17 }
C/C++
18
Fortran
19 Example A.9.1f

20 SUBROUTINE A9(N, M, A, B, Y, Z)

21 INTEGER N, M
22 REAL A(*), B(*), Y(*), Z(*)

23 INTEGER I

24 !$OMP PARALLEL

25 !$OMP DO
26 DO I=2,N
27 B(I) = (A(I) + A(I-1)) / 2.0
28 ENDDO
29 !$OMP END DO NOWAIT

30 !$OMP DO
31 DO I=1,M
32 Y(I) = SQRT(Z(I))
33 ENDDO
34 !$OMP END DO NOWAIT

35 !$OMP END PARALLEL

36 END SUBROUTINE A9
37
Fortran

38 Appendix A Examples 171

1 The following examples show the use of nowait with static scheduling.
C/C++
2 Example A.9.2c
3 #include <math.h>
4 void a92(int n, float *a, float *b, float *c, float *y, float *z)
5 {
6 int i;
7 #pragma omp parallel
8 {
9 #pragma omp for schedule(static) nowait
10 for (i=0; i<n; i++)
11 c[i] = (a[i] + b[i]) / 2.0;
12 #pragma omp for schedule(static) nowait
13 for (i=0; i<n; i++)
14 z[i] = sqrt(c[i]);
15 #pragma omp for schedule(static) nowait
16 for (i=1; i<=n; i++)
17 y[i] = z[i-1] + a[i];
18 }
19 }
C/C++
Fortran
20 Example A.9.2f
21 SUBROUTINE A92(N, A, B, C, Y, Z)
22 INTEGER N
23 REAL A(*), B(*), C(*), Y(*), Z(*)
24 INTEGER I
25 !$OMP PARALLEL
26 !$OMP DO SCHEDULE(STATIC)
27 DO I=1,N
28 C(I) = (A(I) + B(I)) / 2.0
29 ENDDO
30 !$OMP END DO NOWAIT
31 !$OMP DO SCHEDULE(STATIC)
32 DO I=1,N
33 Z(I) = SQRT(C(I))
34 ENDDO
35 !$OMP END DO NOWAIT
36 !$OMP DO SCHEDULE(STATIC)
37 DO I=2,N+1
38 Y(I) = Z(I-1) + A(I)
39 ENDDO
40 !$OMP END DO NOWAIT
41 !$OMP END PARALLEL
42 END SUBROUTINE A92
43
Fortran

44 172 OpenMP API • Version 3.0 May 2008

1

2 A.10 The collapse clause

3 In the next example, the loops over k and j are collapsed and their iteration space is
4 executed by all threads of the current team.
5
Fortran
6 Example A.10.1f

7 subroutine sub()
8 !$omp do collapse(2) private(i,j,k)
9 do k = kl, ku, ks
10 do j = jl, ju, js
11 do i = il, iu, is
12 call bar(a,i,j,k)
13 enddo
14 enddo
15 enddo
16 !$omp end do
17 end subroutine
18
Fortran
19 In the next example, the loops over k and j are collapsed and their iteration space is
20 executed by all threads of the current team. The example prints: 2 3.
21
Fortran
22 Example A.10.2f

23 program test
24 !$omp parallel
25 !$omp do private(j,k) collapse(2) lastprivate(jlast, klast)
26 do k = 1,2
27 do j = 1,3
28 jlast=j
29 klast=k
30 enddo
31 enddo
32 !$omp end do
33 !$omp single
34 print *, klast, jlast
35 !$omp end single
36 !$omp end parallel
37 end program test
38
Fortran

39 Appendix A Examples 173

1 The next example illustrates use of the ordered construct with the collapse
2 construct. Since both loops are collapsed into one, the ordered construct has to be
3 inside all loops associated with the do construct. Since an iteration may not execute
4 more than one ordered region this program would be wrong without the
5 collapse(2) clause. The code prints
6 011
7 012
8 021
9 122
10 131
11 132
12
Fortran
13 Example A.10.3f

14 program test
15 include 'omp_lib.h'
16 !$omp parallel num_threads(2)
17 !$omp do collapse(2) ordered private(j,k) schedule(static,3)
18 do k = 1,3
19 do j = 1,2
20 !$omp ordered
21 print *, omp_get_thread_num(), k, j
22 !$omp end ordered
23 call work(a,j,k)
24 enddo
25 enddo
26 !$omp end do
27 !$omp end parallel
28 end program test
29
Fortran

30

31 A.11 The parallel sections Construct

32 In the following example (for Section 2.5.2 on page 47) routines xaxis, yaxis, and zaxis
33 can be executed concurrently. The first section directive is optional. Note that all
34 section directives need to appear in the parallel sections construct.

35 174 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.11.1c

3 void XAXIS();
4 void YAXIS();
5 void ZAXIS();

6 void a11()
7 {
8 #pragma omp parallel sections
9 {
10 #pragma omp section
11 XAXIS();

12 #pragma omp section

13 YAXIS();
14
15 #pragma omp section
16 ZAXIS();
17 }
18 }
C/C++
19
Fortran
20 Example A.11.1f

21 SUBROUTINE A11()

22 !$OMP PARALLEL SECTIONS

23 !$OMP SECTION
24 CALL XAXIS()

25 !$OMP SECTION
26 CALL YAXIS()

27 !$OMP SECTION
28 CALL ZAXIS()

29 !$OMP END PARALLEL SECTIONS

30 END SUBROUTINE A11

31
Fortran

32 Appendix A Examples 175

1

2 A.12 The single Construct

3 The following example demonstrates the single construct (Section 2.5.3 on page 49).
4 In the example, only one thread prints each of the progress messages. All other threads
5 will skip the single region and stop at the barrier at the end of the single construct
6 until all threads in the team have reached the barrier. If other threads can proceed
7 without waiting for the thread executing the single region, a nowait clause can be
8 specified, as is done in the third single construct in this example. The user must not
9 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 void work1() {}
14 void work2() {}

15 void a12()
16 {
17 #pragma omp parallel
18 {
19 #pragma omp single
20 printf("Beginning work1.\n");

21 work1();

22 #pragma omp single

23 printf("Finishing work1.\n");

24 #pragma omp single nowait

25 printf("Finished work1 and beginning work2.\n");

26 work2();
27 }
28 }
C/C++

29 176 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.12.1f

3 SUBROUTINE WORK1()
4 END SUBROUTINE WORK1
5
6 SUBROUTINE WORK2()
7 END SUBROUTINE WORK2
8
9 PROGRAM A12
10 !$OMP PARALLEL

11 !$OMP SINGLE
12 print *, "Beginning work1."
13 !$OMP END SINGLE

14 CALL WORK1()

15 !$OMP SINGLE
16 print *, "Finishing work1."
17 !$OMP END SINGLE

18 !$OMP SINGLE
19 print *, "Finished work1 and beginning work2."
20 !$OMP END SINGLE NOWAIT

21 CALL WORK2()

22 !$OMP END PARALLEL

23 END PROGRAM A12

24
Fortran

25 A.13 Tasking Constructs

26 The following example shows how to traverse a tree-like structure using explicit tasks.
27 Note that the traverse function should be called from within a parallel region for the
28 different specified tasks to be executed in parallel. Also, note that the tasks will be
29 executed in no specified order because there are no synchronization directives. Thus,
30 assuming that the traversal will be done in post order, as in the sequential code, is
31 wrong.

32 Appendix A Examples 177

1
C/C++
2 Example A.13.1c

3 struct node {
4 struct node *left;
5 struct node *right;
6 };
7 extern void process(struct node *);
8 void traverse( struct node *p ) {
9 if (p->left)
10 #pragma omp task // p is firstprivate by default
11 traverse(p->left);
12 if (p->right)
13 #pragma omp task // p is firstprivate by default
14 traverse(p->right);
15 process(p);
16 }
C/C++
17
Fortran
18 Example A.13.1f

19 RECURSIVE SUBROUTINE traverse ( P )

20 TYPE Node
21 TYPE(Node), POINTER :: left, right
22 END TYPE Node
23 TYPE(Node) :: P
24 IF (associated(P%left)) THEN
25 !$OMP TASK ! P is firstprivate by default
26 call traverse(P%left)
27 !$OMP END TASK
28 ENDIF
29 IF (associated(P%right)) THEN
30 !$OMP TASK ! P is firstprivate by default
31 call traverse(P%right)
32 !$OMP END TASK
33 ENDIF
34 CALL process ( P )
35 END SUBROUTINE
36
Fortran

37 178 OpenMP API • Version 3.0 May 2008

1 In the next example, we force a postorder traversal of the tree by adding a taskwait
2 directive. Now, we can safely assume that the left and right sons have been executed
3 before we process the current node.
4
C/C++
5 Example A.13.2c

6 struct node {
7 struct node *left;
8 struct node *right;
9 };
10 extern void process(struct node *);
11 void postorder_traverse( struct node *p ) {
12 if (p->left)
13 #pragma omp task // p is firstprivate by default
14 postorder_traverse(p->left);
15 if (p->right)
16 #pragma omp task // p is firstprivate by default
17 postorder_traverse(p->right);
18 #pragma omp taskwait
19 process(p);
20 }
C/C++
21
Fortran
22 Example A.13.2f

23 RECURSIVE SUBROUTINE traverse ( P )

24 TYPE Node
25 TYPE(Node), POINTER :: left, right
26 END TYPE Node
27 TYPE(Node) :: P
28 IF (associated(P%left)) THEN
29 !$OMP TASK ! P is firstprivate by default
30 call traverse(P%left)
31 !$OMP END TASK
32 ENDIF
33 IF (associated(P%right)) THEN
34 !$OMP TASK ! P is firstprivate by default
35 call traverse(P%right)
36 !$OMP END TASK
37 ENDIF
38 !$OMP TASKWAIT
39 CALL process ( P )
40 END SUBROUTINE
41
Fortran

42 Appendix A Examples 179

1 The following example demonstrates how to use the task construct to process elements
2 of a linked list in parallel. The pointer p is firstprivate by default on the task construct
3 so it is not necessary to specify it in a firstprivate clause.
4
C/C++
5 Example A.13.3c

6 typedef struct node node;

7 struct node {
8 int data;
9 node * next;
10 };
11
12 void process(node * p)
13 {
14 /* do work here */
15 }
16 void increment_list_items(node * head)
17 {
18 #pragma omp parallel
19 {
20 #pragma omp single
21 {
22 node * p = head;
23 while (p) {
24 #pragma omp task
25 // p is firstprivate by default
26 process(p);
27 p = p->next;
28 }
29 }
30 }
31 }
C/C++

32 180 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.13.3f

3 MODULE LIST
4 TYPE NODE
5 INTEGER :: PAYLOAD
6 TYPE (NODE), POINTER :: NEXT
7 END TYPE NODE
8 CONTAINS
9 SUBROUTINE PROCESS(p)
10 TYPE (NODE), POINTER :: P
11 ! do work here
12 END SUBROUTINE
13 SUBROUTINE INCREMENT_LIST_ITEMS (HEAD)
14 TYPE (NODE), POINTER :: HEAD
15 TYPE (NODE), POINTER :: P
16 !$OMP PARALLEL PRIVATE(P)
17 !$OMP SINGLE
18 P => HEAD
19 DO
20 !$OMP TASK
21 ! P is firstprivate by default
22 CALL PROCESS(P)
23 !$OMP END TASK
24 P => P%NEXT
25 IF ( .NOT. ASSOCIATED (P) ) EXIT
26 END DO
27 !$OMP END SINGLE
28 !$OMP END PARALLEL
29 END SUBROUTINE
30 END MODULE
31
Fortran

32 Appendix A Examples 181

1 This example calculates a Fibonacci number. If a call to this function is encountered by
2 a single thread in a parallel region, a nested task region will be spawned to carry out the
3 computation in parallel.
4
C/C++
5 Example A.13.4c

6 int fib(int n) {
7 int i, j;
8 if (n<2)
9 return n;
10 else {
11 #pragma omp task shared(i)
12 i=fib(n-1);
13 #pragma omp task shared(j)
14 j=fib(n-2);
15 #pragma omp taskwait
16 return i+j;
17 }
18 }
C/C++
19
Fortran
20 Example A.13.4f

21 RECURSIVE INTEGER FUNCTION fib(n)

22 INTEGER n, i, j
23 IF ( n .LT. 2) THEN
24 fib = n
25 ELSE
26 !$OMP TASK SHARED(i)
27 i = fib( n-1 )
28 !$OMP END TASK
29 !$OMP TASK SHARED(j)
30 j = fib( n-2 )
31 !$OMP END TASK
32 !$OMP END TASKWAIT
33 fib = i+j
34 END IF
35 END FUNCTION
36
Fortran
37 Note: Fibonacci number computation is a classic computer science example for showing
38 recursion, but it is also a classic example showing that a simple algorithm may not be
39 efficient. A more efficient method is through the calculation of exponentiation over an
40 integer matrix. Here we used the classic recursion algorithm for illustrative purposes.

41 182 OpenMP API • Version 3.0 May 2008

1 The following example demonstrates a way to generate a large number of tasks with one
2 thread and execute them with the threads in the parallel team. While generating these
3 tasks, the implementation may reach its limit on unassigned tasks. If it does, the
4 implementation is allowed to cause the thread executing the task generating loop to
5 suspend its task at the task scheduling point in the task directive, and start executing
6 unassigned tasks. Once the number of unassigned tasks is sufficiently low, the thread
7 may resume execution of the task generating loop.
8
C/C++
9 Example A.13.5c

10 #define LARGE_NUMBER 10000000

11 double item[LARGE_NUMBER];
12 extern void process(double);
13
14 int main() {
15 #pragma omp parallel
16 {
17 #pragma omp single
18 {
19 int i;
20 for (i=0; i<LARGE_NUMBER; i++)
21 #pragma omp task // i is firstprivate, item is shared
22 process(item[i]);
23 }
24 }
25 }
C/C++
26
Fortran
27 Example A.13.5f

28 real*8 item(10000000)
29 integer i
30
31 !$omp parallel
32 !$omp single ! loop iteration variable i is private
33 do i=1,10000000
34 !$omp task
35 ! i is firstprivate, item is shared
36 call process(item(i))
37 !$omp end task
38 end do
39 !$omp end single
40 !$omp end parallel
41 end
42
Fortran

43 Appendix A Examples 183

1 The following example is the same as the previous one, except that the tasks are
2 generated in an untied task. While generating the tasks, the implementation may reach
3 its limit on unassigned tasks. If it does, the implementation is allowed to cause the
4 thread executing the task generating loop to suspend its task at the task scheduling point
5 in the task directive, and start executing unassigned tasks. If that thread begins
6 execution of a task that takes a long time to complete, the other threads may complete
7 all the other tasks before it is finished.
8 In this case, since the loop is in an untied task, any other thread is eligible to resume the
9 task generating loop. In the previous examples, the other threads would be forced to
10 wait idly until the generating thread finishes its long task, since the task generating loop
11 was in a tied task.
12
C/C++
13 Example A.13.6c

14 #define LARGE_NUMBER 10000000

15 double item[LARGE_NUMBER];
16 extern void process(double);
17 int main() {
18 #pragma omp parallel
19 {
20 #pragma omp single
21 {
22 int i;
23 #pragma omp task untied
24 // i is firstprivate, item is shared
25 {
26 for (i=0; i<LARGE_NUMBER; i++)
27 #pragma omp task
28 process(item[i]);
29 }
30 }
31 }
32 }
C/C++

33 184 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.13.6f

3 real*8 item(10000000)
4 !$omp parallel
5 !$omp single
6 !$omp task untied
7 ! loop iteration variable i is private
8 do i=1,10000000
9 !$omp task ! i is firstprivate, item is shared
10 call process(item(i))
11 !$omp end task
12 end do
13 !$omp end task
14 !$omp end single
15 !$omp end parallel
16 end
17
Fortran
18 The following two examples demonstrate how the scheduling rules illustrated in
19 Section 2.7.1 on page 62 affect the usage of threadprivate variables in tasks. The value
20 of a threadprivate variable will change across task scheduling points if the executing
21 thread executes a part of another schedulable task that modifies the variable. In tied
22 tasks, the user can control where task scheduling points appear in the code.
23 A single thread may execute both of the task regions that modify tp. The parts of these
24 task regions in which tp is modified may be executed in any order so the resulting value
25 of var can be either 1 or 2.

26 Appendix A Examples 185

1
C/C++
2 Example A.13.7c
3 int tp;
4 #pragma omp threadprivate(tp)
5 int var;
6 void work()
7 {
8 #pragma omp task
9 {
10 /* do work here */
11 #pragma omp task
12 {
13 tp = 1;
14 /* do work here */
15 #pragma omp task
16 {
17 /* no modification of tp */
18 }
19 var = tp; //value of tp can be 1 or 2
20 }
21 tp = 2;
22 }
23 }
C/C++
Fortran
24 Example A.13.7f
25 module example
26 integer tp
27 !$omp threadprivate(tp)
28 integer var
29 contains
30 subroutine work
31 use globals
32 !$omp task
33 ! do work here
34 !$omp task
35 tp = 1
36 ! do work here
37 !$omp task
38 ! no modification of tp
39 !$omp end task
40 var = tp ! value of var can be 1 or 2
41 !$omp end task
42 tp = 2
43 !$omp end task
44 end subroutine
45 end module
46
Fortran

47 186 OpenMP API • Version 3.0 May 2008

1 In this example, scheduling rules prohibit a single thread from scheduling a new task
2 that modifies tp while another such task region is suspended. Therefore, the value
3 written will persist across the task scheduling point.
C/C++
4 Example A.13.8c
5 #include <omp.h>
6 int tp;
7 #pragma omp threadprivate(tp)
8 int var;
9 void work()
10 {
11 #pragma omp parallel
12 {
13 /* do work here */
14 #pragma omp task
15 {
16 tp++;
17 /* do work here */
18 #pragma omp task
19 {
20 /* do work here but don't modify tp */
21 }
22 var = tp; //Value does not change after write above
23 }
24 }
25 }
C/C++
Fortran
26 Example A.13.8f
27 module example
28 integer tp
29 !$omp threadprivate(tp)
30 integer var
31 contains
32 subroutine work
33 !$omp parallel
34 ! do work here
35 !$omp task
36 tp = tp + 1
37 ! do work here
38 !$omp task
39 ! do work here but don't modify tp
40 !$omp end task
41 var = tp ! value does not change after write above
42 !$omp end task
43 !$omp end parallel
44 end subroutine
45 end module
46
Fortran

47 Appendix A Examples 187

1 The following two examples demonstrate how the scheduling rules illustrated in
2 Section 2.7.1 on page 62 affect the usage of locks and critical sections in tasks. If a lock
3 is held across a task scheduling point, no attempt should be made to acquire the same
4 lock in any code that may be interleaved. Otherwise, a deadlock is possible.
5 In the example below, suppose the thread executing task 1 defers task 2. When it
6 encounters the task scheduling point at task 3, it could suspend task 1 and begin task 2
7 which will result in a deadlock when it tries to enter critical region 1.
8
C/C++
9 Example A.13.9c

10 void work()
11 {
12 #pragma omp task
13 { //Task 1
14 #pragma omp task
15 { //Task 2
16 #pragma omp critical //Critical region 1
17 {/*do work here */ }
18 }
19 #pragma omp critical //Critical Region 2
20 {
21 //Capture data for the following task
22 #pragma omp task
23 { /* do work here */ } //Task 3
24 }
25 }
26 }
C/C++

27 188 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.13.9f

3 module example
4 contains
5 subroutine work
6 !$omp task
7 ! Task 1
8 !$omp task
9 ! Task 2
10 !$omp critical
11 ! Critical region 1
12 ! do work here
13 !$omp end critical
14 !$omp end task
15 !$omp critical
16 ! Critical region 2
17 ! Capture data for the following task
18 !$omp task
19 !Task 3
20 ! do work here
21 !$omp end task
22 !$omp end critical
23 !$omp end task
24 end subroutine
25 end module
26
Fortran

27 Appendix A Examples 189

1 In the following example, lock is held across a task scheduling point. However,
2 according to the scheduling restrictions outlined in Section 2.7.1 on page 62, the
3 executing thread can't begin executing one of the non-descendant tasks that also acquires
4 lock before the task region is complete. Therefore, no deadlock is possible.
5
C/C++
6 Example A.13.10c

7 #include <omp.h>
8 void work() {
9 omp_lock_t lock;
10 #pragma omp parallel
11 {
12 int i;
13 #pragma omp for
14 for (i = 0; i < 100; i++) {
15 #pragma omp task
16 {
17 // lock is shared by default in the task
18 omp_set_lock(&lock);
19 // Capture data for the following task
20 #pragma omp task
21 // Task Scheduling Point 1
22 { /* do work here */ }
23 omp_unset_lock(&lock);
24 }
25 }
26 }
27 }
C/C++

28 190 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.13.10f

3 module example
4 include 'omp_lib.h'
5 integer (kind=omp_lock_kind) lock
6 integer i
7 contains
8 subroutine work
9 !$omp parallel
10 !$omp do
11 do i=1,100
12 !$omp task
13 ! Outer task
14 call omp_set_lock(lock) ! lock is shared by
15 ! default in the task
16 ! Capture data for the following task
17 !$omp task ! Task Scheduling Point 1
18 ! do work here
19 !$omp end task
20 call omp_unset_lock(lock)
21 !$omp end task
22 end do
23 !$omp end parallel
24 end subroutine
25 end module
26
Fortran
27
Fortran
28

29 A.14 The workshare Construct

30 The following are examples of the workshare construct (see Section 2.5.4 on page
31 51).

32 Appendix A Examples 191

1
Fortran (cont.)
2 In the following example, workshare spreads work across the threads executing the
3 parallel region, and there is a barrier after the last statement. Implementations must
4 enforce Fortran execution rules inside of the workshare block.

5 Example A.14.1f

6 SUBROUTINE A14_1(AA, BB, CC, DD, EE, FF, N)

7 INTEGER N
8 REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N), EE(N,N), FF(N,N)

9 !$OMP PARALLEL
10 !$OMP WORKSHARE
11 AA = BB
12 CC = DD
13 EE = FF
14 !$OMP END WORKSHARE
15 !$OMP END PARALLEL

16 END SUBROUTINE A14_1

17 In the following example, the barrier at the end of the first workshare region is
18 eliminated with a nowait clause. Threads doing CC = DD immediately begin work on
19 EE = FF when they are done with CC = DD.

20 Example A.14.2f

21 SUBROUTINE A14_2(AA, BB, CC, DD, EE, FF, N)

22 INTEGER N
23 REAL AA(N,N), BB(N,N), CC(N,N)
24 REAL DD(N,N), EE(N,N), FF(N,N)

25 !$OMP PARALLEL
26 !$OMP WORKSHARE
27 AA = BB
28 CC = DD
29 !$OMP END WORKSHARE NOWAIT
30 !$OMP WORKSHARE
31 EE = FF
32 !$OMP END WORKSHARE
33 !$OMP END PARALLEL
34 END SUBROUTINE A14_2

35 192 OpenMP API • Version 3.0 May 2008

1
Fortran (cont.)
2 The following example shows the use of an atomic directive inside a workshare
3 construct. The computation of SUM(AA) is workshared, but the update to R is atomic.

4 Example A.14.3f

5 SUBROUTINE A14_3(AA, BB, CC, DD, N)

6 INTEGER N
7 REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N)
8 REAL R

9 R=0
10 !$OMP PARALLEL
11 !$OMP WORKSHARE
12 AA = BB
13 !$OMP ATOMIC
14 R = R + SUM(AA)
15 CC = DD
16 !$OMP END WORKSHARE
17 !$OMP END PARALLEL

18 END SUBROUTINE A14_3

19 Fortran WHERE and FORALL statements are compound statements, made up of a control
20 part and a statement part. When workshare is applied to one of these compound
21 statements, both the control and the statement parts are workshared. The following
22 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

29 Appendix A Examples 193

1
Fortran (cont.)
2 Example A.14.4f

3 SUBROUTINE A14_4(AA, BB, CC, DD, EE, FF, GG, HH, N)

4 INTEGER N
5 REAL AA(N,N), BB(N,N), CC(N,N)
6 REAL DD(N,N), EE(N,N), FF(N,N)
7 REAL GG(N,N), HH(N,N)

8 !$OMP PARALLEL
9 !$OMP WORKSHARE
10 AA = BB
11 CC = DD
12 WHERE (EE .ne. 0) FF = 1 / EE
13 GG = HH
14 !$OMP END WORKSHARE
15 !$OMP END PARALLEL
16
17 END SUBROUTINE A14_4

18 In the following example, an assignment to a shared scalar variable is performed by one

19 thread in a workshare while all other threads in the team wait.

20 Example A.14.5f

21 SUBROUTINE A14_5(AA, BB, CC, DD, N)

22 INTEGER N
23 REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N)

24 INTEGER SHR

25 !$OMP PARALLEL SHARED(SHR)

26 !$OMP WORKSHARE
27 AA = BB
28 SHR = 1
29 CC = DD * SHR
30 !$OMP END WORKSHARE
31 !$OMP END PARALLEL
32
33 END SUBROUTINE A14_5

34 The following example contains an assignment to a private scalar variable, which is

35 performed by one thread in a workshare while all other threads wait. It is non-
36 conforming because the private scalar variable is undefined after the assignment
37 statement.

38 194 OpenMP API • Version 3.0 May 2008

1 Example A.14.6f

2 SUBROUTINE A14_6_WRONG(AA, BB, CC, DD, N)

3 INTEGER N
4 REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N)

5 INTEGER PRI

6 !$OMP PARALLEL PRIVATE(PRI)

7 !$OMP WORKSHARE
8 AA = BB
9 PRI = 1
10 CC = DD * PRI
11 !$OMP END WORKSHARE
12 !$OMP END PARALLEL

13 END SUBROUTINE A14_6_WRONG

14 Fortran execution rules must be enforced inside a workshare construct. In the

15 following example, the same result is produced in the following program fragment
16 regardless of whether the code is executed sequentially or inside an OpenMP program
17 with multiple threads:

18 Example A.14.7f

19 SUBROUTINE A14_7(AA, BB, CC, N)

20 INTEGER N
21 REAL AA(N), BB(N), CC(N)

22 !$OMP PARALLEL
23 !$OMP WORKSHARE
24 AA(1:50) = BB(11:60)
25 CC(11:20) = AA(1:10)
26 !$OMP END WORKSHARE
27 !$OMP END PARALLEL

28 END SUBROUTINE A14_7

29
Fortran

30

31 A.15 The master Construct

32 The following example demonstrates the master construct (Section 2.8.1 on page 63). In
33 the example, the master keeps track of how many iterations have been executed and
34 prints out a progress report. The other threads skip the master region without waiting.

35 Appendix A Examples 195

1
C/C++
2 Example A.15.1c

3 #include <stdio.h>

4 extern float average(float,float,float);

5 void a15( float* x, float* xold, int n, float tol )

6 {
7 int c, i, toobig;
8 float error, y;
9 c = 0;
10 #pragma omp parallel
11 {
12 do{
13 #pragma omp for private(i)
14 for( i = 1; i < n-1; ++i ){
15 xold[i] = x[i];
16 }
17 #pragma omp single
18 {
19 toobig = 0;
20 }
21 #pragma omp for private(i,y,error) reduction(+:toobig)
22 for( i = 1; i < n-1; ++i ){
23 y = x[i];
24 x[i] = average( xold[i-1], x[i], xold[i+1] );
25 error = y - x[i];
26 if( error > tol || error < -tol ) ++toobig;
27 }
28 #pragma omp master
29 {
30 ++c;
31 printf( "iteration %d, toobig=%d\n", c, toobig );
32 }
33 }while( toobig > 0 );
34 }
35 }
C/C++

36 196 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.15.1f

3 SUBROUTINE A15( X, XOLD, N, TOL )

4 REAL X(*), XOLD(*), TOL
5 INTEGER N
6 INTEGER C, I, TOOBIG
7 REAL ERROR, Y, AVERAGE
8 EXTERNAL AVERAGE
9 C = 0
10 TOOBIG = 1
11 !$OMP PARALLEL
12 DO WHILE( TOOBIG > 0 )
13 !$OMP DO PRIVATE(I)
14 DO I = 2, N-1
15 XOLD(I) = X(I)
16 ENDDO
17 !$OMP SINGLE
18 TOOBIG = 0
19 !$OMP END SINGLE
20 !$OMP DO PRIVATE(I,Y,ERROR), REDUCTION(+:TOOBIG)
21 DO I = 2, N-1
22 Y = X(I)
23 X(I) = AVERAGE( XOLD(I-1), X(I), XOLD(I+1) )
24 ERROR = Y-X(I)
25 IF( ERROR > TOL .OR. ERROR < -TOL ) TOOBIG = TOOBIG+1
26 ENDDO
27 !$OMP MASTER
28 C = C + 1
29 PRINT *, 'Iteration ', C, 'TOOBIG=', TOOBIG
30 !$OMP END MASTER
31 ENDDO
32 !$OMP END PARALLEL
33 END SUBROUTINE A15
34
Fortran

35

36 A.16 The critical Construct

37 The following example includes several critical constructs (Section 2.8.2 on page
38 65). The example illustrates a queuing model in which a task is dequeued and worked
39 on. To guard against multiple threads dequeuing the same task, the dequeuing operation
40 must be in a critical region. Because the two queues in this example are
41 independent, they are protected by critical constructs with different names, xaxis
42 and yaxis.

43 Appendix A Examples 197

1
C/C++
2 Example A.16.1c

3 int dequeue(float *a);

4 void work(int i, float *a);

5 void a16(float *x, float *y)

6 {
7 int ix_next, iy_next;

8 #pragma omp parallel shared(x, y) private(ix_next, iy_next)

9 {
10 #pragma omp critical (xaxis)
11 ix_next = dequeue(x);
12 work(ix_next, x);

13 #pragma omp critical (yaxis)

14 iy_next = dequeue(y);
15 work(iy_next, y);
16 }

17 }
C/C++
18
Fortran
19 Example A.16.1f

20 SUBROUTINE A16(X, Y)

21 REAL X(*), Y(*)

22 INTEGER IX_NEXT, IY_NEXT

23 !$OMP PARALLEL SHARED(X, Y) PRIVATE(IX_NEXT, IY_NEXT)

24 !$OMP CRITICAL(XAXIS)
25 CALL DEQUEUE(IX_NEXT, X)
26 !$OMP END CRITICAL(XAXIS)
27 CALL WORK(IX_NEXT, X)

28 !$OMP CRITICAL(YAXIS)
29 CALL DEQUEUE(IY_NEXT,Y)
30 !$OMP END CRITICAL(YAXIS)
31 CALL WORK(IY_NEXT, Y)

32 !$OMP END PARALLEL

33 END SUBROUTINE A16

34
Fortran

35 198 OpenMP API • Version 3.0 May 2008

1

2 A.17 worksharing Constructs Inside a

3 critical Construct
4 The following example demonstrates using a worksharing construct inside a critical
5 construct (see Section 2.8.2 on page 65). This example is conforming because the
6 single region and the critical region are not closely nested (see Section 2.10 on
7 page 104).
8
C/C++
9 Example A.17.1c

10 void a17()
11 {
12 int i = 1;
13 #pragma omp parallel sections
14 {
15 #pragma omp section
16 {
17 #pragma omp critical (name)
18 {
19 #pragma omp parallel
20 {
21 #pragma omp single
22 {
23 i++;
24 }
25 }
26 }
27 }
28 }
29 }
C/C++

30 Appendix A Examples 199

1
Fortran
2 Example A.17.1f

3 SUBROUTINE A17()

4 INTEGER I
5 I = 1

6 !$OMP PARALLEL SECTIONS

7 !$OMP SECTION
8 !$OMP CRITICAL (NAME)
9 !$OMP PARALLEL
10 !$OMP SINGLE
11 I = I + 1
12 !$OMP END SINGLE
13 !$OMP END PARALLEL
14 !$OMP END CRITICAL (NAME)
15 !$OMP END PARALLEL SECTIONS
16 END SUBROUTINE A17
17
Fortran

18

19 A.18 Binding of barrier Regions

20 The binding rules call for a barrier region to bind to the closest enclosing
21 parallel region (see Section 2.8.3 on page 66).
22 In the following example, the call from the main program to sub2 is conforming because
23 the barrier region (in sub3) binds to the parallel region in sub2. The call from
24 the main program to sub1 is conforming because the barrier region binds to the
25 parallel region in subroutine sub2.
26 The call from the main program to sub3 is conforming because the barrier region
27 binds to the implicit inactive parallel region enclosing the sequential part. Also note
28 that the barrier region in sub3 when called from sub2 only synchronizes the team of
29 threads in the enclosing parallel region and not all the threads created in sub1.

30 200 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.18.1c

3 void work(int n) {}

4 void sub3(int n)
5 {
6 work(n);
7 #pragma omp barrier
8 work(n);
9 }

10 void sub2(int k)
11 {
12 #pragma omp parallel shared(k)
13 sub3(k);
14 }

15 void sub1(int n)
16 {
17 int i;
18 #pragma omp parallel private(i) shared(n)
19 {
20 #pragma omp for
21 for (i=0; i<n; i++)
22 sub2(i);
23 }
24 }

25 int main()
26 {
27 sub1(2);
28 sub2(2);
29 sub3(2);
30 return 0;
31 }
C/C++

32 Appendix A Examples 201

1
Fortran
2 Example A.18.1f

3 SUBROUTINE WORK(N)
4 INTEGER N
5 END SUBROUTINE WORK

6 SUBROUTINE SUB3(N)
7 INTEGER N
8 CALL WORK(N)
9 !$OMP BARRIER
10 CALL WORK(N)
11 END SUBROUTINE SUB3

12 SUBROUTINE SUB2(K)
13 INTEGER K
14 !$OMP PARALLEL SHARED(K)
15 CALL SUB3(K)
16 !$OMP END PARALLEL
17 END SUBROUTINE SUB2

18 SUBROUTINE SUB1(N)
19 INTEGER N
20 INTEGER I
21 !$OMP PARALLEL PRIVATE(I) SHARED(N)
22 !$OMP DO
23 DO I = 1, N
24 CALL SUB2(I)
25 END DO
26 !$OMP END PARALLEL
27 END SUBROUTINE SUB1

28 PROGRAM A18
29 CALL SUB1(2)
30 CALL SUB2(2)
31 CALL SUB3(2)
32 END PROGRAM A18
33
Fortran

34

35 A.19 The atomic Construct

36 The following example avoids race conditions (simultaneous updates of an element of x
37 by multiple threads) by using the atomic construct (Section 2.8.5 on page 69).

38 202 OpenMP API • Version 3.0 May 2008

1 The advantage of using the atomic construct in this example is that it allows updates
2 of two different elements of x to occur in parallel. If a critical construct (see
3 Section 2.8.2 on page 65) were used instead, then all updates to elements of x would be
4 executed serially (though not in any guaranteed order).
5 Note that the atomic directive applies only to the statement immediately following it.
6 As a result, elements of y are not updated atomically in this example.
7
C/C++
8 Example A.19.1c

9 float work1(int i)
10 {
11 return 1.0 * i;
12 }

13 float work2(int i)
14 {
15 return 2.0 * i;
16 }

17 void a19(float *x, float *y, int *index, int n)

18 {
19 int i;

20 #pragma omp parallel for shared(x, y, index, n)

21 for (i=0; i<n; i++) {
22 #pragma omp atomic
23 x[index[i]] += work1(i);
24 y[i] += work2(i);
25 }
26 }

27 int main()
28 {
29 float x[1000];
30 float y[10000];
31 int index[10000];
32 int i;

33 for (i = 0; i < 10000; i++) {

34 index[i] = i % 1000;
35 y[i]=0.0;
36 }
37 for (i = 0; i < 1000; i++)
38 x[i] = 0.0;
39 a19(x, y, index, 10000);
40 return 0;
41 }
C/C++

42 Appendix A Examples 203

1
Fortran
2 Example A.19.1f

3 REAL FUNCTION WORK1(I)

4 INTEGER I
5 WORK1 = 1.0 * I
6 RETURN
7 END FUNCTION WORK1

8 REAL FUNCTION WORK2(I)

9 INTEGER I
10 WORK2 = 2.0 * I
11 RETURN
12 END FUNCTION WORK2

13 SUBROUTINE SUBA19(X, Y, INDEX, N)

14 REAL X(*), Y(*)
15 INTEGER INDEX(*), N

16 INTEGER I

17 !$OMP PARALLEL DO SHARED(X, Y, INDEX, N)

18 DO I=1,N
19 !$OMP ATOMIC
20 X(INDEX(I)) = X(INDEX(I)) + WORK1(I)
21 Y(I) = Y(I) + WORK2(I)
22 ENDDO

23 END SUBROUTINE SUBA19

24 PROGRAM A19
25 REAL X(1000), Y(10000)
26 INTEGER INDEX(10000)
27 INTEGER I
28
29 DO I=1,10000
30 INDEX(I) = MOD(I, 1000) + 1
31 Y(I) = 0.0
32 ENDDO
33
34 DO I = 1,1000
35 X(I) = 0.0
36 ENDDO

37 CALL SUBA19(X, Y, INDEX, 10000)

38
39 END PROGRAM A19
40
Fortran

41 204 OpenMP API • Version 3.0 May 2008

1

2 A.20 Restrictions on the atomic Construct

3 The following examples illustrate the restrictions on the atomic construct. For more
4 information, see Section 2.8.5 on page 69.
5
C/C++
6 All atomic references to the storage location of each variable that appears on the left-
7 hand side of an atomic assignment statement throughout the program are required to
8 have a compatible type.
C/C++
9
Fortran
10 All atomic references to the storage location of each variable that appears on the left-
11 hand side of an atomic assignment statement throughout the program are required to
12 have the same type and type parameters.
13
Fortran

14 The following are some non-conforming examples:

15
C/C++
16 Example A.20.1c

17 void a20_1_wrong ()
18 {
19 union {int n; float x;} u;

20 #pragma omp parallel

21 {
22 #pragma omp atomic
23 u.n++;

24 #pragma omp atomic

25 u.x += 1.0;

26 /* Incorrect because the atomic constructs reference the same location

27 through incompatible types */
28 }
29 }
C/C++

30 Appendix A Examples 205

1
Fortran
2 Example A.20.1f

3 SUBROUTINE A20_1_WRONG()
4 INTEGER:: I
5 REAL:: R
6 EQUIVALENCE(I,R)

7 !$OMP PARALLEL
8 !$OMP ATOMIC
9 I = I + 1
10 !$OMP ATOMIC
11 R = R + 1.0
12 ! incorrect because I and R reference the same location
13 ! but have different types
14 !$OMP END PARALLEL
15 END SUBROUTINE A20_1_WRONG
16
Fortran
17
C/C++
18 Example A.20.2c

19 void a20_2_wrong ()
20 {
21 int x;
22 int *i;
23 float *r;

24 i = &x;
25 r = (float *)&x;

26 #pragma omp parallel

27 {
28 #pragma omp atomic
29 *i += 1;

30 #pragma omp atomic

31 *r += 1.0;

32 /* Incorrect because the atomic constructs reference the same location

33 through incompatible types */

34 }
35 }
C/C++

36 206 OpenMP API • Version 3.0 May 2008

1
Fortran
2 The following example is non-conforming because I and R reference the same location
3 but have different types.

4 Example A.20.2f

5 SUBROUTINE SUB()
6 COMMON /BLK/ R
7 REAL R

8 !$OMP ATOMIC
9 R = R + 1.0
10 END SUBROUTINE SUB

11 SUBROUTINE A20_2_WRONG()
12 COMMON /BLK/ I
13 INTEGER I

14 !$OMP PARALLEL

15 !$OMP ATOMIC
16 I = I + 1
17 CALL SUB()
18 !$OMP END PARALLEL
19 END SUBROUTINE A20_2_WRONG

20 Appendix A Examples 207

1 Although the following example might work on some implementations, this is also non-
2 conforming:

3 Example A.20.3f

4 SUBROUTINE A20_3_WRONG
5 INTEGER:: I
6 REAL:: R
7 EQUIVALENCE(I,R)

8 !$OMP PARALLEL
9 !$OMP ATOMIC
10 I = I + 1
11 ! incorrect because I and R reference the same location
12 ! but have different types
13 !$OMP END PARALLEL

14 !$OMP PARALLEL
15 !$OMP ATOMIC
16 R = R + 1.0
17 ! incorrect because I and R reference the same location
18 ! but have different types
19 !$OMP END PARALLEL

20 END SUBROUTINE A20_3_WRONG

21
Fortran

22

23 A.21 The flush Construct with a List

24 The following example uses the flush construct (see Section 2.8.6 on page 72) for
25 point-to-point synchronization of specific variables between pairs of threads:
26
C/C++
27 Example A.21.1c

28 #include <omp.h>
29 #define NUMBER_OF_THREADS 256

30 int synch[NUMBER_OF_THREADS];
31 float work[NUMBER_OF_THREADS];
32 float result[NUMBER_OF_THREADS];

33 float fn1(int i)
34 {
35 return i*2.0;

36 208 OpenMP API • Version 3.0 May 2008

1 }

2 float fn2(float a, float b)

3 {
4 return a + b;
5 }

6 int main()
7 {
8 int iam, neighbor;

9 #pragma omp parallel private(iam,neighbor) shared(work,synch)

10 {
11 iam = omp_get_thread_num();
12 synch[iam] = 0;

13 #pragma omp barrier

14 /*Do computation into my portion of work array */
15 work[iam] = fn1(iam);

16 /* Announce that I am done with my work. The first flush

17 * ensures that my work is made visible before synch.
18 * The second flush ensures that synch is made visible.
19 */

20 #pragma omp flush(work,synch)

21 synch[iam] = 1;
22 #pragma omp flush(synch)

23 /* Wait for neighbor. The first flush ensures that synch is read
24 * from memory, rather than from the temporary view of memory.
25 * The second flush ensures that work is read from memory, and
26 * is done so after the while loop exits.
27 */

28 neighbor = (iam>0 ? iam : omp_get_num_threads()) - 1;

29 while (synch[neighbor] == 0) {
30 #pragma omp flush(synch)
31 }
32
33 #pragma omp flush(work,synch)
34
35 /* Read neighbor’s values of work array */
36 result[iam] = fn2(work[neighbor], work[iam]);
37 }

38 /* output result here */

39 return 0;
40 }
C/C++

41 Appendix A Examples 209

1
Fortran
2 Example A.21.1f
3 REAL FUNCTION FN1(I)
4 INTEGER I
5 FN1 = I * 2.0
6 RETURN
7 END FUNCTION FN1
8 REAL FUNCTION FN2(A, B)
9 REAL A, B
10 FN2 = A + B
11 RETURN
12 END FUNCTION FN2
13
14 PROGRAM A21
15 INCLUDE "omp_lib.h" ! or USE OMP_LIB
16 INTEGER ISYNC(256)
17 REAL WORK(256)
18 REAL RESULT(256)
19 INTEGER IAM, NEIGHBOR

20 !$OMP PARALLEL PRIVATE(IAM, NEIGHBOR) SHARED(WORK, ISYNC)

21 IAM = OMP_GET_THREAD_NUM() + 1
22 ISYNC(IAM) = 0
23 !$OMP BARRIER
24 C Do computation into my portion of work array
25 WORK(IAM) = FN1(IAM)
26 C Announce that I am done with my work.
27 C The first flush ensures that my work is made visible before
28 C synch. The second flush ensures that synch is made visible.
29 !$OMP FLUSH(WORK,ISYNC)
30 ISYNC(IAM) = 1
31 !$OMP FLUSH(ISYNC)
32 C Wait until neighbor is done. The first flush ensures that
33 C synch is read from memory, rather than from the temporary
34 C view of memory. The second flush ensures that work is read
35 C from memory, and is done so after the while loop exits.
36 IF (IAM .EQ. 1) THEN
37 NEIGHBOR = OMP_GET_NUM_THREADS()
38 ELSE
39 NEIGHBOR = IAM - 1
40 ENDIF

41 DO WHILE (ISYNC(NEIGHBOR) .EQ. 0)

42 !$OMP FLUSH(ISYNC)
43 END DO
44 !$OMP FLUSH(WORK, ISYNC)

45 RESULT(IAM) = FN2(WORK(NEIGHBOR), WORK(IAM))

46 !$OMP END PARALLEL
47 END PROGRAM A21
48
Fortran

49 210 OpenMP API • Version 3.0 May 2008

1

2 A.22 The flush Construct without a List

3 The following example (for Section 2.8.6 on page 72) distinguishes the shared variables
4 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 void f1(int *q)

9 {
10 *q = 1;
11 #pragma omp flush
12 /* x, p, and *q are flushed */
13 /* because they are shared and accessible */
14 /* q is not flushed because it is not shared. */
15 }

16 void f2(int *q)

17 {
18 #pragma omp barrier
19 *q = 2;
20 #pragma omp barrier

21 /* a barrier implies a flush */

22 /* x, p, and *q are flushed */
23 /* because they are shared and accessible */
24 /* q is not flushed because it is not shared. */
25 }

26 int g(int n)
27 {
28 int i = 1, j, sum = 0;
29 *p = 1;
30 #pragma omp parallel reduction(+: sum) num_threads(10)
31 {
32 f1(&j);

33 /* i, n and sum were not flushed */

34 /* because they were not accessible in f1 */
35 /* j was flushed because it was accessible */
36 sum += j;

37 f2(&j);

38 /* i, n, and sum were not flushed */

39 /* because they were not accessible in f2 */
40 /* j was flushed because it was accessible */

41 Appendix A Examples 211

1 sum += i + j + *p + n;
2 }
3 return sum;
4 }

5 int main()
6 {
7 int result = g(7);
8 return result;
9 }
C/C++
10
Fortran
11 Example A.22.1f

12 SUBROUTINE F1(Q)
13 COMMON /DATA/ X, P
14 INTEGER, TARGET :: X
15 INTEGER, POINTER :: P
16 INTEGER Q

17 Q = 1
18 !$OMP FLUSH
19 ! X, P and Q are flushed
20 ! because they are shared and accessible
21 END SUBROUTINE F1

22 SUBROUTINE F2(Q)
23 COMMON /DATA/ X, P
24 INTEGER, TARGET :: X
25 INTEGER, POINTER :: P
26 INTEGER Q

27 !$OMP BARRIER
28 Q = 2
29 !$OMP BARRIER
30 ! a barrier implies a flush
31 ! X, P and Q are flushed
32 ! because they are shared and accessible
33 END SUBROUTINE F2

34 INTEGER FUNCTION G(N)

35 COMMON /DATA/ X, P
36 INTEGER, TARGET :: X
37 INTEGER, POINTER :: P
38 INTEGER N
39 INTEGER I, J, SUM

40 I = 1
41 SUM = 0
42 P = 1

43 212 OpenMP API • Version 3.0 May 2008

1 !$OMP PARALLEL REDUCTION(+: SUM) NUM_THREADS(10)
2 CALL F1(J)
3 ! I, N and SUM were not flushed
4 ! because they were not accessible in F1
5 ! J was flushed because it was accessible
6 SUM = SUM + J

7 CALL F2(J)
8 ! I, N, and SUM were not flushed
9 ! because they were not accessible in f2
10 ! J was flushed because it was accessible
11 SUM = SUM + I + J + P + N
12 !$OMP END PARALLEL

13 G = SUM
14 END FUNCTION G

15 PROGRAM A22
16 COMMON /DATA/ X, P
17 INTEGER, TARGET :: X
18 INTEGER, POINTER :: P
19 INTEGER RESULT, G

20 P => X
21 RESULT = G(7)
22 PRINT *, RESULT
23 END PROGRAM A22
24
Fortran

25 Appendix A Examples 213

1
C/C++

2 A.23 Placement of flush, barrier, and

3 taskwait Directives
4 The following example is non-conforming, because the flush, barrier, and
5 taskwait directives cannot be the immediate substatement of an if statement. See
6 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 void a23_wrong()
9 {
10 int a = 1;

11 #pragma omp parallel

12 {
13 if (a != 0)
14 #pragma omp flush(a)
15 /* incorrect as flush cannot be immediate substatement
16 of if statement */

17 if (a != 0)
18 #pragma omp barrier
19 /* incorrect as barrier cannot be immediate substatement
20 of if statement */

21 if (a != 0)
22 #pragma omp taskwait
23 /* incorrect as taskwait cannot be immediate substatement
24 of if statement */

25 }
26 }

27 The following version of the above example is conforming because the flush,
28 barrier, and taskwait directives are enclosed in a compound statement.

29 214 OpenMP API • Version 3.0 May 2008

1 Example A.23.2c

2 void a23()
3 {
4 int a = 1;

5 #pragma omp parallel

6 {
7 if (a != 0) {
8 #pragma omp flush(a)
9 }
10 if (a != 0) {
11 #pragma omp barrier
12 }
13 if (a != 0) {
14 #pragma omp taskwait
15 }
16 }
17 }
C/C++

18

19 A.24 The ordered Clause and the ordered

20 Construct
21 Ordered constructs (Section 2.8.7 on page 75) are useful for sequentially ordering the
22 output from work that is done in parallel. The following program prints out the indices
23 in sequential order:

24 Appendix A Examples 215

1
C/C++
2 Example A.24.1c

3 #include <stdio.h>

4 void work(int k)
5 {
6 #pragma omp ordered
7 printf(" %d\n", k);
8 }

9 void a24(int lb, int ub, int stride)

10 {
11 int i;

12 #pragma omp parallel for ordered schedule(dynamic)

13 for (i=lb; i<ub; i+=stride)
14 work(i);
15 }

16 int main()
17 {
18 a24(0, 100, 5);
19 return 0;
20 }
C/C++

21 216 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.24.1f

3 SUBROUTINE WORK(K)
4 INTEGER k

5 !$OMP ORDERED
6 WRITE(*,*) K
7 !$OMP END ORDERED

8 END SUBROUTINE WORK

9 SUBROUTINE SUBA24(LB, UB, STRIDE)

10 INTEGER LB, UB, STRIDE
11 INTEGER I

12 !$OMP PARALLEL DO ORDERED SCHEDULE(DYNAMIC)

13 DO I=LB,UB,STRIDE
14 CALL WORK(I)
15 END DO
16 !$OMP END PARALLEL DO

17 END SUBROUTINE SUBA24

18 PROGRAM A24
19 CALL SUBA24(1,100,5)
20 END PROGRAM A24
21
Fortran
22 It is possible to have multiple ordered constructs within a loop region with the
23 ordered clause specified. The first example is non-conforming because all iterations
24 execute two ordered regions. An iteration of a loop must not execute more than one
25 ordered region:

26 Appendix A Examples 217

1
C/C++
2 Example A.24.2c

3 void work(int i) {}

4 void a24_wrong(int n)
5 {
6 int i;
7 #pragma omp for ordered
8 for (i=0; i<n; i++) {
9 /* incorrect because an iteration may not execute more than one
10 ordered region */
11 #pragma omp ordered
12 work(i);
13 #pragma omp ordered
14 work(i+1);
15 }
16 }
C/C++
17
Fortran
18 Example A.24.2f

19 SUBROUTINE WORK(I)
20 INTEGER I
21 END SUBROUTINE WORK

22 SUBROUTINE A24_WRONG(N)
23 INTEGER N

24 INTEGER I
25 !$OMP DO ORDERED
26 DO I = 1, N
27 ! incorrect because an iteration may not execute more than one
28 ! ordered region
29 !$OMP ORDERED
30 CALL WORK(I)
31 !$OMP END ORDERED

32 !$OMP ORDERED
33 CALL WORK(I+1)
34 !$OMP END ORDERED
35 END DO
36 END SUBROUTINE A24_WRONG
37
Fortran

38 218 OpenMP API • Version 3.0 May 2008

1 The following is a conforming example with more than one ordered construct. Each
2 iteration will execute only one ordered region:
3
C/C++
4 Example A.24.3c

5 void work(int i) {}
6 void a24_good(int n)
7 {
8 int i;

9 #pragma omp for ordered

10 for (i=0; i<n; i++) {
11 if (i <= 10) {
12 #pragma omp ordered
13 work(i);
14 }

15 if (i > 10) {
16 #pragma omp ordered
17 work(i+1);
18 }
19 }
20 }
C/C++
21
Fortran
22 Example A.24.3f

23 SUBROUTINE A24_GOOD(N)
24 INTEGER N

25 !$OMP DO ORDERED
26 DO I = 1,N
27 IF (I <= 10) THEN
28 !$OMP ORDERED
29 CALL WORK(I)
30 !$OMP END ORDERED
31 ENDIF

32 IF (I > 10) THEN

33 !$OMP ORDERED
34 CALL WORK(I+1)
35 !$OMP END ORDERED
36 ENDIF
37 ENDDO
38 END SUBROUTINE A24_GOOD
39
Fortran

40 Appendix A Examples 219

1

2 A.25 The threadprivate Directive

3 The following examples demonstrate how to use the threadprivate directive
4 (Section 2.9.2 on page 81) to give each thread a separate counter.
5
C/C++
6 Example A.25.1c

7 int counter = 0;
8 #pragma omp threadprivate(counter)

9 int increment_counter()
10 {
11 counter++;
12 return(counter);
13 }
C/C++
14
Fortran
15 Example A.25.1f

16 INTEGER FUNCTION INCREMENT_COUNTER()

17 COMMON/A25_COMMON/COUNTER
18 !$OMP THREADPRIVATE(/A25_COMMON/)

19 COUNTER = COUNTER +1
20 INCREMENT_COUNTER = COUNTER
21 RETURN
22 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 int increment_counter_2()
28 {
29 static int counter = 0;
30 #pragma omp threadprivate(counter)
31 counter++;
32 return(counter);
33 }

34 220 OpenMP API • Version 3.0 May 2008

1 The following example illustrates how modifying a variable that appears in an initializer
2 can cause unspecified behavior, and also how to avoid this problem by using an
3 auxiliary variable and a copy-constructor.

4 Example A.25.3c

5 class T {
6 public:
7 int val;
8 T (int);
9 T (const T&);
10 };

11 T :: T (int v){
12 val = v;
13 }

14 T :: T (const T& t) {
15 val = t.val;
16 }

17 void g(T a, T b){

18 a.val += b.val;
19 }

20 int x = 1;
21 T a(x);
22 const T b_aux(x); /* Capture value of x = 1 */
23 T b(b_aux);
24 #pragma omp threadprivate(a, b)

25 void f(int n) {
26 x++;
27 #pragma omp parallel for
28 /* In each thread:
29 * a is constructed from x (with value 1 or 2?)
30 * b is copy-constructed from b_aux
31 */

32 for (int i=0; i<n; i++) {

33 g(a, b); /* Value of a is unspecified. */
34 }
35 }
C/C++

36 Appendix A Examples 221

1
Fortran
2 The following examples show non-conforming uses and correct uses of the
3 threadprivate directive. For more information, see Section 2.9.2 on page 81 and
4 Section 2.9.4.1 on page 101.
5 The following example is non-conforming because the common block is not declared
6 local to the subroutine that refers to it:

7 Example A.25.2f

8 MODULE A25_MODULE
9 COMMON /T/ A
10 END MODULE A25_MODULE
11
12 SUBROUTINE A25_2_WRONG()
13 USE A25_MODULE
14 !$OMP THREADPRIVATE(/T/)
15 !non-conforming because /T/ not declared in A25_4_WRONG
16 END SUBROUTINE A25_2_WRONG

17 The following example is also non-conforming because the common block is not
18 declared local to the subroutine that refers to it:

19 Example A.25.3f

20 SUBROUTINE A25_3_WRONG()
21 COMMON /T/ A
22 !$OMP THREADPRIVATE(/T/)

23 CONTAINS
24 SUBROUTINE A25_3S_WRONG()
25 !$OMP PARALLEL COPYIN(/T/)
26 !non-conforming because /T/ not declared in A35_5S_WRONG
27 !$OMP END PARALLEL
28 END SUBROUTINE A25_3S_WRONG
29 END SUBROUTINE A25_3_WRONG

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 SUBROUTINE A25_4_GOOD()
5 COMMON /T/ A
6 !$OMP THREADPRIVATE(/T/)

7 CONTAINS
8 SUBROUTINE A25_4S_GOOD()
9 COMMON /T/ A
10 !$OMP THREADPRIVATE(/T/)

11 !$OMP PARALLEL COPYIN(/T/)

12 !$OMP END PARALLEL
13 END SUBROUTINE A25_4S_GOOD
14 END SUBROUTINE A25_4_GOOD

15 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 PROGRAM A25_5_GOOD
5 INTEGER, ALLOCATABLE, SAVE :: A(:)
6 INTEGER, POINTER, SAVE :: PTR
7 INTEGER, SAVE :: I
8 INTEGER, TARGET :: TARG
9 LOGICAL :: FIRSTIN = .TRUE.
10 !$OMP THREADPRIVATE(A, I, PTR)
11
12 ALLOCATE (A(3))
13 A = (/1,2,3/)
14 PTR => TARG
15 I = 5
16
17 !$OMP PARALLEL COPYIN(I, PTR)
18 !$OMP CRITICAL
19 IF (FIRSTIN) THEN
20 TARG = 4 ! Update target of ptr
21 I = I + 10
22 IF (ALLOCATED(A)) A = A + 10
23 FIRSTIN = .FALSE.
24 END IF
25
26 IF (ALLOCATED(A)) THEN
27 PRINT *, 'a = ', A
28 ELSE
29 PRINT *, 'A is not allocated'
30 END IF

31 PRINT *, 'ptr = ', PTR

32 PRINT *, 'i = ', I
33 PRINT *

34 !$OMP END CRITICAL

35 !$OMP END PARALLEL
36 END PROGRAM A25_5_GOOD

37 The above program, if executed by two threads, will print one of the following two sets
38 of output:

39 a = 11 12 13
40 ptr = 4
41 i = 15

42 A is not allocated

43 224 OpenMP API • Version 3.0 May 2008

1 ptr = 4
2 i = 5

3 or

4 A is not allocated
5 ptr = 4
6 i = 15

7 a = 1 2 3
8 ptr = 4
9 i = 5

10 The following is an example of the use of threadprivate for module variables:

11 Example A.25.6f

12 MODULE A25_MODULE6
13 REAL, POINTER :: WORK(:)
14 SAVE WORK
15 !$OMP THREADPRIVATE(WORK)
16 END MODULE A25_MODULE6
17
18 SUBROUTINE SUB1(N)
19 USE A25_MODULE6
20 !$OMP PARALLEL PRIVATE(THE_SUM)
21 ALLOCATE(WORK(N))
22 CALL SUB2(THE_SUM)
23 WRITE(*,*)THE_SUM
24 !$OMP END PARALLEL
25 END SUBROUTINE SUB1
26
27 SUBROUTINE SUB2(THE_SUM)
28 USE A25_MODULE6
29 WORK(:) = 10
30 THE_SUM=SUM(WORK)
31 END SUBROUTINE SUB2
32
33 PROGRAM A25_6_GOOD
34 N = 10
35 CALL SUB1(N)
36 END PROGRAM A25_6_GOOD
37
Fortran
38
C/C++
39 The following example illustrates initialization of threadprivate variables for class-type
40 T. t1 is default constructed, t2 is constructed taking a constructor accepting one
41 argument of integer type, t3 is copy constructed with argument f():

42 Appendix A Examples 225

1 Example A.25.4c

2 static T t1;
3 #pragma omp threadprivate(t1)
4 static T t2( 23 );
5 #pragma omp threadprivate(t2)
6 static T t3 = f();
7 #pragma omp threadprivate(t3)

8 The following example illustrates the use of threadprivate for static class members. The
9 threadprivate directive for a static class member must be placed inside the class
10 definition.

11 Example A.25.5c

12 class T {
13 public:
14 static int i;
15 #pragma omp threadprivate(i)
16 };
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 #include <vector>
24 void iterator_example()
25 {
26 std::vector<int> vec(23);
27 std::vector<int>::iterator it;
28 #pragma omp parallel for default(none) shared(vec)
29 for (it = vec.begin(); it < vec.end(); it++)
30 {
31 // do work with *it //
32 }
33 }
C/C++

34 226 OpenMP API • Version 3.0 May 2008

1
Fortran
2

3 A.27 Fortran Restrictions on shared and

4 private Clauses with Common Blocks
5 When a named common block is specified in a private, firstprivate, or
6 lastprivate clause of a construct, none of its members may be declared in another
7 data-sharing attribute clause on that construct. The following examples illustrate this
8 point. For more information, see Section 2.9.3 on page 85.
9 The following example is conforming:

10 Example A.27.1f

11 SUBROUTINE A27_1_GOOD()
12 COMMON /C/ X,Y
13 REAL X, Y

14 !$OMP PARALLEL PRIVATE (/C/)

15 ! do work here
16 !$OMP END PARALLEL

17 !$OMP PARALLEL SHARED (X,Y)

18 ! do work here
19 !$OMP END PARALLEL
20 END SUBROUTINE A27_1_GOOD

21 Appendix A Examples 227

1
Fortran (cont.)
2 The following example is also conforming:

3 Example A.27.2f

4 SUBROUTINE A27_2_GOOD()
5 COMMON /C/ X,Y
6 REAL X, Y

7 INTEGER I

8 !$OMP PARALLEL
9 !$OMP DO PRIVATE(/C/)
10 DO I=1,1000
11 ! do work here
12 ENDDO
13 !$OMP END DO
14 !
15 !$OMP DO PRIVATE(X)
16 DO I=1,1000
17 ! do work here
18 ENDDO
19 !$OMP END DO
20 !$OMP END PARALLEL
21 END SUBROUTINE A27_2_GOOD

22 The following example is conforming:

23 Example A.27.3f

24 SUBROUTINE A27_3_GOOD()
25 COMMON /C/ X,Y

26 !$OMP PARALLEL PRIVATE (/C/)

27 ! do work here
28 !$OMP END PARALLEL

29 !$OMP PARALLEL SHARED (/C/)

30 ! do work here
31 !$OMP END PARALLEL
32 END SUBROUTINE A27_3_GOOD

33 228 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 SUBROUTINE A27_4_WRONG()
4 COMMON /C/ X,Y
5 ! Incorrect because X is a constituent element of C
6 !$OMP PARALLEL PRIVATE(/C/), SHARED(X)
7 ! do work here
8 !$OMP END PARALLEL
9 END SUBROUTINE A27_4_WRONG

10 The following example is non-conforming because a common block may not be

11 declared both shared and private:

12 Example A.27.5f

13 SUBROUTINE A27_5_WRONG()
14 COMMON /C/ X,Y
15 ! Incorrect: common block C cannot be declared both
16 ! shared and private
17 !$OMP PARALLEL PRIVATE (/C/), SHARED(/C/)
18 ! do work here
19 !$OMP END PARALLEL

20 END SUBROUTINE A27_5_WRONG

21
Fortran

22

23 A.28 The default(none) Clause

24 The following example distinguishes the variables that are affected by the
25 default(none) clause from those that are not. For more information on the
26 default clause, see Section 2.9.3.1 on page 86.

27 Appendix A Examples 229

1
C/C++
2 Example A.28.1c

3 #include <omp.h>
4 int x, y, z[1000];
5 #pragma omp threadprivate(x)

6 void a28(int a) {
7 const int c = 1;
8 int i = 0;

9 #pragma omp parallel default(none) private(a) shared(z)

10 {
11 int j = omp_get_num_threads();
12 /* O.K. - j is declared within parallel region */
13 a = z[j]; /* O.K. - a is listed in private clause */
14 /* - z is listed in shared clause */
15 x = c; /* O.K. - x is threadprivate */
16 /* - c has const-qualified type */
17 z[i] = y; /* Error - cannot reference i or y here */

18 #pragma omp for firstprivate(y)

19 for (i=0; i<10 ; i++) {
20 z[i] = y; /* O.K. - i is the loop iteration variable */
21 /* - y is listed in firstprivate clause */
22 }

23 z[i] = y; /* Error - cannot reference i or y here */

24 }
25 }
C/C++

26 230 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.28.1f

3 SUBROUTINE A28(A)
4 INCLUDE "omp_lib.h" ! or USE OMP_LIB
5
6 INTEGER A

7 INTEGER X, Y, Z(1000)
8 COMMON/BLOCKX/X
9 COMMON/BLOCKY/Y
10 COMMON/BLOCKZ/Z
11 !$OMP THREADPRIVATE(/BLOCKX/)

12 INTEGER I, J
13 i = 1

14 !$OMP PARALLEL DEFAULT(NONE) PRIVATE(A) SHARED(Z) PRIVATE(J)

15 J = OMP_GET_NUM_THREADS();
16 ! O.K. - J is listed in PRIVATE clause
17 A = Z(J) ! O.K. - A is listed in PRIVATE clause
18 ! - Z is listed in SHARED clause
19 X = 1 ! O.K. - X is THREADPRIVATE
20 Z(I) = Y ! Error - cannot reference I or Y here

21 !$OMP DO firstprivate(y)
22 DO I = 1,10
23 Z(I) = Y ! O.K. - I is the loop iteration variable
24 ! Y is listed in FIRSTPRIVATE clause
25 END DO

26 Z(I) = Y ! Error - cannot reference I or Y here

27 !$OMP END PARALLEL
28 END SUBROUTINE A28
29
Fortran

30
Fortran
31

32 A.29 Race Conditions Caused by Implied

33 Copies of Shared Variables in Fortran
34 The following example contains a race condition, because the shared variable, which is
35 an array section, is passed as an actual argument to a routine that has an assumed-size
36 array as its dummy argument (see Section 2.9.3.2 on page 88). The subroutine call
37 passing an array section argument may cause the compiler to copy the argument into a

38 Appendix A Examples 231

1 temporary location prior to the call and copy from the temporary location into the
2 original variable when the subroutine returns. This copying would cause races in the
3 parallel region.

4 Example A.29.1f

5 SUBROUTINE A29
6
7 INCLUDE "omp_lib.h" ! or USE OMP_LIB
8
9 REAL A(20)
10 INTEGER MYTHREAD

11 !$OMP PARALLEL SHARED(A) PRIVATE(MYTHREAD)

12 MYTHREAD = OMP_GET_THREAD_NUM()
13 IF (MYTHREAD .EQ. 0) THEN
14 CALL SUB(A(1:10)) ! compiler may introduce writes to A(6:10)
15 ELSE
16 A(6:10) = 12
17 ENDIF

18 !$OMP END PARALLEL

19 END SUBROUTINE A29

20 SUBROUTINE SUB(X)
21 REAL X(*)
22 X(1:5) = 4
23 END SUBROUTINE SUB
24
Fortran

25

26 A.30 The private Clause

27 In the following example, the values of original list items i and j are retained on exit
28 from the parallel region, while the private list items i and j are modified within the
29 parallel construct. For more information on the private clause, see
30 Section 2.9.3.3 on page 89.

31 232 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.30.1c

3 #include <stdio.h>
4 #include <assert.h>

5 int main()
6 {
7 int i, j;
8 int *ptr_i, *ptr_j;

9 i = 1;
10 j = 2;

11 ptr_i = &i;
12 ptr_j = &j;

13 #pragma omp parallel private(i) firstprivate(j)

14 {
15 i = 3;
16 j = j + 2;
17 assert (*ptr_i == 1 && *ptr_j == 2);
18 }

19 assert(i == 1 && j == 2);

20 return 0;
21 }
C/C++
22
Fortran
23 Example A.30.1f

24 PROGRAM A30
25 INTEGER I, J

26 I = 1
27 J = 2

28 !$OMP PARALLEL PRIVATE(I) FIRSTPRIVATE(J)

29 I = 3
30 J = J + 2
31 !$OMP END PARALLEL

32 PRINT *, I, J ! I .eq. 1 .and. J .eq. 2

33 END PROGRAM A30
34
Fortran

35 Appendix A Examples 233

1 In the following example, all uses of the variable a within the loop construct in the
2 routine f refer to a private list item a, while it is unspecified whether references to a in
3 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 void g(int k) {
8 a = k; /* Accessed in the region but outside of the construct;
9 * therefore unspecified whether original or private list
10 * item is modified. */
11 }

12 void f(int n) {
13 int a = 0;
14
15 #pragma omp parallel for private(a)
16 for (int i=1; i<n; i++) {
17 a = i;
18 g(a*2); /* Private copy of "a" */
19 }
20 }
C/C++

21 234 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.30.2f

3 MODULE A30_2
4 REAL A

5 CONTAINS

6 SUBROUTINE G(K)
7 REAL K
8 A = K ! Accessed in the region but outside of the
9 ! construct; therefore unspecified whether
10 ! original or private list item is modified.
11 END SUBROUTINE G

12 SUBROUTINE F(N)
13 INTEGER N
14 REAL A
15
16 INTEGER I
17 !$OMP PARALLEL DO PRIVATE(A)
18 DO I = 1,N
19 A = I
20 CALL G(A*2)
21 ENDDO
22 !$OMP END PARALLEL DO
23 END SUBROUTINE F

24 END MODULE A30_2

25
Fortran

26

27 A.31 Reprivatization
28 The following example demonstrates the reprivatization of variables (see Section 2.9.3.3
29 on page 89). Private variables can be marked private again in a nested construct.
30 They do not have to be shared in the enclosing parallel region.

31 Appendix A Examples 235

1
C/C++
2 Example A.31.1c

3 #include <assert.h>
4 void a31()
5 {
6 int i, a;

7 #pragma omp parallel private(a)

8 {
9 a = 1;
10 #pragma omp parallel for private(a)
11 for (i=0; i<10; i++)
12 {
13 a = 2;
14 }
15 assert(a == 1);
16 }
17 }
C/C++
18
Fortran
19 Example A.31.1f

20 SUBROUTINE A31()
21 INTEGER I, A

22 !$OMP PARALLEL PRIVATE(A)

23 A = 1
24 !$OMP PARALLEL DO PRIVATE(A)
25 DO I = 1, 10
26 A = 2
27 END DO
28 !$OMP END PARALLEL DO
29 PRINT *, A ! Outer A still has value 1
30 !$OMP END PARALLEL
31 END SUBROUTINE A31
32
Fortran

33 236 OpenMP API • Version 3.0 May 2008

1
Fortran
2

3 A.32 Fortran Restrictions on Storage

4 Association with the private Clause
5 The following non-conforming examples illustrate the implications of the private
6 clause rules with regard to storage association (see Section 2.9.3.3 on page 89).

7 Example A.32.1f

8 SUBROUTINE SUB()
9 COMMON /BLOCK/ X
10 PRINT *,X ! X is undefined
11 END SUBROUTINE SUB

12 PROGRAM A32_1
13 COMMON /BLOCK/ X
14 X = 1.0
15 !$OMP PARALLEL PRIVATE (X)
16 X = 2.0
17 CALL SUB()
18 !$OMP END PARALLEL
19 END PROGRAM A32_1

20 Example A.32.2f

21 PROGRAM A32_2
22 COMMON /BLOCK2/ X
23 X = 1.0

24 !$OMP PARALLEL PRIVATE (X)

25 X = 2.0
26 CALL SUB()
27 !$OMP END PARALLEL

28 CONTAINS

29 SUBROUTINE SUB()
30 COMMON /BLOCK2/ Y

31 PRINT *,X ! X is undefined

32 PRINT *,Y ! Y is undefined
33 END SUBROUTINE SUB

34 END PROGRAM A32_2

35 Appendix A Examples 237

1

2
Fortran (cont.)
3 Example A.32.3f

4 PROGRAM A32_3
5 EQUIVALENCE (X,Y)
6 X = 1.0

7 !$OMP PARALLEL PRIVATE(X)

8 PRINT *,Y ! Y is undefined
9 Y = 10
10 PRINT *,X ! X is undefined
11 !$OMP END PARALLEL
12 END PROGRAM A32_3

13 Example A.32.4f

14 PROGRAM A32_4
15 INTEGER I, J
16 INTEGER A(100), B(100)
17 EQUIVALENCE (A(51), B(1))

18 !$OMP PARALLEL DO DEFAULT(PRIVATE) PRIVATE(I,J) LASTPRIVATE(A)

19 DO I=1,100
20 DO J=1,100
21 B(J) = J - 1
22 ENDDO

23 DO J=1,100
24 A(J) = J ! B becomes undefined at this point
25 ENDDO

26 DO J=1,50
27 B(J) = B(J) + 1 ! B is undefined
28 ! A becomes undefined at this point
29 ENDDO
30 ENDDO
31 !$OMP END PARALLEL DO ! The LASTPRIVATE write for A has
32 ! undefined results

33 PRINT *, B ! B is undefined since the LASTPRIVATE

34 ! write of A was not defined
35 END PROGRAM A32_4

36 238 OpenMP API • Version 3.0 May 2008

1 Example A.32.5f

2 SUBROUTINE SUB1(X)
3 DIMENSION X(10)
4
5 ! This use of X does not conform to the
6 ! specification. It would be legal Fortran 90,
7 ! but the OpenMP private directive allows the
8 ! compiler to break the sequence association that
9 ! A had with the rest of the common block.
10
11 FORALL (I = 1:10) X(I) = I
12 END SUBROUTINE SUB1

13 PROGRAM A32_5
14 COMMON /BLOCK5/ A

15 DIMENSION B(10)
16 EQUIVALENCE (A,B(1))

17 ! the common block has to be at least 10 words

18 A = 0

19 !$OMP PARALLEL PRIVATE(/BLOCK5/)

20
21 ! Without the private clause,
22 ! we would be passing a member of a sequence
23 ! that is at least ten elements long.
24 ! With the private clause, A may no longer be
25 ! sequence-associated.

26 CALL SUB1(A)
27 !$OMP MASTER
28 PRINT *, A
29 !$OMP END MASTER

30 !$OMP END PARALLEL

31 END PROGRAM A32_5
32
Fortran

33 Appendix A Examples 239

1
C/C++

2 A.33 C/C++ Arrays in a firstprivate Clause

3 The following example illustrates the size and value of list items of array or pointer type
4 in a firstprivate clause (Section 2.9.3.4 on page 92). The size of new list items is
5 based on the type of the corresponding original list item, as determined by the base
6 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
15 array is assigned to the corresponding element of the new array. Those of pointer type
16 are initialized as if by assignment from the original item to the new item.

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 void f(int n, int B[n][n], int C[])

5 {
6 int D[2][2] = {1, 2, 3, 4};
7 int E[n][n];

8 assert(n >= 2);

9 E[1][1] = 4;

10 #pragma omp parallel firstprivate(B, C, D, E)

11 {
12 assert(sizeof(B) == sizeof(int (*)[n]));
13 assert(sizeof(C) == sizeof(int*));
14 assert(sizeof(D) == 4 * sizeof(int));
15 assert(sizeof(E) == n * n * sizeof(int));

16 /* Private B and C have values of original B and C. */

17 assert(&B[1][1] == &A[1][1]);
18 assert(&C[3] == &A[1][1]);
19 assert(D[1][1] == 4);
20 assert(E[1][1] == 4);
21 }
22 }

23 int main() {
24 f(2, A, A[0]);
25 return 0;
26 }
C/C++

27

28 A.34 The lastprivate Clause

29 Correct execution sometimes depends on the value that the last iteration of a loop
30 assigns to a variable. Such programs must list all such variables in a lastprivate
31 clause (Section 2.9.3.5 on page 94) so that the values of the variables are the same as
32 when the loop is executed sequentially.

33 Appendix A Examples 241

1
C/C++
2 Example A.34.1c

3 void a34 (int n, float *a, float *b)

4 {
5 int i;

6 #pragma omp parallel

7 {
8 #pragma omp for lastprivate(i)
9 for (i=0; i<n-1; i++)
10 a[i] = b[i] + b[i+1];
11 }

12 a[i]=b[i]; /* i == n-1 here */

13 }
C/C++
14
Fortran
15 Example A.34.1f

16 SUBROUTINE A34(N, A, B)

17 INTEGER N
18 REAL A(*), B(*)
19 INTEGER I

20 !$OMP PARALLEL
21 !$OMP DO LASTPRIVATE(I)

22 DO I=1,N-1
23 A(I) = B(I) + B(I+1)
24 ENDDO

25 !$OMP END PARALLEL

26 A(I) = B(I) ! I has the value of N here

27 END SUBROUTINE A34

28
Fortran

29

30 A.35 The reduction Clause

31 The following example demonstrates the reduction clause (Section 2.9.3.6 on page
32 96):

33 242 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.35.1c

3 void a35_1(float *x, int *y, int n)

4 {
5 int i, b;
6 float a;

7 a = 0.0;
8 b = 0;

9 #pragma omp parallel for private(i) shared(x, y, n) \

10 reduction(+:a) reduction(^:b)
11 for (i=0; i<n; i++) {

12 a += x[i];
13 b ^= y[i];

14 }

15 }
C/C++
16
Fortran
17 Example A.35.1f

18 SUBROUTINE A35_1(A, B, X, Y, N)

19 INTEGER N
20 REAL X(*), Y(*), A, B

21 !$OMP PARALLEL DO PRIVATE(I) SHARED(X, N) REDUCTION(+:A)

22 !$OMP& REDUCTION(MIN:B)

23 DO I=1,N

24 A = A + X(I)

25 B = MIN(B, Y(I))

26 ! Note that some reductions can be expressed in

27 ! other forms. For example, the MIN could be expressed as
28 ! IF (B > Y(I)) B = Y(I)

29 END DO

30 END SUBROUTINE A35_1

31
Fortran

32 Appendix A Examples 243

1 A common implementation of the preceding example is to treat it as if it had been
2 written as follows:
3
C/C++
4 Example A.35.2c

5 void a35_2(float *x, int *y, int n)

6 {
7 int i, b, b_p;
8 float a, a_p;

9 a = 0.0;
10 b = 0;

11 #pragma omp parallel shared(a, b, x, y, n) \

12 private(a_p, b_p)
13 {
14 a_p = 0.0;
15 b_p = 0;

16 #pragma omp for private(i)

17 for (i=0; i<n; i++) {

18 a_p += x[i];
19 b_p ^= y[i];

20 }

21 #pragma omp critical

22 {
23 a += a_p;
24 b ^= b_p;
25 }

26 }
27 }
C/C++

28 244 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.35.2f

3 SUBROUTINE A35_2 (A, B, X, Y, N)

4
5 INTEGER N
6 REAL X(*), Y(*), A, B, A_P, B_P

7 !$OMP PARALLEL SHARED(X, Y, N, A, B) PRIVATE(A_P, B_P)

8 A_P = 0.0
9 B_P = HUGE(B_P)

10 !$OMP DO PRIVATE(I)
11 DO I=1,N
12 A_P = A_P + X(I)
13 B_P = MIN(B_P, Y(I))
14 ENDDO
15 !$OMP END DO

16 !$OMP CRITICAL
17 A = A + A_P
18 B = MIN(B, B_P)
19 !$OMP END CRITICAL

20 !$OMP END PARALLEL

21 END SUBROUTINE A35_2

22 The following program is non-conforming because the reduction is on the intrinsic

23 procedure name MAX but that name has been redefined to be the variable named MAX.

24 Example A.35.3f

25 PROGRAM A35_3_WRONG
26 MAX = HUGE(0)
27 M = 0
28
29 !$OMP PARALLEL DO REDUCTION(MAX: M)
30 ! MAX is no longer the intrinsic so this is non-conforming
31 DO I = 1, 100
32 CALL SUB(M,I)
33 END DO
34
35 END PROGRAM A35_3_WRONG
36
37 SUBROUTINE SUB(M,I)
38 M = MAX(M,I)
39 END SUBROUTINE SUB

40 Appendix A Examples 245

1 The following conforming program performs the reduction using the intrinsic procedure
2 name MAX even though the intrinsic MAX has been renamed to REN.

3 Example A.35.4f

4 MODULE M
5 INTRINSIC MAX
6 END MODULE M

7 PROGRAM A35_4
8 USE M, REN => MAX
9 N = 0
10 !$OMP PARALLEL DO REDUCTION(REN: N) ! still does MAX
11 DO I = 1, 100
12 N = MAX(N,I)
13 END DO
14 END PROGRAM A35_4

15 The following conforming program performs the reduction using intrinsic procedure
16 name MAX even though the intrinsic MAX has been renamed to MIN.

17 Example A.35.5f

18 MODULE MOD
19 INTRINSIC MAX, MIN
20 END MODULE MOD

21 PROGRAM A35_5
22 USE MOD, MIN=>MAX, MAX=>MIN
23 REAL :: R
24 R = -HUGE(0.0)

25 !$OMP PARALLEL DO REDUCTION(MIN: R) ! still does MAX

26 DO I = 1, 1000
27 R = MIN(R, SIN(REAL(I)))
28 END DO
29 PRINT *, R
30 END PROGRAM A35_5
31
Fortran
32 The following example is non-conforming because the initialization (a = 0) of the
33 original list item "a" is not synchronized with the update of "a" as a result of the
34 reduction computation in the for loop. Therefore, the example may print an incorrect
35 value for "a".

36 246 OpenMP API • Version 3.0 May 2008

1 To avoid this problem, the initialization of the original list item "a" should complete
2 before any update of "a" as a result of the reduction clause. This can be achieved by
3 adding an explicit barrier after the assignment a = 0, or by enclosing the assignment a =
4 0 in a single directive (which has an implied barrier), or by initializing "a" before the
5 start of the parallel region.
6
C/C++
7 Example A.35.3c

8 #include <stdio.h>

9 int main (void)

10 {
11 int a, i;

12 #pragma omp parallel shared(a) private(i)

13 {
14 #pragma omp master
15 a = 0;

16 // To avoid race conditions, add a barrier here.

17 #pragma omp for reduction(+:a)

18 for (i = 0; i < 10; i++) {
19 a += i;
20 }

21 #pragma omp single

22 printf ("Sum is %d\n", a);
23 }
24 }
C/C++

25 Appendix A Examples 247

1
Fortran
2 Example A.35.6f

3 INTEGER A, I

4 !$OMP PARALLEL SHARED(A) PRIVATE(I)

5 !$OMP MASTER
6 A = 0
7 !$OMP END MASTER

8 ! To avoid race conditions, add a barrier here.

9 !$OMP DO REDUCTION(+:A)
10 DO I= 0, 9
11 A = A + I
12 END DO

13 !$OMP SINGLE
14 PRINT *, "Sum is ", A
15 !$OMP END SINGLE

16 !$OMP END PARALLEL

17 END
18
Fortran

19

20 A.36 The copyin Clause

21 The copyin clause (see Section 2.9.4.1 on page 101) is used to initialize threadprivate
22 data upon entry to a parallel region. The value of the threadprivate variable in the
23 master thread is copied to the threadprivate variable of each other team member.

24 248 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.36.1c

3 #include <stdlib.h>

4 float* work;
5 int size;
6 float tol;

7 #pragma omp threadprivate(work,size,tol)

8 void build()
9 {
10 int i;
11 work = (float*)malloc( sizeof(float)*size );
12 for( i = 0; i < size; ++i ) work[i] = tol;
13 }

14 void a36( float t, int n )

15 {
16 tol = t;
17 size = n;
18 #pragma omp parallel copyin(tol,size)
19 {
20 build();
21 }
22 }

23
C/C++

24 Appendix A Examples 249

1
Fortran
2 Example A.36.1f

3 MODULE M
4 REAL, POINTER, SAVE :: WORK(:)
5 INTEGER :: SIZE
6 REAL :: TOL
7 !$OMP THREADPRIVATE(WORK,SIZE,TOL)
8 END MODULE M

9 SUBROUTINE A36( T, N )
10 USE M
11 REAL :: T
12 INTEGER :: N
13 TOL = T
14 SIZE = N
15 !$OMP PARALLEL COPYIN(TOL,SIZE)
16 CALL BUILD
17 !$OMP END PARALLEL
18 END SUBROUTINE A36

19 SUBROUTINE BUILD
20 USE M
21 ALLOCATE(WORK(SIZE))
22 WORK = TOL
23 END SUBROUTINE BUILD
24
Fortran

25

26 A.37 The copyprivate Clause

27 The copyprivate clause (see Section 2.9.4.2 on page 102) can be used to broadcast
28 values acquired by a single thread directly to all instances of the private variables in the
29 other threads. In this example, if the routine is called from the sequential part, its
30 behavior is not affected by the presence of the directives. If it is called from a
31 parallel region, then the actual arguments with which a and b are associated must be
32 private. After the input routine has been executed by one thread, no thread leaves the
33 construct until the private variables designated by a, b, x, and y in all threads have
34 become defined with the values read.

35 250 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.37.1c

3 #include <stdio.h>
4 float x, y;
5 #pragma omp threadprivate(x, y)

6 void init(float a, float b ) {

7 #pragma omp single copyprivate(a,b,x,y)
8 {
9 scanf("%f %f %f %f", &a, &b, &x, &y);
10 }
11 }
C/C++
12
Fortran
13 Example A.37.1f

14 SUBROUTINE INIT(A,B)
15 REAL A, B
16 COMMON /XY/ X,Y
17 !$OMP THREADPRIVATE (/XY/)

18 !$OMP SINGLE
19 READ (11) A,B,X,Y
20 !$OMP END SINGLE COPYPRIVATE (A,B,/XY/)

21 END SUBROUTINE INIT

22
Fortran

23 Appendix A Examples 251

1 In contrast to the previous example, suppose the input must be performed by a particular
2 thread, say the master thread. In this case, the copyprivate clause cannot be used to
3 do the broadcast directly, but it can be used to provide access to a temporary shared
4 variable.
5
C/C++
6 Example A.37.2c

7 #include <stdio.h>
8 #include <stdlib.h>

9 float read_next( ) {
10 float * tmp;
11 float return_val;

12 #pragma omp single copyprivate(tmp)

13 {
14 tmp = (float *) malloc(sizeof(float));
15 } /* copies the pointer only */

16 #pragma omp master

17 {
18 scanf("%f", tmp);
19 }

20 #pragma omp barrier

21 return_val = *tmp;
22 #pragma omp barrier

23 #pragma omp single nowait

24 {
25 free(tmp);
26 }

27 return return_val;
28 }
C/C++

29 252 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.37.2f

3 REAL FUNCTION READ_NEXT()

4 REAL, POINTER :: TMP

5 !$OMP SINGLE
6 ALLOCATE (TMP)
7 !$OMP END SINGLE COPYPRIVATE (TMP) ! copies the pointer only

8 !$OMP MASTER
9 READ (11) TMP
10 !$OMP END MASTER

11 !$OMP BARRIER
12 READ_NEXT = TMP
13 !$OMP BARRIER

14 !$OMP SINGLE
15 DEALLOCATE (TMP)
16 !$OMP END SINGLE NOWAIT
17 END FUNCTION READ_NEXT
18
Fortran
19 Suppose that the number of lock variables required within a parallel region cannot
20 easily be determined prior to entering it. The copyprivate clause can be used to
21 provide access to shared lock variables that are allocated within that parallel region.
22
C/C++
23 Example A.37.3c

24 #include <stdio.h>
25 #include <stdlib.h>
26 #include <omp.h>

27 omp_lock_t *new_lock()
28 {
29 omp_lock_t *lock_ptr;

30 #pragma omp single copyprivate(lock_ptr)

31 {
32 lock_ptr = (omp_lock_t *) malloc(sizeof(omp_lock_t));
33 omp_init_lock( lock_ptr );
34 }

35 return lock_ptr;
36 }
C/C++

37 Appendix A Examples 253

1
Fortran
2 Example A.37.3f

3 FUNCTION NEW_LOCK()
4 USE OMP_LIB ! or INCLUDE "omp_lib.h"
5 INTEGER(OMP_LOCK_KIND), POINTER :: NEW_LOCK

6 !$OMP SINGLE
7 ALLOCATE(NEW_LOCK)
8 CALL OMP_INIT_LOCK(NEW_LOCK)
9 !$OMP END SINGLE COPYPRIVATE(NEW_LOCK)
10 END FUNCTION NEW_LOCK

11 Note that the effect of the copyprivate clause on a variable with the allocatable
12 attribute is different than on a variable with the pointer attribute.

13 Example A.37.4f

14 SUBROUTINE S(N)
15 INTEGER N

16 REAL, DIMENSION(:), ALLOCATABLE :: A

17 REAL, DIMENSION(:), POINTER :: B
18
19 ALLOCATE (A(N))
20 !$OMP SINGLE
21 ALLOCATE (B(N))
22 READ (11) A,B
23 !$OMP END SINGLE COPYPRIVATE(A,B)
24 ! Variable A is private and is
25 ! assigned the same value in each thread
26 ! Variable B is shared

27 !$OMP BARRIER
28 !$OMP SINGLE
29 DEALLOCATE (B)
30 !$OMP END SINGLE NOWAIT
31 END SUBROUTINE S
32
Fortran

33 254 OpenMP API • Version 3.0 May 2008

1

2 A.38 Nested Loop Constructs

3 The following example of loop construct nesting (see Section 2.10 on page 104) is
4 conforming because the inner and outer loop regions bind to different parallel
5 regions:
6
C/C++
7 Example A.38.1c

8 void work(int i, int j) {}

9 void good_nesting(int n)
10 {
11 int i, j;
12 #pragma omp parallel default(shared)
13 {
14 #pragma omp for
15 for (i=0; i<n; i++) {
16 #pragma omp parallel shared(i, n)
17 {
18 #pragma omp for
19 for (j=0; j < n; j++)
20 work(i, j);
21 }
22 }
23 }
24 }
C/C++

25 Appendix A Examples 255

1
Fortran
2 Example A.38.1f

3 SUBROUTINE WORK(I, J)
4 INTEGER I, J
5 END SUBROUTINE WORK

6 SUBROUTINE GOOD_NESTING(N)
7 INTEGER N

8 INTEGER I
9 !$OMP PARALLEL DEFAULT(SHARED)
10 !$OMP DO
11 DO I = 1, N
12 !$OMP PARALLEL SHARED(I,N)
13 !$OMP DO
14 DO J = 1, N
15 CALL WORK(I,J)
16 END DO
17 !$OMP END PARALLEL
18 END DO
19 !$OMP END PARALLEL
20 END SUBROUTINE GOOD_NESTING
21
Fortran

22 256 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 void work1(int i, int n)

6 {
7 int j;
8 #pragma omp parallel default(shared)
9 {
10 #pragma omp for
11 for (j=0; j<n; j++)
12 work(i, j);
13 }
14 }

15 void good_nesting2(int n)
16 {
17 int i;
18 #pragma omp parallel default(shared)
19 {
20 #pragma omp for
21 for (i=0; i<n; i++)
22 work1(i, n);
23 }
24 }
C/C++

25 Appendix A Examples 257

1
Fortran
2 Example A.38.2f

3 SUBROUTINE WORK(I, J)
4 INTEGER I, J
5 END SUBROUTINE WORK

6 SUBROUTINE WORK1(I, N)
7 INTEGER J
8 !$OMP PARALLEL DEFAULT(SHARED)
9 !$OMP DO
10 DO J = 1, N
11 CALL WORK(I,J)
12 END DO
13 !$OMP END PARALLEL
14 END SUBROUTINE WORK1

15 SUBROUTINE GOOD_NESTING2(N)
16 INTEGER N
17 !$OMP PARALLEL DEFAULT(SHARED)
18 !$OMP DO
19 DO I = 1, N
20 CALL WORK1(I, N)
21 END DO
22 !$OMP END PARALLEL
23 END SUBROUTINE GOOD_NESTING2
24
Fortran

25

26 A.39 Restrictions on Nesting of Regions

27 The examples in this section illustrate the region nesting rules. For more information on
28 region nesting, see Section 2.10 on page 104.

29 258 OpenMP API • Version 3.0 May 2008

1 The following example is non-conforming because the inner and outer loop regions are
2 closely nested:
3
C/C++
4 Example A.39.1c

5 void work(int i, int j) {}

6 void wrong1(int n)
7 {
8 #pragma omp parallel default(shared)
9 {
10 int i, j;
11 #pragma omp for
12 for (i=0; i<n; i++) {
13 /* incorrect nesting of loop regions */
14 #pragma omp for
15 for (j=0; j<n; j++)
16 work(i, j);
17 }
18 }
19 }
C/C++
20
Fortran
21 Example A.39.1f

22 SUBROUTINE WORK(I, J)
23 INTEGER I, J
24 END SUBROUTINE WORK

25 SUBROUTINE WRONG1(N)
26 INTEGER N

27 INTEGER I,J
28 !$OMP PARALLEL DEFAULT(SHARED)
29 !$OMP DO
30 DO I = 1, N
31 !$OMP DO ! incorrect nesting of loop regions
32 DO J = 1, N
33 CALL WORK(I,J)
34 END DO
35 END DO
36 !$OMP END PARALLEL
37 END SUBROUTINE WRONG1
38
Fortran

39 Appendix A Examples 259

1 The following orphaned version of the preceding example is also non-conforming:
C/C++
2 Example A.39.2c

3 void work(int i, int j) {}

4 void work1(int i, int n)
5 {
6 int j;
7 /* incorrect nesting of loop regions */
8 #pragma omp for
9 for (j=0; j<n; j++)
10 work(i, j);
11 }

12 void wrong2(int n)
13 {
14 #pragma omp parallel default(shared)
15 {
16 int i;
17 #pragma omp for
18 for (i=0; i<n; i++)
19 work1(i, n);
20 }
21 }
C/C++
22
Fortran
23 Example A.39.2f

24 SUBROUTINE WORK1(I,N)
25 INTEGER I, N
26 INTEGER J
27 !$OMP DO ! incorrect nesting of loop regions
28 DO J = 1, N
29 CALL WORK(I,J)
30 END DO
31 END SUBROUTINE WORK1
32 SUBROUTINE WRONG2(N)
33 INTEGER N
34 INTEGER I
35 !$OMP PARALLEL DEFAULT(SHARED)
36 !$OMP DO
37 DO I = 1, N
38 CALL WORK1(I,N)
39 END DO
40 !$OMP END PARALLEL
41 END SUBROUTINE WRONG2
42
Fortran

43 260 OpenMP API • Version 3.0 May 2008

1 The following example is non-conforming because the loop and single regions are
2 closely nested:
3
C/C++
4 Example A.39.3c

5 void work(int i, int j) {}

6 void wrong3(int n)
7 {
8 #pragma omp parallel default(shared)
9 {
10 int i;
11 #pragma omp for
12 for (i=0; i<n; i++) {
13 /* incorrect nesting of regions */
14 #pragma omp single
15 work(i, 0);
16 }
17 }
18 }
C/C++
19
Fortran
20 Example A.39.3f

21 SUBROUTINE WRONG3(N)
22 INTEGER N

23 INTEGER I
24 !$OMP PARALLEL DEFAULT(SHARED)
25 !$OMP DO
26 DO I = 1, N
27 !$OMP SINGLE ! incorrect nesting of regions
28 CALL WORK(I, 1)
29 !$OMP END SINGLE
30 END DO
31 !$OMP END PARALLEL
32 END SUBROUTINE WRONG3
33
Fortran

34 Appendix A Examples 261

1 The following example is non-conforming because a barrier region cannot be closely
2 nested inside a loop region:
3
C/C++
4 Example A.39.4c

5 void work(int i, int j) {}

6 void wrong4(int n)
7 {

8 #pragma omp parallel default(shared)

9 {
10 int i;
11 #pragma omp for
12 for (i=0; i<n; i++) {
13 work(i, 0);
14 /* incorrect nesting of barrier region in a loop region */
15 #pragma omp barrier
16 work(i, 1);
17 }
18 }
19 }
C/C++
20
Fortran
21 Example A.39.4f

22 SUBROUTINE WRONG4(N)
23 INTEGER N

24 INTEGER I
25 !$OMP PARALLEL DEFAULT(SHARED)
26 !$OMP DO
27 DO I = 1, N
28 CALL WORK(I, 1)
29 ! incorrect nesting of barrier region in a loop region
30 !$OMP BARRIER
31 CALL WORK(I, 2)
32 END DO
33 !$OMP END PARALLEL
34 END SUBROUTINE WRONG4
35
Fortran

36 262 OpenMP API • Version 3.0 May 2008

1 The following example is non-conforming because the barrier region cannot be
2 closely nested inside the critical region. If this were permitted, it would result in
3 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 void work(int i, int j) {}

7 void wrong5(int n)
8 {
9 #pragma omp parallel
10 {
11 #pragma omp critical
12 {
13 work(n, 0);
14 /* incorrect nesting of barrier region in a critical region */
15 #pragma omp barrier
16 work(n, 1);
17 }
18 }
19 }
C/C++
20
Fortran
21 Example A.39.5f

22 SUBROUTINE WRONG5(N)
23 INTEGER N

24 !$OMP PARALLEL DEFAULT(SHARED)

25 !$OMP CRITICAL
26 CALL WORK(N,1)
27 ! incorrect nesting of barrier region in a critical region
28 !$OMP BARRIER
29 CALL WORK(N,2)
30 !$OMP END CRITICAL
31 !$OMP END PARALLEL
32 END SUBROUTINE WRONG5
33
Fortran

34 Appendix A Examples 263

1 The following example is non-conforming because the barrier region cannot be
2 closely nested inside the single region. If this were permitted, it would result in
3 deadlock due to the fact that only one thread executes the single region:
4
C/C++
5 Example A.39.6c

6 void work(int i, int j) {}

7 void wrong6(int n)
8 {
9 #pragma omp parallel
10 {
11 #pragma omp single
12 {
13 work(n, 0);
14 /* incorrect nesting of barrier region in a single region */
15 #pragma omp barrier
16 work(n, 1);
17 }
18 }
19 }
C/C++
20
Fortran
21 Example A.39.6f

22 SUBROUTINE WRONG6(N)
23 INTEGER N

24 !$OMP PARALLEL DEFAULT(SHARED)

25 !$OMP SINGLE
26 CALL WORK(N,1)
27 ! incorrect nesting of barrier region in a single region
28 !$OMP BARRIER
29 CALL WORK(N,2)
30 !$OMP END SINGLE
31 !$OMP END PARALLEL
32 END SUBROUTINE WRONG6
33
Fortran

34 264 OpenMP API • Version 3.0 May 2008

1

2 A.40 The omp_set_dynamic and

3 omp_set_num_threads Routines
4 Some programs rely on a fixed, prespecified number of threads to execute correctly.
5 Because the default setting for the dynamic adjustment of the number of threads is
6 implementation defined, such programs can choose to turn off the dynamic threads
7 capability and set the number of threads explicitly to ensure portability. The following
8 example shows how to do this using omp_set_dynamic (Section 3.2.7 on page 117),
9 and omp_set_num_threads (Section 3.2.1 on page 110).
10 In this example, the program executes correctly only if it is executed by 16 threads. If
11 the implementation is not capable of supporting 16 threads, the behavior of this example
12 is implementation defined. Note that the number of threads executing a parallel
13 region remains constant during the region, regardless of the dynamic threads setting.
14 The dynamic threads mechanism determines the number of threads to use at the start of
15 the parallel region and keeps it constant for the duration of the region.
16
C/C++
17 Example A.40.1c

18 #include <omp.h>
19 #include <stdlib.h>

20 void do_by_16(float *x, int iam, int ipoints) {}

21 void a40(float *x, int npoints)

22 {
23 int iam, ipoints;

24 omp_set_dynamic(0);
25 omp_set_num_threads(16);

26 #pragma omp parallel shared(x, npoints) private(iam, ipoints)

27 {
28 if (omp_get_num_threads() != 16)
29 abort();

30 iam = omp_get_thread_num();
31 ipoints = npoints/16;
32 do_by_16(x, iam, ipoints);
33 }
34 }
C/C++

35 Appendix A Examples 265

1
Fortran
2 Example A.40.1f

3 SUBROUTINE DO_BY_16(X, IAM, IPOINTS)

4 REAL X(*)
5 INTEGER IAM, IPOINTS
6 END SUBROUTINE DO_BY_16

7 SUBROUTINE SUBA40(X, NPOINTS)

8 INCLUDE "omp_lib.h" ! or USE OMP_LIB

9 INTEGER NPOINTS
10 REAL X(NPOINTS)

11 INTEGER IAM, IPOINTS

12 CALL OMP_SET_DYNAMIC(.FALSE.)
13 CALL OMP_SET_NUM_THREADS(16)

14 !$OMP PARALLEL SHARED(X,NPOINTS) PRIVATE(IAM, IPOINTS)

15 IF (OMP_GET_NUM_THREADS() .NE. 16) THEN

16 STOP
17 ENDIF

18 IAM = OMP_GET_THREAD_NUM()
19 IPOINTS = NPOINTS/16
20 CALL DO_BY_16(X,IAM,IPOINTS)

21 !$OMP END PARALLEL

22 END SUBROUTINE SUBA40

23
Fortran

24

25 A.41 The omp_get_num_threads Routine

26 In the following example, the omp_get_num_threads call (see Section 3.2.2 on
27 page 111) returns 1 in the sequential part of the code, so np will always be equal to 1. To
28 determine the number of threads that will be deployed for the parallel region, the
29 call should be inside the parallel region.

30 266 OpenMP API • Version 3.0 May 2008

1
C/C++
2 Example A.41.1c

3 #include <omp.h>
4 void work(int i);

5 void incorrect()
6 {
7 int np, i;

8 np = omp_get_num_threads(); /* misplaced */

9 #pragma omp parallel for schedule(static)

10 for (i=0; i < np; i++)
11 work(i);
12 }
C/C++
13
Fortran
14 Example A.41.1f

15 SUBROUTINE WORK(I)
16 INTEGER I
17 I = I + 1
18 END SUBROUTINE WORK

19 SUBROUTINE INCORRECT()
20 INCLUDE "omp_lib.h" ! or USE OMP_LIB
21 INTEGER I, NP

22 NP = OMP_GET_NUM_THREADS() !misplaced: will return 1

23 !$OMP PARALLEL DO SCHEDULE(STATIC)
24 DO I = 0, NP-1
25 CALL WORK(I)
26 ENDDO
27 !$OMP END PARALLEL DO
28 END SUBROUTINE INCORRECT
29
Fortran

30 Appendix A Examples 267

1 The following example shows how to rewrite this program without including a query for
2 the number of threads:
3
C/C++
4 Example A.41.2c

5 #include <omp.h>
6 void work(int i);

7 void correct()
8 {
9 int i;

10 #pragma omp parallel private(i)

11 {
12 i = omp_get_thread_num();
13 work(i);
14 }
15 }
C/C++
16
Fortran
17 Example A.41.2f

18 SUBROUTINE WORK(I)
19 INTEGER I
20
21 I = I + 1

22 END SUBROUTINE WORK

23
24 SUBROUTINE CORRECT()
25 INCLUDE "omp_lib.h" ! or USE OMP_LIB
26 INTEGER I

27 !$OMP PARALLEL PRIVATE(I)

28 I = OMP_GET_THREAD_NUM()
29 CALL WORK(I)
30 !$OMP END PARALLEL

31 END SUBROUTINE CORRECT

32
Fortran

33 268 OpenMP API • Version 3.0 May 2008

1

2 A.42 The omp_init_lock Routine

3 The following example demonstrates how to initialize an array of locks in a parallel
4 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 omp_lock_t *new_locks()
9 {
10 int i;
11 omp_lock_t *lock = new omp_lock_t[1000];

12 #pragma omp parallel for private(i)

13 for (i=0; i<1000; i++)
14 {
15 omp_init_lock(&lock[i]);
16 }
17 return lock;
18 }
C/C++
19
Fortran
20 Example A.42.1f

21 FUNCTION NEW_LOCKS()
22 USE OMP_LIB ! or INCLUDE "omp_lib.h"
23 INTEGER(OMP_LOCK_KIND), DIMENSION(1000) :: NEW_LOCKS

24 INTEGER I

25 !$OMP PARALLEL DO PRIVATE(I)

26 DO I=1,1000
27 CALL OMP_INIT_LOCK(NEW_LOCKS(I))
28 END DO
29 !$OMP END PARALLEL DO
30
31 END FUNCTION NEW_LOCKS
32
Fortran

33 Appendix A Examples 269

1

2 A.43 Ownership of Locks

3 Ownership of locks has changed from OpenMP 2.5 to OpenMP 3.0. In OpenMP 2.5,
4 locks are owned by threads; so a lock released by the omp_unset_lock routine must
5 be owned by the same thread executing the routine. In OpenMP 3.0, on the other hand,
6 locks are owned by task regions; so a lock released by the omp_unset_lock routine
7 in a task region must be owned by the same task region.

8 This change in ownership requires extra care when using locks. The following program
9 is conforming in OpenMP 2.5 because the thread that releases the lock lck in the parallel
10 region is the same thread that acquired the lock in the sequential part of the program
11 (master thread of parallel region and the initial thread are the same). However, it is not
12 conforming in OpenMP 3.0, because the task region that releases the lock lck is different
13 from the task region that acquires the lock.
14
C/C++
15 Example A.43.1c

16 #include <stdlib.h>
17 #include <stdio.h>
18 #include <omp.h>

19 int main()
20 {
21 int x;
22 omp_lock_t lck;
23
24 omp_init_lock (&lck);
25 omp_set_lock (&lck);
26 x = 0;

27 #pragma omp parallel shared (x)

28 {
29 #pragma omp master
30 {
31 x = x + 1;
32 omp_unset_lock (&lck);
33 }

34 /* Some more stuff. */

35 }
36
37 omp_destroy_lock (&lck);
38 }
C/C++

39 270 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.43.1f

3 program lock
4 use omp_lib
5 integer :: x
6 integer (kind=omp_lock_kind) :: lck

7 call omp_init_lock (lck)

8 call omp_set_lock(lck)
9 x = 0

10 !$omp parallel shared (x)

11 !$omp master
12 x = x + 1
13 call omp_unset_lock(lck)
14 !$omp end master
15
16 ! Some more stuff.
17 !$omp end parallel

18 call omp_destroy_lock(lck)
19 end
20
Fortran

21

22 A.44 Simple Lock Routines

23 In the following example (for Section 3.3 on page 134), the lock routines cause the
24 threads to be idle while waiting for entry to the first critical section, but to do other work
25 while waiting for entry to the second. The omp_set_lock function blocks, but the
26 omp_test_lock function does not, allowing the work in skip to be done.

27 Appendix A Examples 271

1
C/C++
2 Note that the argument to the lock routines should have type omp_lock_t, and that
3 there is no need to flush it.

4 Example A.44.1c

5 #include <stdio.h>
6 #include <omp.h>

7 void skip(int i) {}
8 void work(int i) {}

9 int main()
10 {
11 omp_lock_t lck;
12 int id;

13 omp_init_lock(&lck);

14 #pragma omp parallel shared(lck) private(id)

15 {
16 id = omp_get_thread_num();

17 omp_set_lock(&lck);
18 /* only one thread at a time can execute this printf */
19 printf("My thread id is %d.\n", id);
20 omp_unset_lock(&lck);

21 while (! omp_test_lock(&lck)) {
22 skip(id); /* we do not yet have the lock,
23 so we must do something else */
24 }

25 work(id); /* we now have the lock

26 and can do the work */

27 omp_unset_lock(&lck);
28 }

29 omp_destroy_lock(&lck);
30
31 return 0;
32 }
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 SUBROUTINE SKIP(ID)
5 END SUBROUTINE SKIP

6 SUBROUTINE WORK(ID)
7 END SUBROUTINE WORK

8 PROGRAM A44

9 INCLUDE "omp_lib.h" ! or USE OMP_LIB

10 INTEGER(OMP_LOCK_KIND) LCK
11 INTEGER ID

12 CALL OMP_INIT_LOCK(LCK)

13 !$OMP PARALLEL SHARED(LCK) PRIVATE(ID)

14 ID = OMP_GET_THREAD_NUM()
15 CALL OMP_SET_LOCK(LCK)
16 PRINT *, 'My thread id is ', ID
17 CALL OMP_UNSET_LOCK(LCK)

18 DO WHILE (.NOT. OMP_TEST_LOCK(LCK))

19 CALL SKIP(ID) ! We do not yet have the lock
20 ! so we must do something else
21 END DO

22 CALL WORK(ID) ! We now have the lock

23 ! and can do the work

24 CALL OMP_UNSET_LOCK( LCK )

25 !$OMP END PARALLEL

26 CALL OMP_DESTROY_LOCK( LCK )

27 END PROGRAM A44

28
Fortran

29 Appendix A Examples 273

1

2 A.45 Nestable Lock Routines

3 The following example (for Section 3.3 on page 134) demonstrates how a nestable lock
4 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 #include <omp.h>
8 typedef struct {
9 int a,b;
10 omp_nest_lock_t lck; } pair;

11 int work1();
12 int work2();
13 int work3();
14 void incr_a(pair *p, int a)
15 {
16 /* Called only from incr_pair, no need to lock. */
17 p->a += a;
18 }
19 void incr_b(pair *p, int b)
20 {
21 /* Called both from incr_pair and elsewhere, */
22 /* so need a nestable lock. */

23 omp_set_nest_lock(&p->lck);
24 p->b += b;
25 omp_unset_nest_lock(&p->lck);
26 }
27 void incr_pair(pair *p, int a, int b)
28 {
29 omp_set_nest_lock(&p->lck);
30 incr_a(p, a);
31 incr_b(p, b);
32 omp_unset_nest_lock(&p->lck);
33 }
34 void a45(pair *p)
35 {
36 #pragma omp parallel sections
37 {
38 #pragma omp section
39 incr_pair(p, work1(), work2());
40 #pragma omp section
41 incr_b(p, work3());
42 }
43 }
C/C++

44 274 OpenMP API • Version 3.0 May 2008

1
Fortran
2 Example A.45.1f

3 MODULE DATA
4 USE OMP_LIB, ONLY: OMP_NEST_LOCK_KIND
5 TYPE LOCKED_PAIR
6 INTEGER A
7 INTEGER B
8 INTEGER (OMP_NEST_LOCK_KIND) LCK
9 END TYPE
10 END MODULE DATA

11 SUBROUTINE INCR_A(P, A)
12 ! called only from INCR_PAIR, no need to lock
13 USE DATA
14 TYPE(LOCKED_PAIR) :: P
15 INTEGER A
16 P%A = P%A + A
17 END SUBROUTINE INCR_A

18 SUBROUTINE INCR_B(P, B)
19 ! called from both INCR_PAIR and elsewhere,
20 ! so we need a nestable lock
21 USE OMP_LIB ! or INCLUDE "omp_lib.h"
22 USE DATA
23 TYPE(LOCKED_PAIR) :: P
24 INTEGER B
25 CALL OMP_SET_NEST_LOCK(P%LCK)
26 P%B = P%B + B
27 CALL OMP_UNSET_NEST_LOCK(P%LCK)
28 END SUBROUTINE INCR_B

29 SUBROUTINE INCR_PAIR(P, A, B)
30 USE OMP_LIB ! or INCLUDE "omp_lib.h"
31 USE DATA
32 TYPE(LOCKED_PAIR) :: P
33 INTEGER A
34 INTEGER B

35 CALL OMP_SET_NEST_LOCK(P%LCK)
36 CALL INCR_A(P, A)
37 CALL INCR_B(P, B)
38 CALL OMP_UNSET_NEST_LOCK(P%LCK)
39 END SUBROUTINE INCR_PAIR

40 SUBROUTINE A45(P)
41 USE OMP_LIB ! or INCLUDE "omp_lib.h"
42 USE DATA
43 TYPE(LOCKED_PAIR) :: P
44 INTEGER WORK1, WORK2, WORK3
45 EXTERNAL WORK1, WORK2, WORK3

46 Appendix A Examples 275

1 !$OMP PARALLEL SECTIONS

2 !$OMP SECTION
3 CALL INCR_PAIR(P, WORK1(), WORK2())
4 !$OMP SECTION
5 CALL INCR_B(P, WORK3())
6 !$OMP END PARALLEL SECTIONS

7 END SUBROUTINE A45

8
Fortran

9 276 OpenMP API • Version 3.0 May 2008

1
2 APPENDIX B

3 Stubs for Runtime Library

4 Routines
5

6 This section provides stubs for the runtime library routines defined in the OpenMP API.
7 The stubs are provided to enable portability to platforms that do not support the
8 OpenMP API. On these platforms, OpenMP programs must be linked with a library
9 containing these stub routines. The stub routines assume that the directives in the
10 OpenMP program are ignored. As such, they emulate serial semantics.
11 Note that the lock variable that appears in the lock routines must be accessed
12 exclusively through these routines. It should not be initialized or otherwise modified in
13 the user program.
14 In an actual implementation the lock variable might be used to hold the address of an
15 allocated memory block, but here it is used to hold an integer value. Users should not
16 make assumptions about mechanisms used by OpenMP implementations to implement
17 locks based on the scheme used by the stub procedures.
18
Fortran
19 Note – In order to be able to compile the Fortran stubs file, the include file
20 omp_lib.h was split into two files: omp_lib_kinds.h and omp_lib.h and the
21 omp_lib_kinds.h file included where needed. There is no requirement for the
22 implementation to provide separate files.
23
Fortran

24 277
1

2 B.1 C/C++ Stub Routines

3 #include <stdio.h>
4 #include <stdlib.h>
5 #include "omp.h"

6 void omp_set_num_threads(int num_threads)

7 {
8 }

9 int omp_get_num_threads(void)
10 {
11 return 1;
12 }

13 int omp_get_max_threads(void)
14 {
15 return 1;
16 }

17 int omp_get_thread_num(void)
18 {
19 return 0;
20 }

21 int omp_get_num_procs(void)
22 {
23 return 1;
24 }

25 int omp_in_parallel(void)
26 {
27 return 0;
28 }

29 void omp_set_dynamic(int dynamic_threads)

30 {
31 }

32 int omp_get_dynamic(void)
33 {
34 return 0;
35 }

36 void omp_set_nested(int nested)

37 {
38 }

39 278 OpenMP API • Version 3.0 May 2008

1 int omp_get_nested(void)
2 {
3 return 0;
4 }

5 void omp_set_schedule(omp_sched_t kind, int modifier)

6 {
7 }

8 void omp_get_schedule(omp_sched_t *kind, int *modifier)

9 {
10 *kind = omp_sched_static;
11 *modifier = 0;
12 }

13 int omp_get_thread_limit(void)
14 {
15 return 1;
16 }

17 void omp_set_max_active_levels(int max_active_levels)

18 {
19 }

20 int omp_get_max_active_levels(void)
21 {
22 return 0;
23 }

24 int omp_get_level(void)
25 {
26 return 0;
27 }

28 int omp_get_ancestor_thread_num(int level)

29 {
30 if (level == 0)
31 {
32 return 0;
33 }
34 else
35 {
36 return -1;
37 }
38 }

39 Appendix B Stubs for Runtime Library Routines 279

1 int omp_get_team_size(int level)
2 {
3 if (level == 0)
4 {
5 return 1;
6 }
7 else
8 {
9 return -1;
10 }
11 }

12 int omp_get_active_level(void)
13 {
14 return 0;
15 }

16 struct __omp_lock
17 {
18 int lock;
19 };

20 enum { UNLOCKED = -1, INIT, LOCKED };

21 void omp_init_lock(omp_lock_t *arg)

22 {
23 struct __omp_lock *lock = (struct __omp_lock *)arg;
24 lock->lock = UNLOCKED;
25 }

26 void omp_destroy_lock(omp_lock_t *arg)

27 {
28 struct __omp_lock *lock = (struct __omp_lock *)arg;
29 lock->lock = INIT;
30 }

31 void omp_set_lock(omp_lock_t *arg)

32 {
33 struct __omp_lock *lock = (struct __omp_lock *)arg;
34 if (lock->lock == UNLOCKED)
35 {
36 lock->lock = LOCKED;
37 }
38 else if (lock->lock == LOCKED)
39 {
40 fprintf(stderr,
41 "error: deadlock in using lock variable\n");
42 exit(1);
43 }
44 else
45 {
46 fprintf(stderr, "error: lock not initialized\n");
47 exit(1);
48 }

49 280 OpenMP API • Version 3.0 May 2008

1 }

2 void omp_unset_lock(omp_lock_t *arg)

3 {
4 struct __omp_lock *lock = (struct __omp_lock *)arg;
5 if (lock->lock == LOCKED)
6 {
7 lock->lock = UNLOCKED;
8 }
9 else if (lock->lock == UNLOCKED)
10 {
11 fprintf(stderr, "error: lock not set\n");
12 exit(1);
13 }
14 else
15 {
16 fprintf(stderr, "error: lock not initialized\n");
17 exit(1);
18 }
19 }

20 int omp_test_lock(omp_lock_t *arg)

21 {
22 struct __omp_lock *lock = (struct __omp_lock *)arg;
23 if (lock->lock == UNLOCKED)
24 {
25 lock->lock = LOCKED;
26 return 1;
27 }
28 else if (lock->lock == LOCKED)
29 {
30 return 0;
31 }
32 else
33 {
34 fprintf(stderr, "error: lock not initialized\n");
35 exit(1);
36 }
37 }

38 struct __omp_nest_lock
39 {
40 short owner;
41 short count;
42 };

43 enum { NOOWNER = -1, MASTER = 0 };

44 void omp_init_nest_lock(omp_nest_lock_t *arg)

45 {
46 struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
47 nlock->owner = NOOWNER;
48 nlock->count = 0;

49 Appendix B Stubs for Runtime Library Routines 281

1 }

2 void omp_destroy_nest_lock(omp_nest_lock_t *arg)

3 {
4 struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
5 nlock->owner = NOOWNER;
6 nlock->count = UNLOCKED;
7 }

8 void omp_set_nest_lock(omp_nest_lock_t *arg)

9 {
10 struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
11 if (nlock->owner == MASTER && nlock->count >= 1)
12 {
13 nlock->count++;
14 }
15 else if (nlock->owner == NOOWNER && nlock->count == 0)
16 {
17 nlock->owner = MASTER;
18 nlock->count = 1;
19 }
20 else
21 {
22 fprintf(stderr,
23 "error: lock corrupted or not initialized\n");
24 exit(1);
25 }
26 }

27 void omp_unset_nest_lock(omp_nest_lock_t *arg)

28 {
29 struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
30 if (nlock->owner == MASTER && nlock->count >= 1)
31 {
32 nlock->count--;
33 if (nlock->count == 0)
34 {
35 nlock->owner = NOOWNER;
36 }
37 }
38 else if (nlock->owner == NOOWNER && nlock->count == 0)
39 {
40 fprintf(stderr, "error: lock not set\n");
41 exit(1);
42 }
43 else
44 {
45 fprintf(stderr,
46 "error: lock corrupted or not initialized\n");
47 exit(1);

48 282 OpenMP API • Version 3.0 May 2008

1 }
2 }

3 int omp_test_nest_lock(omp_nest_lock_t *arg)

4 {
5 struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg;
6 omp_set_nest_lock(arg);
7 return nlock->count;
8 }

9 double omp_get_wtime(void)
10 {
11 /* This function does not provide a working
12 * wallclock timer. Replace it with a version
13 * customized for the target machine.
14 */
15 return 0.0;
16 }

17 double omp_get_wtick(void)
18 {
19 /* This function does not provide a working
20 * clock tick function. Replace it with
21 * a version customized for the target machine.
22 */
23 return 365. * 86400.;
24 }

25 Appendix B Stubs for Runtime Library Routines 283

1

2 B.2 Fortran Stub Routines

3 C23456
4 subroutine omp_set_num_threads(num_threads)
5 integer num_threads
6 return
7 end subroutine

8 integer function omp_get_num_threads()

9 omp_get_num_threads = 1
10 return
11 end function

12 integer function omp_get_max_threads()

13 omp_get_max_threads = 1
14 return
15 end function

16 integer function omp_get_thread_num()

17 omp_get_thread_num = 0
18 return
19 end function

20 integer function omp_get_num_procs()

21 omp_get_num_procs = 1
22 return
23 end function

24 logical function omp_in_parallel()

25 omp_in_parallel = .false.
26 return
27 end function

28 subroutine omp_set_dynamic(dynamic_threads)
29 logical dynamic_threads
30 return
31 end subroutine

32 logical function omp_get_dynamic()

33 omp_get_dynamic = .false.
34 return
35 end function

36 subroutine omp_set_nested(nested)
37 logical nested
38 return
39 end subroutine

40 284 OpenMP API • Version 3.0 May 2008

1 logical function omp_get_nested()
2 omp_get_nested = .false.
3 return
4 end function

5 subroutine omp_set_schedule(kind, modifier)

6 include 'omp_lib_kinds.h'
7 integer (kind=omp_sched_kind) kind
8 integer modifier
9 return
10 end subroutine

11 subroutine omp_get_schedule(kind, modifier)

12 include 'omp_lib_kinds.h'
13 integer (kind=omp_sched_kind) kind
14 integer modifier

15 kind = omp_sched_static
16 modifier = 0
17 return
18 end subroutine

19 integer function omp_get_thread_limit()

20 omp_get_thread_limit = 1
21 return
22 end function

23 subroutine omp_set_max_active_levels( level )

24 integer level
25 end subroutine

26 integer function omp_get_max_active_levels()

27 omp_get_max_active_levels = 0
28 return
29 end function

30 integer function omp_get_level()

31 omp_get_level = 0
32 return
33 end function

34 integer function omp_get_ancestor_thread_num( level )

35 integer level
36 if ( level .eq. 0 ) then
37 omp_get_ancestor_thread_num = 0
38 else
39 omp_get_ancestor_thread_num = -1
40 end if
41 return
42 end function

43 Appendix B Stubs for Runtime Library Routines 285

1 integer function omp_get_team_size( level )
2 integer level
3 if ( level .eq. 0 ) then
4 omp_get_team_size = 1
5 else
6 omp_get_team_size = -1
7 end if
8 return
9 end function

10 integer function omp_get_active_level()

11 omp_get_active_level = 0
12 return
13 end function

14 subroutine omp_init_lock(lock)
15 ! lock is 0 if the simple lock is not initialized
16 ! -1 if the simple lock is initialized but not set
17 ! 1 if the simple lock is set
18 include 'omp_lib_kinds.h'
19 integer(kind=omp_lock_kind) lock

20 lock = -1
21 return
22 end subroutine

23 subroutine omp_destroy_lock(lock)
24 include 'omp_lib_kinds.h'
25 integer(kind=omp_lock_kind) lock

26 lock = 0
27 return
28 end subroutine

29 subroutine omp_set_lock(lock)
30 include 'omp_lib_kinds.h'
31 integer(kind=omp_lock_kind) lock

32 if (lock .eq. -1) then

33 lock = 1
34 elseif (lock .eq. 1) then
35 print *, 'error: deadlock in using lock variable'
36 stop
37 else
38 print *, 'error: lock not initialized'
39 stop
40 endif

41 return
42 end subroutine

43 286 OpenMP API • Version 3.0 May 2008

1 subroutine omp_unset_lock(lock)
2 include 'omp_lib_kinds.h'
3 integer(kind=omp_lock_kind) lock

4 if (lock .eq. 1) then

5 lock = -1
6 elseif (lock .eq. -1) then
7 print *, 'error: lock not set'
8 stop
9 else
10 print *, 'error: lock not initialized'
11 stop
12 endif

13 return
14 end subroutine

15 logical function omp_test_lock(lock)

16 include 'omp_lib_kinds.h'
17 integer(kind=omp_lock_kind) lock

18 if (lock .eq. -1) then

19 lock = 1
20 omp_test_lock = .true.
21 elseif (lock .eq. 1) then
22 omp_test_lock = .false.
23 else
24 print *, 'error: lock not initialized'
25 stop
26 endif

27 return
28 end function

29 subroutine omp_init_nest_lock(nlock)
30 ! nlock is
31 ! 0 if the nestable lock is not initialized
32 ! -1 if the nestable lock is initialized but not set
33 ! 1 if the nestable lock is set
34 ! no use count is maintained
35 include 'omp_lib_kinds.h'
36 integer(kind=omp_nest_lock_kind) nlock

37 nlock = -1

38 return
39 end subroutine

40 Appendix B Stubs for Runtime Library Routines 287

1 subroutine omp_destroy_nest_lock(nlock)
2 include 'omp_lib_kinds.h'
3 integer(kind=omp_nest_lock_kind) nlock

4 nlock = 0

5 return
6 end subroutine

7 subroutine omp_set_nest_lock(nlock)
8 include 'omp_lib_kinds.h'
9 integer(kind=omp_nest_lock_kind) nlock

10 if (nlock .eq. -1) then

11 nlock = 1
12 elseif (nlock .eq. 0) then
13 print *, 'error: nested lock not initialized'
14 stop
15 else
16 print *, 'error: deadlock using nested lock variable'
17 stop
18 endif

19 return
20 end subroutine

21 subroutine omp_unset_nest_lock(nlock)
22 include 'omp_lib_kinds.h'
23 integer(kind=omp_nest_lock_kind) nlock

24 if (nlock .eq. 1) then

25 nlock = -1
26 elseif (nlock .eq. 0) then
27 print *, 'error: nested lock not initialized'
28 stop
29 else
30 print *, 'error: nested lock not set'
31 stop
32 endif

33 return
34 end subroutine

35 288 OpenMP API • Version 3.0 May 2008

1 integer function omp_test_nest_lock(nlock)
2 include 'omp_lib_kinds.h'
3 integer(kind=omp_nest_lock_kind) nlock

4 if (nlock .eq. -1) then

5 nlock = 1
6 omp_test_nest_lock = 1
7 elseif (nlock .eq. 1) then
8 omp_test_nest_lock = 0
9 else
10 print *, 'error: nested lock not initialized'
11 stop
12 endif

13 return
14 end function

15 double precision function omp_get_wtime()

16 ! this function does not provide a working
17 ! wall clock timer. replace it with a version
18 ! customized for the target machine.

19 omp_get_wtime = 0.0d0

20 return
21 end function

22 double precision function omp_get_wtick()

23 ! this function does not provide a working
24 ! clock tick function. replace it with
25 ! a version customized for the target machine.
26 double precision one_year
27 parameter (one_year=365.d0*86400.d0)

28 omp_get_wtick = one_year

29 return
30 end function

31 Appendix B Stubs for Runtime Library Routines 289

1 290 OpenMP API • Version 3.0 May 2008
1
2 APPENDIX C

3 OpenMP C and C++ Grammar

4

6 C.1 Notation
7 The grammar rules consist of the name for a non-terminal, followed by a colon,
8 followed by replacement alternatives on separate lines.
9 The syntactic expression termopt indicates that the term is optional within the
10 replacement.
11 The syntactic expression termoptseq is equivalent to term-seqopt with the following
12 additional rules:
13 term-seq :

14 term

15 term-seq term

16 term-seq , term

17 291
1

2 C.2 Rules
3 The notation is described in Section 6.1 of the C standard. This grammar appendix
4 shows the extensions to the base language grammar for the OpenMP C and C++
5 directives.

6 /* in C++ (ISO/IEC 14882:1998) */

7 statement-seq:

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 statement:

2 /* standard statements */

3 openmp-construct
4 openmp-construct:
5 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 parallel-construct:

23 parallel-directive structured-block
24 parallel-directive:

25 # pragma omp parallel parallel-clauseoptseq new-line

26 Appendix C OpenMP C and C++ Grammar 293

1 parallel-clause:

2 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 unique-parallel-clause:

9 if ( expression )

10 num_threads ( expression )

11 copyin ( variable-list )
12 for-construct:

13 for-directive iteration-statement
14 for-directive:

15 # pragma omp for for-clauseoptseq new-line

16 for-clause:

17 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 sections-construct:

8 sections-directive section-scope
9 sections-directive:

10 # pragma omp sections sections-clauseoptseq new-line

11 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 { section-sequence }
19 section-sequence:

20 section-directiveopt structured-block

21 section-sequence section-directive structured-block

22 section-directive:

23 # pragma omp section new-line

24 single-construct:

25 single-directive structured-block
26 single-directive:

27 # pragma omp single single-clauseoptseq new-line

28 Appendix C OpenMP C and C++ Grammar 295

1 single-clause:

2 unique-single-clause

3 data-privatization-clause
4 data-privatization-in-clause

5 nowait
6 unique-single-clause:

7 copyprivate ( variable-list )
8 task-construct:

9 task-directive structured-block
10 task-directive:

11 # pragma omp task task-clauseoptseq new-line

12 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 parallel-for-construct:

22 parallel-for-directive iteration-statement
23 parallel-for-directive:

24 # pragma omp parallel for parallel-for-clauseoptseq new-line

25 parallel-for-clause:

26 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 parallel-sections-directive section-scope
8 parallel-sections-directive:

9 # pragma omp parallel sections parallel-sections-clauseoptseq new-line

10 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 master-construct:

19 master-directive structured-block
20 master-directive:

21 # pragma omp master new-line

22 critical-construct:

23 critical-directive structured-block
24 critical-directive:

25 # pragma omp critical region-phraseopt new-line

26 region-phrase:

27 ( identifier )

28 Appendix C OpenMP C and C++ Grammar 297

1 barrier-directive:

2 # pragma omp barrier new-line

3 taskwait-directive:

4 # pragma omp taskwait new-line

5 atomic-construct:

6 atomic-directive expression-statement
7 atomic-directive:

8 # pragma omp atomic new-line

9 flush-directive:

10 # pragma omp flush flush-varsopt new-line

11 flush-vars:

12 ( variable-list )
13 ordered-construct:

14 ordered-directive structured-block
15 ordered-directive:

16 # pragma omp ordered new-line

17 declaration:

18 /* standard declarations */

19 threadprivate-directive
20 threadprivate-directive:

21 # pragma omp threadprivate ( variable-list ) new-line

22 data-default-clause:

23 default ( shared )

24 default ( none )
25 data-privatization-clause:

26 private ( variable-list )
27 data-privatization-in-clause:

28 firstprivate ( variable-list )

29 298 OpenMP API • Version 3.0 May 2008

1 data-privatization-out-clause:

2 lastprivate ( variable-list )
3 data-sharing-clause:

4 shared ( variable-list )
5 data-reduction-clause:

6 reduction ( reduction-operator : variable-list )

7 reduction-operator:

8 One of: + * - & ^ | && ||

9 /* in C */
10 variable-list:

11 identifier

12 variable-list , identifier
13 /* in C++ */
14 variable-list:

15 id-expression

16 variable-list , id-expression

17 Appendix C OpenMP C and C++ Grammar 299

1 300 OpenMP API • Version 3.0 May 2008
1
2 APPENDIX D

3 Interface Declarations
4

5 This appendix gives examples of the C/C++ header file, the Fortran include file and
6 Fortran 90 module that shall be provided by implementations as specified in Chapter 3.
7 It also includes an example of a Fortran 90 generic interface for a library routine.

8 301
1

2 D.1 Example of the omp.h Header File

3 #ifndef _OMP_H_DEF
4 #define _OMP_H_DEF

5 /*
6 * define the lock data types
7 */
8 typedef void *omp_lock_t;

9 typedef void *omp_nest_lock_t;

10 /*
11 * define the schedule kinds
12 */
13 typedef enum omp_sched_t
14 {
15 omp_sched_static = 1,
16 omp_sched_dynamic = 2,
17 omp_sched_guided = 3,
18 omp_sched_auto = 4
19 /* , Add vendor specific schedule constants here */
20 } omp_sched_t;

21 /*
22 * exported OpenMP functions
23 */
24 #ifdef __cplusplus
25 extern "C"
26 {
27 #endif

28 extern void omp_set_num_threads(int num_threads);

29 extern int omp_get_num_threads(void);
30 extern int omp_get_max_threads(void);
31 extern int omp_get_thread_num(void);
32 extern int omp_get_num_procs(void);
33 extern int omp_in_parallel(void);
34 extern void omp_set_dynamic(int dynamic_threads);
35 extern int omp_get_dynamic(void);
36 extern void omp_set_nested(int nested);
37 extern int omp_get_nested(void);
38 extern int omp_get_thread_limit(void);
39 extern void omp_set_max_active_levels(int max_active_levels);
40 extern int omp_get_max_active_levels(void);
41 extern int omp_get_level(void);
42 extern int omp_get_ancestor_thread_num(int level);

43 302 OpenMP API • Version 3.0 May 2008

1 extern int omp_get_team_size(int level);
2 extern int omp_get_active_level(void);
3 extern void omp_set_schedule(omp_sched_t kind, int modifier);
4 extern void omp_get_schedule(omp_sched_t *kind, int *modifier);

5 extern void omp_init_lock(omp_lock_t *lock);

6 extern void omp_destroy_lock(omp_lock_t *lock);
7 extern void omp_set_lock(omp_lock_t *lock);
8 extern void omp_unset_lock(omp_lock_t *lock);
9 extern int omp_test_lock(omp_lock_t *lock);

10 extern void omp_init_nest_lock(omp_nest_lock_t *lock);

11 extern void omp_destroy_nest_lock(omp_nest_lock_t *lock);
12 extern void omp_set_nest_lock(omp_nest_lock_t *lock);
13 extern void omp_unset_nest_lock(omp_nest_lock_t *lock);
14 extern int omp_test_nest_lock(omp_nest_lock_t *lock);

15 extern double omp_get_wtime(void);

16 extern double omp_get_wtick(void);

17 #ifdef __cplusplus
18 }
19 #endif

20 #endif

21 Appendix D Interface Declarations 303

1

2 D.2 Example of an Interface Declaration include

3 File
4 omp_lib_kinds.h:
5 integer omp_lock_kind
6 parameter ( omp_lock_kind = 8 )
7 integer omp_nest_lock_kind
8 parameter ( omp_nest_lock_kind = 8 )

9 integer omp_sched_kind
10 parameter ( omp_sched_kind = 4)

11 integer ( omp_sched_kind ) omp_sched_static

12 parameter ( omp_sched_static = 1 )
13 integer ( omp_sched_kind ) omp_sched_dynamic
14 parameter ( omp_sched_dynamic = 2 )
15 integer ( omp_sched_kind ) omp_sched_guided
16 parameter ( omp_sched_guided = 3 )
17 integer ( omp_sched_kind ) omp_sched_auto
18 parameter ( omp_sched_auto = 4 )

19 omp_lib.h:
20 C default integer type assumed below
21 C default logical type assumed below
22 C OpenMP Fortran API v3.0

23 include 'omp_lib_kinds.h'
24 integer openmp_version
25 parameter ( openmp_version = 200805 )

26 external omp_set_num_threads
27 external omp_get_num_threads
28 integer omp_get_num_threads
29 external omp_get_max_threads
30 integer omp_get_max_threads
31 external omp_get_thread_num
32 integer omp_get_thread_num
33 external omp_get_num_procs
34 integer omp_get_num_procs
35 external omp_in_parallel
36 logical omp_in_parallel
37 external omp_set_dynamic
38 external omp_get_dynamic
39 logical omp_get_dynamic
40 external omp_set_nested
41 external omp_get_nested
42 logical omp_get_nested
43 external omp_set_schedule
44 external omp_get_schedule
45 external omp_get_thread_limit

46 304 OpenMP API • Version 3.0 May 2008

1 integer omp_get_thread_limit
2 external omp_set_max_active_levels
3 external omp_get_max_active_levels
4 integer omp_get_max_active_levels
5 external omp_get_level
6 integer omp_get_level
7 external omp_get_ancestor_thread_num
8 integer omp_get_ancestor_thread_num
9 external omp_get_team_size
10 integer omp_get_team_size
11 external omp_get_active_level
12 integer omp_get_active_level

13 external omp_init_lock
14 external omp_destroy_lock
15 external omp_set_lock
16 external omp_unset_lock
17 external omp_test_lock
18 logical omp_test_lock

19 external omp_init_nest_lock
20 external omp_destroy_nest_lock
21 external omp_set_nest_lock
22 external omp_unset_nest_lock
23 external omp_test_nest_lock
24 integer omp_test_nest_lock

25 external omp_get_wtick
26 double precision omp_get_wtick
27 external omp_get_wtime
28 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 ! the "!" of this comment starts in column 1
5 !23456

6 module omp_lib_kinds
7 integer, parameter :: omp_integer_kind = 4
8 integer, parameter :: omp_logical_kind = 4
9 integer, parameter :: omp_lock_kind = 8
10 integer, parameter :: omp_nest_lock_kind = 8
11 integer, parameter :: omp_sched_kind = 4
12 integer(kind=omp_sched_kind), parameter ::
13 & omp_sched_static = 1
14 integer(kind=omp_sched_kind), parameter ::
15 & omp_sched_dynamic = 2
16 integer(kind=omp_sched_kind), parameter ::
17 & omp_sched_guided = 3
18 integer(kind=omp_sched_kind), parameter ::
19 & omp_sched_auto = 4
20 end module omp_lib_kinds

21 module omp_lib

22 use omp_lib_kinds
23 ! OpenMP Fortran API v3.0
24 integer, parameter :: openmp_version = 200805

25 interface

26 subroutine omp_set_num_threads (number_of_threads_expr)

27 use omp_lib_kinds
28 integer (kind=omp_integer_kind), intent(in) ::
29 & number_of_threads_expr
30 end subroutine omp_set_num_threads

31 function omp_get_num_threads ()
32 use omp_lib_kinds
33 integer (kind=omp_integer_kind) :: omp_get_num_threads
34 end function omp_get_num_threads

35 function omp_get_max_threads ()
36 use omp_lib_kinds
37 integer (kind=omp_integer_kind) :: omp_get_max_threads
38 end function omp_get_max_threads

39 function omp_get_thread_num ()
40 use omp_lib_kinds
41 integer (kind=omp_integer_kind) :: omp_get_thread_num
42 end function omp_get_thread_num

43 306 OpenMP API • Version 3.0 May 2008

1 function omp_get_num_procs ()
2 use omp_lib_kinds
3 integer (kind=omp_integer_kind) :: omp_get_num_procs
4 end function omp_get_num_procs

5 function omp_in_parallel ()
6 use omp_lib_kinds
7 logical (kind=omp_logical_kind) :: omp_in_parallel
8 end function omp_in_parallel

9 subroutine omp_set_dynamic (enable_expr)

10 use omp_lib_kinds
11 logical (kind=omp_logical_kind), intent(in) ::
12 & enable_expr
13 end subroutine omp_set_dynamic

14 function omp_get_dynamic ()
15 use omp_lib_kinds
16 logical (kind=omp_logical_kind) :: omp_get_dynamic
17 end function omp_get_dynamic

18 subroutine omp_set_nested (enable_expr)

19 use omp_lib_kinds
20 logical (kind=omp_logical_kind), intent(in) ::
21 & enable_expr
22 end subroutine omp_set_nested

23 function omp_get_nested ()
24 use omp_lib_kinds
25 logical (kind=omp_logical_kind) :: omp_get_nested
26 end function omp_get_nested

27 subroutine omp_set_schedule (kind, modifier)

28 use omp_lib_kinds
29 integer(kind=omp_sched_kind), intent(in) :: kind
30 integer(kind=omp_integer_kind), intent(in) :: modifier
31 end subroutine omp_set_schedule

32 subroutine omp_get_schedule (kind, modifier)

33 use omp_lib_kinds
34 integer(kind=omp_sched_kind), intent(out) :: kind
35 integer(kind=omp_integer_kind), intent(out)::modifier
36 end subroutine omp_get_schedule

37 function omp_get_thread_limit()
38 use omp_lib_kinds
39 integer (kind=omp_integer_kind) :: omp_get_thread_limit
40 end function omp_get_thread_limit

41 subroutine omp_set_max_active_levels(var)
42 use omp_lib_kinds
43 integer (kind=omp_integer_kind), intent(in) :: var
44 end subroutine omp_set_max_active_levels

45 Appendix D Interface Declarations 307

1 function omp_get_max_active_levels()
2 use omp_lib_kinds
3 integer (kind=omp_integer_kind) ::
4 & omp_get_max_active_levels
5 end function omp_get_max_active_levels

6 function omp_get_level()
7 use omp_lib_kinds
8 integer (kind=omp_integer_kind) :: omp_get_level
9 end function omp_get_level

10 function omp_get_ancestor_thread_num(level)
11 use omp_lib_kinds
12 integer (kind=omp_integer_kind), intent(in) ::
13 & level
14 integer (kind=omp_integer_kind) ::
15 & omp_get_ancestor_thread_num
16 end function omp_get_ancestor_thread_num

17 function omp_get_team_size(level)
18 use omp_lib_kinds
19 integer (kind=omp_integer_kind), intent(in) ::
20 & level
21 integer (kind=omp_integer_kind) :: omp_get_team_size
22 end function omp_get_team_size

23 function omp_get_active_level()
24 use omp_lib_kinds
25 integer (kind=omp_integer_kind) ::
26 & omp_get_active_level
27 end function omp_get_active_level

28 subroutine omp_init_lock (var)

29 use omp_lib_kinds
30 integer (kind=omp_lock_kind), intent(out) :: var
31 end subroutine omp_init_lock

32 subroutine omp_destroy_lock (var)

33 use omp_lib_kinds
34 integer (kind=omp_lock_kind), intent(inout) :: var
35 end subroutine omp_destroy_lock

36 subroutine omp_set_lock (var)

37 use omp_lib_kinds
38 integer (kind=omp_lock_kind), intent(inout) :: var
39 end subroutine omp_set_lock

40 subroutine omp_unset_lock (var)

41 use omp_lib_kinds
42 integer (kind=omp_lock_kind), intent(inout) :: var
43 end subroutine omp_unset_lock

44 308 OpenMP API • Version 3.0 May 2008

1 function omp_test_lock (var)
2 use omp_lib_kinds
3 logical (kind=omp_logical_kind) :: omp_test_lock
4 integer (kind=omp_lock_kind), intent(inout) :: var
5 end function omp_test_lock

6 subroutine omp_init_nest_lock (var)

7 use omp_lib_kinds
8 integer (kind=omp_nest_lock_kind), intent(out) :: var
9 end subroutine omp_init_nest_lock

10 subroutine omp_destroy_nest_lock (var)

11 use omp_lib_kinds
12 integer (kind=omp_nest_lock_kind), intent(inout) :: var
13 end subroutine omp_destroy_nest_lock

14 subroutine omp_set_nest_lock (var)

15 use omp_lib_kinds
16 integer (kind=omp_nest_lock_kind), intent(inout) :: var
17 end subroutine omp_set_nest_lock

18 subroutine omp_unset_nest_lock (var)

19 use omp_lib_kinds
20 integer (kind=omp_nest_lock_kind), intent(inout) :: var
21 end subroutine omp_unset_nest_lock

22 function omp_test_nest_lock (var)

23 use omp_lib_kinds
24 integer (kind=omp_integer_kind) :: omp_test_nest_lock
25 integer (kind=omp_nest_lock_kind), intent(inout) ::
26 & var
27 end function omp_test_nest_lock

28 function omp_get_wtick ()
29 double precision :: omp_get_wtick
30 end function omp_get_wtick

31 function omp_get_wtime ()
32 double precision :: omp_get_wtime
33 end function omp_get_wtime

34 end interface

35 end module omp_lib

36 Appendix D Interface Declarations 309

1

2 D.4 Example of a Generic Interface for a Library

3 Routine
4 Any of the OMP runtime library routines that take an argument may be extended with a
5 generic interface so arguments of different KIND type can be accommodated.
6 Assume an implementation supports both default INTEGER as KIND =
7 OMP_INTEGER_KIND and another INTEGER KIND, KIND = SHORT_INT. Then
8 OMP_SET_NUM_THREADS could be specified in the omp_lib module as the
9 following:
10
11 ! the "!" of this comment starts in column 1
12 interface omp_set_num_threads
13
14 subroutine omp_set_num_threads_1 ( number_of_threads_expr )
15 use omp_lib_kinds
16 integer ( kind=omp_integer_kind ), intent(in) :: &
17 & number_of_threads_expr
18 end subroutine omp_set_num_threads_1
19
20 subroutine omp_set_num_threads_2 ( number_of_threads_expr )
21 use omp_lib_kinds
22 integer ( kind=short_int ), intent(in) :: &
23 & number_of_threads_expr
24 end subroutine omp_set_num_threads_2
25
26 end interface omp_set_num_threads
27

28 310 OpenMP API • Version 3.0 May 2008

1
2 APPENDIX E

3 Implementation Defined
4 Behaviors in OpenMP
5

6 This appendix summarizes the behaviors that are described as implementation defined in
7 this API. Each behavior is cross-referenced back to its description in the main
8 specification. An implementation is required to define and document its behavior in
9 these cases.
10 • Task scheduling points: it is implementation defined where task scheduling points
11 occur in untied task regions (see Section 1.3 on page 11).
12 • Memory model: it is implementation defined as to whether, and in what sizes,
13 memory accesses by multiple threads to the same variable without synchronization
14 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 def-sched-var, stacksize-var, wait-policy-var, thread-limit-var, and max-active-levels-
17 var 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 dynamically adjust the number of threads is provided. Implementations are allowed
20 to deliver fewer threads (but at least one) than indicated in Algorithm 2-1 even if
21 dynamic adjustment is disabled (see Section 2.4.1 on page 35).
22 • Loop directive: the integer type or kind used to compute the iteration count of a
23 collapsed loop is implementation defined. The effect of the schedule(runtime)
24 clause when the run-sched-var ICV is set to auto is implementation defined. See
25 Section 2.5.1 on page 38.
26 • sections construct: the method of scheduling the structured blocks among threads
27 in the team is implementation defined (see Section 2.5.2 on page 47).
28 • single construct: the method of choosing a thread to execute the structured block
29 is implementation defined (see Section 2.5.3 on page 49).

30 311
1 • atomic construct: a compliant implementation may enforce exclusive access
2 between atomic regions which update different storage locations. The
3 circumstances under which this occurs are implementation defined (see Section 2.8.5
4 on page 69).
5 • omp_set_num_threads routine: if the argument is not a positive integer the
6 behavior is implementation defined (see Section 3.2.1 on page 110).
7 • omp_set_schedule routine: the behavior for implementation defined schedule
8 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 parallel region the binding thread set (and binding region, if required) for the
11 omp_set_max_active_levels region is implementation defined and the
12 behavior is implementation defined. If the argument is not a non-negative integer
13 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
15 parallel region the binding thread set (and binding region, if required) for the
16 omp_get_max_active_levels region is implementation defined (see
17 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
20 Section 4.1 on page 146).
21 • OMP_NUM_THREADS environment variable: if the value of the variable is greater
22 than the number of threads the implementation can support or is not a positive integer
23 then the result is implementation defined (see Section 4.2 on page 147).
24 • OMP_DYNAMIC environment variable: if the value is neither true nor false the
25 behavior is implementation defined (see Section 4.3 on page 148).
26 • OMP_NESTED environment variable: if the value is neither true nor false the
27 behavior is implementation defined (see Section 4.4 on page 148).
28 • OMP_STACKSIZE environment variable: if the value does not conform to the
29 specified format or the implementation cannot provide a stack of the specified size
30 then the behavior is implementation defined (see Section 4.5 on page 149).
31 • OMP_WAIT_POLICY environment variable: the details of the ACTIVE and
32 PASSIVE behaviors are implementation defined (see Section 4.6 on page 150).
33 • OMP_MAX_ACTIVE_LEVELS environment variable: if the value is not a non-
34 negative integer or is greater than the number of parallel levels an impementation can
35 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
38 integer, the behavior of the program is implementation defined (see Section 4.8 on
39 page 151).

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 objects of threads (other than the initial thread) to persist between two consecutive
4 active parallel regions do not all hold, the allocation status of an allocatable array in
5 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
8 procedure reference, and back out of the temporary storage into the actual argument
9 storage after the procedure reference. Situations where this occurs other than those
10 specified are implementation defined (see Section 2.9.3.2 on page 88).
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
13 defined whether any of the OpenMP runtime library routines that take an argument
14 are extended with a generic interface so arguments of different KIND type can be
15 accommodated (see Section 3.1 on page 108).
16
Fortran

17 Appendix E Implementation Defined Behaviors in OpenMP 313

1 314 OpenMP API • Version 3.0 May 2008
1
2 APPENDIX F

3 Changes from Version 2.5 to

4 Version 3.0
5

6 This appendix summarizes the major changes between the OpenMP API Version 2.5
7 specification and the OpenMP API Version 3.0 specification.
8 • The concept of tasks has been added to the OpenMP execution model (see
9 Section 1.2.3 on page 8 and Section 1.3 on page 11).
10 • The task construct (see Section 2.7 on page 59) has been added, which provides a
11 mechanism for creating tasks explicitly.
12 • The taskwait construct (see Section 2.8.4 on page 68) has been added, which
13 causes a task to wait for all its child tasks to complete.
14 • The OpenMP memory model now covers atomicity of memory accesses (see
15 Section 1.4.1 on page 13). The description of the behavior of volatile in terms of
16 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
19 3.0, there is one copy of these ICVs per task (see Section 2.3 on page 28). As a result,
20 the omp_set_num_threads, omp_set_nested and omp_set_dynamic
21 runtime library routines now have specified effects when called from inside a
22 parallel region (See Section 3.2.1 on page 110, Section 3.2.7 on page 117 and
23 Section 3.2.9 on page 119).
24 • The definition of active parallel region has been changed: in Version 3.0 a
25 parallel region is active if it is executed by a team consisting of more than one
26 thread (see Section 1.2.2 on page 2).
27 • The rules for determining the number of threads used in a parallel region have
28 been modified (see Section 2.4.1 on page 35).
29 • In Version 3.0, the assignment of iterations to threads in a loop construct with a
30 static schedule kind is deterministic (see Section 2.5.1 on page 38).

31 315
1 • In Version 3.0, a loop construct may be associated with more than one perfectly
2 nested loop. The number of associated loops may be controlled by the collapse
3 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 • The schedule kind auto has been added, which gives the implementation the
7 freedom to choose any possible mapping of iterations in a loop construct to threads in
8 the team (see Section 2.5.1 on page 38).
9 • Fortran assumed-size arrays now have predetermined data-sharing attributes (see
10 Section 2.9.1.1 on page 78).
11 • In Fortran, firstprivate is now permitted as an argument to the default
12 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
15 no attempt is made to reference the original list item inside the parallel region, its
16 value is well defined on exit from the parallel region (see Section 2.9.3.3 on page
17 89).
18 • In Version 3.0, Fortran allocatable arrays may appear in private,
19 firstprivate, lastprivate, reduction, copyin and copyprivate
20 clauses. (see Section 2.9.2 on page 81, Section 2.9.3.3 on page 89, Section 2.9.3.4 on
21 page 92, Section 2.9.3.5 on page 94, Section 2.9.3.6 on page 96, Section 2.9.4.1 on
22 page 101 and Section 2.9.4.2 on page 102).
23 • In Version 3.0, static class members variables may appear in a threadprivate
24 directive (see Section 2.9.2 on page 81).
25 • Version 3.0 makes clear where, and with which arguments, constructors and
26 destructors of private and threadprivate class type variables are called (see
27 Section 2.9.2 on page 81, Section 2.9.3.3 on page 89, Section 2.9.3.4 on page 92,
28 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
30 have been added; these routines respectively set and retrieve the value of the
31 run_sched_var ICV (see Section 3.2.11 on page 121 and Section 3.2.12 on page 123).
32 • The thread-limit-var ICV has been added, which controls the maximium number of
33 threads participating in the OpenMP program. The value of this ICV can be set with
34 the OMP_THREAD_LIMIT environment variable and retrieved with the
35 omp_get_thread_limit runtime library routine (see Section 2.3.1 on page 28,
36 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
39 OMP_MAX_ACTIVE_LEVELS environment variable and the
40 omp_set_max_active_levels runtime library routine, and it can be retrieved

41 316 OpenMP API • Version 3.0 May 2008

1 with the omp_get_max_active_levels runtime library routine (see
2 Section 2.3.1 on page 28, Section 3.2.14 on page 126, Section 3.2.15 on page 127 and
3 Section 4.7 on page 150).
4 • The stacksize-var ICV has been added, which controls the stack size for threads that
5 the OpenMP implementation creates. The value of this ICV can be set with the
6 OMP_STACKSIZE environment variable (see Section 2.3.1 on page 28 and
7 Section 4.5 on page 149).
8 • The wait-policy-var ICV has been added, which controls the desired behavior of
9 waiting threads. The value of this ICV can be set with the OMP_WAIT_POLICY
10 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
12 number of nested parallel regions enclosing the task that contains the call (see
13 Section 3.2.16 on page 129).
14 • The omp_get_ancestor_thread_num runtime library routine has been added,
15 which returns, for a given nested level of the current thread, the thread number of the
16 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
19 ancestor belongs (see Section 3.2.18 on page 131).
20 • The omp_get_active_level runtime library routine has been added, which
21 returns the number of nested, active parallel regions enclosing the task that
22 contains the call (see Section 3.2.19 on page 133).
23 • In Version 3.0, locks are owned by tasks, not by threads (see Section 3.3 on page
24 134).

25 Appendix F Changes from Version 2.5 to Version 3.0 317

1 318 OpenMP API • Version 3.0 May 2008